blob: 19d0d08b396fb1356bc4e834d2aaec2977ee173e [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
2 * Generic hugetlb support.
Nadia Yvette Chambers6d49e352012-12-06 10:39:54 +01003 * (C) Nadia Yvette Chambers, April 2004
Linus Torvalds1da177e2005-04-16 15:20:36 -07004 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07005#include <linux/list.h>
6#include <linux/init.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -07007#include <linux/mm.h>
Alexey Dobriyane1759c22008-10-15 23:50:22 +04008#include <linux/seq_file.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -07009#include <linux/sysctl.h>
10#include <linux/highmem.h>
Andrea Arcangelicddb8a52008-07-28 15:46:29 -070011#include <linux/mmu_notifier.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070012#include <linux/nodemask.h>
David Gibson63551ae2005-06-21 17:14:44 -070013#include <linux/pagemap.h>
Christoph Lameter5da7ca82006-01-06 00:10:46 -080014#include <linux/mempolicy.h>
Gideon Israel Dsouza3b321232014-04-07 15:37:26 -070015#include <linux/compiler.h>
Christoph Lameteraea47ff2006-01-08 01:00:57 -080016#include <linux/cpuset.h>
David Gibson3935baa2006-03-22 00:08:53 -080017#include <linux/mutex.h>
Andi Kleenaa888a72008-07-23 21:27:47 -070018#include <linux/bootmem.h>
Nishanth Aravamudana3437872008-07-23 21:27:44 -070019#include <linux/sysfs.h>
Tejun Heo5a0e3ad2010-03-24 17:04:11 +090020#include <linux/slab.h>
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +090021#include <linux/rmap.h>
Naoya Horiguchifd6a03ed2010-05-28 09:29:21 +090022#include <linux/swap.h>
23#include <linux/swapops.h>
Naoya Horiguchic8721bb2013-09-11 14:22:09 -070024#include <linux/page-isolation.h>
Davidlohr Bueso8382d912014-04-03 14:47:31 -070025#include <linux/jhash.h>
Linus Torvaldsd6606682008-08-06 12:04:54 -070026
David Gibson63551ae2005-06-21 17:14:44 -070027#include <asm/page.h>
28#include <asm/pgtable.h>
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -070029#include <asm/tlb.h>
David Gibson63551ae2005-06-21 17:14:44 -070030
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -070031#include <linux/io.h>
David Gibson63551ae2005-06-21 17:14:44 -070032#include <linux/hugetlb.h>
Aneesh Kumar K.V9dd540e2012-07-31 16:42:15 -070033#include <linux/hugetlb_cgroup.h>
Lee Schermerhorn9a305232009-12-14 17:58:25 -080034#include <linux/node.h>
Nick Piggin7835e982006-03-22 00:08:40 -080035#include "internal.h"
Linus Torvalds1da177e2005-04-16 15:20:36 -070036
Andrey Ryabinin753162c2015-02-10 14:11:36 -080037int hugepages_treat_as_movable;
Andi Kleena5516432008-07-23 21:27:41 -070038
Aneesh Kumar K.Vc3f38a32012-07-31 16:42:10 -070039int hugetlb_max_hstate __read_mostly;
Andi Kleene5ff2152008-07-23 21:27:42 -070040unsigned int default_hstate_idx;
41struct hstate hstates[HUGE_MAX_HSTATE];
Naoya Horiguchi641844f2015-06-24 16:56:59 -070042/*
43 * Minimum page order among possible hugepage sizes, set to a proper value
44 * at boot time.
45 */
46static unsigned int minimum_order __read_mostly = UINT_MAX;
Andi Kleene5ff2152008-07-23 21:27:42 -070047
Jon Tollefson53ba51d2008-07-23 21:27:52 -070048__initdata LIST_HEAD(huge_boot_pages);
49
Andi Kleene5ff2152008-07-23 21:27:42 -070050/* for command line parsing */
51static struct hstate * __initdata parsed_hstate;
52static unsigned long __initdata default_hstate_max_huge_pages;
Nick Piggine11bfbf2008-07-23 21:27:52 -070053static unsigned long __initdata default_hstate_size;
Andi Kleene5ff2152008-07-23 21:27:42 -070054
David Gibson3935baa2006-03-22 00:08:53 -080055/*
Naoya Horiguchi31caf662013-09-11 14:21:59 -070056 * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages,
57 * free_huge_pages, and surplus_huge_pages.
David Gibson3935baa2006-03-22 00:08:53 -080058 */
Aneesh Kumar K.Vc3f38a32012-07-31 16:42:10 -070059DEFINE_SPINLOCK(hugetlb_lock);
Eric Paris0bd0f9f2005-11-21 21:32:28 -080060
Davidlohr Bueso8382d912014-04-03 14:47:31 -070061/*
62 * Serializes faults on the same logical page. This is used to
63 * prevent spurious OOMs when the hugepage pool is fully utilized.
64 */
65static int num_fault_mutexes;
Mike Kravetzc672c7f2015-09-08 15:01:35 -070066struct mutex *hugetlb_fault_mutex_table ____cacheline_aligned_in_smp;
Davidlohr Bueso8382d912014-04-03 14:47:31 -070067
Mike Kravetz7ca02d02015-04-15 16:13:42 -070068/* Forward declaration */
69static int hugetlb_acct_memory(struct hstate *h, long delta);
70
David Gibson90481622012-03-21 16:34:12 -070071static inline void unlock_or_release_subpool(struct hugepage_subpool *spool)
72{
73 bool free = (spool->count == 0) && (spool->used_hpages == 0);
74
75 spin_unlock(&spool->lock);
76
77 /* If no pages are used, and no other handles to the subpool
Mike Kravetz7ca02d02015-04-15 16:13:42 -070078 * remain, give up any reservations mased on minimum size and
79 * free the subpool */
80 if (free) {
81 if (spool->min_hpages != -1)
82 hugetlb_acct_memory(spool->hstate,
83 -spool->min_hpages);
David Gibson90481622012-03-21 16:34:12 -070084 kfree(spool);
Mike Kravetz7ca02d02015-04-15 16:13:42 -070085 }
David Gibson90481622012-03-21 16:34:12 -070086}
87
Mike Kravetz7ca02d02015-04-15 16:13:42 -070088struct hugepage_subpool *hugepage_new_subpool(struct hstate *h, long max_hpages,
89 long min_hpages)
David Gibson90481622012-03-21 16:34:12 -070090{
91 struct hugepage_subpool *spool;
92
Mike Kravetzc6a91822015-04-15 16:13:36 -070093 spool = kzalloc(sizeof(*spool), GFP_KERNEL);
David Gibson90481622012-03-21 16:34:12 -070094 if (!spool)
95 return NULL;
96
97 spin_lock_init(&spool->lock);
98 spool->count = 1;
Mike Kravetz7ca02d02015-04-15 16:13:42 -070099 spool->max_hpages = max_hpages;
100 spool->hstate = h;
101 spool->min_hpages = min_hpages;
102
103 if (min_hpages != -1 && hugetlb_acct_memory(h, min_hpages)) {
104 kfree(spool);
105 return NULL;
106 }
107 spool->rsv_hpages = min_hpages;
David Gibson90481622012-03-21 16:34:12 -0700108
109 return spool;
110}
111
112void hugepage_put_subpool(struct hugepage_subpool *spool)
113{
114 spin_lock(&spool->lock);
115 BUG_ON(!spool->count);
116 spool->count--;
117 unlock_or_release_subpool(spool);
118}
119
Mike Kravetz1c5ecae2015-04-15 16:13:39 -0700120/*
121 * Subpool accounting for allocating and reserving pages.
122 * Return -ENOMEM if there are not enough resources to satisfy the
123 * the request. Otherwise, return the number of pages by which the
124 * global pools must be adjusted (upward). The returned value may
125 * only be different than the passed value (delta) in the case where
126 * a subpool minimum size must be manitained.
127 */
128static long hugepage_subpool_get_pages(struct hugepage_subpool *spool,
David Gibson90481622012-03-21 16:34:12 -0700129 long delta)
130{
Mike Kravetz1c5ecae2015-04-15 16:13:39 -0700131 long ret = delta;
David Gibson90481622012-03-21 16:34:12 -0700132
133 if (!spool)
Mike Kravetz1c5ecae2015-04-15 16:13:39 -0700134 return ret;
David Gibson90481622012-03-21 16:34:12 -0700135
136 spin_lock(&spool->lock);
David Gibson90481622012-03-21 16:34:12 -0700137
Mike Kravetz1c5ecae2015-04-15 16:13:39 -0700138 if (spool->max_hpages != -1) { /* maximum size accounting */
139 if ((spool->used_hpages + delta) <= spool->max_hpages)
140 spool->used_hpages += delta;
141 else {
142 ret = -ENOMEM;
143 goto unlock_ret;
144 }
145 }
146
147 if (spool->min_hpages != -1) { /* minimum size accounting */
148 if (delta > spool->rsv_hpages) {
149 /*
150 * Asking for more reserves than those already taken on
151 * behalf of subpool. Return difference.
152 */
153 ret = delta - spool->rsv_hpages;
154 spool->rsv_hpages = 0;
155 } else {
156 ret = 0; /* reserves already accounted for */
157 spool->rsv_hpages -= delta;
158 }
159 }
160
161unlock_ret:
162 spin_unlock(&spool->lock);
David Gibson90481622012-03-21 16:34:12 -0700163 return ret;
164}
165
Mike Kravetz1c5ecae2015-04-15 16:13:39 -0700166/*
167 * Subpool accounting for freeing and unreserving pages.
168 * Return the number of global page reservations that must be dropped.
169 * The return value may only be different than the passed value (delta)
170 * in the case where a subpool minimum size must be maintained.
171 */
172static long hugepage_subpool_put_pages(struct hugepage_subpool *spool,
David Gibson90481622012-03-21 16:34:12 -0700173 long delta)
174{
Mike Kravetz1c5ecae2015-04-15 16:13:39 -0700175 long ret = delta;
176
David Gibson90481622012-03-21 16:34:12 -0700177 if (!spool)
Mike Kravetz1c5ecae2015-04-15 16:13:39 -0700178 return delta;
David Gibson90481622012-03-21 16:34:12 -0700179
180 spin_lock(&spool->lock);
Mike Kravetz1c5ecae2015-04-15 16:13:39 -0700181
182 if (spool->max_hpages != -1) /* maximum size accounting */
183 spool->used_hpages -= delta;
184
185 if (spool->min_hpages != -1) { /* minimum size accounting */
186 if (spool->rsv_hpages + delta <= spool->min_hpages)
187 ret = 0;
188 else
189 ret = spool->rsv_hpages + delta - spool->min_hpages;
190
191 spool->rsv_hpages += delta;
192 if (spool->rsv_hpages > spool->min_hpages)
193 spool->rsv_hpages = spool->min_hpages;
194 }
195
196 /*
197 * If hugetlbfs_put_super couldn't free spool due to an outstanding
198 * quota reference, free it now.
199 */
David Gibson90481622012-03-21 16:34:12 -0700200 unlock_or_release_subpool(spool);
Mike Kravetz1c5ecae2015-04-15 16:13:39 -0700201
202 return ret;
David Gibson90481622012-03-21 16:34:12 -0700203}
204
205static inline struct hugepage_subpool *subpool_inode(struct inode *inode)
206{
207 return HUGETLBFS_SB(inode->i_sb)->spool;
208}
209
210static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma)
211{
Al Viro496ad9a2013-01-23 17:07:38 -0500212 return subpool_inode(file_inode(vma->vm_file));
David Gibson90481622012-03-21 16:34:12 -0700213}
214
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700215/*
Andy Whitcroft96822902008-07-23 21:27:29 -0700216 * Region tracking -- allows tracking of reservations and instantiated pages
217 * across the pages in a mapping.
Andy Whitcroft84afd992008-07-23 21:27:32 -0700218 *
Mike Kravetz1dd308a2015-06-24 16:57:52 -0700219 * The region data structures are embedded into a resv_map and protected
220 * by a resv_map's lock. The set of regions within the resv_map represent
221 * reservations for huge pages, or huge pages that have already been
222 * instantiated within the map. The from and to elements are huge page
223 * indicies into the associated mapping. from indicates the starting index
224 * of the region. to represents the first index past the end of the region.
225 *
226 * For example, a file region structure with from == 0 and to == 4 represents
227 * four huge pages in a mapping. It is important to note that the to element
228 * represents the first element past the end of the region. This is used in
229 * arithmetic as 4(to) - 0(from) = 4 huge pages in the region.
230 *
231 * Interval notation of the form [from, to) will be used to indicate that
232 * the endpoint from is inclusive and to is exclusive.
Andy Whitcroft96822902008-07-23 21:27:29 -0700233 */
234struct file_region {
235 struct list_head link;
236 long from;
237 long to;
238};
239
Mike Kravetz1dd308a2015-06-24 16:57:52 -0700240/*
241 * Add the huge page range represented by [f, t) to the reserve
Mike Kravetz5e911372015-09-08 15:01:28 -0700242 * map. In the normal case, existing regions will be expanded
243 * to accommodate the specified range. Sufficient regions should
244 * exist for expansion due to the previous call to region_chg
245 * with the same range. However, it is possible that region_del
246 * could have been called after region_chg and modifed the map
247 * in such a way that no region exists to be expanded. In this
248 * case, pull a region descriptor from the cache associated with
249 * the map and use that for the new range.
Mike Kravetzcf3ad202015-06-24 16:57:55 -0700250 *
251 * Return the number of new huge pages added to the map. This
252 * number is greater than or equal to zero.
Mike Kravetz1dd308a2015-06-24 16:57:52 -0700253 */
Joonsoo Kim1406ec92014-04-03 14:47:26 -0700254static long region_add(struct resv_map *resv, long f, long t)
Andy Whitcroft96822902008-07-23 21:27:29 -0700255{
Joonsoo Kim1406ec92014-04-03 14:47:26 -0700256 struct list_head *head = &resv->regions;
Andy Whitcroft96822902008-07-23 21:27:29 -0700257 struct file_region *rg, *nrg, *trg;
Mike Kravetzcf3ad202015-06-24 16:57:55 -0700258 long add = 0;
Andy Whitcroft96822902008-07-23 21:27:29 -0700259
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700260 spin_lock(&resv->lock);
Andy Whitcroft96822902008-07-23 21:27:29 -0700261 /* Locate the region we are either in or before. */
262 list_for_each_entry(rg, head, link)
263 if (f <= rg->to)
264 break;
265
Mike Kravetz5e911372015-09-08 15:01:28 -0700266 /*
267 * If no region exists which can be expanded to include the
268 * specified range, the list must have been modified by an
269 * interleving call to region_del(). Pull a region descriptor
270 * from the cache and use it for this range.
271 */
272 if (&rg->link == head || t < rg->from) {
273 VM_BUG_ON(resv->region_cache_count <= 0);
274
275 resv->region_cache_count--;
276 nrg = list_first_entry(&resv->region_cache, struct file_region,
277 link);
278 list_del(&nrg->link);
279
280 nrg->from = f;
281 nrg->to = t;
282 list_add(&nrg->link, rg->link.prev);
283
284 add += t - f;
285 goto out_locked;
286 }
287
Andy Whitcroft96822902008-07-23 21:27:29 -0700288 /* Round our left edge to the current segment if it encloses us. */
289 if (f > rg->from)
290 f = rg->from;
291
292 /* Check for and consume any regions we now overlap with. */
293 nrg = rg;
294 list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
295 if (&rg->link == head)
296 break;
297 if (rg->from > t)
298 break;
299
300 /* If this area reaches higher then extend our area to
301 * include it completely. If this is not the first area
302 * which we intend to reuse, free it. */
303 if (rg->to > t)
304 t = rg->to;
305 if (rg != nrg) {
Mike Kravetzcf3ad202015-06-24 16:57:55 -0700306 /* Decrement return value by the deleted range.
307 * Another range will span this area so that by
308 * end of routine add will be >= zero
309 */
310 add -= (rg->to - rg->from);
Andy Whitcroft96822902008-07-23 21:27:29 -0700311 list_del(&rg->link);
312 kfree(rg);
313 }
314 }
Mike Kravetzcf3ad202015-06-24 16:57:55 -0700315
316 add += (nrg->from - f); /* Added to beginning of region */
Andy Whitcroft96822902008-07-23 21:27:29 -0700317 nrg->from = f;
Mike Kravetzcf3ad202015-06-24 16:57:55 -0700318 add += t - nrg->to; /* Added to end of region */
Andy Whitcroft96822902008-07-23 21:27:29 -0700319 nrg->to = t;
Mike Kravetzcf3ad202015-06-24 16:57:55 -0700320
Mike Kravetz5e911372015-09-08 15:01:28 -0700321out_locked:
322 resv->adds_in_progress--;
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700323 spin_unlock(&resv->lock);
Mike Kravetzcf3ad202015-06-24 16:57:55 -0700324 VM_BUG_ON(add < 0);
325 return add;
Andy Whitcroft96822902008-07-23 21:27:29 -0700326}
327
Mike Kravetz1dd308a2015-06-24 16:57:52 -0700328/*
329 * Examine the existing reserve map and determine how many
330 * huge pages in the specified range [f, t) are NOT currently
331 * represented. This routine is called before a subsequent
332 * call to region_add that will actually modify the reserve
333 * map to add the specified range [f, t). region_chg does
334 * not change the number of huge pages represented by the
335 * map. However, if the existing regions in the map can not
336 * be expanded to represent the new range, a new file_region
337 * structure is added to the map as a placeholder. This is
338 * so that the subsequent region_add call will have all the
339 * regions it needs and will not fail.
340 *
Mike Kravetz5e911372015-09-08 15:01:28 -0700341 * Upon entry, region_chg will also examine the cache of region descriptors
342 * associated with the map. If there are not enough descriptors cached, one
343 * will be allocated for the in progress add operation.
344 *
345 * Returns the number of huge pages that need to be added to the existing
346 * reservation map for the range [f, t). This number is greater or equal to
347 * zero. -ENOMEM is returned if a new file_region structure or cache entry
348 * is needed and can not be allocated.
Mike Kravetz1dd308a2015-06-24 16:57:52 -0700349 */
Joonsoo Kim1406ec92014-04-03 14:47:26 -0700350static long region_chg(struct resv_map *resv, long f, long t)
Andy Whitcroft96822902008-07-23 21:27:29 -0700351{
Joonsoo Kim1406ec92014-04-03 14:47:26 -0700352 struct list_head *head = &resv->regions;
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700353 struct file_region *rg, *nrg = NULL;
Andy Whitcroft96822902008-07-23 21:27:29 -0700354 long chg = 0;
355
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700356retry:
357 spin_lock(&resv->lock);
Mike Kravetz5e911372015-09-08 15:01:28 -0700358retry_locked:
359 resv->adds_in_progress++;
360
361 /*
362 * Check for sufficient descriptors in the cache to accommodate
363 * the number of in progress add operations.
364 */
365 if (resv->adds_in_progress > resv->region_cache_count) {
366 struct file_region *trg;
367
368 VM_BUG_ON(resv->adds_in_progress - resv->region_cache_count > 1);
369 /* Must drop lock to allocate a new descriptor. */
370 resv->adds_in_progress--;
371 spin_unlock(&resv->lock);
372
373 trg = kmalloc(sizeof(*trg), GFP_KERNEL);
Mike Kravetzdbe409e2015-12-11 13:40:52 -0800374 if (!trg) {
375 kfree(nrg);
Mike Kravetz5e911372015-09-08 15:01:28 -0700376 return -ENOMEM;
Mike Kravetzdbe409e2015-12-11 13:40:52 -0800377 }
Mike Kravetz5e911372015-09-08 15:01:28 -0700378
379 spin_lock(&resv->lock);
380 list_add(&trg->link, &resv->region_cache);
381 resv->region_cache_count++;
382 goto retry_locked;
383 }
384
Andy Whitcroft96822902008-07-23 21:27:29 -0700385 /* Locate the region we are before or in. */
386 list_for_each_entry(rg, head, link)
387 if (f <= rg->to)
388 break;
389
390 /* If we are below the current region then a new region is required.
391 * Subtle, allocate a new region at the position but make it zero
392 * size such that we can guarantee to record the reservation. */
393 if (&rg->link == head || t < rg->from) {
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700394 if (!nrg) {
Mike Kravetz5e911372015-09-08 15:01:28 -0700395 resv->adds_in_progress--;
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700396 spin_unlock(&resv->lock);
397 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
398 if (!nrg)
399 return -ENOMEM;
Andy Whitcroft96822902008-07-23 21:27:29 -0700400
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700401 nrg->from = f;
402 nrg->to = f;
403 INIT_LIST_HEAD(&nrg->link);
404 goto retry;
405 }
406
407 list_add(&nrg->link, rg->link.prev);
408 chg = t - f;
409 goto out_nrg;
Andy Whitcroft96822902008-07-23 21:27:29 -0700410 }
411
412 /* Round our left edge to the current segment if it encloses us. */
413 if (f > rg->from)
414 f = rg->from;
415 chg = t - f;
416
417 /* Check for and consume any regions we now overlap with. */
418 list_for_each_entry(rg, rg->link.prev, link) {
419 if (&rg->link == head)
420 break;
421 if (rg->from > t)
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700422 goto out;
Andy Whitcroft96822902008-07-23 21:27:29 -0700423
Lucas De Marchi25985ed2011-03-30 22:57:33 -0300424 /* We overlap with this area, if it extends further than
Andy Whitcroft96822902008-07-23 21:27:29 -0700425 * us then we must extend ourselves. Account for its
426 * existing reservation. */
427 if (rg->to > t) {
428 chg += rg->to - t;
429 t = rg->to;
430 }
431 chg -= rg->to - rg->from;
432 }
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700433
434out:
435 spin_unlock(&resv->lock);
436 /* We already know we raced and no longer need the new region */
437 kfree(nrg);
438 return chg;
439out_nrg:
440 spin_unlock(&resv->lock);
Andy Whitcroft96822902008-07-23 21:27:29 -0700441 return chg;
442}
443
Mike Kravetz1dd308a2015-06-24 16:57:52 -0700444/*
Mike Kravetz5e911372015-09-08 15:01:28 -0700445 * Abort the in progress add operation. The adds_in_progress field
446 * of the resv_map keeps track of the operations in progress between
447 * calls to region_chg and region_add. Operations are sometimes
448 * aborted after the call to region_chg. In such cases, region_abort
449 * is called to decrement the adds_in_progress counter.
450 *
451 * NOTE: The range arguments [f, t) are not needed or used in this
452 * routine. They are kept to make reading the calling code easier as
453 * arguments will match the associated region_chg call.
454 */
455static void region_abort(struct resv_map *resv, long f, long t)
456{
457 spin_lock(&resv->lock);
458 VM_BUG_ON(!resv->region_cache_count);
459 resv->adds_in_progress--;
460 spin_unlock(&resv->lock);
461}
462
463/*
Mike Kravetzfeba16e2015-09-08 15:01:31 -0700464 * Delete the specified range [f, t) from the reserve map. If the
465 * t parameter is LONG_MAX, this indicates that ALL regions after f
466 * should be deleted. Locate the regions which intersect [f, t)
467 * and either trim, delete or split the existing regions.
468 *
469 * Returns the number of huge pages deleted from the reserve map.
470 * In the normal case, the return value is zero or more. In the
471 * case where a region must be split, a new region descriptor must
472 * be allocated. If the allocation fails, -ENOMEM will be returned.
473 * NOTE: If the parameter t == LONG_MAX, then we will never split
474 * a region and possibly return -ENOMEM. Callers specifying
475 * t == LONG_MAX do not need to check for -ENOMEM error.
Mike Kravetz1dd308a2015-06-24 16:57:52 -0700476 */
Mike Kravetzfeba16e2015-09-08 15:01:31 -0700477static long region_del(struct resv_map *resv, long f, long t)
Andy Whitcroft96822902008-07-23 21:27:29 -0700478{
Joonsoo Kim1406ec92014-04-03 14:47:26 -0700479 struct list_head *head = &resv->regions;
Andy Whitcroft96822902008-07-23 21:27:29 -0700480 struct file_region *rg, *trg;
Mike Kravetzfeba16e2015-09-08 15:01:31 -0700481 struct file_region *nrg = NULL;
482 long del = 0;
Andy Whitcroft96822902008-07-23 21:27:29 -0700483
Mike Kravetzfeba16e2015-09-08 15:01:31 -0700484retry:
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700485 spin_lock(&resv->lock);
Mike Kravetzfeba16e2015-09-08 15:01:31 -0700486 list_for_each_entry_safe(rg, trg, head, link) {
Mike Kravetzdbe409e2015-12-11 13:40:52 -0800487 /*
488 * Skip regions before the range to be deleted. file_region
489 * ranges are normally of the form [from, to). However, there
490 * may be a "placeholder" entry in the map which is of the form
491 * (from, to) with from == to. Check for placeholder entries
492 * at the beginning of the range to be deleted.
493 */
494 if (rg->to <= f && (rg->to != rg->from || rg->to != f))
Mike Kravetzfeba16e2015-09-08 15:01:31 -0700495 continue;
Mike Kravetzdbe409e2015-12-11 13:40:52 -0800496
Mike Kravetzfeba16e2015-09-08 15:01:31 -0700497 if (rg->from >= t)
Andy Whitcroft96822902008-07-23 21:27:29 -0700498 break;
Andy Whitcroft96822902008-07-23 21:27:29 -0700499
Mike Kravetzfeba16e2015-09-08 15:01:31 -0700500 if (f > rg->from && t < rg->to) { /* Must split region */
501 /*
502 * Check for an entry in the cache before dropping
503 * lock and attempting allocation.
504 */
505 if (!nrg &&
506 resv->region_cache_count > resv->adds_in_progress) {
507 nrg = list_first_entry(&resv->region_cache,
508 struct file_region,
509 link);
510 list_del(&nrg->link);
511 resv->region_cache_count--;
512 }
513
514 if (!nrg) {
515 spin_unlock(&resv->lock);
516 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
517 if (!nrg)
518 return -ENOMEM;
519 goto retry;
520 }
521
522 del += t - f;
523
524 /* New entry for end of split region */
525 nrg->from = t;
526 nrg->to = rg->to;
527 INIT_LIST_HEAD(&nrg->link);
528
529 /* Original entry is trimmed */
530 rg->to = f;
531
532 list_add(&nrg->link, &rg->link);
533 nrg = NULL;
534 break;
535 }
536
537 if (f <= rg->from && t >= rg->to) { /* Remove entire region */
538 del += rg->to - rg->from;
539 list_del(&rg->link);
540 kfree(rg);
541 continue;
542 }
543
544 if (f <= rg->from) { /* Trim beginning of region */
545 del += t - rg->from;
546 rg->from = t;
547 } else { /* Trim end of region */
548 del += rg->to - f;
549 rg->to = f;
550 }
Andy Whitcroft96822902008-07-23 21:27:29 -0700551 }
552
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700553 spin_unlock(&resv->lock);
Mike Kravetzfeba16e2015-09-08 15:01:31 -0700554 kfree(nrg);
555 return del;
Andy Whitcroft96822902008-07-23 21:27:29 -0700556}
557
Mike Kravetz1dd308a2015-06-24 16:57:52 -0700558/*
Mike Kravetzb5cec282015-09-08 15:01:41 -0700559 * A rare out of memory error was encountered which prevented removal of
560 * the reserve map region for a page. The huge page itself was free'ed
561 * and removed from the page cache. This routine will adjust the subpool
562 * usage count, and the global reserve count if needed. By incrementing
563 * these counts, the reserve map entry which could not be deleted will
564 * appear as a "reserved" entry instead of simply dangling with incorrect
565 * counts.
566 */
567void hugetlb_fix_reserve_counts(struct inode *inode, bool restore_reserve)
568{
569 struct hugepage_subpool *spool = subpool_inode(inode);
570 long rsv_adjust;
571
572 rsv_adjust = hugepage_subpool_get_pages(spool, 1);
573 if (restore_reserve && rsv_adjust) {
574 struct hstate *h = hstate_inode(inode);
575
576 hugetlb_acct_memory(h, 1);
577 }
578}
579
580/*
Mike Kravetz1dd308a2015-06-24 16:57:52 -0700581 * Count and return the number of huge pages in the reserve map
582 * that intersect with the range [f, t).
