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/memblock.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#include <linux/numa.h>
29
30#include <asm/page.h>
31#include <asm/pgtable.h>
32#include <asm/tlb.h>
33
34#include <linux/io.h>
35#include <linux/hugetlb.h>
36#include <linux/hugetlb_cgroup.h>
37#include <linux/node.h>
38#include <linux/userfaultfd_k.h>
39#include <linux/page_owner.h>
40#include "internal.h"
41
42int hugetlb_max_hstate __read_mostly;
43unsigned int default_hstate_idx;
44struct hstate hstates[HUGE_MAX_HSTATE];
45/*
46 * Minimum page order among possible hugepage sizes, set to a proper value
47 * at boot time.
48 */
49static unsigned int minimum_order __read_mostly = UINT_MAX;
50
51__initdata LIST_HEAD(huge_boot_pages);
52
53/* for command line parsing */
54static struct hstate * __initdata parsed_hstate;
55static unsigned long __initdata default_hstate_max_huge_pages;
56static unsigned long __initdata default_hstate_size;
57static bool __initdata parsed_valid_hugepagesz = true;
58
59/*
60 * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages,
61 * free_huge_pages, and surplus_huge_pages.
62 */
63DEFINE_SPINLOCK(hugetlb_lock);
64
65/*
66 * Serializes faults on the same logical page. This is used to
67 * prevent spurious OOMs when the hugepage pool is fully utilized.
68 */
69static int num_fault_mutexes;
70struct mutex *hugetlb_fault_mutex_table ____cacheline_aligned_in_smp;
71
72/* Forward declaration */
73static int hugetlb_acct_memory(struct hstate *h, long delta);
74
75static inline void unlock_or_release_subpool(struct hugepage_subpool *spool)
76{
77 bool free = (spool->count == 0) && (spool->used_hpages == 0);
78
79 spin_unlock(&spool->lock);
80
81 /* If no pages are used, and no other handles to the subpool
82 * remain, give up any reservations mased on minimum size and
83 * free the subpool */
84 if (free) {
85 if (spool->min_hpages != -1)
86 hugetlb_acct_memory(spool->hstate,
87 -spool->min_hpages);
88 kfree(spool);
89 }
90}
91
92struct hugepage_subpool *hugepage_new_subpool(struct hstate *h, long max_hpages,
93 long min_hpages)
94{
95 struct hugepage_subpool *spool;
96
97 spool = kzalloc(sizeof(*spool), GFP_KERNEL);
98 if (!spool)
99 return NULL;
100
101 spin_lock_init(&spool->lock);
102 spool->count = 1;
103 spool->max_hpages = max_hpages;
104 spool->hstate = h;
105 spool->min_hpages = min_hpages;
106
107 if (min_hpages != -1 && hugetlb_acct_memory(h, min_hpages)) {
108 kfree(spool);
109 return NULL;
110 }
111 spool->rsv_hpages = min_hpages;
112
113 return spool;
114}
115
116void hugepage_put_subpool(struct hugepage_subpool *spool)
117{
118 spin_lock(&spool->lock);
119 BUG_ON(!spool->count);
120 spool->count--;
121 unlock_or_release_subpool(spool);
122}
123
124/*
125 * Subpool accounting for allocating and reserving pages.
126 * Return -ENOMEM if there are not enough resources to satisfy the
127 * the request. Otherwise, return the number of pages by which the
128 * global pools must be adjusted (upward). The returned value may
129 * only be different than the passed value (delta) in the case where
130 * a subpool minimum size must be manitained.
131 */
132static long hugepage_subpool_get_pages(struct hugepage_subpool *spool,
133 long delta)
134{
135 long ret = delta;
136
137 if (!spool)
138 return ret;
139
140 spin_lock(&spool->lock);
141
142 if (spool->max_hpages != -1) { /* maximum size accounting */
143 if ((spool->used_hpages + delta) <= spool->max_hpages)
144 spool->used_hpages += delta;
145 else {
146 ret = -ENOMEM;
147 goto unlock_ret;
148 }
149 }
150
151 /* minimum size accounting */
152 if (spool->min_hpages != -1 && spool->rsv_hpages) {
153 if (delta > spool->rsv_hpages) {
154 /*
155 * Asking for more reserves than those already taken on
156 * behalf of subpool. Return difference.
157 */
158 ret = delta - spool->rsv_hpages;
159 spool->rsv_hpages = 0;
160 } else {
161 ret = 0; /* reserves already accounted for */
162 spool->rsv_hpages -= delta;
163 }
164 }
165
166unlock_ret:
167 spin_unlock(&spool->lock);
168 return ret;
169}
170
171/*
172 * Subpool accounting for freeing and unreserving pages.
173 * Return the number of global page reservations that must be dropped.
174 * The return value may only be different than the passed value (delta)
175 * in the case where a subpool minimum size must be maintained.
176 */
177static long hugepage_subpool_put_pages(struct hugepage_subpool *spool,
178 long delta)
179{
180 long ret = delta;
181
182 if (!spool)
183 return delta;
184
185 spin_lock(&spool->lock);
186
187 if (spool->max_hpages != -1) /* maximum size accounting */
188 spool->used_hpages -= delta;
189
190 /* minimum size accounting */
191 if (spool->min_hpages != -1 && spool->used_hpages < spool->min_hpages) {
192 if (spool->rsv_hpages + delta <= spool->min_hpages)
193 ret = 0;
194 else
195 ret = spool->rsv_hpages + delta - spool->min_hpages;
196
197 spool->rsv_hpages += delta;
198 if (spool->rsv_hpages > spool->min_hpages)
199 spool->rsv_hpages = spool->min_hpages;
200 }
201
202 /*
203 * If hugetlbfs_put_super couldn't free spool due to an outstanding
204 * quota reference, free it now.
205 */
206 unlock_or_release_subpool(spool);
207
208 return ret;
209}
210
211static inline struct hugepage_subpool *subpool_inode(struct inode *inode)
212{
213 return HUGETLBFS_SB(inode->i_sb)->spool;
214}
215
216static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma)
217{
218 return subpool_inode(file_inode(vma->vm_file));
219}
220
221/*
222 * Region tracking -- allows tracking of reservations and instantiated pages
223 * across the pages in a mapping.
224 *
225 * The region data structures are embedded into a resv_map and protected
226 * by a resv_map's lock. The set of regions within the resv_map represent
227 * reservations for huge pages, or huge pages that have already been
228 * instantiated within the map. The from and to elements are huge page
229 * indicies into the associated mapping. from indicates the starting index
230 * of the region. to represents the first index past the end of the region.
231 *
232 * For example, a file region structure with from == 0 and to == 4 represents
233 * four huge pages in a mapping. It is important to note that the to element
234 * represents the first element past the end of the region. This is used in
235 * arithmetic as 4(to) - 0(from) = 4 huge pages in the region.
236 *
237 * Interval notation of the form [from, to) will be used to indicate that
238 * the endpoint from is inclusive and to is exclusive.
239 */
240struct file_region {
241 struct list_head link;
242 long from;
243 long to;
244};
245
246/*
247 * Add the huge page range represented by [f, t) to the reserve
248 * map. In the normal case, existing regions will be expanded
249 * to accommodate the specified range. Sufficient regions should
250 * exist for expansion due to the previous call to region_chg
251 * with the same range. However, it is possible that region_del
252 * could have been called after region_chg and modifed the map
253 * in such a way that no region exists to be expanded. In this
254 * case, pull a region descriptor from the cache associated with
255 * the map and use that for the new range.
256 *
257 * Return the number of new huge pages added to the map. This
258 * number is greater than or equal to zero.
259 */
260static long region_add(struct resv_map *resv, long f, long t)
261{
262 struct list_head *head = &resv->regions;
263 struct file_region *rg, *nrg, *trg;
264 long add = 0;
265
266 spin_lock(&resv->lock);
267 /* Locate the region we are either in or before. */
268 list_for_each_entry(rg, head, link)
269 if (f <= rg->to)
270 break;
271
272 /*
273 * If no region exists which can be expanded to include the
274 * specified range, the list must have been modified by an
275 * interleving call to region_del(). Pull a region descriptor
276 * from the cache and use it for this range.
277 */
278 if (&rg->link == head || t < rg->from) {
279 VM_BUG_ON(resv->region_cache_count <= 0);
280
281 resv->region_cache_count--;
282 nrg = list_first_entry(&resv->region_cache, struct file_region,
283 link);
284 list_del(&nrg->link);
285
286 nrg->from = f;
287 nrg->to = t;
288 list_add(&nrg->link, rg->link.prev);
289
290 add += t - f;
291 goto out_locked;
292 }
293
294 /* Round our left edge to the current segment if it encloses us. */
295 if (f > rg->from)
296 f = rg->from;
297
298 /* Check for and consume any regions we now overlap with. */
299 nrg = rg;
300 list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
301 if (&rg->link == head)
302 break;
303 if (rg->from > t)
304 break;
305
306 /* If this area reaches higher then extend our area to
307 * include it completely. If this is not the first area
308 * which we intend to reuse, free it. */
309 if (rg->to > t)
310 t = rg->to;
311 if (rg != nrg) {
312 /* Decrement return value by the deleted range.
313 * Another range will span this area so that by
314 * end of routine add will be >= zero
315 */
316 add -= (rg->to - rg->from);
317 list_del(&rg->link);
318 kfree(rg);
319 }
320 }
321
322 add += (nrg->from - f); /* Added to beginning of region */
323 nrg->from = f;
324 add += t - nrg->to; /* Added to end of region */
325 nrg->to = t;
326
327out_locked:
328 resv->adds_in_progress--;
329 spin_unlock(&resv->lock);
330 VM_BUG_ON(add < 0);
331 return add;
332}
333
334/*
335 * Examine the existing reserve map and determine how many
336 * huge pages in the specified range [f, t) are NOT currently
337 * represented. This routine is called before a subsequent
338 * call to region_add that will actually modify the reserve
339 * map to add the specified range [f, t). region_chg does
340 * not change the number of huge pages represented by the
341 * map. However, if the existing regions in the map can not
342 * be expanded to represent the new range, a new file_region
343 * structure is added to the map as a placeholder. This is
344 * so that the subsequent region_add call will have all the
345 * regions it needs and will not fail.
346 *
347 * Upon entry, region_chg will also examine the cache of region descriptors
348 * associated with the map. If there are not enough descriptors cached, one
349 * will be allocated for the in progress add operation.
350 *
351 * Returns the number of huge pages that need to be added to the existing
352 * reservation map for the range [f, t). This number is greater or equal to
353 * zero. -ENOMEM is returned if a new file_region structure or cache entry
354 * is needed and can not be allocated.
355 */
356static long region_chg(struct resv_map *resv, long f, long t)
357{
358 struct list_head *head = &resv->regions;
359 struct file_region *rg, *nrg = NULL;
360 long chg = 0;
361
362retry:
363 spin_lock(&resv->lock);
364retry_locked:
365 resv->adds_in_progress++;
366
367 /*
368 * Check for sufficient descriptors in the cache to accommodate
369 * the number of in progress add operations.
370 */
371 if (resv->adds_in_progress > resv->region_cache_count) {
372 struct file_region *trg;
373
374 VM_BUG_ON(resv->adds_in_progress - resv->region_cache_count > 1);
375 /* Must drop lock to allocate a new descriptor. */
376 resv->adds_in_progress--;
377 spin_unlock(&resv->lock);
378
379 trg = kmalloc(sizeof(*trg), GFP_KERNEL);
380 if (!trg) {
381 kfree(nrg);
382 return -ENOMEM;
383 }
384
385 spin_lock(&resv->lock);
386 list_add(&trg->link, &resv->region_cache);
387 resv->region_cache_count++;
388 goto retry_locked;
389 }
390
391 /* Locate the region we are before or in. */
392 list_for_each_entry(rg, head, link)
393 if (f <= rg->to)
394 break;
395
396 /* If we are below the current region then a new region is required.
397 * Subtle, allocate a new region at the position but make it zero
398 * size such that we can guarantee to record the reservation. */
399 if (&rg->link == head || t < rg->from) {
400 if (!nrg) {
401 resv->adds_in_progress--;
402 spin_unlock(&resv->lock);
403 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
404 if (!nrg)
405 return -ENOMEM;
406
407 nrg->from = f;
408 nrg->to = f;
409 INIT_LIST_HEAD(&nrg->link);
410 goto retry;
411 }
412
413 list_add(&nrg->link, rg->link.prev);
414 chg = t - f;
415 goto out_nrg;
416 }
417
418 /* Round our left edge to the current segment if it encloses us. */
419 if (f > rg->from)
420 f = rg->from;
421 chg = t - f;
422
423 /* Check for and consume any regions we now overlap with. */
424 list_for_each_entry(rg, rg->link.prev, link) {
425 if (&rg->link == head)
426 break;
427 if (rg->from > t)
428 goto out;
429
430 /* We overlap with this area, if it extends further than
431 * us then we must extend ourselves. Account for its
432 * existing reservation. */
433 if (rg->to > t) {
434 chg += rg->to - t;
435 t = rg->to;
436 }
437 chg -= rg->to - rg->from;
438 }
439
440out:
441 spin_unlock(&resv->lock);
442 /* We already know we raced and no longer need the new region */
443 kfree(nrg);
444 return chg;
445out_nrg:
446 spin_unlock(&resv->lock);
447 return chg;
448}
449
450/*
451 * Abort the in progress add operation. The adds_in_progress field
452 * of the resv_map keeps track of the operations in progress between
453 * calls to region_chg and region_add. Operations are sometimes
454 * aborted after the call to region_chg. In such cases, region_abort
455 * is called to decrement the adds_in_progress counter.
456 *
457 * NOTE: The range arguments [f, t) are not needed or used in this
458 * routine. They are kept to make reading the calling code easier as
459 * arguments will match the associated region_chg call.
460 */
461static void region_abort(struct resv_map *resv, long f, long t)
462{
463 spin_lock(&resv->lock);
464 VM_BUG_ON(!resv->region_cache_count);
465 resv->adds_in_progress--;
466 spin_unlock(&resv->lock);
467}
468
469/*
470 * Delete the specified range [f, t) from the reserve map. If the
471 * t parameter is LONG_MAX, this indicates that ALL regions after f
472 * should be deleted. Locate the regions which intersect [f, t)
473 * and either trim, delete or split the existing regions.
474 *
475 * Returns the number of huge pages deleted from the reserve map.
476 * In the normal case, the return value is zero or more. In the
477 * case where a region must be split, a new region descriptor must
478 * be allocated. If the allocation fails, -ENOMEM will be returned.
479 * NOTE: If the parameter t == LONG_MAX, then we will never split
480 * a region and possibly return -ENOMEM. Callers specifying
481 * t == LONG_MAX do not need to check for -ENOMEM error.
482 */
483static long region_del(struct resv_map *resv, long f, long t)
484{
485 struct list_head *head = &resv->regions;
486 struct file_region *rg, *trg;
487 struct file_region *nrg = NULL;
488 long del = 0;
489
490retry:
491 spin_lock(&resv->lock);
492 list_for_each_entry_safe(rg, trg, head, link) {
493 /*
494 * Skip regions before the range to be deleted. file_region
495 * ranges are normally of the form [from, to). However, there
496 * may be a "placeholder" entry in the map which is of the form
497 * (from, to) with from == to. Check for placeholder entries
498 * at the beginning of the range to be deleted.
499 */
500 if (rg->to <= f && (rg->to != rg->from || rg->to != f))
501 continue;
502
503 if (rg->from >= t)
504 break;
505
506 if (f > rg->from && t < rg->to) { /* Must split region */
507 /*
508 * Check for an entry in the cache before dropping
509 * lock and attempting allocation.
