blob: 827bb02a43a4e425393c296db89e7a49d6adf02b [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>
7#include <linux/module.h>
8#include <linux/mm.h>
Alexey Dobriyane1759c22008-10-15 23:50:22 +04009#include <linux/seq_file.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070010#include <linux/sysctl.h>
11#include <linux/highmem.h>
Andrea Arcangelicddb8a52008-07-28 15:46:29 -070012#include <linux/mmu_notifier.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070013#include <linux/nodemask.h>
David Gibson63551ae2005-06-21 17:14:44 -070014#include <linux/pagemap.h>
Christoph Lameter5da7ca82006-01-06 00:10:46 -080015#include <linux/mempolicy.h>
Gideon Israel Dsouza3b321232014-04-07 15:37:26 -070016#include <linux/compiler.h>
Christoph Lameteraea47ff2006-01-08 01:00:57 -080017#include <linux/cpuset.h>
David Gibson3935baa2006-03-22 00:08:53 -080018#include <linux/mutex.h>
Andi Kleenaa888a72008-07-23 21:27:47 -070019#include <linux/bootmem.h>
Nishanth Aravamudana3437872008-07-23 21:27:44 -070020#include <linux/sysfs.h>
Tejun Heo5a0e3ad2010-03-24 17:04:11 +090021#include <linux/slab.h>
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +090022#include <linux/rmap.h>
Naoya Horiguchifd6a03ed2010-05-28 09:29:21 +090023#include <linux/swap.h>
24#include <linux/swapops.h>
Naoya Horiguchic8721bb2013-09-11 14:22:09 -070025#include <linux/page-isolation.h>
Davidlohr Bueso8382d912014-04-03 14:47:31 -070026#include <linux/jhash.h>
Linus Torvaldsd6606682008-08-06 12:04:54 -070027
David Gibson63551ae2005-06-21 17:14:44 -070028#include <asm/page.h>
29#include <asm/pgtable.h>
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -070030#include <asm/tlb.h>
David Gibson63551ae2005-06-21 17:14:44 -070031
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -070032#include <linux/io.h>
David Gibson63551ae2005-06-21 17:14:44 -070033#include <linux/hugetlb.h>
Aneesh Kumar K.V9dd540e2012-07-31 16:42:15 -070034#include <linux/hugetlb_cgroup.h>
Lee Schermerhorn9a305232009-12-14 17:58:25 -080035#include <linux/node.h>
Nick Piggin7835e982006-03-22 00:08:40 -080036#include "internal.h"
Linus Torvalds1da177e2005-04-16 15:20:36 -070037
Andrey Ryabinin753162c2015-02-10 14:11:36 -080038int hugepages_treat_as_movable;
Andi Kleena5516432008-07-23 21:27:41 -070039
Aneesh Kumar K.Vc3f38a32012-07-31 16:42:10 -070040int hugetlb_max_hstate __read_mostly;
Andi Kleene5ff2152008-07-23 21:27:42 -070041unsigned int default_hstate_idx;
42struct hstate hstates[HUGE_MAX_HSTATE];
Naoya Horiguchi641844f2015-06-24 16:56:59 -070043/*
44 * Minimum page order among possible hugepage sizes, set to a proper value
45 * at boot time.
46 */
47static unsigned int minimum_order __read_mostly = UINT_MAX;
Andi Kleene5ff2152008-07-23 21:27:42 -070048
Jon Tollefson53ba51d2008-07-23 21:27:52 -070049__initdata LIST_HEAD(huge_boot_pages);
50
Andi Kleene5ff2152008-07-23 21:27:42 -070051/* for command line parsing */
52static struct hstate * __initdata parsed_hstate;
53static unsigned long __initdata default_hstate_max_huge_pages;
Nick Piggine11bfbf2008-07-23 21:27:52 -070054static unsigned long __initdata default_hstate_size;
Andi Kleene5ff2152008-07-23 21:27:42 -070055
David Gibson3935baa2006-03-22 00:08:53 -080056/*
Naoya Horiguchi31caf662013-09-11 14:21:59 -070057 * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages,
58 * free_huge_pages, and surplus_huge_pages.
David Gibson3935baa2006-03-22 00:08:53 -080059 */
Aneesh Kumar K.Vc3f38a32012-07-31 16:42:10 -070060DEFINE_SPINLOCK(hugetlb_lock);
Eric Paris0bd0f9f2005-11-21 21:32:28 -080061
Davidlohr Bueso8382d912014-04-03 14:47:31 -070062/*
63 * Serializes faults on the same logical page. This is used to
64 * prevent spurious OOMs when the hugepage pool is fully utilized.
65 */
66static int num_fault_mutexes;
Mike Kravetzc672c7f2015-09-08 15:01:35 -070067struct mutex *hugetlb_fault_mutex_table ____cacheline_aligned_in_smp;
Davidlohr Bueso8382d912014-04-03 14:47:31 -070068
Mike Kravetz7ca02d02015-04-15 16:13:42 -070069/* Forward declaration */
70static int hugetlb_acct_memory(struct hstate *h, long delta);
71
David Gibson90481622012-03-21 16:34:12 -070072static inline void unlock_or_release_subpool(struct hugepage_subpool *spool)
73{
74 bool free = (spool->count == 0) && (spool->used_hpages == 0);
75
76 spin_unlock(&spool->lock);
77
78 /* If no pages are used, and no other handles to the subpool
Mike Kravetz7ca02d02015-04-15 16:13:42 -070079 * remain, give up any reservations mased on minimum size and
80 * free the subpool */
81 if (free) {
82 if (spool->min_hpages != -1)
83 hugetlb_acct_memory(spool->hstate,
84 -spool->min_hpages);
David Gibson90481622012-03-21 16:34:12 -070085 kfree(spool);
Mike Kravetz7ca02d02015-04-15 16:13:42 -070086 }
David Gibson90481622012-03-21 16:34:12 -070087}
88
Mike Kravetz7ca02d02015-04-15 16:13:42 -070089struct hugepage_subpool *hugepage_new_subpool(struct hstate *h, long max_hpages,
90 long min_hpages)
David Gibson90481622012-03-21 16:34:12 -070091{
92 struct hugepage_subpool *spool;
93
Mike Kravetzc6a91822015-04-15 16:13:36 -070094 spool = kzalloc(sizeof(*spool), GFP_KERNEL);
David Gibson90481622012-03-21 16:34:12 -070095 if (!spool)
96 return NULL;
97
98 spin_lock_init(&spool->lock);
99 spool->count = 1;
Mike Kravetz7ca02d02015-04-15 16:13:42 -0700100 spool->max_hpages = max_hpages;
101 spool->hstate = h;
102 spool->min_hpages = min_hpages;
103
104 if (min_hpages != -1 && hugetlb_acct_memory(h, min_hpages)) {
105 kfree(spool);
106 return NULL;
107 }
108 spool->rsv_hpages = min_hpages;
David Gibson90481622012-03-21 16:34:12 -0700109
110 return spool;
111}
112
113void hugepage_put_subpool(struct hugepage_subpool *spool)
114{
115 spin_lock(&spool->lock);
116 BUG_ON(!spool->count);
117 spool->count--;
118 unlock_or_release_subpool(spool);
119}
120
Mike Kravetz1c5ecae2015-04-15 16:13:39 -0700121/*
122 * Subpool accounting for allocating and reserving pages.
123 * Return -ENOMEM if there are not enough resources to satisfy the
124 * the request. Otherwise, return the number of pages by which the
125 * global pools must be adjusted (upward). The returned value may
126 * only be different than the passed value (delta) in the case where
127 * a subpool minimum size must be manitained.
128 */
129static long hugepage_subpool_get_pages(struct hugepage_subpool *spool,
David Gibson90481622012-03-21 16:34:12 -0700130 long delta)
131{
Mike Kravetz1c5ecae2015-04-15 16:13:39 -0700132 long ret = delta;
David Gibson90481622012-03-21 16:34:12 -0700133
134 if (!spool)
Mike Kravetz1c5ecae2015-04-15 16:13:39 -0700135 return ret;
David Gibson90481622012-03-21 16:34:12 -0700136
137 spin_lock(&spool->lock);
David Gibson90481622012-03-21 16:34:12 -0700138
Mike Kravetz1c5ecae2015-04-15 16:13:39 -0700139 if (spool->max_hpages != -1) { /* maximum size accounting */
140 if ((spool->used_hpages + delta) <= spool->max_hpages)
141 spool->used_hpages += delta;
142 else {
143 ret = -ENOMEM;
144 goto unlock_ret;
145 }
146 }
147
148 if (spool->min_hpages != -1) { /* minimum size accounting */
149 if (delta > spool->rsv_hpages) {
150 /*
151 * Asking for more reserves than those already taken on
152 * behalf of subpool. Return difference.
153 */
154 ret = delta - spool->rsv_hpages;
155 spool->rsv_hpages = 0;
156 } else {
157 ret = 0; /* reserves already accounted for */
158 spool->rsv_hpages -= delta;
159 }
160 }
161
162unlock_ret:
163 spin_unlock(&spool->lock);
David Gibson90481622012-03-21 16:34:12 -0700164 return ret;
165}
166
Mike Kravetz1c5ecae2015-04-15 16:13:39 -0700167/*
168 * Subpool accounting for freeing and unreserving pages.
169 * Return the number of global page reservations that must be dropped.
170 * The return value may only be different than the passed value (delta)
171 * in the case where a subpool minimum size must be maintained.
172 */
173static long hugepage_subpool_put_pages(struct hugepage_subpool *spool,
David Gibson90481622012-03-21 16:34:12 -0700174 long delta)
175{
Mike Kravetz1c5ecae2015-04-15 16:13:39 -0700176 long ret = delta;
177
David Gibson90481622012-03-21 16:34:12 -0700178 if (!spool)
Mike Kravetz1c5ecae2015-04-15 16:13:39 -0700179 return delta;
David Gibson90481622012-03-21 16:34:12 -0700180
181 spin_lock(&spool->lock);
Mike Kravetz1c5ecae2015-04-15 16:13:39 -0700182
183 if (spool->max_hpages != -1) /* maximum size accounting */
184 spool->used_hpages -= delta;
185
186 if (spool->min_hpages != -1) { /* minimum size accounting */
187 if (spool->rsv_hpages + delta <= spool->min_hpages)
188 ret = 0;
189 else
190 ret = spool->rsv_hpages + delta - spool->min_hpages;
191
192 spool->rsv_hpages += delta;
193 if (spool->rsv_hpages > spool->min_hpages)
194 spool->rsv_hpages = spool->min_hpages;
195 }
196
197 /*
198 * If hugetlbfs_put_super couldn't free spool due to an outstanding
199 * quota reference, free it now.
200 */
David Gibson90481622012-03-21 16:34:12 -0700201 unlock_or_release_subpool(spool);
Mike Kravetz1c5ecae2015-04-15 16:13:39 -0700202
203 return ret;
David Gibson90481622012-03-21 16:34:12 -0700204}
205
206static inline struct hugepage_subpool *subpool_inode(struct inode *inode)
207{
208 return HUGETLBFS_SB(inode->i_sb)->spool;
209}
210
211static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma)
212{
Al Viro496ad9a2013-01-23 17:07:38 -0500213 return subpool_inode(file_inode(vma->vm_file));
David Gibson90481622012-03-21 16:34:12 -0700214}
215
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700216/*
Andy Whitcroft96822902008-07-23 21:27:29 -0700217 * Region tracking -- allows tracking of reservations and instantiated pages
218 * across the pages in a mapping.
Andy Whitcroft84afd992008-07-23 21:27:32 -0700219 *
Mike Kravetz1dd308a2015-06-24 16:57:52 -0700220 * The region data structures are embedded into a resv_map and protected
221 * by a resv_map's lock. The set of regions within the resv_map represent
222 * reservations for huge pages, or huge pages that have already been
223 * instantiated within the map. The from and to elements are huge page
224 * indicies into the associated mapping. from indicates the starting index
225 * of the region. to represents the first index past the end of the region.
226 *
227 * For example, a file region structure with from == 0 and to == 4 represents
228 * four huge pages in a mapping. It is important to note that the to element
229 * represents the first element past the end of the region. This is used in
230 * arithmetic as 4(to) - 0(from) = 4 huge pages in the region.
231 *
232 * Interval notation of the form [from, to) will be used to indicate that
233 * the endpoint from is inclusive and to is exclusive.
Andy Whitcroft96822902008-07-23 21:27:29 -0700234 */
235struct file_region {
236 struct list_head link;
237 long from;
238 long to;
239};
240
Mike Kravetz1dd308a2015-06-24 16:57:52 -0700241/*
242 * Add the huge page range represented by [f, t) to the reserve
Mike Kravetz5e911372015-09-08 15:01:28 -0700243 * map. In the normal case, existing regions will be expanded
244 * to accommodate the specified range. Sufficient regions should
245 * exist for expansion due to the previous call to region_chg
246 * with the same range. However, it is possible that region_del
247 * could have been called after region_chg and modifed the map
248 * in such a way that no region exists to be expanded. In this
249 * case, pull a region descriptor from the cache associated with
250 * the map and use that for the new range.
Mike Kravetzcf3ad202015-06-24 16:57:55 -0700251 *
252 * Return the number of new huge pages added to the map. This
253 * number is greater than or equal to zero.
