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; |
a01f4390 MK |
57 | static bool hugetlb_cma_page(struct page *page, unsigned int order) |
58 | { | |
59 | return cma_pages_valid(hugetlb_cma[page_to_nid(page)], page, | |
60 | 1 << order); | |
61 | } | |
62 | #else | |
63 | static bool hugetlb_cma_page(struct page *page, unsigned int order) | |
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 | ||
0db9d74e MA |
258 | /* Helper that removes a struct file_region from the resv_map cache and returns |
259 | * it for use. | |
260 | */ | |
261 | static struct file_region * | |
262 | get_file_region_entry_from_cache(struct resv_map *resv, long from, long to) | |
263 | { | |
3259914f | 264 | struct file_region *nrg; |
0db9d74e MA |
265 | |
266 | VM_BUG_ON(resv->region_cache_count <= 0); | |
267 | ||
268 | resv->region_cache_count--; | |
269 | nrg = list_first_entry(&resv->region_cache, struct file_region, link); | |
0db9d74e MA |
270 | list_del(&nrg->link); |
271 | ||
272 | nrg->from = from; | |
273 | nrg->to = to; | |
274 | ||
275 | return nrg; | |
276 | } | |
277 | ||
075a61d0 MA |
278 | static void copy_hugetlb_cgroup_uncharge_info(struct file_region *nrg, |
279 | struct file_region *rg) | |
280 | { | |
281 | #ifdef CONFIG_CGROUP_HUGETLB | |
282 | nrg->reservation_counter = rg->reservation_counter; | |
283 | nrg->css = rg->css; | |
284 | if (rg->css) | |
285 | css_get(rg->css); | |
286 | #endif | |
287 | } | |
288 | ||
289 | /* Helper that records hugetlb_cgroup uncharge info. */ | |
290 | static void record_hugetlb_cgroup_uncharge_info(struct hugetlb_cgroup *h_cg, | |
291 | struct hstate *h, | |
292 | struct resv_map *resv, | |
293 | struct file_region *nrg) | |
294 | { | |
295 | #ifdef CONFIG_CGROUP_HUGETLB | |
296 | if (h_cg) { | |
297 | nrg->reservation_counter = | |
298 | &h_cg->rsvd_hugepage[hstate_index(h)]; | |
299 | nrg->css = &h_cg->css; | |
d85aecf2 ML |
300 | /* |
301 | * The caller will hold exactly one h_cg->css reference for the | |
302 | * whole contiguous reservation region. But this area might be | |
303 | * scattered when there are already some file_regions reside in | |
304 | * it. As a result, many file_regions may share only one css | |
305 | * reference. In order to ensure that one file_region must hold | |
306 | * exactly one h_cg->css reference, we should do css_get for | |
307 | * each file_region and leave the reference held by caller | |
308 | * untouched. | |
309 | */ | |
310 | css_get(&h_cg->css); | |
075a61d0 MA |
311 | if (!resv->pages_per_hpage) |
312 | resv->pages_per_hpage = pages_per_huge_page(h); | |
313 | /* pages_per_hpage should be the same for all entries in | |
314 | * a resv_map. | |
315 | */ | |
316 | VM_BUG_ON(resv->pages_per_hpage != pages_per_huge_page(h)); | |
317 | } else { | |
318 | nrg->reservation_counter = NULL; | |
319 | nrg->css = NULL; | |
320 | } | |
321 | #endif | |
322 | } | |
323 | ||
d85aecf2 ML |
324 | static void put_uncharge_info(struct file_region *rg) |
325 | { | |
326 | #ifdef CONFIG_CGROUP_HUGETLB | |
327 | if (rg->css) | |
328 | css_put(rg->css); | |
329 | #endif | |
330 | } | |
331 | ||
a9b3f867 MA |
332 | static bool has_same_uncharge_info(struct file_region *rg, |
333 | struct file_region *org) | |
334 | { | |
335 | #ifdef CONFIG_CGROUP_HUGETLB | |
0739eb43 | 336 | return rg->reservation_counter == org->reservation_counter && |
a9b3f867 MA |
337 | rg->css == org->css; |
338 | ||
339 | #else | |
340 | return true; | |
341 | #endif | |
342 | } | |
343 | ||
344 | static void coalesce_file_region(struct resv_map *resv, struct file_region *rg) | |
345 | { | |
3259914f | 346 | struct file_region *nrg, *prg; |
a9b3f867 MA |
347 | |
348 | prg = list_prev_entry(rg, link); | |
349 | if (&prg->link != &resv->regions && prg->to == rg->from && | |
350 | has_same_uncharge_info(prg, rg)) { | |
351 | prg->to = rg->to; | |
352 | ||
353 | list_del(&rg->link); | |
d85aecf2 | 354 | put_uncharge_info(rg); |
a9b3f867 MA |
355 | kfree(rg); |
356 | ||
7db5e7b6 | 357 | rg = prg; |
a9b3f867 MA |
358 | } |
359 | ||
360 | nrg = list_next_entry(rg, link); | |
361 | if (&nrg->link != &resv->regions && nrg->from == rg->to && | |
362 | has_same_uncharge_info(nrg, rg)) { | |
363 | nrg->from = rg->from; | |
364 | ||
365 | list_del(&rg->link); | |
d85aecf2 | 366 | put_uncharge_info(rg); |
a9b3f867 | 367 | kfree(rg); |
a9b3f867 MA |
368 | } |
369 | } | |
370 | ||
2103cf9c | 371 | static inline long |
84448c8e | 372 | hugetlb_resv_map_add(struct resv_map *map, struct list_head *rg, long from, |
2103cf9c PX |
373 | long to, struct hstate *h, struct hugetlb_cgroup *cg, |
374 | long *regions_needed) | |
375 | { | |
376 | struct file_region *nrg; | |
377 | ||
378 | if (!regions_needed) { | |
379 | nrg = get_file_region_entry_from_cache(map, from, to); | |
380 | record_hugetlb_cgroup_uncharge_info(cg, h, map, nrg); | |
84448c8e | 381 | list_add(&nrg->link, rg); |
2103cf9c PX |
382 | coalesce_file_region(map, nrg); |
383 | } else | |
384 | *regions_needed += 1; | |
385 | ||
386 | return to - from; | |
387 | } | |
388 | ||
972a3da3 WY |
389 | /* |
390 | * Must be called with resv->lock held. | |
391 | * | |
392 | * Calling this with regions_needed != NULL will count the number of pages | |
393 | * to be added but will not modify the linked list. And regions_needed will | |
394 | * indicate the number of file_regions needed in the cache to carry out to add | |
395 | * the regions for this range. | |
d75c6af9 MA |
396 | */ |
397 | static long add_reservation_in_range(struct resv_map *resv, long f, long t, | |
075a61d0 | 398 | struct hugetlb_cgroup *h_cg, |
972a3da3 | 399 | struct hstate *h, long *regions_needed) |
d75c6af9 | 400 | { |
0db9d74e | 401 | long add = 0; |
d75c6af9 | 402 | struct list_head *head = &resv->regions; |
0db9d74e | 403 | long last_accounted_offset = f; |
84448c8e JK |
404 | struct file_region *iter, *trg = NULL; |
405 | struct list_head *rg = NULL; | |
d75c6af9 | 406 | |
0db9d74e MA |
407 | if (regions_needed) |
408 | *regions_needed = 0; | |
d75c6af9 | 409 | |
0db9d74e | 410 | /* In this loop, we essentially handle an entry for the range |
84448c8e | 411 | * [last_accounted_offset, iter->from), at every iteration, with some |
0db9d74e MA |
412 | * bounds checking. |
413 | */ | |
84448c8e | 414 | list_for_each_entry_safe(iter, trg, head, link) { |
0db9d74e | 415 | /* Skip irrelevant regions that start before our range. */ |
84448c8e | 416 | if (iter->from < f) { |
0db9d74e MA |
417 | /* If this region ends after the last accounted offset, |
418 | * then we need to update last_accounted_offset. | |
419 | */ | |
84448c8e JK |
420 | if (iter->to > last_accounted_offset) |
421 | last_accounted_offset = iter->to; | |
0db9d74e MA |
422 | continue; |
423 | } | |
d75c6af9 | 424 | |
0db9d74e MA |
425 | /* When we find a region that starts beyond our range, we've |
426 | * finished. | |
427 | */ | |
84448c8e JK |
428 | if (iter->from >= t) { |
429 | rg = iter->link.prev; | |
d75c6af9 | 430 | break; |
84448c8e | 431 | } |
d75c6af9 | 432 | |
84448c8e | 433 | /* Add an entry for last_accounted_offset -> iter->from, and |
0db9d74e MA |
434 | * update last_accounted_offset. |
435 | */ | |
84448c8e JK |
436 | if (iter->from > last_accounted_offset) |
437 | add += hugetlb_resv_map_add(resv, iter->link.prev, | |
2103cf9c | 438 | last_accounted_offset, |
84448c8e | 439 | iter->from, h, h_cg, |
2103cf9c | 440 | regions_needed); |
0db9d74e | 441 | |
84448c8e | 442 | last_accounted_offset = iter->to; |
0db9d74e MA |
443 | } |
444 | ||
445 | /* Handle the case where our range extends beyond | |
446 | * last_accounted_offset. | |
447 | */ | |
84448c8e JK |
448 | if (!rg) |
449 | rg = head->prev; | |
2103cf9c PX |
450 | if (last_accounted_offset < t) |
451 | add += hugetlb_resv_map_add(resv, rg, last_accounted_offset, | |
452 | t, h, h_cg, regions_needed); | |
0db9d74e | 453 | |
0db9d74e MA |
454 | return add; |
455 | } | |
456 | ||
457 | /* Must be called with resv->lock acquired. Will drop lock to allocate entries. | |
458 | */ | |
459 | static int allocate_file_region_entries(struct resv_map *resv, | |
460 | int regions_needed) | |
461 | __must_hold(&resv->lock) | |
462 | { | |
34665341 | 463 | LIST_HEAD(allocated_regions); |
0db9d74e MA |
464 | int to_allocate = 0, i = 0; |
465 | struct file_region *trg = NULL, *rg = NULL; | |
466 | ||
467 | VM_BUG_ON(regions_needed < 0); | |
468 | ||
0db9d74e MA |
469 | /* |
470 | * Check for sufficient descriptors in the cache to accommodate | |
471 | * the number of in progress add operations plus regions_needed. | |
472 | * | |
473 | * This is a while loop because when we drop the lock, some other call | |
474 | * to region_add or region_del may have consumed some region_entries, | |
475 | * so we keep looping here until we finally have enough entries for | |
476 | * (adds_in_progress + regions_needed). | |
477 | */ | |
478 | while (resv->region_cache_count < | |
479 | (resv->adds_in_progress + regions_needed)) { | |
480 | to_allocate = resv->adds_in_progress + regions_needed - | |
481 | resv->region_cache_count; | |
482 | ||
483 | /* At this point, we should have enough entries in the cache | |
f0953a1b | 484 | * for all the existing adds_in_progress. We should only be |
0db9d74e | 485 | * needing to allocate for regions_needed. |
d75c6af9 | 486 | */ |
0db9d74e MA |
487 | VM_BUG_ON(resv->region_cache_count < resv->adds_in_progress); |
488 | ||
489 | spin_unlock(&resv->lock); | |
490 | for (i = 0; i < to_allocate; i++) { | |
491 | trg = kmalloc(sizeof(*trg), GFP_KERNEL); | |
492 | if (!trg) | |
493 | goto out_of_memory; | |
494 | list_add(&trg->link, &allocated_regions); | |
d75c6af9 | 495 | } |
d75c6af9 | 496 | |
0db9d74e MA |
497 | spin_lock(&resv->lock); |
498 | ||
d3ec7b6e WY |
499 | list_splice(&allocated_regions, &resv->region_cache); |
500 | resv->region_cache_count += to_allocate; | |
d75c6af9 MA |
501 | } |
502 | ||
0db9d74e | 503 | return 0; |
d75c6af9 | 504 | |
0db9d74e MA |
505 | out_of_memory: |
506 | list_for_each_entry_safe(rg, trg, &allocated_regions, link) { | |
507 | list_del(&rg->link); | |
508 | kfree(rg); | |
509 | } | |
510 | return -ENOMEM; | |
d75c6af9 MA |
511 | } |
512 | ||
1dd308a7 MK |
513 | /* |
514 | * Add the huge page range represented by [f, t) to the reserve | |
0db9d74e MA |
515 | * map. Regions will be taken from the cache to fill in this range. |
516 | * Sufficient regions should exist in the cache due to the previous | |
517 | * call to region_chg with the same range, but in some cases the cache will not | |
518 | * have sufficient entries due to races with other code doing region_add or | |
519 | * region_del. The extra needed entries will be allocated. | |
cf3ad20b | 520 | * |
0db9d74e MA |
521 | * regions_needed is the out value provided by a previous call to region_chg. |
522 | * | |
523 | * Return the number of new huge pages added to the map. This number is greater | |
524 | * than or equal to zero. If file_region entries needed to be allocated for | |
7c8de358 | 525 | * this operation and we were not able to allocate, it returns -ENOMEM. |
0db9d74e MA |
526 | * region_add of regions of length 1 never allocate file_regions and cannot |
527 | * fail; region_chg will always allocate at least 1 entry and a region_add for | |
528 | * 1 page will only require at most 1 entry. | |
1dd308a7 | 529 | */ |
0db9d74e | 530 | static long region_add(struct resv_map *resv, long f, long t, |
075a61d0 MA |
531 | long in_regions_needed, struct hstate *h, |
532 | struct hugetlb_cgroup *h_cg) | |
96822904 | 533 | { |
0db9d74e | 534 | long add = 0, actual_regions_needed = 0; |
96822904 | 535 | |
7b24d861 | 536 | spin_lock(&resv->lock); |
0db9d74e MA |
537 | retry: |
538 | ||
539 | /* Count how many regions are actually needed to execute this add. */ | |
972a3da3 WY |
540 | add_reservation_in_range(resv, f, t, NULL, NULL, |
541 | &actual_regions_needed); | |
96822904 | 542 | |
5e911373 | 543 | /* |
0db9d74e MA |
544 | * Check for sufficient descriptors in the cache to accommodate |
545 | * this add operation. Note that actual_regions_needed may be greater | |
546 | * than in_regions_needed, as the resv_map may have been modified since | |
547 | * the region_chg call. In this case, we need to make sure that we | |
548 | * allocate extra entries, such that we have enough for all the | |
549 | * existing adds_in_progress, plus the excess needed for this | |
550 | * operation. | |
5e911373 | 551 | */ |
0db9d74e MA |
552 | if (actual_regions_needed > in_regions_needed && |
553 | resv->region_cache_count < | |
554 | resv->adds_in_progress + | |
555 | (actual_regions_needed - in_regions_needed)) { | |
556 | /* region_add operation of range 1 should never need to | |
557 | * allocate file_region entries. | |
558 | */ | |
559 | VM_BUG_ON(t - f <= 1); | |
5e911373 | 560 | |
0db9d74e MA |
561 | if (allocate_file_region_entries( |
562 | resv, actual_regions_needed - in_regions_needed)) { | |
563 | return -ENOMEM; | |
564 | } | |
5e911373 | 565 | |
0db9d74e | 566 | goto retry; |
5e911373 MK |
567 | } |
568 | ||
972a3da3 | 569 | add = add_reservation_in_range(resv, f, t, h_cg, h, NULL); |
0db9d74e MA |
570 | |
571 | resv->adds_in_progress -= in_regions_needed; | |
cf3ad20b | 572 | |
7b24d861 | 573 | spin_unlock(&resv->lock); |
cf3ad20b | 574 | return add; |
96822904 AW |
575 | } |
576 | ||
1dd308a7 MK |
577 | /* |
578 | * Examine the existing reserve map and determine how many | |
579 | * huge pages in the specified range [f, t) are NOT currently | |
580 | * represented. This routine is called before a subsequent | |
581 | * call to region_add that will actually modify the reserve | |
582 | * map to add the specified range [f, t). region_chg does | |
583 | * not change the number of huge pages represented by the | |
0db9d74e MA |
584 | * map. A number of new file_region structures is added to the cache as a |
585 | * placeholder, for the subsequent region_add call to use. At least 1 | |
586 | * file_region structure is added. | |
587 | * | |
588 | * out_regions_needed is the number of regions added to the | |
589 | * resv->adds_in_progress. This value needs to be provided to a follow up call | |
590 | * to region_add or region_abort for proper accounting. | |
5e911373 MK |
591 | * |
592 | * Returns the number of huge pages that need to be added to the existing | |
593 | * reservation map for the range [f, t). This number is greater or equal to | |
594 | * zero. -ENOMEM is returned if a new file_region structure or cache entry | |
595 | * is needed and can not be allocated. | |
1dd308a7 | 596 | */ |
0db9d74e MA |
597 | static long region_chg(struct resv_map *resv, long f, long t, |
598 | long *out_regions_needed) | |
96822904 | 599 | { |
96822904 AW |
600 | long chg = 0; |
601 | ||
7b24d861 | 602 | spin_lock(&resv->lock); |
5e911373 | 603 | |
972a3da3 | 604 | /* Count how many hugepages in this range are NOT represented. */ |
075a61d0 | 605 | chg = add_reservation_in_range(resv, f, t, NULL, NULL, |
972a3da3 | 606 | out_regions_needed); |
5e911373 | 607 | |
0db9d74e MA |
608 | if (*out_regions_needed == 0) |
609 | *out_regions_needed = 1; | |
5e911373 | 610 | |
0db9d74e MA |
611 | if (allocate_file_region_entries(resv, *out_regions_needed)) |
612 | return -ENOMEM; | |
5e911373 | 613 | |
0db9d74e | 614 | resv->adds_in_progress += *out_regions_needed; |
7b24d861 | 615 | |
7b24d861 | 616 | spin_unlock(&resv->lock); |
96822904 AW |
617 | return chg; |
618 | } | |
619 | ||
5e911373 MK |
620 | /* |
621 | * Abort the in progress add operation. The adds_in_progress field | |
622 | * of the resv_map keeps track of the operations in progress between | |
623 | * calls to region_chg and region_add. Operations are sometimes | |
624 | * aborted after the call to region_chg. In such cases, region_abort | |
0db9d74e MA |
625 | * is called to decrement the adds_in_progress counter. regions_needed |
626 | * is the value returned by the region_chg call, it is used to decrement | |
627 | * the adds_in_progress counter. | |
5e911373 MK |
628 | * |
629 | * NOTE: The range arguments [f, t) are not needed or used in this | |
630 | * routine. They are kept to make reading the calling code easier as | |
631 | * arguments will match the associated region_chg call. | |
632 | */ | |
0db9d74e MA |
633 | static void region_abort(struct resv_map *resv, long f, long t, |
634 | long regions_needed) | |
5e911373 MK |
635 | { |
636 | spin_lock(&resv->lock); | |
637 | VM_BUG_ON(!resv->region_cache_count); | |
0db9d74e | 638 | resv->adds_in_progress -= regions_needed; |
5e911373 MK |
639 | spin_unlock(&resv->lock); |
640 | } | |
641 | ||
1dd308a7 | 642 | /* |
feba16e2 MK |
643 | * Delete the specified range [f, t) from the reserve map. If the |
644 | * t parameter is LONG_MAX, this indicates that ALL regions after f | |
645 | * should be deleted. Locate the regions which intersect [f, t) | |
646 | * and either trim, delete or split the existing regions. | |
647 | * | |
648 | * Returns the number of huge pages deleted from the reserve map. | |
649 | * In the normal case, the return value is zero or more. In the | |
650 | * case where a region must be split, a new region descriptor must | |
651 | * be allocated. If the allocation fails, -ENOMEM will be returned. | |
652 | * NOTE: If the parameter t == LONG_MAX, then we will never split | |
653 | * a region and possibly return -ENOMEM. Callers specifying | |
654 | * t == LONG_MAX do not need to check for -ENOMEM error. | |
1dd308a7 | 655 | */ |
feba16e2 | 656 | static long region_del(struct resv_map *resv, long f, long t) |
96822904 | 657 | { |
1406ec9b | 658 | struct list_head *head = &resv->regions; |
96822904 | 659 | struct file_region *rg, *trg; |
feba16e2 MK |
660 | struct file_region *nrg = NULL; |
661 | long del = 0; | |
96822904 | 662 | |
feba16e2 | 663 | retry: |
7b24d861 | 664 | spin_lock(&resv->lock); |
feba16e2 | 665 | list_for_each_entry_safe(rg, trg, head, link) { |
dbe409e4 MK |
666 | /* |
667 | * Skip regions before the range to be deleted. file_region | |
668 | * ranges are normally of the form [from, to). However, there | |
669 | * may be a "placeholder" entry in the map which is of the form | |
670 | * (from, to) with from == to. Check for placeholder entries | |
671 | * at the beginning of the range to be deleted. | |
672 | */ | |
673 | if (rg->to <= f && (rg->to != rg->from || rg->to != f)) | |
feba16e2 | 674 | continue; |
dbe409e4 | 675 | |
feba16e2 | 676 | if (rg->from >= t) |
96822904 | 677 | break; |
96822904 | 678 | |
feba16e2 MK |
679 | if (f > rg->from && t < rg->to) { /* Must split region */ |
680 | /* | |
681 | * Check for an entry in the cache before dropping | |
682 | * lock and attempting allocation. | |
683 | */ | |
684 | if (!nrg && | |
685 | resv->region_cache_count > resv->adds_in_progress) { | |
686 | nrg = list_first_entry(&resv->region_cache, | |
687 | struct file_region, | |
688 | link); | |
689 | list_del(&nrg->link); | |
690 | resv->region_cache_count--; | |
691 | } | |
96822904 | 692 | |
feba16e2 MK |
693 | if (!nrg) { |
694 | spin_unlock(&resv->lock); | |
695 | nrg = kmalloc(sizeof(*nrg), GFP_KERNEL); | |
696 | if (!nrg) | |
697 | return -ENOMEM; | |
698 | goto retry; | |
699 | } | |
700 | ||
701 | del += t - f; | |
79aa925b | 702 | hugetlb_cgroup_uncharge_file_region( |
d85aecf2 | 703 | resv, rg, t - f, false); |
feba16e2 MK |
704 | |
705 | /* New entry for end of split region */ | |
706 | nrg->from = t; | |
707 | nrg->to = rg->to; | |
075a61d0 MA |
708 | |
709 | copy_hugetlb_cgroup_uncharge_info(nrg, rg); | |
710 | ||
feba16e2 MK |
711 | INIT_LIST_HEAD(&nrg->link); |
712 | ||
713 | /* Original entry is trimmed */ | |
714 | rg->to = f; | |
715 | ||
716 | list_add(&nrg->link, &rg->link); | |
717 | nrg = NULL; | |
96822904 | 718 | break; |
feba16e2 MK |
719 | } |
720 | ||
721 | if (f <= rg->from && t >= rg->to) { /* Remove entire region */ | |
722 | del += rg->to - rg->from; | |
075a61d0 | 723 | hugetlb_cgroup_uncharge_file_region(resv, rg, |
d85aecf2 | 724 | rg->to - rg->from, true); |
feba16e2 MK |
725 | list_del(&rg->link); |
726 | kfree(rg); | |
727 | continue; | |
728 | } | |
729 | ||
730 | if (f <= rg->from) { /* Trim beginning of region */ | |
075a61d0 | 731 | hugetlb_cgroup_uncharge_file_region(resv, rg, |
d85aecf2 | 732 | t - rg->from, false); |
075a61d0 | 733 | |
79aa925b MK |
734 | del += t - rg->from; |
735 | rg->from = t; | |
736 | } else { /* Trim end of region */ | |
075a61d0 | 737 | hugetlb_cgroup_uncharge_file_region(resv, rg, |
d85aecf2 | 738 | rg->to - f, false); |
79aa925b MK |
739 | |
740 | del += rg->to - f; | |
741 | rg->to = f; | |
feba16e2 | 742 | } |
96822904 | 743 | } |
7b24d861 | 744 | |
7b24d861 | 745 | spin_unlock(&resv->lock); |
feba16e2 MK |
746 | kfree(nrg); |
747 | return del; | |
96822904 AW |
748 | } |
749 | ||
b5cec28d MK |
750 | /* |
751 | * A rare out of memory error was encountered which prevented removal of | |
752 | * the reserve map region for a page. The huge page itself was free'ed | |
753 | * and removed from the page cache. This routine will adjust the subpool | |
754 | * usage count, and the global reserve count if needed. By incrementing | |
755 | * these counts, the reserve map entry which could not be deleted will | |
756 | * appear as a "reserved" entry instead of simply dangling with incorrect | |
757 | * counts. | |
758 | */ | |
72e2936c | 759 | void hugetlb_fix_reserve_counts(struct inode *inode) |
b5cec28d MK |
760 | { |
761 | struct hugepage_subpool *spool = subpool_inode(inode); | |
762 | long rsv_adjust; | |
da56388c | 763 | bool reserved = false; |
b5cec28d MK |
764 | |
765 | rsv_adjust = hugepage_subpool_get_pages(spool, 1); | |
da56388c | 766 | if (rsv_adjust > 0) { |
b5cec28d MK |
767 | struct hstate *h = hstate_inode(inode); |
768 | ||
da56388c ML |
769 | if (!hugetlb_acct_memory(h, 1)) |
770 | reserved = true; | |
771 | } else if (!rsv_adjust) { | |
772 | reserved = true; | |
b5cec28d | 773 | } |
da56388c ML |
774 | |
775 | if (!reserved) | |
776 | pr_warn("hugetlb: Huge Page Reserved count may go negative.\n"); | |
b5cec28d MK |
777 | } |
778 | ||
1dd308a7 MK |
779 | /* |
780 | * Count and return the number of huge pages in the reserve map | |
781 | * that intersect with the range [f, t). | |
782 | */ | |
1406ec9b | 783 | static long region_count(struct resv_map *resv, long f, long t) |
84afd99b | 784 | { |
1406ec9b | 785 | struct list_head *head = &resv->regions; |
84afd99b AW |
786 | struct file_region *rg; |
787 | long chg = 0; | |
788 | ||
7b24d861 | 789 | spin_lock(&resv->lock); |
84afd99b AW |
790 | /* Locate each segment we overlap with, and count that overlap. */ |
791 | list_for_each_entry(rg, head, link) { | |
f2135a4a WSH |
792 | long seg_from; |
793 | long seg_to; | |
84afd99b AW |
794 | |
795 | if (rg->to <= f) | |
796 | continue; | |
797 | if (rg->from >= t) | |
798 | break; | |
799 | ||
800 | seg_from = max(rg->from, f); | |
801 | seg_to = min(rg->to, t); | |
802 | ||
803 | chg += seg_to - seg_from; | |
804 | } | |
7b24d861 | 805 | spin_unlock(&resv->lock); |
84afd99b AW |
806 | |
807 | return chg; | |
808 | } | |
809 | ||
e7c4b0bf AW |
810 | /* |
811 | * Convert the address within this vma to the page offset within | |
812 | * the mapping, in pagecache page units; huge pages here. | |
813 | */ | |
a5516438 AK |
814 | static pgoff_t vma_hugecache_offset(struct hstate *h, |
815 | struct vm_area_struct *vma, unsigned long address) | |
e7c4b0bf | 816 | { |
a5516438 AK |
817 | return ((address - vma->vm_start) >> huge_page_shift(h)) + |
818 | (vma->vm_pgoff >> huge_page_order(h)); | |
e7c4b0bf AW |
819 | } |
820 | ||
0fe6e20b NH |
821 | pgoff_t linear_hugepage_index(struct vm_area_struct *vma, |
822 | unsigned long address) | |
823 | { | |
824 | return vma_hugecache_offset(hstate_vma(vma), vma, address); | |
825 | } | |
dee41079 | 826 | EXPORT_SYMBOL_GPL(linear_hugepage_index); |
0fe6e20b | 827 | |
08fba699 MG |
828 | /* |
829 | * Return the size of the pages allocated when backing a VMA. In the majority | |
830 | * cases this will be same size as used by the page table entries. | |
831 | */ | |
832 | unsigned long vma_kernel_pagesize(struct vm_area_struct *vma) | |
833 | { | |
05ea8860 DW |
834 | if (vma->vm_ops && vma->vm_ops->pagesize) |
835 | return vma->vm_ops->pagesize(vma); | |
836 | return PAGE_SIZE; | |
08fba699 | 837 | } |
f340ca0f | 838 | EXPORT_SYMBOL_GPL(vma_kernel_pagesize); |
08fba699 | 839 | |
3340289d MG |
840 | /* |
841 | * Return the page size being used by the MMU to back a VMA. In the majority | |
842 | * of cases, the page size used by the kernel matches the MMU size. On | |
09135cc5 DW |
843 | * architectures where it differs, an architecture-specific 'strong' |
844 | * version of this symbol is required. | |
3340289d | 845 | */ |
09135cc5 | 846 | __weak unsigned long vma_mmu_pagesize(struct vm_area_struct *vma) |
3340289d MG |
847 | { |
848 | return vma_kernel_pagesize(vma); | |
849 | } | |
3340289d | 850 | |
84afd99b AW |
851 | /* |
852 | * Flags for MAP_PRIVATE reservations. These are stored in the bottom | |
853 | * bits of the reservation map pointer, which are always clear due to | |
854 | * alignment. | |
855 | */ | |
856 | #define HPAGE_RESV_OWNER (1UL << 0) | |
857 | #define HPAGE_RESV_UNMAPPED (1UL << 1) | |
04f2cbe3 | 858 | #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED) |
84afd99b | 859 | |
a1e78772 MG |
860 | /* |
861 | * These helpers are used to track how many pages are reserved for | |
862 | * faults in a MAP_PRIVATE mapping. Only the process that called mmap() | |
863 | * is guaranteed to have their future faults succeed. | |
864 | * | |
8d9bfb26 | 865 | * With the exception of hugetlb_dup_vma_private() which is called at fork(), |
a1e78772 MG |
866 | * the reserve counters are updated with the hugetlb_lock held. It is safe |
867 | * to reset the VMA at fork() time as it is not in use yet and there is no | |
868 | * chance of the global counters getting corrupted as a result of the values. | |
84afd99b AW |
869 | * |
870 | * The private mapping reservation is represented in a subtly different | |
871 | * manner to a shared mapping. A shared mapping has a region map associated | |
872 | * with the underlying file, this region map represents the backing file | |
873 | * pages which have ever had a reservation assigned which this persists even | |
874 | * after the page is instantiated. A private mapping has a region map | |
875 | * associated with the original mmap which is attached to all VMAs which | |
876 | * reference it, this region map represents those offsets which have consumed | |
877 | * reservation ie. where pages have been instantiated. | |
a1e78772 | 878 | */ |
e7c4b0bf AW |
879 | static unsigned long get_vma_private_data(struct vm_area_struct *vma) |
880 | { | |
881 | return (unsigned long)vma->vm_private_data; | |
882 | } | |
883 | ||
884 | static void set_vma_private_data(struct vm_area_struct *vma, | |
885 | unsigned long value) | |
886 | { | |
887 | vma->vm_private_data = (void *)value; | |
888 | } | |
889 | ||
e9fe92ae MA |
890 | static void |
891 | resv_map_set_hugetlb_cgroup_uncharge_info(struct resv_map *resv_map, | |
892 | struct hugetlb_cgroup *h_cg, | |
893 | struct hstate *h) | |
894 | { | |
895 | #ifdef CONFIG_CGROUP_HUGETLB | |
896 | if (!h_cg || !h) { | |
897 | resv_map->reservation_counter = NULL; | |
898 | resv_map->pages_per_hpage = 0; | |
899 | resv_map->css = NULL; | |
900 | } else { | |
901 | resv_map->reservation_counter = | |
902 | &h_cg->rsvd_hugepage[hstate_index(h)]; | |
903 | resv_map->pages_per_hpage = pages_per_huge_page(h); | |
904 | resv_map->css = &h_cg->css; | |
905 | } | |
906 | #endif | |
907 | } | |
908 | ||
9119a41e | 909 | struct resv_map *resv_map_alloc(void) |
84afd99b AW |
910 | { |
911 | struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL); | |
5e911373 MK |
912 | struct file_region *rg = kmalloc(sizeof(*rg), GFP_KERNEL); |
913 | ||
914 | if (!resv_map || !rg) { | |
915 | kfree(resv_map); | |
916 | kfree(rg); | |
84afd99b | 917 | return NULL; |
5e911373 | 918 | } |
84afd99b AW |
919 | |
920 | kref_init(&resv_map->refs); | |
7b24d861 | 921 | spin_lock_init(&resv_map->lock); |
84afd99b AW |
922 | INIT_LIST_HEAD(&resv_map->regions); |
923 | ||
5e911373 | 924 | resv_map->adds_in_progress = 0; |
e9fe92ae MA |
925 | /* |
926 | * Initialize these to 0. On shared mappings, 0's here indicate these | |
927 | * fields don't do cgroup accounting. On private mappings, these will be | |
928 | * re-initialized to the proper values, to indicate that hugetlb cgroup | |
929 | * reservations are to be un-charged from here. | |
930 | */ | |
931 | resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, NULL, NULL); | |
5e911373 MK |
932 | |
933 | INIT_LIST_HEAD(&resv_map->region_cache); | |
934 | list_add(&rg->link, &resv_map->region_cache); | |
935 | resv_map->region_cache_count = 1; | |
936 | ||
84afd99b AW |
937 | return resv_map; |
938 | } | |
939 | ||
9119a41e | 940 | void resv_map_release(struct kref *ref) |
84afd99b AW |
941 | { |
942 | struct resv_map *resv_map = container_of(ref, struct resv_map, refs); | |
5e911373 MK |
943 | struct list_head *head = &resv_map->region_cache; |
944 | struct file_region *rg, *trg; | |
84afd99b AW |
945 | |
946 | /* Clear out any active regions before we release the map. */ | |
feba16e2 | 947 | region_del(resv_map, 0, LONG_MAX); |
5e911373 MK |
948 | |
949 | /* ... and any entries left in the cache */ | |
950 | list_for_each_entry_safe(rg, trg, head, link) { | |
951 | list_del(&rg->link); | |
952 | kfree(rg); | |
953 | } | |
954 | ||
955 | VM_BUG_ON(resv_map->adds_in_progress); | |
956 | ||
84afd99b AW |
957 | kfree(resv_map); |
958 | } | |
959 | ||
4e35f483 JK |
960 | static inline struct resv_map *inode_resv_map(struct inode *inode) |
961 | { | |
f27a5136 MK |
962 | /* |
963 | * At inode evict time, i_mapping may not point to the original | |
964 | * address space within the inode. This original address space | |
965 | * contains the pointer to the resv_map. So, always use the | |
966 | * address space embedded within the inode. | |
967 | * The VERY common case is inode->mapping == &inode->i_data but, | |
968 | * this may not be true for device special inodes. | |
969 | */ | |
970 | return (struct resv_map *)(&inode->i_data)->private_data; | |
4e35f483 JK |
971 | } |
972 | ||
84afd99b | 973 | static struct resv_map *vma_resv_map(struct vm_area_struct *vma) |
a1e78772 | 974 | { |
81d1b09c | 975 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); |
4e35f483 JK |
976 | if (vma->vm_flags & VM_MAYSHARE) { |
977 | struct address_space *mapping = vma->vm_file->f_mapping; | |
978 | struct inode *inode = mapping->host; | |
979 | ||
980 | return inode_resv_map(inode); | |
981 | ||
982 | } else { | |
84afd99b AW |
983 | return (struct resv_map *)(get_vma_private_data(vma) & |
984 | ~HPAGE_RESV_MASK); | |
4e35f483 | 985 | } |
a1e78772 MG |
986 | } |
987 | ||
84afd99b | 988 | static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map) |
a1e78772 | 989 | { |
81d1b09c SL |
990 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); |
991 | VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma); | |
a1e78772 | 992 | |
84afd99b AW |
993 | set_vma_private_data(vma, (get_vma_private_data(vma) & |
994 | HPAGE_RESV_MASK) | (unsigned long)map); | |
04f2cbe3 MG |
995 | } |
996 | ||
997 | static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags) | |
998 | { | |
81d1b09c SL |
999 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); |
1000 | VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma); | |
e7c4b0bf AW |
1001 | |
1002 | set_vma_private_data(vma, get_vma_private_data(vma) | flags); | |
04f2cbe3 MG |
1003 | } |
1004 | ||
1005 | static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag) | |
1006 | { | |
81d1b09c | 1007 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); |
e7c4b0bf AW |
1008 | |
1009 | return (get_vma_private_data(vma) & flag) != 0; | |
a1e78772 MG |
1010 | } |
1011 | ||
8d9bfb26 | 1012 | void hugetlb_dup_vma_private(struct vm_area_struct *vma) |
a1e78772 | 1013 | { |
81d1b09c | 1014 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); |
8d9bfb26 MK |
1015 | /* |
1016 | * Clear vm_private_data | |
612b8a31 MK |
1017 | * - For shared mappings this is a per-vma semaphore that may be |
1018 | * allocated in a subsequent call to hugetlb_vm_op_open. | |
1019 | * Before clearing, make sure pointer is not associated with vma | |
1020 | * as this will leak the structure. This is the case when called | |
1021 | * via clear_vma_resv_huge_pages() and hugetlb_vm_op_open has already | |
1022 | * been called to allocate a new structure. | |
8d9bfb26 MK |
1023 | * - For MAP_PRIVATE mappings, this is the reserve map which does |
1024 | * not apply to children. Faults generated by the children are | |
1025 | * not guaranteed to succeed, even if read-only. | |
8d9bfb26 | 1026 | */ |
612b8a31 MK |
1027 | if (vma->vm_flags & VM_MAYSHARE) { |
1028 | struct hugetlb_vma_lock *vma_lock = vma->vm_private_data; | |
1029 | ||
1030 | if (vma_lock && vma_lock->vma != vma) | |
1031 | vma->vm_private_data = NULL; | |
1032 | } else | |
1033 | vma->vm_private_data = NULL; | |
a1e78772 MG |
1034 | } |
1035 | ||
550a7d60 MA |
1036 | /* |
1037 | * Reset and decrement one ref on hugepage private reservation. | |
1038 | * Called with mm->mmap_sem writer semaphore held. | |
1039 | * This function should be only used by move_vma() and operate on | |
1040 | * same sized vma. It should never come here with last ref on the | |
1041 | * reservation. | |
1042 | */ | |
1043 | void clear_vma_resv_huge_pages(struct vm_area_struct *vma) | |
1044 | { | |
1045 | /* | |
1046 | * Clear the old hugetlb private page reservation. | |
1047 | * It has already been transferred to new_vma. | |
1048 | * | |
1049 | * During a mremap() operation of a hugetlb vma we call move_vma() | |
1050 | * which copies vma into new_vma and unmaps vma. After the copy | |
1051 | * operation both new_vma and vma share a reference to the resv_map | |
1052 | * struct, and at that point vma is about to be unmapped. We don't | |
1053 | * want to return the reservation to the pool at unmap of vma because | |
1054 | * the reservation still lives on in new_vma, so simply decrement the | |
1055 | * ref here and remove the resv_map reference from this vma. | |
1056 | */ | |
1057 | struct resv_map *reservations = vma_resv_map(vma); | |
1058 | ||
afe041c2 BQM |
1059 | if (reservations && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { |
1060 | resv_map_put_hugetlb_cgroup_uncharge_info(reservations); | |
550a7d60 | 1061 | kref_put(&reservations->refs, resv_map_release); |
afe041c2 | 1062 | } |
550a7d60 | 1063 | |
8d9bfb26 | 1064 | hugetlb_dup_vma_private(vma); |
550a7d60 MA |
1065 | } |
1066 | ||
a1e78772 | 1067 | /* Returns true if the VMA has associated reserve pages */ |
559ec2f8 | 1068 | static bool vma_has_reserves(struct vm_area_struct *vma, long chg) |
a1e78772 | 1069 | { |
af0ed73e JK |
1070 | if (vma->vm_flags & VM_NORESERVE) { |
1071 | /* | |
1072 | * This address is already reserved by other process(chg == 0), | |
1073 | * so, we should decrement reserved count. Without decrementing, | |
1074 | * reserve count remains after releasing inode, because this | |
1075 | * allocated page will go into page cache and is regarded as | |
1076 | * coming from reserved pool in releasing step. Currently, we | |
1077 | * don't have any other solution to deal with this situation | |
1078 | * properly, so add work-around here. | |
1079 | */ | |
1080 | if (vma->vm_flags & VM_MAYSHARE && chg == 0) | |
559ec2f8 | 1081 | return true; |
af0ed73e | 1082 | else |
559ec2f8 | 1083 | return false; |
af0ed73e | 1084 | } |
a63884e9 JK |
1085 | |
1086 | /* Shared mappings always use reserves */ | |
1fb1b0e9 MK |
1087 | if (vma->vm_flags & VM_MAYSHARE) { |
1088 | /* | |
1089 | * We know VM_NORESERVE is not set. Therefore, there SHOULD | |
1090 | * be a region map for all pages. The only situation where | |
1091 | * there is no region map is if a hole was punched via | |
7c8de358 | 1092 | * fallocate. In this case, there really are no reserves to |
1fb1b0e9 MK |
1093 | * use. This situation is indicated if chg != 0. |
1094 | */ | |
1095 | if (chg) | |
1096 | return false; | |
1097 | else | |
1098 | return true; | |
1099 | } | |
a63884e9 JK |
1100 | |
1101 | /* | |
1102 | * Only the process that called mmap() has reserves for | |
1103 | * private mappings. | |
1104 | */ | |
67961f9d MK |
1105 | if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { |
1106 | /* | |
1107 | * Like the shared case above, a hole punch or truncate | |
1108 | * could have been performed on the private mapping. | |
1109 | * Examine the value of chg to determine if reserves | |
1110 | * actually exist or were previously consumed. | |
1111 | * Very Subtle - The value of chg comes from a previous | |
1112 | * call to vma_needs_reserves(). The reserve map for | |
1113 | * private mappings has different (opposite) semantics | |
1114 | * than that of shared mappings. vma_needs_reserves() | |
1115 | * has already taken this difference in semantics into | |
1116 | * account. Therefore, the meaning of chg is the same | |
1117 | * as in the shared case above. Code could easily be | |
1118 | * combined, but keeping it separate draws attention to | |
1119 | * subtle differences. | |
1120 | */ | |
1121 | if (chg) | |
1122 | return false; | |
1123 | else | |
1124 | return true; | |
1125 | } | |
a63884e9 | 1126 | |
559ec2f8 | 1127 | return false; |
a1e78772 MG |
1128 | } |
1129 | ||
a5516438 | 1130 | static void enqueue_huge_page(struct hstate *h, struct page *page) |
1da177e4 LT |
1131 | { |
1132 | int nid = page_to_nid(page); | |
9487ca60 MK |
1133 | |
1134 | lockdep_assert_held(&hugetlb_lock); | |
b65a4eda MK |