583 */
Joonsoo Kim1406ec92014-04-03 14:47:26 -0700584static long region_count(struct resv_map *resv, long f, long t)
Andy Whitcroft84afd992008-07-23 21:27:32 -0700585{
Joonsoo Kim1406ec92014-04-03 14:47:26 -0700586 struct list_head *head = &resv->regions;
Andy Whitcroft84afd992008-07-23 21:27:32 -0700587 struct file_region *rg;
588 long chg = 0;
589
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700590 spin_lock(&resv->lock);
Andy Whitcroft84afd992008-07-23 21:27:32 -0700591 /* Locate each segment we overlap with, and count that overlap. */
592 list_for_each_entry(rg, head, link) {
Wang Sheng-Huif2135a42012-05-29 15:06:17 -0700593 long seg_from;
594 long seg_to;
Andy Whitcroft84afd992008-07-23 21:27:32 -0700595
596 if (rg->to <= f)
597 continue;
598 if (rg->from >= t)
599 break;
600
601 seg_from = max(rg->from, f);
602 seg_to = min(rg->to, t);
603
604 chg += seg_to - seg_from;
605 }
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700606 spin_unlock(&resv->lock);
Andy Whitcroft84afd992008-07-23 21:27:32 -0700607
608 return chg;
609}
610
Andy Whitcroft96822902008-07-23 21:27:29 -0700611/*
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700612 * Convert the address within this vma to the page offset within
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700613 * the mapping, in pagecache page units; huge pages here.
614 */
Andi Kleena5516432008-07-23 21:27:41 -0700615static pgoff_t vma_hugecache_offset(struct hstate *h,
616 struct vm_area_struct *vma, unsigned long address)
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700617{
Andi Kleena5516432008-07-23 21:27:41 -0700618 return ((address - vma->vm_start) >> huge_page_shift(h)) +
619 (vma->vm_pgoff >> huge_page_order(h));
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700620}
621
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +0900622pgoff_t linear_hugepage_index(struct vm_area_struct *vma,
623 unsigned long address)
624{
625 return vma_hugecache_offset(hstate_vma(vma), vma, address);
626}
627
Andy Whitcroft84afd992008-07-23 21:27:32 -0700628/*
Mel Gorman08fba692009-01-06 14:38:53 -0800629 * Return the size of the pages allocated when backing a VMA. In the majority
630 * cases this will be same size as used by the page table entries.
631 */
632unsigned long vma_kernel_pagesize(struct vm_area_struct *vma)
633{
634 struct hstate *hstate;
635
636 if (!is_vm_hugetlb_page(vma))
637 return PAGE_SIZE;
638
639 hstate = hstate_vma(vma);
640
Wanpeng Li2415cf12013-07-03 15:02:43 -0700641 return 1UL << huge_page_shift(hstate);
Mel Gorman08fba692009-01-06 14:38:53 -0800642}
Joerg Roedelf340ca02009-06-19 15:16:22 +0200643EXPORT_SYMBOL_GPL(vma_kernel_pagesize);
Mel Gorman08fba692009-01-06 14:38:53 -0800644
645/*
Mel Gorman33402892009-01-06 14:38:54 -0800646 * Return the page size being used by the MMU to back a VMA. In the majority
647 * of cases, the page size used by the kernel matches the MMU size. On
648 * architectures where it differs, an architecture-specific version of this
649 * function is required.
650 */
651#ifndef vma_mmu_pagesize
652unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
653{
654 return vma_kernel_pagesize(vma);
655}
656#endif
657
658/*
Andy Whitcroft84afd992008-07-23 21:27:32 -0700659 * 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)
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700665#define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
Andy Whitcroft84afd992008-07-23 21:27:32 -0700666
Mel Gormana1e78772008-07-23 21:27:23 -0700667/*
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.
Andy Whitcroft84afd992008-07-23 21:27:32 -0700676 *
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.
Mel Gormana1e78772008-07-23 21:27:23 -0700685 */
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700686static 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
Joonsoo Kim9119a412014-04-03 14:47:25 -0700697struct resv_map *resv_map_alloc(void)
Andy Whitcroft84afd992008-07-23 21:27:32 -0700698{
699 struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
Mike Kravetz5e911372015-09-08 15:01:28 -0700700 struct file_region *rg = kmalloc(sizeof(*rg), GFP_KERNEL);
701
702 if (!resv_map || !rg) {
703 kfree(resv_map);
704 kfree(rg);
Andy Whitcroft84afd992008-07-23 21:27:32 -0700705 return NULL;
Mike Kravetz5e911372015-09-08 15:01:28 -0700706 }
Andy Whitcroft84afd992008-07-23 21:27:32 -0700707
708 kref_init(&resv_map->refs);
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700709 spin_lock_init(&resv_map->lock);
Andy Whitcroft84afd992008-07-23 21:27:32 -0700710 INIT_LIST_HEAD(&resv_map->regions);
711
Mike Kravetz5e911372015-09-08 15:01:28 -0700712 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
Andy Whitcroft84afd992008-07-23 21:27:32 -0700718 return resv_map;
719}
720
Joonsoo Kim9119a412014-04-03 14:47:25 -0700721void resv_map_release(struct kref *ref)
Andy Whitcroft84afd992008-07-23 21:27:32 -0700722{
723 struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
Mike Kravetz5e911372015-09-08 15:01:28 -0700724 struct list_head *head = &resv_map->region_cache;
725 struct file_region *rg, *trg;
Andy Whitcroft84afd992008-07-23 21:27:32 -0700726
727 /* Clear out any active regions before we release the map. */
Mike Kravetzfeba16e2015-09-08 15:01:31 -0700728 region_del(resv_map, 0, LONG_MAX);
Mike Kravetz5e911372015-09-08 15:01:28 -0700729
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
Andy Whitcroft84afd992008-07-23 21:27:32 -0700738 kfree(resv_map);
739}
740
Joonsoo Kim4e35f482014-04-03 14:47:30 -0700741static inline struct resv_map *inode_resv_map(struct inode *inode)
742{
743 return inode->i_mapping->private_data;
744}
745
Andy Whitcroft84afd992008-07-23 21:27:32 -0700746static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
Mel Gormana1e78772008-07-23 21:27:23 -0700747{
Sasha Levin81d1b092014-10-09 15:28:10 -0700748 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
Joonsoo Kim4e35f482014-04-03 14:47:30 -0700749 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 {
Andy Whitcroft84afd992008-07-23 21:27:32 -0700756 return (struct resv_map *)(get_vma_private_data(vma) &
757 ~HPAGE_RESV_MASK);
Joonsoo Kim4e35f482014-04-03 14:47:30 -0700758 }
Mel Gormana1e78772008-07-23 21:27:23 -0700759}
760
Andy Whitcroft84afd992008-07-23 21:27:32 -0700761static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
Mel Gormana1e78772008-07-23 21:27:23 -0700762{
Sasha Levin81d1b092014-10-09 15:28:10 -0700763 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
764 VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
Mel Gormana1e78772008-07-23 21:27:23 -0700765
Andy Whitcroft84afd992008-07-23 21:27:32 -0700766 set_vma_private_data(vma, (get_vma_private_data(vma) &
767 HPAGE_RESV_MASK) | (unsigned long)map);
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700768}
769
770static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
771{
Sasha Levin81d1b092014-10-09 15:28:10 -0700772 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
773 VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700774
775 set_vma_private_data(vma, get_vma_private_data(vma) | flags);
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700776}
777
778static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
779{
Sasha Levin81d1b092014-10-09 15:28:10 -0700780 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700781
782 return (get_vma_private_data(vma) & flag) != 0;
Mel Gormana1e78772008-07-23 21:27:23 -0700783}
784
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700785/* Reset counters to 0 and clear all HPAGE_RESV_* flags */
Mel Gormana1e78772008-07-23 21:27:23 -0700786void reset_vma_resv_huge_pages(struct vm_area_struct *vma)
787{
Sasha Levin81d1b092014-10-09 15:28:10 -0700788 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
Mel Gormanf83a2752009-05-28 14:34:40 -0700789 if (!(vma->vm_flags & VM_MAYSHARE))
Mel Gormana1e78772008-07-23 21:27:23 -0700790 vma->vm_private_data = (void *)0;
791}
792
793/* Returns true if the VMA has associated reserve pages */
Nicholas Krause559ec2f2015-09-04 15:48:27 -0700794static bool vma_has_reserves(struct vm_area_struct *vma, long chg)
Mel Gormana1e78772008-07-23 21:27:23 -0700795{
Joonsoo Kimaf0ed732013-09-11 14:21:18 -0700796 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)
Nicholas Krause559ec2f2015-09-04 15:48:27 -0700807 return true;
Joonsoo Kimaf0ed732013-09-11 14:21:18 -0700808 else
Nicholas Krause559ec2f2015-09-04 15:48:27 -0700809 return false;
Joonsoo Kimaf0ed732013-09-11 14:21:18 -0700810 }
Joonsoo Kima63884e2013-09-11 14:21:07 -0700811
812 /* Shared mappings always use reserves */
Mike Kravetz1fb1b0e2015-09-08 15:01:44 -0700813 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 }
Joonsoo Kima63884e2013-09-11 14:21:07 -0700826
827 /*
828 * Only the process that called mmap() has reserves for
829 * private mappings.
830 */
Mel Gorman7f09ca52008-07-23 21:27:58 -0700831 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER))
Nicholas Krause559ec2f2015-09-04 15:48:27 -0700832 return true;
Joonsoo Kima63884e2013-09-11 14:21:07 -0700833
Nicholas Krause559ec2f2015-09-04 15:48:27 -0700834 return false;
Mel Gormana1e78772008-07-23 21:27:23 -0700835}
836
Andi Kleena5516432008-07-23 21:27:41 -0700837static void enqueue_huge_page(struct hstate *h, struct page *page)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700838{
839 int nid = page_to_nid(page);
Aneesh Kumar K.V0edaecf2012-07-31 16:42:07 -0700840 list_move(&page->lru, &h->hugepage_freelists[nid]);
Andi Kleena5516432008-07-23 21:27:41 -0700841 h->free_huge_pages++;
842 h->free_huge_pages_node[nid]++;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700843}
844
Naoya Horiguchibf50bab2010-09-08 10:19:33 +0900845static struct page *dequeue_huge_page_node(struct hstate *h, int nid)
846{
847 struct page *page;
848
Naoya Horiguchic8721bb2013-09-11 14:22:09 -0700849 list_for_each_entry(page, &h->hugepage_freelists[nid], lru)
850 if (!is_migrate_isolate_page(page))
851 break;
852 /*
853 * if 'non-isolated free hugepage' not found on the list,
854 * the allocation fails.
855 */
856 if (&h->hugepage_freelists[nid] == &page->lru)
Naoya Horiguchibf50bab2010-09-08 10:19:33 +0900857 return NULL;
Aneesh Kumar K.V0edaecf2012-07-31 16:42:07 -0700858 list_move(&page->lru, &h->hugepage_activelist);
Naoya Horiguchia9869b82010-09-08 10:19:37 +0900859 set_page_refcounted(page);
Naoya Horiguchibf50bab2010-09-08 10:19:33 +0900860 h->free_huge_pages--;
861 h->free_huge_pages_node[nid]--;
862 return page;
863}
864
Naoya Horiguchi86cdb462013-09-11 14:22:13 -0700865/* Movability of hugepages depends on migration support. */
866static inline gfp_t htlb_alloc_mask(struct hstate *h)
867{
Naoya Horiguchi100873d2014-06-04 16:10:56 -0700868 if (hugepages_treat_as_movable || hugepage_migration_supported(h))
Naoya Horiguchi86cdb462013-09-11 14:22:13 -0700869 return GFP_HIGHUSER_MOVABLE;
870 else
871 return GFP_HIGHUSER;
872}
873
Andi Kleena5516432008-07-23 21:27:41 -0700874static struct page *dequeue_huge_page_vma(struct hstate *h,
875 struct vm_area_struct *vma,
Joonsoo Kimaf0ed732013-09-11 14:21:18 -0700876 unsigned long address, int avoid_reserve,
877 long chg)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700878{
Konstantin Khlebnikovb1c12cb2012-04-25 16:01:46 -0700879 struct page *page = NULL;
Lee Schermerhorn480eccf2007-09-18 22:46:47 -0700880 struct mempolicy *mpol;
Mel Gorman19770b32008-04-28 02:12:18 -0700881 nodemask_t *nodemask;
Miao Xiec0ff7452010-05-24 14:32:08 -0700882 struct zonelist *zonelist;
Mel Gormandd1a2392008-04-28 02:12:17 -0700883 struct zone *zone;
884 struct zoneref *z;
Mel Gormancc9a6c82012-03-21 16:34:11 -0700885 unsigned int cpuset_mems_cookie;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700886
Mel Gormana1e78772008-07-23 21:27:23 -0700887 /*
888 * A child process with MAP_PRIVATE mappings created by their parent
889 * have no page reserves. This check ensures that reservations are
890 * not "stolen". The child may still get SIGKILLed
891 */
Joonsoo Kimaf0ed732013-09-11 14:21:18 -0700892 if (!vma_has_reserves(vma, chg) &&
Andi Kleena5516432008-07-23 21:27:41 -0700893 h->free_huge_pages - h->resv_huge_pages == 0)
Miao Xiec0ff7452010-05-24 14:32:08 -0700894 goto err;
Mel Gormana1e78772008-07-23 21:27:23 -0700895
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700896 /* If reserves cannot be used, ensure enough pages are in the pool */
Andi Kleena5516432008-07-23 21:27:41 -0700897 if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
Justin P. Mattock6eab04a2011-04-08 19:49:08 -0700898 goto err;
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700899
Joonsoo Kim9966c4b2013-09-11 14:20:50 -0700900retry_cpuset:
Mel Gormand26914d2014-04-03 14:47:24 -0700901 cpuset_mems_cookie = read_mems_allowed_begin();
Joonsoo Kim9966c4b2013-09-11 14:20:50 -0700902 zonelist = huge_zonelist(vma, address,
Naoya Horiguchi86cdb462013-09-11 14:22:13 -0700903 htlb_alloc_mask(h), &mpol, &nodemask);
Joonsoo Kim9966c4b2013-09-11 14:20:50 -0700904
Mel Gorman19770b32008-04-28 02:12:18 -0700905 for_each_zone_zonelist_nodemask(zone, z, zonelist,
906 MAX_NR_ZONES - 1, nodemask) {
Vladimir Davydov344736f2014-10-20 15:50:30 +0400907 if (cpuset_zone_allowed(zone, htlb_alloc_mask(h))) {
Naoya Horiguchibf50bab2010-09-08 10:19:33 +0900908 page = dequeue_huge_page_node(h, zone_to_nid(zone));
909 if (page) {
Joonsoo Kimaf0ed732013-09-11 14:21:18 -0700910 if (avoid_reserve)
911 break;
912 if (!vma_has_reserves(vma, chg))
913 break;
914
Joonsoo Kim07443a82013-09-11 14:21:58 -0700915 SetPagePrivate(page);
Joonsoo Kimaf0ed732013-09-11 14:21:18 -0700916 h->resv_huge_pages--;
Naoya Horiguchibf50bab2010-09-08 10:19:33 +0900917 break;
918 }
Andrew Morton3abf7af2007-07-19 01:49:08 -0700919 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700920 }
Mel Gormancc9a6c82012-03-21 16:34:11 -0700921
922 mpol_cond_put(mpol);
Mel Gormand26914d2014-04-03 14:47:24 -0700923 if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie)))
Mel Gormancc9a6c82012-03-21 16:34:11 -0700924 goto retry_cpuset;
925 return page;
926
Miao Xiec0ff7452010-05-24 14:32:08 -0700927err:
Mel Gormancc9a6c82012-03-21 16:34:11 -0700928 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700929}
930
Luiz Capitulino1cac6f22014-06-04 16:07:11 -0700931/*
932 * common helper functions for hstate_next_node_to_{alloc|free}.
933 * We may have allocated or freed a huge page based on a different
934 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
935 * be outside of *nodes_allowed. Ensure that we use an allowed
936 * node for alloc or free.
937 */
938static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
939{
940 nid = next_node(nid, *nodes_allowed);
941 if (nid == MAX_NUMNODES)
942 nid = first_node(*nodes_allowed);
943 VM_BUG_ON(nid >= MAX_NUMNODES);
944
945 return nid;
946}
947
948static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
949{
950 if (!node_isset(nid, *nodes_allowed))
951 nid = next_node_allowed(nid, nodes_allowed);
952 return nid;
953}
954
955/*
956 * returns the previously saved node ["this node"] from which to
957 * allocate a persistent huge page for the pool and advance the
958 * next node from which to allocate, handling wrap at end of node
959 * mask.
960 */
961static int hstate_next_node_to_alloc(struct hstate *h,
962 nodemask_t *nodes_allowed)
963{
964 int nid;
965
966 VM_BUG_ON(!nodes_allowed);
967
968 nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
969 h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
970
971 return nid;
972}
973
974/*
975 * helper for free_pool_huge_page() - return the previously saved
976 * node ["this node"] from which to free a huge page. Advance the
977 * next node id whether or not we find a free huge page to free so
978 * that the next attempt to free addresses the next node.
979 */
980static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
981{
982 int nid;
983
984 VM_BUG_ON(!nodes_allowed);
985
986 nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
987 h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
988
989 return nid;
990}
991
992#define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \
993 for (nr_nodes = nodes_weight(*mask); \
994 nr_nodes > 0 && \
995 ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \
996 nr_nodes--)
997
998#define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \
999 for (nr_nodes = nodes_weight(*mask); \
1000 nr_nodes > 0 && \
1001 ((node = hstate_next_node_to_free(hs, mask)) || 1); \
1002 nr_nodes--)
1003
Vlastimil Babka080fe202016-02-05 15:36:41 -08001004#if defined(CONFIG_X86_64) && ((defined(CONFIG_MEMORY_ISOLATION) && defined(CONFIG_COMPACTION)) || defined(CONFIG_CMA))
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001005static void destroy_compound_gigantic_page(struct page *page,
Kirill A. Shutemovd00181b2015-11-06 16:29:57 -08001006 unsigned int order)
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001007{
1008 int i;
1009 int nr_pages = 1 << order;
1010 struct page *p = page + 1;
1011
1012 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
Kirill A. Shutemov1d798ca2015-11-06 16:29:54 -08001013 clear_compound_head(p);
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001014 set_page_refcounted(p);
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001015 }
1016
1017 set_compound_order(page, 0);
1018 __ClearPageHead(page);
1019}
1020
Kirill A. Shutemovd00181b2015-11-06 16:29:57 -08001021static void free_gigantic_page(struct page *page, unsigned int order)
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001022{
1023 free_contig_range(page_to_pfn(page), 1 << order);
1024}
1025
1026static int __alloc_gigantic_page(unsigned long start_pfn,
1027 unsigned long nr_pages)
1028{
1029 unsigned long end_pfn = start_pfn + nr_pages;
1030 return alloc_contig_range(start_pfn, end_pfn, MIGRATE_MOVABLE);
1031}
1032
1033static bool pfn_range_valid_gigantic(unsigned long start_pfn,
1034 unsigned long nr_pages)
1035{
1036 unsigned long i, end_pfn = start_pfn + nr_pages;
1037 struct page *page;
1038
1039 for (i = start_pfn; i < end_pfn; i++) {
1040 if (!pfn_valid(i))
1041 return false;
1042
1043 page = pfn_to_page(i);
1044
1045 if (PageReserved(page))
1046 return false;
1047
1048 if (page_count(page) > 0)
1049 return false;
1050
1051 if (PageHuge(page))
1052 return false;
1053 }
1054
1055 return true;
1056}
1057
1058static bool zone_spans_last_pfn(const struct zone *zone,
1059 unsigned long start_pfn, unsigned long nr_pages)
1060{
1061 unsigned long last_pfn = start_pfn + nr_pages - 1;
1062 return zone_spans_pfn(zone, last_pfn);
1063}
1064
Kirill A. Shutemovd00181b2015-11-06 16:29:57 -08001065static struct page *alloc_gigantic_page(int nid, unsigned int order)
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001066{
1067 unsigned long nr_pages = 1 << order;
1068 unsigned long ret, pfn, flags;
1069 struct zone *z;
1070
1071 z = NODE_DATA(nid)->node_zones;
1072 for (; z - NODE_DATA(nid)->node_zones < MAX_NR_ZONES; z++) {
1073 spin_lock_irqsave(&z->lock, flags);
1074
1075 pfn = ALIGN(z->zone_start_pfn, nr_pages);
1076 while (zone_spans_last_pfn(z, pfn, nr_pages)) {
1077 if (pfn_range_valid_gigantic(pfn, nr_pages)) {
1078 /*
1079 * We release the zone lock here because
1080 * alloc_contig_range() will also lock the zone
1081 * at some point. If there's an allocation
1082 * spinning on this lock, it may win the race
1083 * and cause alloc_contig_range() to fail...
1084 */
1085 spin_unlock_irqrestore(&z->lock, flags);
1086 ret = __alloc_gigantic_page(pfn, nr_pages);
1087 if (!ret)
1088 return pfn_to_page(pfn);
1089 spin_lock_irqsave(&z->lock, flags);
1090 }
1091 pfn += nr_pages;
1092 }
1093
1094 spin_unlock_irqrestore(&z->lock, flags);
1095 }
1096
1097 return NULL;
1098}
1099
1100static void prep_new_huge_page(struct hstate *h, struct page *page, int nid);
Kirill A. Shutemovd00181b2015-11-06 16:29:57 -08001101static void prep_compound_gigantic_page(struct page *page, unsigned int order);
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001102
1103static struct page *alloc_fresh_gigantic_page_node(struct hstate *h, int nid)
1104{
1105 struct page *page;
1106
1107 page = alloc_gigantic_page(nid, huge_page_order(h));
1108 if (page) {
1109 prep_compound_gigantic_page(page, huge_page_order(h));
1110 prep_new_huge_page(h, page, nid);
1111 }
1112
1113 return page;
1114}
1115
1116static int alloc_fresh_gigantic_page(struct hstate *h,
1117 nodemask_t *nodes_allowed)
1118{
1119 struct page *page = NULL;
1120 int nr_nodes, node;
1121
1122 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
1123 page = alloc_fresh_gigantic_page_node(h, node);
1124 if (page)
1125 return 1;
1126 }
1127
1128 return 0;
1129}
1130
1131static inline bool gigantic_page_supported(void) { return true; }
1132#else
1133static inline bool gigantic_page_supported(void) { return false; }
Kirill A. Shutemovd00181b2015-11-06 16:29:57 -08001134static inline void free_gigantic_page(struct page *page, unsigned int order) { }
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001135static inline void destroy_compound_gigantic_page(struct page *page,
Kirill A. Shutemovd00181b2015-11-06 16:29:57 -08001136 unsigned int order) { }
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001137static inline int alloc_fresh_gigantic_page(struct hstate *h,
1138 nodemask_t *nodes_allowed) { return 0; }
1139#endif
1140
Andi Kleena5516432008-07-23 21:27:41 -07001141static void update_and_free_page(struct hstate *h, struct page *page)
Adam Litke6af2acb2007-10-16 01:26:16 -07001142{
1143 int i;
Andi Kleena5516432008-07-23 21:27:41 -07001144
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001145 if (hstate_is_gigantic(h) && !gigantic_page_supported())
1146 return;
Andy Whitcroft18229df2008-11-06 12:53:27 -08001147
Andi Kleena5516432008-07-23 21:27:41 -07001148 h->nr_huge_pages--;
1149 h->nr_huge_pages_node[page_to_nid(page)]--;
1150 for (i = 0; i < pages_per_huge_page(h); i++) {
Chris Forbes32f84522011-07-25 17:12:14 -07001151 page[i].flags &= ~(1 << PG_locked | 1 << PG_error |
1152 1 << PG_referenced | 1 << PG_dirty |
Luiz Capitulinoa7407a22014-06-04 16:07:09 -07001153 1 << PG_active | 1 << PG_private |
1154 1 << PG_writeback);
Adam Litke6af2acb2007-10-16 01:26:16 -07001155 }
Sasha Levin309381fea2014-01-23 15:52:54 -08001156 VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page);
Kirill A. Shutemovf1e61552015-11-06 16:29:50 -08001157 set_compound_page_dtor(page, NULL_COMPOUND_DTOR);
Adam Litke6af2acb2007-10-16 01:26:16 -07001158 set_page_refcounted(page);
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001159 if (hstate_is_gigantic(h)) {
1160 destroy_compound_gigantic_page(page, huge_page_order(h));
1161 free_gigantic_page(page, huge_page_order(h));
1162 } else {
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001163 __free_pages(page, huge_page_order(h));
1164 }
Adam Litke6af2acb2007-10-16 01:26:16 -07001165}
1166
Andi Kleene5ff2152008-07-23 21:27:42 -07001167struct hstate *size_to_hstate(unsigned long size)
1168{
1169 struct hstate *h;
1170
1171 for_each_hstate(h) {
1172 if (huge_page_size(h) == size)
1173 return h;
1174 }
1175 return NULL;
1176}
1177
Naoya Horiguchibcc54222015-04-15 16:14:38 -07001178/*
1179 * Test to determine whether the hugepage is "active/in-use" (i.e. being linked
1180 * to hstate->hugepage_activelist.)
1181 *
1182 * This function can be called for tail pages, but never returns true for them.
1183 */
1184bool page_huge_active(struct page *page)
1185{
1186 VM_BUG_ON_PAGE(!PageHuge(page), page);
1187 return PageHead(page) && PagePrivate(&page[1]);
1188}
1189
1190/* never called for tail page */
1191static void set_page_huge_active(struct page *page)
1192{
1193 VM_BUG_ON_PAGE(!PageHeadHuge(page), page);
1194 SetPagePrivate(&page[1]);
1195}
1196
1197static void clear_page_huge_active(struct page *page)
1198{
1199 VM_BUG_ON_PAGE(!PageHeadHuge(page), page);
1200 ClearPagePrivate(&page[1]);
1201}
1202
Atsushi Kumagai8f1d26d2014-07-30 16:08:39 -07001203void free_huge_page(struct page *page)
David Gibson27a85ef2006-03-22 00:08:56 -08001204{
Andi Kleena5516432008-07-23 21:27:41 -07001205 /*
1206 * Can't pass hstate in here because it is called from the
1207 * compound page destructor.
1208 */
Andi Kleene5ff2152008-07-23 21:27:42 -07001209 struct hstate *h = page_hstate(page);
Adam Litke7893d1d2007-10-16 01:26:18 -07001210 int nid = page_to_nid(page);
David Gibson90481622012-03-21 16:34:12 -07001211 struct hugepage_subpool *spool =
1212 (struct hugepage_subpool *)page_private(page);
Joonsoo Kim07443a82013-09-11 14:21:58 -07001213 bool restore_reserve;
David Gibson27a85ef2006-03-22 00:08:56 -08001214
Andy Whitcrofte5df70a2008-02-23 15:23:32 -08001215 set_page_private(page, 0);
Mel Gorman23be7462010-04-23 13:17:56 -04001216 page->mapping = NULL;
Mike Kravetzb4330af2016-02-05 15:36:38 -08001217 VM_BUG_ON_PAGE(page_count(page), page);
1218 VM_BUG_ON_PAGE(page_mapcount(page), page);
Joonsoo Kim07443a82013-09-11 14:21:58 -07001219 restore_reserve = PagePrivate(page);
Joonsoo Kim16c794b2013-10-16 13:46:48 -07001220 ClearPagePrivate(page);
David Gibson27a85ef2006-03-22 00:08:56 -08001221
Mike Kravetz1c5ecae2015-04-15 16:13:39 -07001222 /*
1223 * A return code of zero implies that the subpool will be under its
1224 * minimum size if the reservation is not restored after page is free.
1225 * Therefore, force restore_reserve operation.