510 */
511 if (!nrg &&
512 resv->region_cache_count > resv->adds_in_progress) {
513 nrg = list_first_entry(&resv->region_cache,
514 struct file_region,
515 link);
516 list_del(&nrg->link);
517 resv->region_cache_count--;
518 }
519
520 if (!nrg) {
521 spin_unlock(&resv->lock);
522 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
523 if (!nrg)
524 return -ENOMEM;
525 goto retry;
526 }
527
528 del += t - f;
529
530 /* New entry for end of split region */
531 nrg->from = t;
532 nrg->to = rg->to;
533 INIT_LIST_HEAD(&nrg->link);
534
535 /* Original entry is trimmed */
536 rg->to = f;
537
538 list_add(&nrg->link, &rg->link);
539 nrg = NULL;
540 break;
541 }
542
543 if (f <= rg->from && t >= rg->to) { /* Remove entire region */
544 del += rg->to - rg->from;
545 list_del(&rg->link);
546 kfree(rg);
547 continue;
548 }
549
550 if (f <= rg->from) { /* Trim beginning of region */
551 del += t - rg->from;
552 rg->from = t;
553 } else { /* Trim end of region */
554 del += rg->to - f;
555 rg->to = f;
556 }
557 }
558
559 spin_unlock(&resv->lock);
560 kfree(nrg);
561 return del;
562}
563
564/*
565 * A rare out of memory error was encountered which prevented removal of
566 * the reserve map region for a page. The huge page itself was free'ed
567 * and removed from the page cache. This routine will adjust the subpool
568 * usage count, and the global reserve count if needed. By incrementing
569 * these counts, the reserve map entry which could not be deleted will
570 * appear as a "reserved" entry instead of simply dangling with incorrect
571 * counts.
572 */
573void hugetlb_fix_reserve_counts(struct inode *inode)
574{
575 struct hugepage_subpool *spool = subpool_inode(inode);
576 long rsv_adjust;
577
578 rsv_adjust = hugepage_subpool_get_pages(spool, 1);
579 if (rsv_adjust) {
580 struct hstate *h = hstate_inode(inode);
581
582 hugetlb_acct_memory(h, 1);
583 }
584}
585
586/*
587 * Count and return the number of huge pages in the reserve map
588 * that intersect with the range [f, t).
589 */
590static long region_count(struct resv_map *resv, long f, long t)
591{
592 struct list_head *head = &resv->regions;
593 struct file_region *rg;
594 long chg = 0;
595
596 spin_lock(&resv->lock);
597 /* Locate each segment we overlap with, and count that overlap. */
598 list_for_each_entry(rg, head, link) {
599 long seg_from;
600 long seg_to;
601
602 if (rg->to <= f)
603 continue;
604 if (rg->from >= t)
605 break;
606
607 seg_from = max(rg->from, f);
608 seg_to = min(rg->to, t);
609
610 chg += seg_to - seg_from;
611 }
612 spin_unlock(&resv->lock);
613
614 return chg;
615}
616
617/*
618 * Convert the address within this vma to the page offset within
619 * the mapping, in pagecache page units; huge pages here.
620 */
621static pgoff_t vma_hugecache_offset(struct hstate *h,
622 struct vm_area_struct *vma, unsigned long address)
623{
624 return ((address - vma->vm_start) >> huge_page_shift(h)) +
625 (vma->vm_pgoff >> huge_page_order(h));
626}
627
628pgoff_t linear_hugepage_index(struct vm_area_struct *vma,
629 unsigned long address)
630{
631 return vma_hugecache_offset(hstate_vma(vma), vma, address);
632}
633EXPORT_SYMBOL_GPL(linear_hugepage_index);
634
635/*
636 * Return the size of the pages allocated when backing a VMA. In the majority
637 * cases this will be same size as used by the page table entries.
638 */
639unsigned long vma_kernel_pagesize(struct vm_area_struct *vma)
640{
641 if (vma->vm_ops && vma->vm_ops->pagesize)
642 return vma->vm_ops->pagesize(vma);
643 return PAGE_SIZE;
644}
645EXPORT_SYMBOL_GPL(vma_kernel_pagesize);
646
647/*
648 * Return the page size being used by the MMU to back a VMA. In the majority
649 * of cases, the page size used by the kernel matches the MMU size. On
650 * architectures where it differs, an architecture-specific 'strong'
651 * version of this symbol is required.
652 */
653__weak unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
654{
655 return vma_kernel_pagesize(vma);
656}
657
658/*
659 * Flags for MAP_PRIVATE reservations. These are stored in the bottom
660 * bits of the reservation map pointer, which are always clear due to
661 * alignment.
662 */
663#define HPAGE_RESV_OWNER (1UL << 0)
664#define HPAGE_RESV_UNMAPPED (1UL << 1)
665#define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
666
667/*
668 * These helpers are used to track how many pages are reserved for
669 * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
670 * is guaranteed to have their future faults succeed.
671 *
672 * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
673 * the reserve counters are updated with the hugetlb_lock held. It is safe
674 * to reset the VMA at fork() time as it is not in use yet and there is no
675 * chance of the global counters getting corrupted as a result of the values.
676 *
677 * The private mapping reservation is represented in a subtly different
678 * manner to a shared mapping. A shared mapping has a region map associated
679 * with the underlying file, this region map represents the backing file
680 * pages which have ever had a reservation assigned which this persists even
681 * after the page is instantiated. A private mapping has a region map
682 * associated with the original mmap which is attached to all VMAs which
683 * reference it, this region map represents those offsets which have consumed
684 * reservation ie. where pages have been instantiated.
685 */
686static unsigned long get_vma_private_data(struct vm_area_struct *vma)
687{
688 return (unsigned long)vma->vm_private_data;
689}
690
691static void set_vma_private_data(struct vm_area_struct *vma,
692 unsigned long value)
693{
694 vma->vm_private_data = (void *)value;
695}
696
697struct resv_map *resv_map_alloc(void)
698{
699 struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
700 struct file_region *rg = kmalloc(sizeof(*rg), GFP_KERNEL);
701
702 if (!resv_map || !rg) {
703 kfree(resv_map);
704 kfree(rg);
705 return NULL;
706 }
707
708 kref_init(&resv_map->refs);
709 spin_lock_init(&resv_map->lock);
710 INIT_LIST_HEAD(&resv_map->regions);
711
712 resv_map->adds_in_progress = 0;
713
714 INIT_LIST_HEAD(&resv_map->region_cache);
715 list_add(&rg->link, &resv_map->region_cache);
716 resv_map->region_cache_count = 1;
717
718 return resv_map;
719}
720
721void resv_map_release(struct kref *ref)
722{
723 struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
724 struct list_head *head = &resv_map->region_cache;
725 struct file_region *rg, *trg;
726
727 /* Clear out any active regions before we release the map. */
728 region_del(resv_map, 0, LONG_MAX);
729
730 /* ... and any entries left in the cache */
731 list_for_each_entry_safe(rg, trg, head, link) {
732 list_del(&rg->link);
733 kfree(rg);
734 }
735
736 VM_BUG_ON(resv_map->adds_in_progress);
737
738 kfree(resv_map);
739}
740
741static inline struct resv_map *inode_resv_map(struct inode *inode)
742{
743 return inode->i_mapping->private_data;
744}
745
746static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
747{
748 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
749 if (vma->vm_flags & VM_MAYSHARE) {
750 struct address_space *mapping = vma->vm_file->f_mapping;
751 struct inode *inode = mapping->host;
752
753 return inode_resv_map(inode);
754
755 } else {
756 return (struct resv_map *)(get_vma_private_data(vma) &
757 ~HPAGE_RESV_MASK);
758 }
759}
760
761static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
762{
763 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
764 VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
765
766 set_vma_private_data(vma, (get_vma_private_data(vma) &
767 HPAGE_RESV_MASK) | (unsigned long)map);
768}
769
770static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
771{
772 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
773 VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
774
775 set_vma_private_data(vma, get_vma_private_data(vma) | flags);
776}
777
778static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
779{
780 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
781
782 return (get_vma_private_data(vma) & flag) != 0;
783}
784
785/* Reset counters to 0 and clear all HPAGE_RESV_* flags */
786void reset_vma_resv_huge_pages(struct vm_area_struct *vma)
787{
788 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
789 if (!(vma->vm_flags & VM_MAYSHARE))
790 vma->vm_private_data = (void *)0;
791}
792
793/* Returns true if the VMA has associated reserve pages */
794static bool vma_has_reserves(struct vm_area_struct *vma, long chg)
795{
796 if (vma->vm_flags & VM_NORESERVE) {
797 /*
798 * This address is already reserved by other process(chg == 0),
799 * so, we should decrement reserved count. Without decrementing,
800 * reserve count remains after releasing inode, because this
801 * allocated page will go into page cache and is regarded as
802 * coming from reserved pool in releasing step. Currently, we
803 * don't have any other solution to deal with this situation
804 * properly, so add work-around here.
805 */
806 if (vma->vm_flags & VM_MAYSHARE && chg == 0)
807 return true;
808 else
809 return false;
810 }
811
812 /* Shared mappings always use reserves */
813 if (vma->vm_flags & VM_MAYSHARE) {
814 /*
815 * We know VM_NORESERVE is not set. Therefore, there SHOULD
816 * be a region map for all pages. The only situation where
817 * there is no region map is if a hole was punched via
818 * fallocate. In this case, there really are no reverves to
819 * use. This situation is indicated if chg != 0.
820 */
821 if (chg)
822 return false;
823 else
824 return true;
825 }
826
827 /*
828 * Only the process that called mmap() has reserves for
829 * private mappings.
830 */
831 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
832 /*
833 * Like the shared case above, a hole punch or truncate
834 * could have been performed on the private mapping.
835 * Examine the value of chg to determine if reserves
836 * actually exist or were previously consumed.
837 * Very Subtle - The value of chg comes from a previous
838 * call to vma_needs_reserves(). The reserve map for
839 * private mappings has different (opposite) semantics
840 * than that of shared mappings. vma_needs_reserves()
841 * has already taken this difference in semantics into
842 * account. Therefore, the meaning of chg is the same
843 * as in the shared case above. Code could easily be
844 * combined, but keeping it separate draws attention to
845 * subtle differences.
846 */
847 if (chg)
848 return false;
849 else
850 return true;
851 }
852
853 return false;
854}
855
856static void enqueue_huge_page(struct hstate *h, struct page *page)
857{
858 int nid = page_to_nid(page);
859 list_move(&page->lru, &h->hugepage_freelists[nid]);
860 h->free_huge_pages++;
861 h->free_huge_pages_node[nid]++;
862}
863
864static struct page *dequeue_huge_page_node_exact(struct hstate *h, int nid)
865{
866 struct page *page;
867
868 list_for_each_entry(page, &h->hugepage_freelists[nid], lru)
869 if (!PageHWPoison(page))
870 break;
871 /*
872 * if 'non-isolated free hugepage' not found on the list,
873 * the allocation fails.
874 */
875 if (&h->hugepage_freelists[nid] == &page->lru)
876 return NULL;
877 list_move(&page->lru, &h->hugepage_activelist);
878 set_page_refcounted(page);
879 h->free_huge_pages--;
880 h->free_huge_pages_node[nid]--;
881 return page;
882}
883
884static struct page *dequeue_huge_page_nodemask(struct hstate *h, gfp_t gfp_mask, int nid,
885 nodemask_t *nmask)
886{
887 unsigned int cpuset_mems_cookie;
888 struct zonelist *zonelist;
889 struct zone *zone;
890 struct zoneref *z;
891 int node = NUMA_NO_NODE;
892
893 zonelist = node_zonelist(nid, gfp_mask);
894
895retry_cpuset:
896 cpuset_mems_cookie = read_mems_allowed_begin();
897 for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), nmask) {
898 struct page *page;
899
900 if (!cpuset_zone_allowed(zone, gfp_mask))
901 continue;
902 /*
903 * no need to ask again on the same node. Pool is node rather than
904 * zone aware
905 */
906 if (zone_to_nid(zone) == node)
907 continue;
908 node = zone_to_nid(zone);
909
910 page = dequeue_huge_page_node_exact(h, node);
911 if (page)
912 return page;
913 }
914 if (unlikely(read_mems_allowed_retry(cpuset_mems_cookie)))
915 goto retry_cpuset;
916
917 return NULL;
918}
919
920/* Movability of hugepages depends on migration support. */
921static inline gfp_t htlb_alloc_mask(struct hstate *h)
922{
923 if (hugepage_movable_supported(h))
924 return GFP_HIGHUSER_MOVABLE;
925 else
926 return GFP_HIGHUSER;
927}
928
929static struct page *dequeue_huge_page_vma(struct hstate *h,
930 struct vm_area_struct *vma,
931 unsigned long address, int avoid_reserve,
932 long chg)
933{
934 struct page *page;
935 struct mempolicy *mpol;
936 gfp_t gfp_mask;
937 nodemask_t *nodemask;
938 int nid;
939
940 /*
941 * A child process with MAP_PRIVATE mappings created by their parent
942 * have no page reserves. This check ensures that reservations are
943 * not "stolen". The child may still get SIGKILLed
944 */
945 if (!vma_has_reserves(vma, chg) &&
946 h->free_huge_pages - h->resv_huge_pages == 0)
947 goto err;
948
949 /* If reserves cannot be used, ensure enough pages are in the pool */
950 if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
951 goto err;
952
953 gfp_mask = htlb_alloc_mask(h);
954 nid = huge_node(vma, address, gfp_mask, &mpol, &nodemask);
955 page = dequeue_huge_page_nodemask(h, gfp_mask, nid, nodemask);
956 if (page && !avoid_reserve && vma_has_reserves(vma, chg)) {
957 SetPagePrivate(page);
958 h->resv_huge_pages--;
959 }
960
961 mpol_cond_put(mpol);
962 return page;
963
964err:
965 return NULL;
966}
967
968/*
969 * common helper functions for hstate_next_node_to_{alloc|free}.
970 * We may have allocated or freed a huge page based on a different
971 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
972 * be outside of *nodes_allowed. Ensure that we use an allowed
973 * node for alloc or free.
974 */
975static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
976{
977 nid = next_node_in(nid, *nodes_allowed);
978 VM_BUG_ON(nid >= MAX_NUMNODES);
979
980 return nid;
981}
982
983static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
984{
985 if (!node_isset(nid, *nodes_allowed))
986 nid = next_node_allowed(nid, nodes_allowed);
987 return nid;
988}
989
990/*
991 * returns the previously saved node ["this node"] from which to
992 * allocate a persistent huge page for the pool and advance the
993 * next node from which to allocate, handling wrap at end of node
994 * mask.
995 */
996static int hstate_next_node_to_alloc(struct hstate *h,
997 nodemask_t *nodes_allowed)
998{
999 int nid;
1000
1001 VM_BUG_ON(!nodes_allowed);
1002
1003 nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
1004 h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
1005
1006 return nid;
1007}
1008
1009/*
1010 * helper for free_pool_huge_page() - return the previously saved
1011 * node ["this node"] from which to free a huge page. Advance the
1012 * next node id whether or not we find a free huge page to free so
1013 * that the next attempt to free addresses the next node.
