]> git.proxmox.com Git - mirror_ubuntu-jammy-kernel.git/blame - mm/hugetlb.c
mem-hotplug: fix node spanned pages when we have a node with only ZONE_MOVABLE
[mirror_ubuntu-jammy-kernel.git] / mm / hugetlb.c
CommitLineData
1da177e4
LT
1/*
2 * Generic hugetlb support.
6d49e352 3 * (C) Nadia Yvette Chambers, April 2004
1da177e4 4 */
1da177e4
LT
5#include <linux/list.h>
6#include <linux/init.h>
1da177e4 7#include <linux/mm.h>
e1759c21 8#include <linux/seq_file.h>
1da177e4
LT
9#include <linux/sysctl.h>
10#include <linux/highmem.h>
cddb8a5c 11#include <linux/mmu_notifier.h>
1da177e4 12#include <linux/nodemask.h>
63551ae0 13#include <linux/pagemap.h>
5da7ca86 14#include <linux/mempolicy.h>
3b32123d 15#include <linux/compiler.h>
aea47ff3 16#include <linux/cpuset.h>
3935baa9 17#include <linux/mutex.h>
97ad1087 18#include <linux/memblock.h>
a3437870 19#include <linux/sysfs.h>
5a0e3ad6 20#include <linux/slab.h>
63489f8e 21#include <linux/mmdebug.h>
174cd4b1 22#include <linux/sched/signal.h>
0fe6e20b 23#include <linux/rmap.h>
c6247f72 24#include <linux/string_helpers.h>
fd6a03ed
NH
25#include <linux/swap.h>
26#include <linux/swapops.h>
8382d914 27#include <linux/jhash.h>
98fa15f3 28#include <linux/numa.h>
d6606683 29
63551ae0
DG
30#include <asm/page.h>
31#include <asm/pgtable.h>
24669e58 32#include <asm/tlb.h>
63551ae0 33
24669e58 34#include <linux/io.h>
63551ae0 35#include <linux/hugetlb.h>
9dd540e2 36#include <linux/hugetlb_cgroup.h>
9a305230 37#include <linux/node.h>
1a1aad8a 38#include <linux/userfaultfd_k.h>
ab5ac90a 39#include <linux/page_owner.h>
7835e98b 40#include "internal.h"
1da177e4 41
c3f38a38 42int hugetlb_max_hstate __read_mostly;
e5ff2159
AK
43unsigned int default_hstate_idx;
44struct hstate hstates[HUGE_MAX_HSTATE];
641844f5
NH
45/*
46 * Minimum page order among possible hugepage sizes, set to a proper value
47 * at boot time.
48 */
49static unsigned int minimum_order __read_mostly = UINT_MAX;
e5ff2159 50
53ba51d2
JT
51__initdata LIST_HEAD(huge_boot_pages);
52
e5ff2159
AK
53/* for command line parsing */
54static struct hstate * __initdata parsed_hstate;
55static unsigned long __initdata default_hstate_max_huge_pages;
e11bfbfc 56static unsigned long __initdata default_hstate_size;
9fee021d 57static bool __initdata parsed_valid_hugepagesz = true;
e5ff2159 58
3935baa9 59/*
31caf665
NH
60 * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages,
61 * free_huge_pages, and surplus_huge_pages.
3935baa9 62 */
c3f38a38 63DEFINE_SPINLOCK(hugetlb_lock);
0bd0f9fb 64
8382d914
DB
65/*
66 * Serializes faults on the same logical page. This is used to
67 * prevent spurious OOMs when the hugepage pool is fully utilized.
68 */
69static int num_fault_mutexes;
c672c7f2 70struct mutex *hugetlb_fault_mutex_table ____cacheline_aligned_in_smp;
8382d914 71
7ca02d0a
MK
72/* Forward declaration */
73static int hugetlb_acct_memory(struct hstate *h, long delta);
74
90481622
DG
75static inline void unlock_or_release_subpool(struct hugepage_subpool *spool)
76{
77 bool free = (spool->count == 0) && (spool->used_hpages == 0);
78
79 spin_unlock(&spool->lock);
80
81 /* If no pages are used, and no other handles to the subpool
7ca02d0a
MK
82 * remain, give up any reservations mased on minimum size and
83 * free the subpool */
84 if (free) {
85 if (spool->min_hpages != -1)
86 hugetlb_acct_memory(spool->hstate,
87 -spool->min_hpages);
90481622 88 kfree(spool);
7ca02d0a 89 }
90481622
DG
90}
91
7ca02d0a
MK
92struct hugepage_subpool *hugepage_new_subpool(struct hstate *h, long max_hpages,
93 long min_hpages)
90481622
DG
94{
95 struct hugepage_subpool *spool;
96
c6a91820 97 spool = kzalloc(sizeof(*spool), GFP_KERNEL);
90481622
DG
98 if (!spool)
99 return NULL;
100
101 spin_lock_init(&spool->lock);
102 spool->count = 1;
7ca02d0a
MK
103 spool->max_hpages = max_hpages;
104 spool->hstate = h;
105 spool->min_hpages = min_hpages;
106
107 if (min_hpages != -1 && hugetlb_acct_memory(h, min_hpages)) {
108 kfree(spool);
109 return NULL;
110 }
111 spool->rsv_hpages = min_hpages;
90481622
DG
112
113 return spool;
114}
115
116void hugepage_put_subpool(struct hugepage_subpool *spool)
117{
118 spin_lock(&spool->lock);
119 BUG_ON(!spool->count);
120 spool->count--;
121 unlock_or_release_subpool(spool);
122}
123
1c5ecae3
MK
124/*
125 * Subpool accounting for allocating and reserving pages.
126 * Return -ENOMEM if there are not enough resources to satisfy the
127 * the request. Otherwise, return the number of pages by which the
128 * global pools must be adjusted (upward). The returned value may
129 * only be different than the passed value (delta) in the case where
130 * a subpool minimum size must be manitained.
131 */
132static long hugepage_subpool_get_pages(struct hugepage_subpool *spool,
90481622
DG
133 long delta)
134{
1c5ecae3 135 long ret = delta;
90481622
DG
136
137 if (!spool)
1c5ecae3 138 return ret;
90481622
DG
139
140 spin_lock(&spool->lock);
1c5ecae3
MK
141
142 if (spool->max_hpages != -1) { /* maximum size accounting */
143 if ((spool->used_hpages + delta) <= spool->max_hpages)
144 spool->used_hpages += delta;
145 else {
146 ret = -ENOMEM;
147 goto unlock_ret;
148 }
90481622 149 }
90481622 150
09a95e29
MK
151 /* minimum size accounting */
152 if (spool->min_hpages != -1 && spool->rsv_hpages) {
1c5ecae3
MK
153 if (delta > spool->rsv_hpages) {
154 /*
155 * Asking for more reserves than those already taken on
156 * behalf of subpool. Return difference.
157 */
158 ret = delta - spool->rsv_hpages;
159 spool->rsv_hpages = 0;
160 } else {
161 ret = 0; /* reserves already accounted for */
162 spool->rsv_hpages -= delta;
163 }
164 }
165
166unlock_ret:
167 spin_unlock(&spool->lock);
90481622
DG
168 return ret;
169}
170
1c5ecae3
MK
171/*
172 * Subpool accounting for freeing and unreserving pages.
173 * Return the number of global page reservations that must be dropped.
174 * The return value may only be different than the passed value (delta)
175 * in the case where a subpool minimum size must be maintained.
176 */
177static long hugepage_subpool_put_pages(struct hugepage_subpool *spool,
90481622
DG
178 long delta)
179{
1c5ecae3
MK
180 long ret = delta;
181
90481622 182 if (!spool)
1c5ecae3 183 return delta;
90481622
DG
184
185 spin_lock(&spool->lock);
1c5ecae3
MK
186
187 if (spool->max_hpages != -1) /* maximum size accounting */
188 spool->used_hpages -= delta;
189
09a95e29
MK
190 /* minimum size accounting */
191 if (spool->min_hpages != -1 && spool->used_hpages < spool->min_hpages) {
1c5ecae3
MK
192 if (spool->rsv_hpages + delta <= spool->min_hpages)
193 ret = 0;
194 else
195 ret = spool->rsv_hpages + delta - spool->min_hpages;
196
197 spool->rsv_hpages += delta;
198 if (spool->rsv_hpages > spool->min_hpages)
199 spool->rsv_hpages = spool->min_hpages;
200 }
201
202 /*
203 * If hugetlbfs_put_super couldn't free spool due to an outstanding
204 * quota reference, free it now.
205 */
90481622 206 unlock_or_release_subpool(spool);
1c5ecae3
MK
207
208 return ret;
90481622
DG
209}
210
211static inline struct hugepage_subpool *subpool_inode(struct inode *inode)
212{
213 return HUGETLBFS_SB(inode->i_sb)->spool;
214}
215
216static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma)
217{
496ad9aa 218 return subpool_inode(file_inode(vma->vm_file));
90481622
DG
219}
220
96822904
AW
221/*
222 * Region tracking -- allows tracking of reservations and instantiated pages
223 * across the pages in a mapping.
84afd99b 224 *
1dd308a7
MK
225 * The region data structures are embedded into a resv_map and protected
226 * by a resv_map's lock. The set of regions within the resv_map represent
227 * reservations for huge pages, or huge pages that have already been
228 * instantiated within the map. The from and to elements are huge page
229 * indicies into the associated mapping. from indicates the starting index
230 * of the region. to represents the first index past the end of the region.
231 *
232 * For example, a file region structure with from == 0 and to == 4 represents
233 * four huge pages in a mapping. It is important to note that the to element
234 * represents the first element past the end of the region. This is used in
235 * arithmetic as 4(to) - 0(from) = 4 huge pages in the region.
236 *
237 * Interval notation of the form [from, to) will be used to indicate that
238 * the endpoint from is inclusive and to is exclusive.
96822904
AW
239 */
240struct file_region {
241 struct list_head link;
242 long from;
243 long to;
244};
245
1dd308a7
MK
246/*
247 * Add the huge page range represented by [f, t) to the reserve
5e911373
MK
248 * map. In the normal case, existing regions will be expanded
249 * to accommodate the specified range. Sufficient regions should
250 * exist for expansion due to the previous call to region_chg
251 * with the same range. However, it is possible that region_del
252 * could have been called after region_chg and modifed the map
253 * in such a way that no region exists to be expanded. In this
254 * case, pull a region descriptor from the cache associated with
255 * the map and use that for the new range.
cf3ad20b
MK
256 *
257 * Return the number of new huge pages added to the map. This
258 * number is greater than or equal to zero.
1dd308a7 259 */
1406ec9b 260static long region_add(struct resv_map *resv, long f, long t)
96822904 261{
1406ec9b 262 struct list_head *head = &resv->regions;
96822904 263 struct file_region *rg, *nrg, *trg;
cf3ad20b 264 long add = 0;
96822904 265
7b24d861 266 spin_lock(&resv->lock);
96822904
AW
267 /* Locate the region we are either in or before. */
268 list_for_each_entry(rg, head, link)
269 if (f <= rg->to)
270 break;
271
5e911373
MK
272 /*
273 * If no region exists which can be expanded to include the
274 * specified range, the list must have been modified by an
275 * interleving call to region_del(). Pull a region descriptor
276 * from the cache and use it for this range.
277 */
278 if (&rg->link == head || t < rg->from) {
279 VM_BUG_ON(resv->region_cache_count <= 0);
280
281 resv->region_cache_count--;
282 nrg = list_first_entry(&resv->region_cache, struct file_region,
283 link);
284 list_del(&nrg->link);
285
286 nrg->from = f;
287 nrg->to = t;
288 list_add(&nrg->link, rg->link.prev);
289
290 add += t - f;
291 goto out_locked;
292 }
293
96822904
AW
294 /* Round our left edge to the current segment if it encloses us. */
295 if (f > rg->from)
296 f = rg->from;
297
298 /* Check for and consume any regions we now overlap with. */
299 nrg = rg;
300 list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
301 if (&rg->link == head)
302 break;
303 if (rg->from > t)
304 break;
305
306 /* If this area reaches higher then extend our area to
307 * include it completely. If this is not the first area
308 * which we intend to reuse, free it. */
309 if (rg->to > t)
310 t = rg->to;
311 if (rg != nrg) {
cf3ad20b
MK
312 /* Decrement return value by the deleted range.
313 * Another range will span this area so that by
314 * end of routine add will be >= zero
315 */
316 add -= (rg->to - rg->from);
96822904
AW
317 list_del(&rg->link);
318 kfree(rg);
319 }
320 }
cf3ad20b
MK
321
322 add += (nrg->from - f); /* Added to beginning of region */
96822904 323 nrg->from = f;
cf3ad20b 324 add += t - nrg->to; /* Added to end of region */
96822904 325 nrg->to = t;
cf3ad20b 326
5e911373
MK
327out_locked:
328 resv->adds_in_progress--;
7b24d861 329 spin_unlock(&resv->lock);
cf3ad20b
MK
330 VM_BUG_ON(add < 0);
331 return add;
96822904
AW
332}
333
1dd308a7
MK
334/*
335 * Examine the existing reserve map and determine how many
336 * huge pages in the specified range [f, t) are NOT currently
337 * represented. This routine is called before a subsequent
338 * call to region_add that will actually modify the reserve
339 * map to add the specified range [f, t). region_chg does
340 * not change the number of huge pages represented by the
341 * map. However, if the existing regions in the map can not
342 * be expanded to represent the new range, a new file_region
343 * structure is added to the map as a placeholder. This is
344 * so that the subsequent region_add call will have all the
345 * regions it needs and will not fail.
346 *
5e911373
MK
347 * Upon entry, region_chg will also examine the cache of region descriptors
348 * associated with the map. If there are not enough descriptors cached, one
349 * will be allocated for the in progress add operation.
350 *
351 * Returns the number of huge pages that need to be added to the existing
352 * reservation map for the range [f, t). This number is greater or equal to
353 * zero. -ENOMEM is returned if a new file_region structure or cache entry
354 * is needed and can not be allocated.
1dd308a7 355 */
1406ec9b 356static long region_chg(struct resv_map *resv, long f, long t)
96822904 357{
1406ec9b 358 struct list_head *head = &resv->regions;
7b24d861 359 struct file_region *rg, *nrg = NULL;
96822904
AW
360 long chg = 0;
361
7b24d861
DB
362retry:
363 spin_lock(&resv->lock);
5e911373
MK
364retry_locked:
365 resv->adds_in_progress++;
366
367 /*
368 * Check for sufficient descriptors in the cache to accommodate
369 * the number of in progress add operations.
370 */
371 if (resv->adds_in_progress > resv->region_cache_count) {
372 struct file_region *trg;
373
374 VM_BUG_ON(resv->adds_in_progress - resv->region_cache_count > 1);
375 /* Must drop lock to allocate a new descriptor. */
376 resv->adds_in_progress--;
377 spin_unlock(&resv->lock);
378
379 trg = kmalloc(sizeof(*trg), GFP_KERNEL);
dbe409e4
MK
380 if (!trg) {
381 kfree(nrg);
5e911373 382 return -ENOMEM;
dbe409e4 383 }
5e911373
MK
384
385 spin_lock(&resv->lock);
386 list_add(&trg->link, &resv->region_cache);
387 resv->region_cache_count++;
388 goto retry_locked;
389 }
390
96822904
AW
391 /* Locate the region we are before or in. */
392 list_for_each_entry(rg, head, link)
393 if (f <= rg->to)
394 break;
395
396 /* If we are below the current region then a new region is required.
397 * Subtle, allocate a new region at the position but make it zero
398 * size such that we can guarantee to record the reservation. */
399 if (&rg->link == head || t < rg->from) {
7b24d861 400 if (!nrg) {
5e911373 401 resv->adds_in_progress--;
7b24d861
DB
402 spin_unlock(&resv->lock);
403 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
404 if (!nrg)
405 return -ENOMEM;
406
407 nrg->from = f;
408 nrg->to = f;
409 INIT_LIST_HEAD(&nrg->link);
410 goto retry;
411 }
96822904 412
7b24d861
DB
413 list_add(&nrg->link, rg->link.prev);
414 chg = t - f;
415 goto out_nrg;
96822904
AW
416 }
417
418 /* Round our left edge to the current segment if it encloses us. */
419 if (f > rg->from)
420 f = rg->from;
421 chg = t - f;
422
423 /* Check for and consume any regions we now overlap with. */
424 list_for_each_entry(rg, rg->link.prev, link) {
425 if (&rg->link == head)
426 break;
427 if (rg->from > t)
7b24d861 428 goto out;
96822904 429
25985edc 430 /* We overlap with this area, if it extends further than
96822904
AW
431 * us then we must extend ourselves. Account for its
432 * existing reservation. */
433 if (rg->to > t) {
434 chg += rg->to - t;
435 t = rg->to;
436 }
437 chg -= rg->to - rg->from;
438 }
7b24d861
DB
439
440out:
441 spin_unlock(&resv->lock);
442 /* We already know we raced and no longer need the new region */
443 kfree(nrg);
444 return chg;
445out_nrg:
446 spin_unlock(&resv->lock);
96822904
AW
447 return chg;
448}
449
5e911373
MK
450/*
451 * Abort the in progress add operation. The adds_in_progress field
452 * of the resv_map keeps track of the operations in progress between
453 * calls to region_chg and region_add. Operations are sometimes
454 * aborted after the call to region_chg. In such cases, region_abort
455 * is called to decrement the adds_in_progress counter.
456 *
457 * NOTE: The range arguments [f, t) are not needed or used in this
458 * routine. They are kept to make reading the calling code easier as
459 * arguments will match the associated region_chg call.
460 */
461static void region_abort(struct resv_map *resv, long f, long t)
462{
463 spin_lock(&resv->lock);
464 VM_BUG_ON(!resv->region_cache_count);
465 resv->adds_in_progress--;
466 spin_unlock(&resv->lock);
467}
468
1dd308a7 469/*
feba16e2
MK
470 * Delete the specified range [f, t) from the reserve map. If the
471 * t parameter is LONG_MAX, this indicates that ALL regions after f
472 * should be deleted. Locate the regions which intersect [f, t)
473 * and either trim, delete or split the existing regions.
474 *
475 * Returns the number of huge pages deleted from the reserve map.
476 * In the normal case, the return value is zero or more. In the
477 * case where a region must be split, a new region descriptor must
478 * be allocated. If the allocation fails, -ENOMEM will be returned.
479 * NOTE: If the parameter t == LONG_MAX, then we will never split
480 * a region and possibly return -ENOMEM. Callers specifying
481 * t == LONG_MAX do not need to check for -ENOMEM error.
1dd308a7 482 */
feba16e2 483static long region_del(struct resv_map *resv, long f, long t)
96822904 484{
1406ec9b 485 struct list_head *head = &resv->regions;
96822904 486 struct file_region *rg, *trg;
feba16e2
MK
487 struct file_region *nrg = NULL;
488 long del = 0;
96822904 489
feba16e2 490retry:
7b24d861 491 spin_lock(&resv->lock);
feba16e2 492 list_for_each_entry_safe(rg, trg, head, link) {
dbe409e4
MK
493 /*
494 * Skip regions before the range to be deleted. file_region
495 * ranges are normally of the form [from, to). However, there
496 * may be a "placeholder" entry in the map which is of the form
497 * (from, to) with from == to. Check for placeholder entries
498 * at the beginning of the range to be deleted.
499 */
500 if (rg->to <= f && (rg->to != rg->from || rg->to != f))
feba16e2 501 continue;
dbe409e4 502
feba16e2 503 if (rg->from >= t)
96822904 504 break;
96822904 505
feba16e2
MK
506 if (f > rg->from && t < rg->to) { /* Must split region */
507 /*
508 * Check for an entry in the cache before dropping
509 * lock and attempting allocation.
510 */
511 if (!nrg &&
512 resv->region_cache_count > resv->adds_in_progress) {
513 nrg = list_first_entry(&resv->region_cache,
514 struct file_region,
515 link);
516 list_del(&nrg->link);
517 resv->region_cache_count--;
518 }
96822904 519
feba16e2
MK
520 if (!nrg) {
521 spin_unlock(&resv->lock);
522 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
523 if (!nrg)
524 return -ENOMEM;
525 goto retry;
526 }
527
528 del += t - f;
529
530 /* New entry for end of split region */
531 nrg->from = t;
532 nrg->to = rg->to;
533 INIT_LIST_HEAD(&nrg->link);
534
535 /* Original entry is trimmed */
536 rg->to = f;
537
538 list_add(&nrg->link, &rg->link);
539 nrg = NULL;
96822904 540 break;
feba16e2
MK
541 }
542
543 if (f <= rg->from && t >= rg->to) { /* Remove entire region */
544 del += rg->to - rg->from;
545 list_del(&rg->link);
546 kfree(rg);
547 continue;
548 }
549
550 if (f <= rg->from) { /* Trim beginning of region */
551 del += t - rg->from;
552 rg->from = t;
553 } else { /* Trim end of region */
554 del += rg->to - f;
555 rg->to = f;
556 }
96822904 557 }
7b24d861 558
7b24d861 559 spin_unlock(&resv->lock);
feba16e2
MK
560 kfree(nrg);
561 return del;
96822904
AW
562}
563
b5cec28d
MK
564/*
565 * A rare out of memory error was encountered which prevented removal of
566 * the reserve map region for a page. The huge page itself was free'ed
567 * and removed from the page cache. This routine will adjust the subpool
568 * usage count, and the global reserve count if needed. By incrementing
569 * these counts, the reserve map entry which could not be deleted will
570 * appear as a "reserved" entry instead of simply dangling with incorrect
571 * counts.
572 */
72e2936c 573void hugetlb_fix_reserve_counts(struct inode *inode)
b5cec28d
MK
574{
575 struct hugepage_subpool *spool = subpool_inode(inode);
576 long rsv_adjust;
577
578 rsv_adjust = hugepage_subpool_get_pages(spool, 1);
72e2936c 579 if (rsv_adjust) {
b5cec28d
MK
580 struct hstate *h = hstate_inode(inode);
581
582 hugetlb_acct_memory(h, 1);
583 }
584}
585
1dd308a7
MK
586/*
587 * Count and return the number of huge pages in the reserve map
588 * that intersect with the range [f, t).
589 */
1406ec9b 590static long region_count(struct resv_map *resv, long f, long t)
84afd99b 591{
1406ec9b 592 struct list_head *head = &resv->regions;
84afd99b
AW
593 struct file_region *rg;
594 long chg = 0;
595
7b24d861 596 spin_lock(&resv->lock);
84afd99b
AW
597 /* Locate each segment we overlap with, and count that overlap. */
598 list_for_each_entry(rg, head, link) {
f2135a4a
WSH
599 long seg_from;
600 long seg_to;
84afd99b
AW
601
602 if (rg->to <= f)
603 continue;
604 if (rg->from >= t)
605 break;
606
607 seg_from = max(rg->from, f);
608 seg_to = min(rg->to, t);
609
610 chg += seg_to - seg_from;
611 }
7b24d861 612 spin_unlock(&resv->lock);
84afd99b
AW
613
614 return chg;
615}
616
e7c4b0bf
AW
617/*
618 * Convert the address within this vma to the page offset within
619 * the mapping, in pagecache page units; huge pages here.
620 */
a5516438
AK
621static pgoff_t vma_hugecache_offset(struct hstate *h,
622 struct vm_area_struct *vma, unsigned long address)
e7c4b0bf 623{
a5516438
AK
624 return ((address - vma->vm_start) >> huge_page_shift(h)) +
625 (vma->vm_pgoff >> huge_page_order(h));
e7c4b0bf
AW
626}
627
0fe6e20b
NH
628pgoff_t linear_hugepage_index(struct vm_area_struct *vma,
629 unsigned long address)
630{
631 return vma_hugecache_offset(hstate_vma(vma), vma, address);
632}
dee41079 633EXPORT_SYMBOL_GPL(linear_hugepage_index);
0fe6e20b 634
08fba699
MG
635/*
636 * Return the size of the pages allocated when backing a VMA. In the majority
637 * cases this will be same size as used by the page table entries.
638 */
639unsigned long vma_kernel_pagesize(struct vm_area_struct *vma)
640{
05ea8860
DW
641 if (vma->vm_ops && vma->vm_ops->pagesize)
642 return vma->vm_ops->pagesize(vma);
643 return PAGE_SIZE;
08fba699 644}
f340ca0f 645EXPORT_SYMBOL_GPL(vma_kernel_pagesize);
08fba699 646
3340289d
MG
647/*
648 * Return the page size being used by the MMU to back a VMA. In the majority
649 * of cases, the page size used by the kernel matches the MMU size. On
09135cc5
DW
650 * architectures where it differs, an architecture-specific 'strong'
651 * version of this symbol is required.
3340289d 652 */
09135cc5 653__weak unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
3340289d
MG
654{
655 return vma_kernel_pagesize(vma);
656}
3340289d 657
84afd99b
AW
658/*
659 * Flags for MAP_PRIVATE reservations. These are stored in the bottom
660 * bits of the reservation map pointer, which are always clear due to
661 * alignment.
662 */
663#define HPAGE_RESV_OWNER (1UL << 0)
664#define HPAGE_RESV_UNMAPPED (1UL << 1)
04f2cbe3 665#define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
84afd99b 666
a1e78772
MG
667/*
668 * These helpers are used to track how many pages are reserved for
669 * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
670 * is guaranteed to have their future faults succeed.