1226 */
1227 if (hugepage_subpool_put_pages(spool, 1) == 0)
1228 restore_reserve = true;
1229
David Gibson27a85ef2006-03-22 00:08:56 -08001230 spin_lock(&hugetlb_lock);
Naoya Horiguchibcc54222015-04-15 16:14:38 -07001231 clear_page_huge_active(page);
Aneesh Kumar K.V6d76dcf2012-07-31 16:42:18 -07001232 hugetlb_cgroup_uncharge_page(hstate_index(h),
1233 pages_per_huge_page(h), page);
Joonsoo Kim07443a82013-09-11 14:21:58 -07001234 if (restore_reserve)
1235 h->resv_huge_pages++;
1236
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001237 if (h->surplus_huge_pages_node[nid]) {
Aneesh Kumar K.V0edaecf2012-07-31 16:42:07 -07001238 /* remove the page from active list */
1239 list_del(&page->lru);
Andi Kleena5516432008-07-23 21:27:41 -07001240 update_and_free_page(h, page);
1241 h->surplus_huge_pages--;
1242 h->surplus_huge_pages_node[nid]--;
Adam Litke7893d1d2007-10-16 01:26:18 -07001243 } else {
Will Deacon5d3a5512012-10-08 16:29:32 -07001244 arch_clear_hugepage_flags(page);
Andi Kleena5516432008-07-23 21:27:41 -07001245 enqueue_huge_page(h, page);
Adam Litke7893d1d2007-10-16 01:26:18 -07001246 }
David Gibson27a85ef2006-03-22 00:08:56 -08001247 spin_unlock(&hugetlb_lock);
1248}
1249
Andi Kleena5516432008-07-23 21:27:41 -07001250static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
Andi Kleenb7ba30c2008-07-23 21:27:40 -07001251{
Aneesh Kumar K.V0edaecf2012-07-31 16:42:07 -07001252 INIT_LIST_HEAD(&page->lru);
Kirill A. Shutemovf1e61552015-11-06 16:29:50 -08001253 set_compound_page_dtor(page, HUGETLB_PAGE_DTOR);
Andi Kleenb7ba30c2008-07-23 21:27:40 -07001254 spin_lock(&hugetlb_lock);
Aneesh Kumar K.V9dd540e2012-07-31 16:42:15 -07001255 set_hugetlb_cgroup(page, NULL);
Andi Kleena5516432008-07-23 21:27:41 -07001256 h->nr_huge_pages++;
1257 h->nr_huge_pages_node[nid]++;
Andi Kleenb7ba30c2008-07-23 21:27:40 -07001258 spin_unlock(&hugetlb_lock);
1259 put_page(page); /* free it into the hugepage allocator */
1260}
1261
Kirill A. Shutemovd00181b2015-11-06 16:29:57 -08001262static void prep_compound_gigantic_page(struct page *page, unsigned int order)
Wu Fengguang20a03072009-06-16 15:32:22 -07001263{
1264 int i;
1265 int nr_pages = 1 << order;
1266 struct page *p = page + 1;
1267
1268 /* we rely on prep_new_huge_page to set the destructor */
1269 set_compound_order(page, order);
Andrea Arcangelief5a22b2013-10-16 13:46:56 -07001270 __ClearPageReserved(page);
Kirill A. Shutemovde09d312016-01-15 16:51:42 -08001271 __SetPageHead(page);
Wu Fengguang20a03072009-06-16 15:32:22 -07001272 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
Andrea Arcangelief5a22b2013-10-16 13:46:56 -07001273 /*
1274 * For gigantic hugepages allocated through bootmem at
1275 * boot, it's safer to be consistent with the not-gigantic
1276 * hugepages and clear the PG_reserved bit from all tail pages
1277 * too. Otherwse drivers using get_user_pages() to access tail
1278 * pages may get the reference counting wrong if they see
1279 * PG_reserved set on a tail page (despite the head page not
1280 * having PG_reserved set). Enforcing this consistency between
1281 * head and tail pages allows drivers to optimize away a check
1282 * on the head page when they need know if put_page() is needed
1283 * after get_user_pages().
1284 */
1285 __ClearPageReserved(p);
Youquan Song58a84aa2011-12-08 14:34:18 -08001286 set_page_count(p, 0);
Kirill A. Shutemov1d798ca2015-11-06 16:29:54 -08001287 set_compound_head(p, page);
Wu Fengguang20a03072009-06-16 15:32:22 -07001288 }
Mike Kravetzb4330af2016-02-05 15:36:38 -08001289 atomic_set(compound_mapcount_ptr(page), -1);
Wu Fengguang20a03072009-06-16 15:32:22 -07001290}
1291
Andrew Morton77959122012-10-08 16:34:11 -07001292/*
1293 * PageHuge() only returns true for hugetlbfs pages, but not for normal or
1294 * transparent huge pages. See the PageTransHuge() documentation for more
1295 * details.
1296 */
Wu Fengguang20a03072009-06-16 15:32:22 -07001297int PageHuge(struct page *page)
1298{
Wu Fengguang20a03072009-06-16 15:32:22 -07001299 if (!PageCompound(page))
1300 return 0;
1301
1302 page = compound_head(page);
Kirill A. Shutemovf1e61552015-11-06 16:29:50 -08001303 return page[1].compound_dtor == HUGETLB_PAGE_DTOR;
Wu Fengguang20a03072009-06-16 15:32:22 -07001304}
Naoya Horiguchi43131e12010-05-28 09:29:22 +09001305EXPORT_SYMBOL_GPL(PageHuge);
1306
Andrea Arcangeli27c73ae2013-11-21 14:32:02 -08001307/*
1308 * PageHeadHuge() only returns true for hugetlbfs head page, but not for
1309 * normal or transparent huge pages.
1310 */
1311int PageHeadHuge(struct page *page_head)
1312{
Andrea Arcangeli27c73ae2013-11-21 14:32:02 -08001313 if (!PageHead(page_head))
1314 return 0;
1315
Andrew Morton758f66a2014-01-21 15:48:57 -08001316 return get_compound_page_dtor(page_head) == free_huge_page;
Andrea Arcangeli27c73ae2013-11-21 14:32:02 -08001317}
Andrea Arcangeli27c73ae2013-11-21 14:32:02 -08001318
Zhang Yi13d60f42013-06-25 21:19:31 +08001319pgoff_t __basepage_index(struct page *page)
1320{
1321 struct page *page_head = compound_head(page);
1322 pgoff_t index = page_index(page_head);
1323 unsigned long compound_idx;
1324
1325 if (!PageHuge(page_head))
1326 return page_index(page);
1327
1328 if (compound_order(page_head) >= MAX_ORDER)
1329 compound_idx = page_to_pfn(page) - page_to_pfn(page_head);
1330 else
1331 compound_idx = page - page_head;
1332
1333 return (index << compound_order(page_head)) + compound_idx;
1334}
1335
Andi Kleena5516432008-07-23 21:27:41 -07001336static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001337{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001338 struct page *page;
Joe Jinf96efd52007-07-15 23:38:12 -07001339
Vlastimil Babka96db8002015-09-08 15:03:50 -07001340 page = __alloc_pages_node(nid,
Naoya Horiguchi86cdb462013-09-11 14:22:13 -07001341 htlb_alloc_mask(h)|__GFP_COMP|__GFP_THISNODE|
Nishanth Aravamudan551883a2008-04-29 00:58:26 -07001342 __GFP_REPEAT|__GFP_NOWARN,
Andi Kleena5516432008-07-23 21:27:41 -07001343 huge_page_order(h));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001344 if (page) {
Andi Kleena5516432008-07-23 21:27:41 -07001345 prep_new_huge_page(h, page, nid);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001346 }
Nishanth Aravamudan63b46132007-10-16 01:26:24 -07001347
1348 return page;
1349}
1350
Joonsoo Kimb2261022013-09-11 14:21:00 -07001351static int alloc_fresh_huge_page(struct hstate *h, nodemask_t *nodes_allowed)
1352{
1353 struct page *page;
1354 int nr_nodes, node;
1355 int ret = 0;
1356
1357 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
1358 page = alloc_fresh_huge_page_node(h, node);
1359 if (page) {
1360 ret = 1;
1361 break;
1362 }
1363 }
1364
1365 if (ret)
1366 count_vm_event(HTLB_BUDDY_PGALLOC);
1367 else
1368 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
1369
1370 return ret;
1371}
1372
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001373/*
1374 * Free huge page from pool from next node to free.
1375 * Attempt to keep persistent huge pages more or less
1376 * balanced over allowed nodes.
1377 * Called with hugetlb_lock locked.
1378 */
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001379static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
1380 bool acct_surplus)
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001381{
Joonsoo Kimb2261022013-09-11 14:21:00 -07001382 int nr_nodes, node;
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001383 int ret = 0;
1384
Joonsoo Kimb2261022013-09-11 14:21:00 -07001385 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
Lee Schermerhorn685f3452009-09-21 17:01:23 -07001386 /*
1387 * If we're returning unused surplus pages, only examine
1388 * nodes with surplus pages.
1389 */
Joonsoo Kimb2261022013-09-11 14:21:00 -07001390 if ((!acct_surplus || h->surplus_huge_pages_node[node]) &&
1391 !list_empty(&h->hugepage_freelists[node])) {
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001392 struct page *page =
Joonsoo Kimb2261022013-09-11 14:21:00 -07001393 list_entry(h->hugepage_freelists[node].next,
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001394 struct page, lru);
1395 list_del(&page->lru);
1396 h->free_huge_pages--;
Joonsoo Kimb2261022013-09-11 14:21:00 -07001397 h->free_huge_pages_node[node]--;
Lee Schermerhorn685f3452009-09-21 17:01:23 -07001398 if (acct_surplus) {
1399 h->surplus_huge_pages--;
Joonsoo Kimb2261022013-09-11 14:21:00 -07001400 h->surplus_huge_pages_node[node]--;
Lee Schermerhorn685f3452009-09-21 17:01:23 -07001401 }
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001402 update_and_free_page(h, page);
1403 ret = 1;
Lee Schermerhorn9a76db02009-12-14 17:58:15 -08001404 break;
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001405 }
Joonsoo Kimb2261022013-09-11 14:21:00 -07001406 }
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001407
1408 return ret;
1409}
1410
Naoya Horiguchic8721bb2013-09-11 14:22:09 -07001411/*
1412 * Dissolve a given free hugepage into free buddy pages. This function does
1413 * nothing for in-use (including surplus) hugepages.
1414 */
1415static void dissolve_free_huge_page(struct page *page)
1416{
1417 spin_lock(&hugetlb_lock);
1418 if (PageHuge(page) && !page_count(page)) {
1419 struct hstate *h = page_hstate(page);
1420 int nid = page_to_nid(page);
1421 list_del(&page->lru);
1422 h->free_huge_pages--;
1423 h->free_huge_pages_node[nid]--;
1424 update_and_free_page(h, page);
1425 }
1426 spin_unlock(&hugetlb_lock);
1427}
1428
1429/*
1430 * Dissolve free hugepages in a given pfn range. Used by memory hotplug to
1431 * make specified memory blocks removable from the system.
1432 * Note that start_pfn should aligned with (minimum) hugepage size.
1433 */
1434void dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn)
1435{
Naoya Horiguchic8721bb2013-09-11 14:22:09 -07001436 unsigned long pfn;
Naoya Horiguchic8721bb2013-09-11 14:22:09 -07001437
Li Zhongd0177632014-08-06 16:07:56 -07001438 if (!hugepages_supported())
1439 return;
1440
Naoya Horiguchi641844f2015-06-24 16:56:59 -07001441 VM_BUG_ON(!IS_ALIGNED(start_pfn, 1 << minimum_order));
1442 for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << minimum_order)
Naoya Horiguchic8721bb2013-09-11 14:22:09 -07001443 dissolve_free_huge_page(pfn_to_page(pfn));
1444}
1445
Dave Hansen099730d2015-11-05 18:50:17 -08001446/*
1447 * There are 3 ways this can get called:
1448 * 1. With vma+addr: we use the VMA's memory policy
1449 * 2. With !vma, but nid=NUMA_NO_NODE: We try to allocate a huge
1450 * page from any node, and let the buddy allocator itself figure
1451 * it out.
1452 * 3. With !vma, but nid!=NUMA_NO_NODE. We allocate a huge page
1453 * strictly from 'nid'
1454 */
1455static struct page *__hugetlb_alloc_buddy_huge_page(struct hstate *h,
1456 struct vm_area_struct *vma, unsigned long addr, int nid)
1457{
1458 int order = huge_page_order(h);
1459 gfp_t gfp = htlb_alloc_mask(h)|__GFP_COMP|__GFP_REPEAT|__GFP_NOWARN;
1460 unsigned int cpuset_mems_cookie;
1461
1462 /*
1463 * We need a VMA to get a memory policy. If we do not
Dave Hansene0ec90e2015-11-05 18:50:20 -08001464 * have one, we use the 'nid' argument.
1465 *
1466 * The mempolicy stuff below has some non-inlined bits
1467 * and calls ->vm_ops. That makes it hard to optimize at
1468 * compile-time, even when NUMA is off and it does
1469 * nothing. This helps the compiler optimize it out.
Dave Hansen099730d2015-11-05 18:50:17 -08001470 */
Dave Hansene0ec90e2015-11-05 18:50:20 -08001471 if (!IS_ENABLED(CONFIG_NUMA) || !vma) {
Dave Hansen099730d2015-11-05 18:50:17 -08001472 /*
1473 * If a specific node is requested, make sure to
1474 * get memory from there, but only when a node
1475 * is explicitly specified.
1476 */
1477 if (nid != NUMA_NO_NODE)
1478 gfp |= __GFP_THISNODE;
1479 /*
1480 * Make sure to call something that can handle
1481 * nid=NUMA_NO_NODE
1482 */
1483 return alloc_pages_node(nid, gfp, order);
1484 }
1485
1486 /*
1487 * OK, so we have a VMA. Fetch the mempolicy and try to
Dave Hansene0ec90e2015-11-05 18:50:20 -08001488 * allocate a huge page with it. We will only reach this
1489 * when CONFIG_NUMA=y.
Dave Hansen099730d2015-11-05 18:50:17 -08001490 */
1491 do {
1492 struct page *page;
1493 struct mempolicy *mpol;
1494 struct zonelist *zl;
1495 nodemask_t *nodemask;
1496
1497 cpuset_mems_cookie = read_mems_allowed_begin();
1498 zl = huge_zonelist(vma, addr, gfp, &mpol, &nodemask);
1499 mpol_cond_put(mpol);
1500 page = __alloc_pages_nodemask(gfp, order, zl, nodemask);
1501 if (page)
1502 return page;
1503 } while (read_mems_allowed_retry(cpuset_mems_cookie));
1504
1505 return NULL;
1506}
1507
1508/*
1509 * There are two ways to allocate a huge page:
1510 * 1. When you have a VMA and an address (like a fault)
1511 * 2. When you have no VMA (like when setting /proc/.../nr_hugepages)
1512 *
1513 * 'vma' and 'addr' are only for (1). 'nid' is always NUMA_NO_NODE in
1514 * this case which signifies that the allocation should be done with
1515 * respect for the VMA's memory policy.
1516 *
1517 * For (2), we ignore 'vma' and 'addr' and use 'nid' exclusively. This
1518 * implies that memory policies will not be taken in to account.
1519 */
1520static struct page *__alloc_buddy_huge_page(struct hstate *h,
1521 struct vm_area_struct *vma, unsigned long addr, int nid)
Adam Litke7893d1d2007-10-16 01:26:18 -07001522{
1523 struct page *page;
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001524 unsigned int r_nid;
Adam Litke7893d1d2007-10-16 01:26:18 -07001525
Luiz Capitulinobae7f4a2014-06-04 16:07:08 -07001526 if (hstate_is_gigantic(h))
Andi Kleenaa888a72008-07-23 21:27:47 -07001527 return NULL;
1528
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001529 /*
Dave Hansen099730d2015-11-05 18:50:17 -08001530 * Make sure that anyone specifying 'nid' is not also specifying a VMA.
1531 * This makes sure the caller is picking _one_ of the modes with which
1532 * we can call this function, not both.
1533 */
1534 if (vma || (addr != -1)) {
Dave Hansene0ec90e2015-11-05 18:50:20 -08001535 VM_WARN_ON_ONCE(addr == -1);
1536 VM_WARN_ON_ONCE(nid != NUMA_NO_NODE);
Dave Hansen099730d2015-11-05 18:50:17 -08001537 }
1538 /*
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001539 * Assume we will successfully allocate the surplus page to
1540 * prevent racing processes from causing the surplus to exceed
1541 * overcommit
1542 *
1543 * This however introduces a different race, where a process B
1544 * tries to grow the static hugepage pool while alloc_pages() is
1545 * called by process A. B will only examine the per-node
1546 * counters in determining if surplus huge pages can be
1547 * converted to normal huge pages in adjust_pool_surplus(). A
1548 * won't be able to increment the per-node counter, until the
1549 * lock is dropped by B, but B doesn't drop hugetlb_lock until
1550 * no more huge pages can be converted from surplus to normal
1551 * state (and doesn't try to convert again). Thus, we have a
1552 * case where a surplus huge page exists, the pool is grown, and
1553 * the surplus huge page still exists after, even though it
1554 * should just have been converted to a normal huge page. This
1555 * does not leak memory, though, as the hugepage will be freed
1556 * once it is out of use. It also does not allow the counters to
1557 * go out of whack in adjust_pool_surplus() as we don't modify
1558 * the node values until we've gotten the hugepage and only the
1559 * per-node value is checked there.
1560 */
1561 spin_lock(&hugetlb_lock);
Andi Kleena5516432008-07-23 21:27:41 -07001562 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001563 spin_unlock(&hugetlb_lock);
1564 return NULL;
1565 } else {
Andi Kleena5516432008-07-23 21:27:41 -07001566 h->nr_huge_pages++;
1567 h->surplus_huge_pages++;
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001568 }
1569 spin_unlock(&hugetlb_lock);
1570
Dave Hansen099730d2015-11-05 18:50:17 -08001571 page = __hugetlb_alloc_buddy_huge_page(h, vma, addr, nid);
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001572
1573 spin_lock(&hugetlb_lock);
Adam Litke7893d1d2007-10-16 01:26:18 -07001574 if (page) {
Aneesh Kumar K.V0edaecf2012-07-31 16:42:07 -07001575 INIT_LIST_HEAD(&page->lru);
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001576 r_nid = page_to_nid(page);
Kirill A. Shutemovf1e61552015-11-06 16:29:50 -08001577 set_compound_page_dtor(page, HUGETLB_PAGE_DTOR);
Aneesh Kumar K.V9dd540e2012-07-31 16:42:15 -07001578 set_hugetlb_cgroup(page, NULL);
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001579 /*
1580 * We incremented the global counters already
1581 */
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001582 h->nr_huge_pages_node[r_nid]++;
1583 h->surplus_huge_pages_node[r_nid]++;
Adam Litke3b116302008-04-28 02:13:06 -07001584 __count_vm_event(HTLB_BUDDY_PGALLOC);
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001585 } else {
Andi Kleena5516432008-07-23 21:27:41 -07001586 h->nr_huge_pages--;
1587 h->surplus_huge_pages--;
Adam Litke3b116302008-04-28 02:13:06 -07001588 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
Adam Litke7893d1d2007-10-16 01:26:18 -07001589 }
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001590 spin_unlock(&hugetlb_lock);
Adam Litke7893d1d2007-10-16 01:26:18 -07001591
1592 return page;
1593}
1594
Adam Litkee4e574b2007-10-16 01:26:19 -07001595/*
Dave Hansen099730d2015-11-05 18:50:17 -08001596 * Allocate a huge page from 'nid'. Note, 'nid' may be
1597 * NUMA_NO_NODE, which means that it may be allocated
1598 * anywhere.
1599 */
Dave Hansene0ec90e2015-11-05 18:50:20 -08001600static
Dave Hansen099730d2015-11-05 18:50:17 -08001601struct page *__alloc_buddy_huge_page_no_mpol(struct hstate *h, int nid)
1602{
1603 unsigned long addr = -1;
1604
1605 return __alloc_buddy_huge_page(h, NULL, addr, nid);
1606}
1607
1608/*
1609 * Use the VMA's mpolicy to allocate a huge page from the buddy.
1610 */
Dave Hansene0ec90e2015-11-05 18:50:20 -08001611static
Dave Hansen099730d2015-11-05 18:50:17 -08001612struct page *__alloc_buddy_huge_page_with_mpol(struct hstate *h,
1613 struct vm_area_struct *vma, unsigned long addr)
1614{
1615 return __alloc_buddy_huge_page(h, vma, addr, NUMA_NO_NODE);
1616}
1617
1618/*
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001619 * This allocation function is useful in the context where vma is irrelevant.
1620 * E.g. soft-offlining uses this function because it only cares physical
1621 * address of error page.
1622 */
1623struct page *alloc_huge_page_node(struct hstate *h, int nid)
1624{
Joonsoo Kim4ef91842013-09-11 14:21:51 -07001625 struct page *page = NULL;
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001626
1627 spin_lock(&hugetlb_lock);
Joonsoo Kim4ef91842013-09-11 14:21:51 -07001628 if (h->free_huge_pages - h->resv_huge_pages > 0)
1629 page = dequeue_huge_page_node(h, nid);
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001630 spin_unlock(&hugetlb_lock);
1631
Aneesh Kumar K.V94ae8ba2012-07-31 16:42:35 -07001632 if (!page)
Dave Hansen099730d2015-11-05 18:50:17 -08001633 page = __alloc_buddy_huge_page_no_mpol(h, nid);
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001634
1635 return page;
1636}
1637
1638/*
Lucas De Marchi25985ed2011-03-30 22:57:33 -03001639 * Increase the hugetlb pool such that it can accommodate a reservation
Adam Litkee4e574b2007-10-16 01:26:19 -07001640 * of size 'delta'.
1641 */
Andi Kleena5516432008-07-23 21:27:41 -07001642static int gather_surplus_pages(struct hstate *h, int delta)
Adam Litkee4e574b2007-10-16 01:26:19 -07001643{
1644 struct list_head surplus_list;
1645 struct page *page, *tmp;
1646 int ret, i;
1647 int needed, allocated;
Hillf Danton28073b02012-03-21 16:34:00 -07001648 bool alloc_ok = true;
Adam Litkee4e574b2007-10-16 01:26:19 -07001649
Andi Kleena5516432008-07-23 21:27:41 -07001650 needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
Adam Litkeac09b3a2008-03-04 14:29:38 -08001651 if (needed <= 0) {
Andi Kleena5516432008-07-23 21:27:41 -07001652 h->resv_huge_pages += delta;
Adam Litkee4e574b2007-10-16 01:26:19 -07001653 return 0;
Adam Litkeac09b3a2008-03-04 14:29:38 -08001654 }
Adam Litkee4e574b2007-10-16 01:26:19 -07001655
1656 allocated = 0;
1657 INIT_LIST_HEAD(&surplus_list);
1658
1659 ret = -ENOMEM;
1660retry:
1661 spin_unlock(&hugetlb_lock);
1662 for (i = 0; i < needed; i++) {
Dave Hansen099730d2015-11-05 18:50:17 -08001663 page = __alloc_buddy_huge_page_no_mpol(h, NUMA_NO_NODE);
Hillf Danton28073b02012-03-21 16:34:00 -07001664 if (!page) {
1665 alloc_ok = false;
1666 break;
1667 }
Adam Litkee4e574b2007-10-16 01:26:19 -07001668 list_add(&page->lru, &surplus_list);
1669 }
Hillf Danton28073b02012-03-21 16:34:00 -07001670 allocated += i;
Adam Litkee4e574b2007-10-16 01:26:19 -07001671
1672 /*
1673 * After retaking hugetlb_lock, we need to recalculate 'needed'
1674 * because either resv_huge_pages or free_huge_pages may have changed.
1675 */
1676 spin_lock(&hugetlb_lock);
Andi Kleena5516432008-07-23 21:27:41 -07001677 needed = (h->resv_huge_pages + delta) -
1678 (h->free_huge_pages + allocated);
Hillf Danton28073b02012-03-21 16:34:00 -07001679 if (needed > 0) {
1680 if (alloc_ok)
1681 goto retry;
1682 /*
1683 * We were not able to allocate enough pages to
1684 * satisfy the entire reservation so we free what
1685 * we've allocated so far.
1686 */
1687 goto free;
1688 }
Adam Litkee4e574b2007-10-16 01:26:19 -07001689 /*
1690 * The surplus_list now contains _at_least_ the number of extra pages
Lucas De Marchi25985ed2011-03-30 22:57:33 -03001691 * needed to accommodate the reservation. Add the appropriate number
Adam Litkee4e574b2007-10-16 01:26:19 -07001692 * of pages to the hugetlb pool and free the extras back to the buddy
Adam Litkeac09b3a2008-03-04 14:29:38 -08001693 * allocator. Commit the entire reservation here to prevent another
1694 * process from stealing the pages as they are added to the pool but
1695 * before they are reserved.
Adam Litkee4e574b2007-10-16 01:26:19 -07001696 */
1697 needed += allocated;
Andi Kleena5516432008-07-23 21:27:41 -07001698 h->resv_huge_pages += delta;
Adam Litkee4e574b2007-10-16 01:26:19 -07001699 ret = 0;
Naoya Horiguchia9869b82010-09-08 10:19:37 +09001700
Adam Litke19fc3f02008-04-28 02:12:20 -07001701 /* Free the needed pages to the hugetlb pool */
Adam Litkee4e574b2007-10-16 01:26:19 -07001702 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
Adam Litke19fc3f02008-04-28 02:12:20 -07001703 if ((--needed) < 0)
1704 break;
Naoya Horiguchia9869b82010-09-08 10:19:37 +09001705 /*
1706 * This page is now managed by the hugetlb allocator and has
1707 * no users -- drop the buddy allocator's reference.
1708 */
1709 put_page_testzero(page);
Sasha Levin309381fea2014-01-23 15:52:54 -08001710 VM_BUG_ON_PAGE(page_count(page), page);
Andi Kleena5516432008-07-23 21:27:41 -07001711 enqueue_huge_page(h, page);
Adam Litke19fc3f02008-04-28 02:12:20 -07001712 }
Hillf Danton28073b02012-03-21 16:34:00 -07001713free:
Hillf Dantonb0365c82011-12-28 15:57:16 -08001714 spin_unlock(&hugetlb_lock);
Adam Litke19fc3f02008-04-28 02:12:20 -07001715
1716 /* Free unnecessary surplus pages to the buddy allocator */
Joonsoo Kimc0d934b2013-09-11 14:21:02 -07001717 list_for_each_entry_safe(page, tmp, &surplus_list, lru)
1718 put_page(page);
Naoya Horiguchia9869b82010-09-08 10:19:37 +09001719 spin_lock(&hugetlb_lock);
Adam Litkee4e574b2007-10-16 01:26:19 -07001720
1721 return ret;
1722}
1723
1724/*
1725 * When releasing a hugetlb pool reservation, any surplus pages that were
1726 * allocated to satisfy the reservation must be explicitly freed if they were
1727 * never used.
Lee Schermerhorn685f3452009-09-21 17:01:23 -07001728 * Called with hugetlb_lock held.
Adam Litkee4e574b2007-10-16 01:26:19 -07001729 */
Andi Kleena5516432008-07-23 21:27:41 -07001730static void return_unused_surplus_pages(struct hstate *h,
1731 unsigned long unused_resv_pages)
Adam Litkee4e574b2007-10-16 01:26:19 -07001732{
Adam Litkee4e574b2007-10-16 01:26:19 -07001733 unsigned long nr_pages;
1734
Adam Litkeac09b3a2008-03-04 14:29:38 -08001735 /* Uncommit the reservation */
Andi Kleena5516432008-07-23 21:27:41 -07001736 h->resv_huge_pages -= unused_resv_pages;
Adam Litkeac09b3a2008-03-04 14:29:38 -08001737
Andi Kleenaa888a72008-07-23 21:27:47 -07001738 /* Cannot return gigantic pages currently */
Luiz Capitulinobae7f4a2014-06-04 16:07:08 -07001739 if (hstate_is_gigantic(h))
Andi Kleenaa888a72008-07-23 21:27:47 -07001740 return;
1741
Andi Kleena5516432008-07-23 21:27:41 -07001742 nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
Adam Litkee4e574b2007-10-16 01:26:19 -07001743
Lee Schermerhorn685f3452009-09-21 17:01:23 -07001744 /*
1745 * We want to release as many surplus pages as possible, spread
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -08001746 * evenly across all nodes with memory. Iterate across these nodes
1747 * until we can no longer free unreserved surplus pages. This occurs
1748 * when the nodes with surplus pages have no free pages.