Mike Kravetz1dd308a2015-06-24 16:57:52 -0700254 */
Joonsoo Kim1406ec92014-04-03 14:47:26 -0700255static long region_add(struct resv_map *resv, long f, long t)
Andy Whitcroft96822902008-07-23 21:27:29 -0700256{
Joonsoo Kim1406ec92014-04-03 14:47:26 -0700257 struct list_head *head = &resv->regions;
Andy Whitcroft96822902008-07-23 21:27:29 -0700258 struct file_region *rg, *nrg, *trg;
Mike Kravetzcf3ad202015-06-24 16:57:55 -0700259 long add = 0;
Andy Whitcroft96822902008-07-23 21:27:29 -0700260
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700261 spin_lock(&resv->lock);
Andy Whitcroft96822902008-07-23 21:27:29 -0700262 /* Locate the region we are either in or before. */
263 list_for_each_entry(rg, head, link)
264 if (f <= rg->to)
265 break;
266
Mike Kravetz5e911372015-09-08 15:01:28 -0700267 /*
268 * If no region exists which can be expanded to include the
269 * specified range, the list must have been modified by an
270 * interleving call to region_del(). Pull a region descriptor
271 * from the cache and use it for this range.
272 */
273 if (&rg->link == head || t < rg->from) {
274 VM_BUG_ON(resv->region_cache_count <= 0);
275
276 resv->region_cache_count--;
277 nrg = list_first_entry(&resv->region_cache, struct file_region,
278 link);
279 list_del(&nrg->link);
280
281 nrg->from = f;
282 nrg->to = t;
283 list_add(&nrg->link, rg->link.prev);
284
285 add += t - f;
286 goto out_locked;
287 }
288
Andy Whitcroft96822902008-07-23 21:27:29 -0700289 /* Round our left edge to the current segment if it encloses us. */
290 if (f > rg->from)
291 f = rg->from;
292
293 /* Check for and consume any regions we now overlap with. */
294 nrg = rg;
295 list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
296 if (&rg->link == head)
297 break;
298 if (rg->from > t)
299 break;
300
301 /* If this area reaches higher then extend our area to
302 * include it completely. If this is not the first area
303 * which we intend to reuse, free it. */
304 if (rg->to > t)
305 t = rg->to;
306 if (rg != nrg) {
Mike Kravetzcf3ad202015-06-24 16:57:55 -0700307 /* Decrement return value by the deleted range.
308 * Another range will span this area so that by
309 * end of routine add will be >= zero
310 */
311 add -= (rg->to - rg->from);
Andy Whitcroft96822902008-07-23 21:27:29 -0700312 list_del(&rg->link);
313 kfree(rg);
314 }
315 }
Mike Kravetzcf3ad202015-06-24 16:57:55 -0700316
317 add += (nrg->from - f); /* Added to beginning of region */
Andy Whitcroft96822902008-07-23 21:27:29 -0700318 nrg->from = f;
Mike Kravetzcf3ad202015-06-24 16:57:55 -0700319 add += t - nrg->to; /* Added to end of region */
Andy Whitcroft96822902008-07-23 21:27:29 -0700320 nrg->to = t;
Mike Kravetzcf3ad202015-06-24 16:57:55 -0700321
Mike Kravetz5e911372015-09-08 15:01:28 -0700322out_locked:
323 resv->adds_in_progress--;
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700324 spin_unlock(&resv->lock);
Mike Kravetzcf3ad202015-06-24 16:57:55 -0700325 VM_BUG_ON(add < 0);
326 return add;
Andy Whitcroft96822902008-07-23 21:27:29 -0700327}
328
Mike Kravetz1dd308a2015-06-24 16:57:52 -0700329/*
330 * Examine the existing reserve map and determine how many
331 * huge pages in the specified range [f, t) are NOT currently
332 * represented. This routine is called before a subsequent
333 * call to region_add that will actually modify the reserve
334 * map to add the specified range [f, t). region_chg does
335 * not change the number of huge pages represented by the
336 * map. However, if the existing regions in the map can not
337 * be expanded to represent the new range, a new file_region
338 * structure is added to the map as a placeholder. This is
339 * so that the subsequent region_add call will have all the
340 * regions it needs and will not fail.
341 *
Mike Kravetz5e911372015-09-08 15:01:28 -0700342 * Upon entry, region_chg will also examine the cache of region descriptors
343 * associated with the map. If there are not enough descriptors cached, one
344 * will be allocated for the in progress add operation.
345 *
346 * Returns the number of huge pages that need to be added to the existing
347 * reservation map for the range [f, t). This number is greater or equal to
348 * zero. -ENOMEM is returned if a new file_region structure or cache entry
349 * is needed and can not be allocated.
Mike Kravetz1dd308a2015-06-24 16:57:52 -0700350 */
Joonsoo Kim1406ec92014-04-03 14:47:26 -0700351static long region_chg(struct resv_map *resv, long f, long t)
Andy Whitcroft96822902008-07-23 21:27:29 -0700352{
Joonsoo Kim1406ec92014-04-03 14:47:26 -0700353 struct list_head *head = &resv->regions;
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700354 struct file_region *rg, *nrg = NULL;
Andy Whitcroft96822902008-07-23 21:27:29 -0700355 long chg = 0;
356
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700357retry:
358 spin_lock(&resv->lock);
Mike Kravetz5e911372015-09-08 15:01:28 -0700359retry_locked:
360 resv->adds_in_progress++;
361
362 /*
363 * Check for sufficient descriptors in the cache to accommodate
364 * the number of in progress add operations.
365 */
366 if (resv->adds_in_progress > resv->region_cache_count) {
367 struct file_region *trg;
368
369 VM_BUG_ON(resv->adds_in_progress - resv->region_cache_count > 1);
370 /* Must drop lock to allocate a new descriptor. */
371 resv->adds_in_progress--;
372 spin_unlock(&resv->lock);
373
374 trg = kmalloc(sizeof(*trg), GFP_KERNEL);
375 if (!trg)
376 return -ENOMEM;
377
378 spin_lock(&resv->lock);
379 list_add(&trg->link, &resv->region_cache);
380 resv->region_cache_count++;
381 goto retry_locked;
382 }
383
Andy Whitcroft96822902008-07-23 21:27:29 -0700384 /* Locate the region we are before or in. */
385 list_for_each_entry(rg, head, link)
386 if (f <= rg->to)
387 break;
388
389 /* If we are below the current region then a new region is required.
390 * Subtle, allocate a new region at the position but make it zero
391 * size such that we can guarantee to record the reservation. */
392 if (&rg->link == head || t < rg->from) {
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700393 if (!nrg) {
Mike Kravetz5e911372015-09-08 15:01:28 -0700394 resv->adds_in_progress--;
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700395 spin_unlock(&resv->lock);
396 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
397 if (!nrg)
398 return -ENOMEM;
Andy Whitcroft96822902008-07-23 21:27:29 -0700399
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700400 nrg->from = f;
401 nrg->to = f;
402 INIT_LIST_HEAD(&nrg->link);
403 goto retry;
404 }
405
406 list_add(&nrg->link, rg->link.prev);
407 chg = t - f;
408 goto out_nrg;
Andy Whitcroft96822902008-07-23 21:27:29 -0700409 }
410
411 /* Round our left edge to the current segment if it encloses us. */
412 if (f > rg->from)
413 f = rg->from;
414 chg = t - f;
415
416 /* Check for and consume any regions we now overlap with. */
417 list_for_each_entry(rg, rg->link.prev, link) {
418 if (&rg->link == head)
419 break;
420 if (rg->from > t)
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700421 goto out;
Andy Whitcroft96822902008-07-23 21:27:29 -0700422
Lucas De Marchi25985ed2011-03-30 22:57:33 -0300423 /* We overlap with this area, if it extends further than
Andy Whitcroft96822902008-07-23 21:27:29 -0700424 * us then we must extend ourselves. Account for its
425 * existing reservation. */
426 if (rg->to > t) {
427 chg += rg->to - t;
428 t = rg->to;
429 }
430 chg -= rg->to - rg->from;
431 }
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700432
433out:
434 spin_unlock(&resv->lock);
435 /* We already know we raced and no longer need the new region */
436 kfree(nrg);
437 return chg;
438out_nrg:
439 spin_unlock(&resv->lock);
Andy Whitcroft96822902008-07-23 21:27:29 -0700440 return chg;
441}
442
Mike Kravetz1dd308a2015-06-24 16:57:52 -0700443/*
Mike Kravetz5e911372015-09-08 15:01:28 -0700444 * Abort the in progress add operation. The adds_in_progress field
445 * of the resv_map keeps track of the operations in progress between
446 * calls to region_chg and region_add. Operations are sometimes
447 * aborted after the call to region_chg. In such cases, region_abort
448 * is called to decrement the adds_in_progress counter.
449 *
450 * NOTE: The range arguments [f, t) are not needed or used in this
451 * routine. They are kept to make reading the calling code easier as
452 * arguments will match the associated region_chg call.
453 */
454static void region_abort(struct resv_map *resv, long f, long t)
455{
456 spin_lock(&resv->lock);
457 VM_BUG_ON(!resv->region_cache_count);
458 resv->adds_in_progress--;
459 spin_unlock(&resv->lock);
460}
461
462/*
Mike Kravetzfeba16e2015-09-08 15:01:31 -0700463 * Delete the specified range [f, t) from the reserve map. If the
464 * t parameter is LONG_MAX, this indicates that ALL regions after f
465 * should be deleted. Locate the regions which intersect [f, t)
466 * and either trim, delete or split the existing regions.
467 *
468 * Returns the number of huge pages deleted from the reserve map.
469 * In the normal case, the return value is zero or more. In the
470 * case where a region must be split, a new region descriptor must
471 * be allocated. If the allocation fails, -ENOMEM will be returned.
472 * NOTE: If the parameter t == LONG_MAX, then we will never split
473 * a region and possibly return -ENOMEM. Callers specifying
474 * t == LONG_MAX do not need to check for -ENOMEM error.
Mike Kravetz1dd308a2015-06-24 16:57:52 -0700475 */
Mike Kravetzfeba16e2015-09-08 15:01:31 -0700476static long region_del(struct resv_map *resv, long f, long t)
Andy Whitcroft96822902008-07-23 21:27:29 -0700477{
Joonsoo Kim1406ec92014-04-03 14:47:26 -0700478 struct list_head *head = &resv->regions;
Andy Whitcroft96822902008-07-23 21:27:29 -0700479 struct file_region *rg, *trg;
Mike Kravetzfeba16e2015-09-08 15:01:31 -0700480 struct file_region *nrg = NULL;
481 long del = 0;
Andy Whitcroft96822902008-07-23 21:27:29 -0700482
Mike Kravetzfeba16e2015-09-08 15:01:31 -0700483retry:
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700484 spin_lock(&resv->lock);
Mike Kravetzfeba16e2015-09-08 15:01:31 -0700485 list_for_each_entry_safe(rg, trg, head, link) {
486 if (rg->to <= f)
487 continue;
488 if (rg->from >= t)
Andy Whitcroft96822902008-07-23 21:27:29 -0700489 break;
Andy Whitcroft96822902008-07-23 21:27:29 -0700490
Mike Kravetzfeba16e2015-09-08 15:01:31 -0700491 if (f > rg->from && t < rg->to) { /* Must split region */
492 /*
493 * Check for an entry in the cache before dropping
494 * lock and attempting allocation.
495 */
496 if (!nrg &&
497 resv->region_cache_count > resv->adds_in_progress) {
498 nrg = list_first_entry(&resv->region_cache,
499 struct file_region,
500 link);
501 list_del(&nrg->link);
502 resv->region_cache_count--;
503 }
504
505 if (!nrg) {
506 spin_unlock(&resv->lock);
507 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
508 if (!nrg)
509 return -ENOMEM;
510 goto retry;
511 }
512
513 del += t - f;
514
515 /* New entry for end of split region */
516 nrg->from = t;
517 nrg->to = rg->to;
518 INIT_LIST_HEAD(&nrg->link);
519
520 /* Original entry is trimmed */
521 rg->to = f;
522
523 list_add(&nrg->link, &rg->link);
524 nrg = NULL;
525 break;
526 }
527
528 if (f <= rg->from && t >= rg->to) { /* Remove entire region */
529 del += rg->to - rg->from;
530 list_del(&rg->link);
531 kfree(rg);
532 continue;
533 }
534
535 if (f <= rg->from) { /* Trim beginning of region */
536 del += t - rg->from;
537 rg->from = t;
538 } else { /* Trim end of region */
539 del += rg->to - f;
540 rg->to = f;
541 }
Andy Whitcroft96822902008-07-23 21:27:29 -0700542 }
543
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700544 spin_unlock(&resv->lock);
Mike Kravetzfeba16e2015-09-08 15:01:31 -0700545 kfree(nrg);
546 return del;
Andy Whitcroft96822902008-07-23 21:27:29 -0700547}
548
Mike Kravetz1dd308a2015-06-24 16:57:52 -0700549/*
Mike Kravetzb5cec282015-09-08 15:01:41 -0700550 * A rare out of memory error was encountered which prevented removal of
551 * the reserve map region for a page. The huge page itself was free'ed
552 * and removed from the page cache. This routine will adjust the subpool
553 * usage count, and the global reserve count if needed. By incrementing
554 * these counts, the reserve map entry which could not be deleted will
555 * appear as a "reserved" entry instead of simply dangling with incorrect
556 * counts.
557 */
558void hugetlb_fix_reserve_counts(struct inode *inode, bool restore_reserve)
559{
560 struct hugepage_subpool *spool = subpool_inode(inode);
561 long rsv_adjust;
562
563 rsv_adjust = hugepage_subpool_get_pages(spool, 1);
564 if (restore_reserve && rsv_adjust) {
565 struct hstate *h = hstate_inode(inode);
566
567 hugetlb_acct_memory(h, 1);
568 }
569}
570
571/*
Mike Kravetz1dd308a2015-06-24 16:57:52 -0700572 * Count and return the number of huge pages in the reserve map
573 * that intersect with the range [f, t).