671 *
672 * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
673 * the reserve counters are updated with the hugetlb_lock held. It is safe
674 * to reset the VMA at fork() time as it is not in use yet and there is no
675 * chance of the global counters getting corrupted as a result of the values.
84afd99b
AW
676 *
677 * The private mapping reservation is represented in a subtly different
678 * manner to a shared mapping. A shared mapping has a region map associated
679 * with the underlying file, this region map represents the backing file
680 * pages which have ever had a reservation assigned which this persists even
681 * after the page is instantiated. A private mapping has a region map
682 * associated with the original mmap which is attached to all VMAs which
683 * reference it, this region map represents those offsets which have consumed
684 * reservation ie. where pages have been instantiated.
a1e78772 685 */
e7c4b0bf
AW
686static unsigned long get_vma_private_data(struct vm_area_struct *vma)
687{
688 return (unsigned long)vma->vm_private_data;
689}
690
691static void set_vma_private_data(struct vm_area_struct *vma,
692 unsigned long value)
693{
694 vma->vm_private_data = (void *)value;
695}
696
9119a41e 697struct resv_map *resv_map_alloc(void)
84afd99b
AW
698{
699 struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
5e911373
MK
700 struct file_region *rg = kmalloc(sizeof(*rg), GFP_KERNEL);
701
702 if (!resv_map || !rg) {
703 kfree(resv_map);
704 kfree(rg);
84afd99b 705 return NULL;
5e911373 706 }
84afd99b
AW
707
708 kref_init(&resv_map->refs);
7b24d861 709 spin_lock_init(&resv_map->lock);
84afd99b
AW
710 INIT_LIST_HEAD(&resv_map->regions);
711
5e911373
MK
712 resv_map->adds_in_progress = 0;
713
714 INIT_LIST_HEAD(&resv_map->region_cache);
715 list_add(&rg->link, &resv_map->region_cache);
716 resv_map->region_cache_count = 1;
717
84afd99b
AW
718 return resv_map;
719}
720
9119a41e 721void resv_map_release(struct kref *ref)
84afd99b
AW
722{
723 struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
5e911373
MK
724 struct list_head *head = &resv_map->region_cache;
725 struct file_region *rg, *trg;
84afd99b
AW
726
727 /* Clear out any active regions before we release the map. */
feba16e2 728 region_del(resv_map, 0, LONG_MAX);
5e911373
MK
729
730 /* ... and any entries left in the cache */
731 list_for_each_entry_safe(rg, trg, head, link) {
732 list_del(&rg->link);
733 kfree(rg);
734 }
735
736 VM_BUG_ON(resv_map->adds_in_progress);
737
84afd99b
AW
738 kfree(resv_map);
739}
740
4e35f483
JK
741static inline struct resv_map *inode_resv_map(struct inode *inode)
742{
743 return inode->i_mapping->private_data;
744}
745
84afd99b 746static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
a1e78772 747{
81d1b09c 748 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
4e35f483
JK
749 if (vma->vm_flags & VM_MAYSHARE) {
750 struct address_space *mapping = vma->vm_file->f_mapping;
751 struct inode *inode = mapping->host;
752
753 return inode_resv_map(inode);
754
755 } else {
84afd99b
AW
756 return (struct resv_map *)(get_vma_private_data(vma) &
757 ~HPAGE_RESV_MASK);
4e35f483 758 }
a1e78772
MG
759}
760
84afd99b 761static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
a1e78772 762{
81d1b09c
SL
763 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
764 VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
a1e78772 765
84afd99b
AW
766 set_vma_private_data(vma, (get_vma_private_data(vma) &
767 HPAGE_RESV_MASK) | (unsigned long)map);
04f2cbe3
MG
768}
769
770static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
771{
81d1b09c
SL
772 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
773 VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
e7c4b0bf
AW
774
775 set_vma_private_data(vma, get_vma_private_data(vma) | flags);
04f2cbe3
MG
776}
777
778static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
779{
81d1b09c 780 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
e7c4b0bf
AW
781
782 return (get_vma_private_data(vma) & flag) != 0;
a1e78772
MG
783}
784
04f2cbe3 785/* Reset counters to 0 and clear all HPAGE_RESV_* flags */
a1e78772
MG
786void reset_vma_resv_huge_pages(struct vm_area_struct *vma)
787{
81d1b09c 788 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
f83a275d 789 if (!(vma->vm_flags & VM_MAYSHARE))
a1e78772
MG
790 vma->vm_private_data = (void *)0;
791}
792
793/* Returns true if the VMA has associated reserve pages */
559ec2f8 794static bool vma_has_reserves(struct vm_area_struct *vma, long chg)
a1e78772 795{
af0ed73e
JK
796 if (vma->vm_flags & VM_NORESERVE) {
797 /*
798 * This address is already reserved by other process(chg == 0),
799 * so, we should decrement reserved count. Without decrementing,
800 * reserve count remains after releasing inode, because this
801 * allocated page will go into page cache and is regarded as
802 * coming from reserved pool in releasing step. Currently, we
803 * don't have any other solution to deal with this situation
804 * properly, so add work-around here.
805 */
806 if (vma->vm_flags & VM_MAYSHARE && chg == 0)
559ec2f8 807 return true;
af0ed73e 808 else
559ec2f8 809 return false;
af0ed73e 810 }
a63884e9
JK
811
812 /* Shared mappings always use reserves */
1fb1b0e9
MK
813 if (vma->vm_flags & VM_MAYSHARE) {
814 /*
815 * We know VM_NORESERVE is not set. Therefore, there SHOULD
816 * be a region map for all pages. The only situation where
817 * there is no region map is if a hole was punched via
818 * fallocate. In this case, there really are no reverves to
819 * use. This situation is indicated if chg != 0.
820 */
821 if (chg)
822 return false;
823 else
824 return true;
825 }
a63884e9
JK
826
827 /*
828 * Only the process that called mmap() has reserves for
829 * private mappings.
830 */
67961f9d
MK
831 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
832 /*
833 * Like the shared case above, a hole punch or truncate
834 * could have been performed on the private mapping.
835 * Examine the value of chg to determine if reserves
836 * actually exist or were previously consumed.
837 * Very Subtle - The value of chg comes from a previous
838 * call to vma_needs_reserves(). The reserve map for
839 * private mappings has different (opposite) semantics
840 * than that of shared mappings. vma_needs_reserves()
841 * has already taken this difference in semantics into
842 * account. Therefore, the meaning of chg is the same
843 * as in the shared case above. Code could easily be
844 * combined, but keeping it separate draws attention to
845 * subtle differences.
846 */
847 if (chg)
848 return false;
849 else
850 return true;
851 }
a63884e9 852
559ec2f8 853 return false;
a1e78772
MG
854}
855
a5516438 856static void enqueue_huge_page(struct hstate *h, struct page *page)
1da177e4
LT
857{
858 int nid = page_to_nid(page);
0edaecfa 859 list_move(&page->lru, &h->hugepage_freelists[nid]);
a5516438
AK
860 h->free_huge_pages++;
861 h->free_huge_pages_node[nid]++;
1da177e4
LT
862}
863
94310cbc 864static struct page *dequeue_huge_page_node_exact(struct hstate *h, int nid)
bf50bab2
NH
865{
866 struct page *page;
867
c8721bbb 868 list_for_each_entry(page, &h->hugepage_freelists[nid], lru)
243abd5b 869 if (!PageHWPoison(page))
c8721bbb
NH
870 break;
871 /*
872 * if 'non-isolated free hugepage' not found on the list,
873 * the allocation fails.
874 */
875 if (&h->hugepage_freelists[nid] == &page->lru)
bf50bab2 876 return NULL;
0edaecfa 877 list_move(&page->lru, &h->hugepage_activelist);
a9869b83 878 set_page_refcounted(page);
bf50bab2
NH
879 h->free_huge_pages--;
880 h->free_huge_pages_node[nid]--;
881 return page;
882}
883
3e59fcb0
MH
884static struct page *dequeue_huge_page_nodemask(struct hstate *h, gfp_t gfp_mask, int nid,
885 nodemask_t *nmask)
94310cbc 886{
3e59fcb0
MH
887 unsigned int cpuset_mems_cookie;
888 struct zonelist *zonelist;
889 struct zone *zone;
890 struct zoneref *z;
98fa15f3 891 int node = NUMA_NO_NODE;
94310cbc 892
3e59fcb0
MH
893 zonelist = node_zonelist(nid, gfp_mask);
894
895retry_cpuset:
896 cpuset_mems_cookie = read_mems_allowed_begin();
897 for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), nmask) {
898 struct page *page;
899
900 if (!cpuset_zone_allowed(zone, gfp_mask))
901 continue;
902 /*
903 * no need to ask again on the same node. Pool is node rather than
904 * zone aware
905 */
906 if (zone_to_nid(zone) == node)
907 continue;
908 node = zone_to_nid(zone);
94310cbc 909
94310cbc
AK
910 page = dequeue_huge_page_node_exact(h, node);
911 if (page)
912 return page;
913 }
3e59fcb0
MH
914 if (unlikely(read_mems_allowed_retry(cpuset_mems_cookie)))
915 goto retry_cpuset;
916
94310cbc
AK
917 return NULL;
918}
919
86cdb465
NH
920/* Movability of hugepages depends on migration support. */
921static inline gfp_t htlb_alloc_mask(struct hstate *h)
922{
7ed2c31d 923 if (hugepage_movable_supported(h))
86cdb465
NH
924 return GFP_HIGHUSER_MOVABLE;
925 else
926 return GFP_HIGHUSER;
927}
928
a5516438
AK
929static struct page *dequeue_huge_page_vma(struct hstate *h,
930 struct vm_area_struct *vma,
af0ed73e
JK
931 unsigned long address, int avoid_reserve,
932 long chg)
1da177e4 933{
3e59fcb0 934 struct page *page;
480eccf9 935 struct mempolicy *mpol;
04ec6264 936 gfp_t gfp_mask;
3e59fcb0 937 nodemask_t *nodemask;
04ec6264 938 int nid;
1da177e4 939
a1e78772
MG
940 /*
941 * A child process with MAP_PRIVATE mappings created by their parent
942 * have no page reserves. This check ensures that reservations are
943 * not "stolen". The child may still get SIGKILLed
944 */
af0ed73e 945 if (!vma_has_reserves(vma, chg) &&
a5516438 946 h->free_huge_pages - h->resv_huge_pages == 0)
c0ff7453 947 goto err;
a1e78772 948
04f2cbe3 949 /* If reserves cannot be used, ensure enough pages are in the pool */
a5516438 950 if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
6eab04a8 951 goto err;
04f2cbe3 952
04ec6264
VB
953 gfp_mask = htlb_alloc_mask(h);
954 nid = huge_node(vma, address, gfp_mask, &mpol, &nodemask);
3e59fcb0
MH
955 page = dequeue_huge_page_nodemask(h, gfp_mask, nid, nodemask);
956 if (page && !avoid_reserve && vma_has_reserves(vma, chg)) {
957 SetPagePrivate(page);
958 h->resv_huge_pages--;
1da177e4 959 }
cc9a6c87 960
52cd3b07 961 mpol_cond_put(mpol);
1da177e4 962 return page;
cc9a6c87
MG
963
964err:
cc9a6c87 965 return NULL;
1da177e4
LT
966}
967
1cac6f2c
LC
968/*
969 * common helper functions for hstate_next_node_to_{alloc|free}.
970 * We may have allocated or freed a huge page based on a different
971 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
972 * be outside of *nodes_allowed. Ensure that we use an allowed
973 * node for alloc or free.
974 */
975static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
976{
0edaf86c 977 nid = next_node_in(nid, *nodes_allowed);
1cac6f2c
LC
978 VM_BUG_ON(nid >= MAX_NUMNODES);
979
980 return nid;
981}
982
983static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
984{
985 if (!node_isset(nid, *nodes_allowed))
986 nid = next_node_allowed(nid, nodes_allowed);
987 return nid;
988}
989
990/*
991 * returns the previously saved node ["this node"] from which to
992 * allocate a persistent huge page for the pool and advance the
993 * next node from which to allocate, handling wrap at end of node
994 * mask.
995 */
996static int hstate_next_node_to_alloc(struct hstate *h,
997 nodemask_t *nodes_allowed)
998{
999 int nid;
1000
1001 VM_BUG_ON(!nodes_allowed);
1002
1003 nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
1004 h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
1005
1006 return nid;
1007}
1008
1009/*
1010 * helper for free_pool_huge_page() - return the previously saved
1011 * node ["this node"] from which to free a huge page. Advance the
1012 * next node id whether or not we find a free huge page to free so
1013 * that the next attempt to free addresses the next node.
1014 */
1015static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
1016{
1017 int nid;
1018
1019 VM_BUG_ON(!nodes_allowed);
1020
1021 nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
1022 h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
1023
1024 return nid;
1025}
1026
1027#define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \
1028 for (nr_nodes = nodes_weight(*mask); \
1029 nr_nodes > 0 && \
1030 ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \
1031 nr_nodes--)
1032
1033#define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \
1034 for (nr_nodes = nodes_weight(*mask); \
1035 nr_nodes > 0 && \
1036 ((node = hstate_next_node_to_free(hs, mask)) || 1); \
1037 nr_nodes--)
1038
e1073d1e 1039#ifdef CONFIG_ARCH_HAS_GIGANTIC_PAGE
944d9fec 1040static void destroy_compound_gigantic_page(struct page *page,
d00181b9 1041 unsigned int order)
944d9fec
LC
1042{
1043 int i;
1044 int nr_pages = 1 << order;
1045 struct page *p = page + 1;
1046
c8cc708a 1047 atomic_set(compound_mapcount_ptr(page), 0);
944d9fec 1048 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
1d798ca3 1049 clear_compound_head(p);
944d9fec 1050 set_page_refcounted(p);
944d9fec
LC
1051 }
1052
1053 set_compound_order(page, 0);
1054 __ClearPageHead(page);
1055}
1056
d00181b9 1057static void free_gigantic_page(struct page *page, unsigned int order)
944d9fec
LC
1058{
1059 free_contig_range(page_to_pfn(page), 1 << order);
1060}
1061
4eb0716e 1062#ifdef CONFIG_CONTIG_ALLOC
944d9fec 1063static int __alloc_gigantic_page(unsigned long start_pfn,
79b63f12 1064 unsigned long nr_pages, gfp_t gfp_mask)
944d9fec
LC
1065{
1066 unsigned long end_pfn = start_pfn + nr_pages;
ca96b625 1067 return alloc_contig_range(start_pfn, end_pfn, MIGRATE_MOVABLE,
79b63f12 1068 gfp_mask);
944d9fec
LC
1069}
1070
f44b2dda
JK
1071static bool pfn_range_valid_gigantic(struct zone *z,
1072 unsigned long start_pfn, unsigned long nr_pages)
944d9fec
LC
1073{
1074 unsigned long i, end_pfn = start_pfn + nr_pages;
1075 struct page *page;
1076
1077 for (i = start_pfn; i < end_pfn; i++) {
1078 if (!pfn_valid(i))
1079 return false;
1080
1081 page = pfn_to_page(i);
1082
f44b2dda
JK
1083 if (page_zone(page) != z)
1084 return false;
1085
944d9fec
LC
1086 if (PageReserved(page))
1087 return false;
1088
1089 if (page_count(page) > 0)
1090 return false;
1091
1092 if (PageHuge(page))
1093 return false;
1094 }
1095
1096 return true;
1097}
1098
1099static bool zone_spans_last_pfn(const struct zone *zone,
1100 unsigned long start_pfn, unsigned long nr_pages)
1101{
1102 unsigned long last_pfn = start_pfn + nr_pages - 1;
1103 return zone_spans_pfn(zone, last_pfn);
1104}
1105
d9cc948f
MH
1106static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask,
1107 int nid, nodemask_t *nodemask)
944d9fec 1108{
79b63f12 1109 unsigned int order = huge_page_order(h);
944d9fec
LC
1110 unsigned long nr_pages = 1 << order;
1111 unsigned long ret, pfn, flags;
79b63f12
MH
1112 struct zonelist *zonelist;
1113 struct zone *zone;
1114 struct zoneref *z;
944d9fec 1115
79b63f12 1116 zonelist = node_zonelist(nid, gfp_mask);
d9cc948f 1117 for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), nodemask) {
79b63f12 1118 spin_lock_irqsave(&zone->lock, flags);
944d9fec 1119
79b63f12
MH
1120 pfn = ALIGN(zone->zone_start_pfn, nr_pages);
1121 while (zone_spans_last_pfn(zone, pfn, nr_pages)) {
1122 if (pfn_range_valid_gigantic(zone, pfn, nr_pages)) {
944d9fec
LC
1123 /*
1124 * We release the zone lock here because
1125 * alloc_contig_range() will also lock the zone
1126 * at some point. If there's an allocation
1127 * spinning on this lock, it may win the race
1128 * and cause alloc_contig_range() to fail...
1129 */
79b63f12
MH
1130 spin_unlock_irqrestore(&zone->lock, flags);
1131 ret = __alloc_gigantic_page(pfn, nr_pages, gfp_mask);
944d9fec
LC
1132 if (!ret)
1133 return pfn_to_page(pfn);
79b63f12 1134 spin_lock_irqsave(&zone->lock, flags);
944d9fec
LC
1135 }
1136 pfn += nr_pages;
1137 }
1138
79b63f12 1139 spin_unlock_irqrestore(&zone->lock, flags);
944d9fec
LC
1140 }
1141
1142 return NULL;
1143}
1144
1145static void prep_new_huge_page(struct hstate *h, struct page *page, int nid);
d00181b9 1146static void prep_compound_gigantic_page(struct page *page, unsigned int order);
4eb0716e
AG
1147#else /* !CONFIG_CONTIG_ALLOC */
1148static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask,
1149 int nid, nodemask_t *nodemask)
1150{
1151 return NULL;
1152}
1153#endif /* CONFIG_CONTIG_ALLOC */
944d9fec 1154
e1073d1e 1155#else /* !CONFIG_ARCH_HAS_GIGANTIC_PAGE */
d9cc948f 1156static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask,
4eb0716e
AG
1157 int nid, nodemask_t *nodemask)
1158{
1159 return NULL;
1160}
d00181b9 1161static inline void free_gigantic_page(struct page *page, unsigned int order) { }
944d9fec 1162static inline void destroy_compound_gigantic_page(struct page *page,
d00181b9 1163 unsigned int order) { }
944d9fec
LC
1164#endif
1165
a5516438 1166static void update_and_free_page(struct hstate *h, struct page *page)
6af2acb6
AL
1167{
1168 int i;
a5516438 1169
4eb0716e 1170 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
944d9fec 1171 return;
18229df5 1172
a5516438
AK
1173 h->nr_huge_pages--;
1174 h->nr_huge_pages_node[page_to_nid(page)]--;
1175 for (i = 0; i < pages_per_huge_page(h); i++) {
32f84528
CF
1176 page[i].flags &= ~(1 << PG_locked | 1 << PG_error |
1177 1 << PG_referenced | 1 << PG_dirty |
a7407a27
LC
1178 1 << PG_active | 1 << PG_private |
1179 1 << PG_writeback);
6af2acb6 1180 }
309381fe 1181 VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page);
f1e61557 1182 set_compound_page_dtor(page, NULL_COMPOUND_DTOR);
6af2acb6 1183 set_page_refcounted(page);
944d9fec
LC
1184 if (hstate_is_gigantic(h)) {
1185 destroy_compound_gigantic_page(page, huge_page_order(h));
1186 free_gigantic_page(page, huge_page_order(h));
1187 } else {
944d9fec
LC
1188 __free_pages(page, huge_page_order(h));
1189 }
6af2acb6
AL
1190}
1191
e5ff2159
AK
1192struct hstate *size_to_hstate(unsigned long size)
1193{
1194 struct hstate *h;
1195
1196 for_each_hstate(h) {
1197 if (huge_page_size(h) == size)
1198 return h;
1199 }
1200 return NULL;
1201}
1202
bcc54222
NH
1203/*
1204 * Test to determine whether the hugepage is "active/in-use" (i.e. being linked
1205 * to hstate->hugepage_activelist.)
1206 *
1207 * This function can be called for tail pages, but never returns true for them.
1208 */
1209bool page_huge_active(struct page *page)
1210{
1211 VM_BUG_ON_PAGE(!PageHuge(page), page);
1212 return PageHead(page) && PagePrivate(&page[1]);
1213}
1214
1215/* never called for tail page */
1216static void set_page_huge_active(struct page *page)
1217{
1218 VM_BUG_ON_PAGE(!PageHeadHuge(page), page);
1219 SetPagePrivate(&page[1]);
1220}
1221
1222static void clear_page_huge_active(struct page *page)
1223{
1224 VM_BUG_ON_PAGE(!PageHeadHuge(page), page);
1225 ClearPagePrivate(&page[1]);
1226}
1227
ab5ac90a
MH
1228/*
1229 * Internal hugetlb specific page flag. Do not use outside of the hugetlb
1230 * code
1231 */
1232static inline bool PageHugeTemporary(struct page *page)
1233{
1234 if (!PageHuge(page))
1235 return false;
1236
1237 return (unsigned long)page[2].mapping == -1U;
1238}
1239
1240static inline void SetPageHugeTemporary(struct page *page)
1241{
1242 page[2].mapping = (void *)-1U;
1243}
1244
1245static inline void ClearPageHugeTemporary(struct page *page)
1246{
1247 page[2].mapping = NULL;
1248}
1249
8f1d26d0 1250void free_huge_page(struct page *page)
27a85ef1 1251{
a5516438
AK
1252 /*
1253 * Can't pass hstate in here because it is called from the
1254 * compound page destructor.
1255 */
e5ff2159 1256 struct hstate *h = page_hstate(page);
7893d1d5 1257 int nid = page_to_nid(page);
90481622
DG
1258 struct hugepage_subpool *spool =
1259 (struct hugepage_subpool *)page_private(page);
07443a85 1260 bool restore_reserve;
27a85ef1 1261
b4330afb
MK
1262 VM_BUG_ON_PAGE(page_count(page), page);
1263 VM_BUG_ON_PAGE(page_mapcount(page), page);
8ace22bc
YW
1264
1265 set_page_private(page, 0);
1266 page->mapping = NULL;
07443a85 1267 restore_reserve = PagePrivate(page);
16c794b4 1268 ClearPagePrivate(page);
27a85ef1 1269
1c5ecae3
MK
1270 /*
1271 * A return code of zero implies that the subpool will be under its
1272 * minimum size if the reservation is not restored after page is free.
1273 * Therefore, force restore_reserve operation.
1274 */
1275 if (hugepage_subpool_put_pages(spool, 1) == 0)
1276 restore_reserve = true;
1277
27a85ef1 1278 spin_lock(&hugetlb_lock);
bcc54222 1279 clear_page_huge_active(page);
6d76dcf4
AK
1280 hugetlb_cgroup_uncharge_page(hstate_index(h),
1281 pages_per_huge_page(h), page);
07443a85
JK
1282 if (restore_reserve)
1283 h->resv_huge_pages++;
1284
ab5ac90a
MH
1285 if (PageHugeTemporary(page)) {
1286 list_del(&page->lru);
1287 ClearPageHugeTemporary(page);
1288 update_and_free_page(h, page);
1289 } else if (h->surplus_huge_pages_node[nid]) {
0edaecfa
AK
1290 /* remove the page from active list */
1291 list_del(&page->lru);
a5516438
AK
1292 update_and_free_page(h, page);
1293 h->surplus_huge_pages--;
1294 h->surplus_huge_pages_node[nid]--;
7893d1d5 1295 } else {
5d3a551c 1296 arch_clear_hugepage_flags(page);
a5516438 1297 enqueue_huge_page(h, page);
7893d1d5 1298 }
27a85ef1
DG
1299 spin_unlock(&hugetlb_lock);
1300}
1301
a5516438 1302static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
b7ba30c6 1303{
0edaecfa 1304 INIT_LIST_HEAD(&page->lru);
f1e61557 1305 set_compound_page_dtor(page, HUGETLB_PAGE_DTOR);
b7ba30c6 1306 spin_lock(&hugetlb_lock);
9dd540e2 1307 set_hugetlb_cgroup(page, NULL);
a5516438
AK
1308 h->nr_huge_pages++;
1309 h->nr_huge_pages_node[nid]++;
b7ba30c6 1310 spin_unlock(&hugetlb_lock);
b7ba30c6
AK
1311}
1312
d00181b9 1313static void prep_compound_gigantic_page(struct page *page, unsigned int order)
20a0307c
WF
1314{
1315 int i;
1316 int nr_pages = 1 << order;
1317 struct page *p = page + 1;
1318
1319 /* we rely on prep_new_huge_page to set the destructor */
1320 set_compound_order(page, order);
ef5a22be 1321 __ClearPageReserved(page);
de09d31d 1322 __SetPageHead(page);
20a0307c 1323 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
ef5a22be
AA
1324 /*
1325 * For gigantic hugepages allocated through bootmem at
1326 * boot, it's safer to be consistent with the not-gigantic
1327 * hugepages and clear the PG_reserved bit from all tail pages
1328 * too. Otherwse drivers using get_user_pages() to access tail
1329 * pages may get the reference counting wrong if they see
1330 * PG_reserved set on a tail page (despite the head page not
1331 * having PG_reserved set). Enforcing this consistency between
1332 * head and tail pages allows drivers to optimize away a check
1333 * on the head page when they need know if put_page() is needed
1334 * after get_user_pages().