1749 * free_pool_huge_page() will balance the the freed pages across the
1750 * on-line nodes with memory and will handle the hstate accounting.
Lee Schermerhorn685f3452009-09-21 17:01:23 -07001751 */
1752 while (nr_pages--) {
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08001753 if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1))
Lee Schermerhorn685f3452009-09-21 17:01:23 -07001754 break;
Mizuma, Masayoshi7848a4b2014-04-18 15:07:18 -07001755 cond_resched_lock(&hugetlb_lock);
Adam Litkee4e574b2007-10-16 01:26:19 -07001756 }
1757}
1758
Mike Kravetz5e911372015-09-08 15:01:28 -07001759
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001760/*
Mike Kravetzfeba16e2015-09-08 15:01:31 -07001761 * vma_needs_reservation, vma_commit_reservation and vma_end_reservation
Mike Kravetz5e911372015-09-08 15:01:28 -07001762 * are used by the huge page allocation routines to manage reservations.
Mike Kravetzcf3ad202015-06-24 16:57:55 -07001763 *
1764 * vma_needs_reservation is called to determine if the huge page at addr
1765 * within the vma has an associated reservation. If a reservation is
1766 * needed, the value 1 is returned. The caller is then responsible for
1767 * managing the global reservation and subpool usage counts. After
1768 * the huge page has been allocated, vma_commit_reservation is called
Mike Kravetzfeba16e2015-09-08 15:01:31 -07001769 * to add the page to the reservation map. If the page allocation fails,
1770 * the reservation must be ended instead of committed. vma_end_reservation
1771 * is called in such cases.
Mike Kravetzcf3ad202015-06-24 16:57:55 -07001772 *
1773 * In the normal case, vma_commit_reservation returns the same value
1774 * as the preceding vma_needs_reservation call. The only time this
1775 * is not the case is if a reserve map was changed between calls. It
1776 * is the responsibility of the caller to notice the difference and
1777 * take appropriate action.
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001778 */
Mike Kravetz5e911372015-09-08 15:01:28 -07001779enum vma_resv_mode {
1780 VMA_NEEDS_RESV,
1781 VMA_COMMIT_RESV,
Mike Kravetzfeba16e2015-09-08 15:01:31 -07001782 VMA_END_RESV,
Mike Kravetz5e911372015-09-08 15:01:28 -07001783};
Mike Kravetzcf3ad202015-06-24 16:57:55 -07001784static long __vma_reservation_common(struct hstate *h,
1785 struct vm_area_struct *vma, unsigned long addr,
Mike Kravetz5e911372015-09-08 15:01:28 -07001786 enum vma_resv_mode mode)
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001787{
Joonsoo Kim4e35f482014-04-03 14:47:30 -07001788 struct resv_map *resv;
1789 pgoff_t idx;
Mike Kravetzcf3ad202015-06-24 16:57:55 -07001790 long ret;
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001791
Joonsoo Kim4e35f482014-04-03 14:47:30 -07001792 resv = vma_resv_map(vma);
1793 if (!resv)
Andy Whitcroft84afd992008-07-23 21:27:32 -07001794 return 1;
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001795
Joonsoo Kim4e35f482014-04-03 14:47:30 -07001796 idx = vma_hugecache_offset(h, vma, addr);
Mike Kravetz5e911372015-09-08 15:01:28 -07001797 switch (mode) {
1798 case VMA_NEEDS_RESV:
Mike Kravetzcf3ad202015-06-24 16:57:55 -07001799 ret = region_chg(resv, idx, idx + 1);
Mike Kravetz5e911372015-09-08 15:01:28 -07001800 break;
1801 case VMA_COMMIT_RESV:
1802 ret = region_add(resv, idx, idx + 1);
1803 break;
Mike Kravetzfeba16e2015-09-08 15:01:31 -07001804 case VMA_END_RESV:
Mike Kravetz5e911372015-09-08 15:01:28 -07001805 region_abort(resv, idx, idx + 1);
1806 ret = 0;
1807 break;
1808 default:
1809 BUG();
1810 }
Andy Whitcroft84afd992008-07-23 21:27:32 -07001811
Joonsoo Kim4e35f482014-04-03 14:47:30 -07001812 if (vma->vm_flags & VM_MAYSHARE)
Mike Kravetzcf3ad202015-06-24 16:57:55 -07001813 return ret;
Joonsoo Kim4e35f482014-04-03 14:47:30 -07001814 else
Mike Kravetzcf3ad202015-06-24 16:57:55 -07001815 return ret < 0 ? ret : 0;
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001816}
Mike Kravetzcf3ad202015-06-24 16:57:55 -07001817
1818static long vma_needs_reservation(struct hstate *h,
Andi Kleena5516432008-07-23 21:27:41 -07001819 struct vm_area_struct *vma, unsigned long addr)
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001820{
Mike Kravetz5e911372015-09-08 15:01:28 -07001821 return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV);
Mike Kravetzcf3ad202015-06-24 16:57:55 -07001822}
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001823
Mike Kravetzcf3ad202015-06-24 16:57:55 -07001824static long vma_commit_reservation(struct hstate *h,
1825 struct vm_area_struct *vma, unsigned long addr)
1826{
Mike Kravetz5e911372015-09-08 15:01:28 -07001827 return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV);
1828}
1829
Mike Kravetzfeba16e2015-09-08 15:01:31 -07001830static void vma_end_reservation(struct hstate *h,
Mike Kravetz5e911372015-09-08 15:01:28 -07001831 struct vm_area_struct *vma, unsigned long addr)
1832{
Mike Kravetzfeba16e2015-09-08 15:01:31 -07001833 (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV);
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001834}
1835
Mike Kravetz70c35472015-09-08 15:01:54 -07001836struct page *alloc_huge_page(struct vm_area_struct *vma,
Mel Gorman04f2cbe2008-07-23 21:27:25 -07001837 unsigned long addr, int avoid_reserve)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001838{
David Gibson90481622012-03-21 16:34:12 -07001839 struct hugepage_subpool *spool = subpool_vma(vma);
Andi Kleena5516432008-07-23 21:27:41 -07001840 struct hstate *h = hstate_vma(vma);
Adam Litke348ea202007-11-14 16:59:37 -08001841 struct page *page;
Mike Kravetzd85f69b2015-09-08 15:01:47 -07001842 long map_chg, map_commit;
1843 long gbl_chg;
Aneesh Kumar K.V6d76dcf2012-07-31 16:42:18 -07001844 int ret, idx;
1845 struct hugetlb_cgroup *h_cg;
Adam Litke2fc39ce2007-11-14 16:59:39 -08001846
Aneesh Kumar K.V6d76dcf2012-07-31 16:42:18 -07001847 idx = hstate_index(h);
Mel Gormana1e78772008-07-23 21:27:23 -07001848 /*
Mike Kravetzd85f69b2015-09-08 15:01:47 -07001849 * Examine the region/reserve map to determine if the process
1850 * has a reservation for the page to be allocated. A return
1851 * code of zero indicates a reservation exists (no change).
Mel Gormana1e78772008-07-23 21:27:23 -07001852 */
Mike Kravetzd85f69b2015-09-08 15:01:47 -07001853 map_chg = gbl_chg = vma_needs_reservation(h, vma, addr);
1854 if (map_chg < 0)
Aneesh Kumar K.V76dcee72012-07-31 16:41:57 -07001855 return ERR_PTR(-ENOMEM);
Mike Kravetzd85f69b2015-09-08 15:01:47 -07001856
1857 /*
1858 * Processes that did not create the mapping will have no
1859 * reserves as indicated by the region/reserve map. Check
1860 * that the allocation will not exceed the subpool limit.
1861 * Allocations for MAP_NORESERVE mappings also need to be
1862 * checked against any subpool limit.
1863 */
1864 if (map_chg || avoid_reserve) {
1865 gbl_chg = hugepage_subpool_get_pages(spool, 1);
1866 if (gbl_chg < 0) {
Mike Kravetzfeba16e2015-09-08 15:01:31 -07001867 vma_end_reservation(h, vma, addr);
Aneesh Kumar K.V76dcee72012-07-31 16:41:57 -07001868 return ERR_PTR(-ENOSPC);
Mike Kravetz5e911372015-09-08 15:01:28 -07001869 }
Mel Gormana1e78772008-07-23 21:27:23 -07001870
Mike Kravetzd85f69b2015-09-08 15:01:47 -07001871 /*
1872 * Even though there was no reservation in the region/reserve
1873 * map, there could be reservations associated with the
1874 * subpool that can be used. This would be indicated if the
1875 * return value of hugepage_subpool_get_pages() is zero.
1876 * However, if avoid_reserve is specified we still avoid even
1877 * the subpool reservations.
1878 */
1879 if (avoid_reserve)
1880 gbl_chg = 1;
1881 }
1882
Aneesh Kumar K.V6d76dcf2012-07-31 16:42:18 -07001883 ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg);
Jianyu Zhan8f34af62014-06-04 16:10:36 -07001884 if (ret)
1885 goto out_subpool_put;
1886
Mel Gormana1e78772008-07-23 21:27:23 -07001887 spin_lock(&hugetlb_lock);
Mike Kravetzd85f69b2015-09-08 15:01:47 -07001888 /*
1889 * glb_chg is passed to indicate whether or not a page must be taken
1890 * from the global free pool (global change). gbl_chg == 0 indicates
1891 * a reservation exists for the allocation.
1892 */
1893 page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg);
Joonsoo Kim81a6fca2013-09-11 14:20:58 -07001894 if (!page) {
Aneesh Kumar K.V94ae8ba2012-07-31 16:42:35 -07001895 spin_unlock(&hugetlb_lock);
Dave Hansen099730d2015-11-05 18:50:17 -08001896 page = __alloc_buddy_huge_page_with_mpol(h, vma, addr);
Jianyu Zhan8f34af62014-06-04 16:10:36 -07001897 if (!page)
1898 goto out_uncharge_cgroup;
Naoya Horiguchia88c7692015-12-11 13:40:24 -08001899 if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) {
1900 SetPagePrivate(page);
1901 h->resv_huge_pages--;
1902 }
Aneesh Kumar K.V79dbb232012-07-31 16:42:32 -07001903 spin_lock(&hugetlb_lock);
1904 list_move(&page->lru, &h->hugepage_activelist);
Joonsoo Kim81a6fca2013-09-11 14:20:58 -07001905 /* Fall through */
Mel Gormana1e78772008-07-23 21:27:23 -07001906 }
Joonsoo Kim81a6fca2013-09-11 14:20:58 -07001907 hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page);
1908 spin_unlock(&hugetlb_lock);
Mel Gormana1e78772008-07-23 21:27:23 -07001909
David Gibson90481622012-03-21 16:34:12 -07001910 set_page_private(page, (unsigned long)spool);
Mel Gormana1e78772008-07-23 21:27:23 -07001911
Mike Kravetzd85f69b2015-09-08 15:01:47 -07001912 map_commit = vma_commit_reservation(h, vma, addr);
1913 if (unlikely(map_chg > map_commit)) {
Mike Kravetz33039672015-06-24 16:57:58 -07001914 /*
1915 * The page was added to the reservation map between
1916 * vma_needs_reservation and vma_commit_reservation.
1917 * This indicates a race with hugetlb_reserve_pages.
1918 * Adjust for the subpool count incremented above AND
1919 * in hugetlb_reserve_pages for the same page. Also,
1920 * the reservation count added in hugetlb_reserve_pages
1921 * no longer applies.
1922 */
1923 long rsv_adjust;
1924
1925 rsv_adjust = hugepage_subpool_put_pages(spool, 1);
1926 hugetlb_acct_memory(h, -rsv_adjust);
1927 }
Adam Litke90d8b7e2007-11-14 16:59:42 -08001928 return page;
Jianyu Zhan8f34af62014-06-04 16:10:36 -07001929
1930out_uncharge_cgroup:
1931 hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg);
1932out_subpool_put:
Mike Kravetzd85f69b2015-09-08 15:01:47 -07001933 if (map_chg || avoid_reserve)
Jianyu Zhan8f34af62014-06-04 16:10:36 -07001934 hugepage_subpool_put_pages(spool, 1);
Mike Kravetzfeba16e2015-09-08 15:01:31 -07001935 vma_end_reservation(h, vma, addr);
Jianyu Zhan8f34af62014-06-04 16:10:36 -07001936 return ERR_PTR(-ENOSPC);
David Gibsonb45b5bd2006-03-22 00:08:55 -08001937}
1938
Naoya Horiguchi74060e42013-09-11 14:22:06 -07001939/*
1940 * alloc_huge_page()'s wrapper which simply returns the page if allocation
1941 * succeeds, otherwise NULL. This function is called from new_vma_page(),
1942 * where no ERR_VALUE is expected to be returned.
1943 */
1944struct page *alloc_huge_page_noerr(struct vm_area_struct *vma,
1945 unsigned long addr, int avoid_reserve)
1946{
1947 struct page *page = alloc_huge_page(vma, addr, avoid_reserve);
1948 if (IS_ERR(page))
1949 page = NULL;
1950 return page;
1951}
1952
Cyrill Gorcunov91f47662009-01-06 14:40:33 -08001953int __weak alloc_bootmem_huge_page(struct hstate *h)
Andi Kleenaa888a72008-07-23 21:27:47 -07001954{
1955 struct huge_bootmem_page *m;
Joonsoo Kimb2261022013-09-11 14:21:00 -07001956 int nr_nodes, node;
Andi Kleenaa888a72008-07-23 21:27:47 -07001957
Joonsoo Kimb2261022013-09-11 14:21:00 -07001958 for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) {
Andi Kleenaa888a72008-07-23 21:27:47 -07001959 void *addr;
1960
Grygorii Strashko8b89a112014-01-21 15:50:36 -08001961 addr = memblock_virt_alloc_try_nid_nopanic(
1962 huge_page_size(h), huge_page_size(h),
1963 0, BOOTMEM_ALLOC_ACCESSIBLE, node);
Andi Kleenaa888a72008-07-23 21:27:47 -07001964 if (addr) {
1965 /*
1966 * Use the beginning of the huge page to store the
1967 * huge_bootmem_page struct (until gather_bootmem
1968 * puts them into the mem_map).
1969 */
1970 m = addr;
Cyrill Gorcunov91f47662009-01-06 14:40:33 -08001971 goto found;
Andi Kleenaa888a72008-07-23 21:27:47 -07001972 }
Andi Kleenaa888a72008-07-23 21:27:47 -07001973 }
1974 return 0;
1975
1976found:
Luiz Capitulinodf994ea2014-12-12 16:55:21 -08001977 BUG_ON(!IS_ALIGNED(virt_to_phys(m), huge_page_size(h)));
Andi Kleenaa888a72008-07-23 21:27:47 -07001978 /* Put them into a private list first because mem_map is not up yet */
1979 list_add(&m->list, &huge_boot_pages);
1980 m->hstate = h;
1981 return 1;
1982}
1983
Kirill A. Shutemovd00181b2015-11-06 16:29:57 -08001984static void __init prep_compound_huge_page(struct page *page,
1985 unsigned int order)
Andy Whitcroft18229df2008-11-06 12:53:27 -08001986{
1987 if (unlikely(order > (MAX_ORDER - 1)))
1988 prep_compound_gigantic_page(page, order);
1989 else
1990 prep_compound_page(page, order);
1991}
1992
Andi Kleenaa888a72008-07-23 21:27:47 -07001993/* Put bootmem huge pages into the standard lists after mem_map is up */
1994static void __init gather_bootmem_prealloc(void)
1995{
1996 struct huge_bootmem_page *m;
1997
1998 list_for_each_entry(m, &huge_boot_pages, list) {
Andi Kleenaa888a72008-07-23 21:27:47 -07001999 struct hstate *h = m->hstate;
Becky Bruceee8f2482011-07-25 17:11:50 -07002000 struct page *page;
2001
2002#ifdef CONFIG_HIGHMEM
2003 page = pfn_to_page(m->phys >> PAGE_SHIFT);
Grygorii Strashko8b89a112014-01-21 15:50:36 -08002004 memblock_free_late(__pa(m),
2005 sizeof(struct huge_bootmem_page));
Becky Bruceee8f2482011-07-25 17:11:50 -07002006#else
2007 page = virt_to_page(m);
2008#endif
Andi Kleenaa888a72008-07-23 21:27:47 -07002009 WARN_ON(page_count(page) != 1);
Andy Whitcroft18229df2008-11-06 12:53:27 -08002010 prep_compound_huge_page(page, h->order);
Andrea Arcangelief5a22b2013-10-16 13:46:56 -07002011 WARN_ON(PageReserved(page));
Andi Kleenaa888a72008-07-23 21:27:47 -07002012 prep_new_huge_page(h, page, page_to_nid(page));
Rafael Aquinib0320c72011-06-15 15:08:39 -07002013 /*
2014 * If we had gigantic hugepages allocated at boot time, we need
2015 * to restore the 'stolen' pages to totalram_pages in order to
2016 * fix confusing memory reports from free(1) and another
2017 * side-effects, like CommitLimit going negative.
2018 */
Luiz Capitulinobae7f4a2014-06-04 16:07:08 -07002019 if (hstate_is_gigantic(h))
Jiang Liu3dcc0572013-07-03 15:03:21 -07002020 adjust_managed_page_count(page, 1 << h->order);
Andi Kleenaa888a72008-07-23 21:27:47 -07002021 }
2022}
2023
Andi Kleen8faa8b02008-07-23 21:27:48 -07002024static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002025{
2026 unsigned long i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002027
Andi Kleene5ff2152008-07-23 21:27:42 -07002028 for (i = 0; i < h->max_huge_pages; ++i) {
Luiz Capitulinobae7f4a2014-06-04 16:07:08 -07002029 if (hstate_is_gigantic(h)) {
Andi Kleenaa888a72008-07-23 21:27:47 -07002030 if (!alloc_bootmem_huge_page(h))
2031 break;
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -08002032 } else if (!alloc_fresh_huge_page(h,
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08002033 &node_states[N_MEMORY]))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002034 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002035 }
Andi Kleen8faa8b02008-07-23 21:27:48 -07002036 h->max_huge_pages = i;
Andi Kleene5ff2152008-07-23 21:27:42 -07002037}
2038
2039static void __init hugetlb_init_hstates(void)
2040{
2041 struct hstate *h;
2042
2043 for_each_hstate(h) {
Naoya Horiguchi641844f2015-06-24 16:56:59 -07002044 if (minimum_order > huge_page_order(h))
2045 minimum_order = huge_page_order(h);
2046
Andi Kleen8faa8b02008-07-23 21:27:48 -07002047 /* oversize hugepages were init'ed in early boot */
Luiz Capitulinobae7f4a2014-06-04 16:07:08 -07002048 if (!hstate_is_gigantic(h))
Andi Kleen8faa8b02008-07-23 21:27:48 -07002049 hugetlb_hstate_alloc_pages(h);
Andi Kleene5ff2152008-07-23 21:27:42 -07002050 }
Naoya Horiguchi641844f2015-06-24 16:56:59 -07002051 VM_BUG_ON(minimum_order == UINT_MAX);
Andi Kleene5ff2152008-07-23 21:27:42 -07002052}
2053
Andi Kleen4abd32d2008-07-23 21:27:49 -07002054static char * __init memfmt(char *buf, unsigned long n)
2055{
2056 if (n >= (1UL << 30))
2057 sprintf(buf, "%lu GB", n >> 30);
2058 else if (n >= (1UL << 20))
2059 sprintf(buf, "%lu MB", n >> 20);
2060 else
2061 sprintf(buf, "%lu KB", n >> 10);
2062 return buf;
2063}
2064
Andi Kleene5ff2152008-07-23 21:27:42 -07002065static void __init report_hugepages(void)
2066{
2067 struct hstate *h;
2068
2069 for_each_hstate(h) {
Andi Kleen4abd32d2008-07-23 21:27:49 -07002070 char buf[32];
Andrew Mortonffb22af2013-02-22 16:32:08 -08002071 pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n",
Andi Kleen4abd32d2008-07-23 21:27:49 -07002072 memfmt(buf, huge_page_size(h)),
2073 h->free_huge_pages);
Andi Kleene5ff2152008-07-23 21:27:42 -07002074 }
2075}
2076
Linus Torvalds1da177e2005-04-16 15:20:36 -07002077#ifdef CONFIG_HIGHMEM
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08002078static void try_to_free_low(struct hstate *h, unsigned long count,
2079 nodemask_t *nodes_allowed)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002080{
Christoph Lameter4415cc82006-09-25 23:31:55 -07002081 int i;
2082
Luiz Capitulinobae7f4a2014-06-04 16:07:08 -07002083 if (hstate_is_gigantic(h))
Andi Kleenaa888a72008-07-23 21:27:47 -07002084 return;
2085
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08002086 for_each_node_mask(i, *nodes_allowed) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002087 struct page *page, *next;
Andi Kleena5516432008-07-23 21:27:41 -07002088 struct list_head *freel = &h->hugepage_freelists[i];
2089 list_for_each_entry_safe(page, next, freel, lru) {
2090 if (count >= h->nr_huge_pages)
Adam Litke6b0c8802007-10-16 01:26:23 -07002091 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002092 if (PageHighMem(page))
2093 continue;
2094 list_del(&page->lru);
Andi Kleene5ff2152008-07-23 21:27:42 -07002095 update_and_free_page(h, page);
Andi Kleena5516432008-07-23 21:27:41 -07002096 h->free_huge_pages--;
2097 h->free_huge_pages_node[page_to_nid(page)]--;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002098 }
2099 }
2100}
2101#else
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08002102static inline void try_to_free_low(struct hstate *h, unsigned long count,
2103 nodemask_t *nodes_allowed)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002104{
2105}
2106#endif
2107
Wu Fengguang20a03072009-06-16 15:32:22 -07002108/*
2109 * Increment or decrement surplus_huge_pages. Keep node-specific counters
2110 * balanced by operating on them in a round-robin fashion.
2111 * Returns 1 if an adjustment was made.
2112 */
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08002113static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
2114 int delta)
Wu Fengguang20a03072009-06-16 15:32:22 -07002115{
Joonsoo Kimb2261022013-09-11 14:21:00 -07002116 int nr_nodes, node;
Wu Fengguang20a03072009-06-16 15:32:22 -07002117
2118 VM_BUG_ON(delta != -1 && delta != 1);
Wu Fengguang20a03072009-06-16 15:32:22 -07002119
Joonsoo Kimb2261022013-09-11 14:21:00 -07002120 if (delta < 0) {
2121 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
2122 if (h->surplus_huge_pages_node[node])
2123 goto found;
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07002124 }
Joonsoo Kimb2261022013-09-11 14:21:00 -07002125 } else {
2126 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
2127 if (h->surplus_huge_pages_node[node] <
2128 h->nr_huge_pages_node[node])
2129 goto found;
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07002130 }
Joonsoo Kimb2261022013-09-11 14:21:00 -07002131 }
2132 return 0;
Wu Fengguang20a03072009-06-16 15:32:22 -07002133
Joonsoo Kimb2261022013-09-11 14:21:00 -07002134found:
2135 h->surplus_huge_pages += delta;
2136 h->surplus_huge_pages_node[node] += delta;
2137 return 1;
Wu Fengguang20a03072009-06-16 15:32:22 -07002138}
2139
Andi Kleena5516432008-07-23 21:27:41 -07002140#define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08002141static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count,
2142 nodemask_t *nodes_allowed)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002143{
Adam Litke7893d1d2007-10-16 01:26:18 -07002144 unsigned long min_count, ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002145
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07002146 if (hstate_is_gigantic(h) && !gigantic_page_supported())
Andi Kleenaa888a72008-07-23 21:27:47 -07002147 return h->max_huge_pages;
2148
Adam Litke7893d1d2007-10-16 01:26:18 -07002149 /*
2150 * Increase the pool size
2151 * First take pages out of surplus state. Then make up the
2152 * remaining difference by allocating fresh huge pages.
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08002153 *
Naoya Horiguchid15c7c02015-11-10 14:45:11 -08002154 * We might race with __alloc_buddy_huge_page() here and be unable
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08002155 * to convert a surplus huge page to a normal huge page. That is
2156 * not critical, though, it just means the overall size of the
2157 * pool might be one hugepage larger than it needs to be, but
2158 * within all the constraints specified by the sysctls.
Adam Litke7893d1d2007-10-16 01:26:18 -07002159 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002160 spin_lock(&hugetlb_lock);
Andi Kleena5516432008-07-23 21:27:41 -07002161 while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08002162 if (!adjust_pool_surplus(h, nodes_allowed, -1))
Adam Litke7893d1d2007-10-16 01:26:18 -07002163 break;
2164 }
2165
Andi Kleena5516432008-07-23 21:27:41 -07002166 while (count > persistent_huge_pages(h)) {
Adam Litke7893d1d2007-10-16 01:26:18 -07002167 /*
2168 * If this allocation races such that we no longer need the
2169 * page, free_huge_page will handle it by freeing the page
2170 * and reducing the surplus.
2171 */
2172 spin_unlock(&hugetlb_lock);
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07002173 if (hstate_is_gigantic(h))
2174 ret = alloc_fresh_gigantic_page(h, nodes_allowed);
2175 else
2176 ret = alloc_fresh_huge_page(h, nodes_allowed);
Adam Litke7893d1d2007-10-16 01:26:18 -07002177 spin_lock(&hugetlb_lock);
2178 if (!ret)
2179 goto out;
2180
Mel Gorman536240f22009-12-14 17:59:56 -08002181 /* Bail for signals. Probably ctrl-c from user */
2182 if (signal_pending(current))
2183 goto out;
Adam Litke7893d1d2007-10-16 01:26:18 -07002184 }
Adam Litke7893d1d2007-10-16 01:26:18 -07002185
2186 /*
2187 * Decrease the pool size
2188 * First return free pages to the buddy allocator (being careful
2189 * to keep enough around to satisfy reservations). Then place
2190 * pages into surplus state as needed so the pool will shrink
2191 * to the desired size as pages become free.
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08002192 *
2193 * By placing pages into the surplus state independent of the
2194 * overcommit value, we are allowing the surplus pool size to
2195 * exceed overcommit. There are few sane options here. Since
Naoya Horiguchid15c7c02015-11-10 14:45:11 -08002196 * __alloc_buddy_huge_page() is checking the global counter,
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08002197 * though, we'll note that we're not allowed to exceed surplus
2198 * and won't grow the pool anywhere else. Not until one of the
2199 * sysctls are changed, or the surplus pages go out of use.
