mm/hugetlb: vma_has_reserves() needs to handle fallocate hole punch
[linux-2.6-block.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>
7#include <linux/module.h>
8#include <linux/mm.h>
e1759c21 9#include <linux/seq_file.h>
1da177e4
LT
10#include <linux/sysctl.h>
11#include <linux/highmem.h>
cddb8a5c 12#include <linux/mmu_notifier.h>
1da177e4 13#include <linux/nodemask.h>
63551ae0 14#include <linux/pagemap.h>
5da7ca86 15#include <linux/mempolicy.h>
3b32123d 16#include <linux/compiler.h>
aea47ff3 17#include <linux/cpuset.h>
3935baa9 18#include <linux/mutex.h>
aa888a74 19#include <linux/bootmem.h>
a3437870 20#include <linux/sysfs.h>
5a0e3ad6 21#include <linux/slab.h>
0fe6e20b 22#include <linux/rmap.h>
fd6a03ed
NH
23#include <linux/swap.h>
24#include <linux/swapops.h>
c8721bbb 25#include <linux/page-isolation.h>
8382d914 26#include <linux/jhash.h>
d6606683 27
63551ae0
DG
28#include <asm/page.h>
29#include <asm/pgtable.h>
24669e58 30#include <asm/tlb.h>
63551ae0 31
24669e58 32#include <linux/io.h>
63551ae0 33#include <linux/hugetlb.h>
9dd540e2 34#include <linux/hugetlb_cgroup.h>
9a305230 35#include <linux/node.h>
7835e98b 36#include "internal.h"
1da177e4 37
753162cd 38int hugepages_treat_as_movable;
a5516438 39
c3f38a38 40int hugetlb_max_hstate __read_mostly;
e5ff2159
AK
41unsigned int default_hstate_idx;
42struct hstate hstates[HUGE_MAX_HSTATE];
641844f5
NH
43/*
44 * Minimum page order among possible hugepage sizes, set to a proper value
45 * at boot time.
46 */
47static unsigned int minimum_order __read_mostly = UINT_MAX;
e5ff2159 48
53ba51d2
JT
49__initdata LIST_HEAD(huge_boot_pages);
50
e5ff2159
AK
51/* for command line parsing */
52static struct hstate * __initdata parsed_hstate;
53static unsigned long __initdata default_hstate_max_huge_pages;
e11bfbfc 54static unsigned long __initdata default_hstate_size;
e5ff2159 55
3935baa9 56/*
31caf665
NH
57 * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages,
58 * free_huge_pages, and surplus_huge_pages.
3935baa9 59 */
c3f38a38 60DEFINE_SPINLOCK(hugetlb_lock);
0bd0f9fb 61
8382d914
DB
62/*
63 * Serializes faults on the same logical page. This is used to
64 * prevent spurious OOMs when the hugepage pool is fully utilized.
65 */
66static int num_fault_mutexes;
c672c7f2 67struct mutex *hugetlb_fault_mutex_table ____cacheline_aligned_in_smp;
8382d914 68
7ca02d0a
MK
69/* Forward declaration */
70static int hugetlb_acct_memory(struct hstate *h, long delta);
71
90481622
DG
72static inline void unlock_or_release_subpool(struct hugepage_subpool *spool)
73{
74 bool free = (spool->count == 0) && (spool->used_hpages == 0);
75
76 spin_unlock(&spool->lock);
77
78 /* If no pages are used, and no other handles to the subpool
7ca02d0a
MK
79 * remain, give up any reservations mased on minimum size and
80 * free the subpool */
81 if (free) {
82 if (spool->min_hpages != -1)
83 hugetlb_acct_memory(spool->hstate,
84 -spool->min_hpages);
90481622 85 kfree(spool);
7ca02d0a 86 }
90481622
DG
87}
88
7ca02d0a
MK
89struct hugepage_subpool *hugepage_new_subpool(struct hstate *h, long max_hpages,
90 long min_hpages)
90481622
DG
91{
92 struct hugepage_subpool *spool;
93
c6a91820 94 spool = kzalloc(sizeof(*spool), GFP_KERNEL);
90481622
DG
95 if (!spool)
96 return NULL;
97
98 spin_lock_init(&spool->lock);
99 spool->count = 1;
7ca02d0a
MK
100 spool->max_hpages = max_hpages;
101 spool->hstate = h;
102 spool->min_hpages = min_hpages;
103
104 if (min_hpages != -1 && hugetlb_acct_memory(h, min_hpages)) {
105 kfree(spool);
106 return NULL;
107 }
108 spool->rsv_hpages = min_hpages;
90481622
DG
109
110 return spool;
111}
112
113void hugepage_put_subpool(struct hugepage_subpool *spool)
114{
115 spin_lock(&spool->lock);
116 BUG_ON(!spool->count);
117 spool->count--;
118 unlock_or_release_subpool(spool);
119}
120
1c5ecae3
MK
121/*
122 * Subpool accounting for allocating and reserving pages.
123 * Return -ENOMEM if there are not enough resources to satisfy the
124 * the request. Otherwise, return the number of pages by which the
125 * global pools must be adjusted (upward). The returned value may
126 * only be different than the passed value (delta) in the case where
127 * a subpool minimum size must be manitained.
128 */
129static long hugepage_subpool_get_pages(struct hugepage_subpool *spool,
90481622
DG
130 long delta)
131{
1c5ecae3 132 long ret = delta;
90481622
DG
133
134 if (!spool)
1c5ecae3 135 return ret;
90481622
DG
136
137 spin_lock(&spool->lock);
1c5ecae3
MK
138
139 if (spool->max_hpages != -1) { /* maximum size accounting */
140 if ((spool->used_hpages + delta) <= spool->max_hpages)
141 spool->used_hpages += delta;
142 else {
143 ret = -ENOMEM;
144 goto unlock_ret;
145 }
90481622 146 }
90481622 147
1c5ecae3
MK
148 if (spool->min_hpages != -1) { /* minimum size accounting */
149 if (delta > spool->rsv_hpages) {
150 /*
151 * Asking for more reserves than those already taken on
152 * behalf of subpool. Return difference.
153 */
154 ret = delta - spool->rsv_hpages;
155 spool->rsv_hpages = 0;
156 } else {
157 ret = 0; /* reserves already accounted for */
158 spool->rsv_hpages -= delta;
159 }
160 }
161
162unlock_ret:
163 spin_unlock(&spool->lock);
90481622
DG
164 return ret;
165}
166
1c5ecae3
MK
167/*
168 * Subpool accounting for freeing and unreserving pages.
169 * Return the number of global page reservations that must be dropped.
170 * The return value may only be different than the passed value (delta)
171 * in the case where a subpool minimum size must be maintained.
172 */
173static long hugepage_subpool_put_pages(struct hugepage_subpool *spool,
90481622
DG
174 long delta)
175{
1c5ecae3
MK
176 long ret = delta;
177
90481622 178 if (!spool)
1c5ecae3 179 return delta;
90481622
DG
180
181 spin_lock(&spool->lock);
1c5ecae3
MK
182
183 if (spool->max_hpages != -1) /* maximum size accounting */
184 spool->used_hpages -= delta;
185
186 if (spool->min_hpages != -1) { /* minimum size accounting */
187 if (spool->rsv_hpages + delta <= spool->min_hpages)
188 ret = 0;
189 else
190 ret = spool->rsv_hpages + delta - spool->min_hpages;
191
192 spool->rsv_hpages += delta;
193 if (spool->rsv_hpages > spool->min_hpages)
194 spool->rsv_hpages = spool->min_hpages;
195 }
196
197 /*
198 * If hugetlbfs_put_super couldn't free spool due to an outstanding
199 * quota reference, free it now.
200 */
90481622 201 unlock_or_release_subpool(spool);
1c5ecae3
MK
202
203 return ret;
90481622
DG
204}
205
206static inline struct hugepage_subpool *subpool_inode(struct inode *inode)
207{
208 return HUGETLBFS_SB(inode->i_sb)->spool;
209}
210
211static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma)
212{
496ad9aa 213 return subpool_inode(file_inode(vma->vm_file));
90481622
DG
214}
215
96822904
AW
216/*
217 * Region tracking -- allows tracking of reservations and instantiated pages
218 * across the pages in a mapping.
84afd99b 219 *
1dd308a7
MK
220 * The region data structures are embedded into a resv_map and protected
221 * by a resv_map's lock. The set of regions within the resv_map represent
222 * reservations for huge pages, or huge pages that have already been
223 * instantiated within the map. The from and to elements are huge page
224 * indicies into the associated mapping. from indicates the starting index
225 * of the region. to represents the first index past the end of the region.
226 *
227 * For example, a file region structure with from == 0 and to == 4 represents
228 * four huge pages in a mapping. It is important to note that the to element
229 * represents the first element past the end of the region. This is used in
230 * arithmetic as 4(to) - 0(from) = 4 huge pages in the region.
231 *
232 * Interval notation of the form [from, to) will be used to indicate that
233 * the endpoint from is inclusive and to is exclusive.
96822904
AW
234 */
235struct file_region {
236 struct list_head link;
237 long from;
238 long to;
239};
240
1dd308a7
MK
241/*
242 * Add the huge page range represented by [f, t) to the reserve
5e911373
MK
243 * map. In the normal case, existing regions will be expanded
244 * to accommodate the specified range. Sufficient regions should
245 * exist for expansion due to the previous call to region_chg
246 * with the same range. However, it is possible that region_del
247 * could have been called after region_chg and modifed the map
248 * in such a way that no region exists to be expanded. In this
249 * case, pull a region descriptor from the cache associated with
250 * the map and use that for the new range.
cf3ad20b
MK
251 *
252 * Return the number of new huge pages added to the map. This
253 * number is greater than or equal to zero.
1dd308a7 254 */
1406ec9b 255static long region_add(struct resv_map *resv, long f, long t)
96822904 256{
1406ec9b 257 struct list_head *head = &resv->regions;
96822904 258 struct file_region *rg, *nrg, *trg;
cf3ad20b 259 long add = 0;
96822904 260
7b24d861 261 spin_lock(&resv->lock);
96822904
AW
262 /* Locate the region we are either in or before. */
263 list_for_each_entry(rg, head, link)
264 if (f <= rg->to)
265 break;
266
5e911373
MK
267 /*
268 * If no region exists which can be expanded to include the
269 * specified range, the list must have been modified by an
270 * interleving call to region_del(). Pull a region descriptor
271 * from the cache and use it for this range.
272 */
273 if (&rg->link == head || t < rg->from) {
274 VM_BUG_ON(resv->region_cache_count <= 0);
275
276 resv->region_cache_count--;
277 nrg = list_first_entry(&resv->region_cache, struct file_region,
278 link);
279 list_del(&nrg->link);
280
281 nrg->from = f;
282 nrg->to = t;
283 list_add(&nrg->link, rg->link.prev);
284
285 add += t - f;
286 goto out_locked;
287 }
288
96822904
AW
289 /* Round our left edge to the current segment if it encloses us. */
290 if (f > rg->from)
291 f = rg->from;
292
293 /* Check for and consume any regions we now overlap with. */
294 nrg = rg;
295 list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
296 if (&rg->link == head)
297 break;
298 if (rg->from > t)
299 break;
300
301 /* If this area reaches higher then extend our area to
302 * include it completely. If this is not the first area
303 * which we intend to reuse, free it. */
304 if (rg->to > t)
305 t = rg->to;
306 if (rg != nrg) {
cf3ad20b
MK
307 /* Decrement return value by the deleted range.
308 * Another range will span this area so that by
309 * end of routine add will be >= zero
310 */
311 add -= (rg->to - rg->from);
96822904
AW
312 list_del(&rg->link);
313 kfree(rg);
314 }
315 }
cf3ad20b
MK
316
317 add += (nrg->from - f); /* Added to beginning of region */
96822904 318 nrg->from = f;
cf3ad20b 319 add += t - nrg->to; /* Added to end of region */
96822904 320 nrg->to = t;
cf3ad20b 321
5e911373
MK
322out_locked:
323 resv->adds_in_progress--;
7b24d861 324 spin_unlock(&resv->lock);
cf3ad20b
MK
325 VM_BUG_ON(add < 0);
326 return add;
96822904
AW
327}
328
1dd308a7
MK
329/*
330 * Examine the existing reserve map and determine how many
331 * huge pages in the specified range [f, t) are NOT currently
332 * represented. This routine is called before a subsequent
333 * call to region_add that will actually modify the reserve
334 * map to add the specified range [f, t). region_chg does
335 * not change the number of huge pages represented by the
336 * map. However, if the existing regions in the map can not
337 * be expanded to represent the new range, a new file_region
338 * structure is added to the map as a placeholder. This is
339 * so that the subsequent region_add call will have all the
340 * regions it needs and will not fail.
341 *
5e911373
MK
342 * Upon entry, region_chg will also examine the cache of region descriptors
343 * associated with the map. If there are not enough descriptors cached, one
344 * will be allocated for the in progress add operation.
345 *
346 * Returns the number of huge pages that need to be added to the existing
347 * reservation map for the range [f, t). This number is greater or equal to
348 * zero. -ENOMEM is returned if a new file_region structure or cache entry
349 * is needed and can not be allocated.
1dd308a7 350 */
1406ec9b 351static long region_chg(struct resv_map *resv, long f, long t)
96822904 352{
1406ec9b 353 struct list_head *head = &resv->regions;
7b24d861 354 struct file_region *rg, *nrg = NULL;
96822904
AW
355 long chg = 0;
356
7b24d861
DB
357retry:
358 spin_lock(&resv->lock);
5e911373
MK
359retry_locked:
360 resv->adds_in_progress++;
361
362 /*
363 * Check for sufficient descriptors in the cache to accommodate
364 * the number of in progress add operations.
365 */
366 if (resv->adds_in_progress > resv->region_cache_count) {
367 struct file_region *trg;
368
369 VM_BUG_ON(resv->adds_in_progress - resv->region_cache_count > 1);
370 /* Must drop lock to allocate a new descriptor. */
371 resv->adds_in_progress--;
372 spin_unlock(&resv->lock);
373
374 trg = kmalloc(sizeof(*trg), GFP_KERNEL);
375 if (!trg)
376 return -ENOMEM;
377
378 spin_lock(&resv->lock);
379 list_add(&trg->link, &resv->region_cache);
380 resv->region_cache_count++;
381 goto retry_locked;
382 }
383
96822904
AW
384 /* Locate the region we are before or in. */
385 list_for_each_entry(rg, head, link)
386 if (f <= rg->to)
387 break;
388
389 /* If we are below the current region then a new region is required.
390 * Subtle, allocate a new region at the position but make it zero
391 * size such that we can guarantee to record the reservation. */
392 if (&rg->link == head || t < rg->from) {
7b24d861 393 if (!nrg) {
5e911373 394 resv->adds_in_progress--;
7b24d861
DB
395 spin_unlock(&resv->lock);
396 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
397 if (!nrg)
398 return -ENOMEM;
399
400 nrg->from = f;
401 nrg->to = f;
402 INIT_LIST_HEAD(&nrg->link);
403 goto retry;
404 }
96822904 405
7b24d861
DB
406 list_add(&nrg->link, rg->link.prev);
407 chg = t - f;
408 goto out_nrg;
96822904
AW
409 }
410
411 /* Round our left edge to the current segment if it encloses us. */
412 if (f > rg->from)
413 f = rg->from;
414 chg = t - f;
415
416 /* Check for and consume any regions we now overlap with. */
417 list_for_each_entry(rg, rg->link.prev, link) {
418 if (&rg->link == head)
419 break;
420 if (rg->from > t)
7b24d861 421 goto out;
96822904 422
25985edc 423 /* We overlap with this area, if it extends further than
96822904
AW
424 * us then we must extend ourselves. Account for its
425 * existing reservation. */
426 if (rg->to > t) {
427 chg += rg->to - t;
428 t = rg->to;
429 }
430 chg -= rg->to - rg->from;
431 }
7b24d861
DB
432
433out:
434 spin_unlock(&resv->lock);
435 /* We already know we raced and no longer need the new region */
436 kfree(nrg);
437 return chg;
438out_nrg:
439 spin_unlock(&resv->lock);
96822904
AW
440 return chg;
441}
442
5e911373
MK
443/*
444 * Abort the in progress add operation. The adds_in_progress field
445 * of the resv_map keeps track of the operations in progress between
446 * calls to region_chg and region_add. Operations are sometimes
447 * aborted after the call to region_chg. In such cases, region_abort
448 * is called to decrement the adds_in_progress counter.
449 *
450 * NOTE: The range arguments [f, t) are not needed or used in this
451 * routine. They are kept to make reading the calling code easier as
452 * arguments will match the associated region_chg call.
453 */
454static void region_abort(struct resv_map *resv, long f, long t)
455{
456 spin_lock(&resv->lock);
457 VM_BUG_ON(!resv->region_cache_count);
458 resv->adds_in_progress--;
459 spin_unlock(&resv->lock);
460}
461
1dd308a7 462/*
feba16e2
MK
463 * Delete the specified range [f, t) from the reserve map. If the
464 * t parameter is LONG_MAX, this indicates that ALL regions after f
465 * should be deleted. Locate the regions which intersect [f, t)
466 * and either trim, delete or split the existing regions.
467 *
468 * Returns the number of huge pages deleted from the reserve map.
469 * In the normal case, the return value is zero or more. In the
470 * case where a region must be split, a new region descriptor must
471 * be allocated. If the allocation fails, -ENOMEM will be returned.
472 * NOTE: If the parameter t == LONG_MAX, then we will never split
473 * a region and possibly return -ENOMEM. Callers specifying
474 * t == LONG_MAX do not need to check for -ENOMEM error.
1dd308a7 475 */
feba16e2 476static long region_del(struct resv_map *resv, long f, long t)
96822904 477{
1406ec9b 478 struct list_head *head = &resv->regions;
96822904 479 struct file_region *rg, *trg;
feba16e2
MK
480 struct file_region *nrg = NULL;
481 long del = 0;
96822904 482
feba16e2 483retry:
7b24d861 484 spin_lock(&resv->lock);
feba16e2
MK
485 list_for_each_entry_safe(rg, trg, head, link) {
486 if (rg->to <= f)
487 continue;
488 if (rg->from >= t)
96822904 489 break;
96822904 490
feba16e2
MK
491 if (f > rg->from && t < rg->to) { /* Must split region */
492 /*
493 * Check for an entry in the cache before dropping
494 * lock and attempting allocation.
495 */
496 if (!nrg &&
497 resv->region_cache_count > resv->adds_in_progress) {
498 nrg = list_first_entry(&resv->region_cache,
499 struct file_region,
500 link);
501 list_del(&nrg->link);
502 resv->region_cache_count--;
503 }
96822904 504
feba16e2
MK
505 if (!nrg) {
506 spin_unlock(&resv->lock);
507 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
508 if (!nrg)
509 return -ENOMEM;
510 goto retry;
511 }
512
513 del += t - f;
514
515 /* New entry for end of split region */
516 nrg->from = t;
517 nrg->to = rg->to;
518 INIT_LIST_HEAD(&nrg->link);
519
520 /* Original entry is trimmed */
521 rg->to = f;
522
523 list_add(&nrg->link, &rg->link);
524 nrg = NULL;
96822904 525 break;
feba16e2
MK
526 }
527
528 if (f <= rg->from && t >= rg->to) { /* Remove entire region */
529 del += rg->to - rg->from;
530 list_del(&rg->link);
531 kfree(rg);
532 continue;
533 }
534
535 if (f <= rg->from) { /* Trim beginning of region */
536 del += t - rg->from;
537 rg->from = t;
538 } else { /* Trim end of region */
539 del += rg->to - f;
540 rg->to = f;
541 }
96822904 542 }
7b24d861 543
7b24d861 544 spin_unlock(&resv->lock);
feba16e2
MK
545 kfree(nrg);
546 return del;
96822904
AW
547}
548
b5cec28d
MK
549/*
550 * A rare out of memory error was encountered which prevented removal of
551 * the reserve map region for a page. The huge page itself was free'ed
552 * and removed from the page cache. This routine will adjust the subpool
553 * usage count, and the global reserve count if needed. By incrementing
554 * these counts, the reserve map entry which could not be deleted will
555 * appear as a "reserved" entry instead of simply dangling with incorrect
556 * counts.
557 */
558void hugetlb_fix_reserve_counts(struct inode *inode, bool restore_reserve)
559{
560 struct hugepage_subpool *spool = subpool_inode(inode);
561 long rsv_adjust;
562
563 rsv_adjust = hugepage_subpool_get_pages(spool, 1);
564 if (restore_reserve && rsv_adjust) {
565 struct hstate *h = hstate_inode(inode);
566
567 hugetlb_acct_memory(h, 1);
568 }
569}
570
1dd308a7
MK
571/*
572 * Count and return the number of huge pages in the reserve map
573 * that intersect with the range [f, t).