1335 */
1336 __ClearPageReserved(p);
58a84aa9 1337 set_page_count(p, 0);
1d798ca3 1338 set_compound_head(p, page);
20a0307c 1339 }
b4330afb 1340 atomic_set(compound_mapcount_ptr(page), -1);
20a0307c
WF
1341}
1342
7795912c
AM
1343/*
1344 * PageHuge() only returns true for hugetlbfs pages, but not for normal or
1345 * transparent huge pages. See the PageTransHuge() documentation for more
1346 * details.
1347 */
20a0307c
WF
1348int PageHuge(struct page *page)
1349{
20a0307c
WF
1350 if (!PageCompound(page))
1351 return 0;
1352
1353 page = compound_head(page);
f1e61557 1354 return page[1].compound_dtor == HUGETLB_PAGE_DTOR;
20a0307c 1355}
43131e14
NH
1356EXPORT_SYMBOL_GPL(PageHuge);
1357
27c73ae7
AA
1358/*
1359 * PageHeadHuge() only returns true for hugetlbfs head page, but not for
1360 * normal or transparent huge pages.
1361 */
1362int PageHeadHuge(struct page *page_head)
1363{
27c73ae7
AA
1364 if (!PageHead(page_head))
1365 return 0;
1366
758f66a2 1367 return get_compound_page_dtor(page_head) == free_huge_page;
27c73ae7 1368}
27c73ae7 1369
13d60f4b
ZY
1370pgoff_t __basepage_index(struct page *page)
1371{
1372 struct page *page_head = compound_head(page);
1373 pgoff_t index = page_index(page_head);
1374 unsigned long compound_idx;
1375
1376 if (!PageHuge(page_head))
1377 return page_index(page);
1378
1379 if (compound_order(page_head) >= MAX_ORDER)
1380 compound_idx = page_to_pfn(page) - page_to_pfn(page_head);
1381 else
1382 compound_idx = page - page_head;
1383
1384 return (index << compound_order(page_head)) + compound_idx;
1385}
1386
0c397dae 1387static struct page *alloc_buddy_huge_page(struct hstate *h,
af0fb9df 1388 gfp_t gfp_mask, int nid, nodemask_t *nmask)
1da177e4 1389{
af0fb9df 1390 int order = huge_page_order(h);
1da177e4 1391 struct page *page;
f96efd58 1392
af0fb9df
MH
1393 gfp_mask |= __GFP_COMP|__GFP_RETRY_MAYFAIL|__GFP_NOWARN;
1394 if (nid == NUMA_NO_NODE)
1395 nid = numa_mem_id();
1396 page = __alloc_pages_nodemask(gfp_mask, order, nid, nmask);
1397 if (page)
1398 __count_vm_event(HTLB_BUDDY_PGALLOC);
1399 else
1400 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
63b4613c
NA
1401
1402 return page;
1403}
1404
0c397dae
MH
1405/*
1406 * Common helper to allocate a fresh hugetlb page. All specific allocators
1407 * should use this function to get new hugetlb pages
1408 */
1409static struct page *alloc_fresh_huge_page(struct hstate *h,
1410 gfp_t gfp_mask, int nid, nodemask_t *nmask)
1411{
1412 struct page *page;
1413
1414 if (hstate_is_gigantic(h))
1415 page = alloc_gigantic_page(h, gfp_mask, nid, nmask);
1416 else
1417 page = alloc_buddy_huge_page(h, gfp_mask,
1418 nid, nmask);
1419 if (!page)
1420 return NULL;
1421
1422 if (hstate_is_gigantic(h))
1423 prep_compound_gigantic_page(page, huge_page_order(h));
1424 prep_new_huge_page(h, page, page_to_nid(page));
1425
1426 return page;
1427}
1428
af0fb9df
MH
1429/*
1430 * Allocates a fresh page to the hugetlb allocator pool in the node interleaved
1431 * manner.
1432 */
0c397dae 1433static int alloc_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed)
b2261026
JK
1434{
1435 struct page *page;
1436 int nr_nodes, node;
af0fb9df 1437 gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE;
b2261026
JK
1438
1439 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
0c397dae 1440 page = alloc_fresh_huge_page(h, gfp_mask, node, nodes_allowed);
af0fb9df 1441 if (page)
b2261026 1442 break;
b2261026
JK
1443 }
1444
af0fb9df
MH
1445 if (!page)
1446 return 0;
b2261026 1447
af0fb9df
MH
1448 put_page(page); /* free it into the hugepage allocator */
1449
1450 return 1;
b2261026
JK
1451}
1452
e8c5c824
LS
1453/*
1454 * Free huge page from pool from next node to free.
1455 * Attempt to keep persistent huge pages more or less
1456 * balanced over allowed nodes.
1457 * Called with hugetlb_lock locked.
1458 */
6ae11b27
LS
1459static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
1460 bool acct_surplus)
e8c5c824 1461{
b2261026 1462 int nr_nodes, node;
e8c5c824
LS
1463 int ret = 0;
1464
b2261026 1465 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
685f3457
LS
1466 /*
1467 * If we're returning unused surplus pages, only examine
1468 * nodes with surplus pages.
1469 */
b2261026
JK
1470 if ((!acct_surplus || h->surplus_huge_pages_node[node]) &&
1471 !list_empty(&h->hugepage_freelists[node])) {
e8c5c824 1472 struct page *page =
b2261026 1473 list_entry(h->hugepage_freelists[node].next,
e8c5c824
LS
1474 struct page, lru);
1475 list_del(&page->lru);
1476 h->free_huge_pages--;
b2261026 1477 h->free_huge_pages_node[node]--;
685f3457
LS
1478 if (acct_surplus) {
1479 h->surplus_huge_pages--;
b2261026 1480 h->surplus_huge_pages_node[node]--;
685f3457 1481 }
e8c5c824
LS
1482 update_and_free_page(h, page);
1483 ret = 1;
9a76db09 1484 break;
e8c5c824 1485 }
b2261026 1486 }
e8c5c824
LS
1487
1488 return ret;
1489}
1490
c8721bbb
NH
1491/*
1492 * Dissolve a given free hugepage into free buddy pages. This function does
082d5b6b 1493 * nothing for in-use (including surplus) hugepages. Returns -EBUSY if the
6bc9b564
NH
1494 * dissolution fails because a give page is not a free hugepage, or because
1495 * free hugepages are fully reserved.
c8721bbb 1496 */
c3114a84 1497int dissolve_free_huge_page(struct page *page)
c8721bbb 1498{
6bc9b564 1499 int rc = -EBUSY;
082d5b6b 1500
c8721bbb
NH
1501 spin_lock(&hugetlb_lock);
1502 if (PageHuge(page) && !page_count(page)) {
2247bb33
GS
1503 struct page *head = compound_head(page);
1504 struct hstate *h = page_hstate(head);
1505 int nid = page_to_nid(head);
6bc9b564 1506 if (h->free_huge_pages - h->resv_huge_pages == 0)
082d5b6b 1507 goto out;
c3114a84
AK
1508 /*
1509 * Move PageHWPoison flag from head page to the raw error page,
1510 * which makes any subpages rather than the error page reusable.
1511 */
1512 if (PageHWPoison(head) && page != head) {
1513 SetPageHWPoison(page);
1514 ClearPageHWPoison(head);
1515 }
2247bb33 1516 list_del(&head->lru);
c8721bbb
NH
1517 h->free_huge_pages--;
1518 h->free_huge_pages_node[nid]--;
c1470b33 1519 h->max_huge_pages--;
2247bb33 1520 update_and_free_page(h, head);
6bc9b564 1521 rc = 0;
c8721bbb 1522 }
082d5b6b 1523out:
c8721bbb 1524 spin_unlock(&hugetlb_lock);
082d5b6b 1525 return rc;
c8721bbb
NH
1526}
1527
1528/*
1529 * Dissolve free hugepages in a given pfn range. Used by memory hotplug to
1530 * make specified memory blocks removable from the system.
2247bb33
GS
1531 * Note that this will dissolve a free gigantic hugepage completely, if any
1532 * part of it lies within the given range.
082d5b6b
GS
1533 * Also note that if dissolve_free_huge_page() returns with an error, all
1534 * free hugepages that were dissolved before that error are lost.
c8721bbb 1535 */
082d5b6b 1536int dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn)
c8721bbb 1537{
c8721bbb 1538 unsigned long pfn;
eb03aa00 1539 struct page *page;
082d5b6b 1540 int rc = 0;
c8721bbb 1541
d0177639 1542 if (!hugepages_supported())
082d5b6b 1543 return rc;
d0177639 1544
eb03aa00
GS
1545 for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << minimum_order) {
1546 page = pfn_to_page(pfn);
1547 if (PageHuge(page) && !page_count(page)) {
1548 rc = dissolve_free_huge_page(page);
1549 if (rc)
1550 break;
1551 }
1552 }
082d5b6b
GS
1553
1554 return rc;
c8721bbb
NH
1555}
1556
ab5ac90a
MH
1557/*
1558 * Allocates a fresh surplus page from the page allocator.
1559 */
0c397dae 1560static struct page *alloc_surplus_huge_page(struct hstate *h, gfp_t gfp_mask,
aaf14e40 1561 int nid, nodemask_t *nmask)
7893d1d5 1562{
9980d744 1563 struct page *page = NULL;
7893d1d5 1564
bae7f4ae 1565 if (hstate_is_gigantic(h))
aa888a74
AK
1566 return NULL;
1567
d1c3fb1f 1568 spin_lock(&hugetlb_lock);
9980d744
MH
1569 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages)
1570 goto out_unlock;
d1c3fb1f
NA
1571 spin_unlock(&hugetlb_lock);
1572
0c397dae 1573 page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask);
9980d744 1574 if (!page)
0c397dae 1575 return NULL;
d1c3fb1f
NA
1576
1577 spin_lock(&hugetlb_lock);
9980d744
MH
1578 /*
1579 * We could have raced with the pool size change.
1580 * Double check that and simply deallocate the new page
1581 * if we would end up overcommiting the surpluses. Abuse
1582 * temporary page to workaround the nasty free_huge_page
1583 * codeflow
1584 */
1585 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
1586 SetPageHugeTemporary(page);
2bf753e6 1587 spin_unlock(&hugetlb_lock);
9980d744 1588 put_page(page);
2bf753e6 1589 return NULL;
9980d744 1590 } else {
9980d744 1591 h->surplus_huge_pages++;
4704dea3 1592 h->surplus_huge_pages_node[page_to_nid(page)]++;
7893d1d5 1593 }
9980d744
MH
1594
1595out_unlock:
d1c3fb1f 1596 spin_unlock(&hugetlb_lock);
7893d1d5
AL
1597
1598 return page;
1599}
1600
9a4e9f3b
AK
1601struct page *alloc_migrate_huge_page(struct hstate *h, gfp_t gfp_mask,
1602 int nid, nodemask_t *nmask)
ab5ac90a
MH
1603{
1604 struct page *page;
1605
1606 if (hstate_is_gigantic(h))
1607 return NULL;
1608
0c397dae 1609 page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask);
ab5ac90a
MH
1610 if (!page)
1611 return NULL;
1612
1613 /*
1614 * We do not account these pages as surplus because they are only
1615 * temporary and will be released properly on the last reference
1616 */
ab5ac90a
MH
1617 SetPageHugeTemporary(page);
1618
1619 return page;
1620}
1621
099730d6
DH
1622/*
1623 * Use the VMA's mpolicy to allocate a huge page from the buddy.
1624 */
e0ec90ee 1625static
0c397dae 1626struct page *alloc_buddy_huge_page_with_mpol(struct hstate *h,
099730d6
DH
1627 struct vm_area_struct *vma, unsigned long addr)
1628{
aaf14e40
MH
1629 struct page *page;
1630 struct mempolicy *mpol;
1631 gfp_t gfp_mask = htlb_alloc_mask(h);
1632 int nid;
1633 nodemask_t *nodemask;
1634
1635 nid = huge_node(vma, addr, gfp_mask, &mpol, &nodemask);
0c397dae 1636 page = alloc_surplus_huge_page(h, gfp_mask, nid, nodemask);
aaf14e40
MH
1637 mpol_cond_put(mpol);
1638
1639 return page;
099730d6
DH
1640}
1641
ab5ac90a 1642/* page migration callback function */
bf50bab2
NH
1643struct page *alloc_huge_page_node(struct hstate *h, int nid)
1644{
aaf14e40 1645 gfp_t gfp_mask = htlb_alloc_mask(h);
4ef91848 1646 struct page *page = NULL;
bf50bab2 1647
aaf14e40
MH
1648 if (nid != NUMA_NO_NODE)
1649 gfp_mask |= __GFP_THISNODE;
1650
bf50bab2 1651 spin_lock(&hugetlb_lock);
4ef91848 1652 if (h->free_huge_pages - h->resv_huge_pages > 0)
3e59fcb0 1653 page = dequeue_huge_page_nodemask(h, gfp_mask, nid, NULL);
bf50bab2
NH
1654 spin_unlock(&hugetlb_lock);
1655
94ae8ba7 1656 if (!page)
0c397dae 1657 page = alloc_migrate_huge_page(h, gfp_mask, nid, NULL);
bf50bab2
NH
1658
1659 return page;
1660}
1661
ab5ac90a 1662/* page migration callback function */
3e59fcb0
MH
1663struct page *alloc_huge_page_nodemask(struct hstate *h, int preferred_nid,
1664 nodemask_t *nmask)
4db9b2ef 1665{
aaf14e40 1666 gfp_t gfp_mask = htlb_alloc_mask(h);
4db9b2ef
MH
1667
1668 spin_lock(&hugetlb_lock);
1669 if (h->free_huge_pages - h->resv_huge_pages > 0) {
3e59fcb0
MH
1670 struct page *page;
1671
1672 page = dequeue_huge_page_nodemask(h, gfp_mask, preferred_nid, nmask);
1673 if (page) {
1674 spin_unlock(&hugetlb_lock);
1675 return page;
4db9b2ef
MH
1676 }
1677 }
1678 spin_unlock(&hugetlb_lock);
4db9b2ef 1679
0c397dae 1680 return alloc_migrate_huge_page(h, gfp_mask, preferred_nid, nmask);
4db9b2ef
MH
1681}
1682
ebd63723 1683/* mempolicy aware migration callback */
389c8178
MH
1684struct page *alloc_huge_page_vma(struct hstate *h, struct vm_area_struct *vma,
1685 unsigned long address)
ebd63723
MH
1686{
1687 struct mempolicy *mpol;
1688 nodemask_t *nodemask;
1689 struct page *page;
ebd63723
MH
1690 gfp_t gfp_mask;
1691 int node;
1692
ebd63723
MH
1693 gfp_mask = htlb_alloc_mask(h);
1694 node = huge_node(vma, address, gfp_mask, &mpol, &nodemask);
1695 page = alloc_huge_page_nodemask(h, node, nodemask);
1696 mpol_cond_put(mpol);
1697
1698 return page;
1699}
1700
e4e574b7 1701/*
25985edc 1702 * Increase the hugetlb pool such that it can accommodate a reservation
e4e574b7
AL
1703 * of size 'delta'.
1704 */
a5516438 1705static int gather_surplus_pages(struct hstate *h, int delta)
e4e574b7
AL
1706{
1707 struct list_head surplus_list;
1708 struct page *page, *tmp;
1709 int ret, i;
1710 int needed, allocated;
28073b02 1711 bool alloc_ok = true;
e4e574b7 1712
a5516438 1713 needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
ac09b3a1 1714 if (needed <= 0) {
a5516438 1715 h->resv_huge_pages += delta;
e4e574b7 1716 return 0;
ac09b3a1 1717 }
e4e574b7
AL
1718
1719 allocated = 0;
1720 INIT_LIST_HEAD(&surplus_list);
1721
1722 ret = -ENOMEM;
1723retry:
1724 spin_unlock(&hugetlb_lock);
1725 for (i = 0; i < needed; i++) {
0c397dae 1726 page = alloc_surplus_huge_page(h, htlb_alloc_mask(h),
aaf14e40 1727 NUMA_NO_NODE, NULL);
28073b02
HD
1728 if (!page) {
1729 alloc_ok = false;
1730 break;
1731 }
e4e574b7 1732 list_add(&page->lru, &surplus_list);
69ed779a 1733 cond_resched();
e4e574b7 1734 }
28073b02 1735 allocated += i;
e4e574b7
AL
1736
1737 /*
1738 * After retaking hugetlb_lock, we need to recalculate 'needed'
1739 * because either resv_huge_pages or free_huge_pages may have changed.
1740 */
1741 spin_lock(&hugetlb_lock);
a5516438
AK
1742 needed = (h->resv_huge_pages + delta) -
1743 (h->free_huge_pages + allocated);
28073b02
HD
1744 if (needed > 0) {
1745 if (alloc_ok)
1746 goto retry;
1747 /*
1748 * We were not able to allocate enough pages to
1749 * satisfy the entire reservation so we free what
1750 * we've allocated so far.
1751 */
1752 goto free;
1753 }
e4e574b7
AL
1754 /*
1755 * The surplus_list now contains _at_least_ the number of extra pages
25985edc 1756 * needed to accommodate the reservation. Add the appropriate number
e4e574b7 1757 * of pages to the hugetlb pool and free the extras back to the buddy
ac09b3a1
AL
1758 * allocator. Commit the entire reservation here to prevent another
1759 * process from stealing the pages as they are added to the pool but
1760 * before they are reserved.
e4e574b7
AL
1761 */
1762 needed += allocated;
a5516438 1763 h->resv_huge_pages += delta;
e4e574b7 1764 ret = 0;
a9869b83 1765
19fc3f0a 1766 /* Free the needed pages to the hugetlb pool */
e4e574b7 1767 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
19fc3f0a
AL
1768 if ((--needed) < 0)
1769 break;
a9869b83
NH
1770 /*
1771 * This page is now managed by the hugetlb allocator and has
1772 * no users -- drop the buddy allocator's reference.
1773 */
1774 put_page_testzero(page);
309381fe 1775 VM_BUG_ON_PAGE(page_count(page), page);
a5516438 1776 enqueue_huge_page(h, page);
19fc3f0a 1777 }
28073b02 1778free:
b0365c8d 1779 spin_unlock(&hugetlb_lock);
19fc3f0a
AL
1780
1781 /* Free unnecessary surplus pages to the buddy allocator */
c0d934ba
JK
1782 list_for_each_entry_safe(page, tmp, &surplus_list, lru)
1783 put_page(page);
a9869b83 1784 spin_lock(&hugetlb_lock);
e4e574b7
AL
1785
1786 return ret;
1787}
1788
1789/*
e5bbc8a6
MK
1790 * This routine has two main purposes:
1791 * 1) Decrement the reservation count (resv_huge_pages) by the value passed
1792 * in unused_resv_pages. This corresponds to the prior adjustments made
1793 * to the associated reservation map.
1794 * 2) Free any unused surplus pages that may have been allocated to satisfy
1795 * the reservation. As many as unused_resv_pages may be freed.
1796 *
1797 * Called with hugetlb_lock held. However, the lock could be dropped (and
1798 * reacquired) during calls to cond_resched_lock. Whenever dropping the lock,
1799 * we must make sure nobody else can claim pages we are in the process of
1800 * freeing. Do this by ensuring resv_huge_page always is greater than the
1801 * number of huge pages we plan to free when dropping the lock.
e4e574b7 1802 */
a5516438
AK
1803static void return_unused_surplus_pages(struct hstate *h,
1804 unsigned long unused_resv_pages)
e4e574b7 1805{
e4e574b7
AL
1806 unsigned long nr_pages;
1807
aa888a74 1808 /* Cannot return gigantic pages currently */
bae7f4ae 1809 if (hstate_is_gigantic(h))
e5bbc8a6 1810 goto out;
aa888a74 1811
e5bbc8a6
MK
1812 /*
1813 * Part (or even all) of the reservation could have been backed
1814 * by pre-allocated pages. Only free surplus pages.
1815 */
a5516438 1816 nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
e4e574b7 1817
685f3457
LS
1818 /*
1819 * We want to release as many surplus pages as possible, spread
9b5e5d0f
LS
1820 * evenly across all nodes with memory. Iterate across these nodes
1821 * until we can no longer free unreserved surplus pages. This occurs
1822 * when the nodes with surplus pages have no free pages.
1823 * free_pool_huge_page() will balance the the freed pages across the
1824 * on-line nodes with memory and will handle the hstate accounting.
e5bbc8a6
MK
1825 *
1826 * Note that we decrement resv_huge_pages as we free the pages. If
1827 * we drop the lock, resv_huge_pages will still be sufficiently large
1828 * to cover subsequent pages we may free.
685f3457
LS
1829 */
1830 while (nr_pages--) {
e5bbc8a6
MK
1831 h->resv_huge_pages--;
1832 unused_resv_pages--;
8cebfcd0 1833 if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1))
e5bbc8a6 1834 goto out;
7848a4bf 1835 cond_resched_lock(&hugetlb_lock);
e4e574b7 1836 }
e5bbc8a6
MK
1837
1838out:
1839 /* Fully uncommit the reservation */
1840 h->resv_huge_pages -= unused_resv_pages;
e4e574b7
AL
1841}
1842
5e911373 1843
c37f9fb1 1844/*
feba16e2 1845 * vma_needs_reservation, vma_commit_reservation and vma_end_reservation
5e911373 1846 * are used by the huge page allocation routines to manage reservations.
cf3ad20b
MK
1847 *
1848 * vma_needs_reservation is called to determine if the huge page at addr
1849 * within the vma has an associated reservation. If a reservation is
1850 * needed, the value 1 is returned. The caller is then responsible for
1851 * managing the global reservation and subpool usage counts. After
1852 * the huge page has been allocated, vma_commit_reservation is called
feba16e2
MK
1853 * to add the page to the reservation map. If the page allocation fails,
1854 * the reservation must be ended instead of committed. vma_end_reservation
1855 * is called in such cases.
cf3ad20b
MK
1856 *
1857 * In the normal case, vma_commit_reservation returns the same value
1858 * as the preceding vma_needs_reservation call. The only time this
1859 * is not the case is if a reserve map was changed between calls. It
1860 * is the responsibility of the caller to notice the difference and
1861 * take appropriate action.
96b96a96
MK
1862 *
1863 * vma_add_reservation is used in error paths where a reservation must
1864 * be restored when a newly allocated huge page must be freed. It is
1865 * to be called after calling vma_needs_reservation to determine if a
1866 * reservation exists.
c37f9fb1 1867 */
5e911373
MK
1868enum vma_resv_mode {
1869 VMA_NEEDS_RESV,
1870 VMA_COMMIT_RESV,
feba16e2 1871 VMA_END_RESV,
96b96a96 1872 VMA_ADD_RESV,
5e911373 1873};
cf3ad20b
MK
1874static long __vma_reservation_common(struct hstate *h,
1875 struct vm_area_struct *vma, unsigned long addr,
5e911373 1876 enum vma_resv_mode mode)
c37f9fb1 1877{
4e35f483
JK
1878 struct resv_map *resv;
1879 pgoff_t idx;
cf3ad20b 1880 long ret;
c37f9fb1 1881
4e35f483
JK
1882 resv = vma_resv_map(vma);
1883 if (!resv)
84afd99b 1884 return 1;
c37f9fb1 1885
4e35f483 1886 idx = vma_hugecache_offset(h, vma, addr);
5e911373
MK
1887 switch (mode) {
1888 case VMA_NEEDS_RESV:
cf3ad20b 1889 ret = region_chg(resv, idx, idx + 1);
5e911373
MK
1890 break;
1891 case VMA_COMMIT_RESV:
1892 ret = region_add(resv, idx, idx + 1);
1893 break;
feba16e2 1894 case VMA_END_RESV:
5e911373
MK
1895 region_abort(resv, idx, idx + 1);
1896 ret = 0;
1897 break;
96b96a96
MK
1898 case VMA_ADD_RESV:
1899 if (vma->vm_flags & VM_MAYSHARE)
1900 ret = region_add(resv, idx, idx + 1);
1901 else {
1902 region_abort(resv, idx, idx + 1);
1903 ret = region_del(resv, idx, idx + 1);
1904 }
1905 break;
5e911373
MK
1906 default:
1907 BUG();
1908 }
84afd99b 1909
4e35f483 1910 if (vma->vm_flags & VM_MAYSHARE)
cf3ad20b 1911 return ret;
67961f9d
MK
1912 else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && ret >= 0) {
1913 /*
1914 * In most cases, reserves always exist for private mappings.
1915 * However, a file associated with mapping could have been
1916 * hole punched or truncated after reserves were consumed.
1917 * As subsequent fault on such a range will not use reserves.
1918 * Subtle - The reserve map for private mappings has the
1919 * opposite meaning than that of shared mappings. If NO
1920 * entry is in the reserve map, it means a reservation exists.
1921 * If an entry exists in the reserve map, it means the
1922 * reservation has already been consumed. As a result, the
1923 * return value of this routine is the opposite of the
1924 * value returned from reserve map manipulation routines above.
