Commit | Line | Data |
---|---|---|
457c8996 | 1 | // SPDX-License-Identifier: GPL-2.0-only |
1da177e4 LT |
2 | /* |
3 | * Generic hugetlb support. | |
6d49e352 | 4 | * (C) Nadia Yvette Chambers, April 2004 |
1da177e4 | 5 | */ |
1da177e4 LT |
6 | #include <linux/list.h> |
7 | #include <linux/init.h> | |
1da177e4 | 8 | #include <linux/mm.h> |
e1759c21 | 9 | #include <linux/seq_file.h> |
1da177e4 LT |
10 | #include <linux/sysctl.h> |
11 | #include <linux/highmem.h> | |
cddb8a5c | 12 | #include <linux/mmu_notifier.h> |
1da177e4 | 13 | #include <linux/nodemask.h> |
63551ae0 | 14 | #include <linux/pagemap.h> |
5da7ca86 | 15 | #include <linux/mempolicy.h> |
3b32123d | 16 | #include <linux/compiler.h> |
aea47ff3 | 17 | #include <linux/cpuset.h> |
3935baa9 | 18 | #include <linux/mutex.h> |
97ad1087 | 19 | #include <linux/memblock.h> |
a3437870 | 20 | #include <linux/sysfs.h> |
5a0e3ad6 | 21 | #include <linux/slab.h> |
bbe88753 | 22 | #include <linux/sched/mm.h> |
63489f8e | 23 | #include <linux/mmdebug.h> |
174cd4b1 | 24 | #include <linux/sched/signal.h> |
0fe6e20b | 25 | #include <linux/rmap.h> |
c6247f72 | 26 | #include <linux/string_helpers.h> |
fd6a03ed NH |
27 | #include <linux/swap.h> |
28 | #include <linux/swapops.h> | |
8382d914 | 29 | #include <linux/jhash.h> |
98fa15f3 | 30 | #include <linux/numa.h> |
c77c0a8a | 31 | #include <linux/llist.h> |
cf11e85f | 32 | #include <linux/cma.h> |
8cc5fcbb | 33 | #include <linux/migrate.h> |
f9317f77 | 34 | #include <linux/nospec.h> |
662ce1dc | 35 | #include <linux/delayacct.h> |
b958d4d0 | 36 | #include <linux/memory.h> |
d6606683 | 37 | |
63551ae0 | 38 | #include <asm/page.h> |
ca15ca40 | 39 | #include <asm/pgalloc.h> |
24669e58 | 40 | #include <asm/tlb.h> |
63551ae0 | 41 | |
24669e58 | 42 | #include <linux/io.h> |
63551ae0 | 43 | #include <linux/hugetlb.h> |
9dd540e2 | 44 | #include <linux/hugetlb_cgroup.h> |
9a305230 | 45 | #include <linux/node.h> |
ab5ac90a | 46 | #include <linux/page_owner.h> |
7835e98b | 47 | #include "internal.h" |
f41f2ed4 | 48 | #include "hugetlb_vmemmap.h" |
1da177e4 | 49 | |
c3f38a38 | 50 | int hugetlb_max_hstate __read_mostly; |
e5ff2159 AK |
51 | unsigned int default_hstate_idx; |
52 | struct hstate hstates[HUGE_MAX_HSTATE]; | |
cf11e85f | 53 | |
dbda8fea | 54 | #ifdef CONFIG_CMA |
cf11e85f | 55 | static struct cma *hugetlb_cma[MAX_NUMNODES]; |
38e719ab | 56 | static unsigned long hugetlb_cma_size_in_node[MAX_NUMNODES] __initdata; |
2f6c57d6 | 57 | static bool hugetlb_cma_folio(struct folio *folio, unsigned int order) |
a01f4390 | 58 | { |
2f6c57d6 | 59 | return cma_pages_valid(hugetlb_cma[folio_nid(folio)], &folio->page, |
a01f4390 MK |
60 | 1 << order); |
61 | } | |
62 | #else | |
2f6c57d6 | 63 | static bool hugetlb_cma_folio(struct folio *folio, unsigned int order) |
a01f4390 MK |
64 | { |
65 | return false; | |
66 | } | |
dbda8fea BS |
67 | #endif |
68 | static unsigned long hugetlb_cma_size __initdata; | |
cf11e85f | 69 | |
53ba51d2 JT |
70 | __initdata LIST_HEAD(huge_boot_pages); |
71 | ||
e5ff2159 AK |
72 | /* for command line parsing */ |
73 | static struct hstate * __initdata parsed_hstate; | |
74 | static unsigned long __initdata default_hstate_max_huge_pages; | |
9fee021d | 75 | static bool __initdata parsed_valid_hugepagesz = true; |
282f4214 | 76 | static bool __initdata parsed_default_hugepagesz; |
b5389086 | 77 | static unsigned int default_hugepages_in_node[MAX_NUMNODES] __initdata; |
e5ff2159 | 78 | |
3935baa9 | 79 | /* |
31caf665 NH |
80 | * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages, |
81 | * free_huge_pages, and surplus_huge_pages. | |
3935baa9 | 82 | */ |
c3f38a38 | 83 | DEFINE_SPINLOCK(hugetlb_lock); |
0bd0f9fb | 84 | |
8382d914 DB |
85 | /* |
86 | * Serializes faults on the same logical page. This is used to | |
87 | * prevent spurious OOMs when the hugepage pool is fully utilized. | |
88 | */ | |
89 | static int num_fault_mutexes; | |
c672c7f2 | 90 | struct mutex *hugetlb_fault_mutex_table ____cacheline_aligned_in_smp; |
8382d914 | 91 | |
7ca02d0a MK |
92 | /* Forward declaration */ |
93 | static int hugetlb_acct_memory(struct hstate *h, long delta); | |
8d9bfb26 MK |
94 | static void hugetlb_vma_lock_free(struct vm_area_struct *vma); |
95 | static void hugetlb_vma_lock_alloc(struct vm_area_struct *vma); | |
ecfbd733 | 96 | static void __hugetlb_vma_unlock_write_free(struct vm_area_struct *vma); |
b30c14cd JH |
97 | static void hugetlb_unshare_pmds(struct vm_area_struct *vma, |
98 | unsigned long start, unsigned long end); | |
7ca02d0a | 99 | |
1d88433b | 100 | static inline bool subpool_is_free(struct hugepage_subpool *spool) |
90481622 | 101 | { |
1d88433b ML |
102 | if (spool->count) |
103 | return false; | |
104 | if (spool->max_hpages != -1) | |
105 | return spool->used_hpages == 0; | |
106 | if (spool->min_hpages != -1) | |
107 | return spool->rsv_hpages == spool->min_hpages; | |
108 | ||
109 | return true; | |
110 | } | |
90481622 | 111 | |
db71ef79 MK |
112 | static inline void unlock_or_release_subpool(struct hugepage_subpool *spool, |
113 | unsigned long irq_flags) | |
1d88433b | 114 | { |
db71ef79 | 115 | spin_unlock_irqrestore(&spool->lock, irq_flags); |
90481622 DG |
116 | |
117 | /* If no pages are used, and no other handles to the subpool | |
7c8de358 | 118 | * remain, give up any reservations based on minimum size and |
7ca02d0a | 119 | * free the subpool */ |
1d88433b | 120 | if (subpool_is_free(spool)) { |
7ca02d0a MK |
121 | if (spool->min_hpages != -1) |
122 | hugetlb_acct_memory(spool->hstate, | |
123 | -spool->min_hpages); | |
90481622 | 124 | kfree(spool); |
7ca02d0a | 125 | } |
90481622 DG |
126 | } |
127 | ||
7ca02d0a MK |
128 | struct hugepage_subpool *hugepage_new_subpool(struct hstate *h, long max_hpages, |
129 | long min_hpages) | |
90481622 DG |
130 | { |
131 | struct hugepage_subpool *spool; | |
132 | ||
c6a91820 | 133 | spool = kzalloc(sizeof(*spool), GFP_KERNEL); |
90481622 DG |
134 | if (!spool) |
135 | return NULL; | |
136 | ||
137 | spin_lock_init(&spool->lock); | |
138 | spool->count = 1; | |
7ca02d0a MK |
139 | spool->max_hpages = max_hpages; |
140 | spool->hstate = h; | |
141 | spool->min_hpages = min_hpages; | |
142 | ||
143 | if (min_hpages != -1 && hugetlb_acct_memory(h, min_hpages)) { | |
144 | kfree(spool); | |
145 | return NULL; | |
146 | } | |
147 | spool->rsv_hpages = min_hpages; | |
90481622 DG |
148 | |
149 | return spool; | |
150 | } | |
151 | ||
152 | void hugepage_put_subpool(struct hugepage_subpool *spool) | |
153 | { | |
db71ef79 MK |
154 | unsigned long flags; |
155 | ||
156 | spin_lock_irqsave(&spool->lock, flags); | |
90481622 DG |
157 | BUG_ON(!spool->count); |
158 | spool->count--; | |
db71ef79 | 159 | unlock_or_release_subpool(spool, flags); |
90481622 DG |
160 | } |
161 | ||
1c5ecae3 MK |
162 | /* |
163 | * Subpool accounting for allocating and reserving pages. | |
164 | * Return -ENOMEM if there are not enough resources to satisfy the | |
9e7ee400 | 165 | * request. Otherwise, return the number of pages by which the |
1c5ecae3 MK |
166 | * global pools must be adjusted (upward). The returned value may |
167 | * only be different than the passed value (delta) in the case where | |
7c8de358 | 168 | * a subpool minimum size must be maintained. |
1c5ecae3 MK |
169 | */ |
170 | static long hugepage_subpool_get_pages(struct hugepage_subpool *spool, | |
90481622 DG |
171 | long delta) |
172 | { | |
1c5ecae3 | 173 | long ret = delta; |
90481622 DG |
174 | |
175 | if (!spool) | |
1c5ecae3 | 176 | return ret; |
90481622 | 177 | |
db71ef79 | 178 | spin_lock_irq(&spool->lock); |
1c5ecae3 MK |
179 | |
180 | if (spool->max_hpages != -1) { /* maximum size accounting */ | |
181 | if ((spool->used_hpages + delta) <= spool->max_hpages) | |
182 | spool->used_hpages += delta; | |
183 | else { | |
184 | ret = -ENOMEM; | |
185 | goto unlock_ret; | |
186 | } | |
90481622 | 187 | } |
90481622 | 188 | |
09a95e29 MK |
189 | /* minimum size accounting */ |
190 | if (spool->min_hpages != -1 && spool->rsv_hpages) { | |
1c5ecae3 MK |
191 | if (delta > spool->rsv_hpages) { |
192 | /* | |
193 | * Asking for more reserves than those already taken on | |
194 | * behalf of subpool. Return difference. | |
195 | */ | |
196 | ret = delta - spool->rsv_hpages; | |
197 | spool->rsv_hpages = 0; | |
198 | } else { | |
199 | ret = 0; /* reserves already accounted for */ | |
200 | spool->rsv_hpages -= delta; | |
201 | } | |
202 | } | |
203 | ||
204 | unlock_ret: | |
db71ef79 | 205 | spin_unlock_irq(&spool->lock); |
90481622 DG |
206 | return ret; |
207 | } | |
208 | ||
1c5ecae3 MK |
209 | /* |
210 | * Subpool accounting for freeing and unreserving pages. | |
211 | * Return the number of global page reservations that must be dropped. | |
212 | * The return value may only be different than the passed value (delta) | |
213 | * in the case where a subpool minimum size must be maintained. | |
214 | */ | |
215 | static long hugepage_subpool_put_pages(struct hugepage_subpool *spool, | |
90481622 DG |
216 | long delta) |
217 | { | |
1c5ecae3 | 218 | long ret = delta; |
db71ef79 | 219 | unsigned long flags; |
1c5ecae3 | 220 | |
90481622 | 221 | if (!spool) |
1c5ecae3 | 222 | return delta; |
90481622 | 223 | |
db71ef79 | 224 | spin_lock_irqsave(&spool->lock, flags); |
1c5ecae3 MK |
225 | |
226 | if (spool->max_hpages != -1) /* maximum size accounting */ | |
227 | spool->used_hpages -= delta; | |
228 | ||
09a95e29 MK |
229 | /* minimum size accounting */ |
230 | if (spool->min_hpages != -1 && spool->used_hpages < spool->min_hpages) { | |
1c5ecae3 MK |
231 | if (spool->rsv_hpages + delta <= spool->min_hpages) |
232 | ret = 0; | |
233 | else | |
234 | ret = spool->rsv_hpages + delta - spool->min_hpages; | |
235 | ||
236 | spool->rsv_hpages += delta; | |
237 | if (spool->rsv_hpages > spool->min_hpages) | |
238 | spool->rsv_hpages = spool->min_hpages; | |
239 | } | |
240 | ||
241 | /* | |
242 | * If hugetlbfs_put_super couldn't free spool due to an outstanding | |
243 | * quota reference, free it now. | |
244 | */ | |
db71ef79 | 245 | unlock_or_release_subpool(spool, flags); |
1c5ecae3 MK |
246 | |
247 | return ret; | |
90481622 DG |
248 | } |
249 | ||
250 | static inline struct hugepage_subpool *subpool_inode(struct inode *inode) | |
251 | { | |
252 | return HUGETLBFS_SB(inode->i_sb)->spool; | |
253 | } | |
254 | ||
255 | static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma) | |
256 | { | |
496ad9aa | 257 | return subpool_inode(file_inode(vma->vm_file)); |
90481622 DG |
258 | } |
259 | ||
e700898f MK |
260 | /* |
261 | * hugetlb vma_lock helper routines | |
262 | */ | |
263 | static bool __vma_shareable_lock(struct vm_area_struct *vma) | |
264 | { | |
265 | return vma->vm_flags & (VM_MAYSHARE | VM_SHARED) && | |
266 | vma->vm_private_data; | |
267 | } | |
268 | ||
269 | void hugetlb_vma_lock_read(struct vm_area_struct *vma) | |
270 | { | |
271 | if (__vma_shareable_lock(vma)) { | |
272 | struct hugetlb_vma_lock *vma_lock = vma->vm_private_data; | |
273 | ||
274 | down_read(&vma_lock->rw_sema); | |
275 | } | |
276 | } | |
277 | ||
278 | void hugetlb_vma_unlock_read(struct vm_area_struct *vma) | |
279 | { | |
280 | if (__vma_shareable_lock(vma)) { | |
281 | struct hugetlb_vma_lock *vma_lock = vma->vm_private_data; | |
282 | ||
283 | up_read(&vma_lock->rw_sema); | |
284 | } | |
285 | } | |
286 | ||
287 | void hugetlb_vma_lock_write(struct vm_area_struct *vma) | |
288 | { | |
289 | if (__vma_shareable_lock(vma)) { | |
290 | struct hugetlb_vma_lock *vma_lock = vma->vm_private_data; | |
291 | ||
292 | down_write(&vma_lock->rw_sema); | |
293 | } | |
294 | } | |
295 | ||
296 | void hugetlb_vma_unlock_write(struct vm_area_struct *vma) | |
297 | { | |
298 | if (__vma_shareable_lock(vma)) { | |
299 | struct hugetlb_vma_lock *vma_lock = vma->vm_private_data; | |
300 | ||
301 | up_write(&vma_lock->rw_sema); | |
302 | } | |
303 | } | |
304 | ||
305 | int hugetlb_vma_trylock_write(struct vm_area_struct *vma) | |
306 | { | |
307 | struct hugetlb_vma_lock *vma_lock = vma->vm_private_data; | |
308 | ||
309 | if (!__vma_shareable_lock(vma)) | |
310 | return 1; | |
311 | ||
312 | return down_write_trylock(&vma_lock->rw_sema); | |
313 | } | |
314 | ||
315 | void hugetlb_vma_assert_locked(struct vm_area_struct *vma) | |
316 | { | |
317 | if (__vma_shareable_lock(vma)) { | |
318 | struct hugetlb_vma_lock *vma_lock = vma->vm_private_data; | |
319 | ||
320 | lockdep_assert_held(&vma_lock->rw_sema); | |
321 | } | |
322 | } | |
323 | ||
324 | void hugetlb_vma_lock_release(struct kref *kref) | |
325 | { | |
326 | struct hugetlb_vma_lock *vma_lock = container_of(kref, | |
327 | struct hugetlb_vma_lock, refs); | |
328 | ||
329 | kfree(vma_lock); | |
330 | } | |
331 | ||
332 | static void __hugetlb_vma_unlock_write_put(struct hugetlb_vma_lock *vma_lock) | |
333 | { | |
334 | struct vm_area_struct *vma = vma_lock->vma; | |
335 | ||
336 | /* | |
337 | * vma_lock structure may or not be released as a result of put, | |
338 | * it certainly will no longer be attached to vma so clear pointer. | |
339 | * Semaphore synchronizes access to vma_lock->vma field. | |
340 | */ | |
341 | vma_lock->vma = NULL; | |
342 | vma->vm_private_data = NULL; | |
343 | up_write(&vma_lock->rw_sema); | |
344 | kref_put(&vma_lock->refs, hugetlb_vma_lock_release); | |
345 | } | |
346 | ||
347 | static void __hugetlb_vma_unlock_write_free(struct vm_area_struct *vma) | |
348 | { | |
349 | if (__vma_shareable_lock(vma)) { | |
350 | struct hugetlb_vma_lock *vma_lock = vma->vm_private_data; | |
351 | ||
352 | __hugetlb_vma_unlock_write_put(vma_lock); | |
353 | } | |
354 | } | |
355 | ||
356 | static void hugetlb_vma_lock_free(struct vm_area_struct *vma) | |
357 | { | |
358 | /* | |
359 | * Only present in sharable vmas. | |
360 | */ | |
361 | if (!vma || !__vma_shareable_lock(vma)) | |
362 | return; | |
363 | ||
364 | if (vma->vm_private_data) { | |
365 | struct hugetlb_vma_lock *vma_lock = vma->vm_private_data; | |
366 | ||
367 | down_write(&vma_lock->rw_sema); | |
368 | __hugetlb_vma_unlock_write_put(vma_lock); | |
369 | } | |
370 | } | |
371 | ||
372 | static void hugetlb_vma_lock_alloc(struct vm_area_struct *vma) | |
373 | { | |
374 | struct hugetlb_vma_lock *vma_lock; | |
375 | ||
376 | /* Only establish in (flags) sharable vmas */ | |
377 | if (!vma || !(vma->vm_flags & VM_MAYSHARE)) | |
378 | return; | |
379 | ||
380 | /* Should never get here with non-NULL vm_private_data */ | |
381 | if (vma->vm_private_data) | |
382 | return; | |
383 | ||
384 | vma_lock = kmalloc(sizeof(*vma_lock), GFP_KERNEL); | |
385 | if (!vma_lock) { | |
386 | /* | |
387 | * If we can not allocate structure, then vma can not | |
388 | * participate in pmd sharing. This is only a possible | |
389 | * performance enhancement and memory saving issue. | |
390 | * However, the lock is also used to synchronize page | |
391 | * faults with truncation. If the lock is not present, | |
392 | * unlikely races could leave pages in a file past i_size | |
393 | * until the file is removed. Warn in the unlikely case of | |
394 | * allocation failure. | |
395 | */ | |
396 | pr_warn_once("HugeTLB: unable to allocate vma specific lock\n"); | |
397 | return; | |
398 | } | |
399 | ||
400 | kref_init(&vma_lock->refs); | |
401 | init_rwsem(&vma_lock->rw_sema); | |
402 | vma_lock->vma = vma; | |
403 | vma->vm_private_data = vma_lock; | |
404 | } | |
405 | ||
0db9d74e MA |
406 | /* Helper that removes a struct file_region from the resv_map cache and returns |
407 | * it for use. | |
408 | */ | |
409 | static struct file_region * | |
410 | get_file_region_entry_from_cache(struct resv_map *resv, long from, long to) | |
411 | { | |
3259914f | 412 | struct file_region *nrg; |
0db9d74e MA |
413 | |
414 | VM_BUG_ON(resv->region_cache_count <= 0); | |
415 | ||
416 | resv->region_cache_count--; | |
417 | nrg = list_first_entry(&resv->region_cache, struct file_region, link); | |
0db9d74e MA |
418 | list_del(&nrg->link); |
419 | ||
420 | nrg->from = from; | |
421 | nrg->to = to; | |
422 | ||
423 | return nrg; | |
424 | } | |
425 | ||
075a61d0 MA |
426 | static void copy_hugetlb_cgroup_uncharge_info(struct file_region *nrg, |
427 | struct file_region *rg) | |
428 | { | |
429 | #ifdef CONFIG_CGROUP_HUGETLB | |
430 | nrg->reservation_counter = rg->reservation_counter; | |
431 | nrg->css = rg->css; | |
432 | if (rg->css) | |
433 | css_get(rg->css); | |
434 | #endif | |
435 | } | |
436 | ||
437 | /* Helper that records hugetlb_cgroup uncharge info. */ | |
438 | static void record_hugetlb_cgroup_uncharge_info(struct hugetlb_cgroup *h_cg, | |
439 | struct hstate *h, | |
440 | struct resv_map *resv, | |
441 | struct file_region *nrg) | |
442 | { | |
443 | #ifdef CONFIG_CGROUP_HUGETLB | |
444 | if (h_cg) { | |
445 | nrg->reservation_counter = | |
446 | &h_cg->rsvd_hugepage[hstate_index(h)]; | |
447 | nrg->css = &h_cg->css; | |
d85aecf2 ML |
448 | /* |
449 | * The caller will hold exactly one h_cg->css reference for the | |
450 | * whole contiguous reservation region. But this area might be | |
451 | * scattered when there are already some file_regions reside in | |
452 | * it. As a result, many file_regions may share only one css | |
453 | * reference. In order to ensure that one file_region must hold | |
454 | * exactly one h_cg->css reference, we should do css_get for | |
455 | * each file_region and leave the reference held by caller | |
456 | * untouched. | |
457 | */ | |
458 | css_get(&h_cg->css); | |
075a61d0 MA |
459 | if (!resv->pages_per_hpage) |
460 | resv->pages_per_hpage = pages_per_huge_page(h); | |
461 | /* pages_per_hpage should be the same for all entries in | |
462 | * a resv_map. | |
463 | */ | |
464 | VM_BUG_ON(resv->pages_per_hpage != pages_per_huge_page(h)); | |
465 | } else { | |
466 | nrg->reservation_counter = NULL; | |
467 | nrg->css = NULL; | |
468 | } | |
469 | #endif | |
470 | } | |
471 | ||
d85aecf2 ML |
472 | static void put_uncharge_info(struct file_region *rg) |
473 | { | |
474 | #ifdef CONFIG_CGROUP_HUGETLB | |
475 | if (rg->css) | |
476 | css_put(rg->css); | |
477 | #endif | |
478 | } | |
479 | ||
a9b3f867 MA |
480 | static bool has_same_uncharge_info(struct file_region *rg, |
481 | struct file_region *org) | |
482 | { | |
483 | #ifdef CONFIG_CGROUP_HUGETLB | |
0739eb43 | 484 | return rg->reservation_counter == org->reservation_counter && |
a9b3f867 MA |
485 | rg->css == org->css; |
486 | ||
487 | #else | |
488 | return true; | |
489 | #endif | |
490 | } | |
491 | ||
492 | static void coalesce_file_region(struct resv_map *resv, struct file_region *rg) | |
493 | { | |
3259914f | 494 | struct file_region *nrg, *prg; |
a9b3f867 MA |
495 | |
496 | prg = list_prev_entry(rg, link); | |
497 | if (&prg->link != &resv->regions && prg->to == rg->from && | |
498 | has_same_uncharge_info(prg, rg)) { | |
499 | prg->to = rg->to; | |
500 | ||
501 | list_del(&rg->link); | |
d85aecf2 | 502 | put_uncharge_info(rg); |
a9b3f867 MA |
503 | kfree(rg); |
504 | ||
7db5e7b6 | 505 | rg = prg; |
a9b3f867 MA |
506 | } |
507 | ||
508 | nrg = list_next_entry(rg, link); | |
509 | if (&nrg->link != &resv->regions && nrg->from == rg->to && | |
510 | has_same_uncharge_info(nrg, rg)) { | |
511 | nrg->from = rg->from; | |
512 | ||
513 | list_del(&rg->link); | |
d85aecf2 | 514 | put_uncharge_info(rg); |
a9b3f867 | 515 | kfree(rg); |
a9b3f867 MA |
516 | } |
517 | } | |
518 | ||
2103cf9c | 519 | static inline long |
84448c8e | 520 | hugetlb_resv_map_add(struct resv_map *map, struct list_head *rg, long from, |
2103cf9c PX |
521 | long to, struct hstate *h, struct hugetlb_cgroup *cg, |
522 | long *regions_needed) | |
523 | { | |
524 | struct file_region *nrg; | |
525 | ||
526 | if (!regions_needed) { | |
527 | nrg = get_file_region_entry_from_cache(map, from, to); | |
528 | record_hugetlb_cgroup_uncharge_info(cg, h, map, nrg); | |
84448c8e | 529 | list_add(&nrg->link, rg); |
2103cf9c PX |
530 | coalesce_file_region(map, nrg); |
531 | } else | |
532 | *regions_needed += 1; | |
533 | ||
534 | return to - from; | |
535 | } | |
536 | ||
972a3da3 WY |
537 | /* |
538 | * Must be called with resv->lock held. | |
539 | * | |
540 | * Calling this with regions_needed != NULL will count the number of pages | |
541 | * to be added but will not modify the linked list. And regions_needed will | |
542 | * indicate the number of file_regions needed in the cache to carry out to add | |
543 | * the regions for this range. | |
d75c6af9 MA |
544 | */ |
545 | static long add_reservation_in_range(struct resv_map *resv, long f, long t, | |
075a61d0 | 546 | struct hugetlb_cgroup *h_cg, |
972a3da3 | 547 | struct hstate *h, long *regions_needed) |
d75c6af9 | 548 | { |
0db9d74e | 549 | long add = 0; |
d75c6af9 | 550 | struct list_head *head = &resv->regions; |
0db9d74e | 551 | long last_accounted_offset = f; |
84448c8e JK |
552 | struct file_region *iter, *trg = NULL; |
553 | struct list_head *rg = NULL; | |
d75c6af9 | 554 | |
0db9d74e MA |
555 | if (regions_needed) |
556 | *regions_needed = 0; | |
d75c6af9 | 557 | |
0db9d74e | 558 | /* In this loop, we essentially handle an entry for the range |
84448c8e | 559 | * [last_accounted_offset, iter->from), at every iteration, with some |
0db9d74e MA |
560 | * bounds checking. |
561 | */ | |
84448c8e | 562 | list_for_each_entry_safe(iter, trg, head, link) { |
0db9d74e | 563 | /* Skip irrelevant regions that start before our range. */ |
84448c8e | 564 | if (iter->from < f) { |
0db9d74e MA |
565 | /* If this region ends after the last accounted offset, |
566 | * then we need to update last_accounted_offset. | |
567 | */ | |
84448c8e JK |
568 | if (iter->to > last_accounted_offset) |
569 | last_accounted_offset = iter->to; | |
0db9d74e MA |
570 | continue; |
571 | } | |
d75c6af9 | 572 | |
0db9d74e MA |
573 | /* When we find a region that starts beyond our range, we've |
574 | * finished. | |
575 | */ | |
84448c8e JK |
576 | if (iter->from >= t) { |
577 | rg = iter->link.prev; | |
d75c6af9 | 578 | break; |
84448c8e | 579 | } |
d75c6af9 | 580 | |
84448c8e | 581 | /* Add an entry for last_accounted_offset -> iter->from, and |
0db9d74e MA |
582 | * update last_accounted_offset. |
583 | */ | |
84448c8e JK |
584 | if (iter->from > last_accounted_offset) |
585 | add += hugetlb_resv_map_add(resv, iter->link.prev, | |
2103cf9c | 586 | last_accounted_offset, |
84448c8e | 587 | iter->from, h, h_cg, |
2103cf9c | 588 | regions_needed); |
0db9d74e | 589 | |
84448c8e | 590 | last_accounted_offset = iter->to; |
0db9d74e MA |
591 | } |
592 | ||
593 | /* Handle the case where our range extends beyond | |
594 | * last_accounted_offset. | |
595 | */ | |
84448c8e JK |
596 | if (!rg) |
597 | rg = head->prev; | |
2103cf9c PX |
598 | if (last_accounted_offset < t) |
599 | add += hugetlb_resv_map_add(resv, rg, last_accounted_offset, | |
600 | t, h, h_cg, regions_needed); | |
0db9d74e | 601 | |
0db9d74e MA |
602 | return add; |
603 | } | |
604 | ||
605 | /* Must be called with resv->lock acquired. Will drop lock to allocate entries. | |
606 | */ | |
607 | static int allocate_file_region_entries(struct resv_map *resv, | |
608 | int regions_needed) | |
609 | __must_hold(&resv->lock) | |
610 | { | |
34665341 | 611 | LIST_HEAD(allocated_regions); |
0db9d74e MA |
612 | int to_allocate = 0, i = 0; |
613 | struct file_region *trg = NULL, *rg = NULL; | |
614 | ||
615 | VM_BUG_ON(regions_needed < 0); | |
616 | ||
0db9d74e MA |
617 | /* |
618 | * Check for sufficient descriptors in the cache to accommodate | |
619 | * the number of in progress add operations plus regions_needed. | |
620 | * | |
621 | * This is a while loop because when we drop the lock, some other call | |
622 | * to region_add or region_del may have consumed some region_entries, | |
623 | * so we keep looping here until we finally have enough entries for | |
624 | * (adds_in_progress + regions_needed). | |
625 | */ | |
626 | while (resv->region_cache_count < | |
627 | (resv->adds_in_progress + regions_needed)) { | |
628 | to_allocate = resv->adds_in_progress + regions_needed - | |
629 | resv->region_cache_count; | |
630 | ||
631 | /* At this point, we should have enough entries in the cache | |
f0953a1b | 632 | * for all the existing adds_in_progress. We should only be |
0db9d74e | 633 | * needing to allocate for regions_needed. |
d75c6af9 | 634 | */ |
0db9d74e MA |
635 | VM_BUG_ON(resv->region_cache_count < resv->adds_in_progress); |
636 | ||
637 | spin_unlock(&resv->lock); | |
638 | for (i = 0; i < to_allocate; i++) { | |
639 | trg = kmalloc(sizeof(*trg), GFP_KERNEL); | |
640 | if (!trg) | |
641 | goto out_of_memory; | |
642 | list_add(&trg->link, &allocated_regions); | |
d75c6af9 | 643 | } |
d75c6af9 | 644 | |
0db9d74e MA |
645 | spin_lock(&resv->lock); |
646 | ||
d3ec7b6e WY |
647 | list_splice(&allocated_regions, &resv->region_cache); |
648 | resv->region_cache_count += to_allocate; | |
d75c6af9 MA |
649 | } |
650 | ||
0db9d74e | 651 | return 0; |
d75c6af9 | 652 | |
0db9d74e MA |
653 | out_of_memory: |
654 | list_for_each_entry_safe(rg, trg, &allocated_regions, link) { | |
655 | list_del(&rg->link); | |
656 | kfree(rg); | |
657 | } | |
658 | return -ENOMEM; | |
d75c6af9 MA |
659 | } |
660 | ||
1dd308a7 MK |
661 | /* |
662 | * Add the huge page range represented by [f, t) to the reserve | |
0db9d74e MA |
663 | * map. Regions will be taken from the cache to fill in this range. |
664 | * Sufficient regions should exist in the cache due to the previous | |
665 | * call to region_chg with the same range, but in some cases the cache will not | |
666 | * have sufficient entries due to races with other code doing region_add or | |
667 | * region_del. The extra needed entries will be allocated. | |
cf3ad20b | 668 | * |
0db9d74e MA |
669 | * regions_needed is the out value provided by a previous call to region_chg. |
670 | * | |
671 | * Return the number of new huge pages added to the map. This number is greater | |
672 | * than or equal to zero. If file_region entries needed to be allocated for | |
7c8de358 | 673 | * this operation and we were not able to allocate, it returns -ENOMEM. |
0db9d74e MA |
674 | * region_add of regions of length 1 never allocate file_regions and cannot |
675 | * fail; region_chg will always allocate at least 1 entry and a region_add for | |
676 | * 1 page will only require at most 1 entry. | |
1dd308a7 | 677 | */ |
0db9d74e | 678 | static long region_add(struct resv_map *resv, long f, long t, |
075a61d0 MA |
679 | long in_regions_needed, struct hstate *h, |
680 | struct hugetlb_cgroup *h_cg) | |
96822904 | 681 | { |
0db9d74e | 682 | long add = 0, actual_regions_needed = 0; |
96822904 | 683 | |
7b24d861 | 684 | spin_lock(&resv->lock); |
0db9d74e MA |
685 | retry: |
686 | ||
687 | /* Count how many regions are actually needed to execute this add. */ | |
972a3da3 WY |
688 | add_reservation_in_range(resv, f, t, NULL, NULL, |
689 | &actual_regions_needed); | |
96822904 | 690 | |
5e911373 | 691 | /* |
0db9d74e MA |
692 | * Check for sufficient descriptors in the cache to accommodate |
693 | * this add operation. Note that actual_regions_needed may be greater | |
694 | * than in_regions_needed, as the resv_map may have been modified since | |
695 | * the region_chg call. In this case, we need to make sure that we | |
696 | * allocate extra entries, such that we have enough for all the | |
697 | * existing adds_in_progress, plus the excess needed for this | |
698 | * operation. | |
5e911373 | 699 | */ |
0db9d74e MA |
700 | if (actual_regions_needed > in_regions_needed && |
701 | resv->region_cache_count < | |
702 | resv->adds_in_progress + | |
703 | (actual_regions_needed - in_regions_needed)) { | |
704 | /* region_add operation of range 1 should never need to | |
705 | * allocate file_region entries. | |
706 | */ | |
707 | VM_BUG_ON(t - f <= 1); | |
5e911373 | 708 | |
0db9d74e MA |
709 | if (allocate_file_region_entries( |
710 | resv, actual_regions_needed - in_regions_needed)) { | |
711 | return -ENOMEM; | |
712 | } | |
5e911373 | 713 | |
0db9d74e | 714 | goto retry; |
5e911373 MK |
715 | } |
716 | ||
972a3da3 | 717 | add = add_reservation_in_range(resv, f, t, h_cg, h, NULL); |
0db9d74e MA |
718 | |
719 | resv->adds_in_progress -= in_regions_needed; | |
cf3ad20b | 720 | |
7b24d861 | 721 | spin_unlock(&resv->lock); |
cf3ad20b | 722 | return add; |
96822904 AW |
723 | } |
724 | ||
1dd308a7 MK |
725 | /* |
726 | * Examine the existing reserve map and determine how many | |
727 | * huge pages in the specified range [f, t) are NOT currently | |
728 | * represented. This routine is called before a subsequent | |
729 | * call to region_add that will actually modify the reserve | |
730 | * map to add the specified range [f, t). region_chg does | |
731 | * not change the number of huge pages represented by the | |
0db9d74e MA |
732 | * map. A number of new file_region structures is added to the cache as a |
733 | * placeholder, for the subsequent region_add call to use. At least 1 | |
734 | * file_region structure is added. | |
735 | * | |
736 | * out_regions_needed is the number of regions added to the | |
737 | * resv->adds_in_progress. This value needs to be provided to a follow up call | |
738 | * to region_add or region_abort for proper accounting. | |
5e911373 MK |
739 | * |
740 | * Returns the number of huge pages that need to be added to the existing | |
741 | * reservation map for the range [f, t). This number is greater or equal to | |
742 | * zero. -ENOMEM is returned if a new file_region structure or cache entry | |
743 | * is needed and can not be allocated. | |
1dd308a7 | 744 | */ |
0db9d74e MA |
745 | static long region_chg(struct resv_map *resv, long f, long t, |
746 | long *out_regions_needed) | |
96822904 | 747 | { |
96822904 AW |
748 | long chg = 0; |
749 | ||
7b24d861 | 750 | spin_lock(&resv->lock); |
5e911373 | 751 | |
972a3da3 | 752 | /* Count how many hugepages in this range are NOT represented. */ |
075a61d0 | 753 | chg = add_reservation_in_range(resv, f, t, NULL, NULL, |
972a3da3 | 754 | out_regions_needed); |
5e911373 | 755 | |
0db9d74e MA |
756 | if (*out_regions_needed == 0) |
757 | *out_regions_needed = 1; | |
5e911373 | 758 | |
0db9d74e MA |
759 | if (allocate_file_region_entries(resv, *out_regions_needed)) |
760 | return -ENOMEM; | |
5e911373 | 761 | |
0db9d74e | 762 | resv->adds_in_progress += *out_regions_needed; |
7b24d861 | 763 | |
7b24d861 | 764 | spin_unlock(&resv->lock); |
96822904 AW |
765 | return chg; |
766 | } | |
767 | ||
5e911373 MK |
768 | /* |
769 | * Abort the in progress add operation. The adds_in_progress field | |
770 | * of the resv_map keeps track of the operations in progress between | |
771 | * calls to region_chg and region_add. Operations are sometimes | |
772 | * aborted after the call to region_chg. In such cases, region_abort | |
0db9d74e MA |
773 | * is called to decrement the adds_in_progress counter. regions_needed |
774 | * is the value returned by the region_chg call, it is used to decrement | |
775 | * the adds_in_progress counter. | |
5e911373 MK |
776 | * |
777 | * NOTE: The range arguments [f, t) are not needed or used in this | |
778 | * routine. They are kept to make reading the calling code easier as | |
779 | * arguments will match the associated region_chg call. | |
780 | */ | |
0db9d74e MA |
781 | static void region_abort(struct resv_map *resv, long f, long t, |
782 | long regions_needed) | |
5e911373 MK |
783 | { |
784 | spin_lock(&resv->lock); | |
785 | VM_BUG_ON(!resv->region_cache_count); | |
0db9d74e | 786 | resv->adds_in_progress -= regions_needed; |
5e911373 MK |
787 | spin_unlock(&resv->lock); |
788 | } | |
789 | ||
1dd308a7 | 790 | /* |
feba16e2 MK |
791 | * Delete the specified range [f, t) from the reserve map. If the |
792 | * t parameter is LONG_MAX, this indicates that ALL regions after f | |
793 | * should be deleted. Locate the regions which intersect [f, t) | |
794 | * and either trim, delete or split the existing regions. | |
795 | * | |
796 | * Returns the number of huge pages deleted from the reserve map. | |
797 | * In the normal case, the return value is zero or more. In the | |
798 | * case where a region must be split, a new region descriptor must | |
799 | * be allocated. If the allocation fails, -ENOMEM will be returned. | |
800 | * NOTE: If the parameter t == LONG_MAX, then we will never split | |
801 | * a region and possibly return -ENOMEM. Callers specifying | |
802 | * t == LONG_MAX do not need to check for -ENOMEM error. | |
1dd308a7 | 803 | */ |
feba16e2 | 804 | static long region_del(struct resv_map *resv, long f, long t) |
96822904 | 805 | { |
1406ec9b | 806 | struct list_head *head = &resv->regions; |
96822904 | 807 | struct file_region *rg, *trg; |
feba16e2 MK |
808 | struct file_region *nrg = NULL; |
809 | long del = 0; | |
96822904 | 810 | |
feba16e2 | 811 | retry: |
7b24d861 | 812 | spin_lock(&resv->lock); |
feba16e2 | 813 | list_for_each_entry_safe(rg, trg, head, link) { |
dbe409e4 MK |
814 | /* |
815 | * Skip regions before the range to be deleted. file_region | |
816 | * ranges are normally of the form [from, to). However, there | |
817 | * may be a "placeholder" entry in the map which is of the form | |
818 | * (from, to) with from == to. Check for placeholder entries | |
819 | * at the beginning of the range to be deleted. | |
820 | */ | |
821 | if (rg->to <= f && (rg->to != rg->from || rg->to != f)) | |
feba16e2 | 822 | continue; |
dbe409e4 | 823 | |
feba16e2 | 824 | if (rg->from >= t) |
96822904 | 825 | break; |
96822904 | 826 | |
feba16e2 MK |
827 | if (f > rg->from && t < rg->to) { /* Must split region */ |
828 | /* | |
829 | * Check for an entry in the cache before dropping | |
830 | * lock and attempting allocation. | |
831 | */ | |
832 | if (!nrg && | |
833 | resv->region_cache_count > resv->adds_in_progress) { | |
834 | nrg = list_first_entry(&resv->region_cache, | |
835 | struct file_region, | |
836 | link); | |
837 | list_del(&nrg->link); | |
838 | resv->region_cache_count--; | |
839 | } | |
96822904 | 840 | |
feba16e2 MK |
841 | if (!nrg) { |
842 | spin_unlock(&resv->lock); | |
843 | nrg = kmalloc(sizeof(*nrg), GFP_KERNEL); | |
844 | if (!nrg) | |
845 | return -ENOMEM; | |
846 | goto retry; | |
847 | } | |
848 | ||
849 | del += t - f; | |
79aa925b | 850 | hugetlb_cgroup_uncharge_file_region( |
d85aecf2 | 851 | resv, rg, t - f, false); |
feba16e2 MK |
852 | |
853 | /* New entry for end of split region */ | |
854 | nrg->from = t; | |
855 | nrg->to = rg->to; | |
075a61d0 MA |
856 | |
857 | copy_hugetlb_cgroup_uncharge_info(nrg, rg); | |
858 | ||
feba16e2 MK |
859 | INIT_LIST_HEAD(&nrg->link); |
860 | ||
861 | /* Original entry is trimmed */ | |
862 | rg->to = f; | |
863 | ||
864 | list_add(&nrg->link, &rg->link); | |
865 | nrg = NULL; | |
96822904 | 866 | break; |
feba16e2 MK |
867 | } |
868 | ||
869 | if (f <= rg->from && t >= rg->to) { /* Remove entire region */ | |
870 | del += rg->to - rg->from; | |
075a61d0 | 871 | hugetlb_cgroup_uncharge_file_region(resv, rg, |
d85aecf2 | 872 | rg->to - rg->from, true); |
feba16e2 MK |
873 | list_del(&rg->link); |
874 | kfree(rg); | |
875 | continue; | |
876 | } | |
877 | ||
878 | if (f <= rg->from) { /* Trim beginning of region */ | |
075a61d0 | 879 | hugetlb_cgroup_uncharge_file_region(resv, rg, |
d85aecf2 | 880 | t - rg->from, false); |
075a61d0 | 881 | |
79aa925b MK |
882 | del += t - rg->from; |
883 | rg->from = t; | |
884 | } else { /* Trim end of region */ | |
075a61d0 | 885 | hugetlb_cgroup_uncharge_file_region(resv, rg, |
d85aecf2 | 886 | rg->to - f, false); |
79aa925b MK |
887 | |
888 | del += rg->to - f; | |
889 | rg->to = f; | |
feba16e2 | 890 | } |
96822904 | 891 | } |
7b24d861 | 892 | |
7b24d861 | 893 | spin_unlock(&resv->lock); |
feba16e2 MK |
894 | kfree(nrg); |
895 | return del; | |
96822904 AW |
896 | } |
897 | ||
b5cec28d MK |
898 | /* |
899 | * A rare out of memory error was encountered which prevented removal of | |
900 | * the reserve map region for a page. The huge page itself was free'ed | |
901 | * and removed from the page cache. This routine will adjust the subpool | |
902 | * usage count, and the global reserve count if needed. By incrementing | |
903 | * these counts, the reserve map entry which could not be deleted will | |
904 | * appear as a "reserved" entry instead of simply dangling with incorrect | |
905 | * counts. | |
906 | */ | |
72e2936c | 907 | void hugetlb_fix_reserve_counts(struct inode *inode) |
b5cec28d MK |
908 | { |
909 | struct hugepage_subpool *spool = subpool_inode(inode); | |
910 | long rsv_adjust; | |
da56388c | 911 | bool reserved = false; |
b5cec28d MK |
912 | |
913 | rsv_adjust = hugepage_subpool_get_pages(spool, 1); | |
da56388c | 914 | if (rsv_adjust > 0) { |
b5cec28d MK |
915 | struct hstate *h = hstate_inode(inode); |
916 | ||
da56388c ML |
917 | if (!hugetlb_acct_memory(h, 1)) |
918 | reserved = true; | |
919 | } else if (!rsv_adjust) { | |
920 | reserved = true; | |
b5cec28d | 921 | } |
da56388c ML |
922 | |
923 | if (!reserved) | |
924 | pr_warn("hugetlb: Huge Page Reserved count may go negative.\n"); | |
b5cec28d MK |
925 | } |
926 | ||
1dd308a7 MK |
927 | /* |
928 | * Count and return the number of huge pages in the reserve map | |
929 | * that intersect with the range [f, t). | |
930 | */ | |
1406ec9b | 931 | static long region_count(struct resv_map *resv, long f, long t) |
84afd99b | 932 | { |
1406ec9b | 933 | struct list_head *head = &resv->regions; |
84afd99b AW |
934 | struct file_region *rg; |
935 | long chg = 0; | |
936 | ||
7b24d861 | 937 | spin_lock(&resv->lock); |
84afd99b AW |
938 | /* Locate each segment we overlap with, and count that overlap. */ |
939 | list_for_each_entry(rg, head, link) { | |
f2135a4a WSH |
940 | long seg_from; |
941 | long seg_to; | |
84afd99b AW |
942 | |
943 | if (rg->to <= f) | |
944 | continue; | |
945 | if (rg->from >= t) | |
946 | break; | |
947 | ||
948 | seg_from = max(rg->from, f); | |
949 | seg_to = min(rg->to, t); | |
950 | ||
951 | chg += seg_to - seg_from; | |
952 | } | |
7b24d861 | 953 | spin_unlock(&resv->lock); |
84afd99b AW |
954 | |
955 | return chg; | |
956 | } | |
957 | ||
e7c4b0bf AW |
958 | /* |
959 | * Convert the address within this vma to the page offset within | |
960 | * the mapping, in pagecache page units; huge pages here. | |
961 | */ | |
a5516438 AK |
962 | static pgoff_t vma_hugecache_offset(struct hstate *h, |
963 | struct vm_area_struct *vma, unsigned long address) | |
e7c4b0bf | 964 | { |
a5516438 AK |
965 | return ((address - vma->vm_start) >> huge_page_shift(h)) + |
966 | (vma->vm_pgoff >> huge_page_order(h)); | |
e7c4b0bf AW |
967 | } |
968 | ||
0fe6e20b NH |
969 | pgoff_t linear_hugepage_index(struct vm_area_struct *vma, |
970 | unsigned long address) | |
971 | { | |
972 | return vma_hugecache_offset(hstate_vma(vma), vma, address); | |
973 | } | |
dee41079 | 974 | EXPORT_SYMBOL_GPL(linear_hugepage_index); |
0fe6e20b | 975 | |
08fba699 MG |
976 | /* |
977 | * Return the size of the pages allocated when backing a VMA. In the majority | |
978 | * cases this will be same size as used by the page table entries. | |
979 | */ | |
980 | unsigned long vma_kernel_pagesize(struct vm_area_struct *vma) | |
981 | { | |
05ea8860 DW |
982 | if (vma->vm_ops && vma->vm_ops->pagesize) |