1135 | VM_BUG_ON_PAGE(page_count(page), page); |
1136 | ||
0edaecfa | 1137 | list_move(&page->lru, &h->hugepage_freelists[nid]); |
a5516438 AK |
1138 | h->free_huge_pages++; |
1139 | h->free_huge_pages_node[nid]++; | |
6c037149 | 1140 | SetHPageFreed(page); |
1da177e4 LT |
1141 | } |
1142 | ||
94310cbc | 1143 | static struct page *dequeue_huge_page_node_exact(struct hstate *h, int nid) |
bf50bab2 NH |
1144 | { |
1145 | struct page *page; | |
1a08ae36 | 1146 | bool pin = !!(current->flags & PF_MEMALLOC_PIN); |
bbe88753 | 1147 | |
9487ca60 | 1148 | lockdep_assert_held(&hugetlb_lock); |
bbe88753 | 1149 | list_for_each_entry(page, &h->hugepage_freelists[nid], lru) { |
6077c943 | 1150 | if (pin && !is_longterm_pinnable_page(page)) |
bbe88753 | 1151 | continue; |
bf50bab2 | 1152 | |
6664bfc8 WY |
1153 | if (PageHWPoison(page)) |
1154 | continue; | |
1155 | ||
1156 | list_move(&page->lru, &h->hugepage_activelist); | |
1157 | set_page_refcounted(page); | |
6c037149 | 1158 | ClearHPageFreed(page); |
6664bfc8 WY |
1159 | h->free_huge_pages--; |
1160 | h->free_huge_pages_node[nid]--; | |
1161 | return page; | |
bbe88753 JK |
1162 | } |
1163 | ||
6664bfc8 | 1164 | return NULL; |
bf50bab2 NH |
1165 | } |
1166 | ||
3e59fcb0 MH |
1167 | static struct page *dequeue_huge_page_nodemask(struct hstate *h, gfp_t gfp_mask, int nid, |
1168 | nodemask_t *nmask) | |
94310cbc | 1169 | { |
3e59fcb0 MH |
1170 | unsigned int cpuset_mems_cookie; |
1171 | struct zonelist *zonelist; | |
1172 | struct zone *zone; | |
1173 | struct zoneref *z; | |
98fa15f3 | 1174 | int node = NUMA_NO_NODE; |
94310cbc | 1175 | |
3e59fcb0 MH |
1176 | zonelist = node_zonelist(nid, gfp_mask); |
1177 | ||
1178 | retry_cpuset: | |
1179 | cpuset_mems_cookie = read_mems_allowed_begin(); | |
1180 | for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), nmask) { | |
1181 | struct page *page; | |
1182 | ||
1183 | if (!cpuset_zone_allowed(zone, gfp_mask)) | |
1184 | continue; | |
1185 | /* | |
1186 | * no need to ask again on the same node. Pool is node rather than | |
1187 | * zone aware | |
1188 | */ | |
1189 | if (zone_to_nid(zone) == node) | |
1190 | continue; | |
1191 | node = zone_to_nid(zone); | |
94310cbc | 1192 | |
94310cbc AK |
1193 | page = dequeue_huge_page_node_exact(h, node); |
1194 | if (page) | |
1195 | return page; | |
1196 | } | |
3e59fcb0 MH |
1197 | if (unlikely(read_mems_allowed_retry(cpuset_mems_cookie))) |
1198 | goto retry_cpuset; | |
1199 | ||
94310cbc AK |
1200 | return NULL; |
1201 | } | |
1202 | ||
8346d69d XH |
1203 | static unsigned long available_huge_pages(struct hstate *h) |
1204 | { | |
1205 | return h->free_huge_pages - h->resv_huge_pages; | |
1206 | } | |
1207 | ||
a5516438 AK |
1208 | static struct page *dequeue_huge_page_vma(struct hstate *h, |
1209 | struct vm_area_struct *vma, | |
af0ed73e JK |
1210 | unsigned long address, int avoid_reserve, |
1211 | long chg) | |
1da177e4 | 1212 | { |
cfcaa66f | 1213 | struct page *page = NULL; |
480eccf9 | 1214 | struct mempolicy *mpol; |
04ec6264 | 1215 | gfp_t gfp_mask; |
3e59fcb0 | 1216 | nodemask_t *nodemask; |
04ec6264 | 1217 | int nid; |
1da177e4 | 1218 | |
a1e78772 MG |
1219 | /* |
1220 | * A child process with MAP_PRIVATE mappings created by their parent | |
1221 | * have no page reserves. This check ensures that reservations are | |
1222 | * not "stolen". The child may still get SIGKILLed | |
1223 | */ | |
8346d69d | 1224 | if (!vma_has_reserves(vma, chg) && !available_huge_pages(h)) |
c0ff7453 | 1225 | goto err; |
a1e78772 | 1226 | |
04f2cbe3 | 1227 | /* If reserves cannot be used, ensure enough pages are in the pool */ |
8346d69d | 1228 | if (avoid_reserve && !available_huge_pages(h)) |
6eab04a8 | 1229 | goto err; |
04f2cbe3 | 1230 | |
04ec6264 VB |
1231 | gfp_mask = htlb_alloc_mask(h); |
1232 | nid = huge_node(vma, address, gfp_mask, &mpol, &nodemask); | |
cfcaa66f BW |
1233 | |
1234 | if (mpol_is_preferred_many(mpol)) { | |
1235 | page = dequeue_huge_page_nodemask(h, gfp_mask, nid, nodemask); | |
1236 | ||
1237 | /* Fallback to all nodes if page==NULL */ | |
1238 | nodemask = NULL; | |
1239 | } | |
1240 | ||
1241 | if (!page) | |
1242 | page = dequeue_huge_page_nodemask(h, gfp_mask, nid, nodemask); | |
1243 | ||
3e59fcb0 | 1244 | if (page && !avoid_reserve && vma_has_reserves(vma, chg)) { |
d6995da3 | 1245 | SetHPageRestoreReserve(page); |
3e59fcb0 | 1246 | h->resv_huge_pages--; |
1da177e4 | 1247 | } |
cc9a6c87 | 1248 | |
52cd3b07 | 1249 | mpol_cond_put(mpol); |
1da177e4 | 1250 | return page; |
cc9a6c87 MG |
1251 | |
1252 | err: | |
cc9a6c87 | 1253 | return NULL; |
1da177e4 LT |
1254 | } |
1255 | ||
1cac6f2c LC |
1256 | /* |
1257 | * common helper functions for hstate_next_node_to_{alloc|free}. | |
1258 | * We may have allocated or freed a huge page based on a different | |
1259 | * nodes_allowed previously, so h->next_node_to_{alloc|free} might | |
1260 | * be outside of *nodes_allowed. Ensure that we use an allowed | |
1261 | * node for alloc or free. | |
1262 | */ | |
1263 | static int next_node_allowed(int nid, nodemask_t *nodes_allowed) | |
1264 | { | |
0edaf86c | 1265 | nid = next_node_in(nid, *nodes_allowed); |
1cac6f2c LC |
1266 | VM_BUG_ON(nid >= MAX_NUMNODES); |
1267 | ||
1268 | return nid; | |
1269 | } | |
1270 | ||
1271 | static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed) | |
1272 | { | |
1273 | if (!node_isset(nid, *nodes_allowed)) | |
1274 | nid = next_node_allowed(nid, nodes_allowed); | |
1275 | return nid; | |
1276 | } | |
1277 | ||
1278 | /* | |
1279 | * returns the previously saved node ["this node"] from which to | |
1280 | * allocate a persistent huge page for the pool and advance the | |
1281 | * next node from which to allocate, handling wrap at end of node | |
1282 | * mask. | |
1283 | */ | |
1284 | static int hstate_next_node_to_alloc(struct hstate *h, | |
1285 | nodemask_t *nodes_allowed) | |
1286 | { | |
1287 | int nid; | |
1288 | ||
1289 | VM_BUG_ON(!nodes_allowed); | |
1290 | ||
1291 | nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed); | |
1292 | h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed); | |
1293 | ||
1294 | return nid; | |
1295 | } | |
1296 | ||
1297 | /* | |
10c6ec49 | 1298 | * helper for remove_pool_huge_page() - return the previously saved |
1cac6f2c LC |
1299 | * node ["this node"] from which to free a huge page. Advance the |
1300 | * next node id whether or not we find a free huge page to free so | |
1301 | * that the next attempt to free addresses the next node. | |
1302 | */ | |
1303 | static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed) | |
1304 | { | |
1305 | int nid; | |
1306 | ||
1307 | VM_BUG_ON(!nodes_allowed); | |
1308 | ||
1309 | nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed); | |
1310 | h->next_nid_to_free = next_node_allowed(nid, nodes_allowed); | |
1311 | ||
1312 | return nid; | |
1313 | } | |
1314 | ||
1315 | #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \ | |
1316 | for (nr_nodes = nodes_weight(*mask); \ | |
1317 | nr_nodes > 0 && \ | |
1318 | ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \ | |
1319 | nr_nodes--) | |
1320 | ||
1321 | #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \ | |
1322 | for (nr_nodes = nodes_weight(*mask); \ | |
1323 | nr_nodes > 0 && \ | |
1324 | ((node = hstate_next_node_to_free(hs, mask)) || 1); \ | |
1325 | nr_nodes--) | |
1326 | ||
8531fc6f | 1327 | /* used to demote non-gigantic_huge pages as well */ |
34d9e35b MK |
1328 | static void __destroy_compound_gigantic_page(struct page *page, |
1329 | unsigned int order, bool demote) | |
944d9fec LC |
1330 | { |
1331 | int i; | |
1332 | int nr_pages = 1 << order; | |
14455eab | 1333 | struct page *p; |
944d9fec | 1334 | |
c8cc708a | 1335 | atomic_set(compound_mapcount_ptr(page), 0); |
5291c09b | 1336 | atomic_set(compound_pincount_ptr(page), 0); |
47e29d32 | 1337 | |
14455eab CL |
1338 | for (i = 1; i < nr_pages; i++) { |
1339 | p = nth_page(page, i); | |
a01f4390 | 1340 | p->mapping = NULL; |
1d798ca3 | 1341 | clear_compound_head(p); |
34d9e35b MK |
1342 | if (!demote) |
1343 | set_page_refcounted(p); | |
944d9fec LC |
1344 | } |
1345 | ||
1346 | set_compound_order(page, 0); | |
5232c63f | 1347 | #ifdef CONFIG_64BIT |
ba9c1201 | 1348 | page[1].compound_nr = 0; |
5232c63f | 1349 | #endif |
944d9fec LC |
1350 | __ClearPageHead(page); |
1351 | } | |
1352 | ||
8531fc6f MK |
1353 | static void destroy_compound_hugetlb_page_for_demote(struct page *page, |
1354 | unsigned int order) | |
1355 | { | |
1356 | __destroy_compound_gigantic_page(page, order, true); | |
1357 | } | |
1358 | ||
1359 | #ifdef CONFIG_ARCH_HAS_GIGANTIC_PAGE | |
34d9e35b MK |
1360 | static void destroy_compound_gigantic_page(struct page *page, |
1361 | unsigned int order) | |
1362 | { | |
1363 | __destroy_compound_gigantic_page(page, order, false); | |
1364 | } | |
1365 | ||
d00181b9 | 1366 | static void free_gigantic_page(struct page *page, unsigned int order) |
944d9fec | 1367 | { |
cf11e85f RG |
1368 | /* |
1369 | * If the page isn't allocated using the cma allocator, | |
1370 | * cma_release() returns false. | |
1371 | */ | |
dbda8fea BS |
1372 | #ifdef CONFIG_CMA |
1373 | if (cma_release(hugetlb_cma[page_to_nid(page)], page, 1 << order)) | |
cf11e85f | 1374 | return; |
dbda8fea | 1375 | #endif |
cf11e85f | 1376 | |
944d9fec LC |
1377 | free_contig_range(page_to_pfn(page), 1 << order); |
1378 | } | |
1379 | ||
4eb0716e | 1380 | #ifdef CONFIG_CONTIG_ALLOC |
d9cc948f MH |
1381 | static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask, |
1382 | int nid, nodemask_t *nodemask) | |
944d9fec | 1383 | { |
04adbc3f | 1384 | unsigned long nr_pages = pages_per_huge_page(h); |
953f064a LX |
1385 | if (nid == NUMA_NO_NODE) |
1386 | nid = numa_mem_id(); | |
944d9fec | 1387 | |
dbda8fea BS |
1388 | #ifdef CONFIG_CMA |
1389 | { | |
cf11e85f RG |
1390 | struct page *page; |
1391 | int node; | |
1392 | ||
953f064a LX |
1393 | if (hugetlb_cma[nid]) { |
1394 | page = cma_alloc(hugetlb_cma[nid], nr_pages, | |
1395 | huge_page_order(h), true); | |
cf11e85f RG |
1396 | if (page) |
1397 | return page; | |
1398 | } | |
953f064a LX |
1399 | |
1400 | if (!(gfp_mask & __GFP_THISNODE)) { | |
1401 | for_each_node_mask(node, *nodemask) { | |
1402 | if (node == nid || !hugetlb_cma[node]) | |
1403 | continue; | |
1404 | ||
1405 | page = cma_alloc(hugetlb_cma[node], nr_pages, | |
1406 | huge_page_order(h), true); | |
1407 | if (page) | |
1408 | return page; | |
1409 | } | |
1410 | } | |
cf11e85f | 1411 | } |
dbda8fea | 1412 | #endif |
cf11e85f | 1413 | |
5e27a2df | 1414 | return alloc_contig_pages(nr_pages, gfp_mask, nid, nodemask); |
944d9fec LC |
1415 | } |
1416 | ||
4eb0716e AG |
1417 | #else /* !CONFIG_CONTIG_ALLOC */ |
1418 | static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask, | |
1419 | int nid, nodemask_t *nodemask) | |
1420 | { | |
1421 | return NULL; | |
1422 | } | |
1423 | #endif /* CONFIG_CONTIG_ALLOC */ | |
944d9fec | 1424 | |
e1073d1e | 1425 | #else /* !CONFIG_ARCH_HAS_GIGANTIC_PAGE */ |
d9cc948f | 1426 | static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask, |
4eb0716e AG |
1427 | int nid, nodemask_t *nodemask) |
1428 | { | |
1429 | return NULL; | |
1430 | } | |
d00181b9 | 1431 | static inline void free_gigantic_page(struct page *page, unsigned int order) { } |
944d9fec | 1432 | static inline void destroy_compound_gigantic_page(struct page *page, |
d00181b9 | 1433 | unsigned int order) { } |
944d9fec LC |
1434 | #endif |
1435 | ||
6eb4e88a MK |
1436 | /* |
1437 | * Remove hugetlb page from lists, and update dtor so that page appears | |
34d9e35b MK |
1438 | * as just a compound page. |
1439 | * | |
1440 | * A reference is held on the page, except in the case of demote. | |
6eb4e88a MK |
1441 | * |
1442 | * Must be called with hugetlb lock held. | |
1443 | */ | |
34d9e35b MK |
1444 | static void __remove_hugetlb_page(struct hstate *h, struct page *page, |
1445 | bool adjust_surplus, | |
1446 | bool demote) | |
6eb4e88a MK |
1447 | { |
1448 | int nid = page_to_nid(page); | |
1449 | ||
1450 | VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page); | |
1451 | VM_BUG_ON_PAGE(hugetlb_cgroup_from_page_rsvd(page), page); | |
1452 | ||
9487ca60 | 1453 | lockdep_assert_held(&hugetlb_lock); |
6eb4e88a MK |
1454 | if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) |
1455 | return; | |
1456 | ||
1457 | list_del(&page->lru); | |
1458 | ||
1459 | if (HPageFreed(page)) { | |
1460 | h->free_huge_pages--; | |
1461 | h->free_huge_pages_node[nid]--; | |
1462 | } | |
1463 | if (adjust_surplus) { | |
1464 | h->surplus_huge_pages--; | |
1465 | h->surplus_huge_pages_node[nid]--; | |
1466 | } | |
1467 | ||
e32d20c0 MK |
1468 | /* |
1469 | * Very subtle | |
1470 | * | |
1471 | * For non-gigantic pages set the destructor to the normal compound | |
1472 | * page dtor. This is needed in case someone takes an additional | |
1473 | * temporary ref to the page, and freeing is delayed until they drop | |
1474 | * their reference. | |
1475 | * | |
1476 | * For gigantic pages set the destructor to the null dtor. This | |
1477 | * destructor will never be called. Before freeing the gigantic | |
1478 | * page destroy_compound_gigantic_page will turn the compound page | |
1479 | * into a simple group of pages. After this the destructor does not | |
1480 | * apply. | |
1481 | * | |
1482 | * This handles the case where more than one ref is held when and | |
1483 | * after update_and_free_page is called. | |
34d9e35b MK |
1484 | * |
1485 | * In the case of demote we do not ref count the page as it will soon | |
1486 | * be turned into a page of smaller size. | |
e32d20c0 | 1487 | */ |
34d9e35b MK |
1488 | if (!demote) |
1489 | set_page_refcounted(page); | |
e32d20c0 MK |
1490 | if (hstate_is_gigantic(h)) |
1491 | set_compound_page_dtor(page, NULL_COMPOUND_DTOR); | |
1492 | else | |
1493 | set_compound_page_dtor(page, COMPOUND_PAGE_DTOR); | |
6eb4e88a MK |
1494 | |
1495 | h->nr_huge_pages--; | |
1496 | h->nr_huge_pages_node[nid]--; | |
1497 | } | |
1498 | ||
34d9e35b MK |
1499 | static void remove_hugetlb_page(struct hstate *h, struct page *page, |
1500 | bool adjust_surplus) | |
1501 | { | |
1502 | __remove_hugetlb_page(h, page, adjust_surplus, false); | |
1503 | } | |
1504 | ||
8531fc6f MK |
1505 | static void remove_hugetlb_page_for_demote(struct hstate *h, struct page *page, |
1506 | bool adjust_surplus) | |
1507 | { | |
1508 | __remove_hugetlb_page(h, page, adjust_surplus, true); | |
1509 | } | |
1510 | ||
ad2fa371 MS |
1511 | static void add_hugetlb_page(struct hstate *h, struct page *page, |
1512 | bool adjust_surplus) | |
1513 | { | |
1514 | int zeroed; | |
1515 | int nid = page_to_nid(page); | |
1516 | ||
1517 | VM_BUG_ON_PAGE(!HPageVmemmapOptimized(page), page); | |
1518 | ||
1519 | lockdep_assert_held(&hugetlb_lock); | |
1520 | ||
1521 | INIT_LIST_HEAD(&page->lru); | |
1522 | h->nr_huge_pages++; | |
1523 | h->nr_huge_pages_node[nid]++; | |
1524 | ||
1525 | if (adjust_surplus) { | |
1526 | h->surplus_huge_pages++; | |
1527 | h->surplus_huge_pages_node[nid]++; | |
1528 | } | |
1529 | ||
1530 | set_compound_page_dtor(page, HUGETLB_PAGE_DTOR); | |
1531 | set_page_private(page, 0); | |
a9e1eab2 ML |
1532 | /* |
1533 | * We have to set HPageVmemmapOptimized again as above | |
1534 | * set_page_private(page, 0) cleared it. | |
1535 | */ | |
ad2fa371 MS |
1536 | SetHPageVmemmapOptimized(page); |
1537 | ||
1538 | /* | |
b65a4eda MK |
1539 | * This page is about to be managed by the hugetlb allocator and |
1540 | * should have no users. Drop our reference, and check for others | |
1541 | * just in case. | |
ad2fa371 MS |
1542 | */ |
1543 | zeroed = put_page_testzero(page); | |
b65a4eda MK |
1544 | if (!zeroed) |
1545 | /* | |
1546 | * It is VERY unlikely soneone else has taken a ref on | |
1547 | * the page. In this case, we simply return as the | |
1548 | * hugetlb destructor (free_huge_page) will be called | |
1549 | * when this other ref is dropped. | |
1550 | */ | |
1551 | return; | |
1552 | ||
ad2fa371 MS |
1553 | arch_clear_hugepage_flags(page); |
1554 | enqueue_huge_page(h, page); | |
1555 | } | |
1556 | ||
b65d4adb | 1557 | static void __update_and_free_page(struct hstate *h, struct page *page) |
6af2acb6 AL |
1558 | { |
1559 | int i; | |
14455eab | 1560 | struct page *subpage; |
a5516438 | 1561 | |
4eb0716e | 1562 | if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) |
944d9fec | 1563 | return; |
18229df5 | 1564 | |
161df60e NH |
1565 | /* |
1566 | * If we don't know which subpages are hwpoisoned, we can't free | |
1567 | * the hugepage, so it's leaked intentionally. | |
1568 | */ | |
1569 | if (HPageRawHwpUnreliable(page)) | |
1570 | return; | |
1571 | ||
6213834c | 1572 | if (hugetlb_vmemmap_restore(h, page)) { |
ad2fa371 MS |
1573 | spin_lock_irq(&hugetlb_lock); |
1574 | /* | |
1575 | * If we cannot allocate vmemmap pages, just refuse to free the | |
1576 | * page and put the page back on the hugetlb free list and treat | |
1577 | * as a surplus page. | |
1578 | */ | |
1579 | add_hugetlb_page(h, page, true); | |
1580 | spin_unlock_irq(&hugetlb_lock); | |
1581 | return; | |
1582 | } | |
1583 | ||
161df60e NH |
1584 | /* |
1585 | * Move PageHWPoison flag from head page to the raw error pages, | |
1586 | * which makes any healthy subpages reusable. | |
1587 | */ | |
1588 | if (unlikely(PageHWPoison(page))) | |
1589 | hugetlb_clear_page_hwpoison(page); | |
1590 | ||
14455eab CL |
1591 | for (i = 0; i < pages_per_huge_page(h); i++) { |
1592 | subpage = nth_page(page, i); | |
dbfee5ae | 1593 | subpage->flags &= ~(1 << PG_locked | 1 << PG_error | |
32f84528 | 1594 | 1 << PG_referenced | 1 << PG_dirty | |
a7407a27 LC |
1595 | 1 << PG_active | 1 << PG_private | |
1596 | 1 << PG_writeback); | |
6af2acb6 | 1597 | } |
a01f4390 MK |
1598 | |
1599 | /* | |
1600 | * Non-gigantic pages demoted from CMA allocated gigantic pages | |
1601 | * need to be given back to CMA in free_gigantic_page. | |
1602 | */ | |
1603 | if (hstate_is_gigantic(h) || | |
1604 | hugetlb_cma_page(page, huge_page_order(h))) { | |
944d9fec LC |
1605 | destroy_compound_gigantic_page(page, huge_page_order(h)); |
1606 | free_gigantic_page(page, huge_page_order(h)); | |
1607 | } else { | |
944d9fec LC |
1608 | __free_pages(page, huge_page_order(h)); |
1609 | } | |
6af2acb6 AL |
1610 | } |
1611 | ||
b65d4adb MS |
1612 | /* |
1613 | * As update_and_free_page() can be called under any context, so we cannot | |
1614 | * use GFP_KERNEL to allocate vmemmap pages. However, we can defer the | |
1615 | * actual freeing in a workqueue to prevent from using GFP_ATOMIC to allocate | |
1616 | * the vmemmap pages. | |
1617 | * | |
1618 | * free_hpage_workfn() locklessly retrieves the linked list of pages to be | |
1619 | * freed and frees them one-by-one. As the page->mapping pointer is going | |
1620 | * to be cleared in free_hpage_workfn() anyway, it is reused as the llist_node | |
1621 | * structure of a lockless linked list of huge pages to be freed. | |
1622 | */ | |
1623 | static LLIST_HEAD(hpage_freelist); | |
1624 | ||
1625 | static void free_hpage_workfn(struct work_struct *work) | |
1626 | { | |
1627 | struct llist_node *node; | |
1628 | ||
1629 | node = llist_del_all(&hpage_freelist); | |
1630 | ||
1631 | while (node) { | |
1632 | struct page *page; | |
1633 | struct hstate *h; | |
1634 | ||
1635 | page = container_of((struct address_space **)node, | |
1636 | struct page, mapping); | |
1637 | node = node->next; | |
1638 | page->mapping = NULL; | |
1639 | /* | |
1640 | * The VM_BUG_ON_PAGE(!PageHuge(page), page) in page_hstate() | |
1641 | * is going to trigger because a previous call to | |
1642 | * remove_hugetlb_page() will set_compound_page_dtor(page, | |
1643 | * NULL_COMPOUND_DTOR), so do not use page_hstate() directly. | |
1644 | */ | |
1645 | h = size_to_hstate(page_size(page)); | |
1646 | ||
1647 | __update_and_free_page(h, page); | |
1648 | ||
1649 | cond_resched(); | |
1650 | } | |
1651 | } | |
1652 | static DECLARE_WORK(free_hpage_work, free_hpage_workfn); | |
1653 | ||
1654 | static inline void flush_free_hpage_work(struct hstate *h) | |
1655 | { | |
6213834c | 1656 | if (hugetlb_vmemmap_optimizable(h)) |
b65d4adb MS |
1657 | flush_work(&free_hpage_work); |
1658 | } | |
1659 | ||
1660 | static void update_and_free_page(struct hstate *h, struct page *page, | |
1661 | bool atomic) | |
1662 | { | |
ad2fa371 | 1663 | if (!HPageVmemmapOptimized(page) || !atomic) { |
b65d4adb MS |
1664 | __update_and_free_page(h, page); |
1665 | return; | |
1666 | } | |
1667 | ||
1668 | /* | |
1669 | * Defer freeing to avoid using GFP_ATOMIC to allocate vmemmap pages. | |
1670 | * | |
1671 | * Only call schedule_work() if hpage_freelist is previously | |
1672 | * empty. Otherwise, schedule_work() had been called but the workfn | |
1673 | * hasn't retrieved the list yet. | |
1674 | */ | |
1675 | if (llist_add((struct llist_node *)&page->mapping, &hpage_freelist)) | |
1676 | schedule_work(&free_hpage_work); | |
1677 | } | |
1678 | ||
10c6ec49 MK |
1679 | static void update_and_free_pages_bulk(struct hstate *h, struct list_head *list) |
1680 | { | |
1681 | struct page *page, *t_page; | |
1682 | ||
1683 | list_for_each_entry_safe(page, t_page, list, lru) { | |
b65d4adb | 1684 | update_and_free_page(h, page, false); |
10c6ec49 MK |
1685 | cond_resched(); |
1686 | } | |
1687 | } | |
1688 | ||
e5ff2159 AK |
1689 | struct hstate *size_to_hstate(unsigned long size) |
1690 | { | |
1691 | struct hstate *h; | |
1692 | ||
1693 | for_each_hstate(h) { | |
1694 | if (huge_page_size(h) == size) | |
1695 | return h; | |
1696 | } | |
1697 | return NULL; | |
1698 | } | |
1699 | ||
db71ef79 | 1700 | void free_huge_page(struct page *page) |
27a85ef1 | 1701 | { |
a5516438 AK |
1702 | /* |
1703 | * Can't pass hstate in here because it is called from the | |
1704 | * compound page destructor. | |
1705 | */ | |
e5ff2159 | 1706 | struct hstate *h = page_hstate(page); |
7893d1d5 | 1707 | int nid = page_to_nid(page); |
d6995da3 | 1708 | struct hugepage_subpool *spool = hugetlb_page_subpool(page); |
07443a85 | 1709 | bool restore_reserve; |
db71ef79 | 1710 | unsigned long flags; |
27a85ef1 | 1711 | |
b4330afb MK |
1712 | VM_BUG_ON_PAGE(page_count(page), page); |
1713 | VM_BUG_ON_PAGE(page_mapcount(page), page); | |
8ace22bc | 1714 | |
d6995da3 | 1715 | hugetlb_set_page_subpool(page, NULL); |
78fbe906 DH |
1716 | if (PageAnon(page)) |
1717 | __ClearPageAnonExclusive(page); | |
8ace22bc | 1718 | page->mapping = NULL; |
d6995da3 MK |
1719 | restore_reserve = HPageRestoreReserve(page); |
1720 | ClearHPageRestoreReserve(page); | |
27a85ef1 | 1721 | |
1c5ecae3 | 1722 | /* |
d6995da3 | 1723 | * If HPageRestoreReserve was set on page, page allocation consumed a |
0919e1b6 MK |
1724 | * reservation. If the page was associated with a subpool, there |
1725 | * would have been a page reserved in the subpool before allocation | |
1726 | * via hugepage_subpool_get_pages(). Since we are 'restoring' the | |
6c26d310 | 1727 | * reservation, do not call hugepage_subpool_put_pages() as this will |
0919e1b6 | 1728 | * remove the reserved page from the subpool. |
1c5ecae3 | 1729 | */ |
0919e1b6 MK |
1730 | if (!restore_reserve) { |
1731 | /* | |
1732 | * A return code of zero implies that the subpool will be | |
1733 | * under its minimum size if the reservation is not restored | |
1734 | * after page is free. Therefore, force restore_reserve | |
1735 | * operation. | |
1736 | */ | |
1737 | if (hugepage_subpool_put_pages(spool, 1) == 0) | |
1738 | restore_reserve = true; | |
1739 | } | |
1c5ecae3 | 1740 | |
db71ef79 | 1741 | spin_lock_irqsave(&hugetlb_lock, flags); |
8f251a3d | 1742 | ClearHPageMigratable(page); |
6d76dcf4 AK |
1743 | hugetlb_cgroup_uncharge_page(hstate_index(h), |
1744 | pages_per_huge_page(h), page); | |
08cf9faf MA |
1745 | hugetlb_cgroup_uncharge_page_rsvd(hstate_index(h), |
1746 | pages_per_huge_page(h), page); | |
07443a85 JK |
1747 | if (restore_reserve) |
1748 | h->resv_huge_pages++; | |
1749 | ||
9157c311 | 1750 | if (HPageTemporary(page)) { |
6eb4e88a | 1751 | remove_hugetlb_page(h, page, false); |
db71ef79 | 1752 | spin_unlock_irqrestore(&hugetlb_lock, flags); |
b65d4adb | 1753 | update_and_free_page(h, page, true); |
ab5ac90a | 1754 | } else if (h->surplus_huge_pages_node[nid]) { |
0edaecfa | 1755 | /* remove the page from active list */ |
6eb4e88a | 1756 | remove_hugetlb_page(h, page, true); |
db71ef79 | 1757 | spin_unlock_irqrestore(&hugetlb_lock, flags); |
b65d4adb | 1758 | update_and_free_page(h, page, true); |
7893d1d5 | 1759 | } else { |
5d3a551c | 1760 | arch_clear_hugepage_flags(page); |
a5516438 | 1761 | enqueue_huge_page(h, page); |
db71ef79 | 1762 | spin_unlock_irqrestore(&hugetlb_lock, flags); |
c77c0a8a | 1763 | } |
c77c0a8a WL |
1764 | } |
1765 | ||
d3d99fcc OS |
1766 | /* |
1767 | * Must be called with the hugetlb lock held | |
1768 | */ | |
1769 | static void __prep_account_new_huge_page(struct hstate *h, int nid) | |
1770 | { | |
1771 | lockdep_assert_held(&hugetlb_lock); | |
1772 | h->nr_huge_pages++; | |
1773 | h->nr_huge_pages_node[nid]++; | |
1774 | } | |
1775 | ||
f41f2ed4 | 1776 | static void __prep_new_huge_page(struct hstate *h, struct page *page) |
b7ba30c6 | 1777 | { |
6213834c | 1778 | hugetlb_vmemmap_optimize(h, page); |
0edaecfa | 1779 | INIT_LIST_HEAD(&page->lru); |
f1e61557 | 1780 | set_compound_page_dtor(page, HUGETLB_PAGE_DTOR); |
ff546117 | 1781 | hugetlb_set_page_subpool(page, NULL); |
9dd540e2 | 1782 | set_hugetlb_cgroup(page, NULL); |
1adc4d41 | 1783 | set_hugetlb_cgroup_rsvd(page, NULL); |
d3d99fcc OS |
1784 | } |
1785 | ||
1786 | static void prep_new_huge_page(struct hstate *h, struct page *page, int nid) | |
1787 | { | |
f41f2ed4 | 1788 | __prep_new_huge_page(h, page); |
db71ef79 | 1789 | spin_lock_irq(&hugetlb_lock); |
d3d99fcc | 1790 | __prep_account_new_huge_page(h, nid); |
db71ef79 | 1791 | spin_unlock_irq(&hugetlb_lock); |
b7ba30c6 AK |
1792 | } |
1793 | ||
34d9e35b MK |
1794 | static bool __prep_compound_gigantic_page(struct page *page, unsigned int order, |
1795 | bool demote) | |
20a0307c | 1796 | { |
7118fc29 | 1797 | int i, j; |
20a0307c | 1798 | int nr_pages = 1 << order; |
14455eab | 1799 | struct page *p; |
20a0307c WF |
1800 | |
1801 | /* we rely on prep_new_huge_page to set the destructor */ | |
1802 | set_compound_order(page, order); | |
de09d31d | 1803 | __SetPageHead(page); |
2b21624f | 1804 | for (i = 0; i < nr_pages; i++) { |
14455eab CL |
1805 | p = nth_page(page, i); |
1806 | ||
ef5a22be AA |
1807 | /* |
1808 | * For gigantic hugepages allocated through bootmem at | |
1809 | * boot, it's safer to be consistent with the not-gigantic | |
1810 | * hugepages and clear the PG_reserved bit from all tail pages | |
7c8de358 | 1811 | * too. Otherwise drivers using get_user_pages() to access tail |
ef5a22be AA |
1812 | * pages may get the reference counting wrong if they see |
1813 | * PG_reserved set on a tail page (despite the head page not | |
1814 | * having PG_reserved set). Enforcing this consistency between | |
1815 | * head and tail pages allows drivers to optimize away a check | |
1816 | * on the head page when they need know if put_page() is needed | |
1817 | * after get_user_pages(). | |
1818 | */ | |
1819 | __ClearPageReserved(p); | |
7118fc29 MK |
1820 | /* |
1821 | * Subtle and very unlikely | |
1822 | * | |
1823 | * Gigantic 'page allocators' such as memblock or cma will | |
1824 | * return a set of pages with each page ref counted. We need | |
1825 | * to turn this set of pages into a compound page with tail | |
1826 | * page ref counts set to zero. Code such as speculative page | |
1827 | * cache adding could take a ref on a 'to be' tail page. | |
1828 | * We need to respect any increased ref count, and only set | |
1829 | * the ref count to zero if count is currently 1. If count | |
416d85ed MK |
1830 | * is not 1, we return an error. An error return indicates |
1831 | * the set of pages can not be converted to a gigantic page. | |
1832 | * The caller who allocated the pages should then discard the | |
1833 | * pages using the appropriate free interface. | |
34d9e35b MK |
1834 | * |
1835 | * In the case of demote, the ref count will be zero. | |
7118fc29 | 1836 | */ |
34d9e35b MK |
1837 | if (!demote) { |
1838 | if (!page_ref_freeze(p, 1)) { | |
1839 | pr_warn("HugeTLB page can not be used due to unexpected inflated ref count\n"); | |
1840 | goto out_error; | |
1841 | } | |
1842 | } else { | |
1843 | VM_BUG_ON_PAGE(page_count(p), p); | |
7118fc29 | 1844 | } |
2b21624f MK |
1845 | if (i != 0) |
1846 | set_compound_head(p, page); | |
20a0307c | 1847 | } |
b4330afb | 1848 | atomic_set(compound_mapcount_ptr(page), -1); |
5291c09b | 1849 | atomic_set(compound_pincount_ptr(page), 0); |
7118fc29 MK |
1850 | return true; |
1851 | ||
1852 | out_error: | |
2b21624f MK |
1853 | /* undo page modifications made above */ |
1854 | for (j = 0; j < i; j++) { | |
14455eab | 1855 | p = nth_page(page, j); |
2b21624f MK |
1856 | if (j != 0) |
1857 | clear_compound_head(p); | |
7118fc29 MK |
1858 | set_page_refcounted(p); |
1859 | } | |
1860 | /* need to clear PG_reserved on remaining tail pages */ | |
14455eab CL |
1861 | for (; j < nr_pages; j++) { |
1862 | p = nth_page(page, j); | |
7118fc29 | 1863 | __ClearPageReserved(p); |
14455eab | 1864 | } |
7118fc29 | 1865 | set_compound_order(page, 0); |
5232c63f | 1866 | #ifdef CONFIG_64BIT |
7118fc29 | 1867 | page[1].compound_nr = 0; |
5232c63f | 1868 | #endif |
7118fc29 MK |
1869 | __ClearPageHead(page); |
1870 | return false; | |
20a0307c WF |
1871 | } |
1872 | ||
34d9e35b MK |
1873 | static bool prep_compound_gigantic_page(struct page *page, unsigned int order) |
1874 | { | |
1875 | return __prep_compound_gigantic_page(page, order, false); | |
1876 | } | |
1877 | ||
8531fc6f MK |
1878 | static bool prep_compound_gigantic_page_for_demote(struct page *page, |
1879 | unsigned int order) | |
1880 | { | |
1881 | return __prep_compound_gigantic_page(page, order, true); | |
1882 | } | |
1883 | ||
7795912c AM |
1884 | /* |
1885 | * PageHuge() only returns true for hugetlbfs pages, but not for normal or | |
1886 | * transparent huge pages. See the PageTransHuge() documentation for more | |
1887 | * details. | |
1888 | */ | |
20a0307c WF |
1889 | int PageHuge(struct page *page) |
1890 | { | |
20a0307c WF |
1891 | if (!PageCompound(page)) |
1892 | return 0; | |
1893 | ||
1894 | page = compound_head(page); | |
f1e61557 | 1895 | return page[1].compound_dtor == HUGETLB_PAGE_DTOR; |
20a0307c | 1896 | } |
43131e14 NH |
1897 | EXPORT_SYMBOL_GPL(PageHuge); |
1898 | ||
27c73ae7 AA |
1899 | /* |
1900 | * PageHeadHuge() only returns true for hugetlbfs head page, but not for | |
1901 | * normal or transparent huge pages. | |
1902 | */ | |
1903 | int PageHeadHuge(struct page *page_head) | |
1904 | { | |
27c73ae7 AA |
1905 | if (!PageHead(page_head)) |
1906 | return 0; | |
1907 | ||
d4af73e3 | 1908 | return page_head[1].compound_dtor == HUGETLB_PAGE_DTOR; |
27c73ae7 | 1909 | } |
4e936ecc | 1910 | EXPORT_SYMBOL_GPL(PageHeadHuge); |
27c73ae7 | 1911 | |
c0d0381a MK |
1912 | /* |
1913 | * Find and lock address space (mapping) in write mode. | |
1914 | * | |
336bf30e MK |
1915 | * Upon entry, the page is locked which means that page_mapping() is |
1916 | * stable. Due to locking order, we can only trylock_write. If we can | |
1917 | * not get the lock, simply return NULL to caller. | |
c0d0381a MK |
1918 | */ |
1919 | struct address_space *hugetlb_page_mapping_lock_write(struct page *hpage) | |
1920 | { | |
336bf30e | 1921 | struct address_space *mapping = page_mapping(hpage); |
c0d0381a | 1922 | |
c0d0381a MK |
1923 | if (!mapping) |
1924 | return mapping; | |
1925 | ||
c0d0381a MK |
1926 | if (i_mmap_trylock_write(mapping)) |
1927 | return mapping; | |
1928 | ||
336bf30e | 1929 | return NULL; |
c0d0381a MK |
1930 | } |
1931 | ||
fe19bd3d | 1932 | pgoff_t hugetlb_basepage_index(struct page *page) |
13d60f4b ZY |
1933 | { |
1934 | struct page *page_head = compound_head(page); | |
1935 | pgoff_t index = page_index(page_head); | |
1936 | unsigned long compound_idx; | |
1937 | ||
13d60f4b ZY |
1938 | if (compound_order(page_head) >= MAX_ORDER) |
1939 | compound_idx = page_to_pfn(page) - page_to_pfn(page_head); | |
1940 | else | |
1941 | compound_idx = page - page_head; | |
1942 | ||
1943 | return (index << compound_order(page_head)) + compound_idx; | |
1944 | } | |
1945 | ||
0c397dae | 1946 | static struct page *alloc_buddy_huge_page(struct hstate *h, |
f60858f9 MK |
1947 | gfp_t gfp_mask, int nid, nodemask_t *nmask, |
1948 | nodemask_t *node_alloc_noretry) | |
1da177e4 | 1949 | { |
af0fb9df | 1950 | int order = huge_page_order(h); |
1da177e4 | 1951 | struct page *page; |
f60858f9 | 1952 | bool alloc_try_hard = true; |
2b21624f | 1953 | bool retry = true; |
f96efd58 | 1954 | |
f60858f9 MK |
1955 | /* |
1956 | * By default we always try hard to allocate the page with | |
1957 | * __GFP_RETRY_MAYFAIL flag. However, if we are allocating pages in | |
1958 | * a loop (to adjust global huge page counts) and previous allocation | |
1959 | * failed, do not continue to try hard on the same node. Use the | |
1960 | * node_alloc_noretry bitmap to manage this state information. | |
1961 | */ | |
1962 | if (node_alloc_noretry && node_isset(nid, *node_alloc_noretry)) | |
1963 | alloc_try_hard = false; | |
1964 | gfp_mask |= __GFP_COMP|__GFP_NOWARN; | |
1965 | if (alloc_try_hard) | |
1966 | gfp_mask |= __GFP_RETRY_MAYFAIL; | |
af0fb9df MH |
1967 | if (nid == NUMA_NO_NODE) |
1968 | nid = numa_mem_id(); | |
2b21624f | 1969 | retry: |
84172f4b | 1970 | page = __alloc_pages(gfp_mask, order, nid, nmask); |
2b21624f MK |
1971 | |
1972 | /* Freeze head page */ | |
1973 | if (page && !page_ref_freeze(page, 1)) { | |
1974 | __free_pages(page, order); | |
1975 | if (retry) { /* retry once */ | |
1976 | retry = false; | |
1977 | goto retry; | |
1978 | } | |
1979 | /* WOW! twice in a row. */ | |
1980 | pr_warn("HugeTLB head page unexpected inflated ref count\n"); | |
1981 | page = NULL; | |
1982 | } | |
1983 | ||
af0fb9df MH |
1984 | if (page) |
1985 | __count_vm_event(HTLB_BUDDY_PGALLOC); | |
1986 | else | |
1987 | __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); | |
63b4613c | 1988 | |
f60858f9 MK |
1989 | /* |
1990 | * If we did not specify __GFP_RETRY_MAYFAIL, but still got a page this | |
1991 | * indicates an overall state change. Clear bit so that we resume | |
1992 | * normal 'try hard' allocations. | |
1993 | */ | |
1994 | if (node_alloc_noretry && page && !alloc_try_hard) | |
1995 | node_clear(nid, *node_alloc_noretry); | |
1996 | ||
1997 | /* | |
1998 | * If we tried hard to get a page but failed, set bit so that | |
1999 | * subsequent attempts will not try as hard until there is an | |
2000 | * overall state change. | |
2001 | */ | |
2002 | if (node_alloc_noretry && !page && alloc_try_hard) | |
2003 | node_set(nid, *node_alloc_noretry); | |
2004 | ||
63b4613c NA |
2005 | return page; |
2006 | } | |
2007 | ||
0c397dae MH |
2008 | /* |
2009 | * Common helper to allocate a fresh hugetlb page. All specific allocators | |
2010 | * should use this function to get new hugetlb pages | |
2b21624f MK |
2011 | * |
2012 | * Note that returned page is 'frozen': ref count of head page and all tail | |
2013 | * pages is zero. | |
0c397dae MH |
2014 | */ |
2015 | static struct page *alloc_fresh_huge_page(struct hstate *h, | |
f60858f9 MK |
2016 | gfp_t gfp_mask, int nid, nodemask_t *nmask, |
2017 | nodemask_t *node_alloc_noretry) | |
0c397dae MH |
2018 | { |
2019 | struct page *page; | |
7118fc29 | 2020 | bool retry = false; |
0c397dae | 2021 | |
7118fc29 | 2022 | retry: |
0c397dae MH |
2023 | if (hstate_is_gigantic(h)) |
2024 | page = alloc_gigantic_page(h, gfp_mask, nid, nmask); | |
2025 | else | |
2026 | page = alloc_buddy_huge_page(h, gfp_mask, | |
f60858f9 | 2027 | nid, nmask, node_alloc_noretry); |
0c397dae MH |
2028 | if (!page) |
2029 | return NULL; | |
2030 | ||
7118fc29 MK |
2031 | if (hstate_is_gigantic(h)) { |
2032 | if (!prep_compound_gigantic_page(page, huge_page_order(h))) { | |
2033 | /* | |
2034 | * Rare failure to convert pages to compound page. | |
2035 | * Free pages and try again - ONCE! | |
2036 | */ | |
2037 | free_gigantic_page(page, huge_page_order(h)); | |
2038 | if (!retry) { | |
2039 | retry = true; | |
2040 | goto retry; | |
2041 | } | |
7118fc29 MK |
2042 | return NULL; |
2043 | } | |
2044 | } | |
0c397dae MH |
2045 | prep_new_huge_page(h, page, page_to_nid(page)); |
2046 | ||
2047 | return page; | |
2048 | } | |
2049 | ||
af0fb9df MH |
2050 | /* |
2051 | * Allocates a fresh page to the hugetlb allocator pool in the node interleaved | |
2052 | * manner. | |
2053 | */ | |
f60858f9 MK |
2054 | static int alloc_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed, |
2055 | nodemask_t *node_alloc_noretry) | |
b2261026 JK |
2056 | { |
2057 | struct page *page; | |
2058 | int nr_nodes, node; | |
af0fb9df | 2059 | gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE; |
b2261026 JK |
2060 | |
2061 | for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { | |
f60858f9 MK |
2062 | page = alloc_fresh_huge_page(h, gfp_mask, node, nodes_allowed, |
2063 | node_alloc_noretry); | |
af0fb9df | 2064 | if (page) |
b2261026 | 2065 | break; |
b2261026 JK |
2066 | } |
2067 | ||
af0fb9df MH |
2068 | if (!page) |
2069 | return 0; | |
b2261026 | 2070 | |
2b21624f | 2071 | free_huge_page(page); /* free it into the hugepage allocator */ |
af0fb9df MH |
2072 | |
2073 | return 1; | |
b2261026 JK |
2074 | } |
2075 | ||
e8c5c824 | 2076 | /* |
10c6ec49 MK |
2077 | * Remove huge page from pool from next node to free. Attempt to keep |
2078 | * persistent huge pages more or less balanced over allowed nodes. | |
2079 | * This routine only 'removes' the hugetlb page. The caller must make | |