Adam Litke7893d1d2007-10-16 01:26:18 -07002200 */
Andi Kleena5516432008-07-23 21:27:41 -07002201 min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
Adam Litke6b0c8802007-10-16 01:26:23 -07002202 min_count = max(count, min_count);
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08002203 try_to_free_low(h, min_count, nodes_allowed);
Andi Kleena5516432008-07-23 21:27:41 -07002204 while (min_count < persistent_huge_pages(h)) {
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08002205 if (!free_pool_huge_page(h, nodes_allowed, 0))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002206 break;
Mizuma, Masayoshi55f67142014-04-07 15:37:54 -07002207 cond_resched_lock(&hugetlb_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002208 }
Andi Kleena5516432008-07-23 21:27:41 -07002209 while (count < persistent_huge_pages(h)) {
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08002210 if (!adjust_pool_surplus(h, nodes_allowed, 1))
Adam Litke7893d1d2007-10-16 01:26:18 -07002211 break;
2212 }
2213out:
Andi Kleena5516432008-07-23 21:27:41 -07002214 ret = persistent_huge_pages(h);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002215 spin_unlock(&hugetlb_lock);
Adam Litke7893d1d2007-10-16 01:26:18 -07002216 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002217}
2218
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002219#define HSTATE_ATTR_RO(_name) \
2220 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
2221
2222#define HSTATE_ATTR(_name) \
2223 static struct kobj_attribute _name##_attr = \
2224 __ATTR(_name, 0644, _name##_show, _name##_store)
2225
2226static struct kobject *hugepages_kobj;
2227static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
2228
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002229static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
2230
2231static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002232{
2233 int i;
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002234
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002235 for (i = 0; i < HUGE_MAX_HSTATE; i++)
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002236 if (hstate_kobjs[i] == kobj) {
2237 if (nidp)
2238 *nidp = NUMA_NO_NODE;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002239 return &hstates[i];
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002240 }
2241
2242 return kobj_to_node_hstate(kobj, nidp);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002243}
2244
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002245static ssize_t nr_hugepages_show_common(struct kobject *kobj,
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002246 struct kobj_attribute *attr, char *buf)
2247{
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002248 struct hstate *h;
2249 unsigned long nr_huge_pages;
2250 int nid;
2251
2252 h = kobj_to_hstate(kobj, &nid);
2253 if (nid == NUMA_NO_NODE)
2254 nr_huge_pages = h->nr_huge_pages;
2255 else
2256 nr_huge_pages = h->nr_huge_pages_node[nid];
2257
2258 return sprintf(buf, "%lu\n", nr_huge_pages);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002259}
Eric B Munsonadbe8722011-01-13 15:47:27 -08002260
David Rientjes238d3c12014-08-06 16:06:51 -07002261static ssize_t __nr_hugepages_store_common(bool obey_mempolicy,
2262 struct hstate *h, int nid,
2263 unsigned long count, size_t len)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002264{
2265 int err;
David Rientjesbad44b52009-12-14 17:58:38 -08002266 NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002267
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07002268 if (hstate_is_gigantic(h) && !gigantic_page_supported()) {
Eric B Munsonadbe8722011-01-13 15:47:27 -08002269 err = -EINVAL;
2270 goto out;
2271 }
2272
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002273 if (nid == NUMA_NO_NODE) {
2274 /*
2275 * global hstate attribute
2276 */
2277 if (!(obey_mempolicy &&
2278 init_nodemask_of_mempolicy(nodes_allowed))) {
2279 NODEMASK_FREE(nodes_allowed);
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08002280 nodes_allowed = &node_states[N_MEMORY];
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002281 }
2282 } else if (nodes_allowed) {
2283 /*
2284 * per node hstate attribute: adjust count to global,
2285 * but restrict alloc/free to the specified node.
2286 */
2287 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
2288 init_nodemask_of_node(nodes_allowed, nid);
2289 } else
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08002290 nodes_allowed = &node_states[N_MEMORY];
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002291
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002292 h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002293
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08002294 if (nodes_allowed != &node_states[N_MEMORY])
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002295 NODEMASK_FREE(nodes_allowed);
2296
2297 return len;
Eric B Munsonadbe8722011-01-13 15:47:27 -08002298out:
2299 NODEMASK_FREE(nodes_allowed);
2300 return err;
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002301}
2302
David Rientjes238d3c12014-08-06 16:06:51 -07002303static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
2304 struct kobject *kobj, const char *buf,
2305 size_t len)
2306{
2307 struct hstate *h;
2308 unsigned long count;
2309 int nid;
2310 int err;
2311
2312 err = kstrtoul(buf, 10, &count);
2313 if (err)
2314 return err;
2315
2316 h = kobj_to_hstate(kobj, &nid);
2317 return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len);
2318}
2319
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002320static ssize_t nr_hugepages_show(struct kobject *kobj,
2321 struct kobj_attribute *attr, char *buf)
2322{
2323 return nr_hugepages_show_common(kobj, attr, buf);
2324}
2325
2326static ssize_t nr_hugepages_store(struct kobject *kobj,
2327 struct kobj_attribute *attr, const char *buf, size_t len)
2328{
David Rientjes238d3c12014-08-06 16:06:51 -07002329 return nr_hugepages_store_common(false, kobj, buf, len);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002330}
2331HSTATE_ATTR(nr_hugepages);
2332
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002333#ifdef CONFIG_NUMA
2334
2335/*
2336 * hstate attribute for optionally mempolicy-based constraint on persistent
2337 * huge page alloc/free.
2338 */
2339static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
2340 struct kobj_attribute *attr, char *buf)
2341{
2342 return nr_hugepages_show_common(kobj, attr, buf);
2343}
2344
2345static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
2346 struct kobj_attribute *attr, const char *buf, size_t len)
2347{
David Rientjes238d3c12014-08-06 16:06:51 -07002348 return nr_hugepages_store_common(true, kobj, buf, len);
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002349}
2350HSTATE_ATTR(nr_hugepages_mempolicy);
2351#endif
2352
2353
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002354static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
2355 struct kobj_attribute *attr, char *buf)
2356{
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002357 struct hstate *h = kobj_to_hstate(kobj, NULL);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002358 return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
2359}
Eric B Munsonadbe8722011-01-13 15:47:27 -08002360
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002361static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
2362 struct kobj_attribute *attr, const char *buf, size_t count)
2363{
2364 int err;
2365 unsigned long input;
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002366 struct hstate *h = kobj_to_hstate(kobj, NULL);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002367
Luiz Capitulinobae7f4a2014-06-04 16:07:08 -07002368 if (hstate_is_gigantic(h))
Eric B Munsonadbe8722011-01-13 15:47:27 -08002369 return -EINVAL;
2370
Jingoo Han3dbb95f2013-09-11 14:20:25 -07002371 err = kstrtoul(buf, 10, &input);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002372 if (err)
Eric B Munson73ae31e2011-01-13 15:47:28 -08002373 return err;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002374
2375 spin_lock(&hugetlb_lock);
2376 h->nr_overcommit_huge_pages = input;
2377 spin_unlock(&hugetlb_lock);
2378
2379 return count;
2380}
2381HSTATE_ATTR(nr_overcommit_hugepages);
2382
2383static ssize_t free_hugepages_show(struct kobject *kobj,
2384 struct kobj_attribute *attr, char *buf)
2385{
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002386 struct hstate *h;
2387 unsigned long free_huge_pages;
2388 int nid;
2389
2390 h = kobj_to_hstate(kobj, &nid);
2391 if (nid == NUMA_NO_NODE)
2392 free_huge_pages = h->free_huge_pages;
2393 else
2394 free_huge_pages = h->free_huge_pages_node[nid];
2395
2396 return sprintf(buf, "%lu\n", free_huge_pages);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002397}
2398HSTATE_ATTR_RO(free_hugepages);
2399
2400static ssize_t resv_hugepages_show(struct kobject *kobj,
2401 struct kobj_attribute *attr, char *buf)
2402{
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002403 struct hstate *h = kobj_to_hstate(kobj, NULL);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002404 return sprintf(buf, "%lu\n", h->resv_huge_pages);
2405}
2406HSTATE_ATTR_RO(resv_hugepages);
2407
2408static ssize_t surplus_hugepages_show(struct kobject *kobj,
2409 struct kobj_attribute *attr, char *buf)
2410{
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002411 struct hstate *h;
2412 unsigned long surplus_huge_pages;
2413 int nid;
2414
2415 h = kobj_to_hstate(kobj, &nid);
2416 if (nid == NUMA_NO_NODE)
2417 surplus_huge_pages = h->surplus_huge_pages;
2418 else
2419 surplus_huge_pages = h->surplus_huge_pages_node[nid];
2420
2421 return sprintf(buf, "%lu\n", surplus_huge_pages);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002422}
2423HSTATE_ATTR_RO(surplus_hugepages);
2424
2425static struct attribute *hstate_attrs[] = {
2426 &nr_hugepages_attr.attr,
2427 &nr_overcommit_hugepages_attr.attr,
2428 &free_hugepages_attr.attr,
2429 &resv_hugepages_attr.attr,
2430 &surplus_hugepages_attr.attr,
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002431#ifdef CONFIG_NUMA
2432 &nr_hugepages_mempolicy_attr.attr,
2433#endif
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002434 NULL,
2435};
2436
2437static struct attribute_group hstate_attr_group = {
2438 .attrs = hstate_attrs,
2439};
2440
Jeff Mahoney094e9532010-02-02 13:44:14 -08002441static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
2442 struct kobject **hstate_kobjs,
2443 struct attribute_group *hstate_attr_group)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002444{
2445 int retval;
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07002446 int hi = hstate_index(h);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002447
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002448 hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
2449 if (!hstate_kobjs[hi])
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002450 return -ENOMEM;
2451
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002452 retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002453 if (retval)
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002454 kobject_put(hstate_kobjs[hi]);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002455
2456 return retval;
2457}
2458
2459static void __init hugetlb_sysfs_init(void)
2460{
2461 struct hstate *h;
2462 int err;
2463
2464 hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
2465 if (!hugepages_kobj)
2466 return;
2467
2468 for_each_hstate(h) {
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002469 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
2470 hstate_kobjs, &hstate_attr_group);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002471 if (err)
Andrew Mortonffb22af2013-02-22 16:32:08 -08002472 pr_err("Hugetlb: Unable to add hstate %s", h->name);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002473 }
2474}
2475
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002476#ifdef CONFIG_NUMA
2477
2478/*
2479 * node_hstate/s - associate per node hstate attributes, via their kobjects,
Kay Sievers10fbcf42011-12-21 14:48:43 -08002480 * with node devices in node_devices[] using a parallel array. The array
2481 * index of a node device or _hstate == node id.
2482 * This is here to avoid any static dependency of the node device driver, in
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002483 * the base kernel, on the hugetlb module.
2484 */
2485struct node_hstate {
2486 struct kobject *hugepages_kobj;
2487 struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
2488};
Alexander Kuleshovb4e289a2015-11-05 18:50:14 -08002489static struct node_hstate node_hstates[MAX_NUMNODES];
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002490
2491/*
Kay Sievers10fbcf42011-12-21 14:48:43 -08002492 * A subset of global hstate attributes for node devices
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002493 */
2494static struct attribute *per_node_hstate_attrs[] = {
2495 &nr_hugepages_attr.attr,
2496 &free_hugepages_attr.attr,
2497 &surplus_hugepages_attr.attr,
2498 NULL,
2499};
2500
2501static struct attribute_group per_node_hstate_attr_group = {
2502 .attrs = per_node_hstate_attrs,
2503};
2504
2505/*
Kay Sievers10fbcf42011-12-21 14:48:43 -08002506 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002507 * Returns node id via non-NULL nidp.
2508 */
2509static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
2510{
2511 int nid;
2512
2513 for (nid = 0; nid < nr_node_ids; nid++) {
2514 struct node_hstate *nhs = &node_hstates[nid];
2515 int i;
2516 for (i = 0; i < HUGE_MAX_HSTATE; i++)
2517 if (nhs->hstate_kobjs[i] == kobj) {
2518 if (nidp)
2519 *nidp = nid;
2520 return &hstates[i];
2521 }
2522 }
2523
2524 BUG();
2525 return NULL;
2526}
2527
2528/*
Kay Sievers10fbcf42011-12-21 14:48:43 -08002529 * Unregister hstate attributes from a single node device.
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002530 * No-op if no hstate attributes attached.
2531 */
Claudiu Ghioc3cd8b442013-03-04 12:46:15 +02002532static void hugetlb_unregister_node(struct node *node)
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002533{
2534 struct hstate *h;
Kay Sievers10fbcf42011-12-21 14:48:43 -08002535 struct node_hstate *nhs = &node_hstates[node->dev.id];
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002536
2537 if (!nhs->hugepages_kobj)
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -08002538 return; /* no hstate attributes */
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002539
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07002540 for_each_hstate(h) {
2541 int idx = hstate_index(h);
2542 if (nhs->hstate_kobjs[idx]) {
2543 kobject_put(nhs->hstate_kobjs[idx]);
2544 nhs->hstate_kobjs[idx] = NULL;
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002545 }
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07002546 }
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002547
2548 kobject_put(nhs->hugepages_kobj);
2549 nhs->hugepages_kobj = NULL;
2550}
2551
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002552
2553/*
Kay Sievers10fbcf42011-12-21 14:48:43 -08002554 * Register hstate attributes for a single node device.
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002555 * No-op if attributes already registered.
2556 */
Claudiu Ghioc3cd8b442013-03-04 12:46:15 +02002557static void hugetlb_register_node(struct node *node)
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002558{
2559 struct hstate *h;
Kay Sievers10fbcf42011-12-21 14:48:43 -08002560 struct node_hstate *nhs = &node_hstates[node->dev.id];
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002561 int err;
2562
2563 if (nhs->hugepages_kobj)
2564 return; /* already allocated */
2565
2566 nhs->hugepages_kobj = kobject_create_and_add("hugepages",
Kay Sievers10fbcf42011-12-21 14:48:43 -08002567 &node->dev.kobj);
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002568 if (!nhs->hugepages_kobj)
2569 return;
2570
2571 for_each_hstate(h) {
2572 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
2573 nhs->hstate_kobjs,
2574 &per_node_hstate_attr_group);
2575 if (err) {
Andrew Mortonffb22af2013-02-22 16:32:08 -08002576 pr_err("Hugetlb: Unable to add hstate %s for node %d\n",
2577 h->name, node->dev.id);
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002578 hugetlb_unregister_node(node);
2579 break;
2580 }
2581 }
2582}
2583
2584/*
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -08002585 * hugetlb init time: register hstate attributes for all registered node
Kay Sievers10fbcf42011-12-21 14:48:43 -08002586 * devices of nodes that have memory. All on-line nodes should have
2587 * registered their associated device by this time.
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002588 */
Luiz Capitulino7d9ca002014-12-12 16:55:24 -08002589static void __init hugetlb_register_all_nodes(void)
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002590{
2591 int nid;
2592
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08002593 for_each_node_state(nid, N_MEMORY) {
Wen Congyang87327942012-12-11 16:00:56 -08002594 struct node *node = node_devices[nid];
Kay Sievers10fbcf42011-12-21 14:48:43 -08002595 if (node->dev.id == nid)
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002596 hugetlb_register_node(node);
2597 }
2598
2599 /*
Kay Sievers10fbcf42011-12-21 14:48:43 -08002600 * Let the node device driver know we're here so it can
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002601 * [un]register hstate attributes on node hotplug.
2602 */
2603 register_hugetlbfs_with_node(hugetlb_register_node,
2604 hugetlb_unregister_node);
2605}
2606#else /* !CONFIG_NUMA */
2607
2608static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
2609{
2610 BUG();
2611 if (nidp)
2612 *nidp = -1;
2613 return NULL;
2614}
2615
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002616static void hugetlb_register_all_nodes(void) { }
2617
2618#endif
2619
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002620static int __init hugetlb_init(void)
2621{
Davidlohr Bueso8382d912014-04-03 14:47:31 -07002622 int i;
2623
Nishanth Aravamudan457c1b22014-05-06 12:50:00 -07002624 if (!hugepages_supported())
Benjamin Herrenschmidt0ef89d22008-07-31 00:07:30 -07002625 return 0;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002626
Nick Piggine11bfbf2008-07-23 21:27:52 -07002627 if (!size_to_hstate(default_hstate_size)) {
2628 default_hstate_size = HPAGE_SIZE;
2629 if (!size_to_hstate(default_hstate_size))
2630 hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002631 }
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07002632 default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size));
Vaishali Thakkarf8b74812016-02-17 13:11:26 -08002633 if (default_hstate_max_huge_pages) {
2634 if (!default_hstate.max_huge_pages)
2635 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
2636 }
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002637
2638 hugetlb_init_hstates();
Andi Kleenaa888a72008-07-23 21:27:47 -07002639 gather_bootmem_prealloc();
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002640 report_hugepages();
2641
2642 hugetlb_sysfs_init();
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002643 hugetlb_register_all_nodes();
Jianguo Wu7179e7b2012-12-18 14:23:19 -08002644 hugetlb_cgroup_file_init();
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002645
Davidlohr Bueso8382d912014-04-03 14:47:31 -07002646#ifdef CONFIG_SMP
2647 num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus());
2648#else
2649 num_fault_mutexes = 1;
2650#endif
Mike Kravetzc672c7f2015-09-08 15:01:35 -07002651 hugetlb_fault_mutex_table =
Davidlohr Bueso8382d912014-04-03 14:47:31 -07002652 kmalloc(sizeof(struct mutex) * num_fault_mutexes, GFP_KERNEL);
Mike Kravetzc672c7f2015-09-08 15:01:35 -07002653 BUG_ON(!hugetlb_fault_mutex_table);
Davidlohr Bueso8382d912014-04-03 14:47:31 -07002654
2655 for (i = 0; i < num_fault_mutexes; i++)
Mike Kravetzc672c7f2015-09-08 15:01:35 -07002656 mutex_init(&hugetlb_fault_mutex_table[i]);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002657 return 0;
2658}
Paul Gortmaker3e89e1c2016-01-14 15:21:52 -08002659subsys_initcall(hugetlb_init);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002660
2661/* Should be called on processing a hugepagesz=... option */
Kirill A. Shutemovd00181b2015-11-06 16:29:57 -08002662void __init hugetlb_add_hstate(unsigned int order)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002663{
2664 struct hstate *h;
Andi Kleen8faa8b02008-07-23 21:27:48 -07002665 unsigned long i;
2666
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002667 if (size_to_hstate(PAGE_SIZE << order)) {
Joe Perches598d8092016-03-17 14:19:44 -07002668 pr_warn("hugepagesz= specified twice, ignoring\n");
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002669 return;
2670 }
Aneesh Kumar K.V47d38342012-07-31 16:41:54 -07002671 BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002672 BUG_ON(order == 0);
Aneesh Kumar K.V47d38342012-07-31 16:41:54 -07002673 h = &hstates[hugetlb_max_hstate++];
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002674 h->order = order;
2675 h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
Andi Kleen8faa8b02008-07-23 21:27:48 -07002676 h->nr_huge_pages = 0;
2677 h->free_huge_pages = 0;
2678 for (i = 0; i < MAX_NUMNODES; ++i)
2679 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
Aneesh Kumar K.V0edaecf2012-07-31 16:42:07 -07002680 INIT_LIST_HEAD(&h->hugepage_activelist);
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08002681 h->next_nid_to_alloc = first_node(node_states[N_MEMORY]);
2682 h->next_nid_to_free = first_node(node_states[N_MEMORY]);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002683 snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
2684 huge_page_size(h)/1024);
Andi Kleen8faa8b02008-07-23 21:27:48 -07002685
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002686 parsed_hstate = h;
2687}
2688
Nick Piggine11bfbf2008-07-23 21:27:52 -07002689static int __init hugetlb_nrpages_setup(char *s)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002690{
2691 unsigned long *mhp;
Andi Kleen8faa8b02008-07-23 21:27:48 -07002692 static unsigned long *last_mhp;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002693
2694 /*
Aneesh Kumar K.V47d38342012-07-31 16:41:54 -07002695 * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet,
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002696 * so this hugepages= parameter goes to the "default hstate".
2697 */
Aneesh Kumar K.V47d38342012-07-31 16:41:54 -07002698 if (!hugetlb_max_hstate)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002699 mhp = &default_hstate_max_huge_pages;
2700 else
2701 mhp = &parsed_hstate->max_huge_pages;
2702
Andi Kleen8faa8b02008-07-23 21:27:48 -07002703 if (mhp == last_mhp) {
Joe Perches598d8092016-03-17 14:19:44 -07002704 pr_warn("hugepages= specified twice without interleaving hugepagesz=, ignoring\n");
Andi Kleen8faa8b02008-07-23 21:27:48 -07002705 return 1;
2706 }
2707
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002708 if (sscanf(s, "%lu", mhp) <= 0)
2709 *mhp = 0;
2710
Andi Kleen8faa8b02008-07-23 21:27:48 -07002711 /*
2712 * Global state is always initialized later in hugetlb_init.
2713 * But we need to allocate >= MAX_ORDER hstates here early to still
2714 * use the bootmem allocator.
2715 */
Aneesh Kumar K.V47d38342012-07-31 16:41:54 -07002716 if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER)
Andi Kleen8faa8b02008-07-23 21:27:48 -07002717 hugetlb_hstate_alloc_pages(parsed_hstate);
2718
2719 last_mhp = mhp;
2720
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002721 return 1;
2722}
Nick Piggine11bfbf2008-07-23 21:27:52 -07002723__setup("hugepages=", hugetlb_nrpages_setup);
2724
2725static int __init hugetlb_default_setup(char *s)
2726{
2727 default_hstate_size = memparse(s, &s);
2728 return 1;
2729}
2730__setup("default_hugepagesz=", hugetlb_default_setup);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002731
Nishanth Aravamudan8a213462008-07-25 19:44:37 -07002732static unsigned int cpuset_mems_nr(unsigned int *array)
2733{
2734 int node;
2735 unsigned int nr = 0;
2736
2737 for_each_node_mask(node, cpuset_current_mems_allowed)
2738 nr += array[node];
2739
2740 return nr;
2741}
2742
2743#ifdef CONFIG_SYSCTL
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002744static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
2745 struct ctl_table *table, int write,
2746 void __user *buffer, size_t *length, loff_t *ppos)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002747{
Andi Kleene5ff2152008-07-23 21:27:42 -07002748 struct hstate *h = &default_hstate;
David Rientjes238d3c12014-08-06 16:06:51 -07002749 unsigned long tmp = h->max_huge_pages;
Michal Hocko08d4a242011-01-13 15:47:26 -08002750 int ret;
Andi Kleene5ff2152008-07-23 21:27:42 -07002751
Nishanth Aravamudan457c1b22014-05-06 12:50:00 -07002752 if (!hugepages_supported())
Jan Stancek86613622016-03-09 14:08:35 -08002753 return -EOPNOTSUPP;
Nishanth Aravamudan457c1b22014-05-06 12:50:00 -07002754
Andi Kleene5ff2152008-07-23 21:27:42 -07002755 table->data = &tmp;
2756 table->maxlen = sizeof(unsigned long);
Michal Hocko08d4a242011-01-13 15:47:26 -08002757 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2758 if (ret)
2759 goto out;
Andi Kleene5ff2152008-07-23 21:27:42 -07002760
David Rientjes238d3c12014-08-06 16:06:51 -07002761 if (write)
2762 ret = __nr_hugepages_store_common(obey_mempolicy, h,
2763 NUMA_NO_NODE, tmp, *length);
Michal Hocko08d4a242011-01-13 15:47:26 -08002764out:
2765 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002766}
Mel Gorman396faf02007-07-17 04:03:13 -07002767
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002768int hugetlb_sysctl_handler(struct ctl_table *table, int write,
2769 void __user *buffer, size_t *length, loff_t *ppos)
2770{
2771
2772 return hugetlb_sysctl_handler_common(false, table, write,
2773 buffer, length, ppos);
2774}
2775
2776#ifdef CONFIG_NUMA
2777int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
2778 void __user *buffer, size_t *length, loff_t *ppos)
2779{
2780 return hugetlb_sysctl_handler_common(true, table, write,
2781 buffer, length, ppos);
2782}
2783#endif /* CONFIG_NUMA */
2784
Nishanth Aravamudana3d0c6a2008-02-08 04:18:18 -08002785int hugetlb_overcommit_handler(struct ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07002786 void __user *buffer,
Nishanth Aravamudana3d0c6a2008-02-08 04:18:18 -08002787 size_t *length, loff_t *ppos)
2788{
Andi Kleena5516432008-07-23 21:27:41 -07002789 struct hstate *h = &default_hstate;
Andi Kleene5ff2152008-07-23 21:27:42 -07002790 unsigned long tmp;
Michal Hocko08d4a242011-01-13 15:47:26 -08002791 int ret;
Andi Kleene5ff2152008-07-23 21:27:42 -07002792
Nishanth Aravamudan457c1b22014-05-06 12:50:00 -07002793 if (!hugepages_supported())
Jan Stancek86613622016-03-09 14:08:35 -08002794 return -EOPNOTSUPP;
Nishanth Aravamudan457c1b22014-05-06 12:50:00 -07002795
Petr Holasekc033a932011-03-22 16:33:05 -07002796 tmp = h->nr_overcommit_huge_pages;
Andi Kleene5ff2152008-07-23 21:27:42 -07002797
Luiz Capitulinobae7f4a2014-06-04 16:07:08 -07002798 if (write && hstate_is_gigantic(h))
Eric B Munsonadbe8722011-01-13 15:47:27 -08002799 return -EINVAL;
2800
Andi Kleene5ff2152008-07-23 21:27:42 -07002801 table->data = &tmp;
2802 table->maxlen = sizeof(unsigned long);
Michal Hocko08d4a242011-01-13 15:47:26 -08002803 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2804 if (ret)
2805 goto out;
Andi Kleene5ff2152008-07-23 21:27:42 -07002806
2807 if (write) {
2808 spin_lock(&hugetlb_lock);
2809 h->nr_overcommit_huge_pages = tmp;
2810 spin_unlock(&hugetlb_lock);
2811 }
Michal Hocko08d4a242011-01-13 15:47:26 -08002812out:
2813 return ret;
Nishanth Aravamudana3d0c6a2008-02-08 04:18:18 -08002814}
2815
Linus Torvalds1da177e2005-04-16 15:20:36 -07002816#endif /* CONFIG_SYSCTL */
2817
Alexey Dobriyane1759c22008-10-15 23:50:22 +04002818void hugetlb_report_meminfo(struct seq_file *m)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002819{
Andi Kleena5516432008-07-23 21:27:41 -07002820 struct hstate *h = &default_hstate;
Nishanth Aravamudan457c1b22014-05-06 12:50:00 -07002821 if (!hugepages_supported())
2822 return;
Alexey Dobriyane1759c22008-10-15 23:50:22 +04002823 seq_printf(m,
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002824 "HugePages_Total: %5lu\n"
2825 "HugePages_Free: %5lu\n"
2826 "HugePages_Rsvd: %5lu\n"
2827 "HugePages_Surp: %5lu\n"
2828 "Hugepagesize: %8lu kB\n",
Andi Kleena5516432008-07-23 21:27:41 -07002829 h->nr_huge_pages,
2830 h->free_huge_pages,
2831 h->resv_huge_pages,
2832 h->surplus_huge_pages,
2833 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002834}
2835
2836int hugetlb_report_node_meminfo(int nid, char *buf)
2837{
Andi Kleena5516432008-07-23 21:27:41 -07002838 struct hstate *h = &default_hstate;
Nishanth Aravamudan457c1b22014-05-06 12:50:00 -07002839 if (!hugepages_supported())
2840 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002841 return sprintf(buf,
2842 "Node %d HugePages_Total: %5u\n"
Nishanth Aravamudana1de0912008-03-26 14:37:53 -07002843 "Node %d HugePages_Free: %5u\n"
2844 "Node %d HugePages_Surp: %5u\n",
Andi Kleena5516432008-07-23 21:27:41 -07002845 nid, h->nr_huge_pages_node[nid],
2846 nid, h->free_huge_pages_node[nid],
2847 nid, h->surplus_huge_pages_node[nid]);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002848}
2849
David Rientjes949f7ec2013-04-29 15:07:48 -07002850void hugetlb_show_meminfo(void)
2851{
2852 struct hstate *h;
2853 int nid;
2854
Nishanth Aravamudan457c1b22014-05-06 12:50:00 -07002855 if (!hugepages_supported())
2856 return;
2857
David Rientjes949f7ec2013-04-29 15:07:48 -07002858 for_each_node_state(nid, N_MEMORY)
2859 for_each_hstate(h)
2860 pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
2861 nid,
2862 h->nr_huge_pages_node[nid],
2863 h->free_huge_pages_node[nid],
2864 h->surplus_huge_pages_node[nid],
2865 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
2866}
2867
Naoya Horiguchi5d317b22015-11-05 18:47:14 -08002868void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm)
2869{
2870 seq_printf(m, "HugetlbPages:\t%8lu kB\n",
2871 atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10));
2872}
2873
Linus Torvalds1da177e2005-04-16 15:20:36 -07002874/* Return the number pages of memory we physically have, in PAGE_SIZE units. */
2875unsigned long hugetlb_total_pages(void)
2876{
Wanpeng Lid0028582013-03-22 15:04:40 -07002877 struct hstate *h;
2878 unsigned long nr_total_pages = 0;
2879
2880 for_each_hstate(h)
2881 nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h);
2882 return nr_total_pages;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002883}
Linus Torvalds1da177e2005-04-16 15:20:36 -07002884
Andi Kleena5516432008-07-23 21:27:41 -07002885static int hugetlb_acct_memory(struct hstate *h, long delta)
Mel Gormanfc1b8a72008-07-23 21:27:22 -07002886{
2887 int ret = -ENOMEM;
2888
2889 spin_lock(&hugetlb_lock);
2890 /*
2891 * When cpuset is configured, it breaks the strict hugetlb page
2892 * reservation as the accounting is done on a global variable. Such
2893 * reservation is completely rubbish in the presence of cpuset because
2894 * the reservation is not checked against page availability for the
2895 * current cpuset. Application can still potentially OOM'ed by kernel
2896 * with lack of free htlb page in cpuset that the task is in.