1925 */
1926 if (ret)
1927 return 0;
1928 else
1929 return 1;
1930 }
4e35f483 1931 else
cf3ad20b 1932 return ret < 0 ? ret : 0;
c37f9fb1 1933}
cf3ad20b
MK
1934
1935static long vma_needs_reservation(struct hstate *h,
a5516438 1936 struct vm_area_struct *vma, unsigned long addr)
c37f9fb1 1937{
5e911373 1938 return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV);
cf3ad20b 1939}
84afd99b 1940
cf3ad20b
MK
1941static long vma_commit_reservation(struct hstate *h,
1942 struct vm_area_struct *vma, unsigned long addr)
1943{
5e911373
MK
1944 return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV);
1945}
1946
feba16e2 1947static void vma_end_reservation(struct hstate *h,
5e911373
MK
1948 struct vm_area_struct *vma, unsigned long addr)
1949{
feba16e2 1950 (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV);
c37f9fb1
AW
1951}
1952
96b96a96
MK
1953static long vma_add_reservation(struct hstate *h,
1954 struct vm_area_struct *vma, unsigned long addr)
1955{
1956 return __vma_reservation_common(h, vma, addr, VMA_ADD_RESV);
1957}
1958
1959/*
1960 * This routine is called to restore a reservation on error paths. In the
1961 * specific error paths, a huge page was allocated (via alloc_huge_page)
1962 * and is about to be freed. If a reservation for the page existed,
1963 * alloc_huge_page would have consumed the reservation and set PagePrivate
1964 * in the newly allocated page. When the page is freed via free_huge_page,
1965 * the global reservation count will be incremented if PagePrivate is set.
1966 * However, free_huge_page can not adjust the reserve map. Adjust the
1967 * reserve map here to be consistent with global reserve count adjustments
1968 * to be made by free_huge_page.
1969 */
1970static void restore_reserve_on_error(struct hstate *h,
1971 struct vm_area_struct *vma, unsigned long address,
1972 struct page *page)
1973{
1974 if (unlikely(PagePrivate(page))) {
1975 long rc = vma_needs_reservation(h, vma, address);
1976
1977 if (unlikely(rc < 0)) {
1978 /*
1979 * Rare out of memory condition in reserve map
1980 * manipulation. Clear PagePrivate so that
1981 * global reserve count will not be incremented
1982 * by free_huge_page. This will make it appear
1983 * as though the reservation for this page was
1984 * consumed. This may prevent the task from
1985 * faulting in the page at a later time. This
1986 * is better than inconsistent global huge page
1987 * accounting of reserve counts.
1988 */
1989 ClearPagePrivate(page);
1990 } else if (rc) {
1991 rc = vma_add_reservation(h, vma, address);
1992 if (unlikely(rc < 0))
1993 /*
1994 * See above comment about rare out of
1995 * memory condition.
1996 */
1997 ClearPagePrivate(page);
1998 } else
1999 vma_end_reservation(h, vma, address);
2000 }
2001}
2002
70c3547e 2003struct page *alloc_huge_page(struct vm_area_struct *vma,
04f2cbe3 2004 unsigned long addr, int avoid_reserve)
1da177e4 2005{
90481622 2006 struct hugepage_subpool *spool = subpool_vma(vma);
a5516438 2007 struct hstate *h = hstate_vma(vma);
348ea204 2008 struct page *page;
d85f69b0
MK
2009 long map_chg, map_commit;
2010 long gbl_chg;
6d76dcf4
AK
2011 int ret, idx;
2012 struct hugetlb_cgroup *h_cg;
a1e78772 2013
6d76dcf4 2014 idx = hstate_index(h);
a1e78772 2015 /*
d85f69b0
MK
2016 * Examine the region/reserve map to determine if the process
2017 * has a reservation for the page to be allocated. A return
2018 * code of zero indicates a reservation exists (no change).
a1e78772 2019 */
d85f69b0
MK
2020 map_chg = gbl_chg = vma_needs_reservation(h, vma, addr);
2021 if (map_chg < 0)
76dcee75 2022 return ERR_PTR(-ENOMEM);
d85f69b0
MK
2023
2024 /*
2025 * Processes that did not create the mapping will have no
2026 * reserves as indicated by the region/reserve map. Check
2027 * that the allocation will not exceed the subpool limit.
2028 * Allocations for MAP_NORESERVE mappings also need to be
2029 * checked against any subpool limit.
2030 */
2031 if (map_chg || avoid_reserve) {
2032 gbl_chg = hugepage_subpool_get_pages(spool, 1);
2033 if (gbl_chg < 0) {
feba16e2 2034 vma_end_reservation(h, vma, addr);
76dcee75 2035 return ERR_PTR(-ENOSPC);
5e911373 2036 }
1da177e4 2037
d85f69b0
MK
2038 /*
2039 * Even though there was no reservation in the region/reserve
2040 * map, there could be reservations associated with the
2041 * subpool that can be used. This would be indicated if the
2042 * return value of hugepage_subpool_get_pages() is zero.
2043 * However, if avoid_reserve is specified we still avoid even
2044 * the subpool reservations.
2045 */
2046 if (avoid_reserve)
2047 gbl_chg = 1;
2048 }
2049
6d76dcf4 2050 ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg);
8f34af6f
JZ
2051 if (ret)
2052 goto out_subpool_put;
2053
1da177e4 2054 spin_lock(&hugetlb_lock);
d85f69b0
MK
2055 /*
2056 * glb_chg is passed to indicate whether or not a page must be taken
2057 * from the global free pool (global change). gbl_chg == 0 indicates
2058 * a reservation exists for the allocation.
2059 */
2060 page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg);
81a6fcae 2061 if (!page) {
94ae8ba7 2062 spin_unlock(&hugetlb_lock);
0c397dae 2063 page = alloc_buddy_huge_page_with_mpol(h, vma, addr);
8f34af6f
JZ
2064 if (!page)
2065 goto out_uncharge_cgroup;
a88c7695
NH
2066 if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) {
2067 SetPagePrivate(page);
2068 h->resv_huge_pages--;
2069 }
79dbb236
AK
2070 spin_lock(&hugetlb_lock);
2071 list_move(&page->lru, &h->hugepage_activelist);
81a6fcae 2072 /* Fall through */
68842c9b 2073 }
81a6fcae
JK
2074 hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page);
2075 spin_unlock(&hugetlb_lock);
348ea204 2076
90481622 2077 set_page_private(page, (unsigned long)spool);
90d8b7e6 2078
d85f69b0
MK
2079 map_commit = vma_commit_reservation(h, vma, addr);
2080 if (unlikely(map_chg > map_commit)) {
33039678
MK
2081 /*
2082 * The page was added to the reservation map between
2083 * vma_needs_reservation and vma_commit_reservation.
2084 * This indicates a race with hugetlb_reserve_pages.
2085 * Adjust for the subpool count incremented above AND
2086 * in hugetlb_reserve_pages for the same page. Also,
2087 * the reservation count added in hugetlb_reserve_pages
2088 * no longer applies.
2089 */
2090 long rsv_adjust;
2091
2092 rsv_adjust = hugepage_subpool_put_pages(spool, 1);
2093 hugetlb_acct_memory(h, -rsv_adjust);
2094 }
90d8b7e6 2095 return page;
8f34af6f
JZ
2096
2097out_uncharge_cgroup:
2098 hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg);
2099out_subpool_put:
d85f69b0 2100 if (map_chg || avoid_reserve)
8f34af6f 2101 hugepage_subpool_put_pages(spool, 1);
feba16e2 2102 vma_end_reservation(h, vma, addr);
8f34af6f 2103 return ERR_PTR(-ENOSPC);
b45b5bd6
DG
2104}
2105
e24a1307
AK
2106int alloc_bootmem_huge_page(struct hstate *h)
2107 __attribute__ ((weak, alias("__alloc_bootmem_huge_page")));
2108int __alloc_bootmem_huge_page(struct hstate *h)
aa888a74
AK
2109{
2110 struct huge_bootmem_page *m;
b2261026 2111 int nr_nodes, node;
aa888a74 2112
b2261026 2113 for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) {
aa888a74
AK
2114 void *addr;
2115
eb31d559 2116 addr = memblock_alloc_try_nid_raw(
8b89a116 2117 huge_page_size(h), huge_page_size(h),
97ad1087 2118 0, MEMBLOCK_ALLOC_ACCESSIBLE, node);
aa888a74
AK
2119 if (addr) {
2120 /*
2121 * Use the beginning of the huge page to store the
2122 * huge_bootmem_page struct (until gather_bootmem
2123 * puts them into the mem_map).
2124 */
2125 m = addr;
91f47662 2126 goto found;
aa888a74 2127 }
aa888a74
AK
2128 }
2129 return 0;
2130
2131found:
df994ead 2132 BUG_ON(!IS_ALIGNED(virt_to_phys(m), huge_page_size(h)));
aa888a74 2133 /* Put them into a private list first because mem_map is not up yet */
330d6e48 2134 INIT_LIST_HEAD(&m->list);
aa888a74
AK
2135 list_add(&m->list, &huge_boot_pages);
2136 m->hstate = h;
2137 return 1;
2138}
2139
d00181b9
KS
2140static void __init prep_compound_huge_page(struct page *page,
2141 unsigned int order)
18229df5
AW
2142{
2143 if (unlikely(order > (MAX_ORDER - 1)))
2144 prep_compound_gigantic_page(page, order);
2145 else
2146 prep_compound_page(page, order);
2147}
2148
aa888a74
AK
2149/* Put bootmem huge pages into the standard lists after mem_map is up */
2150static void __init gather_bootmem_prealloc(void)
2151{
2152 struct huge_bootmem_page *m;
2153
2154 list_for_each_entry(m, &huge_boot_pages, list) {
40d18ebf 2155 struct page *page = virt_to_page(m);
aa888a74 2156 struct hstate *h = m->hstate;
ee8f248d 2157
aa888a74 2158 WARN_ON(page_count(page) != 1);
18229df5 2159 prep_compound_huge_page(page, h->order);
ef5a22be 2160 WARN_ON(PageReserved(page));
aa888a74 2161 prep_new_huge_page(h, page, page_to_nid(page));
af0fb9df
MH
2162 put_page(page); /* free it into the hugepage allocator */
2163
b0320c7b
RA
2164 /*
2165 * If we had gigantic hugepages allocated at boot time, we need
2166 * to restore the 'stolen' pages to totalram_pages in order to
2167 * fix confusing memory reports from free(1) and another
2168 * side-effects, like CommitLimit going negative.
2169 */
bae7f4ae 2170 if (hstate_is_gigantic(h))
3dcc0571 2171 adjust_managed_page_count(page, 1 << h->order);
520495fe 2172 cond_resched();
aa888a74
AK
2173 }
2174}
2175
8faa8b07 2176static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
1da177e4
LT
2177{
2178 unsigned long i;
a5516438 2179
e5ff2159 2180 for (i = 0; i < h->max_huge_pages; ++i) {
bae7f4ae 2181 if (hstate_is_gigantic(h)) {
aa888a74
AK
2182 if (!alloc_bootmem_huge_page(h))
2183 break;
0c397dae 2184 } else if (!alloc_pool_huge_page(h,
8cebfcd0 2185 &node_states[N_MEMORY]))
1da177e4 2186 break;
69ed779a 2187 cond_resched();
1da177e4 2188 }
d715cf80
LH
2189 if (i < h->max_huge_pages) {
2190 char buf[32];
2191
c6247f72 2192 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
d715cf80
LH
2193 pr_warn("HugeTLB: allocating %lu of page size %s failed. Only allocated %lu hugepages.\n",
2194 h->max_huge_pages, buf, i);
2195 h->max_huge_pages = i;
2196 }
e5ff2159
AK
2197}
2198
2199static void __init hugetlb_init_hstates(void)
2200{
2201 struct hstate *h;
2202
2203 for_each_hstate(h) {
641844f5
NH
2204 if (minimum_order > huge_page_order(h))
2205 minimum_order = huge_page_order(h);
2206
8faa8b07 2207 /* oversize hugepages were init'ed in early boot */
bae7f4ae 2208 if (!hstate_is_gigantic(h))
8faa8b07 2209 hugetlb_hstate_alloc_pages(h);
e5ff2159 2210 }
641844f5 2211 VM_BUG_ON(minimum_order == UINT_MAX);
e5ff2159
AK
2212}
2213
2214static void __init report_hugepages(void)
2215{
2216 struct hstate *h;
2217
2218 for_each_hstate(h) {
4abd32db 2219 char buf[32];
c6247f72
MW
2220
2221 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
ffb22af5 2222 pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n",
c6247f72 2223 buf, h->free_huge_pages);
e5ff2159
AK
2224 }
2225}
2226
1da177e4 2227#ifdef CONFIG_HIGHMEM
6ae11b27
LS
2228static void try_to_free_low(struct hstate *h, unsigned long count,
2229 nodemask_t *nodes_allowed)
1da177e4 2230{
4415cc8d
CL
2231 int i;
2232
bae7f4ae 2233 if (hstate_is_gigantic(h))
aa888a74
AK
2234 return;
2235
6ae11b27 2236 for_each_node_mask(i, *nodes_allowed) {
1da177e4 2237 struct page *page, *next;
a5516438
AK
2238 struct list_head *freel = &h->hugepage_freelists[i];
2239 list_for_each_entry_safe(page, next, freel, lru) {
2240 if (count >= h->nr_huge_pages)
6b0c880d 2241 return;
1da177e4
LT
2242 if (PageHighMem(page))
2243 continue;
2244 list_del(&page->lru);
e5ff2159 2245 update_and_free_page(h, page);
a5516438
AK
2246 h->free_huge_pages--;
2247 h->free_huge_pages_node[page_to_nid(page)]--;
1da177e4
LT
2248 }
2249 }
2250}
2251#else
6ae11b27
LS
2252static inline void try_to_free_low(struct hstate *h, unsigned long count,
2253 nodemask_t *nodes_allowed)
1da177e4
LT
2254{
2255}
2256#endif
2257
20a0307c
WF
2258/*
2259 * Increment or decrement surplus_huge_pages. Keep node-specific counters
2260 * balanced by operating on them in a round-robin fashion.
2261 * Returns 1 if an adjustment was made.
2262 */
6ae11b27
LS
2263static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
2264 int delta)
20a0307c 2265{
b2261026 2266 int nr_nodes, node;
20a0307c
WF
2267
2268 VM_BUG_ON(delta != -1 && delta != 1);
20a0307c 2269
b2261026
JK
2270 if (delta < 0) {
2271 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
2272 if (h->surplus_huge_pages_node[node])
2273 goto found;
e8c5c824 2274 }
b2261026
JK
2275 } else {
2276 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
2277 if (h->surplus_huge_pages_node[node] <
2278 h->nr_huge_pages_node[node])
2279 goto found;
e8c5c824 2280 }
b2261026
JK
2281 }
2282 return 0;
20a0307c 2283
b2261026
JK
2284found:
2285 h->surplus_huge_pages += delta;
2286 h->surplus_huge_pages_node[node] += delta;
2287 return 1;
20a0307c
WF
2288}
2289
a5516438 2290#define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
4eb0716e
AG
2291static int set_max_huge_pages(struct hstate *h, unsigned long count,
2292 nodemask_t *nodes_allowed)
1da177e4 2293{
7893d1d5 2294 unsigned long min_count, ret;
1da177e4 2295
4eb0716e
AG
2296 spin_lock(&hugetlb_lock);
2297
2298 /*
2299 * Gigantic pages runtime allocation depend on the capability for large
2300 * page range allocation.
2301 * If the system does not provide this feature, return an error when
2302 * the user tries to allocate gigantic pages but let the user free the
2303 * boottime allocated gigantic pages.
2304 */
2305 if (hstate_is_gigantic(h) && !IS_ENABLED(CONFIG_CONTIG_ALLOC)) {
2306 if (count > persistent_huge_pages(h)) {
2307 spin_unlock(&hugetlb_lock);
2308 return -EINVAL;
2309 }
2310 /* Fall through to decrease pool */
2311 }
aa888a74 2312
7893d1d5
AL
2313 /*
2314 * Increase the pool size
2315 * First take pages out of surplus state. Then make up the
2316 * remaining difference by allocating fresh huge pages.
d1c3fb1f 2317 *
0c397dae 2318 * We might race with alloc_surplus_huge_page() here and be unable
d1c3fb1f
NA
2319 * to convert a surplus huge page to a normal huge page. That is
2320 * not critical, though, it just means the overall size of the
2321 * pool might be one hugepage larger than it needs to be, but
2322 * within all the constraints specified by the sysctls.
7893d1d5 2323 */
a5516438 2324 while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
6ae11b27 2325 if (!adjust_pool_surplus(h, nodes_allowed, -1))
7893d1d5
AL
2326 break;
2327 }
2328
a5516438 2329 while (count > persistent_huge_pages(h)) {
7893d1d5
AL
2330 /*
2331 * If this allocation races such that we no longer need the
2332 * page, free_huge_page will handle it by freeing the page
2333 * and reducing the surplus.
2334 */
2335 spin_unlock(&hugetlb_lock);
649920c6
JH
2336
2337 /* yield cpu to avoid soft lockup */
2338 cond_resched();
2339
0c397dae 2340 ret = alloc_pool_huge_page(h, nodes_allowed);
7893d1d5
AL
2341 spin_lock(&hugetlb_lock);
2342 if (!ret)
2343 goto out;
2344
536240f2
MG
2345 /* Bail for signals. Probably ctrl-c from user */
2346 if (signal_pending(current))
2347 goto out;
7893d1d5 2348 }
7893d1d5
AL
2349
2350 /*
2351 * Decrease the pool size
2352 * First return free pages to the buddy allocator (being careful
2353 * to keep enough around to satisfy reservations). Then place
2354 * pages into surplus state as needed so the pool will shrink
2355 * to the desired size as pages become free.
d1c3fb1f
NA
2356 *
2357 * By placing pages into the surplus state independent of the
2358 * overcommit value, we are allowing the surplus pool size to
2359 * exceed overcommit. There are few sane options here. Since
0c397dae 2360 * alloc_surplus_huge_page() is checking the global counter,
d1c3fb1f
NA
2361 * though, we'll note that we're not allowed to exceed surplus
2362 * and won't grow the pool anywhere else. Not until one of the
2363 * sysctls are changed, or the surplus pages go out of use.
7893d1d5 2364 */
a5516438 2365 min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
6b0c880d 2366 min_count = max(count, min_count);
6ae11b27 2367 try_to_free_low(h, min_count, nodes_allowed);
a5516438 2368 while (min_count < persistent_huge_pages(h)) {
6ae11b27 2369 if (!free_pool_huge_page(h, nodes_allowed, 0))
1da177e4 2370 break;
55f67141 2371 cond_resched_lock(&hugetlb_lock);
1da177e4 2372 }
a5516438 2373 while (count < persistent_huge_pages(h)) {
6ae11b27 2374 if (!adjust_pool_surplus(h, nodes_allowed, 1))
7893d1d5
AL
2375 break;
2376 }
2377out:
4eb0716e 2378 h->max_huge_pages = persistent_huge_pages(h);
1da177e4 2379 spin_unlock(&hugetlb_lock);
4eb0716e
AG
2380
2381 return 0;
1da177e4
LT
2382}
2383
a3437870
NA
2384#define HSTATE_ATTR_RO(_name) \
2385 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
2386
2387#define HSTATE_ATTR(_name) \
2388 static struct kobj_attribute _name##_attr = \
2389 __ATTR(_name, 0644, _name##_show, _name##_store)
2390
2391static struct kobject *hugepages_kobj;
2392static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
2393
9a305230
LS
2394static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
2395
2396static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
a3437870
NA
2397{
2398 int i;
9a305230 2399
a3437870 2400 for (i = 0; i < HUGE_MAX_HSTATE; i++)
9a305230
LS
2401 if (hstate_kobjs[i] == kobj) {
2402 if (nidp)
2403 *nidp = NUMA_NO_NODE;
a3437870 2404 return &hstates[i];
9a305230
LS
2405 }
2406
2407 return kobj_to_node_hstate(kobj, nidp);
a3437870
NA
2408}
2409
06808b08 2410static ssize_t nr_hugepages_show_common(struct kobject *kobj,
a3437870
NA
2411 struct kobj_attribute *attr, char *buf)
2412{
9a305230
LS
2413 struct hstate *h;
2414 unsigned long nr_huge_pages;
2415 int nid;
2416
2417 h = kobj_to_hstate(kobj, &nid);
2418 if (nid == NUMA_NO_NODE)
2419 nr_huge_pages = h->nr_huge_pages;
2420 else
2421 nr_huge_pages = h->nr_huge_pages_node[nid];
2422
2423 return sprintf(buf, "%lu\n", nr_huge_pages);
a3437870 2424}
adbe8726 2425
238d3c13
DR
2426static ssize_t __nr_hugepages_store_common(bool obey_mempolicy,
2427 struct hstate *h, int nid,
2428 unsigned long count, size_t len)
a3437870
NA
2429{
2430 int err;
bad44b5b 2431 NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY);
a3437870 2432
4eb0716e 2433 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) {
adbe8726
EM
2434 err = -EINVAL;
2435 goto out;
2436 }
2437
9a305230
LS
2438 if (nid == NUMA_NO_NODE) {
2439 /*
2440 * global hstate attribute
2441 */
2442 if (!(obey_mempolicy &&
2443 init_nodemask_of_mempolicy(nodes_allowed))) {
2444 NODEMASK_FREE(nodes_allowed);
8cebfcd0 2445 nodes_allowed = &node_states[N_MEMORY];
9a305230
LS
2446 }
2447 } else if (nodes_allowed) {
2448 /*
2449 * per node hstate attribute: adjust count to global,
2450 * but restrict alloc/free to the specified node.
2451 */
2452 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
2453 init_nodemask_of_node(nodes_allowed, nid);
2454 } else
8cebfcd0 2455 nodes_allowed = &node_states[N_MEMORY];
9a305230 2456
4eb0716e 2457 err = set_max_huge_pages(h, count, nodes_allowed);
a3437870 2458
4eb0716e 2459out:
8cebfcd0 2460 if (nodes_allowed != &node_states[N_MEMORY])
06808b08
LS
2461 NODEMASK_FREE(nodes_allowed);
2462
4eb0716e 2463 return err ? err : len;
06808b08
LS
2464}
2465
238d3c13
DR
2466static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
2467 struct kobject *kobj, const char *buf,
2468 size_t len)
2469{
2470 struct hstate *h;
2471 unsigned long count;
2472 int nid;
2473 int err;
2474
2475 err = kstrtoul(buf, 10, &count);
2476 if (err)
2477 return err;
2478
2479 h = kobj_to_hstate(kobj, &nid);
2480 return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len);
2481}
2482
06808b08
LS
2483static ssize_t nr_hugepages_show(struct kobject *kobj,
2484 struct kobj_attribute *attr, char *buf)
2485{
2486 return nr_hugepages_show_common(kobj, attr, buf);
2487}
2488
2489static ssize_t nr_hugepages_store(struct kobject *kobj,
2490 struct kobj_attribute *attr, const char *buf, size_t len)
2491{
238d3c13 2492 return nr_hugepages_store_common(false, kobj, buf, len);
a3437870
NA
2493}
2494HSTATE_ATTR(nr_hugepages);
2495
06808b08
LS
2496#ifdef CONFIG_NUMA
2497
2498/*
2499 * hstate attribute for optionally mempolicy-based constraint on persistent
2500 * huge page alloc/free.