574 */
1406ec9b 575static long region_count(struct resv_map *resv, long f, long t)
84afd99b 576{
1406ec9b 577 struct list_head *head = &resv->regions;
84afd99b
AW
578 struct file_region *rg;
579 long chg = 0;
580
7b24d861 581 spin_lock(&resv->lock);
84afd99b
AW
582 /* Locate each segment we overlap with, and count that overlap. */
583 list_for_each_entry(rg, head, link) {
f2135a4a
WSH
584 long seg_from;
585 long seg_to;
84afd99b
AW
586
587 if (rg->to <= f)
588 continue;
589 if (rg->from >= t)
590 break;
591
592 seg_from = max(rg->from, f);
593 seg_to = min(rg->to, t);
594
595 chg += seg_to - seg_from;
596 }
7b24d861 597 spin_unlock(&resv->lock);
84afd99b
AW
598
599 return chg;
600}
601
e7c4b0bf
AW
602/*
603 * Convert the address within this vma to the page offset within
604 * the mapping, in pagecache page units; huge pages here.
605 */
a5516438
AK
606static pgoff_t vma_hugecache_offset(struct hstate *h,
607 struct vm_area_struct *vma, unsigned long address)
e7c4b0bf 608{
a5516438
AK
609 return ((address - vma->vm_start) >> huge_page_shift(h)) +
610 (vma->vm_pgoff >> huge_page_order(h));
e7c4b0bf
AW
611}
612
0fe6e20b
NH
613pgoff_t linear_hugepage_index(struct vm_area_struct *vma,
614 unsigned long address)
615{
616 return vma_hugecache_offset(hstate_vma(vma), vma, address);
617}
618
08fba699
MG
619/*
620 * Return the size of the pages allocated when backing a VMA. In the majority
621 * cases this will be same size as used by the page table entries.
622 */
623unsigned long vma_kernel_pagesize(struct vm_area_struct *vma)
624{
625 struct hstate *hstate;
626
627 if (!is_vm_hugetlb_page(vma))
628 return PAGE_SIZE;
629
630 hstate = hstate_vma(vma);
631
2415cf12 632 return 1UL << huge_page_shift(hstate);
08fba699 633}
f340ca0f 634EXPORT_SYMBOL_GPL(vma_kernel_pagesize);
08fba699 635
3340289d
MG
636/*
637 * Return the page size being used by the MMU to back a VMA. In the majority
638 * of cases, the page size used by the kernel matches the MMU size. On
639 * architectures where it differs, an architecture-specific version of this
640 * function is required.
641 */
642#ifndef vma_mmu_pagesize
643unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
644{
645 return vma_kernel_pagesize(vma);
646}
647#endif
648
84afd99b
AW
649/*
650 * Flags for MAP_PRIVATE reservations. These are stored in the bottom
651 * bits of the reservation map pointer, which are always clear due to
652 * alignment.
653 */
654#define HPAGE_RESV_OWNER (1UL << 0)
655#define HPAGE_RESV_UNMAPPED (1UL << 1)
04f2cbe3 656#define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
84afd99b 657
a1e78772
MG
658/*
659 * These helpers are used to track how many pages are reserved for
660 * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
661 * is guaranteed to have their future faults succeed.
662 *
663 * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
664 * the reserve counters are updated with the hugetlb_lock held. It is safe
665 * to reset the VMA at fork() time as it is not in use yet and there is no
666 * chance of the global counters getting corrupted as a result of the values.
84afd99b
AW
667 *
668 * The private mapping reservation is represented in a subtly different
669 * manner to a shared mapping. A shared mapping has a region map associated
670 * with the underlying file, this region map represents the backing file
671 * pages which have ever had a reservation assigned which this persists even
672 * after the page is instantiated. A private mapping has a region map
673 * associated with the original mmap which is attached to all VMAs which
674 * reference it, this region map represents those offsets which have consumed
675 * reservation ie. where pages have been instantiated.
a1e78772 676 */
e7c4b0bf
AW
677static unsigned long get_vma_private_data(struct vm_area_struct *vma)
678{
679 return (unsigned long)vma->vm_private_data;
680}
681
682static void set_vma_private_data(struct vm_area_struct *vma,
683 unsigned long value)
684{
685 vma->vm_private_data = (void *)value;
686}
687
9119a41e 688struct resv_map *resv_map_alloc(void)
84afd99b
AW
689{
690 struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
5e911373
MK
691 struct file_region *rg = kmalloc(sizeof(*rg), GFP_KERNEL);
692
693 if (!resv_map || !rg) {
694 kfree(resv_map);
695 kfree(rg);
84afd99b 696 return NULL;
5e911373 697 }
84afd99b
AW
698
699 kref_init(&resv_map->refs);
7b24d861 700 spin_lock_init(&resv_map->lock);
84afd99b
AW
701 INIT_LIST_HEAD(&resv_map->regions);
702
5e911373
MK
703 resv_map->adds_in_progress = 0;
704
705 INIT_LIST_HEAD(&resv_map->region_cache);
706 list_add(&rg->link, &resv_map->region_cache);
707 resv_map->region_cache_count = 1;
708
84afd99b
AW
709 return resv_map;
710}
711
9119a41e 712void resv_map_release(struct kref *ref)
84afd99b
AW
713{
714 struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
5e911373
MK
715 struct list_head *head = &resv_map->region_cache;
716 struct file_region *rg, *trg;
84afd99b
AW
717
718 /* Clear out any active regions before we release the map. */
feba16e2 719 region_del(resv_map, 0, LONG_MAX);
5e911373
MK
720
721 /* ... and any entries left in the cache */
722 list_for_each_entry_safe(rg, trg, head, link) {
723 list_del(&rg->link);
724 kfree(rg);
725 }
726
727 VM_BUG_ON(resv_map->adds_in_progress);
728
84afd99b
AW
729 kfree(resv_map);
730}
731
4e35f483
JK
732static inline struct resv_map *inode_resv_map(struct inode *inode)
733{
734 return inode->i_mapping->private_data;
735}
736
84afd99b 737static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
a1e78772 738{
81d1b09c 739 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
4e35f483
JK
740 if (vma->vm_flags & VM_MAYSHARE) {
741 struct address_space *mapping = vma->vm_file->f_mapping;
742 struct inode *inode = mapping->host;
743
744 return inode_resv_map(inode);
745
746 } else {
84afd99b
AW
747 return (struct resv_map *)(get_vma_private_data(vma) &
748 ~HPAGE_RESV_MASK);
4e35f483 749 }
a1e78772
MG
750}
751
84afd99b 752static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
a1e78772 753{
81d1b09c
SL
754 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
755 VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
a1e78772 756
84afd99b
AW
757 set_vma_private_data(vma, (get_vma_private_data(vma) &
758 HPAGE_RESV_MASK) | (unsigned long)map);
04f2cbe3
MG
759}
760
761static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
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);
e7c4b0bf
AW
765
766 set_vma_private_data(vma, get_vma_private_data(vma) | flags);
04f2cbe3
MG
767}
768
769static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
770{
81d1b09c 771 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
e7c4b0bf
AW
772
773 return (get_vma_private_data(vma) & flag) != 0;
a1e78772
MG
774}
775
04f2cbe3 776/* Reset counters to 0 and clear all HPAGE_RESV_* flags */
a1e78772
MG
777void reset_vma_resv_huge_pages(struct vm_area_struct *vma)
778{
81d1b09c 779 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
f83a275d 780 if (!(vma->vm_flags & VM_MAYSHARE))
a1e78772
MG
781 vma->vm_private_data = (void *)0;
782}
783
784/* Returns true if the VMA has associated reserve pages */
559ec2f8 785static bool vma_has_reserves(struct vm_area_struct *vma, long chg)
a1e78772 786{
af0ed73e
JK
787 if (vma->vm_flags & VM_NORESERVE) {
788 /*
789 * This address is already reserved by other process(chg == 0),
790 * so, we should decrement reserved count. Without decrementing,
791 * reserve count remains after releasing inode, because this
792 * allocated page will go into page cache and is regarded as
793 * coming from reserved pool in releasing step. Currently, we
794 * don't have any other solution to deal with this situation
795 * properly, so add work-around here.
796 */
797 if (vma->vm_flags & VM_MAYSHARE && chg == 0)
559ec2f8 798 return true;
af0ed73e 799 else
559ec2f8 800 return false;
af0ed73e 801 }
a63884e9
JK
802
803 /* Shared mappings always use reserves */
1fb1b0e9
MK
804 if (vma->vm_flags & VM_MAYSHARE) {
805 /*
806 * We know VM_NORESERVE is not set. Therefore, there SHOULD
807 * be a region map for all pages. The only situation where
808 * there is no region map is if a hole was punched via
809 * fallocate. In this case, there really are no reverves to
810 * use. This situation is indicated if chg != 0.
811 */
812 if (chg)
813 return false;
814 else
815 return true;
816 }
a63884e9
JK
817
818 /*
819 * Only the process that called mmap() has reserves for
820 * private mappings.
821 */
7f09ca51 822 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER))
559ec2f8 823 return true;
a63884e9 824
559ec2f8 825 return false;
a1e78772
MG
826}
827
a5516438 828static void enqueue_huge_page(struct hstate *h, struct page *page)
1da177e4
LT
829{
830 int nid = page_to_nid(page);
0edaecfa 831 list_move(&page->lru, &h->hugepage_freelists[nid]);
a5516438
AK
832 h->free_huge_pages++;
833 h->free_huge_pages_node[nid]++;
1da177e4
LT
834}
835
bf50bab2
NH
836static struct page *dequeue_huge_page_node(struct hstate *h, int nid)
837{
838 struct page *page;
839
c8721bbb
NH
840 list_for_each_entry(page, &h->hugepage_freelists[nid], lru)
841 if (!is_migrate_isolate_page(page))
842 break;
843 /*
844 * if 'non-isolated free hugepage' not found on the list,
845 * the allocation fails.
846 */
847 if (&h->hugepage_freelists[nid] == &page->lru)
bf50bab2 848 return NULL;
0edaecfa 849 list_move(&page->lru, &h->hugepage_activelist);
a9869b83 850 set_page_refcounted(page);
bf50bab2
NH
851 h->free_huge_pages--;
852 h->free_huge_pages_node[nid]--;
853 return page;
854}
855
86cdb465
NH
856/* Movability of hugepages depends on migration support. */
857static inline gfp_t htlb_alloc_mask(struct hstate *h)
858{
100873d7 859 if (hugepages_treat_as_movable || hugepage_migration_supported(h))
86cdb465
NH
860 return GFP_HIGHUSER_MOVABLE;
861 else
862 return GFP_HIGHUSER;
863}
864
a5516438
AK
865static struct page *dequeue_huge_page_vma(struct hstate *h,
866 struct vm_area_struct *vma,
af0ed73e
JK
867 unsigned long address, int avoid_reserve,
868 long chg)
1da177e4 869{
b1c12cbc 870 struct page *page = NULL;
480eccf9 871 struct mempolicy *mpol;
19770b32 872 nodemask_t *nodemask;
c0ff7453 873 struct zonelist *zonelist;
dd1a239f
MG
874 struct zone *zone;
875 struct zoneref *z;
cc9a6c87 876 unsigned int cpuset_mems_cookie;
1da177e4 877
a1e78772
MG
878 /*
879 * A child process with MAP_PRIVATE mappings created by their parent
880 * have no page reserves. This check ensures that reservations are
881 * not "stolen". The child may still get SIGKILLed
882 */
af0ed73e 883 if (!vma_has_reserves(vma, chg) &&
a5516438 884 h->free_huge_pages - h->resv_huge_pages == 0)
c0ff7453 885 goto err;
a1e78772 886
04f2cbe3 887 /* If reserves cannot be used, ensure enough pages are in the pool */
a5516438 888 if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
6eab04a8 889 goto err;
04f2cbe3 890
9966c4bb 891retry_cpuset:
d26914d1 892 cpuset_mems_cookie = read_mems_allowed_begin();
9966c4bb 893 zonelist = huge_zonelist(vma, address,
86cdb465 894 htlb_alloc_mask(h), &mpol, &nodemask);
9966c4bb 895
19770b32
MG
896 for_each_zone_zonelist_nodemask(zone, z, zonelist,
897 MAX_NR_ZONES - 1, nodemask) {
344736f2 898 if (cpuset_zone_allowed(zone, htlb_alloc_mask(h))) {
bf50bab2
NH
899 page = dequeue_huge_page_node(h, zone_to_nid(zone));
900 if (page) {
af0ed73e
JK
901 if (avoid_reserve)
902 break;
903 if (!vma_has_reserves(vma, chg))
904 break;
905
07443a85 906 SetPagePrivate(page);
af0ed73e 907 h->resv_huge_pages--;
bf50bab2
NH
908 break;
909 }
3abf7afd 910 }
1da177e4 911 }
cc9a6c87 912
52cd3b07 913 mpol_cond_put(mpol);
d26914d1 914 if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie)))
cc9a6c87 915 goto retry_cpuset;
1da177e4 916 return page;
cc9a6c87
MG
917
918err:
cc9a6c87 919 return NULL;
1da177e4
LT
920}
921
1cac6f2c
LC
922/*
923 * common helper functions for hstate_next_node_to_{alloc|free}.
924 * We may have allocated or freed a huge page based on a different
925 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
926 * be outside of *nodes_allowed. Ensure that we use an allowed
927 * node for alloc or free.
928 */
929static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
930{
931 nid = next_node(nid, *nodes_allowed);
932 if (nid == MAX_NUMNODES)
933 nid = first_node(*nodes_allowed);
934 VM_BUG_ON(nid >= MAX_NUMNODES);
935
936 return nid;
937}
938
939static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
940{
941 if (!node_isset(nid, *nodes_allowed))
942 nid = next_node_allowed(nid, nodes_allowed);
943 return nid;
944}
945
946/*
947 * returns the previously saved node ["this node"] from which to
948 * allocate a persistent huge page for the pool and advance the
949 * next node from which to allocate, handling wrap at end of node
950 * mask.
951 */
952static int hstate_next_node_to_alloc(struct hstate *h,
953 nodemask_t *nodes_allowed)
954{
955 int nid;
956
957 VM_BUG_ON(!nodes_allowed);
958
959 nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
960 h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
961
962 return nid;
963}
964
965/*
966 * helper for free_pool_huge_page() - return the previously saved
967 * node ["this node"] from which to free a huge page. Advance the
968 * next node id whether or not we find a free huge page to free so
969 * that the next attempt to free addresses the next node.
970 */
971static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
972{
973 int nid;
974
975 VM_BUG_ON(!nodes_allowed);
976
977 nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
978 h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
979
980 return nid;
981}
982
983#define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \
984 for (nr_nodes = nodes_weight(*mask); \
985 nr_nodes > 0 && \
986 ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \
987 nr_nodes--)
988
989#define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \
990 for (nr_nodes = nodes_weight(*mask); \
991 nr_nodes > 0 && \
992 ((node = hstate_next_node_to_free(hs, mask)) || 1); \
993 nr_nodes--)
994
944d9fec
LC
995#if defined(CONFIG_CMA) && defined(CONFIG_X86_64)
996static void destroy_compound_gigantic_page(struct page *page,
997 unsigned long order)
998{
999 int i;
1000 int nr_pages = 1 << order;
1001 struct page *p = page + 1;
1002
1003 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
1004 __ClearPageTail(p);
1005 set_page_refcounted(p);
1006 p->first_page = NULL;
1007 }
1008
1009 set_compound_order(page, 0);
1010 __ClearPageHead(page);
1011}
1012
1013static void free_gigantic_page(struct page *page, unsigned order)
1014{
1015 free_contig_range(page_to_pfn(page), 1 << order);
1016}
1017
1018static int __alloc_gigantic_page(unsigned long start_pfn,
1019 unsigned long nr_pages)
1020{
1021 unsigned long end_pfn = start_pfn + nr_pages;
1022 return alloc_contig_range(start_pfn, end_pfn, MIGRATE_MOVABLE);
1023}
1024
1025static bool pfn_range_valid_gigantic(unsigned long start_pfn,
1026 unsigned long nr_pages)
1027{
1028 unsigned long i, end_pfn = start_pfn + nr_pages;
1029 struct page *page;
1030
1031 for (i = start_pfn; i < end_pfn; i++) {
1032 if (!pfn_valid(i))
1033 return false;
1034
1035 page = pfn_to_page(i);
1036
1037 if (PageReserved(page))
1038 return false;
1039
1040 if (page_count(page) > 0)
1041 return false;
1042
1043 if (PageHuge(page))
1044 return false;
1045 }
1046
1047 return true;
1048}
1049
1050static bool zone_spans_last_pfn(const struct zone *zone,
1051 unsigned long start_pfn, unsigned long nr_pages)
1052{
1053 unsigned long last_pfn = start_pfn + nr_pages - 1;
1054 return zone_spans_pfn(zone, last_pfn);
1055}
1056
1057static struct page *alloc_gigantic_page(int nid, unsigned order)
1058{
1059 unsigned long nr_pages = 1 << order;
1060 unsigned long ret, pfn, flags;
1061 struct zone *z;
1062
1063 z = NODE_DATA(nid)->node_zones;
1064 for (; z - NODE_DATA(nid)->node_zones < MAX_NR_ZONES; z++) {
1065 spin_lock_irqsave(&z->lock, flags);
1066
1067 pfn = ALIGN(z->zone_start_pfn, nr_pages);
1068 while (zone_spans_last_pfn(z, pfn, nr_pages)) {
1069 if (pfn_range_valid_gigantic(pfn, nr_pages)) {
1070 /*
1071 * We release the zone lock here because
1072 * alloc_contig_range() will also lock the zone
1073 * at some point. If there's an allocation
1074 * spinning on this lock, it may win the race
1075 * and cause alloc_contig_range() to fail...
1076 */
1077 spin_unlock_irqrestore(&z->lock, flags);
1078 ret = __alloc_gigantic_page(pfn, nr_pages);
1079 if (!ret)
1080 return pfn_to_page(pfn);
1081 spin_lock_irqsave(&z->lock, flags);
1082 }
1083 pfn += nr_pages;
1084 }
1085
1086 spin_unlock_irqrestore(&z->lock, flags);
1087 }
1088
1089 return NULL;
1090}
1091
1092static void prep_new_huge_page(struct hstate *h, struct page *page, int nid);
1093static void prep_compound_gigantic_page(struct page *page, unsigned long order);
1094
1095static struct page *alloc_fresh_gigantic_page_node(struct hstate *h, int nid)
1096{
1097 struct page *page;
1098
1099 page = alloc_gigantic_page(nid, huge_page_order(h));
1100 if (page) {
1101 prep_compound_gigantic_page(page, huge_page_order(h));
1102 prep_new_huge_page(h, page, nid);
1103 }
1104
1105 return page;
1106}
1107
1108static int alloc_fresh_gigantic_page(struct hstate *h,
1109 nodemask_t *nodes_allowed)
1110{
1111 struct page *page = NULL;
1112 int nr_nodes, node;
1113
1114 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
1115 page = alloc_fresh_gigantic_page_node(h, node);
1116 if (page)
1117 return 1;
1118 }
1119
1120 return 0;
1121}
1122
1123static inline bool gigantic_page_supported(void) { return true; }
1124#else
1125static inline bool gigantic_page_supported(void) { return false; }
1126static inline void free_gigantic_page(struct page *page, unsigned order) { }
1127static inline void destroy_compound_gigantic_page(struct page *page,
1128 unsigned long order) { }
1129static inline int alloc_fresh_gigantic_page(struct hstate *h,
1130 nodemask_t *nodes_allowed) { return 0; }
1131#endif
1132
a5516438 1133static void update_and_free_page(struct hstate *h, struct page *page)
6af2acb6
AL
1134{
1135 int i;
a5516438 1136
944d9fec
LC
1137 if (hstate_is_gigantic(h) && !gigantic_page_supported())
1138 return;
18229df5 1139
a5516438
AK
1140 h->nr_huge_pages--;
1141 h->nr_huge_pages_node[page_to_nid(page)]--;
1142 for (i = 0; i < pages_per_huge_page(h); i++) {
32f84528
CF
1143 page[i].flags &= ~(1 << PG_locked | 1 << PG_error |
1144 1 << PG_referenced | 1 << PG_dirty |
a7407a27
LC
1145 1 << PG_active | 1 << PG_private |
1146 1 << PG_writeback);
6af2acb6 1147 }
309381fe 1148 VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page);
6af2acb6
AL
1149 set_compound_page_dtor(page, NULL);
1150 set_page_refcounted(page);
944d9fec
LC
1151 if (hstate_is_gigantic(h)) {
1152 destroy_compound_gigantic_page(page, huge_page_order(h));
1153 free_gigantic_page(page, huge_page_order(h));
1154 } else {
944d9fec
LC
1155 __free_pages(page, huge_page_order(h));
1156 }
6af2acb6
AL
1157}
1158
e5ff2159
AK
1159struct hstate *size_to_hstate(unsigned long size)
1160{
1161 struct hstate *h;
1162
1163 for_each_hstate(h) {
1164 if (huge_page_size(h) == size)
1165 return h;
1166 }
1167 return NULL;
1168}
1169
bcc54222
NH
1170/*
1171 * Test to determine whether the hugepage is "active/in-use" (i.e. being linked
1172 * to hstate->hugepage_activelist.)
1173 *
1174 * This function can be called for tail pages, but never returns true for them.