983 | return vma->vm_ops->pagesize(vma); | |
984 | return PAGE_SIZE; | |
08fba699 | 985 | } |
f340ca0f | 986 | EXPORT_SYMBOL_GPL(vma_kernel_pagesize); |
08fba699 | 987 | |
3340289d MG |
988 | /* |
989 | * Return the page size being used by the MMU to back a VMA. In the majority | |
990 | * of cases, the page size used by the kernel matches the MMU size. On | |
09135cc5 DW |
991 | * architectures where it differs, an architecture-specific 'strong' |
992 | * version of this symbol is required. | |
3340289d | 993 | */ |
09135cc5 | 994 | __weak unsigned long vma_mmu_pagesize(struct vm_area_struct *vma) |
3340289d MG |
995 | { |
996 | return vma_kernel_pagesize(vma); | |
997 | } | |
3340289d | 998 | |
84afd99b AW |
999 | /* |
1000 | * Flags for MAP_PRIVATE reservations. These are stored in the bottom | |
1001 | * bits of the reservation map pointer, which are always clear due to | |
1002 | * alignment. | |
1003 | */ | |
1004 | #define HPAGE_RESV_OWNER (1UL << 0) | |
1005 | #define HPAGE_RESV_UNMAPPED (1UL << 1) | |
04f2cbe3 | 1006 | #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED) |
84afd99b | 1007 | |
a1e78772 MG |
1008 | /* |
1009 | * These helpers are used to track how many pages are reserved for | |
1010 | * faults in a MAP_PRIVATE mapping. Only the process that called mmap() | |
1011 | * is guaranteed to have their future faults succeed. | |
1012 | * | |
8d9bfb26 | 1013 | * With the exception of hugetlb_dup_vma_private() which is called at fork(), |
a1e78772 MG |
1014 | * the reserve counters are updated with the hugetlb_lock held. It is safe |
1015 | * to reset the VMA at fork() time as it is not in use yet and there is no | |
1016 | * chance of the global counters getting corrupted as a result of the values. | |
84afd99b AW |
1017 | * |
1018 | * The private mapping reservation is represented in a subtly different | |
1019 | * manner to a shared mapping. A shared mapping has a region map associated | |
1020 | * with the underlying file, this region map represents the backing file | |
1021 | * pages which have ever had a reservation assigned which this persists even | |
1022 | * after the page is instantiated. A private mapping has a region map | |
1023 | * associated with the original mmap which is attached to all VMAs which | |
1024 | * reference it, this region map represents those offsets which have consumed | |
1025 | * reservation ie. where pages have been instantiated. | |
a1e78772 | 1026 | */ |
e7c4b0bf AW |
1027 | static unsigned long get_vma_private_data(struct vm_area_struct *vma) |
1028 | { | |
1029 | return (unsigned long)vma->vm_private_data; | |
1030 | } | |
1031 | ||
1032 | static void set_vma_private_data(struct vm_area_struct *vma, | |
1033 | unsigned long value) | |
1034 | { | |
1035 | vma->vm_private_data = (void *)value; | |
1036 | } | |
1037 | ||
e9fe92ae MA |
1038 | static void |
1039 | resv_map_set_hugetlb_cgroup_uncharge_info(struct resv_map *resv_map, | |
1040 | struct hugetlb_cgroup *h_cg, | |
1041 | struct hstate *h) | |
1042 | { | |
1043 | #ifdef CONFIG_CGROUP_HUGETLB | |
1044 | if (!h_cg || !h) { | |
1045 | resv_map->reservation_counter = NULL; | |
1046 | resv_map->pages_per_hpage = 0; | |
1047 | resv_map->css = NULL; | |
1048 | } else { | |
1049 | resv_map->reservation_counter = | |
1050 | &h_cg->rsvd_hugepage[hstate_index(h)]; | |
1051 | resv_map->pages_per_hpage = pages_per_huge_page(h); | |
1052 | resv_map->css = &h_cg->css; | |
1053 | } | |
1054 | #endif | |
1055 | } | |
1056 | ||
9119a41e | 1057 | struct resv_map *resv_map_alloc(void) |
84afd99b AW |
1058 | { |
1059 | struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL); | |
5e911373 MK |
1060 | struct file_region *rg = kmalloc(sizeof(*rg), GFP_KERNEL); |
1061 | ||
1062 | if (!resv_map || !rg) { | |
1063 | kfree(resv_map); | |
1064 | kfree(rg); | |
84afd99b | 1065 | return NULL; |
5e911373 | 1066 | } |
84afd99b AW |
1067 | |
1068 | kref_init(&resv_map->refs); | |
7b24d861 | 1069 | spin_lock_init(&resv_map->lock); |
84afd99b AW |
1070 | INIT_LIST_HEAD(&resv_map->regions); |
1071 | ||
5e911373 | 1072 | resv_map->adds_in_progress = 0; |
e9fe92ae MA |
1073 | /* |
1074 | * Initialize these to 0. On shared mappings, 0's here indicate these | |
1075 | * fields don't do cgroup accounting. On private mappings, these will be | |
1076 | * re-initialized to the proper values, to indicate that hugetlb cgroup | |
1077 | * reservations are to be un-charged from here. | |
1078 | */ | |
1079 | resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, NULL, NULL); | |
5e911373 MK |
1080 | |
1081 | INIT_LIST_HEAD(&resv_map->region_cache); | |
1082 | list_add(&rg->link, &resv_map->region_cache); | |
1083 | resv_map->region_cache_count = 1; | |
1084 | ||
84afd99b AW |
1085 | return resv_map; |
1086 | } | |
1087 | ||
9119a41e | 1088 | void resv_map_release(struct kref *ref) |
84afd99b AW |
1089 | { |
1090 | struct resv_map *resv_map = container_of(ref, struct resv_map, refs); | |
5e911373 MK |
1091 | struct list_head *head = &resv_map->region_cache; |
1092 | struct file_region *rg, *trg; | |
84afd99b AW |
1093 | |
1094 | /* Clear out any active regions before we release the map. */ | |
feba16e2 | 1095 | region_del(resv_map, 0, LONG_MAX); |
5e911373 MK |
1096 | |
1097 | /* ... and any entries left in the cache */ | |
1098 | list_for_each_entry_safe(rg, trg, head, link) { | |
1099 | list_del(&rg->link); | |
1100 | kfree(rg); | |
1101 | } | |
1102 | ||
1103 | VM_BUG_ON(resv_map->adds_in_progress); | |
1104 | ||
84afd99b AW |
1105 | kfree(resv_map); |
1106 | } | |
1107 | ||
4e35f483 JK |
1108 | static inline struct resv_map *inode_resv_map(struct inode *inode) |
1109 | { | |
f27a5136 MK |
1110 | /* |
1111 | * At inode evict time, i_mapping may not point to the original | |
1112 | * address space within the inode. This original address space | |
1113 | * contains the pointer to the resv_map. So, always use the | |
1114 | * address space embedded within the inode. | |
1115 | * The VERY common case is inode->mapping == &inode->i_data but, | |
1116 | * this may not be true for device special inodes. | |
1117 | */ | |
1118 | return (struct resv_map *)(&inode->i_data)->private_data; | |
4e35f483 JK |
1119 | } |
1120 | ||
84afd99b | 1121 | static struct resv_map *vma_resv_map(struct vm_area_struct *vma) |
a1e78772 | 1122 | { |
81d1b09c | 1123 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); |
4e35f483 JK |
1124 | if (vma->vm_flags & VM_MAYSHARE) { |
1125 | struct address_space *mapping = vma->vm_file->f_mapping; | |
1126 | struct inode *inode = mapping->host; | |
1127 | ||
1128 | return inode_resv_map(inode); | |
1129 | ||
1130 | } else { | |
84afd99b AW |
1131 | return (struct resv_map *)(get_vma_private_data(vma) & |
1132 | ~HPAGE_RESV_MASK); | |
4e35f483 | 1133 | } |
a1e78772 MG |
1134 | } |
1135 | ||
84afd99b | 1136 | static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map) |
a1e78772 | 1137 | { |
81d1b09c SL |
1138 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); |
1139 | VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma); | |
a1e78772 | 1140 | |
84afd99b AW |
1141 | set_vma_private_data(vma, (get_vma_private_data(vma) & |
1142 | HPAGE_RESV_MASK) | (unsigned long)map); | |
04f2cbe3 MG |
1143 | } |
1144 | ||
1145 | static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags) | |
1146 | { | |
81d1b09c SL |
1147 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); |
1148 | VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma); | |
e7c4b0bf AW |
1149 | |
1150 | set_vma_private_data(vma, get_vma_private_data(vma) | flags); | |
04f2cbe3 MG |
1151 | } |
1152 | ||
1153 | static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag) | |
1154 | { | |
81d1b09c | 1155 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); |
e7c4b0bf AW |
1156 | |
1157 | return (get_vma_private_data(vma) & flag) != 0; | |
a1e78772 MG |
1158 | } |
1159 | ||
8d9bfb26 | 1160 | void hugetlb_dup_vma_private(struct vm_area_struct *vma) |
a1e78772 | 1161 | { |
81d1b09c | 1162 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); |
8d9bfb26 MK |
1163 | /* |
1164 | * Clear vm_private_data | |
612b8a31 MK |
1165 | * - For shared mappings this is a per-vma semaphore that may be |
1166 | * allocated in a subsequent call to hugetlb_vm_op_open. | |
1167 | * Before clearing, make sure pointer is not associated with vma | |
1168 | * as this will leak the structure. This is the case when called | |
1169 | * via clear_vma_resv_huge_pages() and hugetlb_vm_op_open has already | |
1170 | * been called to allocate a new structure. | |
8d9bfb26 MK |
1171 | * - For MAP_PRIVATE mappings, this is the reserve map which does |
1172 | * not apply to children. Faults generated by the children are | |
1173 | * not guaranteed to succeed, even if read-only. | |
8d9bfb26 | 1174 | */ |
612b8a31 MK |
1175 | if (vma->vm_flags & VM_MAYSHARE) { |
1176 | struct hugetlb_vma_lock *vma_lock = vma->vm_private_data; | |
1177 | ||
1178 | if (vma_lock && vma_lock->vma != vma) | |
1179 | vma->vm_private_data = NULL; | |
1180 | } else | |
1181 | vma->vm_private_data = NULL; | |
a1e78772 MG |
1182 | } |
1183 | ||
550a7d60 MA |
1184 | /* |
1185 | * Reset and decrement one ref on hugepage private reservation. | |
8651a137 | 1186 | * Called with mm->mmap_lock writer semaphore held. |
550a7d60 MA |
1187 | * This function should be only used by move_vma() and operate on |
1188 | * same sized vma. It should never come here with last ref on the | |
1189 | * reservation. | |
1190 | */ | |
1191 | void clear_vma_resv_huge_pages(struct vm_area_struct *vma) | |
1192 | { | |
1193 | /* | |
1194 | * Clear the old hugetlb private page reservation. | |
1195 | * It has already been transferred to new_vma. | |
1196 | * | |
1197 | * During a mremap() operation of a hugetlb vma we call move_vma() | |
1198 | * which copies vma into new_vma and unmaps vma. After the copy | |
1199 | * operation both new_vma and vma share a reference to the resv_map | |
1200 | * struct, and at that point vma is about to be unmapped. We don't | |
1201 | * want to return the reservation to the pool at unmap of vma because | |
1202 | * the reservation still lives on in new_vma, so simply decrement the | |
1203 | * ref here and remove the resv_map reference from this vma. | |
1204 | */ | |
1205 | struct resv_map *reservations = vma_resv_map(vma); | |
1206 | ||
afe041c2 BQM |
1207 | if (reservations && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { |
1208 | resv_map_put_hugetlb_cgroup_uncharge_info(reservations); | |
550a7d60 | 1209 | kref_put(&reservations->refs, resv_map_release); |
afe041c2 | 1210 | } |
550a7d60 | 1211 | |
8d9bfb26 | 1212 | hugetlb_dup_vma_private(vma); |
550a7d60 MA |
1213 | } |
1214 | ||
a1e78772 | 1215 | /* Returns true if the VMA has associated reserve pages */ |
559ec2f8 | 1216 | static bool vma_has_reserves(struct vm_area_struct *vma, long chg) |
a1e78772 | 1217 | { |
af0ed73e JK |
1218 | if (vma->vm_flags & VM_NORESERVE) { |
1219 | /* | |
1220 | * This address is already reserved by other process(chg == 0), | |
1221 | * so, we should decrement reserved count. Without decrementing, | |
1222 | * reserve count remains after releasing inode, because this | |
1223 | * allocated page will go into page cache and is regarded as | |
1224 | * coming from reserved pool in releasing step. Currently, we | |
1225 | * don't have any other solution to deal with this situation | |
1226 | * properly, so add work-around here. | |
1227 | */ | |
1228 | if (vma->vm_flags & VM_MAYSHARE && chg == 0) | |
559ec2f8 | 1229 | return true; |
af0ed73e | 1230 | else |
559ec2f8 | 1231 | return false; |
af0ed73e | 1232 | } |
a63884e9 JK |
1233 | |
1234 | /* Shared mappings always use reserves */ | |
1fb1b0e9 MK |
1235 | if (vma->vm_flags & VM_MAYSHARE) { |
1236 | /* | |
1237 | * We know VM_NORESERVE is not set. Therefore, there SHOULD | |
1238 | * be a region map for all pages. The only situation where | |
1239 | * there is no region map is if a hole was punched via | |
7c8de358 | 1240 | * fallocate. In this case, there really are no reserves to |
1fb1b0e9 MK |
1241 | * use. This situation is indicated if chg != 0. |
1242 | */ | |
1243 | if (chg) | |
1244 | return false; | |
1245 | else | |
1246 | return true; | |
1247 | } | |
a63884e9 JK |
1248 | |
1249 | /* | |
1250 | * Only the process that called mmap() has reserves for | |
1251 | * private mappings. | |
1252 | */ | |
67961f9d MK |
1253 | if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { |
1254 | /* | |
1255 | * Like the shared case above, a hole punch or truncate | |
1256 | * could have been performed on the private mapping. | |
1257 | * Examine the value of chg to determine if reserves | |
1258 | * actually exist or were previously consumed. | |
1259 | * Very Subtle - The value of chg comes from a previous | |
1260 | * call to vma_needs_reserves(). The reserve map for | |
1261 | * private mappings has different (opposite) semantics | |
1262 | * than that of shared mappings. vma_needs_reserves() | |
1263 | * has already taken this difference in semantics into | |
1264 | * account. Therefore, the meaning of chg is the same | |
1265 | * as in the shared case above. Code could easily be | |
1266 | * combined, but keeping it separate draws attention to | |
1267 | * subtle differences. | |
1268 | */ | |
1269 | if (chg) | |
1270 | return false; | |
1271 | else | |
1272 | return true; | |
1273 | } | |
a63884e9 | 1274 | |
559ec2f8 | 1275 | return false; |
a1e78772 MG |
1276 | } |
1277 | ||
240d67a8 | 1278 | static void enqueue_hugetlb_folio(struct hstate *h, struct folio *folio) |
1da177e4 | 1279 | { |
240d67a8 | 1280 | int nid = folio_nid(folio); |
9487ca60 MK |
1281 | |
1282 | lockdep_assert_held(&hugetlb_lock); | |
240d67a8 | 1283 | VM_BUG_ON_FOLIO(folio_ref_count(folio), folio); |
b65a4eda | 1284 | |
240d67a8 | 1285 | list_move(&folio->lru, &h->hugepage_freelists[nid]); |
a5516438 AK |
1286 | h->free_huge_pages++; |
1287 | h->free_huge_pages_node[nid]++; | |
240d67a8 | 1288 | folio_set_hugetlb_freed(folio); |
1da177e4 LT |
1289 | } |
1290 | ||
94310cbc | 1291 | static struct page *dequeue_huge_page_node_exact(struct hstate *h, int nid) |
bf50bab2 NH |
1292 | { |
1293 | struct page *page; | |
1a08ae36 | 1294 | bool pin = !!(current->flags & PF_MEMALLOC_PIN); |
bbe88753 | 1295 | |
9487ca60 | 1296 | lockdep_assert_held(&hugetlb_lock); |
bbe88753 | 1297 | list_for_each_entry(page, &h->hugepage_freelists[nid], lru) { |
6077c943 | 1298 | if (pin && !is_longterm_pinnable_page(page)) |
bbe88753 | 1299 | continue; |
bf50bab2 | 1300 | |
6664bfc8 WY |
1301 | if (PageHWPoison(page)) |
1302 | continue; | |
1303 | ||
1304 | list_move(&page->lru, &h->hugepage_activelist); | |
1305 | set_page_refcounted(page); | |
6c037149 | 1306 | ClearHPageFreed(page); |
6664bfc8 WY |
1307 | h->free_huge_pages--; |
1308 | h->free_huge_pages_node[nid]--; | |
1309 | return page; | |
bbe88753 JK |
1310 | } |
1311 | ||
6664bfc8 | 1312 | return NULL; |
bf50bab2 NH |
1313 | } |
1314 | ||
3e59fcb0 MH |
1315 | static struct page *dequeue_huge_page_nodemask(struct hstate *h, gfp_t gfp_mask, int nid, |
1316 | nodemask_t *nmask) | |
94310cbc | 1317 | { |
3e59fcb0 MH |
1318 | unsigned int cpuset_mems_cookie; |
1319 | struct zonelist *zonelist; | |
1320 | struct zone *zone; | |
1321 | struct zoneref *z; | |
98fa15f3 | 1322 | int node = NUMA_NO_NODE; |
94310cbc | 1323 | |
3e59fcb0 MH |
1324 | zonelist = node_zonelist(nid, gfp_mask); |
1325 | ||
1326 | retry_cpuset: | |
1327 | cpuset_mems_cookie = read_mems_allowed_begin(); | |
1328 | for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), nmask) { | |
1329 | struct page *page; | |
1330 | ||
1331 | if (!cpuset_zone_allowed(zone, gfp_mask)) | |
1332 | continue; | |
1333 | /* | |
1334 | * no need to ask again on the same node. Pool is node rather than | |
1335 | * zone aware | |
1336 | */ | |
1337 | if (zone_to_nid(zone) == node) | |
1338 | continue; | |
1339 | node = zone_to_nid(zone); | |
94310cbc | 1340 | |
94310cbc AK |
1341 | page = dequeue_huge_page_node_exact(h, node); |
1342 | if (page) | |
1343 | return page; | |
1344 | } | |
3e59fcb0 MH |
1345 | if (unlikely(read_mems_allowed_retry(cpuset_mems_cookie))) |
1346 | goto retry_cpuset; | |
1347 | ||
94310cbc AK |
1348 | return NULL; |
1349 | } | |
1350 | ||
8346d69d XH |
1351 | static unsigned long available_huge_pages(struct hstate *h) |
1352 | { | |
1353 | return h->free_huge_pages - h->resv_huge_pages; | |
1354 | } | |
1355 | ||
a5516438 AK |
1356 | static struct page *dequeue_huge_page_vma(struct hstate *h, |
1357 | struct vm_area_struct *vma, | |
af0ed73e JK |
1358 | unsigned long address, int avoid_reserve, |
1359 | long chg) | |
1da177e4 | 1360 | { |
cfcaa66f | 1361 | struct page *page = NULL; |
480eccf9 | 1362 | struct mempolicy *mpol; |
04ec6264 | 1363 | gfp_t gfp_mask; |
3e59fcb0 | 1364 | nodemask_t *nodemask; |
04ec6264 | 1365 | int nid; |
1da177e4 | 1366 | |
a1e78772 MG |
1367 | /* |
1368 | * A child process with MAP_PRIVATE mappings created by their parent | |
1369 | * have no page reserves. This check ensures that reservations are | |
1370 | * not "stolen". The child may still get SIGKILLed | |
1371 | */ | |
8346d69d | 1372 | if (!vma_has_reserves(vma, chg) && !available_huge_pages(h)) |
c0ff7453 | 1373 | goto err; |
a1e78772 | 1374 | |
04f2cbe3 | 1375 | /* If reserves cannot be used, ensure enough pages are in the pool */ |
8346d69d | 1376 | if (avoid_reserve && !available_huge_pages(h)) |
6eab04a8 | 1377 | goto err; |
04f2cbe3 | 1378 | |
04ec6264 VB |
1379 | gfp_mask = htlb_alloc_mask(h); |
1380 | nid = huge_node(vma, address, gfp_mask, &mpol, &nodemask); | |
cfcaa66f BW |
1381 | |
1382 | if (mpol_is_preferred_many(mpol)) { | |
1383 | page = dequeue_huge_page_nodemask(h, gfp_mask, nid, nodemask); | |
1384 | ||
1385 | /* Fallback to all nodes if page==NULL */ | |
1386 | nodemask = NULL; | |
1387 | } | |
1388 | ||
1389 | if (!page) | |
1390 | page = dequeue_huge_page_nodemask(h, gfp_mask, nid, nodemask); | |
1391 | ||
3e59fcb0 | 1392 | if (page && !avoid_reserve && vma_has_reserves(vma, chg)) { |
d6995da3 | 1393 | SetHPageRestoreReserve(page); |
3e59fcb0 | 1394 | h->resv_huge_pages--; |
1da177e4 | 1395 | } |
cc9a6c87 | 1396 | |
52cd3b07 | 1397 | mpol_cond_put(mpol); |
1da177e4 | 1398 | return page; |
cc9a6c87 MG |
1399 | |
1400 | err: | |
cc9a6c87 | 1401 | return NULL; |
1da177e4 LT |
1402 | } |
1403 | ||
1cac6f2c LC |
1404 | /* |
1405 | * common helper functions for hstate_next_node_to_{alloc|free}. | |
1406 | * We may have allocated or freed a huge page based on a different | |
1407 | * nodes_allowed previously, so h->next_node_to_{alloc|free} might | |
1408 | * be outside of *nodes_allowed. Ensure that we use an allowed | |
1409 | * node for alloc or free. | |
1410 | */ | |
1411 | static int next_node_allowed(int nid, nodemask_t *nodes_allowed) | |
1412 | { | |
0edaf86c | 1413 | nid = next_node_in(nid, *nodes_allowed); |
1cac6f2c LC |
1414 | VM_BUG_ON(nid >= MAX_NUMNODES); |
1415 | ||
1416 | return nid; | |
1417 | } | |
1418 | ||
1419 | static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed) | |
1420 | { | |
1421 | if (!node_isset(nid, *nodes_allowed)) | |
1422 | nid = next_node_allowed(nid, nodes_allowed); | |
1423 | return nid; | |
1424 | } | |
1425 | ||
1426 | /* | |
1427 | * returns the previously saved node ["this node"] from which to | |
1428 | * allocate a persistent huge page for the pool and advance the | |
1429 | * next node from which to allocate, handling wrap at end of node | |
1430 | * mask. | |
1431 | */ | |
1432 | static int hstate_next_node_to_alloc(struct hstate *h, | |
1433 | nodemask_t *nodes_allowed) | |
1434 | { | |
1435 | int nid; | |
1436 | ||
1437 | VM_BUG_ON(!nodes_allowed); | |
1438 | ||
1439 | nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed); | |
1440 | h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed); | |
1441 | ||
1442 | return nid; | |
1443 | } | |
1444 | ||
1445 | /* | |
10c6ec49 | 1446 | * helper for remove_pool_huge_page() - return the previously saved |
1cac6f2c LC |
1447 | * node ["this node"] from which to free a huge page. Advance the |
1448 | * next node id whether or not we find a free huge page to free so | |
1449 | * that the next attempt to free addresses the next node. | |
1450 | */ | |
1451 | static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed) | |
1452 | { | |
1453 | int nid; | |
1454 | ||
1455 | VM_BUG_ON(!nodes_allowed); | |
1456 | ||
1457 | nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed); | |
1458 | h->next_nid_to_free = next_node_allowed(nid, nodes_allowed); | |
1459 | ||
1460 | return nid; | |
1461 | } | |
1462 | ||
1463 | #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \ | |
1464 | for (nr_nodes = nodes_weight(*mask); \ | |
1465 | nr_nodes > 0 && \ | |
1466 | ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \ | |
1467 | nr_nodes--) | |
1468 | ||
1469 | #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \ | |
1470 | for (nr_nodes = nodes_weight(*mask); \ | |
1471 | nr_nodes > 0 && \ | |
1472 | ((node = hstate_next_node_to_free(hs, mask)) || 1); \ | |
1473 | nr_nodes--) | |
1474 | ||
8531fc6f | 1475 | /* used to demote non-gigantic_huge pages as well */ |
911565b8 | 1476 | static void __destroy_compound_gigantic_folio(struct folio *folio, |
34d9e35b | 1477 | unsigned int order, bool demote) |
944d9fec LC |
1478 | { |
1479 | int i; | |
1480 | int nr_pages = 1 << order; | |
14455eab | 1481 | struct page *p; |
944d9fec | 1482 | |
911565b8 SK |
1483 | atomic_set(folio_mapcount_ptr(folio), 0); |
1484 | atomic_set(folio_subpages_mapcount_ptr(folio), 0); | |
1485 | atomic_set(folio_pincount_ptr(folio), 0); | |
47e29d32 | 1486 | |
14455eab | 1487 | for (i = 1; i < nr_pages; i++) { |
911565b8 | 1488 | p = folio_page(folio, i); |
a01f4390 | 1489 | p->mapping = NULL; |
1d798ca3 | 1490 | clear_compound_head(p); |
34d9e35b MK |
1491 | if (!demote) |
1492 | set_page_refcounted(p); | |
944d9fec LC |
1493 | } |
1494 | ||
911565b8 SK |
1495 | folio_set_compound_order(folio, 0); |
1496 | __folio_clear_head(folio); | |
944d9fec LC |
1497 | } |
1498 | ||
911565b8 | 1499 | static void destroy_compound_hugetlb_folio_for_demote(struct folio *folio, |
8531fc6f MK |
1500 | unsigned int order) |
1501 | { | |
911565b8 | 1502 | __destroy_compound_gigantic_folio(folio, order, true); |
8531fc6f MK |
1503 | } |
1504 | ||
1505 | #ifdef CONFIG_ARCH_HAS_GIGANTIC_PAGE | |
911565b8 | 1506 | static void destroy_compound_gigantic_folio(struct folio *folio, |
34d9e35b MK |
1507 | unsigned int order) |
1508 | { | |
911565b8 | 1509 | __destroy_compound_gigantic_folio(folio, order, false); |
34d9e35b MK |
1510 | } |
1511 | ||
7f325a8d | 1512 | static void free_gigantic_folio(struct folio *folio, unsigned int order) |
944d9fec | 1513 | { |
cf11e85f RG |
1514 | /* |
1515 | * If the page isn't allocated using the cma allocator, | |
1516 | * cma_release() returns false. | |
1517 | */ | |
dbda8fea | 1518 | #ifdef CONFIG_CMA |
7f325a8d SK |
1519 | int nid = folio_nid(folio); |
1520 | ||
1521 | if (cma_release(hugetlb_cma[nid], &folio->page, 1 << order)) | |
cf11e85f | 1522 | return; |
dbda8fea | 1523 | #endif |
cf11e85f | 1524 | |
7f325a8d | 1525 | free_contig_range(folio_pfn(folio), 1 << order); |
944d9fec LC |
1526 | } |
1527 | ||
4eb0716e | 1528 | #ifdef CONFIG_CONTIG_ALLOC |
19fc1a7e | 1529 | static struct folio *alloc_gigantic_folio(struct hstate *h, gfp_t gfp_mask, |
d9cc948f | 1530 | int nid, nodemask_t *nodemask) |
944d9fec | 1531 | { |
19fc1a7e | 1532 | struct page *page; |
04adbc3f | 1533 | unsigned long nr_pages = pages_per_huge_page(h); |
953f064a LX |
1534 | if (nid == NUMA_NO_NODE) |
1535 | nid = numa_mem_id(); | |
944d9fec | 1536 | |
dbda8fea BS |
1537 | #ifdef CONFIG_CMA |
1538 | { | |
cf11e85f RG |
1539 | int node; |
1540 | ||
953f064a LX |
1541 | if (hugetlb_cma[nid]) { |
1542 | page = cma_alloc(hugetlb_cma[nid], nr_pages, | |
1543 | huge_page_order(h), true); | |
cf11e85f | 1544 | if (page) |
19fc1a7e | 1545 | return page_folio(page); |
cf11e85f | 1546 | } |
953f064a LX |
1547 | |
1548 | if (!(gfp_mask & __GFP_THISNODE)) { | |
1549 | for_each_node_mask(node, *nodemask) { | |
1550 | if (node == nid || !hugetlb_cma[node]) | |
1551 | continue; | |
1552 | ||
1553 | page = cma_alloc(hugetlb_cma[node], nr_pages, | |
1554 | huge_page_order(h), true); | |
1555 | if (page) | |
19fc1a7e | 1556 | return page_folio(page); |
953f064a LX |
1557 | } |
1558 | } | |
cf11e85f | 1559 | } |
dbda8fea | 1560 | #endif |
cf11e85f | 1561 | |
19fc1a7e SK |
1562 | page = alloc_contig_pages(nr_pages, gfp_mask, nid, nodemask); |
1563 | return page ? page_folio(page) : NULL; | |
944d9fec LC |
1564 | } |
1565 | ||
4eb0716e | 1566 | #else /* !CONFIG_CONTIG_ALLOC */ |
19fc1a7e | 1567 | static struct folio *alloc_gigantic_folio(struct hstate *h, gfp_t gfp_mask, |
4eb0716e AG |
1568 | int nid, nodemask_t *nodemask) |
1569 | { | |
1570 | return NULL; | |
1571 | } | |
1572 | #endif /* CONFIG_CONTIG_ALLOC */ | |
944d9fec | 1573 | |
e1073d1e | 1574 | #else /* !CONFIG_ARCH_HAS_GIGANTIC_PAGE */ |
19fc1a7e | 1575 | static struct folio *alloc_gigantic_folio(struct hstate *h, gfp_t gfp_mask, |
4eb0716e AG |
1576 | int nid, nodemask_t *nodemask) |
1577 | { | |
1578 | return NULL; | |
1579 | } | |
7f325a8d SK |
1580 | static inline void free_gigantic_folio(struct folio *folio, |
1581 | unsigned int order) { } | |
911565b8 | 1582 | static inline void destroy_compound_gigantic_folio(struct folio *folio, |
d00181b9 | 1583 | unsigned int order) { } |
944d9fec LC |
1584 | #endif |
1585 | ||
6eb4e88a | 1586 | /* |
cfd5082b | 1587 | * Remove hugetlb folio from lists, and update dtor so that the folio appears |
34d9e35b MK |
1588 | * as just a compound page. |
1589 | * | |
cfd5082b | 1590 | * A reference is held on the folio, except in the case of demote. |
6eb4e88a MK |
1591 | * |
1592 | * Must be called with hugetlb lock held. | |
1593 | */ | |
cfd5082b | 1594 | static void __remove_hugetlb_folio(struct hstate *h, struct folio *folio, |
34d9e35b MK |
1595 | bool adjust_surplus, |
1596 | bool demote) | |
6eb4e88a | 1597 | { |
cfd5082b | 1598 | int nid = folio_nid(folio); |
6eb4e88a | 1599 | |
f074732d SK |
1600 | VM_BUG_ON_FOLIO(hugetlb_cgroup_from_folio(folio), folio); |
1601 | VM_BUG_ON_FOLIO(hugetlb_cgroup_from_folio_rsvd(folio), folio); | |
6eb4e88a | 1602 | |
9487ca60 | 1603 | lockdep_assert_held(&hugetlb_lock); |
6eb4e88a MK |
1604 | if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) |
1605 | return; | |
1606 | ||
cfd5082b | 1607 | list_del(&folio->lru); |
6eb4e88a | 1608 | |
cfd5082b | 1609 | if (folio_test_hugetlb_freed(folio)) { |
6eb4e88a MK |
1610 | h->free_huge_pages--; |
1611 | h->free_huge_pages_node[nid]--; | |
1612 | } | |
1613 | if (adjust_surplus) { | |
1614 | h->surplus_huge_pages--; | |
1615 | h->surplus_huge_pages_node[nid]--; | |
1616 | } | |
1617 | ||
e32d20c0 MK |
1618 | /* |
1619 | * Very subtle | |
1620 | * | |
1621 | * For non-gigantic pages set the destructor to the normal compound | |
1622 | * page dtor. This is needed in case someone takes an additional | |
1623 | * temporary ref to the page, and freeing is delayed until they drop | |
1624 | * their reference. | |
1625 | * | |
1626 | * For gigantic pages set the destructor to the null dtor. This | |
1627 | * destructor will never be called. Before freeing the gigantic | |
911565b8 SK |
1628 | * page destroy_compound_gigantic_folio will turn the folio into a |
1629 | * simple group of pages. After this the destructor does not | |
e32d20c0 MK |
1630 | * apply. |
1631 | * | |
1632 | * This handles the case where more than one ref is held when and | |
d6ef19e2 | 1633 | * after update_and_free_hugetlb_folio is called. |
34d9e35b MK |
1634 | * |
1635 | * In the case of demote we do not ref count the page as it will soon | |
1636 | * be turned into a page of smaller size. | |
e32d20c0 | 1637 | */ |
34d9e35b | 1638 | if (!demote) |
cfd5082b | 1639 | folio_ref_unfreeze(folio, 1); |
e32d20c0 | 1640 | if (hstate_is_gigantic(h)) |
cfd5082b | 1641 | folio_set_compound_dtor(folio, NULL_COMPOUND_DTOR); |
e32d20c0 | 1642 | else |
cfd5082b | 1643 | folio_set_compound_dtor(folio, COMPOUND_PAGE_DTOR); |
6eb4e88a MK |
1644 | |
1645 | h->nr_huge_pages--; | |
1646 | h->nr_huge_pages_node[nid]--; | |
1647 | } | |
1648 | ||
cfd5082b | 1649 | static void remove_hugetlb_folio(struct hstate *h, struct folio *folio, |
34d9e35b MK |
1650 | bool adjust_surplus) |
1651 | { | |
cfd5082b | 1652 | __remove_hugetlb_folio(h, folio, adjust_surplus, false); |
34d9e35b MK |
1653 | } |
1654 | ||
cfd5082b | 1655 | static void remove_hugetlb_folio_for_demote(struct hstate *h, struct folio *folio, |
8531fc6f MK |
1656 | bool adjust_surplus) |
1657 | { | |
cfd5082b | 1658 | __remove_hugetlb_folio(h, folio, adjust_surplus, true); |
8531fc6f MK |
1659 | } |
1660 | ||
2f6c57d6 | 1661 | static void add_hugetlb_folio(struct hstate *h, struct folio *folio, |
ad2fa371 MS |
1662 | bool adjust_surplus) |
1663 | { | |
1664 | int zeroed; | |
2f6c57d6 | 1665 | int nid = folio_nid(folio); |
ad2fa371 | 1666 | |
2f6c57d6 | 1667 | VM_BUG_ON_FOLIO(!folio_test_hugetlb_vmemmap_optimized(folio), folio); |
ad2fa371 MS |
1668 | |
1669 | lockdep_assert_held(&hugetlb_lock); | |
1670 | ||
2f6c57d6 | 1671 | INIT_LIST_HEAD(&folio->lru); |
ad2fa371 MS |
1672 | h->nr_huge_pages++; |
1673 | h->nr_huge_pages_node[nid]++; | |
1674 | ||
1675 | if (adjust_surplus) { | |
1676 | h->surplus_huge_pages++; | |
1677 | h->surplus_huge_pages_node[nid]++; | |
1678 | } | |
1679 | ||
2f6c57d6 SK |
1680 | folio_set_compound_dtor(folio, HUGETLB_PAGE_DTOR); |
1681 | folio_change_private(folio, NULL); | |
a9e1eab2 | 1682 | /* |
2f6c57d6 SK |
1683 | * We have to set hugetlb_vmemmap_optimized again as above |
1684 | * folio_change_private(folio, NULL) cleared it. | |
a9e1eab2 | 1685 | */ |
2f6c57d6 | 1686 | folio_set_hugetlb_vmemmap_optimized(folio); |
ad2fa371 MS |
1687 | |
1688 | /* | |
2f6c57d6 | 1689 | * This folio is about to be managed by the hugetlb allocator and |
b65a4eda MK |
1690 | * should have no users. Drop our reference, and check for others |
1691 | * just in case. | |
ad2fa371 | 1692 | */ |
2f6c57d6 SK |
1693 | zeroed = folio_put_testzero(folio); |
1694 | if (unlikely(!zeroed)) | |
b65a4eda MK |
1695 | /* |
1696 | * It is VERY unlikely soneone else has taken a ref on | |
1697 | * the page. In this case, we simply return as the | |
1698 | * hugetlb destructor (free_huge_page) will be called | |
1699 | * when this other ref is dropped. | |
1700 | */ | |
1701 | return; | |
1702 | ||
2f6c57d6 | 1703 | arch_clear_hugepage_flags(&folio->page); |
240d67a8 | 1704 | enqueue_hugetlb_folio(h, folio); |
ad2fa371 MS |
1705 | } |
1706 | ||
b65d4adb | 1707 | static void __update_and_free_page(struct hstate *h, struct page *page) |
6af2acb6 AL |
1708 | { |
1709 | int i; | |
911565b8 | 1710 | struct folio *folio = page_folio(page); |
14455eab | 1711 | struct page *subpage; |
a5516438 | 1712 | |
4eb0716e | 1713 | if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) |
944d9fec | 1714 | return; |
18229df5 | 1715 | |
161df60e NH |
1716 | /* |
1717 | * If we don't know which subpages are hwpoisoned, we can't free | |
1718 | * the hugepage, so it's leaked intentionally. | |
1719 | */ | |
7f325a8d | 1720 | if (folio_test_hugetlb_raw_hwp_unreliable(folio)) |
161df60e NH |
1721 | return; |
1722 | ||
6213834c | 1723 | if (hugetlb_vmemmap_restore(h, page)) { |
ad2fa371 MS |
1724 | spin_lock_irq(&hugetlb_lock); |
1725 | /* | |
1726 | * If we cannot allocate vmemmap pages, just refuse to free the | |
1727 | * page and put the page back on the hugetlb free list and treat | |
1728 | * as a surplus page. | |
1729 | */ | |
7f325a8d | 1730 | add_hugetlb_folio(h, folio, true); |
ad2fa371 MS |
1731 | spin_unlock_irq(&hugetlb_lock); |
1732 | return; | |
1733 | } | |
1734 | ||
161df60e NH |
1735 | /* |
1736 | * Move PageHWPoison flag from head page to the raw error pages, | |
1737 | * which makes any healthy subpages reusable. | |
1738 | */ | |
911565b8 SK |
1739 | if (unlikely(folio_test_hwpoison(folio))) |
1740 | hugetlb_clear_page_hwpoison(&folio->page); | |
161df60e | 1741 | |
14455eab | 1742 | for (i = 0; i < pages_per_huge_page(h); i++) { |
7f325a8d | 1743 | subpage = folio_page(folio, i); |
dbfee5ae | 1744 | subpage->flags &= ~(1 << PG_locked | 1 << PG_error | |
32f84528 | 1745 | 1 << PG_referenced | 1 << PG_dirty | |
a7407a27 LC |
1746 | 1 << PG_active | 1 << PG_private | |
1747 | 1 << PG_writeback); | |
6af2acb6 | 1748 | } |
a01f4390 MK |
1749 | |
1750 | /* | |
1751 | * Non-gigantic pages demoted from CMA allocated gigantic pages | |
7f325a8d | 1752 | * need to be given back to CMA in free_gigantic_folio. |
a01f4390 MK |
1753 | */ |
1754 | if (hstate_is_gigantic(h) || | |
2f6c57d6 | 1755 | hugetlb_cma_folio(folio, huge_page_order(h))) { |
911565b8 | 1756 | destroy_compound_gigantic_folio(folio, huge_page_order(h)); |
7f325a8d | 1757 | free_gigantic_folio(folio, huge_page_order(h)); |
944d9fec | 1758 | } else { |
944d9fec LC |
1759 | __free_pages(page, huge_page_order(h)); |
1760 | } | |
6af2acb6 AL |
1761 | } |
1762 | ||
b65d4adb | 1763 | /* |
d6ef19e2 | 1764 | * As update_and_free_hugetlb_folio() can be called under any context, so we cannot |
b65d4adb MS |
1765 | * use GFP_KERNEL to allocate vmemmap pages. However, we can defer the |
1766 | * actual freeing in a workqueue to prevent from using GFP_ATOMIC to allocate | |
1767 | * the vmemmap pages. | |
1768 | * | |
1769 | * free_hpage_workfn() locklessly retrieves the linked list of pages to be | |
1770 | * freed and frees them one-by-one. As the page->mapping pointer is going | |
1771 | * to be cleared in free_hpage_workfn() anyway, it is reused as the llist_node | |
1772 | * structure of a lockless linked list of huge pages to be freed. | |
1773 | */ | |
1774 | static LLIST_HEAD(hpage_freelist); | |
1775 | ||
1776 | static void free_hpage_workfn(struct work_struct *work) | |
1777 | { | |
1778 | struct llist_node *node; | |
1779 | ||
1780 | node = llist_del_all(&hpage_freelist); | |
1781 | ||
1782 | while (node) { | |
1783 | struct page *page; | |
1784 | struct hstate *h; | |
1785 | ||
1786 | page = container_of((struct address_space **)node, | |
1787 | struct page, mapping); | |
1788 | node = node->next; | |
1789 | page->mapping = NULL; | |
1790 | /* | |
1791 | * The VM_BUG_ON_PAGE(!PageHuge(page), page) in page_hstate() | |
1792 | * is going to trigger because a previous call to | |
cfd5082b SK |
1793 | * remove_hugetlb_folio() will call folio_set_compound_dtor |
1794 | * (folio, NULL_COMPOUND_DTOR), so do not use page_hstate() | |
1795 | * directly. | |
b65d4adb MS |
1796 | */ |
1797 | h = size_to_hstate(page_size(page)); | |
1798 | ||
1799 | __update_and_free_page(h, page); | |
1800 | ||
1801 | cond_resched(); | |
1802 | } | |
1803 | } | |
1804 | static DECLARE_WORK(free_hpage_work, free_hpage_workfn); | |
1805 | ||
1806 | static inline void flush_free_hpage_work(struct hstate *h) | |
1807 | { | |
6213834c | 1808 | if (hugetlb_vmemmap_optimizable(h)) |
b65d4adb MS |
1809 | flush_work(&free_hpage_work); |
1810 | } | |
1811 | ||
d6ef19e2 | 1812 | static void update_and_free_hugetlb_folio(struct hstate *h, struct folio *folio, |
b65d4adb MS |
1813 | bool atomic) |
1814 | { | |
d6ef19e2 SK |
1815 | if (!folio_test_hugetlb_vmemmap_optimized(folio) || !atomic) { |
1816 | __update_and_free_page(h, &folio->page); | |
b65d4adb MS |
1817 | return; |
1818 | } | |
1819 | ||
1820 | /* | |
1821 | * Defer freeing to avoid using GFP_ATOMIC to allocate vmemmap pages. | |
1822 | * | |
1823 | * Only call schedule_work() if hpage_freelist is previously | |
1824 | * empty. Otherwise, schedule_work() had been called but the workfn | |
1825 | * hasn't retrieved the list yet. | |
1826 | */ | |
d6ef19e2 | 1827 | if (llist_add((struct llist_node *)&folio->mapping, &hpage_freelist)) |
b65d4adb MS |
1828 | schedule_work(&free_hpage_work); |
1829 | } | |
1830 | ||
10c6ec49 MK |
1831 | static void update_and_free_pages_bulk(struct hstate *h, struct list_head *list) |
1832 | { | |
1833 | struct page *page, *t_page; | |
d6ef19e2 | 1834 | struct folio *folio; |
10c6ec49 MK |
1835 | |
1836 | list_for_each_entry_safe(page, t_page, list, lru) { | |
d6ef19e2 SK |
1837 | folio = page_folio(page); |
1838 | update_and_free_hugetlb_folio(h, folio, false); | |
10c6ec49 MK |
1839 | cond_resched(); |
1840 | } | |
1841 | } | |
1842 | ||
e5ff2159 AK |
1843 | struct hstate *size_to_hstate(unsigned long size) |
1844 | { | |
1845 | struct hstate *h; | |
1846 | ||
1847 | for_each_hstate(h) { | |
1848 | if (huge_page_size(h) == size) | |
1849 | return h; | |
1850 | } | |
1851 | return NULL; | |
1852 | } | |
1853 | ||
db71ef79 | 1854 | void free_huge_page(struct page *page) |
27a85ef1 | 1855 | { |
a5516438 AK |
1856 | /* |
1857 | * Can't pass hstate in here because it is called from the | |
1858 | * compound page destructor. | |
1859 | */ | |
0356c4b9 SK |
1860 | struct folio *folio = page_folio(page); |
1861 | struct hstate *h = folio_hstate(folio); | |
1862 | int nid = folio_nid(folio); | |
1863 | struct hugepage_subpool *spool = hugetlb_folio_subpool(folio); | |
07443a85 | 1864 | bool restore_reserve; |
db71ef79 | 1865 | unsigned long flags; |
27a85ef1 | 1866 | |
0356c4b9 SK |
1867 | VM_BUG_ON_FOLIO(folio_ref_count(folio), folio); |
1868 | VM_BUG_ON_FOLIO(folio_mapcount(folio), folio); | |
8ace22bc | 1869 | |
0356c4b9 SK |
1870 | hugetlb_set_folio_subpool(folio, NULL); |
1871 | if (folio_test_anon(folio)) | |
1872 | __ClearPageAnonExclusive(&folio->page); | |
1873 | folio->mapping = NULL; | |
1874 | restore_reserve = folio_test_hugetlb_restore_reserve(folio); | |
1875 | folio_clear_hugetlb_restore_reserve(folio); | |
27a85ef1 | 1876 | |
1c5ecae3 | 1877 | /* |
d6995da3 | 1878 | * If HPageRestoreReserve was set on page, page allocation consumed a |
0919e1b6 MK |
1879 | * reservation. If the page was associated with a subpool, there |
1880 | * would have been a page reserved in the subpool before allocation | |
1881 | * via hugepage_subpool_get_pages(). Since we are 'restoring' the | |
6c26d310 | 1882 | * reservation, do not call hugepage_subpool_put_pages() as this will |
0919e1b6 | 1883 | * remove the reserved page from the subpool. |
1c5ecae3 | 1884 | */ |
0919e1b6 MK |
1885 | if (!restore_reserve) { |
1886 | /* | |
1887 | * A return code of zero implies that the subpool will be | |
1888 | * under its minimum size if the reservation is not restored | |
1889 | * after page is free. Therefore, force restore_reserve | |
1890 | * operation. | |
1891 | */ | |
1892 | if (hugepage_subpool_put_pages(spool, 1) == 0) | |
1893 | restore_reserve = true; | |
1894 | } | |
1c5ecae3 | 1895 | |
db71ef79 | 1896 | spin_lock_irqsave(&hugetlb_lock, flags); |
0356c4b9 | 1897 | folio_clear_hugetlb_migratable(folio); |
d4ab0316 SK |
1898 | hugetlb_cgroup_uncharge_folio(hstate_index(h), |
1899 | pages_per_huge_page(h), folio); | |
1900 | hugetlb_cgroup_uncharge_folio_rsvd(hstate_index(h), | |
1901 | pages_per_huge_page(h), folio); | |
07443a85 JK |
1902 | if (restore_reserve) |
1903 | h->resv_huge_pages++; | |
1904 | ||
0356c4b9 | 1905 | if (folio_test_hugetlb_temporary(folio)) { |
cfd5082b | 1906 | remove_hugetlb_folio(h, folio, false); |