574 */
Joonsoo Kim1406ec92014-04-03 14:47:26 -0700575static long region_count(struct resv_map *resv, long f, long t)
Andy Whitcroft84afd992008-07-23 21:27:32 -0700576{
Joonsoo Kim1406ec92014-04-03 14:47:26 -0700577 struct list_head *head = &resv->regions;
Andy Whitcroft84afd992008-07-23 21:27:32 -0700578 struct file_region *rg;
579 long chg = 0;
580
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700581 spin_lock(&resv->lock);
Andy Whitcroft84afd992008-07-23 21:27:32 -0700582 /* Locate each segment we overlap with, and count that overlap. */
583 list_for_each_entry(rg, head, link) {
Wang Sheng-Huif2135a42012-05-29 15:06:17 -0700584 long seg_from;
585 long seg_to;
Andy Whitcroft84afd992008-07-23 21:27:32 -0700586
587 if (rg->to <= f)
588 continue;
589 if (rg->from >= t)
590 break;
591
592 seg_from = max(rg->from, f);
593 seg_to = min(rg->to, t);
594
595 chg += seg_to - seg_from;
596 }
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700597 spin_unlock(&resv->lock);
Andy Whitcroft84afd992008-07-23 21:27:32 -0700598
599 return chg;
600}
601
Andy Whitcroft96822902008-07-23 21:27:29 -0700602/*
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700603 * Convert the address within this vma to the page offset within
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700604 * the mapping, in pagecache page units; huge pages here.
605 */
Andi Kleena5516432008-07-23 21:27:41 -0700606static pgoff_t vma_hugecache_offset(struct hstate *h,
607 struct vm_area_struct *vma, unsigned long address)
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700608{
Andi Kleena5516432008-07-23 21:27:41 -0700609 return ((address - vma->vm_start) >> huge_page_shift(h)) +
610 (vma->vm_pgoff >> huge_page_order(h));
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700611}
612
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +0900613pgoff_t linear_hugepage_index(struct vm_area_struct *vma,
614 unsigned long address)
615{
616 return vma_hugecache_offset(hstate_vma(vma), vma, address);
617}
618
Andy Whitcroft84afd992008-07-23 21:27:32 -0700619/*
Mel Gorman08fba692009-01-06 14:38:53 -0800620 * Return the size of the pages allocated when backing a VMA. In the majority
621 * cases this will be same size as used by the page table entries.
622 */
623unsigned long vma_kernel_pagesize(struct vm_area_struct *vma)
624{
625 struct hstate *hstate;
626
627 if (!is_vm_hugetlb_page(vma))
628 return PAGE_SIZE;
629
630 hstate = hstate_vma(vma);
631
Wanpeng Li2415cf12013-07-03 15:02:43 -0700632 return 1UL << huge_page_shift(hstate);
Mel Gorman08fba692009-01-06 14:38:53 -0800633}
Joerg Roedelf340ca02009-06-19 15:16:22 +0200634EXPORT_SYMBOL_GPL(vma_kernel_pagesize);
Mel Gorman08fba692009-01-06 14:38:53 -0800635
636/*
Mel Gorman33402892009-01-06 14:38:54 -0800637 * Return the page size being used by the MMU to back a VMA. In the majority
638 * of cases, the page size used by the kernel matches the MMU size. On
639 * architectures where it differs, an architecture-specific version of this
640 * function is required.
641 */
642#ifndef vma_mmu_pagesize
643unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
644{
645 return vma_kernel_pagesize(vma);
646}
647#endif
648
649/*
Andy Whitcroft84afd992008-07-23 21:27:32 -0700650 * Flags for MAP_PRIVATE reservations. These are stored in the bottom
651 * bits of the reservation map pointer, which are always clear due to
652 * alignment.
653 */
654#define HPAGE_RESV_OWNER (1UL << 0)
655#define HPAGE_RESV_UNMAPPED (1UL << 1)
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700656#define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
Andy Whitcroft84afd992008-07-23 21:27:32 -0700657
Mel Gormana1e78772008-07-23 21:27:23 -0700658/*
659 * These helpers are used to track how many pages are reserved for
660 * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
661 * is guaranteed to have their future faults succeed.
662 *
663 * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
664 * the reserve counters are updated with the hugetlb_lock held. It is safe
665 * to reset the VMA at fork() time as it is not in use yet and there is no
666 * chance of the global counters getting corrupted as a result of the values.
Andy Whitcroft84afd992008-07-23 21:27:32 -0700667 *
668 * The private mapping reservation is represented in a subtly different
669 * manner to a shared mapping. A shared mapping has a region map associated
670 * with the underlying file, this region map represents the backing file
671 * pages which have ever had a reservation assigned which this persists even
672 * after the page is instantiated. A private mapping has a region map
673 * associated with the original mmap which is attached to all VMAs which
674 * reference it, this region map represents those offsets which have consumed
675 * reservation ie. where pages have been instantiated.
Mel Gormana1e78772008-07-23 21:27:23 -0700676 */
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700677static unsigned long get_vma_private_data(struct vm_area_struct *vma)
678{
679 return (unsigned long)vma->vm_private_data;
680}
681
682static void set_vma_private_data(struct vm_area_struct *vma,
683 unsigned long value)
684{
685 vma->vm_private_data = (void *)value;
686}
687
Joonsoo Kim9119a412014-04-03 14:47:25 -0700688struct resv_map *resv_map_alloc(void)
Andy Whitcroft84afd992008-07-23 21:27:32 -0700689{
690 struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
Mike Kravetz5e911372015-09-08 15:01:28 -0700691 struct file_region *rg = kmalloc(sizeof(*rg), GFP_KERNEL);
692
693 if (!resv_map || !rg) {
694 kfree(resv_map);
695 kfree(rg);
Andy Whitcroft84afd992008-07-23 21:27:32 -0700696 return NULL;
Mike Kravetz5e911372015-09-08 15:01:28 -0700697 }
Andy Whitcroft84afd992008-07-23 21:27:32 -0700698
699 kref_init(&resv_map->refs);
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700700 spin_lock_init(&resv_map->lock);
Andy Whitcroft84afd992008-07-23 21:27:32 -0700701 INIT_LIST_HEAD(&resv_map->regions);
702
Mike Kravetz5e911372015-09-08 15:01:28 -0700703 resv_map->adds_in_progress = 0;
704
705 INIT_LIST_HEAD(&resv_map->region_cache);
706 list_add(&rg->link, &resv_map->region_cache);
707 resv_map->region_cache_count = 1;
708
Andy Whitcroft84afd992008-07-23 21:27:32 -0700709 return resv_map;
710}
711
Joonsoo Kim9119a412014-04-03 14:47:25 -0700712void resv_map_release(struct kref *ref)
Andy Whitcroft84afd992008-07-23 21:27:32 -0700713{
714 struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
Mike Kravetz5e911372015-09-08 15:01:28 -0700715 struct list_head *head = &resv_map->region_cache;
716 struct file_region *rg, *trg;
Andy Whitcroft84afd992008-07-23 21:27:32 -0700717
718 /* Clear out any active regions before we release the map. */
Mike Kravetzfeba16e2015-09-08 15:01:31 -0700719 region_del(resv_map, 0, LONG_MAX);
Mike Kravetz5e911372015-09-08 15:01:28 -0700720
721 /* ... and any entries left in the cache */
722 list_for_each_entry_safe(rg, trg, head, link) {
723 list_del(&rg->link);
724 kfree(rg);
725 }
726
727 VM_BUG_ON(resv_map->adds_in_progress);
728
Andy Whitcroft84afd992008-07-23 21:27:32 -0700729 kfree(resv_map);
730}
731
Joonsoo Kim4e35f482014-04-03 14:47:30 -0700732static inline struct resv_map *inode_resv_map(struct inode *inode)
733{
734 return inode->i_mapping->private_data;
735}
736
Andy Whitcroft84afd992008-07-23 21:27:32 -0700737static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
Mel Gormana1e78772008-07-23 21:27:23 -0700738{
Sasha Levin81d1b092014-10-09 15:28:10 -0700739 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
Joonsoo Kim4e35f482014-04-03 14:47:30 -0700740 if (vma->vm_flags & VM_MAYSHARE) {
741 struct address_space *mapping = vma->vm_file->f_mapping;
742 struct inode *inode = mapping->host;
743
744 return inode_resv_map(inode);
745
746 } else {
Andy Whitcroft84afd992008-07-23 21:27:32 -0700747 return (struct resv_map *)(get_vma_private_data(vma) &
748 ~HPAGE_RESV_MASK);
Joonsoo Kim4e35f482014-04-03 14:47:30 -0700749 }
Mel Gormana1e78772008-07-23 21:27:23 -0700750}
751
Andy Whitcroft84afd992008-07-23 21:27:32 -0700752static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
Mel Gormana1e78772008-07-23 21:27:23 -0700753{
Sasha Levin81d1b092014-10-09 15:28:10 -0700754 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
755 VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
Mel Gormana1e78772008-07-23 21:27:23 -0700756
Andy Whitcroft84afd992008-07-23 21:27:32 -0700757 set_vma_private_data(vma, (get_vma_private_data(vma) &
758 HPAGE_RESV_MASK) | (unsigned long)map);
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700759}
760
761static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
762{
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);
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700765
766 set_vma_private_data(vma, get_vma_private_data(vma) | flags);
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700767}
768
769static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
770{
Sasha Levin81d1b092014-10-09 15:28:10 -0700771 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700772
773 return (get_vma_private_data(vma) & flag) != 0;
Mel Gormana1e78772008-07-23 21:27:23 -0700774}
775
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700776/* Reset counters to 0 and clear all HPAGE_RESV_* flags */
Mel Gormana1e78772008-07-23 21:27:23 -0700777void reset_vma_resv_huge_pages(struct vm_area_struct *vma)
778{
Sasha Levin81d1b092014-10-09 15:28:10 -0700779 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
Mel Gormanf83a2752009-05-28 14:34:40 -0700780 if (!(vma->vm_flags & VM_MAYSHARE))
Mel Gormana1e78772008-07-23 21:27:23 -0700781 vma->vm_private_data = (void *)0;
782}
783
784/* Returns true if the VMA has associated reserve pages */
Nicholas Krause559ec2f2015-09-04 15:48:27 -0700785static bool vma_has_reserves(struct vm_area_struct *vma, long chg)
Mel Gormana1e78772008-07-23 21:27:23 -0700786{
Joonsoo Kimaf0ed732013-09-11 14:21:18 -0700787 if (vma->vm_flags & VM_NORESERVE) {
788 /*
789 * This address is already reserved by other process(chg == 0),
790 * so, we should decrement reserved count. Without decrementing,
791 * reserve count remains after releasing inode, because this
792 * allocated page will go into page cache and is regarded as
793 * coming from reserved pool in releasing step. Currently, we
794 * don't have any other solution to deal with this situation
795 * properly, so add work-around here.
796 */
797 if (vma->vm_flags & VM_MAYSHARE && chg == 0)
Nicholas Krause559ec2f2015-09-04 15:48:27 -0700798 return true;
Joonsoo Kimaf0ed732013-09-11 14:21:18 -0700799 else
Nicholas Krause559ec2f2015-09-04 15:48:27 -0700800 return false;
Joonsoo Kimaf0ed732013-09-11 14:21:18 -0700801 }
Joonsoo Kima63884e2013-09-11 14:21:07 -0700802
803 /* Shared mappings always use reserves */
Mike Kravetz1fb1b0e2015-09-08 15:01:44 -0700804 if (vma->vm_flags & VM_MAYSHARE) {
805 /*
806 * We know VM_NORESERVE is not set. Therefore, there SHOULD
807 * be a region map for all pages. The only situation where
808 * there is no region map is if a hole was punched via
809 * fallocate. In this case, there really are no reverves to
810 * use. This situation is indicated if chg != 0.
811 */
812 if (chg)
813 return false;
814 else
815 return true;
816 }
Joonsoo Kima63884e2013-09-11 14:21:07 -0700817
818 /*
819 * Only the process that called mmap() has reserves for
820 * private mappings.
821 */
Mel Gorman7f09ca52008-07-23 21:27:58 -0700822 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER))
Nicholas Krause559ec2f2015-09-04 15:48:27 -0700823 return true;
Joonsoo Kima63884e2013-09-11 14:21:07 -0700824
Nicholas Krause559ec2f2015-09-04 15:48:27 -0700825 return false;
Mel Gormana1e78772008-07-23 21:27:23 -0700826}
827
Andi Kleena5516432008-07-23 21:27:41 -0700828static void enqueue_huge_page(struct hstate *h, struct page *page)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700829{
830 int nid = page_to_nid(page);
Aneesh Kumar K.V0edaecf2012-07-31 16:42:07 -0700831 list_move(&page->lru, &h->hugepage_freelists[nid]);
Andi Kleena5516432008-07-23 21:27:41 -0700832 h->free_huge_pages++;
833 h->free_huge_pages_node[nid]++;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700834}
835
Naoya Horiguchibf50bab2010-09-08 10:19:33 +0900836static struct page *dequeue_huge_page_node(struct hstate *h, int nid)
837{
838 struct page *page;
839
Naoya Horiguchic8721bb2013-09-11 14:22:09 -0700840 list_for_each_entry(page, &h->hugepage_freelists[nid], lru)
841 if (!is_migrate_isolate_page(page))
842 break;
843 /*
844 * if 'non-isolated free hugepage' not found on the list,
845 * the allocation fails.