2897 * Attempt to enforce strict accounting with cpuset is almost
2898 * impossible (or too ugly) because cpuset is too fluid that
2899 * task or memory node can be dynamically moved between cpusets.
2900 *
2901 * The change of semantics for shared hugetlb mapping with cpuset is
2902 * undesirable. However, in order to preserve some of the semantics,
2903 * we fall back to check against current free page availability as
2904 * a best attempt and hopefully to minimize the impact of changing
2905 * semantics that cpuset has.
2906 */
2907 if (delta > 0) {
Andi Kleena5516432008-07-23 21:27:41 -07002908 if (gather_surplus_pages(h, delta) < 0)
Mel Gormanfc1b8a72008-07-23 21:27:22 -07002909 goto out;
2910
Andi Kleena5516432008-07-23 21:27:41 -07002911 if (delta > cpuset_mems_nr(h->free_huge_pages_node)) {
2912 return_unused_surplus_pages(h, delta);
Mel Gormanfc1b8a72008-07-23 21:27:22 -07002913 goto out;
2914 }
2915 }
2916
2917 ret = 0;
2918 if (delta < 0)
Andi Kleena5516432008-07-23 21:27:41 -07002919 return_unused_surplus_pages(h, (unsigned long) -delta);
Mel Gormanfc1b8a72008-07-23 21:27:22 -07002920
2921out:
2922 spin_unlock(&hugetlb_lock);
2923 return ret;
2924}
2925
Andy Whitcroft84afd992008-07-23 21:27:32 -07002926static void hugetlb_vm_op_open(struct vm_area_struct *vma)
2927{
Joonsoo Kimf522c3a2013-09-11 14:21:53 -07002928 struct resv_map *resv = vma_resv_map(vma);
Andy Whitcroft84afd992008-07-23 21:27:32 -07002929
2930 /*
2931 * This new VMA should share its siblings reservation map if present.
2932 * The VMA will only ever have a valid reservation map pointer where
2933 * it is being copied for another still existing VMA. As that VMA
Lucas De Marchi25985ed2011-03-30 22:57:33 -03002934 * has a reference to the reservation map it cannot disappear until
Andy Whitcroft84afd992008-07-23 21:27:32 -07002935 * after this open call completes. It is therefore safe to take a
2936 * new reference here without additional locking.
2937 */
Joonsoo Kim4e35f482014-04-03 14:47:30 -07002938 if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
Joonsoo Kimf522c3a2013-09-11 14:21:53 -07002939 kref_get(&resv->refs);
Andy Whitcroft84afd992008-07-23 21:27:32 -07002940}
2941
Mel Gormana1e78772008-07-23 21:27:23 -07002942static void hugetlb_vm_op_close(struct vm_area_struct *vma)
2943{
Andi Kleena5516432008-07-23 21:27:41 -07002944 struct hstate *h = hstate_vma(vma);
Joonsoo Kimf522c3a2013-09-11 14:21:53 -07002945 struct resv_map *resv = vma_resv_map(vma);
David Gibson90481622012-03-21 16:34:12 -07002946 struct hugepage_subpool *spool = subpool_vma(vma);
Joonsoo Kim4e35f482014-04-03 14:47:30 -07002947 unsigned long reserve, start, end;
Mike Kravetz1c5ecae2015-04-15 16:13:39 -07002948 long gbl_reserve;
Andy Whitcroft84afd992008-07-23 21:27:32 -07002949
Joonsoo Kim4e35f482014-04-03 14:47:30 -07002950 if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER))
2951 return;
Andy Whitcroft84afd992008-07-23 21:27:32 -07002952
Joonsoo Kim4e35f482014-04-03 14:47:30 -07002953 start = vma_hugecache_offset(h, vma, vma->vm_start);
2954 end = vma_hugecache_offset(h, vma, vma->vm_end);
Andy Whitcroft84afd992008-07-23 21:27:32 -07002955
Joonsoo Kim4e35f482014-04-03 14:47:30 -07002956 reserve = (end - start) - region_count(resv, start, end);
Andy Whitcroft84afd992008-07-23 21:27:32 -07002957
Joonsoo Kim4e35f482014-04-03 14:47:30 -07002958 kref_put(&resv->refs, resv_map_release);
2959
2960 if (reserve) {
Mike Kravetz1c5ecae2015-04-15 16:13:39 -07002961 /*
2962 * Decrement reserve counts. The global reserve count may be
2963 * adjusted if the subpool has a minimum size.
2964 */
2965 gbl_reserve = hugepage_subpool_put_pages(spool, reserve);
2966 hugetlb_acct_memory(h, -gbl_reserve);
Andy Whitcroft84afd992008-07-23 21:27:32 -07002967 }
Mel Gormana1e78772008-07-23 21:27:23 -07002968}
2969
Linus Torvalds1da177e2005-04-16 15:20:36 -07002970/*
2971 * We cannot handle pagefaults against hugetlb pages at all. They cause
2972 * handle_mm_fault() to try to instantiate regular-sized pages in the
2973 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
2974 * this far.
2975 */
Nick Piggind0217ac2007-07-19 01:47:03 -07002976static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002977{
2978 BUG();
Nick Piggind0217ac2007-07-19 01:47:03 -07002979 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002980}
2981
Alexey Dobriyanf0f37e22009-09-27 22:29:37 +04002982const struct vm_operations_struct hugetlb_vm_ops = {
Nick Piggind0217ac2007-07-19 01:47:03 -07002983 .fault = hugetlb_vm_op_fault,
Andy Whitcroft84afd992008-07-23 21:27:32 -07002984 .open = hugetlb_vm_op_open,
Mel Gormana1e78772008-07-23 21:27:23 -07002985 .close = hugetlb_vm_op_close,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002986};
2987
David Gibson1e8f8892006-01-06 00:10:44 -08002988static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
2989 int writable)
David Gibson63551ae2005-06-21 17:14:44 -07002990{
2991 pte_t entry;
2992
David Gibson1e8f8892006-01-06 00:10:44 -08002993 if (writable) {
Gerald Schaefer106c9922013-04-29 15:07:23 -07002994 entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page,
2995 vma->vm_page_prot)));
David Gibson63551ae2005-06-21 17:14:44 -07002996 } else {
Gerald Schaefer106c9922013-04-29 15:07:23 -07002997 entry = huge_pte_wrprotect(mk_huge_pte(page,
2998 vma->vm_page_prot));
David Gibson63551ae2005-06-21 17:14:44 -07002999 }
3000 entry = pte_mkyoung(entry);
3001 entry = pte_mkhuge(entry);
Chris Metcalfd9ed9fa2012-04-01 14:01:34 -04003002 entry = arch_make_huge_pte(entry, vma, page, writable);
David Gibson63551ae2005-06-21 17:14:44 -07003003
3004 return entry;
3005}
3006
David Gibson1e8f8892006-01-06 00:10:44 -08003007static void set_huge_ptep_writable(struct vm_area_struct *vma,
3008 unsigned long address, pte_t *ptep)
3009{
3010 pte_t entry;
3011
Gerald Schaefer106c9922013-04-29 15:07:23 -07003012 entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep)));
Chris Forbes32f84522011-07-25 17:12:14 -07003013 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
Russell King4b3073e2009-12-18 16:40:18 +00003014 update_mmu_cache(vma, address, ptep);
David Gibson1e8f8892006-01-06 00:10:44 -08003015}
3016
Naoya Horiguchi4a705fe2014-06-23 13:22:03 -07003017static int is_hugetlb_entry_migration(pte_t pte)
3018{
3019 swp_entry_t swp;
3020
3021 if (huge_pte_none(pte) || pte_present(pte))
3022 return 0;
3023 swp = pte_to_swp_entry(pte);
3024 if (non_swap_entry(swp) && is_migration_entry(swp))
3025 return 1;
3026 else
3027 return 0;
3028}
3029
3030static int is_hugetlb_entry_hwpoisoned(pte_t pte)
3031{
3032 swp_entry_t swp;
3033
3034 if (huge_pte_none(pte) || pte_present(pte))
3035 return 0;
3036 swp = pte_to_swp_entry(pte);
3037 if (non_swap_entry(swp) && is_hwpoison_entry(swp))
3038 return 1;
3039 else
3040 return 0;
3041}
David Gibson1e8f8892006-01-06 00:10:44 -08003042
David Gibson63551ae2005-06-21 17:14:44 -07003043int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
3044 struct vm_area_struct *vma)
3045{
3046 pte_t *src_pte, *dst_pte, entry;
3047 struct page *ptepage;
Hugh Dickins1c598272005-10-19 21:23:43 -07003048 unsigned long addr;
David Gibson1e8f8892006-01-06 00:10:44 -08003049 int cow;
Andi Kleena5516432008-07-23 21:27:41 -07003050 struct hstate *h = hstate_vma(vma);
3051 unsigned long sz = huge_page_size(h);
Andreas Sandberge8569dd2014-01-21 15:49:09 -08003052 unsigned long mmun_start; /* For mmu_notifiers */
3053 unsigned long mmun_end; /* For mmu_notifiers */
3054 int ret = 0;
David Gibson1e8f8892006-01-06 00:10:44 -08003055
3056 cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
David Gibson63551ae2005-06-21 17:14:44 -07003057
Andreas Sandberge8569dd2014-01-21 15:49:09 -08003058 mmun_start = vma->vm_start;
3059 mmun_end = vma->vm_end;
3060 if (cow)
3061 mmu_notifier_invalidate_range_start(src, mmun_start, mmun_end);
3062
Andi Kleena5516432008-07-23 21:27:41 -07003063 for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003064 spinlock_t *src_ptl, *dst_ptl;
Hugh Dickinsc74df322005-10-29 18:16:23 -07003065 src_pte = huge_pte_offset(src, addr);
3066 if (!src_pte)
3067 continue;
Andi Kleena5516432008-07-23 21:27:41 -07003068 dst_pte = huge_pte_alloc(dst, addr, sz);
Andreas Sandberge8569dd2014-01-21 15:49:09 -08003069 if (!dst_pte) {
3070 ret = -ENOMEM;
3071 break;
3072 }
Larry Woodmanc5c99422008-01-24 05:49:25 -08003073
3074 /* If the pagetables are shared don't copy or take references */
3075 if (dst_pte == src_pte)
3076 continue;
3077
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003078 dst_ptl = huge_pte_lock(h, dst, dst_pte);
3079 src_ptl = huge_pte_lockptr(h, src, src_pte);
3080 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
Naoya Horiguchi4a705fe2014-06-23 13:22:03 -07003081 entry = huge_ptep_get(src_pte);
3082 if (huge_pte_none(entry)) { /* skip none entry */
3083 ;
3084 } else if (unlikely(is_hugetlb_entry_migration(entry) ||
3085 is_hugetlb_entry_hwpoisoned(entry))) {
3086 swp_entry_t swp_entry = pte_to_swp_entry(entry);
3087
3088 if (is_write_migration_entry(swp_entry) && cow) {
3089 /*
3090 * COW mappings require pages in both
3091 * parent and child to be set to read.
3092 */
3093 make_migration_entry_read(&swp_entry);
3094 entry = swp_entry_to_pte(swp_entry);
3095 set_huge_pte_at(src, addr, src_pte, entry);
3096 }
3097 set_huge_pte_at(dst, addr, dst_pte, entry);
3098 } else {
Joerg Roedel34ee6452014-11-13 13:46:09 +11003099 if (cow) {
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07003100 huge_ptep_set_wrprotect(src, addr, src_pte);
Joerg Roedel34ee6452014-11-13 13:46:09 +11003101 mmu_notifier_invalidate_range(src, mmun_start,
3102 mmun_end);
3103 }
Naoya Horiguchi0253d632014-07-23 14:00:19 -07003104 entry = huge_ptep_get(src_pte);
Hugh Dickins1c598272005-10-19 21:23:43 -07003105 ptepage = pte_page(entry);
3106 get_page(ptepage);
Kirill A. Shutemov53f92632016-01-15 16:53:42 -08003107 page_dup_rmap(ptepage, true);
Hugh Dickins1c598272005-10-19 21:23:43 -07003108 set_huge_pte_at(dst, addr, dst_pte, entry);
Naoya Horiguchi5d317b22015-11-05 18:47:14 -08003109 hugetlb_count_add(pages_per_huge_page(h), dst);
Hugh Dickins1c598272005-10-19 21:23:43 -07003110 }
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003111 spin_unlock(src_ptl);
3112 spin_unlock(dst_ptl);
David Gibson63551ae2005-06-21 17:14:44 -07003113 }
David Gibson63551ae2005-06-21 17:14:44 -07003114
Andreas Sandberge8569dd2014-01-21 15:49:09 -08003115 if (cow)
3116 mmu_notifier_invalidate_range_end(src, mmun_start, mmun_end);
3117
3118 return ret;
David Gibson63551ae2005-06-21 17:14:44 -07003119}
3120
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07003121void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
3122 unsigned long start, unsigned long end,
3123 struct page *ref_page)
David Gibson63551ae2005-06-21 17:14:44 -07003124{
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07003125 int force_flush = 0;
David Gibson63551ae2005-06-21 17:14:44 -07003126 struct mm_struct *mm = vma->vm_mm;
3127 unsigned long address;
David Gibsonc7546f82005-08-05 11:59:35 -07003128 pte_t *ptep;
David Gibson63551ae2005-06-21 17:14:44 -07003129 pte_t pte;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003130 spinlock_t *ptl;
David Gibson63551ae2005-06-21 17:14:44 -07003131 struct page *page;
Andi Kleena5516432008-07-23 21:27:41 -07003132 struct hstate *h = hstate_vma(vma);
3133 unsigned long sz = huge_page_size(h);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07003134 const unsigned long mmun_start = start; /* For mmu_notifiers */
3135 const unsigned long mmun_end = end; /* For mmu_notifiers */
Andi Kleena5516432008-07-23 21:27:41 -07003136
David Gibson63551ae2005-06-21 17:14:44 -07003137 WARN_ON(!is_vm_hugetlb_page(vma));
Andi Kleena5516432008-07-23 21:27:41 -07003138 BUG_ON(start & ~huge_page_mask(h));
3139 BUG_ON(end & ~huge_page_mask(h));
David Gibson63551ae2005-06-21 17:14:44 -07003140
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07003141 tlb_start_vma(tlb, vma);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07003142 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
Hillf Danton569f48b2014-12-10 15:44:41 -08003143 address = start;
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07003144again:
Hillf Danton569f48b2014-12-10 15:44:41 -08003145 for (; address < end; address += sz) {
David Gibsonc7546f82005-08-05 11:59:35 -07003146 ptep = huge_pte_offset(mm, address);
Adam Litke4c887262005-10-29 18:16:46 -07003147 if (!ptep)
David Gibsonc7546f82005-08-05 11:59:35 -07003148 continue;
3149
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003150 ptl = huge_pte_lock(h, mm, ptep);
Chen, Kenneth W39dde652006-12-06 20:32:03 -08003151 if (huge_pmd_unshare(mm, &address, ptep))
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003152 goto unlock;
Chen, Kenneth W39dde652006-12-06 20:32:03 -08003153
Hillf Danton66293262012-03-23 15:01:48 -07003154 pte = huge_ptep_get(ptep);
3155 if (huge_pte_none(pte))
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003156 goto unlock;
Hillf Danton66293262012-03-23 15:01:48 -07003157
3158 /*
Naoya Horiguchi9fbc1f62015-02-11 15:25:32 -08003159 * Migrating hugepage or HWPoisoned hugepage is already
3160 * unmapped and its refcount is dropped, so just clear pte here.
Hillf Danton66293262012-03-23 15:01:48 -07003161 */
Naoya Horiguchi9fbc1f62015-02-11 15:25:32 -08003162 if (unlikely(!pte_present(pte))) {
Gerald Schaefer106c9922013-04-29 15:07:23 -07003163 huge_pte_clear(mm, address, ptep);
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003164 goto unlock;
Naoya Horiguchi8c4894c2012-12-12 13:52:28 -08003165 }
Hillf Danton66293262012-03-23 15:01:48 -07003166
3167 page = pte_page(pte);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003168 /*
3169 * If a reference page is supplied, it is because a specific
3170 * page is being unmapped, not a range. Ensure the page we
3171 * are about to unmap is the actual page of interest.
3172 */
3173 if (ref_page) {
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003174 if (page != ref_page)
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003175 goto unlock;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003176
3177 /*
3178 * Mark the VMA as having unmapped its page so that
3179 * future faults in this VMA will fail rather than
3180 * looking like data was lost
3181 */
3182 set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
3183 }
3184
David Gibsonc7546f82005-08-05 11:59:35 -07003185 pte = huge_ptep_get_and_clear(mm, address, ptep);
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07003186 tlb_remove_tlb_entry(tlb, ptep, address);
Gerald Schaefer106c9922013-04-29 15:07:23 -07003187 if (huge_pte_dirty(pte))
Ken Chen6649a382007-02-08 14:20:27 -08003188 set_page_dirty(page);
Hillf Danton9e811302012-03-21 16:34:03 -07003189
Naoya Horiguchi5d317b22015-11-05 18:47:14 -08003190 hugetlb_count_sub(pages_per_huge_page(h), mm);
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -08003191 page_remove_rmap(page, true);
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07003192 force_flush = !__tlb_remove_page(tlb, page);
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003193 if (force_flush) {
Hillf Danton569f48b2014-12-10 15:44:41 -08003194 address += sz;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003195 spin_unlock(ptl);
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07003196 break;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003197 }
Hillf Danton9e811302012-03-21 16:34:03 -07003198 /* Bail out after unmapping reference page if supplied */
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003199 if (ref_page) {
3200 spin_unlock(ptl);
Hillf Danton9e811302012-03-21 16:34:03 -07003201 break;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003202 }
3203unlock:
3204 spin_unlock(ptl);
David Gibson63551ae2005-06-21 17:14:44 -07003205 }
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07003206 /*
3207 * mmu_gather ran out of room to batch pages, we break out of
3208 * the PTE lock to avoid doing the potential expensive TLB invalidate
3209 * and page-free while holding it.
3210 */
3211 if (force_flush) {
3212 force_flush = 0;
3213 tlb_flush_mmu(tlb);
3214 if (address < end && !ref_page)
3215 goto again;
Chen, Kenneth Wfe1668a2006-10-04 02:15:24 -07003216 }
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07003217 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07003218 tlb_end_vma(tlb, vma);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003219}
David Gibson63551ae2005-06-21 17:14:44 -07003220
Mel Gormand8333522012-07-31 16:46:20 -07003221void __unmap_hugepage_range_final(struct mmu_gather *tlb,
3222 struct vm_area_struct *vma, unsigned long start,
3223 unsigned long end, struct page *ref_page)
3224{
3225 __unmap_hugepage_range(tlb, vma, start, end, ref_page);
3226
3227 /*
3228 * Clear this flag so that x86's huge_pmd_share page_table_shareable
3229 * test will fail on a vma being torn down, and not grab a page table
3230 * on its way out. We're lucky that the flag has such an appropriate
3231 * name, and can in fact be safely cleared here. We could clear it
3232 * before the __unmap_hugepage_range above, but all that's necessary
Davidlohr Buesoc8c06ef2014-12-12 16:54:24 -08003233 * is to clear it before releasing the i_mmap_rwsem. This works
Mel Gormand8333522012-07-31 16:46:20 -07003234 * because in the context this is called, the VMA is about to be
Davidlohr Buesoc8c06ef2014-12-12 16:54:24 -08003235 * destroyed and the i_mmap_rwsem is held.
Mel Gormand8333522012-07-31 16:46:20 -07003236 */
3237 vma->vm_flags &= ~VM_MAYSHARE;
3238}
3239
Chen, Kenneth W502717f2006-10-11 01:20:46 -07003240void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003241 unsigned long end, struct page *ref_page)
Chen, Kenneth W502717f2006-10-11 01:20:46 -07003242{
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07003243 struct mm_struct *mm;
3244 struct mmu_gather tlb;
3245
3246 mm = vma->vm_mm;
3247
Linus Torvalds2b047252013-08-15 11:42:25 -07003248 tlb_gather_mmu(&tlb, mm, start, end);
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07003249 __unmap_hugepage_range(&tlb, vma, start, end, ref_page);
3250 tlb_finish_mmu(&tlb, start, end);
Chen, Kenneth W502717f2006-10-11 01:20:46 -07003251}
3252
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003253/*
3254 * This is called when the original mapper is failing to COW a MAP_PRIVATE
3255 * mappping it owns the reserve page for. The intention is to unmap the page
3256 * from other VMAs and let the children be SIGKILLed if they are faulting the
3257 * same region.
3258 */
Davidlohr Bueso2f4612a2014-08-06 16:06:45 -07003259static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
3260 struct page *page, unsigned long address)
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003261{
Adam Litke75266742008-11-12 13:24:56 -08003262 struct hstate *h = hstate_vma(vma);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003263 struct vm_area_struct *iter_vma;
3264 struct address_space *mapping;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003265 pgoff_t pgoff;
3266
3267 /*
3268 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
3269 * from page cache lookup which is in HPAGE_SIZE units.
3270 */
Adam Litke75266742008-11-12 13:24:56 -08003271 address = address & huge_page_mask(h);
Michal Hocko36e4f202012-10-08 16:33:31 -07003272 pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) +
3273 vma->vm_pgoff;
Al Viro496ad9a2013-01-23 17:07:38 -05003274 mapping = file_inode(vma->vm_file)->i_mapping;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003275
Mel Gorman4eb2b1d2009-12-14 17:59:53 -08003276 /*
3277 * Take the mapping lock for the duration of the table walk. As
3278 * this mapping should be shared between all the VMAs,
3279 * __unmap_hugepage_range() is called as the lock is already held
3280 */
Davidlohr Bueso83cde9e2014-12-12 16:54:21 -08003281 i_mmap_lock_write(mapping);
Michel Lespinasse6b2dbba2012-10-08 16:31:25 -07003282 vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) {
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003283 /* Do not unmap the current VMA */
3284 if (iter_vma == vma)
3285 continue;
3286
3287 /*
Mel Gorman2f84a892015-10-01 15:36:57 -07003288 * Shared VMAs have their own reserves and do not affect
3289 * MAP_PRIVATE accounting but it is possible that a shared
3290 * VMA is using the same page so check and skip such VMAs.
3291 */
3292 if (iter_vma->vm_flags & VM_MAYSHARE)
3293 continue;
3294
3295 /*
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003296 * Unmap the page from other VMAs without their own reserves.
3297 * They get marked to be SIGKILLed if they fault in these
3298 * areas. This is because a future no-page fault on this VMA
3299 * could insert a zeroed page instead of the data existing
3300 * from the time of fork. This would look like data corruption
3301 */
3302 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07003303 unmap_hugepage_range(iter_vma, address,
3304 address + huge_page_size(h), page);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003305 }
Davidlohr Bueso83cde9e2014-12-12 16:54:21 -08003306 i_mmap_unlock_write(mapping);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003307}
3308
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09003309/*
3310 * Hugetlb_cow() should be called with page lock of the original hugepage held.
Michal Hockoef009b22012-01-10 15:07:21 -08003311 * Called with hugetlb_instantiation_mutex held and pte_page locked so we
3312 * cannot race with other handlers or page migration.
3313 * Keep the pte_same checks anyway to make transition from the mutex easier.
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09003314 */
David Gibson1e8f8892006-01-06 00:10:44 -08003315static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003316 unsigned long address, pte_t *ptep, pte_t pte,
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003317 struct page *pagecache_page, spinlock_t *ptl)
David Gibson1e8f8892006-01-06 00:10:44 -08003318{
Andi Kleena5516432008-07-23 21:27:41 -07003319 struct hstate *h = hstate_vma(vma);
David Gibson1e8f8892006-01-06 00:10:44 -08003320 struct page *old_page, *new_page;
Davidlohr Buesoad4404a2014-08-06 16:06:47 -07003321 int ret = 0, outside_reserve = 0;
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07003322 unsigned long mmun_start; /* For mmu_notifiers */
3323 unsigned long mmun_end; /* For mmu_notifiers */
David Gibson1e8f8892006-01-06 00:10:44 -08003324
3325 old_page = pte_page(pte);
3326
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003327retry_avoidcopy:
David Gibson1e8f8892006-01-06 00:10:44 -08003328 /* If no-one else is actually using this page, avoid the copy
3329 * and just make the page writable */
Joonsoo Kim37a21402013-09-11 14:21:04 -07003330 if (page_mapcount(old_page) == 1 && PageAnon(old_page)) {
3331 page_move_anon_rmap(old_page, vma, address);
David Gibson1e8f8892006-01-06 00:10:44 -08003332 set_huge_ptep_writable(vma, address, ptep);
Nick Piggin83c54072007-07-19 01:47:05 -07003333 return 0;
David Gibson1e8f8892006-01-06 00:10:44 -08003334 }
3335
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003336 /*
3337 * If the process that created a MAP_PRIVATE mapping is about to
3338 * perform a COW due to a shared page count, attempt to satisfy
3339 * the allocation without using the existing reserves. The pagecache
3340 * page is used to determine if the reserve at this address was
3341 * consumed or not. If reserves were used, a partial faulted mapping
3342 * at the time of fork() could consume its reserves on COW instead
3343 * of the full address range.
3344 */
Joonsoo Kim5944d012013-09-11 14:21:55 -07003345 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003346 old_page != pagecache_page)
3347 outside_reserve = 1;
3348
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03003349 get_page(old_page);
Larry Woodmanb76c8cf2009-12-14 17:59:37 -08003350
Davidlohr Buesoad4404a2014-08-06 16:06:47 -07003351 /*
3352 * Drop page table lock as buddy allocator may be called. It will
3353 * be acquired again before returning to the caller, as expected.
3354 */
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003355 spin_unlock(ptl);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003356 new_page = alloc_huge_page(vma, address, outside_reserve);
David Gibson1e8f8892006-01-06 00:10:44 -08003357
Adam Litke2fc39ce2007-11-14 16:59:39 -08003358 if (IS_ERR(new_page)) {
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003359 /*
3360 * If a process owning a MAP_PRIVATE mapping fails to COW,
3361 * it is due to references held by a child and an insufficient
3362 * huge page pool. To guarantee the original mappers
3363 * reliability, unmap the page from child processes. The child
3364 * may get SIGKILLed if it later faults.
3365 */
3366 if (outside_reserve) {
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03003367 put_page(old_page);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003368 BUG_ON(huge_pte_none(pte));
Davidlohr Bueso2f4612a2014-08-06 16:06:45 -07003369 unmap_ref_private(mm, vma, old_page, address);
3370 BUG_ON(huge_pte_none(pte));
3371 spin_lock(ptl);
3372 ptep = huge_pte_offset(mm, address & huge_page_mask(h));
3373 if (likely(ptep &&
3374 pte_same(huge_ptep_get(ptep), pte)))
3375 goto retry_avoidcopy;
3376 /*
3377 * race occurs while re-acquiring page table
3378 * lock, and our job is done.
3379 */
3380 return 0;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003381 }
3382
Davidlohr Buesoad4404a2014-08-06 16:06:47 -07003383 ret = (PTR_ERR(new_page) == -ENOMEM) ?
3384 VM_FAULT_OOM : VM_FAULT_SIGBUS;
3385 goto out_release_old;
David Gibson1e8f8892006-01-06 00:10:44 -08003386 }
3387
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09003388 /*
3389 * When the original hugepage is shared one, it does not have
3390 * anon_vma prepared.