1014 */
1015static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
1016{
1017 int nid;
1018
1019 VM_BUG_ON(!nodes_allowed);
1020
1021 nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
1022 h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
1023
1024 return nid;
1025}
1026
1027#define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \
1028 for (nr_nodes = nodes_weight(*mask); \
1029 nr_nodes > 0 && \
1030 ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \
1031 nr_nodes--)
1032
1033#define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \
1034 for (nr_nodes = nodes_weight(*mask); \
1035 nr_nodes > 0 && \
1036 ((node = hstate_next_node_to_free(hs, mask)) || 1); \
1037 nr_nodes--)
1038
1039#ifdef CONFIG_ARCH_HAS_GIGANTIC_PAGE
1040static void destroy_compound_gigantic_page(struct page *page,
1041 unsigned int order)
1042{
1043 int i;
1044 int nr_pages = 1 << order;
1045 struct page *p = page + 1;
1046
1047 atomic_set(compound_mapcount_ptr(page), 0);
1048 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
1049 clear_compound_head(p);
1050 set_page_refcounted(p);
1051 }
1052
1053 set_compound_order(page, 0);
1054 __ClearPageHead(page);
1055}
1056
1057static void free_gigantic_page(struct page *page, unsigned int order)
1058{
1059 free_contig_range(page_to_pfn(page), 1 << order);
1060}
1061
1062static int __alloc_gigantic_page(unsigned long start_pfn,
1063 unsigned long nr_pages, gfp_t gfp_mask)
1064{
1065 unsigned long end_pfn = start_pfn + nr_pages;
1066 return alloc_contig_range(start_pfn, end_pfn, MIGRATE_MOVABLE,
1067 gfp_mask);
1068}
1069
1070static bool pfn_range_valid_gigantic(struct zone *z,
1071 unsigned long start_pfn, unsigned long nr_pages)
1072{
1073 unsigned long i, end_pfn = start_pfn + nr_pages;
1074 struct page *page;
1075
1076 for (i = start_pfn; i < end_pfn; i++) {
1077 if (!pfn_valid(i))
1078 return false;
1079
1080 page = pfn_to_page(i);
1081
1082 if (page_zone(page) != z)
1083 return false;
1084
1085 if (PageReserved(page))
1086 return false;
1087
1088 if (page_count(page) > 0)
1089 return false;
1090
1091 if (PageHuge(page))
1092 return false;
1093 }
1094
1095 return true;
1096}
1097
1098static bool zone_spans_last_pfn(const struct zone *zone,
1099 unsigned long start_pfn, unsigned long nr_pages)
1100{
1101 unsigned long last_pfn = start_pfn + nr_pages - 1;
1102 return zone_spans_pfn(zone, last_pfn);
1103}
1104
1105static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask,
1106 int nid, nodemask_t *nodemask)
1107{
1108 unsigned int order = huge_page_order(h);
1109 unsigned long nr_pages = 1 << order;
1110 unsigned long ret, pfn, flags;
1111 struct zonelist *zonelist;
1112 struct zone *zone;
1113 struct zoneref *z;
1114
1115 zonelist = node_zonelist(nid, gfp_mask);
1116 for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), nodemask) {
1117 spin_lock_irqsave(&zone->lock, flags);
1118
1119 pfn = ALIGN(zone->zone_start_pfn, nr_pages);
1120 while (zone_spans_last_pfn(zone, pfn, nr_pages)) {
1121 if (pfn_range_valid_gigantic(zone, pfn, nr_pages)) {
1122 /*
1123 * We release the zone lock here because
1124 * alloc_contig_range() will also lock the zone
1125 * at some point. If there's an allocation
1126 * spinning on this lock, it may win the race
1127 * and cause alloc_contig_range() to fail...
1128 */
1129 spin_unlock_irqrestore(&zone->lock, flags);
1130 ret = __alloc_gigantic_page(pfn, nr_pages, gfp_mask);
1131 if (!ret)
1132 return pfn_to_page(pfn);
1133 spin_lock_irqsave(&zone->lock, flags);
1134 }
1135 pfn += nr_pages;
1136 }
1137
1138 spin_unlock_irqrestore(&zone->lock, flags);
1139 }
1140
1141 return NULL;
1142}
1143
1144static void prep_new_huge_page(struct hstate *h, struct page *page, int nid);
1145static void prep_compound_gigantic_page(struct page *page, unsigned int order);
1146
1147#else /* !CONFIG_ARCH_HAS_GIGANTIC_PAGE */
1148static inline bool gigantic_page_supported(void) { return false; }
1149static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask,
1150 int nid, nodemask_t *nodemask) { return NULL; }
1151static inline void free_gigantic_page(struct page *page, unsigned int order) { }
1152static inline void destroy_compound_gigantic_page(struct page *page,
1153 unsigned int order) { }
1154#endif
1155
1156static void update_and_free_page(struct hstate *h, struct page *page)
1157{
1158 int i;
1159
1160 if (hstate_is_gigantic(h) && !gigantic_page_supported())
1161 return;
1162
1163 h->nr_huge_pages--;
1164 h->nr_huge_pages_node[page_to_nid(page)]--;
1165 for (i = 0; i < pages_per_huge_page(h); i++) {
1166 page[i].flags &= ~(1 << PG_locked | 1 << PG_error |
1167 1 << PG_referenced | 1 << PG_dirty |
1168 1 << PG_active | 1 << PG_private |
1169 1 << PG_writeback);
1170 }
1171 VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page);
1172 set_compound_page_dtor(page, NULL_COMPOUND_DTOR);
1173 set_page_refcounted(page);
1174 if (hstate_is_gigantic(h)) {
1175 destroy_compound_gigantic_page(page, huge_page_order(h));
1176 free_gigantic_page(page, huge_page_order(h));
1177 } else {
1178 __free_pages(page, huge_page_order(h));
1179 }
1180}
1181
1182struct hstate *size_to_hstate(unsigned long size)
1183{
1184 struct hstate *h;
1185
1186 for_each_hstate(h) {
1187 if (huge_page_size(h) == size)
1188 return h;
1189 }
1190 return NULL;
1191}
1192
1193/*
1194 * Test to determine whether the hugepage is "active/in-use" (i.e. being linked
1195 * to hstate->hugepage_activelist.)
1196 *
1197 * This function can be called for tail pages, but never returns true for them.
1198 */
1199bool page_huge_active(struct page *page)
1200{
1201 VM_BUG_ON_PAGE(!PageHuge(page), page);
1202 return PageHead(page) && PagePrivate(&page[1]);
1203}
1204
1205/* never called for tail page */
1206static void set_page_huge_active(struct page *page)
1207{
1208 VM_BUG_ON_PAGE(!PageHeadHuge(page), page);
1209 SetPagePrivate(&page[1]);
1210}
1211
1212static void clear_page_huge_active(struct page *page)
1213{
1214 VM_BUG_ON_PAGE(!PageHeadHuge(page), page);
1215 ClearPagePrivate(&page[1]);
1216}
1217
1218/*
1219 * Internal hugetlb specific page flag. Do not use outside of the hugetlb
1220 * code
1221 */
1222static inline bool PageHugeTemporary(struct page *page)
1223{
1224 if (!PageHuge(page))
1225 return false;
1226
1227 return (unsigned long)page[2].mapping == -1U;
1228}
1229
1230static inline void SetPageHugeTemporary(struct page *page)
1231{
1232 page[2].mapping = (void *)-1U;
1233}
1234
1235static inline void ClearPageHugeTemporary(struct page *page)
1236{
1237 page[2].mapping = NULL;
1238}
1239
1240void free_huge_page(struct page *page)
1241{
1242 /*
1243 * Can't pass hstate in here because it is called from the
1244 * compound page destructor.
1245 */
1246 struct hstate *h = page_hstate(page);
1247 int nid = page_to_nid(page);
1248 struct hugepage_subpool *spool =
1249 (struct hugepage_subpool *)page_private(page);
1250 bool restore_reserve;
1251
1252 VM_BUG_ON_PAGE(page_count(page), page);
1253 VM_BUG_ON_PAGE(page_mapcount(page), page);
1254
1255 set_page_private(page, 0);
1256 page->mapping = NULL;
1257 restore_reserve = PagePrivate(page);
1258 ClearPagePrivate(page);
1259
1260 /*
1261 * A return code of zero implies that the subpool will be under its
1262 * minimum size if the reservation is not restored after page is free.
1263 * Therefore, force restore_reserve operation.
1264 */
1265 if (hugepage_subpool_put_pages(spool, 1) == 0)
1266 restore_reserve = true;
1267
1268 spin_lock(&hugetlb_lock);
1269 clear_page_huge_active(page);
1270 hugetlb_cgroup_uncharge_page(hstate_index(h),
1271 pages_per_huge_page(h), page);
1272 if (restore_reserve)
1273 h->resv_huge_pages++;
1274
1275 if (PageHugeTemporary(page)) {
1276 list_del(&page->lru);
1277 ClearPageHugeTemporary(page);
1278 update_and_free_page(h, page);
1279 } else if (h->surplus_huge_pages_node[nid]) {
1280 /* remove the page from active list */
1281 list_del(&page->lru);
1282 update_and_free_page(h, page);
1283 h->surplus_huge_pages--;
1284 h->surplus_huge_pages_node[nid]--;
1285 } else {
1286 arch_clear_hugepage_flags(page);
1287 enqueue_huge_page(h, page);
1288 }
1289 spin_unlock(&hugetlb_lock);
1290}
1291
1292static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
1293{
1294 INIT_LIST_HEAD(&page->lru);
1295 set_compound_page_dtor(page, HUGETLB_PAGE_DTOR);
1296 spin_lock(&hugetlb_lock);
1297 set_hugetlb_cgroup(page, NULL);
1298 h->nr_huge_pages++;
1299 h->nr_huge_pages_node[nid]++;
1300 spin_unlock(&hugetlb_lock);
1301}
1302
1303static void prep_compound_gigantic_page(struct page *page, unsigned int order)
1304{
1305 int i;
1306 int nr_pages = 1 << order;
1307 struct page *p = page + 1;
1308
1309 /* we rely on prep_new_huge_page to set the destructor */
1310 set_compound_order(page, order);
1311 __ClearPageReserved(page);
1312 __SetPageHead(page);
1313 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
1314 /*
1315 * For gigantic hugepages allocated through bootmem at
1316 * boot, it's safer to be consistent with the not-gigantic
1317 * hugepages and clear the PG_reserved bit from all tail pages
1318 * too. Otherwse drivers using get_user_pages() to access tail
1319 * pages may get the reference counting wrong if they see
1320 * PG_reserved set on a tail page (despite the head page not
1321 * having PG_reserved set). Enforcing this consistency between
1322 * head and tail pages allows drivers to optimize away a check
1323 * on the head page when they need know if put_page() is needed
1324 * after get_user_pages().
1325 */
1326 __ClearPageReserved(p);
1327 set_page_count(p, 0);
1328 set_compound_head(p, page);
1329 }
1330 atomic_set(compound_mapcount_ptr(page), -1);
1331}
1332
1333/*
1334 * PageHuge() only returns true for hugetlbfs pages, but not for normal or
1335 * transparent huge pages. See the PageTransHuge() documentation for more
1336 * details.
1337 */
1338int PageHuge(struct page *page)
1339{
1340 if (!PageCompound(page))
1341 return 0;
1342
1343 page = compound_head(page);
1344 return page[1].compound_dtor == HUGETLB_PAGE_DTOR;
1345}
1346EXPORT_SYMBOL_GPL(PageHuge);
1347
1348/*
1349 * PageHeadHuge() only returns true for hugetlbfs head page, but not for
1350 * normal or transparent huge pages.
1351 */
1352int PageHeadHuge(struct page *page_head)
1353{
1354 if (!PageHead(page_head))
1355 return 0;
1356
1357 return get_compound_page_dtor(page_head) == free_huge_page;
1358}
1359
1360pgoff_t __basepage_index(struct page *page)
1361{
1362 struct page *page_head = compound_head(page);
1363 pgoff_t index = page_index(page_head);
1364 unsigned long compound_idx;
1365
1366 if (!PageHuge(page_head))
1367 return page_index(page);
1368
1369 if (compound_order(page_head) >= MAX_ORDER)
1370 compound_idx = page_to_pfn(page) - page_to_pfn(page_head);
1371 else
1372 compound_idx = page - page_head;
1373
1374 return (index << compound_order(page_head)) + compound_idx;
1375}
1376
1377static struct page *alloc_buddy_huge_page(struct hstate *h,
1378 gfp_t gfp_mask, int nid, nodemask_t *nmask)
1379{
1380 int order = huge_page_order(h);
1381 struct page *page;
1382
1383 gfp_mask |= __GFP_COMP|__GFP_RETRY_MAYFAIL|__GFP_NOWARN;
1384 if (nid == NUMA_NO_NODE)
1385 nid = numa_mem_id();
1386 page = __alloc_pages_nodemask(gfp_mask, order, nid, nmask);
1387 if (page)
1388 __count_vm_event(HTLB_BUDDY_PGALLOC);
1389 else
1390 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
1391
1392 return page;
1393}
1394
1395/*
1396 * Common helper to allocate a fresh hugetlb page. All specific allocators
1397 * should use this function to get new hugetlb pages
1398 */
1399static struct page *alloc_fresh_huge_page(struct hstate *h,
1400 gfp_t gfp_mask, int nid, nodemask_t *nmask)
1401{
1402 struct page *page;
1403
1404 if (hstate_is_gigantic(h))
1405 page = alloc_gigantic_page(h, gfp_mask, nid, nmask);
1406 else
1407 page = alloc_buddy_huge_page(h, gfp_mask,
1408 nid, nmask);
1409 if (!page)
1410 return NULL;
1411
1412 if (hstate_is_gigantic(h))
1413 prep_compound_gigantic_page(page, huge_page_order(h));
1414 prep_new_huge_page(h, page, page_to_nid(page));
1415
1416 return page;
1417}
1418
1419/*
1420 * Allocates a fresh page to the hugetlb allocator pool in the node interleaved
1421 * manner.
1422 */
1423static int alloc_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed)
1424{
1425 struct page *page;
1426 int nr_nodes, node;
1427 gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE;
1428
1429 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
1430 page = alloc_fresh_huge_page(h, gfp_mask, node, nodes_allowed);
1431 if (page)
1432 break;
1433 }
1434
1435 if (!page)
1436 return 0;
1437
1438 put_page(page); /* free it into the hugepage allocator */
1439
1440 return 1;
1441}
1442
1443/*
1444 * Free huge page from pool from next node to free.
1445 * Attempt to keep persistent huge pages more or less
1446 * balanced over allowed nodes.
1447 * Called with hugetlb_lock locked.
1448 */
1449static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
1450 bool acct_surplus)
1451{
1452 int nr_nodes, node;
1453 int ret = 0;
1454
1455 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
1456 /*
1457 * If we're returning unused surplus pages, only examine
1458 * nodes with surplus pages.
1459 */
1460 if ((!acct_surplus || h->surplus_huge_pages_node[node]) &&
1461 !list_empty(&h->hugepage_freelists[node])) {
1462 struct page *page =
1463 list_entry(h->hugepage_freelists[node].next,
1464 struct page, lru);
1465 list_del(&page->lru);
1466 h->free_huge_pages--;
1467 h->free_huge_pages_node[node]--;
1468 if (acct_surplus) {
1469 h->surplus_huge_pages--;
1470 h->surplus_huge_pages_node[node]--;
1471 }
1472 update_and_free_page(h, page);
1473 ret = 1;
1474 break;
1475 }
1476 }
1477
1478 return ret;
1479}
1480
1481/*
1482 * Dissolve a given free hugepage into free buddy pages. This function does
1483 * nothing for in-use (including surplus) hugepages. Returns -EBUSY if the
1484 * dissolution fails because a give page is not a free hugepage, or because
1485 * free hugepages are fully reserved.
1486 */
1487int dissolve_free_huge_page(struct page *page)
1488{
1489 int rc = -EBUSY;
1490
1491 spin_lock(&hugetlb_lock);
1492 if (PageHuge(page) && !page_count(page)) {
1493 struct page *head = compound_head(page);
1494 struct hstate *h = page_hstate(head);
1495 int nid = page_to_nid(head);
1496 if (h->free_huge_pages - h->resv_huge_pages == 0)
1497 goto out;
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 rc = 0;
1512 }
1513out:
1514 spin_unlock(&hugetlb_lock);
1515 return rc;
1516}
1517
1518/*
1519 * Dissolve free hugepages in a given pfn range. Used by memory hotplug to
1520 * make specified memory blocks removable from the system.