2080 | * an additional call to free the page to low level allocators. | |
e8c5c824 LS |
2081 | * Called with hugetlb_lock locked. |
2082 | */ | |
10c6ec49 MK |
2083 | static struct page *remove_pool_huge_page(struct hstate *h, |
2084 | nodemask_t *nodes_allowed, | |
2085 | bool acct_surplus) | |
e8c5c824 | 2086 | { |
b2261026 | 2087 | int nr_nodes, node; |
10c6ec49 | 2088 | struct page *page = NULL; |
e8c5c824 | 2089 | |
9487ca60 | 2090 | lockdep_assert_held(&hugetlb_lock); |
b2261026 | 2091 | for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { |
685f3457 LS |
2092 | /* |
2093 | * If we're returning unused surplus pages, only examine | |
2094 | * nodes with surplus pages. | |
2095 | */ | |
b2261026 JK |
2096 | if ((!acct_surplus || h->surplus_huge_pages_node[node]) && |
2097 | !list_empty(&h->hugepage_freelists[node])) { | |
10c6ec49 | 2098 | page = list_entry(h->hugepage_freelists[node].next, |
e8c5c824 | 2099 | struct page, lru); |
6eb4e88a | 2100 | remove_hugetlb_page(h, page, acct_surplus); |
9a76db09 | 2101 | break; |
e8c5c824 | 2102 | } |
b2261026 | 2103 | } |
e8c5c824 | 2104 | |
10c6ec49 | 2105 | return page; |
e8c5c824 LS |
2106 | } |
2107 | ||
c8721bbb NH |
2108 | /* |
2109 | * Dissolve a given free hugepage into free buddy pages. This function does | |
faf53def NH |
2110 | * nothing for in-use hugepages and non-hugepages. |
2111 | * This function returns values like below: | |
2112 | * | |
ad2fa371 MS |
2113 | * -ENOMEM: failed to allocate vmemmap pages to free the freed hugepages |
2114 | * when the system is under memory pressure and the feature of | |
2115 | * freeing unused vmemmap pages associated with each hugetlb page | |
2116 | * is enabled. | |
2117 | * -EBUSY: failed to dissolved free hugepages or the hugepage is in-use | |
2118 | * (allocated or reserved.) | |
2119 | * 0: successfully dissolved free hugepages or the page is not a | |
2120 | * hugepage (considered as already dissolved) | |
c8721bbb | 2121 | */ |
c3114a84 | 2122 | int dissolve_free_huge_page(struct page *page) |
c8721bbb | 2123 | { |
6bc9b564 | 2124 | int rc = -EBUSY; |
082d5b6b | 2125 | |
7ffddd49 | 2126 | retry: |
faf53def NH |
2127 | /* Not to disrupt normal path by vainly holding hugetlb_lock */ |
2128 | if (!PageHuge(page)) | |
2129 | return 0; | |
2130 | ||
db71ef79 | 2131 | spin_lock_irq(&hugetlb_lock); |
faf53def NH |
2132 | if (!PageHuge(page)) { |
2133 | rc = 0; | |
2134 | goto out; | |
2135 | } | |
2136 | ||
2137 | if (!page_count(page)) { | |
2247bb33 GS |
2138 | struct page *head = compound_head(page); |
2139 | struct hstate *h = page_hstate(head); | |
8346d69d | 2140 | if (!available_huge_pages(h)) |
082d5b6b | 2141 | goto out; |
7ffddd49 MS |
2142 | |
2143 | /* | |
2144 | * We should make sure that the page is already on the free list | |
2145 | * when it is dissolved. | |
2146 | */ | |
6c037149 | 2147 | if (unlikely(!HPageFreed(head))) { |
db71ef79 | 2148 | spin_unlock_irq(&hugetlb_lock); |
7ffddd49 MS |
2149 | cond_resched(); |
2150 | ||
2151 | /* | |
2152 | * Theoretically, we should return -EBUSY when we | |
2153 | * encounter this race. In fact, we have a chance | |
2154 | * to successfully dissolve the page if we do a | |
2155 | * retry. Because the race window is quite small. | |
2156 | * If we seize this opportunity, it is an optimization | |
2157 | * for increasing the success rate of dissolving page. | |
2158 | */ | |
2159 | goto retry; | |
2160 | } | |
2161 | ||
0c5da357 | 2162 | remove_hugetlb_page(h, head, false); |
c1470b33 | 2163 | h->max_huge_pages--; |
db71ef79 | 2164 | spin_unlock_irq(&hugetlb_lock); |
ad2fa371 MS |
2165 | |
2166 | /* | |
2167 | * Normally update_and_free_page will allocate required vmemmmap | |
2168 | * before freeing the page. update_and_free_page will fail to | |
2169 | * free the page if it can not allocate required vmemmap. We | |
2170 | * need to adjust max_huge_pages if the page is not freed. | |
2171 | * Attempt to allocate vmemmmap here so that we can take | |
2172 | * appropriate action on failure. | |
2173 | */ | |
6213834c | 2174 | rc = hugetlb_vmemmap_restore(h, head); |
ad2fa371 | 2175 | if (!rc) { |
ad2fa371 MS |
2176 | update_and_free_page(h, head, false); |
2177 | } else { | |
2178 | spin_lock_irq(&hugetlb_lock); | |
2179 | add_hugetlb_page(h, head, false); | |
2180 | h->max_huge_pages++; | |
2181 | spin_unlock_irq(&hugetlb_lock); | |
2182 | } | |
2183 | ||
2184 | return rc; | |
c8721bbb | 2185 | } |
082d5b6b | 2186 | out: |
db71ef79 | 2187 | spin_unlock_irq(&hugetlb_lock); |
082d5b6b | 2188 | return rc; |
c8721bbb NH |
2189 | } |
2190 | ||
2191 | /* | |
2192 | * Dissolve free hugepages in a given pfn range. Used by memory hotplug to | |
2193 | * make specified memory blocks removable from the system. | |
2247bb33 GS |
2194 | * Note that this will dissolve a free gigantic hugepage completely, if any |
2195 | * part of it lies within the given range. | |
082d5b6b GS |
2196 | * Also note that if dissolve_free_huge_page() returns with an error, all |
2197 | * free hugepages that were dissolved before that error are lost. | |
c8721bbb | 2198 | */ |
082d5b6b | 2199 | int dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn) |
c8721bbb | 2200 | { |
c8721bbb | 2201 | unsigned long pfn; |
eb03aa00 | 2202 | struct page *page; |
082d5b6b | 2203 | int rc = 0; |
dc2628f3 MS |
2204 | unsigned int order; |
2205 | struct hstate *h; | |
c8721bbb | 2206 | |
d0177639 | 2207 | if (!hugepages_supported()) |
082d5b6b | 2208 | return rc; |
d0177639 | 2209 | |
dc2628f3 MS |
2210 | order = huge_page_order(&default_hstate); |
2211 | for_each_hstate(h) | |
2212 | order = min(order, huge_page_order(h)); | |
2213 | ||
2214 | for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << order) { | |
eb03aa00 | 2215 | page = pfn_to_page(pfn); |
faf53def NH |
2216 | rc = dissolve_free_huge_page(page); |
2217 | if (rc) | |
2218 | break; | |
eb03aa00 | 2219 | } |
082d5b6b GS |
2220 | |
2221 | return rc; | |
c8721bbb NH |
2222 | } |
2223 | ||
ab5ac90a MH |
2224 | /* |
2225 | * Allocates a fresh surplus page from the page allocator. | |
2226 | */ | |
0c397dae | 2227 | static struct page *alloc_surplus_huge_page(struct hstate *h, gfp_t gfp_mask, |
2b21624f | 2228 | int nid, nodemask_t *nmask) |
7893d1d5 | 2229 | { |
9980d744 | 2230 | struct page *page = NULL; |
7893d1d5 | 2231 | |
bae7f4ae | 2232 | if (hstate_is_gigantic(h)) |
aa888a74 AK |
2233 | return NULL; |
2234 | ||
db71ef79 | 2235 | spin_lock_irq(&hugetlb_lock); |
9980d744 MH |
2236 | if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) |
2237 | goto out_unlock; | |
db71ef79 | 2238 | spin_unlock_irq(&hugetlb_lock); |
d1c3fb1f | 2239 | |
f60858f9 | 2240 | page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL); |
9980d744 | 2241 | if (!page) |
0c397dae | 2242 | return NULL; |
d1c3fb1f | 2243 | |
db71ef79 | 2244 | spin_lock_irq(&hugetlb_lock); |
9980d744 MH |
2245 | /* |
2246 | * We could have raced with the pool size change. | |
2247 | * Double check that and simply deallocate the new page | |
2248 | * if we would end up overcommiting the surpluses. Abuse | |
2249 | * temporary page to workaround the nasty free_huge_page | |
2250 | * codeflow | |
2251 | */ | |
2252 | if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) { | |
9157c311 | 2253 | SetHPageTemporary(page); |
db71ef79 | 2254 | spin_unlock_irq(&hugetlb_lock); |
2b21624f | 2255 | free_huge_page(page); |
2bf753e6 | 2256 | return NULL; |
7893d1d5 | 2257 | } |
9980d744 | 2258 | |
b65a4eda MK |
2259 | h->surplus_huge_pages++; |
2260 | h->surplus_huge_pages_node[page_to_nid(page)]++; | |
2261 | ||
9980d744 | 2262 | out_unlock: |
db71ef79 | 2263 | spin_unlock_irq(&hugetlb_lock); |
7893d1d5 AL |
2264 | |
2265 | return page; | |
2266 | } | |
2267 | ||
bbe88753 | 2268 | static struct page *alloc_migrate_huge_page(struct hstate *h, gfp_t gfp_mask, |
9a4e9f3b | 2269 | int nid, nodemask_t *nmask) |
ab5ac90a MH |
2270 | { |
2271 | struct page *page; | |
2272 | ||
2273 | if (hstate_is_gigantic(h)) | |
2274 | return NULL; | |
2275 | ||
f60858f9 | 2276 | page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL); |
ab5ac90a MH |
2277 | if (!page) |
2278 | return NULL; | |
2279 | ||
2b21624f MK |
2280 | /* fresh huge pages are frozen */ |
2281 | set_page_refcounted(page); | |
2282 | ||
ab5ac90a MH |
2283 | /* |
2284 | * We do not account these pages as surplus because they are only | |
2285 | * temporary and will be released properly on the last reference | |
2286 | */ | |
9157c311 | 2287 | SetHPageTemporary(page); |
ab5ac90a MH |
2288 | |
2289 | return page; | |
2290 | } | |
2291 | ||
099730d6 DH |
2292 | /* |
2293 | * Use the VMA's mpolicy to allocate a huge page from the buddy. | |
2294 | */ | |
e0ec90ee | 2295 | static |
0c397dae | 2296 | struct page *alloc_buddy_huge_page_with_mpol(struct hstate *h, |
099730d6 DH |
2297 | struct vm_area_struct *vma, unsigned long addr) |
2298 | { | |
cfcaa66f | 2299 | struct page *page = NULL; |
aaf14e40 MH |
2300 | struct mempolicy *mpol; |
2301 | gfp_t gfp_mask = htlb_alloc_mask(h); | |
2302 | int nid; | |
2303 | nodemask_t *nodemask; | |
2304 | ||
2305 | nid = huge_node(vma, addr, gfp_mask, &mpol, &nodemask); | |
cfcaa66f BW |
2306 | if (mpol_is_preferred_many(mpol)) { |
2307 | gfp_t gfp = gfp_mask | __GFP_NOWARN; | |
2308 | ||
2309 | gfp &= ~(__GFP_DIRECT_RECLAIM | __GFP_NOFAIL); | |
2b21624f | 2310 | page = alloc_surplus_huge_page(h, gfp, nid, nodemask); |
aaf14e40 | 2311 | |
cfcaa66f BW |
2312 | /* Fallback to all nodes if page==NULL */ |
2313 | nodemask = NULL; | |
2314 | } | |
2315 | ||
2316 | if (!page) | |
2b21624f | 2317 | page = alloc_surplus_huge_page(h, gfp_mask, nid, nodemask); |
cfcaa66f | 2318 | mpol_cond_put(mpol); |
aaf14e40 | 2319 | return page; |
099730d6 DH |
2320 | } |
2321 | ||
ab5ac90a | 2322 | /* page migration callback function */ |
3e59fcb0 | 2323 | struct page *alloc_huge_page_nodemask(struct hstate *h, int preferred_nid, |
d92bbc27 | 2324 | nodemask_t *nmask, gfp_t gfp_mask) |
4db9b2ef | 2325 | { |
db71ef79 | 2326 | spin_lock_irq(&hugetlb_lock); |
8346d69d | 2327 | if (available_huge_pages(h)) { |
3e59fcb0 MH |
2328 | struct page *page; |
2329 | ||
2330 | page = dequeue_huge_page_nodemask(h, gfp_mask, preferred_nid, nmask); | |
2331 | if (page) { | |
db71ef79 | 2332 | spin_unlock_irq(&hugetlb_lock); |
3e59fcb0 | 2333 | return page; |
4db9b2ef MH |
2334 | } |
2335 | } | |
db71ef79 | 2336 | spin_unlock_irq(&hugetlb_lock); |
4db9b2ef | 2337 | |
0c397dae | 2338 | return alloc_migrate_huge_page(h, gfp_mask, preferred_nid, nmask); |
4db9b2ef MH |
2339 | } |
2340 | ||
ebd63723 | 2341 | /* mempolicy aware migration callback */ |
389c8178 MH |
2342 | struct page *alloc_huge_page_vma(struct hstate *h, struct vm_area_struct *vma, |
2343 | unsigned long address) | |
ebd63723 MH |
2344 | { |
2345 | struct mempolicy *mpol; | |
2346 | nodemask_t *nodemask; | |
2347 | struct page *page; | |
ebd63723 MH |
2348 | gfp_t gfp_mask; |
2349 | int node; | |
2350 | ||
ebd63723 MH |
2351 | gfp_mask = htlb_alloc_mask(h); |
2352 | node = huge_node(vma, address, gfp_mask, &mpol, &nodemask); | |
d92bbc27 | 2353 | page = alloc_huge_page_nodemask(h, node, nodemask, gfp_mask); |
ebd63723 MH |
2354 | mpol_cond_put(mpol); |
2355 | ||
2356 | return page; | |
2357 | } | |
2358 | ||
e4e574b7 | 2359 | /* |
25985edc | 2360 | * Increase the hugetlb pool such that it can accommodate a reservation |
e4e574b7 AL |
2361 | * of size 'delta'. |
2362 | */ | |
0a4f3d1b | 2363 | static int gather_surplus_pages(struct hstate *h, long delta) |
1b2a1e7b | 2364 | __must_hold(&hugetlb_lock) |
e4e574b7 | 2365 | { |
34665341 | 2366 | LIST_HEAD(surplus_list); |
e4e574b7 | 2367 | struct page *page, *tmp; |
0a4f3d1b LX |
2368 | int ret; |
2369 | long i; | |
2370 | long needed, allocated; | |
28073b02 | 2371 | bool alloc_ok = true; |
e4e574b7 | 2372 | |
9487ca60 | 2373 | lockdep_assert_held(&hugetlb_lock); |
a5516438 | 2374 | needed = (h->resv_huge_pages + delta) - h->free_huge_pages; |
ac09b3a1 | 2375 | if (needed <= 0) { |
a5516438 | 2376 | h->resv_huge_pages += delta; |
e4e574b7 | 2377 | return 0; |
ac09b3a1 | 2378 | } |
e4e574b7 AL |
2379 | |
2380 | allocated = 0; | |
e4e574b7 AL |
2381 | |
2382 | ret = -ENOMEM; | |
2383 | retry: | |
db71ef79 | 2384 | spin_unlock_irq(&hugetlb_lock); |
e4e574b7 | 2385 | for (i = 0; i < needed; i++) { |
0c397dae | 2386 | page = alloc_surplus_huge_page(h, htlb_alloc_mask(h), |
2b21624f | 2387 | NUMA_NO_NODE, NULL); |
28073b02 HD |
2388 | if (!page) { |
2389 | alloc_ok = false; | |
2390 | break; | |
2391 | } | |
e4e574b7 | 2392 | list_add(&page->lru, &surplus_list); |
69ed779a | 2393 | cond_resched(); |
e4e574b7 | 2394 | } |
28073b02 | 2395 | allocated += i; |
e4e574b7 AL |
2396 | |
2397 | /* | |
2398 | * After retaking hugetlb_lock, we need to recalculate 'needed' | |
2399 | * because either resv_huge_pages or free_huge_pages may have changed. | |
2400 | */ | |
db71ef79 | 2401 | spin_lock_irq(&hugetlb_lock); |
a5516438 AK |
2402 | needed = (h->resv_huge_pages + delta) - |
2403 | (h->free_huge_pages + allocated); | |
28073b02 HD |
2404 | if (needed > 0) { |
2405 | if (alloc_ok) | |
2406 | goto retry; | |
2407 | /* | |
2408 | * We were not able to allocate enough pages to | |
2409 | * satisfy the entire reservation so we free what | |
2410 | * we've allocated so far. | |
2411 | */ | |
2412 | goto free; | |
2413 | } | |
e4e574b7 AL |
2414 | /* |
2415 | * The surplus_list now contains _at_least_ the number of extra pages | |
25985edc | 2416 | * needed to accommodate the reservation. Add the appropriate number |
e4e574b7 | 2417 | * of pages to the hugetlb pool and free the extras back to the buddy |
ac09b3a1 AL |
2418 | * allocator. Commit the entire reservation here to prevent another |
2419 | * process from stealing the pages as they are added to the pool but | |
2420 | * before they are reserved. | |
e4e574b7 AL |
2421 | */ |
2422 | needed += allocated; | |
a5516438 | 2423 | h->resv_huge_pages += delta; |
e4e574b7 | 2424 | ret = 0; |
a9869b83 | 2425 | |
19fc3f0a | 2426 | /* Free the needed pages to the hugetlb pool */ |
e4e574b7 | 2427 | list_for_each_entry_safe(page, tmp, &surplus_list, lru) { |
19fc3f0a AL |
2428 | if ((--needed) < 0) |
2429 | break; | |
b65a4eda | 2430 | /* Add the page to the hugetlb allocator */ |
a5516438 | 2431 | enqueue_huge_page(h, page); |
19fc3f0a | 2432 | } |
28073b02 | 2433 | free: |
db71ef79 | 2434 | spin_unlock_irq(&hugetlb_lock); |
19fc3f0a | 2435 | |
b65a4eda MK |
2436 | /* |
2437 | * Free unnecessary surplus pages to the buddy allocator. | |
2438 | * Pages have no ref count, call free_huge_page directly. | |
2439 | */ | |
c0d934ba | 2440 | list_for_each_entry_safe(page, tmp, &surplus_list, lru) |
b65a4eda | 2441 | free_huge_page(page); |
db71ef79 | 2442 | spin_lock_irq(&hugetlb_lock); |
e4e574b7 AL |
2443 | |
2444 | return ret; | |
2445 | } | |
2446 | ||
2447 | /* | |
e5bbc8a6 MK |
2448 | * This routine has two main purposes: |
2449 | * 1) Decrement the reservation count (resv_huge_pages) by the value passed | |
2450 | * in unused_resv_pages. This corresponds to the prior adjustments made | |
2451 | * to the associated reservation map. | |
2452 | * 2) Free any unused surplus pages that may have been allocated to satisfy | |
2453 | * the reservation. As many as unused_resv_pages may be freed. | |
e4e574b7 | 2454 | */ |
a5516438 AK |
2455 | static void return_unused_surplus_pages(struct hstate *h, |
2456 | unsigned long unused_resv_pages) | |
e4e574b7 | 2457 | { |
e4e574b7 | 2458 | unsigned long nr_pages; |
10c6ec49 MK |
2459 | struct page *page; |
2460 | LIST_HEAD(page_list); | |
2461 | ||
9487ca60 | 2462 | lockdep_assert_held(&hugetlb_lock); |
10c6ec49 MK |
2463 | /* Uncommit the reservation */ |
2464 | h->resv_huge_pages -= unused_resv_pages; | |
e4e574b7 | 2465 | |
c0531714 | 2466 | if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) |
e5bbc8a6 | 2467 | goto out; |
aa888a74 | 2468 | |
e5bbc8a6 MK |
2469 | /* |
2470 | * Part (or even all) of the reservation could have been backed | |
2471 | * by pre-allocated pages. Only free surplus pages. | |
2472 | */ | |
a5516438 | 2473 | nr_pages = min(unused_resv_pages, h->surplus_huge_pages); |
e4e574b7 | 2474 | |
685f3457 LS |
2475 | /* |
2476 | * We want to release as many surplus pages as possible, spread | |
9b5e5d0f LS |
2477 | * evenly across all nodes with memory. Iterate across these nodes |
2478 | * until we can no longer free unreserved surplus pages. This occurs | |
2479 | * when the nodes with surplus pages have no free pages. | |
10c6ec49 | 2480 | * remove_pool_huge_page() will balance the freed pages across the |
9b5e5d0f | 2481 | * on-line nodes with memory and will handle the hstate accounting. |
685f3457 LS |
2482 | */ |
2483 | while (nr_pages--) { | |
10c6ec49 MK |
2484 | page = remove_pool_huge_page(h, &node_states[N_MEMORY], 1); |
2485 | if (!page) | |
e5bbc8a6 | 2486 | goto out; |
10c6ec49 MK |
2487 | |
2488 | list_add(&page->lru, &page_list); | |
e4e574b7 | 2489 | } |
e5bbc8a6 MK |
2490 | |
2491 | out: | |
db71ef79 | 2492 | spin_unlock_irq(&hugetlb_lock); |
10c6ec49 | 2493 | update_and_free_pages_bulk(h, &page_list); |
db71ef79 | 2494 | spin_lock_irq(&hugetlb_lock); |
e4e574b7 AL |
2495 | } |
2496 | ||
5e911373 | 2497 | |
c37f9fb1 | 2498 | /* |
feba16e2 | 2499 | * vma_needs_reservation, vma_commit_reservation and vma_end_reservation |
5e911373 | 2500 | * are used by the huge page allocation routines to manage reservations. |
cf3ad20b MK |
2501 | * |
2502 | * vma_needs_reservation is called to determine if the huge page at addr | |
2503 | * within the vma has an associated reservation. If a reservation is | |
2504 | * needed, the value 1 is returned. The caller is then responsible for | |
2505 | * managing the global reservation and subpool usage counts. After | |
2506 | * the huge page has been allocated, vma_commit_reservation is called | |
feba16e2 MK |
2507 | * to add the page to the reservation map. If the page allocation fails, |
2508 | * the reservation must be ended instead of committed. vma_end_reservation | |
2509 | * is called in such cases. | |
cf3ad20b MK |
2510 | * |
2511 | * In the normal case, vma_commit_reservation returns the same value | |
2512 | * as the preceding vma_needs_reservation call. The only time this | |
2513 | * is not the case is if a reserve map was changed between calls. It | |
2514 | * is the responsibility of the caller to notice the difference and | |
2515 | * take appropriate action. | |
96b96a96 MK |
2516 | * |
2517 | * vma_add_reservation is used in error paths where a reservation must | |
2518 | * be restored when a newly allocated huge page must be freed. It is | |
2519 | * to be called after calling vma_needs_reservation to determine if a | |
2520 | * reservation exists. | |
846be085 MK |
2521 | * |
2522 | * vma_del_reservation is used in error paths where an entry in the reserve | |
2523 | * map was created during huge page allocation and must be removed. It is to | |
2524 | * be called after calling vma_needs_reservation to determine if a reservation | |
2525 | * exists. | |
c37f9fb1 | 2526 | */ |
5e911373 MK |
2527 | enum vma_resv_mode { |
2528 | VMA_NEEDS_RESV, | |
2529 | VMA_COMMIT_RESV, | |
feba16e2 | 2530 | VMA_END_RESV, |
96b96a96 | 2531 | VMA_ADD_RESV, |
846be085 | 2532 | VMA_DEL_RESV, |
5e911373 | 2533 | }; |
cf3ad20b MK |
2534 | static long __vma_reservation_common(struct hstate *h, |
2535 | struct vm_area_struct *vma, unsigned long addr, | |
5e911373 | 2536 | enum vma_resv_mode mode) |
c37f9fb1 | 2537 | { |
4e35f483 JK |
2538 | struct resv_map *resv; |
2539 | pgoff_t idx; | |
cf3ad20b | 2540 | long ret; |
0db9d74e | 2541 | long dummy_out_regions_needed; |
c37f9fb1 | 2542 | |
4e35f483 JK |
2543 | resv = vma_resv_map(vma); |
2544 | if (!resv) | |
84afd99b | 2545 | return 1; |
c37f9fb1 | 2546 | |
4e35f483 | 2547 | idx = vma_hugecache_offset(h, vma, addr); |
5e911373 MK |
2548 | switch (mode) { |
2549 | case VMA_NEEDS_RESV: | |
0db9d74e MA |
2550 | ret = region_chg(resv, idx, idx + 1, &dummy_out_regions_needed); |
2551 | /* We assume that vma_reservation_* routines always operate on | |
2552 | * 1 page, and that adding to resv map a 1 page entry can only | |
2553 | * ever require 1 region. | |
2554 | */ | |
2555 | VM_BUG_ON(dummy_out_regions_needed != 1); | |
5e911373 MK |
2556 | break; |
2557 | case VMA_COMMIT_RESV: | |
075a61d0 | 2558 | ret = region_add(resv, idx, idx + 1, 1, NULL, NULL); |
0db9d74e MA |
2559 | /* region_add calls of range 1 should never fail. */ |
2560 | VM_BUG_ON(ret < 0); | |
5e911373 | 2561 | break; |
feba16e2 | 2562 | case VMA_END_RESV: |
0db9d74e | 2563 | region_abort(resv, idx, idx + 1, 1); |
5e911373 MK |
2564 | ret = 0; |
2565 | break; | |
96b96a96 | 2566 | case VMA_ADD_RESV: |
0db9d74e | 2567 | if (vma->vm_flags & VM_MAYSHARE) { |
075a61d0 | 2568 | ret = region_add(resv, idx, idx + 1, 1, NULL, NULL); |
0db9d74e MA |
2569 | /* region_add calls of range 1 should never fail. */ |
2570 | VM_BUG_ON(ret < 0); | |
2571 | } else { | |
2572 | region_abort(resv, idx, idx + 1, 1); | |
96b96a96 MK |
2573 | ret = region_del(resv, idx, idx + 1); |
2574 | } | |
2575 | break; | |
846be085 MK |
2576 | case VMA_DEL_RESV: |
2577 | if (vma->vm_flags & VM_MAYSHARE) { | |
2578 | region_abort(resv, idx, idx + 1, 1); | |
2579 | ret = region_del(resv, idx, idx + 1); | |
2580 | } else { | |
2581 | ret = region_add(resv, idx, idx + 1, 1, NULL, NULL); | |
2582 | /* region_add calls of range 1 should never fail. */ | |
2583 | VM_BUG_ON(ret < 0); | |
2584 | } | |
2585 | break; | |
5e911373 MK |
2586 | default: |
2587 | BUG(); | |
2588 | } | |
84afd99b | 2589 | |
846be085 | 2590 | if (vma->vm_flags & VM_MAYSHARE || mode == VMA_DEL_RESV) |
cf3ad20b | 2591 | return ret; |
bf3d12b9 ML |
2592 | /* |
2593 | * We know private mapping must have HPAGE_RESV_OWNER set. | |
2594 | * | |
2595 | * In most cases, reserves always exist for private mappings. | |
2596 | * However, a file associated with mapping could have been | |
2597 | * hole punched or truncated after reserves were consumed. | |
2598 | * As subsequent fault on such a range will not use reserves. | |
2599 | * Subtle - The reserve map for private mappings has the | |
2600 | * opposite meaning than that of shared mappings. If NO | |
2601 | * entry is in the reserve map, it means a reservation exists. | |
2602 | * If an entry exists in the reserve map, it means the | |
2603 | * reservation has already been consumed. As a result, the | |
2604 | * return value of this routine is the opposite of the | |
2605 | * value returned from reserve map manipulation routines above. | |
2606 | */ | |
2607 | if (ret > 0) | |
2608 | return 0; | |
2609 | if (ret == 0) | |
2610 | return 1; | |
2611 | return ret; | |
c37f9fb1 | 2612 | } |
cf3ad20b MK |
2613 | |
2614 | static long vma_needs_reservation(struct hstate *h, | |
a5516438 | 2615 | struct vm_area_struct *vma, unsigned long addr) |
c37f9fb1 | 2616 | { |
5e911373 | 2617 | return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV); |
cf3ad20b | 2618 | } |
84afd99b | 2619 | |
cf3ad20b MK |
2620 | static long vma_commit_reservation(struct hstate *h, |
2621 | struct vm_area_struct *vma, unsigned long addr) | |
2622 | { | |
5e911373 MK |
2623 | return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV); |
2624 | } | |
2625 | ||
feba16e2 | 2626 | static void vma_end_reservation(struct hstate *h, |
5e911373 MK |
2627 | struct vm_area_struct *vma, unsigned long addr) |
2628 | { | |
feba16e2 | 2629 | (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV); |
c37f9fb1 AW |
2630 | } |
2631 | ||
96b96a96 MK |
2632 | static long vma_add_reservation(struct hstate *h, |
2633 | struct vm_area_struct *vma, unsigned long addr) | |
2634 | { | |
2635 | return __vma_reservation_common(h, vma, addr, VMA_ADD_RESV); | |
2636 | } | |
2637 | ||
846be085 MK |
2638 | static long vma_del_reservation(struct hstate *h, |
2639 | struct vm_area_struct *vma, unsigned long addr) | |
2640 | { | |
2641 | return __vma_reservation_common(h, vma, addr, VMA_DEL_RESV); | |
2642 | } | |
2643 | ||
96b96a96 | 2644 | /* |
846be085 MK |
2645 | * This routine is called to restore reservation information on error paths. |
2646 | * It should ONLY be called for pages allocated via alloc_huge_page(), and | |
2647 | * the hugetlb mutex should remain held when calling this routine. | |
2648 | * | |
2649 | * It handles two specific cases: | |
2650 | * 1) A reservation was in place and the page consumed the reservation. | |
2651 | * HPageRestoreReserve is set in the page. | |
2652 | * 2) No reservation was in place for the page, so HPageRestoreReserve is | |
2653 | * not set. However, alloc_huge_page always updates the reserve map. | |
2654 | * | |
2655 | * In case 1, free_huge_page later in the error path will increment the | |
2656 | * global reserve count. But, free_huge_page does not have enough context | |
2657 | * to adjust the reservation map. This case deals primarily with private | |
2658 | * mappings. Adjust the reserve map here to be consistent with global | |
2659 | * reserve count adjustments to be made by free_huge_page. Make sure the | |
2660 | * reserve map indicates there is a reservation present. | |
2661 | * | |
2662 | * In case 2, simply undo reserve map modifications done by alloc_huge_page. | |
96b96a96 | 2663 | */ |
846be085 MK |
2664 | void restore_reserve_on_error(struct hstate *h, struct vm_area_struct *vma, |
2665 | unsigned long address, struct page *page) | |
96b96a96 | 2666 | { |
846be085 | 2667 | long rc = vma_needs_reservation(h, vma, address); |
96b96a96 | 2668 | |
846be085 MK |
2669 | if (HPageRestoreReserve(page)) { |
2670 | if (unlikely(rc < 0)) | |
96b96a96 MK |
2671 | /* |
2672 | * Rare out of memory condition in reserve map | |
d6995da3 | 2673 | * manipulation. Clear HPageRestoreReserve so that |
96b96a96 MK |
2674 | * global reserve count will not be incremented |
2675 | * by free_huge_page. This will make it appear | |
2676 | * as though the reservation for this page was | |
2677 | * consumed. This may prevent the task from | |
2678 | * faulting in the page at a later time. This | |
2679 | * is better than inconsistent global huge page | |
2680 | * accounting of reserve counts. | |
2681 | */ | |
d6995da3 | 2682 | ClearHPageRestoreReserve(page); |
846be085 MK |
2683 | else if (rc) |
2684 | (void)vma_add_reservation(h, vma, address); | |
2685 | else | |
2686 | vma_end_reservation(h, vma, address); | |
2687 | } else { | |
2688 | if (!rc) { | |
2689 | /* | |
2690 | * This indicates there is an entry in the reserve map | |
c7b1850d | 2691 | * not added by alloc_huge_page. We know it was added |
846be085 MK |
2692 | * before the alloc_huge_page call, otherwise |
2693 | * HPageRestoreReserve would be set on the page. | |
2694 | * Remove the entry so that a subsequent allocation | |
2695 | * does not consume a reservation. | |
2696 | */ | |
2697 | rc = vma_del_reservation(h, vma, address); | |
2698 | if (rc < 0) | |
96b96a96 | 2699 | /* |
846be085 MK |
2700 | * VERY rare out of memory condition. Since |
2701 | * we can not delete the entry, set | |
2702 | * HPageRestoreReserve so that the reserve | |
2703 | * count will be incremented when the page | |
2704 | * is freed. This reserve will be consumed | |
2705 | * on a subsequent allocation. | |
96b96a96 | 2706 | */ |
846be085 MK |
2707 | SetHPageRestoreReserve(page); |
2708 | } else if (rc < 0) { | |
2709 | /* | |
2710 | * Rare out of memory condition from | |
2711 | * vma_needs_reservation call. Memory allocation is | |
2712 | * only attempted if a new entry is needed. Therefore, | |
2713 | * this implies there is not an entry in the | |
2714 | * reserve map. | |
2715 | * | |
2716 | * For shared mappings, no entry in the map indicates | |
2717 | * no reservation. We are done. | |
2718 | */ | |
2719 | if (!(vma->vm_flags & VM_MAYSHARE)) | |
2720 | /* | |
2721 | * For private mappings, no entry indicates | |
2722 | * a reservation is present. Since we can | |
2723 | * not add an entry, set SetHPageRestoreReserve | |
2724 | * on the page so reserve count will be | |
2725 | * incremented when freed. This reserve will | |
2726 | * be consumed on a subsequent allocation. | |
2727 | */ | |
2728 | SetHPageRestoreReserve(page); | |
96b96a96 | 2729 | } else |
846be085 MK |
2730 | /* |
2731 | * No reservation present, do nothing | |
2732 | */ | |
2733 | vma_end_reservation(h, vma, address); | |
96b96a96 MK |
2734 | } |
2735 | } | |
2736 | ||
369fa227 OS |
2737 | /* |
2738 | * alloc_and_dissolve_huge_page - Allocate a new page and dissolve the old one | |
2739 | * @h: struct hstate old page belongs to | |
2740 | * @old_page: Old page to dissolve | |
ae37c7ff | 2741 | * @list: List to isolate the page in case we need to |
369fa227 OS |
2742 | * Returns 0 on success, otherwise negated error. |
2743 | */ | |
ae37c7ff OS |
2744 | static int alloc_and_dissolve_huge_page(struct hstate *h, struct page *old_page, |
2745 | struct list_head *list) | |
369fa227 OS |
2746 | { |
2747 | gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE; | |
2748 | int nid = page_to_nid(old_page); | |
2749 | struct page *new_page; | |
2750 | int ret = 0; | |
2751 | ||
2752 | /* | |
2753 | * Before dissolving the page, we need to allocate a new one for the | |
f41f2ed4 MS |
2754 | * pool to remain stable. Here, we allocate the page and 'prep' it |
2755 | * by doing everything but actually updating counters and adding to | |
2756 | * the pool. This simplifies and let us do most of the processing | |
2757 | * under the lock. | |
369fa227 OS |
2758 | */ |
2759 | new_page = alloc_buddy_huge_page(h, gfp_mask, nid, NULL, NULL); | |
2760 | if (!new_page) | |
2761 | return -ENOMEM; | |
f41f2ed4 | 2762 | __prep_new_huge_page(h, new_page); |
369fa227 OS |
2763 | |
2764 | retry: | |
2765 | spin_lock_irq(&hugetlb_lock); | |
2766 | if (!PageHuge(old_page)) { | |
2767 | /* | |
2768 | * Freed from under us. Drop new_page too. | |
2769 | */ | |
2770 | goto free_new; | |
2771 | } else if (page_count(old_page)) { | |
2772 | /* | |
ae37c7ff OS |
2773 | * Someone has grabbed the page, try to isolate it here. |
2774 | * Fail with -EBUSY if not possible. | |
369fa227 | 2775 | */ |
ae37c7ff | 2776 | spin_unlock_irq(&hugetlb_lock); |
7ce82f4c | 2777 | ret = isolate_hugetlb(old_page, list); |
ae37c7ff | 2778 | spin_lock_irq(&hugetlb_lock); |
369fa227 OS |
2779 | goto free_new; |
2780 | } else if (!HPageFreed(old_page)) { | |
2781 | /* | |
2782 | * Page's refcount is 0 but it has not been enqueued in the | |
2783 | * freelist yet. Race window is small, so we can succeed here if | |
2784 | * we retry. | |
2785 | */ | |
2786 | spin_unlock_irq(&hugetlb_lock); | |
2787 | cond_resched(); | |
2788 | goto retry; | |
2789 | } else { | |
2790 | /* | |
2791 | * Ok, old_page is still a genuine free hugepage. Remove it from | |
2792 | * the freelist and decrease the counters. These will be | |
2793 | * incremented again when calling __prep_account_new_huge_page() | |
2794 | * and enqueue_huge_page() for new_page. The counters will remain | |
2795 | * stable since this happens under the lock. | |
2796 | */ | |
2797 | remove_hugetlb_page(h, old_page, false); | |
2798 | ||
2799 | /* | |
b65a4eda MK |
2800 | * Ref count on new page is already zero as it was dropped |
2801 | * earlier. It can be directly added to the pool free list. | |
369fa227 | 2802 | */ |
369fa227 | 2803 | __prep_account_new_huge_page(h, nid); |
369fa227 OS |
2804 | enqueue_huge_page(h, new_page); |
2805 | ||
2806 | /* | |
2807 | * Pages have been replaced, we can safely free the old one. | |
2808 | */ | |
2809 | spin_unlock_irq(&hugetlb_lock); | |
b65d4adb | 2810 | update_and_free_page(h, old_page, false); |
369fa227 OS |
2811 | } |
2812 | ||
2813 | return ret; | |
2814 | ||
2815 | free_new: | |
2816 | spin_unlock_irq(&hugetlb_lock); | |
b65a4eda MK |
2817 | /* Page has a zero ref count, but needs a ref to be freed */ |
2818 | set_page_refcounted(new_page); | |
b65d4adb | 2819 | update_and_free_page(h, new_page, false); |
369fa227 OS |
2820 | |
2821 | return ret; | |
2822 | } | |
2823 | ||
ae37c7ff | 2824 | int isolate_or_dissolve_huge_page(struct page *page, struct list_head *list) |
369fa227 OS |
2825 | { |
2826 | struct hstate *h; | |
2827 | struct page *head; | |
ae37c7ff | 2828 | int ret = -EBUSY; |
369fa227 OS |
2829 | |
2830 | /* | |
2831 | * The page might have been dissolved from under our feet, so make sure | |
2832 | * to carefully check the state under the lock. | |
2833 | * Return success when racing as if we dissolved the page ourselves. | |
2834 | */ | |
2835 | spin_lock_irq(&hugetlb_lock); | |
2836 | if (PageHuge(page)) { | |
2837 | head = compound_head(page); | |
2838 | h = page_hstate(head); | |
2839 | } else { | |
2840 | spin_unlock_irq(&hugetlb_lock); | |
2841 | return 0; | |
2842 | } | |
2843 | spin_unlock_irq(&hugetlb_lock); | |
2844 | ||
2845 | /* | |
2846 | * Fence off gigantic pages as there is a cyclic dependency between | |
2847 | * alloc_contig_range and them. Return -ENOMEM as this has the effect | |
2848 | * of bailing out right away without further retrying. | |
2849 | */ | |
2850 | if (hstate_is_gigantic(h)) | |
2851 | return -ENOMEM; | |
2852 | ||
7ce82f4c | 2853 | if (page_count(head) && !isolate_hugetlb(head, list)) |
ae37c7ff OS |
2854 | ret = 0; |
2855 | else if (!page_count(head)) | |
2856 | ret = alloc_and_dissolve_huge_page(h, head, list); | |
2857 | ||
2858 | return ret; | |
369fa227 OS |
2859 | } |
2860 | ||
70c3547e | 2861 | struct page *alloc_huge_page(struct vm_area_struct *vma, |
04f2cbe3 | 2862 | unsigned long addr, int avoid_reserve) |
1da177e4 | 2863 | { |
90481622 | 2864 | struct hugepage_subpool *spool = subpool_vma(vma); |
a5516438 | 2865 | struct hstate *h = hstate_vma(vma); |
348ea204 | 2866 | struct page *page; |
d85f69b0 MK |
2867 | long map_chg, map_commit; |
2868 | long gbl_chg; | |
6d76dcf4 AK |
2869 | int ret, idx; |
2870 | struct hugetlb_cgroup *h_cg; | |
08cf9faf | 2871 | bool deferred_reserve; |
a1e78772 | 2872 | |
6d76dcf4 | 2873 | idx = hstate_index(h); |
a1e78772 | 2874 | /* |
d85f69b0 MK |
2875 | * Examine the region/reserve map to determine if the process |
2876 | * has a reservation for the page to be allocated. A return | |
2877 | * code of zero indicates a reservation exists (no change). | |
a1e78772 | 2878 | */ |
d85f69b0 MK |
2879 | map_chg = gbl_chg = vma_needs_reservation(h, vma, addr); |
2880 | if (map_chg < 0) | |
76dcee75 | 2881 | return ERR_PTR(-ENOMEM); |
d85f69b0 MK |
2882 | |
2883 | /* | |
2884 | * Processes that did not create the mapping will have no | |
2885 | * reserves as indicated by the region/reserve map. Check | |
2886 | * that the allocation will not exceed the subpool limit. | |
2887 | * Allocations for MAP_NORESERVE mappings also need to be | |
2888 | * checked against any subpool limit. | |
2889 | */ | |
2890 | if (map_chg || avoid_reserve) { | |
2891 | gbl_chg = hugepage_subpool_get_pages(spool, 1); | |
2892 | if (gbl_chg < 0) { | |
feba16e2 | 2893 | vma_end_reservation(h, vma, addr); |
76dcee75 | 2894 | return ERR_PTR(-ENOSPC); |
5e911373 | 2895 | } |
1da177e4 | 2896 | |
d85f69b0 MK |
2897 | /* |
2898 | * Even though there was no reservation in the region/reserve | |
2899 | * map, there could be reservations associated with the | |
2900 | * subpool that can be used. This would be indicated if the | |
2901 | * return value of hugepage_subpool_get_pages() is zero. | |
2902 | * However, if avoid_reserve is specified we still avoid even | |
2903 | * the subpool reservations. | |
2904 | */ | |
2905 | if (avoid_reserve) | |
2906 | gbl_chg = 1; | |
2907 | } | |
2908 | ||
08cf9faf MA |
2909 | /* If this allocation is not consuming a reservation, charge it now. |
2910 | */ | |
6501fe5f | 2911 | deferred_reserve = map_chg || avoid_reserve; |
08cf9faf MA |
2912 | if (deferred_reserve) { |
2913 | ret = hugetlb_cgroup_charge_cgroup_rsvd( | |
2914 | idx, pages_per_huge_page(h), &h_cg); | |
2915 | if (ret) | |
2916 | goto out_subpool_put; | |
2917 | } | |
2918 | ||
6d76dcf4 | 2919 | ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg); |
8f34af6f | 2920 | if (ret) |
08cf9faf | 2921 | goto out_uncharge_cgroup_reservation; |
8f34af6f | 2922 | |
db71ef79 | 2923 | spin_lock_irq(&hugetlb_lock); |
d85f69b0 MK |
2924 | /* |
2925 | * glb_chg is passed to indicate whether or not a page must be taken | |
2926 | * from the global free pool (global change). gbl_chg == 0 indicates | |
2927 | * a reservation exists for the allocation. | |
2928 | */ | |
2929 | page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg); | |
81a6fcae | 2930 | if (!page) { |
db71ef79 | 2931 | spin_unlock_irq(&hugetlb_lock); |
0c397dae | 2932 | page = alloc_buddy_huge_page_with_mpol(h, vma, addr); |
8f34af6f JZ |
2933 | if (!page) |
2934 | goto out_uncharge_cgroup; | |
12df140f | 2935 | spin_lock_irq(&hugetlb_lock); |
a88c7695 | 2936 | if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) { |
d6995da3 | 2937 | SetHPageRestoreReserve(page); |
a88c7695 NH |
2938 | h->resv_huge_pages--; |
2939 | } | |
15a8d68e | 2940 | list_add(&page->lru, &h->hugepage_activelist); |
2b21624f | 2941 | set_page_refcounted(page); |
81a6fcae | 2942 | /* Fall through */ |
68842c9b | 2943 | } |
81a6fcae | 2944 | hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page); |
08cf9faf MA |
2945 | /* If allocation is not consuming a reservation, also store the |
2946 | * hugetlb_cgroup pointer on the page. | |
2947 | */ | |
2948 | if (deferred_reserve) { | |
2949 | hugetlb_cgroup_commit_charge_rsvd(idx, pages_per_huge_page(h), | |
2950 | h_cg, page); | |
2951 | } | |
2952 | ||
db71ef79 | 2953 | spin_unlock_irq(&hugetlb_lock); |
348ea204 | 2954 | |
d6995da3 | 2955 | hugetlb_set_page_subpool(page, spool); |
90d8b7e6 | 2956 | |
d85f69b0 MK |
2957 | map_commit = vma_commit_reservation(h, vma, addr); |