db71ef79 | 1907 | spin_unlock_irqrestore(&hugetlb_lock, flags); |
d6ef19e2 | 1908 | update_and_free_hugetlb_folio(h, folio, true); |
ab5ac90a | 1909 | } else if (h->surplus_huge_pages_node[nid]) { |
0edaecfa | 1910 | /* remove the page from active list */ |
cfd5082b | 1911 | remove_hugetlb_folio(h, folio, true); |
db71ef79 | 1912 | spin_unlock_irqrestore(&hugetlb_lock, flags); |
d6ef19e2 | 1913 | update_and_free_hugetlb_folio(h, folio, true); |
7893d1d5 | 1914 | } else { |
5d3a551c | 1915 | arch_clear_hugepage_flags(page); |
240d67a8 | 1916 | enqueue_hugetlb_folio(h, folio); |
db71ef79 | 1917 | spin_unlock_irqrestore(&hugetlb_lock, flags); |
c77c0a8a | 1918 | } |
c77c0a8a WL |
1919 | } |
1920 | ||
d3d99fcc OS |
1921 | /* |
1922 | * Must be called with the hugetlb lock held | |
1923 | */ | |
1924 | static void __prep_account_new_huge_page(struct hstate *h, int nid) | |
1925 | { | |
1926 | lockdep_assert_held(&hugetlb_lock); | |
1927 | h->nr_huge_pages++; | |
1928 | h->nr_huge_pages_node[nid]++; | |
1929 | } | |
1930 | ||
de656ed3 | 1931 | static void __prep_new_hugetlb_folio(struct hstate *h, struct folio *folio) |
b7ba30c6 | 1932 | { |
de656ed3 SK |
1933 | hugetlb_vmemmap_optimize(h, &folio->page); |
1934 | INIT_LIST_HEAD(&folio->lru); | |
9fd33058 | 1935 | folio_set_compound_dtor(folio, HUGETLB_PAGE_DTOR); |
de656ed3 SK |
1936 | hugetlb_set_folio_subpool(folio, NULL); |
1937 | set_hugetlb_cgroup(folio, NULL); | |
1938 | set_hugetlb_cgroup_rsvd(folio, NULL); | |
d3d99fcc OS |
1939 | } |
1940 | ||
d1c60955 | 1941 | static void prep_new_hugetlb_folio(struct hstate *h, struct folio *folio, int nid) |
d3d99fcc | 1942 | { |
de656ed3 | 1943 | __prep_new_hugetlb_folio(h, folio); |
db71ef79 | 1944 | spin_lock_irq(&hugetlb_lock); |
d3d99fcc | 1945 | __prep_account_new_huge_page(h, nid); |
db71ef79 | 1946 | spin_unlock_irq(&hugetlb_lock); |
b7ba30c6 AK |
1947 | } |
1948 | ||
d1c60955 SK |
1949 | static bool __prep_compound_gigantic_folio(struct folio *folio, |
1950 | unsigned int order, bool demote) | |
20a0307c | 1951 | { |
7118fc29 | 1952 | int i, j; |
20a0307c | 1953 | int nr_pages = 1 << order; |
14455eab | 1954 | struct page *p; |
20a0307c | 1955 | |
d1c60955 SK |
1956 | __folio_clear_reserved(folio); |
1957 | __folio_set_head(folio); | |
c45bc55a SK |
1958 | /* we rely on prep_new_hugetlb_folio to set the destructor */ |
1959 | folio_set_compound_order(folio, order); | |
2b21624f | 1960 | for (i = 0; i < nr_pages; i++) { |
d1c60955 | 1961 | p = folio_page(folio, i); |
14455eab | 1962 | |
ef5a22be AA |
1963 | /* |
1964 | * For gigantic hugepages allocated through bootmem at | |
1965 | * boot, it's safer to be consistent with the not-gigantic | |
1966 | * hugepages and clear the PG_reserved bit from all tail pages | |
7c8de358 | 1967 | * too. Otherwise drivers using get_user_pages() to access tail |
ef5a22be AA |
1968 | * pages may get the reference counting wrong if they see |
1969 | * PG_reserved set on a tail page (despite the head page not | |
1970 | * having PG_reserved set). Enforcing this consistency between | |
1971 | * head and tail pages allows drivers to optimize away a check | |
1972 | * on the head page when they need know if put_page() is needed | |
1973 | * after get_user_pages(). | |
1974 | */ | |
7fb0728a MK |
1975 | if (i != 0) /* head page cleared above */ |
1976 | __ClearPageReserved(p); | |
7118fc29 MK |
1977 | /* |
1978 | * Subtle and very unlikely | |
1979 | * | |
1980 | * Gigantic 'page allocators' such as memblock or cma will | |
1981 | * return a set of pages with each page ref counted. We need | |
1982 | * to turn this set of pages into a compound page with tail | |
1983 | * page ref counts set to zero. Code such as speculative page | |
1984 | * cache adding could take a ref on a 'to be' tail page. | |
1985 | * We need to respect any increased ref count, and only set | |
1986 | * the ref count to zero if count is currently 1. If count | |
416d85ed MK |
1987 | * is not 1, we return an error. An error return indicates |
1988 | * the set of pages can not be converted to a gigantic page. | |
1989 | * The caller who allocated the pages should then discard the | |
1990 | * pages using the appropriate free interface. | |
34d9e35b MK |
1991 | * |
1992 | * In the case of demote, the ref count will be zero. | |
7118fc29 | 1993 | */ |
34d9e35b MK |
1994 | if (!demote) { |
1995 | if (!page_ref_freeze(p, 1)) { | |
1996 | pr_warn("HugeTLB page can not be used due to unexpected inflated ref count\n"); | |
1997 | goto out_error; | |
1998 | } | |
1999 | } else { | |
2000 | VM_BUG_ON_PAGE(page_count(p), p); | |
7118fc29 | 2001 | } |
2b21624f | 2002 | if (i != 0) |
d1c60955 | 2003 | set_compound_head(p, &folio->page); |
20a0307c | 2004 | } |
d1c60955 SK |
2005 | atomic_set(folio_mapcount_ptr(folio), -1); |
2006 | atomic_set(folio_subpages_mapcount_ptr(folio), 0); | |
2007 | atomic_set(folio_pincount_ptr(folio), 0); | |
7118fc29 MK |
2008 | return true; |
2009 | ||
2010 | out_error: | |
2b21624f MK |
2011 | /* undo page modifications made above */ |
2012 | for (j = 0; j < i; j++) { | |
d1c60955 | 2013 | p = folio_page(folio, j); |
2b21624f MK |
2014 | if (j != 0) |
2015 | clear_compound_head(p); | |
7118fc29 MK |
2016 | set_page_refcounted(p); |
2017 | } | |
2018 | /* need to clear PG_reserved on remaining tail pages */ | |
14455eab | 2019 | for (; j < nr_pages; j++) { |
d1c60955 | 2020 | p = folio_page(folio, j); |
7118fc29 | 2021 | __ClearPageReserved(p); |
14455eab | 2022 | } |
d1c60955 SK |
2023 | folio_set_compound_order(folio, 0); |
2024 | __folio_clear_head(folio); | |
7118fc29 | 2025 | return false; |
20a0307c WF |
2026 | } |
2027 | ||
d1c60955 SK |
2028 | static bool prep_compound_gigantic_folio(struct folio *folio, |
2029 | unsigned int order) | |
34d9e35b | 2030 | { |
d1c60955 | 2031 | return __prep_compound_gigantic_folio(folio, order, false); |
34d9e35b MK |
2032 | } |
2033 | ||
d1c60955 | 2034 | static bool prep_compound_gigantic_folio_for_demote(struct folio *folio, |
8531fc6f MK |
2035 | unsigned int order) |
2036 | { | |
d1c60955 | 2037 | return __prep_compound_gigantic_folio(folio, order, true); |
8531fc6f MK |
2038 | } |
2039 | ||
7795912c AM |
2040 | /* |
2041 | * PageHuge() only returns true for hugetlbfs pages, but not for normal or | |
2042 | * transparent huge pages. See the PageTransHuge() documentation for more | |
2043 | * details. | |
2044 | */ | |
20a0307c WF |
2045 | int PageHuge(struct page *page) |
2046 | { | |
20a0307c WF |
2047 | if (!PageCompound(page)) |
2048 | return 0; | |
2049 | ||
2050 | page = compound_head(page); | |
f1e61557 | 2051 | return page[1].compound_dtor == HUGETLB_PAGE_DTOR; |
20a0307c | 2052 | } |
43131e14 NH |
2053 | EXPORT_SYMBOL_GPL(PageHuge); |
2054 | ||
27c73ae7 AA |
2055 | /* |
2056 | * PageHeadHuge() only returns true for hugetlbfs head page, but not for | |
2057 | * normal or transparent huge pages. | |
2058 | */ | |
2059 | int PageHeadHuge(struct page *page_head) | |
2060 | { | |
27c73ae7 AA |
2061 | if (!PageHead(page_head)) |
2062 | return 0; | |
2063 | ||
d4af73e3 | 2064 | return page_head[1].compound_dtor == HUGETLB_PAGE_DTOR; |
27c73ae7 | 2065 | } |
4e936ecc | 2066 | EXPORT_SYMBOL_GPL(PageHeadHuge); |
27c73ae7 | 2067 | |
c0d0381a MK |
2068 | /* |
2069 | * Find and lock address space (mapping) in write mode. | |
2070 | * | |
336bf30e MK |
2071 | * Upon entry, the page is locked which means that page_mapping() is |
2072 | * stable. Due to locking order, we can only trylock_write. If we can | |
2073 | * not get the lock, simply return NULL to caller. | |
c0d0381a MK |
2074 | */ |
2075 | struct address_space *hugetlb_page_mapping_lock_write(struct page *hpage) | |
2076 | { | |
336bf30e | 2077 | struct address_space *mapping = page_mapping(hpage); |
c0d0381a | 2078 | |
c0d0381a MK |
2079 | if (!mapping) |
2080 | return mapping; | |
2081 | ||
c0d0381a MK |
2082 | if (i_mmap_trylock_write(mapping)) |
2083 | return mapping; | |
2084 | ||
336bf30e | 2085 | return NULL; |
c0d0381a MK |
2086 | } |
2087 | ||
fe19bd3d | 2088 | pgoff_t hugetlb_basepage_index(struct page *page) |
13d60f4b ZY |
2089 | { |
2090 | struct page *page_head = compound_head(page); | |
2091 | pgoff_t index = page_index(page_head); | |
2092 | unsigned long compound_idx; | |
2093 | ||
13d60f4b ZY |
2094 | if (compound_order(page_head) >= MAX_ORDER) |
2095 | compound_idx = page_to_pfn(page) - page_to_pfn(page_head); | |
2096 | else | |
2097 | compound_idx = page - page_head; | |
2098 | ||
2099 | return (index << compound_order(page_head)) + compound_idx; | |
2100 | } | |
2101 | ||
19fc1a7e | 2102 | static struct folio *alloc_buddy_hugetlb_folio(struct hstate *h, |
f60858f9 MK |
2103 | gfp_t gfp_mask, int nid, nodemask_t *nmask, |
2104 | nodemask_t *node_alloc_noretry) | |
1da177e4 | 2105 | { |
af0fb9df | 2106 | int order = huge_page_order(h); |
1da177e4 | 2107 | struct page *page; |
f60858f9 | 2108 | bool alloc_try_hard = true; |
2b21624f | 2109 | bool retry = true; |
f96efd58 | 2110 | |
f60858f9 MK |
2111 | /* |
2112 | * By default we always try hard to allocate the page with | |
2113 | * __GFP_RETRY_MAYFAIL flag. However, if we are allocating pages in | |
2114 | * a loop (to adjust global huge page counts) and previous allocation | |
2115 | * failed, do not continue to try hard on the same node. Use the | |
2116 | * node_alloc_noretry bitmap to manage this state information. | |
2117 | */ | |
2118 | if (node_alloc_noretry && node_isset(nid, *node_alloc_noretry)) | |
2119 | alloc_try_hard = false; | |
2120 | gfp_mask |= __GFP_COMP|__GFP_NOWARN; | |
2121 | if (alloc_try_hard) | |
2122 | gfp_mask |= __GFP_RETRY_MAYFAIL; | |
af0fb9df MH |
2123 | if (nid == NUMA_NO_NODE) |
2124 | nid = numa_mem_id(); | |
2b21624f | 2125 | retry: |
84172f4b | 2126 | page = __alloc_pages(gfp_mask, order, nid, nmask); |
2b21624f MK |
2127 | |
2128 | /* Freeze head page */ | |
2129 | if (page && !page_ref_freeze(page, 1)) { | |
2130 | __free_pages(page, order); | |
2131 | if (retry) { /* retry once */ | |
2132 | retry = false; | |
2133 | goto retry; | |
2134 | } | |
2135 | /* WOW! twice in a row. */ | |
2136 | pr_warn("HugeTLB head page unexpected inflated ref count\n"); | |
2137 | page = NULL; | |
2138 | } | |
2139 | ||
f60858f9 MK |
2140 | /* |
2141 | * If we did not specify __GFP_RETRY_MAYFAIL, but still got a page this | |
2142 | * indicates an overall state change. Clear bit so that we resume | |
2143 | * normal 'try hard' allocations. | |
2144 | */ | |
2145 | if (node_alloc_noretry && page && !alloc_try_hard) | |
2146 | node_clear(nid, *node_alloc_noretry); | |
2147 | ||
2148 | /* | |
2149 | * If we tried hard to get a page but failed, set bit so that | |
2150 | * subsequent attempts will not try as hard until there is an | |
2151 | * overall state change. | |
2152 | */ | |
2153 | if (node_alloc_noretry && !page && alloc_try_hard) | |
2154 | node_set(nid, *node_alloc_noretry); | |
2155 | ||
19fc1a7e SK |
2156 | if (!page) { |
2157 | __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); | |
2158 | return NULL; | |
2159 | } | |
2160 | ||
2161 | __count_vm_event(HTLB_BUDDY_PGALLOC); | |
2162 | return page_folio(page); | |
63b4613c NA |
2163 | } |
2164 | ||
0c397dae MH |
2165 | /* |
2166 | * Common helper to allocate a fresh hugetlb page. All specific allocators | |
2167 | * should use this function to get new hugetlb pages | |
2b21624f MK |
2168 | * |
2169 | * Note that returned page is 'frozen': ref count of head page and all tail | |
2170 | * pages is zero. | |
0c397dae | 2171 | */ |
19fc1a7e | 2172 | static struct folio *alloc_fresh_hugetlb_folio(struct hstate *h, |
f60858f9 MK |
2173 | gfp_t gfp_mask, int nid, nodemask_t *nmask, |
2174 | nodemask_t *node_alloc_noretry) | |
0c397dae | 2175 | { |
7f325a8d | 2176 | struct folio *folio; |
7118fc29 | 2177 | bool retry = false; |
0c397dae | 2178 | |
7118fc29 | 2179 | retry: |
0c397dae | 2180 | if (hstate_is_gigantic(h)) |
19fc1a7e | 2181 | folio = alloc_gigantic_folio(h, gfp_mask, nid, nmask); |
0c397dae | 2182 | else |
19fc1a7e | 2183 | folio = alloc_buddy_hugetlb_folio(h, gfp_mask, |
f60858f9 | 2184 | nid, nmask, node_alloc_noretry); |
19fc1a7e | 2185 | if (!folio) |
0c397dae | 2186 | return NULL; |
7118fc29 | 2187 | if (hstate_is_gigantic(h)) { |
d1c60955 | 2188 | if (!prep_compound_gigantic_folio(folio, huge_page_order(h))) { |
7118fc29 MK |
2189 | /* |
2190 | * Rare failure to convert pages to compound page. | |
2191 | * Free pages and try again - ONCE! | |
2192 | */ | |
7f325a8d | 2193 | free_gigantic_folio(folio, huge_page_order(h)); |
7118fc29 MK |
2194 | if (!retry) { |
2195 | retry = true; | |
2196 | goto retry; | |
2197 | } | |
7118fc29 MK |
2198 | return NULL; |
2199 | } | |
2200 | } | |
d1c60955 | 2201 | prep_new_hugetlb_folio(h, folio, folio_nid(folio)); |
0c397dae | 2202 | |
19fc1a7e | 2203 | return folio; |
0c397dae MH |
2204 | } |
2205 | ||
af0fb9df MH |
2206 | /* |
2207 | * Allocates a fresh page to the hugetlb allocator pool in the node interleaved | |
2208 | * manner. | |
2209 | */ | |
f60858f9 MK |
2210 | static int alloc_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed, |
2211 | nodemask_t *node_alloc_noretry) | |
b2261026 | 2212 | { |
19fc1a7e | 2213 | struct folio *folio; |
b2261026 | 2214 | int nr_nodes, node; |
af0fb9df | 2215 | gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE; |
b2261026 JK |
2216 | |
2217 | for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { | |
19fc1a7e SK |
2218 | folio = alloc_fresh_hugetlb_folio(h, gfp_mask, node, |
2219 | nodes_allowed, node_alloc_noretry); | |
2220 | if (folio) { | |
2221 | free_huge_page(&folio->page); /* free it into the hugepage allocator */ | |
2222 | return 1; | |
2223 | } | |
b2261026 JK |
2224 | } |
2225 | ||
19fc1a7e | 2226 | return 0; |
b2261026 JK |
2227 | } |
2228 | ||
e8c5c824 | 2229 | /* |
10c6ec49 MK |
2230 | * Remove huge page from pool from next node to free. Attempt to keep |
2231 | * persistent huge pages more or less balanced over allowed nodes. | |
2232 | * This routine only 'removes' the hugetlb page. The caller must make | |
2233 | * an additional call to free the page to low level allocators. | |
e8c5c824 LS |
2234 | * Called with hugetlb_lock locked. |
2235 | */ | |
10c6ec49 MK |
2236 | static struct page *remove_pool_huge_page(struct hstate *h, |
2237 | nodemask_t *nodes_allowed, | |
2238 | bool acct_surplus) | |
e8c5c824 | 2239 | { |
b2261026 | 2240 | int nr_nodes, node; |
10c6ec49 | 2241 | struct page *page = NULL; |
cfd5082b | 2242 | struct folio *folio; |
e8c5c824 | 2243 | |
9487ca60 | 2244 | lockdep_assert_held(&hugetlb_lock); |
b2261026 | 2245 | for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { |
685f3457 LS |
2246 | /* |
2247 | * If we're returning unused surplus pages, only examine | |
2248 | * nodes with surplus pages. | |
2249 | */ | |
b2261026 JK |
2250 | if ((!acct_surplus || h->surplus_huge_pages_node[node]) && |
2251 | !list_empty(&h->hugepage_freelists[node])) { | |
10c6ec49 | 2252 | page = list_entry(h->hugepage_freelists[node].next, |
e8c5c824 | 2253 | struct page, lru); |
cfd5082b SK |
2254 | folio = page_folio(page); |
2255 | remove_hugetlb_folio(h, folio, acct_surplus); | |
9a76db09 | 2256 | break; |
e8c5c824 | 2257 | } |
b2261026 | 2258 | } |
e8c5c824 | 2259 | |
10c6ec49 | 2260 | return page; |
e8c5c824 LS |
2261 | } |
2262 | ||
c8721bbb NH |
2263 | /* |
2264 | * Dissolve a given free hugepage into free buddy pages. This function does | |
faf53def NH |
2265 | * nothing for in-use hugepages and non-hugepages. |
2266 | * This function returns values like below: | |
2267 | * | |
ad2fa371 MS |
2268 | * -ENOMEM: failed to allocate vmemmap pages to free the freed hugepages |
2269 | * when the system is under memory pressure and the feature of | |
2270 | * freeing unused vmemmap pages associated with each hugetlb page | |
2271 | * is enabled. | |
2272 | * -EBUSY: failed to dissolved free hugepages or the hugepage is in-use | |
2273 | * (allocated or reserved.) | |
2274 | * 0: successfully dissolved free hugepages or the page is not a | |
2275 | * hugepage (considered as already dissolved) | |
c8721bbb | 2276 | */ |
c3114a84 | 2277 | int dissolve_free_huge_page(struct page *page) |
c8721bbb | 2278 | { |
6bc9b564 | 2279 | int rc = -EBUSY; |
1a7cdab5 | 2280 | struct folio *folio = page_folio(page); |
082d5b6b | 2281 | |
7ffddd49 | 2282 | retry: |
faf53def | 2283 | /* Not to disrupt normal path by vainly holding hugetlb_lock */ |
1a7cdab5 | 2284 | if (!folio_test_hugetlb(folio)) |
faf53def NH |
2285 | return 0; |
2286 | ||
db71ef79 | 2287 | spin_lock_irq(&hugetlb_lock); |
1a7cdab5 | 2288 | if (!folio_test_hugetlb(folio)) { |
faf53def NH |
2289 | rc = 0; |
2290 | goto out; | |
2291 | } | |
2292 | ||
1a7cdab5 SK |
2293 | if (!folio_ref_count(folio)) { |
2294 | struct hstate *h = folio_hstate(folio); | |
8346d69d | 2295 | if (!available_huge_pages(h)) |
082d5b6b | 2296 | goto out; |
7ffddd49 MS |
2297 | |
2298 | /* | |
2299 | * We should make sure that the page is already on the free list | |
2300 | * when it is dissolved. | |
2301 | */ | |
1a7cdab5 | 2302 | if (unlikely(!folio_test_hugetlb_freed(folio))) { |
db71ef79 | 2303 | spin_unlock_irq(&hugetlb_lock); |
7ffddd49 MS |
2304 | cond_resched(); |
2305 | ||
2306 | /* | |
2307 | * Theoretically, we should return -EBUSY when we | |
2308 | * encounter this race. In fact, we have a chance | |
2309 | * to successfully dissolve the page if we do a | |
2310 | * retry. Because the race window is quite small. | |
2311 | * If we seize this opportunity, it is an optimization | |
2312 | * for increasing the success rate of dissolving page. | |
2313 | */ | |
2314 | goto retry; | |
2315 | } | |
2316 | ||
cfd5082b | 2317 | remove_hugetlb_folio(h, folio, false); |
c1470b33 | 2318 | h->max_huge_pages--; |
db71ef79 | 2319 | spin_unlock_irq(&hugetlb_lock); |
ad2fa371 MS |
2320 | |
2321 | /* | |
d6ef19e2 SK |
2322 | * Normally update_and_free_hugtlb_folio will allocate required vmemmmap |
2323 | * before freeing the page. update_and_free_hugtlb_folio will fail to | |
ad2fa371 MS |
2324 | * free the page if it can not allocate required vmemmap. We |
2325 | * need to adjust max_huge_pages if the page is not freed. | |
2326 | * Attempt to allocate vmemmmap here so that we can take | |
2327 | * appropriate action on failure. | |
2328 | */ | |
1a7cdab5 | 2329 | rc = hugetlb_vmemmap_restore(h, &folio->page); |
ad2fa371 | 2330 | if (!rc) { |
d6ef19e2 | 2331 | update_and_free_hugetlb_folio(h, folio, false); |
ad2fa371 MS |
2332 | } else { |
2333 | spin_lock_irq(&hugetlb_lock); | |
2f6c57d6 | 2334 | add_hugetlb_folio(h, folio, false); |
ad2fa371 MS |
2335 | h->max_huge_pages++; |
2336 | spin_unlock_irq(&hugetlb_lock); | |
2337 | } | |
2338 | ||
2339 | return rc; | |
c8721bbb | 2340 | } |
082d5b6b | 2341 | out: |
db71ef79 | 2342 | spin_unlock_irq(&hugetlb_lock); |
082d5b6b | 2343 | return rc; |
c8721bbb NH |
2344 | } |
2345 | ||
2346 | /* | |
2347 | * Dissolve free hugepages in a given pfn range. Used by memory hotplug to | |
2348 | * make specified memory blocks removable from the system. | |
2247bb33 GS |
2349 | * Note that this will dissolve a free gigantic hugepage completely, if any |
2350 | * part of it lies within the given range. | |
082d5b6b GS |
2351 | * Also note that if dissolve_free_huge_page() returns with an error, all |
2352 | * free hugepages that were dissolved before that error are lost. | |
c8721bbb | 2353 | */ |
082d5b6b | 2354 | int dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn) |
c8721bbb | 2355 | { |
c8721bbb | 2356 | unsigned long pfn; |
eb03aa00 | 2357 | struct page *page; |
082d5b6b | 2358 | int rc = 0; |
dc2628f3 MS |
2359 | unsigned int order; |
2360 | struct hstate *h; | |
c8721bbb | 2361 | |
d0177639 | 2362 | if (!hugepages_supported()) |
082d5b6b | 2363 | return rc; |
d0177639 | 2364 | |
dc2628f3 MS |
2365 | order = huge_page_order(&default_hstate); |
2366 | for_each_hstate(h) | |
2367 | order = min(order, huge_page_order(h)); | |
2368 | ||
2369 | for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << order) { | |
eb03aa00 | 2370 | page = pfn_to_page(pfn); |
faf53def NH |
2371 | rc = dissolve_free_huge_page(page); |
2372 | if (rc) | |
2373 | break; | |
eb03aa00 | 2374 | } |
082d5b6b GS |
2375 | |
2376 | return rc; | |
c8721bbb NH |
2377 | } |
2378 | ||
ab5ac90a MH |
2379 | /* |
2380 | * Allocates a fresh surplus page from the page allocator. | |
2381 | */ | |
0c397dae | 2382 | static struct page *alloc_surplus_huge_page(struct hstate *h, gfp_t gfp_mask, |
2b21624f | 2383 | int nid, nodemask_t *nmask) |
7893d1d5 | 2384 | { |
19fc1a7e | 2385 | struct folio *folio = NULL; |
7893d1d5 | 2386 | |
bae7f4ae | 2387 | if (hstate_is_gigantic(h)) |
aa888a74 AK |
2388 | return NULL; |
2389 | ||
db71ef79 | 2390 | spin_lock_irq(&hugetlb_lock); |
9980d744 MH |
2391 | if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) |
2392 | goto out_unlock; | |
db71ef79 | 2393 | spin_unlock_irq(&hugetlb_lock); |
d1c3fb1f | 2394 | |
19fc1a7e SK |
2395 | folio = alloc_fresh_hugetlb_folio(h, gfp_mask, nid, nmask, NULL); |
2396 | if (!folio) | |
0c397dae | 2397 | return NULL; |
d1c3fb1f | 2398 | |
db71ef79 | 2399 | spin_lock_irq(&hugetlb_lock); |
9980d744 MH |
2400 | /* |
2401 | * We could have raced with the pool size change. | |
2402 | * Double check that and simply deallocate the new page | |
2403 | * if we would end up overcommiting the surpluses. Abuse | |
2404 | * temporary page to workaround the nasty free_huge_page | |
2405 | * codeflow | |
2406 | */ | |
2407 | if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) { | |
19fc1a7e | 2408 | folio_set_hugetlb_temporary(folio); |
db71ef79 | 2409 | spin_unlock_irq(&hugetlb_lock); |
19fc1a7e | 2410 | free_huge_page(&folio->page); |
2bf753e6 | 2411 | return NULL; |
7893d1d5 | 2412 | } |
9980d744 | 2413 | |
b65a4eda | 2414 | h->surplus_huge_pages++; |
19fc1a7e | 2415 | h->surplus_huge_pages_node[folio_nid(folio)]++; |
b65a4eda | 2416 | |
9980d744 | 2417 | out_unlock: |
db71ef79 | 2418 | spin_unlock_irq(&hugetlb_lock); |
7893d1d5 | 2419 | |
19fc1a7e | 2420 | return &folio->page; |
7893d1d5 AL |
2421 | } |
2422 | ||
bbe88753 | 2423 | static struct page *alloc_migrate_huge_page(struct hstate *h, gfp_t gfp_mask, |
9a4e9f3b | 2424 | int nid, nodemask_t *nmask) |
ab5ac90a | 2425 | { |
19fc1a7e | 2426 | struct folio *folio; |
ab5ac90a MH |
2427 | |
2428 | if (hstate_is_gigantic(h)) | |
2429 | return NULL; | |
2430 | ||
19fc1a7e SK |
2431 | folio = alloc_fresh_hugetlb_folio(h, gfp_mask, nid, nmask, NULL); |
2432 | if (!folio) | |
ab5ac90a MH |
2433 | return NULL; |
2434 | ||
2b21624f | 2435 | /* fresh huge pages are frozen */ |
19fc1a7e | 2436 | folio_ref_unfreeze(folio, 1); |
ab5ac90a MH |
2437 | /* |
2438 | * We do not account these pages as surplus because they are only | |
2439 | * temporary and will be released properly on the last reference | |
2440 | */ | |
19fc1a7e | 2441 | folio_set_hugetlb_temporary(folio); |
ab5ac90a | 2442 | |
19fc1a7e | 2443 | return &folio->page; |
ab5ac90a MH |
2444 | } |
2445 | ||
099730d6 DH |
2446 | /* |
2447 | * Use the VMA's mpolicy to allocate a huge page from the buddy. | |
2448 | */ | |
e0ec90ee | 2449 | static |
0c397dae | 2450 | struct page *alloc_buddy_huge_page_with_mpol(struct hstate *h, |
099730d6 DH |
2451 | struct vm_area_struct *vma, unsigned long addr) |
2452 | { | |
cfcaa66f | 2453 | struct page *page = NULL; |
aaf14e40 MH |
2454 | struct mempolicy *mpol; |
2455 | gfp_t gfp_mask = htlb_alloc_mask(h); | |
2456 | int nid; | |
2457 | nodemask_t *nodemask; | |
2458 | ||
2459 | nid = huge_node(vma, addr, gfp_mask, &mpol, &nodemask); | |
cfcaa66f BW |
2460 | if (mpol_is_preferred_many(mpol)) { |
2461 | gfp_t gfp = gfp_mask | __GFP_NOWARN; | |
2462 | ||
2463 | gfp &= ~(__GFP_DIRECT_RECLAIM | __GFP_NOFAIL); | |
2b21624f | 2464 | page = alloc_surplus_huge_page(h, gfp, nid, nodemask); |
aaf14e40 | 2465 | |
cfcaa66f BW |
2466 | /* Fallback to all nodes if page==NULL */ |
2467 | nodemask = NULL; | |
2468 | } | |
2469 | ||
2470 | if (!page) | |
2b21624f | 2471 | page = alloc_surplus_huge_page(h, gfp_mask, nid, nodemask); |
cfcaa66f | 2472 | mpol_cond_put(mpol); |
aaf14e40 | 2473 | return page; |
099730d6 DH |
2474 | } |
2475 | ||
ab5ac90a | 2476 | /* page migration callback function */ |
3e59fcb0 | 2477 | struct page *alloc_huge_page_nodemask(struct hstate *h, int preferred_nid, |
d92bbc27 | 2478 | nodemask_t *nmask, gfp_t gfp_mask) |
4db9b2ef | 2479 | { |
db71ef79 | 2480 | spin_lock_irq(&hugetlb_lock); |
8346d69d | 2481 | if (available_huge_pages(h)) { |
3e59fcb0 MH |
2482 | struct page *page; |
2483 | ||
2484 | page = dequeue_huge_page_nodemask(h, gfp_mask, preferred_nid, nmask); | |
2485 | if (page) { | |
db71ef79 | 2486 | spin_unlock_irq(&hugetlb_lock); |
3e59fcb0 | 2487 | return page; |
4db9b2ef MH |
2488 | } |
2489 | } | |
db71ef79 | 2490 | spin_unlock_irq(&hugetlb_lock); |
4db9b2ef | 2491 | |
0c397dae | 2492 | return alloc_migrate_huge_page(h, gfp_mask, preferred_nid, nmask); |
4db9b2ef MH |
2493 | } |
2494 | ||
ebd63723 | 2495 | /* mempolicy aware migration callback */ |
389c8178 MH |
2496 | struct page *alloc_huge_page_vma(struct hstate *h, struct vm_area_struct *vma, |
2497 | unsigned long address) | |
ebd63723 MH |
2498 | { |
2499 | struct mempolicy *mpol; | |
2500 | nodemask_t *nodemask; | |
2501 | struct page *page; | |
ebd63723 MH |
2502 | gfp_t gfp_mask; |
2503 | int node; | |
2504 | ||
ebd63723 MH |
2505 | gfp_mask = htlb_alloc_mask(h); |
2506 | node = huge_node(vma, address, gfp_mask, &mpol, &nodemask); | |
d92bbc27 | 2507 | page = alloc_huge_page_nodemask(h, node, nodemask, gfp_mask); |
ebd63723 MH |
2508 | mpol_cond_put(mpol); |
2509 | ||
2510 | return page; | |
2511 | } | |
2512 | ||
e4e574b7 | 2513 | /* |
25985edc | 2514 | * Increase the hugetlb pool such that it can accommodate a reservation |
e4e574b7 AL |
2515 | * of size 'delta'. |
2516 | */ | |
0a4f3d1b | 2517 | static int gather_surplus_pages(struct hstate *h, long delta) |
1b2a1e7b | 2518 | __must_hold(&hugetlb_lock) |
e4e574b7 | 2519 | { |
34665341 | 2520 | LIST_HEAD(surplus_list); |
e4e574b7 | 2521 | struct page *page, *tmp; |
0a4f3d1b LX |
2522 | int ret; |
2523 | long i; | |
2524 | long needed, allocated; | |
28073b02 | 2525 | bool alloc_ok = true; |
e4e574b7 | 2526 | |
9487ca60 | 2527 | lockdep_assert_held(&hugetlb_lock); |
a5516438 | 2528 | needed = (h->resv_huge_pages + delta) - h->free_huge_pages; |
ac09b3a1 | 2529 | if (needed <= 0) { |
a5516438 | 2530 | h->resv_huge_pages += delta; |
e4e574b7 | 2531 | return 0; |
ac09b3a1 | 2532 | } |
e4e574b7 AL |
2533 | |
2534 | allocated = 0; | |
e4e574b7 AL |
2535 | |
2536 | ret = -ENOMEM; | |
2537 | retry: | |
db71ef79 | 2538 | spin_unlock_irq(&hugetlb_lock); |
e4e574b7 | 2539 | for (i = 0; i < needed; i++) { |
0c397dae | 2540 | page = alloc_surplus_huge_page(h, htlb_alloc_mask(h), |
2b21624f | 2541 | NUMA_NO_NODE, NULL); |
28073b02 HD |
2542 | if (!page) { |
2543 | alloc_ok = false; | |
2544 | break; | |
2545 | } | |
e4e574b7 | 2546 | list_add(&page->lru, &surplus_list); |
69ed779a | 2547 | cond_resched(); |
e4e574b7 | 2548 | } |
28073b02 | 2549 | allocated += i; |
e4e574b7 AL |
2550 | |
2551 | /* | |
2552 | * After retaking hugetlb_lock, we need to recalculate 'needed' | |
2553 | * because either resv_huge_pages or free_huge_pages may have changed. | |
2554 | */ | |
db71ef79 | 2555 | spin_lock_irq(&hugetlb_lock); |
a5516438 AK |
2556 | needed = (h->resv_huge_pages + delta) - |
2557 | (h->free_huge_pages + allocated); | |
28073b02 HD |
2558 | if (needed > 0) { |
2559 | if (alloc_ok) | |
2560 | goto retry; | |
2561 | /* | |
2562 | * We were not able to allocate enough pages to | |
2563 | * satisfy the entire reservation so we free what | |
2564 | * we've allocated so far. | |
2565 | */ | |
2566 | goto free; | |
2567 | } | |
e4e574b7 AL |
2568 | /* |
2569 | * The surplus_list now contains _at_least_ the number of extra pages | |
25985edc | 2570 | * needed to accommodate the reservation. Add the appropriate number |
e4e574b7 | 2571 | * of pages to the hugetlb pool and free the extras back to the buddy |
ac09b3a1 AL |
2572 | * allocator. Commit the entire reservation here to prevent another |
2573 | * process from stealing the pages as they are added to the pool but | |
2574 | * before they are reserved. | |
e4e574b7 AL |
2575 | */ |
2576 | needed += allocated; | |
a5516438 | 2577 | h->resv_huge_pages += delta; |
e4e574b7 | 2578 | ret = 0; |
a9869b83 | 2579 | |
19fc3f0a | 2580 | /* Free the needed pages to the hugetlb pool */ |
e4e574b7 | 2581 | list_for_each_entry_safe(page, tmp, &surplus_list, lru) { |
19fc3f0a AL |
2582 | if ((--needed) < 0) |
2583 | break; | |
b65a4eda | 2584 | /* Add the page to the hugetlb allocator */ |
240d67a8 | 2585 | enqueue_hugetlb_folio(h, page_folio(page)); |
19fc3f0a | 2586 | } |
28073b02 | 2587 | free: |
db71ef79 | 2588 | spin_unlock_irq(&hugetlb_lock); |
19fc3f0a | 2589 | |
b65a4eda MK |
2590 | /* |
2591 | * Free unnecessary surplus pages to the buddy allocator. | |
2592 | * Pages have no ref count, call free_huge_page directly. | |
2593 | */ | |
c0d934ba | 2594 | list_for_each_entry_safe(page, tmp, &surplus_list, lru) |
b65a4eda | 2595 | free_huge_page(page); |
db71ef79 | 2596 | spin_lock_irq(&hugetlb_lock); |
e4e574b7 AL |
2597 | |
2598 | return ret; | |
2599 | } | |
2600 | ||
2601 | /* | |
e5bbc8a6 MK |
2602 | * This routine has two main purposes: |
2603 | * 1) Decrement the reservation count (resv_huge_pages) by the value passed | |
2604 | * in unused_resv_pages. This corresponds to the prior adjustments made | |
2605 | * to the associated reservation map. | |
2606 | * 2) Free any unused surplus pages that may have been allocated to satisfy | |
2607 | * the reservation. As many as unused_resv_pages may be freed. | |
e4e574b7 | 2608 | */ |
a5516438 AK |
2609 | static void return_unused_surplus_pages(struct hstate *h, |
2610 | unsigned long unused_resv_pages) | |
e4e574b7 | 2611 | { |
e4e574b7 | 2612 | unsigned long nr_pages; |
10c6ec49 MK |
2613 | struct page *page; |
2614 | LIST_HEAD(page_list); | |
2615 | ||
9487ca60 | 2616 | lockdep_assert_held(&hugetlb_lock); |
10c6ec49 MK |
2617 | /* Uncommit the reservation */ |
2618 | h->resv_huge_pages -= unused_resv_pages; | |
e4e574b7 | 2619 | |
c0531714 | 2620 | if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) |
e5bbc8a6 | 2621 | goto out; |
aa888a74 | 2622 | |
e5bbc8a6 MK |
2623 | /* |
2624 | * Part (or even all) of the reservation could have been backed | |
2625 | * by pre-allocated pages. Only free surplus pages. | |
2626 | */ | |
a5516438 | 2627 | nr_pages = min(unused_resv_pages, h->surplus_huge_pages); |
e4e574b7 | 2628 | |
685f3457 LS |
2629 | /* |
2630 | * We want to release as many surplus pages as possible, spread | |
9b5e5d0f LS |
2631 | * evenly across all nodes with memory. Iterate across these nodes |
2632 | * until we can no longer free unreserved surplus pages. This occurs | |
2633 | * when the nodes with surplus pages have no free pages. | |
10c6ec49 | 2634 | * remove_pool_huge_page() will balance the freed pages across the |
9b5e5d0f | 2635 | * on-line nodes with memory and will handle the hstate accounting. |
685f3457 LS |
2636 | */ |
2637 | while (nr_pages--) { | |
10c6ec49 MK |
2638 | page = remove_pool_huge_page(h, &node_states[N_MEMORY], 1); |
2639 | if (!page) | |
e5bbc8a6 | 2640 | goto out; |
10c6ec49 MK |
2641 | |
2642 | list_add(&page->lru, &page_list); | |
e4e574b7 | 2643 | } |
e5bbc8a6 MK |
2644 | |
2645 | out: | |
db71ef79 | 2646 | spin_unlock_irq(&hugetlb_lock); |
10c6ec49 | 2647 | update_and_free_pages_bulk(h, &page_list); |
db71ef79 | 2648 | spin_lock_irq(&hugetlb_lock); |
e4e574b7 AL |
2649 | } |
2650 | ||
5e911373 | 2651 | |
c37f9fb1 | 2652 | /* |
feba16e2 | 2653 | * vma_needs_reservation, vma_commit_reservation and vma_end_reservation |
5e911373 | 2654 | * are used by the huge page allocation routines to manage reservations. |
cf3ad20b MK |
2655 | * |
2656 | * vma_needs_reservation is called to determine if the huge page at addr | |
2657 | * within the vma has an associated reservation. If a reservation is | |
2658 | * needed, the value 1 is returned. The caller is then responsible for | |
2659 | * managing the global reservation and subpool usage counts. After | |
2660 | * the huge page has been allocated, vma_commit_reservation is called | |
feba16e2 MK |
2661 | * to add the page to the reservation map. If the page allocation fails, |
2662 | * the reservation must be ended instead of committed. vma_end_reservation | |
2663 | * is called in such cases. | |
cf3ad20b MK |
2664 | * |
2665 | * In the normal case, vma_commit_reservation returns the same value | |
2666 | * as the preceding vma_needs_reservation call. The only time this | |
2667 | * is not the case is if a reserve map was changed between calls. It | |
2668 | * is the responsibility of the caller to notice the difference and | |
2669 | * take appropriate action. | |
96b96a96 MK |
2670 | * |
2671 | * vma_add_reservation is used in error paths where a reservation must | |
2672 | * be restored when a newly allocated huge page must be freed. It is | |
2673 | * to be called after calling vma_needs_reservation to determine if a | |
2674 | * reservation exists. | |
846be085 MK |
2675 | * |
2676 | * vma_del_reservation is used in error paths where an entry in the reserve | |
2677 | * map was created during huge page allocation and must be removed. It is to | |
2678 | * be called after calling vma_needs_reservation to determine if a reservation | |
2679 | * exists. | |
c37f9fb1 | 2680 | */ |
5e911373 MK |
2681 | enum vma_resv_mode { |
2682 | VMA_NEEDS_RESV, | |
2683 | VMA_COMMIT_RESV, | |
feba16e2 | 2684 | VMA_END_RESV, |
96b96a96 | 2685 | VMA_ADD_RESV, |
846be085 | 2686 | VMA_DEL_RESV, |
5e911373 | 2687 | }; |
cf3ad20b MK |
2688 | static long __vma_reservation_common(struct hstate *h, |
2689 | struct vm_area_struct *vma, unsigned long addr, | |
5e911373 | 2690 | enum vma_resv_mode mode) |
c37f9fb1 | 2691 | { |
4e35f483 JK |
2692 | struct resv_map *resv; |
2693 | pgoff_t idx; | |
cf3ad20b | 2694 | long ret; |
0db9d74e | 2695 | long dummy_out_regions_needed; |
c37f9fb1 | 2696 | |
4e35f483 JK |
2697 | resv = vma_resv_map(vma); |
2698 | if (!resv) | |
84afd99b | 2699 | return 1; |
c37f9fb1 | 2700 | |
4e35f483 | 2701 | idx = vma_hugecache_offset(h, vma, addr); |
5e911373 MK |
2702 | switch (mode) { |
2703 | case VMA_NEEDS_RESV: | |
0db9d74e MA |
2704 | ret = region_chg(resv, idx, idx + 1, &dummy_out_regions_needed); |
2705 | /* We assume that vma_reservation_* routines always operate on | |
2706 | * 1 page, and that adding to resv map a 1 page entry can only | |
2707 | * ever require 1 region. | |
2708 | */ | |
2709 | VM_BUG_ON(dummy_out_regions_needed != 1); | |
5e911373 MK |
2710 | break; |
2711 | case VMA_COMMIT_RESV: | |
075a61d0 | 2712 | ret = region_add(resv, idx, idx + 1, 1, NULL, NULL); |
0db9d74e MA |
2713 | /* region_add calls of range 1 should never fail. */ |
2714 | VM_BUG_ON(ret < 0); | |
5e911373 | 2715 | break; |
feba16e2 | 2716 | case VMA_END_RESV: |
0db9d74e | 2717 | region_abort(resv, idx, idx + 1, 1); |
5e911373 MK |
2718 | ret = 0; |
2719 | break; | |
96b96a96 | 2720 | case VMA_ADD_RESV: |
0db9d74e | 2721 | if (vma->vm_flags & VM_MAYSHARE) { |
075a61d0 | 2722 | ret = region_add(resv, idx, idx + 1, 1, NULL, NULL); |
0db9d74e MA |
2723 | /* region_add calls of range 1 should never fail. */ |
2724 | VM_BUG_ON(ret < 0); | |
2725 | } else { | |
2726 | region_abort(resv, idx, idx + 1, 1); | |
96b96a96 MK |
2727 | ret = region_del(resv, idx, idx + 1); |
2728 | } | |
2729 | break; | |
846be085 MK |
2730 | case VMA_DEL_RESV: |
2731 | if (vma->vm_flags & VM_MAYSHARE) { | |
2732 | region_abort(resv, idx, idx + 1, 1); | |
2733 | ret = region_del(resv, idx, idx + 1); | |
2734 | } else { | |
2735 | ret = region_add(resv, idx, idx + 1, 1, NULL, NULL); | |
2736 | /* region_add calls of range 1 should never fail. */ | |
2737 | VM_BUG_ON(ret < 0); | |
2738 | } | |
2739 | break; | |
5e911373 MK |
2740 | default: |
2741 | BUG(); | |
2742 | } | |
84afd99b | 2743 | |
846be085 | 2744 | if (vma->vm_flags & VM_MAYSHARE || mode == VMA_DEL_RESV) |
cf3ad20b | 2745 | return ret; |
bf3d12b9 ML |
2746 | /* |
2747 | * We know private mapping must have HPAGE_RESV_OWNER set. | |
2748 | * | |
2749 | * In most cases, reserves always exist for private mappings. | |
2750 | * However, a file associated with mapping could have been | |
2751 | * hole punched or truncated after reserves were consumed. | |
2752 | * As subsequent fault on such a range will not use reserves. | |
2753 | * Subtle - The reserve map for private mappings has the | |
2754 | * opposite meaning than that of shared mappings. If NO | |
2755 | * entry is in the reserve map, it means a reservation exists. | |
2756 | * If an entry exists in the reserve map, it means the | |
2757 | * reservation has already been consumed. As a result, the | |
2758 | * return value of this routine is the opposite of the | |
2759 | * value returned from reserve map manipulation routines above. | |
2760 | */ | |
2761 | if (ret > 0) | |
2762 | return 0; | |
2763 | if (ret == 0) | |
2764 | return 1; | |
2765 | return ret; | |
c37f9fb1 | 2766 | } |
cf3ad20b MK |
2767 | |
2768 | static long vma_needs_reservation(struct hstate *h, | |
a5516438 | 2769 | struct vm_area_struct *vma, unsigned long addr) |
c37f9fb1 | 2770 | { |
5e911373 | 2771 | return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV); |
cf3ad20b | 2772 | } |
84afd99b | 2773 | |
cf3ad20b MK |
2774 | static long vma_commit_reservation(struct hstate *h, |
2775 | struct vm_area_struct *vma, unsigned long addr) | |
2776 | { | |
5e911373 MK |
2777 | return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV); |
2778 | } | |
2779 | ||
feba16e2 | 2780 | static void vma_end_reservation(struct hstate *h, |
5e911373 MK |
2781 | struct vm_area_struct *vma, unsigned long addr) |
2782 | { | |
feba16e2 | 2783 | (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV); |
c37f9fb1 AW |
2784 | } |
2785 | ||
96b96a96 MK |
2786 | static long vma_add_reservation(struct hstate *h, |
2787 | struct vm_area_struct *vma, unsigned long addr) | |
2788 | { | |
2789 | return __vma_reservation_common(h, vma, addr, VMA_ADD_RESV); | |
2790 | } | |
2791 | ||
846be085 MK |
2792 | static long vma_del_reservation(struct hstate *h, |
2793 | struct vm_area_struct *vma, unsigned long addr) | |
2794 | { | |
2795 | return __vma_reservation_common(h, vma, addr, VMA_DEL_RESV); | |
2796 | } | |
2797 | ||
96b96a96 | 2798 | /* |
846be085 MK |
2799 | * This routine is called to restore reservation information on error paths. |
2800 | * It should ONLY be called for pages allocated via alloc_huge_page(), and | |
2801 | * the hugetlb mutex should remain held when calling this routine. | |
2802 | * | |
2803 | * It handles two specific cases: | |
2804 | * 1) A reservation was in place and the page consumed the reservation. | |
2805 | * HPageRestoreReserve is set in the page. | |
2806 | * 2) No reservation was in place for the page, so HPageRestoreReserve is | |
2807 | * not set. However, alloc_huge_page always updates the reserve map. | |
2808 | * | |