1521 * Note that this will dissolve a free gigantic hugepage completely, if any
1522 * part of it lies within the given range.
1523 * Also note that if dissolve_free_huge_page() returns with an error, all
1524 * free hugepages that were dissolved before that error are lost.
1525 */
1526int dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn)
1527{
1528 unsigned long pfn;
1529 struct page *page;
1530 int rc = 0;
1531
1532 if (!hugepages_supported())
1533 return rc;
1534
1535 for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << minimum_order) {
1536 page = pfn_to_page(pfn);
1537 if (PageHuge(page) && !page_count(page)) {
1538 rc = dissolve_free_huge_page(page);
1539 if (rc)
1540 break;
1541 }
1542 }
1543
1544 return rc;
1545}
1546
1547/*
1548 * Allocates a fresh surplus page from the page allocator.
1549 */
1550static struct page *alloc_surplus_huge_page(struct hstate *h, gfp_t gfp_mask,
1551 int nid, nodemask_t *nmask)
1552{
1553 struct page *page = NULL;
1554
1555 if (hstate_is_gigantic(h))
1556 return NULL;
1557
1558 spin_lock(&hugetlb_lock);
1559 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages)
1560 goto out_unlock;
1561 spin_unlock(&hugetlb_lock);
1562
1563 page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask);
1564 if (!page)
1565 return NULL;
1566
1567 spin_lock(&hugetlb_lock);
1568 /*
1569 * We could have raced with the pool size change.
1570 * Double check that and simply deallocate the new page
1571 * if we would end up overcommiting the surpluses. Abuse
1572 * temporary page to workaround the nasty free_huge_page
1573 * codeflow
1574 */
1575 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
1576 SetPageHugeTemporary(page);
1577 put_page(page);
1578 page = NULL;
1579 } else {
1580 h->surplus_huge_pages++;
1581 h->surplus_huge_pages_node[page_to_nid(page)]++;
1582 }
1583
1584out_unlock:
1585 spin_unlock(&hugetlb_lock);
1586
1587 return page;
1588}
1589
1590struct page *alloc_migrate_huge_page(struct hstate *h, gfp_t gfp_mask,
1591 int nid, nodemask_t *nmask)
1592{
1593 struct page *page;
1594
1595 if (hstate_is_gigantic(h))
1596 return NULL;
1597
1598 page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask);
1599 if (!page)
1600 return NULL;
1601
1602 /*
1603 * We do not account these pages as surplus because they are only
1604 * temporary and will be released properly on the last reference
1605 */
1606 SetPageHugeTemporary(page);
1607
1608 return page;
1609}
1610
1611/*
1612 * Use the VMA's mpolicy to allocate a huge page from the buddy.
1613 */
1614static
1615struct page *alloc_buddy_huge_page_with_mpol(struct hstate *h,
1616 struct vm_area_struct *vma, unsigned long addr)
1617{
1618 struct page *page;
1619 struct mempolicy *mpol;
1620 gfp_t gfp_mask = htlb_alloc_mask(h);
1621 int nid;
1622 nodemask_t *nodemask;
1623
1624 nid = huge_node(vma, addr, gfp_mask, &mpol, &nodemask);
1625 page = alloc_surplus_huge_page(h, gfp_mask, nid, nodemask);
1626 mpol_cond_put(mpol);
1627
1628 return page;
1629}
1630
1631/* page migration callback function */
1632struct page *alloc_huge_page_node(struct hstate *h, int nid)
1633{
1634 gfp_t gfp_mask = htlb_alloc_mask(h);
1635 struct page *page = NULL;
1636
1637 if (nid != NUMA_NO_NODE)
1638 gfp_mask |= __GFP_THISNODE;
1639
1640 spin_lock(&hugetlb_lock);
1641 if (h->free_huge_pages - h->resv_huge_pages > 0)
1642 page = dequeue_huge_page_nodemask(h, gfp_mask, nid, NULL);
1643 spin_unlock(&hugetlb_lock);
1644
1645 if (!page)
1646 page = alloc_migrate_huge_page(h, gfp_mask, nid, NULL);
1647
1648 return page;
1649}
1650
1651/* page migration callback function */
1652struct page *alloc_huge_page_nodemask(struct hstate *h, int preferred_nid,
1653 nodemask_t *nmask)
1654{
1655 gfp_t gfp_mask = htlb_alloc_mask(h);
1656
1657 spin_lock(&hugetlb_lock);
1658 if (h->free_huge_pages - h->resv_huge_pages > 0) {
1659 struct page *page;
1660
1661 page = dequeue_huge_page_nodemask(h, gfp_mask, preferred_nid, nmask);
1662 if (page) {
1663 spin_unlock(&hugetlb_lock);
1664 return page;
1665 }
1666 }
1667 spin_unlock(&hugetlb_lock);
1668
1669 return alloc_migrate_huge_page(h, gfp_mask, preferred_nid, nmask);
1670}
1671
1672/* mempolicy aware migration callback */
1673struct page *alloc_huge_page_vma(struct hstate *h, struct vm_area_struct *vma,
1674 unsigned long address)
1675{
1676 struct mempolicy *mpol;
1677 nodemask_t *nodemask;
1678 struct page *page;
1679 gfp_t gfp_mask;
1680 int node;
1681
1682 gfp_mask = htlb_alloc_mask(h);
1683 node = huge_node(vma, address, gfp_mask, &mpol, &nodemask);
1684 page = alloc_huge_page_nodemask(h, node, nodemask);
1685 mpol_cond_put(mpol);
1686
1687 return page;
1688}
1689
1690/*
1691 * Increase the hugetlb pool such that it can accommodate a reservation
1692 * of size 'delta'.
1693 */
1694static int gather_surplus_pages(struct hstate *h, int delta)
1695{
1696 struct list_head surplus_list;
1697 struct page *page, *tmp;
1698 int ret, i;
1699 int needed, allocated;
1700 bool alloc_ok = true;
1701
1702 needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
1703 if (needed <= 0) {
1704 h->resv_huge_pages += delta;
1705 return 0;
1706 }
1707
1708 allocated = 0;
1709 INIT_LIST_HEAD(&surplus_list);
1710
1711 ret = -ENOMEM;
1712retry:
1713 spin_unlock(&hugetlb_lock);
1714 for (i = 0; i < needed; i++) {
1715 page = alloc_surplus_huge_page(h, htlb_alloc_mask(h),
1716 NUMA_NO_NODE, NULL);
1717 if (!page) {
1718 alloc_ok = false;
1719 break;
1720 }
1721 list_add(&page->lru, &surplus_list);
1722 cond_resched();
1723 }
1724 allocated += i;
1725
1726 /*
1727 * After retaking hugetlb_lock, we need to recalculate 'needed'
1728 * because either resv_huge_pages or free_huge_pages may have changed.
1729 */
1730 spin_lock(&hugetlb_lock);
1731 needed = (h->resv_huge_pages + delta) -
1732 (h->free_huge_pages + allocated);
1733 if (needed > 0) {
1734 if (alloc_ok)
1735 goto retry;
1736 /*
1737 * We were not able to allocate enough pages to
1738 * satisfy the entire reservation so we free what
1739 * we've allocated so far.
1740 */
1741 goto free;
1742 }
1743 /*
1744 * The surplus_list now contains _at_least_ the number of extra pages
1745 * needed to accommodate the reservation. Add the appropriate number
1746 * of pages to the hugetlb pool and free the extras back to the buddy
1747 * allocator. Commit the entire reservation here to prevent another
1748 * process from stealing the pages as they are added to the pool but
1749 * before they are reserved.
1750 */
1751 needed += allocated;
1752 h->resv_huge_pages += delta;
1753 ret = 0;
1754
1755 /* Free the needed pages to the hugetlb pool */
1756 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
1757 if ((--needed) < 0)
1758 break;
1759 /*
1760 * This page is now managed by the hugetlb allocator and has
1761 * no users -- drop the buddy allocator's reference.
1762 */
1763 put_page_testzero(page);
1764 VM_BUG_ON_PAGE(page_count(page), page);
1765 enqueue_huge_page(h, page);
1766 }
1767free:
1768 spin_unlock(&hugetlb_lock);
1769
1770 /* Free unnecessary surplus pages to the buddy allocator */
1771 list_for_each_entry_safe(page, tmp, &surplus_list, lru)
1772 put_page(page);
1773 spin_lock(&hugetlb_lock);
1774
1775 return ret;
1776}
1777
1778/*
1779 * This routine has two main purposes:
1780 * 1) Decrement the reservation count (resv_huge_pages) by the value passed
1781 * in unused_resv_pages. This corresponds to the prior adjustments made
1782 * to the associated reservation map.
1783 * 2) Free any unused surplus pages that may have been allocated to satisfy
1784 * the reservation. As many as unused_resv_pages may be freed.
1785 *
1786 * Called with hugetlb_lock held. However, the lock could be dropped (and
1787 * reacquired) during calls to cond_resched_lock. Whenever dropping the lock,
1788 * we must make sure nobody else can claim pages we are in the process of
1789 * freeing. Do this by ensuring resv_huge_page always is greater than the
1790 * number of huge pages we plan to free when dropping the lock.
1791 */
1792static void return_unused_surplus_pages(struct hstate *h,
1793 unsigned long unused_resv_pages)
1794{
1795 unsigned long nr_pages;
1796
1797 /* Cannot return gigantic pages currently */
1798 if (hstate_is_gigantic(h))
1799 goto out;
1800
1801 /*
1802 * Part (or even all) of the reservation could have been backed
1803 * by pre-allocated pages. Only free surplus pages.
1804 */
1805 nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
1806
1807 /*
1808 * We want to release as many surplus pages as possible, spread
1809 * evenly across all nodes with memory. Iterate across these nodes
1810 * until we can no longer free unreserved surplus pages. This occurs
1811 * when the nodes with surplus pages have no free pages.
1812 * free_pool_huge_page() will balance the the freed pages across the
1813 * on-line nodes with memory and will handle the hstate accounting.
1814 *
1815 * Note that we decrement resv_huge_pages as we free the pages. If
1816 * we drop the lock, resv_huge_pages will still be sufficiently large
1817 * to cover subsequent pages we may free.
1818 */
1819 while (nr_pages--) {
1820 h->resv_huge_pages--;
1821 unused_resv_pages--;
1822 if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1))
1823 goto out;
1824 cond_resched_lock(&hugetlb_lock);
1825 }
1826
1827out:
1828 /* Fully uncommit the reservation */
1829 h->resv_huge_pages -= unused_resv_pages;
1830}
1831
1832
1833/*
1834 * vma_needs_reservation, vma_commit_reservation and vma_end_reservation
1835 * are used by the huge page allocation routines to manage reservations.
1836 *
1837 * vma_needs_reservation is called to determine if the huge page at addr
1838 * within the vma has an associated reservation. If a reservation is
1839 * needed, the value 1 is returned. The caller is then responsible for
1840 * managing the global reservation and subpool usage counts. After
1841 * the huge page has been allocated, vma_commit_reservation is called
1842 * to add the page to the reservation map. If the page allocation fails,
1843 * the reservation must be ended instead of committed. vma_end_reservation
1844 * is called in such cases.
1845 *
1846 * In the normal case, vma_commit_reservation returns the same value
1847 * as the preceding vma_needs_reservation call. The only time this
1848 * is not the case is if a reserve map was changed between calls. It
1849 * is the responsibility of the caller to notice the difference and
1850 * take appropriate action.
1851 *
1852 * vma_add_reservation is used in error paths where a reservation must
1853 * be restored when a newly allocated huge page must be freed. It is
1854 * to be called after calling vma_needs_reservation to determine if a
1855 * reservation exists.
1856 */
1857enum vma_resv_mode {
1858 VMA_NEEDS_RESV,
1859 VMA_COMMIT_RESV,
1860 VMA_END_RESV,
1861 VMA_ADD_RESV,
1862};
1863static long __vma_reservation_common(struct hstate *h,
1864 struct vm_area_struct *vma, unsigned long addr,
1865 enum vma_resv_mode mode)
1866{
1867 struct resv_map *resv;
1868 pgoff_t idx;
1869 long ret;
1870
1871 resv = vma_resv_map(vma);
1872 if (!resv)
1873 return 1;
1874
1875 idx = vma_hugecache_offset(h, vma, addr);
1876 switch (mode) {
1877 case VMA_NEEDS_RESV:
1878 ret = region_chg(resv, idx, idx + 1);
1879 break;
1880 case VMA_COMMIT_RESV:
1881 ret = region_add(resv, idx, idx + 1);
1882 break;
1883 case VMA_END_RESV:
1884 region_abort(resv, idx, idx + 1);
1885 ret = 0;
1886 break;
1887 case VMA_ADD_RESV:
1888 if (vma->vm_flags & VM_MAYSHARE)
1889 ret = region_add(resv, idx, idx + 1);
1890 else {
1891 region_abort(resv, idx, idx + 1);
1892 ret = region_del(resv, idx, idx + 1);
1893 }
1894 break;
1895 default:
1896 BUG();
1897 }
1898
1899 if (vma->vm_flags & VM_MAYSHARE)
1900 return ret;
1901 else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && ret >= 0) {
1902 /*
1903 * In most cases, reserves always exist for private mappings.
1904 * However, a file associated with mapping could have been
1905 * hole punched or truncated after reserves were consumed.
1906 * As subsequent fault on such a range will not use reserves.
1907 * Subtle - The reserve map for private mappings has the
1908 * opposite meaning than that of shared mappings. If NO
1909 * entry is in the reserve map, it means a reservation exists.
1910 * If an entry exists in the reserve map, it means the
1911 * reservation has already been consumed. As a result, the
1912 * return value of this routine is the opposite of the
1913 * value returned from reserve map manipulation routines above.
1914 */
1915 if (ret)
1916 return 0;
1917 else
1918 return 1;
1919 }
1920 else
1921 return ret < 0 ? ret : 0;
1922}
1923
1924static long vma_needs_reservation(struct hstate *h,
1925 struct vm_area_struct *vma, unsigned long addr)
1926{
1927 return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV);
1928}
1929
1930static long vma_commit_reservation(struct hstate *h,
1931 struct vm_area_struct *vma, unsigned long addr)
1932{
1933 return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV);
1934}
1935
1936static void vma_end_reservation(struct hstate *h,
1937 struct vm_area_struct *vma, unsigned long addr)
1938{
1939 (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV);
1940}
1941
1942static long vma_add_reservation(struct hstate *h,
1943 struct vm_area_struct *vma, unsigned long addr)
1944{
1945 return __vma_reservation_common(h, vma, addr, VMA_ADD_RESV);
1946}
1947
1948/*
1949 * This routine is called to restore a reservation on error paths. In the
1950 * specific error paths, a huge page was allocated (via alloc_huge_page)
1951 * and is about to be freed. If a reservation for the page existed,
1952 * alloc_huge_page would have consumed the reservation and set PagePrivate
1953 * in the newly allocated page. When the page is freed via free_huge_page,
1954 * the global reservation count will be incremented if PagePrivate is set.
1955 * However, free_huge_page can not adjust the reserve map. Adjust the
1956 * reserve map here to be consistent with global reserve count adjustments
1957 * to be made by free_huge_page.