2501 */
2502static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
2503 struct kobj_attribute *attr, char *buf)
2504{
2505 return nr_hugepages_show_common(kobj, attr, buf);
2506}
2507
2508static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
2509 struct kobj_attribute *attr, const char *buf, size_t len)
2510{
238d3c13 2511 return nr_hugepages_store_common(true, kobj, buf, len);
06808b08
LS
2512}
2513HSTATE_ATTR(nr_hugepages_mempolicy);
2514#endif
2515
2516
a3437870
NA
2517static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
2518 struct kobj_attribute *attr, char *buf)
2519{
9a305230 2520 struct hstate *h = kobj_to_hstate(kobj, NULL);
a3437870
NA
2521 return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
2522}
adbe8726 2523
a3437870
NA
2524static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
2525 struct kobj_attribute *attr, const char *buf, size_t count)
2526{
2527 int err;
2528 unsigned long input;
9a305230 2529 struct hstate *h = kobj_to_hstate(kobj, NULL);
a3437870 2530
bae7f4ae 2531 if (hstate_is_gigantic(h))
adbe8726
EM
2532 return -EINVAL;
2533
3dbb95f7 2534 err = kstrtoul(buf, 10, &input);
a3437870 2535 if (err)
73ae31e5 2536 return err;
a3437870
NA
2537
2538 spin_lock(&hugetlb_lock);
2539 h->nr_overcommit_huge_pages = input;
2540 spin_unlock(&hugetlb_lock);
2541
2542 return count;
2543}
2544HSTATE_ATTR(nr_overcommit_hugepages);
2545
2546static ssize_t free_hugepages_show(struct kobject *kobj,
2547 struct kobj_attribute *attr, char *buf)
2548{
9a305230
LS
2549 struct hstate *h;
2550 unsigned long free_huge_pages;
2551 int nid;
2552
2553 h = kobj_to_hstate(kobj, &nid);
2554 if (nid == NUMA_NO_NODE)
2555 free_huge_pages = h->free_huge_pages;
2556 else
2557 free_huge_pages = h->free_huge_pages_node[nid];
2558
2559 return sprintf(buf, "%lu\n", free_huge_pages);
a3437870
NA
2560}
2561HSTATE_ATTR_RO(free_hugepages);
2562
2563static ssize_t resv_hugepages_show(struct kobject *kobj,
2564 struct kobj_attribute *attr, char *buf)
2565{
9a305230 2566 struct hstate *h = kobj_to_hstate(kobj, NULL);
a3437870
NA
2567 return sprintf(buf, "%lu\n", h->resv_huge_pages);
2568}
2569HSTATE_ATTR_RO(resv_hugepages);
2570
2571static ssize_t surplus_hugepages_show(struct kobject *kobj,
2572 struct kobj_attribute *attr, char *buf)
2573{
9a305230
LS
2574 struct hstate *h;
2575 unsigned long surplus_huge_pages;
2576 int nid;
2577
2578 h = kobj_to_hstate(kobj, &nid);
2579 if (nid == NUMA_NO_NODE)
2580 surplus_huge_pages = h->surplus_huge_pages;
2581 else
2582 surplus_huge_pages = h->surplus_huge_pages_node[nid];
2583
2584 return sprintf(buf, "%lu\n", surplus_huge_pages);
a3437870
NA
2585}
2586HSTATE_ATTR_RO(surplus_hugepages);
2587
2588static struct attribute *hstate_attrs[] = {
2589 &nr_hugepages_attr.attr,
2590 &nr_overcommit_hugepages_attr.attr,
2591 &free_hugepages_attr.attr,
2592 &resv_hugepages_attr.attr,
2593 &surplus_hugepages_attr.attr,
06808b08
LS
2594#ifdef CONFIG_NUMA
2595 &nr_hugepages_mempolicy_attr.attr,
2596#endif
a3437870
NA
2597 NULL,
2598};
2599
67e5ed96 2600static const struct attribute_group hstate_attr_group = {
a3437870
NA
2601 .attrs = hstate_attrs,
2602};
2603
094e9539
JM
2604static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
2605 struct kobject **hstate_kobjs,
67e5ed96 2606 const struct attribute_group *hstate_attr_group)
a3437870
NA
2607{
2608 int retval;
972dc4de 2609 int hi = hstate_index(h);
a3437870 2610
9a305230
LS
2611 hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
2612 if (!hstate_kobjs[hi])
a3437870
NA
2613 return -ENOMEM;
2614
9a305230 2615 retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
a3437870 2616 if (retval)
9a305230 2617 kobject_put(hstate_kobjs[hi]);
a3437870
NA
2618
2619 return retval;
2620}
2621
2622static void __init hugetlb_sysfs_init(void)
2623{
2624 struct hstate *h;
2625 int err;
2626
2627 hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
2628 if (!hugepages_kobj)
2629 return;
2630
2631 for_each_hstate(h) {
9a305230
LS
2632 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
2633 hstate_kobjs, &hstate_attr_group);
a3437870 2634 if (err)
ffb22af5 2635 pr_err("Hugetlb: Unable to add hstate %s", h->name);
a3437870
NA
2636 }
2637}
2638
9a305230
LS
2639#ifdef CONFIG_NUMA
2640
2641/*
2642 * node_hstate/s - associate per node hstate attributes, via their kobjects,
10fbcf4c
KS
2643 * with node devices in node_devices[] using a parallel array. The array
2644 * index of a node device or _hstate == node id.
2645 * This is here to avoid any static dependency of the node device driver, in
9a305230
LS
2646 * the base kernel, on the hugetlb module.
2647 */
2648struct node_hstate {
2649 struct kobject *hugepages_kobj;
2650 struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
2651};
b4e289a6 2652static struct node_hstate node_hstates[MAX_NUMNODES];
9a305230
LS
2653
2654/*
10fbcf4c 2655 * A subset of global hstate attributes for node devices
9a305230
LS
2656 */
2657static struct attribute *per_node_hstate_attrs[] = {
2658 &nr_hugepages_attr.attr,
2659 &free_hugepages_attr.attr,
2660 &surplus_hugepages_attr.attr,
2661 NULL,
2662};
2663
67e5ed96 2664static const struct attribute_group per_node_hstate_attr_group = {
9a305230
LS
2665 .attrs = per_node_hstate_attrs,
2666};
2667
2668/*
10fbcf4c 2669 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
9a305230
LS
2670 * Returns node id via non-NULL nidp.
2671 */
2672static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
2673{
2674 int nid;
2675
2676 for (nid = 0; nid < nr_node_ids; nid++) {
2677 struct node_hstate *nhs = &node_hstates[nid];
2678 int i;
2679 for (i = 0; i < HUGE_MAX_HSTATE; i++)
2680 if (nhs->hstate_kobjs[i] == kobj) {
2681 if (nidp)
2682 *nidp = nid;
2683 return &hstates[i];
2684 }
2685 }
2686
2687 BUG();
2688 return NULL;
2689}
2690
2691/*
10fbcf4c 2692 * Unregister hstate attributes from a single node device.
9a305230
LS
2693 * No-op if no hstate attributes attached.
2694 */
3cd8b44f 2695static void hugetlb_unregister_node(struct node *node)
9a305230
LS
2696{
2697 struct hstate *h;
10fbcf4c 2698 struct node_hstate *nhs = &node_hstates[node->dev.id];
9a305230
LS
2699
2700 if (!nhs->hugepages_kobj)
9b5e5d0f 2701 return; /* no hstate attributes */
9a305230 2702
972dc4de
AK
2703 for_each_hstate(h) {
2704 int idx = hstate_index(h);
2705 if (nhs->hstate_kobjs[idx]) {
2706 kobject_put(nhs->hstate_kobjs[idx]);
2707 nhs->hstate_kobjs[idx] = NULL;
9a305230 2708 }
972dc4de 2709 }
9a305230
LS
2710
2711 kobject_put(nhs->hugepages_kobj);
2712 nhs->hugepages_kobj = NULL;
2713}
2714
9a305230
LS
2715
2716/*
10fbcf4c 2717 * Register hstate attributes for a single node device.
9a305230
LS
2718 * No-op if attributes already registered.
2719 */
3cd8b44f 2720static void hugetlb_register_node(struct node *node)
9a305230
LS
2721{
2722 struct hstate *h;
10fbcf4c 2723 struct node_hstate *nhs = &node_hstates[node->dev.id];
9a305230
LS
2724 int err;
2725
2726 if (nhs->hugepages_kobj)
2727 return; /* already allocated */
2728
2729 nhs->hugepages_kobj = kobject_create_and_add("hugepages",
10fbcf4c 2730 &node->dev.kobj);
9a305230
LS
2731 if (!nhs->hugepages_kobj)
2732 return;
2733
2734 for_each_hstate(h) {
2735 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
2736 nhs->hstate_kobjs,
2737 &per_node_hstate_attr_group);
2738 if (err) {
ffb22af5
AM
2739 pr_err("Hugetlb: Unable to add hstate %s for node %d\n",
2740 h->name, node->dev.id);
9a305230
LS
2741 hugetlb_unregister_node(node);
2742 break;
2743 }
2744 }
2745}
2746
2747/*
9b5e5d0f 2748 * hugetlb init time: register hstate attributes for all registered node
10fbcf4c
KS
2749 * devices of nodes that have memory. All on-line nodes should have
2750 * registered their associated device by this time.
9a305230 2751 */
7d9ca000 2752static void __init hugetlb_register_all_nodes(void)
9a305230
LS
2753{
2754 int nid;
2755
8cebfcd0 2756 for_each_node_state(nid, N_MEMORY) {
8732794b 2757 struct node *node = node_devices[nid];
10fbcf4c 2758 if (node->dev.id == nid)
9a305230
LS
2759 hugetlb_register_node(node);
2760 }
2761
2762 /*
10fbcf4c 2763 * Let the node device driver know we're here so it can
9a305230
LS
2764 * [un]register hstate attributes on node hotplug.
2765 */
2766 register_hugetlbfs_with_node(hugetlb_register_node,
2767 hugetlb_unregister_node);
2768}
2769#else /* !CONFIG_NUMA */
2770
2771static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
2772{
2773 BUG();
2774 if (nidp)
2775 *nidp = -1;
2776 return NULL;
2777}
2778
9a305230
LS
2779static void hugetlb_register_all_nodes(void) { }
2780
2781#endif
2782
a3437870
NA
2783static int __init hugetlb_init(void)
2784{
8382d914
DB
2785 int i;
2786
457c1b27 2787 if (!hugepages_supported())
0ef89d25 2788 return 0;
a3437870 2789
e11bfbfc 2790 if (!size_to_hstate(default_hstate_size)) {
d715cf80
LH
2791 if (default_hstate_size != 0) {
2792 pr_err("HugeTLB: unsupported default_hugepagesz %lu. Reverting to %lu\n",
2793 default_hstate_size, HPAGE_SIZE);
2794 }
2795
e11bfbfc
NP
2796 default_hstate_size = HPAGE_SIZE;
2797 if (!size_to_hstate(default_hstate_size))
2798 hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
a3437870 2799 }
972dc4de 2800 default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size));
f8b74815
VT
2801 if (default_hstate_max_huge_pages) {
2802 if (!default_hstate.max_huge_pages)
2803 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
2804 }
a3437870
NA
2805
2806 hugetlb_init_hstates();
aa888a74 2807 gather_bootmem_prealloc();
a3437870
NA
2808 report_hugepages();
2809
2810 hugetlb_sysfs_init();
9a305230 2811 hugetlb_register_all_nodes();
7179e7bf 2812 hugetlb_cgroup_file_init();
9a305230 2813
8382d914
DB
2814#ifdef CONFIG_SMP
2815 num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus());
2816#else
2817 num_fault_mutexes = 1;
2818#endif
c672c7f2 2819 hugetlb_fault_mutex_table =
6da2ec56
KC
2820 kmalloc_array(num_fault_mutexes, sizeof(struct mutex),
2821 GFP_KERNEL);
c672c7f2 2822 BUG_ON(!hugetlb_fault_mutex_table);
8382d914
DB
2823
2824 for (i = 0; i < num_fault_mutexes; i++)
c672c7f2 2825 mutex_init(&hugetlb_fault_mutex_table[i]);
a3437870
NA
2826 return 0;
2827}
3e89e1c5 2828subsys_initcall(hugetlb_init);
a3437870
NA
2829
2830/* Should be called on processing a hugepagesz=... option */
9fee021d
VT
2831void __init hugetlb_bad_size(void)
2832{
2833 parsed_valid_hugepagesz = false;
2834}
2835
d00181b9 2836void __init hugetlb_add_hstate(unsigned int order)
a3437870
NA
2837{
2838 struct hstate *h;
8faa8b07
AK
2839 unsigned long i;
2840
a3437870 2841 if (size_to_hstate(PAGE_SIZE << order)) {
598d8091 2842 pr_warn("hugepagesz= specified twice, ignoring\n");
a3437870
NA
2843 return;
2844 }
47d38344 2845 BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE);
a3437870 2846 BUG_ON(order == 0);
47d38344 2847 h = &hstates[hugetlb_max_hstate++];
a3437870
NA
2848 h->order = order;
2849 h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
8faa8b07
AK
2850 h->nr_huge_pages = 0;
2851 h->free_huge_pages = 0;
2852 for (i = 0; i < MAX_NUMNODES; ++i)
2853 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
0edaecfa 2854 INIT_LIST_HEAD(&h->hugepage_activelist);
54f18d35
AM
2855 h->next_nid_to_alloc = first_memory_node;
2856 h->next_nid_to_free = first_memory_node;
a3437870
NA
2857 snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
2858 huge_page_size(h)/1024);
8faa8b07 2859
a3437870
NA
2860 parsed_hstate = h;
2861}
2862
e11bfbfc 2863static int __init hugetlb_nrpages_setup(char *s)
a3437870
NA
2864{
2865 unsigned long *mhp;
8faa8b07 2866 static unsigned long *last_mhp;
a3437870 2867
9fee021d
VT
2868 if (!parsed_valid_hugepagesz) {
2869 pr_warn("hugepages = %s preceded by "
2870 "an unsupported hugepagesz, ignoring\n", s);
2871 parsed_valid_hugepagesz = true;
2872 return 1;
2873 }
a3437870 2874 /*
47d38344 2875 * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet,
a3437870
NA
2876 * so this hugepages= parameter goes to the "default hstate".
2877 */
9fee021d 2878 else if (!hugetlb_max_hstate)
a3437870
NA
2879 mhp = &default_hstate_max_huge_pages;
2880 else
2881 mhp = &parsed_hstate->max_huge_pages;
2882
8faa8b07 2883 if (mhp == last_mhp) {
598d8091 2884 pr_warn("hugepages= specified twice without interleaving hugepagesz=, ignoring\n");
8faa8b07
AK
2885 return 1;
2886 }
2887
a3437870
NA
2888 if (sscanf(s, "%lu", mhp) <= 0)
2889 *mhp = 0;
2890
8faa8b07
AK
2891 /*
2892 * Global state is always initialized later in hugetlb_init.
2893 * But we need to allocate >= MAX_ORDER hstates here early to still
2894 * use the bootmem allocator.
2895 */
47d38344 2896 if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER)
8faa8b07
AK
2897 hugetlb_hstate_alloc_pages(parsed_hstate);
2898
2899 last_mhp = mhp;
2900
a3437870
NA
2901 return 1;
2902}
e11bfbfc
NP
2903__setup("hugepages=", hugetlb_nrpages_setup);
2904
2905static int __init hugetlb_default_setup(char *s)
2906{
2907 default_hstate_size = memparse(s, &s);
2908 return 1;
2909}
2910__setup("default_hugepagesz=", hugetlb_default_setup);
a3437870 2911
8a213460
NA
2912static unsigned int cpuset_mems_nr(unsigned int *array)
2913{
2914 int node;
2915 unsigned int nr = 0;
2916
2917 for_each_node_mask(node, cpuset_current_mems_allowed)
2918 nr += array[node];
2919
2920 return nr;
2921}
2922
2923#ifdef CONFIG_SYSCTL
06808b08
LS
2924static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
2925 struct ctl_table *table, int write,
2926 void __user *buffer, size_t *length, loff_t *ppos)
1da177e4 2927{
e5ff2159 2928 struct hstate *h = &default_hstate;
238d3c13 2929 unsigned long tmp = h->max_huge_pages;
08d4a246 2930 int ret;
e5ff2159 2931
457c1b27 2932 if (!hugepages_supported())
86613628 2933 return -EOPNOTSUPP;
457c1b27 2934
e5ff2159
AK
2935 table->data = &tmp;
2936 table->maxlen = sizeof(unsigned long);
08d4a246
MH
2937 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2938 if (ret)
2939 goto out;
e5ff2159 2940
238d3c13
DR
2941 if (write)
2942 ret = __nr_hugepages_store_common(obey_mempolicy, h,
2943 NUMA_NO_NODE, tmp, *length);
08d4a246
MH
2944out:
2945 return ret;
1da177e4 2946}
396faf03 2947
06808b08
LS
2948int hugetlb_sysctl_handler(struct ctl_table *table, int write,
2949 void __user *buffer, size_t *length, loff_t *ppos)
2950{
2951
2952 return hugetlb_sysctl_handler_common(false, table, write,
2953 buffer, length, ppos);
2954}
2955
2956#ifdef CONFIG_NUMA
2957int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
2958 void __user *buffer, size_t *length, loff_t *ppos)
2959{
2960 return hugetlb_sysctl_handler_common(true, table, write,
2961 buffer, length, ppos);
2962}
2963#endif /* CONFIG_NUMA */
2964
a3d0c6aa 2965int hugetlb_overcommit_handler(struct ctl_table *table, int write,
8d65af78 2966 void __user *buffer,
a3d0c6aa
NA
2967 size_t *length, loff_t *ppos)
2968{
a5516438 2969 struct hstate *h = &default_hstate;
e5ff2159 2970 unsigned long tmp;
08d4a246 2971 int ret;
e5ff2159 2972
457c1b27 2973 if (!hugepages_supported())
86613628 2974 return -EOPNOTSUPP;
457c1b27 2975
c033a93c 2976 tmp = h->nr_overcommit_huge_pages;
e5ff2159 2977
bae7f4ae 2978 if (write && hstate_is_gigantic(h))
adbe8726
EM
2979 return -EINVAL;
2980
e5ff2159
AK
2981 table->data = &tmp;
2982 table->maxlen = sizeof(unsigned long);
08d4a246
MH
2983 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2984 if (ret)
2985 goto out;
e5ff2159
AK
2986
2987 if (write) {
2988 spin_lock(&hugetlb_lock);
2989 h->nr_overcommit_huge_pages = tmp;
2990 spin_unlock(&hugetlb_lock);
2991 }
08d4a246
MH
2992out:
2993 return ret;
a3d0c6aa
NA
2994}
2995
1da177e4
LT
2996#endif /* CONFIG_SYSCTL */
2997
e1759c21 2998void hugetlb_report_meminfo(struct seq_file *m)
1da177e4 2999{
fcb2b0c5
RG
3000 struct hstate *h;
3001 unsigned long total = 0;
3002
457c1b27
NA
3003 if (!hugepages_supported())
3004 return;
fcb2b0c5
RG
3005
3006 for_each_hstate(h) {
3007 unsigned long count = h->nr_huge_pages;
3008
3009 total += (PAGE_SIZE << huge_page_order(h)) * count;
3010
3011 if (h == &default_hstate)
3012 seq_printf(m,
3013 "HugePages_Total: %5lu\n"
3014 "HugePages_Free: %5lu\n"
3015 "HugePages_Rsvd: %5lu\n"
3016 "HugePages_Surp: %5lu\n"
3017 "Hugepagesize: %8lu kB\n",
3018 count,
3019 h->free_huge_pages,
3020 h->resv_huge_pages,
3021 h->surplus_huge_pages,
3022 (PAGE_SIZE << huge_page_order(h)) / 1024);
3023 }
3024
3025 seq_printf(m, "Hugetlb: %8lu kB\n", total / 1024);
1da177e4
LT
3026}
3027
3028int hugetlb_report_node_meminfo(int nid, char *buf)
3029{
a5516438 3030 struct hstate *h = &default_hstate;
457c1b27
NA
3031 if (!hugepages_supported())
3032 return 0;
1da177e4
LT
3033 return sprintf(buf,
3034 "Node %d HugePages_Total: %5u\n"
a1de0919
NA
3035 "Node %d HugePages_Free: %5u\n"
3036 "Node %d HugePages_Surp: %5u\n",
a5516438
AK
3037 nid, h->nr_huge_pages_node[nid],
3038 nid, h->free_huge_pages_node[nid],
3039 nid, h->surplus_huge_pages_node[nid]);
1da177e4
LT
3040}
3041
949f7ec5
DR
3042void hugetlb_show_meminfo(void)
3043{
3044 struct hstate *h;
3045 int nid;
3046
457c1b27
NA
3047 if (!hugepages_supported())
3048 return;
3049
949f7ec5
DR
3050 for_each_node_state(nid, N_MEMORY)
3051 for_each_hstate(h)
3052 pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
3053 nid,
3054 h->nr_huge_pages_node[nid],
3055 h->free_huge_pages_node[nid],
3056 h->surplus_huge_pages_node[nid],
3057 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
3058}
3059
5d317b2b
NH
3060void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm)
3061{
3062 seq_printf(m, "HugetlbPages:\t%8lu kB\n",
3063 atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10));
3064}
3065
1da177e4
LT
3066/* Return the number pages of memory we physically have, in PAGE_SIZE units. */
3067unsigned long hugetlb_total_pages(void)
3068{
d0028588
WL
3069 struct hstate *h;
3070 unsigned long nr_total_pages = 0;
3071
3072 for_each_hstate(h)
3073 nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h);
3074 return nr_total_pages;
1da177e4 3075}
1da177e4 3076
a5516438 3077static int hugetlb_acct_memory(struct hstate *h, long delta)
fc1b8a73
MG
3078{
3079 int ret = -ENOMEM;
3080
3081 spin_lock(&hugetlb_lock);
3082 /*
3083 * When cpuset is configured, it breaks the strict hugetlb page
3084 * reservation as the accounting is done on a global variable. Such
3085 * reservation is completely rubbish in the presence of cpuset because
3086 * the reservation is not checked against page availability for the
3087 * current cpuset. Application can still potentially OOM'ed by kernel
3088 * with lack of free htlb page in cpuset that the task is in.
3089 * Attempt to enforce strict accounting with cpuset is almost
3090 * impossible (or too ugly) because cpuset is too fluid that
3091 * task or memory node can be dynamically moved between cpusets.
3092 *
3093 * The change of semantics for shared hugetlb mapping with cpuset is
3094 * undesirable. However, in order to preserve some of the semantics,
3095 * we fall back to check against current free page availability as
3096 * a best attempt and hopefully to minimize the impact of changing
3097 * semantics that cpuset has.
3098 */
3099 if (delta > 0) {
a5516438 3100 if (gather_surplus_pages(h, delta) < 0)
fc1b8a73
MG
3101 goto out;
3102
a5516438
AK
3103 if (delta > cpuset_mems_nr(h->free_huge_pages_node)) {
3104 return_unused_surplus_pages(h, delta);
fc1b8a73
MG
3105 goto out;
3106 }
3107 }
3108
3109 ret = 0;
3110 if (delta < 0)
a5516438 3111 return_unused_surplus_pages(h, (unsigned long) -delta);
fc1b8a73
MG
3112
3113out:
3114 spin_unlock(&hugetlb_lock);
3115 return ret;
3116}
3117
84afd99b
AW
3118static void hugetlb_vm_op_open(struct vm_area_struct *vma)
3119{
f522c3ac 3120 struct resv_map *resv = vma_resv_map(vma);
84afd99b
AW
3121
3122 /*
3123 * This new VMA should share its siblings reservation map if present.
3124 * The VMA will only ever have a valid reservation map pointer where
3125 * it is being copied for another still existing VMA. As that VMA
25985edc 3126 * has a reference to the reservation map it cannot disappear until
84afd99b
AW
3127 * after this open call completes. It is therefore safe to take a
3128 * new reference here without additional locking.
3129 */
4e35f483 3130 if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
f522c3ac 3131 kref_get(&resv->refs);
84afd99b
AW
3132}
3133
a1e78772
MG
3134static void hugetlb_vm_op_close(struct vm_area_struct *vma)
3135{
a5516438 3136 struct hstate *h = hstate_vma(vma);
f522c3ac 3137 struct resv_map *resv = vma_resv_map(vma);
90481622 3138 struct hugepage_subpool *spool = subpool_vma(vma);
4e35f483 3139 unsigned long reserve, start, end;
1c5ecae3 3140 long gbl_reserve;
84afd99b 3141
4e35f483
JK
3142 if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER))
3143 return;
84afd99b 3144
4e35f483
JK
3145 start = vma_hugecache_offset(h, vma, vma->vm_start);
3146 end = vma_hugecache_offset(h, vma, vma->vm_end);
84afd99b 3147
4e35f483 3148 reserve = (end - start) - region_count(resv, start, end);
84afd99b 3149
4e35f483
JK
3150 kref_put(&resv->refs, resv_map_release);
3151
3152 if (reserve) {
1c5ecae3
MK
3153 /*
3154 * Decrement reserve counts. The global reserve count may be
3155 * adjusted if the subpool has a minimum size.
3156 */
3157 gbl_reserve = hugepage_subpool_put_pages(spool, reserve);
3158 hugetlb_acct_memory(h, -gbl_reserve);
84afd99b 3159 }
a1e78772
MG
3160}
3161
31383c68
DW
3162static int hugetlb_vm_op_split(struct vm_area_struct *vma, unsigned long addr)
3163{
3164 if (addr & ~(huge_page_mask(hstate_vma(vma))))
3165 return -EINVAL;
3166 return 0;
3167}
3168
05ea8860
DW
3169static unsigned long hugetlb_vm_op_pagesize(struct vm_area_struct *vma)
3170{
3171 struct hstate *hstate = hstate_vma(vma);
3172
3173 return 1UL << huge_page_shift(hstate);
3174}
3175
1da177e4
LT
3176/*
3177 * We cannot handle pagefaults against hugetlb pages at all. They cause
3178 * handle_mm_fault() to try to instantiate regular-sized pages in the
3179 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
3180 * this far.
3181 */
b3ec9f33 3182static vm_fault_t hugetlb_vm_op_fault(struct vm_fault *vmf)
1da177e4
LT
3183{
3184 BUG();
d0217ac0 3185 return 0;
1da177e4
LT
3186}
3187
eec3636a
JC
3188/*
3189 * When a new function is introduced to vm_operations_struct and added
3190 * to hugetlb_vm_ops, please consider adding the function to shm_vm_ops.
3191 * This is because under System V memory model, mappings created via
3192 * shmget/shmat with "huge page" specified are backed by hugetlbfs files,
3193 * their original vm_ops are overwritten with shm_vm_ops.