846 */
847 if (&h->hugepage_freelists[nid] == &page->lru)
Naoya Horiguchibf50bab2010-09-08 10:19:33 +0900848 return NULL;
Aneesh Kumar K.V0edaecf2012-07-31 16:42:07 -0700849 list_move(&page->lru, &h->hugepage_activelist);
Naoya Horiguchia9869b82010-09-08 10:19:37 +0900850 set_page_refcounted(page);
Naoya Horiguchibf50bab2010-09-08 10:19:33 +0900851 h->free_huge_pages--;
852 h->free_huge_pages_node[nid]--;
853 return page;
854}
855
Naoya Horiguchi86cdb462013-09-11 14:22:13 -0700856/* Movability of hugepages depends on migration support. */
857static inline gfp_t htlb_alloc_mask(struct hstate *h)
858{
Naoya Horiguchi100873d2014-06-04 16:10:56 -0700859 if (hugepages_treat_as_movable || hugepage_migration_supported(h))
Naoya Horiguchi86cdb462013-09-11 14:22:13 -0700860 return GFP_HIGHUSER_MOVABLE;
861 else
862 return GFP_HIGHUSER;
863}
864
Andi Kleena5516432008-07-23 21:27:41 -0700865static struct page *dequeue_huge_page_vma(struct hstate *h,
866 struct vm_area_struct *vma,
Joonsoo Kimaf0ed732013-09-11 14:21:18 -0700867 unsigned long address, int avoid_reserve,
868 long chg)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700869{
Konstantin Khlebnikovb1c12cb2012-04-25 16:01:46 -0700870 struct page *page = NULL;
Lee Schermerhorn480eccf2007-09-18 22:46:47 -0700871 struct mempolicy *mpol;
Mel Gorman19770b32008-04-28 02:12:18 -0700872 nodemask_t *nodemask;
Miao Xiec0ff7452010-05-24 14:32:08 -0700873 struct zonelist *zonelist;
Mel Gormandd1a2392008-04-28 02:12:17 -0700874 struct zone *zone;
875 struct zoneref *z;
Mel Gormancc9a6c82012-03-21 16:34:11 -0700876 unsigned int cpuset_mems_cookie;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700877
Mel Gormana1e78772008-07-23 21:27:23 -0700878 /*
879 * A child process with MAP_PRIVATE mappings created by their parent
880 * have no page reserves. This check ensures that reservations are
881 * not "stolen". The child may still get SIGKILLed
882 */
Joonsoo Kimaf0ed732013-09-11 14:21:18 -0700883 if (!vma_has_reserves(vma, chg) &&
Andi Kleena5516432008-07-23 21:27:41 -0700884 h->free_huge_pages - h->resv_huge_pages == 0)
Miao Xiec0ff7452010-05-24 14:32:08 -0700885 goto err;
Mel Gormana1e78772008-07-23 21:27:23 -0700886
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700887 /* If reserves cannot be used, ensure enough pages are in the pool */
Andi Kleena5516432008-07-23 21:27:41 -0700888 if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
Justin P. Mattock6eab04a2011-04-08 19:49:08 -0700889 goto err;
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700890
Joonsoo Kim9966c4b2013-09-11 14:20:50 -0700891retry_cpuset:
Mel Gormand26914d2014-04-03 14:47:24 -0700892 cpuset_mems_cookie = read_mems_allowed_begin();
Joonsoo Kim9966c4b2013-09-11 14:20:50 -0700893 zonelist = huge_zonelist(vma, address,
Naoya Horiguchi86cdb462013-09-11 14:22:13 -0700894 htlb_alloc_mask(h), &mpol, &nodemask);
Joonsoo Kim9966c4b2013-09-11 14:20:50 -0700895
Mel Gorman19770b32008-04-28 02:12:18 -0700896 for_each_zone_zonelist_nodemask(zone, z, zonelist,
897 MAX_NR_ZONES - 1, nodemask) {
Vladimir Davydov344736f2014-10-20 15:50:30 +0400898 if (cpuset_zone_allowed(zone, htlb_alloc_mask(h))) {
Naoya Horiguchibf50bab2010-09-08 10:19:33 +0900899 page = dequeue_huge_page_node(h, zone_to_nid(zone));
900 if (page) {
Joonsoo Kimaf0ed732013-09-11 14:21:18 -0700901 if (avoid_reserve)
902 break;
903 if (!vma_has_reserves(vma, chg))
904 break;
905
Joonsoo Kim07443a82013-09-11 14:21:58 -0700906 SetPagePrivate(page);
Joonsoo Kimaf0ed732013-09-11 14:21:18 -0700907 h->resv_huge_pages--;
Naoya Horiguchibf50bab2010-09-08 10:19:33 +0900908 break;
909 }
Andrew Morton3abf7af2007-07-19 01:49:08 -0700910 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700911 }
Mel Gormancc9a6c82012-03-21 16:34:11 -0700912
913 mpol_cond_put(mpol);
Mel Gormand26914d2014-04-03 14:47:24 -0700914 if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie)))
Mel Gormancc9a6c82012-03-21 16:34:11 -0700915 goto retry_cpuset;
916 return page;
917
Miao Xiec0ff7452010-05-24 14:32:08 -0700918err:
Mel Gormancc9a6c82012-03-21 16:34:11 -0700919 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700920}
921
Luiz Capitulino1cac6f22014-06-04 16:07:11 -0700922/*
923 * common helper functions for hstate_next_node_to_{alloc|free}.
924 * We may have allocated or freed a huge page based on a different
925 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
926 * be outside of *nodes_allowed. Ensure that we use an allowed
927 * node for alloc or free.
928 */
929static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
930{
931 nid = next_node(nid, *nodes_allowed);
932 if (nid == MAX_NUMNODES)
933 nid = first_node(*nodes_allowed);
934 VM_BUG_ON(nid >= MAX_NUMNODES);
935
936 return nid;
937}
938
939static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
940{
941 if (!node_isset(nid, *nodes_allowed))
942 nid = next_node_allowed(nid, nodes_allowed);
943 return nid;
944}
945
946/*
947 * returns the previously saved node ["this node"] from which to
948 * allocate a persistent huge page for the pool and advance the
949 * next node from which to allocate, handling wrap at end of node
950 * mask.
951 */
952static int hstate_next_node_to_alloc(struct hstate *h,
953 nodemask_t *nodes_allowed)
954{
955 int nid;
956
957 VM_BUG_ON(!nodes_allowed);
958
959 nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
960 h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
961
962 return nid;
963}
964
965/*
966 * helper for free_pool_huge_page() - return the previously saved
967 * node ["this node"] from which to free a huge page. Advance the
968 * next node id whether or not we find a free huge page to free so
969 * that the next attempt to free addresses the next node.
970 */
971static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
972{
973 int nid;
974
975 VM_BUG_ON(!nodes_allowed);
976
977 nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
978 h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
979
980 return nid;
981}
982
983#define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \
984 for (nr_nodes = nodes_weight(*mask); \
985 nr_nodes > 0 && \
986 ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \
987 nr_nodes--)
988
989#define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \
990 for (nr_nodes = nodes_weight(*mask); \
991 nr_nodes > 0 && \
992 ((node = hstate_next_node_to_free(hs, mask)) || 1); \
993 nr_nodes--)
994
Luiz Capitulino944d9fe2014-06-04 16:07:13 -0700995#if defined(CONFIG_CMA) && defined(CONFIG_X86_64)
996static void destroy_compound_gigantic_page(struct page *page,
Kirill A. Shutemovd00181b2015-11-06 16:29:57 -0800997 unsigned int order)
Luiz Capitulino944d9fe2014-06-04 16:07:13 -0700998{
999 int i;
1000 int nr_pages = 1 << order;
1001 struct page *p = page + 1;
1002
1003 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
Kirill A. Shutemov1d798ca2015-11-06 16:29:54 -08001004 clear_compound_head(p);
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001005 set_page_refcounted(p);
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001006 }
1007
1008 set_compound_order(page, 0);
1009 __ClearPageHead(page);
1010}
1011
Kirill A. Shutemovd00181b2015-11-06 16:29:57 -08001012static void free_gigantic_page(struct page *page, unsigned int order)
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001013{
1014 free_contig_range(page_to_pfn(page), 1 << order);
1015}
1016
1017static int __alloc_gigantic_page(unsigned long start_pfn,
1018 unsigned long nr_pages)
1019{
1020 unsigned long end_pfn = start_pfn + nr_pages;
1021 return alloc_contig_range(start_pfn, end_pfn, MIGRATE_MOVABLE);
1022}
1023
1024static bool pfn_range_valid_gigantic(unsigned long start_pfn,
1025 unsigned long nr_pages)
1026{
1027 unsigned long i, end_pfn = start_pfn + nr_pages;
1028 struct page *page;
1029
1030 for (i = start_pfn; i < end_pfn; i++) {
1031 if (!pfn_valid(i))
1032 return false;
1033
1034 page = pfn_to_page(i);
1035
1036 if (PageReserved(page))
1037 return false;
1038
1039 if (page_count(page) > 0)
1040 return false;
1041
1042 if (PageHuge(page))
1043 return false;
1044 }
1045
1046 return true;
1047}
1048
1049static bool zone_spans_last_pfn(const struct zone *zone,
1050 unsigned long start_pfn, unsigned long nr_pages)
1051{
1052 unsigned long last_pfn = start_pfn + nr_pages - 1;
1053 return zone_spans_pfn(zone, last_pfn);
1054}
1055
Kirill A. Shutemovd00181b2015-11-06 16:29:57 -08001056static struct page *alloc_gigantic_page(int nid, unsigned int order)
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001057{
1058 unsigned long nr_pages = 1 << order;
1059 unsigned long ret, pfn, flags;
1060 struct zone *z;
1061
1062 z = NODE_DATA(nid)->node_zones;
1063 for (; z - NODE_DATA(nid)->node_zones < MAX_NR_ZONES; z++) {
1064 spin_lock_irqsave(&z->lock, flags);
1065
1066 pfn = ALIGN(z->zone_start_pfn, nr_pages);
1067 while (zone_spans_last_pfn(z, pfn, nr_pages)) {
1068 if (pfn_range_valid_gigantic(pfn, nr_pages)) {
1069 /*
1070 * We release the zone lock here because
1071 * alloc_contig_range() will also lock the zone
1072 * at some point. If there's an allocation
1073 * spinning on this lock, it may win the race
1074 * and cause alloc_contig_range() to fail...
1075 */
1076 spin_unlock_irqrestore(&z->lock, flags);
1077 ret = __alloc_gigantic_page(pfn, nr_pages);
1078 if (!ret)
1079 return pfn_to_page(pfn);
1080 spin_lock_irqsave(&z->lock, flags);
1081 }
1082 pfn += nr_pages;
1083 }
1084
1085 spin_unlock_irqrestore(&z->lock, flags);
1086 }
1087
1088 return NULL;
1089}
1090
1091static void prep_new_huge_page(struct hstate *h, struct page *page, int nid);
Kirill A. Shutemovd00181b2015-11-06 16:29:57 -08001092static void prep_compound_gigantic_page(struct page *page, unsigned int order);
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001093
1094static struct page *alloc_fresh_gigantic_page_node(struct hstate *h, int nid)
1095{
1096 struct page *page;
1097
1098 page = alloc_gigantic_page(nid, huge_page_order(h));
1099 if (page) {
1100 prep_compound_gigantic_page(page, huge_page_order(h));
1101 prep_new_huge_page(h, page, nid);
1102 }
1103
1104 return page;
1105}
1106
1107static int alloc_fresh_gigantic_page(struct hstate *h,
1108 nodemask_t *nodes_allowed)
1109{
1110 struct page *page = NULL;
1111 int nr_nodes, node;
1112
1113 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
1114 page = alloc_fresh_gigantic_page_node(h, node);
1115 if (page)
1116 return 1;
1117 }
1118
1119 return 0;
1120}
1121
1122static inline bool gigantic_page_supported(void) { return true; }
1123#else
1124static inline bool gigantic_page_supported(void) { return false; }
Kirill A. Shutemovd00181b2015-11-06 16:29:57 -08001125static inline void free_gigantic_page(struct page *page, unsigned int order) { }
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001126static inline void destroy_compound_gigantic_page(struct page *page,
Kirill A. Shutemovd00181b2015-11-06 16:29:57 -08001127 unsigned int order) { }
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001128static inline int alloc_fresh_gigantic_page(struct hstate *h,
1129 nodemask_t *nodes_allowed) { return 0; }
1130#endif
1131
Andi Kleena5516432008-07-23 21:27:41 -07001132static void update_and_free_page(struct hstate *h, struct page *page)
Adam Litke6af2acb2007-10-16 01:26:16 -07001133{
1134 int i;
Andi Kleena5516432008-07-23 21:27:41 -07001135
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001136 if (hstate_is_gigantic(h) && !gigantic_page_supported())
1137 return;
Andy Whitcroft18229df2008-11-06 12:53:27 -08001138
Andi Kleena5516432008-07-23 21:27:41 -07001139 h->nr_huge_pages--;
1140 h->nr_huge_pages_node[page_to_nid(page)]--;
1141 for (i = 0; i < pages_per_huge_page(h); i++) {
Chris Forbes32f84522011-07-25 17:12:14 -07001142 page[i].flags &= ~(1 << PG_locked | 1 << PG_error |
1143 1 << PG_referenced | 1 << PG_dirty |
Luiz Capitulinoa7407a22014-06-04 16:07:09 -07001144 1 << PG_active | 1 << PG_private |
1145 1 << PG_writeback);
Adam Litke6af2acb2007-10-16 01:26:16 -07001146 }
Sasha Levin309381fea2014-01-23 15:52:54 -08001147 VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page);
Kirill A. Shutemovf1e61552015-11-06 16:29:50 -08001148 set_compound_page_dtor(page, NULL_COMPOUND_DTOR);
Adam Litke6af2acb2007-10-16 01:26:16 -07001149 set_page_refcounted(page);
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001150 if (hstate_is_gigantic(h)) {
1151 destroy_compound_gigantic_page(page, huge_page_order(h));
1152 free_gigantic_page(page, huge_page_order(h));
1153 } else {
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001154 __free_pages(page, huge_page_order(h));
1155 }
Adam Litke6af2acb2007-10-16 01:26:16 -07001156}
1157
Andi Kleene5ff2152008-07-23 21:27:42 -07001158struct hstate *size_to_hstate(unsigned long size)
1159{
1160 struct hstate *h;
1161
1162 for_each_hstate(h) {
1163 if (huge_page_size(h) == size)
1164 return h;
1165 }
1166 return NULL;
1167}
1168
Naoya Horiguchibcc54222015-04-15 16:14:38 -07001169/*
1170 * Test to determine whether the hugepage is "active/in-use" (i.e. being linked
1171 * to hstate->hugepage_activelist.)