1958 */
1959static void restore_reserve_on_error(struct hstate *h,
1960 struct vm_area_struct *vma, unsigned long address,
1961 struct page *page)
1962{
1963 if (unlikely(PagePrivate(page))) {
1964 long rc = vma_needs_reservation(h, vma, address);
1965
1966 if (unlikely(rc < 0)) {
1967 /*
1968 * Rare out of memory condition in reserve map
1969 * manipulation. Clear PagePrivate so that
1970 * global reserve count will not be incremented
1971 * by free_huge_page. This will make it appear
1972 * as though the reservation for this page was
1973 * consumed. This may prevent the task from
1974 * faulting in the page at a later time. This
1975 * is better than inconsistent global huge page
1976 * accounting of reserve counts.
1977 */
1978 ClearPagePrivate(page);
1979 } else if (rc) {
1980 rc = vma_add_reservation(h, vma, address);
1981 if (unlikely(rc < 0))
1982 /*
1983 * See above comment about rare out of
1984 * memory condition.
1985 */
1986 ClearPagePrivate(page);
1987 } else
1988 vma_end_reservation(h, vma, address);
1989 }
1990}
1991
1992struct page *alloc_huge_page(struct vm_area_struct *vma,
1993 unsigned long addr, int avoid_reserve)
1994{
1995 struct hugepage_subpool *spool = subpool_vma(vma);
1996 struct hstate *h = hstate_vma(vma);
1997 struct page *page;
1998 long map_chg, map_commit;
1999 long gbl_chg;
2000 int ret, idx;
2001 struct hugetlb_cgroup *h_cg;
2002
2003 idx = hstate_index(h);
2004 /*
2005 * Examine the region/reserve map to determine if the process
2006 * has a reservation for the page to be allocated. A return
2007 * code of zero indicates a reservation exists (no change).
2008 */
2009 map_chg = gbl_chg = vma_needs_reservation(h, vma, addr);
2010 if (map_chg < 0)
2011 return ERR_PTR(-ENOMEM);
2012
2013 /*
2014 * Processes that did not create the mapping will have no
2015 * reserves as indicated by the region/reserve map. Check
2016 * that the allocation will not exceed the subpool limit.
2017 * Allocations for MAP_NORESERVE mappings also need to be
2018 * checked against any subpool limit.
2019 */
2020 if (map_chg || avoid_reserve) {
2021 gbl_chg = hugepage_subpool_get_pages(spool, 1);
2022 if (gbl_chg < 0) {
2023 vma_end_reservation(h, vma, addr);
2024 return ERR_PTR(-ENOSPC);
2025 }
2026
2027 /*
2028 * Even though there was no reservation in the region/reserve
2029 * map, there could be reservations associated with the
2030 * subpool that can be used. This would be indicated if the
2031 * return value of hugepage_subpool_get_pages() is zero.
2032 * However, if avoid_reserve is specified we still avoid even
2033 * the subpool reservations.
2034 */
2035 if (avoid_reserve)
2036 gbl_chg = 1;
2037 }
2038
2039 ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg);
2040 if (ret)
2041 goto out_subpool_put;
2042
2043 spin_lock(&hugetlb_lock);
2044 /*
2045 * glb_chg is passed to indicate whether or not a page must be taken
2046 * from the global free pool (global change). gbl_chg == 0 indicates
2047 * a reservation exists for the allocation.
2048 */
2049 page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg);
2050 if (!page) {
2051 spin_unlock(&hugetlb_lock);
2052 page = alloc_buddy_huge_page_with_mpol(h, vma, addr);
2053 if (!page)
2054 goto out_uncharge_cgroup;
2055 if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) {
2056 SetPagePrivate(page);
2057 h->resv_huge_pages--;
2058 }
2059 spin_lock(&hugetlb_lock);
2060 list_move(&page->lru, &h->hugepage_activelist);
2061 /* Fall through */
2062 }
2063 hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page);
2064 spin_unlock(&hugetlb_lock);
2065
2066 set_page_private(page, (unsigned long)spool);
2067
2068 map_commit = vma_commit_reservation(h, vma, addr);
2069 if (unlikely(map_chg > map_commit)) {
2070 /*
2071 * The page was added to the reservation map between
2072 * vma_needs_reservation and vma_commit_reservation.
2073 * This indicates a race with hugetlb_reserve_pages.
2074 * Adjust for the subpool count incremented above AND
2075 * in hugetlb_reserve_pages for the same page. Also,
2076 * the reservation count added in hugetlb_reserve_pages
2077 * no longer applies.
2078 */
2079 long rsv_adjust;
2080
2081 rsv_adjust = hugepage_subpool_put_pages(spool, 1);
2082 hugetlb_acct_memory(h, -rsv_adjust);
2083 }
2084 return page;
2085
2086out_uncharge_cgroup:
2087 hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg);
2088out_subpool_put:
2089 if (map_chg || avoid_reserve)
2090 hugepage_subpool_put_pages(spool, 1);
2091 vma_end_reservation(h, vma, addr);
2092 return ERR_PTR(-ENOSPC);
2093}
2094
2095int alloc_bootmem_huge_page(struct hstate *h)
2096 __attribute__ ((weak, alias("__alloc_bootmem_huge_page")));
2097int __alloc_bootmem_huge_page(struct hstate *h)
2098{
2099 struct huge_bootmem_page *m;
2100 int nr_nodes, node;
2101
2102 for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) {
2103 void *addr;
2104
2105 addr = memblock_alloc_try_nid_raw(
2106 huge_page_size(h), huge_page_size(h),
2107 0, MEMBLOCK_ALLOC_ACCESSIBLE, node);
2108 if (addr) {
2109 /*
2110 * Use the beginning of the huge page to store the
2111 * huge_bootmem_page struct (until gather_bootmem
2112 * puts them into the mem_map).
2113 */
2114 m = addr;
2115 goto found;
2116 }
2117 }
2118 return 0;
2119
2120found:
2121 BUG_ON(!IS_ALIGNED(virt_to_phys(m), huge_page_size(h)));
2122 /* Put them into a private list first because mem_map is not up yet */
2123 INIT_LIST_HEAD(&m->list);
2124 list_add(&m->list, &huge_boot_pages);
2125 m->hstate = h;
2126 return 1;
2127}
2128
2129static void __init prep_compound_huge_page(struct page *page,
2130 unsigned int order)
2131{
2132 if (unlikely(order > (MAX_ORDER - 1)))
2133 prep_compound_gigantic_page(page, order);
2134 else
2135 prep_compound_page(page, order);
2136}
2137
2138/* Put bootmem huge pages into the standard lists after mem_map is up */
2139static void __init gather_bootmem_prealloc(void)
2140{
2141 struct huge_bootmem_page *m;
2142
2143 list_for_each_entry(m, &huge_boot_pages, list) {
2144 struct page *page = virt_to_page(m);
2145 struct hstate *h = m->hstate;
2146
2147 WARN_ON(page_count(page) != 1);
2148 prep_compound_huge_page(page, h->order);
2149 WARN_ON(PageReserved(page));
2150 prep_new_huge_page(h, page, page_to_nid(page));
2151 put_page(page); /* free it into the hugepage allocator */
2152
2153 /*
2154 * If we had gigantic hugepages allocated at boot time, we need
2155 * to restore the 'stolen' pages to totalram_pages in order to
2156 * fix confusing memory reports from free(1) and another
2157 * side-effects, like CommitLimit going negative.
2158 */
2159 if (hstate_is_gigantic(h))
2160 adjust_managed_page_count(page, 1 << h->order);
2161 cond_resched();
2162 }
2163}
2164
2165static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
2166{
2167 unsigned long i;
2168
2169 for (i = 0; i < h->max_huge_pages; ++i) {
2170 if (hstate_is_gigantic(h)) {
2171 if (!alloc_bootmem_huge_page(h))
2172 break;
2173 } else if (!alloc_pool_huge_page(h,
2174 &node_states[N_MEMORY]))
2175 break;
2176 cond_resched();
2177 }
2178 if (i < h->max_huge_pages) {
2179 char buf[32];
2180
2181 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
2182 pr_warn("HugeTLB: allocating %lu of page size %s failed. Only allocated %lu hugepages.\n",
2183 h->max_huge_pages, buf, i);
2184 h->max_huge_pages = i;
2185 }
2186}
2187
2188static void __init hugetlb_init_hstates(void)
2189{
2190 struct hstate *h;
2191
2192 for_each_hstate(h) {
2193 if (minimum_order > huge_page_order(h))
2194 minimum_order = huge_page_order(h);
2195
2196 /* oversize hugepages were init'ed in early boot */
2197 if (!hstate_is_gigantic(h))
2198 hugetlb_hstate_alloc_pages(h);
2199 }
2200 VM_BUG_ON(minimum_order == UINT_MAX);
2201}
2202
2203static void __init report_hugepages(void)
2204{
2205 struct hstate *h;
2206
2207 for_each_hstate(h) {
2208 char buf[32];
2209
2210 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
2211 pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n",
2212 buf, h->free_huge_pages);
2213 }
2214}
2215
2216#ifdef CONFIG_HIGHMEM
2217static void try_to_free_low(struct hstate *h, unsigned long count,
2218 nodemask_t *nodes_allowed)
2219{
2220 int i;
2221
2222 if (hstate_is_gigantic(h))
2223 return;
2224
2225 for_each_node_mask(i, *nodes_allowed) {
2226 struct page *page, *next;
2227 struct list_head *freel = &h->hugepage_freelists[i];
2228 list_for_each_entry_safe(page, next, freel, lru) {
2229 if (count >= h->nr_huge_pages)
2230 return;
2231 if (PageHighMem(page))
2232 continue;
2233 list_del(&page->lru);
2234 update_and_free_page(h, page);
2235 h->free_huge_pages--;
2236 h->free_huge_pages_node[page_to_nid(page)]--;
2237 }
2238 }
2239}
2240#else
2241static inline void try_to_free_low(struct hstate *h, unsigned long count,
2242 nodemask_t *nodes_allowed)
2243{
2244}
2245#endif
2246
2247/*
2248 * Increment or decrement surplus_huge_pages. Keep node-specific counters
2249 * balanced by operating on them in a round-robin fashion.
2250 * Returns 1 if an adjustment was made.
2251 */
2252static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
2253 int delta)
2254{
2255 int nr_nodes, node;
2256
2257 VM_BUG_ON(delta != -1 && delta != 1);
2258
2259 if (delta < 0) {
2260 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
2261 if (h->surplus_huge_pages_node[node])
2262 goto found;
2263 }
2264 } else {
2265 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
2266 if (h->surplus_huge_pages_node[node] <
2267 h->nr_huge_pages_node[node])
2268 goto found;
2269 }
2270 }
2271 return 0;
2272
2273found:
2274 h->surplus_huge_pages += delta;
2275 h->surplus_huge_pages_node[node] += delta;
2276 return 1;
2277}
2278
2279#define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
2280static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count,
2281 nodemask_t *nodes_allowed)
2282{
2283 unsigned long min_count, ret;
2284
2285 if (hstate_is_gigantic(h) && !gigantic_page_supported())
2286 return h->max_huge_pages;
2287
2288 /*
2289 * Increase the pool size
2290 * First take pages out of surplus state. Then make up the
2291 * remaining difference by allocating fresh huge pages.
2292 *
2293 * We might race with alloc_surplus_huge_page() here and be unable
2294 * to convert a surplus huge page to a normal huge page. That is
2295 * not critical, though, it just means the overall size of the
2296 * pool might be one hugepage larger than it needs to be, but
2297 * within all the constraints specified by the sysctls.
2298 */
2299 spin_lock(&hugetlb_lock);
2300 while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
2301 if (!adjust_pool_surplus(h, nodes_allowed, -1))
2302 break;
2303 }
2304
2305 while (count > persistent_huge_pages(h)) {
2306 /*
2307 * If this allocation races such that we no longer need the
2308 * page, free_huge_page will handle it by freeing the page
2309 * and reducing the surplus.
2310 */
2311 spin_unlock(&hugetlb_lock);
2312
2313 /* yield cpu to avoid soft lockup */
2314 cond_resched();
2315
2316 ret = alloc_pool_huge_page(h, nodes_allowed);
2317 spin_lock(&hugetlb_lock);
2318 if (!ret)
2319 goto out;
2320
2321 /* Bail for signals. Probably ctrl-c from user */
2322 if (signal_pending(current))
2323 goto out;
2324 }
2325
2326 /*
2327 * Decrease the pool size
2328 * First return free pages to the buddy allocator (being careful
2329 * to keep enough around to satisfy reservations). Then place
2330 * pages into surplus state as needed so the pool will shrink
2331 * to the desired size as pages become free.
2332 *
2333 * By placing pages into the surplus state independent of the
2334 * overcommit value, we are allowing the surplus pool size to
2335 * exceed overcommit. There are few sane options here. Since
2336 * alloc_surplus_huge_page() is checking the global counter,
2337 * though, we'll note that we're not allowed to exceed surplus
2338 * and won't grow the pool anywhere else. Not until one of the
2339 * sysctls are changed, or the surplus pages go out of use.
2340 */
2341 min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
2342 min_count = max(count, min_count);
2343 try_to_free_low(h, min_count, nodes_allowed);
2344 while (min_count < persistent_huge_pages(h)) {
2345 if (!free_pool_huge_page(h, nodes_allowed, 0))
2346 break;
2347 cond_resched_lock(&hugetlb_lock);
2348 }
2349 while (count < persistent_huge_pages(h)) {
2350 if (!adjust_pool_surplus(h, nodes_allowed, 1))
2351 break;
2352 }
2353out:
2354 ret = persistent_huge_pages(h);
2355 spin_unlock(&hugetlb_lock);
2356 return ret;
2357}
2358
2359#define HSTATE_ATTR_RO(_name) \
2360 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
2361
2362#define HSTATE_ATTR(_name) \
2363 static struct kobj_attribute _name##_attr = \
2364 __ATTR(_name, 0644, _name##_show, _name##_store)
2365
2366static struct kobject *hugepages_kobj;
2367static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
2368
2369static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
2370
2371static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
2372{
2373 int i;
2374
2375 for (i = 0; i < HUGE_MAX_HSTATE; i++)
2376 if (hstate_kobjs[i] == kobj) {
2377 if (nidp)
2378 *nidp = NUMA_NO_NODE;
2379 return &hstates[i];
2380 }
2381
2382 return kobj_to_node_hstate(kobj, nidp);
2383}
2384
2385static ssize_t nr_hugepages_show_common(struct kobject *kobj,
2386 struct kobj_attribute *attr, char *buf)
2387{
2388 struct hstate *h;
2389 unsigned long nr_huge_pages;
2390 int nid;
2391
2392 h = kobj_to_hstate(kobj, &nid);
2393 if (nid == NUMA_NO_NODE)
2394 nr_huge_pages = h->nr_huge_pages;
2395 else
2396 nr_huge_pages = h->nr_huge_pages_node[nid];
2397
2398 return sprintf(buf, "%lu\n", nr_huge_pages);
2399}
2400
2401static ssize_t __nr_hugepages_store_common(bool obey_mempolicy,
2402 struct hstate *h, int nid,
2403 unsigned long count, size_t len)
2404{
2405 int err;
2406 NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY);
2407
2408 if (hstate_is_gigantic(h) && !gigantic_page_supported()) {
2409 err = -EINVAL;
2410 goto out;
2411 }
2412
2413 if (nid == NUMA_NO_NODE) {
2414 /*
2415 * global hstate attribute
2416 */
2417 if (!(obey_mempolicy &&
2418 init_nodemask_of_mempolicy(nodes_allowed))) {
2419 NODEMASK_FREE(nodes_allowed);
2420 nodes_allowed = &node_states[N_MEMORY];
2421 }
2422 } else if (nodes_allowed) {
2423 /*
2424 * per node hstate attribute: adjust count to global,
2425 * but restrict alloc/free to the specified node.