2809 | * In case 1, free_huge_page later in the error path will increment the | |
2810 | * global reserve count. But, free_huge_page does not have enough context | |
2811 | * to adjust the reservation map. This case deals primarily with private | |
2812 | * mappings. Adjust the reserve map here to be consistent with global | |
2813 | * reserve count adjustments to be made by free_huge_page. Make sure the | |
2814 | * reserve map indicates there is a reservation present. | |
2815 | * | |
2816 | * In case 2, simply undo reserve map modifications done by alloc_huge_page. | |
96b96a96 | 2817 | */ |
846be085 MK |
2818 | void restore_reserve_on_error(struct hstate *h, struct vm_area_struct *vma, |
2819 | unsigned long address, struct page *page) | |
96b96a96 | 2820 | { |
846be085 | 2821 | long rc = vma_needs_reservation(h, vma, address); |
96b96a96 | 2822 | |
846be085 MK |
2823 | if (HPageRestoreReserve(page)) { |
2824 | if (unlikely(rc < 0)) | |
96b96a96 MK |
2825 | /* |
2826 | * Rare out of memory condition in reserve map | |
d6995da3 | 2827 | * manipulation. Clear HPageRestoreReserve so that |
96b96a96 MK |
2828 | * global reserve count will not be incremented |
2829 | * by free_huge_page. This will make it appear | |
2830 | * as though the reservation for this page was | |
2831 | * consumed. This may prevent the task from | |
2832 | * faulting in the page at a later time. This | |
2833 | * is better than inconsistent global huge page | |
2834 | * accounting of reserve counts. | |
2835 | */ | |
d6995da3 | 2836 | ClearHPageRestoreReserve(page); |
846be085 MK |
2837 | else if (rc) |
2838 | (void)vma_add_reservation(h, vma, address); | |
2839 | else | |
2840 | vma_end_reservation(h, vma, address); | |
2841 | } else { | |
2842 | if (!rc) { | |
2843 | /* | |
2844 | * This indicates there is an entry in the reserve map | |
c7b1850d | 2845 | * not added by alloc_huge_page. We know it was added |
846be085 MK |
2846 | * before the alloc_huge_page call, otherwise |
2847 | * HPageRestoreReserve would be set on the page. | |
2848 | * Remove the entry so that a subsequent allocation | |
2849 | * does not consume a reservation. | |
2850 | */ | |
2851 | rc = vma_del_reservation(h, vma, address); | |
2852 | if (rc < 0) | |
96b96a96 | 2853 | /* |
846be085 MK |
2854 | * VERY rare out of memory condition. Since |
2855 | * we can not delete the entry, set | |
2856 | * HPageRestoreReserve so that the reserve | |
2857 | * count will be incremented when the page | |
2858 | * is freed. This reserve will be consumed | |
2859 | * on a subsequent allocation. | |
96b96a96 | 2860 | */ |
846be085 MK |
2861 | SetHPageRestoreReserve(page); |
2862 | } else if (rc < 0) { | |
2863 | /* | |
2864 | * Rare out of memory condition from | |
2865 | * vma_needs_reservation call. Memory allocation is | |
2866 | * only attempted if a new entry is needed. Therefore, | |
2867 | * this implies there is not an entry in the | |
2868 | * reserve map. | |
2869 | * | |
2870 | * For shared mappings, no entry in the map indicates | |
2871 | * no reservation. We are done. | |
2872 | */ | |
2873 | if (!(vma->vm_flags & VM_MAYSHARE)) | |
2874 | /* | |
2875 | * For private mappings, no entry indicates | |
2876 | * a reservation is present. Since we can | |
2877 | * not add an entry, set SetHPageRestoreReserve | |
2878 | * on the page so reserve count will be | |
2879 | * incremented when freed. This reserve will | |
2880 | * be consumed on a subsequent allocation. | |
2881 | */ | |
2882 | SetHPageRestoreReserve(page); | |
96b96a96 | 2883 | } else |
846be085 MK |
2884 | /* |
2885 | * No reservation present, do nothing | |
2886 | */ | |
2887 | vma_end_reservation(h, vma, address); | |
96b96a96 MK |
2888 | } |
2889 | } | |
2890 | ||
369fa227 | 2891 | /* |
19fc1a7e SK |
2892 | * alloc_and_dissolve_hugetlb_folio - Allocate a new folio and dissolve |
2893 | * the old one | |
369fa227 | 2894 | * @h: struct hstate old page belongs to |
19fc1a7e | 2895 | * @old_folio: Old folio to dissolve |
ae37c7ff | 2896 | * @list: List to isolate the page in case we need to |
369fa227 OS |
2897 | * Returns 0 on success, otherwise negated error. |
2898 | */ | |
19fc1a7e SK |
2899 | static int alloc_and_dissolve_hugetlb_folio(struct hstate *h, |
2900 | struct folio *old_folio, struct list_head *list) | |
369fa227 OS |
2901 | { |
2902 | gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE; | |
de656ed3 | 2903 | int nid = folio_nid(old_folio); |
de656ed3 | 2904 | struct folio *new_folio; |
369fa227 OS |
2905 | int ret = 0; |
2906 | ||
2907 | /* | |
19fc1a7e SK |
2908 | * Before dissolving the folio, we need to allocate a new one for the |
2909 | * pool to remain stable. Here, we allocate the folio and 'prep' it | |
f41f2ed4 MS |
2910 | * by doing everything but actually updating counters and adding to |
2911 | * the pool. This simplifies and let us do most of the processing | |
2912 | * under the lock. | |
369fa227 | 2913 | */ |
19fc1a7e SK |
2914 | new_folio = alloc_buddy_hugetlb_folio(h, gfp_mask, nid, NULL, NULL); |
2915 | if (!new_folio) | |
369fa227 | 2916 | return -ENOMEM; |
de656ed3 | 2917 | __prep_new_hugetlb_folio(h, new_folio); |
369fa227 OS |
2918 | |
2919 | retry: | |
2920 | spin_lock_irq(&hugetlb_lock); | |
de656ed3 | 2921 | if (!folio_test_hugetlb(old_folio)) { |
369fa227 | 2922 | /* |
19fc1a7e | 2923 | * Freed from under us. Drop new_folio too. |
369fa227 OS |
2924 | */ |
2925 | goto free_new; | |
de656ed3 | 2926 | } else if (folio_ref_count(old_folio)) { |
369fa227 | 2927 | /* |
19fc1a7e | 2928 | * Someone has grabbed the folio, try to isolate it here. |
ae37c7ff | 2929 | * Fail with -EBUSY if not possible. |
369fa227 | 2930 | */ |
ae37c7ff | 2931 | spin_unlock_irq(&hugetlb_lock); |
19fc1a7e | 2932 | ret = isolate_hugetlb(&old_folio->page, list); |
ae37c7ff | 2933 | spin_lock_irq(&hugetlb_lock); |
369fa227 | 2934 | goto free_new; |
de656ed3 | 2935 | } else if (!folio_test_hugetlb_freed(old_folio)) { |
369fa227 | 2936 | /* |
19fc1a7e | 2937 | * Folio's refcount is 0 but it has not been enqueued in the |
369fa227 OS |
2938 | * freelist yet. Race window is small, so we can succeed here if |
2939 | * we retry. | |
2940 | */ | |
2941 | spin_unlock_irq(&hugetlb_lock); | |
2942 | cond_resched(); | |
2943 | goto retry; | |
2944 | } else { | |
2945 | /* | |
19fc1a7e | 2946 | * Ok, old_folio is still a genuine free hugepage. Remove it from |
369fa227 OS |
2947 | * the freelist and decrease the counters. These will be |
2948 | * incremented again when calling __prep_account_new_huge_page() | |
240d67a8 SK |
2949 | * and enqueue_hugetlb_folio() for new_folio. The counters will |
2950 | * remain stable since this happens under the lock. | |
369fa227 | 2951 | */ |
cfd5082b | 2952 | remove_hugetlb_folio(h, old_folio, false); |
369fa227 OS |
2953 | |
2954 | /* | |
19fc1a7e | 2955 | * Ref count on new_folio is already zero as it was dropped |
b65a4eda | 2956 | * earlier. It can be directly added to the pool free list. |
369fa227 | 2957 | */ |
369fa227 | 2958 | __prep_account_new_huge_page(h, nid); |
240d67a8 | 2959 | enqueue_hugetlb_folio(h, new_folio); |
369fa227 OS |
2960 | |
2961 | /* | |
19fc1a7e | 2962 | * Folio has been replaced, we can safely free the old one. |
369fa227 OS |
2963 | */ |
2964 | spin_unlock_irq(&hugetlb_lock); | |
d6ef19e2 | 2965 | update_and_free_hugetlb_folio(h, old_folio, false); |
369fa227 OS |
2966 | } |
2967 | ||
2968 | return ret; | |
2969 | ||
2970 | free_new: | |
2971 | spin_unlock_irq(&hugetlb_lock); | |
19fc1a7e | 2972 | /* Folio has a zero ref count, but needs a ref to be freed */ |
de656ed3 | 2973 | folio_ref_unfreeze(new_folio, 1); |
d6ef19e2 | 2974 | update_and_free_hugetlb_folio(h, new_folio, false); |
369fa227 OS |
2975 | |
2976 | return ret; | |
2977 | } | |
2978 | ||
ae37c7ff | 2979 | int isolate_or_dissolve_huge_page(struct page *page, struct list_head *list) |
369fa227 OS |
2980 | { |
2981 | struct hstate *h; | |
d5e33bd8 | 2982 | struct folio *folio = page_folio(page); |
ae37c7ff | 2983 | int ret = -EBUSY; |
369fa227 OS |
2984 | |
2985 | /* | |
2986 | * The page might have been dissolved from under our feet, so make sure | |
2987 | * to carefully check the state under the lock. | |
2988 | * Return success when racing as if we dissolved the page ourselves. | |
2989 | */ | |
2990 | spin_lock_irq(&hugetlb_lock); | |
d5e33bd8 SK |
2991 | if (folio_test_hugetlb(folio)) { |
2992 | h = folio_hstate(folio); | |
369fa227 OS |
2993 | } else { |
2994 | spin_unlock_irq(&hugetlb_lock); | |
2995 | return 0; | |
2996 | } | |
2997 | spin_unlock_irq(&hugetlb_lock); | |
2998 | ||
2999 | /* | |
3000 | * Fence off gigantic pages as there is a cyclic dependency between | |
3001 | * alloc_contig_range and them. Return -ENOMEM as this has the effect | |
3002 | * of bailing out right away without further retrying. | |
3003 | */ | |
3004 | if (hstate_is_gigantic(h)) | |
3005 | return -ENOMEM; | |
3006 | ||
d5e33bd8 | 3007 | if (folio_ref_count(folio) && !isolate_hugetlb(&folio->page, list)) |
ae37c7ff | 3008 | ret = 0; |
d5e33bd8 | 3009 | else if (!folio_ref_count(folio)) |
19fc1a7e | 3010 | ret = alloc_and_dissolve_hugetlb_folio(h, folio, list); |
ae37c7ff OS |
3011 | |
3012 | return ret; | |
369fa227 OS |
3013 | } |
3014 | ||
70c3547e | 3015 | struct page *alloc_huge_page(struct vm_area_struct *vma, |
04f2cbe3 | 3016 | unsigned long addr, int avoid_reserve) |
1da177e4 | 3017 | { |
90481622 | 3018 | struct hugepage_subpool *spool = subpool_vma(vma); |
a5516438 | 3019 | struct hstate *h = hstate_vma(vma); |
348ea204 | 3020 | struct page *page; |
d4ab0316 | 3021 | struct folio *folio; |
d85f69b0 MK |
3022 | long map_chg, map_commit; |
3023 | long gbl_chg; | |
6d76dcf4 AK |
3024 | int ret, idx; |
3025 | struct hugetlb_cgroup *h_cg; | |
08cf9faf | 3026 | bool deferred_reserve; |
a1e78772 | 3027 | |
6d76dcf4 | 3028 | idx = hstate_index(h); |
a1e78772 | 3029 | /* |
d85f69b0 MK |
3030 | * Examine the region/reserve map to determine if the process |
3031 | * has a reservation for the page to be allocated. A return | |
3032 | * code of zero indicates a reservation exists (no change). | |
a1e78772 | 3033 | */ |
d85f69b0 MK |
3034 | map_chg = gbl_chg = vma_needs_reservation(h, vma, addr); |
3035 | if (map_chg < 0) | |
76dcee75 | 3036 | return ERR_PTR(-ENOMEM); |
d85f69b0 MK |
3037 | |
3038 | /* | |
3039 | * Processes that did not create the mapping will have no | |
3040 | * reserves as indicated by the region/reserve map. Check | |
3041 | * that the allocation will not exceed the subpool limit. | |
3042 | * Allocations for MAP_NORESERVE mappings also need to be | |
3043 | * checked against any subpool limit. | |
3044 | */ | |
3045 | if (map_chg || avoid_reserve) { | |
3046 | gbl_chg = hugepage_subpool_get_pages(spool, 1); | |
3047 | if (gbl_chg < 0) { | |
feba16e2 | 3048 | vma_end_reservation(h, vma, addr); |
76dcee75 | 3049 | return ERR_PTR(-ENOSPC); |
5e911373 | 3050 | } |
1da177e4 | 3051 | |
d85f69b0 MK |
3052 | /* |
3053 | * Even though there was no reservation in the region/reserve | |
3054 | * map, there could be reservations associated with the | |
3055 | * subpool that can be used. This would be indicated if the | |
3056 | * return value of hugepage_subpool_get_pages() is zero. | |
3057 | * However, if avoid_reserve is specified we still avoid even | |
3058 | * the subpool reservations. | |
3059 | */ | |
3060 | if (avoid_reserve) | |
3061 | gbl_chg = 1; | |
3062 | } | |
3063 | ||
08cf9faf MA |
3064 | /* If this allocation is not consuming a reservation, charge it now. |
3065 | */ | |
6501fe5f | 3066 | deferred_reserve = map_chg || avoid_reserve; |
08cf9faf MA |
3067 | if (deferred_reserve) { |
3068 | ret = hugetlb_cgroup_charge_cgroup_rsvd( | |
3069 | idx, pages_per_huge_page(h), &h_cg); | |
3070 | if (ret) | |
3071 | goto out_subpool_put; | |
3072 | } | |
3073 | ||
6d76dcf4 | 3074 | ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg); |
8f34af6f | 3075 | if (ret) |
08cf9faf | 3076 | goto out_uncharge_cgroup_reservation; |
8f34af6f | 3077 | |
db71ef79 | 3078 | spin_lock_irq(&hugetlb_lock); |
d85f69b0 MK |
3079 | /* |
3080 | * glb_chg is passed to indicate whether or not a page must be taken | |
3081 | * from the global free pool (global change). gbl_chg == 0 indicates | |
3082 | * a reservation exists for the allocation. | |
3083 | */ | |
3084 | page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg); | |
81a6fcae | 3085 | if (!page) { |
db71ef79 | 3086 | spin_unlock_irq(&hugetlb_lock); |
0c397dae | 3087 | page = alloc_buddy_huge_page_with_mpol(h, vma, addr); |
8f34af6f JZ |
3088 | if (!page) |
3089 | goto out_uncharge_cgroup; | |
12df140f | 3090 | spin_lock_irq(&hugetlb_lock); |
a88c7695 | 3091 | if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) { |
d6995da3 | 3092 | SetHPageRestoreReserve(page); |
a88c7695 NH |
3093 | h->resv_huge_pages--; |
3094 | } | |
15a8d68e | 3095 | list_add(&page->lru, &h->hugepage_activelist); |
2b21624f | 3096 | set_page_refcounted(page); |
81a6fcae | 3097 | /* Fall through */ |
68842c9b | 3098 | } |
d4ab0316 | 3099 | folio = page_folio(page); |
81a6fcae | 3100 | hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page); |
08cf9faf MA |
3101 | /* If allocation is not consuming a reservation, also store the |
3102 | * hugetlb_cgroup pointer on the page. | |
3103 | */ | |
3104 | if (deferred_reserve) { | |
3105 | hugetlb_cgroup_commit_charge_rsvd(idx, pages_per_huge_page(h), | |
3106 | h_cg, page); | |
3107 | } | |
3108 | ||
db71ef79 | 3109 | spin_unlock_irq(&hugetlb_lock); |
348ea204 | 3110 | |
d6995da3 | 3111 | hugetlb_set_page_subpool(page, spool); |
90d8b7e6 | 3112 | |
d85f69b0 MK |
3113 | map_commit = vma_commit_reservation(h, vma, addr); |
3114 | if (unlikely(map_chg > map_commit)) { | |
33039678 MK |
3115 | /* |
3116 | * The page was added to the reservation map between | |
3117 | * vma_needs_reservation and vma_commit_reservation. | |
3118 | * This indicates a race with hugetlb_reserve_pages. | |
3119 | * Adjust for the subpool count incremented above AND | |
3120 | * in hugetlb_reserve_pages for the same page. Also, | |
3121 | * the reservation count added in hugetlb_reserve_pages | |
3122 | * no longer applies. | |
3123 | */ | |
3124 | long rsv_adjust; | |
3125 | ||
3126 | rsv_adjust = hugepage_subpool_put_pages(spool, 1); | |
3127 | hugetlb_acct_memory(h, -rsv_adjust); | |
79aa925b | 3128 | if (deferred_reserve) |
d4ab0316 SK |
3129 | hugetlb_cgroup_uncharge_folio_rsvd(hstate_index(h), |
3130 | pages_per_huge_page(h), folio); | |
33039678 | 3131 | } |
90d8b7e6 | 3132 | return page; |
8f34af6f JZ |
3133 | |
3134 | out_uncharge_cgroup: | |
3135 | hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg); | |
08cf9faf MA |
3136 | out_uncharge_cgroup_reservation: |
3137 | if (deferred_reserve) | |
3138 | hugetlb_cgroup_uncharge_cgroup_rsvd(idx, pages_per_huge_page(h), | |
3139 | h_cg); | |
8f34af6f | 3140 | out_subpool_put: |
d85f69b0 | 3141 | if (map_chg || avoid_reserve) |
8f34af6f | 3142 | hugepage_subpool_put_pages(spool, 1); |
feba16e2 | 3143 | vma_end_reservation(h, vma, addr); |
8f34af6f | 3144 | return ERR_PTR(-ENOSPC); |
b45b5bd6 DG |
3145 | } |
3146 | ||
b5389086 | 3147 | int alloc_bootmem_huge_page(struct hstate *h, int nid) |
e24a1307 | 3148 | __attribute__ ((weak, alias("__alloc_bootmem_huge_page"))); |
b5389086 | 3149 | int __alloc_bootmem_huge_page(struct hstate *h, int nid) |
aa888a74 | 3150 | { |
b5389086 | 3151 | struct huge_bootmem_page *m = NULL; /* initialize for clang */ |
b2261026 | 3152 | int nr_nodes, node; |
aa888a74 | 3153 | |
b5389086 ZY |
3154 | /* do node specific alloc */ |
3155 | if (nid != NUMA_NO_NODE) { | |
3156 | m = memblock_alloc_try_nid_raw(huge_page_size(h), huge_page_size(h), | |
3157 | 0, MEMBLOCK_ALLOC_ACCESSIBLE, nid); | |
3158 | if (!m) | |
3159 | return 0; | |
3160 | goto found; | |
3161 | } | |
3162 | /* allocate from next node when distributing huge pages */ | |
b2261026 | 3163 | for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) { |
b5389086 | 3164 | m = memblock_alloc_try_nid_raw( |
8b89a116 | 3165 | huge_page_size(h), huge_page_size(h), |
97ad1087 | 3166 | 0, MEMBLOCK_ALLOC_ACCESSIBLE, node); |
b5389086 ZY |
3167 | /* |
3168 | * Use the beginning of the huge page to store the | |
3169 | * huge_bootmem_page struct (until gather_bootmem | |
3170 | * puts them into the mem_map). | |
3171 | */ | |
3172 | if (!m) | |
3173 | return 0; | |
3174 | goto found; | |
aa888a74 | 3175 | } |
aa888a74 AK |
3176 | |
3177 | found: | |
aa888a74 | 3178 | /* Put them into a private list first because mem_map is not up yet */ |
330d6e48 | 3179 | INIT_LIST_HEAD(&m->list); |
aa888a74 AK |
3180 | list_add(&m->list, &huge_boot_pages); |
3181 | m->hstate = h; | |
3182 | return 1; | |
3183 | } | |
3184 | ||
48b8d744 MK |
3185 | /* |
3186 | * Put bootmem huge pages into the standard lists after mem_map is up. | |
3187 | * Note: This only applies to gigantic (order > MAX_ORDER) pages. | |
3188 | */ | |
aa888a74 AK |
3189 | static void __init gather_bootmem_prealloc(void) |
3190 | { | |
3191 | struct huge_bootmem_page *m; | |
3192 | ||
3193 | list_for_each_entry(m, &huge_boot_pages, list) { | |
40d18ebf | 3194 | struct page *page = virt_to_page(m); |
7f325a8d | 3195 | struct folio *folio = page_folio(page); |
aa888a74 | 3196 | struct hstate *h = m->hstate; |
ee8f248d | 3197 | |
48b8d744 | 3198 | VM_BUG_ON(!hstate_is_gigantic(h)); |
d1c60955 SK |
3199 | WARN_ON(folio_ref_count(folio) != 1); |
3200 | if (prep_compound_gigantic_folio(folio, huge_page_order(h))) { | |
3201 | WARN_ON(folio_test_reserved(folio)); | |
3202 | prep_new_hugetlb_folio(h, folio, folio_nid(folio)); | |
2b21624f | 3203 | free_huge_page(page); /* add to the hugepage allocator */ |
7118fc29 | 3204 | } else { |
416d85ed | 3205 | /* VERY unlikely inflated ref count on a tail page */ |
7f325a8d | 3206 | free_gigantic_folio(folio, huge_page_order(h)); |
7118fc29 | 3207 | } |
af0fb9df | 3208 | |
b0320c7b | 3209 | /* |
48b8d744 MK |
3210 | * We need to restore the 'stolen' pages to totalram_pages |
3211 | * in order to fix confusing memory reports from free(1) and | |
3212 | * other side-effects, like CommitLimit going negative. | |
b0320c7b | 3213 | */ |
48b8d744 | 3214 | adjust_managed_page_count(page, pages_per_huge_page(h)); |
520495fe | 3215 | cond_resched(); |
aa888a74 AK |
3216 | } |
3217 | } | |
b5389086 ZY |
3218 | static void __init hugetlb_hstate_alloc_pages_onenode(struct hstate *h, int nid) |
3219 | { | |
3220 | unsigned long i; | |
3221 | char buf[32]; | |
3222 | ||
3223 | for (i = 0; i < h->max_huge_pages_node[nid]; ++i) { | |
3224 | if (hstate_is_gigantic(h)) { | |
3225 | if (!alloc_bootmem_huge_page(h, nid)) | |
3226 | break; | |
3227 | } else { | |
19fc1a7e | 3228 | struct folio *folio; |
b5389086 ZY |
3229 | gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE; |
3230 | ||
19fc1a7e | 3231 | folio = alloc_fresh_hugetlb_folio(h, gfp_mask, nid, |
b5389086 | 3232 | &node_states[N_MEMORY], NULL); |
19fc1a7e | 3233 | if (!folio) |
b5389086 | 3234 | break; |
19fc1a7e | 3235 | free_huge_page(&folio->page); /* free it into the hugepage allocator */ |
b5389086 ZY |
3236 | } |
3237 | cond_resched(); | |
3238 | } | |
3239 | if (i == h->max_huge_pages_node[nid]) | |
3240 | return; | |
3241 | ||
3242 | string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32); | |
3243 | pr_warn("HugeTLB: allocating %u of page size %s failed node%d. Only allocated %lu hugepages.\n", | |
3244 | h->max_huge_pages_node[nid], buf, nid, i); | |
3245 | h->max_huge_pages -= (h->max_huge_pages_node[nid] - i); | |
3246 | h->max_huge_pages_node[nid] = i; | |
3247 | } | |
aa888a74 | 3248 | |
8faa8b07 | 3249 | static void __init hugetlb_hstate_alloc_pages(struct hstate *h) |
1da177e4 LT |
3250 | { |
3251 | unsigned long i; | |
f60858f9 | 3252 | nodemask_t *node_alloc_noretry; |
b5389086 ZY |
3253 | bool node_specific_alloc = false; |
3254 | ||
3255 | /* skip gigantic hugepages allocation if hugetlb_cma enabled */ | |
3256 | if (hstate_is_gigantic(h) && hugetlb_cma_size) { | |
3257 | pr_warn_once("HugeTLB: hugetlb_cma is enabled, skip boot time allocation\n"); | |
3258 | return; | |
3259 | } | |
3260 | ||
3261 | /* do node specific alloc */ | |
0a7a0f6f | 3262 | for_each_online_node(i) { |
b5389086 ZY |
3263 | if (h->max_huge_pages_node[i] > 0) { |
3264 | hugetlb_hstate_alloc_pages_onenode(h, i); | |
3265 | node_specific_alloc = true; | |
3266 | } | |
3267 | } | |
f60858f9 | 3268 | |
b5389086 ZY |
3269 | if (node_specific_alloc) |
3270 | return; | |
3271 | ||
3272 | /* below will do all node balanced alloc */ | |
f60858f9 MK |
3273 | if (!hstate_is_gigantic(h)) { |
3274 | /* | |
3275 | * Bit mask controlling how hard we retry per-node allocations. | |
3276 | * Ignore errors as lower level routines can deal with | |
3277 | * node_alloc_noretry == NULL. If this kmalloc fails at boot | |
3278 | * time, we are likely in bigger trouble. | |
3279 | */ | |
3280 | node_alloc_noretry = kmalloc(sizeof(*node_alloc_noretry), | |
3281 | GFP_KERNEL); | |
3282 | } else { | |
3283 | /* allocations done at boot time */ | |
3284 | node_alloc_noretry = NULL; | |
3285 | } | |
3286 | ||
3287 | /* bit mask controlling how hard we retry per-node allocations */ | |
3288 | if (node_alloc_noretry) | |
3289 | nodes_clear(*node_alloc_noretry); | |
a5516438 | 3290 | |
e5ff2159 | 3291 | for (i = 0; i < h->max_huge_pages; ++i) { |
bae7f4ae | 3292 | if (hstate_is_gigantic(h)) { |
b5389086 | 3293 | if (!alloc_bootmem_huge_page(h, NUMA_NO_NODE)) |
aa888a74 | 3294 | break; |
0c397dae | 3295 | } else if (!alloc_pool_huge_page(h, |
f60858f9 MK |
3296 | &node_states[N_MEMORY], |
3297 | node_alloc_noretry)) | |
1da177e4 | 3298 | break; |
69ed779a | 3299 | cond_resched(); |
1da177e4 | 3300 | } |
d715cf80 LH |
3301 | if (i < h->max_huge_pages) { |
3302 | char buf[32]; | |
3303 | ||
c6247f72 | 3304 | string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32); |
d715cf80 LH |
3305 | pr_warn("HugeTLB: allocating %lu of page size %s failed. Only allocated %lu hugepages.\n", |
3306 | h->max_huge_pages, buf, i); | |
3307 | h->max_huge_pages = i; | |
3308 | } | |
f60858f9 | 3309 | kfree(node_alloc_noretry); |
e5ff2159 AK |
3310 | } |
3311 | ||
3312 | static void __init hugetlb_init_hstates(void) | |
3313 | { | |
79dfc695 | 3314 | struct hstate *h, *h2; |
e5ff2159 AK |
3315 | |
3316 | for_each_hstate(h) { | |
8faa8b07 | 3317 | /* oversize hugepages were init'ed in early boot */ |
bae7f4ae | 3318 | if (!hstate_is_gigantic(h)) |
8faa8b07 | 3319 | hugetlb_hstate_alloc_pages(h); |
79dfc695 MK |
3320 | |
3321 | /* | |
3322 | * Set demote order for each hstate. Note that | |
3323 | * h->demote_order is initially 0. | |
3324 | * - We can not demote gigantic pages if runtime freeing | |
3325 | * is not supported, so skip this. | |
a01f4390 MK |
3326 | * - If CMA allocation is possible, we can not demote |
3327 | * HUGETLB_PAGE_ORDER or smaller size pages. | |
79dfc695 MK |
3328 | */ |
3329 | if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) | |
3330 | continue; | |
a01f4390 MK |
3331 | if (hugetlb_cma_size && h->order <= HUGETLB_PAGE_ORDER) |
3332 | continue; | |
79dfc695 MK |
3333 | for_each_hstate(h2) { |
3334 | if (h2 == h) | |
3335 | continue; | |
3336 | if (h2->order < h->order && | |
3337 | h2->order > h->demote_order) | |
3338 | h->demote_order = h2->order; | |
3339 | } | |
e5ff2159 AK |
3340 | } |
3341 | } | |
3342 | ||
3343 | static void __init report_hugepages(void) | |
3344 | { | |
3345 | struct hstate *h; | |
3346 | ||
3347 | for_each_hstate(h) { | |
4abd32db | 3348 | char buf[32]; |
c6247f72 MW |
3349 | |
3350 | string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32); | |
6213834c | 3351 | pr_info("HugeTLB: registered %s page size, pre-allocated %ld pages\n", |
c6247f72 | 3352 | buf, h->free_huge_pages); |
6213834c MS |
3353 | pr_info("HugeTLB: %d KiB vmemmap can be freed for a %s page\n", |
3354 | hugetlb_vmemmap_optimizable_size(h) / SZ_1K, buf); | |
e5ff2159 AK |
3355 | } |
3356 | } | |
3357 | ||
1da177e4 | 3358 | #ifdef CONFIG_HIGHMEM |
6ae11b27 LS |
3359 | static void try_to_free_low(struct hstate *h, unsigned long count, |
3360 | nodemask_t *nodes_allowed) | |
1da177e4 | 3361 | { |
4415cc8d | 3362 | int i; |
1121828a | 3363 | LIST_HEAD(page_list); |
4415cc8d | 3364 | |
9487ca60 | 3365 | lockdep_assert_held(&hugetlb_lock); |
bae7f4ae | 3366 | if (hstate_is_gigantic(h)) |
aa888a74 AK |
3367 | return; |
3368 | ||
1121828a MK |
3369 | /* |
3370 | * Collect pages to be freed on a list, and free after dropping lock | |
3371 | */ | |
6ae11b27 | 3372 | for_each_node_mask(i, *nodes_allowed) { |
10c6ec49 | 3373 | struct page *page, *next; |
a5516438 AK |
3374 | struct list_head *freel = &h->hugepage_freelists[i]; |
3375 | list_for_each_entry_safe(page, next, freel, lru) { | |
3376 | if (count >= h->nr_huge_pages) | |
1121828a | 3377 | goto out; |
1da177e4 LT |
3378 | if (PageHighMem(page)) |
3379 | continue; | |
cfd5082b | 3380 | remove_hugetlb_folio(h, page_folio(page), false); |
1121828a | 3381 | list_add(&page->lru, &page_list); |
1da177e4 LT |
3382 | } |
3383 | } | |
1121828a MK |
3384 | |
3385 | out: | |
db71ef79 | 3386 | spin_unlock_irq(&hugetlb_lock); |
10c6ec49 | 3387 | update_and_free_pages_bulk(h, &page_list); |
db71ef79 | 3388 | spin_lock_irq(&hugetlb_lock); |
1da177e4 LT |
3389 | } |
3390 | #else | |
6ae11b27 LS |
3391 | static inline void try_to_free_low(struct hstate *h, unsigned long count, |
3392 | nodemask_t *nodes_allowed) | |
1da177e4 LT |
3393 | { |
3394 | } | |
3395 | #endif | |
3396 | ||
20a0307c WF |
3397 | /* |
3398 | * Increment or decrement surplus_huge_pages. Keep node-specific counters | |
3399 | * balanced by operating on them in a round-robin fashion. | |
3400 | * Returns 1 if an adjustment was made. | |
3401 | */ | |
6ae11b27 LS |
3402 | static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed, |
3403 | int delta) | |
20a0307c | 3404 | { |
b2261026 | 3405 | int nr_nodes, node; |
20a0307c | 3406 | |
9487ca60 | 3407 | lockdep_assert_held(&hugetlb_lock); |
20a0307c | 3408 | VM_BUG_ON(delta != -1 && delta != 1); |
20a0307c | 3409 | |
b2261026 JK |
3410 | if (delta < 0) { |
3411 | for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { | |
3412 | if (h->surplus_huge_pages_node[node]) | |
3413 | goto found; | |
e8c5c824 | 3414 | } |
b2261026 JK |
3415 | } else { |
3416 | for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { | |
3417 | if (h->surplus_huge_pages_node[node] < | |
3418 | h->nr_huge_pages_node[node]) | |
3419 | goto found; | |
e8c5c824 | 3420 | } |
b2261026 JK |
3421 | } |
3422 | return 0; | |
20a0307c | 3423 | |
b2261026 JK |
3424 | found: |
3425 | h->surplus_huge_pages += delta; | |
3426 | h->surplus_huge_pages_node[node] += delta; | |
3427 | return 1; | |
20a0307c WF |
3428 | } |
3429 | ||
a5516438 | 3430 | #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages) |
fd875dca | 3431 | static int set_max_huge_pages(struct hstate *h, unsigned long count, int nid, |
4eb0716e | 3432 | nodemask_t *nodes_allowed) |
1da177e4 | 3433 | { |
7893d1d5 | 3434 | unsigned long min_count, ret; |
10c6ec49 MK |
3435 | struct page *page; |
3436 | LIST_HEAD(page_list); | |
f60858f9 MK |
3437 | NODEMASK_ALLOC(nodemask_t, node_alloc_noretry, GFP_KERNEL); |
3438 | ||
3439 | /* | |
3440 | * Bit mask controlling how hard we retry per-node allocations. | |
3441 | * If we can not allocate the bit mask, do not attempt to allocate | |
3442 | * the requested huge pages. | |
3443 | */ | |
3444 | if (node_alloc_noretry) | |
3445 | nodes_clear(*node_alloc_noretry); | |
3446 | else | |
3447 | return -ENOMEM; | |
1da177e4 | 3448 | |
29383967 MK |
3449 | /* |
3450 | * resize_lock mutex prevents concurrent adjustments to number of | |
3451 | * pages in hstate via the proc/sysfs interfaces. | |
3452 | */ | |
3453 | mutex_lock(&h->resize_lock); | |
b65d4adb | 3454 | flush_free_hpage_work(h); |
db71ef79 | 3455 | spin_lock_irq(&hugetlb_lock); |
4eb0716e | 3456 | |
fd875dca MK |
3457 | /* |
3458 | * Check for a node specific request. | |
3459 | * Changing node specific huge page count may require a corresponding | |
3460 | * change to the global count. In any case, the passed node mask | |
3461 | * (nodes_allowed) will restrict alloc/free to the specified node. | |
3462 | */ | |
3463 | if (nid != NUMA_NO_NODE) { | |
3464 | unsigned long old_count = count; | |
3465 | ||
3466 | count += h->nr_huge_pages - h->nr_huge_pages_node[nid]; | |
3467 | /* | |
3468 | * User may have specified a large count value which caused the | |
3469 | * above calculation to overflow. In this case, they wanted | |
3470 | * to allocate as many huge pages as possible. Set count to | |
3471 | * largest possible value to align with their intention. | |
3472 | */ | |
3473 | if (count < old_count) | |
3474 | count = ULONG_MAX; | |
3475 | } | |
3476 | ||
4eb0716e AG |
3477 | /* |
3478 | * Gigantic pages runtime allocation depend on the capability for large | |
3479 | * page range allocation. | |
3480 | * If the system does not provide this feature, return an error when | |
3481 | * the user tries to allocate gigantic pages but let the user free the | |
3482 | * boottime allocated gigantic pages. | |
3483 | */ | |
3484 | if (hstate_is_gigantic(h) && !IS_ENABLED(CONFIG_CONTIG_ALLOC)) { | |
3485 | if (count > persistent_huge_pages(h)) { | |
db71ef79 | 3486 | spin_unlock_irq(&hugetlb_lock); |
29383967 | 3487 | mutex_unlock(&h->resize_lock); |
f60858f9 | 3488 | NODEMASK_FREE(node_alloc_noretry); |
4eb0716e AG |
3489 | return -EINVAL; |
3490 | } | |
3491 | /* Fall through to decrease pool */ | |
3492 | } | |
aa888a74 | 3493 | |
7893d1d5 AL |
3494 | /* |
3495 | * Increase the pool size | |
3496 | * First take pages out of surplus state. Then make up the | |
3497 | * remaining difference by allocating fresh huge pages. | |
d1c3fb1f | 3498 | * |
0c397dae | 3499 | * We might race with alloc_surplus_huge_page() here and be unable |
d1c3fb1f NA |
3500 | * to convert a surplus huge page to a normal huge page. That is |
3501 | * not critical, though, it just means the overall size of the | |
3502 | * pool might be one hugepage larger than it needs to be, but | |
3503 | * within all the constraints specified by the sysctls. | |
7893d1d5 | 3504 | */ |
a5516438 | 3505 | while (h->surplus_huge_pages && count > persistent_huge_pages(h)) { |
6ae11b27 | 3506 | if (!adjust_pool_surplus(h, nodes_allowed, -1)) |
7893d1d5 AL |
3507 | break; |
3508 | } | |
3509 | ||
a5516438 | 3510 | while (count > persistent_huge_pages(h)) { |
7893d1d5 AL |
3511 | /* |
3512 | * If this allocation races such that we no longer need the | |
3513 | * page, free_huge_page will handle it by freeing the page | |
3514 | * and reducing the surplus. | |
3515 | */ | |
db71ef79 | 3516 | spin_unlock_irq(&hugetlb_lock); |
649920c6 JH |
3517 | |
3518 | /* yield cpu to avoid soft lockup */ | |
3519 | cond_resched(); | |
3520 | ||
f60858f9 MK |
3521 | ret = alloc_pool_huge_page(h, nodes_allowed, |
3522 | node_alloc_noretry); | |
db71ef79 | 3523 | spin_lock_irq(&hugetlb_lock); |
7893d1d5 AL |
3524 | if (!ret) |
3525 | goto out; | |
3526 | ||
536240f2 MG |
3527 | /* Bail for signals. Probably ctrl-c from user */ |
3528 | if (signal_pending(current)) | |
3529 | goto out; | |
7893d1d5 | 3530 | } |
7893d1d5 AL |
3531 | |
3532 | /* | |
3533 | * Decrease the pool size | |
3534 | * First return free pages to the buddy allocator (being careful | |
3535 | * to keep enough around to satisfy reservations). Then place | |
3536 | * pages into surplus state as needed so the pool will shrink | |
3537 | * to the desired size as pages become free. | |
d1c3fb1f NA |
3538 | * |
3539 | * By placing pages into the surplus state independent of the | |
3540 | * overcommit value, we are allowing the surplus pool size to | |
3541 | * exceed overcommit. There are few sane options here. Since | |
0c397dae | 3542 | * alloc_surplus_huge_page() is checking the global counter, |
d1c3fb1f NA |
3543 | * though, we'll note that we're not allowed to exceed surplus |
3544 | * and won't grow the pool anywhere else. Not until one of the | |
3545 | * sysctls are changed, or the surplus pages go out of use. | |
7893d1d5 | 3546 | */ |
a5516438 | 3547 | min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages; |
6b0c880d | 3548 | min_count = max(count, min_count); |
6ae11b27 | 3549 | try_to_free_low(h, min_count, nodes_allowed); |
10c6ec49 MK |
3550 | |
3551 | /* | |
3552 | * Collect pages to be removed on list without dropping lock | |
3553 | */ | |
a5516438 | 3554 | while (min_count < persistent_huge_pages(h)) { |
10c6ec49 MK |
3555 | page = remove_pool_huge_page(h, nodes_allowed, 0); |
3556 | if (!page) | |
1da177e4 | 3557 | break; |
10c6ec49 MK |
3558 | |
3559 | list_add(&page->lru, &page_list); | |
1da177e4 | 3560 | } |
10c6ec49 | 3561 | /* free the pages after dropping lock */ |
db71ef79 | 3562 | spin_unlock_irq(&hugetlb_lock); |
10c6ec49 | 3563 | update_and_free_pages_bulk(h, &page_list); |
b65d4adb | 3564 | flush_free_hpage_work(h); |
db71ef79 | 3565 | spin_lock_irq(&hugetlb_lock); |
10c6ec49 | 3566 | |
a5516438 | 3567 | while (count < persistent_huge_pages(h)) { |
6ae11b27 | 3568 | if (!adjust_pool_surplus(h, nodes_allowed, 1)) |
7893d1d5 AL |
3569 | break; |
3570 | } | |
3571 | out: | |
4eb0716e | 3572 | h->max_huge_pages = persistent_huge_pages(h); |
db71ef79 | 3573 | spin_unlock_irq(&hugetlb_lock); |
29383967 | 3574 | mutex_unlock(&h->resize_lock); |
4eb0716e | 3575 | |
f60858f9 MK |
3576 | NODEMASK_FREE(node_alloc_noretry); |
3577 | ||
4eb0716e | 3578 | return 0; |
1da177e4 LT |
3579 | } |
3580 | ||
8531fc6f MK |
3581 | static int demote_free_huge_page(struct hstate *h, struct page *page) |
3582 | { | |
3583 | int i, nid = page_to_nid(page); | |
3584 | struct hstate *target_hstate; | |
911565b8 | 3585 | struct folio *folio = page_folio(page); |
31731452 | 3586 | struct page *subpage; |
8531fc6f MK |
3587 | int rc = 0; |
3588 | ||
3589 | target_hstate = size_to_hstate(PAGE_SIZE << h->demote_order); | |
3590 | ||
cfd5082b | 3591 | remove_hugetlb_folio_for_demote(h, folio, false); |
8531fc6f MK |
3592 | spin_unlock_irq(&hugetlb_lock); |
3593 | ||
6213834c | 3594 | rc = hugetlb_vmemmap_restore(h, page); |
8531fc6f MK |
3595 | if (rc) { |
3596 | /* Allocation of vmemmmap failed, we can not demote page */ | |
3597 | spin_lock_irq(&hugetlb_lock); | |
3598 | set_page_refcounted(page); | |
2f6c57d6 | 3599 | add_hugetlb_folio(h, page_folio(page), false); |
8531fc6f MK |
3600 | return rc; |
3601 | } | |
3602 | ||
3603 | /* | |
911565b8 | 3604 | * Use destroy_compound_hugetlb_folio_for_demote for all huge page |
8531fc6f MK |
3605 | * sizes as it will not ref count pages. |
3606 | */ | |
911565b8 | 3607 | destroy_compound_hugetlb_folio_for_demote(folio, huge_page_order(h)); |
8531fc6f MK |
3608 | |
3609 | /* | |
3610 | * Taking target hstate mutex synchronizes with set_max_huge_pages. | |
3611 | * Without the mutex, pages added to target hstate could be marked | |
3612 | * as surplus. | |
3613 | * | |
3614 | * Note that we already hold h->resize_lock. To prevent deadlock, | |
3615 | * use the convention of always taking larger size hstate mutex first. | |
3616 | */ | |
3617 | mutex_lock(&target_hstate->resize_lock); | |
3618 | for (i = 0; i < pages_per_huge_page(h); | |
3619 | i += pages_per_huge_page(target_hstate)) { | |
31731452 | 3620 | subpage = nth_page(page, i); |
d1c60955 | 3621 | folio = page_folio(subpage); |
8531fc6f | 3622 | if (hstate_is_gigantic(target_hstate)) |
d1c60955 | 3623 | prep_compound_gigantic_folio_for_demote(folio, |
8531fc6f MK |
3624 | target_hstate->order); |
3625 | else | |
31731452 DB |
3626 | prep_compound_page(subpage, target_hstate->order); |
3627 | set_page_private(subpage, 0); | |
d1c60955 | 3628 | prep_new_hugetlb_folio(target_hstate, folio, nid); |
2b21624f | 3629 | free_huge_page(subpage); |
8531fc6f MK |
3630 | } |
3631 | mutex_unlock(&target_hstate->resize_lock); | |
3632 | ||
3633 | spin_lock_irq(&hugetlb_lock); | |
3634 | ||
3635 | /* | |
3636 | * Not absolutely necessary, but for consistency update max_huge_pages | |
3637 | * based on pool changes for the demoted page. | |
3638 | */ | |
3639 | h->max_huge_pages--; | |
a43a83c7 ML |
3640 | target_hstate->max_huge_pages += |
3641 | pages_per_huge_page(h) / pages_per_huge_page(target_hstate); | |
8531fc6f MK |
3642 | |
3643 | return rc; | |
3644 | } | |
3645 | ||
79dfc695 MK |
3646 | static int demote_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed) |
3647 | __must_hold(&hugetlb_lock) | |
3648 | { | |
8531fc6f MK |
3649 | int nr_nodes, node; |
3650 | struct page *page; | |
79dfc695 MK |
3651 | |
3652 | lockdep_assert_held(&hugetlb_lock); | |
3653 | ||
3654 | /* We should never get here if no demote order */ | |
3655 | if (!h->demote_order) { | |
3656 | pr_warn("HugeTLB: NULL demote order passed to demote_pool_huge_page.\n"); | |
3657 | return -EINVAL; /* internal error */ | |
3658 | } | |
3659 | ||
8531fc6f | 3660 | for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { |
5a317412 MK |
3661 | list_for_each_entry(page, &h->hugepage_freelists[node], lru) { |
3662 | if (PageHWPoison(page)) | |
3663 | continue; | |
3664 | ||
3665 | return demote_free_huge_page(h, page); | |
8531fc6f MK |
3666 | } |
3667 | } | |
3668 | ||
5a317412 MK |
3669 | /* |
3670 | * Only way to get here is if all pages on free lists are poisoned. | |
3671 | * Return -EBUSY so that caller will not retry. | |
3672 | */ | |
3673 | return -EBUSY; | |
79dfc695 MK |
3674 | } |
3675 | ||
a3437870 NA |
3676 | #define HSTATE_ATTR_RO(_name) \ |
3677 | static struct kobj_attribute _name##_attr = __ATTR_RO(_name) | |
3678 | ||
79dfc695 MK |
3679 | #define HSTATE_ATTR_WO(_name) \ |
3680 | static struct kobj_attribute _name##_attr = __ATTR_WO(_name) | |
3681 | ||
a3437870 | 3682 | #define HSTATE_ATTR(_name) \ |
98bc26ac | 3683 | static struct kobj_attribute _name##_attr = __ATTR_RW(_name) |
a3437870 NA |
3684 | |
3685 | static struct kobject *hugepages_kobj; | |
3686 | static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; | |
3687 | ||
9a305230 LS |
3688 | static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp); |
3689 | ||
3690 | static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp) | |
a3437870 NA |
3691 | { |
3692 | int i; | |
9a305230 | 3693 | |
a3437870 | 3694 | for (i = 0; i < HUGE_MAX_HSTATE; i++) |
9a305230 LS |
3695 | if (hstate_kobjs[i] == kobj) { |
3696 | if (nidp) | |
3697 | *nidp = NUMA_NO_NODE; | |
a3437870 | 3698 | return &hstates[i]; |
9a305230 LS |
3699 | } |
3700 | ||
3701 | return kobj_to_node_hstate(kobj, nidp); | |
a3437870 NA |
3702 | } |
3703 | ||
06808b08 | 3704 | static ssize_t nr_hugepages_show_common(struct kobject *kobj, |
a3437870 NA |
3705 | struct kobj_attribute *attr, char *buf) |
3706 | { | |
9a305230 LS |
3707 | struct hstate *h; |
3708 | unsigned long nr_huge_pages; | |
3709 | int nid; | |
3710 | ||
3711 | h = kobj_to_hstate(kobj, &nid); | |
3712 | if (nid == NUMA_NO_NODE) | |
3713 | nr_huge_pages = h->nr_huge_pages; | |
3714 | else | |
3715 | nr_huge_pages = h->nr_huge_pages_node[nid]; | |
3716 | ||
ae7a927d | 3717 | return sysfs_emit(buf, "%lu\n", nr_huge_pages); |
a3437870 | 3718 | } |
adbe8726 | 3719 | |