3194 */
f0f37e2f 3195const struct vm_operations_struct hugetlb_vm_ops = {
d0217ac0 3196 .fault = hugetlb_vm_op_fault,
84afd99b 3197 .open = hugetlb_vm_op_open,
a1e78772 3198 .close = hugetlb_vm_op_close,
31383c68 3199 .split = hugetlb_vm_op_split,
05ea8860 3200 .pagesize = hugetlb_vm_op_pagesize,
1da177e4
LT
3201};
3202
1e8f889b
DG
3203static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
3204 int writable)
63551ae0
DG
3205{
3206 pte_t entry;
3207
1e8f889b 3208 if (writable) {
106c992a
GS
3209 entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page,
3210 vma->vm_page_prot)));
63551ae0 3211 } else {
106c992a
GS
3212 entry = huge_pte_wrprotect(mk_huge_pte(page,
3213 vma->vm_page_prot));
63551ae0
DG
3214 }
3215 entry = pte_mkyoung(entry);
3216 entry = pte_mkhuge(entry);
d9ed9faa 3217 entry = arch_make_huge_pte(entry, vma, page, writable);
63551ae0
DG
3218
3219 return entry;
3220}
3221
1e8f889b
DG
3222static void set_huge_ptep_writable(struct vm_area_struct *vma,
3223 unsigned long address, pte_t *ptep)
3224{
3225 pte_t entry;
3226
106c992a 3227 entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep)));
32f84528 3228 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
4b3073e1 3229 update_mmu_cache(vma, address, ptep);
1e8f889b
DG
3230}
3231
d5ed7444 3232bool is_hugetlb_entry_migration(pte_t pte)
4a705fef
NH
3233{
3234 swp_entry_t swp;
3235
3236 if (huge_pte_none(pte) || pte_present(pte))
d5ed7444 3237 return false;
4a705fef
NH
3238 swp = pte_to_swp_entry(pte);
3239 if (non_swap_entry(swp) && is_migration_entry(swp))
d5ed7444 3240 return true;
4a705fef 3241 else
d5ed7444 3242 return false;
4a705fef
NH
3243}
3244
3245static int is_hugetlb_entry_hwpoisoned(pte_t pte)
3246{
3247 swp_entry_t swp;
3248
3249 if (huge_pte_none(pte) || pte_present(pte))
3250 return 0;
3251 swp = pte_to_swp_entry(pte);
3252 if (non_swap_entry(swp) && is_hwpoison_entry(swp))
3253 return 1;
3254 else
3255 return 0;
3256}
1e8f889b 3257
63551ae0
DG
3258int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
3259 struct vm_area_struct *vma)
3260{
5e41540c 3261 pte_t *src_pte, *dst_pte, entry, dst_entry;
63551ae0 3262 struct page *ptepage;
1c59827d 3263 unsigned long addr;
1e8f889b 3264 int cow;
a5516438
AK
3265 struct hstate *h = hstate_vma(vma);
3266 unsigned long sz = huge_page_size(h);
ac46d4f3 3267 struct mmu_notifier_range range;
e8569dd2 3268 int ret = 0;
1e8f889b
DG
3269
3270 cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
63551ae0 3271
ac46d4f3
JG
3272 if (cow) {
3273 mmu_notifier_range_init(&range, src, vma->vm_start,
3274 vma->vm_end);
3275 mmu_notifier_invalidate_range_start(&range);
3276 }
e8569dd2 3277
a5516438 3278 for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
cb900f41 3279 spinlock_t *src_ptl, *dst_ptl;
7868a208 3280 src_pte = huge_pte_offset(src, addr, sz);
c74df32c
HD
3281 if (!src_pte)
3282 continue;
a5516438 3283 dst_pte = huge_pte_alloc(dst, addr, sz);
e8569dd2
AS
3284 if (!dst_pte) {
3285 ret = -ENOMEM;
3286 break;
3287 }
c5c99429 3288
5e41540c
MK
3289 /*
3290 * If the pagetables are shared don't copy or take references.
3291 * dst_pte == src_pte is the common case of src/dest sharing.
3292 *
3293 * However, src could have 'unshared' and dst shares with
3294 * another vma. If dst_pte !none, this implies sharing.
3295 * Check here before taking page table lock, and once again
3296 * after taking the lock below.
3297 */
3298 dst_entry = huge_ptep_get(dst_pte);
3299 if ((dst_pte == src_pte) || !huge_pte_none(dst_entry))
c5c99429
LW
3300 continue;
3301
cb900f41
KS
3302 dst_ptl = huge_pte_lock(h, dst, dst_pte);
3303 src_ptl = huge_pte_lockptr(h, src, src_pte);
3304 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
4a705fef 3305 entry = huge_ptep_get(src_pte);
5e41540c
MK
3306 dst_entry = huge_ptep_get(dst_pte);
3307 if (huge_pte_none(entry) || !huge_pte_none(dst_entry)) {
3308 /*
3309 * Skip if src entry none. Also, skip in the
3310 * unlikely case dst entry !none as this implies
3311 * sharing with another vma.
3312 */
4a705fef
NH
3313 ;
3314 } else if (unlikely(is_hugetlb_entry_migration(entry) ||
3315 is_hugetlb_entry_hwpoisoned(entry))) {
3316 swp_entry_t swp_entry = pte_to_swp_entry(entry);
3317
3318 if (is_write_migration_entry(swp_entry) && cow) {
3319 /*
3320 * COW mappings require pages in both
3321 * parent and child to be set to read.
3322 */
3323 make_migration_entry_read(&swp_entry);
3324 entry = swp_entry_to_pte(swp_entry);
e5251fd4
PA
3325 set_huge_swap_pte_at(src, addr, src_pte,
3326 entry, sz);
4a705fef 3327 }
e5251fd4 3328 set_huge_swap_pte_at(dst, addr, dst_pte, entry, sz);
4a705fef 3329 } else {
34ee645e 3330 if (cow) {
0f10851e
JG
3331 /*
3332 * No need to notify as we are downgrading page
3333 * table protection not changing it to point
3334 * to a new page.
3335 *
ad56b738 3336 * See Documentation/vm/mmu_notifier.rst
0f10851e 3337 */
7f2e9525 3338 huge_ptep_set_wrprotect(src, addr, src_pte);
34ee645e 3339 }
0253d634 3340 entry = huge_ptep_get(src_pte);
1c59827d
HD
3341 ptepage = pte_page(entry);
3342 get_page(ptepage);
53f9263b 3343 page_dup_rmap(ptepage, true);
1c59827d 3344 set_huge_pte_at(dst, addr, dst_pte, entry);
5d317b2b 3345 hugetlb_count_add(pages_per_huge_page(h), dst);
1c59827d 3346 }
cb900f41
KS
3347 spin_unlock(src_ptl);
3348 spin_unlock(dst_ptl);
63551ae0 3349 }
63551ae0 3350
e8569dd2 3351 if (cow)
ac46d4f3 3352 mmu_notifier_invalidate_range_end(&range);
e8569dd2
AS
3353
3354 return ret;
63551ae0
DG
3355}
3356
24669e58
AK
3357void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
3358 unsigned long start, unsigned long end,
3359 struct page *ref_page)
63551ae0
DG
3360{
3361 struct mm_struct *mm = vma->vm_mm;
3362 unsigned long address;
c7546f8f 3363 pte_t *ptep;
63551ae0 3364 pte_t pte;
cb900f41 3365 spinlock_t *ptl;
63551ae0 3366 struct page *page;
a5516438
AK
3367 struct hstate *h = hstate_vma(vma);
3368 unsigned long sz = huge_page_size(h);
ac46d4f3 3369 struct mmu_notifier_range range;
a5516438 3370
63551ae0 3371 WARN_ON(!is_vm_hugetlb_page(vma));
a5516438
AK
3372 BUG_ON(start & ~huge_page_mask(h));
3373 BUG_ON(end & ~huge_page_mask(h));
63551ae0 3374
07e32661
AK
3375 /*
3376 * This is a hugetlb vma, all the pte entries should point
3377 * to huge page.
3378 */
ed6a7935 3379 tlb_change_page_size(tlb, sz);
24669e58 3380 tlb_start_vma(tlb, vma);
dff11abe
MK
3381
3382 /*
3383 * If sharing possible, alert mmu notifiers of worst case.
3384 */
ac46d4f3
JG
3385 mmu_notifier_range_init(&range, mm, start, end);
3386 adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end);
3387 mmu_notifier_invalidate_range_start(&range);
569f48b8 3388 address = start;
569f48b8 3389 for (; address < end; address += sz) {
7868a208 3390 ptep = huge_pte_offset(mm, address, sz);
4c887265 3391 if (!ptep)
c7546f8f
DG
3392 continue;
3393
cb900f41 3394 ptl = huge_pte_lock(h, mm, ptep);
31d49da5
AK
3395 if (huge_pmd_unshare(mm, &address, ptep)) {
3396 spin_unlock(ptl);
dff11abe
MK
3397 /*
3398 * We just unmapped a page of PMDs by clearing a PUD.
3399 * The caller's TLB flush range should cover this area.
3400 */
31d49da5
AK
3401 continue;
3402 }
39dde65c 3403
6629326b 3404 pte = huge_ptep_get(ptep);
31d49da5
AK
3405 if (huge_pte_none(pte)) {
3406 spin_unlock(ptl);
3407 continue;
3408 }
6629326b
HD
3409
3410 /*
9fbc1f63
NH
3411 * Migrating hugepage or HWPoisoned hugepage is already
3412 * unmapped and its refcount is dropped, so just clear pte here.
6629326b 3413 */
9fbc1f63 3414 if (unlikely(!pte_present(pte))) {
9386fac3 3415 huge_pte_clear(mm, address, ptep, sz);
31d49da5
AK
3416 spin_unlock(ptl);
3417 continue;
8c4894c6 3418 }
6629326b
HD
3419
3420 page = pte_page(pte);
04f2cbe3
MG
3421 /*
3422 * If a reference page is supplied, it is because a specific
3423 * page is being unmapped, not a range. Ensure the page we
3424 * are about to unmap is the actual page of interest.
3425 */
3426 if (ref_page) {
31d49da5
AK
3427 if (page != ref_page) {
3428 spin_unlock(ptl);
3429 continue;
3430 }
04f2cbe3
MG
3431 /*
3432 * Mark the VMA as having unmapped its page so that
3433 * future faults in this VMA will fail rather than
3434 * looking like data was lost
3435 */
3436 set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
3437 }
3438
c7546f8f 3439 pte = huge_ptep_get_and_clear(mm, address, ptep);
b528e4b6 3440 tlb_remove_huge_tlb_entry(h, tlb, ptep, address);
106c992a 3441 if (huge_pte_dirty(pte))
6649a386 3442 set_page_dirty(page);
9e81130b 3443
5d317b2b 3444 hugetlb_count_sub(pages_per_huge_page(h), mm);
d281ee61 3445 page_remove_rmap(page, true);
31d49da5 3446
cb900f41 3447 spin_unlock(ptl);
e77b0852 3448 tlb_remove_page_size(tlb, page, huge_page_size(h));
31d49da5
AK
3449 /*
3450 * Bail out after unmapping reference page if supplied
3451 */
3452 if (ref_page)
3453 break;
fe1668ae 3454 }
ac46d4f3 3455 mmu_notifier_invalidate_range_end(&range);
24669e58 3456 tlb_end_vma(tlb, vma);
1da177e4 3457}
63551ae0 3458
d833352a
MG
3459void __unmap_hugepage_range_final(struct mmu_gather *tlb,
3460 struct vm_area_struct *vma, unsigned long start,
3461 unsigned long end, struct page *ref_page)
3462{
3463 __unmap_hugepage_range(tlb, vma, start, end, ref_page);
3464
3465 /*
3466 * Clear this flag so that x86's huge_pmd_share page_table_shareable
3467 * test will fail on a vma being torn down, and not grab a page table
3468 * on its way out. We're lucky that the flag has such an appropriate
3469 * name, and can in fact be safely cleared here. We could clear it
3470 * before the __unmap_hugepage_range above, but all that's necessary
c8c06efa 3471 * is to clear it before releasing the i_mmap_rwsem. This works
d833352a 3472 * because in the context this is called, the VMA is about to be
c8c06efa 3473 * destroyed and the i_mmap_rwsem is held.
d833352a
MG
3474 */
3475 vma->vm_flags &= ~VM_MAYSHARE;
3476}
3477
502717f4 3478void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
04f2cbe3 3479 unsigned long end, struct page *ref_page)
502717f4 3480{
24669e58
AK
3481 struct mm_struct *mm;
3482 struct mmu_gather tlb;
dff11abe
MK
3483 unsigned long tlb_start = start;
3484 unsigned long tlb_end = end;
3485
3486 /*
3487 * If shared PMDs were possibly used within this vma range, adjust
3488 * start/end for worst case tlb flushing.
3489 * Note that we can not be sure if PMDs are shared until we try to
3490 * unmap pages. However, we want to make sure TLB flushing covers
3491 * the largest possible range.
3492 */
3493 adjust_range_if_pmd_sharing_possible(vma, &tlb_start, &tlb_end);
24669e58
AK
3494
3495 mm = vma->vm_mm;
3496
dff11abe 3497 tlb_gather_mmu(&tlb, mm, tlb_start, tlb_end);
24669e58 3498 __unmap_hugepage_range(&tlb, vma, start, end, ref_page);
dff11abe 3499 tlb_finish_mmu(&tlb, tlb_start, tlb_end);
502717f4
KC
3500}
3501
04f2cbe3
MG
3502/*
3503 * This is called when the original mapper is failing to COW a MAP_PRIVATE
3504 * mappping it owns the reserve page for. The intention is to unmap the page
3505 * from other VMAs and let the children be SIGKILLed if they are faulting the
3506 * same region.
3507 */
2f4612af
DB
3508static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
3509 struct page *page, unsigned long address)
04f2cbe3 3510{
7526674d 3511 struct hstate *h = hstate_vma(vma);
04f2cbe3
MG
3512 struct vm_area_struct *iter_vma;
3513 struct address_space *mapping;
04f2cbe3
MG
3514 pgoff_t pgoff;
3515
3516 /*
3517 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
3518 * from page cache lookup which is in HPAGE_SIZE units.
3519 */
7526674d 3520 address = address & huge_page_mask(h);
36e4f20a
MH
3521 pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) +
3522 vma->vm_pgoff;
93c76a3d 3523 mapping = vma->vm_file->f_mapping;
04f2cbe3 3524
4eb2b1dc
MG
3525 /*
3526 * Take the mapping lock for the duration of the table walk. As
3527 * this mapping should be shared between all the VMAs,
3528 * __unmap_hugepage_range() is called as the lock is already held
3529 */
83cde9e8 3530 i_mmap_lock_write(mapping);
6b2dbba8 3531 vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) {
04f2cbe3
MG
3532 /* Do not unmap the current VMA */
3533 if (iter_vma == vma)
3534 continue;
3535
2f84a899
MG
3536 /*
3537 * Shared VMAs have their own reserves and do not affect
3538 * MAP_PRIVATE accounting but it is possible that a shared
3539 * VMA is using the same page so check and skip such VMAs.
3540 */
3541 if (iter_vma->vm_flags & VM_MAYSHARE)
3542 continue;
3543
04f2cbe3
MG
3544 /*
3545 * Unmap the page from other VMAs without their own reserves.
3546 * They get marked to be SIGKILLed if they fault in these
3547 * areas. This is because a future no-page fault on this VMA
3548 * could insert a zeroed page instead of the data existing
3549 * from the time of fork. This would look like data corruption
3550 */
3551 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
24669e58
AK
3552 unmap_hugepage_range(iter_vma, address,
3553 address + huge_page_size(h), page);
04f2cbe3 3554 }
83cde9e8 3555 i_mmap_unlock_write(mapping);
04f2cbe3
MG
3556}
3557
0fe6e20b
NH
3558/*
3559 * Hugetlb_cow() should be called with page lock of the original hugepage held.
ef009b25
MH
3560 * Called with hugetlb_instantiation_mutex held and pte_page locked so we
3561 * cannot race with other handlers or page migration.
3562 * Keep the pte_same checks anyway to make transition from the mutex easier.
0fe6e20b 3563 */
2b740303 3564static vm_fault_t hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
974e6d66 3565 unsigned long address, pte_t *ptep,
3999f52e 3566 struct page *pagecache_page, spinlock_t *ptl)
1e8f889b 3567{
3999f52e 3568 pte_t pte;
a5516438 3569 struct hstate *h = hstate_vma(vma);
1e8f889b 3570 struct page *old_page, *new_page;
2b740303
SJ
3571 int outside_reserve = 0;
3572 vm_fault_t ret = 0;
974e6d66 3573 unsigned long haddr = address & huge_page_mask(h);
ac46d4f3 3574 struct mmu_notifier_range range;
1e8f889b 3575
3999f52e 3576 pte = huge_ptep_get(ptep);
1e8f889b
DG
3577 old_page = pte_page(pte);
3578
04f2cbe3 3579retry_avoidcopy:
1e8f889b
DG
3580 /* If no-one else is actually using this page, avoid the copy
3581 * and just make the page writable */
37a2140d 3582 if (page_mapcount(old_page) == 1 && PageAnon(old_page)) {
5a49973d 3583 page_move_anon_rmap(old_page, vma);
5b7a1d40 3584 set_huge_ptep_writable(vma, haddr, ptep);
83c54070 3585 return 0;
1e8f889b
DG
3586 }
3587
04f2cbe3
MG
3588 /*
3589 * If the process that created a MAP_PRIVATE mapping is about to
3590 * perform a COW due to a shared page count, attempt to satisfy
3591 * the allocation without using the existing reserves. The pagecache
3592 * page is used to determine if the reserve at this address was
3593 * consumed or not. If reserves were used, a partial faulted mapping
3594 * at the time of fork() could consume its reserves on COW instead
3595 * of the full address range.
3596 */
5944d011 3597 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
04f2cbe3
MG
3598 old_page != pagecache_page)
3599 outside_reserve = 1;
3600
09cbfeaf 3601 get_page(old_page);
b76c8cfb 3602
ad4404a2
DB
3603 /*
3604 * Drop page table lock as buddy allocator may be called. It will
3605 * be acquired again before returning to the caller, as expected.
3606 */
cb900f41 3607 spin_unlock(ptl);
5b7a1d40 3608 new_page = alloc_huge_page(vma, haddr, outside_reserve);
1e8f889b 3609
2fc39cec 3610 if (IS_ERR(new_page)) {
04f2cbe3
MG
3611 /*
3612 * If a process owning a MAP_PRIVATE mapping fails to COW,
3613 * it is due to references held by a child and an insufficient
3614 * huge page pool. To guarantee the original mappers
3615 * reliability, unmap the page from child processes. The child
3616 * may get SIGKILLed if it later faults.
3617 */
3618 if (outside_reserve) {
09cbfeaf 3619 put_page(old_page);
04f2cbe3 3620 BUG_ON(huge_pte_none(pte));
5b7a1d40 3621 unmap_ref_private(mm, vma, old_page, haddr);
2f4612af
DB
3622 BUG_ON(huge_pte_none(pte));
3623 spin_lock(ptl);
5b7a1d40 3624 ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
2f4612af
DB
3625 if (likely(ptep &&
3626 pte_same(huge_ptep_get(ptep), pte)))
3627 goto retry_avoidcopy;
3628 /*
3629 * race occurs while re-acquiring page table
3630 * lock, and our job is done.
3631 */
3632 return 0;
04f2cbe3
MG
3633 }
3634
2b740303 3635 ret = vmf_error(PTR_ERR(new_page));
ad4404a2 3636 goto out_release_old;
1e8f889b
DG
3637 }
3638
0fe6e20b
NH
3639 /*
3640 * When the original hugepage is shared one, it does not have
3641 * anon_vma prepared.
3642 */
44e2aa93 3643 if (unlikely(anon_vma_prepare(vma))) {
ad4404a2
DB
3644 ret = VM_FAULT_OOM;
3645 goto out_release_all;
44e2aa93 3646 }
0fe6e20b 3647
974e6d66 3648 copy_user_huge_page(new_page, old_page, address, vma,
47ad8475 3649 pages_per_huge_page(h));
0ed361de 3650 __SetPageUptodate(new_page);
1e8f889b 3651
ac46d4f3
JG
3652 mmu_notifier_range_init(&range, mm, haddr, haddr + huge_page_size(h));
3653 mmu_notifier_invalidate_range_start(&range);
ad4404a2 3654
b76c8cfb 3655 /*
cb900f41 3656 * Retake the page table lock to check for racing updates
b76c8cfb
LW
3657 * before the page tables are altered
3658 */
cb900f41 3659 spin_lock(ptl);
5b7a1d40 3660 ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
a9af0c5d 3661 if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) {
07443a85
JK
3662 ClearPagePrivate(new_page);
3663
1e8f889b 3664 /* Break COW */
5b7a1d40 3665 huge_ptep_clear_flush(vma, haddr, ptep);
ac46d4f3 3666 mmu_notifier_invalidate_range(mm, range.start, range.end);
5b7a1d40 3667 set_huge_pte_at(mm, haddr, ptep,
1e8f889b 3668 make_huge_pte(vma, new_page, 1));
d281ee61 3669 page_remove_rmap(old_page, true);
5b7a1d40 3670 hugepage_add_new_anon_rmap(new_page, vma, haddr);
cb6acd01 3671 set_page_huge_active(new_page);
1e8f889b
DG
3672 /* Make the old page be freed below */
3673 new_page = old_page;
3674 }
cb900f41 3675 spin_unlock(ptl);
ac46d4f3 3676 mmu_notifier_invalidate_range_end(&range);
ad4404a2 3677out_release_all:
5b7a1d40 3678 restore_reserve_on_error(h, vma, haddr, new_page);
09cbfeaf 3679 put_page(new_page);
ad4404a2 3680out_release_old:
09cbfeaf 3681 put_page(old_page);
8312034f 3682
ad4404a2
DB
3683 spin_lock(ptl); /* Caller expects lock to be held */
3684 return ret;
1e8f889b
DG
3685}
3686
04f2cbe3 3687/* Return the pagecache page at a given address within a VMA */
a5516438
AK
3688static struct page *hugetlbfs_pagecache_page(struct hstate *h,
3689 struct vm_area_struct *vma, unsigned long address)
04f2cbe3
MG
3690{
3691 struct address_space *mapping;
e7c4b0bf 3692 pgoff_t idx;
04f2cbe3
MG
3693
3694 mapping = vma->vm_file->f_mapping;
a5516438 3695 idx = vma_hugecache_offset(h, vma, address);
04f2cbe3
MG
3696
3697 return find_lock_page(mapping, idx);
3698}
3699
3ae77f43
HD
3700/*
3701 * Return whether there is a pagecache page to back given address within VMA.
3702 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
3703 */
3704static bool hugetlbfs_pagecache_present(struct hstate *h,
2a15efc9
HD
3705 struct vm_area_struct *vma, unsigned long address)
3706{
3707 struct address_space *mapping;
3708 pgoff_t idx;
3709 struct page *page;
3710
3711 mapping = vma->vm_file->f_mapping;
3712 idx = vma_hugecache_offset(h, vma, address);
3713
3714 page = find_get_page(mapping, idx);
3715 if (page)
3716 put_page(page);
3717 return page != NULL;
3718}
3719
ab76ad54
MK
3720int huge_add_to_page_cache(struct page *page, struct address_space *mapping,
3721 pgoff_t idx)
3722{
3723 struct inode *inode = mapping->host;
3724 struct hstate *h = hstate_inode(inode);
3725 int err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
3726
3727 if (err)
3728 return err;
3729 ClearPagePrivate(page);
3730
22146c3c
MK
3731 /*
3732 * set page dirty so that it will not be removed from cache/file
3733 * by non-hugetlbfs specific code paths.
3734 */
3735 set_page_dirty(page);
3736
ab76ad54
MK
3737 spin_lock(&inode->i_lock);
3738 inode->i_blocks += blocks_per_huge_page(h);
3739 spin_unlock(&inode->i_lock);
3740 return 0;
3741}
3742
2b740303
SJ
3743static vm_fault_t hugetlb_no_page(struct mm_struct *mm,
3744 struct vm_area_struct *vma,
3745 struct address_space *mapping, pgoff_t idx,
3746 unsigned long address, pte_t *ptep, unsigned int flags)
ac9b9c66 3747{
a5516438 3748 struct hstate *h = hstate_vma(vma);
2b740303 3749 vm_fault_t ret = VM_FAULT_SIGBUS;
409eb8c2 3750 int anon_rmap = 0;
4c887265 3751 unsigned long size;
4c887265 3752 struct page *page;
1e8f889b 3753 pte_t new_pte;
cb900f41 3754 spinlock_t *ptl;
285b8dca 3755 unsigned long haddr = address & huge_page_mask(h);
cb6acd01 3756 bool new_page = false;
4c887265 3757
04f2cbe3
MG
3758 /*
3759 * Currently, we are forced to kill the process in the event the
3760 * original mapper has unmapped pages from the child due to a failed
25985edc 3761 * COW. Warn that such a situation has occurred as it may not be obvious
04f2cbe3
MG
3762 */
3763 if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
910154d5 3764 pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n",
ffb22af5 3765 current->pid);
04f2cbe3
MG
3766 return ret;
3767 }
3768
4c887265 3769 /*
e7c58097
MK
3770 * Use page lock to guard against racing truncation
3771 * before we get page_table_lock.