3391 */
Dean Nelson44e2aa92010-10-26 14:22:08 -07003392 if (unlikely(anon_vma_prepare(vma))) {
Davidlohr Buesoad4404a2014-08-06 16:06:47 -07003393 ret = VM_FAULT_OOM;
3394 goto out_release_all;
Dean Nelson44e2aa92010-10-26 14:22:08 -07003395 }
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09003396
Andrea Arcangeli47ad8472011-01-13 15:46:47 -08003397 copy_user_huge_page(new_page, old_page, address, vma,
3398 pages_per_huge_page(h));
Nick Piggin0ed361d2008-02-04 22:29:34 -08003399 __SetPageUptodate(new_page);
Naoya Horiguchibcc54222015-04-15 16:14:38 -07003400 set_page_huge_active(new_page);
David Gibson1e8f8892006-01-06 00:10:44 -08003401
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07003402 mmun_start = address & huge_page_mask(h);
3403 mmun_end = mmun_start + huge_page_size(h);
3404 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
Davidlohr Buesoad4404a2014-08-06 16:06:47 -07003405
Larry Woodmanb76c8cf2009-12-14 17:59:37 -08003406 /*
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003407 * Retake the page table lock to check for racing updates
Larry Woodmanb76c8cf2009-12-14 17:59:37 -08003408 * before the page tables are altered
3409 */
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003410 spin_lock(ptl);
Andi Kleena5516432008-07-23 21:27:41 -07003411 ptep = huge_pte_offset(mm, address & huge_page_mask(h));
Naoya Horiguchia9af0c52014-04-07 15:36:54 -07003412 if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) {
Joonsoo Kim07443a82013-09-11 14:21:58 -07003413 ClearPagePrivate(new_page);
3414
David Gibson1e8f8892006-01-06 00:10:44 -08003415 /* Break COW */
Gerald Schaefer8fe627e2008-04-28 02:13:28 -07003416 huge_ptep_clear_flush(vma, address, ptep);
Joerg Roedel34ee6452014-11-13 13:46:09 +11003417 mmu_notifier_invalidate_range(mm, mmun_start, mmun_end);
David Gibson1e8f8892006-01-06 00:10:44 -08003418 set_huge_pte_at(mm, address, ptep,
3419 make_huge_pte(vma, new_page, 1));
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -08003420 page_remove_rmap(old_page, true);
Naoya Horiguchicd67f0d2010-09-10 13:23:04 +09003421 hugepage_add_new_anon_rmap(new_page, vma, address);
David Gibson1e8f8892006-01-06 00:10:44 -08003422 /* Make the old page be freed below */
3423 new_page = old_page;
3424 }
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003425 spin_unlock(ptl);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07003426 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
Davidlohr Buesoad4404a2014-08-06 16:06:47 -07003427out_release_all:
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03003428 put_page(new_page);
Davidlohr Buesoad4404a2014-08-06 16:06:47 -07003429out_release_old:
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03003430 put_page(old_page);
Joonsoo Kim83120342013-09-11 14:21:57 -07003431
Davidlohr Buesoad4404a2014-08-06 16:06:47 -07003432 spin_lock(ptl); /* Caller expects lock to be held */
3433 return ret;
David Gibson1e8f8892006-01-06 00:10:44 -08003434}
3435
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003436/* Return the pagecache page at a given address within a VMA */
Andi Kleena5516432008-07-23 21:27:41 -07003437static struct page *hugetlbfs_pagecache_page(struct hstate *h,
3438 struct vm_area_struct *vma, unsigned long address)
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003439{
3440 struct address_space *mapping;
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -07003441 pgoff_t idx;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003442
3443 mapping = vma->vm_file->f_mapping;
Andi Kleena5516432008-07-23 21:27:41 -07003444 idx = vma_hugecache_offset(h, vma, address);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003445
3446 return find_lock_page(mapping, idx);
3447}
3448
Hugh Dickins3ae77f42009-09-21 17:03:33 -07003449/*
3450 * Return whether there is a pagecache page to back given address within VMA.
3451 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
3452 */
3453static bool hugetlbfs_pagecache_present(struct hstate *h,
Hugh Dickins2a15efc2009-09-21 17:03:27 -07003454 struct vm_area_struct *vma, unsigned long address)
3455{
3456 struct address_space *mapping;
3457 pgoff_t idx;
3458 struct page *page;
3459
3460 mapping = vma->vm_file->f_mapping;
3461 idx = vma_hugecache_offset(h, vma, address);
3462
3463 page = find_get_page(mapping, idx);
3464 if (page)
3465 put_page(page);
3466 return page != NULL;
3467}
3468
Mike Kravetzab76ad52015-09-08 15:01:50 -07003469int huge_add_to_page_cache(struct page *page, struct address_space *mapping,
3470 pgoff_t idx)
3471{
3472 struct inode *inode = mapping->host;
3473 struct hstate *h = hstate_inode(inode);
3474 int err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
3475
3476 if (err)
3477 return err;
3478 ClearPagePrivate(page);
3479
3480 spin_lock(&inode->i_lock);
3481 inode->i_blocks += blocks_per_huge_page(h);
3482 spin_unlock(&inode->i_lock);
3483 return 0;
3484}
3485
Robert P. J. Daya1ed3dd2007-07-17 04:03:33 -07003486static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
Davidlohr Bueso8382d912014-04-03 14:47:31 -07003487 struct address_space *mapping, pgoff_t idx,
3488 unsigned long address, pte_t *ptep, unsigned int flags)
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01003489{
Andi Kleena5516432008-07-23 21:27:41 -07003490 struct hstate *h = hstate_vma(vma);
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01003491 int ret = VM_FAULT_SIGBUS;
Hillf Danton409eb8c2012-01-20 14:34:13 -08003492 int anon_rmap = 0;
Adam Litke4c887262005-10-29 18:16:46 -07003493 unsigned long size;
Adam Litke4c887262005-10-29 18:16:46 -07003494 struct page *page;
David Gibson1e8f8892006-01-06 00:10:44 -08003495 pte_t new_pte;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003496 spinlock_t *ptl;
Adam Litke4c887262005-10-29 18:16:46 -07003497
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003498 /*
3499 * Currently, we are forced to kill the process in the event the
3500 * original mapper has unmapped pages from the child due to a failed
Lucas De Marchi25985ed2011-03-30 22:57:33 -03003501 * COW. Warn that such a situation has occurred as it may not be obvious
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003502 */
3503 if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
Geoffrey Thomas910154d2016-03-09 14:08:04 -08003504 pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n",
Andrew Mortonffb22af2013-02-22 16:32:08 -08003505 current->pid);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003506 return ret;
3507 }
3508
Adam Litke4c887262005-10-29 18:16:46 -07003509 /*
3510 * Use page lock to guard against racing truncation
3511 * before we get page_table_lock.
3512 */
Christoph Lameter6bda6662006-01-06 00:10:49 -08003513retry:
3514 page = find_lock_page(mapping, idx);
3515 if (!page) {
Andi Kleena5516432008-07-23 21:27:41 -07003516 size = i_size_read(mapping->host) >> huge_page_shift(h);
Hugh Dickinsebed4bf2006-10-28 10:38:43 -07003517 if (idx >= size)
3518 goto out;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003519 page = alloc_huge_page(vma, address, 0);
Adam Litke2fc39ce2007-11-14 16:59:39 -08003520 if (IS_ERR(page)) {
Aneesh Kumar K.V76dcee72012-07-31 16:41:57 -07003521 ret = PTR_ERR(page);
3522 if (ret == -ENOMEM)
3523 ret = VM_FAULT_OOM;
3524 else
3525 ret = VM_FAULT_SIGBUS;
Christoph Lameter6bda6662006-01-06 00:10:49 -08003526 goto out;
3527 }
Andrea Arcangeli47ad8472011-01-13 15:46:47 -08003528 clear_huge_page(page, address, pages_per_huge_page(h));
Nick Piggin0ed361d2008-02-04 22:29:34 -08003529 __SetPageUptodate(page);
Naoya Horiguchibcc54222015-04-15 16:14:38 -07003530 set_page_huge_active(page);
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01003531
Mel Gormanf83a2752009-05-28 14:34:40 -07003532 if (vma->vm_flags & VM_MAYSHARE) {
Mike Kravetzab76ad52015-09-08 15:01:50 -07003533 int err = huge_add_to_page_cache(page, mapping, idx);
Christoph Lameter6bda6662006-01-06 00:10:49 -08003534 if (err) {
3535 put_page(page);
Christoph Lameter6bda6662006-01-06 00:10:49 -08003536 if (err == -EEXIST)
3537 goto retry;
3538 goto out;
3539 }
Mel Gorman23be7462010-04-23 13:17:56 -04003540 } else {
Christoph Lameter6bda6662006-01-06 00:10:49 -08003541 lock_page(page);
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09003542 if (unlikely(anon_vma_prepare(vma))) {
3543 ret = VM_FAULT_OOM;
3544 goto backout_unlocked;
3545 }
Hillf Danton409eb8c2012-01-20 14:34:13 -08003546 anon_rmap = 1;
Mel Gorman23be7462010-04-23 13:17:56 -04003547 }
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09003548 } else {
Naoya Horiguchi998b4382010-09-08 10:19:32 +09003549 /*
3550 * If memory error occurs between mmap() and fault, some process
3551 * don't have hwpoisoned swap entry for errored virtual address.
3552 * So we need to block hugepage fault by PG_hwpoison bit check.
3553 */
3554 if (unlikely(PageHWPoison(page))) {
Chris Forbes32f84522011-07-25 17:12:14 -07003555 ret = VM_FAULT_HWPOISON |
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07003556 VM_FAULT_SET_HINDEX(hstate_index(h));
Naoya Horiguchi998b4382010-09-08 10:19:32 +09003557 goto backout_unlocked;
3558 }
Christoph Lameter6bda6662006-01-06 00:10:49 -08003559 }
David Gibson1e8f8892006-01-06 00:10:44 -08003560
Andy Whitcroft57303d82008-08-12 15:08:47 -07003561 /*
3562 * If we are going to COW a private mapping later, we examine the
3563 * pending reservations for this page now. This will ensure that
3564 * any allocations necessary to record that reservation occur outside
3565 * the spinlock.
3566 */
Mike Kravetz5e911372015-09-08 15:01:28 -07003567 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
Andy Whitcroft2b267362008-08-12 15:08:49 -07003568 if (vma_needs_reservation(h, vma, address) < 0) {
3569 ret = VM_FAULT_OOM;
3570 goto backout_unlocked;
3571 }
Mike Kravetz5e911372015-09-08 15:01:28 -07003572 /* Just decrements count, does not deallocate */
Mike Kravetzfeba16e2015-09-08 15:01:31 -07003573 vma_end_reservation(h, vma, address);
Mike Kravetz5e911372015-09-08 15:01:28 -07003574 }
Andy Whitcroft57303d82008-08-12 15:08:47 -07003575
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003576 ptl = huge_pte_lockptr(h, mm, ptep);
3577 spin_lock(ptl);
Andi Kleena5516432008-07-23 21:27:41 -07003578 size = i_size_read(mapping->host) >> huge_page_shift(h);
Adam Litke4c887262005-10-29 18:16:46 -07003579 if (idx >= size)
3580 goto backout;
3581
Nick Piggin83c54072007-07-19 01:47:05 -07003582 ret = 0;
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07003583 if (!huge_pte_none(huge_ptep_get(ptep)))
Adam Litke4c887262005-10-29 18:16:46 -07003584 goto backout;
3585
Joonsoo Kim07443a82013-09-11 14:21:58 -07003586 if (anon_rmap) {
3587 ClearPagePrivate(page);
Hillf Danton409eb8c2012-01-20 14:34:13 -08003588 hugepage_add_new_anon_rmap(page, vma, address);
Choi Gi-yongac714902014-04-07 15:37:36 -07003589 } else
Kirill A. Shutemov53f92632016-01-15 16:53:42 -08003590 page_dup_rmap(page, true);
David Gibson1e8f8892006-01-06 00:10:44 -08003591 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
3592 && (vma->vm_flags & VM_SHARED)));
3593 set_huge_pte_at(mm, address, ptep, new_pte);
3594
Naoya Horiguchi5d317b22015-11-05 18:47:14 -08003595 hugetlb_count_add(pages_per_huge_page(h), mm);
Hugh Dickins788c7df2009-06-23 13:49:05 +01003596 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
David Gibson1e8f8892006-01-06 00:10:44 -08003597 /* Optimization, do the COW without a second fault */
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003598 ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page, ptl);
David Gibson1e8f8892006-01-06 00:10:44 -08003599 }
3600
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003601 spin_unlock(ptl);
Adam Litke4c887262005-10-29 18:16:46 -07003602 unlock_page(page);
3603out:
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01003604 return ret;
Adam Litke4c887262005-10-29 18:16:46 -07003605
3606backout:
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003607 spin_unlock(ptl);
Andy Whitcroft2b267362008-08-12 15:08:49 -07003608backout_unlocked:
Adam Litke4c887262005-10-29 18:16:46 -07003609 unlock_page(page);
3610 put_page(page);
3611 goto out;
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01003612}
3613
Davidlohr Bueso8382d912014-04-03 14:47:31 -07003614#ifdef CONFIG_SMP
Mike Kravetzc672c7f2015-09-08 15:01:35 -07003615u32 hugetlb_fault_mutex_hash(struct hstate *h, struct mm_struct *mm,
Davidlohr Bueso8382d912014-04-03 14:47:31 -07003616 struct vm_area_struct *vma,
3617 struct address_space *mapping,
3618 pgoff_t idx, unsigned long address)
3619{
3620 unsigned long key[2];
3621 u32 hash;
3622
3623 if (vma->vm_flags & VM_SHARED) {
3624 key[0] = (unsigned long) mapping;
3625 key[1] = idx;
3626 } else {
3627 key[0] = (unsigned long) mm;
3628 key[1] = address >> huge_page_shift(h);
3629 }
3630
3631 hash = jhash2((u32 *)&key, sizeof(key)/sizeof(u32), 0);
3632
3633 return hash & (num_fault_mutexes - 1);
3634}
3635#else
3636/*
3637 * For uniprocesor systems we always use a single mutex, so just
3638 * return 0 and avoid the hashing overhead.
3639 */
Mike Kravetzc672c7f2015-09-08 15:01:35 -07003640u32 hugetlb_fault_mutex_hash(struct hstate *h, struct mm_struct *mm,
Davidlohr Bueso8382d912014-04-03 14:47:31 -07003641 struct vm_area_struct *vma,
3642 struct address_space *mapping,
3643 pgoff_t idx, unsigned long address)
3644{
3645 return 0;
3646}
3647#endif
3648
Adam Litke86e52162006-01-06 00:10:43 -08003649int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
Hugh Dickins788c7df2009-06-23 13:49:05 +01003650 unsigned long address, unsigned int flags)
Adam Litke86e52162006-01-06 00:10:43 -08003651{
Davidlohr Bueso8382d912014-04-03 14:47:31 -07003652 pte_t *ptep, entry;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003653 spinlock_t *ptl;
David Gibson1e8f8892006-01-06 00:10:44 -08003654 int ret;
Davidlohr Bueso8382d912014-04-03 14:47:31 -07003655 u32 hash;
3656 pgoff_t idx;
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09003657 struct page *page = NULL;
Andy Whitcroft57303d82008-08-12 15:08:47 -07003658 struct page *pagecache_page = NULL;
Andi Kleena5516432008-07-23 21:27:41 -07003659 struct hstate *h = hstate_vma(vma);
Davidlohr Bueso8382d912014-04-03 14:47:31 -07003660 struct address_space *mapping;
Naoya Horiguchi0f792cf2015-02-11 15:25:25 -08003661 int need_wait_lock = 0;
Adam Litke86e52162006-01-06 00:10:43 -08003662
KAMEZAWA Hiroyuki1e16a532012-01-10 15:07:22 -08003663 address &= huge_page_mask(h);
3664
Naoya Horiguchifd6a03ed2010-05-28 09:29:21 +09003665 ptep = huge_pte_offset(mm, address);
3666 if (ptep) {
3667 entry = huge_ptep_get(ptep);
Naoya Horiguchi290408d2010-09-08 10:19:35 +09003668 if (unlikely(is_hugetlb_entry_migration(entry))) {
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003669 migration_entry_wait_huge(vma, mm, ptep);
Naoya Horiguchi290408d2010-09-08 10:19:35 +09003670 return 0;
3671 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
Chris Forbes32f84522011-07-25 17:12:14 -07003672 return VM_FAULT_HWPOISON_LARGE |
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07003673 VM_FAULT_SET_HINDEX(hstate_index(h));
Naoya Horiguchi0d777df2015-12-11 13:40:49 -08003674 } else {
3675 ptep = huge_pte_alloc(mm, address, huge_page_size(h));
3676 if (!ptep)
3677 return VM_FAULT_OOM;
Naoya Horiguchifd6a03ed2010-05-28 09:29:21 +09003678 }
3679
Davidlohr Bueso8382d912014-04-03 14:47:31 -07003680 mapping = vma->vm_file->f_mapping;
3681 idx = vma_hugecache_offset(h, vma, address);
3682
David Gibson3935baa2006-03-22 00:08:53 -08003683 /*
3684 * Serialize hugepage allocation and instantiation, so that we don't
3685 * get spurious allocation failures if two CPUs race to instantiate
3686 * the same page in the page cache.
3687 */
Mike Kravetzc672c7f2015-09-08 15:01:35 -07003688 hash = hugetlb_fault_mutex_hash(h, mm, vma, mapping, idx, address);
3689 mutex_lock(&hugetlb_fault_mutex_table[hash]);
Davidlohr Bueso8382d912014-04-03 14:47:31 -07003690
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07003691 entry = huge_ptep_get(ptep);
3692 if (huge_pte_none(entry)) {
Davidlohr Bueso8382d912014-04-03 14:47:31 -07003693 ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags);
David Gibsonb4d1d992008-10-15 22:01:11 -07003694 goto out_mutex;
David Gibson3935baa2006-03-22 00:08:53 -08003695 }
Adam Litke86e52162006-01-06 00:10:43 -08003696
Nick Piggin83c54072007-07-19 01:47:05 -07003697 ret = 0;
David Gibson1e8f8892006-01-06 00:10:44 -08003698
Andy Whitcroft57303d82008-08-12 15:08:47 -07003699 /*
Naoya Horiguchi0f792cf2015-02-11 15:25:25 -08003700 * entry could be a migration/hwpoison entry at this point, so this
3701 * check prevents the kernel from going below assuming that we have
3702 * a active hugepage in pagecache. This goto expects the 2nd page fault,
3703 * and is_hugetlb_entry_(migration|hwpoisoned) check will properly
3704 * handle it.
3705 */
3706 if (!pte_present(entry))
3707 goto out_mutex;
3708
3709 /*
Andy Whitcroft57303d82008-08-12 15:08:47 -07003710 * If we are going to COW the mapping later, we examine the pending
3711 * reservations for this page now. This will ensure that any
3712 * allocations necessary to record that reservation occur outside the
3713 * spinlock. For private mappings, we also lookup the pagecache
3714 * page now as it is used to determine if a reservation has been
3715 * consumed.
3716 */
Gerald Schaefer106c9922013-04-29 15:07:23 -07003717 if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) {
Andy Whitcroft2b267362008-08-12 15:08:49 -07003718 if (vma_needs_reservation(h, vma, address) < 0) {
3719 ret = VM_FAULT_OOM;
David Gibsonb4d1d992008-10-15 22:01:11 -07003720 goto out_mutex;
Andy Whitcroft2b267362008-08-12 15:08:49 -07003721 }
Mike Kravetz5e911372015-09-08 15:01:28 -07003722 /* Just decrements count, does not deallocate */
Mike Kravetzfeba16e2015-09-08 15:01:31 -07003723 vma_end_reservation(h, vma, address);
Andy Whitcroft57303d82008-08-12 15:08:47 -07003724
Mel Gormanf83a2752009-05-28 14:34:40 -07003725 if (!(vma->vm_flags & VM_MAYSHARE))
Andy Whitcroft57303d82008-08-12 15:08:47 -07003726 pagecache_page = hugetlbfs_pagecache_page(h,
3727 vma, address);
3728 }
3729
Naoya Horiguchi0f792cf2015-02-11 15:25:25 -08003730 ptl = huge_pte_lock(h, mm, ptep);
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09003731
David Gibson1e8f8892006-01-06 00:10:44 -08003732 /* Check for a racing update before calling hugetlb_cow */
David Gibsonb4d1d992008-10-15 22:01:11 -07003733 if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003734 goto out_ptl;
David Gibsonb4d1d992008-10-15 22:01:11 -07003735
Naoya Horiguchi0f792cf2015-02-11 15:25:25 -08003736 /*
3737 * hugetlb_cow() requires page locks of pte_page(entry) and
3738 * pagecache_page, so here we need take the former one
3739 * when page != pagecache_page or !pagecache_page.
3740 */
3741 page = pte_page(entry);
3742 if (page != pagecache_page)
3743 if (!trylock_page(page)) {
3744 need_wait_lock = 1;
3745 goto out_ptl;
3746 }
3747
3748 get_page(page);
David Gibsonb4d1d992008-10-15 22:01:11 -07003749
Hugh Dickins788c7df2009-06-23 13:49:05 +01003750 if (flags & FAULT_FLAG_WRITE) {
Gerald Schaefer106c9922013-04-29 15:07:23 -07003751 if (!huge_pte_write(entry)) {
Andy Whitcroft57303d82008-08-12 15:08:47 -07003752 ret = hugetlb_cow(mm, vma, address, ptep, entry,
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003753 pagecache_page, ptl);
Naoya Horiguchi0f792cf2015-02-11 15:25:25 -08003754 goto out_put_page;
David Gibsonb4d1d992008-10-15 22:01:11 -07003755 }
Gerald Schaefer106c9922013-04-29 15:07:23 -07003756 entry = huge_pte_mkdirty(entry);
David Gibsonb4d1d992008-10-15 22:01:11 -07003757 }
3758 entry = pte_mkyoung(entry);
Hugh Dickins788c7df2009-06-23 13:49:05 +01003759 if (huge_ptep_set_access_flags(vma, address, ptep, entry,
3760 flags & FAULT_FLAG_WRITE))
Russell King4b3073e2009-12-18 16:40:18 +00003761 update_mmu_cache(vma, address, ptep);
Naoya Horiguchi0f792cf2015-02-11 15:25:25 -08003762out_put_page:
3763 if (page != pagecache_page)
3764 unlock_page(page);
3765 put_page(page);
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003766out_ptl:
3767 spin_unlock(ptl);
Andy Whitcroft57303d82008-08-12 15:08:47 -07003768
3769 if (pagecache_page) {
3770 unlock_page(pagecache_page);
3771 put_page(pagecache_page);
3772 }
David Gibsonb4d1d992008-10-15 22:01:11 -07003773out_mutex:
Mike Kravetzc672c7f2015-09-08 15:01:35 -07003774 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
Naoya Horiguchi0f792cf2015-02-11 15:25:25 -08003775 /*
3776 * Generally it's safe to hold refcount during waiting page lock. But
3777 * here we just wait to defer the next page fault to avoid busy loop and
3778 * the page is not used after unlocked before returning from the current
3779 * page fault. So we are safe from accessing freed page, even if we wait
3780 * here without taking refcount.
3781 */
3782 if (need_wait_lock)
3783 wait_on_page_locked(page);
David Gibson1e8f8892006-01-06 00:10:44 -08003784 return ret;
Adam Litke86e52162006-01-06 00:10:43 -08003785}
3786
Michel Lespinasse28a35712013-02-22 16:35:55 -08003787long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
3788 struct page **pages, struct vm_area_struct **vmas,
3789 unsigned long *position, unsigned long *nr_pages,
3790 long i, unsigned int flags)
David Gibson63551ae2005-06-21 17:14:44 -07003791{
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08003792 unsigned long pfn_offset;
3793 unsigned long vaddr = *position;
Michel Lespinasse28a35712013-02-22 16:35:55 -08003794 unsigned long remainder = *nr_pages;
Andi Kleena5516432008-07-23 21:27:41 -07003795 struct hstate *h = hstate_vma(vma);
David Gibson63551ae2005-06-21 17:14:44 -07003796
David Gibson63551ae2005-06-21 17:14:44 -07003797 while (vaddr < vma->vm_end && remainder) {
Adam Litke4c887262005-10-29 18:16:46 -07003798 pte_t *pte;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003799 spinlock_t *ptl = NULL;
Hugh Dickins2a15efc2009-09-21 17:03:27 -07003800 int absent;
Adam Litke4c887262005-10-29 18:16:46 -07003801 struct page *page;
3802
3803 /*
David Rientjes02057962015-04-14 15:48:24 -07003804 * If we have a pending SIGKILL, don't keep faulting pages and
3805 * potentially allocating memory.
3806 */
3807 if (unlikely(fatal_signal_pending(current))) {
3808 remainder = 0;
3809 break;
3810 }
3811
3812 /*
Adam Litke4c887262005-10-29 18:16:46 -07003813 * Some archs (sparc64, sh*) have multiple pte_ts to
Hugh Dickins2a15efc2009-09-21 17:03:27 -07003814 * each hugepage. We have to make sure we get the
Adam Litke4c887262005-10-29 18:16:46 -07003815 * first, for the page indexing below to work.
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003816 *
3817 * Note that page table lock is not held when pte is null.
Adam Litke4c887262005-10-29 18:16:46 -07003818 */
Andi Kleena5516432008-07-23 21:27:41 -07003819 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h));
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003820 if (pte)
3821 ptl = huge_pte_lock(h, mm, pte);
Hugh Dickins2a15efc2009-09-21 17:03:27 -07003822 absent = !pte || huge_pte_none(huge_ptep_get(pte));
Adam Litke4c887262005-10-29 18:16:46 -07003823
Hugh Dickins2a15efc2009-09-21 17:03:27 -07003824 /*
3825 * When coredumping, it suits get_dump_page if we just return
Hugh Dickins3ae77f42009-09-21 17:03:33 -07003826 * an error where there's an empty slot with no huge pagecache
3827 * to back it. This way, we avoid allocating a hugepage, and
3828 * the sparse dumpfile avoids allocating disk blocks, but its
3829 * huge holes still show up with zeroes where they need to be.
Hugh Dickins2a15efc2009-09-21 17:03:27 -07003830 */
Hugh Dickins3ae77f42009-09-21 17:03:33 -07003831 if (absent && (flags & FOLL_DUMP) &&
3832 !hugetlbfs_pagecache_present(h, vma, vaddr)) {
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003833 if (pte)
3834 spin_unlock(ptl);
Hugh Dickins2a15efc2009-09-21 17:03:27 -07003835 remainder = 0;
3836 break;
3837 }
3838
Naoya Horiguchi9cc3a5b2013-04-17 15:58:30 -07003839 /*
3840 * We need call hugetlb_fault for both hugepages under migration
3841 * (in which case hugetlb_fault waits for the migration,) and
3842 * hwpoisoned hugepages (in which case we need to prevent the
3843 * caller from accessing to them.) In order to do this, we use
3844 * here is_swap_pte instead of is_hugetlb_entry_migration and
3845 * is_hugetlb_entry_hwpoisoned. This is because it simply covers
3846 * both cases, and because we can't follow correct pages
3847 * directly from any kind of swap entries.
3848 */
3849 if (absent || is_swap_pte(huge_ptep_get(pte)) ||
Gerald Schaefer106c9922013-04-29 15:07:23 -07003850 ((flags & FOLL_WRITE) &&
3851 !huge_pte_write(huge_ptep_get(pte)))) {
Adam Litke4c887262005-10-29 18:16:46 -07003852 int ret;
3853
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003854 if (pte)
3855 spin_unlock(ptl);
Hugh Dickins2a15efc2009-09-21 17:03:27 -07003856 ret = hugetlb_fault(mm, vma, vaddr,
3857 (flags & FOLL_WRITE) ? FAULT_FLAG_WRITE : 0);
Adam Litkea89182c2007-08-22 14:01:51 -07003858 if (!(ret & VM_FAULT_ERROR))
Adam Litke4c887262005-10-29 18:16:46 -07003859 continue;
3860
3861 remainder = 0;
Adam Litke4c887262005-10-29 18:16:46 -07003862 break;
3863 }
David Gibson63551ae2005-06-21 17:14:44 -07003864
Andi Kleena5516432008-07-23 21:27:41 -07003865 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07003866 page = pte_page(huge_ptep_get(pte));
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08003867same_page:
Chen, Kenneth Wd6692182006-03-31 02:29:57 -08003868 if (pages) {
Hugh Dickins2a15efc2009-09-21 17:03:27 -07003869 pages[i] = mem_map_offset(page, pfn_offset);
Kirill A. Shutemovddc58f22016-01-15 16:52:56 -08003870 get_page(pages[i]);
Chen, Kenneth Wd6692182006-03-31 02:29:57 -08003871 }
David Gibson63551ae2005-06-21 17:14:44 -07003872
3873 if (vmas)
3874 vmas[i] = vma;
3875
3876 vaddr += PAGE_SIZE;
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08003877 ++pfn_offset;
David Gibson63551ae2005-06-21 17:14:44 -07003878 --remainder;
3879 ++i;
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08003880 if (vaddr < vma->vm_end && remainder &&
Andi Kleena5516432008-07-23 21:27:41 -07003881 pfn_offset < pages_per_huge_page(h)) {
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08003882 /*
3883 * We use pfn_offset to avoid touching the pageframes
3884 * of this compound page.