2958 | if (unlikely(map_chg > map_commit)) { | |
33039678 MK |
2959 | /* |
2960 | * The page was added to the reservation map between | |
2961 | * vma_needs_reservation and vma_commit_reservation. | |
2962 | * This indicates a race with hugetlb_reserve_pages. | |
2963 | * Adjust for the subpool count incremented above AND | |
2964 | * in hugetlb_reserve_pages for the same page. Also, | |
2965 | * the reservation count added in hugetlb_reserve_pages | |
2966 | * no longer applies. | |
2967 | */ | |
2968 | long rsv_adjust; | |
2969 | ||
2970 | rsv_adjust = hugepage_subpool_put_pages(spool, 1); | |
2971 | hugetlb_acct_memory(h, -rsv_adjust); | |
79aa925b MK |
2972 | if (deferred_reserve) |
2973 | hugetlb_cgroup_uncharge_page_rsvd(hstate_index(h), | |
2974 | pages_per_huge_page(h), page); | |
33039678 | 2975 | } |
90d8b7e6 | 2976 | return page; |
8f34af6f JZ |
2977 | |
2978 | out_uncharge_cgroup: | |
2979 | hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg); | |
08cf9faf MA |
2980 | out_uncharge_cgroup_reservation: |
2981 | if (deferred_reserve) | |
2982 | hugetlb_cgroup_uncharge_cgroup_rsvd(idx, pages_per_huge_page(h), | |
2983 | h_cg); | |
8f34af6f | 2984 | out_subpool_put: |
d85f69b0 | 2985 | if (map_chg || avoid_reserve) |
8f34af6f | 2986 | hugepage_subpool_put_pages(spool, 1); |
feba16e2 | 2987 | vma_end_reservation(h, vma, addr); |
8f34af6f | 2988 | return ERR_PTR(-ENOSPC); |
b45b5bd6 DG |
2989 | } |
2990 | ||
b5389086 | 2991 | int alloc_bootmem_huge_page(struct hstate *h, int nid) |
e24a1307 | 2992 | __attribute__ ((weak, alias("__alloc_bootmem_huge_page"))); |
b5389086 | 2993 | int __alloc_bootmem_huge_page(struct hstate *h, int nid) |
aa888a74 | 2994 | { |
b5389086 | 2995 | struct huge_bootmem_page *m = NULL; /* initialize for clang */ |
b2261026 | 2996 | int nr_nodes, node; |
aa888a74 | 2997 | |
b5389086 ZY |
2998 | /* do node specific alloc */ |
2999 | if (nid != NUMA_NO_NODE) { | |
3000 | m = memblock_alloc_try_nid_raw(huge_page_size(h), huge_page_size(h), | |
3001 | 0, MEMBLOCK_ALLOC_ACCESSIBLE, nid); | |
3002 | if (!m) | |
3003 | return 0; | |
3004 | goto found; | |
3005 | } | |
3006 | /* allocate from next node when distributing huge pages */ | |
b2261026 | 3007 | for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) { |
b5389086 | 3008 | m = memblock_alloc_try_nid_raw( |
8b89a116 | 3009 | huge_page_size(h), huge_page_size(h), |
97ad1087 | 3010 | 0, MEMBLOCK_ALLOC_ACCESSIBLE, node); |
b5389086 ZY |
3011 | /* |
3012 | * Use the beginning of the huge page to store the | |
3013 | * huge_bootmem_page struct (until gather_bootmem | |
3014 | * puts them into the mem_map). | |
3015 | */ | |
3016 | if (!m) | |
3017 | return 0; | |
3018 | goto found; | |
aa888a74 | 3019 | } |
aa888a74 AK |
3020 | |
3021 | found: | |
aa888a74 | 3022 | /* Put them into a private list first because mem_map is not up yet */ |
330d6e48 | 3023 | INIT_LIST_HEAD(&m->list); |
aa888a74 AK |
3024 | list_add(&m->list, &huge_boot_pages); |
3025 | m->hstate = h; | |
3026 | return 1; | |
3027 | } | |
3028 | ||
48b8d744 MK |
3029 | /* |
3030 | * Put bootmem huge pages into the standard lists after mem_map is up. | |
3031 | * Note: This only applies to gigantic (order > MAX_ORDER) pages. | |
3032 | */ | |
aa888a74 AK |
3033 | static void __init gather_bootmem_prealloc(void) |
3034 | { | |
3035 | struct huge_bootmem_page *m; | |
3036 | ||
3037 | list_for_each_entry(m, &huge_boot_pages, list) { | |
40d18ebf | 3038 | struct page *page = virt_to_page(m); |
aa888a74 | 3039 | struct hstate *h = m->hstate; |
ee8f248d | 3040 | |
48b8d744 | 3041 | VM_BUG_ON(!hstate_is_gigantic(h)); |
aa888a74 | 3042 | WARN_ON(page_count(page) != 1); |
7118fc29 MK |
3043 | if (prep_compound_gigantic_page(page, huge_page_order(h))) { |
3044 | WARN_ON(PageReserved(page)); | |
3045 | prep_new_huge_page(h, page, page_to_nid(page)); | |
2b21624f | 3046 | free_huge_page(page); /* add to the hugepage allocator */ |
7118fc29 | 3047 | } else { |
416d85ed | 3048 | /* VERY unlikely inflated ref count on a tail page */ |
7118fc29 | 3049 | free_gigantic_page(page, huge_page_order(h)); |
7118fc29 | 3050 | } |
af0fb9df | 3051 | |
b0320c7b | 3052 | /* |
48b8d744 MK |
3053 | * We need to restore the 'stolen' pages to totalram_pages |
3054 | * in order to fix confusing memory reports from free(1) and | |
3055 | * other side-effects, like CommitLimit going negative. | |
b0320c7b | 3056 | */ |
48b8d744 | 3057 | adjust_managed_page_count(page, pages_per_huge_page(h)); |
520495fe | 3058 | cond_resched(); |
aa888a74 AK |
3059 | } |
3060 | } | |
b5389086 ZY |
3061 | static void __init hugetlb_hstate_alloc_pages_onenode(struct hstate *h, int nid) |
3062 | { | |
3063 | unsigned long i; | |
3064 | char buf[32]; | |
3065 | ||
3066 | for (i = 0; i < h->max_huge_pages_node[nid]; ++i) { | |
3067 | if (hstate_is_gigantic(h)) { | |
3068 | if (!alloc_bootmem_huge_page(h, nid)) | |
3069 | break; | |
3070 | } else { | |
3071 | struct page *page; | |
3072 | gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE; | |
3073 | ||
3074 | page = alloc_fresh_huge_page(h, gfp_mask, nid, | |
3075 | &node_states[N_MEMORY], NULL); | |
3076 | if (!page) | |
3077 | break; | |
2b21624f | 3078 | free_huge_page(page); /* free it into the hugepage allocator */ |
b5389086 ZY |
3079 | } |
3080 | cond_resched(); | |
3081 | } | |
3082 | if (i == h->max_huge_pages_node[nid]) | |
3083 | return; | |
3084 | ||
3085 | string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32); | |
3086 | pr_warn("HugeTLB: allocating %u of page size %s failed node%d. Only allocated %lu hugepages.\n", | |
3087 | h->max_huge_pages_node[nid], buf, nid, i); | |
3088 | h->max_huge_pages -= (h->max_huge_pages_node[nid] - i); | |
3089 | h->max_huge_pages_node[nid] = i; | |
3090 | } | |
aa888a74 | 3091 | |
8faa8b07 | 3092 | static void __init hugetlb_hstate_alloc_pages(struct hstate *h) |
1da177e4 LT |
3093 | { |
3094 | unsigned long i; | |
f60858f9 | 3095 | nodemask_t *node_alloc_noretry; |
b5389086 ZY |
3096 | bool node_specific_alloc = false; |
3097 | ||
3098 | /* skip gigantic hugepages allocation if hugetlb_cma enabled */ | |
3099 | if (hstate_is_gigantic(h) && hugetlb_cma_size) { | |
3100 | pr_warn_once("HugeTLB: hugetlb_cma is enabled, skip boot time allocation\n"); | |
3101 | return; | |
3102 | } | |
3103 | ||
3104 | /* do node specific alloc */ | |
0a7a0f6f | 3105 | for_each_online_node(i) { |
b5389086 ZY |
3106 | if (h->max_huge_pages_node[i] > 0) { |
3107 | hugetlb_hstate_alloc_pages_onenode(h, i); | |
3108 | node_specific_alloc = true; | |
3109 | } | |
3110 | } | |
f60858f9 | 3111 | |
b5389086 ZY |
3112 | if (node_specific_alloc) |
3113 | return; | |
3114 | ||
3115 | /* below will do all node balanced alloc */ | |
f60858f9 MK |
3116 | if (!hstate_is_gigantic(h)) { |
3117 | /* | |
3118 | * Bit mask controlling how hard we retry per-node allocations. | |
3119 | * Ignore errors as lower level routines can deal with | |
3120 | * node_alloc_noretry == NULL. If this kmalloc fails at boot | |
3121 | * time, we are likely in bigger trouble. | |
3122 | */ | |
3123 | node_alloc_noretry = kmalloc(sizeof(*node_alloc_noretry), | |
3124 | GFP_KERNEL); | |
3125 | } else { | |
3126 | /* allocations done at boot time */ | |
3127 | node_alloc_noretry = NULL; | |
3128 | } | |
3129 | ||
3130 | /* bit mask controlling how hard we retry per-node allocations */ | |
3131 | if (node_alloc_noretry) | |
3132 | nodes_clear(*node_alloc_noretry); | |
a5516438 | 3133 | |
e5ff2159 | 3134 | for (i = 0; i < h->max_huge_pages; ++i) { |
bae7f4ae | 3135 | if (hstate_is_gigantic(h)) { |
b5389086 | 3136 | if (!alloc_bootmem_huge_page(h, NUMA_NO_NODE)) |
aa888a74 | 3137 | break; |
0c397dae | 3138 | } else if (!alloc_pool_huge_page(h, |
f60858f9 MK |
3139 | &node_states[N_MEMORY], |
3140 | node_alloc_noretry)) | |
1da177e4 | 3141 | break; |
69ed779a | 3142 | cond_resched(); |
1da177e4 | 3143 | } |
d715cf80 LH |
3144 | if (i < h->max_huge_pages) { |
3145 | char buf[32]; | |
3146 | ||
c6247f72 | 3147 | string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32); |
d715cf80 LH |
3148 | pr_warn("HugeTLB: allocating %lu of page size %s failed. Only allocated %lu hugepages.\n", |
3149 | h->max_huge_pages, buf, i); | |
3150 | h->max_huge_pages = i; | |
3151 | } | |
f60858f9 | 3152 | kfree(node_alloc_noretry); |
e5ff2159 AK |
3153 | } |
3154 | ||
3155 | static void __init hugetlb_init_hstates(void) | |
3156 | { | |
79dfc695 | 3157 | struct hstate *h, *h2; |
e5ff2159 AK |
3158 | |
3159 | for_each_hstate(h) { | |
8faa8b07 | 3160 | /* oversize hugepages were init'ed in early boot */ |
bae7f4ae | 3161 | if (!hstate_is_gigantic(h)) |
8faa8b07 | 3162 | hugetlb_hstate_alloc_pages(h); |
79dfc695 MK |
3163 | |
3164 | /* | |
3165 | * Set demote order for each hstate. Note that | |
3166 | * h->demote_order is initially 0. | |
3167 | * - We can not demote gigantic pages if runtime freeing | |
3168 | * is not supported, so skip this. | |
a01f4390 MK |
3169 | * - If CMA allocation is possible, we can not demote |
3170 | * HUGETLB_PAGE_ORDER or smaller size pages. | |
79dfc695 MK |
3171 | */ |
3172 | if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) | |
3173 | continue; | |
a01f4390 MK |
3174 | if (hugetlb_cma_size && h->order <= HUGETLB_PAGE_ORDER) |
3175 | continue; | |
79dfc695 MK |
3176 | for_each_hstate(h2) { |
3177 | if (h2 == h) | |
3178 | continue; | |
3179 | if (h2->order < h->order && | |
3180 | h2->order > h->demote_order) | |
3181 | h->demote_order = h2->order; | |
3182 | } | |
e5ff2159 AK |
3183 | } |
3184 | } | |
3185 | ||
3186 | static void __init report_hugepages(void) | |
3187 | { | |
3188 | struct hstate *h; | |
3189 | ||
3190 | for_each_hstate(h) { | |
4abd32db | 3191 | char buf[32]; |
c6247f72 MW |
3192 | |
3193 | string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32); | |
6213834c | 3194 | pr_info("HugeTLB: registered %s page size, pre-allocated %ld pages\n", |
c6247f72 | 3195 | buf, h->free_huge_pages); |
6213834c MS |
3196 | pr_info("HugeTLB: %d KiB vmemmap can be freed for a %s page\n", |
3197 | hugetlb_vmemmap_optimizable_size(h) / SZ_1K, buf); | |
e5ff2159 AK |
3198 | } |
3199 | } | |
3200 | ||
1da177e4 | 3201 | #ifdef CONFIG_HIGHMEM |
6ae11b27 LS |
3202 | static void try_to_free_low(struct hstate *h, unsigned long count, |
3203 | nodemask_t *nodes_allowed) | |
1da177e4 | 3204 | { |
4415cc8d | 3205 | int i; |
1121828a | 3206 | LIST_HEAD(page_list); |
4415cc8d | 3207 | |
9487ca60 | 3208 | lockdep_assert_held(&hugetlb_lock); |
bae7f4ae | 3209 | if (hstate_is_gigantic(h)) |
aa888a74 AK |
3210 | return; |
3211 | ||
1121828a MK |
3212 | /* |
3213 | * Collect pages to be freed on a list, and free after dropping lock | |
3214 | */ | |
6ae11b27 | 3215 | for_each_node_mask(i, *nodes_allowed) { |
10c6ec49 | 3216 | struct page *page, *next; |
a5516438 AK |
3217 | struct list_head *freel = &h->hugepage_freelists[i]; |
3218 | list_for_each_entry_safe(page, next, freel, lru) { | |
3219 | if (count >= h->nr_huge_pages) | |
1121828a | 3220 | goto out; |
1da177e4 LT |
3221 | if (PageHighMem(page)) |
3222 | continue; | |
6eb4e88a | 3223 | remove_hugetlb_page(h, page, false); |
1121828a | 3224 | list_add(&page->lru, &page_list); |
1da177e4 LT |
3225 | } |
3226 | } | |
1121828a MK |
3227 | |
3228 | out: | |
db71ef79 | 3229 | spin_unlock_irq(&hugetlb_lock); |
10c6ec49 | 3230 | update_and_free_pages_bulk(h, &page_list); |
db71ef79 | 3231 | spin_lock_irq(&hugetlb_lock); |
1da177e4 LT |
3232 | } |
3233 | #else | |
6ae11b27 LS |
3234 | static inline void try_to_free_low(struct hstate *h, unsigned long count, |
3235 | nodemask_t *nodes_allowed) | |
1da177e4 LT |
3236 | { |
3237 | } | |
3238 | #endif | |
3239 | ||
20a0307c WF |
3240 | /* |
3241 | * Increment or decrement surplus_huge_pages. Keep node-specific counters | |
3242 | * balanced by operating on them in a round-robin fashion. | |
3243 | * Returns 1 if an adjustment was made. | |
3244 | */ | |
6ae11b27 LS |
3245 | static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed, |
3246 | int delta) | |
20a0307c | 3247 | { |
b2261026 | 3248 | int nr_nodes, node; |
20a0307c | 3249 | |
9487ca60 | 3250 | lockdep_assert_held(&hugetlb_lock); |
20a0307c | 3251 | VM_BUG_ON(delta != -1 && delta != 1); |
20a0307c | 3252 | |
b2261026 JK |
3253 | if (delta < 0) { |
3254 | for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { | |
3255 | if (h->surplus_huge_pages_node[node]) | |
3256 | goto found; | |
e8c5c824 | 3257 | } |
b2261026 JK |
3258 | } else { |
3259 | for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { | |
3260 | if (h->surplus_huge_pages_node[node] < | |
3261 | h->nr_huge_pages_node[node]) | |
3262 | goto found; | |
e8c5c824 | 3263 | } |
b2261026 JK |
3264 | } |
3265 | return 0; | |
20a0307c | 3266 | |
b2261026 JK |
3267 | found: |
3268 | h->surplus_huge_pages += delta; | |
3269 | h->surplus_huge_pages_node[node] += delta; | |
3270 | return 1; | |
20a0307c WF |
3271 | } |
3272 | ||
a5516438 | 3273 | #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages) |
fd875dca | 3274 | static int set_max_huge_pages(struct hstate *h, unsigned long count, int nid, |
4eb0716e | 3275 | nodemask_t *nodes_allowed) |
1da177e4 | 3276 | { |
7893d1d5 | 3277 | unsigned long min_count, ret; |
10c6ec49 MK |
3278 | struct page *page; |
3279 | LIST_HEAD(page_list); | |
f60858f9 MK |
3280 | NODEMASK_ALLOC(nodemask_t, node_alloc_noretry, GFP_KERNEL); |
3281 | ||
3282 | /* | |
3283 | * Bit mask controlling how hard we retry per-node allocations. | |
3284 | * If we can not allocate the bit mask, do not attempt to allocate | |
3285 | * the requested huge pages. | |
3286 | */ | |
3287 | if (node_alloc_noretry) | |
3288 | nodes_clear(*node_alloc_noretry); | |
3289 | else | |
3290 | return -ENOMEM; | |
1da177e4 | 3291 | |
29383967 MK |
3292 | /* |
3293 | * resize_lock mutex prevents concurrent adjustments to number of | |
3294 | * pages in hstate via the proc/sysfs interfaces. | |
3295 | */ | |
3296 | mutex_lock(&h->resize_lock); | |
b65d4adb | 3297 | flush_free_hpage_work(h); |
db71ef79 | 3298 | spin_lock_irq(&hugetlb_lock); |
4eb0716e | 3299 | |
fd875dca MK |
3300 | /* |
3301 | * Check for a node specific request. | |
3302 | * Changing node specific huge page count may require a corresponding | |
3303 | * change to the global count. In any case, the passed node mask | |
3304 | * (nodes_allowed) will restrict alloc/free to the specified node. | |
3305 | */ | |
3306 | if (nid != NUMA_NO_NODE) { | |
3307 | unsigned long old_count = count; | |
3308 | ||
3309 | count += h->nr_huge_pages - h->nr_huge_pages_node[nid]; | |
3310 | /* | |
3311 | * User may have specified a large count value which caused the | |
3312 | * above calculation to overflow. In this case, they wanted | |
3313 | * to allocate as many huge pages as possible. Set count to | |
3314 | * largest possible value to align with their intention. | |
3315 | */ | |
3316 | if (count < old_count) | |
3317 | count = ULONG_MAX; | |
3318 | } | |
3319 | ||
4eb0716e AG |
3320 | /* |
3321 | * Gigantic pages runtime allocation depend on the capability for large | |
3322 | * page range allocation. | |
3323 | * If the system does not provide this feature, return an error when | |
3324 | * the user tries to allocate gigantic pages but let the user free the | |
3325 | * boottime allocated gigantic pages. | |
3326 | */ | |
3327 | if (hstate_is_gigantic(h) && !IS_ENABLED(CONFIG_CONTIG_ALLOC)) { | |
3328 | if (count > persistent_huge_pages(h)) { | |
db71ef79 | 3329 | spin_unlock_irq(&hugetlb_lock); |
29383967 | 3330 | mutex_unlock(&h->resize_lock); |
f60858f9 | 3331 | NODEMASK_FREE(node_alloc_noretry); |
4eb0716e AG |
3332 | return -EINVAL; |
3333 | } | |
3334 | /* Fall through to decrease pool */ | |
3335 | } | |
aa888a74 | 3336 | |
7893d1d5 AL |
3337 | /* |
3338 | * Increase the pool size | |
3339 | * First take pages out of surplus state. Then make up the | |
3340 | * remaining difference by allocating fresh huge pages. | |
d1c3fb1f | 3341 | * |
0c397dae | 3342 | * We might race with alloc_surplus_huge_page() here and be unable |
d1c3fb1f NA |
3343 | * to convert a surplus huge page to a normal huge page. That is |
3344 | * not critical, though, it just means the overall size of the | |
3345 | * pool might be one hugepage larger than it needs to be, but | |
3346 | * within all the constraints specified by the sysctls. | |
7893d1d5 | 3347 | */ |
a5516438 | 3348 | while (h->surplus_huge_pages && count > persistent_huge_pages(h)) { |
6ae11b27 | 3349 | if (!adjust_pool_surplus(h, nodes_allowed, -1)) |
7893d1d5 AL |
3350 | break; |
3351 | } | |
3352 | ||
a5516438 | 3353 | while (count > persistent_huge_pages(h)) { |
7893d1d5 AL |
3354 | /* |
3355 | * If this allocation races such that we no longer need the | |
3356 | * page, free_huge_page will handle it by freeing the page | |
3357 | * and reducing the surplus. | |
3358 | */ | |
db71ef79 | 3359 | spin_unlock_irq(&hugetlb_lock); |
649920c6 JH |
3360 | |
3361 | /* yield cpu to avoid soft lockup */ | |
3362 | cond_resched(); | |
3363 | ||
f60858f9 MK |
3364 | ret = alloc_pool_huge_page(h, nodes_allowed, |
3365 | node_alloc_noretry); | |
db71ef79 | 3366 | spin_lock_irq(&hugetlb_lock); |
7893d1d5 AL |
3367 | if (!ret) |
3368 | goto out; | |
3369 | ||
536240f2 MG |
3370 | /* Bail for signals. Probably ctrl-c from user */ |
3371 | if (signal_pending(current)) | |
3372 | goto out; | |
7893d1d5 | 3373 | } |
7893d1d5 AL |
3374 | |
3375 | /* | |
3376 | * Decrease the pool size | |
3377 | * First return free pages to the buddy allocator (being careful | |
3378 | * to keep enough around to satisfy reservations). Then place | |
3379 | * pages into surplus state as needed so the pool will shrink | |
3380 | * to the desired size as pages become free. | |
d1c3fb1f NA |
3381 | * |
3382 | * By placing pages into the surplus state independent of the | |
3383 | * overcommit value, we are allowing the surplus pool size to | |
3384 | * exceed overcommit. There are few sane options here. Since | |
0c397dae | 3385 | * alloc_surplus_huge_page() is checking the global counter, |
d1c3fb1f NA |
3386 | * though, we'll note that we're not allowed to exceed surplus |
3387 | * and won't grow the pool anywhere else. Not until one of the | |
3388 | * sysctls are changed, or the surplus pages go out of use. | |
7893d1d5 | 3389 | */ |
a5516438 | 3390 | min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages; |
6b0c880d | 3391 | min_count = max(count, min_count); |
6ae11b27 | 3392 | try_to_free_low(h, min_count, nodes_allowed); |
10c6ec49 MK |
3393 | |
3394 | /* | |
3395 | * Collect pages to be removed on list without dropping lock | |
3396 | */ | |
a5516438 | 3397 | while (min_count < persistent_huge_pages(h)) { |
10c6ec49 MK |
3398 | page = remove_pool_huge_page(h, nodes_allowed, 0); |
3399 | if (!page) | |
1da177e4 | 3400 | break; |
10c6ec49 MK |
3401 | |
3402 | list_add(&page->lru, &page_list); | |
1da177e4 | 3403 | } |
10c6ec49 | 3404 | /* free the pages after dropping lock */ |
db71ef79 | 3405 | spin_unlock_irq(&hugetlb_lock); |
10c6ec49 | 3406 | update_and_free_pages_bulk(h, &page_list); |
b65d4adb | 3407 | flush_free_hpage_work(h); |
db71ef79 | 3408 | spin_lock_irq(&hugetlb_lock); |
10c6ec49 | 3409 | |
a5516438 | 3410 | while (count < persistent_huge_pages(h)) { |
6ae11b27 | 3411 | if (!adjust_pool_surplus(h, nodes_allowed, 1)) |
7893d1d5 AL |
3412 | break; |
3413 | } | |
3414 | out: | |
4eb0716e | 3415 | h->max_huge_pages = persistent_huge_pages(h); |
db71ef79 | 3416 | spin_unlock_irq(&hugetlb_lock); |
29383967 | 3417 | mutex_unlock(&h->resize_lock); |
4eb0716e | 3418 | |
f60858f9 MK |
3419 | NODEMASK_FREE(node_alloc_noretry); |
3420 | ||
4eb0716e | 3421 | return 0; |
1da177e4 LT |
3422 | } |
3423 | ||
8531fc6f MK |
3424 | static int demote_free_huge_page(struct hstate *h, struct page *page) |
3425 | { | |
3426 | int i, nid = page_to_nid(page); | |
3427 | struct hstate *target_hstate; | |
31731452 | 3428 | struct page *subpage; |
8531fc6f MK |
3429 | int rc = 0; |
3430 | ||
3431 | target_hstate = size_to_hstate(PAGE_SIZE << h->demote_order); | |
3432 | ||
3433 | remove_hugetlb_page_for_demote(h, page, false); | |
3434 | spin_unlock_irq(&hugetlb_lock); | |
3435 | ||
6213834c | 3436 | rc = hugetlb_vmemmap_restore(h, page); |
8531fc6f MK |
3437 | if (rc) { |
3438 | /* Allocation of vmemmmap failed, we can not demote page */ | |
3439 | spin_lock_irq(&hugetlb_lock); | |
3440 | set_page_refcounted(page); | |
3441 | add_hugetlb_page(h, page, false); | |
3442 | return rc; | |
3443 | } | |
3444 | ||
3445 | /* | |
3446 | * Use destroy_compound_hugetlb_page_for_demote for all huge page | |
3447 | * sizes as it will not ref count pages. | |
3448 | */ | |
3449 | destroy_compound_hugetlb_page_for_demote(page, huge_page_order(h)); | |
3450 | ||
3451 | /* | |
3452 | * Taking target hstate mutex synchronizes with set_max_huge_pages. | |
3453 | * Without the mutex, pages added to target hstate could be marked | |
3454 | * as surplus. | |
3455 | * | |
3456 | * Note that we already hold h->resize_lock. To prevent deadlock, | |
3457 | * use the convention of always taking larger size hstate mutex first. | |
3458 | */ | |
3459 | mutex_lock(&target_hstate->resize_lock); | |
3460 | for (i = 0; i < pages_per_huge_page(h); | |
3461 | i += pages_per_huge_page(target_hstate)) { | |
31731452 | 3462 | subpage = nth_page(page, i); |
8531fc6f | 3463 | if (hstate_is_gigantic(target_hstate)) |
31731452 | 3464 | prep_compound_gigantic_page_for_demote(subpage, |
8531fc6f MK |
3465 | target_hstate->order); |
3466 | else | |
31731452 DB |
3467 | prep_compound_page(subpage, target_hstate->order); |
3468 | set_page_private(subpage, 0); | |
31731452 | 3469 | prep_new_huge_page(target_hstate, subpage, nid); |
2b21624f | 3470 | free_huge_page(subpage); |
8531fc6f MK |
3471 | } |
3472 | mutex_unlock(&target_hstate->resize_lock); | |
3473 | ||
3474 | spin_lock_irq(&hugetlb_lock); | |
3475 | ||
3476 | /* | |
3477 | * Not absolutely necessary, but for consistency update max_huge_pages | |
3478 | * based on pool changes for the demoted page. | |
3479 | */ | |
3480 | h->max_huge_pages--; | |
a43a83c7 ML |
3481 | target_hstate->max_huge_pages += |
3482 | pages_per_huge_page(h) / pages_per_huge_page(target_hstate); | |
8531fc6f MK |
3483 | |
3484 | return rc; | |
3485 | } | |
3486 | ||
79dfc695 MK |
3487 | static int demote_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed) |
3488 | __must_hold(&hugetlb_lock) | |
3489 | { | |
8531fc6f MK |
3490 | int nr_nodes, node; |
3491 | struct page *page; | |
79dfc695 MK |
3492 | |
3493 | lockdep_assert_held(&hugetlb_lock); | |
3494 | ||
3495 | /* We should never get here if no demote order */ | |
3496 | if (!h->demote_order) { | |
3497 | pr_warn("HugeTLB: NULL demote order passed to demote_pool_huge_page.\n"); | |
3498 | return -EINVAL; /* internal error */ | |
3499 | } | |
3500 | ||
8531fc6f | 3501 | for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { |
5a317412 MK |
3502 | list_for_each_entry(page, &h->hugepage_freelists[node], lru) { |
3503 | if (PageHWPoison(page)) | |
3504 | continue; | |
3505 | ||
3506 | return demote_free_huge_page(h, page); | |
8531fc6f MK |
3507 | } |
3508 | } | |
3509 | ||
5a317412 MK |
3510 | /* |
3511 | * Only way to get here is if all pages on free lists are poisoned. | |
3512 | * Return -EBUSY so that caller will not retry. | |
3513 | */ | |
3514 | return -EBUSY; | |
79dfc695 MK |
3515 | } |
3516 | ||
a3437870 NA |
3517 | #define HSTATE_ATTR_RO(_name) \ |
3518 | static struct kobj_attribute _name##_attr = __ATTR_RO(_name) | |
3519 | ||
79dfc695 MK |
3520 | #define HSTATE_ATTR_WO(_name) \ |
3521 | static struct kobj_attribute _name##_attr = __ATTR_WO(_name) | |
3522 | ||
a3437870 | 3523 | #define HSTATE_ATTR(_name) \ |
98bc26ac | 3524 | static struct kobj_attribute _name##_attr = __ATTR_RW(_name) |
a3437870 NA |
3525 | |
3526 | static struct kobject *hugepages_kobj; | |
3527 | static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; | |
3528 | ||
9a305230 LS |
3529 | static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp); |
3530 | ||
3531 | static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp) | |
a3437870 NA |
3532 | { |
3533 | int i; | |
9a305230 | 3534 | |
a3437870 | 3535 | for (i = 0; i < HUGE_MAX_HSTATE; i++) |
9a305230 LS |
3536 | if (hstate_kobjs[i] == kobj) { |
3537 | if (nidp) | |
3538 | *nidp = NUMA_NO_NODE; | |
a3437870 | 3539 | return &hstates[i]; |
9a305230 LS |
3540 | } |
3541 | ||
3542 | return kobj_to_node_hstate(kobj, nidp); | |
a3437870 NA |
3543 | } |
3544 | ||
06808b08 | 3545 | static ssize_t nr_hugepages_show_common(struct kobject *kobj, |
a3437870 NA |
3546 | struct kobj_attribute *attr, char *buf) |
3547 | { | |
9a305230 LS |
3548 | struct hstate *h; |
3549 | unsigned long nr_huge_pages; | |
3550 | int nid; | |
3551 | ||
3552 | h = kobj_to_hstate(kobj, &nid); | |
3553 | if (nid == NUMA_NO_NODE) | |
3554 | nr_huge_pages = h->nr_huge_pages; | |
3555 | else | |
3556 | nr_huge_pages = h->nr_huge_pages_node[nid]; | |
3557 | ||
ae7a927d | 3558 | return sysfs_emit(buf, "%lu\n", nr_huge_pages); |
a3437870 | 3559 | } |
adbe8726 | 3560 | |
238d3c13 DR |
3561 | static ssize_t __nr_hugepages_store_common(bool obey_mempolicy, |
3562 | struct hstate *h, int nid, | |
3563 | unsigned long count, size_t len) | |
a3437870 NA |
3564 | { |
3565 | int err; | |
2d0adf7e | 3566 | nodemask_t nodes_allowed, *n_mask; |
a3437870 | 3567 | |
2d0adf7e OS |
3568 | if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) |
3569 | return -EINVAL; | |
adbe8726 | 3570 | |
9a305230 LS |
3571 | if (nid == NUMA_NO_NODE) { |
3572 | /* | |
3573 | * global hstate attribute | |
3574 | */ | |
3575 | if (!(obey_mempolicy && | |
2d0adf7e OS |
3576 | init_nodemask_of_mempolicy(&nodes_allowed))) |
3577 | n_mask = &node_states[N_MEMORY]; | |
3578 | else | |
3579 | n_mask = &nodes_allowed; | |
3580 | } else { | |
9a305230 | 3581 | /* |
fd875dca MK |
3582 | * Node specific request. count adjustment happens in |
3583 | * set_max_huge_pages() after acquiring hugetlb_lock. | |
9a305230 | 3584 | */ |
2d0adf7e OS |
3585 | init_nodemask_of_node(&nodes_allowed, nid); |
3586 | n_mask = &nodes_allowed; | |
fd875dca | 3587 | } |
9a305230 | 3588 | |
2d0adf7e | 3589 | err = set_max_huge_pages(h, count, nid, n_mask); |
06808b08 | 3590 | |
4eb0716e | 3591 | return err ? err : len; |
06808b08 LS |
3592 | } |
3593 | ||
238d3c13 DR |
3594 | static ssize_t nr_hugepages_store_common(bool obey_mempolicy, |
3595 | struct kobject *kobj, const char *buf, | |
3596 | size_t len) | |
3597 | { | |
3598 | struct hstate *h; | |
3599 | unsigned long count; | |
3600 | int nid; | |
3601 | int err; | |
3602 | ||
3603 | err = kstrtoul(buf, 10, &count); | |
3604 | if (err) | |
3605 | return err; | |
3606 | ||
3607 | h = kobj_to_hstate(kobj, &nid); | |
3608 | return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len); | |
3609 | } | |
3610 | ||
06808b08 LS |
3611 | static ssize_t nr_hugepages_show(struct kobject *kobj, |
3612 | struct kobj_attribute *attr, char *buf) | |
3613 | { | |
3614 | return nr_hugepages_show_common(kobj, attr, buf); | |
3615 | } | |
3616 | ||
3617 | static ssize_t nr_hugepages_store(struct kobject *kobj, | |
3618 | struct kobj_attribute *attr, const char *buf, size_t len) | |
3619 | { | |
238d3c13 | 3620 | return nr_hugepages_store_common(false, kobj, buf, len); |
a3437870 NA |
3621 | } |
3622 | HSTATE_ATTR(nr_hugepages); | |
3623 | ||
06808b08 LS |
3624 | #ifdef CONFIG_NUMA |
3625 | ||
3626 | /* | |
3627 | * hstate attribute for optionally mempolicy-based constraint on persistent | |
3628 | * huge page alloc/free. | |
3629 | */ | |
3630 | static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj, | |
ae7a927d JP |
3631 | struct kobj_attribute *attr, |
3632 | char *buf) | |
06808b08 LS |
3633 | { |
3634 | return nr_hugepages_show_common(kobj, attr, buf); | |
3635 | } | |
3636 | ||
3637 | static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj, | |
3638 | struct kobj_attribute *attr, const char *buf, size_t len) | |
3639 | { | |
238d3c13 | 3640 | return nr_hugepages_store_common(true, kobj, buf, len); |
06808b08 LS |
3641 | } |
3642 | HSTATE_ATTR(nr_hugepages_mempolicy); | |
3643 | #endif | |
3644 | ||
3645 | ||
a3437870 NA |
3646 | static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj, |
3647 | struct kobj_attribute *attr, char *buf) | |
3648 | { | |
9a305230 | 3649 | struct hstate *h = kobj_to_hstate(kobj, NULL); |
ae7a927d | 3650 | return sysfs_emit(buf, "%lu\n", h->nr_overcommit_huge_pages); |
a3437870 | 3651 | } |
adbe8726 | 3652 | |
a3437870 NA |
3653 | static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj, |
3654 | struct kobj_attribute *attr, const char *buf, size_t count) | |
3655 | { | |
3656 | int err; | |
3657 | unsigned long input; | |
9a305230 | 3658 | struct hstate *h = kobj_to_hstate(kobj, NULL); |
a3437870 | 3659 | |
bae7f4ae | 3660 | if (hstate_is_gigantic(h)) |
adbe8726 EM |
3661 | return -EINVAL; |
3662 | ||
3dbb95f7 | 3663 | err = kstrtoul(buf, 10, &input); |
a3437870 | 3664 | if (err) |
73ae31e5 | 3665 | return err; |
a3437870 | 3666 | |
db71ef79 | 3667 | spin_lock_irq(&hugetlb_lock); |
a3437870 | 3668 | h->nr_overcommit_huge_pages = input; |
db71ef79 | 3669 | spin_unlock_irq(&hugetlb_lock); |
a3437870 NA |
3670 | |
3671 | return count; | |
3672 | } | |
3673 | HSTATE_ATTR(nr_overcommit_hugepages); | |
3674 | ||
3675 | static ssize_t free_hugepages_show(struct kobject *kobj, | |
3676 | struct kobj_attribute *attr, char *buf) | |
3677 | { | |
9a305230 LS |
3678 | struct hstate *h; |
3679 | unsigned long free_huge_pages; | |
3680 | int nid; | |
3681 | ||
3682 | h = kobj_to_hstate(kobj, &nid); | |
3683 | if (nid == NUMA_NO_NODE) | |
3684 | free_huge_pages = h->free_huge_pages; | |
3685 | else | |
3686 | free_huge_pages = h->free_huge_pages_node[nid]; | |
3687 | ||
ae7a927d | 3688 | return sysfs_emit(buf, "%lu\n", free_huge_pages); |
a3437870 NA |
3689 | } |
3690 | HSTATE_ATTR_RO(free_hugepages); | |
3691 | ||
3692 | static ssize_t resv_hugepages_show(struct kobject *kobj, | |
3693 | struct kobj_attribute *attr, char *buf) | |
3694 | { | |
9a305230 | 3695 | struct hstate *h = kobj_to_hstate(kobj, NULL); |
ae7a927d | 3696 | return sysfs_emit(buf, "%lu\n", h->resv_huge_pages); |
a3437870 NA |
3697 | } |
3698 | HSTATE_ATTR_RO(resv_hugepages); | |
3699 | ||
3700 | static ssize_t surplus_hugepages_show(struct kobject *kobj, | |
3701 | struct kobj_attribute *attr, char *buf) | |
3702 | { | |
9a305230 LS |
3703 | struct hstate *h; |
3704 | unsigned long surplus_huge_pages; | |
3705 | int nid; | |
3706 | ||
3707 | h = kobj_to_hstate(kobj, &nid); | |
3708 | if (nid == NUMA_NO_NODE) | |
3709 | surplus_huge_pages = h->surplus_huge_pages; | |
3710 | else | |
3711 | surplus_huge_pages = h->surplus_huge_pages_node[nid]; | |
3712 | ||
ae7a927d | 3713 | return sysfs_emit(buf, "%lu\n", surplus_huge_pages); |
a3437870 NA |
3714 | } |
3715 | HSTATE_ATTR_RO(surplus_hugepages); | |
3716 | ||
79dfc695 MK |
3717 | static ssize_t demote_store(struct kobject *kobj, |
3718 | struct kobj_attribute *attr, const char *buf, size_t len) | |
3719 | { | |
3720 | unsigned long nr_demote; | |
3721 | unsigned long nr_available; | |
3722 | nodemask_t nodes_allowed, *n_mask; | |
3723 | struct hstate *h; | |
8eeda55f | 3724 | int err; |
79dfc695 MK |
3725 | int nid; |
3726 | ||
3727 | err = kstrtoul(buf, 10, &nr_demote); | |
3728 | if (err) | |
3729 | return err; | |
3730 | h = kobj_to_hstate(kobj, &nid); | |
3731 | ||
3732 | if (nid != NUMA_NO_NODE) { | |
3733 | init_nodemask_of_node(&nodes_allowed, nid); | |
3734 | n_mask = &nodes_allowed; | |
3735 | } else { | |
3736 | n_mask = &node_states[N_MEMORY]; | |
3737 | } | |
3738 | ||
3739 | /* Synchronize with other sysfs operations modifying huge pages */ | |
3740 | mutex_lock(&h->resize_lock); | |
3741 | spin_lock_irq(&hugetlb_lock); | |
3742 | ||
3743 | while (nr_demote) { | |
3744 | /* | |
3745 | * Check for available pages to demote each time thorough the | |
3746 | * loop as demote_pool_huge_page will drop hugetlb_lock. | |
79dfc695 MK |
3747 | */ |
3748 | if (nid != NUMA_NO_NODE) | |
3749 | nr_available = h->free_huge_pages_node[nid]; | |
3750 | else | |
3751 | nr_available = h->free_huge_pages; | |
3752 | nr_available -= h->resv_huge_pages; | |
3753 | if (!nr_available) | |
3754 | break; | |
3755 | ||
3756 | err = demote_pool_huge_page(h, n_mask); | |
3757 | if (err) | |
3758 | break; | |
3759 | ||
3760 | nr_demote--; | |
3761 | } | |
3762 | ||
3763 | spin_unlock_irq(&hugetlb_lock); | |
3764 | mutex_unlock(&h->resize_lock); | |
3765 | ||
3766 | if (err) | |
3767 | return err; | |
3768 | return len; | |
3769 | } | |
3770 | HSTATE_ATTR_WO(demote); | |
3771 | ||
3772 | static ssize_t demote_size_show(struct kobject *kobj, | |
3773 | struct kobj_attribute *attr, char *buf) | |
3774 | { | |
12658abf | 3775 | struct hstate *h = kobj_to_hstate(kobj, NULL); |
79dfc695 MK |
3776 | unsigned long demote_size = (PAGE_SIZE << h->demote_order) / SZ_1K; |
3777 | ||
3778 | return sysfs_emit(buf, "%lukB\n", demote_size); | |
3779 | } | |
3780 | ||
3781 | static ssize_t demote_size_store(struct kobject *kobj, | |
3782 | struct kobj_attribute *attr, | |
3783 | const char *buf, size_t count) | |
3784 | { | |
3785 | struct hstate *h, *demote_hstate; | |
3786 | unsigned long demote_size; | |
3787 | unsigned int demote_order; | |
79dfc695 MK |
3788 | |
3789 | demote_size = (unsigned long)memparse(buf, NULL); | |
3790 | ||
3791 | demote_hstate = size_to_hstate(demote_size); | |
3792 | if (!demote_hstate) | |
3793 | return -EINVAL; | |
3794 | demote_order = demote_hstate->order; | |
a01f4390 MK |
3795 | if (demote_order < HUGETLB_PAGE_ORDER) |
3796 | return -EINVAL; | |
79dfc695 MK |
3797 | |
3798 | /* demote order must be smaller than hstate order */ | |
12658abf | 3799 | h = kobj_to_hstate(kobj, NULL); |
79dfc695 MK |
3800 | if (demote_order >= h->order) |
3801 | return -EINVAL; | |
3802 | ||
3803 | /* resize_lock synchronizes access to demote size and writes */ | |
3804 | mutex_lock(&h->resize_lock); | |
3805 | h->demote_order = demote_order; | |
3806 | mutex_unlock(&h->resize_lock); | |
3807 | ||
3808 | return count; | |
3809 | } | |
3810 | HSTATE_ATTR(demote_size); | |
3811 | ||
a3437870 NA |
3812 | static struct attribute *hstate_attrs[] = { |
3813 | &nr_hugepages_attr.attr, | |
3814 | &nr_overcommit_hugepages_attr.attr, | |
3815 | &free_hugepages_attr.attr, | |
3816 | &resv_hugepages_attr.attr, | |
3817 | &surplus_hugepages_attr.attr, | |
06808b08 LS |
3818 | #ifdef CONFIG_NUMA |
3819 | &nr_hugepages_mempolicy_attr.attr, | |
3820 | #endif | |
a3437870 NA |
3821 | NULL, |
3822 | }; | |
3823 | ||
67e5ed96 | 3824 | static const struct attribute_group hstate_attr_group = { |
a3437870 NA |
3825 | .attrs = hstate_attrs, |
3826 | }; | |
3827 | ||
79dfc695 MK |
3828 | static struct attribute *hstate_demote_attrs[] = { |
3829 | &demote_size_attr.attr, | |
3830 | &demote_attr.attr, | |
3831 | NULL, | |
3832 | }; | |
3833 | ||
3834 | static const struct attribute_group hstate_demote_attr_group = { | |
3835 | .attrs = hstate_demote_attrs, | |
3836 | }; | |
3837 | ||
094e9539 JM |
3838 | static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent, |
3839 | struct kobject **hstate_kobjs, | |
67e5ed96 | 3840 | const struct attribute_group *hstate_attr_group) |
a3437870 NA |
3841 | { |
3842 | int retval; | |
972dc4de | 3843 | int hi = hstate_index(h); |
a3437870 | 3844 | |
9a305230 LS |
3845 | hstate_kobjs[hi] = kobject_create_and_add(h->name, parent); |
3846 | if (!hstate_kobjs[hi]) | |
a3437870 NA |
3847 | return -ENOMEM; |
3848 | ||
9a305230 | 3849 | retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group); |
cc2205a6 | 3850 | if (retval) { |
9a305230 | 3851 | kobject_put(hstate_kobjs[hi]); |
cc2205a6 | 3852 | hstate_kobjs[hi] = NULL; |
3a6bdda0 | 3853 | return retval; |
cc2205a6 | 3854 | } |
a3437870 | 3855 | |
79dfc695 | 3856 | if (h->demote_order) { |
01088a60 ML |
3857 | retval = sysfs_create_group(hstate_kobjs[hi], |
3858 | &hstate_demote_attr_group); | |
3859 | if (retval) { | |
79dfc695 | 3860 | pr_warn("HugeTLB unable to create demote interfaces for %s\n", h->name); |
01088a60 ML |
3861 | sysfs_remove_group(hstate_kobjs[hi], hstate_attr_group); |
3862 | kobject_put(hstate_kobjs[hi]); | |
3863 | hstate_kobjs[hi] = NULL; | |
3864 | return retval; | |
3865 | } | |
79dfc695 MK |
3866 | } |
3867 | ||
01088a60 | 3868 | return 0; |
a3437870 NA |
3869 | } |
3870 | ||
9a305230 | 3871 | #ifdef CONFIG_NUMA |
a4a00b45 | 3872 | static bool hugetlb_sysfs_initialized __ro_after_init; |
9a305230 LS |
3873 | |
3874 | /* | |
3875 | * node_hstate/s - associate per node hstate attributes, via their kobjects, | |
10fbcf4c KS |
3876 | * with node devices in node_devices[] using a parallel array. The array |
3877 | * index of a node device or _hstate == node id. | |
3878 | * This is here to avoid any static dependency of the node device driver, in | |
9a305230 LS |
3879 | * the base kernel, on the hugetlb module. |
3880 | */ | |
3881 | struct node_hstate { | |
3882 | struct kobject *hugepages_kobj; | |
3883 | struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; | |
3884 | }; | |
b4e289a6 | 3885 | static struct node_hstate node_hstates[MAX_NUMNODES]; |
9a305230 LS |
3886 | |
3887 | /* | |
10fbcf4c | 3888 | * A subset of global hstate attributes for node devices |
9a305230 LS |
3889 | */ |
3890 | static struct attribute *per_node_hstate_attrs[] = { | |
3891 | &nr_hugepages_attr.attr, | |
3892 | &free_hugepages_attr.attr, | |
3893 | &surplus_hugepages_attr.attr, | |
3894 | NULL, | |
3895 | }; | |
3896 | ||
67e5ed96 | 3897 | static const struct attribute_group per_node_hstate_attr_group = { |
9a305230 LS |
3898 | .attrs = per_node_hstate_attrs, |