4c887265 3772 */
6bda666a
CL
3773retry:
3774 page = find_lock_page(mapping, idx);
3775 if (!page) {
e7c58097
MK
3776 size = i_size_read(mapping->host) >> huge_page_shift(h);
3777 if (idx >= size)
3778 goto out;
3779
1a1aad8a
MK
3780 /*
3781 * Check for page in userfault range
3782 */
3783 if (userfaultfd_missing(vma)) {
3784 u32 hash;
3785 struct vm_fault vmf = {
3786 .vma = vma,
285b8dca 3787 .address = haddr,
1a1aad8a
MK
3788 .flags = flags,
3789 /*
3790 * Hard to debug if it ends up being
3791 * used by a callee that assumes
3792 * something about the other
3793 * uninitialized fields... same as in
3794 * memory.c
3795 */
3796 };
3797
3798 /*
ddeaab32
MK
3799 * hugetlb_fault_mutex must be dropped before
3800 * handling userfault. Reacquire after handling
3801 * fault to make calling code simpler.
1a1aad8a
MK
3802 */
3803 hash = hugetlb_fault_mutex_hash(h, mm, vma, mapping,
285b8dca 3804 idx, haddr);
1a1aad8a
MK
3805 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
3806 ret = handle_userfault(&vmf, VM_UFFD_MISSING);
3807 mutex_lock(&hugetlb_fault_mutex_table[hash]);
3808 goto out;
3809 }
3810
285b8dca 3811 page = alloc_huge_page(vma, haddr, 0);
2fc39cec 3812 if (IS_ERR(page)) {
2b740303 3813 ret = vmf_error(PTR_ERR(page));
6bda666a
CL
3814 goto out;
3815 }
47ad8475 3816 clear_huge_page(page, address, pages_per_huge_page(h));
0ed361de 3817 __SetPageUptodate(page);
cb6acd01 3818 new_page = true;
ac9b9c66 3819
f83a275d 3820 if (vma->vm_flags & VM_MAYSHARE) {
ab76ad54 3821 int err = huge_add_to_page_cache(page, mapping, idx);
6bda666a
CL
3822 if (err) {
3823 put_page(page);
6bda666a
CL
3824 if (err == -EEXIST)
3825 goto retry;
3826 goto out;
3827 }
23be7468 3828 } else {
6bda666a 3829 lock_page(page);
0fe6e20b
NH
3830 if (unlikely(anon_vma_prepare(vma))) {
3831 ret = VM_FAULT_OOM;
3832 goto backout_unlocked;
3833 }
409eb8c2 3834 anon_rmap = 1;
23be7468 3835 }
0fe6e20b 3836 } else {
998b4382
NH
3837 /*
3838 * If memory error occurs between mmap() and fault, some process
3839 * don't have hwpoisoned swap entry for errored virtual address.
3840 * So we need to block hugepage fault by PG_hwpoison bit check.
3841 */
3842 if (unlikely(PageHWPoison(page))) {
32f84528 3843 ret = VM_FAULT_HWPOISON |
972dc4de 3844 VM_FAULT_SET_HINDEX(hstate_index(h));
998b4382
NH
3845 goto backout_unlocked;
3846 }
6bda666a 3847 }
1e8f889b 3848
57303d80
AW
3849 /*
3850 * If we are going to COW a private mapping later, we examine the
3851 * pending reservations for this page now. This will ensure that
3852 * any allocations necessary to record that reservation occur outside
3853 * the spinlock.
3854 */
5e911373 3855 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
285b8dca 3856 if (vma_needs_reservation(h, vma, haddr) < 0) {
2b26736c
AW
3857 ret = VM_FAULT_OOM;
3858 goto backout_unlocked;
3859 }
5e911373 3860 /* Just decrements count, does not deallocate */
285b8dca 3861 vma_end_reservation(h, vma, haddr);
5e911373 3862 }
57303d80 3863
8bea8052 3864 ptl = huge_pte_lock(h, mm, ptep);
e7c58097
MK
3865 size = i_size_read(mapping->host) >> huge_page_shift(h);
3866 if (idx >= size)
3867 goto backout;
4c887265 3868
83c54070 3869 ret = 0;
7f2e9525 3870 if (!huge_pte_none(huge_ptep_get(ptep)))
4c887265
AL
3871 goto backout;
3872
07443a85
JK
3873 if (anon_rmap) {
3874 ClearPagePrivate(page);
285b8dca 3875 hugepage_add_new_anon_rmap(page, vma, haddr);
ac714904 3876 } else
53f9263b 3877 page_dup_rmap(page, true);
1e8f889b
DG
3878 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
3879 && (vma->vm_flags & VM_SHARED)));
285b8dca 3880 set_huge_pte_at(mm, haddr, ptep, new_pte);
1e8f889b 3881
5d317b2b 3882 hugetlb_count_add(pages_per_huge_page(h), mm);
788c7df4 3883 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
1e8f889b 3884 /* Optimization, do the COW without a second fault */
974e6d66 3885 ret = hugetlb_cow(mm, vma, address, ptep, page, ptl);
1e8f889b
DG
3886 }
3887
cb900f41 3888 spin_unlock(ptl);
cb6acd01
MK
3889
3890 /*
3891 * Only make newly allocated pages active. Existing pages found
3892 * in the pagecache could be !page_huge_active() if they have been
3893 * isolated for migration.
3894 */
3895 if (new_page)
3896 set_page_huge_active(page);
3897
4c887265
AL
3898 unlock_page(page);
3899out:
ac9b9c66 3900 return ret;
4c887265
AL
3901
3902backout:
cb900f41 3903 spin_unlock(ptl);
2b26736c 3904backout_unlocked:
4c887265 3905 unlock_page(page);
285b8dca 3906 restore_reserve_on_error(h, vma, haddr, page);
4c887265
AL
3907 put_page(page);
3908 goto out;
ac9b9c66
HD
3909}
3910
8382d914 3911#ifdef CONFIG_SMP
c672c7f2 3912u32 hugetlb_fault_mutex_hash(struct hstate *h, struct mm_struct *mm,
8382d914
DB
3913 struct vm_area_struct *vma,
3914 struct address_space *mapping,
3915 pgoff_t idx, unsigned long address)
3916{
3917 unsigned long key[2];
3918 u32 hash;
3919
3920 if (vma->vm_flags & VM_SHARED) {
3921 key[0] = (unsigned long) mapping;
3922 key[1] = idx;
3923 } else {
3924 key[0] = (unsigned long) mm;
3925 key[1] = address >> huge_page_shift(h);
3926 }
3927
3928 hash = jhash2((u32 *)&key, sizeof(key)/sizeof(u32), 0);
3929
3930 return hash & (num_fault_mutexes - 1);
3931}
3932#else
3933/*
3934 * For uniprocesor systems we always use a single mutex, so just
3935 * return 0 and avoid the hashing overhead.
3936 */
c672c7f2 3937u32 hugetlb_fault_mutex_hash(struct hstate *h, struct mm_struct *mm,
8382d914
DB
3938 struct vm_area_struct *vma,
3939 struct address_space *mapping,
3940 pgoff_t idx, unsigned long address)
3941{
3942 return 0;
3943}
3944#endif
3945
2b740303 3946vm_fault_t hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
788c7df4 3947 unsigned long address, unsigned int flags)
86e5216f 3948{
8382d914 3949 pte_t *ptep, entry;
cb900f41 3950 spinlock_t *ptl;
2b740303 3951 vm_fault_t ret;
8382d914
DB
3952 u32 hash;
3953 pgoff_t idx;
0fe6e20b 3954 struct page *page = NULL;
57303d80 3955 struct page *pagecache_page = NULL;
a5516438 3956 struct hstate *h = hstate_vma(vma);
8382d914 3957 struct address_space *mapping;
0f792cf9 3958 int need_wait_lock = 0;
285b8dca 3959 unsigned long haddr = address & huge_page_mask(h);
86e5216f 3960
285b8dca 3961 ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
fd6a03ed
NH
3962 if (ptep) {
3963 entry = huge_ptep_get(ptep);
290408d4 3964 if (unlikely(is_hugetlb_entry_migration(entry))) {
cb900f41 3965 migration_entry_wait_huge(vma, mm, ptep);
290408d4
NH
3966 return 0;
3967 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
32f84528 3968 return VM_FAULT_HWPOISON_LARGE |
972dc4de 3969 VM_FAULT_SET_HINDEX(hstate_index(h));
ddeaab32
MK
3970 } else {
3971 ptep = huge_pte_alloc(mm, haddr, huge_page_size(h));
3972 if (!ptep)
3973 return VM_FAULT_OOM;
fd6a03ed
NH
3974 }
3975
8382d914 3976 mapping = vma->vm_file->f_mapping;
ddeaab32 3977 idx = vma_hugecache_offset(h, vma, haddr);
8382d914 3978
3935baa9
DG
3979 /*
3980 * Serialize hugepage allocation and instantiation, so that we don't
3981 * get spurious allocation failures if two CPUs race to instantiate
3982 * the same page in the page cache.
3983 */
285b8dca 3984 hash = hugetlb_fault_mutex_hash(h, mm, vma, mapping, idx, haddr);
c672c7f2 3985 mutex_lock(&hugetlb_fault_mutex_table[hash]);
8382d914 3986
7f2e9525
GS
3987 entry = huge_ptep_get(ptep);
3988 if (huge_pte_none(entry)) {
8382d914 3989 ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags);
b4d1d99f 3990 goto out_mutex;
3935baa9 3991 }
86e5216f 3992
83c54070 3993 ret = 0;
1e8f889b 3994
0f792cf9
NH
3995 /*
3996 * entry could be a migration/hwpoison entry at this point, so this
3997 * check prevents the kernel from going below assuming that we have
3998 * a active hugepage in pagecache. This goto expects the 2nd page fault,
3999 * and is_hugetlb_entry_(migration|hwpoisoned) check will properly
4000 * handle it.
4001 */
4002 if (!pte_present(entry))
4003 goto out_mutex;
4004
57303d80
AW
4005 /*
4006 * If we are going to COW the mapping later, we examine the pending
4007 * reservations for this page now. This will ensure that any
4008 * allocations necessary to record that reservation occur outside the
4009 * spinlock. For private mappings, we also lookup the pagecache
4010 * page now as it is used to determine if a reservation has been
4011 * consumed.
4012 */
106c992a 4013 if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) {
285b8dca 4014 if (vma_needs_reservation(h, vma, haddr) < 0) {
2b26736c 4015 ret = VM_FAULT_OOM;
b4d1d99f 4016 goto out_mutex;
2b26736c 4017 }
5e911373 4018 /* Just decrements count, does not deallocate */
285b8dca 4019 vma_end_reservation(h, vma, haddr);
57303d80 4020
f83a275d 4021 if (!(vma->vm_flags & VM_MAYSHARE))
57303d80 4022 pagecache_page = hugetlbfs_pagecache_page(h,
285b8dca 4023 vma, haddr);
57303d80
AW
4024 }
4025
0f792cf9
NH
4026 ptl = huge_pte_lock(h, mm, ptep);
4027
4028 /* Check for a racing update before calling hugetlb_cow */
4029 if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
4030 goto out_ptl;
4031
56c9cfb1
NH
4032 /*
4033 * hugetlb_cow() requires page locks of pte_page(entry) and
4034 * pagecache_page, so here we need take the former one
4035 * when page != pagecache_page or !pagecache_page.
56c9cfb1
NH
4036 */
4037 page = pte_page(entry);
4038 if (page != pagecache_page)
0f792cf9
NH
4039 if (!trylock_page(page)) {
4040 need_wait_lock = 1;
4041 goto out_ptl;
4042 }
b4d1d99f 4043
0f792cf9 4044 get_page(page);
b4d1d99f 4045
788c7df4 4046 if (flags & FAULT_FLAG_WRITE) {
106c992a 4047 if (!huge_pte_write(entry)) {
974e6d66 4048 ret = hugetlb_cow(mm, vma, address, ptep,
3999f52e 4049 pagecache_page, ptl);
0f792cf9 4050 goto out_put_page;
b4d1d99f 4051 }
106c992a 4052 entry = huge_pte_mkdirty(entry);
b4d1d99f
DG
4053 }
4054 entry = pte_mkyoung(entry);
285b8dca 4055 if (huge_ptep_set_access_flags(vma, haddr, ptep, entry,
788c7df4 4056 flags & FAULT_FLAG_WRITE))
285b8dca 4057 update_mmu_cache(vma, haddr, ptep);
0f792cf9
NH
4058out_put_page:
4059 if (page != pagecache_page)
4060 unlock_page(page);
4061 put_page(page);
cb900f41
KS
4062out_ptl:
4063 spin_unlock(ptl);
57303d80
AW
4064
4065 if (pagecache_page) {
4066 unlock_page(pagecache_page);
4067 put_page(pagecache_page);
4068 }
b4d1d99f 4069out_mutex:
c672c7f2 4070 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
0f792cf9
NH
4071 /*
4072 * Generally it's safe to hold refcount during waiting page lock. But
4073 * here we just wait to defer the next page fault to avoid busy loop and
4074 * the page is not used after unlocked before returning from the current
4075 * page fault. So we are safe from accessing freed page, even if we wait
4076 * here without taking refcount.
4077 */
4078 if (need_wait_lock)
4079 wait_on_page_locked(page);
1e8f889b 4080 return ret;
86e5216f
AL
4081}
4082
8fb5debc
MK
4083/*
4084 * Used by userfaultfd UFFDIO_COPY. Based on mcopy_atomic_pte with
4085 * modifications for huge pages.
4086 */
4087int hugetlb_mcopy_atomic_pte(struct mm_struct *dst_mm,
4088 pte_t *dst_pte,
4089 struct vm_area_struct *dst_vma,
4090 unsigned long dst_addr,
4091 unsigned long src_addr,
4092 struct page **pagep)
4093{
1e392147
AA
4094 struct address_space *mapping;
4095 pgoff_t idx;
4096 unsigned long size;
1c9e8def 4097 int vm_shared = dst_vma->vm_flags & VM_SHARED;
8fb5debc
MK
4098 struct hstate *h = hstate_vma(dst_vma);
4099 pte_t _dst_pte;
4100 spinlock_t *ptl;
4101 int ret;
4102 struct page *page;
4103
4104 if (!*pagep) {
4105 ret = -ENOMEM;
4106 page = alloc_huge_page(dst_vma, dst_addr, 0);
4107 if (IS_ERR(page))
4108 goto out;
4109
4110 ret = copy_huge_page_from_user(page,
4111 (const void __user *) src_addr,
810a56b9 4112 pages_per_huge_page(h), false);
8fb5debc
MK
4113
4114 /* fallback to copy_from_user outside mmap_sem */
4115 if (unlikely(ret)) {
9e368259 4116 ret = -ENOENT;
8fb5debc
MK
4117 *pagep = page;
4118 /* don't free the page */
4119 goto out;
4120 }
4121 } else {
4122 page = *pagep;
4123 *pagep = NULL;
4124 }
4125
4126 /*
4127 * The memory barrier inside __SetPageUptodate makes sure that
4128 * preceding stores to the page contents become visible before
4129 * the set_pte_at() write.
4130 */
4131 __SetPageUptodate(page);
8fb5debc 4132
1e392147
AA
4133 mapping = dst_vma->vm_file->f_mapping;
4134 idx = vma_hugecache_offset(h, dst_vma, dst_addr);
4135
1c9e8def
MK
4136 /*
4137 * If shared, add to page cache
4138 */
4139 if (vm_shared) {
1e392147
AA
4140 size = i_size_read(mapping->host) >> huge_page_shift(h);
4141 ret = -EFAULT;
4142 if (idx >= size)
4143 goto out_release_nounlock;
1c9e8def 4144
1e392147
AA
4145 /*
4146 * Serialization between remove_inode_hugepages() and
4147 * huge_add_to_page_cache() below happens through the
4148 * hugetlb_fault_mutex_table that here must be hold by
4149 * the caller.
4150 */
1c9e8def
MK
4151 ret = huge_add_to_page_cache(page, mapping, idx);
4152 if (ret)
4153 goto out_release_nounlock;
4154 }
4155
8fb5debc
MK
4156 ptl = huge_pte_lockptr(h, dst_mm, dst_pte);
4157 spin_lock(ptl);
4158
1e392147
AA
4159 /*
4160 * Recheck the i_size after holding PT lock to make sure not
4161 * to leave any page mapped (as page_mapped()) beyond the end
4162 * of the i_size (remove_inode_hugepages() is strict about
4163 * enforcing that). If we bail out here, we'll also leave a
4164 * page in the radix tree in the vm_shared case beyond the end
4165 * of the i_size, but remove_inode_hugepages() will take care
4166 * of it as soon as we drop the hugetlb_fault_mutex_table.
4167 */
4168 size = i_size_read(mapping->host) >> huge_page_shift(h);
4169 ret = -EFAULT;
4170 if (idx >= size)
4171 goto out_release_unlock;
4172
8fb5debc
MK
4173 ret = -EEXIST;
4174 if (!huge_pte_none(huge_ptep_get(dst_pte)))
4175 goto out_release_unlock;
4176
1c9e8def
MK
4177 if (vm_shared) {
4178 page_dup_rmap(page, true);
4179 } else {
4180 ClearPagePrivate(page);
4181 hugepage_add_new_anon_rmap(page, dst_vma, dst_addr);
4182 }
8fb5debc
MK
4183
4184 _dst_pte = make_huge_pte(dst_vma, page, dst_vma->vm_flags & VM_WRITE);
4185 if (dst_vma->vm_flags & VM_WRITE)
4186 _dst_pte = huge_pte_mkdirty(_dst_pte);
4187 _dst_pte = pte_mkyoung(_dst_pte);
4188
4189 set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte);
4190
4191 (void)huge_ptep_set_access_flags(dst_vma, dst_addr, dst_pte, _dst_pte,
4192 dst_vma->vm_flags & VM_WRITE);
4193 hugetlb_count_add(pages_per_huge_page(h), dst_mm);
4194
4195 /* No need to invalidate - it was non-present before */
4196 update_mmu_cache(dst_vma, dst_addr, dst_pte);
4197
4198 spin_unlock(ptl);
cb6acd01 4199 set_page_huge_active(page);
1c9e8def
MK
4200 if (vm_shared)
4201 unlock_page(page);
8fb5debc
MK
4202 ret = 0;
4203out:
4204 return ret;
4205out_release_unlock:
4206 spin_unlock(ptl);
1c9e8def
MK
4207 if (vm_shared)
4208 unlock_page(page);
5af10dfd 4209out_release_nounlock:
8fb5debc
MK
4210 put_page(page);
4211 goto out;
4212}
4213
28a35716
ML
4214long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
4215 struct page **pages, struct vm_area_struct **vmas,
4216 unsigned long *position, unsigned long *nr_pages,
87ffc118 4217 long i, unsigned int flags, int *nonblocking)
63551ae0 4218{
d5d4b0aa
KC
4219 unsigned long pfn_offset;
4220 unsigned long vaddr = *position;
28a35716 4221 unsigned long remainder = *nr_pages;
a5516438 4222 struct hstate *h = hstate_vma(vma);
2be7cfed 4223 int err = -EFAULT;
63551ae0 4224
63551ae0 4225 while (vaddr < vma->vm_end && remainder) {
4c887265 4226 pte_t *pte;
cb900f41 4227 spinlock_t *ptl = NULL;
2a15efc9 4228 int absent;
4c887265 4229 struct page *page;
63551ae0 4230
02057967
DR
4231 /*
4232 * If we have a pending SIGKILL, don't keep faulting pages and
4233 * potentially allocating memory.
4234 */
fa45f116 4235 if (fatal_signal_pending(current)) {
02057967
DR
4236 remainder = 0;
4237 break;
4238 }
4239
4c887265
AL
4240 /*
4241 * Some archs (sparc64, sh*) have multiple pte_ts to
2a15efc9 4242 * each hugepage. We have to make sure we get the
4c887265 4243 * first, for the page indexing below to work.
cb900f41
KS
4244 *
4245 * Note that page table lock is not held when pte is null.
4c887265 4246 */
7868a208
PA
4247 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h),
4248 huge_page_size(h));
cb900f41
KS
4249 if (pte)
4250 ptl = huge_pte_lock(h, mm, pte);
2a15efc9
HD
4251 absent = !pte || huge_pte_none(huge_ptep_get(pte));
4252
4253 /*
4254 * When coredumping, it suits get_dump_page if we just return
3ae77f43
HD
4255 * an error where there's an empty slot with no huge pagecache
4256 * to back it. This way, we avoid allocating a hugepage, and
4257 * the sparse dumpfile avoids allocating disk blocks, but its
4258 * huge holes still show up with zeroes where they need to be.
2a15efc9 4259 */
3ae77f43
HD
4260 if (absent && (flags & FOLL_DUMP) &&
4261 !hugetlbfs_pagecache_present(h, vma, vaddr)) {
cb900f41
KS
4262 if (pte)
4263 spin_unlock(ptl);
2a15efc9
HD
4264 remainder = 0;
4265 break;
4266 }
63551ae0 4267
9cc3a5bd
NH
4268 /*
4269 * We need call hugetlb_fault for both hugepages under migration
4270 * (in which case hugetlb_fault waits for the migration,) and
4271 * hwpoisoned hugepages (in which case we need to prevent the
4272 * caller from accessing to them.) In order to do this, we use
4273 * here is_swap_pte instead of is_hugetlb_entry_migration and
4274 * is_hugetlb_entry_hwpoisoned. This is because it simply covers
4275 * both cases, and because we can't follow correct pages
4276 * directly from any kind of swap entries.
4277 */
4278 if (absent || is_swap_pte(huge_ptep_get(pte)) ||
106c992a
GS
4279 ((flags & FOLL_WRITE) &&
4280 !huge_pte_write(huge_ptep_get(pte)))) {
2b740303 4281 vm_fault_t ret;
87ffc118 4282 unsigned int fault_flags = 0;
63551ae0 4283
cb900f41
KS
4284 if (pte)
4285 spin_unlock(ptl);
87ffc118
AA
4286 if (flags & FOLL_WRITE)
4287 fault_flags |= FAULT_FLAG_WRITE;
4288 if (nonblocking)
4289 fault_flags |= FAULT_FLAG_ALLOW_RETRY;
4290 if (flags & FOLL_NOWAIT)
4291 fault_flags |= FAULT_FLAG_ALLOW_RETRY |
4292 FAULT_FLAG_RETRY_NOWAIT;
4293 if (flags & FOLL_TRIED) {
4294 VM_WARN_ON_ONCE(fault_flags &
4295 FAULT_FLAG_ALLOW_RETRY);
4296 fault_flags |= FAULT_FLAG_TRIED;
4297 }
4298 ret = hugetlb_fault(mm, vma, vaddr, fault_flags);
4299 if (ret & VM_FAULT_ERROR) {
2be7cfed 4300 err = vm_fault_to_errno(ret, flags);
87ffc118
AA
4301 remainder = 0;
4302 break;
4303 }
4304 if (ret & VM_FAULT_RETRY) {
1ac25013
AA
4305 if (nonblocking &&
4306 !(fault_flags & FAULT_FLAG_RETRY_NOWAIT))
87ffc118
AA
4307 *nonblocking = 0;
4308 *nr_pages = 0;
4309 /*
4310 * VM_FAULT_RETRY must not return an
4311 * error, it will return zero
4312 * instead.
4313 *
4314 * No need to update "position" as the
4315 * caller will not check it after
4316 * *nr_pages is set to 0.
4317 */
4318 return i;
4319 }
4320 continue;
4c887265
AL
4321 }
4322
a5516438 4323 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
7f2e9525 4324 page = pte_page(huge_ptep_get(pte));
8fde12ca
LT
4325
4326 /*
4327 * Instead of doing 'try_get_page()' below in the same_page
4328 * loop, just check the count once here.
4329 */
4330 if (unlikely(page_count(page) <= 0)) {
4331 if (pages) {
4332 spin_unlock(ptl);
4333 remainder = 0;
4334 err = -ENOMEM;
4335 break;
4336 }
4337 }
d5d4b0aa 4338same_page:
d6692183 4339 if (pages) {
2a15efc9 4340 pages[i] = mem_map_offset(page, pfn_offset);
ddc58f27 4341 get_page(pages[i]);
d6692183 4342 }
63551ae0
DG
4343
4344 if (vmas)
4345 vmas[i] = vma;
4346
4347 vaddr += PAGE_SIZE;
d5d4b0aa 4348 ++pfn_offset;
63551ae0
DG
4349 --remainder;
4350 ++i;
d5d4b0aa 4351 if (vaddr < vma->vm_end && remainder &&
a5516438 4352 pfn_offset < pages_per_huge_page(h)) {
d5d4b0aa
KC
4353 /*
4354 * We use pfn_offset to avoid touching the pageframes
4355 * of this compound page.
4356 */
4357 goto same_page;
4358 }
cb900f41 4359 spin_unlock(ptl);
63551ae0 4360 }
28a35716 4361 *nr_pages = remainder;
87ffc118
AA
4362 /*
4363 * setting position is actually required only if remainder is
4364 * not zero but it's faster not to add a "if (remainder)"
4365 * branch.
4366 */
63551ae0
DG
4367 *position = vaddr;
4368
2be7cfed 4369 return i ? i : err;
63551ae0 4370}
8f860591 4371
5491ae7b
AK
4372#ifndef __HAVE_ARCH_FLUSH_HUGETLB_TLB_RANGE
4373/*
4374 * ARCHes with special requirements for evicting HUGETLB backing TLB entries can
4375 * implement this.