1175 */
1176bool page_huge_active(struct page *page)
1177{
1178 VM_BUG_ON_PAGE(!PageHuge(page), page);
1179 return PageHead(page) && PagePrivate(&page[1]);
1180}
1181
1182/* never called for tail page */
1183static void set_page_huge_active(struct page *page)
1184{
1185 VM_BUG_ON_PAGE(!PageHeadHuge(page), page);
1186 SetPagePrivate(&page[1]);
1187}
1188
1189static void clear_page_huge_active(struct page *page)
1190{
1191 VM_BUG_ON_PAGE(!PageHeadHuge(page), page);
1192 ClearPagePrivate(&page[1]);
1193}
1194
8f1d26d0 1195void free_huge_page(struct page *page)
27a85ef1 1196{
a5516438
AK
1197 /*
1198 * Can't pass hstate in here because it is called from the
1199 * compound page destructor.
1200 */
e5ff2159 1201 struct hstate *h = page_hstate(page);
7893d1d5 1202 int nid = page_to_nid(page);
90481622
DG
1203 struct hugepage_subpool *spool =
1204 (struct hugepage_subpool *)page_private(page);
07443a85 1205 bool restore_reserve;
27a85ef1 1206
e5df70ab 1207 set_page_private(page, 0);
23be7468 1208 page->mapping = NULL;
7893d1d5 1209 BUG_ON(page_count(page));
0fe6e20b 1210 BUG_ON(page_mapcount(page));
07443a85 1211 restore_reserve = PagePrivate(page);
16c794b4 1212 ClearPagePrivate(page);
27a85ef1 1213
1c5ecae3
MK
1214 /*
1215 * A return code of zero implies that the subpool will be under its
1216 * minimum size if the reservation is not restored after page is free.
1217 * Therefore, force restore_reserve operation.
1218 */
1219 if (hugepage_subpool_put_pages(spool, 1) == 0)
1220 restore_reserve = true;
1221
27a85ef1 1222 spin_lock(&hugetlb_lock);
bcc54222 1223 clear_page_huge_active(page);
6d76dcf4
AK
1224 hugetlb_cgroup_uncharge_page(hstate_index(h),
1225 pages_per_huge_page(h), page);
07443a85
JK
1226 if (restore_reserve)
1227 h->resv_huge_pages++;
1228
944d9fec 1229 if (h->surplus_huge_pages_node[nid]) {
0edaecfa
AK
1230 /* remove the page from active list */
1231 list_del(&page->lru);
a5516438
AK
1232 update_and_free_page(h, page);
1233 h->surplus_huge_pages--;
1234 h->surplus_huge_pages_node[nid]--;
7893d1d5 1235 } else {
5d3a551c 1236 arch_clear_hugepage_flags(page);
a5516438 1237 enqueue_huge_page(h, page);
7893d1d5 1238 }
27a85ef1
DG
1239 spin_unlock(&hugetlb_lock);
1240}
1241
a5516438 1242static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
b7ba30c6 1243{
0edaecfa 1244 INIT_LIST_HEAD(&page->lru);
b7ba30c6
AK
1245 set_compound_page_dtor(page, free_huge_page);
1246 spin_lock(&hugetlb_lock);
9dd540e2 1247 set_hugetlb_cgroup(page, NULL);
a5516438
AK
1248 h->nr_huge_pages++;
1249 h->nr_huge_pages_node[nid]++;
b7ba30c6
AK
1250 spin_unlock(&hugetlb_lock);
1251 put_page(page); /* free it into the hugepage allocator */
1252}
1253
2906dd52 1254static void prep_compound_gigantic_page(struct page *page, unsigned long order)
20a0307c
WF
1255{
1256 int i;
1257 int nr_pages = 1 << order;
1258 struct page *p = page + 1;
1259
1260 /* we rely on prep_new_huge_page to set the destructor */
1261 set_compound_order(page, order);
1262 __SetPageHead(page);
ef5a22be 1263 __ClearPageReserved(page);
20a0307c 1264 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
ef5a22be
AA
1265 /*
1266 * For gigantic hugepages allocated through bootmem at
1267 * boot, it's safer to be consistent with the not-gigantic
1268 * hugepages and clear the PG_reserved bit from all tail pages
1269 * too. Otherwse drivers using get_user_pages() to access tail
1270 * pages may get the reference counting wrong if they see
1271 * PG_reserved set on a tail page (despite the head page not
1272 * having PG_reserved set). Enforcing this consistency between
1273 * head and tail pages allows drivers to optimize away a check
1274 * on the head page when they need know if put_page() is needed
1275 * after get_user_pages().
1276 */
1277 __ClearPageReserved(p);
58a84aa9 1278 set_page_count(p, 0);
20a0307c 1279 p->first_page = page;
44fc8057
DR
1280 /* Make sure p->first_page is always valid for PageTail() */
1281 smp_wmb();
1282 __SetPageTail(p);
20a0307c
WF
1283 }
1284}
1285
7795912c
AM
1286/*
1287 * PageHuge() only returns true for hugetlbfs pages, but not for normal or
1288 * transparent huge pages. See the PageTransHuge() documentation for more
1289 * details.
1290 */
20a0307c
WF
1291int PageHuge(struct page *page)
1292{
20a0307c
WF
1293 if (!PageCompound(page))
1294 return 0;
1295
1296 page = compound_head(page);
758f66a2 1297 return get_compound_page_dtor(page) == free_huge_page;
20a0307c 1298}
43131e14
NH
1299EXPORT_SYMBOL_GPL(PageHuge);
1300
27c73ae7
AA
1301/*
1302 * PageHeadHuge() only returns true for hugetlbfs head page, but not for
1303 * normal or transparent huge pages.
1304 */
1305int PageHeadHuge(struct page *page_head)
1306{
27c73ae7
AA
1307 if (!PageHead(page_head))
1308 return 0;
1309
758f66a2 1310 return get_compound_page_dtor(page_head) == free_huge_page;
27c73ae7 1311}
27c73ae7 1312
13d60f4b
ZY
1313pgoff_t __basepage_index(struct page *page)
1314{
1315 struct page *page_head = compound_head(page);
1316 pgoff_t index = page_index(page_head);
1317 unsigned long compound_idx;
1318
1319 if (!PageHuge(page_head))
1320 return page_index(page);
1321
1322 if (compound_order(page_head) >= MAX_ORDER)
1323 compound_idx = page_to_pfn(page) - page_to_pfn(page_head);
1324 else
1325 compound_idx = page - page_head;
1326
1327 return (index << compound_order(page_head)) + compound_idx;
1328}
1329
a5516438 1330static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid)
1da177e4 1331{
1da177e4 1332 struct page *page;
f96efd58 1333
6484eb3e 1334 page = alloc_pages_exact_node(nid,
86cdb465 1335 htlb_alloc_mask(h)|__GFP_COMP|__GFP_THISNODE|
551883ae 1336 __GFP_REPEAT|__GFP_NOWARN,
a5516438 1337 huge_page_order(h));
1da177e4 1338 if (page) {
a5516438 1339 prep_new_huge_page(h, page, nid);
1da177e4 1340 }
63b4613c
NA
1341
1342 return page;
1343}
1344
b2261026
JK
1345static int alloc_fresh_huge_page(struct hstate *h, nodemask_t *nodes_allowed)
1346{
1347 struct page *page;
1348 int nr_nodes, node;
1349 int ret = 0;
1350
1351 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
1352 page = alloc_fresh_huge_page_node(h, node);
1353 if (page) {
1354 ret = 1;
1355 break;
1356 }
1357 }
1358
1359 if (ret)
1360 count_vm_event(HTLB_BUDDY_PGALLOC);
1361 else
1362 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
1363
1364 return ret;
1365}
1366
e8c5c824
LS
1367/*
1368 * Free huge page from pool from next node to free.
1369 * Attempt to keep persistent huge pages more or less
1370 * balanced over allowed nodes.
1371 * Called with hugetlb_lock locked.
1372 */
6ae11b27
LS
1373static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
1374 bool acct_surplus)
e8c5c824 1375{
b2261026 1376 int nr_nodes, node;
e8c5c824
LS
1377 int ret = 0;
1378
b2261026 1379 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
685f3457
LS
1380 /*
1381 * If we're returning unused surplus pages, only examine
1382 * nodes with surplus pages.
1383 */
b2261026
JK
1384 if ((!acct_surplus || h->surplus_huge_pages_node[node]) &&
1385 !list_empty(&h->hugepage_freelists[node])) {
e8c5c824 1386 struct page *page =
b2261026 1387 list_entry(h->hugepage_freelists[node].next,
e8c5c824
LS
1388 struct page, lru);
1389 list_del(&page->lru);
1390 h->free_huge_pages--;
b2261026 1391 h->free_huge_pages_node[node]--;
685f3457
LS
1392 if (acct_surplus) {
1393 h->surplus_huge_pages--;
b2261026 1394 h->surplus_huge_pages_node[node]--;
685f3457 1395 }
e8c5c824
LS
1396 update_and_free_page(h, page);
1397 ret = 1;
9a76db09 1398 break;
e8c5c824 1399 }
b2261026 1400 }
e8c5c824
LS
1401
1402 return ret;
1403}
1404
c8721bbb
NH
1405/*
1406 * Dissolve a given free hugepage into free buddy pages. This function does
1407 * nothing for in-use (including surplus) hugepages.
1408 */
1409static void dissolve_free_huge_page(struct page *page)
1410{
1411 spin_lock(&hugetlb_lock);
1412 if (PageHuge(page) && !page_count(page)) {
1413 struct hstate *h = page_hstate(page);
1414 int nid = page_to_nid(page);
1415 list_del(&page->lru);
1416 h->free_huge_pages--;
1417 h->free_huge_pages_node[nid]--;
1418 update_and_free_page(h, page);
1419 }
1420 spin_unlock(&hugetlb_lock);
1421}
1422
1423/*
1424 * Dissolve free hugepages in a given pfn range. Used by memory hotplug to
1425 * make specified memory blocks removable from the system.
1426 * Note that start_pfn should aligned with (minimum) hugepage size.
1427 */
1428void dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn)
1429{
c8721bbb 1430 unsigned long pfn;
c8721bbb 1431
d0177639
LZ
1432 if (!hugepages_supported())
1433 return;
1434
641844f5
NH
1435 VM_BUG_ON(!IS_ALIGNED(start_pfn, 1 << minimum_order));
1436 for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << minimum_order)
c8721bbb
NH
1437 dissolve_free_huge_page(pfn_to_page(pfn));
1438}
1439
bf50bab2 1440static struct page *alloc_buddy_huge_page(struct hstate *h, int nid)
7893d1d5
AL
1441{
1442 struct page *page;
bf50bab2 1443 unsigned int r_nid;
7893d1d5 1444
bae7f4ae 1445 if (hstate_is_gigantic(h))
aa888a74
AK
1446 return NULL;
1447
d1c3fb1f
NA
1448 /*
1449 * Assume we will successfully allocate the surplus page to
1450 * prevent racing processes from causing the surplus to exceed
1451 * overcommit
1452 *
1453 * This however introduces a different race, where a process B
1454 * tries to grow the static hugepage pool while alloc_pages() is
1455 * called by process A. B will only examine the per-node
1456 * counters in determining if surplus huge pages can be
1457 * converted to normal huge pages in adjust_pool_surplus(). A
1458 * won't be able to increment the per-node counter, until the
1459 * lock is dropped by B, but B doesn't drop hugetlb_lock until
1460 * no more huge pages can be converted from surplus to normal
1461 * state (and doesn't try to convert again). Thus, we have a
1462 * case where a surplus huge page exists, the pool is grown, and
1463 * the surplus huge page still exists after, even though it
1464 * should just have been converted to a normal huge page. This
1465 * does not leak memory, though, as the hugepage will be freed
1466 * once it is out of use. It also does not allow the counters to
1467 * go out of whack in adjust_pool_surplus() as we don't modify
1468 * the node values until we've gotten the hugepage and only the
1469 * per-node value is checked there.
1470 */
1471 spin_lock(&hugetlb_lock);
a5516438 1472 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
d1c3fb1f
NA
1473 spin_unlock(&hugetlb_lock);
1474 return NULL;
1475 } else {
a5516438
AK
1476 h->nr_huge_pages++;
1477 h->surplus_huge_pages++;
d1c3fb1f
NA
1478 }
1479 spin_unlock(&hugetlb_lock);
1480
bf50bab2 1481 if (nid == NUMA_NO_NODE)
86cdb465 1482 page = alloc_pages(htlb_alloc_mask(h)|__GFP_COMP|
bf50bab2
NH
1483 __GFP_REPEAT|__GFP_NOWARN,
1484 huge_page_order(h));
1485 else
1486 page = alloc_pages_exact_node(nid,
86cdb465 1487 htlb_alloc_mask(h)|__GFP_COMP|__GFP_THISNODE|
bf50bab2 1488 __GFP_REPEAT|__GFP_NOWARN, huge_page_order(h));
d1c3fb1f
NA
1489
1490 spin_lock(&hugetlb_lock);
7893d1d5 1491 if (page) {
0edaecfa 1492 INIT_LIST_HEAD(&page->lru);
bf50bab2 1493 r_nid = page_to_nid(page);
7893d1d5 1494 set_compound_page_dtor(page, free_huge_page);
9dd540e2 1495 set_hugetlb_cgroup(page, NULL);
d1c3fb1f
NA
1496 /*
1497 * We incremented the global counters already
1498 */
bf50bab2
NH
1499 h->nr_huge_pages_node[r_nid]++;
1500 h->surplus_huge_pages_node[r_nid]++;
3b116300 1501 __count_vm_event(HTLB_BUDDY_PGALLOC);
d1c3fb1f 1502 } else {
a5516438
AK
1503 h->nr_huge_pages--;
1504 h->surplus_huge_pages--;
3b116300 1505 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
7893d1d5 1506 }
d1c3fb1f 1507 spin_unlock(&hugetlb_lock);
7893d1d5
AL
1508
1509 return page;
1510}
1511
bf50bab2
NH
1512/*
1513 * This allocation function is useful in the context where vma is irrelevant.
1514 * E.g. soft-offlining uses this function because it only cares physical
1515 * address of error page.
1516 */
1517struct page *alloc_huge_page_node(struct hstate *h, int nid)
1518{
4ef91848 1519 struct page *page = NULL;
bf50bab2
NH
1520
1521 spin_lock(&hugetlb_lock);
4ef91848
JK
1522 if (h->free_huge_pages - h->resv_huge_pages > 0)
1523 page = dequeue_huge_page_node(h, nid);
bf50bab2
NH
1524 spin_unlock(&hugetlb_lock);
1525
94ae8ba7 1526 if (!page)
bf50bab2
NH
1527 page = alloc_buddy_huge_page(h, nid);
1528
1529 return page;
1530}
1531
e4e574b7 1532/*
25985edc 1533 * Increase the hugetlb pool such that it can accommodate a reservation
e4e574b7
AL
1534 * of size 'delta'.
1535 */
a5516438 1536static int gather_surplus_pages(struct hstate *h, int delta)
e4e574b7
AL
1537{
1538 struct list_head surplus_list;
1539 struct page *page, *tmp;
1540 int ret, i;
1541 int needed, allocated;
28073b02 1542 bool alloc_ok = true;
e4e574b7 1543
a5516438 1544 needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
ac09b3a1 1545 if (needed <= 0) {
a5516438 1546 h->resv_huge_pages += delta;
e4e574b7 1547 return 0;
ac09b3a1 1548 }
e4e574b7
AL
1549
1550 allocated = 0;
1551 INIT_LIST_HEAD(&surplus_list);
1552
1553 ret = -ENOMEM;
1554retry:
1555 spin_unlock(&hugetlb_lock);
1556 for (i = 0; i < needed; i++) {
bf50bab2 1557 page = alloc_buddy_huge_page(h, NUMA_NO_NODE);
28073b02
HD
1558 if (!page) {
1559 alloc_ok = false;
1560 break;
1561 }
e4e574b7
AL
1562 list_add(&page->lru, &surplus_list);
1563 }
28073b02 1564 allocated += i;
e4e574b7
AL
1565
1566 /*
1567 * After retaking hugetlb_lock, we need to recalculate 'needed'
1568 * because either resv_huge_pages or free_huge_pages may have changed.
1569 */
1570 spin_lock(&hugetlb_lock);
a5516438
AK
1571 needed = (h->resv_huge_pages + delta) -
1572 (h->free_huge_pages + allocated);
28073b02
HD
1573 if (needed > 0) {
1574 if (alloc_ok)
1575 goto retry;
1576 /*
1577 * We were not able to allocate enough pages to
1578 * satisfy the entire reservation so we free what
1579 * we've allocated so far.
1580 */
1581 goto free;
1582 }
e4e574b7
AL
1583 /*
1584 * The surplus_list now contains _at_least_ the number of extra pages
25985edc 1585 * needed to accommodate the reservation. Add the appropriate number
e4e574b7 1586 * of pages to the hugetlb pool and free the extras back to the buddy
ac09b3a1
AL
1587 * allocator. Commit the entire reservation here to prevent another
1588 * process from stealing the pages as they are added to the pool but
1589 * before they are reserved.
e4e574b7
AL
1590 */
1591 needed += allocated;
a5516438 1592 h->resv_huge_pages += delta;
e4e574b7 1593 ret = 0;
a9869b83 1594
19fc3f0a 1595 /* Free the needed pages to the hugetlb pool */
e4e574b7 1596 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
19fc3f0a
AL
1597 if ((--needed) < 0)
1598 break;
a9869b83
NH
1599 /*
1600 * This page is now managed by the hugetlb allocator and has
1601 * no users -- drop the buddy allocator's reference.
1602 */
1603 put_page_testzero(page);
309381fe 1604 VM_BUG_ON_PAGE(page_count(page), page);
a5516438 1605 enqueue_huge_page(h, page);
19fc3f0a 1606 }
28073b02 1607free:
b0365c8d 1608 spin_unlock(&hugetlb_lock);
19fc3f0a
AL
1609
1610 /* Free unnecessary surplus pages to the buddy allocator */
c0d934ba
JK
1611 list_for_each_entry_safe(page, tmp, &surplus_list, lru)
1612 put_page(page);
a9869b83 1613 spin_lock(&hugetlb_lock);
e4e574b7
AL
1614
1615 return ret;
1616}
1617
1618/*
1619 * When releasing a hugetlb pool reservation, any surplus pages that were
1620 * allocated to satisfy the reservation must be explicitly freed if they were
1621 * never used.
685f3457 1622 * Called with hugetlb_lock held.
e4e574b7 1623 */
a5516438
AK
1624static void return_unused_surplus_pages(struct hstate *h,
1625 unsigned long unused_resv_pages)
e4e574b7 1626{
e4e574b7
AL
1627 unsigned long nr_pages;
1628
ac09b3a1 1629 /* Uncommit the reservation */
a5516438 1630 h->resv_huge_pages -= unused_resv_pages;
ac09b3a1 1631
aa888a74 1632 /* Cannot return gigantic pages currently */
bae7f4ae 1633 if (hstate_is_gigantic(h))
aa888a74
AK
1634 return;
1635
a5516438 1636 nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
e4e574b7 1637
685f3457
LS
1638 /*
1639 * We want to release as many surplus pages as possible, spread
9b5e5d0f
LS
1640 * evenly across all nodes with memory. Iterate across these nodes
1641 * until we can no longer free unreserved surplus pages. This occurs
1642 * when the nodes with surplus pages have no free pages.
1643 * free_pool_huge_page() will balance the the freed pages across the
1644 * on-line nodes with memory and will handle the hstate accounting.