1172 *
1173 * This function can be called for tail pages, but never returns true for them.
1174 */
1175bool page_huge_active(struct page *page)
1176{
1177 VM_BUG_ON_PAGE(!PageHuge(page), page);
1178 return PageHead(page) && PagePrivate(&page[1]);
1179}
1180
1181/* never called for tail page */
1182static void set_page_huge_active(struct page *page)
1183{
1184 VM_BUG_ON_PAGE(!PageHeadHuge(page), page);
1185 SetPagePrivate(&page[1]);
1186}
1187
1188static void clear_page_huge_active(struct page *page)
1189{
1190 VM_BUG_ON_PAGE(!PageHeadHuge(page), page);
1191 ClearPagePrivate(&page[1]);
1192}
1193
Atsushi Kumagai8f1d26d2014-07-30 16:08:39 -07001194void free_huge_page(struct page *page)
David Gibson27a85ef2006-03-22 00:08:56 -08001195{
Andi Kleena5516432008-07-23 21:27:41 -07001196 /*
1197 * Can't pass hstate in here because it is called from the
1198 * compound page destructor.
1199 */
Andi Kleene5ff2152008-07-23 21:27:42 -07001200 struct hstate *h = page_hstate(page);
Adam Litke7893d1d2007-10-16 01:26:18 -07001201 int nid = page_to_nid(page);
David Gibson90481622012-03-21 16:34:12 -07001202 struct hugepage_subpool *spool =
1203 (struct hugepage_subpool *)page_private(page);
Joonsoo Kim07443a82013-09-11 14:21:58 -07001204 bool restore_reserve;
David Gibson27a85ef2006-03-22 00:08:56 -08001205
Andy Whitcrofte5df70a2008-02-23 15:23:32 -08001206 set_page_private(page, 0);
Mel Gorman23be7462010-04-23 13:17:56 -04001207 page->mapping = NULL;
Adam Litke7893d1d2007-10-16 01:26:18 -07001208 BUG_ON(page_count(page));
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09001209 BUG_ON(page_mapcount(page));
Joonsoo Kim07443a82013-09-11 14:21:58 -07001210 restore_reserve = PagePrivate(page);
Joonsoo Kim16c794b2013-10-16 13:46:48 -07001211 ClearPagePrivate(page);
David Gibson27a85ef2006-03-22 00:08:56 -08001212
Mike Kravetz1c5ecae2015-04-15 16:13:39 -07001213 /*
1214 * A return code of zero implies that the subpool will be under its
1215 * minimum size if the reservation is not restored after page is free.
1216 * Therefore, force restore_reserve operation.
1217 */
1218 if (hugepage_subpool_put_pages(spool, 1) == 0)
1219 restore_reserve = true;
1220
David Gibson27a85ef2006-03-22 00:08:56 -08001221 spin_lock(&hugetlb_lock);
Naoya Horiguchibcc54222015-04-15 16:14:38 -07001222 clear_page_huge_active(page);
Aneesh Kumar K.V6d76dcf2012-07-31 16:42:18 -07001223 hugetlb_cgroup_uncharge_page(hstate_index(h),
1224 pages_per_huge_page(h), page);
Joonsoo Kim07443a82013-09-11 14:21:58 -07001225 if (restore_reserve)
1226 h->resv_huge_pages++;
1227
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001228 if (h->surplus_huge_pages_node[nid]) {
Aneesh Kumar K.V0edaecf2012-07-31 16:42:07 -07001229 /* remove the page from active list */
1230 list_del(&page->lru);
Andi Kleena5516432008-07-23 21:27:41 -07001231 update_and_free_page(h, page);
1232 h->surplus_huge_pages--;
1233 h->surplus_huge_pages_node[nid]--;
Adam Litke7893d1d2007-10-16 01:26:18 -07001234 } else {
Will Deacon5d3a5512012-10-08 16:29:32 -07001235 arch_clear_hugepage_flags(page);
Andi Kleena5516432008-07-23 21:27:41 -07001236 enqueue_huge_page(h, page);
Adam Litke7893d1d2007-10-16 01:26:18 -07001237 }
David Gibson27a85ef2006-03-22 00:08:56 -08001238 spin_unlock(&hugetlb_lock);
1239}
1240
Andi Kleena5516432008-07-23 21:27:41 -07001241static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
Andi Kleenb7ba30c2008-07-23 21:27:40 -07001242{
Aneesh Kumar K.V0edaecf2012-07-31 16:42:07 -07001243 INIT_LIST_HEAD(&page->lru);
Kirill A. Shutemovf1e61552015-11-06 16:29:50 -08001244 set_compound_page_dtor(page, HUGETLB_PAGE_DTOR);
Andi Kleenb7ba30c2008-07-23 21:27:40 -07001245 spin_lock(&hugetlb_lock);
Aneesh Kumar K.V9dd540e2012-07-31 16:42:15 -07001246 set_hugetlb_cgroup(page, NULL);
Andi Kleena5516432008-07-23 21:27:41 -07001247 h->nr_huge_pages++;
1248 h->nr_huge_pages_node[nid]++;
Andi Kleenb7ba30c2008-07-23 21:27:40 -07001249 spin_unlock(&hugetlb_lock);
1250 put_page(page); /* free it into the hugepage allocator */
1251}
1252
Kirill A. Shutemovd00181b2015-11-06 16:29:57 -08001253static void prep_compound_gigantic_page(struct page *page, unsigned int order)
Wu Fengguang20a03072009-06-16 15:32:22 -07001254{
1255 int i;
1256 int nr_pages = 1 << order;
1257 struct page *p = page + 1;
1258
1259 /* we rely on prep_new_huge_page to set the destructor */
1260 set_compound_order(page, order);
1261 __SetPageHead(page);
Andrea Arcangelief5a22b2013-10-16 13:46:56 -07001262 __ClearPageReserved(page);
Wu Fengguang20a03072009-06-16 15:32:22 -07001263 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
Andrea Arcangelief5a22b2013-10-16 13:46:56 -07001264 /*
1265 * For gigantic hugepages allocated through bootmem at
1266 * boot, it's safer to be consistent with the not-gigantic
1267 * hugepages and clear the PG_reserved bit from all tail pages
1268 * too. Otherwse drivers using get_user_pages() to access tail
1269 * pages may get the reference counting wrong if they see
1270 * PG_reserved set on a tail page (despite the head page not
1271 * having PG_reserved set). Enforcing this consistency between
1272 * head and tail pages allows drivers to optimize away a check
1273 * on the head page when they need know if put_page() is needed
1274 * after get_user_pages().
1275 */
1276 __ClearPageReserved(p);
Youquan Song58a84aa2011-12-08 14:34:18 -08001277 set_page_count(p, 0);
Kirill A. Shutemov1d798ca2015-11-06 16:29:54 -08001278 set_compound_head(p, page);
Wu Fengguang20a03072009-06-16 15:32:22 -07001279 }
1280}
1281
Andrew Morton77959122012-10-08 16:34:11 -07001282/*
1283 * PageHuge() only returns true for hugetlbfs pages, but not for normal or
1284 * transparent huge pages. See the PageTransHuge() documentation for more
1285 * details.
1286 */
Wu Fengguang20a03072009-06-16 15:32:22 -07001287int PageHuge(struct page *page)
1288{
Wu Fengguang20a03072009-06-16 15:32:22 -07001289 if (!PageCompound(page))
1290 return 0;
1291
1292 page = compound_head(page);
Kirill A. Shutemovf1e61552015-11-06 16:29:50 -08001293 return page[1].compound_dtor == HUGETLB_PAGE_DTOR;
Wu Fengguang20a03072009-06-16 15:32:22 -07001294}
Naoya Horiguchi43131e12010-05-28 09:29:22 +09001295EXPORT_SYMBOL_GPL(PageHuge);
1296
Andrea Arcangeli27c73ae2013-11-21 14:32:02 -08001297/*
1298 * PageHeadHuge() only returns true for hugetlbfs head page, but not for
1299 * normal or transparent huge pages.
1300 */
1301int PageHeadHuge(struct page *page_head)
1302{
Andrea Arcangeli27c73ae2013-11-21 14:32:02 -08001303 if (!PageHead(page_head))
1304 return 0;
1305
Andrew Morton758f66a2014-01-21 15:48:57 -08001306 return get_compound_page_dtor(page_head) == free_huge_page;
Andrea Arcangeli27c73ae2013-11-21 14:32:02 -08001307}
Andrea Arcangeli27c73ae2013-11-21 14:32:02 -08001308
Zhang Yi13d60f42013-06-25 21:19:31 +08001309pgoff_t __basepage_index(struct page *page)
1310{
1311 struct page *page_head = compound_head(page);
1312 pgoff_t index = page_index(page_head);
1313 unsigned long compound_idx;
1314
1315 if (!PageHuge(page_head))
1316 return page_index(page);
1317
1318 if (compound_order(page_head) >= MAX_ORDER)
1319 compound_idx = page_to_pfn(page) - page_to_pfn(page_head);
1320 else
1321 compound_idx = page - page_head;
1322
1323 return (index << compound_order(page_head)) + compound_idx;
1324}
1325
Andi Kleena5516432008-07-23 21:27:41 -07001326static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001327{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001328 struct page *page;
Joe Jinf96efd52007-07-15 23:38:12 -07001329
Vlastimil Babka96db8002015-09-08 15:03:50 -07001330 page = __alloc_pages_node(nid,
Naoya Horiguchi86cdb462013-09-11 14:22:13 -07001331 htlb_alloc_mask(h)|__GFP_COMP|__GFP_THISNODE|
Nishanth Aravamudan551883a2008-04-29 00:58:26 -07001332 __GFP_REPEAT|__GFP_NOWARN,
Andi Kleena5516432008-07-23 21:27:41 -07001333 huge_page_order(h));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001334 if (page) {
Andi Kleena5516432008-07-23 21:27:41 -07001335 prep_new_huge_page(h, page, nid);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001336 }
Nishanth Aravamudan63b46132007-10-16 01:26:24 -07001337
1338 return page;
1339}
1340
Joonsoo Kimb2261022013-09-11 14:21:00 -07001341static int alloc_fresh_huge_page(struct hstate *h, nodemask_t *nodes_allowed)
1342{
1343 struct page *page;
1344 int nr_nodes, node;
1345 int ret = 0;
1346
1347 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
1348 page = alloc_fresh_huge_page_node(h, node);
1349 if (page) {
1350 ret = 1;
1351 break;
1352 }
1353 }
1354
1355 if (ret)
1356 count_vm_event(HTLB_BUDDY_PGALLOC);
1357 else
1358 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
1359
1360 return ret;
1361}
1362
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001363/*
1364 * Free huge page from pool from next node to free.
1365 * Attempt to keep persistent huge pages more or less
1366 * balanced over allowed nodes.
1367 * Called with hugetlb_lock locked.
1368 */
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001369static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
1370 bool acct_surplus)
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001371{
Joonsoo Kimb2261022013-09-11 14:21:00 -07001372 int nr_nodes, node;
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001373 int ret = 0;
1374
Joonsoo Kimb2261022013-09-11 14:21:00 -07001375 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
Lee Schermerhorn685f3452009-09-21 17:01:23 -07001376 /*
1377 * If we're returning unused surplus pages, only examine
1378 * nodes with surplus pages.
1379 */
Joonsoo Kimb2261022013-09-11 14:21:00 -07001380 if ((!acct_surplus || h->surplus_huge_pages_node[node]) &&
1381 !list_empty(&h->hugepage_freelists[node])) {
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001382 struct page *page =
Joonsoo Kimb2261022013-09-11 14:21:00 -07001383 list_entry(h->hugepage_freelists[node].next,
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001384 struct page, lru);
1385 list_del(&page->lru);
1386 h->free_huge_pages--;
Joonsoo Kimb2261022013-09-11 14:21:00 -07001387 h->free_huge_pages_node[node]--;
Lee Schermerhorn685f3452009-09-21 17:01:23 -07001388 if (acct_surplus) {
1389 h->surplus_huge_pages--;
Joonsoo Kimb2261022013-09-11 14:21:00 -07001390 h->surplus_huge_pages_node[node]--;
Lee Schermerhorn685f3452009-09-21 17:01:23 -07001391 }
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001392 update_and_free_page(h, page);
1393 ret = 1;
Lee Schermerhorn9a76db02009-12-14 17:58:15 -08001394 break;
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001395 }
Joonsoo Kimb2261022013-09-11 14:21:00 -07001396 }
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001397
1398 return ret;
1399}
1400
Naoya Horiguchic8721bb2013-09-11 14:22:09 -07001401/*
1402 * Dissolve a given free hugepage into free buddy pages. This function does
1403 * nothing for in-use (including surplus) hugepages.