3899 | }; | |
3900 | ||
3901 | /* | |
10fbcf4c | 3902 | * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj. |
9a305230 LS |
3903 | * Returns node id via non-NULL nidp. |
3904 | */ | |
3905 | static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) | |
3906 | { | |
3907 | int nid; | |
3908 | ||
3909 | for (nid = 0; nid < nr_node_ids; nid++) { | |
3910 | struct node_hstate *nhs = &node_hstates[nid]; | |
3911 | int i; | |
3912 | for (i = 0; i < HUGE_MAX_HSTATE; i++) | |
3913 | if (nhs->hstate_kobjs[i] == kobj) { | |
3914 | if (nidp) | |
3915 | *nidp = nid; | |
3916 | return &hstates[i]; | |
3917 | } | |
3918 | } | |
3919 | ||
3920 | BUG(); | |
3921 | return NULL; | |
3922 | } | |
3923 | ||
3924 | /* | |
10fbcf4c | 3925 | * Unregister hstate attributes from a single node device. |
9a305230 LS |
3926 | * No-op if no hstate attributes attached. |
3927 | */ | |
a4a00b45 | 3928 | void hugetlb_unregister_node(struct node *node) |
9a305230 LS |
3929 | { |
3930 | struct hstate *h; | |
10fbcf4c | 3931 | struct node_hstate *nhs = &node_hstates[node->dev.id]; |
9a305230 LS |
3932 | |
3933 | if (!nhs->hugepages_kobj) | |
9b5e5d0f | 3934 | return; /* no hstate attributes */ |
9a305230 | 3935 | |
972dc4de AK |
3936 | for_each_hstate(h) { |
3937 | int idx = hstate_index(h); | |
01088a60 ML |
3938 | struct kobject *hstate_kobj = nhs->hstate_kobjs[idx]; |
3939 | ||
3940 | if (!hstate_kobj) | |
3941 | continue; | |
3942 | if (h->demote_order) | |
3943 | sysfs_remove_group(hstate_kobj, &hstate_demote_attr_group); | |
3944 | sysfs_remove_group(hstate_kobj, &per_node_hstate_attr_group); | |
3945 | kobject_put(hstate_kobj); | |
3946 | nhs->hstate_kobjs[idx] = NULL; | |
972dc4de | 3947 | } |
9a305230 LS |
3948 | |
3949 | kobject_put(nhs->hugepages_kobj); | |
3950 | nhs->hugepages_kobj = NULL; | |
3951 | } | |
3952 | ||
9a305230 LS |
3953 | |
3954 | /* | |
10fbcf4c | 3955 | * Register hstate attributes for a single node device. |
9a305230 LS |
3956 | * No-op if attributes already registered. |
3957 | */ | |
a4a00b45 | 3958 | void hugetlb_register_node(struct node *node) |
9a305230 LS |
3959 | { |
3960 | struct hstate *h; | |
10fbcf4c | 3961 | struct node_hstate *nhs = &node_hstates[node->dev.id]; |
9a305230 LS |
3962 | int err; |
3963 | ||
a4a00b45 MS |
3964 | if (!hugetlb_sysfs_initialized) |
3965 | return; | |
3966 | ||
9a305230 LS |
3967 | if (nhs->hugepages_kobj) |
3968 | return; /* already allocated */ | |
3969 | ||
3970 | nhs->hugepages_kobj = kobject_create_and_add("hugepages", | |
10fbcf4c | 3971 | &node->dev.kobj); |
9a305230 LS |
3972 | if (!nhs->hugepages_kobj) |
3973 | return; | |
3974 | ||
3975 | for_each_hstate(h) { | |
3976 | err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj, | |
3977 | nhs->hstate_kobjs, | |
3978 | &per_node_hstate_attr_group); | |
3979 | if (err) { | |
282f4214 | 3980 | pr_err("HugeTLB: Unable to add hstate %s for node %d\n", |
ffb22af5 | 3981 | h->name, node->dev.id); |
9a305230 LS |
3982 | hugetlb_unregister_node(node); |
3983 | break; | |
3984 | } | |
3985 | } | |
3986 | } | |
3987 | ||
3988 | /* | |
9b5e5d0f | 3989 | * hugetlb init time: register hstate attributes for all registered node |
10fbcf4c KS |
3990 | * devices of nodes that have memory. All on-line nodes should have |
3991 | * registered their associated device by this time. | |
9a305230 | 3992 | */ |
7d9ca000 | 3993 | static void __init hugetlb_register_all_nodes(void) |
9a305230 LS |
3994 | { |
3995 | int nid; | |
3996 | ||
a4a00b45 | 3997 | for_each_online_node(nid) |
b958d4d0 | 3998 | hugetlb_register_node(node_devices[nid]); |
9a305230 LS |
3999 | } |
4000 | #else /* !CONFIG_NUMA */ | |
4001 | ||
4002 | static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) | |
4003 | { | |
4004 | BUG(); | |
4005 | if (nidp) | |
4006 | *nidp = -1; | |
4007 | return NULL; | |
4008 | } | |
4009 | ||
9a305230 LS |
4010 | static void hugetlb_register_all_nodes(void) { } |
4011 | ||
4012 | #endif | |
4013 | ||
263b8998 ML |
4014 | #ifdef CONFIG_CMA |
4015 | static void __init hugetlb_cma_check(void); | |
4016 | #else | |
4017 | static inline __init void hugetlb_cma_check(void) | |
4018 | { | |
4019 | } | |
4020 | #endif | |
4021 | ||
a4a00b45 MS |
4022 | static void __init hugetlb_sysfs_init(void) |
4023 | { | |
4024 | struct hstate *h; | |
4025 | int err; | |
4026 | ||
4027 | hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj); | |
4028 | if (!hugepages_kobj) | |
4029 | return; | |
4030 | ||
4031 | for_each_hstate(h) { | |
4032 | err = hugetlb_sysfs_add_hstate(h, hugepages_kobj, | |
4033 | hstate_kobjs, &hstate_attr_group); | |
4034 | if (err) | |
4035 | pr_err("HugeTLB: Unable to add hstate %s", h->name); | |
4036 | } | |
4037 | ||
4038 | #ifdef CONFIG_NUMA | |
4039 | hugetlb_sysfs_initialized = true; | |
4040 | #endif | |
4041 | hugetlb_register_all_nodes(); | |
4042 | } | |
4043 | ||
a3437870 NA |
4044 | static int __init hugetlb_init(void) |
4045 | { | |
8382d914 DB |
4046 | int i; |
4047 | ||
d6995da3 MK |
4048 | BUILD_BUG_ON(sizeof_field(struct page, private) * BITS_PER_BYTE < |
4049 | __NR_HPAGEFLAGS); | |
4050 | ||
c2833a5b MK |
4051 | if (!hugepages_supported()) { |
4052 | if (hugetlb_max_hstate || default_hstate_max_huge_pages) | |
4053 | pr_warn("HugeTLB: huge pages not supported, ignoring associated command-line parameters\n"); | |
0ef89d25 | 4054 | return 0; |
c2833a5b | 4055 | } |
a3437870 | 4056 | |
282f4214 MK |
4057 | /* |
4058 | * Make sure HPAGE_SIZE (HUGETLB_PAGE_ORDER) hstate exists. Some | |
4059 | * architectures depend on setup being done here. | |
4060 | */ | |
4061 | hugetlb_add_hstate(HUGETLB_PAGE_ORDER); | |
4062 | if (!parsed_default_hugepagesz) { | |
4063 | /* | |
4064 | * If we did not parse a default huge page size, set | |
4065 | * default_hstate_idx to HPAGE_SIZE hstate. And, if the | |
4066 | * number of huge pages for this default size was implicitly | |
4067 | * specified, set that here as well. | |
4068 | * Note that the implicit setting will overwrite an explicit | |
4069 | * setting. A warning will be printed in this case. | |
4070 | */ | |
4071 | default_hstate_idx = hstate_index(size_to_hstate(HPAGE_SIZE)); | |
4072 | if (default_hstate_max_huge_pages) { | |
4073 | if (default_hstate.max_huge_pages) { | |
4074 | char buf[32]; | |
4075 | ||
4076 | string_get_size(huge_page_size(&default_hstate), | |
4077 | 1, STRING_UNITS_2, buf, 32); | |
4078 | pr_warn("HugeTLB: Ignoring hugepages=%lu associated with %s page size\n", | |
4079 | default_hstate.max_huge_pages, buf); | |
4080 | pr_warn("HugeTLB: Using hugepages=%lu for number of default huge pages\n", | |
4081 | default_hstate_max_huge_pages); | |
4082 | } | |
4083 | default_hstate.max_huge_pages = | |
4084 | default_hstate_max_huge_pages; | |
b5389086 | 4085 | |
0a7a0f6f | 4086 | for_each_online_node(i) |
b5389086 ZY |
4087 | default_hstate.max_huge_pages_node[i] = |
4088 | default_hugepages_in_node[i]; | |
d715cf80 | 4089 | } |
f8b74815 | 4090 | } |
a3437870 | 4091 | |
cf11e85f | 4092 | hugetlb_cma_check(); |
a3437870 | 4093 | hugetlb_init_hstates(); |
aa888a74 | 4094 | gather_bootmem_prealloc(); |
a3437870 NA |
4095 | report_hugepages(); |
4096 | ||
4097 | hugetlb_sysfs_init(); | |
7179e7bf | 4098 | hugetlb_cgroup_file_init(); |
9a305230 | 4099 | |
8382d914 DB |
4100 | #ifdef CONFIG_SMP |
4101 | num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus()); | |
4102 | #else | |
4103 | num_fault_mutexes = 1; | |
4104 | #endif | |
c672c7f2 | 4105 | hugetlb_fault_mutex_table = |
6da2ec56 KC |
4106 | kmalloc_array(num_fault_mutexes, sizeof(struct mutex), |
4107 | GFP_KERNEL); | |
c672c7f2 | 4108 | BUG_ON(!hugetlb_fault_mutex_table); |
8382d914 DB |
4109 | |
4110 | for (i = 0; i < num_fault_mutexes; i++) | |
c672c7f2 | 4111 | mutex_init(&hugetlb_fault_mutex_table[i]); |
a3437870 NA |
4112 | return 0; |
4113 | } | |
3e89e1c5 | 4114 | subsys_initcall(hugetlb_init); |
a3437870 | 4115 | |
ae94da89 MK |
4116 | /* Overwritten by architectures with more huge page sizes */ |
4117 | bool __init __attribute((weak)) arch_hugetlb_valid_size(unsigned long size) | |
9fee021d | 4118 | { |
ae94da89 | 4119 | return size == HPAGE_SIZE; |
9fee021d VT |
4120 | } |
4121 | ||
d00181b9 | 4122 | void __init hugetlb_add_hstate(unsigned int order) |
a3437870 NA |
4123 | { |
4124 | struct hstate *h; | |
8faa8b07 AK |
4125 | unsigned long i; |
4126 | ||
a3437870 | 4127 | if (size_to_hstate(PAGE_SIZE << order)) { |
a3437870 NA |
4128 | return; |
4129 | } | |
47d38344 | 4130 | BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE); |
a3437870 | 4131 | BUG_ON(order == 0); |
47d38344 | 4132 | h = &hstates[hugetlb_max_hstate++]; |
29383967 | 4133 | mutex_init(&h->resize_lock); |
a3437870 | 4134 | h->order = order; |
aca78307 | 4135 | h->mask = ~(huge_page_size(h) - 1); |
8faa8b07 AK |
4136 | for (i = 0; i < MAX_NUMNODES; ++i) |
4137 | INIT_LIST_HEAD(&h->hugepage_freelists[i]); | |
0edaecfa | 4138 | INIT_LIST_HEAD(&h->hugepage_activelist); |
54f18d35 AM |
4139 | h->next_nid_to_alloc = first_memory_node; |
4140 | h->next_nid_to_free = first_memory_node; | |
a3437870 | 4141 | snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB", |
c2c3a60a | 4142 | huge_page_size(h)/SZ_1K); |
8faa8b07 | 4143 | |
a3437870 NA |
4144 | parsed_hstate = h; |
4145 | } | |
4146 | ||
b5389086 ZY |
4147 | bool __init __weak hugetlb_node_alloc_supported(void) |
4148 | { | |
4149 | return true; | |
4150 | } | |
f87442f4 PL |
4151 | |
4152 | static void __init hugepages_clear_pages_in_node(void) | |
4153 | { | |
4154 | if (!hugetlb_max_hstate) { | |
4155 | default_hstate_max_huge_pages = 0; | |
4156 | memset(default_hugepages_in_node, 0, | |
10395680 | 4157 | sizeof(default_hugepages_in_node)); |
f87442f4 PL |
4158 | } else { |
4159 | parsed_hstate->max_huge_pages = 0; | |
4160 | memset(parsed_hstate->max_huge_pages_node, 0, | |
10395680 | 4161 | sizeof(parsed_hstate->max_huge_pages_node)); |
f87442f4 PL |
4162 | } |
4163 | } | |
4164 | ||
282f4214 MK |
4165 | /* |
4166 | * hugepages command line processing | |
4167 | * hugepages normally follows a valid hugepagsz or default_hugepagsz | |
4168 | * specification. If not, ignore the hugepages value. hugepages can also | |
4169 | * be the first huge page command line option in which case it implicitly | |
4170 | * specifies the number of huge pages for the default size. | |
4171 | */ | |
4172 | static int __init hugepages_setup(char *s) | |
a3437870 NA |
4173 | { |
4174 | unsigned long *mhp; | |
8faa8b07 | 4175 | static unsigned long *last_mhp; |
b5389086 ZY |
4176 | int node = NUMA_NO_NODE; |
4177 | int count; | |
4178 | unsigned long tmp; | |
4179 | char *p = s; | |
a3437870 | 4180 | |
9fee021d | 4181 | if (!parsed_valid_hugepagesz) { |
282f4214 | 4182 | pr_warn("HugeTLB: hugepages=%s does not follow a valid hugepagesz, ignoring\n", s); |
9fee021d | 4183 | parsed_valid_hugepagesz = true; |
f81f6e4b | 4184 | return 1; |
9fee021d | 4185 | } |
282f4214 | 4186 | |
a3437870 | 4187 | /* |
282f4214 MK |
4188 | * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter |
4189 | * yet, so this hugepages= parameter goes to the "default hstate". | |
4190 | * Otherwise, it goes with the previously parsed hugepagesz or | |
4191 | * default_hugepagesz. | |
a3437870 | 4192 | */ |
9fee021d | 4193 | else if (!hugetlb_max_hstate) |
a3437870 NA |
4194 | mhp = &default_hstate_max_huge_pages; |
4195 | else | |
4196 | mhp = &parsed_hstate->max_huge_pages; | |
4197 | ||
8faa8b07 | 4198 | if (mhp == last_mhp) { |
282f4214 | 4199 | pr_warn("HugeTLB: hugepages= specified twice without interleaving hugepagesz=, ignoring hugepages=%s\n", s); |
f81f6e4b | 4200 | return 1; |
8faa8b07 AK |
4201 | } |
4202 | ||
b5389086 ZY |
4203 | while (*p) { |
4204 | count = 0; | |
4205 | if (sscanf(p, "%lu%n", &tmp, &count) != 1) | |
4206 | goto invalid; | |
4207 | /* Parameter is node format */ | |
4208 | if (p[count] == ':') { | |
4209 | if (!hugetlb_node_alloc_supported()) { | |
4210 | pr_warn("HugeTLB: architecture can't support node specific alloc, ignoring!\n"); | |
f81f6e4b | 4211 | return 1; |
b5389086 | 4212 | } |
0a7a0f6f | 4213 | if (tmp >= MAX_NUMNODES || !node_online(tmp)) |
e79ce983 | 4214 | goto invalid; |
0a7a0f6f | 4215 | node = array_index_nospec(tmp, MAX_NUMNODES); |
b5389086 | 4216 | p += count + 1; |
b5389086 ZY |
4217 | /* Parse hugepages */ |
4218 | if (sscanf(p, "%lu%n", &tmp, &count) != 1) | |
4219 | goto invalid; | |
4220 | if (!hugetlb_max_hstate) | |
4221 | default_hugepages_in_node[node] = tmp; | |
4222 | else | |
4223 | parsed_hstate->max_huge_pages_node[node] = tmp; | |
4224 | *mhp += tmp; | |
4225 | /* Go to parse next node*/ | |
4226 | if (p[count] == ',') | |
4227 | p += count + 1; | |
4228 | else | |
4229 | break; | |
4230 | } else { | |
4231 | if (p != s) | |
4232 | goto invalid; | |
4233 | *mhp = tmp; | |
4234 | break; | |
4235 | } | |
4236 | } | |
a3437870 | 4237 | |
8faa8b07 AK |
4238 | /* |
4239 | * Global state is always initialized later in hugetlb_init. | |
04adbc3f | 4240 | * But we need to allocate gigantic hstates here early to still |
8faa8b07 AK |
4241 | * use the bootmem allocator. |
4242 | */ | |
04adbc3f | 4243 | if (hugetlb_max_hstate && hstate_is_gigantic(parsed_hstate)) |
8faa8b07 AK |
4244 | hugetlb_hstate_alloc_pages(parsed_hstate); |
4245 | ||
4246 | last_mhp = mhp; | |
4247 | ||
a3437870 | 4248 | return 1; |
b5389086 ZY |
4249 | |
4250 | invalid: | |
4251 | pr_warn("HugeTLB: Invalid hugepages parameter %s\n", p); | |
f87442f4 | 4252 | hugepages_clear_pages_in_node(); |
f81f6e4b | 4253 | return 1; |
a3437870 | 4254 | } |
282f4214 | 4255 | __setup("hugepages=", hugepages_setup); |
e11bfbfc | 4256 | |
282f4214 MK |
4257 | /* |
4258 | * hugepagesz command line processing | |
4259 | * A specific huge page size can only be specified once with hugepagesz. | |
4260 | * hugepagesz is followed by hugepages on the command line. The global | |
4261 | * variable 'parsed_valid_hugepagesz' is used to determine if prior | |
4262 | * hugepagesz argument was valid. | |
4263 | */ | |
359f2544 | 4264 | static int __init hugepagesz_setup(char *s) |
e11bfbfc | 4265 | { |
359f2544 | 4266 | unsigned long size; |
282f4214 MK |
4267 | struct hstate *h; |
4268 | ||
4269 | parsed_valid_hugepagesz = false; | |
359f2544 MK |
4270 | size = (unsigned long)memparse(s, NULL); |
4271 | ||
4272 | if (!arch_hugetlb_valid_size(size)) { | |
282f4214 | 4273 | pr_err("HugeTLB: unsupported hugepagesz=%s\n", s); |
f81f6e4b | 4274 | return 1; |
359f2544 MK |
4275 | } |
4276 | ||
282f4214 MK |
4277 | h = size_to_hstate(size); |
4278 | if (h) { | |
4279 | /* | |
4280 | * hstate for this size already exists. This is normally | |
4281 | * an error, but is allowed if the existing hstate is the | |
4282 | * default hstate. More specifically, it is only allowed if | |
4283 | * the number of huge pages for the default hstate was not | |
4284 | * previously specified. | |
4285 | */ | |
4286 | if (!parsed_default_hugepagesz || h != &default_hstate || | |
4287 | default_hstate.max_huge_pages) { | |
4288 | pr_warn("HugeTLB: hugepagesz=%s specified twice, ignoring\n", s); | |
f81f6e4b | 4289 | return 1; |
282f4214 MK |
4290 | } |
4291 | ||
4292 | /* | |
4293 | * No need to call hugetlb_add_hstate() as hstate already | |
4294 | * exists. But, do set parsed_hstate so that a following | |
4295 | * hugepages= parameter will be applied to this hstate. | |
4296 | */ | |
4297 | parsed_hstate = h; | |
4298 | parsed_valid_hugepagesz = true; | |
4299 | return 1; | |
38237830 MK |
4300 | } |
4301 | ||
359f2544 | 4302 | hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT); |
282f4214 | 4303 | parsed_valid_hugepagesz = true; |
e11bfbfc NP |
4304 | return 1; |
4305 | } | |
359f2544 MK |
4306 | __setup("hugepagesz=", hugepagesz_setup); |
4307 | ||
282f4214 MK |
4308 | /* |
4309 | * default_hugepagesz command line input | |
4310 | * Only one instance of default_hugepagesz allowed on command line. | |
4311 | */ | |
ae94da89 | 4312 | static int __init default_hugepagesz_setup(char *s) |
e11bfbfc | 4313 | { |
ae94da89 | 4314 | unsigned long size; |
b5389086 | 4315 | int i; |
ae94da89 | 4316 | |
282f4214 | 4317 | parsed_valid_hugepagesz = false; |
282f4214 MK |
4318 | if (parsed_default_hugepagesz) { |
4319 | pr_err("HugeTLB: default_hugepagesz previously specified, ignoring %s\n", s); | |
f81f6e4b | 4320 | return 1; |
282f4214 MK |
4321 | } |
4322 | ||
ae94da89 MK |
4323 | size = (unsigned long)memparse(s, NULL); |
4324 | ||
4325 | if (!arch_hugetlb_valid_size(size)) { | |
282f4214 | 4326 | pr_err("HugeTLB: unsupported default_hugepagesz=%s\n", s); |
f81f6e4b | 4327 | return 1; |
ae94da89 MK |
4328 | } |
4329 | ||
282f4214 MK |
4330 | hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT); |
4331 | parsed_valid_hugepagesz = true; | |
4332 | parsed_default_hugepagesz = true; | |
4333 | default_hstate_idx = hstate_index(size_to_hstate(size)); | |
4334 | ||
4335 | /* | |
4336 | * The number of default huge pages (for this size) could have been | |
4337 | * specified as the first hugetlb parameter: hugepages=X. If so, | |
4338 | * then default_hstate_max_huge_pages is set. If the default huge | |
4339 | * page size is gigantic (>= MAX_ORDER), then the pages must be | |
4340 | * allocated here from bootmem allocator. | |
4341 | */ | |
4342 | if (default_hstate_max_huge_pages) { | |
4343 | default_hstate.max_huge_pages = default_hstate_max_huge_pages; | |
0a7a0f6f | 4344 | for_each_online_node(i) |
b5389086 ZY |
4345 | default_hstate.max_huge_pages_node[i] = |
4346 | default_hugepages_in_node[i]; | |
282f4214 MK |
4347 | if (hstate_is_gigantic(&default_hstate)) |
4348 | hugetlb_hstate_alloc_pages(&default_hstate); | |
4349 | default_hstate_max_huge_pages = 0; | |
4350 | } | |
4351 | ||
e11bfbfc NP |
4352 | return 1; |
4353 | } | |
ae94da89 | 4354 | __setup("default_hugepagesz=", default_hugepagesz_setup); |
a3437870 | 4355 | |
d2226ebd FT |
4356 | static nodemask_t *policy_mbind_nodemask(gfp_t gfp) |
4357 | { | |
4358 | #ifdef CONFIG_NUMA | |
4359 | struct mempolicy *mpol = get_task_policy(current); | |
4360 | ||
4361 | /* | |
4362 | * Only enforce MPOL_BIND policy which overlaps with cpuset policy | |
4363 | * (from policy_nodemask) specifically for hugetlb case | |
4364 | */ | |
4365 | if (mpol->mode == MPOL_BIND && | |
4366 | (apply_policy_zone(mpol, gfp_zone(gfp)) && | |
4367 | cpuset_nodemask_valid_mems_allowed(&mpol->nodes))) | |
4368 | return &mpol->nodes; | |
4369 | #endif | |
4370 | return NULL; | |
4371 | } | |
4372 | ||
8ca39e68 | 4373 | static unsigned int allowed_mems_nr(struct hstate *h) |
8a213460 NA |
4374 | { |
4375 | int node; | |
4376 | unsigned int nr = 0; | |
d2226ebd | 4377 | nodemask_t *mbind_nodemask; |
8ca39e68 MS |
4378 | unsigned int *array = h->free_huge_pages_node; |
4379 | gfp_t gfp_mask = htlb_alloc_mask(h); | |
4380 | ||
d2226ebd | 4381 | mbind_nodemask = policy_mbind_nodemask(gfp_mask); |
8ca39e68 | 4382 | for_each_node_mask(node, cpuset_current_mems_allowed) { |
d2226ebd | 4383 | if (!mbind_nodemask || node_isset(node, *mbind_nodemask)) |
8ca39e68 MS |
4384 | nr += array[node]; |
4385 | } | |
8a213460 NA |
4386 | |
4387 | return nr; | |
4388 | } | |
4389 | ||
4390 | #ifdef CONFIG_SYSCTL | |
17743798 MS |
4391 | static int proc_hugetlb_doulongvec_minmax(struct ctl_table *table, int write, |
4392 | void *buffer, size_t *length, | |
4393 | loff_t *ppos, unsigned long *out) | |
4394 | { | |
4395 | struct ctl_table dup_table; | |
4396 | ||
4397 | /* | |
4398 | * In order to avoid races with __do_proc_doulongvec_minmax(), we | |
4399 | * can duplicate the @table and alter the duplicate of it. | |
4400 | */ | |
4401 | dup_table = *table; | |
4402 | dup_table.data = out; | |
4403 | ||
4404 | return proc_doulongvec_minmax(&dup_table, write, buffer, length, ppos); | |
4405 | } | |
4406 | ||
06808b08 LS |
4407 | static int hugetlb_sysctl_handler_common(bool obey_mempolicy, |
4408 | struct ctl_table *table, int write, | |
32927393 | 4409 | void *buffer, size_t *length, loff_t *ppos) |
1da177e4 | 4410 | { |
e5ff2159 | 4411 | struct hstate *h = &default_hstate; |
238d3c13 | 4412 | unsigned long tmp = h->max_huge_pages; |
08d4a246 | 4413 | int ret; |
e5ff2159 | 4414 | |
457c1b27 | 4415 | if (!hugepages_supported()) |
86613628 | 4416 | return -EOPNOTSUPP; |
457c1b27 | 4417 | |
17743798 MS |
4418 | ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos, |
4419 | &tmp); | |
08d4a246 MH |
4420 | if (ret) |
4421 | goto out; | |
e5ff2159 | 4422 | |
238d3c13 DR |
4423 | if (write) |
4424 | ret = __nr_hugepages_store_common(obey_mempolicy, h, | |
4425 | NUMA_NO_NODE, tmp, *length); | |
08d4a246 MH |
4426 | out: |
4427 | return ret; | |
1da177e4 | 4428 | } |
396faf03 | 4429 | |
06808b08 | 4430 | int hugetlb_sysctl_handler(struct ctl_table *table, int write, |
32927393 | 4431 | void *buffer, size_t *length, loff_t *ppos) |
06808b08 LS |
4432 | { |
4433 | ||
4434 | return hugetlb_sysctl_handler_common(false, table, write, | |
4435 | buffer, length, ppos); | |
4436 | } | |
4437 | ||
4438 | #ifdef CONFIG_NUMA | |
4439 | int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write, | |
32927393 | 4440 | void *buffer, size_t *length, loff_t *ppos) |
06808b08 LS |
4441 | { |
4442 | return hugetlb_sysctl_handler_common(true, table, write, | |
4443 | buffer, length, ppos); | |
4444 | } | |
4445 | #endif /* CONFIG_NUMA */ | |
4446 | ||
a3d0c6aa | 4447 | int hugetlb_overcommit_handler(struct ctl_table *table, int write, |
32927393 | 4448 | void *buffer, size_t *length, loff_t *ppos) |
a3d0c6aa | 4449 | { |
a5516438 | 4450 | struct hstate *h = &default_hstate; |
e5ff2159 | 4451 | unsigned long tmp; |
08d4a246 | 4452 | int ret; |
e5ff2159 | 4453 | |
457c1b27 | 4454 | if (!hugepages_supported()) |
86613628 | 4455 | return -EOPNOTSUPP; |
457c1b27 | 4456 | |
c033a93c | 4457 | tmp = h->nr_overcommit_huge_pages; |
e5ff2159 | 4458 | |
bae7f4ae | 4459 | if (write && hstate_is_gigantic(h)) |
adbe8726 EM |
4460 | return -EINVAL; |
4461 | ||
17743798 MS |
4462 | ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos, |
4463 | &tmp); | |
08d4a246 MH |
4464 | if (ret) |
4465 | goto out; | |
e5ff2159 AK |
4466 | |
4467 | if (write) { | |
db71ef79 | 4468 | spin_lock_irq(&hugetlb_lock); |
e5ff2159 | 4469 | h->nr_overcommit_huge_pages = tmp; |
db71ef79 | 4470 | spin_unlock_irq(&hugetlb_lock); |
e5ff2159 | 4471 | } |
08d4a246 MH |
4472 | out: |
4473 | return ret; | |
a3d0c6aa NA |
4474 | } |
4475 | ||
1da177e4 LT |
4476 | #endif /* CONFIG_SYSCTL */ |
4477 | ||
e1759c21 | 4478 | void hugetlb_report_meminfo(struct seq_file *m) |
1da177e4 | 4479 | { |
fcb2b0c5 RG |
4480 | struct hstate *h; |
4481 | unsigned long total = 0; | |
4482 | ||
457c1b27 NA |
4483 | if (!hugepages_supported()) |
4484 | return; | |
fcb2b0c5 RG |
4485 | |
4486 | for_each_hstate(h) { | |
4487 | unsigned long count = h->nr_huge_pages; | |
4488 | ||
aca78307 | 4489 | total += huge_page_size(h) * count; |
fcb2b0c5 RG |
4490 | |
4491 | if (h == &default_hstate) | |
4492 | seq_printf(m, | |
4493 | "HugePages_Total: %5lu\n" | |
4494 | "HugePages_Free: %5lu\n" | |
4495 | "HugePages_Rsvd: %5lu\n" | |
4496 | "HugePages_Surp: %5lu\n" | |
4497 | "Hugepagesize: %8lu kB\n", | |
4498 | count, | |
4499 | h->free_huge_pages, | |
4500 | h->resv_huge_pages, | |
4501 | h->surplus_huge_pages, | |
aca78307 | 4502 | huge_page_size(h) / SZ_1K); |
fcb2b0c5 RG |
4503 | } |
4504 | ||
aca78307 | 4505 | seq_printf(m, "Hugetlb: %8lu kB\n", total / SZ_1K); |
1da177e4 LT |
4506 | } |
4507 | ||
7981593b | 4508 | int hugetlb_report_node_meminfo(char *buf, int len, int nid) |
1da177e4 | 4509 | { |
a5516438 | 4510 | struct hstate *h = &default_hstate; |
7981593b | 4511 | |
457c1b27 NA |
4512 | if (!hugepages_supported()) |
4513 | return 0; | |
7981593b JP |
4514 | |
4515 | return sysfs_emit_at(buf, len, | |
4516 | "Node %d HugePages_Total: %5u\n" | |
4517 | "Node %d HugePages_Free: %5u\n" | |
4518 | "Node %d HugePages_Surp: %5u\n", | |
4519 | nid, h->nr_huge_pages_node[nid], | |
4520 | nid, h->free_huge_pages_node[nid], | |
4521 | nid, h->surplus_huge_pages_node[nid]); | |
1da177e4 LT |
4522 | } |
4523 | ||
dcadcf1c | 4524 | void hugetlb_show_meminfo_node(int nid) |
949f7ec5 DR |
4525 | { |
4526 | struct hstate *h; | |
949f7ec5 | 4527 | |
457c1b27 NA |
4528 | if (!hugepages_supported()) |
4529 | return; | |
4530 | ||
dcadcf1c GL |
4531 | for_each_hstate(h) |
4532 | printk("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n", | |
4533 | nid, | |
4534 | h->nr_huge_pages_node[nid], | |
4535 | h->free_huge_pages_node[nid], | |
4536 | h->surplus_huge_pages_node[nid], | |
4537 | huge_page_size(h) / SZ_1K); | |
949f7ec5 DR |
4538 | } |
4539 | ||
5d317b2b NH |
4540 | void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm) |
4541 | { | |
4542 | seq_printf(m, "HugetlbPages:\t%8lu kB\n", | |
4543 | atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10)); | |
4544 | } | |
4545 | ||
1da177e4 LT |
4546 | /* Return the number pages of memory we physically have, in PAGE_SIZE units. */ |
4547 | unsigned long hugetlb_total_pages(void) | |
4548 | { | |
d0028588 WL |
4549 | struct hstate *h; |
4550 | unsigned long nr_total_pages = 0; | |
4551 | ||
4552 | for_each_hstate(h) | |
4553 | nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h); | |
4554 | return nr_total_pages; | |
1da177e4 | 4555 | } |
1da177e4 | 4556 | |
a5516438 | 4557 | static int hugetlb_acct_memory(struct hstate *h, long delta) |
fc1b8a73 MG |
4558 | { |
4559 | int ret = -ENOMEM; | |
4560 | ||
0aa7f354 ML |
4561 | if (!delta) |
4562 | return 0; | |
4563 | ||
db71ef79 | 4564 | spin_lock_irq(&hugetlb_lock); |
fc1b8a73 MG |
4565 | /* |
4566 | * When cpuset is configured, it breaks the strict hugetlb page | |
4567 | * reservation as the accounting is done on a global variable. Such | |
4568 | * reservation is completely rubbish in the presence of cpuset because | |
4569 | * the reservation is not checked against page availability for the | |
4570 | * current cpuset. Application can still potentially OOM'ed by kernel | |
4571 | * with lack of free htlb page in cpuset that the task is in. | |
4572 | * Attempt to enforce strict accounting with cpuset is almost | |
4573 | * impossible (or too ugly) because cpuset is too fluid that | |
4574 | * task or memory node can be dynamically moved between cpusets. | |
4575 | * | |
4576 | * The change of semantics for shared hugetlb mapping with cpuset is | |
4577 | * undesirable. However, in order to preserve some of the semantics, | |
4578 | * we fall back to check against current free page availability as | |
4579 | * a best attempt and hopefully to minimize the impact of changing | |
4580 | * semantics that cpuset has. | |
8ca39e68 MS |
4581 | * |
4582 | * Apart from cpuset, we also have memory policy mechanism that | |
4583 | * also determines from which node the kernel will allocate memory | |
4584 | * in a NUMA system. So similar to cpuset, we also should consider | |
4585 | * the memory policy of the current task. Similar to the description | |
4586 | * above. | |
fc1b8a73 MG |
4587 | */ |
4588 | if (delta > 0) { | |
a5516438 | 4589 | if (gather_surplus_pages(h, delta) < 0) |
fc1b8a73 MG |
4590 | goto out; |
4591 | ||
8ca39e68 | 4592 | if (delta > allowed_mems_nr(h)) { |
a5516438 | 4593 | return_unused_surplus_pages(h, delta); |
fc1b8a73 MG |
4594 | goto out; |
4595 | } | |
4596 | } | |
4597 | ||
4598 | ret = 0; | |
4599 | if (delta < 0) | |
a5516438 | 4600 | return_unused_surplus_pages(h, (unsigned long) -delta); |
fc1b8a73 MG |
4601 | |
4602 | out: | |
db71ef79 | 4603 | spin_unlock_irq(&hugetlb_lock); |
fc1b8a73 MG |
4604 | return ret; |
4605 | } | |
4606 | ||
84afd99b AW |
4607 | static void hugetlb_vm_op_open(struct vm_area_struct *vma) |
4608 | { | |
f522c3ac | 4609 | struct resv_map *resv = vma_resv_map(vma); |
84afd99b AW |
4610 | |
4611 | /* | |
612b8a31 | 4612 | * HPAGE_RESV_OWNER indicates a private mapping. |
84afd99b AW |
4613 | * This new VMA should share its siblings reservation map if present. |
4614 | * The VMA will only ever have a valid reservation map pointer where | |
4615 | * it is being copied for another still existing VMA. As that VMA | |
25985edc | 4616 | * has a reference to the reservation map it cannot disappear until |
84afd99b AW |
4617 | * after this open call completes. It is therefore safe to take a |
4618 | * new reference here without additional locking. | |
4619 | */ | |
09a26e83 MK |
4620 | if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { |
4621 | resv_map_dup_hugetlb_cgroup_uncharge_info(resv); | |
f522c3ac | 4622 | kref_get(&resv->refs); |
09a26e83 | 4623 | } |
8d9bfb26 | 4624 | |
131a79b4 MK |
4625 | /* |
4626 | * vma_lock structure for sharable mappings is vma specific. | |
612b8a31 MK |
4627 | * Clear old pointer (if copied via vm_area_dup) and allocate |
4628 | * new structure. Before clearing, make sure vma_lock is not | |
4629 | * for this vma. | |
131a79b4 MK |
4630 | */ |
4631 | if (vma->vm_flags & VM_MAYSHARE) { | |
612b8a31 MK |
4632 | struct hugetlb_vma_lock *vma_lock = vma->vm_private_data; |
4633 | ||
4634 | if (vma_lock) { | |
4635 | if (vma_lock->vma != vma) { | |
4636 | vma->vm_private_data = NULL; | |
4637 | hugetlb_vma_lock_alloc(vma); | |
4638 | } else | |
4639 | pr_warn("HugeTLB: vma_lock already exists in %s.\n", __func__); | |
4640 | } else | |
4641 | hugetlb_vma_lock_alloc(vma); | |
131a79b4 | 4642 | } |
84afd99b AW |
4643 | } |
4644 | ||
a1e78772 MG |
4645 | static void hugetlb_vm_op_close(struct vm_area_struct *vma) |
4646 | { | |
a5516438 | 4647 | struct hstate *h = hstate_vma(vma); |
8d9bfb26 | 4648 | struct resv_map *resv; |
90481622 | 4649 | struct hugepage_subpool *spool = subpool_vma(vma); |
4e35f483 | 4650 | unsigned long reserve, start, end; |
1c5ecae3 | 4651 | long gbl_reserve; |
84afd99b | 4652 | |
8d9bfb26 MK |
4653 | hugetlb_vma_lock_free(vma); |
4654 | ||
4655 | resv = vma_resv_map(vma); | |
4e35f483 JK |
4656 | if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER)) |
4657 | return; | |
84afd99b | 4658 | |
4e35f483 JK |
4659 | start = vma_hugecache_offset(h, vma, vma->vm_start); |
4660 | end = vma_hugecache_offset(h, vma, vma->vm_end); | |
84afd99b | 4661 | |
4e35f483 | 4662 | reserve = (end - start) - region_count(resv, start, end); |
e9fe92ae | 4663 | hugetlb_cgroup_uncharge_counter(resv, start, end); |
4e35f483 | 4664 | if (reserve) { |
1c5ecae3 MK |
4665 | /* |
4666 | * Decrement reserve counts. The global reserve count may be | |
4667 | * adjusted if the subpool has a minimum size. | |
4668 | */ | |
4669 | gbl_reserve = hugepage_subpool_put_pages(spool, reserve); | |
4670 | hugetlb_acct_memory(h, -gbl_reserve); | |
84afd99b | 4671 | } |
e9fe92ae MA |
4672 | |
4673 | kref_put(&resv->refs, resv_map_release); | |
a1e78772 MG |
4674 | } |
4675 | ||
31383c68 DW |
4676 | static int hugetlb_vm_op_split(struct vm_area_struct *vma, unsigned long addr) |
4677 | { | |
4678 | if (addr & ~(huge_page_mask(hstate_vma(vma)))) | |
4679 | return -EINVAL; | |
4680 | return 0; | |
4681 | } | |
4682 | ||
05ea8860 DW |
4683 | static unsigned long hugetlb_vm_op_pagesize(struct vm_area_struct *vma) |
4684 | { | |
aca78307 | 4685 | return huge_page_size(hstate_vma(vma)); |
05ea8860 DW |
4686 | } |
4687 | ||
1da177e4 LT |
4688 | /* |
4689 | * We cannot handle pagefaults against hugetlb pages at all. They cause | |
4690 | * handle_mm_fault() to try to instantiate regular-sized pages in the | |
6c26d310 | 4691 | * hugepage VMA. do_page_fault() is supposed to trap this, so BUG is we get |
1da177e4 LT |
4692 | * this far. |
4693 | */ | |
b3ec9f33 | 4694 | static vm_fault_t hugetlb_vm_op_fault(struct vm_fault *vmf) |
1da177e4 LT |
4695 | { |
4696 | BUG(); | |
d0217ac0 | 4697 | return 0; |
1da177e4 LT |
4698 | } |
4699 | ||
eec3636a JC |
4700 | /* |
4701 | * When a new function is introduced to vm_operations_struct and added | |
4702 | * to hugetlb_vm_ops, please consider adding the function to shm_vm_ops. | |
4703 | * This is because under System V memory model, mappings created via | |
4704 | * shmget/shmat with "huge page" specified are backed by hugetlbfs files, | |
4705 | * their original vm_ops are overwritten with shm_vm_ops. | |
4706 | */ | |
f0f37e2f | 4707 | const struct vm_operations_struct hugetlb_vm_ops = { |
d0217ac0 | 4708 | .fault = hugetlb_vm_op_fault, |
84afd99b | 4709 | .open = hugetlb_vm_op_open, |
a1e78772 | 4710 | .close = hugetlb_vm_op_close, |
dd3b614f | 4711 | .may_split = hugetlb_vm_op_split, |
05ea8860 | 4712 | .pagesize = hugetlb_vm_op_pagesize, |
1da177e4 LT |
4713 | }; |
4714 | ||
1e8f889b DG |
4715 | static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page, |
4716 | int writable) | |
63551ae0 DG |
4717 | { |
4718 | pte_t entry; | |
79c1c594 | 4719 | unsigned int shift = huge_page_shift(hstate_vma(vma)); |
63551ae0 | 4720 | |
1e8f889b | 4721 | if (writable) { |
106c992a GS |
4722 | entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page, |
4723 | vma->vm_page_prot))); | |
63551ae0 | 4724 | } else { |
106c992a GS |
4725 | entry = huge_pte_wrprotect(mk_huge_pte(page, |
4726 | vma->vm_page_prot)); | |
63551ae0 DG |
4727 | } |
4728 | entry = pte_mkyoung(entry); | |
79c1c594 | 4729 | entry = arch_make_huge_pte(entry, shift, vma->vm_flags); |
63551ae0 DG |
4730 | |
4731 | return entry; | |
4732 | } | |
4733 | ||
1e8f889b DG |
4734 | static void set_huge_ptep_writable(struct vm_area_struct *vma, |
4735 | unsigned long address, pte_t *ptep) | |
4736 | { | |
4737 | pte_t entry; | |
4738 | ||
106c992a | 4739 | entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep))); |
32f84528 | 4740 | if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) |
4b3073e1 | 4741 | update_mmu_cache(vma, address, ptep); |
1e8f889b DG |
4742 | } |
4743 | ||
d5ed7444 | 4744 | bool is_hugetlb_entry_migration(pte_t pte) |
4a705fef NH |
4745 | { |
4746 | swp_entry_t swp; | |
4747 | ||
4748 | if (huge_pte_none(pte) || pte_present(pte)) | |
d5ed7444 | 4749 | return false; |
4a705fef | 4750 | swp = pte_to_swp_entry(pte); |
d79d176a | 4751 | if (is_migration_entry(swp)) |
d5ed7444 | 4752 | return true; |
4a705fef | 4753 | else |
d5ed7444 | 4754 | return false; |
4a705fef NH |
4755 | } |
4756 | ||
3e5c3600 | 4757 | static bool is_hugetlb_entry_hwpoisoned(pte_t pte) |
4a705fef NH |
4758 | { |
4759 | swp_entry_t swp; | |
4760 | ||
4761 | if (huge_pte_none(pte) || pte_present(pte)) | |
3e5c3600 | 4762 | return false; |
4a705fef | 4763 | swp = pte_to_swp_entry(pte); |
d79d176a | 4764 | if (is_hwpoison_entry(swp)) |
3e5c3600 | 4765 | return true; |
4a705fef | 4766 | else |
3e5c3600 | 4767 | return false; |
4a705fef | 4768 | } |
1e8f889b | 4769 | |
4eae4efa PX |
4770 | static void |
4771 | hugetlb_install_page(struct vm_area_struct *vma, pte_t *ptep, unsigned long addr, | |
4772 | struct page *new_page) | |
4773 | { | |
4774 | __SetPageUptodate(new_page); | |
4eae4efa | 4775 | hugepage_add_new_anon_rmap(new_page, vma, addr); |
1eba86c0 | 4776 | set_huge_pte_at(vma->vm_mm, addr, ptep, make_huge_pte(vma, new_page, 1)); |
4eae4efa PX |
4777 | hugetlb_count_add(pages_per_huge_page(hstate_vma(vma)), vma->vm_mm); |
4778 | ClearHPageRestoreReserve(new_page); | |
4779 | SetHPageMigratable(new_page); | |
4780 | } | |
4781 | ||
63551ae0 | 4782 | int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src, |
bc70fbf2 PX |
4783 | struct vm_area_struct *dst_vma, |
4784 | struct vm_area_struct *src_vma) | |
63551ae0 | 4785 | { |
3aa4ed80 | 4786 | pte_t *src_pte, *dst_pte, entry; |
63551ae0 | 4787 | struct page *ptepage; |
1c59827d | 4788 | unsigned long addr; |
bc70fbf2 PX |
4789 | bool cow = is_cow_mapping(src_vma->vm_flags); |
4790 | struct hstate *h = hstate_vma(src_vma); | |
a5516438 | 4791 | unsigned long sz = huge_page_size(h); |
4eae4efa | 4792 | unsigned long npages = pages_per_huge_page(h); |
ac46d4f3 | 4793 | struct mmu_notifier_range range; |
e95a9851 | 4794 | unsigned long last_addr_mask; |
e8569dd2 | 4795 | int ret = 0; |
1e8f889b | 4796 | |
ac46d4f3 | 4797 | if (cow) { |
bc70fbf2 PX |
4798 | mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, src_vma, src, |
4799 | src_vma->vm_start, | |
4800 | src_vma->vm_end); | |
ac46d4f3 | 4801 | mmu_notifier_invalidate_range_start(&range); |
623a1ddf DH |
4802 | mmap_assert_write_locked(src); |
4803 | raw_write_seqcount_begin(&src->write_protect_seq); | |
40549ba8 MK |
4804 | } else { |
4805 | /* | |
4806 | * For shared mappings the vma lock must be held before | |
4807 | * calling huge_pte_offset in the src vma. Otherwise, the | |
4808 | * returned ptep could go away if part of a shared pmd and | |
4809 | * another thread calls huge_pmd_unshare. | |
4810 | */ | |
4811 | hugetlb_vma_lock_read(src_vma); | |
ac46d4f3 | 4812 | } |
e8569dd2 | 4813 | |
e95a9851 | 4814 | last_addr_mask = hugetlb_mask_last_page(h); |
bc70fbf2 | 4815 | for (addr = src_vma->vm_start; addr < src_vma->vm_end; addr += sz) { |
cb900f41 | 4816 | spinlock_t *src_ptl, *dst_ptl; |
7868a208 | 4817 | src_pte = huge_pte_offset(src, addr, sz); |
e95a9851 MK |
4818 | if (!src_pte) { |
4819 | addr |= last_addr_mask; | |
c74df32c | 4820 | continue; |
e95a9851 | 4821 | } |
bc70fbf2 | 4822 | dst_pte = huge_pte_alloc(dst, dst_vma, addr, sz); |
e8569dd2 AS |
4823 | if (!dst_pte) { |
4824 | ret = -ENOMEM; | |
4825 | break; | |
4826 | } | |
c5c99429 | 4827 | |
5e41540c MK |
4828 | /* |
4829 | * If the pagetables are shared don't copy or take references. | |
5e41540c | 4830 | * |
3aa4ed80 | 4831 | * dst_pte == src_pte is the common case of src/dest sharing. |
5e41540c | 4832 | * However, src could have 'unshared' and dst shares with |
3aa4ed80 ML |
4833 | * another vma. So page_count of ptep page is checked instead |
4834 | * to reliably determine whether pte is shared. | |
5e41540c | 4835 | */ |
3aa4ed80 | 4836 | if (page_count(virt_to_page(dst_pte)) > 1) { |