685f3457
LS
1645 */
1646 while (nr_pages--) {
8cebfcd0 1647 if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1))
685f3457 1648 break;
7848a4bf 1649 cond_resched_lock(&hugetlb_lock);
e4e574b7
AL
1650 }
1651}
1652
5e911373 1653
c37f9fb1 1654/*
feba16e2 1655 * vma_needs_reservation, vma_commit_reservation and vma_end_reservation
5e911373 1656 * are used by the huge page allocation routines to manage reservations.
cf3ad20b
MK
1657 *
1658 * vma_needs_reservation is called to determine if the huge page at addr
1659 * within the vma has an associated reservation. If a reservation is
1660 * needed, the value 1 is returned. The caller is then responsible for
1661 * managing the global reservation and subpool usage counts. After
1662 * the huge page has been allocated, vma_commit_reservation is called
feba16e2
MK
1663 * to add the page to the reservation map. If the page allocation fails,
1664 * the reservation must be ended instead of committed. vma_end_reservation
1665 * is called in such cases.
cf3ad20b
MK
1666 *
1667 * In the normal case, vma_commit_reservation returns the same value
1668 * as the preceding vma_needs_reservation call. The only time this
1669 * is not the case is if a reserve map was changed between calls. It
1670 * is the responsibility of the caller to notice the difference and
1671 * take appropriate action.
c37f9fb1 1672 */
5e911373
MK
1673enum vma_resv_mode {
1674 VMA_NEEDS_RESV,
1675 VMA_COMMIT_RESV,
feba16e2 1676 VMA_END_RESV,
5e911373 1677};
cf3ad20b
MK
1678static long __vma_reservation_common(struct hstate *h,
1679 struct vm_area_struct *vma, unsigned long addr,
5e911373 1680 enum vma_resv_mode mode)
c37f9fb1 1681{
4e35f483
JK
1682 struct resv_map *resv;
1683 pgoff_t idx;
cf3ad20b 1684 long ret;
c37f9fb1 1685
4e35f483
JK
1686 resv = vma_resv_map(vma);
1687 if (!resv)
84afd99b 1688 return 1;
c37f9fb1 1689
4e35f483 1690 idx = vma_hugecache_offset(h, vma, addr);
5e911373
MK
1691 switch (mode) {
1692 case VMA_NEEDS_RESV:
cf3ad20b 1693 ret = region_chg(resv, idx, idx + 1);
5e911373
MK
1694 break;
1695 case VMA_COMMIT_RESV:
1696 ret = region_add(resv, idx, idx + 1);
1697 break;
feba16e2 1698 case VMA_END_RESV:
5e911373
MK
1699 region_abort(resv, idx, idx + 1);
1700 ret = 0;
1701 break;
1702 default:
1703 BUG();
1704 }
84afd99b 1705
4e35f483 1706 if (vma->vm_flags & VM_MAYSHARE)
cf3ad20b 1707 return ret;
4e35f483 1708 else
cf3ad20b 1709 return ret < 0 ? ret : 0;
c37f9fb1 1710}
cf3ad20b
MK
1711
1712static long vma_needs_reservation(struct hstate *h,
a5516438 1713 struct vm_area_struct *vma, unsigned long addr)
c37f9fb1 1714{
5e911373 1715 return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV);
cf3ad20b 1716}
84afd99b 1717
cf3ad20b
MK
1718static long vma_commit_reservation(struct hstate *h,
1719 struct vm_area_struct *vma, unsigned long addr)
1720{
5e911373
MK
1721 return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV);
1722}
1723
feba16e2 1724static void vma_end_reservation(struct hstate *h,
5e911373
MK
1725 struct vm_area_struct *vma, unsigned long addr)
1726{
feba16e2 1727 (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV);
c37f9fb1
AW
1728}
1729
a1e78772 1730static struct page *alloc_huge_page(struct vm_area_struct *vma,
04f2cbe3 1731 unsigned long addr, int avoid_reserve)
1da177e4 1732{
90481622 1733 struct hugepage_subpool *spool = subpool_vma(vma);
a5516438 1734 struct hstate *h = hstate_vma(vma);
348ea204 1735 struct page *page;
33039678 1736 long chg, commit;
6d76dcf4
AK
1737 int ret, idx;
1738 struct hugetlb_cgroup *h_cg;
a1e78772 1739
6d76dcf4 1740 idx = hstate_index(h);
a1e78772 1741 /*
90481622
DG
1742 * Processes that did not create the mapping will have no
1743 * reserves and will not have accounted against subpool
1744 * limit. Check that the subpool limit can be made before
1745 * satisfying the allocation MAP_NORESERVE mappings may also
1746 * need pages and subpool limit allocated allocated if no reserve
1747 * mapping overlaps.
a1e78772 1748 */
a5516438 1749 chg = vma_needs_reservation(h, vma, addr);
c37f9fb1 1750 if (chg < 0)
76dcee75 1751 return ERR_PTR(-ENOMEM);
8bb3f12e 1752 if (chg || avoid_reserve)
5e911373 1753 if (hugepage_subpool_get_pages(spool, 1) < 0) {
feba16e2 1754 vma_end_reservation(h, vma, addr);
76dcee75 1755 return ERR_PTR(-ENOSPC);
5e911373 1756 }
1da177e4 1757
6d76dcf4 1758 ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg);
8f34af6f
JZ
1759 if (ret)
1760 goto out_subpool_put;
1761
1da177e4 1762 spin_lock(&hugetlb_lock);
af0ed73e 1763 page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, chg);
81a6fcae 1764 if (!page) {
94ae8ba7 1765 spin_unlock(&hugetlb_lock);
bf50bab2 1766 page = alloc_buddy_huge_page(h, NUMA_NO_NODE);
8f34af6f
JZ
1767 if (!page)
1768 goto out_uncharge_cgroup;
1769
79dbb236
AK
1770 spin_lock(&hugetlb_lock);
1771 list_move(&page->lru, &h->hugepage_activelist);
81a6fcae 1772 /* Fall through */
68842c9b 1773 }
81a6fcae
JK
1774 hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page);
1775 spin_unlock(&hugetlb_lock);
348ea204 1776
90481622 1777 set_page_private(page, (unsigned long)spool);
90d8b7e6 1778
33039678
MK
1779 commit = vma_commit_reservation(h, vma, addr);
1780 if (unlikely(chg > commit)) {
1781 /*
1782 * The page was added to the reservation map between
1783 * vma_needs_reservation and vma_commit_reservation.
1784 * This indicates a race with hugetlb_reserve_pages.
1785 * Adjust for the subpool count incremented above AND
1786 * in hugetlb_reserve_pages for the same page. Also,
1787 * the reservation count added in hugetlb_reserve_pages
1788 * no longer applies.
1789 */
1790 long rsv_adjust;
1791
1792 rsv_adjust = hugepage_subpool_put_pages(spool, 1);
1793 hugetlb_acct_memory(h, -rsv_adjust);
1794 }
90d8b7e6 1795 return page;
8f34af6f
JZ
1796
1797out_uncharge_cgroup:
1798 hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg);
1799out_subpool_put:
1800 if (chg || avoid_reserve)
1801 hugepage_subpool_put_pages(spool, 1);
feba16e2 1802 vma_end_reservation(h, vma, addr);
8f34af6f 1803 return ERR_PTR(-ENOSPC);
b45b5bd6
DG
1804}
1805
74060e4d
NH
1806/*
1807 * alloc_huge_page()'s wrapper which simply returns the page if allocation
1808 * succeeds, otherwise NULL. This function is called from new_vma_page(),
1809 * where no ERR_VALUE is expected to be returned.
1810 */
1811struct page *alloc_huge_page_noerr(struct vm_area_struct *vma,
1812 unsigned long addr, int avoid_reserve)
1813{
1814 struct page *page = alloc_huge_page(vma, addr, avoid_reserve);
1815 if (IS_ERR(page))
1816 page = NULL;
1817 return page;
1818}
1819
91f47662 1820int __weak alloc_bootmem_huge_page(struct hstate *h)
aa888a74
AK
1821{
1822 struct huge_bootmem_page *m;
b2261026 1823 int nr_nodes, node;
aa888a74 1824
b2261026 1825 for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) {
aa888a74
AK
1826 void *addr;
1827
8b89a116
GS
1828 addr = memblock_virt_alloc_try_nid_nopanic(
1829 huge_page_size(h), huge_page_size(h),
1830 0, BOOTMEM_ALLOC_ACCESSIBLE, node);
aa888a74
AK
1831 if (addr) {
1832 /*
1833 * Use the beginning of the huge page to store the
1834 * huge_bootmem_page struct (until gather_bootmem
1835 * puts them into the mem_map).
1836 */
1837 m = addr;
91f47662 1838 goto found;
aa888a74 1839 }
aa888a74
AK
1840 }
1841 return 0;
1842
1843found:
df994ead 1844 BUG_ON(!IS_ALIGNED(virt_to_phys(m), huge_page_size(h)));
aa888a74
AK
1845 /* Put them into a private list first because mem_map is not up yet */
1846 list_add(&m->list, &huge_boot_pages);
1847 m->hstate = h;
1848 return 1;
1849}
1850
f412c97a 1851static void __init prep_compound_huge_page(struct page *page, int order)
18229df5
AW
1852{
1853 if (unlikely(order > (MAX_ORDER - 1)))
1854 prep_compound_gigantic_page(page, order);
1855 else
1856 prep_compound_page(page, order);
1857}
1858
aa888a74
AK
1859/* Put bootmem huge pages into the standard lists after mem_map is up */
1860static void __init gather_bootmem_prealloc(void)
1861{
1862 struct huge_bootmem_page *m;
1863
1864 list_for_each_entry(m, &huge_boot_pages, list) {
aa888a74 1865 struct hstate *h = m->hstate;
ee8f248d
BB
1866 struct page *page;
1867
1868#ifdef CONFIG_HIGHMEM
1869 page = pfn_to_page(m->phys >> PAGE_SHIFT);
8b89a116
GS
1870 memblock_free_late(__pa(m),
1871 sizeof(struct huge_bootmem_page));
ee8f248d
BB
1872#else
1873 page = virt_to_page(m);
1874#endif
aa888a74 1875 WARN_ON(page_count(page) != 1);
18229df5 1876 prep_compound_huge_page(page, h->order);
ef5a22be 1877 WARN_ON(PageReserved(page));
aa888a74 1878 prep_new_huge_page(h, page, page_to_nid(page));
b0320c7b
RA
1879 /*
1880 * If we had gigantic hugepages allocated at boot time, we need
1881 * to restore the 'stolen' pages to totalram_pages in order to
1882 * fix confusing memory reports from free(1) and another
1883 * side-effects, like CommitLimit going negative.
1884 */
bae7f4ae 1885 if (hstate_is_gigantic(h))
3dcc0571 1886 adjust_managed_page_count(page, 1 << h->order);
aa888a74
AK
1887 }
1888}
1889
8faa8b07 1890static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
1da177e4
LT
1891{
1892 unsigned long i;
a5516438 1893
e5ff2159 1894 for (i = 0; i < h->max_huge_pages; ++i) {
bae7f4ae 1895 if (hstate_is_gigantic(h)) {
aa888a74
AK
1896 if (!alloc_bootmem_huge_page(h))
1897 break;
9b5e5d0f 1898 } else if (!alloc_fresh_huge_page(h,
8cebfcd0 1899 &node_states[N_MEMORY]))
1da177e4 1900 break;
1da177e4 1901 }
8faa8b07 1902 h->max_huge_pages = i;
e5ff2159
AK
1903}
1904
1905static void __init hugetlb_init_hstates(void)
1906{
1907 struct hstate *h;
1908
1909 for_each_hstate(h) {
641844f5
NH
1910 if (minimum_order > huge_page_order(h))
1911 minimum_order = huge_page_order(h);
1912
8faa8b07 1913 /* oversize hugepages were init'ed in early boot */
bae7f4ae 1914 if (!hstate_is_gigantic(h))
8faa8b07 1915 hugetlb_hstate_alloc_pages(h);
e5ff2159 1916 }
641844f5 1917 VM_BUG_ON(minimum_order == UINT_MAX);
e5ff2159
AK
1918}
1919
4abd32db
AK
1920static char * __init memfmt(char *buf, unsigned long n)
1921{
1922 if (n >= (1UL << 30))
1923 sprintf(buf, "%lu GB", n >> 30);
1924 else if (n >= (1UL << 20))
1925 sprintf(buf, "%lu MB", n >> 20);
1926 else
1927 sprintf(buf, "%lu KB", n >> 10);
1928 return buf;
1929}
1930
e5ff2159
AK
1931static void __init report_hugepages(void)
1932{
1933 struct hstate *h;
1934
1935 for_each_hstate(h) {
4abd32db 1936 char buf[32];
ffb22af5 1937 pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n",
4abd32db
AK
1938 memfmt(buf, huge_page_size(h)),
1939 h->free_huge_pages);
e5ff2159
AK
1940 }
1941}
1942
1da177e4 1943#ifdef CONFIG_HIGHMEM
6ae11b27
LS
1944static void try_to_free_low(struct hstate *h, unsigned long count,
1945 nodemask_t *nodes_allowed)
1da177e4 1946{
4415cc8d
CL
1947 int i;
1948
bae7f4ae 1949 if (hstate_is_gigantic(h))
aa888a74
AK
1950 return;
1951
6ae11b27 1952 for_each_node_mask(i, *nodes_allowed) {
1da177e4 1953 struct page *page, *next;
a5516438
AK
1954 struct list_head *freel = &h->hugepage_freelists[i];
1955 list_for_each_entry_safe(page, next, freel, lru) {
1956 if (count >= h->nr_huge_pages)
6b0c880d 1957 return;
1da177e4
LT
1958 if (PageHighMem(page))
1959 continue;
1960 list_del(&page->lru);
e5ff2159 1961 update_and_free_page(h, page);
a5516438
AK
1962 h->free_huge_pages--;
1963 h->free_huge_pages_node[page_to_nid(page)]--;
1da177e4
LT
1964 }
1965 }
1966}
1967#else
6ae11b27
LS
1968static inline void try_to_free_low(struct hstate *h, unsigned long count,
1969 nodemask_t *nodes_allowed)
1da177e4
LT
1970{
1971}
1972#endif
1973
20a0307c
WF
1974/*
1975 * Increment or decrement surplus_huge_pages. Keep node-specific counters
1976 * balanced by operating on them in a round-robin fashion.
1977 * Returns 1 if an adjustment was made.
1978 */
6ae11b27
LS
1979static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
1980 int delta)
20a0307c 1981{
b2261026 1982 int nr_nodes, node;
20a0307c
WF
1983
1984 VM_BUG_ON(delta != -1 && delta != 1);
20a0307c 1985
b2261026
JK
1986 if (delta < 0) {
1987 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
1988 if (h->surplus_huge_pages_node[node])
1989 goto found;
e8c5c824 1990 }
b2261026
JK
1991 } else {
1992 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
1993 if (h->surplus_huge_pages_node[node] <
1994 h->nr_huge_pages_node[node])
1995 goto found;
e8c5c824 1996 }
b2261026
JK
1997 }
1998 return 0;
20a0307c 1999
b2261026
JK
2000found:
2001 h->surplus_huge_pages += delta;
2002 h->surplus_huge_pages_node[node] += delta;
2003 return 1;
20a0307c
WF
2004}
2005
a5516438 2006#define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
6ae11b27
LS
2007static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count,
2008 nodemask_t *nodes_allowed)
1da177e4 2009{
7893d1d5 2010 unsigned long min_count, ret;
1da177e4 2011
944d9fec 2012 if (hstate_is_gigantic(h) && !gigantic_page_supported())
aa888a74
AK
2013 return h->max_huge_pages;
2014
7893d1d5
AL
2015 /*
2016 * Increase the pool size
2017 * First take pages out of surplus state. Then make up the
2018 * remaining difference by allocating fresh huge pages.
d1c3fb1f
NA
2019 *
2020 * We might race with alloc_buddy_huge_page() here and be unable
2021 * to convert a surplus huge page to a normal huge page. That is
2022 * not critical, though, it just means the overall size of the
2023 * pool might be one hugepage larger than it needs to be, but
2024 * within all the constraints specified by the sysctls.
7893d1d5 2025 */
1da177e4 2026 spin_lock(&hugetlb_lock);
a5516438 2027 while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
6ae11b27 2028 if (!adjust_pool_surplus(h, nodes_allowed, -1))
7893d1d5
AL
2029 break;
2030 }
2031
a5516438 2032 while (count > persistent_huge_pages(h)) {
7893d1d5
AL
2033 /*
2034 * If this allocation races such that we no longer need the
2035 * page, free_huge_page will handle it by freeing the page
2036 * and reducing the surplus.
2037 */
2038 spin_unlock(&hugetlb_lock);
944d9fec
LC
2039 if (hstate_is_gigantic(h))
2040 ret = alloc_fresh_gigantic_page(h, nodes_allowed);
2041 else
2042 ret = alloc_fresh_huge_page(h, nodes_allowed);
7893d1d5
AL
2043 spin_lock(&hugetlb_lock);
2044 if (!ret)
2045 goto out;
2046
536240f2
MG
2047 /* Bail for signals. Probably ctrl-c from user */
2048 if (signal_pending(current))
2049 goto out;
7893d1d5 2050 }
7893d1d5
AL
2051
2052 /*
2053 * Decrease the pool size
2054 * First return free pages to the buddy allocator (being careful
2055 * to keep enough around to satisfy reservations). Then place
2056 * pages into surplus state as needed so the pool will shrink
2057 * to the desired size as pages become free.
d1c3fb1f
NA
2058 *
2059 * By placing pages into the surplus state independent of the
2060 * overcommit value, we are allowing the surplus pool size to
2061 * exceed overcommit. There are few sane options here. Since
2062 * alloc_buddy_huge_page() is checking the global counter,
2063 * though, we'll note that we're not allowed to exceed surplus
2064 * and won't grow the pool anywhere else. Not until one of the
2065 * sysctls are changed, or the surplus pages go out of use.
7893d1d5 2066 */
a5516438 2067 min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
6b0c880d 2068 min_count = max(count, min_count);
6ae11b27 2069 try_to_free_low(h, min_count, nodes_allowed);
a5516438 2070 while (min_count < persistent_huge_pages(h)) {
6ae11b27 2071 if (!free_pool_huge_page(h, nodes_allowed, 0))
1da177e4 2072 break;
55f67141 2073 cond_resched_lock(&hugetlb_lock);
1da177e4 2074 }
a5516438 2075 while (count < persistent_huge_pages(h)) {
6ae11b27 2076 if (!adjust_pool_surplus(h, nodes_allowed, 1))
7893d1d5
AL
2077 break;
2078 }
2079out:
a5516438 2080 ret = persistent_huge_pages(h);
1da177e4 2081 spin_unlock(&hugetlb_lock);
7893d1d5 2082 return ret;
1da177e4
LT
2083}
2084
a3437870
NA
2085#define HSTATE_ATTR_RO(_name) \
2086 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
2087
2088#define HSTATE_ATTR(_name) \
2089 static struct kobj_attribute _name##_attr = \
2090 __ATTR(_name, 0644, _name##_show, _name##_store)
2091
2092static struct kobject *hugepages_kobj;
2093static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
2094
9a305230
LS
2095static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
2096
2097static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
a3437870
NA
2098{
2099 int i;
9a305230 2100
a3437870 2101 for (i = 0; i < HUGE_MAX_HSTATE; i++)
9a305230
LS
2102 if (hstate_kobjs[i] == kobj) {
2103 if (nidp)
2104 *nidp = NUMA_NO_NODE;
a3437870 2105 return &hstates[i];
9a305230
LS
2106 }
2107
2108 return kobj_to_node_hstate(kobj, nidp);
a3437870
NA
2109}
2110
06808b08 2111static ssize_t nr_hugepages_show_common(struct kobject *kobj,
a3437870
NA
2112 struct kobj_attribute *attr, char *buf)
2113{
9a305230
LS
2114 struct hstate *h;
2115 unsigned long nr_huge_pages;
2116 int nid;
2117
2118 h = kobj_to_hstate(kobj, &nid);
2119 if (nid == NUMA_NO_NODE)
2120 nr_huge_pages = h->nr_huge_pages;
2121 else
2122 nr_huge_pages = h->nr_huge_pages_node[nid];
2123
2124 return sprintf(buf, "%lu\n", nr_huge_pages);
a3437870 2125}
adbe8726 2126
238d3c13
DR
2127static ssize_t __nr_hugepages_store_common(bool obey_mempolicy,
2128 struct hstate *h, int nid,
2129 unsigned long count, size_t len)
a3437870
NA
2130{
2131 int err;
bad44b5b 2132 NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY);
a3437870 2133
944d9fec 2134 if (hstate_is_gigantic(h) && !gigantic_page_supported()) {
adbe8726
EM
2135 err = -EINVAL;
2136 goto out;
2137 }
2138
9a305230
LS
2139 if (nid == NUMA_NO_NODE) {
2140 /*
2141 * global hstate attribute
2142 */
2143 if (!(obey_mempolicy &&
2144 init_nodemask_of_mempolicy(nodes_allowed))) {
2145 NODEMASK_FREE(nodes_allowed);
8cebfcd0 2146 nodes_allowed = &node_states[N_MEMORY];
9a305230
LS
2147 }
2148 } else if (nodes_allowed) {
2149 /*
2150 * per node hstate attribute: adjust count to global,
2151 * but restrict alloc/free to the specified node.
2152 */
2153 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
2154 init_nodemask_of_node(nodes_allowed, nid);
2155 } else
8cebfcd0 2156 nodes_allowed = &node_states[N_MEMORY];
9a305230 2157
06808b08 2158 h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed);
a3437870 2159
8cebfcd0 2160 if (nodes_allowed != &node_states[N_MEMORY])
06808b08
LS
2161 NODEMASK_FREE(nodes_allowed);
2162
2163 return len;
adbe8726
EM
2164out:
2165 NODEMASK_FREE(nodes_allowed);
2166 return err;
06808b08
LS
2167}
2168
238d3c13
DR
2169static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
2170 struct kobject *kobj, const char *buf,
2171 size_t len)
2172{
2173 struct hstate *h;
2174 unsigned long count;
2175 int nid;
2176 int err;
2177
2178 err = kstrtoul(buf, 10, &count);
2179 if (err)
2180 return err;
2181
2182 h = kobj_to_hstate(kobj, &nid);
2183 return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len);
2184}
2185
06808b08
LS
2186static ssize_t nr_hugepages_show(struct kobject *kobj,
2187 struct kobj_attribute *attr, char *buf)
2188{
2189 return nr_hugepages_show_common(kobj, attr, buf);
2190}
2191
2192static ssize_t nr_hugepages_store(struct kobject *kobj,
2193 struct kobj_attribute *attr, const char *buf, size_t len)
2194{
238d3c13 2195 return nr_hugepages_store_common(false, kobj, buf, len);
a3437870
NA
2196}
2197HSTATE_ATTR(nr_hugepages);
2198
06808b08
LS
2199#ifdef CONFIG_NUMA
2200
2201/*
2202 * hstate attribute for optionally mempolicy-based constraint on persistent
2203 * huge page alloc/free.