1404 */
1405static void dissolve_free_huge_page(struct page *page)
1406{
1407 spin_lock(&hugetlb_lock);
1408 if (PageHuge(page) && !page_count(page)) {
1409 struct hstate *h = page_hstate(page);
1410 int nid = page_to_nid(page);
1411 list_del(&page->lru);
1412 h->free_huge_pages--;
1413 h->free_huge_pages_node[nid]--;
1414 update_and_free_page(h, page);
1415 }
1416 spin_unlock(&hugetlb_lock);
1417}
1418
1419/*
1420 * Dissolve free hugepages in a given pfn range. Used by memory hotplug to
1421 * make specified memory blocks removable from the system.
1422 * Note that start_pfn should aligned with (minimum) hugepage size.
1423 */
1424void dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn)
1425{
Naoya Horiguchic8721bb2013-09-11 14:22:09 -07001426 unsigned long pfn;
Naoya Horiguchic8721bb2013-09-11 14:22:09 -07001427
Li Zhongd0177632014-08-06 16:07:56 -07001428 if (!hugepages_supported())
1429 return;
1430
Naoya Horiguchi641844f2015-06-24 16:56:59 -07001431 VM_BUG_ON(!IS_ALIGNED(start_pfn, 1 << minimum_order));
1432 for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << minimum_order)
Naoya Horiguchic8721bb2013-09-11 14:22:09 -07001433 dissolve_free_huge_page(pfn_to_page(pfn));
1434}
1435
Dave Hansen099730d2015-11-05 18:50:17 -08001436/*
1437 * There are 3 ways this can get called:
1438 * 1. With vma+addr: we use the VMA's memory policy
1439 * 2. With !vma, but nid=NUMA_NO_NODE: We try to allocate a huge
1440 * page from any node, and let the buddy allocator itself figure
1441 * it out.
1442 * 3. With !vma, but nid!=NUMA_NO_NODE. We allocate a huge page
1443 * strictly from 'nid'
1444 */
1445static struct page *__hugetlb_alloc_buddy_huge_page(struct hstate *h,
1446 struct vm_area_struct *vma, unsigned long addr, int nid)
1447{
1448 int order = huge_page_order(h);
1449 gfp_t gfp = htlb_alloc_mask(h)|__GFP_COMP|__GFP_REPEAT|__GFP_NOWARN;
1450 unsigned int cpuset_mems_cookie;
1451
1452 /*
1453 * We need a VMA to get a memory policy. If we do not
Dave Hansene0ec90e2015-11-05 18:50:20 -08001454 * have one, we use the 'nid' argument.
1455 *
1456 * The mempolicy stuff below has some non-inlined bits
1457 * and calls ->vm_ops. That makes it hard to optimize at
1458 * compile-time, even when NUMA is off and it does
1459 * nothing. This helps the compiler optimize it out.
Dave Hansen099730d2015-11-05 18:50:17 -08001460 */
Dave Hansene0ec90e2015-11-05 18:50:20 -08001461 if (!IS_ENABLED(CONFIG_NUMA) || !vma) {
Dave Hansen099730d2015-11-05 18:50:17 -08001462 /*
1463 * If a specific node is requested, make sure to
1464 * get memory from there, but only when a node
1465 * is explicitly specified.
1466 */
1467 if (nid != NUMA_NO_NODE)
1468 gfp |= __GFP_THISNODE;
1469 /*
1470 * Make sure to call something that can handle
1471 * nid=NUMA_NO_NODE
1472 */
1473 return alloc_pages_node(nid, gfp, order);
1474 }
1475
1476 /*
1477 * OK, so we have a VMA. Fetch the mempolicy and try to
Dave Hansene0ec90e2015-11-05 18:50:20 -08001478 * allocate a huge page with it. We will only reach this
1479 * when CONFIG_NUMA=y.
Dave Hansen099730d2015-11-05 18:50:17 -08001480 */
1481 do {
1482 struct page *page;
1483 struct mempolicy *mpol;
1484 struct zonelist *zl;
1485 nodemask_t *nodemask;
1486
1487 cpuset_mems_cookie = read_mems_allowed_begin();
1488 zl = huge_zonelist(vma, addr, gfp, &mpol, &nodemask);
1489 mpol_cond_put(mpol);
1490 page = __alloc_pages_nodemask(gfp, order, zl, nodemask);
1491 if (page)
1492 return page;
1493 } while (read_mems_allowed_retry(cpuset_mems_cookie));
1494
1495 return NULL;
1496}
1497
1498/*
1499 * There are two ways to allocate a huge page:
1500 * 1. When you have a VMA and an address (like a fault)
1501 * 2. When you have no VMA (like when setting /proc/.../nr_hugepages)
1502 *
1503 * 'vma' and 'addr' are only for (1). 'nid' is always NUMA_NO_NODE in
1504 * this case which signifies that the allocation should be done with
1505 * respect for the VMA's memory policy.
1506 *
1507 * For (2), we ignore 'vma' and 'addr' and use 'nid' exclusively. This
1508 * implies that memory policies will not be taken in to account.
1509 */
1510static struct page *__alloc_buddy_huge_page(struct hstate *h,
1511 struct vm_area_struct *vma, unsigned long addr, int nid)
Adam Litke7893d1d2007-10-16 01:26:18 -07001512{
1513 struct page *page;
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001514 unsigned int r_nid;
Adam Litke7893d1d2007-10-16 01:26:18 -07001515
Luiz Capitulinobae7f4a2014-06-04 16:07:08 -07001516 if (hstate_is_gigantic(h))
Andi Kleenaa888a72008-07-23 21:27:47 -07001517 return NULL;
1518
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001519 /*
Dave Hansen099730d2015-11-05 18:50:17 -08001520 * Make sure that anyone specifying 'nid' is not also specifying a VMA.
1521 * This makes sure the caller is picking _one_ of the modes with which
1522 * we can call this function, not both.
1523 */
1524 if (vma || (addr != -1)) {
Dave Hansene0ec90e2015-11-05 18:50:20 -08001525 VM_WARN_ON_ONCE(addr == -1);
1526 VM_WARN_ON_ONCE(nid != NUMA_NO_NODE);
Dave Hansen099730d2015-11-05 18:50:17 -08001527 }
1528 /*
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001529 * Assume we will successfully allocate the surplus page to
1530 * prevent racing processes from causing the surplus to exceed
1531 * overcommit
1532 *
1533 * This however introduces a different race, where a process B
1534 * tries to grow the static hugepage pool while alloc_pages() is
1535 * called by process A. B will only examine the per-node
1536 * counters in determining if surplus huge pages can be
1537 * converted to normal huge pages in adjust_pool_surplus(). A
1538 * won't be able to increment the per-node counter, until the
1539 * lock is dropped by B, but B doesn't drop hugetlb_lock until
1540 * no more huge pages can be converted from surplus to normal
1541 * state (and doesn't try to convert again). Thus, we have a
1542 * case where a surplus huge page exists, the pool is grown, and
1543 * the surplus huge page still exists after, even though it
1544 * should just have been converted to a normal huge page. This
1545 * does not leak memory, though, as the hugepage will be freed
1546 * once it is out of use. It also does not allow the counters to
1547 * go out of whack in adjust_pool_surplus() as we don't modify
1548 * the node values until we've gotten the hugepage and only the
1549 * per-node value is checked there.
1550 */
1551 spin_lock(&hugetlb_lock);
Andi Kleena5516432008-07-23 21:27:41 -07001552 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001553 spin_unlock(&hugetlb_lock);
1554 return NULL;
1555 } else {
Andi Kleena5516432008-07-23 21:27:41 -07001556 h->nr_huge_pages++;
1557 h->surplus_huge_pages++;
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001558 }
1559 spin_unlock(&hugetlb_lock);
1560
Dave Hansen099730d2015-11-05 18:50:17 -08001561 page = __hugetlb_alloc_buddy_huge_page(h, vma, addr, nid);
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001562
1563 spin_lock(&hugetlb_lock);
Adam Litke7893d1d2007-10-16 01:26:18 -07001564 if (page) {
Aneesh Kumar K.V0edaecf2012-07-31 16:42:07 -07001565 INIT_LIST_HEAD(&page->lru);
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001566 r_nid = page_to_nid(page);
Kirill A. Shutemovf1e61552015-11-06 16:29:50 -08001567 set_compound_page_dtor(page, HUGETLB_PAGE_DTOR);
Aneesh Kumar K.V9dd540e2012-07-31 16:42:15 -07001568 set_hugetlb_cgroup(page, NULL);
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001569 /*
1570 * We incremented the global counters already
1571 */
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001572 h->nr_huge_pages_node[r_nid]++;
1573 h->surplus_huge_pages_node[r_nid]++;
Adam Litke3b116302008-04-28 02:13:06 -07001574 __count_vm_event(HTLB_BUDDY_PGALLOC);
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001575 } else {
Andi Kleena5516432008-07-23 21:27:41 -07001576 h->nr_huge_pages--;
1577 h->surplus_huge_pages--;
Adam Litke3b116302008-04-28 02:13:06 -07001578 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
Adam Litke7893d1d2007-10-16 01:26:18 -07001579 }
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001580 spin_unlock(&hugetlb_lock);
Adam Litke7893d1d2007-10-16 01:26:18 -07001581
1582 return page;
1583}
1584
Adam Litkee4e574b2007-10-16 01:26:19 -07001585/*
Dave Hansen099730d2015-11-05 18:50:17 -08001586 * Allocate a huge page from 'nid'. Note, 'nid' may be
1587 * NUMA_NO_NODE, which means that it may be allocated
1588 * anywhere.
1589 */
Dave Hansene0ec90e2015-11-05 18:50:20 -08001590static
Dave Hansen099730d2015-11-05 18:50:17 -08001591struct page *__alloc_buddy_huge_page_no_mpol(struct hstate *h, int nid)
1592{
1593 unsigned long addr = -1;
1594
1595 return __alloc_buddy_huge_page(h, NULL, addr, nid);
1596}
1597
1598/*
1599 * Use the VMA's mpolicy to allocate a huge page from the buddy.
1600 */
Dave Hansene0ec90e2015-11-05 18:50:20 -08001601static
Dave Hansen099730d2015-11-05 18:50:17 -08001602struct page *__alloc_buddy_huge_page_with_mpol(struct hstate *h,
1603 struct vm_area_struct *vma, unsigned long addr)
1604{
1605 return __alloc_buddy_huge_page(h, vma, addr, NUMA_NO_NODE);
1606}
1607
1608/*
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001609 * This allocation function is useful in the context where vma is irrelevant.
1610 * E.g. soft-offlining uses this function because it only cares physical
1611 * address of error page.
1612 */
1613struct page *alloc_huge_page_node(struct hstate *h, int nid)
1614{
Joonsoo Kim4ef91842013-09-11 14:21:51 -07001615 struct page *page = NULL;
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001616
1617 spin_lock(&hugetlb_lock);
Joonsoo Kim4ef91842013-09-11 14:21:51 -07001618 if (h->free_huge_pages - h->resv_huge_pages > 0)
1619 page = dequeue_huge_page_node(h, nid);
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001620 spin_unlock(&hugetlb_lock);
1621
Aneesh Kumar K.V94ae8ba2012-07-31 16:42:35 -07001622 if (!page)
Dave Hansen099730d2015-11-05 18:50:17 -08001623 page = __alloc_buddy_huge_page_no_mpol(h, nid);
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001624
1625 return page;
1626}
1627
1628/*
Lucas De Marchi25985ed2011-03-30 22:57:33 -03001629 * Increase the hugetlb pool such that it can accommodate a reservation
Adam Litkee4e574b2007-10-16 01:26:19 -07001630 * of size 'delta'.
1631 */
Andi Kleena5516432008-07-23 21:27:41 -07001632static int gather_surplus_pages(struct hstate *h, int delta)
Adam Litkee4e574b2007-10-16 01:26:19 -07001633{
1634 struct list_head surplus_list;
1635 struct page *page, *tmp;
1636 int ret, i;
1637 int needed, allocated;
Hillf Danton28073b02012-03-21 16:34:00 -07001638 bool alloc_ok = true;
Adam Litkee4e574b2007-10-16 01:26:19 -07001639
Andi Kleena5516432008-07-23 21:27:41 -07001640 needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
Adam Litkeac09b3a2008-03-04 14:29:38 -08001641 if (needed <= 0) {
Andi Kleena5516432008-07-23 21:27:41 -07001642 h->resv_huge_pages += delta;
Adam Litkee4e574b2007-10-16 01:26:19 -07001643 return 0;
Adam Litkeac09b3a2008-03-04 14:29:38 -08001644 }
Adam Litkee4e574b2007-10-16 01:26:19 -07001645
1646 allocated = 0;
1647 INIT_LIST_HEAD(&surplus_list);
1648
1649 ret = -ENOMEM;
1650retry:
1651 spin_unlock(&hugetlb_lock);
1652 for (i = 0; i < needed; i++) {
Dave Hansen099730d2015-11-05 18:50:17 -08001653 page = __alloc_buddy_huge_page_no_mpol(h, NUMA_NO_NODE);
Hillf Danton28073b02012-03-21 16:34:00 -07001654 if (!page) {
1655 alloc_ok = false;
1656 break;
1657 }
Adam Litkee4e574b2007-10-16 01:26:19 -07001658 list_add(&page->lru, &surplus_list);
1659 }
Hillf Danton28073b02012-03-21 16:34:00 -07001660 allocated += i;
Adam Litkee4e574b2007-10-16 01:26:19 -07001661
1662 /*
1663 * After retaking hugetlb_lock, we need to recalculate 'needed'
1664 * because either resv_huge_pages or free_huge_pages may have changed.