238d3c13 DR |
3720 | static ssize_t __nr_hugepages_store_common(bool obey_mempolicy, |
3721 | struct hstate *h, int nid, | |
3722 | unsigned long count, size_t len) | |
a3437870 NA |
3723 | { |
3724 | int err; | |
2d0adf7e | 3725 | nodemask_t nodes_allowed, *n_mask; |
a3437870 | 3726 | |
2d0adf7e OS |
3727 | if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) |
3728 | return -EINVAL; | |
adbe8726 | 3729 | |
9a305230 LS |
3730 | if (nid == NUMA_NO_NODE) { |
3731 | /* | |
3732 | * global hstate attribute | |
3733 | */ | |
3734 | if (!(obey_mempolicy && | |
2d0adf7e OS |
3735 | init_nodemask_of_mempolicy(&nodes_allowed))) |
3736 | n_mask = &node_states[N_MEMORY]; | |
3737 | else | |
3738 | n_mask = &nodes_allowed; | |
3739 | } else { | |
9a305230 | 3740 | /* |
fd875dca MK |
3741 | * Node specific request. count adjustment happens in |
3742 | * set_max_huge_pages() after acquiring hugetlb_lock. | |
9a305230 | 3743 | */ |
2d0adf7e OS |
3744 | init_nodemask_of_node(&nodes_allowed, nid); |
3745 | n_mask = &nodes_allowed; | |
fd875dca | 3746 | } |
9a305230 | 3747 | |
2d0adf7e | 3748 | err = set_max_huge_pages(h, count, nid, n_mask); |
06808b08 | 3749 | |
4eb0716e | 3750 | return err ? err : len; |
06808b08 LS |
3751 | } |
3752 | ||
238d3c13 DR |
3753 | static ssize_t nr_hugepages_store_common(bool obey_mempolicy, |
3754 | struct kobject *kobj, const char *buf, | |
3755 | size_t len) | |
3756 | { | |
3757 | struct hstate *h; | |
3758 | unsigned long count; | |
3759 | int nid; | |
3760 | int err; | |
3761 | ||
3762 | err = kstrtoul(buf, 10, &count); | |
3763 | if (err) | |
3764 | return err; | |
3765 | ||
3766 | h = kobj_to_hstate(kobj, &nid); | |
3767 | return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len); | |
3768 | } | |
3769 | ||
06808b08 LS |
3770 | static ssize_t nr_hugepages_show(struct kobject *kobj, |
3771 | struct kobj_attribute *attr, char *buf) | |
3772 | { | |
3773 | return nr_hugepages_show_common(kobj, attr, buf); | |
3774 | } | |
3775 | ||
3776 | static ssize_t nr_hugepages_store(struct kobject *kobj, | |
3777 | struct kobj_attribute *attr, const char *buf, size_t len) | |
3778 | { | |
238d3c13 | 3779 | return nr_hugepages_store_common(false, kobj, buf, len); |
a3437870 NA |
3780 | } |
3781 | HSTATE_ATTR(nr_hugepages); | |
3782 | ||
06808b08 LS |
3783 | #ifdef CONFIG_NUMA |
3784 | ||
3785 | /* | |
3786 | * hstate attribute for optionally mempolicy-based constraint on persistent | |
3787 | * huge page alloc/free. | |
3788 | */ | |
3789 | static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj, | |
ae7a927d JP |
3790 | struct kobj_attribute *attr, |
3791 | char *buf) | |
06808b08 LS |
3792 | { |
3793 | return nr_hugepages_show_common(kobj, attr, buf); | |
3794 | } | |
3795 | ||
3796 | static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj, | |
3797 | struct kobj_attribute *attr, const char *buf, size_t len) | |
3798 | { | |
238d3c13 | 3799 | return nr_hugepages_store_common(true, kobj, buf, len); |
06808b08 LS |
3800 | } |
3801 | HSTATE_ATTR(nr_hugepages_mempolicy); | |
3802 | #endif | |
3803 | ||
3804 | ||
a3437870 NA |
3805 | static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj, |
3806 | struct kobj_attribute *attr, char *buf) | |
3807 | { | |
9a305230 | 3808 | struct hstate *h = kobj_to_hstate(kobj, NULL); |
ae7a927d | 3809 | return sysfs_emit(buf, "%lu\n", h->nr_overcommit_huge_pages); |
a3437870 | 3810 | } |
adbe8726 | 3811 | |
a3437870 NA |
3812 | static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj, |
3813 | struct kobj_attribute *attr, const char *buf, size_t count) | |
3814 | { | |
3815 | int err; | |
3816 | unsigned long input; | |
9a305230 | 3817 | struct hstate *h = kobj_to_hstate(kobj, NULL); |
a3437870 | 3818 | |
bae7f4ae | 3819 | if (hstate_is_gigantic(h)) |
adbe8726 EM |
3820 | return -EINVAL; |
3821 | ||
3dbb95f7 | 3822 | err = kstrtoul(buf, 10, &input); |
a3437870 | 3823 | if (err) |
73ae31e5 | 3824 | return err; |
a3437870 | 3825 | |
db71ef79 | 3826 | spin_lock_irq(&hugetlb_lock); |
a3437870 | 3827 | h->nr_overcommit_huge_pages = input; |
db71ef79 | 3828 | spin_unlock_irq(&hugetlb_lock); |
a3437870 NA |
3829 | |
3830 | return count; | |
3831 | } | |
3832 | HSTATE_ATTR(nr_overcommit_hugepages); | |
3833 | ||
3834 | static ssize_t free_hugepages_show(struct kobject *kobj, | |
3835 | struct kobj_attribute *attr, char *buf) | |
3836 | { | |
9a305230 LS |
3837 | struct hstate *h; |
3838 | unsigned long free_huge_pages; | |
3839 | int nid; | |
3840 | ||
3841 | h = kobj_to_hstate(kobj, &nid); | |
3842 | if (nid == NUMA_NO_NODE) | |
3843 | free_huge_pages = h->free_huge_pages; | |
3844 | else | |
3845 | free_huge_pages = h->free_huge_pages_node[nid]; | |
3846 | ||
ae7a927d | 3847 | return sysfs_emit(buf, "%lu\n", free_huge_pages); |
a3437870 NA |
3848 | } |
3849 | HSTATE_ATTR_RO(free_hugepages); | |
3850 | ||
3851 | static ssize_t resv_hugepages_show(struct kobject *kobj, | |
3852 | struct kobj_attribute *attr, char *buf) | |
3853 | { | |
9a305230 | 3854 | struct hstate *h = kobj_to_hstate(kobj, NULL); |
ae7a927d | 3855 | return sysfs_emit(buf, "%lu\n", h->resv_huge_pages); |
a3437870 NA |
3856 | } |
3857 | HSTATE_ATTR_RO(resv_hugepages); | |
3858 | ||
3859 | static ssize_t surplus_hugepages_show(struct kobject *kobj, | |
3860 | struct kobj_attribute *attr, char *buf) | |
3861 | { | |
9a305230 LS |
3862 | struct hstate *h; |
3863 | unsigned long surplus_huge_pages; | |
3864 | int nid; | |
3865 | ||
3866 | h = kobj_to_hstate(kobj, &nid); | |
3867 | if (nid == NUMA_NO_NODE) | |
3868 | surplus_huge_pages = h->surplus_huge_pages; | |
3869 | else | |
3870 | surplus_huge_pages = h->surplus_huge_pages_node[nid]; | |
3871 | ||
ae7a927d | 3872 | return sysfs_emit(buf, "%lu\n", surplus_huge_pages); |
a3437870 NA |
3873 | } |
3874 | HSTATE_ATTR_RO(surplus_hugepages); | |
3875 | ||
79dfc695 MK |
3876 | static ssize_t demote_store(struct kobject *kobj, |
3877 | struct kobj_attribute *attr, const char *buf, size_t len) | |
3878 | { | |
3879 | unsigned long nr_demote; | |
3880 | unsigned long nr_available; | |
3881 | nodemask_t nodes_allowed, *n_mask; | |
3882 | struct hstate *h; | |
8eeda55f | 3883 | int err; |
79dfc695 MK |
3884 | int nid; |
3885 | ||
3886 | err = kstrtoul(buf, 10, &nr_demote); | |
3887 | if (err) | |
3888 | return err; | |
3889 | h = kobj_to_hstate(kobj, &nid); | |
3890 | ||
3891 | if (nid != NUMA_NO_NODE) { | |
3892 | init_nodemask_of_node(&nodes_allowed, nid); | |
3893 | n_mask = &nodes_allowed; | |
3894 | } else { | |
3895 | n_mask = &node_states[N_MEMORY]; | |
3896 | } | |
3897 | ||
3898 | /* Synchronize with other sysfs operations modifying huge pages */ | |
3899 | mutex_lock(&h->resize_lock); | |
3900 | spin_lock_irq(&hugetlb_lock); | |
3901 | ||
3902 | while (nr_demote) { | |
3903 | /* | |
3904 | * Check for available pages to demote each time thorough the | |
3905 | * loop as demote_pool_huge_page will drop hugetlb_lock. | |
79dfc695 MK |
3906 | */ |
3907 | if (nid != NUMA_NO_NODE) | |
3908 | nr_available = h->free_huge_pages_node[nid]; | |
3909 | else | |
3910 | nr_available = h->free_huge_pages; | |
3911 | nr_available -= h->resv_huge_pages; | |
3912 | if (!nr_available) | |
3913 | break; | |
3914 | ||
3915 | err = demote_pool_huge_page(h, n_mask); | |
3916 | if (err) | |
3917 | break; | |
3918 | ||
3919 | nr_demote--; | |
3920 | } | |
3921 | ||
3922 | spin_unlock_irq(&hugetlb_lock); | |
3923 | mutex_unlock(&h->resize_lock); | |
3924 | ||
3925 | if (err) | |
3926 | return err; | |
3927 | return len; | |
3928 | } | |
3929 | HSTATE_ATTR_WO(demote); | |
3930 | ||
3931 | static ssize_t demote_size_show(struct kobject *kobj, | |
3932 | struct kobj_attribute *attr, char *buf) | |
3933 | { | |
12658abf | 3934 | struct hstate *h = kobj_to_hstate(kobj, NULL); |
79dfc695 MK |
3935 | unsigned long demote_size = (PAGE_SIZE << h->demote_order) / SZ_1K; |
3936 | ||
3937 | return sysfs_emit(buf, "%lukB\n", demote_size); | |
3938 | } | |
3939 | ||
3940 | static ssize_t demote_size_store(struct kobject *kobj, | |
3941 | struct kobj_attribute *attr, | |
3942 | const char *buf, size_t count) | |
3943 | { | |
3944 | struct hstate *h, *demote_hstate; | |
3945 | unsigned long demote_size; | |
3946 | unsigned int demote_order; | |
79dfc695 MK |
3947 | |
3948 | demote_size = (unsigned long)memparse(buf, NULL); | |
3949 | ||
3950 | demote_hstate = size_to_hstate(demote_size); | |
3951 | if (!demote_hstate) | |
3952 | return -EINVAL; | |
3953 | demote_order = demote_hstate->order; | |
a01f4390 MK |
3954 | if (demote_order < HUGETLB_PAGE_ORDER) |
3955 | return -EINVAL; | |
79dfc695 MK |
3956 | |
3957 | /* demote order must be smaller than hstate order */ | |
12658abf | 3958 | h = kobj_to_hstate(kobj, NULL); |
79dfc695 MK |
3959 | if (demote_order >= h->order) |
3960 | return -EINVAL; | |
3961 | ||
3962 | /* resize_lock synchronizes access to demote size and writes */ | |
3963 | mutex_lock(&h->resize_lock); | |
3964 | h->demote_order = demote_order; | |
3965 | mutex_unlock(&h->resize_lock); | |
3966 | ||
3967 | return count; | |
3968 | } | |
3969 | HSTATE_ATTR(demote_size); | |
3970 | ||
a3437870 NA |
3971 | static struct attribute *hstate_attrs[] = { |
3972 | &nr_hugepages_attr.attr, | |
3973 | &nr_overcommit_hugepages_attr.attr, | |
3974 | &free_hugepages_attr.attr, | |
3975 | &resv_hugepages_attr.attr, | |
3976 | &surplus_hugepages_attr.attr, | |
06808b08 LS |
3977 | #ifdef CONFIG_NUMA |
3978 | &nr_hugepages_mempolicy_attr.attr, | |
3979 | #endif | |
a3437870 NA |
3980 | NULL, |
3981 | }; | |
3982 | ||
67e5ed96 | 3983 | static const struct attribute_group hstate_attr_group = { |
a3437870 NA |
3984 | .attrs = hstate_attrs, |
3985 | }; | |
3986 | ||
79dfc695 MK |
3987 | static struct attribute *hstate_demote_attrs[] = { |
3988 | &demote_size_attr.attr, | |
3989 | &demote_attr.attr, | |
3990 | NULL, | |
3991 | }; | |
3992 | ||
3993 | static const struct attribute_group hstate_demote_attr_group = { | |
3994 | .attrs = hstate_demote_attrs, | |
3995 | }; | |
3996 | ||
094e9539 JM |
3997 | static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent, |
3998 | struct kobject **hstate_kobjs, | |
67e5ed96 | 3999 | const struct attribute_group *hstate_attr_group) |
a3437870 NA |
4000 | { |
4001 | int retval; | |
972dc4de | 4002 | int hi = hstate_index(h); |
a3437870 | 4003 | |
9a305230 LS |
4004 | hstate_kobjs[hi] = kobject_create_and_add(h->name, parent); |
4005 | if (!hstate_kobjs[hi]) | |
a3437870 NA |
4006 | return -ENOMEM; |
4007 | ||
9a305230 | 4008 | retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group); |
cc2205a6 | 4009 | if (retval) { |
9a305230 | 4010 | kobject_put(hstate_kobjs[hi]); |
cc2205a6 | 4011 | hstate_kobjs[hi] = NULL; |
3a6bdda0 | 4012 | return retval; |
cc2205a6 | 4013 | } |
a3437870 | 4014 | |
79dfc695 | 4015 | if (h->demote_order) { |
01088a60 ML |
4016 | retval = sysfs_create_group(hstate_kobjs[hi], |
4017 | &hstate_demote_attr_group); | |
4018 | if (retval) { | |
79dfc695 | 4019 | pr_warn("HugeTLB unable to create demote interfaces for %s\n", h->name); |
01088a60 ML |
4020 | sysfs_remove_group(hstate_kobjs[hi], hstate_attr_group); |
4021 | kobject_put(hstate_kobjs[hi]); | |
4022 | hstate_kobjs[hi] = NULL; | |
4023 | return retval; | |
4024 | } | |
79dfc695 MK |
4025 | } |
4026 | ||
01088a60 | 4027 | return 0; |
a3437870 NA |
4028 | } |
4029 | ||
9a305230 | 4030 | #ifdef CONFIG_NUMA |
a4a00b45 | 4031 | static bool hugetlb_sysfs_initialized __ro_after_init; |
9a305230 LS |
4032 | |
4033 | /* | |
4034 | * node_hstate/s - associate per node hstate attributes, via their kobjects, | |
10fbcf4c KS |
4035 | * with node devices in node_devices[] using a parallel array. The array |
4036 | * index of a node device or _hstate == node id. | |
4037 | * This is here to avoid any static dependency of the node device driver, in | |
9a305230 LS |
4038 | * the base kernel, on the hugetlb module. |
4039 | */ | |
4040 | struct node_hstate { | |
4041 | struct kobject *hugepages_kobj; | |
4042 | struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; | |
4043 | }; | |
b4e289a6 | 4044 | static struct node_hstate node_hstates[MAX_NUMNODES]; |
9a305230 LS |
4045 | |
4046 | /* | |
10fbcf4c | 4047 | * A subset of global hstate attributes for node devices |
9a305230 LS |
4048 | */ |
4049 | static struct attribute *per_node_hstate_attrs[] = { | |
4050 | &nr_hugepages_attr.attr, | |
4051 | &free_hugepages_attr.attr, | |
4052 | &surplus_hugepages_attr.attr, | |
4053 | NULL, | |
4054 | }; | |
4055 | ||
67e5ed96 | 4056 | static const struct attribute_group per_node_hstate_attr_group = { |
9a305230 LS |
4057 | .attrs = per_node_hstate_attrs, |
4058 | }; | |
4059 | ||
4060 | /* | |
10fbcf4c | 4061 | * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj. |
9a305230 LS |
4062 | * Returns node id via non-NULL nidp. |
4063 | */ | |
4064 | static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) | |
4065 | { | |
4066 | int nid; | |
4067 | ||
4068 | for (nid = 0; nid < nr_node_ids; nid++) { | |
4069 | struct node_hstate *nhs = &node_hstates[nid]; | |
4070 | int i; | |
4071 | for (i = 0; i < HUGE_MAX_HSTATE; i++) | |
4072 | if (nhs->hstate_kobjs[i] == kobj) { | |
4073 | if (nidp) | |
4074 | *nidp = nid; | |
4075 | return &hstates[i]; | |
4076 | } | |
4077 | } | |
4078 | ||
4079 | BUG(); | |
4080 | return NULL; | |
4081 | } | |
4082 | ||
4083 | /* | |
10fbcf4c | 4084 | * Unregister hstate attributes from a single node device. |
9a305230 LS |
4085 | * No-op if no hstate attributes attached. |
4086 | */ | |
a4a00b45 | 4087 | void hugetlb_unregister_node(struct node *node) |
9a305230 LS |
4088 | { |
4089 | struct hstate *h; | |
10fbcf4c | 4090 | struct node_hstate *nhs = &node_hstates[node->dev.id]; |
9a305230 LS |
4091 | |
4092 | if (!nhs->hugepages_kobj) | |
9b5e5d0f | 4093 | return; /* no hstate attributes */ |
9a305230 | 4094 | |
972dc4de AK |
4095 | for_each_hstate(h) { |
4096 | int idx = hstate_index(h); | |
01088a60 ML |
4097 | struct kobject *hstate_kobj = nhs->hstate_kobjs[idx]; |
4098 | ||
4099 | if (!hstate_kobj) | |
4100 | continue; | |
4101 | if (h->demote_order) | |
4102 | sysfs_remove_group(hstate_kobj, &hstate_demote_attr_group); | |
4103 | sysfs_remove_group(hstate_kobj, &per_node_hstate_attr_group); | |
4104 | kobject_put(hstate_kobj); | |
4105 | nhs->hstate_kobjs[idx] = NULL; | |
972dc4de | 4106 | } |
9a305230 LS |
4107 | |
4108 | kobject_put(nhs->hugepages_kobj); | |
4109 | nhs->hugepages_kobj = NULL; | |
4110 | } | |
4111 | ||
9a305230 LS |
4112 | |
4113 | /* | |
10fbcf4c | 4114 | * Register hstate attributes for a single node device. |
9a305230 LS |
4115 | * No-op if attributes already registered. |
4116 | */ | |
a4a00b45 | 4117 | void hugetlb_register_node(struct node *node) |
9a305230 LS |
4118 | { |
4119 | struct hstate *h; | |
10fbcf4c | 4120 | struct node_hstate *nhs = &node_hstates[node->dev.id]; |
9a305230 LS |
4121 | int err; |
4122 | ||
a4a00b45 MS |
4123 | if (!hugetlb_sysfs_initialized) |
4124 | return; | |
4125 | ||
9a305230 LS |
4126 | if (nhs->hugepages_kobj) |
4127 | return; /* already allocated */ | |
4128 | ||
4129 | nhs->hugepages_kobj = kobject_create_and_add("hugepages", | |
10fbcf4c | 4130 | &node->dev.kobj); |
9a305230 LS |
4131 | if (!nhs->hugepages_kobj) |
4132 | return; | |
4133 | ||
4134 | for_each_hstate(h) { | |
4135 | err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj, | |
4136 | nhs->hstate_kobjs, | |
4137 | &per_node_hstate_attr_group); | |
4138 | if (err) { | |
282f4214 | 4139 | pr_err("HugeTLB: Unable to add hstate %s for node %d\n", |
ffb22af5 | 4140 | h->name, node->dev.id); |
9a305230 LS |
4141 | hugetlb_unregister_node(node); |
4142 | break; | |
4143 | } | |
4144 | } | |
4145 | } | |
4146 | ||
4147 | /* | |
9b5e5d0f | 4148 | * hugetlb init time: register hstate attributes for all registered node |
10fbcf4c KS |
4149 | * devices of nodes that have memory. All on-line nodes should have |
4150 | * registered their associated device by this time. | |
9a305230 | 4151 | */ |
7d9ca000 | 4152 | static void __init hugetlb_register_all_nodes(void) |
9a305230 LS |
4153 | { |
4154 | int nid; | |
4155 | ||
a4a00b45 | 4156 | for_each_online_node(nid) |
b958d4d0 | 4157 | hugetlb_register_node(node_devices[nid]); |
9a305230 LS |
4158 | } |
4159 | #else /* !CONFIG_NUMA */ | |
4160 | ||
4161 | static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) | |
4162 | { | |
4163 | BUG(); | |
4164 | if (nidp) | |
4165 | *nidp = -1; | |
4166 | return NULL; | |
4167 | } | |
4168 | ||
9a305230 LS |
4169 | static void hugetlb_register_all_nodes(void) { } |
4170 | ||
4171 | #endif | |
4172 | ||
263b8998 ML |
4173 | #ifdef CONFIG_CMA |
4174 | static void __init hugetlb_cma_check(void); | |
4175 | #else | |
4176 | static inline __init void hugetlb_cma_check(void) | |
4177 | { | |
4178 | } | |
4179 | #endif | |
4180 | ||
a4a00b45 MS |
4181 | static void __init hugetlb_sysfs_init(void) |
4182 | { | |
4183 | struct hstate *h; | |
4184 | int err; | |
4185 | ||
4186 | hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj); | |
4187 | if (!hugepages_kobj) | |
4188 | return; | |
4189 | ||
4190 | for_each_hstate(h) { | |
4191 | err = hugetlb_sysfs_add_hstate(h, hugepages_kobj, | |
4192 | hstate_kobjs, &hstate_attr_group); | |
4193 | if (err) | |
4194 | pr_err("HugeTLB: Unable to add hstate %s", h->name); | |
4195 | } | |
4196 | ||
4197 | #ifdef CONFIG_NUMA | |
4198 | hugetlb_sysfs_initialized = true; | |
4199 | #endif | |
4200 | hugetlb_register_all_nodes(); | |
4201 | } | |
4202 | ||
a3437870 NA |
4203 | static int __init hugetlb_init(void) |
4204 | { | |
8382d914 DB |
4205 | int i; |
4206 | ||
d6995da3 MK |
4207 | BUILD_BUG_ON(sizeof_field(struct page, private) * BITS_PER_BYTE < |
4208 | __NR_HPAGEFLAGS); | |
4209 | ||
c2833a5b MK |
4210 | if (!hugepages_supported()) { |
4211 | if (hugetlb_max_hstate || default_hstate_max_huge_pages) | |
4212 | pr_warn("HugeTLB: huge pages not supported, ignoring associated command-line parameters\n"); | |
0ef89d25 | 4213 | return 0; |
c2833a5b | 4214 | } |
a3437870 | 4215 | |
282f4214 MK |
4216 | /* |
4217 | * Make sure HPAGE_SIZE (HUGETLB_PAGE_ORDER) hstate exists. Some | |
4218 | * architectures depend on setup being done here. | |
4219 | */ | |
4220 | hugetlb_add_hstate(HUGETLB_PAGE_ORDER); | |
4221 | if (!parsed_default_hugepagesz) { | |
4222 | /* | |
4223 | * If we did not parse a default huge page size, set | |
4224 | * default_hstate_idx to HPAGE_SIZE hstate. And, if the | |
4225 | * number of huge pages for this default size was implicitly | |
4226 | * specified, set that here as well. | |
4227 | * Note that the implicit setting will overwrite an explicit | |
4228 | * setting. A warning will be printed in this case. | |
4229 | */ | |
4230 | default_hstate_idx = hstate_index(size_to_hstate(HPAGE_SIZE)); | |
4231 | if (default_hstate_max_huge_pages) { | |
4232 | if (default_hstate.max_huge_pages) { | |
4233 | char buf[32]; | |
4234 | ||
4235 | string_get_size(huge_page_size(&default_hstate), | |
4236 | 1, STRING_UNITS_2, buf, 32); | |
4237 | pr_warn("HugeTLB: Ignoring hugepages=%lu associated with %s page size\n", | |
4238 | default_hstate.max_huge_pages, buf); | |
4239 | pr_warn("HugeTLB: Using hugepages=%lu for number of default huge pages\n", | |
4240 | default_hstate_max_huge_pages); | |
4241 | } | |
4242 | default_hstate.max_huge_pages = | |
4243 | default_hstate_max_huge_pages; | |
b5389086 | 4244 | |
0a7a0f6f | 4245 | for_each_online_node(i) |
b5389086 ZY |
4246 | default_hstate.max_huge_pages_node[i] = |
4247 | default_hugepages_in_node[i]; | |
d715cf80 | 4248 | } |
f8b74815 | 4249 | } |
a3437870 | 4250 | |
cf11e85f | 4251 | hugetlb_cma_check(); |
a3437870 | 4252 | hugetlb_init_hstates(); |
aa888a74 | 4253 | gather_bootmem_prealloc(); |
a3437870 NA |
4254 | report_hugepages(); |
4255 | ||
4256 | hugetlb_sysfs_init(); | |
7179e7bf | 4257 | hugetlb_cgroup_file_init(); |
9a305230 | 4258 | |
8382d914 DB |
4259 | #ifdef CONFIG_SMP |
4260 | num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus()); | |
4261 | #else | |
4262 | num_fault_mutexes = 1; | |
4263 | #endif | |
c672c7f2 | 4264 | hugetlb_fault_mutex_table = |
6da2ec56 KC |
4265 | kmalloc_array(num_fault_mutexes, sizeof(struct mutex), |
4266 | GFP_KERNEL); | |
c672c7f2 | 4267 | BUG_ON(!hugetlb_fault_mutex_table); |
8382d914 DB |
4268 | |
4269 | for (i = 0; i < num_fault_mutexes; i++) | |
c672c7f2 | 4270 | mutex_init(&hugetlb_fault_mutex_table[i]); |
a3437870 NA |
4271 | return 0; |
4272 | } | |
3e89e1c5 | 4273 | subsys_initcall(hugetlb_init); |
a3437870 | 4274 | |
ae94da89 MK |
4275 | /* Overwritten by architectures with more huge page sizes */ |
4276 | bool __init __attribute((weak)) arch_hugetlb_valid_size(unsigned long size) | |
9fee021d | 4277 | { |
ae94da89 | 4278 | return size == HPAGE_SIZE; |
9fee021d VT |
4279 | } |
4280 | ||
d00181b9 | 4281 | void __init hugetlb_add_hstate(unsigned int order) |
a3437870 NA |
4282 | { |
4283 | struct hstate *h; | |
8faa8b07 AK |
4284 | unsigned long i; |
4285 | ||
a3437870 | 4286 | if (size_to_hstate(PAGE_SIZE << order)) { |
a3437870 NA |
4287 | return; |
4288 | } | |
47d38344 | 4289 | BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE); |
a3437870 | 4290 | BUG_ON(order == 0); |
47d38344 | 4291 | h = &hstates[hugetlb_max_hstate++]; |
29383967 | 4292 | mutex_init(&h->resize_lock); |
a3437870 | 4293 | h->order = order; |
aca78307 | 4294 | h->mask = ~(huge_page_size(h) - 1); |
8faa8b07 AK |
4295 | for (i = 0; i < MAX_NUMNODES; ++i) |
4296 | INIT_LIST_HEAD(&h->hugepage_freelists[i]); | |
0edaecfa | 4297 | INIT_LIST_HEAD(&h->hugepage_activelist); |
54f18d35 AM |
4298 | h->next_nid_to_alloc = first_memory_node; |
4299 | h->next_nid_to_free = first_memory_node; | |
a3437870 | 4300 | snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB", |
c2c3a60a | 4301 | huge_page_size(h)/SZ_1K); |
8faa8b07 | 4302 | |
a3437870 NA |
4303 | parsed_hstate = h; |
4304 | } | |
4305 | ||
b5389086 ZY |
4306 | bool __init __weak hugetlb_node_alloc_supported(void) |
4307 | { | |
4308 | return true; | |
4309 | } | |
f87442f4 PL |
4310 | |
4311 | static void __init hugepages_clear_pages_in_node(void) | |
4312 | { | |
4313 | if (!hugetlb_max_hstate) { | |
4314 | default_hstate_max_huge_pages = 0; | |
4315 | memset(default_hugepages_in_node, 0, | |
10395680 | 4316 | sizeof(default_hugepages_in_node)); |
f87442f4 PL |
4317 | } else { |
4318 | parsed_hstate->max_huge_pages = 0; | |
4319 | memset(parsed_hstate->max_huge_pages_node, 0, | |
10395680 | 4320 | sizeof(parsed_hstate->max_huge_pages_node)); |
f87442f4 PL |
4321 | } |
4322 | } | |
4323 | ||
282f4214 MK |
4324 | /* |
4325 | * hugepages command line processing | |
4326 | * hugepages normally follows a valid hugepagsz or default_hugepagsz | |
4327 | * specification. If not, ignore the hugepages value. hugepages can also | |
4328 | * be the first huge page command line option in which case it implicitly | |
4329 | * specifies the number of huge pages for the default size. | |
4330 | */ | |
4331 | static int __init hugepages_setup(char *s) | |
a3437870 NA |
4332 | { |
4333 | unsigned long *mhp; | |
8faa8b07 | 4334 | static unsigned long *last_mhp; |
b5389086 ZY |
4335 | int node = NUMA_NO_NODE; |
4336 | int count; | |
4337 | unsigned long tmp; | |
4338 | char *p = s; | |
a3437870 | 4339 | |
9fee021d | 4340 | if (!parsed_valid_hugepagesz) { |
282f4214 | 4341 | pr_warn("HugeTLB: hugepages=%s does not follow a valid hugepagesz, ignoring\n", s); |
9fee021d | 4342 | parsed_valid_hugepagesz = true; |
f81f6e4b | 4343 | return 1; |
9fee021d | 4344 | } |
282f4214 | 4345 | |
a3437870 | 4346 | /* |
282f4214 MK |
4347 | * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter |
4348 | * yet, so this hugepages= parameter goes to the "default hstate". | |
4349 | * Otherwise, it goes with the previously parsed hugepagesz or | |
4350 | * default_hugepagesz. | |
a3437870 | 4351 | */ |
9fee021d | 4352 | else if (!hugetlb_max_hstate) |
a3437870 NA |
4353 | mhp = &default_hstate_max_huge_pages; |
4354 | else | |
4355 | mhp = &parsed_hstate->max_huge_pages; | |
4356 | ||
8faa8b07 | 4357 | if (mhp == last_mhp) { |
282f4214 | 4358 | pr_warn("HugeTLB: hugepages= specified twice without interleaving hugepagesz=, ignoring hugepages=%s\n", s); |
f81f6e4b | 4359 | return 1; |
8faa8b07 AK |
4360 | } |
4361 | ||
b5389086 ZY |
4362 | while (*p) { |
4363 | count = 0; | |
4364 | if (sscanf(p, "%lu%n", &tmp, &count) != 1) | |
4365 | goto invalid; | |
4366 | /* Parameter is node format */ | |
4367 | if (p[count] == ':') { | |
4368 | if (!hugetlb_node_alloc_supported()) { | |
4369 | pr_warn("HugeTLB: architecture can't support node specific alloc, ignoring!\n"); | |
f81f6e4b | 4370 | return 1; |
b5389086 | 4371 | } |
0a7a0f6f | 4372 | if (tmp >= MAX_NUMNODES || !node_online(tmp)) |
e79ce983 | 4373 | goto invalid; |
0a7a0f6f | 4374 | node = array_index_nospec(tmp, MAX_NUMNODES); |
b5389086 | 4375 | p += count + 1; |
b5389086 ZY |
4376 | /* Parse hugepages */ |
4377 | if (sscanf(p, "%lu%n", &tmp, &count) != 1) | |
4378 | goto invalid; | |
4379 | if (!hugetlb_max_hstate) | |
4380 | default_hugepages_in_node[node] = tmp; | |
4381 | else | |
4382 | parsed_hstate->max_huge_pages_node[node] = tmp; | |
4383 | *mhp += tmp; | |
4384 | /* Go to parse next node*/ | |
4385 | if (p[count] == ',') | |
4386 | p += count + 1; | |
4387 | else | |
4388 | break; | |
4389 | } else { | |
4390 | if (p != s) | |
4391 | goto invalid; | |
4392 | *mhp = tmp; | |
4393 | break; | |
4394 | } | |
4395 | } | |
a3437870 | 4396 | |
8faa8b07 AK |
4397 | /* |
4398 | * Global state is always initialized later in hugetlb_init. | |
04adbc3f | 4399 | * But we need to allocate gigantic hstates here early to still |
8faa8b07 AK |
4400 | * use the bootmem allocator. |
4401 | */ | |
04adbc3f | 4402 | if (hugetlb_max_hstate && hstate_is_gigantic(parsed_hstate)) |
8faa8b07 AK |
4403 | hugetlb_hstate_alloc_pages(parsed_hstate); |
4404 | ||
4405 | last_mhp = mhp; | |
4406 | ||
a3437870 | 4407 | return 1; |
b5389086 ZY |
4408 | |
4409 | invalid: | |
4410 | pr_warn("HugeTLB: Invalid hugepages parameter %s\n", p); | |
f87442f4 | 4411 | hugepages_clear_pages_in_node(); |
f81f6e4b | 4412 | return 1; |
a3437870 | 4413 | } |
282f4214 | 4414 | __setup("hugepages=", hugepages_setup); |
e11bfbfc | 4415 | |
282f4214 MK |
4416 | /* |
4417 | * hugepagesz command line processing | |
4418 | * A specific huge page size can only be specified once with hugepagesz. | |
4419 | * hugepagesz is followed by hugepages on the command line. The global | |
4420 | * variable 'parsed_valid_hugepagesz' is used to determine if prior | |
4421 | * hugepagesz argument was valid. | |
4422 | */ | |
359f2544 | 4423 | static int __init hugepagesz_setup(char *s) |
e11bfbfc | 4424 | { |
359f2544 | 4425 | unsigned long size; |
282f4214 MK |
4426 | struct hstate *h; |
4427 | ||
4428 | parsed_valid_hugepagesz = false; | |
359f2544 MK |
4429 | size = (unsigned long)memparse(s, NULL); |
4430 | ||
4431 | if (!arch_hugetlb_valid_size(size)) { | |
282f4214 | 4432 | pr_err("HugeTLB: unsupported hugepagesz=%s\n", s); |
f81f6e4b | 4433 | return 1; |
359f2544 MK |
4434 | } |
4435 | ||
282f4214 MK |
4436 | h = size_to_hstate(size); |
4437 | if (h) { | |
4438 | /* | |
4439 | * hstate for this size already exists. This is normally | |
4440 | * an error, but is allowed if the existing hstate is the | |
4441 | * default hstate. More specifically, it is only allowed if | |
4442 | * the number of huge pages for the default hstate was not | |
4443 | * previously specified. | |
4444 | */ | |
4445 | if (!parsed_default_hugepagesz || h != &default_hstate || | |
4446 | default_hstate.max_huge_pages) { | |
4447 | pr_warn("HugeTLB: hugepagesz=%s specified twice, ignoring\n", s); | |
f81f6e4b | 4448 | return 1; |
282f4214 MK |
4449 | } |
4450 | ||
4451 | /* | |
4452 | * No need to call hugetlb_add_hstate() as hstate already | |
4453 | * exists. But, do set parsed_hstate so that a following | |
4454 | * hugepages= parameter will be applied to this hstate. | |
4455 | */ | |
4456 | parsed_hstate = h; | |
4457 | parsed_valid_hugepagesz = true; | |
4458 | return 1; | |
38237830 MK |
4459 | } |
4460 | ||
359f2544 | 4461 | hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT); |
282f4214 | 4462 | parsed_valid_hugepagesz = true; |
e11bfbfc NP |
4463 | return 1; |
4464 | } | |
359f2544 MK |
4465 | __setup("hugepagesz=", hugepagesz_setup); |
4466 | ||
282f4214 MK |
4467 | /* |
4468 | * default_hugepagesz command line input | |
4469 | * Only one instance of default_hugepagesz allowed on command line. | |
4470 | */ | |
ae94da89 | 4471 | static int __init default_hugepagesz_setup(char *s) |
e11bfbfc | 4472 | { |
ae94da89 | 4473 | unsigned long size; |
b5389086 | 4474 | int i; |
ae94da89 | 4475 | |
282f4214 | 4476 | parsed_valid_hugepagesz = false; |
282f4214 MK |
4477 | if (parsed_default_hugepagesz) { |
4478 | pr_err("HugeTLB: default_hugepagesz previously specified, ignoring %s\n", s); | |
f81f6e4b | 4479 | return 1; |
282f4214 MK |
4480 | } |
4481 | ||
ae94da89 MK |
4482 | size = (unsigned long)memparse(s, NULL); |
4483 | ||
4484 | if (!arch_hugetlb_valid_size(size)) { | |
282f4214 | 4485 | pr_err("HugeTLB: unsupported default_hugepagesz=%s\n", s); |
f81f6e4b | 4486 | return 1; |
ae94da89 MK |
4487 | } |
4488 | ||
282f4214 MK |
4489 | hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT); |
4490 | parsed_valid_hugepagesz = true; | |
4491 | parsed_default_hugepagesz = true; | |
4492 | default_hstate_idx = hstate_index(size_to_hstate(size)); | |
4493 | ||
4494 | /* | |
4495 | * The number of default huge pages (for this size) could have been | |
4496 | * specified as the first hugetlb parameter: hugepages=X. If so, | |
4497 | * then default_hstate_max_huge_pages is set. If the default huge | |
4498 | * page size is gigantic (>= MAX_ORDER), then the pages must be | |
4499 | * allocated here from bootmem allocator. | |
4500 | */ | |
4501 | if (default_hstate_max_huge_pages) { | |
4502 | default_hstate.max_huge_pages = default_hstate_max_huge_pages; | |
0a7a0f6f | 4503 | for_each_online_node(i) |
b5389086 ZY |
4504 | default_hstate.max_huge_pages_node[i] = |
4505 | default_hugepages_in_node[i]; | |
282f4214 MK |
4506 | if (hstate_is_gigantic(&default_hstate)) |
4507 | hugetlb_hstate_alloc_pages(&default_hstate); | |
4508 | default_hstate_max_huge_pages = 0; | |
4509 | } | |
4510 | ||
e11bfbfc NP |
4511 | return 1; |
4512 | } | |
ae94da89 | 4513 | __setup("default_hugepagesz=", default_hugepagesz_setup); |
a3437870 | 4514 | |
d2226ebd FT |
4515 | static nodemask_t *policy_mbind_nodemask(gfp_t gfp) |
4516 | { | |
4517 | #ifdef CONFIG_NUMA | |
4518 | struct mempolicy *mpol = get_task_policy(current); | |
4519 | ||
4520 | /* | |
4521 | * Only enforce MPOL_BIND policy which overlaps with cpuset policy | |
4522 | * (from policy_nodemask) specifically for hugetlb case | |
4523 | */ | |
4524 | if (mpol->mode == MPOL_BIND && | |
4525 | (apply_policy_zone(mpol, gfp_zone(gfp)) && | |
4526 | cpuset_nodemask_valid_mems_allowed(&mpol->nodes))) | |
4527 | return &mpol->nodes; | |
4528 | #endif | |
4529 | return NULL; | |
4530 | } | |
4531 | ||
8ca39e68 | 4532 | static unsigned int allowed_mems_nr(struct hstate *h) |
8a213460 NA |
4533 | { |
4534 | int node; | |
4535 | unsigned int nr = 0; | |
d2226ebd | 4536 | nodemask_t *mbind_nodemask; |
8ca39e68 MS |
4537 | unsigned int *array = h->free_huge_pages_node; |
4538 | gfp_t gfp_mask = htlb_alloc_mask(h); | |
4539 | ||
d2226ebd | 4540 | mbind_nodemask = policy_mbind_nodemask(gfp_mask); |
8ca39e68 | 4541 | for_each_node_mask(node, cpuset_current_mems_allowed) { |
d2226ebd | 4542 | if (!mbind_nodemask || node_isset(node, *mbind_nodemask)) |
8ca39e68 MS |
4543 | nr += array[node]; |
4544 | } | |
8a213460 NA |
4545 | |
4546 | return nr; | |
4547 | } | |
4548 | ||
4549 | #ifdef CONFIG_SYSCTL | |
17743798 MS |
4550 | static int proc_hugetlb_doulongvec_minmax(struct ctl_table *table, int write, |
4551 | void *buffer, size_t *length, | |
4552 | loff_t *ppos, unsigned long *out) | |
4553 | { | |
4554 | struct ctl_table dup_table; | |
4555 | ||
4556 | /* | |
4557 | * In order to avoid races with __do_proc_doulongvec_minmax(), we | |
4558 | * can duplicate the @table and alter the duplicate of it. | |
4559 | */ | |
4560 | dup_table = *table; | |
4561 | dup_table.data = out; | |
4562 | ||
4563 | return proc_doulongvec_minmax(&dup_table, write, buffer, length, ppos); | |
4564 | } | |
4565 | ||
06808b08 LS |
4566 | static int hugetlb_sysctl_handler_common(bool obey_mempolicy, |
4567 | struct ctl_table *table, int write, | |
32927393 | 4568 | void *buffer, size_t *length, loff_t *ppos) |
1da177e4 | 4569 | { |
e5ff2159 | 4570 | struct hstate *h = &default_hstate; |
238d3c13 | 4571 | unsigned long tmp = h->max_huge_pages; |
08d4a246 | 4572 | int ret; |
e5ff2159 | 4573 | |
457c1b27 | 4574 | if (!hugepages_supported()) |
86613628 | 4575 | return -EOPNOTSUPP; |
457c1b27 | 4576 | |
17743798 MS |
4577 | ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos, |
4578 | &tmp); | |
08d4a246 MH |
4579 | if (ret) |
4580 | goto out; | |
e5ff2159 | 4581 | |
238d3c13 DR |
4582 | if (write) |
4583 | ret = __nr_hugepages_store_common(obey_mempolicy, h, | |
4584 | NUMA_NO_NODE, tmp, *length); | |
08d4a246 MH |
4585 | out: |
4586 | return ret; | |
1da177e4 | 4587 | } |
396faf03 | 4588 | |
06808b08 | 4589 | int hugetlb_sysctl_handler(struct ctl_table *table, int write, |
32927393 | 4590 | void *buffer, size_t *length, loff_t *ppos) |
06808b08 LS |
4591 | { |
4592 | ||
4593 | return hugetlb_sysctl_handler_common(false, table, write, | |
4594 | buffer, length, ppos); | |
4595 | } | |
4596 | ||
4597 | #ifdef CONFIG_NUMA | |
4598 | int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write, | |
32927393 | 4599 | void *buffer, size_t *length, loff_t *ppos) |
06808b08 LS |
4600 | { |
4601 | return hugetlb_sysctl_handler_common(true, table, write, | |
4602 | buffer, length, ppos); | |
4603 | } | |
4604 | #endif /* CONFIG_NUMA */ | |
4605 | ||
a3d0c6aa | 4606 | int hugetlb_overcommit_handler(struct ctl_table *table, int write, |
32927393 | 4607 | void *buffer, size_t *length, loff_t *ppos) |
a3d0c6aa | 4608 | { |
a5516438 | 4609 | struct hstate *h = &default_hstate; |
e5ff2159 | 4610 | unsigned long tmp; |
08d4a246 | 4611 | int ret; |
e5ff2159 | 4612 | |
457c1b27 | 4613 | if (!hugepages_supported()) |
86613628 | 4614 | return -EOPNOTSUPP; |
457c1b27 | 4615 | |
c033a93c | 4616 | tmp = h->nr_overcommit_huge_pages; |
e5ff2159 | 4617 | |
bae7f4ae | 4618 | if (write && hstate_is_gigantic(h)) |
adbe8726 EM |
4619 | return -EINVAL; |
4620 | ||
17743798 MS |
4621 | ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos, |
4622 | &tmp); | |
08d4a246 MH |
4623 | if (ret) |
4624 | goto out; | |
e5ff2159 AK |
4625 | |
4626 | if (write) { | |
db71ef79 | 4627 | spin_lock_irq(&hugetlb_lock); |
e5ff2159 | 4628 | h->nr_overcommit_huge_pages = tmp; |
db71ef79 | 4629 | spin_unlock_irq(&hugetlb_lock); |
e5ff2159 | 4630 | } |
08d4a246 MH |
4631 | out: |
4632 | return ret; | |
a3d0c6aa NA |
4633 | } |
4634 | ||
1da177e4 LT |
4635 | #endif /* CONFIG_SYSCTL */ |
4636 | ||
e1759c21 | 4637 | void hugetlb_report_meminfo(struct seq_file *m) |
1da177e4 | 4638 | { |
fcb2b0c5 RG |
4639 | struct hstate *h; |
4640 | unsigned long total = 0; | |
4641 | ||
457c1b27 NA |
4642 | if (!hugepages_supported()) |
4643 | return; | |
fcb2b0c5 RG |
4644 | |
4645 | for_each_hstate(h) { | |
4646 | unsigned long count = h->nr_huge_pages; | |
4647 | ||
aca78307 | 4648 | total += huge_page_size(h) * count; |
fcb2b0c5 RG |
4649 | |
4650 | if (h == &default_hstate) | |
4651 | seq_printf(m, | |
4652 | "HugePages_Total: %5lu\n" | |
4653 | "HugePages_Free: %5lu\n" | |
4654 | "HugePages_Rsvd: %5lu\n" | |
4655 | "HugePages_Surp: %5lu\n" | |
4656 | "Hugepagesize: %8lu kB\n", | |
4657 | count, | |
4658 | h->free_huge_pages, | |
4659 | h->resv_huge_pages, | |
4660 | h->surplus_huge_pages, | |
aca78307 | 4661 | huge_page_size(h) / SZ_1K); |
fcb2b0c5 RG |
4662 | } |
4663 | ||
aca78307 | 4664 | seq_printf(m, "Hugetlb: %8lu kB\n", total / SZ_1K); |
1da177e4 LT |
4665 | } |
4666 | ||
7981593b | 4667 | int hugetlb_report_node_meminfo(char *buf, int len, int nid) |
1da177e4 | 4668 | { |
a5516438 | 4669 | struct hstate *h = &default_hstate; |
7981593b | 4670 | |
457c1b27 NA |
4671 | if (!hugepages_supported()) |
4672 | return 0; | |
7981593b JP |
4673 | |
4674 | return sysfs_emit_at(buf, len, | |
4675 | "Node %d HugePages_Total: %5u\n" | |
4676 | "Node %d HugePages_Free: %5u\n" | |
4677 | "Node %d HugePages_Surp: %5u\n", | |
4678 | nid, h->nr_huge_pages_node[nid], | |
4679 | nid, h->free_huge_pages_node[nid], | |
4680 | nid, h->surplus_huge_pages_node[nid]); | |
1da177e4 LT |
4681 | } |
4682 | ||
dcadcf1c | 4683 | void hugetlb_show_meminfo_node(int nid) |
949f7ec5 DR |
4684 | { |
4685 | struct hstate *h; | |
949f7ec5 | 4686 | |
457c1b27 NA |
4687 | if (!hugepages_supported()) |
4688 | return; | |
4689 | ||
dcadcf1c GL |
4690 | for_each_hstate(h) |
4691 | printk("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n", | |
4692 | nid, | |
4693 | h->nr_huge_pages_node[nid], | |
4694 | h->free_huge_pages_node[nid], | |
4695 | h->surplus_huge_pages_node[nid], | |
4696 | huge_page_size(h) / SZ_1K); | |
949f7ec5 DR |
4697 | } |
4698 | ||
5d317b2b NH |
4699 | void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm) |
4700 | { | |
4701 | seq_printf(m, "HugetlbPages:\t%8lu kB\n", | |
4702 | atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10)); | |
4703 | } | |
4704 | ||
1da177e4 LT |
4705 | /* Return the number pages of memory we physically have, in PAGE_SIZE units. */ |
4706 | unsigned long hugetlb_total_pages(void) | |
4707 | { | |
d0028588 WL |
4708 | struct hstate *h; |
4709 | unsigned long nr_total_pages = 0; | |
4710 | ||
4711 | for_each_hstate(h) | |
4712 | nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h); | |
4713 | return nr_total_pages; | |
1da177e4 | 4714 | } |
1da177e4 | 4715 | |
a5516438 | 4716 | static int hugetlb_acct_memory(struct hstate *h, long delta) |
fc1b8a73 MG |
4717 | { |
4718 | int ret = -ENOMEM; | |
4719 | ||
0aa7f354 ML |
4720 | if (!delta) |
4721 | return 0; | |
4722 | ||
db71ef79 | 4723 | spin_lock_irq(&hugetlb_lock); |
fc1b8a73 MG |
4724 | /* |
4725 | * When cpuset is configured, it breaks the strict hugetlb page | |
4726 | * reservation as the accounting is done on a global variable. Such | |
4727 | * reservation is completely rubbish in the presence of cpuset because | |
4728 | * the reservation is not checked against page availability for the | |
4729 | * current cpuset. Application can still potentially OOM'ed by kernel | |
4730 | * with lack of free htlb page in cpuset that the task is in. | |
4731 | * Attempt to enforce strict accounting with cpuset is almost | |
4732 | * impossible (or too ugly) because cpuset is too fluid that | |
4733 | * task or memory node can be dynamically moved between cpusets. | |
4734 | * | |
4735 | * The change of semantics for shared hugetlb mapping with cpuset is | |
4736 | * undesirable. However, in order to preserve some of the semantics, | |
4737 | * we fall back to check against current free page availability as | |
