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