2426 */
2427 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
2428 init_nodemask_of_node(nodes_allowed, nid);
2429 } else
2430 nodes_allowed = &node_states[N_MEMORY];
2431
2432 h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed);
2433
2434 if (nodes_allowed != &node_states[N_MEMORY])
2435 NODEMASK_FREE(nodes_allowed);
2436
2437 return len;
2438out:
2439 NODEMASK_FREE(nodes_allowed);
2440 return err;
2441}
2442
2443static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
2444 struct kobject *kobj, const char *buf,
2445 size_t len)
2446{
2447 struct hstate *h;
2448 unsigned long count;
2449 int nid;
2450 int err;
2451
2452 err = kstrtoul(buf, 10, &count);
2453 if (err)
2454 return err;
2455
2456 h = kobj_to_hstate(kobj, &nid);
2457 return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len);
2458}
2459
2460static ssize_t nr_hugepages_show(struct kobject *kobj,
2461 struct kobj_attribute *attr, char *buf)
2462{
2463 return nr_hugepages_show_common(kobj, attr, buf);
2464}
2465
2466static ssize_t nr_hugepages_store(struct kobject *kobj,
2467 struct kobj_attribute *attr, const char *buf, size_t len)
2468{
2469 return nr_hugepages_store_common(false, kobj, buf, len);
2470}
2471HSTATE_ATTR(nr_hugepages);
2472
2473#ifdef CONFIG_NUMA
2474
2475/*
2476 * hstate attribute for optionally mempolicy-based constraint on persistent
2477 * huge page alloc/free.
2478 */
2479static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
2480 struct kobj_attribute *attr, char *buf)
2481{
2482 return nr_hugepages_show_common(kobj, attr, buf);
2483}
2484
2485static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
2486 struct kobj_attribute *attr, const char *buf, size_t len)
2487{
2488 return nr_hugepages_store_common(true, kobj, buf, len);
2489}
2490HSTATE_ATTR(nr_hugepages_mempolicy);
2491#endif
2492
2493
2494static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
2495 struct kobj_attribute *attr, char *buf)
2496{
2497 struct hstate *h = kobj_to_hstate(kobj, NULL);
2498 return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
2499}
2500
2501static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
2502 struct kobj_attribute *attr, const char *buf, size_t count)
2503{
2504 int err;
2505 unsigned long input;
2506 struct hstate *h = kobj_to_hstate(kobj, NULL);
2507
2508 if (hstate_is_gigantic(h))
2509 return -EINVAL;
2510
2511 err = kstrtoul(buf, 10, &input);
2512 if (err)
2513 return err;
2514
2515 spin_lock(&hugetlb_lock);
2516 h->nr_overcommit_huge_pages = input;
2517 spin_unlock(&hugetlb_lock);
2518
2519 return count;
2520}
2521HSTATE_ATTR(nr_overcommit_hugepages);
2522
2523static ssize_t free_hugepages_show(struct kobject *kobj,
2524 struct kobj_attribute *attr, char *buf)
2525{
2526 struct hstate *h;
2527 unsigned long free_huge_pages;
2528 int nid;
2529
2530 h = kobj_to_hstate(kobj, &nid);
2531 if (nid == NUMA_NO_NODE)
2532 free_huge_pages = h->free_huge_pages;
2533 else
2534 free_huge_pages = h->free_huge_pages_node[nid];
2535
2536 return sprintf(buf, "%lu\n", free_huge_pages);
2537}
2538HSTATE_ATTR_RO(free_hugepages);
2539
2540static ssize_t resv_hugepages_show(struct kobject *kobj,
2541 struct kobj_attribute *attr, char *buf)
2542{
2543 struct hstate *h = kobj_to_hstate(kobj, NULL);
2544 return sprintf(buf, "%lu\n", h->resv_huge_pages);
2545}
2546HSTATE_ATTR_RO(resv_hugepages);
2547
2548static ssize_t surplus_hugepages_show(struct kobject *kobj,
2549 struct kobj_attribute *attr, char *buf)
2550{
2551 struct hstate *h;
2552 unsigned long surplus_huge_pages;
2553 int nid;
2554
2555 h = kobj_to_hstate(kobj, &nid);
2556 if (nid == NUMA_NO_NODE)
2557 surplus_huge_pages = h->surplus_huge_pages;
2558 else
2559 surplus_huge_pages = h->surplus_huge_pages_node[nid];
2560
2561 return sprintf(buf, "%lu\n", surplus_huge_pages);
2562}
2563HSTATE_ATTR_RO(surplus_hugepages);
2564
2565static struct attribute *hstate_attrs[] = {
2566 &nr_hugepages_attr.attr,
2567 &nr_overcommit_hugepages_attr.attr,
2568 &free_hugepages_attr.attr,
2569 &resv_hugepages_attr.attr,
2570 &surplus_hugepages_attr.attr,
2571#ifdef CONFIG_NUMA
2572 &nr_hugepages_mempolicy_attr.attr,
2573#endif
2574 NULL,
2575};
2576
2577static const struct attribute_group hstate_attr_group = {
2578 .attrs = hstate_attrs,
2579};
2580
2581static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
2582 struct kobject **hstate_kobjs,
2583 const struct attribute_group *hstate_attr_group)
2584{
2585 int retval;
2586 int hi = hstate_index(h);
2587
2588 hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
2589 if (!hstate_kobjs[hi])
2590 return -ENOMEM;
2591
2592 retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
2593 if (retval)
2594 kobject_put(hstate_kobjs[hi]);
2595
2596 return retval;
2597}
2598
2599static void __init hugetlb_sysfs_init(void)
2600{
2601 struct hstate *h;
2602 int err;
2603
2604 hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
2605 if (!hugepages_kobj)
2606 return;
2607
2608 for_each_hstate(h) {
2609 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
2610 hstate_kobjs, &hstate_attr_group);
2611 if (err)
2612 pr_err("Hugetlb: Unable to add hstate %s", h->name);
2613 }
2614}
2615
2616#ifdef CONFIG_NUMA
2617
2618/*
2619 * node_hstate/s - associate per node hstate attributes, via their kobjects,
2620 * with node devices in node_devices[] using a parallel array. The array
2621 * index of a node device or _hstate == node id.
2622 * This is here to avoid any static dependency of the node device driver, in
2623 * the base kernel, on the hugetlb module.
2624 */
2625struct node_hstate {
2626 struct kobject *hugepages_kobj;
2627 struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
2628};
2629static struct node_hstate node_hstates[MAX_NUMNODES];
2630
2631/*
2632 * A subset of global hstate attributes for node devices
2633 */
2634static struct attribute *per_node_hstate_attrs[] = {
2635 &nr_hugepages_attr.attr,
2636 &free_hugepages_attr.attr,
2637 &surplus_hugepages_attr.attr,
2638 NULL,
2639};
2640
2641static const struct attribute_group per_node_hstate_attr_group = {
2642 .attrs = per_node_hstate_attrs,
2643};
2644
2645/*
2646 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
2647 * Returns node id via non-NULL nidp.
2648 */
2649static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
2650{
2651 int nid;
2652
2653 for (nid = 0; nid < nr_node_ids; nid++) {
2654 struct node_hstate *nhs = &node_hstates[nid];
2655 int i;
2656 for (i = 0; i < HUGE_MAX_HSTATE; i++)
2657 if (nhs->hstate_kobjs[i] == kobj) {
2658 if (nidp)
2659 *nidp = nid;
2660 return &hstates[i];
2661 }
2662 }
2663
2664 BUG();
2665 return NULL;
2666}
2667
2668/*
2669 * Unregister hstate attributes from a single node device.
2670 * No-op if no hstate attributes attached.
2671 */
2672static void hugetlb_unregister_node(struct node *node)
2673{
2674 struct hstate *h;
2675 struct node_hstate *nhs = &node_hstates[node->dev.id];
2676
2677 if (!nhs->hugepages_kobj)
2678 return; /* no hstate attributes */
2679
2680 for_each_hstate(h) {
2681 int idx = hstate_index(h);
2682 if (nhs->hstate_kobjs[idx]) {
2683 kobject_put(nhs->hstate_kobjs[idx]);
2684 nhs->hstate_kobjs[idx] = NULL;
2685 }
2686 }
2687
2688 kobject_put(nhs->hugepages_kobj);
2689 nhs->hugepages_kobj = NULL;
2690}
2691
2692
2693/*
2694 * Register hstate attributes for a single node device.
2695 * No-op if attributes already registered.
2696 */
2697static void hugetlb_register_node(struct node *node)
2698{
2699 struct hstate *h;
2700 struct node_hstate *nhs = &node_hstates[node->dev.id];
2701 int err;
2702
2703 if (nhs->hugepages_kobj)
2704 return; /* already allocated */
2705
2706 nhs->hugepages_kobj = kobject_create_and_add("hugepages",
2707 &node->dev.kobj);
2708 if (!nhs->hugepages_kobj)
2709 return;
2710
2711 for_each_hstate(h) {
2712 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
2713 nhs->hstate_kobjs,
2714 &per_node_hstate_attr_group);
2715 if (err) {
2716 pr_err("Hugetlb: Unable to add hstate %s for node %d\n",
2717 h->name, node->dev.id);
2718 hugetlb_unregister_node(node);
2719 break;
2720 }
2721 }
2722}
2723
2724/*
2725 * hugetlb init time: register hstate attributes for all registered node
2726 * devices of nodes that have memory. All on-line nodes should have
2727 * registered their associated device by this time.
2728 */
2729static void __init hugetlb_register_all_nodes(void)
2730{
2731 int nid;
2732
2733 for_each_node_state(nid, N_MEMORY) {
2734 struct node *node = node_devices[nid];
2735 if (node->dev.id == nid)
2736 hugetlb_register_node(node);
2737 }
2738
2739 /*
2740 * Let the node device driver know we're here so it can
2741 * [un]register hstate attributes on node hotplug.
2742 */
2743 register_hugetlbfs_with_node(hugetlb_register_node,
2744 hugetlb_unregister_node);
2745}
2746#else /* !CONFIG_NUMA */
2747
2748static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
2749{
2750 BUG();
2751 if (nidp)
2752 *nidp = -1;
2753 return NULL;
2754}
2755
2756static void hugetlb_register_all_nodes(void) { }
2757
2758#endif
2759
2760static int __init hugetlb_init(void)
2761{
2762 int i;
2763
2764 if (!hugepages_supported())
2765 return 0;
2766
2767 if (!size_to_hstate(default_hstate_size)) {
2768 if (default_hstate_size != 0) {
2769 pr_err("HugeTLB: unsupported default_hugepagesz %lu. Reverting to %lu\n",
2770 default_hstate_size, HPAGE_SIZE);
2771 }
2772
2773 default_hstate_size = HPAGE_SIZE;
2774 if (!size_to_hstate(default_hstate_size))
2775 hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
2776 }
2777 default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size));
2778 if (default_hstate_max_huge_pages) {
2779 if (!default_hstate.max_huge_pages)
2780 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
2781 }
2782
2783 hugetlb_init_hstates();
2784 gather_bootmem_prealloc();
2785 report_hugepages();
2786
2787 hugetlb_sysfs_init();
2788 hugetlb_register_all_nodes();
2789 hugetlb_cgroup_file_init();
2790
2791#ifdef CONFIG_SMP
2792 num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus());
2793#else
2794 num_fault_mutexes = 1;
2795#endif
2796 hugetlb_fault_mutex_table =
2797 kmalloc_array(num_fault_mutexes, sizeof(struct mutex),
2798 GFP_KERNEL);
2799 BUG_ON(!hugetlb_fault_mutex_table);
2800
2801 for (i = 0; i < num_fault_mutexes; i++)
2802 mutex_init(&hugetlb_fault_mutex_table[i]);
2803 return 0;
2804}
2805subsys_initcall(hugetlb_init);
2806
2807/* Should be called on processing a hugepagesz=... option */
2808void __init hugetlb_bad_size(void)
2809{
2810 parsed_valid_hugepagesz = false;
2811}
2812
2813void __init hugetlb_add_hstate(unsigned int order)
2814{
2815 struct hstate *h;
2816 unsigned long i;
2817
2818 if (size_to_hstate(PAGE_SIZE << order)) {
2819 pr_warn("hugepagesz= specified twice, ignoring\n");
2820 return;
2821 }
2822 BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE);
2823 BUG_ON(order == 0);
2824 h = &hstates[hugetlb_max_hstate++];
2825 h->order = order;
2826 h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
2827 h->nr_huge_pages = 0;
2828 h->free_huge_pages = 0;
2829 for (i = 0; i < MAX_NUMNODES; ++i)
2830 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
2831 INIT_LIST_HEAD(&h->hugepage_activelist);
2832 h->next_nid_to_alloc = first_memory_node;
2833 h->next_nid_to_free = first_memory_node;
2834 snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
2835 huge_page_size(h)/1024);
2836
2837 parsed_hstate = h;
2838}
2839
2840static int __init hugetlb_nrpages_setup(char *s)
2841{
2842 unsigned long *mhp;
2843 static unsigned long *last_mhp;
2844
2845 if (!parsed_valid_hugepagesz) {
2846 pr_warn("hugepages = %s preceded by "
2847 "an unsupported hugepagesz, ignoring\n", s);
2848 parsed_valid_hugepagesz = true;
2849 return 1;
2850 }
2851 /*
2852 * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet,
2853 * so this hugepages= parameter goes to the "default hstate".
2854 */
2855 else if (!hugetlb_max_hstate)
2856 mhp = &default_hstate_max_huge_pages;
2857 else
2858 mhp = &parsed_hstate->max_huge_pages;
2859
2860 if (mhp == last_mhp) {
2861 pr_warn("hugepages= specified twice without interleaving hugepagesz=, ignoring\n");
2862 return 1;
2863 }
2864
2865 if (sscanf(s, "%lu", mhp) <= 0)
2866 *mhp = 0;
2867
2868 /*
2869 * Global state is always initialized later in hugetlb_init.
2870 * But we need to allocate >= MAX_ORDER hstates here early to still
2871 * use the bootmem allocator.