1665 */
1666 spin_lock(&hugetlb_lock);
Andi Kleena5516432008-07-23 21:27:41 -07001667 needed = (h->resv_huge_pages + delta) -
1668 (h->free_huge_pages + allocated);
Hillf Danton28073b02012-03-21 16:34:00 -07001669 if (needed > 0) {
1670 if (alloc_ok)
1671 goto retry;
1672 /*
1673 * We were not able to allocate enough pages to
1674 * satisfy the entire reservation so we free what
1675 * we've allocated so far.
1676 */
1677 goto free;
1678 }
Adam Litkee4e574b2007-10-16 01:26:19 -07001679 /*
1680 * The surplus_list now contains _at_least_ the number of extra pages
Lucas De Marchi25985ed2011-03-30 22:57:33 -03001681 * needed to accommodate the reservation. Add the appropriate number
Adam Litkee4e574b2007-10-16 01:26:19 -07001682 * of pages to the hugetlb pool and free the extras back to the buddy
Adam Litkeac09b3a2008-03-04 14:29:38 -08001683 * allocator. Commit the entire reservation here to prevent another
1684 * process from stealing the pages as they are added to the pool but
1685 * before they are reserved.
Adam Litkee4e574b2007-10-16 01:26:19 -07001686 */
1687 needed += allocated;
Andi Kleena5516432008-07-23 21:27:41 -07001688 h->resv_huge_pages += delta;
Adam Litkee4e574b2007-10-16 01:26:19 -07001689 ret = 0;
Naoya Horiguchia9869b82010-09-08 10:19:37 +09001690
Adam Litke19fc3f02008-04-28 02:12:20 -07001691 /* Free the needed pages to the hugetlb pool */
Adam Litkee4e574b2007-10-16 01:26:19 -07001692 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
Adam Litke19fc3f02008-04-28 02:12:20 -07001693 if ((--needed) < 0)
1694 break;
Naoya Horiguchia9869b82010-09-08 10:19:37 +09001695 /*
1696 * This page is now managed by the hugetlb allocator and has
1697 * no users -- drop the buddy allocator's reference.
1698 */
1699 put_page_testzero(page);
Sasha Levin309381fea2014-01-23 15:52:54 -08001700 VM_BUG_ON_PAGE(page_count(page), page);
Andi Kleena5516432008-07-23 21:27:41 -07001701 enqueue_huge_page(h, page);
Adam Litke19fc3f02008-04-28 02:12:20 -07001702 }
Hillf Danton28073b02012-03-21 16:34:00 -07001703free:
Hillf Dantonb0365c82011-12-28 15:57:16 -08001704 spin_unlock(&hugetlb_lock);
Adam Litke19fc3f02008-04-28 02:12:20 -07001705
1706 /* Free unnecessary surplus pages to the buddy allocator */
Joonsoo Kimc0d934b2013-09-11 14:21:02 -07001707 list_for_each_entry_safe(page, tmp, &surplus_list, lru)
1708 put_page(page);
Naoya Horiguchia9869b82010-09-08 10:19:37 +09001709 spin_lock(&hugetlb_lock);
Adam Litkee4e574b2007-10-16 01:26:19 -07001710
1711 return ret;
1712}
1713
1714/*
1715 * When releasing a hugetlb pool reservation, any surplus pages that were
1716 * allocated to satisfy the reservation must be explicitly freed if they were
1717 * never used.
Lee Schermerhorn685f3452009-09-21 17:01:23 -07001718 * Called with hugetlb_lock held.
Adam Litkee4e574b2007-10-16 01:26:19 -07001719 */
Andi Kleena5516432008-07-23 21:27:41 -07001720static void return_unused_surplus_pages(struct hstate *h,
1721 unsigned long unused_resv_pages)
Adam Litkee4e574b2007-10-16 01:26:19 -07001722{
Adam Litkee4e574b2007-10-16 01:26:19 -07001723 unsigned long nr_pages;
1724
Adam Litkeac09b3a2008-03-04 14:29:38 -08001725 /* Uncommit the reservation */
Andi Kleena5516432008-07-23 21:27:41 -07001726 h->resv_huge_pages -= unused_resv_pages;
Adam Litkeac09b3a2008-03-04 14:29:38 -08001727
Andi Kleenaa888a72008-07-23 21:27:47 -07001728 /* Cannot return gigantic pages currently */
Luiz Capitulinobae7f4a2014-06-04 16:07:08 -07001729 if (hstate_is_gigantic(h))
Andi Kleenaa888a72008-07-23 21:27:47 -07001730 return;
1731
Andi Kleena5516432008-07-23 21:27:41 -07001732 nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
Adam Litkee4e574b2007-10-16 01:26:19 -07001733
Lee Schermerhorn685f3452009-09-21 17:01:23 -07001734 /*
1735 * We want to release as many surplus pages as possible, spread
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -08001736 * evenly across all nodes with memory. Iterate across these nodes
1737 * until we can no longer free unreserved surplus pages. This occurs
1738 * when the nodes with surplus pages have no free pages.
1739 * free_pool_huge_page() will balance the the freed pages across the
1740 * on-line nodes with memory and will handle the hstate accounting.
Lee Schermerhorn685f3452009-09-21 17:01:23 -07001741 */
1742 while (nr_pages--) {
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08001743 if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1))
Lee Schermerhorn685f3452009-09-21 17:01:23 -07001744 break;
Mizuma, Masayoshi7848a4b2014-04-18 15:07:18 -07001745 cond_resched_lock(&hugetlb_lock);
Adam Litkee4e574b2007-10-16 01:26:19 -07001746 }
1747}
1748
Mike Kravetz5e911372015-09-08 15:01:28 -07001749
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001750/*
Mike Kravetzfeba16e2015-09-08 15:01:31 -07001751 * vma_needs_reservation, vma_commit_reservation and vma_end_reservation
Mike Kravetz5e911372015-09-08 15:01:28 -07001752 * are used by the huge page allocation routines to manage reservations.
Mike Kravetzcf3ad202015-06-24 16:57:55 -07001753 *
1754 * vma_needs_reservation is called to determine if the huge page at addr
1755 * within the vma has an associated reservation. If a reservation is
1756 * needed, the value 1 is returned. The caller is then responsible for
1757 * managing the global reservation and subpool usage counts. After
1758 * the huge page has been allocated, vma_commit_reservation is called
Mike Kravetzfeba16e2015-09-08 15:01:31 -07001759 * to add the page to the reservation map. If the page allocation fails,
1760 * the reservation must be ended instead of committed. vma_end_reservation
1761 * is called in such cases.
Mike Kravetzcf3ad202015-06-24 16:57:55 -07001762 *
1763 * In the normal case, vma_commit_reservation returns the same value
1764 * as the preceding vma_needs_reservation call. The only time this
1765 * is not the case is if a reserve map was changed between calls. It
1766 * is the responsibility of the caller to notice the difference and
1767 * take appropriate action.
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001768 */
Mike Kravetz5e911372015-09-08 15:01:28 -07001769enum vma_resv_mode {
1770 VMA_NEEDS_RESV,
1771 VMA_COMMIT_RESV,
Mike Kravetzfeba16e2015-09-08 15:01:31 -07001772 VMA_END_RESV,
Mike Kravetz5e911372015-09-08 15:01:28 -07001773};
Mike Kravetzcf3ad202015-06-24 16:57:55 -07001774static long __vma_reservation_common(struct hstate *h,
1775 struct vm_area_struct *vma, unsigned long addr,
Mike Kravetz5e911372015-09-08 15:01:28 -07001776 enum vma_resv_mode mode)
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001777{
Joonsoo Kim4e35f482014-04-03 14:47:30 -07001778 struct resv_map *resv;
1779 pgoff_t idx;
Mike Kravetzcf3ad202015-06-24 16:57:55 -07001780 long ret;
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001781
Joonsoo Kim4e35f482014-04-03 14:47:30 -07001782 resv = vma_resv_map(vma);
1783 if (!resv)
Andy Whitcroft84afd992008-07-23 21:27:32 -07001784 return 1;
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001785
Joonsoo Kim4e35f482014-04-03 14:47:30 -07001786 idx = vma_hugecache_offset(h, vma, addr);
Mike Kravetz5e911372015-09-08 15:01:28 -07001787 switch (mode) {
1788 case VMA_NEEDS_RESV:
Mike Kravetzcf3ad202015-06-24 16:57:55 -07001789 ret = region_chg(resv, idx, idx + 1);
Mike Kravetz5e911372015-09-08 15:01:28 -07001790 break;
1791 case VMA_COMMIT_RESV:
1792 ret = region_add(resv, idx, idx + 1);
1793 break;
Mike Kravetzfeba16e2015-09-08 15:01:31 -07001794 case VMA_END_RESV:
Mike Kravetz5e911372015-09-08 15:01:28 -07001795 region_abort(resv, idx, idx + 1);
1796 ret = 0;
1797 break;
1798 default:
1799 BUG();
1800 }
Andy Whitcroft84afd992008-07-23 21:27:32 -07001801
Joonsoo Kim4e35f482014-04-03 14:47:30 -07001802 if (vma->vm_flags & VM_MAYSHARE)
Mike Kravetzcf3ad202015-06-24 16:57:55 -07001803 return ret;
Joonsoo Kim4e35f482014-04-03 14:47:30 -07001804 else
Mike Kravetzcf3ad202015-06-24 16:57:55 -07001805 return ret < 0 ? ret : 0;
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001806}
Mike Kravetzcf3ad202015-06-24 16:57:55 -07001807
1808static long vma_needs_reservation(struct hstate *h,
Andi Kleena5516432008-07-23 21:27:41 -07001809 struct vm_area_struct *vma, unsigned long addr)
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001810{
Mike Kravetz5e911372015-09-08 15:01:28 -07001811 return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV);
Mike Kravetzcf3ad202015-06-24 16:57:55 -07001812}
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001813
Mike Kravetzcf3ad202015-06-24 16:57:55 -07001814static long vma_commit_reservation(struct hstate *h,
1815 struct vm_area_struct *vma, unsigned long addr)
1816{
Mike Kravetz5e911372015-09-08 15:01:28 -07001817 return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV);
1818}
1819
Mike Kravetzfeba16e2015-09-08 15:01:31 -07001820static void vma_end_reservation(struct hstate *h,
Mike Kravetz5e911372015-09-08 15:01:28 -07001821 struct vm_area_struct *vma, unsigned long addr)
1822{
Mike Kravetzfeba16e2015-09-08 15:01:31 -07001823 (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV);
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001824}
1825
Mike Kravetz70c35472015-09-08 15:01:54 -07001826struct page *alloc_huge_page(struct vm_area_struct *vma,
Mel Gorman04f2cbe2008-07-23 21:27:25 -07001827 unsigned long addr, int avoid_reserve)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001828{
David Gibson90481622012-03-21 16:34:12 -07001829 struct hugepage_subpool *spool = subpool_vma(vma);
Andi Kleena5516432008-07-23 21:27:41 -07001830 struct hstate *h = hstate_vma(vma);
Adam Litke348ea202007-11-14 16:59:37 -08001831 struct page *page;
Mike Kravetzd85f69b2015-09-08 15:01:47 -07001832 long map_chg, map_commit;
1833 long gbl_chg;
Aneesh Kumar K.V6d76dcf2012-07-31 16:42:18 -07001834 int ret, idx;
1835 struct hugetlb_cgroup *h_cg;
Adam Litke2fc39ce2007-11-14 16:59:39 -08001836
Aneesh Kumar K.V6d76dcf2012-07-31 16:42:18 -07001837 idx = hstate_index(h);
Mel Gormana1e78772008-07-23 21:27:23 -07001838 /*
Mike Kravetzd85f69b2015-09-08 15:01:47 -07001839 * Examine the region/reserve map to determine if the process
1840 * has a reservation for the page to be allocated. A return
1841 * code of zero indicates a reservation exists (no change).
Mel Gormana1e78772008-07-23 21:27:23 -07001842 */
Mike Kravetzd85f69b2015-09-08 15:01:47 -07001843 map_chg = gbl_chg = vma_needs_reservation(h, vma, addr);
1844 if (map_chg < 0)
Aneesh Kumar K.V76dcee72012-07-31 16:41:57 -07001845 return ERR_PTR(-ENOMEM);
Mike Kravetzd85f69b2015-09-08 15:01:47 -07001846
1847 /*
1848 * Processes that did not create the mapping will have no
1849 * reserves as indicated by the region/reserve map. Check
1850 * that the allocation will not exceed the subpool limit.
1851 * Allocations for MAP_NORESERVE mappings also need to be
1852 * checked against any subpool limit.
1853 */
1854 if (map_chg || avoid_reserve) {
1855 gbl_chg = hugepage_subpool_get_pages(spool, 1);
1856 if (gbl_chg < 0) {
Mike Kravetzfeba16e2015-09-08 15:01:31 -07001857 vma_end_reservation(h, vma, addr);
Aneesh Kumar K.V76dcee72012-07-31 16:41:57 -07001858 return ERR_PTR(-ENOSPC);
Mike Kravetz5e911372015-09-08 15:01:28 -07001859 }
Mel Gormana1e78772008-07-23 21:27:23 -07001860
Mike Kravetzd85f69b2015-09-08 15:01:47 -07001861 /*
1862 * Even though there was no reservation in the region/reserve
1863 * map, there could be reservations associated with the
1864 * subpool that can be used. This would be indicated if the
1865 * return value of hugepage_subpool_get_pages() is zero.
1866 * However, if avoid_reserve is specified we still avoid even
1867 * the subpool reservations.