2204 */
2205static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
2206 struct kobj_attribute *attr, char *buf)
2207{
2208 return nr_hugepages_show_common(kobj, attr, buf);
2209}
2210
2211static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
2212 struct kobj_attribute *attr, const char *buf, size_t len)
2213{
238d3c13 2214 return nr_hugepages_store_common(true, kobj, buf, len);
06808b08
LS
2215}
2216HSTATE_ATTR(nr_hugepages_mempolicy);
2217#endif
2218
2219
a3437870
NA
2220static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
2221 struct kobj_attribute *attr, char *buf)
2222{
9a305230 2223 struct hstate *h = kobj_to_hstate(kobj, NULL);
a3437870
NA
2224 return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
2225}
adbe8726 2226
a3437870
NA
2227static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
2228 struct kobj_attribute *attr, const char *buf, size_t count)
2229{
2230 int err;
2231 unsigned long input;
9a305230 2232 struct hstate *h = kobj_to_hstate(kobj, NULL);
a3437870 2233
bae7f4ae 2234 if (hstate_is_gigantic(h))
adbe8726
EM
2235 return -EINVAL;
2236
3dbb95f7 2237 err = kstrtoul(buf, 10, &input);
a3437870 2238 if (err)
73ae31e5 2239 return err;
a3437870
NA
2240
2241 spin_lock(&hugetlb_lock);
2242 h->nr_overcommit_huge_pages = input;
2243 spin_unlock(&hugetlb_lock);
2244
2245 return count;
2246}
2247HSTATE_ATTR(nr_overcommit_hugepages);
2248
2249static ssize_t free_hugepages_show(struct kobject *kobj,
2250 struct kobj_attribute *attr, char *buf)
2251{
9a305230
LS
2252 struct hstate *h;
2253 unsigned long free_huge_pages;
2254 int nid;
2255
2256 h = kobj_to_hstate(kobj, &nid);
2257 if (nid == NUMA_NO_NODE)
2258 free_huge_pages = h->free_huge_pages;
2259 else
2260 free_huge_pages = h->free_huge_pages_node[nid];
2261
2262 return sprintf(buf, "%lu\n", free_huge_pages);
a3437870
NA
2263}
2264HSTATE_ATTR_RO(free_hugepages);
2265
2266static ssize_t resv_hugepages_show(struct kobject *kobj,
2267 struct kobj_attribute *attr, char *buf)
2268{
9a305230 2269 struct hstate *h = kobj_to_hstate(kobj, NULL);
a3437870
NA
2270 return sprintf(buf, "%lu\n", h->resv_huge_pages);
2271}
2272HSTATE_ATTR_RO(resv_hugepages);
2273
2274static ssize_t surplus_hugepages_show(struct kobject *kobj,
2275 struct kobj_attribute *attr, char *buf)
2276{
9a305230
LS
2277 struct hstate *h;
2278 unsigned long surplus_huge_pages;
2279 int nid;
2280
2281 h = kobj_to_hstate(kobj, &nid);
2282 if (nid == NUMA_NO_NODE)
2283 surplus_huge_pages = h->surplus_huge_pages;
2284 else
2285 surplus_huge_pages = h->surplus_huge_pages_node[nid];
2286
2287 return sprintf(buf, "%lu\n", surplus_huge_pages);
a3437870
NA
2288}
2289HSTATE_ATTR_RO(surplus_hugepages);
2290
2291static struct attribute *hstate_attrs[] = {
2292 &nr_hugepages_attr.attr,
2293 &nr_overcommit_hugepages_attr.attr,
2294 &free_hugepages_attr.attr,
2295 &resv_hugepages_attr.attr,
2296 &surplus_hugepages_attr.attr,
06808b08
LS
2297#ifdef CONFIG_NUMA
2298 &nr_hugepages_mempolicy_attr.attr,
2299#endif
a3437870
NA
2300 NULL,
2301};
2302
2303static struct attribute_group hstate_attr_group = {
2304 .attrs = hstate_attrs,
2305};
2306
094e9539
JM
2307static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
2308 struct kobject **hstate_kobjs,
2309 struct attribute_group *hstate_attr_group)
a3437870
NA
2310{
2311 int retval;
972dc4de 2312 int hi = hstate_index(h);
a3437870 2313
9a305230
LS
2314 hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
2315 if (!hstate_kobjs[hi])
a3437870
NA
2316 return -ENOMEM;
2317
9a305230 2318 retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
a3437870 2319 if (retval)
9a305230 2320 kobject_put(hstate_kobjs[hi]);
a3437870
NA
2321
2322 return retval;
2323}
2324
2325static void __init hugetlb_sysfs_init(void)
2326{
2327 struct hstate *h;
2328 int err;
2329
2330 hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
2331 if (!hugepages_kobj)
2332 return;
2333
2334 for_each_hstate(h) {
9a305230
LS
2335 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
2336 hstate_kobjs, &hstate_attr_group);
a3437870 2337 if (err)
ffb22af5 2338 pr_err("Hugetlb: Unable to add hstate %s", h->name);
a3437870
NA
2339 }
2340}
2341
9a305230
LS
2342#ifdef CONFIG_NUMA
2343
2344/*
2345 * node_hstate/s - associate per node hstate attributes, via their kobjects,
10fbcf4c
KS
2346 * with node devices in node_devices[] using a parallel array. The array
2347 * index of a node device or _hstate == node id.
2348 * This is here to avoid any static dependency of the node device driver, in
9a305230
LS
2349 * the base kernel, on the hugetlb module.
2350 */
2351struct node_hstate {
2352 struct kobject *hugepages_kobj;
2353 struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
2354};
2355struct node_hstate node_hstates[MAX_NUMNODES];
2356
2357/*
10fbcf4c 2358 * A subset of global hstate attributes for node devices
9a305230
LS
2359 */
2360static struct attribute *per_node_hstate_attrs[] = {
2361 &nr_hugepages_attr.attr,
2362 &free_hugepages_attr.attr,
2363 &surplus_hugepages_attr.attr,
2364 NULL,
2365};
2366
2367static struct attribute_group per_node_hstate_attr_group = {
2368 .attrs = per_node_hstate_attrs,
2369};
2370
2371/*
10fbcf4c 2372 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
9a305230
LS
2373 * Returns node id via non-NULL nidp.
2374 */
2375static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
2376{
2377 int nid;
2378
2379 for (nid = 0; nid < nr_node_ids; nid++) {
2380 struct node_hstate *nhs = &node_hstates[nid];
2381 int i;
2382 for (i = 0; i < HUGE_MAX_HSTATE; i++)
2383 if (nhs->hstate_kobjs[i] == kobj) {
2384 if (nidp)
2385 *nidp = nid;
2386 return &hstates[i];
2387 }
2388 }
2389
2390 BUG();
2391 return NULL;
2392}
2393
2394/*
10fbcf4c 2395 * Unregister hstate attributes from a single node device.
9a305230
LS
2396 * No-op if no hstate attributes attached.
2397 */
3cd8b44f 2398static void hugetlb_unregister_node(struct node *node)
9a305230
LS
2399{
2400 struct hstate *h;
10fbcf4c 2401 struct node_hstate *nhs = &node_hstates[node->dev.id];
9a305230
LS
2402
2403 if (!nhs->hugepages_kobj)
9b5e5d0f 2404 return; /* no hstate attributes */
9a305230 2405
972dc4de
AK
2406 for_each_hstate(h) {
2407 int idx = hstate_index(h);
2408 if (nhs->hstate_kobjs[idx]) {
2409 kobject_put(nhs->hstate_kobjs[idx]);
2410 nhs->hstate_kobjs[idx] = NULL;
9a305230 2411 }
972dc4de 2412 }
9a305230
LS
2413
2414 kobject_put(nhs->hugepages_kobj);
2415 nhs->hugepages_kobj = NULL;
2416}
2417
2418/*
10fbcf4c 2419 * hugetlb module exit: unregister hstate attributes from node devices
9a305230
LS
2420 * that have them.
2421 */
2422static void hugetlb_unregister_all_nodes(void)
2423{
2424 int nid;
2425
2426 /*
10fbcf4c 2427 * disable node device registrations.
9a305230
LS
2428 */
2429 register_hugetlbfs_with_node(NULL, NULL);
2430
2431 /*
2432 * remove hstate attributes from any nodes that have them.
2433 */
2434 for (nid = 0; nid < nr_node_ids; nid++)
8732794b 2435 hugetlb_unregister_node(node_devices[nid]);
9a305230
LS
2436}
2437
2438/*
10fbcf4c 2439 * Register hstate attributes for a single node device.
9a305230
LS
2440 * No-op if attributes already registered.
2441 */
3cd8b44f 2442static void hugetlb_register_node(struct node *node)
9a305230
LS
2443{
2444 struct hstate *h;
10fbcf4c 2445 struct node_hstate *nhs = &node_hstates[node->dev.id];
9a305230
LS
2446 int err;
2447
2448 if (nhs->hugepages_kobj)
2449 return; /* already allocated */
2450
2451 nhs->hugepages_kobj = kobject_create_and_add("hugepages",
10fbcf4c 2452 &node->dev.kobj);
9a305230
LS
2453 if (!nhs->hugepages_kobj)
2454 return;
2455
2456 for_each_hstate(h) {
2457 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
2458 nhs->hstate_kobjs,
2459 &per_node_hstate_attr_group);
2460 if (err) {
ffb22af5
AM
2461 pr_err("Hugetlb: Unable to add hstate %s for node %d\n",
2462 h->name, node->dev.id);
9a305230
LS
2463 hugetlb_unregister_node(node);
2464 break;
2465 }
2466 }
2467}
2468
2469/*
9b5e5d0f 2470 * hugetlb init time: register hstate attributes for all registered node
10fbcf4c
KS
2471 * devices of nodes that have memory. All on-line nodes should have
2472 * registered their associated device by this time.
9a305230 2473 */
7d9ca000 2474static void __init hugetlb_register_all_nodes(void)
9a305230
LS
2475{
2476 int nid;
2477
8cebfcd0 2478 for_each_node_state(nid, N_MEMORY) {
8732794b 2479 struct node *node = node_devices[nid];
10fbcf4c 2480 if (node->dev.id == nid)
9a305230
LS
2481 hugetlb_register_node(node);
2482 }
2483
2484 /*
10fbcf4c 2485 * Let the node device driver know we're here so it can
9a305230
LS
2486 * [un]register hstate attributes on node hotplug.
2487 */
2488 register_hugetlbfs_with_node(hugetlb_register_node,
2489 hugetlb_unregister_node);
2490}
2491#else /* !CONFIG_NUMA */
2492
2493static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
2494{
2495 BUG();
2496 if (nidp)
2497 *nidp = -1;
2498 return NULL;
2499}
2500
2501static void hugetlb_unregister_all_nodes(void) { }
2502
2503static void hugetlb_register_all_nodes(void) { }
2504
2505#endif
2506
a3437870
NA
2507static void __exit hugetlb_exit(void)
2508{
2509 struct hstate *h;
2510
9a305230
LS
2511 hugetlb_unregister_all_nodes();
2512
a3437870 2513 for_each_hstate(h) {
972dc4de 2514 kobject_put(hstate_kobjs[hstate_index(h)]);
a3437870
NA
2515 }
2516
2517 kobject_put(hugepages_kobj);
c672c7f2 2518 kfree(hugetlb_fault_mutex_table);
a3437870
NA
2519}
2520module_exit(hugetlb_exit);
2521
2522static int __init hugetlb_init(void)
2523{
8382d914
DB
2524 int i;
2525
457c1b27 2526 if (!hugepages_supported())
0ef89d25 2527 return 0;
a3437870 2528
e11bfbfc
NP
2529 if (!size_to_hstate(default_hstate_size)) {
2530 default_hstate_size = HPAGE_SIZE;
2531 if (!size_to_hstate(default_hstate_size))
2532 hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
a3437870 2533 }
972dc4de 2534 default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size));
e11bfbfc
NP
2535 if (default_hstate_max_huge_pages)
2536 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
a3437870
NA
2537
2538 hugetlb_init_hstates();
aa888a74 2539 gather_bootmem_prealloc();
a3437870
NA
2540 report_hugepages();
2541
2542 hugetlb_sysfs_init();
9a305230 2543 hugetlb_register_all_nodes();
7179e7bf 2544 hugetlb_cgroup_file_init();
9a305230 2545
8382d914
DB
2546#ifdef CONFIG_SMP
2547 num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus());
2548#else
2549 num_fault_mutexes = 1;
2550#endif
c672c7f2 2551 hugetlb_fault_mutex_table =
8382d914 2552 kmalloc(sizeof(struct mutex) * num_fault_mutexes, GFP_KERNEL);
c672c7f2 2553 BUG_ON(!hugetlb_fault_mutex_table);
8382d914
DB
2554
2555 for (i = 0; i < num_fault_mutexes; i++)
c672c7f2 2556 mutex_init(&hugetlb_fault_mutex_table[i]);
a3437870
NA
2557 return 0;
2558}
2559module_init(hugetlb_init);
2560
2561/* Should be called on processing a hugepagesz=... option */
2562void __init hugetlb_add_hstate(unsigned order)
2563{
2564 struct hstate *h;
8faa8b07
AK
2565 unsigned long i;
2566
a3437870 2567 if (size_to_hstate(PAGE_SIZE << order)) {
ffb22af5 2568 pr_warning("hugepagesz= specified twice, ignoring\n");
a3437870
NA
2569 return;
2570 }
47d38344 2571 BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE);
a3437870 2572 BUG_ON(order == 0);
47d38344 2573 h = &hstates[hugetlb_max_hstate++];
a3437870
NA
2574 h->order = order;
2575 h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
8faa8b07
AK
2576 h->nr_huge_pages = 0;
2577 h->free_huge_pages = 0;
2578 for (i = 0; i < MAX_NUMNODES; ++i)
2579 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
0edaecfa 2580 INIT_LIST_HEAD(&h->hugepage_activelist);
8cebfcd0
LJ
2581 h->next_nid_to_alloc = first_node(node_states[N_MEMORY]);
2582 h->next_nid_to_free = first_node(node_states[N_MEMORY]);
a3437870
NA
2583 snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
2584 huge_page_size(h)/1024);
8faa8b07 2585
a3437870
NA
2586 parsed_hstate = h;
2587}
2588
e11bfbfc 2589static int __init hugetlb_nrpages_setup(char *s)
a3437870
NA
2590{
2591 unsigned long *mhp;
8faa8b07 2592 static unsigned long *last_mhp;
a3437870
NA
2593
2594 /*
47d38344 2595 * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet,
a3437870
NA
2596 * so this hugepages= parameter goes to the "default hstate".
2597 */
47d38344 2598 if (!hugetlb_max_hstate)
a3437870
NA
2599 mhp = &default_hstate_max_huge_pages;
2600 else
2601 mhp = &parsed_hstate->max_huge_pages;
2602
8faa8b07 2603 if (mhp == last_mhp) {
ffb22af5
AM
2604 pr_warning("hugepages= specified twice without "
2605 "interleaving hugepagesz=, ignoring\n");
8faa8b07
AK
2606 return 1;
2607 }
2608
a3437870
NA
2609 if (sscanf(s, "%lu", mhp) <= 0)
2610 *mhp = 0;
2611
8faa8b07
AK
2612 /*
2613 * Global state is always initialized later in hugetlb_init.
2614 * But we need to allocate >= MAX_ORDER hstates here early to still
2615 * use the bootmem allocator.
2616 */
47d38344 2617 if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER)
8faa8b07
AK
2618 hugetlb_hstate_alloc_pages(parsed_hstate);
2619
2620 last_mhp = mhp;
2621
a3437870
NA
2622 return 1;
2623}
e11bfbfc
NP
2624__setup("hugepages=", hugetlb_nrpages_setup);
2625
2626static int __init hugetlb_default_setup(char *s)
2627{
2628 default_hstate_size = memparse(s, &s);
2629 return 1;
2630}
2631__setup("default_hugepagesz=", hugetlb_default_setup);
a3437870 2632
8a213460
NA
2633static unsigned int cpuset_mems_nr(unsigned int *array)
2634{
2635 int node;
2636 unsigned int nr = 0;
2637
2638 for_each_node_mask(node, cpuset_current_mems_allowed)
2639 nr += array[node];
2640
2641 return nr;
2642}
2643
2644#ifdef CONFIG_SYSCTL
06808b08
LS
2645static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
2646 struct ctl_table *table, int write,
2647 void __user *buffer, size_t *length, loff_t *ppos)
1da177e4 2648{
e5ff2159 2649 struct hstate *h = &default_hstate;
238d3c13 2650 unsigned long tmp = h->max_huge_pages;
08d4a246 2651 int ret;
e5ff2159 2652
457c1b27
NA
2653 if (!hugepages_supported())
2654 return -ENOTSUPP;
2655
e5ff2159
AK
2656 table->data = &tmp;
2657 table->maxlen = sizeof(unsigned long);
08d4a246
MH
2658 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2659 if (ret)
2660 goto out;
e5ff2159 2661
238d3c13
DR
2662 if (write)
2663 ret = __nr_hugepages_store_common(obey_mempolicy, h,
2664 NUMA_NO_NODE, tmp, *length);
08d4a246
MH
2665out:
2666 return ret;
1da177e4 2667}
396faf03 2668
06808b08
LS
2669int hugetlb_sysctl_handler(struct ctl_table *table, int write,
2670 void __user *buffer, size_t *length, loff_t *ppos)
2671{
2672
2673 return hugetlb_sysctl_handler_common(false, table, write,
2674 buffer, length, ppos);
2675}
2676
2677#ifdef CONFIG_NUMA
2678int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
2679 void __user *buffer, size_t *length, loff_t *ppos)
2680{
2681 return hugetlb_sysctl_handler_common(true, table, write,
2682 buffer, length, ppos);
2683}
2684#endif /* CONFIG_NUMA */
2685
a3d0c6aa 2686int hugetlb_overcommit_handler(struct ctl_table *table, int write,
8d65af78 2687 void __user *buffer,
a3d0c6aa
NA
2688 size_t *length, loff_t *ppos)
2689{
a5516438 2690 struct hstate *h = &default_hstate;
e5ff2159 2691 unsigned long tmp;
08d4a246 2692 int ret;
e5ff2159 2693
457c1b27
NA
2694 if (!hugepages_supported())
2695 return -ENOTSUPP;
2696
c033a93c 2697 tmp = h->nr_overcommit_huge_pages;
e5ff2159 2698
bae7f4ae 2699 if (write && hstate_is_gigantic(h))
adbe8726
EM
2700 return -EINVAL;
2701
e5ff2159
AK
2702 table->data = &tmp;
2703 table->maxlen = sizeof(unsigned long);
08d4a246
MH
2704 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2705 if (ret)
2706 goto out;
e5ff2159
AK
2707
2708 if (write) {
2709 spin_lock(&hugetlb_lock);
2710 h->nr_overcommit_huge_pages = tmp;
2711 spin_unlock(&hugetlb_lock);
2712 }
08d4a246
MH
2713out:
2714 return ret;
a3d0c6aa
NA
2715}
2716
1da177e4
LT
2717#endif /* CONFIG_SYSCTL */
2718
e1759c21 2719void hugetlb_report_meminfo(struct seq_file *m)
1da177e4 2720{
a5516438 2721 struct hstate *h = &default_hstate;
457c1b27
NA
2722 if (!hugepages_supported())
2723 return;
e1759c21 2724 seq_printf(m,
4f98a2fe
RR
2725 "HugePages_Total: %5lu\n"
2726 "HugePages_Free: %5lu\n"
2727 "HugePages_Rsvd: %5lu\n"
2728 "HugePages_Surp: %5lu\n"
2729 "Hugepagesize: %8lu kB\n",
a5516438
AK
2730 h->nr_huge_pages,
2731 h->free_huge_pages,
2732 h->resv_huge_pages,
2733 h->surplus_huge_pages,
2734 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
1da177e4
LT
2735}
2736
2737int hugetlb_report_node_meminfo(int nid, char *buf)
2738{
a5516438 2739 struct hstate *h = &default_hstate;
457c1b27
NA
2740 if (!hugepages_supported())
2741 return 0;
1da177e4
LT
2742 return sprintf(buf,
2743 "Node %d HugePages_Total: %5u\n"
a1de0919
NA
2744 "Node %d HugePages_Free: %5u\n"
2745 "Node %d HugePages_Surp: %5u\n",
a5516438
AK
2746 nid, h->nr_huge_pages_node[nid],
2747 nid, h->free_huge_pages_node[nid],
2748 nid, h->surplus_huge_pages_node[nid]);
1da177e4
LT
2749}
2750
949f7ec5
DR
2751void hugetlb_show_meminfo(void)
2752{
2753 struct hstate *h;
2754 int nid;
2755
457c1b27
NA
2756 if (!hugepages_supported())
2757 return;
2758
949f7ec5
DR
2759 for_each_node_state(nid, N_MEMORY)
2760 for_each_hstate(h)
2761 pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
2762 nid,
2763 h->nr_huge_pages_node[nid],
2764 h->free_huge_pages_node[nid],
2765 h->surplus_huge_pages_node[nid],
2766 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
2767}
2768
1da177e4
LT
2769/* Return the number pages of memory we physically have, in PAGE_SIZE units. */
2770unsigned long hugetlb_total_pages(void)
2771{
d0028588
WL
2772 struct hstate *h;
2773 unsigned long nr_total_pages = 0;
2774
2775 for_each_hstate(h)
2776 nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h);
2777 return nr_total_pages;
1da177e4 2778}
1da177e4 2779
a5516438 2780static int hugetlb_acct_memory(struct hstate *h, long delta)
fc1b8a73
MG
2781{
2782 int ret = -ENOMEM;
2783
2784 spin_lock(&hugetlb_lock);
2785 /*
2786 * When cpuset is configured, it breaks the strict hugetlb page
2787 * reservation as the accounting is done on a global variable. Such
2788 * reservation is completely rubbish in the presence of cpuset because
2789 * the reservation is not checked against page availability for the
2790 * current cpuset. Application can still potentially OOM'ed by kernel
2791 * with lack of free htlb page in cpuset that the task is in.