2872 */
2873 if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER)
2874 hugetlb_hstate_alloc_pages(parsed_hstate);
2875
2876 last_mhp = mhp;
2877
2878 return 1;
2879}
2880__setup("hugepages=", hugetlb_nrpages_setup);
2881
2882static int __init hugetlb_default_setup(char *s)
2883{
2884 default_hstate_size = memparse(s, &s);
2885 return 1;
2886}
2887__setup("default_hugepagesz=", hugetlb_default_setup);
2888
2889static unsigned int cpuset_mems_nr(unsigned int *array)
2890{
2891 int node;
2892 unsigned int nr = 0;
2893
2894 for_each_node_mask(node, cpuset_current_mems_allowed)
2895 nr += array[node];
2896
2897 return nr;
2898}
2899
2900#ifdef CONFIG_SYSCTL
2901static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
2902 struct ctl_table *table, int write,
2903 void __user *buffer, size_t *length, loff_t *ppos)
2904{
2905 struct hstate *h = &default_hstate;
2906 unsigned long tmp = h->max_huge_pages;
2907 int ret;
2908
2909 if (!hugepages_supported())
2910 return -EOPNOTSUPP;
2911
2912 table->data = &tmp;
2913 table->maxlen = sizeof(unsigned long);
2914 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2915 if (ret)
2916 goto out;
2917
2918 if (write)
2919 ret = __nr_hugepages_store_common(obey_mempolicy, h,
2920 NUMA_NO_NODE, tmp, *length);
2921out:
2922 return ret;
2923}
2924
2925int hugetlb_sysctl_handler(struct ctl_table *table, int write,
2926 void __user *buffer, size_t *length, loff_t *ppos)
2927{
2928
2929 return hugetlb_sysctl_handler_common(false, table, write,
2930 buffer, length, ppos);
2931}
2932
2933#ifdef CONFIG_NUMA
2934int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
2935 void __user *buffer, size_t *length, loff_t *ppos)
2936{
2937 return hugetlb_sysctl_handler_common(true, table, write,
2938 buffer, length, ppos);
2939}
2940#endif /* CONFIG_NUMA */
2941
2942int hugetlb_overcommit_handler(struct ctl_table *table, int write,
2943 void __user *buffer,
2944 size_t *length, loff_t *ppos)
2945{
2946 struct hstate *h = &default_hstate;
2947 unsigned long tmp;
2948 int ret;
2949
2950 if (!hugepages_supported())
2951 return -EOPNOTSUPP;
2952
2953 tmp = h->nr_overcommit_huge_pages;
2954
2955 if (write && hstate_is_gigantic(h))
2956 return -EINVAL;
2957
2958 table->data = &tmp;
2959 table->maxlen = sizeof(unsigned long);
2960 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2961 if (ret)
2962 goto out;
2963
2964 if (write) {
2965 spin_lock(&hugetlb_lock);
2966 h->nr_overcommit_huge_pages = tmp;
2967 spin_unlock(&hugetlb_lock);
2968 }
2969out:
2970 return ret;
2971}
2972
2973#endif /* CONFIG_SYSCTL */
2974
2975void hugetlb_report_meminfo(struct seq_file *m)
2976{
2977 struct hstate *h;
2978 unsigned long total = 0;
2979
2980 if (!hugepages_supported())
2981 return;
2982
2983 for_each_hstate(h) {
2984 unsigned long count = h->nr_huge_pages;
2985
2986 total += (PAGE_SIZE << huge_page_order(h)) * count;
2987
2988 if (h == &default_hstate)
2989 seq_printf(m,
2990 "HugePages_Total: %5lu\n"
2991 "HugePages_Free: %5lu\n"
2992 "HugePages_Rsvd: %5lu\n"
2993 "HugePages_Surp: %5lu\n"
2994 "Hugepagesize: %8lu kB\n",
2995 count,
2996 h->free_huge_pages,
2997 h->resv_huge_pages,
2998 h->surplus_huge_pages,
2999 (PAGE_SIZE << huge_page_order(h)) / 1024);
3000 }
3001
3002 seq_printf(m, "Hugetlb: %8lu kB\n", total / 1024);
3003}
3004
3005int hugetlb_report_node_meminfo(int nid, char *buf)
3006{
3007 struct hstate *h = &default_hstate;
3008 if (!hugepages_supported())
3009 return 0;
3010 return sprintf(buf,
3011 "Node %d HugePages_Total: %5u\n"
3012 "Node %d HugePages_Free: %5u\n"
3013 "Node %d HugePages_Surp: %5u\n",
3014 nid, h->nr_huge_pages_node[nid],
3015 nid, h->free_huge_pages_node[nid],
3016 nid, h->surplus_huge_pages_node[nid]);
3017}
3018
3019void hugetlb_show_meminfo(void)
3020{
3021 struct hstate *h;
3022 int nid;
3023
3024 if (!hugepages_supported())
3025 return;
3026
3027 for_each_node_state(nid, N_MEMORY)
3028 for_each_hstate(h)
3029 pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
3030 nid,
3031 h->nr_huge_pages_node[nid],
3032 h->free_huge_pages_node[nid],
3033 h->surplus_huge_pages_node[nid],
3034 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
3035}
3036
3037void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm)
3038{
3039 seq_printf(m, "HugetlbPages:\t%8lu kB\n",
3040 atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10));
3041}
3042
3043/* Return the number pages of memory we physically have, in PAGE_SIZE units. */
3044unsigned long hugetlb_total_pages(void)
3045{
3046 struct hstate *h;
3047 unsigned long nr_total_pages = 0;
3048
3049 for_each_hstate(h)
3050 nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h);
3051 return nr_total_pages;
3052}
3053
3054static int hugetlb_acct_memory(struct hstate *h, long delta)
3055{
3056 int ret = -ENOMEM;
3057
3058 spin_lock(&hugetlb_lock);
3059 /*
3060 * When cpuset is configured, it breaks the strict hugetlb page
3061 * reservation as the accounting is done on a global variable. Such
3062 * reservation is completely rubbish in the presence of cpuset because
3063 * the reservation is not checked against page availability for the
3064 * current cpuset. Application can still potentially OOM'ed by kernel
3065 * with lack of free htlb page in cpuset that the task is in.
3066 * Attempt to enforce strict accounting with cpuset is almost
3067 * impossible (or too ugly) because cpuset is too fluid that
3068 * task or memory node can be dynamically moved between cpusets.
3069 *
3070 * The change of semantics for shared hugetlb mapping with cpuset is
3071 * undesirable. However, in order to preserve some of the semantics,
3072 * we fall back to check against current free page availability as
3073 * a best attempt and hopefully to minimize the impact of changing
3074 * semantics that cpuset has.
3075 */
3076 if (delta > 0) {
3077 if (gather_surplus_pages(h, delta) < 0)
3078 goto out;
3079
3080 if (delta > cpuset_mems_nr(h->free_huge_pages_node)) {
3081 return_unused_surplus_pages(h, delta);
3082 goto out;
3083 }
3084 }
3085
3086 ret = 0;
3087 if (delta < 0)
3088 return_unused_surplus_pages(h, (unsigned long) -delta);
3089
3090out:
3091 spin_unlock(&hugetlb_lock);
3092 return ret;
3093}
3094
3095static void hugetlb_vm_op_open(struct vm_area_struct *vma)
3096{
3097 struct resv_map *resv = vma_resv_map(vma);
3098
3099 /*
3100 * This new VMA should share its siblings reservation map if present.
3101 * The VMA will only ever have a valid reservation map pointer where
3102 * it is being copied for another still existing VMA. As that VMA
3103 * has a reference to the reservation map it cannot disappear until
3104 * after this open call completes. It is therefore safe to take a
3105 * new reference here without additional locking.
3106 */
3107 if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
3108 kref_get(&resv->refs);
3109}
3110
3111static void hugetlb_vm_op_close(struct vm_area_struct *vma)
3112{
3113 struct hstate *h = hstate_vma(vma);
3114 struct resv_map *resv = vma_resv_map(vma);
3115 struct hugepage_subpool *spool = subpool_vma(vma);
3116 unsigned long reserve, start, end;
3117 long gbl_reserve;
3118
3119 if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER))
3120 return;
3121
3122 start = vma_hugecache_offset(h, vma, vma->vm_start);
3123 end = vma_hugecache_offset(h, vma, vma->vm_end);
3124
3125 reserve = (end - start) - region_count(resv, start, end);
3126
3127 kref_put(&resv->refs, resv_map_release);
3128
3129 if (reserve) {
3130 /*
3131 * Decrement reserve counts. The global reserve count may be
3132 * adjusted if the subpool has a minimum size.
3133 */
3134 gbl_reserve = hugepage_subpool_put_pages(spool, reserve);
3135 hugetlb_acct_memory(h, -gbl_reserve);
3136 }
3137}
3138
3139static int hugetlb_vm_op_split(struct vm_area_struct *vma, unsigned long addr)
3140{
3141 if (addr & ~(huge_page_mask(hstate_vma(vma))))
3142 return -EINVAL;
3143 return 0;
3144}
3145
3146static unsigned long hugetlb_vm_op_pagesize(struct vm_area_struct *vma)
3147{
3148 struct hstate *hstate = hstate_vma(vma);
3149
3150 return 1UL << huge_page_shift(hstate);
3151}
3152
3153/*
3154 * We cannot handle pagefaults against hugetlb pages at all. They cause
3155 * handle_mm_fault() to try to instantiate regular-sized pages in the
3156 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
3157 * this far.
3158 */
3159static vm_fault_t hugetlb_vm_op_fault(struct vm_fault *vmf)
3160{
3161 BUG();
3162 return 0;
3163}
3164
3165/*
3166 * When a new function is introduced to vm_operations_struct and added
3167 * to hugetlb_vm_ops, please consider adding the function to shm_vm_ops.
3168 * This is because under System V memory model, mappings created via
3169 * shmget/shmat with "huge page" specified are backed by hugetlbfs files,
3170 * their original vm_ops are overwritten with shm_vm_ops.
3171 */
3172const struct vm_operations_struct hugetlb_vm_ops = {
3173 .fault = hugetlb_vm_op_fault,
3174 .open = hugetlb_vm_op_open,
3175 .close = hugetlb_vm_op_close,
3176 .split = hugetlb_vm_op_split,
3177 .pagesize = hugetlb_vm_op_pagesize,
3178};
3179
3180static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
3181 int writable)
3182{
3183 pte_t entry;
3184
3185 if (writable) {
3186 entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page,
3187 vma->vm_page_prot)));
3188 } else {
3189 entry = huge_pte_wrprotect(mk_huge_pte(page,
3190 vma->vm_page_prot));
3191 }
3192 entry = pte_mkyoung(entry);
3193 entry = pte_mkhuge(entry);
3194 entry = arch_make_huge_pte(entry, vma, page, writable);
3195
3196 return entry;
3197}
3198
3199static void set_huge_ptep_writable(struct vm_area_struct *vma,
3200 unsigned long address, pte_t *ptep)
3201{
3202 pte_t entry;
3203
3204 entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep)));
3205 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
3206 update_mmu_cache(vma, address, ptep);
3207}
3208
3209bool is_hugetlb_entry_migration(pte_t pte)
3210{
3211 swp_entry_t swp;
3212
3213 if (huge_pte_none(pte) || pte_present(pte))
3214 return false;
3215 swp = pte_to_swp_entry(pte);
3216 if (non_swap_entry(swp) && is_migration_entry(swp))
3217 return true;
3218 else
3219 return false;
3220}
3221
3222static int is_hugetlb_entry_hwpoisoned(pte_t pte)
3223{
3224 swp_entry_t swp;
3225
3226 if (huge_pte_none(pte) || pte_present(pte))
3227 return 0;
3228 swp = pte_to_swp_entry(pte);
3229 if (non_swap_entry(swp) && is_hwpoison_entry(swp))
3230 return 1;
3231 else
3232 return 0;
3233}
3234
3235int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
3236 struct vm_area_struct *vma)
3237{
3238 pte_t *src_pte, *dst_pte, entry, dst_entry;
3239 struct page *ptepage;
3240 unsigned long addr;
3241 int cow;
3242 struct hstate *h = hstate_vma(vma);
3243 unsigned long sz = huge_page_size(h);
3244 struct mmu_notifier_range range;
3245 int ret = 0;
3246
3247 cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
3248
3249 if (cow) {
3250 mmu_notifier_range_init(&range, src, vma->vm_start,
3251 vma->vm_end);
3252 mmu_notifier_invalidate_range_start(&range);
3253 }
3254
3255 for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
3256 spinlock_t *src_ptl, *dst_ptl;
3257 src_pte = huge_pte_offset(src, addr, sz);
3258 if (!src_pte)
3259 continue;
3260 dst_pte = huge_pte_alloc(dst, addr, sz);
3261 if (!dst_pte) {
3262 ret = -ENOMEM;
3263 break;
3264 }
3265
3266 /*
3267 * If the pagetables are shared don't copy or take references.
3268 * dst_pte == src_pte is the common case of src/dest sharing.
3269 *
3270 * However, src could have 'unshared' and dst shares with
3271 * another vma. If dst_pte !none, this implies sharing.
3272 * Check here before taking page table lock, and once again
3273 * after taking the lock below.
3274 */
3275 dst_entry = huge_ptep_get(dst_pte);
3276 if ((dst_pte == src_pte) || !huge_pte_none(dst_entry))
3277 continue;
3278
3279 dst_ptl = huge_pte_lock(h, dst, dst_pte);
3280 src_ptl = huge_pte_lockptr(h, src, src_pte);
3281 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
3282 entry = huge_ptep_get(src_pte);
3283 dst_entry = huge_ptep_get(dst_pte);
3284 if (huge_pte_none(entry) || !huge_pte_none(dst_entry)) {
3285 /*
3286 * Skip if src entry none. Also, skip in the
3287 * unlikely case dst entry !none as this implies
3288 * sharing with another vma.
3289 */
3290 ;
3291 } else if (unlikely(is_hugetlb_entry_migration(entry) ||
3292 is_hugetlb_entry_hwpoisoned(entry))) {
3293 swp_entry_t swp_entry = pte_to_swp_entry(entry);
3294
3295 if (is_write_migration_entry(swp_entry) && cow) {
3296 /*
3297 * COW mappings require pages in both
3298 * parent and child to be set to read.
3299 */
3300 make_migration_entry_read(&swp_entry);
3301 entry = swp_entry_to_pte(swp_entry);
3302 set_huge_swap_pte_at(src, addr, src_pte,
3303 entry, sz);
3304 }
3305 set_huge_swap_pte_at(dst, addr, dst_pte, entry, sz);
3306 } else {
3307 if (cow) {
3308 /*
3309 * No need to notify as we are downgrading page
3310 * table protection not changing it to point
3311 * to a new page.
3312 *
3313 * See Documentation/vm/mmu_notifier.rst
3314 */
3315 huge_ptep_set_wrprotect(src, addr, src_pte);
3316 }
3317 entry = huge_ptep_get(src_pte);
3318 ptepage = pte_page(entry);
3319 get_page(ptepage);
3320 page_dup_rmap(ptepage, true);
3321 set_huge_pte_at(dst, addr, dst_pte, entry);
3322 hugetlb_count_add(pages_per_huge_page(h), dst);
3323 }
3324 spin_unlock(src_ptl);
3325 spin_unlock(dst_ptl);
3326 }
3327
3328 if (cow)
3329 mmu_notifier_invalidate_range_end(&range);
3330
3331 return ret;
3332}
3333
3334void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
3335 unsigned long start, unsigned long end,
3336 struct page *ref_page)
3337{
3338 struct mm_struct *mm = vma->vm_mm;
3339 unsigned long address;
3340 pte_t *ptep;
3341 pte_t pte;
3342 spinlock_t *ptl;
3343 struct page *page;
3344 struct hstate *h = hstate_vma(vma);
3345 unsigned long sz = huge_page_size(h);
3346 struct mmu_notifier_range range;
3347
3348 WARN_ON(!is_vm_hugetlb_page(vma));
3349 BUG_ON(start & ~huge_page_mask(h));
3350 BUG_ON(end & ~huge_page_mask(h));
3351
3352 /*
3353 * This is a hugetlb vma, all the pte entries should point
3354 * to huge page.
3355 */
3356 tlb_remove_check_page_size_change(tlb, sz);
3357 tlb_start_vma(tlb, vma);
3358
3359 /*
3360 * If sharing possible, alert mmu notifiers of worst case.
3361 */
3362 mmu_notifier_range_init(&range, mm, start, end);
3363 adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end);
3364 mmu_notifier_invalidate_range_start(&range);
3365 address = start;
3366 for (; address < end; address += sz) {
3367 ptep = huge_pte_offset(mm, address, sz);
3368 if (!ptep)
3369 continue;
3370
3371 ptl = huge_pte_lock(h, mm, ptep);
3372 if (huge_pmd_unshare(mm, &address, ptep)) {
3373 spin_unlock(ptl);
3374 /*
3375 * We just unmapped a page of PMDs by clearing a PUD.
3376 * The caller's TLB flush range should cover this area.
3377 */
3378 continue;
3379 }
3380
3381 pte = huge_ptep_get(ptep);
3382 if (huge_pte_none(pte)) {
3383 spin_unlock(ptl);
3384 continue;
3385 }
3386
3387 /*
3388 * Migrating hugepage or HWPoisoned hugepage is already
3389 * unmapped and its refcount is dropped, so just clear pte here.
