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