e95a9851 | 4837 | addr |= last_addr_mask; |
c5c99429 | 4838 | continue; |
e95a9851 | 4839 | } |
c5c99429 | 4840 | |
cb900f41 KS |
4841 | dst_ptl = huge_pte_lock(h, dst, dst_pte); |
4842 | src_ptl = huge_pte_lockptr(h, src, src_pte); | |
4843 | spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); | |
4a705fef | 4844 | entry = huge_ptep_get(src_pte); |
4eae4efa | 4845 | again: |
3aa4ed80 | 4846 | if (huge_pte_none(entry)) { |
5e41540c | 4847 | /* |
3aa4ed80 | 4848 | * Skip if src entry none. |
5e41540c | 4849 | */ |
4a705fef | 4850 | ; |
c2cb0dcc NH |
4851 | } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry))) { |
4852 | bool uffd_wp = huge_pte_uffd_wp(entry); | |
4853 | ||
4854 | if (!userfaultfd_wp(dst_vma) && uffd_wp) | |
4855 | entry = huge_pte_clear_uffd_wp(entry); | |
4856 | set_huge_pte_at(dst, addr, dst_pte, entry); | |
4857 | } else if (unlikely(is_hugetlb_entry_migration(entry))) { | |
4a705fef | 4858 | swp_entry_t swp_entry = pte_to_swp_entry(entry); |
bc70fbf2 | 4859 | bool uffd_wp = huge_pte_uffd_wp(entry); |
4a705fef | 4860 | |
6c287605 | 4861 | if (!is_readable_migration_entry(swp_entry) && cow) { |
4a705fef NH |
4862 | /* |
4863 | * COW mappings require pages in both | |
4864 | * parent and child to be set to read. | |
4865 | */ | |
4dd845b5 AP |
4866 | swp_entry = make_readable_migration_entry( |
4867 | swp_offset(swp_entry)); | |
4a705fef | 4868 | entry = swp_entry_to_pte(swp_entry); |
bc70fbf2 PX |
4869 | if (userfaultfd_wp(src_vma) && uffd_wp) |
4870 | entry = huge_pte_mkuffd_wp(entry); | |
18f39629 | 4871 | set_huge_pte_at(src, addr, src_pte, entry); |
4a705fef | 4872 | } |
bc70fbf2 PX |
4873 | if (!userfaultfd_wp(dst_vma) && uffd_wp) |
4874 | entry = huge_pte_clear_uffd_wp(entry); | |
18f39629 | 4875 | set_huge_pte_at(dst, addr, dst_pte, entry); |
bc70fbf2 PX |
4876 | } else if (unlikely(is_pte_marker(entry))) { |
4877 | /* | |
4878 | * We copy the pte marker only if the dst vma has | |
4879 | * uffd-wp enabled. | |
4880 | */ | |
4881 | if (userfaultfd_wp(dst_vma)) | |
4882 | set_huge_pte_at(dst, addr, dst_pte, entry); | |
4a705fef | 4883 | } else { |
4eae4efa PX |
4884 | entry = huge_ptep_get(src_pte); |
4885 | ptepage = pte_page(entry); | |
4886 | get_page(ptepage); | |
4887 | ||
4888 | /* | |
fb3d824d DH |
4889 | * Failing to duplicate the anon rmap is a rare case |
4890 | * where we see pinned hugetlb pages while they're | |
4891 | * prone to COW. We need to do the COW earlier during | |
4892 | * fork. | |
4eae4efa PX |
4893 | * |
4894 | * When pre-allocating the page or copying data, we | |
4895 | * need to be without the pgtable locks since we could | |
4896 | * sleep during the process. | |
4897 | */ | |
fb3d824d DH |
4898 | if (!PageAnon(ptepage)) { |
4899 | page_dup_file_rmap(ptepage, true); | |
bc70fbf2 PX |
4900 | } else if (page_try_dup_anon_rmap(ptepage, true, |
4901 | src_vma)) { | |
4eae4efa PX |
4902 | pte_t src_pte_old = entry; |
4903 | struct page *new; | |
4904 | ||
4905 | spin_unlock(src_ptl); | |
4906 | spin_unlock(dst_ptl); | |
4907 | /* Do not use reserve as it's private owned */ | |
bc70fbf2 | 4908 | new = alloc_huge_page(dst_vma, addr, 1); |
4eae4efa PX |
4909 | if (IS_ERR(new)) { |
4910 | put_page(ptepage); | |
4911 | ret = PTR_ERR(new); | |
4912 | break; | |
4913 | } | |
bc70fbf2 | 4914 | copy_user_huge_page(new, ptepage, addr, dst_vma, |
4eae4efa PX |
4915 | npages); |
4916 | put_page(ptepage); | |
4917 | ||
4918 | /* Install the new huge page if src pte stable */ | |
4919 | dst_ptl = huge_pte_lock(h, dst, dst_pte); | |
4920 | src_ptl = huge_pte_lockptr(h, src, src_pte); | |
4921 | spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); | |
4922 | entry = huge_ptep_get(src_pte); | |
4923 | if (!pte_same(src_pte_old, entry)) { | |
bc70fbf2 | 4924 | restore_reserve_on_error(h, dst_vma, addr, |
846be085 | 4925 | new); |
4eae4efa | 4926 | put_page(new); |
3aa4ed80 | 4927 | /* huge_ptep of dst_pte won't change as in child */ |
4eae4efa PX |
4928 | goto again; |
4929 | } | |
bc70fbf2 | 4930 | hugetlb_install_page(dst_vma, dst_pte, addr, new); |
4eae4efa PX |
4931 | spin_unlock(src_ptl); |
4932 | spin_unlock(dst_ptl); | |
4933 | continue; | |
4934 | } | |
4935 | ||
34ee645e | 4936 | if (cow) { |
0f10851e JG |
4937 | /* |
4938 | * No need to notify as we are downgrading page | |
4939 | * table protection not changing it to point | |
4940 | * to a new page. | |
4941 | * | |
ee65728e | 4942 | * See Documentation/mm/mmu_notifier.rst |
0f10851e | 4943 | */ |
7f2e9525 | 4944 | huge_ptep_set_wrprotect(src, addr, src_pte); |
84894e1c | 4945 | entry = huge_pte_wrprotect(entry); |
34ee645e | 4946 | } |
4eae4efa | 4947 | |
1c59827d | 4948 | set_huge_pte_at(dst, addr, dst_pte, entry); |
4eae4efa | 4949 | hugetlb_count_add(npages, dst); |
1c59827d | 4950 | } |
cb900f41 KS |
4951 | spin_unlock(src_ptl); |
4952 | spin_unlock(dst_ptl); | |
63551ae0 | 4953 | } |
63551ae0 | 4954 | |
623a1ddf DH |
4955 | if (cow) { |
4956 | raw_write_seqcount_end(&src->write_protect_seq); | |
ac46d4f3 | 4957 | mmu_notifier_invalidate_range_end(&range); |
40549ba8 MK |
4958 | } else { |
4959 | hugetlb_vma_unlock_read(src_vma); | |
623a1ddf | 4960 | } |
e8569dd2 AS |
4961 | |
4962 | return ret; | |
63551ae0 DG |
4963 | } |
4964 | ||
550a7d60 | 4965 | static void move_huge_pte(struct vm_area_struct *vma, unsigned long old_addr, |
db110a99 | 4966 | unsigned long new_addr, pte_t *src_pte, pte_t *dst_pte) |
550a7d60 MA |
4967 | { |
4968 | struct hstate *h = hstate_vma(vma); | |
4969 | struct mm_struct *mm = vma->vm_mm; | |
550a7d60 | 4970 | spinlock_t *src_ptl, *dst_ptl; |
db110a99 | 4971 | pte_t pte; |
550a7d60 | 4972 | |
550a7d60 MA |
4973 | dst_ptl = huge_pte_lock(h, mm, dst_pte); |
4974 | src_ptl = huge_pte_lockptr(h, mm, src_pte); | |
4975 | ||
4976 | /* | |
4977 | * We don't have to worry about the ordering of src and dst ptlocks | |
4978 | * because exclusive mmap_sem (or the i_mmap_lock) prevents deadlock. | |
4979 | */ | |
4980 | if (src_ptl != dst_ptl) | |
4981 | spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); | |
4982 | ||
4983 | pte = huge_ptep_get_and_clear(mm, old_addr, src_pte); | |
4984 | set_huge_pte_at(mm, new_addr, dst_pte, pte); | |
4985 | ||
4986 | if (src_ptl != dst_ptl) | |
4987 | spin_unlock(src_ptl); | |
4988 | spin_unlock(dst_ptl); | |
4989 | } | |
4990 | ||
4991 | int move_hugetlb_page_tables(struct vm_area_struct *vma, | |
4992 | struct vm_area_struct *new_vma, | |
4993 | unsigned long old_addr, unsigned long new_addr, | |
4994 | unsigned long len) | |
4995 | { | |
4996 | struct hstate *h = hstate_vma(vma); | |
4997 | struct address_space *mapping = vma->vm_file->f_mapping; | |
4998 | unsigned long sz = huge_page_size(h); | |
4999 | struct mm_struct *mm = vma->vm_mm; | |
5000 | unsigned long old_end = old_addr + len; | |
e95a9851 | 5001 | unsigned long last_addr_mask; |
550a7d60 MA |
5002 | pte_t *src_pte, *dst_pte; |
5003 | struct mmu_notifier_range range; | |
3d0b95cd | 5004 | bool shared_pmd = false; |
550a7d60 MA |
5005 | |
5006 | mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, old_addr, | |
5007 | old_end); | |
5008 | adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end); | |
3d0b95cd BW |
5009 | /* |
5010 | * In case of shared PMDs, we should cover the maximum possible | |
5011 | * range. | |
5012 | */ | |
5013 | flush_cache_range(vma, range.start, range.end); | |
5014 | ||
550a7d60 | 5015 | mmu_notifier_invalidate_range_start(&range); |
e95a9851 | 5016 | last_addr_mask = hugetlb_mask_last_page(h); |
550a7d60 | 5017 | /* Prevent race with file truncation */ |
40549ba8 | 5018 | hugetlb_vma_lock_write(vma); |
550a7d60 MA |
5019 | i_mmap_lock_write(mapping); |
5020 | for (; old_addr < old_end; old_addr += sz, new_addr += sz) { | |
5021 | src_pte = huge_pte_offset(mm, old_addr, sz); | |
e95a9851 MK |
5022 | if (!src_pte) { |
5023 | old_addr |= last_addr_mask; | |
5024 | new_addr |= last_addr_mask; | |
550a7d60 | 5025 | continue; |
e95a9851 | 5026 | } |
550a7d60 MA |
5027 | if (huge_pte_none(huge_ptep_get(src_pte))) |
5028 | continue; | |
5029 | ||
4ddb4d91 | 5030 | if (huge_pmd_unshare(mm, vma, old_addr, src_pte)) { |
3d0b95cd | 5031 | shared_pmd = true; |
4ddb4d91 MK |
5032 | old_addr |= last_addr_mask; |
5033 | new_addr |= last_addr_mask; | |
550a7d60 | 5034 | continue; |
3d0b95cd | 5035 | } |
550a7d60 MA |
5036 | |
5037 | dst_pte = huge_pte_alloc(mm, new_vma, new_addr, sz); | |
5038 | if (!dst_pte) | |
5039 | break; | |
5040 | ||
db110a99 | 5041 | move_huge_pte(vma, old_addr, new_addr, src_pte, dst_pte); |
550a7d60 | 5042 | } |
3d0b95cd BW |
5043 | |
5044 | if (shared_pmd) | |
5045 | flush_tlb_range(vma, range.start, range.end); | |
5046 | else | |
5047 | flush_tlb_range(vma, old_end - len, old_end); | |
550a7d60 | 5048 | mmu_notifier_invalidate_range_end(&range); |
13e4ad2c | 5049 | i_mmap_unlock_write(mapping); |
40549ba8 | 5050 | hugetlb_vma_unlock_write(vma); |
550a7d60 MA |
5051 | |
5052 | return len + old_addr - old_end; | |
5053 | } | |
5054 | ||
73c54763 PX |
5055 | static void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma, |
5056 | unsigned long start, unsigned long end, | |
05e90bd0 | 5057 | struct page *ref_page, zap_flags_t zap_flags) |
63551ae0 DG |
5058 | { |
5059 | struct mm_struct *mm = vma->vm_mm; | |
5060 | unsigned long address; | |
c7546f8f | 5061 | pte_t *ptep; |
63551ae0 | 5062 | pte_t pte; |
cb900f41 | 5063 | spinlock_t *ptl; |
63551ae0 | 5064 | struct page *page; |
a5516438 AK |
5065 | struct hstate *h = hstate_vma(vma); |
5066 | unsigned long sz = huge_page_size(h); | |
ac46d4f3 | 5067 | struct mmu_notifier_range range; |
e95a9851 | 5068 | unsigned long last_addr_mask; |
a4a118f2 | 5069 | bool force_flush = false; |
a5516438 | 5070 | |
63551ae0 | 5071 | WARN_ON(!is_vm_hugetlb_page(vma)); |
a5516438 AK |
5072 | BUG_ON(start & ~huge_page_mask(h)); |
5073 | BUG_ON(end & ~huge_page_mask(h)); | |
63551ae0 | 5074 | |
07e32661 AK |
5075 | /* |
5076 | * This is a hugetlb vma, all the pte entries should point | |
5077 | * to huge page. | |
5078 | */ | |
ed6a7935 | 5079 | tlb_change_page_size(tlb, sz); |
24669e58 | 5080 | tlb_start_vma(tlb, vma); |
dff11abe MK |
5081 | |
5082 | /* | |
5083 | * If sharing possible, alert mmu notifiers of worst case. | |
5084 | */ | |
6f4f13e8 JG |
5085 | mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma, mm, start, |
5086 | end); | |
ac46d4f3 JG |
5087 | adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end); |
5088 | mmu_notifier_invalidate_range_start(&range); | |
e95a9851 | 5089 | last_addr_mask = hugetlb_mask_last_page(h); |
569f48b8 | 5090 | address = start; |
569f48b8 | 5091 | for (; address < end; address += sz) { |
7868a208 | 5092 | ptep = huge_pte_offset(mm, address, sz); |
e95a9851 MK |
5093 | if (!ptep) { |
5094 | address |= last_addr_mask; | |
c7546f8f | 5095 | continue; |
e95a9851 | 5096 | } |
c7546f8f | 5097 | |
cb900f41 | 5098 | ptl = huge_pte_lock(h, mm, ptep); |
4ddb4d91 | 5099 | if (huge_pmd_unshare(mm, vma, address, ptep)) { |
31d49da5 | 5100 | spin_unlock(ptl); |
a4a118f2 NA |
5101 | tlb_flush_pmd_range(tlb, address & PUD_MASK, PUD_SIZE); |
5102 | force_flush = true; | |
4ddb4d91 | 5103 | address |= last_addr_mask; |
31d49da5 AK |
5104 | continue; |
5105 | } | |
39dde65c | 5106 | |
6629326b | 5107 | pte = huge_ptep_get(ptep); |
31d49da5 AK |
5108 | if (huge_pte_none(pte)) { |
5109 | spin_unlock(ptl); | |
5110 | continue; | |
5111 | } | |
6629326b HD |
5112 | |
5113 | /* | |
9fbc1f63 NH |
5114 | * Migrating hugepage or HWPoisoned hugepage is already |
5115 | * unmapped and its refcount is dropped, so just clear pte here. | |
6629326b | 5116 | */ |
9fbc1f63 | 5117 | if (unlikely(!pte_present(pte))) { |
515778e2 | 5118 | #ifdef CONFIG_PTE_MARKER_UFFD_WP |
05e90bd0 PX |
5119 | /* |
5120 | * If the pte was wr-protected by uffd-wp in any of the | |
5121 | * swap forms, meanwhile the caller does not want to | |
5122 | * drop the uffd-wp bit in this zap, then replace the | |
5123 | * pte with a marker. | |
5124 | */ | |
5125 | if (pte_swp_uffd_wp_any(pte) && | |
5126 | !(zap_flags & ZAP_FLAG_DROP_MARKER)) | |
5127 | set_huge_pte_at(mm, address, ptep, | |
5128 | make_pte_marker(PTE_MARKER_UFFD_WP)); | |
5129 | else | |
515778e2 | 5130 | #endif |
05e90bd0 | 5131 | huge_pte_clear(mm, address, ptep, sz); |
31d49da5 AK |
5132 | spin_unlock(ptl); |
5133 | continue; | |
8c4894c6 | 5134 | } |
6629326b HD |
5135 | |
5136 | page = pte_page(pte); | |
04f2cbe3 MG |
5137 | /* |
5138 | * If a reference page is supplied, it is because a specific | |
5139 | * page is being unmapped, not a range. Ensure the page we | |
5140 | * are about to unmap is the actual page of interest. | |
5141 | */ | |
5142 | if (ref_page) { | |
31d49da5 AK |
5143 | if (page != ref_page) { |
5144 | spin_unlock(ptl); | |
5145 | continue; | |
5146 | } | |
04f2cbe3 MG |
5147 | /* |
5148 | * Mark the VMA as having unmapped its page so that | |
5149 | * future faults in this VMA will fail rather than | |
5150 | * looking like data was lost | |
5151 | */ | |
5152 | set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED); | |
5153 | } | |
5154 | ||
c7546f8f | 5155 | pte = huge_ptep_get_and_clear(mm, address, ptep); |
b528e4b6 | 5156 | tlb_remove_huge_tlb_entry(h, tlb, ptep, address); |
106c992a | 5157 | if (huge_pte_dirty(pte)) |
6649a386 | 5158 | set_page_dirty(page); |
515778e2 | 5159 | #ifdef CONFIG_PTE_MARKER_UFFD_WP |
05e90bd0 PX |
5160 | /* Leave a uffd-wp pte marker if needed */ |
5161 | if (huge_pte_uffd_wp(pte) && | |
5162 | !(zap_flags & ZAP_FLAG_DROP_MARKER)) | |
5163 | set_huge_pte_at(mm, address, ptep, | |
5164 | make_pte_marker(PTE_MARKER_UFFD_WP)); | |
515778e2 | 5165 | #endif |
5d317b2b | 5166 | hugetlb_count_sub(pages_per_huge_page(h), mm); |
cea86fe2 | 5167 | page_remove_rmap(page, vma, true); |
31d49da5 | 5168 | |
cb900f41 | 5169 | spin_unlock(ptl); |
e77b0852 | 5170 | tlb_remove_page_size(tlb, page, huge_page_size(h)); |
31d49da5 AK |
5171 | /* |
5172 | * Bail out after unmapping reference page if supplied | |
5173 | */ | |
5174 | if (ref_page) | |
5175 | break; | |
fe1668ae | 5176 | } |
ac46d4f3 | 5177 | mmu_notifier_invalidate_range_end(&range); |
24669e58 | 5178 | tlb_end_vma(tlb, vma); |
a4a118f2 NA |
5179 | |
5180 | /* | |
5181 | * If we unshared PMDs, the TLB flush was not recorded in mmu_gather. We | |
5182 | * could defer the flush until now, since by holding i_mmap_rwsem we | |
5183 | * guaranteed that the last refernece would not be dropped. But we must | |
5184 | * do the flushing before we return, as otherwise i_mmap_rwsem will be | |
5185 | * dropped and the last reference to the shared PMDs page might be | |
5186 | * dropped as well. | |
5187 | * | |
5188 | * In theory we could defer the freeing of the PMD pages as well, but | |
5189 | * huge_pmd_unshare() relies on the exact page_count for the PMD page to | |
5190 | * detect sharing, so we cannot defer the release of the page either. | |
5191 | * Instead, do flush now. | |
5192 | */ | |
5193 | if (force_flush) | |
5194 | tlb_flush_mmu_tlbonly(tlb); | |
1da177e4 | 5195 | } |
63551ae0 | 5196 | |
d833352a MG |
5197 | void __unmap_hugepage_range_final(struct mmu_gather *tlb, |
5198 | struct vm_area_struct *vma, unsigned long start, | |
05e90bd0 PX |
5199 | unsigned long end, struct page *ref_page, |
5200 | zap_flags_t zap_flags) | |
d833352a | 5201 | { |
131a79b4 MK |
5202 | hugetlb_vma_lock_write(vma); |
5203 | i_mmap_lock_write(vma->vm_file->f_mapping); | |
5204 | ||
05e90bd0 | 5205 | __unmap_hugepage_range(tlb, vma, start, end, ref_page, zap_flags); |
d833352a MG |
5206 | |
5207 | /* | |
131a79b4 MK |
5208 | * Unlock and free the vma lock before releasing i_mmap_rwsem. When |
5209 | * the vma_lock is freed, this makes the vma ineligible for pmd | |
5210 | * sharing. And, i_mmap_rwsem is required to set up pmd sharing. | |
5211 | * This is important as page tables for this unmapped range will | |
5212 | * be asynchrously deleted. If the page tables are shared, there | |
5213 | * will be issues when accessed by someone else. | |
d833352a | 5214 | */ |
ecfbd733 | 5215 | __hugetlb_vma_unlock_write_free(vma); |
131a79b4 MK |
5216 | |
5217 | i_mmap_unlock_write(vma->vm_file->f_mapping); | |
d833352a MG |
5218 | } |
5219 | ||
502717f4 | 5220 | void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start, |
05e90bd0 PX |
5221 | unsigned long end, struct page *ref_page, |
5222 | zap_flags_t zap_flags) | |
502717f4 | 5223 | { |
24669e58 | 5224 | struct mmu_gather tlb; |
dff11abe | 5225 | |
a72afd87 | 5226 | tlb_gather_mmu(&tlb, vma->vm_mm); |
05e90bd0 | 5227 | __unmap_hugepage_range(&tlb, vma, start, end, ref_page, zap_flags); |
ae8eba8b | 5228 | tlb_finish_mmu(&tlb); |
502717f4 CK |
5229 | } |
5230 | ||
04f2cbe3 MG |
5231 | /* |
5232 | * This is called when the original mapper is failing to COW a MAP_PRIVATE | |
578b7725 | 5233 | * mapping it owns the reserve page for. The intention is to unmap the page |
04f2cbe3 MG |
5234 | * from other VMAs and let the children be SIGKILLed if they are faulting the |
5235 | * same region. | |
5236 | */ | |
2f4612af DB |
5237 | static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma, |
5238 | struct page *page, unsigned long address) | |
04f2cbe3 | 5239 | { |
7526674d | 5240 | struct hstate *h = hstate_vma(vma); |
04f2cbe3 MG |
5241 | struct vm_area_struct *iter_vma; |
5242 | struct address_space *mapping; | |
04f2cbe3 MG |
5243 | pgoff_t pgoff; |
5244 | ||
5245 | /* | |
5246 | * vm_pgoff is in PAGE_SIZE units, hence the different calculation | |
5247 | * from page cache lookup which is in HPAGE_SIZE units. | |
5248 | */ | |
7526674d | 5249 | address = address & huge_page_mask(h); |
36e4f20a MH |
5250 | pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) + |
5251 | vma->vm_pgoff; | |
93c76a3d | 5252 | mapping = vma->vm_file->f_mapping; |
04f2cbe3 | 5253 | |
4eb2b1dc MG |
5254 | /* |
5255 | * Take the mapping lock for the duration of the table walk. As | |
5256 | * this mapping should be shared between all the VMAs, | |
5257 | * __unmap_hugepage_range() is called as the lock is already held | |
5258 | */ | |
83cde9e8 | 5259 | i_mmap_lock_write(mapping); |
6b2dbba8 | 5260 | vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) { |
04f2cbe3 MG |
5261 | /* Do not unmap the current VMA */ |
5262 | if (iter_vma == vma) | |
5263 | continue; | |
5264 | ||
2f84a899 MG |
5265 | /* |
5266 | * Shared VMAs have their own reserves and do not affect | |
5267 | * MAP_PRIVATE accounting but it is possible that a shared | |
5268 | * VMA is using the same page so check and skip such VMAs. | |
5269 | */ | |
5270 | if (iter_vma->vm_flags & VM_MAYSHARE) | |
5271 | continue; | |
5272 | ||
04f2cbe3 MG |
5273 | /* |
5274 | * Unmap the page from other VMAs without their own reserves. | |
5275 | * They get marked to be SIGKILLed if they fault in these | |
5276 | * areas. This is because a future no-page fault on this VMA | |
5277 | * could insert a zeroed page instead of the data existing | |
5278 | * from the time of fork. This would look like data corruption | |
5279 | */ | |
5280 | if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER)) | |
24669e58 | 5281 | unmap_hugepage_range(iter_vma, address, |
05e90bd0 | 5282 | address + huge_page_size(h), page, 0); |
04f2cbe3 | 5283 | } |
83cde9e8 | 5284 | i_mmap_unlock_write(mapping); |
04f2cbe3 MG |
5285 | } |
5286 | ||
0fe6e20b | 5287 | /* |
c89357e2 | 5288 | * hugetlb_wp() should be called with page lock of the original hugepage held. |
aa6d2e8c | 5289 | * Called with hugetlb_fault_mutex_table held and pte_page locked so we |
ef009b25 MH |
5290 | * cannot race with other handlers or page migration. |
5291 | * Keep the pte_same checks anyway to make transition from the mutex easier. | |
0fe6e20b | 5292 | */ |
c89357e2 DH |
5293 | static vm_fault_t hugetlb_wp(struct mm_struct *mm, struct vm_area_struct *vma, |
5294 | unsigned long address, pte_t *ptep, unsigned int flags, | |
3999f52e | 5295 | struct page *pagecache_page, spinlock_t *ptl) |
1e8f889b | 5296 | { |
c89357e2 | 5297 | const bool unshare = flags & FAULT_FLAG_UNSHARE; |
3999f52e | 5298 | pte_t pte; |
a5516438 | 5299 | struct hstate *h = hstate_vma(vma); |
1e8f889b | 5300 | struct page *old_page, *new_page; |
2b740303 SJ |
5301 | int outside_reserve = 0; |
5302 | vm_fault_t ret = 0; | |
974e6d66 | 5303 | unsigned long haddr = address & huge_page_mask(h); |
ac46d4f3 | 5304 | struct mmu_notifier_range range; |
1e8f889b | 5305 | |
c89357e2 DH |
5306 | VM_BUG_ON(unshare && (flags & FOLL_WRITE)); |
5307 | VM_BUG_ON(!unshare && !(flags & FOLL_WRITE)); | |
5308 | ||
1d8d1464 DH |
5309 | /* |
5310 | * hugetlb does not support FOLL_FORCE-style write faults that keep the | |
5311 | * PTE mapped R/O such as maybe_mkwrite() would do. | |
5312 | */ | |
5313 | if (WARN_ON_ONCE(!unshare && !(vma->vm_flags & VM_WRITE))) | |
5314 | return VM_FAULT_SIGSEGV; | |
5315 | ||
5316 | /* Let's take out MAP_SHARED mappings first. */ | |
5317 | if (vma->vm_flags & VM_MAYSHARE) { | |
5318 | if (unlikely(unshare)) | |
5319 | return 0; | |
5320 | set_huge_ptep_writable(vma, haddr, ptep); | |
5321 | return 0; | |
5322 | } | |
5323 | ||
3999f52e | 5324 | pte = huge_ptep_get(ptep); |
1e8f889b DG |
5325 | old_page = pte_page(pte); |
5326 | ||
662ce1dc YY |
5327 | delayacct_wpcopy_start(); |
5328 | ||
04f2cbe3 | 5329 | retry_avoidcopy: |
c89357e2 DH |
5330 | /* |
5331 | * If no-one else is actually using this page, we're the exclusive | |
5332 | * owner and can reuse this page. | |
5333 | */ | |
37a2140d | 5334 | if (page_mapcount(old_page) == 1 && PageAnon(old_page)) { |
c89357e2 DH |
5335 | if (!PageAnonExclusive(old_page)) |
5336 | page_move_anon_rmap(old_page, vma); | |
5337 | if (likely(!unshare)) | |
5338 | set_huge_ptep_writable(vma, haddr, ptep); | |
662ce1dc YY |
5339 | |
5340 | delayacct_wpcopy_end(); | |
83c54070 | 5341 | return 0; |
1e8f889b | 5342 | } |
6c287605 DH |
5343 | VM_BUG_ON_PAGE(PageAnon(old_page) && PageAnonExclusive(old_page), |
5344 | old_page); | |
1e8f889b | 5345 | |
04f2cbe3 MG |
5346 | /* |
5347 | * If the process that created a MAP_PRIVATE mapping is about to | |
5348 | * perform a COW due to a shared page count, attempt to satisfy | |
5349 | * the allocation without using the existing reserves. The pagecache | |
5350 | * page is used to determine if the reserve at this address was | |
5351 | * consumed or not. If reserves were used, a partial faulted mapping | |
5352 | * at the time of fork() could consume its reserves on COW instead | |
5353 | * of the full address range. | |
5354 | */ | |
5944d011 | 5355 | if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && |
04f2cbe3 MG |
5356 | old_page != pagecache_page) |
5357 | outside_reserve = 1; | |
5358 | ||
09cbfeaf | 5359 | get_page(old_page); |
b76c8cfb | 5360 | |
ad4404a2 DB |
5361 | /* |
5362 | * Drop page table lock as buddy allocator may be called. It will | |
5363 | * be acquired again before returning to the caller, as expected. | |
5364 | */ | |
cb900f41 | 5365 | spin_unlock(ptl); |
5b7a1d40 | 5366 | new_page = alloc_huge_page(vma, haddr, outside_reserve); |
1e8f889b | 5367 | |
2fc39cec | 5368 | if (IS_ERR(new_page)) { |
04f2cbe3 MG |
5369 | /* |
5370 | * If a process owning a MAP_PRIVATE mapping fails to COW, | |
5371 | * it is due to references held by a child and an insufficient | |
5372 | * huge page pool. To guarantee the original mappers | |
5373 | * reliability, unmap the page from child processes. The child | |
5374 | * may get SIGKILLed if it later faults. | |
5375 | */ | |
5376 | if (outside_reserve) { | |
40549ba8 MK |
5377 | struct address_space *mapping = vma->vm_file->f_mapping; |
5378 | pgoff_t idx; | |
5379 | u32 hash; | |
5380 | ||
09cbfeaf | 5381 | put_page(old_page); |
40549ba8 MK |
5382 | /* |
5383 | * Drop hugetlb_fault_mutex and vma_lock before | |
5384 | * unmapping. unmapping needs to hold vma_lock | |
5385 | * in write mode. Dropping vma_lock in read mode | |
5386 | * here is OK as COW mappings do not interact with | |
5387 | * PMD sharing. | |
5388 | * | |
5389 | * Reacquire both after unmap operation. | |
5390 | */ | |
5391 | idx = vma_hugecache_offset(h, vma, haddr); | |
5392 | hash = hugetlb_fault_mutex_hash(mapping, idx); | |
5393 | hugetlb_vma_unlock_read(vma); | |
5394 | mutex_unlock(&hugetlb_fault_mutex_table[hash]); | |
5395 | ||
5b7a1d40 | 5396 | unmap_ref_private(mm, vma, old_page, haddr); |
40549ba8 MK |
5397 | |
5398 | mutex_lock(&hugetlb_fault_mutex_table[hash]); | |
5399 | hugetlb_vma_lock_read(vma); | |
2f4612af | 5400 | spin_lock(ptl); |
5b7a1d40 | 5401 | ptep = huge_pte_offset(mm, haddr, huge_page_size(h)); |
2f4612af DB |
5402 | if (likely(ptep && |
5403 | pte_same(huge_ptep_get(ptep), pte))) | |
5404 | goto retry_avoidcopy; | |
5405 | /* | |
5406 | * race occurs while re-acquiring page table | |
5407 | * lock, and our job is done. | |
5408 | */ | |
662ce1dc | 5409 | delayacct_wpcopy_end(); |
2f4612af | 5410 | return 0; |
04f2cbe3 MG |
5411 | } |
5412 | ||
2b740303 | 5413 | ret = vmf_error(PTR_ERR(new_page)); |
ad4404a2 | 5414 | goto out_release_old; |
1e8f889b DG |
5415 | } |
5416 | ||
0fe6e20b NH |
5417 | /* |
5418 | * When the original hugepage is shared one, it does not have | |
5419 | * anon_vma prepared. | |
5420 | */ | |
44e2aa93 | 5421 | if (unlikely(anon_vma_prepare(vma))) { |
ad4404a2 DB |
5422 | ret = VM_FAULT_OOM; |
5423 | goto out_release_all; | |
44e2aa93 | 5424 | } |
0fe6e20b | 5425 | |
974e6d66 | 5426 | copy_user_huge_page(new_page, old_page, address, vma, |
47ad8475 | 5427 | pages_per_huge_page(h)); |
0ed361de | 5428 | __SetPageUptodate(new_page); |
1e8f889b | 5429 | |
7269f999 | 5430 | mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, haddr, |
6f4f13e8 | 5431 | haddr + huge_page_size(h)); |
ac46d4f3 | 5432 | mmu_notifier_invalidate_range_start(&range); |
ad4404a2 | 5433 | |
b76c8cfb | 5434 | /* |
cb900f41 | 5435 | * Retake the page table lock to check for racing updates |
b76c8cfb LW |
5436 | * before the page tables are altered |
5437 | */ | |
cb900f41 | 5438 | spin_lock(ptl); |
5b7a1d40 | 5439 | ptep = huge_pte_offset(mm, haddr, huge_page_size(h)); |
a9af0c5d | 5440 | if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) { |
d6995da3 | 5441 | ClearHPageRestoreReserve(new_page); |
07443a85 | 5442 | |
c89357e2 | 5443 | /* Break COW or unshare */ |
5b7a1d40 | 5444 | huge_ptep_clear_flush(vma, haddr, ptep); |
ac46d4f3 | 5445 | mmu_notifier_invalidate_range(mm, range.start, range.end); |
cea86fe2 | 5446 | page_remove_rmap(old_page, vma, true); |
5b7a1d40 | 5447 | hugepage_add_new_anon_rmap(new_page, vma, haddr); |
1eba86c0 | 5448 | set_huge_pte_at(mm, haddr, ptep, |
c89357e2 | 5449 | make_huge_pte(vma, new_page, !unshare)); |
8f251a3d | 5450 | SetHPageMigratable(new_page); |
1e8f889b DG |
5451 | /* Make the old page be freed below */ |
5452 | new_page = old_page; | |
5453 | } | |
cb900f41 | 5454 | spin_unlock(ptl); |
ac46d4f3 | 5455 | mmu_notifier_invalidate_range_end(&range); |
ad4404a2 | 5456 | out_release_all: |
c89357e2 DH |
5457 | /* |
5458 | * No restore in case of successful pagetable update (Break COW or | |
5459 | * unshare) | |
5460 | */ | |
c7b1850d MK |
5461 | if (new_page != old_page) |
5462 | restore_reserve_on_error(h, vma, haddr, new_page); | |
09cbfeaf | 5463 | put_page(new_page); |
ad4404a2 | 5464 | out_release_old: |
09cbfeaf | 5465 | put_page(old_page); |
8312034f | 5466 | |
ad4404a2 | 5467 | spin_lock(ptl); /* Caller expects lock to be held */ |
662ce1dc YY |
5468 | |
5469 | delayacct_wpcopy_end(); | |
ad4404a2 | 5470 | return ret; |
1e8f889b DG |
5471 | } |
5472 | ||
3ae77f43 HD |
5473 | /* |
5474 | * Return whether there is a pagecache page to back given address within VMA. | |
5475 | * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page. | |
5476 | */ | |
5477 | static bool hugetlbfs_pagecache_present(struct hstate *h, | |
2a15efc9 HD |
5478 | struct vm_area_struct *vma, unsigned long address) |
5479 | { | |
5480 | struct address_space *mapping; | |
5481 | pgoff_t idx; | |
5482 | struct page *page; | |
5483 | ||
5484 | mapping = vma->vm_file->f_mapping; | |
5485 | idx = vma_hugecache_offset(h, vma, address); | |
5486 | ||
5487 | page = find_get_page(mapping, idx); | |
5488 | if (page) | |
5489 | put_page(page); | |
5490 | return page != NULL; | |
5491 | } | |
5492 | ||
7e1813d4 | 5493 | int hugetlb_add_to_page_cache(struct page *page, struct address_space *mapping, |
ab76ad54 MK |
5494 | pgoff_t idx) |
5495 | { | |
d9ef44de | 5496 | struct folio *folio = page_folio(page); |
ab76ad54 MK |
5497 | struct inode *inode = mapping->host; |
5498 | struct hstate *h = hstate_inode(inode); | |
d9ef44de | 5499 | int err; |
ab76ad54 | 5500 | |
d9ef44de MWO |
5501 | __folio_set_locked(folio); |
5502 | err = __filemap_add_folio(mapping, folio, idx, GFP_KERNEL, NULL); | |
5503 | ||
5504 | if (unlikely(err)) { | |
5505 | __folio_clear_locked(folio); | |
ab76ad54 | 5506 | return err; |
d9ef44de | 5507 | } |
d6995da3 | 5508 | ClearHPageRestoreReserve(page); |
ab76ad54 | 5509 | |
22146c3c | 5510 | /* |
d9ef44de | 5511 | * mark folio dirty so that it will not be removed from cache/file |
22146c3c MK |
5512 | * by non-hugetlbfs specific code paths. |
5513 | */ | |
d9ef44de | 5514 | folio_mark_dirty(folio); |
22146c3c | 5515 | |
ab76ad54 MK |
5516 | spin_lock(&inode->i_lock); |
5517 | inode->i_blocks += blocks_per_huge_page(h); | |
5518 | spin_unlock(&inode->i_lock); | |
5519 | return 0; | |
5520 | } | |
5521 | ||
7677f7fd AR |
5522 | static inline vm_fault_t hugetlb_handle_userfault(struct vm_area_struct *vma, |
5523 | struct address_space *mapping, | |
5524 | pgoff_t idx, | |
5525 | unsigned int flags, | |
5526 | unsigned long haddr, | |
824ddc60 | 5527 | unsigned long addr, |
7677f7fd AR |
5528 | unsigned long reason) |
5529 | { | |
7677f7fd AR |
5530 | u32 hash; |
5531 | struct vm_fault vmf = { | |
5532 | .vma = vma, | |
5533 | .address = haddr, | |
824ddc60 | 5534 | .real_address = addr, |
7677f7fd AR |
5535 | .flags = flags, |
5536 | ||
5537 | /* | |
5538 | * Hard to debug if it ends up being | |
5539 | * used by a callee that assumes | |
5540 | * something about the other | |
5541 | * uninitialized fields... same as in | |
5542 | * memory.c | |
5543 | */ | |
5544 | }; | |
5545 | ||
5546 | /* | |
958f32ce LS |
5547 | * vma_lock and hugetlb_fault_mutex must be dropped before handling |
5548 | * userfault. Also mmap_lock could be dropped due to handling | |
5549 | * userfault, any vma operation should be careful from here. | |
7677f7fd | 5550 | */ |
40549ba8 | 5551 | hugetlb_vma_unlock_read(vma); |
7677f7fd AR |
5552 | hash = hugetlb_fault_mutex_hash(mapping, idx); |
5553 | mutex_unlock(&hugetlb_fault_mutex_table[hash]); | |
958f32ce | 5554 | return handle_userfault(&vmf, reason); |
7677f7fd AR |
5555 | } |
5556 | ||
2ea7ff1e PX |
5557 | /* |
5558 | * Recheck pte with pgtable lock. Returns true if pte didn't change, or | |
5559 | * false if pte changed or is changing. | |
5560 | */ | |
5561 | static bool hugetlb_pte_stable(struct hstate *h, struct mm_struct *mm, | |
5562 | pte_t *ptep, pte_t old_pte) | |
5563 | { | |
5564 | spinlock_t *ptl; | |
5565 | bool same; | |
5566 | ||
5567 | ptl = huge_pte_lock(h, mm, ptep); | |
5568 | same = pte_same(huge_ptep_get(ptep), old_pte); | |
5569 | spin_unlock(ptl); | |
5570 | ||
5571 | return same; | |
5572 | } | |
5573 | ||
2b740303 SJ |
5574 | static vm_fault_t hugetlb_no_page(struct mm_struct *mm, |
5575 | struct vm_area_struct *vma, | |
5576 | struct address_space *mapping, pgoff_t idx, | |
c64e912c PX |
5577 | unsigned long address, pte_t *ptep, |
5578 | pte_t old_pte, unsigned int flags) | |
ac9b9c66 | 5579 | { |
a5516438 | 5580 | struct hstate *h = hstate_vma(vma); |
2b740303 | 5581 | vm_fault_t ret = VM_FAULT_SIGBUS; |
409eb8c2 | 5582 | int anon_rmap = 0; |
4c887265 | 5583 | unsigned long size; |
4c887265 | 5584 | struct page *page; |
1e8f889b | 5585 | pte_t new_pte; |
cb900f41 | 5586 | spinlock_t *ptl; |
285b8dca | 5587 | unsigned long haddr = address & huge_page_mask(h); |
c7b1850d | 5588 | bool new_page, new_pagecache_page = false; |
958f32ce | 5589 | u32 hash = hugetlb_fault_mutex_hash(mapping, idx); |
4c887265 | 5590 | |
04f2cbe3 MG |
5591 | /* |
5592 | * Currently, we are forced to kill the process in the event the | |
5593 | * original mapper has unmapped pages from the child due to a failed | |
c89357e2 DH |
5594 | * COW/unsharing. Warn that such a situation has occurred as it may not |
5595 | * be obvious. | |
04f2cbe3 MG |
5596 | */ |
5597 | if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) { | |
910154d5 | 5598 | pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n", |
ffb22af5 | 5599 | current->pid); |
958f32ce | 5600 | goto out; |
04f2cbe3 MG |
5601 | } |
5602 | ||
4c887265 | 5603 | /* |
188a3972 MK |
5604 | * Use page lock to guard against racing truncation |
5605 | * before we get page_table_lock. | |
4c887265 | 5606 | */ |
c7b1850d | 5607 | new_page = false; |
6bda666a CL |
5608 | page = find_lock_page(mapping, idx); |
5609 | if (!page) { | |
188a3972 MK |
5610 | size = i_size_read(mapping->host) >> huge_page_shift(h); |
5611 | if (idx >= size) | |
5612 | goto out; | |
7677f7fd | 5613 | /* Check for page in userfault range */ |
2ea7ff1e PX |
5614 | if (userfaultfd_missing(vma)) { |
5615 | /* | |
5616 | * Since hugetlb_no_page() was examining pte | |
5617 | * without pgtable lock, we need to re-test under | |
5618 | * lock because the pte may not be stable and could | |
5619 | * have changed from under us. Try to detect | |
5620 | * either changed or during-changing ptes and retry | |
5621 | * properly when needed. | |
5622 | * | |
5623 | * Note that userfaultfd is actually fine with | |
5624 | * false positives (e.g. caused by pte changed), | |
5625 | * but not wrong logical events (e.g. caused by | |
5626 | * reading a pte during changing). The latter can | |
5627 | * confuse the userspace, so the strictness is very | |
5628 | * much preferred. E.g., MISSING event should | |
5629 | * never happen on the page after UFFDIO_COPY has | |
5630 | * correctly installed the page and returned. | |
5631 | */ | |
5632 | if (!hugetlb_pte_stable(h, mm, ptep, old_pte)) { | |
5633 | ret = 0; | |
5634 | goto out; | |
5635 | } | |
5636 | ||
5637 | return hugetlb_handle_userfault(vma, mapping, idx, flags, | |
5638 | haddr, address, | |
5639 | VM_UFFD_MISSING); | |
5640 | } | |
1a1aad8a | 5641 | |
285b8dca | 5642 | page = alloc_huge_page(vma, haddr, 0); |
2fc39cec | 5643 | if (IS_ERR(page)) { |
4643d67e MK |
5644 | /* |
5645 | * Returning error will result in faulting task being | |
5646 | * sent SIGBUS. The hugetlb fault mutex prevents two | |
5647 | * tasks from racing to fault in the same page which | |
5648 | * could result in false unable to allocate errors. | |
5649 | * Page migration does not take the fault mutex, but | |
5650 | * does a clear then write of pte's under page table | |
5651 | * lock. Page fault code could race with migration, | |
5652 | * notice the clear pte and try to allocate a page | |
5653 | * here. Before returning error, get ptl and make | |
5654 | * sure there really is no pte entry. | |
5655 | */ | |
f9bf6c03 | 5656 | if (hugetlb_pte_stable(h, mm, ptep, old_pte)) |
d83e6c8a | 5657 | ret = vmf_error(PTR_ERR(page)); |
f9bf6c03 PX |
5658 | else |
5659 | ret = 0; | |
6bda666a CL |
5660 | goto out; |
5661 | } | |
47ad8475 | 5662 | clear_huge_page(page, address, pages_per_huge_page(h)); |
0ed361de | 5663 | __SetPageUptodate(page); |
cb6acd01 | 5664 | new_page = true; |
ac9b9c66 | 5665 | |
f83a275d | 5666 | if (vma->vm_flags & VM_MAYSHARE) { |
7e1813d4 | 5667 | int err = hugetlb_add_to_page_cache(page, mapping, idx); |
6bda666a | 5668 | if (err) { |
3a5497a2 ML |
5669 | /* |
5670 | * err can't be -EEXIST which implies someone | |
5671 | * else consumed the reservation since hugetlb | |
5672 | * fault mutex is held when add a hugetlb page | |
5673 | * to the page cache. So it's safe to call | |
5674 | * restore_reserve_on_error() here. | |
5675 | */ | |
5676 | restore_reserve_on_error(h, vma, haddr, page); | |
6bda666a | 5677 | put_page(page); |
6bda666a CL |
5678 | goto out; |
5679 | } | |
c7b1850d | 5680 | new_pagecache_page = true; |
23be7468 | 5681 | } else { |
6bda666a | 5682 | lock_page(page); |
0fe6e20b NH |
5683 | if (unlikely(anon_vma_prepare(vma))) { |
5684 | ret = VM_FAULT_OOM; | |
5685 | goto backout_unlocked; | |
5686 | } | |
409eb8c2 | 5687 | anon_rmap = 1; |
23be7468 | 5688 | } |
0fe6e20b | 5689 | } else { |
998b4382 NH |
5690 | /* |
5691 | * If memory error occurs between mmap() and fault, some process | |
5692 | * don't have hwpoisoned swap entry for errored virtual address. | |
5693 | * So we need to block hugepage fault by PG_hwpoison bit check. | |
5694 | */ | |
5695 | if (unlikely(PageHWPoison(page))) { | |
0eb98f15 | 5696 | ret = VM_FAULT_HWPOISON_LARGE | |
972dc4de | 5697 | VM_FAULT_SET_HINDEX(hstate_index(h)); |
998b4382 NH |
5698 | goto backout_unlocked; |
5699 | } | |
7677f7fd AR |
5700 | |
5701 | /* Check for page in userfault range. */ | |
5702 | if (userfaultfd_minor(vma)) { | |
5703 | unlock_page(page); | |
5704 | put_page(page); | |
2ea7ff1e PX |
5705 | /* See comment in userfaultfd_missing() block above */ |
5706 | if (!hugetlb_pte_stable(h, mm, ptep, old_pte)) { | |
5707 | ret = 0; | |
5708 | goto out; | |
5709 | } | |
5710 | return hugetlb_handle_userfault(vma, mapping, idx, flags, | |
5711 | haddr, address, | |
5712 | VM_UFFD_MINOR); | |
7677f7fd | 5713 | } |
6bda666a | 5714 | } |
1e8f889b | 5715 | |
57303d80 AW |
5716 | /* |
5717 | * If we are going to COW a private mapping later, we examine the | |
5718 | * pending reservations for this page now. This will ensure that | |
5719 | * any allocations necessary to record that reservation occur outside | |
5720 | * the spinlock. | |
5721 | */ | |