2792 * Attempt to enforce strict accounting with cpuset is almost
2793 * impossible (or too ugly) because cpuset is too fluid that
2794 * task or memory node can be dynamically moved between cpusets.
2795 *
2796 * The change of semantics for shared hugetlb mapping with cpuset is
2797 * undesirable. However, in order to preserve some of the semantics,
2798 * we fall back to check against current free page availability as
2799 * a best attempt and hopefully to minimize the impact of changing
2800 * semantics that cpuset has.
2801 */
2802 if (delta > 0) {
a5516438 2803 if (gather_surplus_pages(h, delta) < 0)
fc1b8a73
MG
2804 goto out;
2805
a5516438
AK
2806 if (delta > cpuset_mems_nr(h->free_huge_pages_node)) {
2807 return_unused_surplus_pages(h, delta);
fc1b8a73
MG
2808 goto out;
2809 }
2810 }
2811
2812 ret = 0;
2813 if (delta < 0)
a5516438 2814 return_unused_surplus_pages(h, (unsigned long) -delta);
fc1b8a73
MG
2815
2816out:
2817 spin_unlock(&hugetlb_lock);
2818 return ret;
2819}
2820
84afd99b
AW
2821static void hugetlb_vm_op_open(struct vm_area_struct *vma)
2822{
f522c3ac 2823 struct resv_map *resv = vma_resv_map(vma);
84afd99b
AW
2824
2825 /*
2826 * This new VMA should share its siblings reservation map if present.
2827 * The VMA will only ever have a valid reservation map pointer where
2828 * it is being copied for another still existing VMA. As that VMA
25985edc 2829 * has a reference to the reservation map it cannot disappear until
84afd99b
AW
2830 * after this open call completes. It is therefore safe to take a
2831 * new reference here without additional locking.
2832 */
4e35f483 2833 if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
f522c3ac 2834 kref_get(&resv->refs);
84afd99b
AW
2835}
2836
a1e78772
MG
2837static void hugetlb_vm_op_close(struct vm_area_struct *vma)
2838{
a5516438 2839 struct hstate *h = hstate_vma(vma);
f522c3ac 2840 struct resv_map *resv = vma_resv_map(vma);
90481622 2841 struct hugepage_subpool *spool = subpool_vma(vma);
4e35f483 2842 unsigned long reserve, start, end;
1c5ecae3 2843 long gbl_reserve;
84afd99b 2844
4e35f483
JK
2845 if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER))
2846 return;
84afd99b 2847
4e35f483
JK
2848 start = vma_hugecache_offset(h, vma, vma->vm_start);
2849 end = vma_hugecache_offset(h, vma, vma->vm_end);
84afd99b 2850
4e35f483 2851 reserve = (end - start) - region_count(resv, start, end);
84afd99b 2852
4e35f483
JK
2853 kref_put(&resv->refs, resv_map_release);
2854
2855 if (reserve) {
1c5ecae3
MK
2856 /*
2857 * Decrement reserve counts. The global reserve count may be
2858 * adjusted if the subpool has a minimum size.
2859 */
2860 gbl_reserve = hugepage_subpool_put_pages(spool, reserve);
2861 hugetlb_acct_memory(h, -gbl_reserve);
84afd99b 2862 }
a1e78772
MG
2863}
2864
1da177e4
LT
2865/*
2866 * We cannot handle pagefaults against hugetlb pages at all. They cause
2867 * handle_mm_fault() to try to instantiate regular-sized pages in the
2868 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
2869 * this far.
2870 */
d0217ac0 2871static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1da177e4
LT
2872{
2873 BUG();
d0217ac0 2874 return 0;
1da177e4
LT
2875}
2876
f0f37e2f 2877const struct vm_operations_struct hugetlb_vm_ops = {
d0217ac0 2878 .fault = hugetlb_vm_op_fault,
84afd99b 2879 .open = hugetlb_vm_op_open,
a1e78772 2880 .close = hugetlb_vm_op_close,
1da177e4
LT
2881};
2882
1e8f889b
DG
2883static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
2884 int writable)
63551ae0
DG
2885{
2886 pte_t entry;
2887
1e8f889b 2888 if (writable) {
106c992a
GS
2889 entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page,
2890 vma->vm_page_prot)));
63551ae0 2891 } else {
106c992a
GS
2892 entry = huge_pte_wrprotect(mk_huge_pte(page,
2893 vma->vm_page_prot));
63551ae0
DG
2894 }
2895 entry = pte_mkyoung(entry);
2896 entry = pte_mkhuge(entry);
d9ed9faa 2897 entry = arch_make_huge_pte(entry, vma, page, writable);
63551ae0
DG
2898
2899 return entry;
2900}
2901
1e8f889b
DG
2902static void set_huge_ptep_writable(struct vm_area_struct *vma,
2903 unsigned long address, pte_t *ptep)
2904{
2905 pte_t entry;
2906
106c992a 2907 entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep)));
32f84528 2908 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
4b3073e1 2909 update_mmu_cache(vma, address, ptep);
1e8f889b
DG
2910}
2911
4a705fef
NH
2912static int is_hugetlb_entry_migration(pte_t pte)
2913{
2914 swp_entry_t swp;
2915
2916 if (huge_pte_none(pte) || pte_present(pte))
2917 return 0;
2918 swp = pte_to_swp_entry(pte);
2919 if (non_swap_entry(swp) && is_migration_entry(swp))
2920 return 1;
2921 else
2922 return 0;
2923}
2924
2925static int is_hugetlb_entry_hwpoisoned(pte_t pte)
2926{
2927 swp_entry_t swp;
2928
2929 if (huge_pte_none(pte) || pte_present(pte))
2930 return 0;
2931 swp = pte_to_swp_entry(pte);
2932 if (non_swap_entry(swp) && is_hwpoison_entry(swp))
2933 return 1;
2934 else
2935 return 0;
2936}
1e8f889b 2937
63551ae0
DG
2938int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
2939 struct vm_area_struct *vma)
2940{
2941 pte_t *src_pte, *dst_pte, entry;
2942 struct page *ptepage;
1c59827d 2943 unsigned long addr;
1e8f889b 2944 int cow;
a5516438
AK
2945 struct hstate *h = hstate_vma(vma);
2946 unsigned long sz = huge_page_size(h);
e8569dd2
AS
2947 unsigned long mmun_start; /* For mmu_notifiers */
2948 unsigned long mmun_end; /* For mmu_notifiers */
2949 int ret = 0;
1e8f889b
DG
2950
2951 cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
63551ae0 2952
e8569dd2
AS
2953 mmun_start = vma->vm_start;
2954 mmun_end = vma->vm_end;
2955 if (cow)
2956 mmu_notifier_invalidate_range_start(src, mmun_start, mmun_end);
2957
a5516438 2958 for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
cb900f41 2959 spinlock_t *src_ptl, *dst_ptl;
c74df32c
HD
2960 src_pte = huge_pte_offset(src, addr);
2961 if (!src_pte)
2962 continue;
a5516438 2963 dst_pte = huge_pte_alloc(dst, addr, sz);
e8569dd2
AS
2964 if (!dst_pte) {
2965 ret = -ENOMEM;
2966 break;
2967 }
c5c99429
LW
2968
2969 /* If the pagetables are shared don't copy or take references */
2970 if (dst_pte == src_pte)
2971 continue;
2972
cb900f41
KS
2973 dst_ptl = huge_pte_lock(h, dst, dst_pte);
2974 src_ptl = huge_pte_lockptr(h, src, src_pte);
2975 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
4a705fef
NH
2976 entry = huge_ptep_get(src_pte);
2977 if (huge_pte_none(entry)) { /* skip none entry */
2978 ;
2979 } else if (unlikely(is_hugetlb_entry_migration(entry) ||
2980 is_hugetlb_entry_hwpoisoned(entry))) {
2981 swp_entry_t swp_entry = pte_to_swp_entry(entry);
2982
2983 if (is_write_migration_entry(swp_entry) && cow) {
2984 /*
2985 * COW mappings require pages in both
2986 * parent and child to be set to read.
2987 */
2988 make_migration_entry_read(&swp_entry);
2989 entry = swp_entry_to_pte(swp_entry);
2990 set_huge_pte_at(src, addr, src_pte, entry);
2991 }
2992 set_huge_pte_at(dst, addr, dst_pte, entry);
2993 } else {
34ee645e 2994 if (cow) {
7f2e9525 2995 huge_ptep_set_wrprotect(src, addr, src_pte);
34ee645e
JR
2996 mmu_notifier_invalidate_range(src, mmun_start,
2997 mmun_end);
2998 }
0253d634 2999 entry = huge_ptep_get(src_pte);
1c59827d
HD
3000 ptepage = pte_page(entry);
3001 get_page(ptepage);
0fe6e20b 3002 page_dup_rmap(ptepage);
1c59827d
HD
3003 set_huge_pte_at(dst, addr, dst_pte, entry);
3004 }
cb900f41
KS
3005 spin_unlock(src_ptl);
3006 spin_unlock(dst_ptl);
63551ae0 3007 }
63551ae0 3008
e8569dd2
AS
3009 if (cow)
3010 mmu_notifier_invalidate_range_end(src, mmun_start, mmun_end);
3011
3012 return ret;
63551ae0
DG
3013}
3014
24669e58
AK
3015void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
3016 unsigned long start, unsigned long end,
3017 struct page *ref_page)
63551ae0 3018{
24669e58 3019 int force_flush = 0;
63551ae0
DG
3020 struct mm_struct *mm = vma->vm_mm;
3021 unsigned long address;
c7546f8f 3022 pte_t *ptep;
63551ae0 3023 pte_t pte;
cb900f41 3024 spinlock_t *ptl;
63551ae0 3025 struct page *page;
a5516438
AK
3026 struct hstate *h = hstate_vma(vma);
3027 unsigned long sz = huge_page_size(h);
2ec74c3e
SG
3028 const unsigned long mmun_start = start; /* For mmu_notifiers */
3029 const unsigned long mmun_end = end; /* For mmu_notifiers */
a5516438 3030
63551ae0 3031 WARN_ON(!is_vm_hugetlb_page(vma));
a5516438
AK
3032 BUG_ON(start & ~huge_page_mask(h));
3033 BUG_ON(end & ~huge_page_mask(h));
63551ae0 3034
24669e58 3035 tlb_start_vma(tlb, vma);
2ec74c3e 3036 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
569f48b8 3037 address = start;
24669e58 3038again:
569f48b8 3039 for (; address < end; address += sz) {
c7546f8f 3040 ptep = huge_pte_offset(mm, address);
4c887265 3041 if (!ptep)
c7546f8f
DG
3042 continue;
3043
cb900f41 3044 ptl = huge_pte_lock(h, mm, ptep);
39dde65c 3045 if (huge_pmd_unshare(mm, &address, ptep))
cb900f41 3046 goto unlock;
39dde65c 3047
6629326b
HD
3048 pte = huge_ptep_get(ptep);
3049 if (huge_pte_none(pte))
cb900f41 3050 goto unlock;
6629326b
HD
3051
3052 /*
9fbc1f63
NH
3053 * Migrating hugepage or HWPoisoned hugepage is already
3054 * unmapped and its refcount is dropped, so just clear pte here.
6629326b 3055 */
9fbc1f63 3056 if (unlikely(!pte_present(pte))) {
106c992a 3057 huge_pte_clear(mm, address, ptep);
cb900f41 3058 goto unlock;
8c4894c6 3059 }
6629326b
HD
3060
3061 page = pte_page(pte);
04f2cbe3
MG
3062 /*
3063 * If a reference page is supplied, it is because a specific
3064 * page is being unmapped, not a range. Ensure the page we
3065 * are about to unmap is the actual page of interest.
3066 */
3067 if (ref_page) {
04f2cbe3 3068 if (page != ref_page)
cb900f41 3069 goto unlock;
04f2cbe3
MG
3070
3071 /*
3072 * Mark the VMA as having unmapped its page so that
3073 * future faults in this VMA will fail rather than
3074 * looking like data was lost
3075 */
3076 set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
3077 }
3078
c7546f8f 3079 pte = huge_ptep_get_and_clear(mm, address, ptep);
24669e58 3080 tlb_remove_tlb_entry(tlb, ptep, address);
106c992a 3081 if (huge_pte_dirty(pte))
6649a386 3082 set_page_dirty(page);
9e81130b 3083
24669e58
AK
3084 page_remove_rmap(page);
3085 force_flush = !__tlb_remove_page(tlb, page);
cb900f41 3086 if (force_flush) {
569f48b8 3087 address += sz;
cb900f41 3088 spin_unlock(ptl);
24669e58 3089 break;
cb900f41 3090 }
9e81130b 3091 /* Bail out after unmapping reference page if supplied */
cb900f41
KS
3092 if (ref_page) {
3093 spin_unlock(ptl);
9e81130b 3094 break;
cb900f41
KS
3095 }
3096unlock:
3097 spin_unlock(ptl);
63551ae0 3098 }
24669e58
AK
3099 /*
3100 * mmu_gather ran out of room to batch pages, we break out of
3101 * the PTE lock to avoid doing the potential expensive TLB invalidate
3102 * and page-free while holding it.
3103 */
3104 if (force_flush) {
3105 force_flush = 0;
3106 tlb_flush_mmu(tlb);
3107 if (address < end && !ref_page)
3108 goto again;
fe1668ae 3109 }
2ec74c3e 3110 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
24669e58 3111 tlb_end_vma(tlb, vma);
1da177e4 3112}
63551ae0 3113
d833352a
MG
3114void __unmap_hugepage_range_final(struct mmu_gather *tlb,
3115 struct vm_area_struct *vma, unsigned long start,
3116 unsigned long end, struct page *ref_page)
3117{
3118 __unmap_hugepage_range(tlb, vma, start, end, ref_page);
3119
3120 /*
3121 * Clear this flag so that x86's huge_pmd_share page_table_shareable
3122 * test will fail on a vma being torn down, and not grab a page table
3123 * on its way out. We're lucky that the flag has such an appropriate
3124 * name, and can in fact be safely cleared here. We could clear it
3125 * before the __unmap_hugepage_range above, but all that's necessary
c8c06efa 3126 * is to clear it before releasing the i_mmap_rwsem. This works
d833352a 3127 * because in the context this is called, the VMA is about to be
c8c06efa 3128 * destroyed and the i_mmap_rwsem is held.
d833352a
MG
3129 */
3130 vma->vm_flags &= ~VM_MAYSHARE;
3131}
3132
502717f4 3133void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
04f2cbe3 3134 unsigned long end, struct page *ref_page)
502717f4 3135{
24669e58
AK
3136 struct mm_struct *mm;
3137 struct mmu_gather tlb;
3138
3139 mm = vma->vm_mm;
3140
2b047252 3141 tlb_gather_mmu(&tlb, mm, start, end);
24669e58
AK
3142 __unmap_hugepage_range(&tlb, vma, start, end, ref_page);
3143 tlb_finish_mmu(&tlb, start, end);
502717f4
CK
3144}
3145
04f2cbe3
MG
3146/*
3147 * This is called when the original mapper is failing to COW a MAP_PRIVATE
3148 * mappping it owns the reserve page for. The intention is to unmap the page
3149 * from other VMAs and let the children be SIGKILLed if they are faulting the
3150 * same region.
3151 */
2f4612af
DB
3152static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
3153 struct page *page, unsigned long address)
04f2cbe3 3154{
7526674d 3155 struct hstate *h = hstate_vma(vma);
04f2cbe3
MG
3156 struct vm_area_struct *iter_vma;
3157 struct address_space *mapping;
04f2cbe3
MG
3158 pgoff_t pgoff;
3159
3160 /*
3161 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
3162 * from page cache lookup which is in HPAGE_SIZE units.
3163 */
7526674d 3164 address = address & huge_page_mask(h);
36e4f20a
MH
3165 pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) +
3166 vma->vm_pgoff;
496ad9aa 3167 mapping = file_inode(vma->vm_file)->i_mapping;
04f2cbe3 3168
4eb2b1dc
MG
3169 /*
3170 * Take the mapping lock for the duration of the table walk. As
3171 * this mapping should be shared between all the VMAs,
3172 * __unmap_hugepage_range() is called as the lock is already held
3173 */
83cde9e8 3174 i_mmap_lock_write(mapping);
6b2dbba8 3175 vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) {
04f2cbe3
MG
3176 /* Do not unmap the current VMA */
3177 if (iter_vma == vma)
3178 continue;
3179
3180 /*
3181 * Unmap the page from other VMAs without their own reserves.
3182 * They get marked to be SIGKILLed if they fault in these
3183 * areas. This is because a future no-page fault on this VMA
3184 * could insert a zeroed page instead of the data existing
3185 * from the time of fork. This would look like data corruption
3186 */
3187 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
24669e58
AK
3188 unmap_hugepage_range(iter_vma, address,
3189 address + huge_page_size(h), page);
04f2cbe3 3190 }
83cde9e8 3191 i_mmap_unlock_write(mapping);
04f2cbe3
MG
3192}
3193
0fe6e20b
NH
3194/*
3195 * Hugetlb_cow() should be called with page lock of the original hugepage held.
ef009b25
MH
3196 * Called with hugetlb_instantiation_mutex held and pte_page locked so we
3197 * cannot race with other handlers or page migration.
3198 * Keep the pte_same checks anyway to make transition from the mutex easier.
0fe6e20b 3199 */
1e8f889b 3200static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
04f2cbe3 3201 unsigned long address, pte_t *ptep, pte_t pte,
cb900f41 3202 struct page *pagecache_page, spinlock_t *ptl)
1e8f889b 3203{
a5516438 3204 struct hstate *h = hstate_vma(vma);
1e8f889b 3205 struct page *old_page, *new_page;
ad4404a2 3206 int ret = 0, outside_reserve = 0;
2ec74c3e
SG
3207 unsigned long mmun_start; /* For mmu_notifiers */
3208 unsigned long mmun_end; /* For mmu_notifiers */
1e8f889b
DG
3209
3210 old_page = pte_page(pte);
3211
04f2cbe3 3212retry_avoidcopy:
1e8f889b
DG
3213 /* If no-one else is actually using this page, avoid the copy
3214 * and just make the page writable */
37a2140d
JK
3215 if (page_mapcount(old_page) == 1 && PageAnon(old_page)) {
3216 page_move_anon_rmap(old_page, vma, address);
1e8f889b 3217 set_huge_ptep_writable(vma, address, ptep);
83c54070 3218 return 0;
1e8f889b
DG
3219 }
3220
04f2cbe3
MG
3221 /*
3222 * If the process that created a MAP_PRIVATE mapping is about to
3223 * perform a COW due to a shared page count, attempt to satisfy
3224 * the allocation without using the existing reserves. The pagecache
3225 * page is used to determine if the reserve at this address was
3226 * consumed or not. If reserves were used, a partial faulted mapping
3227 * at the time of fork() could consume its reserves on COW instead
3228 * of the full address range.
3229 */
5944d011 3230 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
04f2cbe3
MG
3231 old_page != pagecache_page)
3232 outside_reserve = 1;
3233
1e8f889b 3234 page_cache_get(old_page);
b76c8cfb 3235
ad4404a2
DB
3236 /*
3237 * Drop page table lock as buddy allocator may be called. It will
3238 * be acquired again before returning to the caller, as expected.
3239 */
cb900f41 3240 spin_unlock(ptl);
04f2cbe3 3241 new_page = alloc_huge_page(vma, address, outside_reserve);
1e8f889b 3242
2fc39cec 3243 if (IS_ERR(new_page)) {
04f2cbe3
MG
3244 /*
3245 * If a process owning a MAP_PRIVATE mapping fails to COW,
3246 * it is due to references held by a child and an insufficient
3247 * huge page pool. To guarantee the original mappers
3248 * reliability, unmap the page from child processes. The child
3249 * may get SIGKILLed if it later faults.
3250 */
3251 if (outside_reserve) {
ad4404a2 3252 page_cache_release(old_page);
04f2cbe3 3253 BUG_ON(huge_pte_none(pte));
2f4612af
DB
3254 unmap_ref_private(mm, vma, old_page, address);
3255 BUG_ON(huge_pte_none(pte));
3256 spin_lock(ptl);
3257 ptep = huge_pte_offset(mm, address & huge_page_mask(h));
3258 if (likely(ptep &&
3259 pte_same(huge_ptep_get(ptep), pte)))
3260 goto retry_avoidcopy;
3261 /*
3262 * race occurs while re-acquiring page table
3263 * lock, and our job is done.