4738 | * a best attempt and hopefully to minimize the impact of changing | |
4739 | * semantics that cpuset has. | |
8ca39e68 MS |
4740 | * |
4741 | * Apart from cpuset, we also have memory policy mechanism that | |
4742 | * also determines from which node the kernel will allocate memory | |
4743 | * in a NUMA system. So similar to cpuset, we also should consider | |
4744 | * the memory policy of the current task. Similar to the description | |
4745 | * above. | |
fc1b8a73 MG |
4746 | */ |
4747 | if (delta > 0) { | |
a5516438 | 4748 | if (gather_surplus_pages(h, delta) < 0) |
fc1b8a73 MG |
4749 | goto out; |
4750 | ||
8ca39e68 | 4751 | if (delta > allowed_mems_nr(h)) { |
a5516438 | 4752 | return_unused_surplus_pages(h, delta); |
fc1b8a73 MG |
4753 | goto out; |
4754 | } | |
4755 | } | |
4756 | ||
4757 | ret = 0; | |
4758 | if (delta < 0) | |
a5516438 | 4759 | return_unused_surplus_pages(h, (unsigned long) -delta); |
fc1b8a73 MG |
4760 | |
4761 | out: | |
db71ef79 | 4762 | spin_unlock_irq(&hugetlb_lock); |
fc1b8a73 MG |
4763 | return ret; |
4764 | } | |
4765 | ||
84afd99b AW |
4766 | static void hugetlb_vm_op_open(struct vm_area_struct *vma) |
4767 | { | |
f522c3ac | 4768 | struct resv_map *resv = vma_resv_map(vma); |
84afd99b AW |
4769 | |
4770 | /* | |
612b8a31 | 4771 | * HPAGE_RESV_OWNER indicates a private mapping. |
84afd99b AW |
4772 | * This new VMA should share its siblings reservation map if present. |
4773 | * The VMA will only ever have a valid reservation map pointer where | |
4774 | * it is being copied for another still existing VMA. As that VMA | |
25985edc | 4775 | * has a reference to the reservation map it cannot disappear until |
84afd99b AW |
4776 | * after this open call completes. It is therefore safe to take a |
4777 | * new reference here without additional locking. | |
4778 | */ | |
09a26e83 MK |
4779 | if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { |
4780 | resv_map_dup_hugetlb_cgroup_uncharge_info(resv); | |
f522c3ac | 4781 | kref_get(&resv->refs); |
09a26e83 | 4782 | } |
8d9bfb26 | 4783 | |
131a79b4 MK |
4784 | /* |
4785 | * vma_lock structure for sharable mappings is vma specific. | |
612b8a31 MK |
4786 | * Clear old pointer (if copied via vm_area_dup) and allocate |
4787 | * new structure. Before clearing, make sure vma_lock is not | |
4788 | * for this vma. | |
131a79b4 MK |
4789 | */ |
4790 | if (vma->vm_flags & VM_MAYSHARE) { | |
612b8a31 MK |
4791 | struct hugetlb_vma_lock *vma_lock = vma->vm_private_data; |
4792 | ||
4793 | if (vma_lock) { | |
4794 | if (vma_lock->vma != vma) { | |
4795 | vma->vm_private_data = NULL; | |
4796 | hugetlb_vma_lock_alloc(vma); | |
4797 | } else | |
4798 | pr_warn("HugeTLB: vma_lock already exists in %s.\n", __func__); | |
4799 | } else | |
4800 | hugetlb_vma_lock_alloc(vma); | |
131a79b4 | 4801 | } |
84afd99b AW |
4802 | } |
4803 | ||
a1e78772 MG |
4804 | static void hugetlb_vm_op_close(struct vm_area_struct *vma) |
4805 | { | |
a5516438 | 4806 | struct hstate *h = hstate_vma(vma); |
8d9bfb26 | 4807 | struct resv_map *resv; |
90481622 | 4808 | struct hugepage_subpool *spool = subpool_vma(vma); |
4e35f483 | 4809 | unsigned long reserve, start, end; |
1c5ecae3 | 4810 | long gbl_reserve; |
84afd99b | 4811 | |
8d9bfb26 MK |
4812 | hugetlb_vma_lock_free(vma); |
4813 | ||
4814 | resv = vma_resv_map(vma); | |
4e35f483 JK |
4815 | if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER)) |
4816 | return; | |
84afd99b | 4817 | |
4e35f483 JK |
4818 | start = vma_hugecache_offset(h, vma, vma->vm_start); |
4819 | end = vma_hugecache_offset(h, vma, vma->vm_end); | |
84afd99b | 4820 | |
4e35f483 | 4821 | reserve = (end - start) - region_count(resv, start, end); |
e9fe92ae | 4822 | hugetlb_cgroup_uncharge_counter(resv, start, end); |
4e35f483 | 4823 | if (reserve) { |
1c5ecae3 MK |
4824 | /* |
4825 | * Decrement reserve counts. The global reserve count may be | |
4826 | * adjusted if the subpool has a minimum size. | |
4827 | */ | |
4828 | gbl_reserve = hugepage_subpool_put_pages(spool, reserve); | |
4829 | hugetlb_acct_memory(h, -gbl_reserve); | |
84afd99b | 4830 | } |
e9fe92ae MA |
4831 | |
4832 | kref_put(&resv->refs, resv_map_release); | |
a1e78772 MG |
4833 | } |
4834 | ||
31383c68 DW |
4835 | static int hugetlb_vm_op_split(struct vm_area_struct *vma, unsigned long addr) |
4836 | { | |
4837 | if (addr & ~(huge_page_mask(hstate_vma(vma)))) | |
4838 | return -EINVAL; | |
b30c14cd JH |
4839 | |
4840 | /* | |
4841 | * PMD sharing is only possible for PUD_SIZE-aligned address ranges | |
4842 | * in HugeTLB VMAs. If we will lose PUD_SIZE alignment due to this | |
4843 | * split, unshare PMDs in the PUD_SIZE interval surrounding addr now. | |
4844 | */ | |
4845 | if (addr & ~PUD_MASK) { | |
4846 | /* | |
4847 | * hugetlb_vm_op_split is called right before we attempt to | |
4848 | * split the VMA. We will need to unshare PMDs in the old and | |
4849 | * new VMAs, so let's unshare before we split. | |
4850 | */ | |
4851 | unsigned long floor = addr & PUD_MASK; | |
4852 | unsigned long ceil = floor + PUD_SIZE; | |
4853 | ||
4854 | if (floor >= vma->vm_start && ceil <= vma->vm_end) | |
4855 | hugetlb_unshare_pmds(vma, floor, ceil); | |
4856 | } | |
4857 | ||
31383c68 DW |
4858 | return 0; |
4859 | } | |
4860 | ||
05ea8860 DW |
4861 | static unsigned long hugetlb_vm_op_pagesize(struct vm_area_struct *vma) |
4862 | { | |
aca78307 | 4863 | return huge_page_size(hstate_vma(vma)); |
05ea8860 DW |
4864 | } |
4865 | ||
1da177e4 LT |
4866 | /* |
4867 | * We cannot handle pagefaults against hugetlb pages at all. They cause | |
4868 | * handle_mm_fault() to try to instantiate regular-sized pages in the | |
6c26d310 | 4869 | * hugepage VMA. do_page_fault() is supposed to trap this, so BUG is we get |
1da177e4 LT |
4870 | * this far. |
4871 | */ | |
b3ec9f33 | 4872 | static vm_fault_t hugetlb_vm_op_fault(struct vm_fault *vmf) |
1da177e4 LT |
4873 | { |
4874 | BUG(); | |
d0217ac0 | 4875 | return 0; |
1da177e4 LT |
4876 | } |
4877 | ||
eec3636a JC |
4878 | /* |
4879 | * When a new function is introduced to vm_operations_struct and added | |
4880 | * to hugetlb_vm_ops, please consider adding the function to shm_vm_ops. | |
4881 | * This is because under System V memory model, mappings created via | |
4882 | * shmget/shmat with "huge page" specified are backed by hugetlbfs files, | |
4883 | * their original vm_ops are overwritten with shm_vm_ops. | |
4884 | */ | |
f0f37e2f | 4885 | const struct vm_operations_struct hugetlb_vm_ops = { |
d0217ac0 | 4886 | .fault = hugetlb_vm_op_fault, |
84afd99b | 4887 | .open = hugetlb_vm_op_open, |
a1e78772 | 4888 | .close = hugetlb_vm_op_close, |
dd3b614f | 4889 | .may_split = hugetlb_vm_op_split, |
05ea8860 | 4890 | .pagesize = hugetlb_vm_op_pagesize, |
1da177e4 LT |
4891 | }; |
4892 | ||
1e8f889b DG |
4893 | static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page, |
4894 | int writable) | |
63551ae0 DG |
4895 | { |
4896 | pte_t entry; | |
79c1c594 | 4897 | unsigned int shift = huge_page_shift(hstate_vma(vma)); |
63551ae0 | 4898 | |
1e8f889b | 4899 | if (writable) { |
106c992a GS |
4900 | entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page, |
4901 | vma->vm_page_prot))); | |
63551ae0 | 4902 | } else { |
106c992a GS |
4903 | entry = huge_pte_wrprotect(mk_huge_pte(page, |
4904 | vma->vm_page_prot)); | |
63551ae0 DG |
4905 | } |
4906 | entry = pte_mkyoung(entry); | |
79c1c594 | 4907 | entry = arch_make_huge_pte(entry, shift, vma->vm_flags); |
63551ae0 DG |
4908 | |
4909 | return entry; | |
4910 | } | |
4911 | ||
1e8f889b DG |
4912 | static void set_huge_ptep_writable(struct vm_area_struct *vma, |
4913 | unsigned long address, pte_t *ptep) | |
4914 | { | |
4915 | pte_t entry; | |
4916 | ||
106c992a | 4917 | entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep))); |
32f84528 | 4918 | if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) |
4b3073e1 | 4919 | update_mmu_cache(vma, address, ptep); |
1e8f889b DG |
4920 | } |
4921 | ||
d5ed7444 | 4922 | bool is_hugetlb_entry_migration(pte_t pte) |
4a705fef NH |
4923 | { |
4924 | swp_entry_t swp; | |
4925 | ||
4926 | if (huge_pte_none(pte) || pte_present(pte)) | |
d5ed7444 | 4927 | return false; |
4a705fef | 4928 | swp = pte_to_swp_entry(pte); |
d79d176a | 4929 | if (is_migration_entry(swp)) |
d5ed7444 | 4930 | return true; |
4a705fef | 4931 | else |
d5ed7444 | 4932 | return false; |
4a705fef NH |
4933 | } |
4934 | ||
3e5c3600 | 4935 | static bool is_hugetlb_entry_hwpoisoned(pte_t pte) |
4a705fef NH |
4936 | { |
4937 | swp_entry_t swp; | |
4938 | ||
4939 | if (huge_pte_none(pte) || pte_present(pte)) | |
3e5c3600 | 4940 | return false; |
4a705fef | 4941 | swp = pte_to_swp_entry(pte); |
d79d176a | 4942 | if (is_hwpoison_entry(swp)) |
3e5c3600 | 4943 | return true; |
4a705fef | 4944 | else |
3e5c3600 | 4945 | return false; |
4a705fef | 4946 | } |
1e8f889b | 4947 | |
4eae4efa PX |
4948 | static void |
4949 | hugetlb_install_page(struct vm_area_struct *vma, pte_t *ptep, unsigned long addr, | |
4950 | struct page *new_page) | |
4951 | { | |
4952 | __SetPageUptodate(new_page); | |
4eae4efa | 4953 | hugepage_add_new_anon_rmap(new_page, vma, addr); |
1eba86c0 | 4954 | set_huge_pte_at(vma->vm_mm, addr, ptep, make_huge_pte(vma, new_page, 1)); |
4eae4efa | 4955 | hugetlb_count_add(pages_per_huge_page(hstate_vma(vma)), vma->vm_mm); |
4eae4efa PX |
4956 | SetHPageMigratable(new_page); |
4957 | } | |
4958 | ||
63551ae0 | 4959 | int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src, |
bc70fbf2 PX |
4960 | struct vm_area_struct *dst_vma, |
4961 | struct vm_area_struct *src_vma) | |
63551ae0 | 4962 | { |
3aa4ed80 | 4963 | pte_t *src_pte, *dst_pte, entry; |
63551ae0 | 4964 | struct page *ptepage; |
1c59827d | 4965 | unsigned long addr; |
bc70fbf2 PX |
4966 | bool cow = is_cow_mapping(src_vma->vm_flags); |
4967 | struct hstate *h = hstate_vma(src_vma); | |
a5516438 | 4968 | unsigned long sz = huge_page_size(h); |
4eae4efa | 4969 | unsigned long npages = pages_per_huge_page(h); |
ac46d4f3 | 4970 | struct mmu_notifier_range range; |
e95a9851 | 4971 | unsigned long last_addr_mask; |
e8569dd2 | 4972 | int ret = 0; |
1e8f889b | 4973 | |
ac46d4f3 | 4974 | if (cow) { |
bc70fbf2 PX |
4975 | mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, src_vma, src, |
4976 | src_vma->vm_start, | |
4977 | src_vma->vm_end); | |
ac46d4f3 | 4978 | mmu_notifier_invalidate_range_start(&range); |
623a1ddf DH |
4979 | mmap_assert_write_locked(src); |
4980 | raw_write_seqcount_begin(&src->write_protect_seq); | |
40549ba8 MK |
4981 | } else { |
4982 | /* | |
4983 | * For shared mappings the vma lock must be held before | |
4984 | * calling huge_pte_offset in the src vma. Otherwise, the | |
4985 | * returned ptep could go away if part of a shared pmd and | |
4986 | * another thread calls huge_pmd_unshare. | |
4987 | */ | |
4988 | hugetlb_vma_lock_read(src_vma); | |
ac46d4f3 | 4989 | } |
e8569dd2 | 4990 | |
e95a9851 | 4991 | last_addr_mask = hugetlb_mask_last_page(h); |
bc70fbf2 | 4992 | for (addr = src_vma->vm_start; addr < src_vma->vm_end; addr += sz) { |
cb900f41 | 4993 | spinlock_t *src_ptl, *dst_ptl; |
7868a208 | 4994 | src_pte = huge_pte_offset(src, addr, sz); |
e95a9851 MK |
4995 | if (!src_pte) { |
4996 | addr |= last_addr_mask; | |
c74df32c | 4997 | continue; |
e95a9851 | 4998 | } |
bc70fbf2 | 4999 | dst_pte = huge_pte_alloc(dst, dst_vma, addr, sz); |
e8569dd2 AS |
5000 | if (!dst_pte) { |
5001 | ret = -ENOMEM; | |
5002 | break; | |
5003 | } | |
c5c99429 | 5004 | |
5e41540c MK |
5005 | /* |
5006 | * If the pagetables are shared don't copy or take references. | |
5e41540c | 5007 | * |
3aa4ed80 | 5008 | * dst_pte == src_pte is the common case of src/dest sharing. |
5e41540c | 5009 | * However, src could have 'unshared' and dst shares with |
3aa4ed80 ML |
5010 | * another vma. So page_count of ptep page is checked instead |
5011 | * to reliably determine whether pte is shared. | |
5e41540c | 5012 | */ |
3aa4ed80 | 5013 | if (page_count(virt_to_page(dst_pte)) > 1) { |
e95a9851 | 5014 | addr |= last_addr_mask; |
c5c99429 | 5015 | continue; |
e95a9851 | 5016 | } |
c5c99429 | 5017 | |
cb900f41 KS |
5018 | dst_ptl = huge_pte_lock(h, dst, dst_pte); |
5019 | src_ptl = huge_pte_lockptr(h, src, src_pte); | |
5020 | spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); | |
4a705fef | 5021 | entry = huge_ptep_get(src_pte); |
4eae4efa | 5022 | again: |
3aa4ed80 | 5023 | if (huge_pte_none(entry)) { |
5e41540c | 5024 | /* |
3aa4ed80 | 5025 | * Skip if src entry none. |
5e41540c | 5026 | */ |
4a705fef | 5027 | ; |
c2cb0dcc NH |
5028 | } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry))) { |
5029 | bool uffd_wp = huge_pte_uffd_wp(entry); | |
5030 | ||
5031 | if (!userfaultfd_wp(dst_vma) && uffd_wp) | |
5032 | entry = huge_pte_clear_uffd_wp(entry); | |
5033 | set_huge_pte_at(dst, addr, dst_pte, entry); | |
5034 | } else if (unlikely(is_hugetlb_entry_migration(entry))) { | |
4a705fef | 5035 | swp_entry_t swp_entry = pte_to_swp_entry(entry); |
bc70fbf2 | 5036 | bool uffd_wp = huge_pte_uffd_wp(entry); |
4a705fef | 5037 | |
6c287605 | 5038 | if (!is_readable_migration_entry(swp_entry) && cow) { |
4a705fef NH |
5039 | /* |
5040 | * COW mappings require pages in both | |
5041 | * parent and child to be set to read. | |
5042 | */ | |
4dd845b5 AP |
5043 | swp_entry = make_readable_migration_entry( |
5044 | swp_offset(swp_entry)); | |
4a705fef | 5045 | entry = swp_entry_to_pte(swp_entry); |
bc70fbf2 PX |
5046 | if (userfaultfd_wp(src_vma) && uffd_wp) |
5047 | entry = huge_pte_mkuffd_wp(entry); | |
18f39629 | 5048 | set_huge_pte_at(src, addr, src_pte, entry); |
4a705fef | 5049 | } |
bc70fbf2 PX |
5050 | if (!userfaultfd_wp(dst_vma) && uffd_wp) |
5051 | entry = huge_pte_clear_uffd_wp(entry); | |
18f39629 | 5052 | set_huge_pte_at(dst, addr, dst_pte, entry); |
bc70fbf2 PX |
5053 | } else if (unlikely(is_pte_marker(entry))) { |
5054 | /* | |
5055 | * We copy the pte marker only if the dst vma has | |
5056 | * uffd-wp enabled. | |
5057 | */ | |
5058 | if (userfaultfd_wp(dst_vma)) | |
5059 | set_huge_pte_at(dst, addr, dst_pte, entry); | |
4a705fef | 5060 | } else { |
4eae4efa PX |
5061 | entry = huge_ptep_get(src_pte); |
5062 | ptepage = pte_page(entry); | |
5063 | get_page(ptepage); | |
5064 | ||
5065 | /* | |
fb3d824d DH |
5066 | * Failing to duplicate the anon rmap is a rare case |
5067 | * where we see pinned hugetlb pages while they're | |
5068 | * prone to COW. We need to do the COW earlier during | |
5069 | * fork. | |
4eae4efa PX |
5070 | * |
5071 | * When pre-allocating the page or copying data, we | |
5072 | * need to be without the pgtable locks since we could | |
5073 | * sleep during the process. | |
5074 | */ | |
fb3d824d DH |
5075 | if (!PageAnon(ptepage)) { |
5076 | page_dup_file_rmap(ptepage, true); | |
bc70fbf2 PX |
5077 | } else if (page_try_dup_anon_rmap(ptepage, true, |
5078 | src_vma)) { | |
4eae4efa PX |
5079 | pte_t src_pte_old = entry; |
5080 | struct page *new; | |
5081 | ||
5082 | spin_unlock(src_ptl); | |
5083 | spin_unlock(dst_ptl); | |
5084 | /* Do not use reserve as it's private owned */ | |
bc70fbf2 | 5085 | new = alloc_huge_page(dst_vma, addr, 1); |
4eae4efa PX |
5086 | if (IS_ERR(new)) { |
5087 | put_page(ptepage); | |
5088 | ret = PTR_ERR(new); | |
5089 | break; | |
5090 | } | |
bc70fbf2 | 5091 | copy_user_huge_page(new, ptepage, addr, dst_vma, |
4eae4efa PX |
5092 | npages); |
5093 | put_page(ptepage); | |
5094 | ||
5095 | /* Install the new huge page if src pte stable */ | |
5096 | dst_ptl = huge_pte_lock(h, dst, dst_pte); | |
5097 | src_ptl = huge_pte_lockptr(h, src, src_pte); | |
5098 | spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); | |
5099 | entry = huge_ptep_get(src_pte); | |
5100 | if (!pte_same(src_pte_old, entry)) { | |
bc70fbf2 | 5101 | restore_reserve_on_error(h, dst_vma, addr, |
846be085 | 5102 | new); |
4eae4efa | 5103 | put_page(new); |
3aa4ed80 | 5104 | /* huge_ptep of dst_pte won't change as in child */ |
4eae4efa PX |
5105 | goto again; |
5106 | } | |
bc70fbf2 | 5107 | hugetlb_install_page(dst_vma, dst_pte, addr, new); |
4eae4efa PX |
5108 | spin_unlock(src_ptl); |
5109 | spin_unlock(dst_ptl); | |
5110 | continue; | |
5111 | } | |
5112 | ||
34ee645e | 5113 | if (cow) { |
0f10851e JG |
5114 | /* |
5115 | * No need to notify as we are downgrading page | |
5116 | * table protection not changing it to point | |
5117 | * to a new page. | |
5118 | * | |
ee65728e | 5119 | * See Documentation/mm/mmu_notifier.rst |
0f10851e | 5120 | */ |
7f2e9525 | 5121 | huge_ptep_set_wrprotect(src, addr, src_pte); |
84894e1c | 5122 | entry = huge_pte_wrprotect(entry); |
34ee645e | 5123 | } |
4eae4efa | 5124 | |
1c59827d | 5125 | set_huge_pte_at(dst, addr, dst_pte, entry); |
4eae4efa | 5126 | hugetlb_count_add(npages, dst); |
1c59827d | 5127 | } |
cb900f41 KS |
5128 | spin_unlock(src_ptl); |
5129 | spin_unlock(dst_ptl); | |
63551ae0 | 5130 | } |
63551ae0 | 5131 | |
623a1ddf DH |
5132 | if (cow) { |
5133 | raw_write_seqcount_end(&src->write_protect_seq); | |
ac46d4f3 | 5134 | mmu_notifier_invalidate_range_end(&range); |
40549ba8 MK |
5135 | } else { |
5136 | hugetlb_vma_unlock_read(src_vma); | |
623a1ddf | 5137 | } |
e8569dd2 AS |
5138 | |
5139 | return ret; | |
63551ae0 DG |
5140 | } |
5141 | ||
550a7d60 | 5142 | static void move_huge_pte(struct vm_area_struct *vma, unsigned long old_addr, |
db110a99 | 5143 | unsigned long new_addr, pte_t *src_pte, pte_t *dst_pte) |
550a7d60 MA |
5144 | { |
5145 | struct hstate *h = hstate_vma(vma); | |
5146 | struct mm_struct *mm = vma->vm_mm; | |
550a7d60 | 5147 | spinlock_t *src_ptl, *dst_ptl; |
db110a99 | 5148 | pte_t pte; |
550a7d60 | 5149 | |
550a7d60 MA |
5150 | dst_ptl = huge_pte_lock(h, mm, dst_pte); |
5151 | src_ptl = huge_pte_lockptr(h, mm, src_pte); | |
5152 | ||
5153 | /* | |
5154 | * We don't have to worry about the ordering of src and dst ptlocks | |
8651a137 | 5155 | * because exclusive mmap_lock (or the i_mmap_lock) prevents deadlock. |
550a7d60 MA |
5156 | */ |
5157 | if (src_ptl != dst_ptl) | |
5158 | spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); | |
5159 | ||
5160 | pte = huge_ptep_get_and_clear(mm, old_addr, src_pte); | |
5161 | set_huge_pte_at(mm, new_addr, dst_pte, pte); | |
5162 | ||
5163 | if (src_ptl != dst_ptl) | |
5164 | spin_unlock(src_ptl); | |
5165 | spin_unlock(dst_ptl); | |
5166 | } | |
5167 | ||
5168 | int move_hugetlb_page_tables(struct vm_area_struct *vma, | |
5169 | struct vm_area_struct *new_vma, | |
5170 | unsigned long old_addr, unsigned long new_addr, | |
5171 | unsigned long len) | |
5172 | { | |
5173 | struct hstate *h = hstate_vma(vma); | |
5174 | struct address_space *mapping = vma->vm_file->f_mapping; | |
5175 | unsigned long sz = huge_page_size(h); | |
5176 | struct mm_struct *mm = vma->vm_mm; | |
5177 | unsigned long old_end = old_addr + len; | |
e95a9851 | 5178 | unsigned long last_addr_mask; |
550a7d60 MA |
5179 | pte_t *src_pte, *dst_pte; |
5180 | struct mmu_notifier_range range; | |
3d0b95cd | 5181 | bool shared_pmd = false; |
550a7d60 MA |
5182 | |
5183 | mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, old_addr, | |
5184 | old_end); | |
5185 | adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end); | |
3d0b95cd BW |
5186 | /* |
5187 | * In case of shared PMDs, we should cover the maximum possible | |
5188 | * range. | |
5189 | */ | |
5190 | flush_cache_range(vma, range.start, range.end); | |
5191 | ||
550a7d60 | 5192 | mmu_notifier_invalidate_range_start(&range); |
e95a9851 | 5193 | last_addr_mask = hugetlb_mask_last_page(h); |
550a7d60 | 5194 | /* Prevent race with file truncation */ |
40549ba8 | 5195 | hugetlb_vma_lock_write(vma); |
550a7d60 MA |
5196 | i_mmap_lock_write(mapping); |
5197 | for (; old_addr < old_end; old_addr += sz, new_addr += sz) { | |
5198 | src_pte = huge_pte_offset(mm, old_addr, sz); | |
e95a9851 MK |
5199 | if (!src_pte) { |
5200 | old_addr |= last_addr_mask; | |
5201 | new_addr |= last_addr_mask; | |
550a7d60 | 5202 | continue; |
e95a9851 | 5203 | } |
550a7d60 MA |
5204 | if (huge_pte_none(huge_ptep_get(src_pte))) |
5205 | continue; | |
5206 | ||
4ddb4d91 | 5207 | if (huge_pmd_unshare(mm, vma, old_addr, src_pte)) { |
3d0b95cd | 5208 | shared_pmd = true; |
4ddb4d91 MK |
5209 | old_addr |= last_addr_mask; |
5210 | new_addr |= last_addr_mask; | |
550a7d60 | 5211 | continue; |
3d0b95cd | 5212 | } |
550a7d60 MA |
5213 | |
5214 | dst_pte = huge_pte_alloc(mm, new_vma, new_addr, sz); | |
5215 | if (!dst_pte) | |
5216 | break; | |
5217 | ||
db110a99 | 5218 | move_huge_pte(vma, old_addr, new_addr, src_pte, dst_pte); |
550a7d60 | 5219 | } |
3d0b95cd BW |
5220 | |
5221 | if (shared_pmd) | |
5222 | flush_tlb_range(vma, range.start, range.end); | |
5223 | else | |
5224 | flush_tlb_range(vma, old_end - len, old_end); | |
550a7d60 | 5225 | mmu_notifier_invalidate_range_end(&range); |
13e4ad2c | 5226 | i_mmap_unlock_write(mapping); |
40549ba8 | 5227 | hugetlb_vma_unlock_write(vma); |
550a7d60 MA |
5228 | |
5229 | return len + old_addr - old_end; | |
5230 | } | |
5231 | ||
73c54763 PX |
5232 | static void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma, |
5233 | unsigned long start, unsigned long end, | |
05e90bd0 | 5234 | struct page *ref_page, zap_flags_t zap_flags) |
63551ae0 DG |
5235 | { |
5236 | struct mm_struct *mm = vma->vm_mm; | |
5237 | unsigned long address; | |
c7546f8f | 5238 | pte_t *ptep; |
63551ae0 | 5239 | pte_t pte; |
cb900f41 | 5240 | spinlock_t *ptl; |
63551ae0 | 5241 | struct page *page; |
a5516438 AK |
5242 | struct hstate *h = hstate_vma(vma); |
5243 | unsigned long sz = huge_page_size(h); | |
e95a9851 | 5244 | unsigned long last_addr_mask; |
a4a118f2 | 5245 | bool force_flush = false; |
a5516438 | 5246 | |
63551ae0 | 5247 | WARN_ON(!is_vm_hugetlb_page(vma)); |
a5516438 AK |
5248 | BUG_ON(start & ~huge_page_mask(h)); |
5249 | BUG_ON(end & ~huge_page_mask(h)); | |
63551ae0 | 5250 | |
07e32661 AK |
5251 | /* |
5252 | * This is a hugetlb vma, all the pte entries should point | |
5253 | * to huge page. | |
5254 | */ | |
ed6a7935 | 5255 | tlb_change_page_size(tlb, sz); |
24669e58 | 5256 | tlb_start_vma(tlb, vma); |
dff11abe | 5257 | |
e95a9851 | 5258 | last_addr_mask = hugetlb_mask_last_page(h); |
569f48b8 | 5259 | address = start; |
569f48b8 | 5260 | for (; address < end; address += sz) { |
7868a208 | 5261 | ptep = huge_pte_offset(mm, address, sz); |
e95a9851 MK |
5262 | if (!ptep) { |
5263 | address |= last_addr_mask; | |
c7546f8f | 5264 | continue; |
e95a9851 | 5265 | } |
c7546f8f | 5266 | |
cb900f41 | 5267 | ptl = huge_pte_lock(h, mm, ptep); |
4ddb4d91 | 5268 | if (huge_pmd_unshare(mm, vma, address, ptep)) { |
31d49da5 | 5269 | spin_unlock(ptl); |
a4a118f2 NA |
5270 | tlb_flush_pmd_range(tlb, address & PUD_MASK, PUD_SIZE); |
5271 | force_flush = true; | |
4ddb4d91 | 5272 | address |= last_addr_mask; |
31d49da5 AK |
5273 | continue; |
5274 | } | |
39dde65c | 5275 | |
6629326b | 5276 | pte = huge_ptep_get(ptep); |
31d49da5 AK |
5277 | if (huge_pte_none(pte)) { |
5278 | spin_unlock(ptl); | |
5279 | continue; | |
5280 | } | |
6629326b HD |
5281 | |
5282 | /* | |
9fbc1f63 NH |
5283 | * Migrating hugepage or HWPoisoned hugepage is already |
5284 | * unmapped and its refcount is dropped, so just clear pte here. | |
6629326b | 5285 | */ |
9fbc1f63 | 5286 | if (unlikely(!pte_present(pte))) { |
05e90bd0 PX |
5287 | /* |
5288 | * If the pte was wr-protected by uffd-wp in any of the | |
5289 | * swap forms, meanwhile the caller does not want to | |
5290 | * drop the uffd-wp bit in this zap, then replace the | |
5291 | * pte with a marker. | |
5292 | */ | |
5293 | if (pte_swp_uffd_wp_any(pte) && | |
5294 | !(zap_flags & ZAP_FLAG_DROP_MARKER)) | |
5295 | set_huge_pte_at(mm, address, ptep, | |
5296 | make_pte_marker(PTE_MARKER_UFFD_WP)); | |
5297 | else | |
5298 | huge_pte_clear(mm, address, ptep, sz); | |
31d49da5 AK |
5299 | spin_unlock(ptl); |
5300 | continue; | |
8c4894c6 | 5301 | } |
6629326b HD |
5302 | |
5303 | page = pte_page(pte); | |
04f2cbe3 MG |
5304 | /* |
5305 | * If a reference page is supplied, it is because a specific | |
5306 | * page is being unmapped, not a range. Ensure the page we | |
5307 | * are about to unmap is the actual page of interest. | |
5308 | */ | |
5309 | if (ref_page) { | |
31d49da5 AK |
5310 | if (page != ref_page) { |
5311 | spin_unlock(ptl); | |
5312 | continue; | |
5313 | } | |
04f2cbe3 MG |
5314 | /* |
5315 | * Mark the VMA as having unmapped its page so that | |
5316 | * future faults in this VMA will fail rather than | |
5317 | * looking like data was lost | |
5318 | */ | |
5319 | set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED); | |
5320 | } | |
5321 | ||
c7546f8f | 5322 | pte = huge_ptep_get_and_clear(mm, address, ptep); |
b528e4b6 | 5323 | tlb_remove_huge_tlb_entry(h, tlb, ptep, address); |
106c992a | 5324 | if (huge_pte_dirty(pte)) |
6649a386 | 5325 | set_page_dirty(page); |
05e90bd0 PX |
5326 | /* Leave a uffd-wp pte marker if needed */ |
5327 | if (huge_pte_uffd_wp(pte) && | |
5328 | !(zap_flags & ZAP_FLAG_DROP_MARKER)) | |
5329 | set_huge_pte_at(mm, address, ptep, | |
5330 | make_pte_marker(PTE_MARKER_UFFD_WP)); | |
5d317b2b | 5331 | hugetlb_count_sub(pages_per_huge_page(h), mm); |
cea86fe2 | 5332 | page_remove_rmap(page, vma, true); |
31d49da5 | 5333 | |
cb900f41 | 5334 | spin_unlock(ptl); |
e77b0852 | 5335 | tlb_remove_page_size(tlb, page, huge_page_size(h)); |
31d49da5 AK |
5336 | /* |
5337 | * Bail out after unmapping reference page if supplied | |
5338 | */ | |
5339 | if (ref_page) | |
5340 | break; | |
fe1668ae | 5341 | } |
24669e58 | 5342 | tlb_end_vma(tlb, vma); |
a4a118f2 NA |
5343 | |
5344 | /* | |
5345 | * If we unshared PMDs, the TLB flush was not recorded in mmu_gather. We | |
5346 | * could defer the flush until now, since by holding i_mmap_rwsem we | |
5347 | * guaranteed that the last refernece would not be dropped. But we must | |
5348 | * do the flushing before we return, as otherwise i_mmap_rwsem will be | |
5349 | * dropped and the last reference to the shared PMDs page might be | |
5350 | * dropped as well. | |
5351 | * | |
5352 | * In theory we could defer the freeing of the PMD pages as well, but | |
5353 | * huge_pmd_unshare() relies on the exact page_count for the PMD page to | |
5354 | * detect sharing, so we cannot defer the release of the page either. | |
5355 | * Instead, do flush now. | |
5356 | */ | |
5357 | if (force_flush) | |
5358 | tlb_flush_mmu_tlbonly(tlb); | |
1da177e4 | 5359 | } |
63551ae0 | 5360 | |
d833352a MG |
5361 | void __unmap_hugepage_range_final(struct mmu_gather *tlb, |
5362 | struct vm_area_struct *vma, unsigned long start, | |
05e90bd0 PX |
5363 | unsigned long end, struct page *ref_page, |
5364 | zap_flags_t zap_flags) | |
d833352a | 5365 | { |
131a79b4 MK |
5366 | hugetlb_vma_lock_write(vma); |
5367 | i_mmap_lock_write(vma->vm_file->f_mapping); | |
5368 | ||
369258ce | 5369 | /* mmu notification performed in caller */ |
05e90bd0 | 5370 | __unmap_hugepage_range(tlb, vma, start, end, ref_page, zap_flags); |
d833352a | 5371 | |
04ada095 MK |
5372 | if (zap_flags & ZAP_FLAG_UNMAP) { /* final unmap */ |
5373 | /* | |
5374 | * Unlock and free the vma lock before releasing i_mmap_rwsem. | |
5375 | * When the vma_lock is freed, this makes the vma ineligible | |
5376 | * for pmd sharing. And, i_mmap_rwsem is required to set up | |
5377 | * pmd sharing. This is important as page tables for this | |
5378 | * unmapped range will be asynchrously deleted. If the page | |
5379 | * tables are shared, there will be issues when accessed by | |
5380 | * someone else. | |
5381 | */ | |
5382 | __hugetlb_vma_unlock_write_free(vma); | |
5383 | i_mmap_unlock_write(vma->vm_file->f_mapping); | |
5384 | } else { | |
5385 | i_mmap_unlock_write(vma->vm_file->f_mapping); | |
5386 | hugetlb_vma_unlock_write(vma); | |
5387 | } | |
d833352a MG |
5388 | } |
5389 | ||
502717f4 | 5390 | void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start, |
05e90bd0 PX |
5391 | unsigned long end, struct page *ref_page, |
5392 | zap_flags_t zap_flags) | |
502717f4 | 5393 | { |
369258ce | 5394 | struct mmu_notifier_range range; |
24669e58 | 5395 | struct mmu_gather tlb; |
dff11abe | 5396 | |
369258ce MK |
5397 | mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm, |
5398 | start, end); | |
5399 | adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end); | |
5400 | mmu_notifier_invalidate_range_start(&range); | |
a72afd87 | 5401 | tlb_gather_mmu(&tlb, vma->vm_mm); |
369258ce | 5402 | |
05e90bd0 | 5403 | __unmap_hugepage_range(&tlb, vma, start, end, ref_page, zap_flags); |
369258ce MK |
5404 | |
5405 | mmu_notifier_invalidate_range_end(&range); | |
ae8eba8b | 5406 | tlb_finish_mmu(&tlb); |
502717f4 CK |
5407 | } |
5408 | ||
04f2cbe3 MG |
5409 | /* |
5410 | * This is called when the original mapper is failing to COW a MAP_PRIVATE | |
578b7725 | 5411 | * mapping it owns the reserve page for. The intention is to unmap the page |
04f2cbe3 MG |
5412 | * from other VMAs and let the children be SIGKILLed if they are faulting the |
5413 | * same region. | |
5414 | */ | |
2f4612af DB |
5415 | static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma, |
5416 | struct page *page, unsigned long address) | |
04f2cbe3 | 5417 | { |
7526674d | 5418 | struct hstate *h = hstate_vma(vma); |
04f2cbe3 MG |
5419 | struct vm_area_struct *iter_vma; |
5420 | struct address_space *mapping; | |
04f2cbe3 MG |
5421 | pgoff_t pgoff; |
5422 | ||
5423 | /* | |
5424 | * vm_pgoff is in PAGE_SIZE units, hence the different calculation | |
5425 | * from page cache lookup which is in HPAGE_SIZE units. | |
5426 | */ | |
7526674d | 5427 | address = address & huge_page_mask(h); |
36e4f20a MH |
5428 | pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) + |
5429 | vma->vm_pgoff; | |
93c76a3d | 5430 | mapping = vma->vm_file->f_mapping; |
04f2cbe3 | 5431 | |
4eb2b1dc MG |
5432 | /* |
5433 | * Take the mapping lock for the duration of the table walk. As | |
5434 | * this mapping should be shared between all the VMAs, | |
5435 | * __unmap_hugepage_range() is called as the lock is already held | |
5436 | */ | |
83cde9e8 | 5437 | i_mmap_lock_write(mapping); |
6b2dbba8 | 5438 | vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) { |
04f2cbe3 MG |
5439 | /* Do not unmap the current VMA */ |
5440 | if (iter_vma == vma) | |
5441 | continue; | |
5442 | ||
2f84a899 MG |
5443 | /* |
5444 | * Shared VMAs have their own reserves and do not affect | |
5445 | * MAP_PRIVATE accounting but it is possible that a shared | |
5446 | * VMA is using the same page so check and skip such VMAs. | |
5447 | */ | |
5448 | if (iter_vma->vm_flags & VM_MAYSHARE) | |
5449 | continue; | |
5450 | ||
04f2cbe3 MG |
5451 | /* |
5452 | * Unmap the page from other VMAs without their own reserves. | |
5453 | * They get marked to be SIGKILLed if they fault in these | |
5454 | * areas. This is because a future no-page fault on this VMA | |
5455 | * could insert a zeroed page instead of the data existing | |
5456 | * from the time of fork. This would look like data corruption | |
5457 | */ | |
5458 | if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER)) | |
24669e58 | 5459 | unmap_hugepage_range(iter_vma, address, |
05e90bd0 | 5460 | address + huge_page_size(h), page, 0); |
04f2cbe3 | 5461 | } |
83cde9e8 | 5462 | i_mmap_unlock_write(mapping); |
04f2cbe3 MG |
5463 | } |
5464 | ||
0fe6e20b | 5465 | /* |
c89357e2 | 5466 | * hugetlb_wp() should be called with page lock of the original hugepage held. |
aa6d2e8c | 5467 | * Called with hugetlb_fault_mutex_table held and pte_page locked so we |
ef009b25 MH |
5468 | * cannot race with other handlers or page migration. |
5469 | * Keep the pte_same checks anyway to make transition from the mutex easier. | |
0fe6e20b | 5470 | */ |
c89357e2 DH |
5471 | static vm_fault_t hugetlb_wp(struct mm_struct *mm, struct vm_area_struct *vma, |
5472 | unsigned long address, pte_t *ptep, unsigned int flags, | |
3999f52e | 5473 | struct page *pagecache_page, spinlock_t *ptl) |
1e8f889b | 5474 | { |
c89357e2 | 5475 | const bool unshare = flags & FAULT_FLAG_UNSHARE; |
3999f52e | 5476 | pte_t pte; |
a5516438 | 5477 | struct hstate *h = hstate_vma(vma); |
1e8f889b | 5478 | struct page *old_page, *new_page; |
2b740303 SJ |
5479 | int outside_reserve = 0; |
5480 | vm_fault_t ret = 0; | |
974e6d66 | 5481 | unsigned long haddr = address & huge_page_mask(h); |
ac46d4f3 | 5482 | struct mmu_notifier_range range; |
1e8f889b | 5483 | |
1d8d1464 DH |
5484 | /* |
5485 | * hugetlb does not support FOLL_FORCE-style write faults that keep the | |
5486 | * PTE mapped R/O such as maybe_mkwrite() would do. | |
5487 | */ | |
5488 | if (WARN_ON_ONCE(!unshare && !(vma->vm_flags & VM_WRITE))) | |
5489 | return VM_FAULT_SIGSEGV; | |
5490 | ||
5491 | /* Let's take out MAP_SHARED mappings first. */ | |
5492 | if (vma->vm_flags & VM_MAYSHARE) { | |
1d8d1464 DH |
5493 | set_huge_ptep_writable(vma, haddr, ptep); |
5494 | return 0; | |
5495 | } | |
5496 | ||
3999f52e | 5497 | pte = huge_ptep_get(ptep); |
1e8f889b DG |
5498 | old_page = pte_page(pte); |
5499 | ||
662ce1dc YY |
5500 | delayacct_wpcopy_start(); |
5501 | ||
04f2cbe3 | 5502 | retry_avoidcopy: |
c89357e2 DH |
5503 | /* |
5504 | * If no-one else is actually using this page, we're the exclusive | |
5505 | * owner and can reuse this page. | |
5506 | */ | |
37a2140d | 5507 | if (page_mapcount(old_page) == 1 && PageAnon(old_page)) { |
c89357e2 DH |
5508 | if (!PageAnonExclusive(old_page)) |
5509 | page_move_anon_rmap(old_page, vma); | |
5510 | if (likely(!unshare)) | |
5511 | set_huge_ptep_writable(vma, haddr, ptep); | |
662ce1dc YY |
5512 | |
5513 | delayacct_wpcopy_end(); | |
83c54070 | 5514 | return 0; |
1e8f889b | 5515 | } |
6c287605 DH |
5516 | VM_BUG_ON_PAGE(PageAnon(old_page) && PageAnonExclusive(old_page), |
5517 | old_page); | |
1e8f889b | 5518 | |
04f2cbe3 MG |
5519 | /* |
5520 | * If the process that created a MAP_PRIVATE mapping is about to | |
5521 | * perform a COW due to a shared page count, attempt to satisfy | |
5522 | * the allocation without using the existing reserves. The pagecache | |
5523 | * page is used to determine if the reserve at this address was | |
5524 | * consumed or not. If reserves were used, a partial faulted mapping | |
5525 | * at the time of fork() could consume its reserves on COW instead | |
5526 | * of the full address range. | |
5527 | */ | |
5944d011 | 5528 | if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && |
04f2cbe3 MG |
5529 | old_page != pagecache_page) |
5530 | outside_reserve = 1; | |
5531 | ||
09cbfeaf | 5532 | get_page(old_page); |
b76c8cfb | 5533 | |
ad4404a2 DB |
5534 | /* |
5535 | * Drop page table lock as buddy allocator may be called. It will | |
5536 | * be acquired again before returning to the caller, as expected. | |
5537 | */ | |
cb900f41 | 5538 | spin_unlock(ptl); |
5b7a1d40 | 5539 | new_page = alloc_huge_page(vma, haddr, outside_reserve); |
1e8f889b | 5540 | |
2fc39cec | 5541 | if (IS_ERR(new_page)) { |
04f2cbe3 MG |
5542 | /* |
5543 | * If a process owning a MAP_PRIVATE mapping fails to COW, | |
5544 | * it is due to references held by a child and an insufficient | |
5545 | * huge page pool. To guarantee the original mappers | |
5546 | * reliability, unmap the page from child processes. The child | |
5547 | * may get SIGKILLed if it later faults. | |
5548 | */ | |
5549 | if (outside_reserve) { | |
40549ba8 MK |
5550 | struct address_space *mapping = vma->vm_file->f_mapping; |
5551 | pgoff_t idx; | |
5552 | u32 hash; | |
5553 | ||
09cbfeaf | 5554 | put_page(old_page); |
40549ba8 MK |
5555 | /* |
5556 | * Drop hugetlb_fault_mutex and vma_lock before | |
5557 | * unmapping. unmapping needs to hold vma_lock | |
5558 | * in write mode. Dropping vma_lock in read mode | |
5559 | * here is OK as COW mappings do not interact with | |
5560 | * PMD sharing. | |
5561 | * | |
5562 | * Reacquire both after unmap operation. | |
5563 | */ | |
5564 | idx = vma_hugecache_offset(h, vma, haddr); | |
5565 | hash = hugetlb_fault_mutex_hash(mapping, idx); | |
5566 | hugetlb_vma_unlock_read(vma); | |
5567 | mutex_unlock(&hugetlb_fault_mutex_table[hash]); | |
5568 | ||
5b7a1d40 | 5569 | unmap_ref_private(mm, vma, old_page, haddr); |
40549ba8 MK |
5570 | |
5571 | mutex_lock(&hugetlb_fault_mutex_table[hash]); | |
5572 | hugetlb_vma_lock_read(vma); | |
2f4612af | 5573 | spin_lock(ptl); |
5b7a1d40 | 5574 | ptep = huge_pte_offset(mm, haddr, huge_page_size(h)); |
2f4612af DB |
5575 | if (likely(ptep && |
5576 | pte_same(huge_ptep_get(ptep), pte))) | |
5577 | goto retry_avoidcopy; | |
5578 | /* | |
5579 | * race occurs while re-acquiring page table | |
5580 | * lock, and our job is done. | |
5581 | */ | |
662ce1dc | 5582 | delayacct_wpcopy_end(); |
2f4612af | 5583 | return 0; |
04f2cbe3 MG |
5584 | } |
5585 | ||
2b740303 | 5586 | ret = vmf_error(PTR_ERR(new_page)); |
ad4404a2 | 5587 | goto out_release_old; |
1e8f889b DG |
5588 | } |
5589 | ||
0fe6e20b NH |
5590 | /* |
5591 | * When the original hugepage is shared one, it does not have | |
5592 | * anon_vma prepared. | |
5593 | */ | |
44e2aa93 | 5594 | if (unlikely(anon_vma_prepare(vma))) { |
ad4404a2 DB |
5595 | ret = VM_FAULT_OOM; |
5596 | goto out_release_all; | |
44e2aa93 | 5597 | } |
0fe6e20b | 5598 | |
974e6d66 | 5599 | copy_user_huge_page(new_page, old_page, address, vma, |
47ad8475 | 5600 | pages_per_huge_page(h)); |
0ed361de | 5601 | __SetPageUptodate(new_page); |
1e8f889b | 5602 | |
7269f999 | 5603 | mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, haddr, |
6f4f13e8 | 5604 | haddr + huge_page_size(h)); |
ac46d4f3 | 5605 | mmu_notifier_invalidate_range_start(&range); |
ad4404a2 | 5606 | |
b76c8cfb | 5607 | /* |
cb900f41 | 5608 | * Retake the page table lock to check for racing updates |
b76c8cfb LW |
5609 | * before the page tables are altered |
5610 | */ | |
cb900f41 | 5611 | spin_lock(ptl); |
5b7a1d40 | 5612 | ptep = huge_pte_offset(mm, haddr, huge_page_size(h)); |
a9af0c5d | 5613 | if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) { |
c89357e2 | 5614 | /* Break COW or unshare */ |
5b7a1d40 | 5615 | huge_ptep_clear_flush(vma, haddr, ptep); |
ac46d4f3 | 5616 | mmu_notifier_invalidate_range(mm, range.start, range.end); |
cea86fe2 | 5617 | page_remove_rmap(old_page, vma, true); |
5b7a1d40 | 5618 | hugepage_add_new_anon_rmap(new_page, vma, haddr); |
1eba86c0 | 5619 | set_huge_pte_at(mm, haddr, ptep, |
c89357e2 | 5620 | make_huge_pte(vma, new_page, !unshare)); |
8f251a3d | 5621 | SetHPageMigratable(new_page); |
1e8f889b DG |
5622 | /* Make the old page be freed below */ |
5623 | new_page = old_page; | |
5624 | } | |
cb900f41 | 5625 | spin_unlock(ptl); |
ac46d4f3 | 5626 | mmu_notifier_invalidate_range_end(&range); |
ad4404a2 | 5627 | out_release_all: |
c89357e2 DH |
5628 | /* |
5629 | * No restore in case of successful pagetable update (Break COW or | |
5630 | * unshare) | |
5631 | */ | |
c7b1850d MK |
5632 | if (new_page != old_page) |
5633 | restore_reserve_on_error(h, vma, haddr, new_page); | |
09cbfeaf | 5634 | put_page(new_page); |
ad4404a2 | 5635 | out_release_old: |
09cbfeaf | 5636 | put_page(old_page); |
8312034f | 5637 | |
ad4404a2 | 5638 | spin_lock(ptl); /* Caller expects lock to be held */ |
662ce1dc YY |
5639 | |