5e911373 | 5722 | if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { |
285b8dca | 5723 | if (vma_needs_reservation(h, vma, haddr) < 0) { |
2b26736c AW |
5724 | ret = VM_FAULT_OOM; |
5725 | goto backout_unlocked; | |
5726 | } | |
5e911373 | 5727 | /* Just decrements count, does not deallocate */ |
285b8dca | 5728 | vma_end_reservation(h, vma, haddr); |
5e911373 | 5729 | } |
57303d80 | 5730 | |
8bea8052 | 5731 | ptl = huge_pte_lock(h, mm, ptep); |
83c54070 | 5732 | ret = 0; |
c64e912c PX |
5733 | /* If pte changed from under us, retry */ |
5734 | if (!pte_same(huge_ptep_get(ptep), old_pte)) | |
4c887265 AL |
5735 | goto backout; |
5736 | ||
07443a85 | 5737 | if (anon_rmap) { |
d6995da3 | 5738 | ClearHPageRestoreReserve(page); |
285b8dca | 5739 | hugepage_add_new_anon_rmap(page, vma, haddr); |
ac714904 | 5740 | } else |
fb3d824d | 5741 | page_dup_file_rmap(page, true); |
1e8f889b DG |
5742 | new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE) |
5743 | && (vma->vm_flags & VM_SHARED))); | |
c64e912c PX |
5744 | /* |
5745 | * If this pte was previously wr-protected, keep it wr-protected even | |
5746 | * if populated. | |
5747 | */ | |
5748 | if (unlikely(pte_marker_uffd_wp(old_pte))) | |
5749 | new_pte = huge_pte_wrprotect(huge_pte_mkuffd_wp(new_pte)); | |
285b8dca | 5750 | set_huge_pte_at(mm, haddr, ptep, new_pte); |
1e8f889b | 5751 | |
5d317b2b | 5752 | hugetlb_count_add(pages_per_huge_page(h), mm); |
788c7df4 | 5753 | if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { |
1e8f889b | 5754 | /* Optimization, do the COW without a second fault */ |
c89357e2 | 5755 | ret = hugetlb_wp(mm, vma, address, ptep, flags, page, ptl); |
1e8f889b DG |
5756 | } |
5757 | ||
cb900f41 | 5758 | spin_unlock(ptl); |
cb6acd01 MK |
5759 | |
5760 | /* | |
8f251a3d MK |
5761 | * Only set HPageMigratable in newly allocated pages. Existing pages |
5762 | * found in the pagecache may not have HPageMigratableset if they have | |
5763 | * been isolated for migration. | |
cb6acd01 MK |
5764 | */ |
5765 | if (new_page) | |
8f251a3d | 5766 | SetHPageMigratable(page); |
cb6acd01 | 5767 | |
4c887265 AL |
5768 | unlock_page(page); |
5769 | out: | |
958f32ce LS |
5770 | hugetlb_vma_unlock_read(vma); |
5771 | mutex_unlock(&hugetlb_fault_mutex_table[hash]); | |
ac9b9c66 | 5772 | return ret; |
4c887265 AL |
5773 | |
5774 | backout: | |
cb900f41 | 5775 | spin_unlock(ptl); |
2b26736c | 5776 | backout_unlocked: |
c7b1850d MK |
5777 | if (new_page && !new_pagecache_page) |
5778 | restore_reserve_on_error(h, vma, haddr, page); | |
fa27759a MK |
5779 | |
5780 | unlock_page(page); | |
4c887265 AL |
5781 | put_page(page); |
5782 | goto out; | |
ac9b9c66 HD |
5783 | } |
5784 | ||
8382d914 | 5785 | #ifdef CONFIG_SMP |
188b04a7 | 5786 | u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx) |
8382d914 DB |
5787 | { |
5788 | unsigned long key[2]; | |
5789 | u32 hash; | |
5790 | ||
1b426bac MK |
5791 | key[0] = (unsigned long) mapping; |
5792 | key[1] = idx; | |
8382d914 | 5793 | |
55254636 | 5794 | hash = jhash2((u32 *)&key, sizeof(key)/(sizeof(u32)), 0); |
8382d914 DB |
5795 | |
5796 | return hash & (num_fault_mutexes - 1); | |
5797 | } | |
5798 | #else | |
5799 | /* | |
6c26d310 | 5800 | * For uniprocessor systems we always use a single mutex, so just |
8382d914 DB |
5801 | * return 0 and avoid the hashing overhead. |
5802 | */ | |
188b04a7 | 5803 | u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx) |
8382d914 DB |
5804 | { |
5805 | return 0; | |
5806 | } | |
5807 | #endif | |
5808 | ||
2b740303 | 5809 | vm_fault_t hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma, |
788c7df4 | 5810 | unsigned long address, unsigned int flags) |
86e5216f | 5811 | { |
8382d914 | 5812 | pte_t *ptep, entry; |
cb900f41 | 5813 | spinlock_t *ptl; |
2b740303 | 5814 | vm_fault_t ret; |
8382d914 DB |
5815 | u32 hash; |
5816 | pgoff_t idx; | |
0fe6e20b | 5817 | struct page *page = NULL; |
57303d80 | 5818 | struct page *pagecache_page = NULL; |
a5516438 | 5819 | struct hstate *h = hstate_vma(vma); |
8382d914 | 5820 | struct address_space *mapping; |
0f792cf9 | 5821 | int need_wait_lock = 0; |
285b8dca | 5822 | unsigned long haddr = address & huge_page_mask(h); |
86e5216f | 5823 | |
285b8dca | 5824 | ptep = huge_pte_offset(mm, haddr, huge_page_size(h)); |
fd6a03ed | 5825 | if (ptep) { |
40549ba8 MK |
5826 | /* |
5827 | * Since we hold no locks, ptep could be stale. That is | |
5828 | * OK as we are only making decisions based on content and | |
5829 | * not actually modifying content here. | |
5830 | */ | |
fd6a03ed | 5831 | entry = huge_ptep_get(ptep); |
290408d4 | 5832 | if (unlikely(is_hugetlb_entry_migration(entry))) { |
ad1ac596 | 5833 | migration_entry_wait_huge(vma, ptep); |
290408d4 NH |
5834 | return 0; |
5835 | } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry))) | |
32f84528 | 5836 | return VM_FAULT_HWPOISON_LARGE | |
972dc4de | 5837 | VM_FAULT_SET_HINDEX(hstate_index(h)); |
fd6a03ed NH |
5838 | } |
5839 | ||
3935baa9 DG |
5840 | /* |
5841 | * Serialize hugepage allocation and instantiation, so that we don't | |
5842 | * get spurious allocation failures if two CPUs race to instantiate | |
5843 | * the same page in the page cache. | |
5844 | */ | |
40549ba8 MK |
5845 | mapping = vma->vm_file->f_mapping; |
5846 | idx = vma_hugecache_offset(h, vma, haddr); | |
188b04a7 | 5847 | hash = hugetlb_fault_mutex_hash(mapping, idx); |
c672c7f2 | 5848 | mutex_lock(&hugetlb_fault_mutex_table[hash]); |
8382d914 | 5849 | |
40549ba8 MK |
5850 | /* |
5851 | * Acquire vma lock before calling huge_pte_alloc and hold | |
5852 | * until finished with ptep. This prevents huge_pmd_unshare from | |
5853 | * being called elsewhere and making the ptep no longer valid. | |
5854 | * | |
5855 | * ptep could have already be assigned via huge_pte_offset. That | |
5856 | * is OK, as huge_pte_alloc will return the same value unless | |
5857 | * something has changed. | |
5858 | */ | |
5859 | hugetlb_vma_lock_read(vma); | |
5860 | ptep = huge_pte_alloc(mm, vma, haddr, huge_page_size(h)); | |
5861 | if (!ptep) { | |
5862 | hugetlb_vma_unlock_read(vma); | |
5863 | mutex_unlock(&hugetlb_fault_mutex_table[hash]); | |
5864 | return VM_FAULT_OOM; | |
5865 | } | |
5866 | ||
7f2e9525 | 5867 | entry = huge_ptep_get(ptep); |
c64e912c | 5868 | /* PTE markers should be handled the same way as none pte */ |
958f32ce LS |
5869 | if (huge_pte_none_mostly(entry)) |
5870 | /* | |
5871 | * hugetlb_no_page will drop vma lock and hugetlb fault | |
5872 | * mutex internally, which make us return immediately. | |
5873 | */ | |
5874 | return hugetlb_no_page(mm, vma, mapping, idx, address, ptep, | |
c64e912c | 5875 | entry, flags); |
86e5216f | 5876 | |
83c54070 | 5877 | ret = 0; |
1e8f889b | 5878 | |
0f792cf9 NH |
5879 | /* |
5880 | * entry could be a migration/hwpoison entry at this point, so this | |
5881 | * check prevents the kernel from going below assuming that we have | |
7c8de358 EP |
5882 | * an active hugepage in pagecache. This goto expects the 2nd page |
5883 | * fault, and is_hugetlb_entry_(migration|hwpoisoned) check will | |
5884 | * properly handle it. | |
0f792cf9 NH |
5885 | */ |
5886 | if (!pte_present(entry)) | |
5887 | goto out_mutex; | |
5888 | ||
57303d80 | 5889 | /* |
c89357e2 DH |
5890 | * If we are going to COW/unshare the mapping later, we examine the |
5891 | * pending reservations for this page now. This will ensure that any | |
57303d80 | 5892 | * allocations necessary to record that reservation occur outside the |
1d8d1464 DH |
5893 | * spinlock. Also lookup the pagecache page now as it is used to |
5894 | * determine if a reservation has been consumed. | |
57303d80 | 5895 | */ |
c89357e2 | 5896 | if ((flags & (FAULT_FLAG_WRITE|FAULT_FLAG_UNSHARE)) && |
1d8d1464 | 5897 | !(vma->vm_flags & VM_MAYSHARE) && !huge_pte_write(entry)) { |
285b8dca | 5898 | if (vma_needs_reservation(h, vma, haddr) < 0) { |
2b26736c | 5899 | ret = VM_FAULT_OOM; |
b4d1d99f | 5900 | goto out_mutex; |
2b26736c | 5901 | } |
5e911373 | 5902 | /* Just decrements count, does not deallocate */ |
285b8dca | 5903 | vma_end_reservation(h, vma, haddr); |
57303d80 | 5904 | |
29be8426 | 5905 | pagecache_page = find_lock_page(mapping, idx); |
57303d80 AW |
5906 | } |
5907 | ||
0f792cf9 NH |
5908 | ptl = huge_pte_lock(h, mm, ptep); |
5909 | ||
c89357e2 | 5910 | /* Check for a racing update before calling hugetlb_wp() */ |
0f792cf9 NH |
5911 | if (unlikely(!pte_same(entry, huge_ptep_get(ptep)))) |
5912 | goto out_ptl; | |
5913 | ||
166f3ecc PX |
5914 | /* Handle userfault-wp first, before trying to lock more pages */ |
5915 | if (userfaultfd_wp(vma) && huge_pte_uffd_wp(huge_ptep_get(ptep)) && | |
5916 | (flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) { | |
5917 | struct vm_fault vmf = { | |
5918 | .vma = vma, | |
5919 | .address = haddr, | |
5920 | .real_address = address, | |
5921 | .flags = flags, | |
5922 | }; | |
5923 | ||
5924 | spin_unlock(ptl); | |
5925 | if (pagecache_page) { | |
5926 | unlock_page(pagecache_page); | |
5927 | put_page(pagecache_page); | |
5928 | } | |
40549ba8 | 5929 | hugetlb_vma_unlock_read(vma); |
166f3ecc | 5930 | mutex_unlock(&hugetlb_fault_mutex_table[hash]); |
166f3ecc PX |
5931 | return handle_userfault(&vmf, VM_UFFD_WP); |
5932 | } | |
5933 | ||
56c9cfb1 | 5934 | /* |
c89357e2 | 5935 | * hugetlb_wp() requires page locks of pte_page(entry) and |
56c9cfb1 NH |
5936 | * pagecache_page, so here we need take the former one |
5937 | * when page != pagecache_page or !pagecache_page. | |
56c9cfb1 NH |
5938 | */ |
5939 | page = pte_page(entry); | |
5940 | if (page != pagecache_page) | |
0f792cf9 NH |
5941 | if (!trylock_page(page)) { |
5942 | need_wait_lock = 1; | |
5943 | goto out_ptl; | |
5944 | } | |
b4d1d99f | 5945 | |
0f792cf9 | 5946 | get_page(page); |
b4d1d99f | 5947 | |
c89357e2 | 5948 | if (flags & (FAULT_FLAG_WRITE|FAULT_FLAG_UNSHARE)) { |
106c992a | 5949 | if (!huge_pte_write(entry)) { |
c89357e2 DH |
5950 | ret = hugetlb_wp(mm, vma, address, ptep, flags, |
5951 | pagecache_page, ptl); | |
0f792cf9 | 5952 | goto out_put_page; |
c89357e2 DH |
5953 | } else if (likely(flags & FAULT_FLAG_WRITE)) { |
5954 | entry = huge_pte_mkdirty(entry); | |
b4d1d99f | 5955 | } |
b4d1d99f DG |
5956 | } |
5957 | entry = pte_mkyoung(entry); | |
285b8dca | 5958 | if (huge_ptep_set_access_flags(vma, haddr, ptep, entry, |
788c7df4 | 5959 | flags & FAULT_FLAG_WRITE)) |
285b8dca | 5960 | update_mmu_cache(vma, haddr, ptep); |
0f792cf9 NH |
5961 | out_put_page: |
5962 | if (page != pagecache_page) | |
5963 | unlock_page(page); | |
5964 | put_page(page); | |
cb900f41 KS |
5965 | out_ptl: |
5966 | spin_unlock(ptl); | |
57303d80 AW |
5967 | |
5968 | if (pagecache_page) { | |
5969 | unlock_page(pagecache_page); | |
5970 | put_page(pagecache_page); | |
5971 | } | |
b4d1d99f | 5972 | out_mutex: |
40549ba8 | 5973 | hugetlb_vma_unlock_read(vma); |
c672c7f2 | 5974 | mutex_unlock(&hugetlb_fault_mutex_table[hash]); |
0f792cf9 NH |
5975 | /* |
5976 | * Generally it's safe to hold refcount during waiting page lock. But | |
5977 | * here we just wait to defer the next page fault to avoid busy loop and | |
5978 | * the page is not used after unlocked before returning from the current | |
5979 | * page fault. So we are safe from accessing freed page, even if we wait | |
5980 | * here without taking refcount. | |
5981 | */ | |
5982 | if (need_wait_lock) | |
5983 | wait_on_page_locked(page); | |
1e8f889b | 5984 | return ret; |
86e5216f AL |
5985 | } |
5986 | ||
714c1891 | 5987 | #ifdef CONFIG_USERFAULTFD |
8fb5debc MK |
5988 | /* |
5989 | * Used by userfaultfd UFFDIO_COPY. Based on mcopy_atomic_pte with | |
5990 | * modifications for huge pages. | |
5991 | */ | |
5992 | int hugetlb_mcopy_atomic_pte(struct mm_struct *dst_mm, | |
5993 | pte_t *dst_pte, | |
5994 | struct vm_area_struct *dst_vma, | |
5995 | unsigned long dst_addr, | |
5996 | unsigned long src_addr, | |
f6191471 | 5997 | enum mcopy_atomic_mode mode, |
6041c691 PX |
5998 | struct page **pagep, |
5999 | bool wp_copy) | |
8fb5debc | 6000 | { |
f6191471 | 6001 | bool is_continue = (mode == MCOPY_ATOMIC_CONTINUE); |
8cc5fcbb MA |
6002 | struct hstate *h = hstate_vma(dst_vma); |
6003 | struct address_space *mapping = dst_vma->vm_file->f_mapping; | |
6004 | pgoff_t idx = vma_hugecache_offset(h, dst_vma, dst_addr); | |
1e392147 | 6005 | unsigned long size; |
1c9e8def | 6006 | int vm_shared = dst_vma->vm_flags & VM_SHARED; |
8fb5debc MK |
6007 | pte_t _dst_pte; |
6008 | spinlock_t *ptl; | |
8cc5fcbb | 6009 | int ret = -ENOMEM; |
8fb5debc | 6010 | struct page *page; |
f6191471 | 6011 | int writable; |
cc30042d | 6012 | bool page_in_pagecache = false; |
8fb5debc | 6013 | |
f6191471 AR |
6014 | if (is_continue) { |
6015 | ret = -EFAULT; | |
6016 | page = find_lock_page(mapping, idx); | |
6017 | if (!page) | |
6018 | goto out; | |
cc30042d | 6019 | page_in_pagecache = true; |
f6191471 | 6020 | } else if (!*pagep) { |
d84cf06e MA |
6021 | /* If a page already exists, then it's UFFDIO_COPY for |
6022 | * a non-missing case. Return -EEXIST. | |
6023 | */ | |
6024 | if (vm_shared && | |
6025 | hugetlbfs_pagecache_present(h, dst_vma, dst_addr)) { | |
6026 | ret = -EEXIST; | |
6027 | goto out; | |
6028 | } | |
6029 | ||
8fb5debc | 6030 | page = alloc_huge_page(dst_vma, dst_addr, 0); |
d84cf06e MA |
6031 | if (IS_ERR(page)) { |
6032 | ret = -ENOMEM; | |
8fb5debc | 6033 | goto out; |
d84cf06e | 6034 | } |
8fb5debc MK |
6035 | |
6036 | ret = copy_huge_page_from_user(page, | |
6037 | (const void __user *) src_addr, | |
810a56b9 | 6038 | pages_per_huge_page(h), false); |
8fb5debc | 6039 | |
c1e8d7c6 | 6040 | /* fallback to copy_from_user outside mmap_lock */ |
8fb5debc | 6041 | if (unlikely(ret)) { |
9e368259 | 6042 | ret = -ENOENT; |
8cc5fcbb MA |
6043 | /* Free the allocated page which may have |
6044 | * consumed a reservation. | |
6045 | */ | |
6046 | restore_reserve_on_error(h, dst_vma, dst_addr, page); | |
6047 | put_page(page); | |
6048 | ||
6049 | /* Allocate a temporary page to hold the copied | |
6050 | * contents. | |
6051 | */ | |
6052 | page = alloc_huge_page_vma(h, dst_vma, dst_addr); | |
6053 | if (!page) { | |
6054 | ret = -ENOMEM; | |
6055 | goto out; | |
6056 | } | |
8fb5debc | 6057 | *pagep = page; |
8cc5fcbb MA |
6058 | /* Set the outparam pagep and return to the caller to |
6059 | * copy the contents outside the lock. Don't free the | |
6060 | * page. | |
6061 | */ | |
8fb5debc MK |
6062 | goto out; |
6063 | } | |
6064 | } else { | |
8cc5fcbb MA |
6065 | if (vm_shared && |
6066 | hugetlbfs_pagecache_present(h, dst_vma, dst_addr)) { | |
6067 | put_page(*pagep); | |
6068 | ret = -EEXIST; | |
6069 | *pagep = NULL; | |
6070 | goto out; | |
6071 | } | |
6072 | ||
6073 | page = alloc_huge_page(dst_vma, dst_addr, 0); | |
6074 | if (IS_ERR(page)) { | |
da9a298f | 6075 | put_page(*pagep); |
8cc5fcbb MA |
6076 | ret = -ENOMEM; |
6077 | *pagep = NULL; | |
6078 | goto out; | |
6079 | } | |
34892366 MS |
6080 | copy_user_huge_page(page, *pagep, dst_addr, dst_vma, |
6081 | pages_per_huge_page(h)); | |
8cc5fcbb | 6082 | put_page(*pagep); |
8fb5debc MK |
6083 | *pagep = NULL; |
6084 | } | |
6085 | ||
6086 | /* | |
6087 | * The memory barrier inside __SetPageUptodate makes sure that | |
6088 | * preceding stores to the page contents become visible before | |
6089 | * the set_pte_at() write. | |
6090 | */ | |
6091 | __SetPageUptodate(page); | |
8fb5debc | 6092 | |
f6191471 AR |
6093 | /* Add shared, newly allocated pages to the page cache. */ |
6094 | if (vm_shared && !is_continue) { | |
1e392147 AA |
6095 | size = i_size_read(mapping->host) >> huge_page_shift(h); |
6096 | ret = -EFAULT; | |
6097 | if (idx >= size) | |
6098 | goto out_release_nounlock; | |
1c9e8def | 6099 | |
1e392147 AA |
6100 | /* |
6101 | * Serialization between remove_inode_hugepages() and | |
7e1813d4 | 6102 | * hugetlb_add_to_page_cache() below happens through the |
1e392147 AA |
6103 | * hugetlb_fault_mutex_table that here must be hold by |
6104 | * the caller. | |
6105 | */ | |
7e1813d4 | 6106 | ret = hugetlb_add_to_page_cache(page, mapping, idx); |
1c9e8def MK |
6107 | if (ret) |
6108 | goto out_release_nounlock; | |
cc30042d | 6109 | page_in_pagecache = true; |
1c9e8def MK |
6110 | } |
6111 | ||
bcc66543 | 6112 | ptl = huge_pte_lock(h, dst_mm, dst_pte); |
8fb5debc | 6113 | |
8625147c JH |
6114 | ret = -EIO; |
6115 | if (PageHWPoison(page)) | |
6116 | goto out_release_unlock; | |
6117 | ||
6041c691 PX |
6118 | /* |
6119 | * We allow to overwrite a pte marker: consider when both MISSING|WP | |
6120 | * registered, we firstly wr-protect a none pte which has no page cache | |
6121 | * page backing it, then access the page. | |
6122 | */ | |
fa27759a | 6123 | ret = -EEXIST; |
6041c691 | 6124 | if (!huge_pte_none_mostly(huge_ptep_get(dst_pte))) |
8fb5debc MK |
6125 | goto out_release_unlock; |
6126 | ||
ab74ef70 | 6127 | if (page_in_pagecache) { |
fb3d824d | 6128 | page_dup_file_rmap(page, true); |
1c9e8def | 6129 | } else { |
d6995da3 | 6130 | ClearHPageRestoreReserve(page); |
1c9e8def MK |
6131 | hugepage_add_new_anon_rmap(page, dst_vma, dst_addr); |
6132 | } | |
8fb5debc | 6133 | |
6041c691 PX |
6134 | /* |
6135 | * For either: (1) CONTINUE on a non-shared VMA, or (2) UFFDIO_COPY | |
6136 | * with wp flag set, don't set pte write bit. | |
6137 | */ | |
6138 | if (wp_copy || (is_continue && !vm_shared)) | |
f6191471 AR |
6139 | writable = 0; |
6140 | else | |
6141 | writable = dst_vma->vm_flags & VM_WRITE; | |
6142 | ||
6143 | _dst_pte = make_huge_pte(dst_vma, page, writable); | |
6041c691 PX |
6144 | /* |
6145 | * Always mark UFFDIO_COPY page dirty; note that this may not be | |
6146 | * extremely important for hugetlbfs for now since swapping is not | |
6147 | * supported, but we should still be clear in that this page cannot be | |
6148 | * thrown away at will, even if write bit not set. | |
6149 | */ | |
6150 | _dst_pte = huge_pte_mkdirty(_dst_pte); | |
8fb5debc MK |
6151 | _dst_pte = pte_mkyoung(_dst_pte); |
6152 | ||
6041c691 PX |
6153 | if (wp_copy) |
6154 | _dst_pte = huge_pte_mkuffd_wp(_dst_pte); | |
6155 | ||
8fb5debc MK |
6156 | set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte); |
6157 | ||
8fb5debc MK |
6158 | hugetlb_count_add(pages_per_huge_page(h), dst_mm); |
6159 | ||
6160 | /* No need to invalidate - it was non-present before */ | |
6161 | update_mmu_cache(dst_vma, dst_addr, dst_pte); | |
6162 | ||
6163 | spin_unlock(ptl); | |
f6191471 AR |
6164 | if (!is_continue) |
6165 | SetHPageMigratable(page); | |
6166 | if (vm_shared || is_continue) | |
1c9e8def | 6167 | unlock_page(page); |
8fb5debc MK |
6168 | ret = 0; |
6169 | out: | |
6170 | return ret; | |
6171 | out_release_unlock: | |
6172 | spin_unlock(ptl); | |
f6191471 | 6173 | if (vm_shared || is_continue) |
1c9e8def | 6174 | unlock_page(page); |
5af10dfd | 6175 | out_release_nounlock: |
cc30042d | 6176 | if (!page_in_pagecache) |
c7b1850d | 6177 | restore_reserve_on_error(h, dst_vma, dst_addr, page); |
8fb5debc MK |
6178 | put_page(page); |
6179 | goto out; | |
6180 | } | |
714c1891 | 6181 | #endif /* CONFIG_USERFAULTFD */ |
8fb5debc | 6182 | |
82e5d378 JM |
6183 | static void record_subpages_vmas(struct page *page, struct vm_area_struct *vma, |
6184 | int refs, struct page **pages, | |
6185 | struct vm_area_struct **vmas) | |
6186 | { | |
6187 | int nr; | |
6188 | ||
6189 | for (nr = 0; nr < refs; nr++) { | |
6190 | if (likely(pages)) | |
14455eab | 6191 | pages[nr] = nth_page(page, nr); |
82e5d378 JM |
6192 | if (vmas) |
6193 | vmas[nr] = vma; | |
6194 | } | |
6195 | } | |
6196 | ||
a7f22660 DH |
6197 | static inline bool __follow_hugetlb_must_fault(unsigned int flags, pte_t *pte, |
6198 | bool *unshare) | |
6199 | { | |
6200 | pte_t pteval = huge_ptep_get(pte); | |
6201 | ||
6202 | *unshare = false; | |
6203 | if (is_swap_pte(pteval)) | |
6204 | return true; | |
6205 | if (huge_pte_write(pteval)) | |
6206 | return false; | |
6207 | if (flags & FOLL_WRITE) | |
6208 | return true; | |
6209 | if (gup_must_unshare(flags, pte_page(pteval))) { | |
6210 | *unshare = true; | |
6211 | return true; | |
6212 | } | |
6213 | return false; | |
6214 | } | |
6215 | ||
28a35716 ML |
6216 | long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma, |
6217 | struct page **pages, struct vm_area_struct **vmas, | |
6218 | unsigned long *position, unsigned long *nr_pages, | |
4f6da934 | 6219 | long i, unsigned int flags, int *locked) |
63551ae0 | 6220 | { |
d5d4b0aa CK |
6221 | unsigned long pfn_offset; |
6222 | unsigned long vaddr = *position; | |
28a35716 | 6223 | unsigned long remainder = *nr_pages; |
a5516438 | 6224 | struct hstate *h = hstate_vma(vma); |
0fa5bc40 | 6225 | int err = -EFAULT, refs; |
63551ae0 | 6226 | |
63551ae0 | 6227 | while (vaddr < vma->vm_end && remainder) { |
4c887265 | 6228 | pte_t *pte; |
cb900f41 | 6229 | spinlock_t *ptl = NULL; |
a7f22660 | 6230 | bool unshare = false; |
2a15efc9 | 6231 | int absent; |
4c887265 | 6232 | struct page *page; |
63551ae0 | 6233 | |
02057967 DR |
6234 | /* |
6235 | * If we have a pending SIGKILL, don't keep faulting pages and | |
6236 | * potentially allocating memory. | |
6237 | */ | |
fa45f116 | 6238 | if (fatal_signal_pending(current)) { |
02057967 DR |
6239 | remainder = 0; |
6240 | break; | |
6241 | } | |
6242 | ||
4c887265 AL |
6243 | /* |
6244 | * Some archs (sparc64, sh*) have multiple pte_ts to | |
2a15efc9 | 6245 | * each hugepage. We have to make sure we get the |
4c887265 | 6246 | * first, for the page indexing below to work. |
cb900f41 KS |
6247 | * |
6248 | * Note that page table lock is not held when pte is null. | |
4c887265 | 6249 | */ |
7868a208 PA |
6250 | pte = huge_pte_offset(mm, vaddr & huge_page_mask(h), |
6251 | huge_page_size(h)); | |
cb900f41 KS |
6252 | if (pte) |
6253 | ptl = huge_pte_lock(h, mm, pte); | |
2a15efc9 HD |
6254 | absent = !pte || huge_pte_none(huge_ptep_get(pte)); |
6255 | ||
6256 | /* | |
6257 | * When coredumping, it suits get_dump_page if we just return | |
3ae77f43 HD |
6258 | * an error where there's an empty slot with no huge pagecache |
6259 | * to back it. This way, we avoid allocating a hugepage, and | |
6260 | * the sparse dumpfile avoids allocating disk blocks, but its | |
6261 | * huge holes still show up with zeroes where they need to be. | |
2a15efc9 | 6262 | */ |
3ae77f43 HD |
6263 | if (absent && (flags & FOLL_DUMP) && |
6264 | !hugetlbfs_pagecache_present(h, vma, vaddr)) { | |
cb900f41 KS |
6265 | if (pte) |
6266 | spin_unlock(ptl); | |
2a15efc9 HD |
6267 | remainder = 0; |
6268 | break; | |
6269 | } | |
63551ae0 | 6270 | |
9cc3a5bd NH |
6271 | /* |
6272 | * We need call hugetlb_fault for both hugepages under migration | |
6273 | * (in which case hugetlb_fault waits for the migration,) and | |
6274 | * hwpoisoned hugepages (in which case we need to prevent the | |
6275 | * caller from accessing to them.) In order to do this, we use | |
6276 | * here is_swap_pte instead of is_hugetlb_entry_migration and | |
6277 | * is_hugetlb_entry_hwpoisoned. This is because it simply covers | |
6278 | * both cases, and because we can't follow correct pages | |
6279 | * directly from any kind of swap entries. | |
6280 | */ | |
a7f22660 DH |
6281 | if (absent || |
6282 | __follow_hugetlb_must_fault(flags, pte, &unshare)) { | |
2b740303 | 6283 | vm_fault_t ret; |
87ffc118 | 6284 | unsigned int fault_flags = 0; |
63551ae0 | 6285 | |
cb900f41 KS |
6286 | if (pte) |
6287 | spin_unlock(ptl); | |
87ffc118 AA |
6288 | if (flags & FOLL_WRITE) |
6289 | fault_flags |= FAULT_FLAG_WRITE; | |
a7f22660 DH |
6290 | else if (unshare) |
6291 | fault_flags |= FAULT_FLAG_UNSHARE; | |
4f6da934 | 6292 | if (locked) |
71335f37 PX |
6293 | fault_flags |= FAULT_FLAG_ALLOW_RETRY | |
6294 | FAULT_FLAG_KILLABLE; | |
87ffc118 AA |
6295 | if (flags & FOLL_NOWAIT) |
6296 | fault_flags |= FAULT_FLAG_ALLOW_RETRY | | |
6297 | FAULT_FLAG_RETRY_NOWAIT; | |
6298 | if (flags & FOLL_TRIED) { | |
4426e945 PX |
6299 | /* |
6300 | * Note: FAULT_FLAG_ALLOW_RETRY and | |
6301 | * FAULT_FLAG_TRIED can co-exist | |
6302 | */ | |
87ffc118 AA |
6303 | fault_flags |= FAULT_FLAG_TRIED; |
6304 | } | |
6305 | ret = hugetlb_fault(mm, vma, vaddr, fault_flags); | |
6306 | if (ret & VM_FAULT_ERROR) { | |
2be7cfed | 6307 | err = vm_fault_to_errno(ret, flags); |
87ffc118 AA |
6308 | remainder = 0; |
6309 | break; | |
6310 | } | |
6311 | if (ret & VM_FAULT_RETRY) { | |
4f6da934 | 6312 | if (locked && |
1ac25013 | 6313 | !(fault_flags & FAULT_FLAG_RETRY_NOWAIT)) |
4f6da934 | 6314 | *locked = 0; |
87ffc118 AA |
6315 | *nr_pages = 0; |
6316 | /* | |
6317 | * VM_FAULT_RETRY must not return an | |
6318 | * error, it will return zero | |
6319 | * instead. | |
6320 | * | |
6321 | * No need to update "position" as the | |
6322 | * caller will not check it after | |
6323 | * *nr_pages is set to 0. | |
6324 | */ | |
6325 | return i; | |
6326 | } | |
6327 | continue; | |
4c887265 AL |
6328 | } |
6329 | ||
a5516438 | 6330 | pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT; |
7f2e9525 | 6331 | page = pte_page(huge_ptep_get(pte)); |
8fde12ca | 6332 | |
b6a2619c DH |
6333 | VM_BUG_ON_PAGE((flags & FOLL_PIN) && PageAnon(page) && |
6334 | !PageAnonExclusive(page), page); | |
6335 | ||
acbfb087 ZL |
6336 | /* |
6337 | * If subpage information not requested, update counters | |
6338 | * and skip the same_page loop below. | |
6339 | */ | |
6340 | if (!pages && !vmas && !pfn_offset && | |
6341 | (vaddr + huge_page_size(h) < vma->vm_end) && | |
6342 | (remainder >= pages_per_huge_page(h))) { | |
6343 | vaddr += huge_page_size(h); | |
6344 | remainder -= pages_per_huge_page(h); | |
6345 | i += pages_per_huge_page(h); | |
6346 | spin_unlock(ptl); | |
6347 | continue; | |
6348 | } | |
6349 | ||
d08af0a5 JM |
6350 | /* vaddr may not be aligned to PAGE_SIZE */ |
6351 | refs = min3(pages_per_huge_page(h) - pfn_offset, remainder, | |
6352 | (vma->vm_end - ALIGN_DOWN(vaddr, PAGE_SIZE)) >> PAGE_SHIFT); | |
0fa5bc40 | 6353 | |
82e5d378 | 6354 | if (pages || vmas) |
14455eab | 6355 | record_subpages_vmas(nth_page(page, pfn_offset), |
82e5d378 JM |
6356 | vma, refs, |
6357 | likely(pages) ? pages + i : NULL, | |
6358 | vmas ? vmas + i : NULL); | |
63551ae0 | 6359 | |
82e5d378 | 6360 | if (pages) { |
0fa5bc40 | 6361 | /* |
822951d8 | 6362 | * try_grab_folio() should always succeed here, |
0fa5bc40 JM |
6363 | * because: a) we hold the ptl lock, and b) we've just |
6364 | * checked that the huge page is present in the page | |
6365 | * tables. If the huge page is present, then the tail | |
6366 | * pages must also be present. The ptl prevents the | |
6367 | * head page and tail pages from being rearranged in | |
6368 | * any way. So this page must be available at this | |
6369 | * point, unless the page refcount overflowed: | |
6370 | */ | |
822951d8 MWO |
6371 | if (WARN_ON_ONCE(!try_grab_folio(pages[i], refs, |
6372 | flags))) { | |
0fa5bc40 JM |
6373 | spin_unlock(ptl); |
6374 | remainder = 0; | |
6375 | err = -ENOMEM; | |
6376 | break; | |
6377 | } | |
d5d4b0aa | 6378 | } |
82e5d378 JM |
6379 | |
6380 | vaddr += (refs << PAGE_SHIFT); | |
6381 | remainder -= refs; | |
6382 | i += refs; | |
6383 | ||
cb900f41 | 6384 | spin_unlock(ptl); |
63551ae0 | 6385 | } |
28a35716 | 6386 | *nr_pages = remainder; |
87ffc118 AA |
6387 | /* |
6388 | * setting position is actually required only if remainder is | |
6389 | * not zero but it's faster not to add a "if (remainder)" | |
6390 | * branch. | |
6391 | */ | |
63551ae0 DG |
6392 | *position = vaddr; |
6393 | ||
2be7cfed | 6394 | return i ? i : err; |
63551ae0 | 6395 | } |
8f860591 | 6396 | |
7da4d641 | 6397 | unsigned long hugetlb_change_protection(struct vm_area_struct *vma, |
5a90d5a1 PX |
6398 | unsigned long address, unsigned long end, |
6399 | pgprot_t newprot, unsigned long cp_flags) | |
8f860591 ZY |
6400 | { |
6401 | struct mm_struct *mm = vma->vm_mm; | |
6402 | unsigned long start = address; | |
6403 | pte_t *ptep; | |
6404 | pte_t pte; | |
a5516438 | 6405 | struct hstate *h = hstate_vma(vma); |
60dfaad6 | 6406 | unsigned long pages = 0, psize = huge_page_size(h); |
dff11abe | 6407 | bool shared_pmd = false; |
ac46d4f3 | 6408 | struct mmu_notifier_range range; |
e95a9851 | 6409 | unsigned long last_addr_mask; |
5a90d5a1 PX |
6410 | bool uffd_wp = cp_flags & MM_CP_UFFD_WP; |
6411 | bool uffd_wp_resolve = cp_flags & MM_CP_UFFD_WP_RESOLVE; | |
dff11abe MK |
6412 | |
6413 | /* | |
6414 | * In the case of shared PMDs, the area to flush could be beyond | |
ac46d4f3 | 6415 | * start/end. Set range.start/range.end to cover the maximum possible |
dff11abe MK |
6416 | * range if PMD sharing is possible. |
6417 | */ | |
7269f999 JG |
6418 | mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_VMA, |
6419 | 0, vma, mm, start, end); | |
ac46d4f3 | 6420 | adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end); |
8f860591 ZY |
6421 | |
6422 | BUG_ON(address >= end); | |
ac46d4f3 | 6423 | flush_cache_range(vma, range.start, range.end); |
8f860591 | 6424 | |
ac46d4f3 | 6425 | mmu_notifier_invalidate_range_start(&range); |
40549ba8 | 6426 | hugetlb_vma_lock_write(vma); |
83cde9e8 | 6427 | i_mmap_lock_write(vma->vm_file->f_mapping); |
40549ba8 | 6428 | last_addr_mask = hugetlb_mask_last_page(h); |
60dfaad6 | 6429 | for (; address < end; address += psize) { |
cb900f41 | 6430 | spinlock_t *ptl; |
60dfaad6 | 6431 | ptep = huge_pte_offset(mm, address, psize); |
e95a9851 MK |
6432 | if (!ptep) { |
6433 | address |= last_addr_mask; | |
8f860591 | 6434 | continue; |
e95a9851 | 6435 | } |
cb900f41 | 6436 | ptl = huge_pte_lock(h, mm, ptep); |
4ddb4d91 | 6437 | if (huge_pmd_unshare(mm, vma, address, ptep)) { |
60dfaad6 PX |
6438 | /* |
6439 | * When uffd-wp is enabled on the vma, unshare | |
6440 | * shouldn't happen at all. Warn about it if it | |
6441 | * happened due to some reason. | |
6442 | */ | |
6443 | WARN_ON_ONCE(uffd_wp || uffd_wp_resolve); | |
7da4d641 | 6444 | pages++; |
cb900f41 | 6445 | spin_unlock(ptl); |
dff11abe | 6446 | shared_pmd = true; |
4ddb4d91 | 6447 | address |= last_addr_mask; |
39dde65c | 6448 | continue; |
7da4d641 | 6449 | } |
a8bda28d NH |
6450 | pte = huge_ptep_get(ptep); |
6451 | if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) { | |
6452 | spin_unlock(ptl); | |
6453 | continue; | |
6454 | } | |
6455 | if (unlikely(is_hugetlb_entry_migration(pte))) { | |
6456 | swp_entry_t entry = pte_to_swp_entry(pte); | |
6c287605 | 6457 | struct page *page = pfn_swap_entry_to_page(entry); |
a8bda28d | 6458 | |
6c287605 | 6459 | if (!is_readable_migration_entry(entry)) { |
a8bda28d NH |
6460 | pte_t newpte; |
6461 | ||
6c287605 DH |
6462 | if (PageAnon(page)) |
6463 | entry = make_readable_exclusive_migration_entry( | |
6464 | swp_offset(entry)); | |
6465 | else | |
6466 | entry = make_readable_migration_entry( | |
6467 | swp_offset(entry)); | |
a8bda28d | 6468 | newpte = swp_entry_to_pte(entry); |
5a90d5a1 PX |
6469 | if (uffd_wp) |
6470 | newpte = pte_swp_mkuffd_wp(newpte); | |
6471 | else if (uffd_wp_resolve) | |
6472 | newpte = pte_swp_clear_uffd_wp(newpte); | |
18f39629 | 6473 | set_huge_pte_at(mm, address, ptep, newpte); |
a8bda28d NH |
6474 | pages++; |
6475 | } | |
6476 | spin_unlock(ptl); | |
6477 | continue; | |
6478 | } | |
60dfaad6 PX |
6479 | if (unlikely(pte_marker_uffd_wp(pte))) { |
6480 | /* | |
6481 | * This is changing a non-present pte into a none pte, | |
6482 | * no need for huge_ptep_modify_prot_start/commit(). | |
6483 | */ | |
6484 | if (uffd_wp_resolve) | |
6485 | huge_pte_clear(mm, address, ptep, psize); | |
6486 | } | |
a8bda28d | 6487 | if (!huge_pte_none(pte)) { |
023bdd00 | 6488 | pte_t old_pte; |
79c1c594 | 6489 | unsigned int shift = huge_page_shift(hstate_vma(vma)); |
023bdd00 AK |
6490 | |
6491 | old_pte = huge_ptep_modify_prot_start(vma, address, ptep); | |
16785bd7 | 6492 | pte = huge_pte_modify(old_pte, newprot); |
79c1c594 | 6493 | pte = arch_make_huge_pte(pte, shift, vma->vm_flags); |
5a90d5a1 PX |
6494 | if (uffd_wp) |
6495 | pte = huge_pte_mkuffd_wp(huge_pte_wrprotect(pte)); | |
6496 | else if (uffd_wp_resolve) | |
6497 | pte = huge_pte_clear_uffd_wp(pte); | |
023bdd00 | 6498 | huge_ptep_modify_prot_commit(vma, address, ptep, old_pte, pte); |
7da4d641 | 6499 | pages++; |
60dfaad6 PX |
6500 | } else { |
6501 | /* None pte */ | |
6502 | if (unlikely(uffd_wp)) | |
6503 | /* Safe to modify directly (none->non-present). */ | |
6504 | set_huge_pte_at(mm, address, ptep, | |
6505 | make_pte_marker(PTE_MARKER_UFFD_WP)); | |
8f860591 | 6506 | } |
cb900f41 | 6507 | spin_unlock(ptl); |
8f860591 | 6508 | } |
d833352a | 6509 | /* |
c8c06efa | 6510 | * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare |
d833352a | 6511 | * may have cleared our pud entry and done put_page on the page table: |
c8c06efa | 6512 | * once we release i_mmap_rwsem, another task can do the final put_page |
dff11abe MK |
6513 | * and that page table be reused and filled with junk. If we actually |
6514 | * did unshare a page of pmds, flush the range corresponding to the pud. | |
d833352a | 6515 | */ |
dff11abe | 6516 | if (shared_pmd) |
ac46d4f3 | 6517 | flush_hugetlb_tlb_range(vma, range.start, range.end); |
dff11abe MK |
6518 | else |
6519 | flush_hugetlb_tlb_range(vma, start, end); | |
0f10851e JG |
6520 | /* |
6521 | * No need to call mmu_notifier_invalidate_range() we are downgrading | |
6522 | * page table protection not changing it to point to a new page. | |
6523 | * | |
ee65728e | 6524 | * See Documentation/mm/mmu_notifier.rst |
0f10851e | 6525 | */ |
83cde9e8 | 6526 | i_mmap_unlock_write(vma->vm_file->f_mapping); |
40549ba8 | 6527 | hugetlb_vma_unlock_write(vma); |
ac46d4f3 | 6528 | mmu_notifier_invalidate_range_end(&range); |
7da4d641 PZ |
6529 | |
6530 | return pages << h->order; | |
8f860591 ZY |
6531 | } |
6532 | ||
33b8f84a MK |
6533 | /* Return true if reservation was successful, false otherwise. */ |
6534 | bool hugetlb_reserve_pages(struct inode *inode, | |
a1e78772 | 6535 | long from, long to, |
5a6fe125 | 6536 | struct vm_area_struct *vma, |
ca16d140 | 6537 | vm_flags_t vm_flags) |
e4e574b7 | 6538 | { |
33b8f84a | 6539 | long chg, add = -1; |
a5516438 | 6540 | struct hstate *h = hstate_inode(inode); |
90481622 | 6541 | struct hugepage_subpool *spool = subpool_inode(inode); |
9119a41e | 6542 | struct resv_map *resv_map; |
075a61d0 | 6543 | struct hugetlb_cgroup *h_cg = NULL; |
0db9d74e | 6544 | long gbl_reserve, regions_needed = 0; |
e4e574b7 | 6545 | |
63489f8e MK |
6546 | /* This should never happen */ |
6547 | if (from > to) { | |
6548 | VM_WARN(1, "%s called with a negative range\n", __func__); | |
33b8f84a | 6549 | return false; |
63489f8e MK |
6550 | } |
6551 | ||
8d9bfb26 MK |
6552 | /* |
6553 | * vma specific semaphore used for pmd sharing synchronization | |
6554 | */ | |
6555 | hugetlb_vma_lock_alloc(vma); | |
6556 | ||
17c9d12e MG |
6557 | /* |
6558 | * Only apply hugepage reservation if asked. At fault time, an | |
6559 | * attempt will be made for VM_NORESERVE to allocate a page | |
90481622 | 6560 | * without using reserves |
17c9d12e | 6561 | */ |
ca16d140 | 6562 | if (vm_flags & VM_NORESERVE) |
33b8f84a | 6563 | return true; |
17c9d12e | 6564 | |
a1e78772 MG |
6565 | /* |
6566 | * Shared mappings base their reservation on the number of pages that | |
6567 | * are already allocated on behalf of the file. Private mappings need | |
6568 | * to reserve the full area even if read-only as mprotect() may be | |
6569 | * called to make the mapping read-write. Assume !vma is a shm mapping | |
6570 | */ | |
9119a41e | 6571 | if (!vma || vma->vm_flags & VM_MAYSHARE) { |
f27a5136 MK |
6572 | /* |
6573 | * resv_map can not be NULL as hugetlb_reserve_pages is only | |
6574 | * called for inodes for which resv_maps were created (see | |
6575 | * hugetlbfs_get_inode). | |
6576 | */ | |
4e35f483 | 6577 | resv_map = inode_resv_map(inode); |
9119a41e | 6578 | |
0db9d74e | 6579 | chg = region_chg(resv_map, from, to, ®ions_needed); |
9119a41e | 6580 | } else { |
e9fe92ae | 6581 | /* Private mapping. */ |
9119a41e | 6582 | resv_map = resv_map_alloc(); |
17c9d12e | 6583 | if (!resv_map) |
8d9bfb26 | 6584 | goto out_err; |
17c9d12e | 6585 | |
a1e78772 | 6586 | chg = to - from; |
84afd99b | 6587 | |
17c9d12e MG |
6588 | set_vma_resv_map(vma, resv_map); |
6589 | set_vma_resv_flags(vma, HPAGE_RESV_OWNER); | |
6590 | } | |
6591 | ||
33b8f84a | 6592 | if (chg < 0) |
c50ac050 | 6593 | goto out_err; |
8a630112 | 6594 | |
33b8f84a MK |
6595 | if (hugetlb_cgroup_charge_cgroup_rsvd(hstate_index(h), |
6596 | chg * pages_per_huge_page(h), &h_cg) < 0) | |
075a61d0 | 6597 | goto out_err; |
075a61d0 MA |
6598 | |
6599 | if (vma && !(vma->vm_flags & VM_MAYSHARE) && h_cg) { | |
6600 | /* For private mappings, the hugetlb_cgroup uncharge info hangs | |
6601 | * of the resv_map. | |
6602 | */ | |
6603 | resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, h_cg, h); | |
6604 | } | |
6605 | ||
1c5ecae3 MK |
6606 | /* |
6607 | * There must be enough pages in the subpool for the mapping. If | |
6608 | * the subpool has a minimum size, there may be some global | |
6609 | * reservations already in place (gbl_reserve). | |
6610 | */ | |
6611 | gbl_reserve = hugepage_subpool_get_pages(spool, chg); | |
33b8f84a | 6612 | if (gbl_reserve < 0) |
075a61d0 | 6613 | goto out_uncharge_cgroup; |
5a6fe125 MG |
6614 | |
6615 | /* | |
17c9d12e | 6616 | * Check enough hugepages are available for the reservation. |
90481622 | 6617 | * Hand the pages back to the subpool if there are not |
5a6fe125 | 6618 | */ |
33b8f84a | 6619 | if (hugetlb_acct_memory(h, gbl_reserve) < 0) |
075a61d0 | 6620 | goto out_put_pages; |
17c9d12e MG |
6621 | |
6622 | /* | |
6623 | * Account for the reservations made. Shared mappings record regions | |
6624 | * that have reservations as they are shared by multiple VMAs. | |