3264 */
3265 return 0;
04f2cbe3
MG
3266 }
3267
ad4404a2
DB
3268 ret = (PTR_ERR(new_page) == -ENOMEM) ?
3269 VM_FAULT_OOM : VM_FAULT_SIGBUS;
3270 goto out_release_old;
1e8f889b
DG
3271 }
3272
0fe6e20b
NH
3273 /*
3274 * When the original hugepage is shared one, it does not have
3275 * anon_vma prepared.
3276 */
44e2aa93 3277 if (unlikely(anon_vma_prepare(vma))) {
ad4404a2
DB
3278 ret = VM_FAULT_OOM;
3279 goto out_release_all;
44e2aa93 3280 }
0fe6e20b 3281
47ad8475
AA
3282 copy_user_huge_page(new_page, old_page, address, vma,
3283 pages_per_huge_page(h));
0ed361de 3284 __SetPageUptodate(new_page);
bcc54222 3285 set_page_huge_active(new_page);
1e8f889b 3286
2ec74c3e
SG
3287 mmun_start = address & huge_page_mask(h);
3288 mmun_end = mmun_start + huge_page_size(h);
3289 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
ad4404a2 3290
b76c8cfb 3291 /*
cb900f41 3292 * Retake the page table lock to check for racing updates
b76c8cfb
LW
3293 * before the page tables are altered
3294 */
cb900f41 3295 spin_lock(ptl);
a5516438 3296 ptep = huge_pte_offset(mm, address & huge_page_mask(h));
a9af0c5d 3297 if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) {
07443a85
JK
3298 ClearPagePrivate(new_page);
3299
1e8f889b 3300 /* Break COW */
8fe627ec 3301 huge_ptep_clear_flush(vma, address, ptep);
34ee645e 3302 mmu_notifier_invalidate_range(mm, mmun_start, mmun_end);
1e8f889b
DG
3303 set_huge_pte_at(mm, address, ptep,
3304 make_huge_pte(vma, new_page, 1));
0fe6e20b 3305 page_remove_rmap(old_page);
cd67f0d2 3306 hugepage_add_new_anon_rmap(new_page, vma, address);
1e8f889b
DG
3307 /* Make the old page be freed below */
3308 new_page = old_page;
3309 }
cb900f41 3310 spin_unlock(ptl);
2ec74c3e 3311 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
ad4404a2 3312out_release_all:
1e8f889b 3313 page_cache_release(new_page);
ad4404a2 3314out_release_old:
1e8f889b 3315 page_cache_release(old_page);
8312034f 3316
ad4404a2
DB
3317 spin_lock(ptl); /* Caller expects lock to be held */
3318 return ret;
1e8f889b
DG
3319}
3320
04f2cbe3 3321/* Return the pagecache page at a given address within a VMA */
a5516438
AK
3322static struct page *hugetlbfs_pagecache_page(struct hstate *h,
3323 struct vm_area_struct *vma, unsigned long address)
04f2cbe3
MG
3324{
3325 struct address_space *mapping;
e7c4b0bf 3326 pgoff_t idx;
04f2cbe3
MG
3327
3328 mapping = vma->vm_file->f_mapping;
a5516438 3329 idx = vma_hugecache_offset(h, vma, address);
04f2cbe3
MG
3330
3331 return find_lock_page(mapping, idx);
3332}
3333
3ae77f43
HD
3334/*
3335 * Return whether there is a pagecache page to back given address within VMA.
3336 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
3337 */
3338static bool hugetlbfs_pagecache_present(struct hstate *h,
2a15efc9
HD
3339 struct vm_area_struct *vma, unsigned long address)
3340{
3341 struct address_space *mapping;
3342 pgoff_t idx;
3343 struct page *page;
3344
3345 mapping = vma->vm_file->f_mapping;
3346 idx = vma_hugecache_offset(h, vma, address);
3347
3348 page = find_get_page(mapping, idx);
3349 if (page)
3350 put_page(page);
3351 return page != NULL;
3352}
3353
a1ed3dda 3354static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
8382d914
DB
3355 struct address_space *mapping, pgoff_t idx,
3356 unsigned long address, pte_t *ptep, unsigned int flags)
ac9b9c66 3357{
a5516438 3358 struct hstate *h = hstate_vma(vma);
ac9b9c66 3359 int ret = VM_FAULT_SIGBUS;
409eb8c2 3360 int anon_rmap = 0;
4c887265 3361 unsigned long size;
4c887265 3362 struct page *page;
1e8f889b 3363 pte_t new_pte;
cb900f41 3364 spinlock_t *ptl;
4c887265 3365
04f2cbe3
MG
3366 /*
3367 * Currently, we are forced to kill the process in the event the
3368 * original mapper has unmapped pages from the child due to a failed
25985edc 3369 * COW. Warn that such a situation has occurred as it may not be obvious
04f2cbe3
MG
3370 */
3371 if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
ffb22af5
AM
3372 pr_warning("PID %d killed due to inadequate hugepage pool\n",
3373 current->pid);
04f2cbe3
MG
3374 return ret;
3375 }
3376
4c887265
AL
3377 /*
3378 * Use page lock to guard against racing truncation
3379 * before we get page_table_lock.
3380 */
6bda666a
CL
3381retry:
3382 page = find_lock_page(mapping, idx);
3383 if (!page) {
a5516438 3384 size = i_size_read(mapping->host) >> huge_page_shift(h);
ebed4bfc
HD
3385 if (idx >= size)
3386 goto out;
04f2cbe3 3387 page = alloc_huge_page(vma, address, 0);
2fc39cec 3388 if (IS_ERR(page)) {
76dcee75
AK
3389 ret = PTR_ERR(page);
3390 if (ret == -ENOMEM)
3391 ret = VM_FAULT_OOM;
3392 else
3393 ret = VM_FAULT_SIGBUS;
6bda666a
CL
3394 goto out;
3395 }
47ad8475 3396 clear_huge_page(page, address, pages_per_huge_page(h));
0ed361de 3397 __SetPageUptodate(page);
bcc54222 3398 set_page_huge_active(page);
ac9b9c66 3399
f83a275d 3400 if (vma->vm_flags & VM_MAYSHARE) {
6bda666a 3401 int err;
45c682a6 3402 struct inode *inode = mapping->host;
6bda666a
CL
3403
3404 err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
3405 if (err) {
3406 put_page(page);
6bda666a
CL
3407 if (err == -EEXIST)
3408 goto retry;
3409 goto out;
3410 }
07443a85 3411 ClearPagePrivate(page);
45c682a6
KC
3412
3413 spin_lock(&inode->i_lock);
a5516438 3414 inode->i_blocks += blocks_per_huge_page(h);
45c682a6 3415 spin_unlock(&inode->i_lock);
23be7468 3416 } else {
6bda666a 3417 lock_page(page);
0fe6e20b
NH
3418 if (unlikely(anon_vma_prepare(vma))) {
3419 ret = VM_FAULT_OOM;
3420 goto backout_unlocked;
3421 }
409eb8c2 3422 anon_rmap = 1;
23be7468 3423 }
0fe6e20b 3424 } else {
998b4382
NH
3425 /*
3426 * If memory error occurs between mmap() and fault, some process
3427 * don't have hwpoisoned swap entry for errored virtual address.
3428 * So we need to block hugepage fault by PG_hwpoison bit check.
3429 */
3430 if (unlikely(PageHWPoison(page))) {
32f84528 3431 ret = VM_FAULT_HWPOISON |
972dc4de 3432 VM_FAULT_SET_HINDEX(hstate_index(h));
998b4382
NH
3433 goto backout_unlocked;
3434 }
6bda666a 3435 }
1e8f889b 3436
57303d80
AW
3437 /*
3438 * If we are going to COW a private mapping later, we examine the
3439 * pending reservations for this page now. This will ensure that
3440 * any allocations necessary to record that reservation occur outside
3441 * the spinlock.
3442 */
5e911373 3443 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
2b26736c
AW
3444 if (vma_needs_reservation(h, vma, address) < 0) {
3445 ret = VM_FAULT_OOM;
3446 goto backout_unlocked;
3447 }
5e911373 3448 /* Just decrements count, does not deallocate */
feba16e2 3449 vma_end_reservation(h, vma, address);
5e911373 3450 }
57303d80 3451
cb900f41
KS
3452 ptl = huge_pte_lockptr(h, mm, ptep);
3453 spin_lock(ptl);
a5516438 3454 size = i_size_read(mapping->host) >> huge_page_shift(h);
4c887265
AL
3455 if (idx >= size)
3456 goto backout;
3457
83c54070 3458 ret = 0;
7f2e9525 3459 if (!huge_pte_none(huge_ptep_get(ptep)))
4c887265
AL
3460 goto backout;
3461
07443a85
JK
3462 if (anon_rmap) {
3463 ClearPagePrivate(page);
409eb8c2 3464 hugepage_add_new_anon_rmap(page, vma, address);
ac714904 3465 } else
409eb8c2 3466 page_dup_rmap(page);
1e8f889b
DG
3467 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
3468 && (vma->vm_flags & VM_SHARED)));
3469 set_huge_pte_at(mm, address, ptep, new_pte);
3470
788c7df4 3471 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
1e8f889b 3472 /* Optimization, do the COW without a second fault */
cb900f41 3473 ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page, ptl);
1e8f889b
DG
3474 }
3475
cb900f41 3476 spin_unlock(ptl);
4c887265
AL
3477 unlock_page(page);
3478out:
ac9b9c66 3479 return ret;
4c887265
AL
3480
3481backout:
cb900f41 3482 spin_unlock(ptl);
2b26736c 3483backout_unlocked:
4c887265
AL
3484 unlock_page(page);
3485 put_page(page);
3486 goto out;
ac9b9c66
HD
3487}
3488
8382d914 3489#ifdef CONFIG_SMP
c672c7f2 3490u32 hugetlb_fault_mutex_hash(struct hstate *h, struct mm_struct *mm,
8382d914
DB
3491 struct vm_area_struct *vma,
3492 struct address_space *mapping,
3493 pgoff_t idx, unsigned long address)
3494{
3495 unsigned long key[2];
3496 u32 hash;
3497
3498 if (vma->vm_flags & VM_SHARED) {
3499 key[0] = (unsigned long) mapping;
3500 key[1] = idx;
3501 } else {
3502 key[0] = (unsigned long) mm;
3503 key[1] = address >> huge_page_shift(h);
3504 }
3505
3506 hash = jhash2((u32 *)&key, sizeof(key)/sizeof(u32), 0);
3507
3508 return hash & (num_fault_mutexes - 1);
3509}
3510#else
3511/*
3512 * For uniprocesor systems we always use a single mutex, so just
3513 * return 0 and avoid the hashing overhead.
3514 */
c672c7f2 3515u32 hugetlb_fault_mutex_hash(struct hstate *h, struct mm_struct *mm,
8382d914
DB
3516 struct vm_area_struct *vma,
3517 struct address_space *mapping,
3518 pgoff_t idx, unsigned long address)
3519{
3520 return 0;
3521}
3522#endif
3523
86e5216f 3524int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
788c7df4 3525 unsigned long address, unsigned int flags)
86e5216f 3526{
8382d914 3527 pte_t *ptep, entry;
cb900f41 3528 spinlock_t *ptl;
1e8f889b 3529 int ret;
8382d914
DB
3530 u32 hash;
3531 pgoff_t idx;
0fe6e20b 3532 struct page *page = NULL;
57303d80 3533 struct page *pagecache_page = NULL;
a5516438 3534 struct hstate *h = hstate_vma(vma);
8382d914 3535 struct address_space *mapping;
0f792cf9 3536 int need_wait_lock = 0;
86e5216f 3537
1e16a539
KH
3538 address &= huge_page_mask(h);
3539
fd6a03ed
NH
3540 ptep = huge_pte_offset(mm, address);
3541 if (ptep) {
3542 entry = huge_ptep_get(ptep);
290408d4 3543 if (unlikely(is_hugetlb_entry_migration(entry))) {
cb900f41 3544 migration_entry_wait_huge(vma, mm, ptep);
290408d4
NH
3545 return 0;
3546 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
32f84528 3547 return VM_FAULT_HWPOISON_LARGE |
972dc4de 3548 VM_FAULT_SET_HINDEX(hstate_index(h));
fd6a03ed
NH
3549 }
3550
a5516438 3551 ptep = huge_pte_alloc(mm, address, huge_page_size(h));
86e5216f
AL
3552 if (!ptep)
3553 return VM_FAULT_OOM;
3554
8382d914
DB
3555 mapping = vma->vm_file->f_mapping;
3556 idx = vma_hugecache_offset(h, vma, address);
3557
3935baa9
DG
3558 /*
3559 * Serialize hugepage allocation and instantiation, so that we don't
3560 * get spurious allocation failures if two CPUs race to instantiate
3561 * the same page in the page cache.
3562 */
c672c7f2
MK
3563 hash = hugetlb_fault_mutex_hash(h, mm, vma, mapping, idx, address);
3564 mutex_lock(&hugetlb_fault_mutex_table[hash]);
8382d914 3565
7f2e9525
GS
3566 entry = huge_ptep_get(ptep);
3567 if (huge_pte_none(entry)) {
8382d914 3568 ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags);
b4d1d99f 3569 goto out_mutex;
3935baa9 3570 }
86e5216f 3571
83c54070 3572 ret = 0;
1e8f889b 3573
0f792cf9
NH
3574 /*
3575 * entry could be a migration/hwpoison entry at this point, so this
3576 * check prevents the kernel from going below assuming that we have
3577 * a active hugepage in pagecache. This goto expects the 2nd page fault,
3578 * and is_hugetlb_entry_(migration|hwpoisoned) check will properly
3579 * handle it.
3580 */
3581 if (!pte_present(entry))
3582 goto out_mutex;
3583
57303d80
AW
3584 /*
3585 * If we are going to COW the mapping later, we examine the pending
3586 * reservations for this page now. This will ensure that any
3587 * allocations necessary to record that reservation occur outside the
3588 * spinlock. For private mappings, we also lookup the pagecache
3589 * page now as it is used to determine if a reservation has been
3590 * consumed.
3591 */
106c992a 3592 if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) {
2b26736c
AW
3593 if (vma_needs_reservation(h, vma, address) < 0) {
3594 ret = VM_FAULT_OOM;
b4d1d99f 3595 goto out_mutex;
2b26736c 3596 }
5e911373 3597 /* Just decrements count, does not deallocate */
feba16e2 3598 vma_end_reservation(h, vma, address);
57303d80 3599
f83a275d 3600 if (!(vma->vm_flags & VM_MAYSHARE))
57303d80
AW
3601 pagecache_page = hugetlbfs_pagecache_page(h,
3602 vma, address);
3603 }
3604
0f792cf9
NH
3605 ptl = huge_pte_lock(h, mm, ptep);
3606
3607 /* Check for a racing update before calling hugetlb_cow */
3608 if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
3609 goto out_ptl;
3610
56c9cfb1
NH
3611 /*
3612 * hugetlb_cow() requires page locks of pte_page(entry) and
3613 * pagecache_page, so here we need take the former one
3614 * when page != pagecache_page or !pagecache_page.
56c9cfb1
NH
3615 */
3616 page = pte_page(entry);
3617 if (page != pagecache_page)
0f792cf9
NH
3618 if (!trylock_page(page)) {
3619 need_wait_lock = 1;
3620 goto out_ptl;
3621 }
b4d1d99f 3622
0f792cf9 3623 get_page(page);
b4d1d99f 3624
788c7df4 3625 if (flags & FAULT_FLAG_WRITE) {
106c992a 3626 if (!huge_pte_write(entry)) {
57303d80 3627 ret = hugetlb_cow(mm, vma, address, ptep, entry,
cb900f41 3628 pagecache_page, ptl);
0f792cf9 3629 goto out_put_page;
b4d1d99f 3630 }
106c992a 3631 entry = huge_pte_mkdirty(entry);
b4d1d99f
DG
3632 }
3633 entry = pte_mkyoung(entry);
788c7df4
HD
3634 if (huge_ptep_set_access_flags(vma, address, ptep, entry,
3635 flags & FAULT_FLAG_WRITE))
4b3073e1 3636 update_mmu_cache(vma, address, ptep);
0f792cf9
NH
3637out_put_page:
3638 if (page != pagecache_page)
3639 unlock_page(page);
3640 put_page(page);
cb900f41
KS
3641out_ptl:
3642 spin_unlock(ptl);
57303d80
AW
3643
3644 if (pagecache_page) {
3645 unlock_page(pagecache_page);
3646 put_page(pagecache_page);
3647 }
b4d1d99f 3648out_mutex:
c672c7f2 3649 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
0f792cf9
NH
3650 /*
3651 * Generally it's safe to hold refcount during waiting page lock. But
3652 * here we just wait to defer the next page fault to avoid busy loop and
3653 * the page is not used after unlocked before returning from the current
3654 * page fault. So we are safe from accessing freed page, even if we wait
3655 * here without taking refcount.
3656 */
3657 if (need_wait_lock)
3658 wait_on_page_locked(page);
1e8f889b 3659 return ret;
86e5216f
AL
3660}
3661
28a35716
ML
3662long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
3663 struct page **pages, struct vm_area_struct **vmas,
3664 unsigned long *position, unsigned long *nr_pages,
3665 long i, unsigned int flags)
63551ae0 3666{
d5d4b0aa
CK
3667 unsigned long pfn_offset;
3668 unsigned long vaddr = *position;
28a35716 3669 unsigned long remainder = *nr_pages;
a5516438 3670 struct hstate *h = hstate_vma(vma);
63551ae0 3671
63551ae0 3672 while (vaddr < vma->vm_end && remainder) {
4c887265 3673 pte_t *pte;
cb900f41 3674 spinlock_t *ptl = NULL;
2a15efc9 3675 int absent;
4c887265 3676 struct page *page;
63551ae0 3677
02057967
DR
3678 /*
3679 * If we have a pending SIGKILL, don't keep faulting pages and
3680 * potentially allocating memory.
3681 */
3682 if (unlikely(fatal_signal_pending(current))) {
3683 remainder = 0;
3684 break;
3685 }
3686
4c887265
AL
3687 /*
3688 * Some archs (sparc64, sh*) have multiple pte_ts to
2a15efc9 3689 * each hugepage. We have to make sure we get the
4c887265 3690 * first, for the page indexing below to work.
cb900f41
KS
3691 *
3692 * Note that page table lock is not held when pte is null.
4c887265 3693 */
a5516438 3694 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h));
cb900f41
KS
3695 if (pte)
3696 ptl = huge_pte_lock(h, mm, pte);
2a15efc9
HD
3697 absent = !pte || huge_pte_none(huge_ptep_get(pte));
3698
3699 /*
3700 * When coredumping, it suits get_dump_page if we just return
3ae77f43
HD
3701 * an error where there's an empty slot with no huge pagecache
3702 * to back it. This way, we avoid allocating a hugepage, and
3703 * the sparse dumpfile avoids allocating disk blocks, but its
3704 * huge holes still show up with zeroes where they need to be.
2a15efc9 3705 */
3ae77f43
HD
3706 if (absent && (flags & FOLL_DUMP) &&
3707 !hugetlbfs_pagecache_present(h, vma, vaddr)) {
cb900f41
KS
3708 if (pte)
3709 spin_unlock(ptl);
2a15efc9
HD
3710 remainder = 0;
3711 break;
3712 }
63551ae0 3713
9cc3a5bd
NH
3714 /*
3715 * We need call hugetlb_fault for both hugepages under migration
3716 * (in which case hugetlb_fault waits for the migration,) and
3717 * hwpoisoned hugepages (in which case we need to prevent the
3718 * caller from accessing to them.) In order to do this, we use
3719 * here is_swap_pte instead of is_hugetlb_entry_migration and
3720 * is_hugetlb_entry_hwpoisoned. This is because it simply covers
3721 * both cases, and because we can't follow correct pages
3722 * directly from any kind of swap entries.
3723 */
3724 if (absent || is_swap_pte(huge_ptep_get(pte)) ||
106c992a
GS
3725 ((flags & FOLL_WRITE) &&
3726 !huge_pte_write(huge_ptep_get(pte)))) {
4c887265 3727 int ret;
63551ae0 3728
cb900f41
KS
3729 if (pte)
3730 spin_unlock(ptl);
2a15efc9
HD
3731 ret = hugetlb_fault(mm, vma, vaddr,
3732 (flags & FOLL_WRITE) ? FAULT_FLAG_WRITE : 0);
a89182c7 3733 if (!(ret & VM_FAULT_ERROR))
4c887265 3734 continue;
63551ae0 3735
4c887265 3736 remainder = 0;
4c887265
AL
3737 break;
3738 }
3739
a5516438 3740 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
7f2e9525 3741 page = pte_page(huge_ptep_get(pte));
d5d4b0aa 3742same_page:
d6692183 3743 if (pages) {
2a15efc9 3744 pages[i] = mem_map_offset(page, pfn_offset);
a0368d4e 3745 get_page_foll(pages[i]);
d6692183 3746 }
63551ae0
DG
3747
3748 if (vmas)
3749 vmas[i] = vma;
3750
3751 vaddr += PAGE_SIZE;
d5d4b0aa 3752 ++pfn_offset;
63551ae0
DG
3753 --remainder;
3754 ++i;
d5d4b0aa 3755 if (vaddr < vma->vm_end && remainder &&
a5516438 3756 pfn_offset < pages_per_huge_page(h)) {
d5d4b0aa
CK
3757 /*
3758 * We use pfn_offset to avoid touching the pageframes
3759 * of this compound page.