5640 | delayacct_wpcopy_end(); | |
ad4404a2 | 5641 | return ret; |
1e8f889b DG |
5642 | } |
5643 | ||
3ae77f43 HD |
5644 | /* |
5645 | * Return whether there is a pagecache page to back given address within VMA. | |
5646 | * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page. | |
5647 | */ | |
5648 | static bool hugetlbfs_pagecache_present(struct hstate *h, | |
2a15efc9 HD |
5649 | struct vm_area_struct *vma, unsigned long address) |
5650 | { | |
5651 | struct address_space *mapping; | |
5652 | pgoff_t idx; | |
5653 | struct page *page; | |
5654 | ||
5655 | mapping = vma->vm_file->f_mapping; | |
5656 | idx = vma_hugecache_offset(h, vma, address); | |
5657 | ||
5658 | page = find_get_page(mapping, idx); | |
5659 | if (page) | |
5660 | put_page(page); | |
5661 | return page != NULL; | |
5662 | } | |
5663 | ||
7e1813d4 | 5664 | int hugetlb_add_to_page_cache(struct page *page, struct address_space *mapping, |
ab76ad54 MK |
5665 | pgoff_t idx) |
5666 | { | |
d9ef44de | 5667 | struct folio *folio = page_folio(page); |
ab76ad54 MK |
5668 | struct inode *inode = mapping->host; |
5669 | struct hstate *h = hstate_inode(inode); | |
d9ef44de | 5670 | int err; |
ab76ad54 | 5671 | |
d9ef44de MWO |
5672 | __folio_set_locked(folio); |
5673 | err = __filemap_add_folio(mapping, folio, idx, GFP_KERNEL, NULL); | |
5674 | ||
5675 | if (unlikely(err)) { | |
5676 | __folio_clear_locked(folio); | |
ab76ad54 | 5677 | return err; |
d9ef44de | 5678 | } |
d6995da3 | 5679 | ClearHPageRestoreReserve(page); |
ab76ad54 | 5680 | |
22146c3c | 5681 | /* |
d9ef44de | 5682 | * mark folio dirty so that it will not be removed from cache/file |
22146c3c MK |
5683 | * by non-hugetlbfs specific code paths. |
5684 | */ | |
d9ef44de | 5685 | folio_mark_dirty(folio); |
22146c3c | 5686 | |
ab76ad54 MK |
5687 | spin_lock(&inode->i_lock); |
5688 | inode->i_blocks += blocks_per_huge_page(h); | |
5689 | spin_unlock(&inode->i_lock); | |
5690 | return 0; | |
5691 | } | |
5692 | ||
7677f7fd AR |
5693 | static inline vm_fault_t hugetlb_handle_userfault(struct vm_area_struct *vma, |
5694 | struct address_space *mapping, | |
5695 | pgoff_t idx, | |
5696 | unsigned int flags, | |
5697 | unsigned long haddr, | |
824ddc60 | 5698 | unsigned long addr, |
7677f7fd AR |
5699 | unsigned long reason) |
5700 | { | |
7677f7fd AR |
5701 | u32 hash; |
5702 | struct vm_fault vmf = { | |
5703 | .vma = vma, | |
5704 | .address = haddr, | |
824ddc60 | 5705 | .real_address = addr, |
7677f7fd AR |
5706 | .flags = flags, |
5707 | ||
5708 | /* | |
5709 | * Hard to debug if it ends up being | |
5710 | * used by a callee that assumes | |
5711 | * something about the other | |
5712 | * uninitialized fields... same as in | |
5713 | * memory.c | |
5714 | */ | |
5715 | }; | |
5716 | ||
5717 | /* | |
958f32ce LS |
5718 | * vma_lock and hugetlb_fault_mutex must be dropped before handling |
5719 | * userfault. Also mmap_lock could be dropped due to handling | |
5720 | * userfault, any vma operation should be careful from here. | |
7677f7fd | 5721 | */ |
40549ba8 | 5722 | hugetlb_vma_unlock_read(vma); |
7677f7fd AR |
5723 | hash = hugetlb_fault_mutex_hash(mapping, idx); |
5724 | mutex_unlock(&hugetlb_fault_mutex_table[hash]); | |
958f32ce | 5725 | return handle_userfault(&vmf, reason); |
7677f7fd AR |
5726 | } |
5727 | ||
2ea7ff1e PX |
5728 | /* |
5729 | * Recheck pte with pgtable lock. Returns true if pte didn't change, or | |
5730 | * false if pte changed or is changing. | |
5731 | */ | |
5732 | static bool hugetlb_pte_stable(struct hstate *h, struct mm_struct *mm, | |
5733 | pte_t *ptep, pte_t old_pte) | |
5734 | { | |
5735 | spinlock_t *ptl; | |
5736 | bool same; | |
5737 | ||
5738 | ptl = huge_pte_lock(h, mm, ptep); | |
5739 | same = pte_same(huge_ptep_get(ptep), old_pte); | |
5740 | spin_unlock(ptl); | |
5741 | ||
5742 | return same; | |
5743 | } | |
5744 | ||
2b740303 SJ |
5745 | static vm_fault_t hugetlb_no_page(struct mm_struct *mm, |
5746 | struct vm_area_struct *vma, | |
5747 | struct address_space *mapping, pgoff_t idx, | |
c64e912c PX |
5748 | unsigned long address, pte_t *ptep, |
5749 | pte_t old_pte, unsigned int flags) | |
ac9b9c66 | 5750 | { |
a5516438 | 5751 | struct hstate *h = hstate_vma(vma); |
2b740303 | 5752 | vm_fault_t ret = VM_FAULT_SIGBUS; |
409eb8c2 | 5753 | int anon_rmap = 0; |
4c887265 | 5754 | unsigned long size; |
4c887265 | 5755 | struct page *page; |
1e8f889b | 5756 | pte_t new_pte; |
cb900f41 | 5757 | spinlock_t *ptl; |
285b8dca | 5758 | unsigned long haddr = address & huge_page_mask(h); |
c7b1850d | 5759 | bool new_page, new_pagecache_page = false; |
958f32ce | 5760 | u32 hash = hugetlb_fault_mutex_hash(mapping, idx); |
4c887265 | 5761 | |
04f2cbe3 MG |
5762 | /* |
5763 | * Currently, we are forced to kill the process in the event the | |
5764 | * original mapper has unmapped pages from the child due to a failed | |
c89357e2 DH |
5765 | * COW/unsharing. Warn that such a situation has occurred as it may not |
5766 | * be obvious. | |
04f2cbe3 MG |
5767 | */ |
5768 | if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) { | |
910154d5 | 5769 | pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n", |
ffb22af5 | 5770 | current->pid); |
958f32ce | 5771 | goto out; |
04f2cbe3 MG |
5772 | } |
5773 | ||
4c887265 | 5774 | /* |
188a3972 MK |
5775 | * Use page lock to guard against racing truncation |
5776 | * before we get page_table_lock. | |
4c887265 | 5777 | */ |
c7b1850d | 5778 | new_page = false; |
6bda666a CL |
5779 | page = find_lock_page(mapping, idx); |
5780 | if (!page) { | |
188a3972 MK |
5781 | size = i_size_read(mapping->host) >> huge_page_shift(h); |
5782 | if (idx >= size) | |
5783 | goto out; | |
7677f7fd | 5784 | /* Check for page in userfault range */ |
2ea7ff1e PX |
5785 | if (userfaultfd_missing(vma)) { |
5786 | /* | |
5787 | * Since hugetlb_no_page() was examining pte | |
5788 | * without pgtable lock, we need to re-test under | |
5789 | * lock because the pte may not be stable and could | |
5790 | * have changed from under us. Try to detect | |
5791 | * either changed or during-changing ptes and retry | |
5792 | * properly when needed. | |
5793 | * | |
5794 | * Note that userfaultfd is actually fine with | |
5795 | * false positives (e.g. caused by pte changed), | |
5796 | * but not wrong logical events (e.g. caused by | |
5797 | * reading a pte during changing). The latter can | |
5798 | * confuse the userspace, so the strictness is very | |
5799 | * much preferred. E.g., MISSING event should | |
5800 | * never happen on the page after UFFDIO_COPY has | |
5801 | * correctly installed the page and returned. | |
5802 | */ | |
5803 | if (!hugetlb_pte_stable(h, mm, ptep, old_pte)) { | |
5804 | ret = 0; | |
5805 | goto out; | |
5806 | } | |
5807 | ||
5808 | return hugetlb_handle_userfault(vma, mapping, idx, flags, | |
5809 | haddr, address, | |
5810 | VM_UFFD_MISSING); | |
5811 | } | |
1a1aad8a | 5812 | |
285b8dca | 5813 | page = alloc_huge_page(vma, haddr, 0); |
2fc39cec | 5814 | if (IS_ERR(page)) { |
4643d67e MK |
5815 | /* |
5816 | * Returning error will result in faulting task being | |
5817 | * sent SIGBUS. The hugetlb fault mutex prevents two | |
5818 | * tasks from racing to fault in the same page which | |
5819 | * could result in false unable to allocate errors. | |
5820 | * Page migration does not take the fault mutex, but | |
5821 | * does a clear then write of pte's under page table | |
5822 | * lock. Page fault code could race with migration, | |
5823 | * notice the clear pte and try to allocate a page | |
5824 | * here. Before returning error, get ptl and make | |
5825 | * sure there really is no pte entry. | |
5826 | */ | |
f9bf6c03 | 5827 | if (hugetlb_pte_stable(h, mm, ptep, old_pte)) |
d83e6c8a | 5828 | ret = vmf_error(PTR_ERR(page)); |
f9bf6c03 PX |
5829 | else |
5830 | ret = 0; | |
6bda666a CL |
5831 | goto out; |
5832 | } | |
47ad8475 | 5833 | clear_huge_page(page, address, pages_per_huge_page(h)); |
0ed361de | 5834 | __SetPageUptodate(page); |
cb6acd01 | 5835 | new_page = true; |
ac9b9c66 | 5836 | |
f83a275d | 5837 | if (vma->vm_flags & VM_MAYSHARE) { |
7e1813d4 | 5838 | int err = hugetlb_add_to_page_cache(page, mapping, idx); |
6bda666a | 5839 | if (err) { |
3a5497a2 ML |
5840 | /* |
5841 | * err can't be -EEXIST which implies someone | |
5842 | * else consumed the reservation since hugetlb | |
5843 | * fault mutex is held when add a hugetlb page | |
5844 | * to the page cache. So it's safe to call | |
5845 | * restore_reserve_on_error() here. | |
5846 | */ | |
5847 | restore_reserve_on_error(h, vma, haddr, page); | |
6bda666a | 5848 | put_page(page); |
6bda666a CL |
5849 | goto out; |
5850 | } | |
c7b1850d | 5851 | new_pagecache_page = true; |
23be7468 | 5852 | } else { |
6bda666a | 5853 | lock_page(page); |
0fe6e20b NH |
5854 | if (unlikely(anon_vma_prepare(vma))) { |
5855 | ret = VM_FAULT_OOM; | |
5856 | goto backout_unlocked; | |
5857 | } | |
409eb8c2 | 5858 | anon_rmap = 1; |
23be7468 | 5859 | } |
0fe6e20b | 5860 | } else { |
998b4382 NH |
5861 | /* |
5862 | * If memory error occurs between mmap() and fault, some process | |
5863 | * don't have hwpoisoned swap entry for errored virtual address. | |
5864 | * So we need to block hugepage fault by PG_hwpoison bit check. | |
5865 | */ | |
5866 | if (unlikely(PageHWPoison(page))) { | |
0eb98f15 | 5867 | ret = VM_FAULT_HWPOISON_LARGE | |
972dc4de | 5868 | VM_FAULT_SET_HINDEX(hstate_index(h)); |
998b4382 NH |
5869 | goto backout_unlocked; |
5870 | } | |
7677f7fd AR |
5871 | |
5872 | /* Check for page in userfault range. */ | |
5873 | if (userfaultfd_minor(vma)) { | |
5874 | unlock_page(page); | |
5875 | put_page(page); | |
2ea7ff1e PX |
5876 | /* See comment in userfaultfd_missing() block above */ |
5877 | if (!hugetlb_pte_stable(h, mm, ptep, old_pte)) { | |
5878 | ret = 0; | |
5879 | goto out; | |
5880 | } | |
5881 | return hugetlb_handle_userfault(vma, mapping, idx, flags, | |
5882 | haddr, address, | |
5883 | VM_UFFD_MINOR); | |
7677f7fd | 5884 | } |
6bda666a | 5885 | } |
1e8f889b | 5886 | |
57303d80 AW |
5887 | /* |
5888 | * If we are going to COW a private mapping later, we examine the | |
5889 | * pending reservations for this page now. This will ensure that | |
5890 | * any allocations necessary to record that reservation occur outside | |
5891 | * the spinlock. | |
5892 | */ | |
5e911373 | 5893 | if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { |
285b8dca | 5894 | if (vma_needs_reservation(h, vma, haddr) < 0) { |
2b26736c AW |
5895 | ret = VM_FAULT_OOM; |
5896 | goto backout_unlocked; | |
5897 | } | |
5e911373 | 5898 | /* Just decrements count, does not deallocate */ |
285b8dca | 5899 | vma_end_reservation(h, vma, haddr); |
5e911373 | 5900 | } |
57303d80 | 5901 | |
8bea8052 | 5902 | ptl = huge_pte_lock(h, mm, ptep); |
83c54070 | 5903 | ret = 0; |
c64e912c PX |
5904 | /* If pte changed from under us, retry */ |
5905 | if (!pte_same(huge_ptep_get(ptep), old_pte)) | |
4c887265 AL |
5906 | goto backout; |
5907 | ||
4781593d | 5908 | if (anon_rmap) |
285b8dca | 5909 | hugepage_add_new_anon_rmap(page, vma, haddr); |
4781593d | 5910 | else |
fb3d824d | 5911 | page_dup_file_rmap(page, true); |
1e8f889b DG |
5912 | new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE) |
5913 | && (vma->vm_flags & VM_SHARED))); | |
c64e912c PX |
5914 | /* |
5915 | * If this pte was previously wr-protected, keep it wr-protected even | |
5916 | * if populated. | |
5917 | */ | |
5918 | if (unlikely(pte_marker_uffd_wp(old_pte))) | |
5919 | new_pte = huge_pte_wrprotect(huge_pte_mkuffd_wp(new_pte)); | |
285b8dca | 5920 | set_huge_pte_at(mm, haddr, ptep, new_pte); |
1e8f889b | 5921 | |
5d317b2b | 5922 | hugetlb_count_add(pages_per_huge_page(h), mm); |
788c7df4 | 5923 | if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { |
1e8f889b | 5924 | /* Optimization, do the COW without a second fault */ |
c89357e2 | 5925 | ret = hugetlb_wp(mm, vma, address, ptep, flags, page, ptl); |
1e8f889b DG |
5926 | } |
5927 | ||
cb900f41 | 5928 | spin_unlock(ptl); |
cb6acd01 MK |
5929 | |
5930 | /* | |
8f251a3d MK |
5931 | * Only set HPageMigratable in newly allocated pages. Existing pages |
5932 | * found in the pagecache may not have HPageMigratableset if they have | |
5933 | * been isolated for migration. | |
cb6acd01 MK |
5934 | */ |
5935 | if (new_page) | |
8f251a3d | 5936 | SetHPageMigratable(page); |
cb6acd01 | 5937 | |
4c887265 AL |
5938 | unlock_page(page); |
5939 | out: | |
958f32ce LS |
5940 | hugetlb_vma_unlock_read(vma); |
5941 | mutex_unlock(&hugetlb_fault_mutex_table[hash]); | |
ac9b9c66 | 5942 | return ret; |
4c887265 AL |
5943 | |
5944 | backout: | |
cb900f41 | 5945 | spin_unlock(ptl); |
2b26736c | 5946 | backout_unlocked: |
c7b1850d MK |
5947 | if (new_page && !new_pagecache_page) |
5948 | restore_reserve_on_error(h, vma, haddr, page); | |
fa27759a MK |
5949 | |
5950 | unlock_page(page); | |
4c887265 AL |
5951 | put_page(page); |
5952 | goto out; | |
ac9b9c66 HD |
5953 | } |
5954 | ||
8382d914 | 5955 | #ifdef CONFIG_SMP |
188b04a7 | 5956 | u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx) |
8382d914 DB |
5957 | { |
5958 | unsigned long key[2]; | |
5959 | u32 hash; | |
5960 | ||
1b426bac MK |
5961 | key[0] = (unsigned long) mapping; |
5962 | key[1] = idx; | |
8382d914 | 5963 | |
55254636 | 5964 | hash = jhash2((u32 *)&key, sizeof(key)/(sizeof(u32)), 0); |
8382d914 DB |
5965 | |
5966 | return hash & (num_fault_mutexes - 1); | |
5967 | } | |
5968 | #else | |
5969 | /* | |
6c26d310 | 5970 | * For uniprocessor systems we always use a single mutex, so just |
8382d914 DB |
5971 | * return 0 and avoid the hashing overhead. |
5972 | */ | |
188b04a7 | 5973 | u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx) |
8382d914 DB |
5974 | { |
5975 | return 0; | |
5976 | } | |
5977 | #endif | |
5978 | ||
2b740303 | 5979 | vm_fault_t hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma, |
788c7df4 | 5980 | unsigned long address, unsigned int flags) |
86e5216f | 5981 | { |
8382d914 | 5982 | pte_t *ptep, entry; |
cb900f41 | 5983 | spinlock_t *ptl; |
2b740303 | 5984 | vm_fault_t ret; |
8382d914 DB |
5985 | u32 hash; |
5986 | pgoff_t idx; | |
0fe6e20b | 5987 | struct page *page = NULL; |
57303d80 | 5988 | struct page *pagecache_page = NULL; |
a5516438 | 5989 | struct hstate *h = hstate_vma(vma); |
8382d914 | 5990 | struct address_space *mapping; |
0f792cf9 | 5991 | int need_wait_lock = 0; |
285b8dca | 5992 | unsigned long haddr = address & huge_page_mask(h); |
86e5216f | 5993 | |
285b8dca | 5994 | ptep = huge_pte_offset(mm, haddr, huge_page_size(h)); |
fd6a03ed | 5995 | if (ptep) { |
40549ba8 MK |
5996 | /* |
5997 | * Since we hold no locks, ptep could be stale. That is | |
5998 | * OK as we are only making decisions based on content and | |
5999 | * not actually modifying content here. | |
6000 | */ | |
fd6a03ed | 6001 | entry = huge_ptep_get(ptep); |
290408d4 | 6002 | if (unlikely(is_hugetlb_entry_migration(entry))) { |
ad1ac596 | 6003 | migration_entry_wait_huge(vma, ptep); |
290408d4 NH |
6004 | return 0; |
6005 | } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry))) | |
32f84528 | 6006 | return VM_FAULT_HWPOISON_LARGE | |
972dc4de | 6007 | VM_FAULT_SET_HINDEX(hstate_index(h)); |
fd6a03ed NH |
6008 | } |
6009 | ||
3935baa9 DG |
6010 | /* |
6011 | * Serialize hugepage allocation and instantiation, so that we don't | |
6012 | * get spurious allocation failures if two CPUs race to instantiate | |
6013 | * the same page in the page cache. | |
6014 | */ | |
40549ba8 MK |
6015 | mapping = vma->vm_file->f_mapping; |
6016 | idx = vma_hugecache_offset(h, vma, haddr); | |
188b04a7 | 6017 | hash = hugetlb_fault_mutex_hash(mapping, idx); |
c672c7f2 | 6018 | mutex_lock(&hugetlb_fault_mutex_table[hash]); |
8382d914 | 6019 | |
40549ba8 MK |
6020 | /* |
6021 | * Acquire vma lock before calling huge_pte_alloc and hold | |
6022 | * until finished with ptep. This prevents huge_pmd_unshare from | |
6023 | * being called elsewhere and making the ptep no longer valid. | |
6024 | * | |
6025 | * ptep could have already be assigned via huge_pte_offset. That | |
6026 | * is OK, as huge_pte_alloc will return the same value unless | |
6027 | * something has changed. | |
6028 | */ | |
6029 | hugetlb_vma_lock_read(vma); | |
6030 | ptep = huge_pte_alloc(mm, vma, haddr, huge_page_size(h)); | |
6031 | if (!ptep) { | |
6032 | hugetlb_vma_unlock_read(vma); | |
6033 | mutex_unlock(&hugetlb_fault_mutex_table[hash]); | |
6034 | return VM_FAULT_OOM; | |
6035 | } | |
6036 | ||
7f2e9525 | 6037 | entry = huge_ptep_get(ptep); |
c64e912c | 6038 | /* PTE markers should be handled the same way as none pte */ |
958f32ce LS |
6039 | if (huge_pte_none_mostly(entry)) |
6040 | /* | |
6041 | * hugetlb_no_page will drop vma lock and hugetlb fault | |
6042 | * mutex internally, which make us return immediately. | |
6043 | */ | |
6044 | return hugetlb_no_page(mm, vma, mapping, idx, address, ptep, | |
c64e912c | 6045 | entry, flags); |
86e5216f | 6046 | |
83c54070 | 6047 | ret = 0; |
1e8f889b | 6048 | |
0f792cf9 NH |
6049 | /* |
6050 | * entry could be a migration/hwpoison entry at this point, so this | |
6051 | * check prevents the kernel from going below assuming that we have | |
7c8de358 EP |
6052 | * an active hugepage in pagecache. This goto expects the 2nd page |
6053 | * fault, and is_hugetlb_entry_(migration|hwpoisoned) check will | |
6054 | * properly handle it. | |
0f792cf9 NH |
6055 | */ |
6056 | if (!pte_present(entry)) | |
6057 | goto out_mutex; | |
6058 | ||
57303d80 | 6059 | /* |
c89357e2 DH |
6060 | * If we are going to COW/unshare the mapping later, we examine the |
6061 | * pending reservations for this page now. This will ensure that any | |
57303d80 | 6062 | * allocations necessary to record that reservation occur outside the |
1d8d1464 DH |
6063 | * spinlock. Also lookup the pagecache page now as it is used to |
6064 | * determine if a reservation has been consumed. | |
57303d80 | 6065 | */ |
c89357e2 | 6066 | if ((flags & (FAULT_FLAG_WRITE|FAULT_FLAG_UNSHARE)) && |
1d8d1464 | 6067 | !(vma->vm_flags & VM_MAYSHARE) && !huge_pte_write(entry)) { |
285b8dca | 6068 | if (vma_needs_reservation(h, vma, haddr) < 0) { |
2b26736c | 6069 | ret = VM_FAULT_OOM; |
b4d1d99f | 6070 | goto out_mutex; |
2b26736c | 6071 | } |
5e911373 | 6072 | /* Just decrements count, does not deallocate */ |
285b8dca | 6073 | vma_end_reservation(h, vma, haddr); |
57303d80 | 6074 | |
29be8426 | 6075 | pagecache_page = find_lock_page(mapping, idx); |
57303d80 AW |
6076 | } |
6077 | ||
0f792cf9 NH |
6078 | ptl = huge_pte_lock(h, mm, ptep); |
6079 | ||
c89357e2 | 6080 | /* Check for a racing update before calling hugetlb_wp() */ |
0f792cf9 NH |
6081 | if (unlikely(!pte_same(entry, huge_ptep_get(ptep)))) |
6082 | goto out_ptl; | |
6083 | ||
166f3ecc PX |
6084 | /* Handle userfault-wp first, before trying to lock more pages */ |
6085 | if (userfaultfd_wp(vma) && huge_pte_uffd_wp(huge_ptep_get(ptep)) && | |
6086 | (flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) { | |
6087 | struct vm_fault vmf = { | |
6088 | .vma = vma, | |
6089 | .address = haddr, | |
6090 | .real_address = address, | |
6091 | .flags = flags, | |
6092 | }; | |
6093 | ||
6094 | spin_unlock(ptl); | |
6095 | if (pagecache_page) { | |
6096 | unlock_page(pagecache_page); | |
6097 | put_page(pagecache_page); | |
6098 | } | |
40549ba8 | 6099 | hugetlb_vma_unlock_read(vma); |
166f3ecc | 6100 | mutex_unlock(&hugetlb_fault_mutex_table[hash]); |
166f3ecc PX |
6101 | return handle_userfault(&vmf, VM_UFFD_WP); |
6102 | } | |
6103 | ||
56c9cfb1 | 6104 | /* |
c89357e2 | 6105 | * hugetlb_wp() requires page locks of pte_page(entry) and |
56c9cfb1 NH |
6106 | * pagecache_page, so here we need take the former one |
6107 | * when page != pagecache_page or !pagecache_page. | |
56c9cfb1 NH |
6108 | */ |
6109 | page = pte_page(entry); | |
6110 | if (page != pagecache_page) | |
0f792cf9 NH |
6111 | if (!trylock_page(page)) { |
6112 | need_wait_lock = 1; | |
6113 | goto out_ptl; | |
6114 | } | |
b4d1d99f | 6115 | |
0f792cf9 | 6116 | get_page(page); |
b4d1d99f | 6117 | |
c89357e2 | 6118 | if (flags & (FAULT_FLAG_WRITE|FAULT_FLAG_UNSHARE)) { |
106c992a | 6119 | if (!huge_pte_write(entry)) { |
c89357e2 DH |
6120 | ret = hugetlb_wp(mm, vma, address, ptep, flags, |
6121 | pagecache_page, ptl); | |
0f792cf9 | 6122 | goto out_put_page; |
c89357e2 DH |
6123 | } else if (likely(flags & FAULT_FLAG_WRITE)) { |
6124 | entry = huge_pte_mkdirty(entry); | |
b4d1d99f | 6125 | } |
b4d1d99f DG |
6126 | } |
6127 | entry = pte_mkyoung(entry); | |
285b8dca | 6128 | if (huge_ptep_set_access_flags(vma, haddr, ptep, entry, |
788c7df4 | 6129 | flags & FAULT_FLAG_WRITE)) |
285b8dca | 6130 | update_mmu_cache(vma, haddr, ptep); |
0f792cf9 NH |
6131 | out_put_page: |
6132 | if (page != pagecache_page) | |
6133 | unlock_page(page); | |
6134 | put_page(page); | |
cb900f41 KS |
6135 | out_ptl: |
6136 | spin_unlock(ptl); | |
57303d80 AW |
6137 | |
6138 | if (pagecache_page) { | |
6139 | unlock_page(pagecache_page); | |
6140 | put_page(pagecache_page); | |
6141 | } | |
b4d1d99f | 6142 | out_mutex: |
40549ba8 | 6143 | hugetlb_vma_unlock_read(vma); |
c672c7f2 | 6144 | mutex_unlock(&hugetlb_fault_mutex_table[hash]); |
0f792cf9 NH |
6145 | /* |
6146 | * Generally it's safe to hold refcount during waiting page lock. But | |
6147 | * here we just wait to defer the next page fault to avoid busy loop and | |
6148 | * the page is not used after unlocked before returning from the current | |
6149 | * page fault. So we are safe from accessing freed page, even if we wait | |
6150 | * here without taking refcount. | |
6151 | */ | |
6152 | if (need_wait_lock) | |
6153 | wait_on_page_locked(page); | |
1e8f889b | 6154 | return ret; |
86e5216f AL |
6155 | } |
6156 | ||
714c1891 | 6157 | #ifdef CONFIG_USERFAULTFD |
8fb5debc MK |
6158 | /* |
6159 | * Used by userfaultfd UFFDIO_COPY. Based on mcopy_atomic_pte with | |
6160 | * modifications for huge pages. | |
6161 | */ | |
6162 | int hugetlb_mcopy_atomic_pte(struct mm_struct *dst_mm, | |
6163 | pte_t *dst_pte, | |
6164 | struct vm_area_struct *dst_vma, | |
6165 | unsigned long dst_addr, | |
6166 | unsigned long src_addr, | |
f6191471 | 6167 | enum mcopy_atomic_mode mode, |
6041c691 PX |
6168 | struct page **pagep, |
6169 | bool wp_copy) | |
8fb5debc | 6170 | { |
f6191471 | 6171 | bool is_continue = (mode == MCOPY_ATOMIC_CONTINUE); |
8cc5fcbb MA |
6172 | struct hstate *h = hstate_vma(dst_vma); |
6173 | struct address_space *mapping = dst_vma->vm_file->f_mapping; | |
6174 | pgoff_t idx = vma_hugecache_offset(h, dst_vma, dst_addr); | |
1e392147 | 6175 | unsigned long size; |
1c9e8def | 6176 | int vm_shared = dst_vma->vm_flags & VM_SHARED; |
8fb5debc MK |
6177 | pte_t _dst_pte; |
6178 | spinlock_t *ptl; | |
8cc5fcbb | 6179 | int ret = -ENOMEM; |
8fb5debc | 6180 | struct page *page; |
f6191471 | 6181 | int writable; |
cc30042d | 6182 | bool page_in_pagecache = false; |
8fb5debc | 6183 | |
f6191471 AR |
6184 | if (is_continue) { |
6185 | ret = -EFAULT; | |
6186 | page = find_lock_page(mapping, idx); | |
6187 | if (!page) | |
6188 | goto out; | |
cc30042d | 6189 | page_in_pagecache = true; |
f6191471 | 6190 | } else if (!*pagep) { |
d84cf06e MA |
6191 | /* If a page already exists, then it's UFFDIO_COPY for |
6192 | * a non-missing case. Return -EEXIST. | |
6193 | */ | |
6194 | if (vm_shared && | |
6195 | hugetlbfs_pagecache_present(h, dst_vma, dst_addr)) { | |
6196 | ret = -EEXIST; | |
6197 | goto out; | |
6198 | } | |
6199 | ||
8fb5debc | 6200 | page = alloc_huge_page(dst_vma, dst_addr, 0); |
d84cf06e MA |
6201 | if (IS_ERR(page)) { |
6202 | ret = -ENOMEM; | |
8fb5debc | 6203 | goto out; |
d84cf06e | 6204 | } |
8fb5debc MK |
6205 | |
6206 | ret = copy_huge_page_from_user(page, | |
6207 | (const void __user *) src_addr, | |
810a56b9 | 6208 | pages_per_huge_page(h), false); |
8fb5debc | 6209 | |
c1e8d7c6 | 6210 | /* fallback to copy_from_user outside mmap_lock */ |
8fb5debc | 6211 | if (unlikely(ret)) { |
9e368259 | 6212 | ret = -ENOENT; |
8cc5fcbb MA |
6213 | /* Free the allocated page which may have |
6214 | * consumed a reservation. | |
6215 | */ | |
6216 | restore_reserve_on_error(h, dst_vma, dst_addr, page); | |
6217 | put_page(page); | |
6218 | ||
6219 | /* Allocate a temporary page to hold the copied | |
6220 | * contents. | |
6221 | */ | |
6222 | page = alloc_huge_page_vma(h, dst_vma, dst_addr); | |
6223 | if (!page) { | |
6224 | ret = -ENOMEM; | |
6225 | goto out; | |
6226 | } | |
8fb5debc | 6227 | *pagep = page; |
8cc5fcbb MA |
6228 | /* Set the outparam pagep and return to the caller to |
6229 | * copy the contents outside the lock. Don't free the | |
6230 | * page. | |
6231 | */ | |
8fb5debc MK |
6232 | goto out; |
6233 | } | |
6234 | } else { | |
8cc5fcbb MA |
6235 | if (vm_shared && |
6236 | hugetlbfs_pagecache_present(h, dst_vma, dst_addr)) { | |
6237 | put_page(*pagep); | |
6238 | ret = -EEXIST; | |
6239 | *pagep = NULL; | |
6240 | goto out; | |
6241 | } | |
6242 | ||
6243 | page = alloc_huge_page(dst_vma, dst_addr, 0); | |
6244 | if (IS_ERR(page)) { | |
da9a298f | 6245 | put_page(*pagep); |
8cc5fcbb MA |
6246 | ret = -ENOMEM; |
6247 | *pagep = NULL; | |
6248 | goto out; | |
6249 | } | |
34892366 MS |
6250 | copy_user_huge_page(page, *pagep, dst_addr, dst_vma, |
6251 | pages_per_huge_page(h)); | |
8cc5fcbb | 6252 | put_page(*pagep); |
8fb5debc MK |
6253 | *pagep = NULL; |
6254 | } | |
6255 | ||
6256 | /* | |
6257 | * The memory barrier inside __SetPageUptodate makes sure that | |
6258 | * preceding stores to the page contents become visible before | |
6259 | * the set_pte_at() write. | |
6260 | */ | |
6261 | __SetPageUptodate(page); | |
8fb5debc | 6262 | |
f6191471 AR |
6263 | /* Add shared, newly allocated pages to the page cache. */ |
6264 | if (vm_shared && !is_continue) { | |
1e392147 AA |
6265 | size = i_size_read(mapping->host) >> huge_page_shift(h); |
6266 | ret = -EFAULT; | |
6267 | if (idx >= size) | |
6268 | goto out_release_nounlock; | |
1c9e8def | 6269 | |
1e392147 AA |
6270 | /* |
6271 | * Serialization between remove_inode_hugepages() and | |
7e1813d4 | 6272 | * hugetlb_add_to_page_cache() below happens through the |
1e392147 AA |
6273 | * hugetlb_fault_mutex_table that here must be hold by |
6274 | * the caller. | |
6275 | */ | |
7e1813d4 | 6276 | ret = hugetlb_add_to_page_cache(page, mapping, idx); |
1c9e8def MK |
6277 | if (ret) |
6278 | goto out_release_nounlock; | |
cc30042d | 6279 | page_in_pagecache = true; |
1c9e8def MK |
6280 | } |
6281 | ||
bcc66543 | 6282 | ptl = huge_pte_lock(h, dst_mm, dst_pte); |
8fb5debc | 6283 | |
8625147c JH |
6284 | ret = -EIO; |
6285 | if (PageHWPoison(page)) | |
6286 | goto out_release_unlock; | |
6287 | ||
6041c691 PX |
6288 | /* |
6289 | * We allow to overwrite a pte marker: consider when both MISSING|WP | |
6290 | * registered, we firstly wr-protect a none pte which has no page cache | |
6291 | * page backing it, then access the page. | |
6292 | */ | |
fa27759a | 6293 | ret = -EEXIST; |
6041c691 | 6294 | if (!huge_pte_none_mostly(huge_ptep_get(dst_pte))) |
8fb5debc MK |
6295 | goto out_release_unlock; |
6296 | ||
4781593d | 6297 | if (page_in_pagecache) |
fb3d824d | 6298 | page_dup_file_rmap(page, true); |
4781593d | 6299 | else |
1c9e8def | 6300 | hugepage_add_new_anon_rmap(page, dst_vma, dst_addr); |
8fb5debc | 6301 | |
6041c691 PX |
6302 | /* |
6303 | * For either: (1) CONTINUE on a non-shared VMA, or (2) UFFDIO_COPY | |
6304 | * with wp flag set, don't set pte write bit. | |
6305 | */ | |
6306 | if (wp_copy || (is_continue && !vm_shared)) | |
f6191471 AR |
6307 | writable = 0; |
6308 | else | |
6309 | writable = dst_vma->vm_flags & VM_WRITE; | |
6310 | ||
6311 | _dst_pte = make_huge_pte(dst_vma, page, writable); | |
6041c691 PX |
6312 | /* |
6313 | * Always mark UFFDIO_COPY page dirty; note that this may not be | |
6314 | * extremely important for hugetlbfs for now since swapping is not | |
6315 | * supported, but we should still be clear in that this page cannot be | |
6316 | * thrown away at will, even if write bit not set. | |
6317 | */ | |
6318 | _dst_pte = huge_pte_mkdirty(_dst_pte); | |
8fb5debc MK |
6319 | _dst_pte = pte_mkyoung(_dst_pte); |
6320 | ||
6041c691 PX |
6321 | if (wp_copy) |
6322 | _dst_pte = huge_pte_mkuffd_wp(_dst_pte); | |
6323 | ||
8fb5debc MK |
6324 | set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte); |
6325 | ||
8fb5debc MK |
6326 | hugetlb_count_add(pages_per_huge_page(h), dst_mm); |
6327 | ||
6328 | /* No need to invalidate - it was non-present before */ | |
6329 | update_mmu_cache(dst_vma, dst_addr, dst_pte); | |
6330 | ||
6331 | spin_unlock(ptl); | |
f6191471 AR |
6332 | if (!is_continue) |
6333 | SetHPageMigratable(page); | |
6334 | if (vm_shared || is_continue) | |
1c9e8def | 6335 | unlock_page(page); |
8fb5debc MK |
6336 | ret = 0; |
6337 | out: | |
6338 | return ret; | |
6339 | out_release_unlock: | |
6340 | spin_unlock(ptl); | |
f6191471 | 6341 | if (vm_shared || is_continue) |
1c9e8def | 6342 | unlock_page(page); |
5af10dfd | 6343 | out_release_nounlock: |
cc30042d | 6344 | if (!page_in_pagecache) |
c7b1850d | 6345 | restore_reserve_on_error(h, dst_vma, dst_addr, page); |
8fb5debc MK |
6346 | put_page(page); |
6347 | goto out; | |
6348 | } | |
714c1891 | 6349 | #endif /* CONFIG_USERFAULTFD */ |
8fb5debc | 6350 | |
82e5d378 JM |
6351 | static void record_subpages_vmas(struct page *page, struct vm_area_struct *vma, |
6352 | int refs, struct page **pages, | |
6353 | struct vm_area_struct **vmas) | |
6354 | { | |
6355 | int nr; | |
6356 | ||
6357 | for (nr = 0; nr < refs; nr++) { | |
6358 | if (likely(pages)) | |
14455eab | 6359 | pages[nr] = nth_page(page, nr); |
82e5d378 JM |
6360 | if (vmas) |
6361 | vmas[nr] = vma; | |
6362 | } | |
6363 | } | |
6364 | ||
84209e87 DH |
6365 | static inline bool __follow_hugetlb_must_fault(struct vm_area_struct *vma, |
6366 | unsigned int flags, pte_t *pte, | |
a7f22660 DH |
6367 | bool *unshare) |
6368 | { | |
6369 | pte_t pteval = huge_ptep_get(pte); | |
6370 | ||
6371 | *unshare = false; | |
6372 | if (is_swap_pte(pteval)) | |
6373 | return true; | |
6374 | if (huge_pte_write(pteval)) | |
6375 | return false; | |
6376 | if (flags & FOLL_WRITE) | |
6377 | return true; | |
84209e87 | 6378 | if (gup_must_unshare(vma, flags, pte_page(pteval))) { |
a7f22660 DH |
6379 | *unshare = true; |
6380 | return true; | |
6381 | } | |
6382 | return false; | |
6383 | } | |
6384 | ||
57a196a5 MK |
6385 | struct page *hugetlb_follow_page_mask(struct vm_area_struct *vma, |
6386 | unsigned long address, unsigned int flags) | |
6387 | { | |
6388 | struct hstate *h = hstate_vma(vma); | |
6389 | struct mm_struct *mm = vma->vm_mm; | |
6390 | unsigned long haddr = address & huge_page_mask(h); | |
6391 | struct page *page = NULL; | |
6392 | spinlock_t *ptl; | |
6393 | pte_t *pte, entry; | |
6394 | ||
6395 | /* | |
6396 | * FOLL_PIN is not supported for follow_page(). Ordinary GUP goes via | |
6397 | * follow_hugetlb_page(). | |
6398 | */ | |
6399 | if (WARN_ON_ONCE(flags & FOLL_PIN)) | |
6400 | return NULL; | |
6401 | ||
6402 | retry: | |
6403 | pte = huge_pte_offset(mm, haddr, huge_page_size(h)); | |
6404 | if (!pte) | |
6405 | return NULL; | |
6406 | ||
6407 | ptl = huge_pte_lock(h, mm, pte); | |
6408 | entry = huge_ptep_get(pte); | |
6409 | if (pte_present(entry)) { | |
6410 | page = pte_page(entry) + | |
6411 | ((address & ~huge_page_mask(h)) >> PAGE_SHIFT); | |
6412 | /* | |
6413 | * Note that page may be a sub-page, and with vmemmap | |
6414 | * optimizations the page struct may be read only. | |
6415 | * try_grab_page() will increase the ref count on the | |
6416 | * head page, so this will be OK. | |
6417 | * | |
e2ca6ba6 LT |
6418 | * try_grab_page() should always be able to get the page here, |
6419 | * because we hold the ptl lock and have verified pte_present(). | |
57a196a5 | 6420 | */ |
e2ca6ba6 | 6421 | if (try_grab_page(page, flags)) { |
57a196a5 MK |
6422 | page = NULL; |
6423 | goto out; | |
6424 | } | |
6425 | } else { | |
6426 | if (is_hugetlb_entry_migration(entry)) { | |
6427 | spin_unlock(ptl); | |
6428 | __migration_entry_wait_huge(pte, ptl); | |
6429 | goto retry; | |
6430 | } | |
6431 | /* | |
6432 | * hwpoisoned entry is treated as no_page_table in | |
6433 | * follow_page_mask(). | |
6434 | */ | |
6435 | } | |
6436 | out: | |
6437 | spin_unlock(ptl); | |
6438 | return page; | |
6439 | } | |
6440 | ||
28a35716 ML |
6441 | long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma, |
6442 | struct page **pages, struct vm_area_struct **vmas, | |
6443 | unsigned long *position, unsigned long *nr_pages, | |
4f6da934 | 6444 | long i, unsigned int flags, int *locked) |
63551ae0 | 6445 | { |
d5d4b0aa CK |
6446 | unsigned long pfn_offset; |
6447 | unsigned long vaddr = *position; | |
28a35716 | 6448 | unsigned long remainder = *nr_pages; |
a5516438 | 6449 | struct hstate *h = hstate_vma(vma); |
0fa5bc40 | 6450 | int err = -EFAULT, refs; |
63551ae0 | 6451 | |
63551ae0 | 6452 | while (vaddr < vma->vm_end && remainder) { |
4c887265 | 6453 | pte_t *pte; |
cb900f41 | 6454 | spinlock_t *ptl = NULL; |
a7f22660 | 6455 | bool unshare = false; |
2a15efc9 | 6456 | int absent; |
4c887265 | 6457 | struct page *page; |
63551ae0 | 6458 | |
02057967 DR |
6459 | /* |
6460 | * If we have a pending SIGKILL, don't keep faulting pages and | |
6461 | * potentially allocating memory. | |
6462 | */ | |
fa45f116 | 6463 | if (fatal_signal_pending(current)) { |
02057967 DR |
6464 | remainder = 0; |
6465 | break; | |
6466 | } | |
6467 | ||
4c887265 AL |
6468 | /* |
6469 | * Some archs (sparc64, sh*) have multiple pte_ts to | |
2a15efc9 | 6470 | * each hugepage. We have to make sure we get the |
4c887265 | 6471 | * first, for the page indexing below to work. |
cb900f41 KS |
6472 | * |
6473 | * Note that page table lock is not held when pte is null. | |
4c887265 | 6474 | */ |
7868a208 PA |
6475 | pte = huge_pte_offset(mm, vaddr & huge_page_mask(h), |
6476 | huge_page_size(h)); | |
cb900f41 KS |
6477 | if (pte) |
6478 | ptl = huge_pte_lock(h, mm, pte); | |
2a15efc9 HD |
6479 | absent = !pte || huge_pte_none(huge_ptep_get(pte)); |
6480 | ||
6481 | /* | |
6482 | * When coredumping, it suits get_dump_page if we just return | |
3ae77f43 HD |
6483 | * an error where there's an empty slot with no huge pagecache |
6484 | * to back it. This way, we avoid allocating a hugepage, and | |
6485 | * the sparse dumpfile avoids allocating disk blocks, but its | |
6486 | * huge holes still show up with zeroes where they need to be. | |
2a15efc9 | 6487 | */ |
3ae77f43 HD |
6488 | if (absent && (flags & FOLL_DUMP) && |
6489 | !hugetlbfs_pagecache_present(h, vma, vaddr)) { | |
cb900f41 KS |
6490 | if (pte) |
6491 | spin_unlock(ptl); | |
2a15efc9 HD |
6492 | remainder = 0; |
6493 | break; | |
6494 | } | |
63551ae0 | 6495 | |
9cc3a5bd NH |
6496 | /* |
6497 | * We need call hugetlb_fault for both hugepages under migration | |
6498 | * (in which case hugetlb_fault waits for the migration,) and | |
6499 | * hwpoisoned hugepages (in which case we need to prevent the | |
6500 | * caller from accessing to them.) In order to do this, we use | |
6501 | * here is_swap_pte instead of is_hugetlb_entry_migration and | |
6502 | * is_hugetlb_entry_hwpoisoned. This is because it simply covers | |
6503 | * both cases, and because we can't follow correct pages | |
6504 | * directly from any kind of swap entries. | |
6505 | */ | |
a7f22660 | 6506 | if (absent || |
84209e87 | 6507 | __follow_hugetlb_must_fault(vma, flags, pte, &unshare)) { |
2b740303 | 6508 | vm_fault_t ret; |
87ffc118 | 6509 | unsigned int fault_flags = 0; |
63551ae0 | 6510 | |
cb900f41 KS |
6511 | if (pte) |
6512 | spin_unlock(ptl); | |
87ffc118 AA |
6513 | if (flags & FOLL_WRITE) |
6514 | fault_flags |= FAULT_FLAG_WRITE; | |
a7f22660 DH |
6515 | else if (unshare) |
6516 | fault_flags |= FAULT_FLAG_UNSHARE; | |
93c5c61d | 6517 | if (locked) { |
71335f37 PX |
6518 | fault_flags |= FAULT_FLAG_ALLOW_RETRY | |
6519 | FAULT_FLAG_KILLABLE; | |
93c5c61d PX |
6520 | if (flags & FOLL_INTERRUPTIBLE) |
6521 | fault_flags |= FAULT_FLAG_INTERRUPTIBLE; | |
6522 | } | |
87ffc118 AA |
6523 | if (flags & FOLL_NOWAIT) |
6524 | fault_flags |= FAULT_FLAG_ALLOW_RETRY | | |
6525 | FAULT_FLAG_RETRY_NOWAIT; | |
6526 | if (flags & FOLL_TRIED) { | |
4426e945 PX |
6527 | /* |
6528 | * Note: FAULT_FLAG_ALLOW_RETRY and | |
6529 | * FAULT_FLAG_TRIED can co-exist | |
6530 | */ | |
87ffc118 AA |
6531 | fault_flags |= FAULT_FLAG_TRIED; |
6532 | } | |
6533 | ret = hugetlb_fault(mm, vma, vaddr, fault_flags); | |
6534 | if (ret & VM_FAULT_ERROR) { | |
2be7cfed | 6535 | err = vm_fault_to_errno(ret, flags); |
87ffc118 AA |
6536 | remainder = 0; |
6537 | break; | |
6538 | } | |
6539 | if (ret & VM_FAULT_RETRY) { | |
4f6da934 | 6540 | if (locked && |
1ac25013 | 6541 | !(fault_flags & FAULT_FLAG_RETRY_NOWAIT)) |
4f6da934 | 6542 | *locked = 0; |
87ffc118 AA |
6543 | *nr_pages = 0; |
6544 | /* | |
6545 | * VM_FAULT_RETRY must not return an | |
6546 | * error, it will return zero | |
6547 | * instead. | |
6548 | * | |
6549 | * No need to update "position" as the | |
6550 | * caller will not check it after | |
6551 | * *nr_pages is set to 0. | |
6552 | */ | |
6553 | return i; | |
6554 | } | |
6555 | continue; | |
4c887265 AL |
6556 | } |
6557 | ||
a5516438 | 6558 | pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT; |
7f2e9525 | 6559 | page = pte_page(huge_ptep_get(pte)); |
8fde12ca | 6560 | |
b6a2619c DH |
6561 | VM_BUG_ON_PAGE((flags & FOLL_PIN) && PageAnon(page) && |
6562 | !PageAnonExclusive(page), page); | |
6563 | ||
acbfb087 ZL |
6564 | /* |
6565 | * If subpage information not requested, update counters | |
6566 | * and skip the same_page loop below. | |
6567 | */ | |
6568 | if (!pages && !vmas && !pfn_offset && | |
6569 | (vaddr + huge_page_size(h) < vma->vm_end) && | |
6570 | (remainder >= pages_per_huge_page(h))) { | |
6571 | vaddr += huge_page_size(h); | |
6572 | remainder -= pages_per_huge_page(h); | |
6573 | i += pages_per_huge_page(h); | |
6574 | spin_unlock(ptl); | |
6575 | continue; | |
6576 | } | |
6577 | ||
d08af0a5 JM |
6578 | /* vaddr may not be aligned to PAGE_SIZE */ |
6579 | refs = min3(pages_per_huge_page(h) - pfn_offset, remainder, | |
6580 | (vma->vm_end - ALIGN_DOWN(vaddr, PAGE_SIZE)) >> PAGE_SHIFT); | |
0fa5bc40 | 6581 | |
82e5d378 | 6582 | if (pages || vmas) |
14455eab | 6583 | record_subpages_vmas(nth_page(page, pfn_offset), |
82e5d378 JM |
6584 | vma, refs, |
6585 | likely(pages) ? pages + i : NULL, | |
6586 | vmas ? vmas + i : NULL); | |
63551ae0 | 6587 | |
82e5d378 | 6588 | if (pages) { |
0fa5bc40 | 6589 | /* |
822951d8 | 6590 | * try_grab_folio() should always succeed here, |
0fa5bc40 JM |
6591 | * because: a) we hold the ptl lock, and b) we've just |
6592 | * checked that the huge page is present in the page | |
6593 | * tables. If the huge page is present, then the tail | |
6594 | * pages must also be present. The ptl prevents the | |
6595 | * head page and tail pages from being rearranged in | |
4003f107 LG |
6596 | * any way. As this is hugetlb, the pages will never |
6597 | * be p2pdma or not longterm pinable. So this page | |
6598 | * must be available at this point, unless the page | |