3885 */
3886 goto same_page;
3887 }
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003888 spin_unlock(ptl);
David Gibson63551ae2005-06-21 17:14:44 -07003889 }
Michel Lespinasse28a35712013-02-22 16:35:55 -08003890 *nr_pages = remainder;
David Gibson63551ae2005-06-21 17:14:44 -07003891 *position = vaddr;
3892
Hugh Dickins2a15efc2009-09-21 17:03:27 -07003893 return i ? i : -EFAULT;
David Gibson63551ae2005-06-21 17:14:44 -07003894}
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003895
Peter Zijlstra7da4d642012-11-19 03:14:23 +01003896unsigned long hugetlb_change_protection(struct vm_area_struct *vma,
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003897 unsigned long address, unsigned long end, pgprot_t newprot)
3898{
3899 struct mm_struct *mm = vma->vm_mm;
3900 unsigned long start = address;
3901 pte_t *ptep;
3902 pte_t pte;
Andi Kleena5516432008-07-23 21:27:41 -07003903 struct hstate *h = hstate_vma(vma);
Peter Zijlstra7da4d642012-11-19 03:14:23 +01003904 unsigned long pages = 0;
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003905
3906 BUG_ON(address >= end);
3907 flush_cache_range(vma, address, end);
3908
Rik van Riela5338092014-04-07 15:36:57 -07003909 mmu_notifier_invalidate_range_start(mm, start, end);
Davidlohr Bueso83cde9e2014-12-12 16:54:21 -08003910 i_mmap_lock_write(vma->vm_file->f_mapping);
Andi Kleena5516432008-07-23 21:27:41 -07003911 for (; address < end; address += huge_page_size(h)) {
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003912 spinlock_t *ptl;
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003913 ptep = huge_pte_offset(mm, address);
3914 if (!ptep)
3915 continue;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003916 ptl = huge_pte_lock(h, mm, ptep);
Peter Zijlstra7da4d642012-11-19 03:14:23 +01003917 if (huge_pmd_unshare(mm, &address, ptep)) {
3918 pages++;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003919 spin_unlock(ptl);
Chen, Kenneth W39dde652006-12-06 20:32:03 -08003920 continue;
Peter Zijlstra7da4d642012-11-19 03:14:23 +01003921 }
Naoya Horiguchia8bda282015-02-11 15:25:28 -08003922 pte = huge_ptep_get(ptep);
3923 if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) {
3924 spin_unlock(ptl);
3925 continue;
3926 }
3927 if (unlikely(is_hugetlb_entry_migration(pte))) {
3928 swp_entry_t entry = pte_to_swp_entry(pte);
3929
3930 if (is_write_migration_entry(entry)) {
3931 pte_t newpte;
3932
3933 make_migration_entry_read(&entry);
3934 newpte = swp_entry_to_pte(entry);
3935 set_huge_pte_at(mm, address, ptep, newpte);
3936 pages++;
3937 }
3938 spin_unlock(ptl);
3939 continue;
3940 }
3941 if (!huge_pte_none(pte)) {
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003942 pte = huge_ptep_get_and_clear(mm, address, ptep);
Gerald Schaefer106c9922013-04-29 15:07:23 -07003943 pte = pte_mkhuge(huge_pte_modify(pte, newprot));
Tony Lube7517d2013-02-04 14:28:46 -08003944 pte = arch_make_huge_pte(pte, vma, NULL, 0);
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003945 set_huge_pte_at(mm, address, ptep, pte);
Peter Zijlstra7da4d642012-11-19 03:14:23 +01003946 pages++;
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003947 }
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003948 spin_unlock(ptl);
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003949 }
Mel Gormand8333522012-07-31 16:46:20 -07003950 /*
Davidlohr Buesoc8c06ef2014-12-12 16:54:24 -08003951 * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare
Mel Gormand8333522012-07-31 16:46:20 -07003952 * may have cleared our pud entry and done put_page on the page table:
Davidlohr Buesoc8c06ef2014-12-12 16:54:24 -08003953 * once we release i_mmap_rwsem, another task can do the final put_page
Mel Gormand8333522012-07-31 16:46:20 -07003954 * and that page table be reused and filled with junk.
3955 */
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003956 flush_tlb_range(vma, start, end);
Joerg Roedel34ee6452014-11-13 13:46:09 +11003957 mmu_notifier_invalidate_range(mm, start, end);
Davidlohr Bueso83cde9e2014-12-12 16:54:21 -08003958 i_mmap_unlock_write(vma->vm_file->f_mapping);
Rik van Riela5338092014-04-07 15:36:57 -07003959 mmu_notifier_invalidate_range_end(mm, start, end);
Peter Zijlstra7da4d642012-11-19 03:14:23 +01003960
3961 return pages << h->order;
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003962}
3963
Mel Gormana1e78772008-07-23 21:27:23 -07003964int hugetlb_reserve_pages(struct inode *inode,
3965 long from, long to,
Mel Gorman5a6fe122009-02-10 14:02:27 +00003966 struct vm_area_struct *vma,
KOSAKI Motohiroca16d142011-05-26 19:16:19 +09003967 vm_flags_t vm_flags)
Adam Litkee4e574b2007-10-16 01:26:19 -07003968{
Mel Gorman17c9d122009-02-11 16:34:16 +00003969 long ret, chg;
Andi Kleena5516432008-07-23 21:27:41 -07003970 struct hstate *h = hstate_inode(inode);
David Gibson90481622012-03-21 16:34:12 -07003971 struct hugepage_subpool *spool = subpool_inode(inode);
Joonsoo Kim9119a412014-04-03 14:47:25 -07003972 struct resv_map *resv_map;
Mike Kravetz1c5ecae2015-04-15 16:13:39 -07003973 long gbl_reserve;
Adam Litkee4e574b2007-10-16 01:26:19 -07003974
Mel Gormana1e78772008-07-23 21:27:23 -07003975 /*
Mel Gorman17c9d122009-02-11 16:34:16 +00003976 * Only apply hugepage reservation if asked. At fault time, an
3977 * attempt will be made for VM_NORESERVE to allocate a page
David Gibson90481622012-03-21 16:34:12 -07003978 * without using reserves
Mel Gorman17c9d122009-02-11 16:34:16 +00003979 */
KOSAKI Motohiroca16d142011-05-26 19:16:19 +09003980 if (vm_flags & VM_NORESERVE)
Mel Gorman17c9d122009-02-11 16:34:16 +00003981 return 0;
3982
3983 /*
Mel Gormana1e78772008-07-23 21:27:23 -07003984 * Shared mappings base their reservation on the number of pages that
3985 * are already allocated on behalf of the file. Private mappings need
3986 * to reserve the full area even if read-only as mprotect() may be
3987 * called to make the mapping read-write. Assume !vma is a shm mapping
3988 */
Joonsoo Kim9119a412014-04-03 14:47:25 -07003989 if (!vma || vma->vm_flags & VM_MAYSHARE) {
Joonsoo Kim4e35f482014-04-03 14:47:30 -07003990 resv_map = inode_resv_map(inode);
Joonsoo Kim9119a412014-04-03 14:47:25 -07003991
Joonsoo Kim1406ec92014-04-03 14:47:26 -07003992 chg = region_chg(resv_map, from, to);
Joonsoo Kim9119a412014-04-03 14:47:25 -07003993
3994 } else {
3995 resv_map = resv_map_alloc();
Mel Gorman5a6fe122009-02-10 14:02:27 +00003996 if (!resv_map)
3997 return -ENOMEM;
3998
Mel Gorman17c9d122009-02-11 16:34:16 +00003999 chg = to - from;
4000
Mel Gorman5a6fe122009-02-10 14:02:27 +00004001 set_vma_resv_map(vma, resv_map);
4002 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
4003 }
4004
Dave Hansenc50ac052012-05-29 15:06:46 -07004005 if (chg < 0) {
4006 ret = chg;
4007 goto out_err;
4008 }
Mel Gorman17c9d122009-02-11 16:34:16 +00004009
Mike Kravetz1c5ecae2015-04-15 16:13:39 -07004010 /*
4011 * There must be enough pages in the subpool for the mapping. If
4012 * the subpool has a minimum size, there may be some global
4013 * reservations already in place (gbl_reserve).
4014 */
4015 gbl_reserve = hugepage_subpool_get_pages(spool, chg);
4016 if (gbl_reserve < 0) {
Dave Hansenc50ac052012-05-29 15:06:46 -07004017 ret = -ENOSPC;
4018 goto out_err;
4019 }
Mel Gorman17c9d122009-02-11 16:34:16 +00004020
4021 /*
4022 * Check enough hugepages are available for the reservation.
David Gibson90481622012-03-21 16:34:12 -07004023 * Hand the pages back to the subpool if there are not
Mel Gorman17c9d122009-02-11 16:34:16 +00004024 */
Mike Kravetz1c5ecae2015-04-15 16:13:39 -07004025 ret = hugetlb_acct_memory(h, gbl_reserve);
Mel Gorman17c9d122009-02-11 16:34:16 +00004026 if (ret < 0) {
Mike Kravetz1c5ecae2015-04-15 16:13:39 -07004027 /* put back original number of pages, chg */
4028 (void)hugepage_subpool_put_pages(spool, chg);
Dave Hansenc50ac052012-05-29 15:06:46 -07004029 goto out_err;
Mel Gorman17c9d122009-02-11 16:34:16 +00004030 }
4031
4032 /*
4033 * Account for the reservations made. Shared mappings record regions
4034 * that have reservations as they are shared by multiple VMAs.
4035 * When the last VMA disappears, the region map says how much
4036 * the reservation was and the page cache tells how much of
4037 * the reservation was consumed. Private mappings are per-VMA and
4038 * only the consumed reservations are tracked. When the VMA
4039 * disappears, the original reservation is the VMA size and the
4040 * consumed reservations are stored in the map. Hence, nothing
4041 * else has to be done for private mappings here
4042 */
Mike Kravetz33039672015-06-24 16:57:58 -07004043 if (!vma || vma->vm_flags & VM_MAYSHARE) {
4044 long add = region_add(resv_map, from, to);
4045
4046 if (unlikely(chg > add)) {
4047 /*
4048 * pages in this range were added to the reserve
4049 * map between region_chg and region_add. This
4050 * indicates a race with alloc_huge_page. Adjust
4051 * the subpool and reserve counts modified above
4052 * based on the difference.
4053 */
4054 long rsv_adjust;
4055
4056 rsv_adjust = hugepage_subpool_put_pages(spool,
4057 chg - add);
4058 hugetlb_acct_memory(h, -rsv_adjust);
4059 }
4060 }
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -07004061 return 0;
Dave Hansenc50ac052012-05-29 15:06:46 -07004062out_err:
Mike Kravetz5e911372015-09-08 15:01:28 -07004063 if (!vma || vma->vm_flags & VM_MAYSHARE)
4064 region_abort(resv_map, from, to);
Joonsoo Kimf031dd22014-04-03 14:47:28 -07004065 if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
4066 kref_put(&resv_map->refs, resv_map_release);
Dave Hansenc50ac052012-05-29 15:06:46 -07004067 return ret;
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -07004068}
4069
Mike Kravetzb5cec282015-09-08 15:01:41 -07004070long hugetlb_unreserve_pages(struct inode *inode, long start, long end,
4071 long freed)
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -07004072{
Andi Kleena5516432008-07-23 21:27:41 -07004073 struct hstate *h = hstate_inode(inode);
Joonsoo Kim4e35f482014-04-03 14:47:30 -07004074 struct resv_map *resv_map = inode_resv_map(inode);
Joonsoo Kim9119a412014-04-03 14:47:25 -07004075 long chg = 0;
David Gibson90481622012-03-21 16:34:12 -07004076 struct hugepage_subpool *spool = subpool_inode(inode);
Mike Kravetz1c5ecae2015-04-15 16:13:39 -07004077 long gbl_reserve;
Ken Chen45c682a2007-11-14 16:59:44 -08004078
Mike Kravetzb5cec282015-09-08 15:01:41 -07004079 if (resv_map) {
4080 chg = region_del(resv_map, start, end);
4081 /*
4082 * region_del() can fail in the rare case where a region
4083 * must be split and another region descriptor can not be
4084 * allocated. If end == LONG_MAX, it will not fail.
4085 */
4086 if (chg < 0)
4087 return chg;
4088 }
4089
Ken Chen45c682a2007-11-14 16:59:44 -08004090 spin_lock(&inode->i_lock);
Eric Sandeene4c6f8b2009-07-29 15:02:16 -07004091 inode->i_blocks -= (blocks_per_huge_page(h) * freed);
Ken Chen45c682a2007-11-14 16:59:44 -08004092 spin_unlock(&inode->i_lock);
4093
Mike Kravetz1c5ecae2015-04-15 16:13:39 -07004094 /*
4095 * If the subpool has a minimum size, the number of global
4096 * reservations to be released may be adjusted.
4097 */
4098 gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed));
4099 hugetlb_acct_memory(h, -gbl_reserve);
Mike Kravetzb5cec282015-09-08 15:01:41 -07004100
4101 return 0;
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -07004102}
Naoya Horiguchi93f70f92010-05-28 09:29:20 +09004103
Steve Capper3212b532013-04-23 12:35:02 +01004104#ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
4105static unsigned long page_table_shareable(struct vm_area_struct *svma,
4106 struct vm_area_struct *vma,
4107 unsigned long addr, pgoff_t idx)
4108{
4109 unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) +
4110 svma->vm_start;
4111 unsigned long sbase = saddr & PUD_MASK;
4112 unsigned long s_end = sbase + PUD_SIZE;
4113
4114 /* Allow segments to share if only one is marked locked */
Eric B Munsonde60f5f2015-11-05 18:51:36 -08004115 unsigned long vm_flags = vma->vm_flags & VM_LOCKED_CLEAR_MASK;
4116 unsigned long svm_flags = svma->vm_flags & VM_LOCKED_CLEAR_MASK;
Steve Capper3212b532013-04-23 12:35:02 +01004117
4118 /*
4119 * match the virtual addresses, permission and the alignment of the
4120 * page table page.
4121 */
4122 if (pmd_index(addr) != pmd_index(saddr) ||
4123 vm_flags != svm_flags ||
4124 sbase < svma->vm_start || svma->vm_end < s_end)
4125 return 0;
4126
4127 return saddr;
4128}
4129
Nicholas Krause31aafb42015-09-04 15:47:58 -07004130static bool vma_shareable(struct vm_area_struct *vma, unsigned long addr)
Steve Capper3212b532013-04-23 12:35:02 +01004131{
4132 unsigned long base = addr & PUD_MASK;
4133 unsigned long end = base + PUD_SIZE;
4134
4135 /*
4136 * check on proper vm_flags and page table alignment
4137 */
4138 if (vma->vm_flags & VM_MAYSHARE &&
4139 vma->vm_start <= base && end <= vma->vm_end)
Nicholas Krause31aafb42015-09-04 15:47:58 -07004140 return true;
4141 return false;
Steve Capper3212b532013-04-23 12:35:02 +01004142}
4143
4144/*
4145 * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
4146 * and returns the corresponding pte. While this is not necessary for the
4147 * !shared pmd case because we can allocate the pmd later as well, it makes the
4148 * code much cleaner. pmd allocation is essential for the shared case because
Davidlohr Buesoc8c06ef2014-12-12 16:54:24 -08004149 * pud has to be populated inside the same i_mmap_rwsem section - otherwise
Steve Capper3212b532013-04-23 12:35:02 +01004150 * racing tasks could either miss the sharing (see huge_pte_offset) or select a
4151 * bad pmd for sharing.
4152 */
4153pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
4154{
4155 struct vm_area_struct *vma = find_vma(mm, addr);
4156 struct address_space *mapping = vma->vm_file->f_mapping;
4157 pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) +
4158 vma->vm_pgoff;
4159 struct vm_area_struct *svma;
4160 unsigned long saddr;
4161 pte_t *spte = NULL;
4162 pte_t *pte;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08004163 spinlock_t *ptl;
Steve Capper3212b532013-04-23 12:35:02 +01004164
4165 if (!vma_shareable(vma, addr))
4166 return (pte_t *)pmd_alloc(mm, pud, addr);
4167
Davidlohr Bueso83cde9e2014-12-12 16:54:21 -08004168 i_mmap_lock_write(mapping);
Steve Capper3212b532013-04-23 12:35:02 +01004169 vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) {
4170 if (svma == vma)
4171 continue;
4172
4173 saddr = page_table_shareable(svma, vma, addr, idx);
4174 if (saddr) {
4175 spte = huge_pte_offset(svma->vm_mm, saddr);
4176 if (spte) {
Kirill A. Shutemovdc6c9a32015-02-11 15:26:50 -08004177 mm_inc_nr_pmds(mm);
Steve Capper3212b532013-04-23 12:35:02 +01004178 get_page(virt_to_page(spte));
4179 break;
4180 }
4181 }
4182 }
4183
4184 if (!spte)
4185 goto out;
4186
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08004187 ptl = huge_pte_lockptr(hstate_vma(vma), mm, spte);
4188 spin_lock(ptl);
Kirill A. Shutemovdc6c9a32015-02-11 15:26:50 -08004189 if (pud_none(*pud)) {
Steve Capper3212b532013-04-23 12:35:02 +01004190 pud_populate(mm, pud,
4191 (pmd_t *)((unsigned long)spte & PAGE_MASK));
Kirill A. Shutemovdc6c9a32015-02-11 15:26:50 -08004192 } else {
Steve Capper3212b532013-04-23 12:35:02 +01004193 put_page(virt_to_page(spte));
Kirill A. Shutemovdc6c9a32015-02-11 15:26:50 -08004194 mm_inc_nr_pmds(mm);
4195 }
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08004196 spin_unlock(ptl);
Steve Capper3212b532013-04-23 12:35:02 +01004197out:
4198 pte = (pte_t *)pmd_alloc(mm, pud, addr);
Davidlohr Bueso83cde9e2014-12-12 16:54:21 -08004199 i_mmap_unlock_write(mapping);
Steve Capper3212b532013-04-23 12:35:02 +01004200 return pte;
4201}
4202
4203/*
4204 * unmap huge page backed by shared pte.
4205 *
4206 * Hugetlb pte page is ref counted at the time of mapping. If pte is shared
4207 * indicated by page_count > 1, unmap is achieved by clearing pud and
4208 * decrementing the ref count. If count == 1, the pte page is not shared.
4209 *
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08004210 * called with page table lock held.
Steve Capper3212b532013-04-23 12:35:02 +01004211 *
4212 * returns: 1 successfully unmapped a shared pte page
4213 * 0 the underlying pte page is not shared, or it is the last user
4214 */
4215int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
4216{
4217 pgd_t *pgd = pgd_offset(mm, *addr);
4218 pud_t *pud = pud_offset(pgd, *addr);
4219
4220 BUG_ON(page_count(virt_to_page(ptep)) == 0);
4221 if (page_count(virt_to_page(ptep)) == 1)
4222 return 0;
4223
4224 pud_clear(pud);
4225 put_page(virt_to_page(ptep));
Kirill A. Shutemovdc6c9a32015-02-11 15:26:50 -08004226 mm_dec_nr_pmds(mm);
Steve Capper3212b532013-04-23 12:35:02 +01004227 *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE;
4228 return 1;
4229}
Steve Capper9e5fc742013-04-30 08:02:03 +01004230#define want_pmd_share() (1)
4231#else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
4232pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
4233{
4234 return NULL;
4235}
Zhang Zhene81f2d22015-06-24 16:56:13 -07004236
4237int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
4238{
4239 return 0;
4240}
Steve Capper9e5fc742013-04-30 08:02:03 +01004241#define want_pmd_share() (0)
Steve Capper3212b532013-04-23 12:35:02 +01004242#endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
4243
Steve Capper9e5fc742013-04-30 08:02:03 +01004244#ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
4245pte_t *huge_pte_alloc(struct mm_struct *mm,
4246 unsigned long addr, unsigned long sz)
4247{
4248 pgd_t *pgd;
4249 pud_t *pud;
4250 pte_t *pte = NULL;
4251
4252 pgd = pgd_offset(mm, addr);
4253 pud = pud_alloc(mm, pgd, addr);
4254 if (pud) {
4255 if (sz == PUD_SIZE) {
4256 pte = (pte_t *)pud;
4257 } else {
4258 BUG_ON(sz != PMD_SIZE);
4259 if (want_pmd_share() && pud_none(*pud))
4260 pte = huge_pmd_share(mm, addr, pud);
4261 else
4262 pte = (pte_t *)pmd_alloc(mm, pud, addr);
4263 }
4264 }
4265 BUG_ON(pte && !pte_none(*pte) && !pte_huge(*pte));
4266
4267 return pte;
4268}
4269
4270pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr)
4271{
4272 pgd_t *pgd;
4273 pud_t *pud;
4274 pmd_t *pmd = NULL;
4275
4276 pgd = pgd_offset(mm, addr);
4277 if (pgd_present(*pgd)) {
4278 pud = pud_offset(pgd, addr);
4279 if (pud_present(*pud)) {
4280 if (pud_huge(*pud))
4281 return (pte_t *)pud;
4282 pmd = pmd_offset(pud, addr);
4283 }
4284 }
4285 return (pte_t *) pmd;
4286}
4287
Naoya Horiguchi61f77ed2015-02-11 15:25:15 -08004288#endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
4289
4290/*
4291 * These functions are overwritable if your architecture needs its own
4292 * behavior.
4293 */
4294struct page * __weak
4295follow_huge_addr(struct mm_struct *mm, unsigned long address,
4296 int write)
4297{
4298 return ERR_PTR(-EINVAL);
4299}
4300
4301struct page * __weak
Steve Capper9e5fc742013-04-30 08:02:03 +01004302follow_huge_pmd(struct mm_struct *mm, unsigned long address,
Naoya Horiguchie66f17f2015-02-11 15:25:22 -08004303 pmd_t *pmd, int flags)
Steve Capper9e5fc742013-04-30 08:02:03 +01004304{
Naoya Horiguchie66f17f2015-02-11 15:25:22 -08004305 struct page *page = NULL;
4306 spinlock_t *ptl;
4307retry:
4308 ptl = pmd_lockptr(mm, pmd);
4309 spin_lock(ptl);
4310 /*
4311 * make sure that the address range covered by this pmd is not
4312 * unmapped from other threads.
4313 */
4314 if (!pmd_huge(*pmd))
4315 goto out;
4316 if (pmd_present(*pmd)) {
Gerald Schaefer97534122015-04-14 15:42:30 -07004317 page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT);
Naoya Horiguchie66f17f2015-02-11 15:25:22 -08004318 if (flags & FOLL_GET)
4319 get_page(page);
4320 } else {
4321 if (is_hugetlb_entry_migration(huge_ptep_get((pte_t *)pmd))) {
4322 spin_unlock(ptl);
4323 __migration_entry_wait(mm, (pte_t *)pmd, ptl);
4324 goto retry;
4325 }
4326 /*
4327 * hwpoisoned entry is treated as no_page_table in
4328 * follow_page_mask().
4329 */
4330 }
4331out:
4332 spin_unlock(ptl);
Steve Capper9e5fc742013-04-30 08:02:03 +01004333 return page;
4334}
4335
Naoya Horiguchi61f77ed2015-02-11 15:25:15 -08004336struct page * __weak
Steve Capper9e5fc742013-04-30 08:02:03 +01004337follow_huge_pud(struct mm_struct *mm, unsigned long address,
Naoya Horiguchie66f17f2015-02-11 15:25:22 -08004338 pud_t *pud, int flags)
Steve Capper9e5fc742013-04-30 08:02:03 +01004339{
Naoya Horiguchie66f17f2015-02-11 15:25:22 -08004340 if (flags & FOLL_GET)
4341 return NULL;
Steve Capper9e5fc742013-04-30 08:02:03 +01004342
Naoya Horiguchie66f17f2015-02-11 15:25:22 -08004343 return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT);
Steve Capper9e5fc742013-04-30 08:02:03 +01004344}
4345
Andi Kleend5bd9102010-09-27 09:00:12 +02004346#ifdef CONFIG_MEMORY_FAILURE
4347
Naoya Horiguchi93f70f92010-05-28 09:29:20 +09004348/*
4349 * This function is called from memory failure code.
4350 * Assume the caller holds page lock of the head page.
4351 */
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +09004352int dequeue_hwpoisoned_huge_page(struct page *hpage)
Naoya Horiguchi93f70f92010-05-28 09:29:20 +09004353{
4354 struct hstate *h = page_hstate(hpage);
4355 int nid = page_to_nid(hpage);
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +09004356 int ret = -EBUSY;
Naoya Horiguchi93f70f92010-05-28 09:29:20 +09004357
4358 spin_lock(&hugetlb_lock);
Naoya Horiguchi7e1f0492015-04-15 16:14:41 -07004359 /*
4360 * Just checking !page_huge_active is not enough, because that could be
4361 * an isolated/hwpoisoned hugepage (which have >0 refcount).
4362 */
4363 if (!page_huge_active(hpage) && !page_count(hpage)) {
Naoya Horiguchi56f2fb12012-12-12 13:52:33 -08004364 /*
4365 * Hwpoisoned hugepage isn't linked to activelist or freelist,
4366 * but dangling hpage->lru can trigger list-debug warnings
4367 * (this happens when we call unpoison_memory() on it),
4368 * so let it point to itself with list_del_init().
4369 */
4370 list_del_init(&hpage->lru);
Naoya Horiguchi8c6c2ec2010-09-08 10:19:38 +09004371 set_page_refcounted(hpage);
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +09004372 h->free_huge_pages--;
4373 h->free_huge_pages_node[nid]--;
4374 ret = 0;
4375 }
Naoya Horiguchi93f70f92010-05-28 09:29:20 +09004376 spin_unlock(&hugetlb_lock);
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +09004377 return ret;
Naoya Horiguchi93f70f92010-05-28 09:29:20 +09004378}
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +09004379#endif
Naoya Horiguchi31caf662013-09-11 14:21:59 -07004380
4381bool isolate_huge_page(struct page *page, struct list_head *list)
4382{
Naoya Horiguchibcc54222015-04-15 16:14:38 -07004383 bool ret = true;
4384
Sasha Levin309381fea2014-01-23 15:52:54 -08004385 VM_BUG_ON_PAGE(!PageHead(page), page);
Naoya Horiguchi31caf662013-09-11 14:21:59 -07004386 spin_lock(&hugetlb_lock);
Naoya Horiguchibcc54222015-04-15 16:14:38 -07004387 if (!page_huge_active(page) || !get_page_unless_zero(page)) {
4388 ret = false;
4389 goto unlock;
4390 }
4391 clear_page_huge_active(page);
Naoya Horiguchi31caf662013-09-11 14:21:59 -07004392 list_move_tail(&page->lru, list);
Naoya Horiguchibcc54222015-04-15 16:14:38 -07004393unlock:
Naoya Horiguchi31caf662013-09-11 14:21:59 -07004394 spin_unlock(&hugetlb_lock);
Naoya Horiguchibcc54222015-04-15 16:14:38 -07004395 return ret;
Naoya Horiguchi31caf662013-09-11 14:21:59 -07004396}
4397
4398void putback_active_hugepage(struct page *page)
4399{
Sasha Levin309381fea2014-01-23 15:52:54 -08004400 VM_BUG_ON_PAGE(!PageHead(page), page);
Naoya Horiguchi31caf662013-09-11 14:21:59 -07004401 spin_lock(&hugetlb_lock);
Naoya Horiguchibcc54222015-04-15 16:14:38 -07004402 set_page_huge_active(page);
Naoya Horiguchi31caf662013-09-11 14:21:59 -07004403 list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist);
4404 spin_unlock(&hugetlb_lock);
4405 put_page(page);
4406}