3760 */
3761 goto same_page;
3762 }
cb900f41 3763 spin_unlock(ptl);
63551ae0 3764 }
28a35716 3765 *nr_pages = remainder;
63551ae0
DG
3766 *position = vaddr;
3767
2a15efc9 3768 return i ? i : -EFAULT;
63551ae0 3769}
8f860591 3770
7da4d641 3771unsigned long hugetlb_change_protection(struct vm_area_struct *vma,
8f860591
ZY
3772 unsigned long address, unsigned long end, pgprot_t newprot)
3773{
3774 struct mm_struct *mm = vma->vm_mm;
3775 unsigned long start = address;
3776 pte_t *ptep;
3777 pte_t pte;
a5516438 3778 struct hstate *h = hstate_vma(vma);
7da4d641 3779 unsigned long pages = 0;
8f860591
ZY
3780
3781 BUG_ON(address >= end);
3782 flush_cache_range(vma, address, end);
3783
a5338093 3784 mmu_notifier_invalidate_range_start(mm, start, end);
83cde9e8 3785 i_mmap_lock_write(vma->vm_file->f_mapping);
a5516438 3786 for (; address < end; address += huge_page_size(h)) {
cb900f41 3787 spinlock_t *ptl;
8f860591
ZY
3788 ptep = huge_pte_offset(mm, address);
3789 if (!ptep)
3790 continue;
cb900f41 3791 ptl = huge_pte_lock(h, mm, ptep);
7da4d641
PZ
3792 if (huge_pmd_unshare(mm, &address, ptep)) {
3793 pages++;
cb900f41 3794 spin_unlock(ptl);
39dde65c 3795 continue;
7da4d641 3796 }
a8bda28d
NH
3797 pte = huge_ptep_get(ptep);
3798 if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) {
3799 spin_unlock(ptl);
3800 continue;
3801 }
3802 if (unlikely(is_hugetlb_entry_migration(pte))) {
3803 swp_entry_t entry = pte_to_swp_entry(pte);
3804
3805 if (is_write_migration_entry(entry)) {
3806 pte_t newpte;
3807
3808 make_migration_entry_read(&entry);
3809 newpte = swp_entry_to_pte(entry);
3810 set_huge_pte_at(mm, address, ptep, newpte);
3811 pages++;
3812 }
3813 spin_unlock(ptl);
3814 continue;
3815 }
3816 if (!huge_pte_none(pte)) {
8f860591 3817 pte = huge_ptep_get_and_clear(mm, address, ptep);
106c992a 3818 pte = pte_mkhuge(huge_pte_modify(pte, newprot));
be7517d6 3819 pte = arch_make_huge_pte(pte, vma, NULL, 0);
8f860591 3820 set_huge_pte_at(mm, address, ptep, pte);
7da4d641 3821 pages++;
8f860591 3822 }
cb900f41 3823 spin_unlock(ptl);
8f860591 3824 }
d833352a 3825 /*
c8c06efa 3826 * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare
d833352a 3827 * may have cleared our pud entry and done put_page on the page table:
c8c06efa 3828 * once we release i_mmap_rwsem, another task can do the final put_page
d833352a
MG
3829 * and that page table be reused and filled with junk.
3830 */
8f860591 3831 flush_tlb_range(vma, start, end);
34ee645e 3832 mmu_notifier_invalidate_range(mm, start, end);
83cde9e8 3833 i_mmap_unlock_write(vma->vm_file->f_mapping);
a5338093 3834 mmu_notifier_invalidate_range_end(mm, start, end);
7da4d641
PZ
3835
3836 return pages << h->order;
8f860591
ZY
3837}
3838
a1e78772
MG
3839int hugetlb_reserve_pages(struct inode *inode,
3840 long from, long to,
5a6fe125 3841 struct vm_area_struct *vma,
ca16d140 3842 vm_flags_t vm_flags)
e4e574b7 3843{
17c9d12e 3844 long ret, chg;
a5516438 3845 struct hstate *h = hstate_inode(inode);
90481622 3846 struct hugepage_subpool *spool = subpool_inode(inode);
9119a41e 3847 struct resv_map *resv_map;
1c5ecae3 3848 long gbl_reserve;
e4e574b7 3849
17c9d12e
MG
3850 /*
3851 * Only apply hugepage reservation if asked. At fault time, an
3852 * attempt will be made for VM_NORESERVE to allocate a page
90481622 3853 * without using reserves
17c9d12e 3854 */
ca16d140 3855 if (vm_flags & VM_NORESERVE)
17c9d12e
MG
3856 return 0;
3857
a1e78772
MG
3858 /*
3859 * Shared mappings base their reservation on the number of pages that
3860 * are already allocated on behalf of the file. Private mappings need
3861 * to reserve the full area even if read-only as mprotect() may be
3862 * called to make the mapping read-write. Assume !vma is a shm mapping
3863 */
9119a41e 3864 if (!vma || vma->vm_flags & VM_MAYSHARE) {
4e35f483 3865 resv_map = inode_resv_map(inode);
9119a41e 3866
1406ec9b 3867 chg = region_chg(resv_map, from, to);
9119a41e
JK
3868
3869 } else {
3870 resv_map = resv_map_alloc();
17c9d12e
MG
3871 if (!resv_map)
3872 return -ENOMEM;
3873
a1e78772 3874 chg = to - from;
84afd99b 3875
17c9d12e
MG
3876 set_vma_resv_map(vma, resv_map);
3877 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
3878 }
3879
c50ac050
DH
3880 if (chg < 0) {
3881 ret = chg;
3882 goto out_err;
3883 }
8a630112 3884
1c5ecae3
MK
3885 /*
3886 * There must be enough pages in the subpool for the mapping. If
3887 * the subpool has a minimum size, there may be some global
3888 * reservations already in place (gbl_reserve).
3889 */
3890 gbl_reserve = hugepage_subpool_get_pages(spool, chg);
3891 if (gbl_reserve < 0) {
c50ac050
DH
3892 ret = -ENOSPC;
3893 goto out_err;
3894 }
5a6fe125
MG
3895
3896 /*
17c9d12e 3897 * Check enough hugepages are available for the reservation.
90481622 3898 * Hand the pages back to the subpool if there are not
5a6fe125 3899 */
1c5ecae3 3900 ret = hugetlb_acct_memory(h, gbl_reserve);
68842c9b 3901 if (ret < 0) {
1c5ecae3
MK
3902 /* put back original number of pages, chg */
3903 (void)hugepage_subpool_put_pages(spool, chg);
c50ac050 3904 goto out_err;
68842c9b 3905 }
17c9d12e
MG
3906
3907 /*
3908 * Account for the reservations made. Shared mappings record regions
3909 * that have reservations as they are shared by multiple VMAs.
3910 * When the last VMA disappears, the region map says how much
3911 * the reservation was and the page cache tells how much of
3912 * the reservation was consumed. Private mappings are per-VMA and
3913 * only the consumed reservations are tracked. When the VMA
3914 * disappears, the original reservation is the VMA size and the
3915 * consumed reservations are stored in the map. Hence, nothing
3916 * else has to be done for private mappings here
3917 */
33039678
MK
3918 if (!vma || vma->vm_flags & VM_MAYSHARE) {
3919 long add = region_add(resv_map, from, to);
3920
3921 if (unlikely(chg > add)) {
3922 /*
3923 * pages in this range were added to the reserve
3924 * map between region_chg and region_add. This
3925 * indicates a race with alloc_huge_page. Adjust
3926 * the subpool and reserve counts modified above
3927 * based on the difference.
3928 */
3929 long rsv_adjust;
3930
3931 rsv_adjust = hugepage_subpool_put_pages(spool,
3932 chg - add);
3933 hugetlb_acct_memory(h, -rsv_adjust);
3934 }
3935 }
a43a8c39 3936 return 0;
c50ac050 3937out_err:
5e911373
MK
3938 if (!vma || vma->vm_flags & VM_MAYSHARE)
3939 region_abort(resv_map, from, to);
f031dd27
JK
3940 if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
3941 kref_put(&resv_map->refs, resv_map_release);
c50ac050 3942 return ret;
a43a8c39
CK
3943}
3944
b5cec28d
MK
3945long hugetlb_unreserve_pages(struct inode *inode, long start, long end,
3946 long freed)
a43a8c39 3947{
a5516438 3948 struct hstate *h = hstate_inode(inode);
4e35f483 3949 struct resv_map *resv_map = inode_resv_map(inode);
9119a41e 3950 long chg = 0;
90481622 3951 struct hugepage_subpool *spool = subpool_inode(inode);
1c5ecae3 3952 long gbl_reserve;
45c682a6 3953
b5cec28d
MK
3954 if (resv_map) {
3955 chg = region_del(resv_map, start, end);
3956 /*
3957 * region_del() can fail in the rare case where a region
3958 * must be split and another region descriptor can not be
3959 * allocated. If end == LONG_MAX, it will not fail.
3960 */
3961 if (chg < 0)
3962 return chg;
3963 }
3964
45c682a6 3965 spin_lock(&inode->i_lock);
e4c6f8be 3966 inode->i_blocks -= (blocks_per_huge_page(h) * freed);
45c682a6
KC
3967 spin_unlock(&inode->i_lock);
3968
1c5ecae3
MK
3969 /*
3970 * If the subpool has a minimum size, the number of global
3971 * reservations to be released may be adjusted.
3972 */
3973 gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed));
3974 hugetlb_acct_memory(h, -gbl_reserve);
b5cec28d
MK
3975
3976 return 0;
a43a8c39 3977}
93f70f90 3978
3212b535
SC
3979#ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
3980static unsigned long page_table_shareable(struct vm_area_struct *svma,
3981 struct vm_area_struct *vma,
3982 unsigned long addr, pgoff_t idx)
3983{
3984 unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) +
3985 svma->vm_start;
3986 unsigned long sbase = saddr & PUD_MASK;
3987 unsigned long s_end = sbase + PUD_SIZE;
3988
3989 /* Allow segments to share if only one is marked locked */
3990 unsigned long vm_flags = vma->vm_flags & ~VM_LOCKED;
3991 unsigned long svm_flags = svma->vm_flags & ~VM_LOCKED;
3992
3993 /*
3994 * match the virtual addresses, permission and the alignment of the
3995 * page table page.
3996 */
3997 if (pmd_index(addr) != pmd_index(saddr) ||
3998 vm_flags != svm_flags ||
3999 sbase < svma->vm_start || svma->vm_end < s_end)
4000 return 0;
4001
4002 return saddr;
4003}
4004
31aafb45 4005static bool vma_shareable(struct vm_area_struct *vma, unsigned long addr)
3212b535
SC
4006{
4007 unsigned long base = addr & PUD_MASK;
4008 unsigned long end = base + PUD_SIZE;
4009
4010 /*
4011 * check on proper vm_flags and page table alignment
4012 */
4013 if (vma->vm_flags & VM_MAYSHARE &&
4014 vma->vm_start <= base && end <= vma->vm_end)
31aafb45
NK
4015 return true;
4016 return false;
3212b535
SC
4017}
4018
4019/*
4020 * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
4021 * and returns the corresponding pte. While this is not necessary for the
4022 * !shared pmd case because we can allocate the pmd later as well, it makes the
4023 * code much cleaner. pmd allocation is essential for the shared case because
c8c06efa 4024 * pud has to be populated inside the same i_mmap_rwsem section - otherwise
3212b535
SC
4025 * racing tasks could either miss the sharing (see huge_pte_offset) or select a
4026 * bad pmd for sharing.
4027 */
4028pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
4029{
4030 struct vm_area_struct *vma = find_vma(mm, addr);
4031 struct address_space *mapping = vma->vm_file->f_mapping;
4032 pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) +
4033 vma->vm_pgoff;
4034 struct vm_area_struct *svma;
4035 unsigned long saddr;
4036 pte_t *spte = NULL;
4037 pte_t *pte;
cb900f41 4038 spinlock_t *ptl;
3212b535
SC
4039
4040 if (!vma_shareable(vma, addr))
4041 return (pte_t *)pmd_alloc(mm, pud, addr);
4042
83cde9e8 4043 i_mmap_lock_write(mapping);
3212b535
SC
4044 vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) {
4045 if (svma == vma)
4046 continue;
4047
4048 saddr = page_table_shareable(svma, vma, addr, idx);
4049 if (saddr) {
4050 spte = huge_pte_offset(svma->vm_mm, saddr);
4051 if (spte) {
dc6c9a35 4052 mm_inc_nr_pmds(mm);
3212b535
SC
4053 get_page(virt_to_page(spte));
4054 break;
4055 }
4056 }
4057 }
4058
4059 if (!spte)
4060 goto out;
4061
cb900f41
KS
4062 ptl = huge_pte_lockptr(hstate_vma(vma), mm, spte);
4063 spin_lock(ptl);
dc6c9a35 4064 if (pud_none(*pud)) {
3212b535
SC
4065 pud_populate(mm, pud,
4066 (pmd_t *)((unsigned long)spte & PAGE_MASK));
dc6c9a35 4067 } else {
3212b535 4068 put_page(virt_to_page(spte));
dc6c9a35
KS
4069 mm_inc_nr_pmds(mm);
4070 }
cb900f41 4071 spin_unlock(ptl);
3212b535
SC
4072out:
4073 pte = (pte_t *)pmd_alloc(mm, pud, addr);
83cde9e8 4074 i_mmap_unlock_write(mapping);
3212b535
SC
4075 return pte;
4076}
4077
4078/*
4079 * unmap huge page backed by shared pte.
4080 *
4081 * Hugetlb pte page is ref counted at the time of mapping. If pte is shared
4082 * indicated by page_count > 1, unmap is achieved by clearing pud and
4083 * decrementing the ref count. If count == 1, the pte page is not shared.
4084 *
cb900f41 4085 * called with page table lock held.
3212b535
SC
4086 *
4087 * returns: 1 successfully unmapped a shared pte page
4088 * 0 the underlying pte page is not shared, or it is the last user
4089 */
4090int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
4091{
4092 pgd_t *pgd = pgd_offset(mm, *addr);
4093 pud_t *pud = pud_offset(pgd, *addr);
4094
4095 BUG_ON(page_count(virt_to_page(ptep)) == 0);
4096 if (page_count(virt_to_page(ptep)) == 1)
4097 return 0;
4098
4099 pud_clear(pud);
4100 put_page(virt_to_page(ptep));
dc6c9a35 4101 mm_dec_nr_pmds(mm);
3212b535
SC
4102 *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE;
4103 return 1;
4104}
9e5fc74c
SC
4105#define want_pmd_share() (1)
4106#else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
4107pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
4108{
4109 return NULL;
4110}
e81f2d22
ZZ
4111
4112int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
4113{
4114 return 0;
4115}
9e5fc74c 4116#define want_pmd_share() (0)
3212b535
SC
4117#endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
4118
9e5fc74c
SC
4119#ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
4120pte_t *huge_pte_alloc(struct mm_struct *mm,
4121 unsigned long addr, unsigned long sz)
4122{
4123 pgd_t *pgd;
4124 pud_t *pud;
4125 pte_t *pte = NULL;
4126
4127 pgd = pgd_offset(mm, addr);
4128 pud = pud_alloc(mm, pgd, addr);
4129 if (pud) {
4130 if (sz == PUD_SIZE) {
4131 pte = (pte_t *)pud;
4132 } else {
4133 BUG_ON(sz != PMD_SIZE);
4134 if (want_pmd_share() && pud_none(*pud))
4135 pte = huge_pmd_share(mm, addr, pud);
4136 else
4137 pte = (pte_t *)pmd_alloc(mm, pud, addr);
4138 }
4139 }
4140 BUG_ON(pte && !pte_none(*pte) && !pte_huge(*pte));
4141
4142 return pte;
4143}
4144
4145pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr)
4146{
4147 pgd_t *pgd;
4148 pud_t *pud;
4149 pmd_t *pmd = NULL;
4150
4151 pgd = pgd_offset(mm, addr);
4152 if (pgd_present(*pgd)) {
4153 pud = pud_offset(pgd, addr);
4154 if (pud_present(*pud)) {
4155 if (pud_huge(*pud))
4156 return (pte_t *)pud;
4157 pmd = pmd_offset(pud, addr);
4158 }
4159 }
4160 return (pte_t *) pmd;
4161}
4162
61f77eda
NH
4163#endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
4164
4165/*
4166 * These functions are overwritable if your architecture needs its own
4167 * behavior.
4168 */
4169struct page * __weak
4170follow_huge_addr(struct mm_struct *mm, unsigned long address,
4171 int write)
4172{
4173 return ERR_PTR(-EINVAL);
4174}
4175
4176struct page * __weak
9e5fc74c 4177follow_huge_pmd(struct mm_struct *mm, unsigned long address,
e66f17ff 4178 pmd_t *pmd, int flags)
9e5fc74c 4179{
e66f17ff
NH
4180 struct page *page = NULL;
4181 spinlock_t *ptl;
4182retry:
4183 ptl = pmd_lockptr(mm, pmd);
4184 spin_lock(ptl);
4185 /*
4186 * make sure that the address range covered by this pmd is not
4187 * unmapped from other threads.
4188 */
4189 if (!pmd_huge(*pmd))
4190 goto out;
4191 if (pmd_present(*pmd)) {
97534127 4192 page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT);
e66f17ff
NH
4193 if (flags & FOLL_GET)
4194 get_page(page);
4195 } else {
4196 if (is_hugetlb_entry_migration(huge_ptep_get((pte_t *)pmd))) {
4197 spin_unlock(ptl);
4198 __migration_entry_wait(mm, (pte_t *)pmd, ptl);
4199 goto retry;
4200 }
4201 /*
4202 * hwpoisoned entry is treated as no_page_table in
4203 * follow_page_mask().
4204 */
4205 }
4206out:
4207 spin_unlock(ptl);
9e5fc74c
SC
4208 return page;
4209}
4210
61f77eda 4211struct page * __weak
9e5fc74c 4212follow_huge_pud(struct mm_struct *mm, unsigned long address,
e66f17ff 4213 pud_t *pud, int flags)
9e5fc74c 4214{
e66f17ff
NH
4215 if (flags & FOLL_GET)
4216 return NULL;
9e5fc74c 4217
e66f17ff 4218 return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT);
9e5fc74c
SC
4219}
4220
d5bd9106
AK
4221#ifdef CONFIG_MEMORY_FAILURE
4222
93f70f90
NH
4223/*
4224 * This function is called from memory failure code.
4225 * Assume the caller holds page lock of the head page.
4226 */
6de2b1aa 4227int dequeue_hwpoisoned_huge_page(struct page *hpage)
93f70f90
NH
4228{
4229 struct hstate *h = page_hstate(hpage);
4230 int nid = page_to_nid(hpage);
6de2b1aa 4231 int ret = -EBUSY;
93f70f90
NH
4232
4233 spin_lock(&hugetlb_lock);
7e1f049e
NH
4234 /*
4235 * Just checking !page_huge_active is not enough, because that could be
4236 * an isolated/hwpoisoned hugepage (which have >0 refcount).
4237 */
4238 if (!page_huge_active(hpage) && !page_count(hpage)) {
56f2fb14
NH
4239 /*
4240 * Hwpoisoned hugepage isn't linked to activelist or freelist,
4241 * but dangling hpage->lru can trigger list-debug warnings
4242 * (this happens when we call unpoison_memory() on it),
4243 * so let it point to itself with list_del_init().
4244 */
4245 list_del_init(&hpage->lru);
8c6c2ecb 4246 set_page_refcounted(hpage);
6de2b1aa
NH
4247 h->free_huge_pages--;
4248 h->free_huge_pages_node[nid]--;
4249 ret = 0;
4250 }
93f70f90 4251 spin_unlock(&hugetlb_lock);
6de2b1aa 4252 return ret;
93f70f90 4253}
6de2b1aa 4254#endif
31caf665
NH
4255
4256bool isolate_huge_page(struct page *page, struct list_head *list)
4257{
bcc54222
NH
4258 bool ret = true;
4259
309381fe 4260 VM_BUG_ON_PAGE(!PageHead(page), page);
31caf665 4261 spin_lock(&hugetlb_lock);
bcc54222
NH
4262 if (!page_huge_active(page) || !get_page_unless_zero(page)) {
4263 ret = false;
4264 goto unlock;
4265 }
4266 clear_page_huge_active(page);
31caf665 4267 list_move_tail(&page->lru, list);
bcc54222 4268unlock:
31caf665 4269 spin_unlock(&hugetlb_lock);
bcc54222 4270 return ret;
31caf665
NH
4271}
4272
4273void putback_active_hugepage(struct page *page)
4274{
309381fe 4275 VM_BUG_ON_PAGE(!PageHead(page), page);
31caf665 4276 spin_lock(&hugetlb_lock);
bcc54222 4277 set_page_huge_active(page);
31caf665
NH
4278 list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist);
4279 spin_unlock(&hugetlb_lock);
4280 put_page(page);
4281}