3390 */
3391 if (unlikely(!pte_present(pte))) {
3392 huge_pte_clear(mm, address, ptep, sz);
3393 spin_unlock(ptl);
3394 continue;
3395 }
3396
3397 page = pte_page(pte);
3398 /*
3399 * If a reference page is supplied, it is because a specific
3400 * page is being unmapped, not a range. Ensure the page we
3401 * are about to unmap is the actual page of interest.
3402 */
3403 if (ref_page) {
3404 if (page != ref_page) {
3405 spin_unlock(ptl);
3406 continue;
3407 }
3408 /*
3409 * Mark the VMA as having unmapped its page so that
3410 * future faults in this VMA will fail rather than
3411 * looking like data was lost
3412 */
3413 set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
3414 }
3415
3416 pte = huge_ptep_get_and_clear(mm, address, ptep);
3417 tlb_remove_huge_tlb_entry(h, tlb, ptep, address);
3418 if (huge_pte_dirty(pte))
3419 set_page_dirty(page);
3420
3421 hugetlb_count_sub(pages_per_huge_page(h), mm);
3422 page_remove_rmap(page, true);
3423
3424 spin_unlock(ptl);
3425 tlb_remove_page_size(tlb, page, huge_page_size(h));
3426 /*
3427 * Bail out after unmapping reference page if supplied
3428 */
3429 if (ref_page)
3430 break;
3431 }
3432 mmu_notifier_invalidate_range_end(&range);
3433 tlb_end_vma(tlb, vma);
3434}
3435
3436void __unmap_hugepage_range_final(struct mmu_gather *tlb,
3437 struct vm_area_struct *vma, unsigned long start,
3438 unsigned long end, struct page *ref_page)
3439{
3440 __unmap_hugepage_range(tlb, vma, start, end, ref_page);
3441
3442 /*
3443 * Clear this flag so that x86's huge_pmd_share page_table_shareable
3444 * test will fail on a vma being torn down, and not grab a page table
3445 * on its way out. We're lucky that the flag has such an appropriate
3446 * name, and can in fact be safely cleared here. We could clear it
3447 * before the __unmap_hugepage_range above, but all that's necessary
3448 * is to clear it before releasing the i_mmap_rwsem. This works
3449 * because in the context this is called, the VMA is about to be
3450 * destroyed and the i_mmap_rwsem is held.
3451 */
3452 vma->vm_flags &= ~VM_MAYSHARE;
3453}
3454
3455void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
3456 unsigned long end, struct page *ref_page)
3457{
3458 struct mm_struct *mm;
3459 struct mmu_gather tlb;
3460 unsigned long tlb_start = start;
3461 unsigned long tlb_end = end;
3462
3463 /*
3464 * If shared PMDs were possibly used within this vma range, adjust
3465 * start/end for worst case tlb flushing.
3466 * Note that we can not be sure if PMDs are shared until we try to
3467 * unmap pages. However, we want to make sure TLB flushing covers
3468 * the largest possible range.
3469 */
3470 adjust_range_if_pmd_sharing_possible(vma, &tlb_start, &tlb_end);
3471
3472 mm = vma->vm_mm;
3473
3474 tlb_gather_mmu(&tlb, mm, tlb_start, tlb_end);
3475 __unmap_hugepage_range(&tlb, vma, start, end, ref_page);
3476 tlb_finish_mmu(&tlb, tlb_start, tlb_end);
3477}
3478
3479/*
3480 * This is called when the original mapper is failing to COW a MAP_PRIVATE
3481 * mappping it owns the reserve page for. The intention is to unmap the page
3482 * from other VMAs and let the children be SIGKILLed if they are faulting the
3483 * same region.
3484 */
3485static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
3486 struct page *page, unsigned long address)
3487{
3488 struct hstate *h = hstate_vma(vma);
3489 struct vm_area_struct *iter_vma;
3490 struct address_space *mapping;
3491 pgoff_t pgoff;
3492
3493 /*
3494 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
3495 * from page cache lookup which is in HPAGE_SIZE units.
3496 */
3497 address = address & huge_page_mask(h);
3498 pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) +
3499 vma->vm_pgoff;
3500 mapping = vma->vm_file->f_mapping;
3501
3502 /*
3503 * Take the mapping lock for the duration of the table walk. As
3504 * this mapping should be shared between all the VMAs,
3505 * __unmap_hugepage_range() is called as the lock is already held
3506 */
3507 i_mmap_lock_write(mapping);
3508 vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) {
3509 /* Do not unmap the current VMA */
3510 if (iter_vma == vma)
3511 continue;
3512
3513 /*
3514 * Shared VMAs have their own reserves and do not affect
3515 * MAP_PRIVATE accounting but it is possible that a shared
3516 * VMA is using the same page so check and skip such VMAs.
3517 */
3518 if (iter_vma->vm_flags & VM_MAYSHARE)
3519 continue;
3520
3521 /*
3522 * Unmap the page from other VMAs without their own reserves.
3523 * They get marked to be SIGKILLed if they fault in these
3524 * areas. This is because a future no-page fault on this VMA
3525 * could insert a zeroed page instead of the data existing
3526 * from the time of fork. This would look like data corruption
3527 */
3528 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
3529 unmap_hugepage_range(iter_vma, address,
3530 address + huge_page_size(h), page);
3531 }
3532 i_mmap_unlock_write(mapping);
3533}
3534
3535/*
3536 * Hugetlb_cow() should be called with page lock of the original hugepage held.
3537 * Called with hugetlb_instantiation_mutex held and pte_page locked so we
3538 * cannot race with other handlers or page migration.
3539 * Keep the pte_same checks anyway to make transition from the mutex easier.
3540 */
3541static vm_fault_t hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
3542 unsigned long address, pte_t *ptep,
3543 struct page *pagecache_page, spinlock_t *ptl)
3544{
3545 pte_t pte;
3546 struct hstate *h = hstate_vma(vma);
3547 struct page *old_page, *new_page;
3548 int outside_reserve = 0;
3549 vm_fault_t ret = 0;
3550 unsigned long haddr = address & huge_page_mask(h);
3551 struct mmu_notifier_range range;
3552
3553 pte = huge_ptep_get(ptep);
3554 old_page = pte_page(pte);
3555
3556retry_avoidcopy:
3557 /* If no-one else is actually using this page, avoid the copy
3558 * and just make the page writable */
3559 if (page_mapcount(old_page) == 1 && PageAnon(old_page)) {
3560 page_move_anon_rmap(old_page, vma);
3561 set_huge_ptep_writable(vma, haddr, ptep);
3562 return 0;
3563 }
3564
3565 /*
3566 * If the process that created a MAP_PRIVATE mapping is about to
3567 * perform a COW due to a shared page count, attempt to satisfy
3568 * the allocation without using the existing reserves. The pagecache
3569 * page is used to determine if the reserve at this address was
3570 * consumed or not. If reserves were used, a partial faulted mapping
3571 * at the time of fork() could consume its reserves on COW instead
3572 * of the full address range.
3573 */
3574 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
3575 old_page != pagecache_page)
3576 outside_reserve = 1;
3577
3578 get_page(old_page);
3579
3580 /*
3581 * Drop page table lock as buddy allocator may be called. It will
3582 * be acquired again before returning to the caller, as expected.
3583 */
3584 spin_unlock(ptl);
3585 new_page = alloc_huge_page(vma, haddr, outside_reserve);
3586
3587 if (IS_ERR(new_page)) {
3588 /*
3589 * If a process owning a MAP_PRIVATE mapping fails to COW,
3590 * it is due to references held by a child and an insufficient
3591 * huge page pool. To guarantee the original mappers
3592 * reliability, unmap the page from child processes. The child
3593 * may get SIGKILLed if it later faults.
3594 */
3595 if (outside_reserve) {
3596 put_page(old_page);
3597 BUG_ON(huge_pte_none(pte));
3598 unmap_ref_private(mm, vma, old_page, haddr);
3599 BUG_ON(huge_pte_none(pte));
3600 spin_lock(ptl);
3601 ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
3602 if (likely(ptep &&
3603 pte_same(huge_ptep_get(ptep), pte)))
3604 goto retry_avoidcopy;
3605 /*
3606 * race occurs while re-acquiring page table
3607 * lock, and our job is done.
3608 */
3609 return 0;
3610 }
3611
3612 ret = vmf_error(PTR_ERR(new_page));
3613 goto out_release_old;
3614 }
3615
3616 /*
3617 * When the original hugepage is shared one, it does not have
3618 * anon_vma prepared.
3619 */
3620 if (unlikely(anon_vma_prepare(vma))) {
3621 ret = VM_FAULT_OOM;
3622 goto out_release_all;
3623 }
3624
3625 copy_user_huge_page(new_page, old_page, address, vma,
3626 pages_per_huge_page(h));
3627 __SetPageUptodate(new_page);
3628
3629 mmu_notifier_range_init(&range, mm, haddr, haddr + huge_page_size(h));
3630 mmu_notifier_invalidate_range_start(&range);
3631
3632 /*
3633 * Retake the page table lock to check for racing updates
3634 * before the page tables are altered
3635 */
3636 spin_lock(ptl);
3637 ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
3638 if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) {
3639 ClearPagePrivate(new_page);
3640
3641 /* Break COW */
3642 huge_ptep_clear_flush(vma, haddr, ptep);
3643 mmu_notifier_invalidate_range(mm, range.start, range.end);
3644 set_huge_pte_at(mm, haddr, ptep,
3645 make_huge_pte(vma, new_page, 1));
3646 page_remove_rmap(old_page, true);
3647 hugepage_add_new_anon_rmap(new_page, vma, haddr);
3648 set_page_huge_active(new_page);
3649 /* Make the old page be freed below */
3650 new_page = old_page;
3651 }
3652 spin_unlock(ptl);
3653 mmu_notifier_invalidate_range_end(&range);
3654out_release_all:
3655 restore_reserve_on_error(h, vma, haddr, new_page);
3656 put_page(new_page);
3657out_release_old:
3658 put_page(old_page);
3659
3660 spin_lock(ptl); /* Caller expects lock to be held */
3661 return ret;
3662}
3663
3664/* Return the pagecache page at a given address within a VMA */
3665static struct page *hugetlbfs_pagecache_page(struct hstate *h,
3666 struct vm_area_struct *vma, unsigned long address)
3667{
3668 struct address_space *mapping;
3669 pgoff_t idx;
3670
3671 mapping = vma->vm_file->f_mapping;
3672 idx = vma_hugecache_offset(h, vma, address);
3673
3674 return find_lock_page(mapping, idx);
3675}
3676
3677/*
3678 * Return whether there is a pagecache page to back given address within VMA.
3679 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
3680 */
3681static bool hugetlbfs_pagecache_present(struct hstate *h,
3682 struct vm_area_struct *vma, unsigned long address)
3683{
3684 struct address_space *mapping;
3685 pgoff_t idx;
3686 struct page *page;
3687
3688 mapping = vma->vm_file->f_mapping;
3689 idx = vma_hugecache_offset(h, vma, address);
3690
3691 page = find_get_page(mapping, idx);
3692 if (page)
3693 put_page(page);
3694 return page != NULL;
3695}
3696
3697int huge_add_to_page_cache(struct page *page, struct address_space *mapping,
3698 pgoff_t idx)
3699{
3700 struct inode *inode = mapping->host;
3701 struct hstate *h = hstate_inode(inode);
3702 int err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
3703
3704 if (err)
3705 return err;
3706 ClearPagePrivate(page);
3707
3708 /*
3709 * set page dirty so that it will not be removed from cache/file
3710 * by non-hugetlbfs specific code paths.
3711 */
3712 set_page_dirty(page);
3713
3714 spin_lock(&inode->i_lock);
3715 inode->i_blocks += blocks_per_huge_page(h);
3716 spin_unlock(&inode->i_lock);
3717 return 0;
3718}
3719
3720static vm_fault_t hugetlb_no_page(struct mm_struct *mm,
3721 struct vm_area_struct *vma,
3722 struct address_space *mapping, pgoff_t idx,
3723 unsigned long address, pte_t *ptep, unsigned int flags)
3724{
3725 struct hstate *h = hstate_vma(vma);
3726 vm_fault_t ret = VM_FAULT_SIGBUS;
3727 int anon_rmap = 0;
3728 unsigned long size;
3729 struct page *page;
3730 pte_t new_pte;
3731 spinlock_t *ptl;
3732 unsigned long haddr = address & huge_page_mask(h);
3733 bool new_page = false;
3734
3735 /*
3736 * Currently, we are forced to kill the process in the event the
3737 * original mapper has unmapped pages from the child due to a failed
3738 * COW. Warn that such a situation has occurred as it may not be obvious
3739 */
3740 if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
3741 pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n",
3742 current->pid);
3743 return ret;
3744 }
3745
3746 /*
3747 * Use page lock to guard against racing truncation
3748 * before we get page_table_lock.
3749 */
3750retry:
3751 page = find_lock_page(mapping, idx);
3752 if (!page) {
3753 size = i_size_read(mapping->host) >> huge_page_shift(h);
3754 if (idx >= size)
3755 goto out;
3756
3757 /*
3758 * Check for page in userfault range
3759 */
3760 if (userfaultfd_missing(vma)) {
3761 u32 hash;
3762 struct vm_fault vmf = {
3763 .vma = vma,
3764 .address = haddr,
3765 .flags = flags,
3766 /*
3767 * Hard to debug if it ends up being
3768 * used by a callee that assumes
3769 * something about the other
3770 * uninitialized fields... same as in
3771 * memory.c
3772 */
3773 };
3774
3775 /*
3776 * hugetlb_fault_mutex must be dropped before
3777 * handling userfault. Reacquire after handling
3778 * fault to make calling code simpler.
3779 */
3780 hash = hugetlb_fault_mutex_hash(h, mm, vma, mapping,
3781 idx, haddr);
3782 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
3783 ret = handle_userfault(&vmf, VM_UFFD_MISSING);
3784 mutex_lock(&hugetlb_fault_mutex_table[hash]);
3785 goto out;
3786 }
3787
3788 page = alloc_huge_page(vma, haddr, 0);
3789 if (IS_ERR(page)) {
3790 ret = vmf_error(PTR_ERR(page));
3791 goto out;
3792 }
3793 clear_huge_page(page, address, pages_per_huge_page(h));
3794 __SetPageUptodate(page);
3795 new_page = true;
3796
3797 if (vma->vm_flags & VM_MAYSHARE) {
3798 int err = huge_add_to_page_cache(page, mapping, idx);
3799 if (err) {
3800 put_page(page);
3801 if (err == -EEXIST)
3802 goto retry;
3803 goto out;
3804 }
3805 } else {
3806 lock_page(page);
3807 if (unlikely(anon_vma_prepare(vma))) {
3808 ret = VM_FAULT_OOM;
3809 goto backout_unlocked;
3810 }
3811 anon_rmap = 1;
3812 }
3813 } else {
3814 /*
3815 * If memory error occurs between mmap() and fault, some process
3816 * don't have hwpoisoned swap entry for errored virtual address.
3817 * So we need to block hugepage fault by PG_hwpoison bit check.
3818 */
3819 if (unlikely(PageHWPoison(page))) {
3820 ret = VM_FAULT_HWPOISON |
3821 VM_FAULT_SET_HINDEX(hstate_index(h));
3822 goto backout_unlocked;
3823 }
3824 }
3825
3826 /*
3827 * If we are going to COW a private mapping later, we examine the
3828 * pending reservations for this page now. This will ensure that
3829 * any allocations necessary to record that reservation occur outside
3830 * the spinlock.
3831 */
3832 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
3833 if (vma_needs_reservation(h, vma, haddr) < 0) {
3834 ret = VM_FAULT_OOM;
3835 goto