6599 | * refcount overflowed: | |
0fa5bc40 | 6600 | */ |
822951d8 MWO |
6601 | if (WARN_ON_ONCE(!try_grab_folio(pages[i], refs, |
6602 | flags))) { | |
0fa5bc40 JM |
6603 | spin_unlock(ptl); |
6604 | remainder = 0; | |
6605 | err = -ENOMEM; | |
6606 | break; | |
6607 | } | |
d5d4b0aa | 6608 | } |
82e5d378 JM |
6609 | |
6610 | vaddr += (refs << PAGE_SHIFT); | |
6611 | remainder -= refs; | |
6612 | i += refs; | |
6613 | ||
cb900f41 | 6614 | spin_unlock(ptl); |
63551ae0 | 6615 | } |
28a35716 | 6616 | *nr_pages = remainder; |
87ffc118 AA |
6617 | /* |
6618 | * setting position is actually required only if remainder is | |
6619 | * not zero but it's faster not to add a "if (remainder)" | |
6620 | * branch. | |
6621 | */ | |
63551ae0 DG |
6622 | *position = vaddr; |
6623 | ||
2be7cfed | 6624 | return i ? i : err; |
63551ae0 | 6625 | } |
8f860591 | 6626 | |
7da4d641 | 6627 | unsigned long hugetlb_change_protection(struct vm_area_struct *vma, |
5a90d5a1 PX |
6628 | unsigned long address, unsigned long end, |
6629 | pgprot_t newprot, unsigned long cp_flags) | |
8f860591 ZY |
6630 | { |
6631 | struct mm_struct *mm = vma->vm_mm; | |
6632 | unsigned long start = address; | |
6633 | pte_t *ptep; | |
6634 | pte_t pte; | |
a5516438 | 6635 | struct hstate *h = hstate_vma(vma); |
60dfaad6 | 6636 | unsigned long pages = 0, psize = huge_page_size(h); |
dff11abe | 6637 | bool shared_pmd = false; |
ac46d4f3 | 6638 | struct mmu_notifier_range range; |
e95a9851 | 6639 | unsigned long last_addr_mask; |
5a90d5a1 PX |
6640 | bool uffd_wp = cp_flags & MM_CP_UFFD_WP; |
6641 | bool uffd_wp_resolve = cp_flags & MM_CP_UFFD_WP_RESOLVE; | |
dff11abe MK |
6642 | |
6643 | /* | |
6644 | * In the case of shared PMDs, the area to flush could be beyond | |
ac46d4f3 | 6645 | * start/end. Set range.start/range.end to cover the maximum possible |
dff11abe MK |
6646 | * range if PMD sharing is possible. |
6647 | */ | |
7269f999 JG |
6648 | mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_VMA, |
6649 | 0, vma, mm, start, end); | |
ac46d4f3 | 6650 | adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end); |
8f860591 ZY |
6651 | |
6652 | BUG_ON(address >= end); | |
ac46d4f3 | 6653 | flush_cache_range(vma, range.start, range.end); |
8f860591 | 6654 | |
ac46d4f3 | 6655 | mmu_notifier_invalidate_range_start(&range); |
40549ba8 | 6656 | hugetlb_vma_lock_write(vma); |
83cde9e8 | 6657 | i_mmap_lock_write(vma->vm_file->f_mapping); |
40549ba8 | 6658 | last_addr_mask = hugetlb_mask_last_page(h); |
60dfaad6 | 6659 | for (; address < end; address += psize) { |
cb900f41 | 6660 | spinlock_t *ptl; |
60dfaad6 | 6661 | ptep = huge_pte_offset(mm, address, psize); |
e95a9851 | 6662 | if (!ptep) { |
fed15f13 PX |
6663 | if (!uffd_wp) { |
6664 | address |= last_addr_mask; | |
6665 | continue; | |
6666 | } | |
6667 | /* | |
6668 | * Userfaultfd wr-protect requires pgtable | |
6669 | * pre-allocations to install pte markers. | |
6670 | */ | |
6671 | ptep = huge_pte_alloc(mm, vma, address, psize); | |
6672 | if (!ptep) | |
6673 | break; | |
e95a9851 | 6674 | } |
cb900f41 | 6675 | ptl = huge_pte_lock(h, mm, ptep); |
4ddb4d91 | 6676 | if (huge_pmd_unshare(mm, vma, address, ptep)) { |
60dfaad6 PX |
6677 | /* |
6678 | * When uffd-wp is enabled on the vma, unshare | |
6679 | * shouldn't happen at all. Warn about it if it | |
6680 | * happened due to some reason. | |
6681 | */ | |
6682 | WARN_ON_ONCE(uffd_wp || uffd_wp_resolve); | |
7da4d641 | 6683 | pages++; |
cb900f41 | 6684 | spin_unlock(ptl); |
dff11abe | 6685 | shared_pmd = true; |
4ddb4d91 | 6686 | address |= last_addr_mask; |
39dde65c | 6687 | continue; |
7da4d641 | 6688 | } |
a8bda28d NH |
6689 | pte = huge_ptep_get(ptep); |
6690 | if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) { | |
0e678153 DH |
6691 | /* Nothing to do. */ |
6692 | } else if (unlikely(is_hugetlb_entry_migration(pte))) { | |
a8bda28d | 6693 | swp_entry_t entry = pte_to_swp_entry(pte); |
6c287605 | 6694 | struct page *page = pfn_swap_entry_to_page(entry); |
44f86392 | 6695 | pte_t newpte = pte; |
a8bda28d | 6696 | |
44f86392 | 6697 | if (is_writable_migration_entry(entry)) { |
6c287605 DH |
6698 | if (PageAnon(page)) |
6699 | entry = make_readable_exclusive_migration_entry( | |
6700 | swp_offset(entry)); | |
6701 | else | |
6702 | entry = make_readable_migration_entry( | |
6703 | swp_offset(entry)); | |
a8bda28d | 6704 | newpte = swp_entry_to_pte(entry); |
a8bda28d NH |
6705 | pages++; |
6706 | } | |
44f86392 DH |
6707 | |
6708 | if (uffd_wp) | |
6709 | newpte = pte_swp_mkuffd_wp(newpte); | |
6710 | else if (uffd_wp_resolve) | |
6711 | newpte = pte_swp_clear_uffd_wp(newpte); | |
6712 | if (!pte_same(pte, newpte)) | |
6713 | set_huge_pte_at(mm, address, ptep, newpte); | |
0e678153 DH |
6714 | } else if (unlikely(is_pte_marker(pte))) { |
6715 | /* No other markers apply for now. */ | |
6716 | WARN_ON_ONCE(!pte_marker_uffd_wp(pte)); | |
60dfaad6 | 6717 | if (uffd_wp_resolve) |
0e678153 | 6718 | /* Safe to modify directly (non-present->none). */ |
60dfaad6 | 6719 | huge_pte_clear(mm, address, ptep, psize); |
0e678153 | 6720 | } else if (!huge_pte_none(pte)) { |
023bdd00 | 6721 | pte_t old_pte; |
79c1c594 | 6722 | unsigned int shift = huge_page_shift(hstate_vma(vma)); |
023bdd00 AK |
6723 | |
6724 | old_pte = huge_ptep_modify_prot_start(vma, address, ptep); | |
16785bd7 | 6725 | pte = huge_pte_modify(old_pte, newprot); |
79c1c594 | 6726 | pte = arch_make_huge_pte(pte, shift, vma->vm_flags); |
5a90d5a1 PX |
6727 | if (uffd_wp) |
6728 | pte = huge_pte_mkuffd_wp(huge_pte_wrprotect(pte)); | |
6729 | else if (uffd_wp_resolve) | |
6730 | pte = huge_pte_clear_uffd_wp(pte); | |
023bdd00 | 6731 | huge_ptep_modify_prot_commit(vma, address, ptep, old_pte, pte); |
7da4d641 | 6732 | pages++; |
60dfaad6 PX |
6733 | } else { |
6734 | /* None pte */ | |
6735 | if (unlikely(uffd_wp)) | |
6736 | /* Safe to modify directly (none->non-present). */ | |
6737 | set_huge_pte_at(mm, address, ptep, | |
6738 | make_pte_marker(PTE_MARKER_UFFD_WP)); | |
8f860591 | 6739 | } |
cb900f41 | 6740 | spin_unlock(ptl); |
8f860591 | 6741 | } |
d833352a | 6742 | /* |
c8c06efa | 6743 | * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare |
d833352a | 6744 | * may have cleared our pud entry and done put_page on the page table: |
c8c06efa | 6745 | * once we release i_mmap_rwsem, another task can do the final put_page |
dff11abe MK |
6746 | * and that page table be reused and filled with junk. If we actually |
6747 | * did unshare a page of pmds, flush the range corresponding to the pud. | |
d833352a | 6748 | */ |
dff11abe | 6749 | if (shared_pmd) |
ac46d4f3 | 6750 | flush_hugetlb_tlb_range(vma, range.start, range.end); |
dff11abe MK |
6751 | else |
6752 | flush_hugetlb_tlb_range(vma, start, end); | |
0f10851e JG |
6753 | /* |
6754 | * No need to call mmu_notifier_invalidate_range() we are downgrading | |
6755 | * page table protection not changing it to point to a new page. | |
6756 | * | |
ee65728e | 6757 | * See Documentation/mm/mmu_notifier.rst |
0f10851e | 6758 | */ |
83cde9e8 | 6759 | i_mmap_unlock_write(vma->vm_file->f_mapping); |
40549ba8 | 6760 | hugetlb_vma_unlock_write(vma); |
ac46d4f3 | 6761 | mmu_notifier_invalidate_range_end(&range); |
7da4d641 PZ |
6762 | |
6763 | return pages << h->order; | |
8f860591 ZY |
6764 | } |
6765 | ||
33b8f84a MK |
6766 | /* Return true if reservation was successful, false otherwise. */ |
6767 | bool hugetlb_reserve_pages(struct inode *inode, | |
a1e78772 | 6768 | long from, long to, |
5a6fe125 | 6769 | struct vm_area_struct *vma, |
ca16d140 | 6770 | vm_flags_t vm_flags) |
e4e574b7 | 6771 | { |
33b8f84a | 6772 | long chg, add = -1; |
a5516438 | 6773 | struct hstate *h = hstate_inode(inode); |
90481622 | 6774 | struct hugepage_subpool *spool = subpool_inode(inode); |
9119a41e | 6775 | struct resv_map *resv_map; |
075a61d0 | 6776 | struct hugetlb_cgroup *h_cg = NULL; |
0db9d74e | 6777 | long gbl_reserve, regions_needed = 0; |
e4e574b7 | 6778 | |
63489f8e MK |
6779 | /* This should never happen */ |
6780 | if (from > to) { | |
6781 | VM_WARN(1, "%s called with a negative range\n", __func__); | |
33b8f84a | 6782 | return false; |
63489f8e MK |
6783 | } |
6784 | ||
8d9bfb26 | 6785 | /* |
e700898f MK |
6786 | * vma specific semaphore used for pmd sharing and fault/truncation |
6787 | * synchronization | |
8d9bfb26 MK |
6788 | */ |
6789 | hugetlb_vma_lock_alloc(vma); | |
6790 | ||
17c9d12e MG |
6791 | /* |
6792 | * Only apply hugepage reservation if asked. At fault time, an | |
6793 | * attempt will be made for VM_NORESERVE to allocate a page | |
90481622 | 6794 | * without using reserves |
17c9d12e | 6795 | */ |
ca16d140 | 6796 | if (vm_flags & VM_NORESERVE) |
33b8f84a | 6797 | return true; |
17c9d12e | 6798 | |
a1e78772 MG |
6799 | /* |
6800 | * Shared mappings base their reservation on the number of pages that | |
6801 | * are already allocated on behalf of the file. Private mappings need | |
6802 | * to reserve the full area even if read-only as mprotect() may be | |
6803 | * called to make the mapping read-write. Assume !vma is a shm mapping | |
6804 | */ | |
9119a41e | 6805 | if (!vma || vma->vm_flags & VM_MAYSHARE) { |
f27a5136 MK |
6806 | /* |
6807 | * resv_map can not be NULL as hugetlb_reserve_pages is only | |
6808 | * called for inodes for which resv_maps were created (see | |
6809 | * hugetlbfs_get_inode). | |
6810 | */ | |
4e35f483 | 6811 | resv_map = inode_resv_map(inode); |
9119a41e | 6812 | |
0db9d74e | 6813 | chg = region_chg(resv_map, from, to, ®ions_needed); |
9119a41e | 6814 | } else { |
e9fe92ae | 6815 | /* Private mapping. */ |
9119a41e | 6816 | resv_map = resv_map_alloc(); |
17c9d12e | 6817 | if (!resv_map) |
8d9bfb26 | 6818 | goto out_err; |
17c9d12e | 6819 | |
a1e78772 | 6820 | chg = to - from; |
84afd99b | 6821 | |
17c9d12e MG |
6822 | set_vma_resv_map(vma, resv_map); |
6823 | set_vma_resv_flags(vma, HPAGE_RESV_OWNER); | |
6824 | } | |
6825 | ||
33b8f84a | 6826 | if (chg < 0) |
c50ac050 | 6827 | goto out_err; |
8a630112 | 6828 | |
33b8f84a MK |
6829 | if (hugetlb_cgroup_charge_cgroup_rsvd(hstate_index(h), |
6830 | chg * pages_per_huge_page(h), &h_cg) < 0) | |
075a61d0 | 6831 | goto out_err; |
075a61d0 MA |
6832 | |
6833 | if (vma && !(vma->vm_flags & VM_MAYSHARE) && h_cg) { | |
6834 | /* For private mappings, the hugetlb_cgroup uncharge info hangs | |
6835 | * of the resv_map. | |
6836 | */ | |
6837 | resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, h_cg, h); | |
6838 | } | |
6839 | ||
1c5ecae3 MK |
6840 | /* |
6841 | * There must be enough pages in the subpool for the mapping. If | |
6842 | * the subpool has a minimum size, there may be some global | |
6843 | * reservations already in place (gbl_reserve). | |
6844 | */ | |
6845 | gbl_reserve = hugepage_subpool_get_pages(spool, chg); | |
33b8f84a | 6846 | if (gbl_reserve < 0) |
075a61d0 | 6847 | goto out_uncharge_cgroup; |
5a6fe125 MG |
6848 | |
6849 | /* | |
17c9d12e | 6850 | * Check enough hugepages are available for the reservation. |
90481622 | 6851 | * Hand the pages back to the subpool if there are not |
5a6fe125 | 6852 | */ |
33b8f84a | 6853 | if (hugetlb_acct_memory(h, gbl_reserve) < 0) |
075a61d0 | 6854 | goto out_put_pages; |
17c9d12e MG |
6855 | |
6856 | /* | |
6857 | * Account for the reservations made. Shared mappings record regions | |
6858 | * that have reservations as they are shared by multiple VMAs. | |
6859 | * When the last VMA disappears, the region map says how much | |
6860 | * the reservation was and the page cache tells how much of | |
6861 | * the reservation was consumed. Private mappings are per-VMA and | |
6862 | * only the consumed reservations are tracked. When the VMA | |
6863 | * disappears, the original reservation is the VMA size and the | |
6864 | * consumed reservations are stored in the map. Hence, nothing | |
6865 | * else has to be done for private mappings here | |
6866 | */ | |
33039678 | 6867 | if (!vma || vma->vm_flags & VM_MAYSHARE) { |
075a61d0 | 6868 | add = region_add(resv_map, from, to, regions_needed, h, h_cg); |
0db9d74e MA |
6869 | |
6870 | if (unlikely(add < 0)) { | |
6871 | hugetlb_acct_memory(h, -gbl_reserve); | |
075a61d0 | 6872 | goto out_put_pages; |
0db9d74e | 6873 | } else if (unlikely(chg > add)) { |
33039678 MK |
6874 | /* |
6875 | * pages in this range were added to the reserve | |
6876 | * map between region_chg and region_add. This | |
6877 | * indicates a race with alloc_huge_page. Adjust | |
6878 | * the subpool and reserve counts modified above | |
6879 | * based on the difference. | |
6880 | */ | |
6881 | long rsv_adjust; | |
6882 | ||
d85aecf2 ML |
6883 | /* |
6884 | * hugetlb_cgroup_uncharge_cgroup_rsvd() will put the | |
6885 | * reference to h_cg->css. See comment below for detail. | |
6886 | */ | |
075a61d0 MA |
6887 | hugetlb_cgroup_uncharge_cgroup_rsvd( |
6888 | hstate_index(h), | |
6889 | (chg - add) * pages_per_huge_page(h), h_cg); | |
6890 | ||
33039678 MK |
6891 | rsv_adjust = hugepage_subpool_put_pages(spool, |
6892 | chg - add); | |
6893 | hugetlb_acct_memory(h, -rsv_adjust); | |
d85aecf2 ML |
6894 | } else if (h_cg) { |
6895 | /* | |
6896 | * The file_regions will hold their own reference to | |
6897 | * h_cg->css. So we should release the reference held | |
6898 | * via hugetlb_cgroup_charge_cgroup_rsvd() when we are | |
6899 | * done. | |
6900 | */ | |
6901 | hugetlb_cgroup_put_rsvd_cgroup(h_cg); | |
33039678 MK |
6902 | } |
6903 | } | |
33b8f84a MK |
6904 | return true; |
6905 | ||
075a61d0 MA |
6906 | out_put_pages: |
6907 | /* put back original number of pages, chg */ | |
6908 | (void)hugepage_subpool_put_pages(spool, chg); | |
6909 | out_uncharge_cgroup: | |
6910 | hugetlb_cgroup_uncharge_cgroup_rsvd(hstate_index(h), | |
6911 | chg * pages_per_huge_page(h), h_cg); | |
c50ac050 | 6912 | out_err: |
8d9bfb26 | 6913 | hugetlb_vma_lock_free(vma); |
5e911373 | 6914 | if (!vma || vma->vm_flags & VM_MAYSHARE) |
0db9d74e MA |
6915 | /* Only call region_abort if the region_chg succeeded but the |
6916 | * region_add failed or didn't run. | |
6917 | */ | |
6918 | if (chg >= 0 && add < 0) | |
6919 | region_abort(resv_map, from, to, regions_needed); | |
f031dd27 JK |
6920 | if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) |
6921 | kref_put(&resv_map->refs, resv_map_release); | |
33b8f84a | 6922 | return false; |
a43a8c39 CK |
6923 | } |
6924 | ||
b5cec28d MK |
6925 | long hugetlb_unreserve_pages(struct inode *inode, long start, long end, |
6926 | long freed) | |
a43a8c39 | 6927 | { |
a5516438 | 6928 | struct hstate *h = hstate_inode(inode); |
4e35f483 | 6929 | struct resv_map *resv_map = inode_resv_map(inode); |
9119a41e | 6930 | long chg = 0; |
90481622 | 6931 | struct hugepage_subpool *spool = subpool_inode(inode); |
1c5ecae3 | 6932 | long gbl_reserve; |
45c682a6 | 6933 | |
f27a5136 MK |
6934 | /* |
6935 | * Since this routine can be called in the evict inode path for all | |
6936 | * hugetlbfs inodes, resv_map could be NULL. | |
6937 | */ | |
b5cec28d MK |
6938 | if (resv_map) { |
6939 | chg = region_del(resv_map, start, end); | |
6940 | /* | |
6941 | * region_del() can fail in the rare case where a region | |
6942 | * must be split and another region descriptor can not be | |
6943 | * allocated. If end == LONG_MAX, it will not fail. | |
6944 | */ | |
6945 | if (chg < 0) | |
6946 | return chg; | |
6947 | } | |
6948 | ||
45c682a6 | 6949 | spin_lock(&inode->i_lock); |
e4c6f8be | 6950 | inode->i_blocks -= (blocks_per_huge_page(h) * freed); |
45c682a6 KC |
6951 | spin_unlock(&inode->i_lock); |
6952 | ||
1c5ecae3 MK |
6953 | /* |
6954 | * If the subpool has a minimum size, the number of global | |
6955 | * reservations to be released may be adjusted. | |
dddf31a4 ML |
6956 | * |
6957 | * Note that !resv_map implies freed == 0. So (chg - freed) | |
6958 | * won't go negative. | |
1c5ecae3 MK |
6959 | */ |
6960 | gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed)); | |
6961 | hugetlb_acct_memory(h, -gbl_reserve); | |
b5cec28d MK |
6962 | |
6963 | return 0; | |
a43a8c39 | 6964 | } |
93f70f90 | 6965 | |
3212b535 SC |
6966 | #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE |
6967 | static unsigned long page_table_shareable(struct vm_area_struct *svma, | |
6968 | struct vm_area_struct *vma, | |
6969 | unsigned long addr, pgoff_t idx) | |
6970 | { | |
6971 | unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) + | |
6972 | svma->vm_start; | |
6973 | unsigned long sbase = saddr & PUD_MASK; | |
6974 | unsigned long s_end = sbase + PUD_SIZE; | |
6975 | ||
6976 | /* Allow segments to share if only one is marked locked */ | |
de60f5f1 EM |
6977 | unsigned long vm_flags = vma->vm_flags & VM_LOCKED_CLEAR_MASK; |
6978 | unsigned long svm_flags = svma->vm_flags & VM_LOCKED_CLEAR_MASK; | |
3212b535 SC |
6979 | |
6980 | /* | |
6981 | * match the virtual addresses, permission and the alignment of the | |
6982 | * page table page. | |
131a79b4 MK |
6983 | * |
6984 | * Also, vma_lock (vm_private_data) is required for sharing. | |
3212b535 SC |
6985 | */ |
6986 | if (pmd_index(addr) != pmd_index(saddr) || | |
6987 | vm_flags != svm_flags || | |
131a79b4 MK |
6988 | !range_in_vma(svma, sbase, s_end) || |
6989 | !svma->vm_private_data) | |
3212b535 SC |
6990 | return 0; |
6991 | ||
6992 | return saddr; | |
6993 | } | |
6994 | ||
bbff39cc | 6995 | bool want_pmd_share(struct vm_area_struct *vma, unsigned long addr) |
3212b535 | 6996 | { |
bbff39cc MK |
6997 | unsigned long start = addr & PUD_MASK; |
6998 | unsigned long end = start + PUD_SIZE; | |
6999 | ||
8d9bfb26 MK |
7000 | #ifdef CONFIG_USERFAULTFD |
7001 | if (uffd_disable_huge_pmd_share(vma)) | |
7002 | return false; | |
7003 | #endif | |
3212b535 SC |
7004 | /* |
7005 | * check on proper vm_flags and page table alignment | |
7006 | */ | |
8d9bfb26 MK |
7007 | if (!(vma->vm_flags & VM_MAYSHARE)) |
7008 | return false; | |
bbff39cc | 7009 | if (!vma->vm_private_data) /* vma lock required for sharing */ |
8d9bfb26 MK |
7010 | return false; |
7011 | if (!range_in_vma(vma, start, end)) | |
7012 | return false; | |
7013 | return true; | |
7014 | } | |
7015 | ||
017b1660 MK |
7016 | /* |
7017 | * Determine if start,end range within vma could be mapped by shared pmd. | |
7018 | * If yes, adjust start and end to cover range associated with possible | |
7019 | * shared pmd mappings. | |
7020 | */ | |
7021 | void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma, | |
7022 | unsigned long *start, unsigned long *end) | |
7023 | { | |
a1ba9da8 LX |
7024 | unsigned long v_start = ALIGN(vma->vm_start, PUD_SIZE), |
7025 | v_end = ALIGN_DOWN(vma->vm_end, PUD_SIZE); | |
017b1660 | 7026 | |
a1ba9da8 | 7027 | /* |
f0953a1b IM |
7028 | * vma needs to span at least one aligned PUD size, and the range |
7029 | * must be at least partially within in. | |
a1ba9da8 LX |
7030 | */ |
7031 | if (!(vma->vm_flags & VM_MAYSHARE) || !(v_end > v_start) || | |
7032 | (*end <= v_start) || (*start >= v_end)) | |
017b1660 MK |
7033 | return; |
7034 | ||
75802ca6 | 7035 | /* Extend the range to be PUD aligned for a worst case scenario */ |
a1ba9da8 LX |
7036 | if (*start > v_start) |
7037 | *start = ALIGN_DOWN(*start, PUD_SIZE); | |
017b1660 | 7038 | |
a1ba9da8 LX |
7039 | if (*end < v_end) |
7040 | *end = ALIGN(*end, PUD_SIZE); | |
017b1660 MK |
7041 | } |
7042 | ||
3212b535 SC |
7043 | /* |
7044 | * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc() | |
7045 | * and returns the corresponding pte. While this is not necessary for the | |
7046 | * !shared pmd case because we can allocate the pmd later as well, it makes the | |
3a47c54f MK |
7047 | * code much cleaner. pmd allocation is essential for the shared case because |
7048 | * pud has to be populated inside the same i_mmap_rwsem section - otherwise | |
7049 | * racing tasks could either miss the sharing (see huge_pte_offset) or select a | |
7050 | * bad pmd for sharing. | |
3212b535 | 7051 | */ |
aec44e0f PX |
7052 | pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct *vma, |
7053 | unsigned long addr, pud_t *pud) | |
3212b535 | 7054 | { |
3212b535 SC |
7055 | struct address_space *mapping = vma->vm_file->f_mapping; |
7056 | pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) + | |
7057 | vma->vm_pgoff; | |
7058 | struct vm_area_struct *svma; | |
7059 | unsigned long saddr; | |
7060 | pte_t *spte = NULL; | |
7061 | pte_t *pte; | |
cb900f41 | 7062 | spinlock_t *ptl; |
3212b535 | 7063 | |
3a47c54f | 7064 | i_mmap_lock_read(mapping); |
3212b535 SC |
7065 | vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) { |
7066 | if (svma == vma) | |
7067 | continue; | |
7068 | ||
7069 | saddr = page_table_shareable(svma, vma, addr, idx); | |
7070 | if (saddr) { | |
7868a208 PA |
7071 | spte = huge_pte_offset(svma->vm_mm, saddr, |
7072 | vma_mmu_pagesize(svma)); | |
3212b535 SC |
7073 | if (spte) { |
7074 | get_page(virt_to_page(spte)); | |
7075 | break; | |
7076 | } | |
7077 | } | |
7078 | } | |
7079 | ||
7080 | if (!spte) | |
7081 | goto out; | |
7082 | ||
8bea8052 | 7083 | ptl = huge_pte_lock(hstate_vma(vma), mm, spte); |
dc6c9a35 | 7084 | if (pud_none(*pud)) { |
3212b535 SC |
7085 | pud_populate(mm, pud, |
7086 | (pmd_t *)((unsigned long)spte & PAGE_MASK)); | |
c17b1f42 | 7087 | mm_inc_nr_pmds(mm); |
dc6c9a35 | 7088 | } else { |
3212b535 | 7089 | put_page(virt_to_page(spte)); |
dc6c9a35 | 7090 | } |
cb900f41 | 7091 | spin_unlock(ptl); |
3212b535 SC |
7092 | out: |
7093 | pte = (pte_t *)pmd_alloc(mm, pud, addr); | |
3a47c54f | 7094 | i_mmap_unlock_read(mapping); |
3212b535 SC |
7095 | return pte; |
7096 | } | |
7097 | ||
7098 | /* | |
7099 | * unmap huge page backed by shared pte. | |
7100 | * | |
7101 | * Hugetlb pte page is ref counted at the time of mapping. If pte is shared | |
7102 | * indicated by page_count > 1, unmap is achieved by clearing pud and | |
7103 | * decrementing the ref count. If count == 1, the pte page is not shared. | |
7104 | * | |
3a47c54f | 7105 | * Called with page table lock held. |
3212b535 SC |
7106 | * |
7107 | * returns: 1 successfully unmapped a shared pte page | |
7108 | * 0 the underlying pte page is not shared, or it is the last user | |
7109 | */ | |
34ae204f | 7110 | int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma, |
4ddb4d91 | 7111 | unsigned long addr, pte_t *ptep) |
3212b535 | 7112 | { |
4ddb4d91 MK |
7113 | pgd_t *pgd = pgd_offset(mm, addr); |
7114 | p4d_t *p4d = p4d_offset(pgd, addr); | |
7115 | pud_t *pud = pud_offset(p4d, addr); | |
3212b535 | 7116 | |
34ae204f | 7117 | i_mmap_assert_write_locked(vma->vm_file->f_mapping); |
40549ba8 | 7118 | hugetlb_vma_assert_locked(vma); |
3212b535 SC |
7119 | BUG_ON(page_count(virt_to_page(ptep)) == 0); |
7120 | if (page_count(virt_to_page(ptep)) == 1) | |
7121 | return 0; | |
7122 | ||
7123 | pud_clear(pud); | |
7124 | put_page(virt_to_page(ptep)); | |
dc6c9a35 | 7125 | mm_dec_nr_pmds(mm); |
3212b535 SC |
7126 | return 1; |
7127 | } | |
c1991e07 | 7128 | |
9e5fc74c | 7129 | #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ |
8d9bfb26 | 7130 | |
aec44e0f PX |
7131 | pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct *vma, |
7132 | unsigned long addr, pud_t *pud) | |
9e5fc74c SC |
7133 | { |
7134 | return NULL; | |
7135 | } | |
e81f2d22 | 7136 | |
34ae204f | 7137 | int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma, |
4ddb4d91 | 7138 | unsigned long addr, pte_t *ptep) |
e81f2d22 ZZ |
7139 | { |
7140 | return 0; | |
7141 | } | |
017b1660 MK |
7142 | |
7143 | void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma, | |
7144 | unsigned long *start, unsigned long *end) | |
7145 | { | |
7146 | } | |
c1991e07 PX |
7147 | |
7148 | bool want_pmd_share(struct vm_area_struct *vma, unsigned long addr) | |
7149 | { | |
7150 | return false; | |
7151 | } | |
3212b535 SC |
7152 | #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ |
7153 | ||
9e5fc74c | 7154 | #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB |
aec44e0f | 7155 | pte_t *huge_pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma, |
9e5fc74c SC |
7156 | unsigned long addr, unsigned long sz) |
7157 | { | |
7158 | pgd_t *pgd; | |
c2febafc | 7159 | p4d_t *p4d; |
9e5fc74c SC |
7160 | pud_t *pud; |
7161 | pte_t *pte = NULL; | |
7162 | ||
7163 | pgd = pgd_offset(mm, addr); | |
f4f0a3d8 KS |
7164 | p4d = p4d_alloc(mm, pgd, addr); |
7165 | if (!p4d) | |
7166 | return NULL; | |
c2febafc | 7167 | pud = pud_alloc(mm, p4d, addr); |
9e5fc74c SC |
7168 | if (pud) { |
7169 | if (sz == PUD_SIZE) { | |
7170 | pte = (pte_t *)pud; | |
7171 | } else { | |
7172 | BUG_ON(sz != PMD_SIZE); | |
c1991e07 | 7173 | if (want_pmd_share(vma, addr) && pud_none(*pud)) |
aec44e0f | 7174 | pte = huge_pmd_share(mm, vma, addr, pud); |
9e5fc74c SC |
7175 | else |
7176 | pte = (pte_t *)pmd_alloc(mm, pud, addr); | |
7177 | } | |
7178 | } | |
4e666314 | 7179 | BUG_ON(pte && pte_present(*pte) && !pte_huge(*pte)); |
9e5fc74c SC |
7180 | |
7181 | return pte; | |
7182 | } | |
7183 | ||
9b19df29 PA |
7184 | /* |
7185 | * huge_pte_offset() - Walk the page table to resolve the hugepage | |
7186 | * entry at address @addr | |
7187 | * | |
8ac0b81a LX |
7188 | * Return: Pointer to page table entry (PUD or PMD) for |
7189 | * address @addr, or NULL if a !p*d_present() entry is encountered and the | |
9b19df29 PA |
7190 | * size @sz doesn't match the hugepage size at this level of the page |
7191 | * table. | |
7192 | */ | |
7868a208 PA |
7193 | pte_t *huge_pte_offset(struct mm_struct *mm, |
7194 | unsigned long addr, unsigned long sz) | |
9e5fc74c SC |
7195 | { |
7196 | pgd_t *pgd; | |
c2febafc | 7197 | p4d_t *p4d; |
8ac0b81a LX |
7198 | pud_t *pud; |
7199 | pmd_t *pmd; | |
9e5fc74c SC |
7200 | |
7201 | pgd = pgd_offset(mm, addr); | |
c2febafc KS |
7202 | if (!pgd_present(*pgd)) |
7203 | return NULL; | |
7204 | p4d = p4d_offset(pgd, addr); | |
7205 | if (!p4d_present(*p4d)) | |
7206 | return NULL; | |
9b19df29 | 7207 | |
c2febafc | 7208 | pud = pud_offset(p4d, addr); |
8ac0b81a LX |
7209 | if (sz == PUD_SIZE) |
7210 | /* must be pud huge, non-present or none */ | |
c2febafc | 7211 | return (pte_t *)pud; |
8ac0b81a | 7212 | if (!pud_present(*pud)) |
9b19df29 | 7213 | return NULL; |
8ac0b81a | 7214 | /* must have a valid entry and size to go further */ |
9b19df29 | 7215 | |
8ac0b81a LX |
7216 | pmd = pmd_offset(pud, addr); |
7217 | /* must be pmd huge, non-present or none */ | |
7218 | return (pte_t *)pmd; | |
9e5fc74c SC |
7219 | } |
7220 | ||
e95a9851 MK |
7221 | /* |
7222 | * Return a mask that can be used to update an address to the last huge | |
7223 | * page in a page table page mapping size. Used to skip non-present | |
7224 | * page table entries when linearly scanning address ranges. Architectures | |
7225 | * with unique huge page to page table relationships can define their own | |
7226 | * version of this routine. | |
7227 | */ | |
7228 | unsigned long hugetlb_mask_last_page(struct hstate *h) | |
7229 | { | |
7230 | unsigned long hp_size = huge_page_size(h); | |
7231 | ||
7232 | if (hp_size == PUD_SIZE) | |
7233 | return P4D_SIZE - PUD_SIZE; | |
7234 | else if (hp_size == PMD_SIZE) | |
7235 | return PUD_SIZE - PMD_SIZE; | |
7236 | else | |
7237 | return 0UL; | |
7238 | } | |
7239 | ||
7240 | #else | |
7241 | ||
7242 | /* See description above. Architectures can provide their own version. */ | |
7243 | __weak unsigned long hugetlb_mask_last_page(struct hstate *h) | |
7244 | { | |
4ddb4d91 MK |
7245 | #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE |
7246 | if (huge_page_size(h) == PMD_SIZE) | |
7247 | return PUD_SIZE - PMD_SIZE; | |
7248 | #endif | |
e95a9851 MK |
7249 | return 0UL; |
7250 | } | |
7251 | ||
61f77eda NH |
7252 | #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */ |
7253 | ||
7254 | /* | |
7255 | * These functions are overwritable if your architecture needs its own | |
7256 | * behavior. | |
7257 | */ | |
7ce82f4c | 7258 | int isolate_hugetlb(struct page *page, struct list_head *list) |
31caf665 | 7259 | { |
7ce82f4c | 7260 | int ret = 0; |
bcc54222 | 7261 | |
db71ef79 | 7262 | spin_lock_irq(&hugetlb_lock); |
8f251a3d MK |
7263 | if (!PageHeadHuge(page) || |
7264 | !HPageMigratable(page) || | |
0eb2df2b | 7265 | !get_page_unless_zero(page)) { |
7ce82f4c | 7266 | ret = -EBUSY; |
bcc54222 NH |
7267 | goto unlock; |
7268 | } | |
8f251a3d | 7269 | ClearHPageMigratable(page); |
31caf665 | 7270 | list_move_tail(&page->lru, list); |
bcc54222 | 7271 | unlock: |
db71ef79 | 7272 | spin_unlock_irq(&hugetlb_lock); |
bcc54222 | 7273 | return ret; |
31caf665 NH |
7274 | } |
7275 | ||
e591ef7d | 7276 | int get_hwpoison_huge_page(struct page *page, bool *hugetlb, bool unpoison) |
25182f05 NH |
7277 | { |
7278 | int ret = 0; | |
7279 | ||
7280 | *hugetlb = false; | |
7281 | spin_lock_irq(&hugetlb_lock); | |
7282 | if (PageHeadHuge(page)) { | |
7283 | *hugetlb = true; | |
b283d983 NH |
7284 | if (HPageFreed(page)) |
7285 | ret = 0; | |
e591ef7d | 7286 | else if (HPageMigratable(page) || unpoison) |
25182f05 | 7287 | ret = get_page_unless_zero(page); |
0ed950d1 NH |
7288 | else |
7289 | ret = -EBUSY; | |
25182f05 NH |
7290 | } |
7291 | spin_unlock_irq(&hugetlb_lock); | |
7292 | return ret; | |
7293 | } | |
7294 | ||
e591ef7d NH |
7295 | int get_huge_page_for_hwpoison(unsigned long pfn, int flags, |
7296 | bool *migratable_cleared) | |
405ce051 NH |
7297 | { |
7298 | int ret; | |
7299 | ||
7300 | spin_lock_irq(&hugetlb_lock); | |
e591ef7d | 7301 | ret = __get_huge_page_for_hwpoison(pfn, flags, migratable_cleared); |
405ce051 NH |
7302 | spin_unlock_irq(&hugetlb_lock); |
7303 | return ret; | |
7304 | } | |
7305 | ||
31caf665 NH |
7306 | void putback_active_hugepage(struct page *page) |
7307 | { | |
db71ef79 | 7308 | spin_lock_irq(&hugetlb_lock); |
8f251a3d | 7309 | SetHPageMigratable(page); |
31caf665 | 7310 | list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist); |
db71ef79 | 7311 | spin_unlock_irq(&hugetlb_lock); |
31caf665 NH |
7312 | put_page(page); |
7313 | } | |
ab5ac90a | 7314 | |
345c62d1 | 7315 | void move_hugetlb_state(struct folio *old_folio, struct folio *new_folio, int reason) |
ab5ac90a | 7316 | { |
345c62d1 | 7317 | struct hstate *h = folio_hstate(old_folio); |
ab5ac90a | 7318 | |
345c62d1 SK |
7319 | hugetlb_cgroup_migrate(old_folio, new_folio); |
7320 | set_page_owner_migrate_reason(&new_folio->page, reason); | |
ab5ac90a MH |
7321 | |
7322 | /* | |
345c62d1 | 7323 | * transfer temporary state of the new hugetlb folio. This is |
ab5ac90a MH |
7324 | * reverse to other transitions because the newpage is going to |
7325 | * be final while the old one will be freed so it takes over | |
7326 | * the temporary status. | |
7327 | * | |
7328 | * Also note that we have to transfer the per-node surplus state | |
7329 | * here as well otherwise the global surplus count will not match | |
7330 | * the per-node's. | |
7331 | */ | |
345c62d1 SK |
7332 | if (folio_test_hugetlb_temporary(new_folio)) { |
7333 | int old_nid = folio_nid(old_folio); | |
7334 | int new_nid = folio_nid(new_folio); | |
7335 | ||
345c62d1 SK |
7336 | folio_set_hugetlb_temporary(old_folio); |
7337 | folio_clear_hugetlb_temporary(new_folio); | |
ab5ac90a | 7338 | |
ab5ac90a | 7339 | |
5af1ab1d ML |
7340 | /* |
7341 | * There is no need to transfer the per-node surplus state | |
7342 | * when we do not cross the node. | |
7343 | */ | |
7344 | if (new_nid == old_nid) | |
7345 | return; | |
db71ef79 | 7346 | spin_lock_irq(&hugetlb_lock); |
ab5ac90a MH |
7347 | if (h->surplus_huge_pages_node[old_nid]) { |
7348 | h->surplus_huge_pages_node[old_nid]--; | |
7349 | h->surplus_huge_pages_node[new_nid]++; | |
7350 | } | |
db71ef79 | 7351 | spin_unlock_irq(&hugetlb_lock); |
ab5ac90a MH |
7352 | } |
7353 | } | |
cf11e85f | 7354 | |
b30c14cd JH |
7355 | static void hugetlb_unshare_pmds(struct vm_area_struct *vma, |
7356 | unsigned long start, | |
7357 | unsigned long end) | |
6dfeaff9 PX |
7358 | { |
7359 | struct hstate *h = hstate_vma(vma); | |
7360 | unsigned long sz = huge_page_size(h); | |
7361 | struct mm_struct *mm = vma->vm_mm; | |
7362 | struct mmu_notifier_range range; | |
b30c14cd | 7363 | unsigned long address; |
6dfeaff9 PX |
7364 | spinlock_t *ptl; |
7365 | pte_t *ptep; | |
7366 | ||
7367 | if (!(vma->vm_flags & VM_MAYSHARE)) | |
7368 | return; | |
7369 | ||
6dfeaff9 PX |
7370 | if (start >= end) |
7371 | return; | |
7372 | ||
9c8bbfac | 7373 | flush_cache_range(vma, start, end); |
6dfeaff9 PX |
7374 | /* |
7375 | * No need to call adjust_range_if_pmd_sharing_possible(), because | |
7376 | * we have already done the PUD_SIZE alignment. | |
7377 | */ | |
7378 | mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, | |
7379 | start, end); | |
7380 | mmu_notifier_invalidate_range_start(&range); | |
40549ba8 | 7381 | hugetlb_vma_lock_write(vma); |
6dfeaff9 PX |
7382 | i_mmap_lock_write(vma->vm_file->f_mapping); |
7383 | for (address = start; address < end; address += PUD_SIZE) { | |
6dfeaff9 PX |
7384 | ptep = huge_pte_offset(mm, address, sz); |
7385 | if (!ptep) | |
7386 | continue; | |
7387 | ptl = huge_pte_lock(h, mm, ptep); | |
4ddb4d91 | 7388 | huge_pmd_unshare(mm, vma, address, ptep); |
6dfeaff9 PX |
7389 | spin_unlock(ptl); |
7390 | } | |
7391 | flush_hugetlb_tlb_range(vma, start, end); | |
7392 | i_mmap_unlock_write(vma->vm_file->f_mapping); | |
40549ba8 | 7393 | hugetlb_vma_unlock_write(vma); |
6dfeaff9 PX |
7394 | /* |
7395 | * No need to call mmu_notifier_invalidate_range(), see | |
ee65728e | 7396 | * Documentation/mm/mmu_notifier.rst. |
6dfeaff9 PX |
7397 | */ |
7398 | mmu_notifier_invalidate_range_end(&range); | |
7399 | } | |
7400 | ||
b30c14cd JH |
7401 | /* |
7402 | * This function will unconditionally remove all the shared pmd pgtable entries | |
7403 | * within the specific vma for a hugetlbfs memory range. | |
7404 | */ | |
7405 | void hugetlb_unshare_all_pmds(struct vm_area_struct *vma) | |
7406 | { | |
7407 | hugetlb_unshare_pmds(vma, ALIGN(vma->vm_start, PUD_SIZE), | |
7408 | ALIGN_DOWN(vma->vm_end, PUD_SIZE)); | |
7409 | } | |
7410 | ||
cf11e85f | 7411 | #ifdef CONFIG_CMA |
cf11e85f RG |
7412 | static bool cma_reserve_called __initdata; |
7413 | ||
7414 | static int __init cmdline_parse_hugetlb_cma(char *p) | |
7415 | { | |
38e719ab BW |
7416 | int nid, count = 0; |
7417 | unsigned long tmp; | |
7418 | char *s = p; | |
7419 | ||
7420 | while (*s) { | |
7421 | if (sscanf(s, "%lu%n", &tmp, &count) != 1) | |
7422 | break; | |
7423 | ||
7424 | if (s[count] == ':') { | |
f9317f77 | 7425 | if (tmp >= MAX_NUMNODES) |
38e719ab | 7426 | break; |
f9317f77 | 7427 | nid = array_index_nospec(tmp, MAX_NUMNODES); |
38e719ab BW |
7428 | |
7429 | s += count + 1; | |
7430 | tmp = memparse(s, &s); | |
7431 | hugetlb_cma_size_in_node[nid] = tmp; | |
7432 | hugetlb_cma_size += tmp; | |
7433 | ||
7434 | /* | |
7435 | * Skip the separator if have one, otherwise | |
7436 | * break the parsing. | |
7437 | */ | |
7438 | if (*s == ',') | |
7439 | s++; | |
7440 | else | |
7441 | break; | |
7442 | } else { | |
7443 | hugetlb_cma_size = memparse(p, &p); | |
7444 | break; | |
7445 | } | |
7446 | } | |
7447 | ||
cf11e85f RG |
7448 | return 0; |
7449 | } | |
7450 | ||
7451 | early_param("hugetlb_cma", cmdline_parse_hugetlb_cma); | |
7452 | ||
7453 | void __init hugetlb_cma_reserve(int order) | |
7454 | { | |
7455 | unsigned long size, reserved, per_node; | |
38e719ab | 7456 | bool node_specific_cma_alloc = false; |
cf11e85f RG |
7457 | int nid; |
7458 | ||
7459 | cma_reserve_called = true; | |
7460 | ||
38e719ab BW |
7461 | if (!hugetlb_cma_size) |
7462 | return; | |
7463 | ||
7464 | for (nid = 0; nid < MAX_NUMNODES; nid++) { | |
7465 | if (hugetlb_cma_size_in_node[nid] == 0) | |
7466 | continue; | |
7467 | ||
30a51400 | 7468 | if (!node_online(nid)) { |
38e719ab BW |
7469 | pr_warn("hugetlb_cma: invalid node %d specified\n", nid); |
7470 | hugetlb_cma_size -= hugetlb_cma_size_in_node[nid]; | |
7471 | hugetlb_cma_size_in_node[nid] = 0; | |
7472 | continue; | |
7473 | } | |
7474 | ||
7475 | if (hugetlb_cma_size_in_node[nid] < (PAGE_SIZE << order)) { | |
7476 | pr_warn("hugetlb_cma: cma area of node %d should be at least %lu MiB\n", | |
7477 | nid, (PAGE_SIZE << order) / SZ_1M); | |
7478 | hugetlb_cma_size -= hugetlb_cma_size_in_node[nid]; | |
7479 | hugetlb_cma_size_in_node[nid] = 0; | |
7480 | } else { | |
7481 | node_specific_cma_alloc = true; | |
7482 | } | |
7483 | } | |
7484 | ||
7485 | /* Validate the CMA size again in case some invalid nodes specified. */ | |
cf11e85f RG |
7486 | if (!hugetlb_cma_size) |
7487 | return; | |
7488 | ||
7489 | if (hugetlb_cma_size < (PAGE_SIZE << order)) { | |
7490 | pr_warn("hugetlb_cma: cma area should be at least %lu MiB\n", | |
7491 | (PAGE_SIZE << order) / SZ_1M); | |
a01f4390 | 7492 | hugetlb_cma_size = 0; |
cf11e85f RG |
7493 | return; |
7494 | } | |
7495 | ||
38e719ab BW |
7496 | if (!node_specific_cma_alloc) { |
7497 | /* | |
7498 | * If 3 GB area is requested on a machine with 4 numa nodes, | |
7499 | * let's allocate 1 GB on first three nodes and ignore the last one. | |
7500 | */ | |
7501 | per_node = DIV_ROUND_UP(hugetlb_cma_size, nr_online_nodes); | |
7502 | pr_info("hugetlb_cma: reserve %lu MiB, up to %lu MiB per node\n", | |
7503 | hugetlb_cma_size / SZ_1M, per_node / SZ_1M); | |
7504 | } | |
cf11e85f RG |
7505 | |
7506 | reserved = 0; | |
30a51400 | 7507 | for_each_online_node(nid) { |
cf11e85f | 7508 | int res; |
2281f797 | 7509 | char name[CMA_MAX_NAME]; |
cf11e85f | 7510 | |
38e719ab BW |
7511 | if (node_specific_cma_alloc) { |
7512 | if (hugetlb_cma_size_in_node[nid] == 0) | |
7513 | continue; | |
7514 | ||
7515 | size = hugetlb_cma_size_in_node[nid]; | |
7516 | } else { | |
7517 | size = min(per_node, hugetlb_cma_size - reserved); | |
7518 | } | |
7519 | ||
cf11e85f RG |
7520 | size = round_up(size, PAGE_SIZE << order); |
7521 | ||
2281f797 | 7522 | snprintf(name, sizeof(name), "hugetlb%d", nid); |
a01f4390 MK |
7523 | /* |
7524 | * Note that 'order per bit' is based on smallest size that | |
7525 | * may be returned to CMA allocator in the case of | |
7526 | * huge page demotion. | |
7527 | */ | |
7528 | res = cma_declare_contiguous_nid(0, size, 0, | |
7529 | PAGE_SIZE << HUGETLB_PAGE_ORDER, | |
29d0f41d | 7530 | 0, false, name, |
cf11e85f RG |
7531 | &hugetlb_cma[nid], nid); |
7532 | if (res) { | |
7533 | pr_warn("hugetlb_cma: reservation failed: err %d, node %d", | |
7534 | res, nid); | |
7535 | continue; | |
7536 | } | |
7537 | ||
7538 | reserved += size; | |
7539 | pr_info("hugetlb_cma: reserved %lu MiB on node %d\n", | |
7540 | size / SZ_1M, nid); | |
7541 | ||
7542 | if (reserved >= hugetlb_cma_size) | |
7543 | break; | |
7544 | } | |
a01f4390 MK |
7545 | |
7546 | if (!reserved) | |
7547 | /* | |
7548 | * hugetlb_cma_size is used to determine if allocations from | |
7549 | * cma are possible. Set to zero if no cma regions are set up. | |
7550 | */ | |
7551 | hugetlb_cma_size = 0; | |
cf11e85f RG |
7552 | } |
7553 | ||
263b8998 | 7554 | static void __init hugetlb_cma_check(void) |
cf11e85f RG |
7555 | { |
7556 | if (!hugetlb_cma_size || cma_reserve_called) | |
7557 | return; | |
7558 | ||
7559 | pr_warn("hugetlb_cma: the option isn't supported by current arch\n"); | |
7560 | } | |
7561 | ||
7562 | #endif /* CONFIG_CMA */ |