4376 */
4377#define flush_hugetlb_tlb_range(vma, addr, end) flush_tlb_range(vma, addr, end)
4378#endif
4379
7da4d641 4380unsigned long hugetlb_change_protection(struct vm_area_struct *vma,
8f860591
ZY
4381 unsigned long address, unsigned long end, pgprot_t newprot)
4382{
4383 struct mm_struct *mm = vma->vm_mm;
4384 unsigned long start = address;
4385 pte_t *ptep;
4386 pte_t pte;
a5516438 4387 struct hstate *h = hstate_vma(vma);
7da4d641 4388 unsigned long pages = 0;
dff11abe 4389 bool shared_pmd = false;
ac46d4f3 4390 struct mmu_notifier_range range;
dff11abe
MK
4391
4392 /*
4393 * In the case of shared PMDs, the area to flush could be beyond
ac46d4f3 4394 * start/end. Set range.start/range.end to cover the maximum possible
dff11abe
MK
4395 * range if PMD sharing is possible.
4396 */
ac46d4f3
JG
4397 mmu_notifier_range_init(&range, mm, start, end);
4398 adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end);
8f860591
ZY
4399
4400 BUG_ON(address >= end);
ac46d4f3 4401 flush_cache_range(vma, range.start, range.end);
8f860591 4402
ac46d4f3 4403 mmu_notifier_invalidate_range_start(&range);
83cde9e8 4404 i_mmap_lock_write(vma->vm_file->f_mapping);
a5516438 4405 for (; address < end; address += huge_page_size(h)) {
cb900f41 4406 spinlock_t *ptl;
7868a208 4407 ptep = huge_pte_offset(mm, address, huge_page_size(h));
8f860591
ZY
4408 if (!ptep)
4409 continue;
cb900f41 4410 ptl = huge_pte_lock(h, mm, ptep);
7da4d641
PZ
4411 if (huge_pmd_unshare(mm, &address, ptep)) {
4412 pages++;
cb900f41 4413 spin_unlock(ptl);
dff11abe 4414 shared_pmd = true;
39dde65c 4415 continue;
7da4d641 4416 }
a8bda28d
NH
4417 pte = huge_ptep_get(ptep);
4418 if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) {
4419 spin_unlock(ptl);
4420 continue;
4421 }
4422 if (unlikely(is_hugetlb_entry_migration(pte))) {
4423 swp_entry_t entry = pte_to_swp_entry(pte);
4424
4425 if (is_write_migration_entry(entry)) {
4426 pte_t newpte;
4427
4428 make_migration_entry_read(&entry);
4429 newpte = swp_entry_to_pte(entry);
e5251fd4
PA
4430 set_huge_swap_pte_at(mm, address, ptep,
4431 newpte, huge_page_size(h));
a8bda28d
NH
4432 pages++;
4433 }
4434 spin_unlock(ptl);
4435 continue;
4436 }
4437 if (!huge_pte_none(pte)) {
023bdd00
AK
4438 pte_t old_pte;
4439
4440 old_pte = huge_ptep_modify_prot_start(vma, address, ptep);
4441 pte = pte_mkhuge(huge_pte_modify(old_pte, newprot));
be7517d6 4442 pte = arch_make_huge_pte(pte, vma, NULL, 0);
023bdd00 4443 huge_ptep_modify_prot_commit(vma, address, ptep, old_pte, pte);
7da4d641 4444 pages++;
8f860591 4445 }
cb900f41 4446 spin_unlock(ptl);
8f860591 4447 }
d833352a 4448 /*
c8c06efa 4449 * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare
d833352a 4450 * may have cleared our pud entry and done put_page on the page table:
c8c06efa 4451 * once we release i_mmap_rwsem, another task can do the final put_page
dff11abe
MK
4452 * and that page table be reused and filled with junk. If we actually
4453 * did unshare a page of pmds, flush the range corresponding to the pud.
d833352a 4454 */
dff11abe 4455 if (shared_pmd)
ac46d4f3 4456 flush_hugetlb_tlb_range(vma, range.start, range.end);
dff11abe
MK
4457 else
4458 flush_hugetlb_tlb_range(vma, start, end);
0f10851e
JG
4459 /*
4460 * No need to call mmu_notifier_invalidate_range() we are downgrading
4461 * page table protection not changing it to point to a new page.
4462 *
ad56b738 4463 * See Documentation/vm/mmu_notifier.rst
0f10851e 4464 */
83cde9e8 4465 i_mmap_unlock_write(vma->vm_file->f_mapping);
ac46d4f3 4466 mmu_notifier_invalidate_range_end(&range);
7da4d641
PZ
4467
4468 return pages << h->order;
8f860591
ZY
4469}
4470
a1e78772
MG
4471int hugetlb_reserve_pages(struct inode *inode,
4472 long from, long to,
5a6fe125 4473 struct vm_area_struct *vma,
ca16d140 4474 vm_flags_t vm_flags)
e4e574b7 4475{
17c9d12e 4476 long ret, chg;
a5516438 4477 struct hstate *h = hstate_inode(inode);
90481622 4478 struct hugepage_subpool *spool = subpool_inode(inode);
9119a41e 4479 struct resv_map *resv_map;
1c5ecae3 4480 long gbl_reserve;
e4e574b7 4481
63489f8e
MK
4482 /* This should never happen */
4483 if (from > to) {
4484 VM_WARN(1, "%s called with a negative range\n", __func__);
4485 return -EINVAL;
4486 }
4487
17c9d12e
MG
4488 /*
4489 * Only apply hugepage reservation if asked. At fault time, an
4490 * attempt will be made for VM_NORESERVE to allocate a page
90481622 4491 * without using reserves
17c9d12e 4492 */
ca16d140 4493 if (vm_flags & VM_NORESERVE)
17c9d12e
MG
4494 return 0;
4495
a1e78772
MG
4496 /*
4497 * Shared mappings base their reservation on the number of pages that
4498 * are already allocated on behalf of the file. Private mappings need
4499 * to reserve the full area even if read-only as mprotect() may be
4500 * called to make the mapping read-write. Assume !vma is a shm mapping
4501 */
9119a41e 4502 if (!vma || vma->vm_flags & VM_MAYSHARE) {
4e35f483 4503 resv_map = inode_resv_map(inode);
9119a41e 4504
1406ec9b 4505 chg = region_chg(resv_map, from, to);
9119a41e
JK
4506
4507 } else {
4508 resv_map = resv_map_alloc();
17c9d12e
MG
4509 if (!resv_map)
4510 return -ENOMEM;
4511
a1e78772 4512 chg = to - from;
84afd99b 4513
17c9d12e
MG
4514 set_vma_resv_map(vma, resv_map);
4515 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
4516 }
4517
c50ac050
DH
4518 if (chg < 0) {
4519 ret = chg;
4520 goto out_err;
4521 }
8a630112 4522
1c5ecae3
MK
4523 /*
4524 * There must be enough pages in the subpool for the mapping. If
4525 * the subpool has a minimum size, there may be some global
4526 * reservations already in place (gbl_reserve).
4527 */
4528 gbl_reserve = hugepage_subpool_get_pages(spool, chg);
4529 if (gbl_reserve < 0) {
c50ac050
DH
4530 ret = -ENOSPC;
4531 goto out_err;
4532 }
5a6fe125
MG
4533
4534 /*
17c9d12e 4535 * Check enough hugepages are available for the reservation.
90481622 4536 * Hand the pages back to the subpool if there are not
5a6fe125 4537 */
1c5ecae3 4538 ret = hugetlb_acct_memory(h, gbl_reserve);
68842c9b 4539 if (ret < 0) {
1c5ecae3
MK
4540 /* put back original number of pages, chg */
4541 (void)hugepage_subpool_put_pages(spool, chg);
c50ac050 4542 goto out_err;
68842c9b 4543 }
17c9d12e
MG
4544
4545 /*
4546 * Account for the reservations made. Shared mappings record regions
4547 * that have reservations as they are shared by multiple VMAs.
4548 * When the last VMA disappears, the region map says how much
4549 * the reservation was and the page cache tells how much of
4550 * the reservation was consumed. Private mappings are per-VMA and
4551 * only the consumed reservations are tracked. When the VMA
4552 * disappears, the original reservation is the VMA size and the
4553 * consumed reservations are stored in the map. Hence, nothing
4554 * else has to be done for private mappings here
4555 */
33039678
MK
4556 if (!vma || vma->vm_flags & VM_MAYSHARE) {
4557 long add = region_add(resv_map, from, to);
4558
4559 if (unlikely(chg > add)) {
4560 /*
4561 * pages in this range were added to the reserve
4562 * map between region_chg and region_add. This
4563 * indicates a race with alloc_huge_page. Adjust
4564 * the subpool and reserve counts modified above
4565 * based on the difference.
4566 */
4567 long rsv_adjust;
4568
4569 rsv_adjust = hugepage_subpool_put_pages(spool,
4570 chg - add);
4571 hugetlb_acct_memory(h, -rsv_adjust);
4572 }
4573 }
a43a8c39 4574 return 0;
c50ac050 4575out_err:
5e911373 4576 if (!vma || vma->vm_flags & VM_MAYSHARE)
ff8c0c53
MK
4577 /* Don't call region_abort if region_chg failed */
4578 if (chg >= 0)
4579 region_abort(resv_map, from, to);
f031dd27
JK
4580 if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
4581 kref_put(&resv_map->refs, resv_map_release);
c50ac050 4582 return ret;
a43a8c39
KC
4583}
4584
b5cec28d
MK
4585long hugetlb_unreserve_pages(struct inode *inode, long start, long end,
4586 long freed)
a43a8c39 4587{
a5516438 4588 struct hstate *h = hstate_inode(inode);
4e35f483 4589 struct resv_map *resv_map = inode_resv_map(inode);
9119a41e 4590 long chg = 0;
90481622 4591 struct hugepage_subpool *spool = subpool_inode(inode);
1c5ecae3 4592 long gbl_reserve;
45c682a6 4593
b5cec28d
MK
4594 if (resv_map) {
4595 chg = region_del(resv_map, start, end);
4596 /*
4597 * region_del() can fail in the rare case where a region
4598 * must be split and another region descriptor can not be
4599 * allocated. If end == LONG_MAX, it will not fail.
4600 */
4601 if (chg < 0)
4602 return chg;
4603 }
4604
45c682a6 4605 spin_lock(&inode->i_lock);
e4c6f8be 4606 inode->i_blocks -= (blocks_per_huge_page(h) * freed);
45c682a6
KC
4607 spin_unlock(&inode->i_lock);
4608
1c5ecae3
MK
4609 /*
4610 * If the subpool has a minimum size, the number of global
4611 * reservations to be released may be adjusted.
4612 */
4613 gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed));
4614 hugetlb_acct_memory(h, -gbl_reserve);
b5cec28d
MK
4615
4616 return 0;
a43a8c39 4617}
93f70f90 4618
3212b535
SC
4619#ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
4620static unsigned long page_table_shareable(struct vm_area_struct *svma,
4621 struct vm_area_struct *vma,
4622 unsigned long addr, pgoff_t idx)
4623{
4624 unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) +
4625 svma->vm_start;
4626 unsigned long sbase = saddr & PUD_MASK;
4627 unsigned long s_end = sbase + PUD_SIZE;
4628
4629 /* Allow segments to share if only one is marked locked */
de60f5f1
EM
4630 unsigned long vm_flags = vma->vm_flags & VM_LOCKED_CLEAR_MASK;
4631 unsigned long svm_flags = svma->vm_flags & VM_LOCKED_CLEAR_MASK;
3212b535
SC
4632
4633 /*
4634 * match the virtual addresses, permission and the alignment of the
4635 * page table page.
4636 */
4637 if (pmd_index(addr) != pmd_index(saddr) ||
4638 vm_flags != svm_flags ||
4639 sbase < svma->vm_start || svma->vm_end < s_end)
4640 return 0;
4641
4642 return saddr;
4643}
4644
31aafb45 4645static bool vma_shareable(struct vm_area_struct *vma, unsigned long addr)
3212b535
SC
4646{
4647 unsigned long base = addr & PUD_MASK;
4648 unsigned long end = base + PUD_SIZE;
4649
4650 /*
4651 * check on proper vm_flags and page table alignment
4652 */
017b1660 4653 if (vma->vm_flags & VM_MAYSHARE && range_in_vma(vma, base, end))
31aafb45
NK
4654 return true;
4655 return false;
3212b535
SC
4656}
4657
017b1660
MK
4658/*
4659 * Determine if start,end range within vma could be mapped by shared pmd.
4660 * If yes, adjust start and end to cover range associated with possible
4661 * shared pmd mappings.
4662 */
4663void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma,
4664 unsigned long *start, unsigned long *end)
4665{
4666 unsigned long check_addr = *start;
4667
4668 if (!(vma->vm_flags & VM_MAYSHARE))
4669 return;
4670
4671 for (check_addr = *start; check_addr < *end; check_addr += PUD_SIZE) {
4672 unsigned long a_start = check_addr & PUD_MASK;
4673 unsigned long a_end = a_start + PUD_SIZE;
4674
4675 /*
4676 * If sharing is possible, adjust start/end if necessary.
4677 */
4678 if (range_in_vma(vma, a_start, a_end)) {
4679 if (a_start < *start)
4680 *start = a_start;
4681 if (a_end > *end)
4682 *end = a_end;
4683 }
4684 }
4685}
4686
3212b535
SC
4687/*
4688 * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
4689 * and returns the corresponding pte. While this is not necessary for the
4690 * !shared pmd case because we can allocate the pmd later as well, it makes the
ddeaab32
MK
4691 * code much cleaner. pmd allocation is essential for the shared case because
4692 * pud has to be populated inside the same i_mmap_rwsem section - otherwise
4693 * racing tasks could either miss the sharing (see huge_pte_offset) or select a
4694 * bad pmd for sharing.
3212b535
SC
4695 */
4696pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
4697{
4698 struct vm_area_struct *vma = find_vma(mm, addr);
4699 struct address_space *mapping = vma->vm_file->f_mapping;
4700 pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) +
4701 vma->vm_pgoff;
4702 struct vm_area_struct *svma;
4703 unsigned long saddr;
4704 pte_t *spte = NULL;
4705 pte_t *pte;
cb900f41 4706 spinlock_t *ptl;
3212b535
SC
4707
4708 if (!vma_shareable(vma, addr))
4709 return (pte_t *)pmd_alloc(mm, pud, addr);
4710
ddeaab32 4711 i_mmap_lock_write(mapping);
3212b535
SC
4712 vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) {
4713 if (svma == vma)
4714 continue;
4715
4716 saddr = page_table_shareable(svma, vma, addr, idx);
4717 if (saddr) {
7868a208
PA
4718 spte = huge_pte_offset(svma->vm_mm, saddr,
4719 vma_mmu_pagesize(svma));
3212b535
SC
4720 if (spte) {
4721 get_page(virt_to_page(spte));
4722 break;
4723 }
4724 }
4725 }
4726
4727 if (!spte)
4728 goto out;
4729
8bea8052 4730 ptl = huge_pte_lock(hstate_vma(vma), mm, spte);
dc6c9a35 4731 if (pud_none(*pud)) {
3212b535
SC
4732 pud_populate(mm, pud,
4733 (pmd_t *)((unsigned long)spte & PAGE_MASK));
c17b1f42 4734 mm_inc_nr_pmds(mm);
dc6c9a35 4735 } else {
3212b535 4736 put_page(virt_to_page(spte));
dc6c9a35 4737 }
cb900f41 4738 spin_unlock(ptl);
3212b535
SC
4739out:
4740 pte = (pte_t *)pmd_alloc(mm, pud, addr);
ddeaab32 4741 i_mmap_unlock_write(mapping);
3212b535
SC
4742 return pte;
4743}
4744
4745/*
4746 * unmap huge page backed by shared pte.
4747 *
4748 * Hugetlb pte page is ref counted at the time of mapping. If pte is shared
4749 * indicated by page_count > 1, unmap is achieved by clearing pud and
4750 * decrementing the ref count. If count == 1, the pte page is not shared.
4751 *
ddeaab32 4752 * called with page table lock held.
3212b535
SC
4753 *
4754 * returns: 1 successfully unmapped a shared pte page
4755 * 0 the underlying pte page is not shared, or it is the last user
4756 */
4757int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
4758{
4759 pgd_t *pgd = pgd_offset(mm, *addr);
c2febafc
KS
4760 p4d_t *p4d = p4d_offset(pgd, *addr);
4761 pud_t *pud = pud_offset(p4d, *addr);
3212b535
SC
4762
4763 BUG_ON(page_count(virt_to_page(ptep)) == 0);
4764 if (page_count(virt_to_page(ptep)) == 1)
4765 return 0;
4766
4767 pud_clear(pud);
4768 put_page(virt_to_page(ptep));
dc6c9a35 4769 mm_dec_nr_pmds(mm);
3212b535
SC
4770 *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE;
4771 return 1;
4772}
9e5fc74c
SC
4773#define want_pmd_share() (1)
4774#else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
4775pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
4776{
4777 return NULL;
4778}
e81f2d22
ZZ
4779
4780int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
4781{
4782 return 0;
4783}
017b1660
MK
4784
4785void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma,
4786 unsigned long *start, unsigned long *end)
4787{
4788}
9e5fc74c 4789#define want_pmd_share() (0)
3212b535
SC
4790#endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
4791
9e5fc74c
SC
4792#ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
4793pte_t *huge_pte_alloc(struct mm_struct *mm,
4794 unsigned long addr, unsigned long sz)
4795{
4796 pgd_t *pgd;
c2febafc 4797 p4d_t *p4d;
9e5fc74c
SC
4798 pud_t *pud;
4799 pte_t *pte = NULL;
4800
4801 pgd = pgd_offset(mm, addr);
f4f0a3d8
KS
4802 p4d = p4d_alloc(mm, pgd, addr);
4803 if (!p4d)
4804 return NULL;
c2febafc 4805 pud = pud_alloc(mm, p4d, addr);
9e5fc74c
SC
4806 if (pud) {
4807 if (sz == PUD_SIZE) {
4808 pte = (pte_t *)pud;
4809 } else {
4810 BUG_ON(sz != PMD_SIZE);
4811 if (want_pmd_share() && pud_none(*pud))
4812 pte = huge_pmd_share(mm, addr, pud);
4813 else
4814 pte = (pte_t *)pmd_alloc(mm, pud, addr);
4815 }
4816 }
4e666314 4817 BUG_ON(pte && pte_present(*pte) && !pte_huge(*pte));
9e5fc74c
SC
4818
4819 return pte;
4820}
4821
9b19df29
PA
4822/*
4823 * huge_pte_offset() - Walk the page table to resolve the hugepage
4824 * entry at address @addr
4825 *
4826 * Return: Pointer to page table or swap entry (PUD or PMD) for
4827 * address @addr, or NULL if a p*d_none() entry is encountered and the
4828 * size @sz doesn't match the hugepage size at this level of the page
4829 * table.
4830 */
7868a208
PA
4831pte_t *huge_pte_offset(struct mm_struct *mm,
4832 unsigned long addr, unsigned long sz)
9e5fc74c
SC
4833{
4834 pgd_t *pgd;
c2febafc 4835 p4d_t *p4d;
9e5fc74c 4836 pud_t *pud;
c2febafc 4837 pmd_t *pmd;
9e5fc74c
SC
4838
4839 pgd = pgd_offset(mm, addr);
c2febafc
KS
4840 if (!pgd_present(*pgd))
4841 return NULL;
4842 p4d = p4d_offset(pgd, addr);
4843 if (!p4d_present(*p4d))
4844 return NULL;
9b19df29 4845
c2febafc 4846 pud = pud_offset(p4d, addr);
9b19df29 4847 if (sz != PUD_SIZE && pud_none(*pud))
c2febafc 4848 return NULL;
9b19df29
PA
4849 /* hugepage or swap? */
4850 if (pud_huge(*pud) || !pud_present(*pud))
c2febafc 4851 return (pte_t *)pud;
9b19df29 4852
c2febafc 4853 pmd = pmd_offset(pud, addr);
9b19df29
PA
4854 if (sz != PMD_SIZE && pmd_none(*pmd))
4855 return NULL;
4856 /* hugepage or swap? */
4857 if (pmd_huge(*pmd) || !pmd_present(*pmd))
4858 return (pte_t *)pmd;
4859
4860 return NULL;
9e5fc74c
SC
4861}
4862
61f77eda
NH
4863#endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
4864
4865/*
4866 * These functions are overwritable if your architecture needs its own
4867 * behavior.
4868 */
4869struct page * __weak
4870follow_huge_addr(struct mm_struct *mm, unsigned long address,
4871 int write)
4872{
4873 return ERR_PTR(-EINVAL);
4874}
4875
4dc71451
AK
4876struct page * __weak
4877follow_huge_pd(struct vm_area_struct *vma,
4878 unsigned long address, hugepd_t hpd, int flags, int pdshift)
4879{
4880 WARN(1, "hugepd follow called with no support for hugepage directory format\n");
4881 return NULL;
4882}
4883
61f77eda 4884struct page * __weak
9e5fc74c 4885follow_huge_pmd(struct mm_struct *mm, unsigned long address,
e66f17ff 4886 pmd_t *pmd, int flags)
9e5fc74c 4887{
e66f17ff
NH
4888 struct page *page = NULL;
4889 spinlock_t *ptl;
c9d398fa 4890 pte_t pte;
e66f17ff
NH
4891retry:
4892 ptl = pmd_lockptr(mm, pmd);
4893 spin_lock(ptl);
4894 /*
4895 * make sure that the address range covered by this pmd is not
4896 * unmapped from other threads.
4897 */
4898 if (!pmd_huge(*pmd))
4899 goto out;
c9d398fa
NH
4900 pte = huge_ptep_get((pte_t *)pmd);
4901 if (pte_present(pte)) {
97534127 4902 page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT);
e66f17ff
NH
4903 if (flags & FOLL_GET)
4904 get_page(page);
4905 } else {
c9d398fa 4906 if (is_hugetlb_entry_migration(pte)) {
e66f17ff
NH
4907 spin_unlock(ptl);
4908 __migration_entry_wait(mm, (pte_t *)pmd, ptl);
4909 goto retry;
4910 }
4911 /*
4912 * hwpoisoned entry is treated as no_page_table in
4913 * follow_page_mask().
4914 */
4915 }
4916out:
4917 spin_unlock(ptl);
9e5fc74c
SC
4918 return page;
4919}
4920
61f77eda 4921struct page * __weak
9e5fc74c 4922follow_huge_pud(struct mm_struct *mm, unsigned long address,
e66f17ff 4923 pud_t *pud, int flags)
9e5fc74c 4924{
e66f17ff
NH
4925 if (flags & FOLL_GET)
4926 return NULL;
9e5fc74c 4927
e66f17ff 4928 return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT);
9e5fc74c
SC
4929}
4930
faaa5b62
AK
4931struct page * __weak
4932follow_huge_pgd(struct mm_struct *mm, unsigned long address, pgd_t *pgd, int flags)
4933{
4934 if (flags & FOLL_GET)
4935 return NULL;
4936
4937 return pte_page(*(pte_t *)pgd) + ((address & ~PGDIR_MASK) >> PAGE_SHIFT);
4938}
4939
31caf665
NH
4940bool isolate_huge_page(struct page *page, struct list_head *list)
4941{
bcc54222
NH
4942 bool ret = true;
4943
309381fe 4944 VM_BUG_ON_PAGE(!PageHead(page), page);
31caf665 4945 spin_lock(&hugetlb_lock);
bcc54222
NH
4946 if (!page_huge_active(page) || !get_page_unless_zero(page)) {
4947 ret = false;
4948 goto unlock;
4949 }
4950 clear_page_huge_active(page);
31caf665 4951 list_move_tail(&page->lru, list);
bcc54222 4952unlock:
31caf665 4953 spin_unlock(&hugetlb_lock);
bcc54222 4954 return ret;
31caf665
NH
4955}
4956
4957void putback_active_hugepage(struct page *page)
4958{
309381fe 4959 VM_BUG_ON_PAGE(!PageHead(page), page);
31caf665 4960 spin_lock(&hugetlb_lock);
bcc54222 4961 set_page_huge_active(page);
31caf665
NH
4962 list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist);
4963 spin_unlock(&hugetlb_lock);
4964 put_page(page);
4965}
ab5ac90a
MH
4966
4967void move_hugetlb_state(struct page *oldpage, struct page *newpage, int reason)
4968{
4969 struct hstate *h = page_hstate(oldpage);
4970
4971 hugetlb_cgroup_migrate(oldpage, newpage);
4972 set_page_owner_migrate_reason(newpage, reason);
4973
4974 /*
4975 * transfer temporary state of the new huge page. This is
4976 * reverse to other transitions because the newpage is going to
4977 * be final while the old one will be freed so it takes over
4978 * the temporary status.
4979 *
4980 * Also note that we have to transfer the per-node surplus state
4981 * here as well otherwise the global surplus count will not match
4982 * the per-node's.
4983 */
4984 if (PageHugeTemporary(newpage)) {
4985 int old_nid = page_to_nid(oldpage);
4986 int new_nid = page_to_nid(newpage);
4987
4988 SetPageHugeTemporary(oldpage);
4989 ClearPageHugeTemporary(newpage);
4990
4991 spin_lock(&hugetlb_lock);
4992 if (h->surplus_huge_pages_node[old_nid]) {
4993 h->surplus_huge_pages_node[old_nid]--;
4994 h->surplus_huge_pages_node[new_nid]++;
4995 }
4996 spin_unlock(&hugetlb_lock);
4997 }
4998}