6625 | * When the last VMA disappears, the region map says how much | |
6626 | * the reservation was and the page cache tells how much of | |
6627 | * the reservation was consumed. Private mappings are per-VMA and | |
6628 | * only the consumed reservations are tracked. When the VMA | |
6629 | * disappears, the original reservation is the VMA size and the | |
6630 | * consumed reservations are stored in the map. Hence, nothing | |
6631 | * else has to be done for private mappings here | |
6632 | */ | |
33039678 | 6633 | if (!vma || vma->vm_flags & VM_MAYSHARE) { |
075a61d0 | 6634 | add = region_add(resv_map, from, to, regions_needed, h, h_cg); |
0db9d74e MA |
6635 | |
6636 | if (unlikely(add < 0)) { | |
6637 | hugetlb_acct_memory(h, -gbl_reserve); | |
075a61d0 | 6638 | goto out_put_pages; |
0db9d74e | 6639 | } else if (unlikely(chg > add)) { |
33039678 MK |
6640 | /* |
6641 | * pages in this range were added to the reserve | |
6642 | * map between region_chg and region_add. This | |
6643 | * indicates a race with alloc_huge_page. Adjust | |
6644 | * the subpool and reserve counts modified above | |
6645 | * based on the difference. | |
6646 | */ | |
6647 | long rsv_adjust; | |
6648 | ||
d85aecf2 ML |
6649 | /* |
6650 | * hugetlb_cgroup_uncharge_cgroup_rsvd() will put the | |
6651 | * reference to h_cg->css. See comment below for detail. | |
6652 | */ | |
075a61d0 MA |
6653 | hugetlb_cgroup_uncharge_cgroup_rsvd( |
6654 | hstate_index(h), | |
6655 | (chg - add) * pages_per_huge_page(h), h_cg); | |
6656 | ||
33039678 MK |
6657 | rsv_adjust = hugepage_subpool_put_pages(spool, |
6658 | chg - add); | |
6659 | hugetlb_acct_memory(h, -rsv_adjust); | |
d85aecf2 ML |
6660 | } else if (h_cg) { |
6661 | /* | |
6662 | * The file_regions will hold their own reference to | |
6663 | * h_cg->css. So we should release the reference held | |
6664 | * via hugetlb_cgroup_charge_cgroup_rsvd() when we are | |
6665 | * done. | |
6666 | */ | |
6667 | hugetlb_cgroup_put_rsvd_cgroup(h_cg); | |
33039678 MK |
6668 | } |
6669 | } | |
33b8f84a MK |
6670 | return true; |
6671 | ||
075a61d0 MA |
6672 | out_put_pages: |
6673 | /* put back original number of pages, chg */ | |
6674 | (void)hugepage_subpool_put_pages(spool, chg); | |
6675 | out_uncharge_cgroup: | |
6676 | hugetlb_cgroup_uncharge_cgroup_rsvd(hstate_index(h), | |
6677 | chg * pages_per_huge_page(h), h_cg); | |
c50ac050 | 6678 | out_err: |
8d9bfb26 | 6679 | hugetlb_vma_lock_free(vma); |
5e911373 | 6680 | if (!vma || vma->vm_flags & VM_MAYSHARE) |
0db9d74e MA |
6681 | /* Only call region_abort if the region_chg succeeded but the |
6682 | * region_add failed or didn't run. | |
6683 | */ | |
6684 | if (chg >= 0 && add < 0) | |
6685 | region_abort(resv_map, from, to, regions_needed); | |
f031dd27 JK |
6686 | if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) |
6687 | kref_put(&resv_map->refs, resv_map_release); | |
33b8f84a | 6688 | return false; |
a43a8c39 CK |
6689 | } |
6690 | ||
b5cec28d MK |
6691 | long hugetlb_unreserve_pages(struct inode *inode, long start, long end, |
6692 | long freed) | |
a43a8c39 | 6693 | { |
a5516438 | 6694 | struct hstate *h = hstate_inode(inode); |
4e35f483 | 6695 | struct resv_map *resv_map = inode_resv_map(inode); |
9119a41e | 6696 | long chg = 0; |
90481622 | 6697 | struct hugepage_subpool *spool = subpool_inode(inode); |
1c5ecae3 | 6698 | long gbl_reserve; |
45c682a6 | 6699 | |
f27a5136 MK |
6700 | /* |
6701 | * Since this routine can be called in the evict inode path for all | |
6702 | * hugetlbfs inodes, resv_map could be NULL. | |
6703 | */ | |
b5cec28d MK |
6704 | if (resv_map) { |
6705 | chg = region_del(resv_map, start, end); | |
6706 | /* | |
6707 | * region_del() can fail in the rare case where a region | |
6708 | * must be split and another region descriptor can not be | |
6709 | * allocated. If end == LONG_MAX, it will not fail. | |
6710 | */ | |
6711 | if (chg < 0) | |
6712 | return chg; | |
6713 | } | |
6714 | ||
45c682a6 | 6715 | spin_lock(&inode->i_lock); |
e4c6f8be | 6716 | inode->i_blocks -= (blocks_per_huge_page(h) * freed); |
45c682a6 KC |
6717 | spin_unlock(&inode->i_lock); |
6718 | ||
1c5ecae3 MK |
6719 | /* |
6720 | * If the subpool has a minimum size, the number of global | |
6721 | * reservations to be released may be adjusted. | |
dddf31a4 ML |
6722 | * |
6723 | * Note that !resv_map implies freed == 0. So (chg - freed) | |
6724 | * won't go negative. | |
1c5ecae3 MK |
6725 | */ |
6726 | gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed)); | |
6727 | hugetlb_acct_memory(h, -gbl_reserve); | |
b5cec28d MK |
6728 | |
6729 | return 0; | |
a43a8c39 | 6730 | } |
93f70f90 | 6731 | |
3212b535 SC |
6732 | #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE |
6733 | static unsigned long page_table_shareable(struct vm_area_struct *svma, | |
6734 | struct vm_area_struct *vma, | |
6735 | unsigned long addr, pgoff_t idx) | |
6736 | { | |
6737 | unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) + | |
6738 | svma->vm_start; | |
6739 | unsigned long sbase = saddr & PUD_MASK; | |
6740 | unsigned long s_end = sbase + PUD_SIZE; | |
6741 | ||
6742 | /* Allow segments to share if only one is marked locked */ | |
de60f5f1 EM |
6743 | unsigned long vm_flags = vma->vm_flags & VM_LOCKED_CLEAR_MASK; |
6744 | unsigned long svm_flags = svma->vm_flags & VM_LOCKED_CLEAR_MASK; | |
3212b535 SC |
6745 | |
6746 | /* | |
6747 | * match the virtual addresses, permission and the alignment of the | |
6748 | * page table page. | |
131a79b4 MK |
6749 | * |
6750 | * Also, vma_lock (vm_private_data) is required for sharing. | |
3212b535 SC |
6751 | */ |
6752 | if (pmd_index(addr) != pmd_index(saddr) || | |
6753 | vm_flags != svm_flags || | |
131a79b4 MK |
6754 | !range_in_vma(svma, sbase, s_end) || |
6755 | !svma->vm_private_data) | |
3212b535 SC |
6756 | return 0; |
6757 | ||
6758 | return saddr; | |
6759 | } | |
6760 | ||
bbff39cc | 6761 | bool want_pmd_share(struct vm_area_struct *vma, unsigned long addr) |
3212b535 | 6762 | { |
bbff39cc MK |
6763 | unsigned long start = addr & PUD_MASK; |
6764 | unsigned long end = start + PUD_SIZE; | |
6765 | ||
8d9bfb26 MK |
6766 | #ifdef CONFIG_USERFAULTFD |
6767 | if (uffd_disable_huge_pmd_share(vma)) | |
6768 | return false; | |
6769 | #endif | |
3212b535 SC |
6770 | /* |
6771 | * check on proper vm_flags and page table alignment | |
6772 | */ | |
8d9bfb26 MK |
6773 | if (!(vma->vm_flags & VM_MAYSHARE)) |
6774 | return false; | |
bbff39cc | 6775 | if (!vma->vm_private_data) /* vma lock required for sharing */ |
8d9bfb26 MK |
6776 | return false; |
6777 | if (!range_in_vma(vma, start, end)) | |
6778 | return false; | |
6779 | return true; | |
6780 | } | |
6781 | ||
017b1660 MK |
6782 | /* |
6783 | * Determine if start,end range within vma could be mapped by shared pmd. | |
6784 | * If yes, adjust start and end to cover range associated with possible | |
6785 | * shared pmd mappings. | |
6786 | */ | |
6787 | void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma, | |
6788 | unsigned long *start, unsigned long *end) | |
6789 | { | |
a1ba9da8 LX |
6790 | unsigned long v_start = ALIGN(vma->vm_start, PUD_SIZE), |
6791 | v_end = ALIGN_DOWN(vma->vm_end, PUD_SIZE); | |
017b1660 | 6792 | |
a1ba9da8 | 6793 | /* |
f0953a1b IM |
6794 | * vma needs to span at least one aligned PUD size, and the range |
6795 | * must be at least partially within in. | |
a1ba9da8 LX |
6796 | */ |
6797 | if (!(vma->vm_flags & VM_MAYSHARE) || !(v_end > v_start) || | |
6798 | (*end <= v_start) || (*start >= v_end)) | |
017b1660 MK |
6799 | return; |
6800 | ||
75802ca6 | 6801 | /* Extend the range to be PUD aligned for a worst case scenario */ |
a1ba9da8 LX |
6802 | if (*start > v_start) |
6803 | *start = ALIGN_DOWN(*start, PUD_SIZE); | |
017b1660 | 6804 | |
a1ba9da8 LX |
6805 | if (*end < v_end) |
6806 | *end = ALIGN(*end, PUD_SIZE); | |
017b1660 MK |
6807 | } |
6808 | ||
8d9bfb26 MK |
6809 | static bool __vma_shareable_flags_pmd(struct vm_area_struct *vma) |
6810 | { | |
6811 | return vma->vm_flags & (VM_MAYSHARE | VM_SHARED) && | |
6812 | vma->vm_private_data; | |
6813 | } | |
6814 | ||
6815 | void hugetlb_vma_lock_read(struct vm_area_struct *vma) | |
6816 | { | |
6817 | if (__vma_shareable_flags_pmd(vma)) { | |
6818 | struct hugetlb_vma_lock *vma_lock = vma->vm_private_data; | |
6819 | ||
6820 | down_read(&vma_lock->rw_sema); | |
6821 | } | |
6822 | } | |
6823 | ||
6824 | void hugetlb_vma_unlock_read(struct vm_area_struct *vma) | |
6825 | { | |
6826 | if (__vma_shareable_flags_pmd(vma)) { | |
6827 | struct hugetlb_vma_lock *vma_lock = vma->vm_private_data; | |
6828 | ||
6829 | up_read(&vma_lock->rw_sema); | |
6830 | } | |
6831 | } | |
6832 | ||
6833 | void hugetlb_vma_lock_write(struct vm_area_struct *vma) | |
6834 | { | |
6835 | if (__vma_shareable_flags_pmd(vma)) { | |
6836 | struct hugetlb_vma_lock *vma_lock = vma->vm_private_data; | |
6837 | ||
6838 | down_write(&vma_lock->rw_sema); | |
6839 | } | |
6840 | } | |
6841 | ||
6842 | void hugetlb_vma_unlock_write(struct vm_area_struct *vma) | |
6843 | { | |
6844 | if (__vma_shareable_flags_pmd(vma)) { | |
6845 | struct hugetlb_vma_lock *vma_lock = vma->vm_private_data; | |
6846 | ||
6847 | up_write(&vma_lock->rw_sema); | |
6848 | } | |
6849 | } | |
6850 | ||
6851 | int hugetlb_vma_trylock_write(struct vm_area_struct *vma) | |
6852 | { | |
6853 | struct hugetlb_vma_lock *vma_lock = vma->vm_private_data; | |
6854 | ||
6855 | if (!__vma_shareable_flags_pmd(vma)) | |
6856 | return 1; | |
6857 | ||
6858 | return down_write_trylock(&vma_lock->rw_sema); | |
6859 | } | |
6860 | ||
6861 | void hugetlb_vma_assert_locked(struct vm_area_struct *vma) | |
6862 | { | |
6863 | if (__vma_shareable_flags_pmd(vma)) { | |
6864 | struct hugetlb_vma_lock *vma_lock = vma->vm_private_data; | |
6865 | ||
6866 | lockdep_assert_held(&vma_lock->rw_sema); | |
6867 | } | |
6868 | } | |
6869 | ||
6870 | void hugetlb_vma_lock_release(struct kref *kref) | |
6871 | { | |
6872 | struct hugetlb_vma_lock *vma_lock = container_of(kref, | |
6873 | struct hugetlb_vma_lock, refs); | |
6874 | ||
6875 | kfree(vma_lock); | |
6876 | } | |
6877 | ||
acfac378 | 6878 | static void __hugetlb_vma_unlock_write_put(struct hugetlb_vma_lock *vma_lock) |
ecfbd733 MK |
6879 | { |
6880 | struct vm_area_struct *vma = vma_lock->vma; | |
6881 | ||
6882 | /* | |
6883 | * vma_lock structure may or not be released as a result of put, | |
6884 | * it certainly will no longer be attached to vma so clear pointer. | |
6885 | * Semaphore synchronizes access to vma_lock->vma field. | |
6886 | */ | |
6887 | vma_lock->vma = NULL; | |
6888 | vma->vm_private_data = NULL; | |
6889 | up_write(&vma_lock->rw_sema); | |
6890 | kref_put(&vma_lock->refs, hugetlb_vma_lock_release); | |
6891 | } | |
6892 | ||
6893 | static void __hugetlb_vma_unlock_write_free(struct vm_area_struct *vma) | |
6894 | { | |
6895 | if (__vma_shareable_flags_pmd(vma)) { | |
6896 | struct hugetlb_vma_lock *vma_lock = vma->vm_private_data; | |
6897 | ||
6898 | __hugetlb_vma_unlock_write_put(vma_lock); | |
6899 | } | |
6900 | } | |
6901 | ||
8d9bfb26 MK |
6902 | static void hugetlb_vma_lock_free(struct vm_area_struct *vma) |
6903 | { | |
6904 | /* | |
131a79b4 | 6905 | * Only present in sharable vmas. |
8d9bfb26 | 6906 | */ |
131a79b4 | 6907 | if (!vma || !__vma_shareable_flags_pmd(vma)) |
8d9bfb26 MK |
6908 | return; |
6909 | ||
6910 | if (vma->vm_private_data) { | |
6911 | struct hugetlb_vma_lock *vma_lock = vma->vm_private_data; | |
6912 | ||
ecfbd733 MK |
6913 | down_write(&vma_lock->rw_sema); |
6914 | __hugetlb_vma_unlock_write_put(vma_lock); | |
8d9bfb26 MK |
6915 | } |
6916 | } | |
6917 | ||
6918 | static void hugetlb_vma_lock_alloc(struct vm_area_struct *vma) | |
6919 | { | |
6920 | struct hugetlb_vma_lock *vma_lock; | |
6921 | ||
6922 | /* Only establish in (flags) sharable vmas */ | |
6923 | if (!vma || !(vma->vm_flags & VM_MAYSHARE)) | |
6924 | return; | |
6925 | ||
6926 | /* Should never get here with non-NULL vm_private_data */ | |
6927 | if (vma->vm_private_data) | |
6928 | return; | |
6929 | ||
8d9bfb26 | 6930 | vma_lock = kmalloc(sizeof(*vma_lock), GFP_KERNEL); |
bbff39cc | 6931 | if (!vma_lock) { |
8d9bfb26 MK |
6932 | /* |
6933 | * If we can not allocate structure, then vma can not | |
bbff39cc MK |
6934 | * participate in pmd sharing. This is only a possible |
6935 | * performance enhancement and memory saving issue. | |
6936 | * However, the lock is also used to synchronize page | |
6937 | * faults with truncation. If the lock is not present, | |
6938 | * unlikely races could leave pages in a file past i_size | |
6939 | * until the file is removed. Warn in the unlikely case of | |
6940 | * allocation failure. | |
8d9bfb26 | 6941 | */ |
bbff39cc | 6942 | pr_warn_once("HugeTLB: unable to allocate vma specific lock\n"); |
8d9bfb26 | 6943 | return; |
bbff39cc | 6944 | } |
8d9bfb26 MK |
6945 | |
6946 | kref_init(&vma_lock->refs); | |
6947 | init_rwsem(&vma_lock->rw_sema); | |
6948 | vma_lock->vma = vma; | |
6949 | vma->vm_private_data = vma_lock; | |
6950 | } | |
6951 | ||
3212b535 SC |
6952 | /* |
6953 | * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc() | |
6954 | * and returns the corresponding pte. While this is not necessary for the | |
6955 | * !shared pmd case because we can allocate the pmd later as well, it makes the | |
3a47c54f MK |
6956 | * code much cleaner. pmd allocation is essential for the shared case because |
6957 | * pud has to be populated inside the same i_mmap_rwsem section - otherwise | |
6958 | * racing tasks could either miss the sharing (see huge_pte_offset) or select a | |
6959 | * bad pmd for sharing. | |
3212b535 | 6960 | */ |
aec44e0f PX |
6961 | pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct *vma, |
6962 | unsigned long addr, pud_t *pud) | |
3212b535 | 6963 | { |
3212b535 SC |
6964 | struct address_space *mapping = vma->vm_file->f_mapping; |
6965 | pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) + | |
6966 | vma->vm_pgoff; | |
6967 | struct vm_area_struct *svma; | |
6968 | unsigned long saddr; | |
6969 | pte_t *spte = NULL; | |
6970 | pte_t *pte; | |
cb900f41 | 6971 | spinlock_t *ptl; |
3212b535 | 6972 | |
3a47c54f | 6973 | i_mmap_lock_read(mapping); |
3212b535 SC |
6974 | vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) { |
6975 | if (svma == vma) | |
6976 | continue; | |
6977 | ||
6978 | saddr = page_table_shareable(svma, vma, addr, idx); | |
6979 | if (saddr) { | |
7868a208 PA |
6980 | spte = huge_pte_offset(svma->vm_mm, saddr, |
6981 | vma_mmu_pagesize(svma)); | |
3212b535 SC |
6982 | if (spte) { |
6983 | get_page(virt_to_page(spte)); | |
6984 | break; | |
6985 | } | |
6986 | } | |
6987 | } | |
6988 | ||
6989 | if (!spte) | |
6990 | goto out; | |
6991 | ||
8bea8052 | 6992 | ptl = huge_pte_lock(hstate_vma(vma), mm, spte); |
dc6c9a35 | 6993 | if (pud_none(*pud)) { |
3212b535 SC |
6994 | pud_populate(mm, pud, |
6995 | (pmd_t *)((unsigned long)spte & PAGE_MASK)); | |
c17b1f42 | 6996 | mm_inc_nr_pmds(mm); |
dc6c9a35 | 6997 | } else { |
3212b535 | 6998 | put_page(virt_to_page(spte)); |
dc6c9a35 | 6999 | } |
cb900f41 | 7000 | spin_unlock(ptl); |
3212b535 SC |
7001 | out: |
7002 | pte = (pte_t *)pmd_alloc(mm, pud, addr); | |
3a47c54f | 7003 | i_mmap_unlock_read(mapping); |
3212b535 SC |
7004 | return pte; |
7005 | } | |
7006 | ||
7007 | /* | |
7008 | * unmap huge page backed by shared pte. | |
7009 | * | |
7010 | * Hugetlb pte page is ref counted at the time of mapping. If pte is shared | |
7011 | * indicated by page_count > 1, unmap is achieved by clearing pud and | |
7012 | * decrementing the ref count. If count == 1, the pte page is not shared. | |
7013 | * | |
3a47c54f | 7014 | * Called with page table lock held. |
3212b535 SC |
7015 | * |
7016 | * returns: 1 successfully unmapped a shared pte page | |
7017 | * 0 the underlying pte page is not shared, or it is the last user | |
7018 | */ | |
34ae204f | 7019 | int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma, |
4ddb4d91 | 7020 | unsigned long addr, pte_t *ptep) |
3212b535 | 7021 | { |
4ddb4d91 MK |
7022 | pgd_t *pgd = pgd_offset(mm, addr); |
7023 | p4d_t *p4d = p4d_offset(pgd, addr); | |
7024 | pud_t *pud = pud_offset(p4d, addr); | |
3212b535 | 7025 | |
34ae204f | 7026 | i_mmap_assert_write_locked(vma->vm_file->f_mapping); |
40549ba8 | 7027 | hugetlb_vma_assert_locked(vma); |
3212b535 SC |
7028 | BUG_ON(page_count(virt_to_page(ptep)) == 0); |
7029 | if (page_count(virt_to_page(ptep)) == 1) | |
7030 | return 0; | |
7031 | ||
7032 | pud_clear(pud); | |
7033 | put_page(virt_to_page(ptep)); | |
dc6c9a35 | 7034 | mm_dec_nr_pmds(mm); |
3212b535 SC |
7035 | return 1; |
7036 | } | |
c1991e07 | 7037 | |
9e5fc74c | 7038 | #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ |
8d9bfb26 | 7039 | |
40549ba8 MK |
7040 | void hugetlb_vma_lock_read(struct vm_area_struct *vma) |
7041 | { | |
7042 | } | |
7043 | ||
7044 | void hugetlb_vma_unlock_read(struct vm_area_struct *vma) | |
7045 | { | |
7046 | } | |
7047 | ||
7048 | void hugetlb_vma_lock_write(struct vm_area_struct *vma) | |
7049 | { | |
7050 | } | |
7051 | ||
7052 | void hugetlb_vma_unlock_write(struct vm_area_struct *vma) | |
7053 | { | |
7054 | } | |
7055 | ||
7056 | int hugetlb_vma_trylock_write(struct vm_area_struct *vma) | |
7057 | { | |
7058 | return 1; | |
7059 | } | |
7060 | ||
7061 | void hugetlb_vma_assert_locked(struct vm_area_struct *vma) | |
7062 | { | |
7063 | } | |
7064 | ||
8d9bfb26 MK |
7065 | void hugetlb_vma_lock_release(struct kref *kref) |
7066 | { | |
7067 | } | |
7068 | ||
ecfbd733 MK |
7069 | static void __hugetlb_vma_unlock_write_free(struct vm_area_struct *vma) |
7070 | { | |
7071 | } | |
7072 | ||
8d9bfb26 MK |
7073 | static void hugetlb_vma_lock_free(struct vm_area_struct *vma) |
7074 | { | |
7075 | } | |
7076 | ||
7077 | static void hugetlb_vma_lock_alloc(struct vm_area_struct *vma) | |
7078 | { | |
7079 | } | |
7080 | ||
aec44e0f PX |
7081 | pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct *vma, |
7082 | unsigned long addr, pud_t *pud) | |
9e5fc74c SC |
7083 | { |
7084 | return NULL; | |
7085 | } | |
e81f2d22 | 7086 | |
34ae204f | 7087 | int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma, |
4ddb4d91 | 7088 | unsigned long addr, pte_t *ptep) |
e81f2d22 ZZ |
7089 | { |
7090 | return 0; | |
7091 | } | |
017b1660 MK |
7092 | |
7093 | void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma, | |
7094 | unsigned long *start, unsigned long *end) | |
7095 | { | |
7096 | } | |
c1991e07 PX |
7097 | |
7098 | bool want_pmd_share(struct vm_area_struct *vma, unsigned long addr) | |
7099 | { | |
7100 | return false; | |
7101 | } | |
3212b535 SC |
7102 | #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ |
7103 | ||
9e5fc74c | 7104 | #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB |
aec44e0f | 7105 | pte_t *huge_pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma, |
9e5fc74c SC |
7106 | unsigned long addr, unsigned long sz) |
7107 | { | |
7108 | pgd_t *pgd; | |
c2febafc | 7109 | p4d_t *p4d; |
9e5fc74c SC |
7110 | pud_t *pud; |
7111 | pte_t *pte = NULL; | |
7112 | ||
7113 | pgd = pgd_offset(mm, addr); | |
f4f0a3d8 KS |
7114 | p4d = p4d_alloc(mm, pgd, addr); |
7115 | if (!p4d) | |
7116 | return NULL; | |
c2febafc | 7117 | pud = pud_alloc(mm, p4d, addr); |
9e5fc74c SC |
7118 | if (pud) { |
7119 | if (sz == PUD_SIZE) { | |
7120 | pte = (pte_t *)pud; | |
7121 | } else { | |
7122 | BUG_ON(sz != PMD_SIZE); | |
c1991e07 | 7123 | if (want_pmd_share(vma, addr) && pud_none(*pud)) |
aec44e0f | 7124 | pte = huge_pmd_share(mm, vma, addr, pud); |
9e5fc74c SC |
7125 | else |
7126 | pte = (pte_t *)pmd_alloc(mm, pud, addr); | |
7127 | } | |
7128 | } | |
4e666314 | 7129 | BUG_ON(pte && pte_present(*pte) && !pte_huge(*pte)); |
9e5fc74c SC |
7130 | |
7131 | return pte; | |
7132 | } | |
7133 | ||
9b19df29 PA |
7134 | /* |
7135 | * huge_pte_offset() - Walk the page table to resolve the hugepage | |
7136 | * entry at address @addr | |
7137 | * | |
8ac0b81a LX |
7138 | * Return: Pointer to page table entry (PUD or PMD) for |
7139 | * address @addr, or NULL if a !p*d_present() entry is encountered and the | |
9b19df29 PA |
7140 | * size @sz doesn't match the hugepage size at this level of the page |
7141 | * table. | |
7142 | */ | |
7868a208 PA |
7143 | pte_t *huge_pte_offset(struct mm_struct *mm, |
7144 | unsigned long addr, unsigned long sz) | |
9e5fc74c SC |
7145 | { |
7146 | pgd_t *pgd; | |
c2febafc | 7147 | p4d_t *p4d; |
8ac0b81a LX |
7148 | pud_t *pud; |
7149 | pmd_t *pmd; | |
9e5fc74c SC |
7150 | |
7151 | pgd = pgd_offset(mm, addr); | |
c2febafc KS |
7152 | if (!pgd_present(*pgd)) |
7153 | return NULL; | |
7154 | p4d = p4d_offset(pgd, addr); | |
7155 | if (!p4d_present(*p4d)) | |
7156 | return NULL; | |
9b19df29 | 7157 | |
c2febafc | 7158 | pud = pud_offset(p4d, addr); |
8ac0b81a LX |
7159 | if (sz == PUD_SIZE) |
7160 | /* must be pud huge, non-present or none */ | |
c2febafc | 7161 | return (pte_t *)pud; |
8ac0b81a | 7162 | if (!pud_present(*pud)) |
9b19df29 | 7163 | return NULL; |
8ac0b81a | 7164 | /* must have a valid entry and size to go further */ |
9b19df29 | 7165 | |
8ac0b81a LX |
7166 | pmd = pmd_offset(pud, addr); |
7167 | /* must be pmd huge, non-present or none */ | |
7168 | return (pte_t *)pmd; | |
9e5fc74c SC |
7169 | } |
7170 | ||
e95a9851 MK |
7171 | /* |
7172 | * Return a mask that can be used to update an address to the last huge | |
7173 | * page in a page table page mapping size. Used to skip non-present | |
7174 | * page table entries when linearly scanning address ranges. Architectures | |
7175 | * with unique huge page to page table relationships can define their own | |
7176 | * version of this routine. | |
7177 | */ | |
7178 | unsigned long hugetlb_mask_last_page(struct hstate *h) | |
7179 | { | |
7180 | unsigned long hp_size = huge_page_size(h); | |
7181 | ||
7182 | if (hp_size == PUD_SIZE) | |
7183 | return P4D_SIZE - PUD_SIZE; | |
7184 | else if (hp_size == PMD_SIZE) | |
7185 | return PUD_SIZE - PMD_SIZE; | |
7186 | else | |
7187 | return 0UL; | |
7188 | } | |
7189 | ||
7190 | #else | |
7191 | ||
7192 | /* See description above. Architectures can provide their own version. */ | |
7193 | __weak unsigned long hugetlb_mask_last_page(struct hstate *h) | |
7194 | { | |
4ddb4d91 MK |
7195 | #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE |
7196 | if (huge_page_size(h) == PMD_SIZE) | |
7197 | return PUD_SIZE - PMD_SIZE; | |
7198 | #endif | |
e95a9851 MK |
7199 | return 0UL; |
7200 | } | |
7201 | ||
61f77eda NH |
7202 | #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */ |
7203 | ||
7204 | /* | |
7205 | * These functions are overwritable if your architecture needs its own | |
7206 | * behavior. | |
7207 | */ | |
7208 | struct page * __weak | |
7209 | follow_huge_addr(struct mm_struct *mm, unsigned long address, | |
7210 | int write) | |
7211 | { | |
7212 | return ERR_PTR(-EINVAL); | |
7213 | } | |
7214 | ||
4dc71451 AK |
7215 | struct page * __weak |
7216 | follow_huge_pd(struct vm_area_struct *vma, | |
7217 | unsigned long address, hugepd_t hpd, int flags, int pdshift) | |
7218 | { | |
7219 | WARN(1, "hugepd follow called with no support for hugepage directory format\n"); | |
7220 | return NULL; | |
7221 | } | |
7222 | ||
61f77eda | 7223 | struct page * __weak |
fac35ba7 | 7224 | follow_huge_pmd_pte(struct vm_area_struct *vma, unsigned long address, int flags) |
9e5fc74c | 7225 | { |
fac35ba7 BW |
7226 | struct hstate *h = hstate_vma(vma); |
7227 | struct mm_struct *mm = vma->vm_mm; | |
e66f17ff NH |
7228 | struct page *page = NULL; |
7229 | spinlock_t *ptl; | |
fac35ba7 | 7230 | pte_t *ptep, pte; |
3faa52c0 | 7231 | |
8909691b DH |
7232 | /* |
7233 | * FOLL_PIN is not supported for follow_page(). Ordinary GUP goes via | |
7234 | * follow_hugetlb_page(). | |
7235 | */ | |
7236 | if (WARN_ON_ONCE(flags & FOLL_PIN)) | |
3faa52c0 JH |
7237 | return NULL; |
7238 | ||
e66f17ff | 7239 | retry: |
fac35ba7 BW |
7240 | ptep = huge_pte_offset(mm, address, huge_page_size(h)); |
7241 | if (!ptep) | |
7242 | return NULL; | |
7243 | ||
7244 | ptl = huge_pte_lock(h, mm, ptep); | |
7245 | pte = huge_ptep_get(ptep); | |
c9d398fa | 7246 | if (pte_present(pte)) { |
fac35ba7 BW |
7247 | page = pte_page(pte) + |
7248 | ((address & ~huge_page_mask(h)) >> PAGE_SHIFT); | |
3faa52c0 JH |
7249 | /* |
7250 | * try_grab_page() should always succeed here, because: a) we | |
7251 | * hold the pmd (ptl) lock, and b) we've just checked that the | |
7252 | * huge pmd (head) page is present in the page tables. The ptl | |
7253 | * prevents the head page and tail pages from being rearranged | |
7254 | * in any way. So this page must be available at this point, | |
7255 | * unless the page refcount overflowed: | |
7256 | */ | |
7257 | if (WARN_ON_ONCE(!try_grab_page(page, flags))) { | |
7258 | page = NULL; | |
7259 | goto out; | |
7260 | } | |
e66f17ff | 7261 | } else { |
c9d398fa | 7262 | if (is_hugetlb_entry_migration(pte)) { |
e66f17ff | 7263 | spin_unlock(ptl); |
fac35ba7 | 7264 | __migration_entry_wait_huge(ptep, ptl); |
e66f17ff NH |
7265 | goto retry; |
7266 | } | |
7267 | /* | |
7268 | * hwpoisoned entry is treated as no_page_table in | |
7269 | * follow_page_mask(). | |
7270 | */ | |
7271 | } | |
7272 | out: | |
7273 | spin_unlock(ptl); | |
9e5fc74c SC |
7274 | return page; |
7275 | } | |
7276 | ||
61f77eda | 7277 | struct page * __weak |
9e5fc74c | 7278 | follow_huge_pud(struct mm_struct *mm, unsigned long address, |
e66f17ff | 7279 | pud_t *pud, int flags) |
9e5fc74c | 7280 | { |
3a194f3f NH |
7281 | struct page *page = NULL; |
7282 | spinlock_t *ptl; | |
7283 | pte_t pte; | |
7284 | ||
7285 | if (WARN_ON_ONCE(flags & FOLL_PIN)) | |
e66f17ff | 7286 | return NULL; |
9e5fc74c | 7287 | |
3a194f3f NH |
7288 | retry: |
7289 | ptl = huge_pte_lock(hstate_sizelog(PUD_SHIFT), mm, (pte_t *)pud); | |
7290 | if (!pud_huge(*pud)) | |
7291 | goto out; | |
7292 | pte = huge_ptep_get((pte_t *)pud); | |
7293 | if (pte_present(pte)) { | |
7294 | page = pud_page(*pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT); | |
7295 | if (WARN_ON_ONCE(!try_grab_page(page, flags))) { | |
7296 | page = NULL; | |
7297 | goto out; | |
7298 | } | |
7299 | } else { | |
7300 | if (is_hugetlb_entry_migration(pte)) { | |
7301 | spin_unlock(ptl); | |
7302 | __migration_entry_wait(mm, (pte_t *)pud, ptl); | |
7303 | goto retry; | |
7304 | } | |
7305 | /* | |
7306 | * hwpoisoned entry is treated as no_page_table in | |
7307 | * follow_page_mask(). | |
7308 | */ | |
7309 | } | |
7310 | out: | |
7311 | spin_unlock(ptl); | |
7312 | return page; | |
9e5fc74c SC |
7313 | } |
7314 | ||
faaa5b62 AK |
7315 | struct page * __weak |
7316 | follow_huge_pgd(struct mm_struct *mm, unsigned long address, pgd_t *pgd, int flags) | |
7317 | { | |
3faa52c0 | 7318 | if (flags & (FOLL_GET | FOLL_PIN)) |
faaa5b62 AK |
7319 | return NULL; |
7320 | ||
7321 | return pte_page(*(pte_t *)pgd) + ((address & ~PGDIR_MASK) >> PAGE_SHIFT); | |
7322 | } | |
7323 | ||
7ce82f4c | 7324 | int isolate_hugetlb(struct page *page, struct list_head *list) |
31caf665 | 7325 | { |
7ce82f4c | 7326 | int ret = 0; |
bcc54222 | 7327 | |
db71ef79 | 7328 | spin_lock_irq(&hugetlb_lock); |
8f251a3d MK |
7329 | if (!PageHeadHuge(page) || |
7330 | !HPageMigratable(page) || | |
0eb2df2b | 7331 | !get_page_unless_zero(page)) { |
7ce82f4c | 7332 | ret = -EBUSY; |
bcc54222 NH |
7333 | goto unlock; |
7334 | } | |
8f251a3d | 7335 | ClearHPageMigratable(page); |
31caf665 | 7336 | list_move_tail(&page->lru, list); |
bcc54222 | 7337 | unlock: |
db71ef79 | 7338 | spin_unlock_irq(&hugetlb_lock); |
bcc54222 | 7339 | return ret; |
31caf665 NH |
7340 | } |
7341 | ||
25182f05 NH |
7342 | int get_hwpoison_huge_page(struct page *page, bool *hugetlb) |
7343 | { | |
7344 | int ret = 0; | |
7345 | ||
7346 | *hugetlb = false; | |
7347 | spin_lock_irq(&hugetlb_lock); | |
7348 | if (PageHeadHuge(page)) { | |
7349 | *hugetlb = true; | |
b283d983 NH |
7350 | if (HPageFreed(page)) |
7351 | ret = 0; | |
7352 | else if (HPageMigratable(page)) | |
25182f05 | 7353 | ret = get_page_unless_zero(page); |
0ed950d1 NH |
7354 | else |
7355 | ret = -EBUSY; | |
25182f05 NH |
7356 | } |
7357 | spin_unlock_irq(&hugetlb_lock); | |
7358 | return ret; | |
7359 | } | |
7360 | ||
405ce051 NH |
7361 | int get_huge_page_for_hwpoison(unsigned long pfn, int flags) |
7362 | { | |
7363 | int ret; | |
7364 | ||
7365 | spin_lock_irq(&hugetlb_lock); | |
7366 | ret = __get_huge_page_for_hwpoison(pfn, flags); | |
7367 | spin_unlock_irq(&hugetlb_lock); | |
7368 | return ret; | |
7369 | } | |
7370 | ||
31caf665 NH |
7371 | void putback_active_hugepage(struct page *page) |
7372 | { | |
db71ef79 | 7373 | spin_lock_irq(&hugetlb_lock); |
8f251a3d | 7374 | SetHPageMigratable(page); |
31caf665 | 7375 | list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist); |
db71ef79 | 7376 | spin_unlock_irq(&hugetlb_lock); |
31caf665 NH |
7377 | put_page(page); |
7378 | } | |
ab5ac90a MH |
7379 | |
7380 | void move_hugetlb_state(struct page *oldpage, struct page *newpage, int reason) | |
7381 | { | |
7382 | struct hstate *h = page_hstate(oldpage); | |
7383 | ||
7384 | hugetlb_cgroup_migrate(oldpage, newpage); | |
7385 | set_page_owner_migrate_reason(newpage, reason); | |
7386 | ||
7387 | /* | |
7388 | * transfer temporary state of the new huge page. This is | |
7389 | * reverse to other transitions because the newpage is going to | |
7390 | * be final while the old one will be freed so it takes over | |
7391 | * the temporary status. | |
7392 | * | |
7393 | * Also note that we have to transfer the per-node surplus state | |
7394 | * here as well otherwise the global surplus count will not match | |
7395 | * the per-node's. | |
7396 | */ | |
9157c311 | 7397 | if (HPageTemporary(newpage)) { |
ab5ac90a MH |
7398 | int old_nid = page_to_nid(oldpage); |
7399 | int new_nid = page_to_nid(newpage); | |
7400 | ||
9157c311 MK |
7401 | SetHPageTemporary(oldpage); |
7402 | ClearHPageTemporary(newpage); | |
ab5ac90a | 7403 | |
5af1ab1d ML |
7404 | /* |
7405 | * There is no need to transfer the per-node surplus state | |
7406 | * when we do not cross the node. | |
7407 | */ | |
7408 | if (new_nid == old_nid) | |
7409 | return; | |
db71ef79 | 7410 | spin_lock_irq(&hugetlb_lock); |
ab5ac90a MH |
7411 | if (h->surplus_huge_pages_node[old_nid]) { |
7412 | h->surplus_huge_pages_node[old_nid]--; | |
7413 | h->surplus_huge_pages_node[new_nid]++; | |
7414 | } | |
db71ef79 | 7415 | spin_unlock_irq(&hugetlb_lock); |
ab5ac90a MH |
7416 | } |
7417 | } | |
cf11e85f | 7418 | |
6dfeaff9 PX |
7419 | /* |
7420 | * This function will unconditionally remove all the shared pmd pgtable entries | |
7421 | * within the specific vma for a hugetlbfs memory range. | |
7422 | */ | |
7423 | void hugetlb_unshare_all_pmds(struct vm_area_struct *vma) | |
7424 | { | |
7425 | struct hstate *h = hstate_vma(vma); | |
7426 | unsigned long sz = huge_page_size(h); | |
7427 | struct mm_struct *mm = vma->vm_mm; | |
7428 | struct mmu_notifier_range range; | |
7429 | unsigned long address, start, end; | |
7430 | spinlock_t *ptl; | |
7431 | pte_t *ptep; | |
7432 | ||
7433 | if (!(vma->vm_flags & VM_MAYSHARE)) | |
7434 | return; | |
7435 | ||
7436 | start = ALIGN(vma->vm_start, PUD_SIZE); | |
7437 | end = ALIGN_DOWN(vma->vm_end, PUD_SIZE); | |
7438 | ||
7439 | if (start >= end) | |
7440 | return; | |
7441 | ||
9c8bbfac | 7442 | flush_cache_range(vma, start, end); |
6dfeaff9 PX |
7443 | /* |
7444 | * No need to call adjust_range_if_pmd_sharing_possible(), because | |
7445 | * we have already done the PUD_SIZE alignment. | |
7446 | */ | |
7447 | mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, | |
7448 | start, end); | |
7449 | mmu_notifier_invalidate_range_start(&range); | |
40549ba8 | 7450 | hugetlb_vma_lock_write(vma); |
6dfeaff9 PX |
7451 | i_mmap_lock_write(vma->vm_file->f_mapping); |
7452 | for (address = start; address < end; address += PUD_SIZE) { | |
6dfeaff9 PX |
7453 | ptep = huge_pte_offset(mm, address, sz); |
7454 | if (!ptep) | |
7455 | continue; | |
7456 | ptl = huge_pte_lock(h, mm, ptep); | |
4ddb4d91 | 7457 | huge_pmd_unshare(mm, vma, address, ptep); |
6dfeaff9 PX |
7458 | spin_unlock(ptl); |
7459 | } | |
7460 | flush_hugetlb_tlb_range(vma, start, end); | |
7461 | i_mmap_unlock_write(vma->vm_file->f_mapping); | |
40549ba8 | 7462 | hugetlb_vma_unlock_write(vma); |
6dfeaff9 PX |
7463 | /* |
7464 | * No need to call mmu_notifier_invalidate_range(), see | |
ee65728e | 7465 | * Documentation/mm/mmu_notifier.rst. |
6dfeaff9 PX |
7466 | */ |
7467 | mmu_notifier_invalidate_range_end(&range); | |
7468 | } | |
7469 | ||
cf11e85f | 7470 | #ifdef CONFIG_CMA |
cf11e85f RG |
7471 | static bool cma_reserve_called __initdata; |
7472 | ||
7473 | static int __init cmdline_parse_hugetlb_cma(char *p) | |
7474 | { | |
38e719ab BW |
7475 | int nid, count = 0; |
7476 | unsigned long tmp; | |
7477 | char *s = p; | |
7478 | ||
7479 | while (*s) { | |
7480 | if (sscanf(s, "%lu%n", &tmp, &count) != 1) | |
7481 | break; | |
7482 | ||
7483 | if (s[count] == ':') { | |
f9317f77 | 7484 | if (tmp >= MAX_NUMNODES) |
38e719ab | 7485 | break; |
f9317f77 | 7486 | nid = array_index_nospec(tmp, MAX_NUMNODES); |
38e719ab BW |
7487 | |
7488 | s += count + 1; | |
7489 | tmp = memparse(s, &s); | |
7490 | hugetlb_cma_size_in_node[nid] = tmp; | |
7491 | hugetlb_cma_size += tmp; | |
7492 | ||
7493 | /* | |
7494 | * Skip the separator if have one, otherwise | |
7495 | * break the parsing. | |
7496 | */ | |
7497 | if (*s == ',') | |
7498 | s++; | |
7499 | else | |
7500 | break; | |
7501 | } else { | |
7502 | hugetlb_cma_size = memparse(p, &p); | |
7503 | break; | |
7504 | } | |
7505 | } | |
7506 | ||
cf11e85f RG |
7507 | return 0; |
7508 | } | |
7509 | ||
7510 | early_param("hugetlb_cma", cmdline_parse_hugetlb_cma); | |
7511 | ||
7512 | void __init hugetlb_cma_reserve(int order) | |
7513 | { | |
7514 | unsigned long size, reserved, per_node; | |
38e719ab | 7515 | bool node_specific_cma_alloc = false; |
cf11e85f RG |
7516 | int nid; |
7517 | ||
7518 | cma_reserve_called = true; | |
7519 | ||
38e719ab BW |
7520 | if (!hugetlb_cma_size) |
7521 | return; | |
7522 | ||
7523 | for (nid = 0; nid < MAX_NUMNODES; nid++) { | |
7524 | if (hugetlb_cma_size_in_node[nid] == 0) | |
7525 | continue; | |
7526 | ||
30a51400 | 7527 | if (!node_online(nid)) { |
38e719ab BW |
7528 | pr_warn("hugetlb_cma: invalid node %d specified\n", nid); |
7529 | hugetlb_cma_size -= hugetlb_cma_size_in_node[nid]; | |
7530 | hugetlb_cma_size_in_node[nid] = 0; | |
7531 | continue; | |
7532 | } | |
7533 | ||
7534 | if (hugetlb_cma_size_in_node[nid] < (PAGE_SIZE << order)) { | |
7535 | pr_warn("hugetlb_cma: cma area of node %d should be at least %lu MiB\n", | |
7536 | nid, (PAGE_SIZE << order) / SZ_1M); | |
7537 | hugetlb_cma_size -= hugetlb_cma_size_in_node[nid]; | |
7538 | hugetlb_cma_size_in_node[nid] = 0; | |
7539 | } else { | |
7540 | node_specific_cma_alloc = true; | |
7541 | } | |
7542 | } | |
7543 | ||
7544 | /* Validate the CMA size again in case some invalid nodes specified. */ | |
cf11e85f RG |
7545 | if (!hugetlb_cma_size) |
7546 | return; | |
7547 | ||
7548 | if (hugetlb_cma_size < (PAGE_SIZE << order)) { | |
7549 | pr_warn("hugetlb_cma: cma area should be at least %lu MiB\n", | |
7550 | (PAGE_SIZE << order) / SZ_1M); | |
a01f4390 | 7551 | hugetlb_cma_size = 0; |
cf11e85f RG |
7552 | return; |
7553 | } | |
7554 | ||
38e719ab BW |
7555 | if (!node_specific_cma_alloc) { |
7556 | /* | |
7557 | * If 3 GB area is requested on a machine with 4 numa nodes, | |
7558 | * let's allocate 1 GB on first three nodes and ignore the last one. | |
7559 | */ | |
7560 | per_node = DIV_ROUND_UP(hugetlb_cma_size, nr_online_nodes); | |
7561 | pr_info("hugetlb_cma: reserve %lu MiB, up to %lu MiB per node\n", | |
7562 | hugetlb_cma_size / SZ_1M, per_node / SZ_1M); | |
7563 | } | |
cf11e85f RG |
7564 | |
7565 | reserved = 0; | |
30a51400 | 7566 | for_each_online_node(nid) { |
cf11e85f | 7567 | int res; |
2281f797 | 7568 | char name[CMA_MAX_NAME]; |
cf11e85f | 7569 | |
38e719ab BW |
7570 | if (node_specific_cma_alloc) { |
7571 | if (hugetlb_cma_size_in_node[nid] == 0) | |
7572 | continue; | |
7573 | ||
7574 | size = hugetlb_cma_size_in_node[nid]; | |
7575 | } else { | |
7576 | size = min(per_node, hugetlb_cma_size - reserved); | |
7577 | } | |
7578 | ||
cf11e85f RG |
7579 | size = round_up(size, PAGE_SIZE << order); |
7580 | ||
2281f797 | 7581 | snprintf(name, sizeof(name), "hugetlb%d", nid); |
a01f4390 MK |
7582 | /* |
7583 | * Note that 'order per bit' is based on smallest size that | |
7584 | * may be returned to CMA allocator in the case of | |
7585 | * huge page demotion. | |
7586 | */ | |
7587 | res = cma_declare_contiguous_nid(0, size, 0, | |
7588 | PAGE_SIZE << HUGETLB_PAGE_ORDER, | |
29d0f41d | 7589 | 0, false, name, |
cf11e85f RG |
7590 | &hugetlb_cma[nid], nid); |
7591 | if (res) { | |
7592 | pr_warn("hugetlb_cma: reservation failed: err %d, node %d", | |
7593 | res, nid); | |
7594 | continue; | |
7595 | } | |
7596 | ||
7597 | reserved += size; | |
7598 | pr_info("hugetlb_cma: reserved %lu MiB on node %d\n", | |
7599 | size / SZ_1M, nid); | |
7600 | ||
7601 | if (reserved >= hugetlb_cma_size) | |
7602 | break; | |
7603 | } | |
a01f4390 MK |
7604 | |
7605 | if (!reserved) | |
7606 | /* | |
7607 | * hugetlb_cma_size is used to determine if allocations from | |
7608 | * cma are possible. Set to zero if no cma regions are set up. | |
7609 | */ | |
7610 | hugetlb_cma_size = 0; | |
cf11e85f RG |
7611 | } |
7612 | ||
263b8998 | 7613 | static void __init hugetlb_cma_check(void) |
cf11e85f RG |
7614 | { |
7615 | if (!hugetlb_cma_size || cma_reserve_called) | |
7616 | return; | |
7617 | ||
7618 | pr_warn("hugetlb_cma: the option isn't supported by current arch\n"); | |
7619 | } | |
7620 | ||
7621 | #endif /* CONFIG_CMA */ |