1868 */
1869 if (avoid_reserve)
1870 gbl_chg = 1;
1871 }
1872
Aneesh Kumar K.V6d76dcf2012-07-31 16:42:18 -07001873 ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg);
Jianyu Zhan8f34af62014-06-04 16:10:36 -07001874 if (ret)
1875 goto out_subpool_put;
1876
Mel Gormana1e78772008-07-23 21:27:23 -07001877 spin_lock(&hugetlb_lock);
Mike Kravetzd85f69b2015-09-08 15:01:47 -07001878 /*
1879 * glb_chg is passed to indicate whether or not a page must be taken
1880 * from the global free pool (global change). gbl_chg == 0 indicates
1881 * a reservation exists for the allocation.
1882 */
1883 page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg);
Joonsoo Kim81a6fca2013-09-11 14:20:58 -07001884 if (!page) {
Aneesh Kumar K.V94ae8ba2012-07-31 16:42:35 -07001885 spin_unlock(&hugetlb_lock);
Dave Hansen099730d2015-11-05 18:50:17 -08001886 page = __alloc_buddy_huge_page_with_mpol(h, vma, addr);
Jianyu Zhan8f34af62014-06-04 16:10:36 -07001887 if (!page)
1888 goto out_uncharge_cgroup;
1889
Aneesh Kumar K.V79dbb232012-07-31 16:42:32 -07001890 spin_lock(&hugetlb_lock);
1891 list_move(&page->lru, &h->hugepage_activelist);
Joonsoo Kim81a6fca2013-09-11 14:20:58 -07001892 /* Fall through */
Mel Gormana1e78772008-07-23 21:27:23 -07001893 }
Joonsoo Kim81a6fca2013-09-11 14:20:58 -07001894 hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page);
1895 spin_unlock(&hugetlb_lock);
Mel Gormana1e78772008-07-23 21:27:23 -07001896
David Gibson90481622012-03-21 16:34:12 -07001897 set_page_private(page, (unsigned long)spool);
Mel Gormana1e78772008-07-23 21:27:23 -07001898
Mike Kravetzd85f69b2015-09-08 15:01:47 -07001899 map_commit = vma_commit_reservation(h, vma, addr);
1900 if (unlikely(map_chg > map_commit)) {
Mike Kravetz33039672015-06-24 16:57:58 -07001901 /*
1902 * The page was added to the reservation map between
1903 * vma_needs_reservation and vma_commit_reservation.
1904 * This indicates a race with hugetlb_reserve_pages.
1905 * Adjust for the subpool count incremented above AND
1906 * in hugetlb_reserve_pages for the same page. Also,
1907 * the reservation count added in hugetlb_reserve_pages
1908 * no longer applies.
1909 */
1910 long rsv_adjust;
1911
1912 rsv_adjust = hugepage_subpool_put_pages(spool, 1);
1913 hugetlb_acct_memory(h, -rsv_adjust);
1914 }
Adam Litke90d8b7e2007-11-14 16:59:42 -08001915 return page;
Jianyu Zhan8f34af62014-06-04 16:10:36 -07001916
1917out_uncharge_cgroup:
1918 hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg);
1919out_subpool_put:
Mike Kravetzd85f69b2015-09-08 15:01:47 -07001920 if (map_chg || avoid_reserve)
Jianyu Zhan8f34af62014-06-04 16:10:36 -07001921 hugepage_subpool_put_pages(spool, 1);
Mike Kravetzfeba16e2015-09-08 15:01:31 -07001922 vma_end_reservation(h, vma, addr);
Jianyu Zhan8f34af62014-06-04 16:10:36 -07001923 return ERR_PTR(-ENOSPC);
David Gibsonb45b5bd2006-03-22 00:08:55 -08001924}
1925
Naoya Horiguchi74060e42013-09-11 14:22:06 -07001926/*
1927 * alloc_huge_page()'s wrapper which simply returns the page if allocation
1928 * succeeds, otherwise NULL. This function is called from new_vma_page(),
1929 * where no ERR_VALUE is expected to be returned.
1930 */
1931struct page *alloc_huge_page_noerr(struct vm_area_struct *vma,
1932 unsigned long addr, int avoid_reserve)
1933{
1934 struct page *page = alloc_huge_page(vma, addr, avoid_reserve);
1935 if (IS_ERR(page))
1936 page = NULL;
1937 return page;
1938}
1939
Cyrill Gorcunov91f47662009-01-06 14:40:33 -08001940int __weak alloc_bootmem_huge_page(struct hstate *h)
Andi Kleenaa888a72008-07-23 21:27:47 -07001941{
1942 struct huge_bootmem_page *m;
Joonsoo Kimb2261022013-09-11 14:21:00 -07001943 int nr_nodes, node;
Andi Kleenaa888a72008-07-23 21:27:47 -07001944
Joonsoo Kimb2261022013-09-11 14:21:00 -07001945 for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) {
Andi Kleenaa888a72008-07-23 21:27:47 -07001946 void *addr;
1947
Grygorii Strashko8b89a112014-01-21 15:50:36 -08001948 addr = memblock_virt_alloc_try_nid_nopanic(
1949 huge_page_size(h), huge_page_size(h),
1950 0, BOOTMEM_ALLOC_ACCESSIBLE, node);
Andi Kleenaa888a72008-07-23 21:27:47 -07001951 if (addr) {
1952 /*
1953 * Use the beginning of the huge page to store the
1954 * huge_bootmem_page struct (until gather_bootmem
1955 * puts them into the mem_map).
1956 */
1957 m = addr;
Cyrill Gorcunov91f47662009-01-06 14:40:33 -08001958 goto found;
Andi Kleenaa888a72008-07-23 21:27:47 -07001959 }
Andi Kleenaa888a72008-07-23 21:27:47 -07001960 }
1961 return 0;
1962
1963found:
Luiz Capitulinodf994ea2014-12-12 16:55:21 -08001964 BUG_ON(!IS_ALIGNED(virt_to_phys(m), huge_page_size(h)));
Andi Kleenaa888a72008-07-23 21:27:47 -07001965 /* Put them into a private list first because mem_map is not up yet */
1966 list_add(&m->list, &huge_boot_pages);
1967 m->hstate = h;
1968 return 1;
1969}
1970
Kirill A. Shutemovd00181b2015-11-06 16:29:57 -08001971static void __init prep_compound_huge_page(struct page *page,
1972 unsigned int order)
Andy Whitcroft18229df2008-11-06 12:53:27 -08001973{
1974 if (unlikely(order > (MAX_ORDER - 1)))
1975 prep_compound_gigantic_page(page, order);
1976 else
1977 prep_compound_page(page, order);
1978}
1979
Andi Kleenaa888a72008-07-23 21:27:47 -07001980/* Put bootmem huge pages into the standard lists after mem_map is up */
1981static void __init gather_bootmem_prealloc(void)
1982{
1983 struct huge_bootmem_page *m;
1984
1985 list_for_each_entry(m, &huge_boot_pages, list) {
Andi Kleenaa888a72008-07-23 21:27:47 -07001986 struct hstate *h = m->hstate;
Becky Bruceee8f2482011-07-25 17:11:50 -07001987 struct page *page;
1988
1989#ifdef CONFIG_HIGHMEM
1990 page = pfn_to_page(m->phys >> PAGE_SHIFT);
Grygorii Strashko8b89a112014-01-21 15:50:36 -08001991 memblock_free_late(__pa(m),
1992 sizeof(struct huge_bootmem_page));
Becky Bruceee8f2482011-07-25 17:11:50 -07001993#else
1994 page = virt_to_page(m);
1995#endif
Andi Kleenaa888a72008-07-23 21:27:47 -07001996 WARN_ON(page_count(page) != 1);
Andy Whitcroft18229df2008-11-06 12:53:27 -08001997 prep_compound_huge_page(page, h->order);
Andrea Arcangelief5a22b2013-10-16 13:46:56 -07001998 WARN_ON(PageReserved(page));
Andi Kleenaa888a72008-07-23 21:27:47 -07001999 prep_new_huge_page(h, page, page_to_nid(page));
Rafael Aquinib0320c72011-06-15 15:08:39 -07002000 /*
2001 * If we had gigantic hugepages allocated at boot time, we need
2002 * to restore the 'stolen' pages to totalram_pages in order to
2003 * fix confusing memory reports from free(1) and another
2004 * side-effects, like CommitLimit going negative.
2005 */
Luiz Capitulinobae7f4a2014-06-04 16:07:08 -07002006 if (hstate_is_gigantic(h))
Jiang Liu3dcc0572013-07-03 15:03:21 -07002007 adjust_managed_page_count(page, 1 << h->order);
Andi Kleenaa888a72008-07-23 21:27:47 -07002008 }
2009}
2010
Andi Kleen8faa8b02008-07-23 21:27:48 -07002011static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002012{
2013 unsigned long i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002014
Andi Kleene5ff2152008-07-23 21:27:42 -07002015 for (i = 0; i < h->max_huge_pages; ++i) {
Luiz Capitulinobae7f4a2014-06-04 16:07:08 -07002016 if (hstate_is_gigantic(h)) {
Andi Kleenaa888a72008-07-23 21:27:47 -07002017 if (!alloc_bootmem_huge_page(h))
2018 break;
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -08002019 } else if (!alloc_fresh_huge_page(h,
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08002020 &node_states[N_MEMORY]))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002021 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002022 }
Andi Kleen8faa8b02008-07-23 21:27:48 -07002023 h->max_huge_pages = i;
Andi Kleene5ff2152008-07-23 21:27:42 -07002024}
2025
2026static void __init hugetlb_init_hstates(void)
2027{
2028 struct hstate *h;
2029
2030 for_each_hstate(h) {
Naoya Horiguchi641844f2015-06-24 16:56:59 -07002031 if (minimum_order > huge_page_order(h))
2032 minimum_order = huge_page_order(h);
2033
Andi Kleen8faa8b02008-07-23 21:27:48 -07002034 /* oversize hugepages were init'ed in early boot */
Luiz Capitulinobae7f4a2014-06-04 16:07:08 -07002035 if (!hstate_is_gigantic(h))
Andi Kleen8faa8b02008-07-23 21:27:48 -07002036 hugetlb_hstate_alloc_pages(h);
Andi Kleene5ff2152008-07-23 21:27:42 -07002037 }
Naoya Horiguchi641844f2015-06-24 16:56:59 -07002038 VM_BUG_ON(minimum_order == UINT_MAX);
Andi Kleene5ff2152008-07-23 21:27:42 -07002039}
2040
Andi Kleen4abd32d2008-07-23 21:27:49 -07002041static char * __init memfmt(char *buf, unsigned long n)
2042{
2043 if (n >= (1UL << 30))
2044 sprintf(buf, "%lu GB", n >> 30);
2045 else if (n >= (1UL << 20))
2046 sprintf(buf, "%lu MB", n >> 20);
2047 else
2048 sprintf(buf, "%lu KB", n >> 10);
2049 return buf;
2050}
2051
Andi Kleene5ff2152008-07-23 21:27:42 -07002052static void __init report_hugepages(void)
2053{
2054 struct hstate *h;
2055
2056 for_each_hstate(h) {
Andi Kleen4abd32d2008-07-23 21:27:49 -07002057 char buf[32];
Andrew Mortonffb22af2013-02-22 16:32:08 -08002058 pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n",
Andi Kleen4abd32d2008-07-23 21:27:49 -07002059 memfmt(buf, huge_page_size(h)),
2060 h->free_huge_pages);
Andi Kleene5ff2152008-07-23 21:27:42 -07002061 }
2062}
2063
Linus Torvalds1da177e2005-04-16 15:20:36 -07002064#ifdef CONFIG_HIGHMEM
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08002065static void try_to_free_low(struct hstate *h, unsigned long count,
2066 nodemask_t *nodes_allowed)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002067{
Christoph Lameter4415cc82006-09-25 23:31:55 -07002068 int i;
2069
Luiz Capitulinobae7f4a2014-06-04 16:07:08 -07002070 if (hstate_is_gigantic(h))
Andi Kleenaa888a72008-07-23 21:27:47 -07002071 return;
2072
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08002073 for_each_node_mask(i, *nodes_allowed) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002074 struct page *page, *next;
Andi Kleena5516432008-07-23 21:27:41 -07002075 struct list_head *freel = &h->hugepage_freelists[i];
2076 list_for_each_entry_safe(page, next, freel, lru) {
2077 if (count >= h->nr_huge_pages)
Adam Litke6b0c8802007-10-16 01:26:23 -07002078 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002079 if (PageHighMem(page))
2080 continue;
2081 list_del(&page->lru);
Andi Kleene5ff2152008-07-23 21:27:42 -07002082 update_and_free_page(h, page);
Andi Kleena5516432008-07-23 21:27:41 -07002083 h->free_huge_pages--;
2084 h->free_huge_pages_node[page_to_nid(page)]--;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002085 }
2086 }
2087}
2088#else
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08002089static inline void try_to_free_low(struct hstate *h, unsigned long count,
2090 nodemask_t *nodes_allowed)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002091{
2092}
2093#endif
2094
Wu Fengguang20a03072009-06-16 15:32:22 -07002095/*
2096 * Increment or decrement surplus_huge_pages. Keep node-specific counters
2097 * balanced by operating on them in a round-robin fashion.
2098 * Returns 1 if an adjustment was made.
2099 */
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08002100static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
2101 int delta)
Wu Fengguang20a03072009-06-16 15:32:22 -07002102{
Joonsoo Kimb2261022013-09-11 14:21:00 -07002103 int nr_nodes, node;
Wu Fengguang20a03072009-06-16 15:32:22 -07002104
2105 VM_BUG_ON(delta != -1 && delta != 1);
Wu Fengguang20a03072009-06-16 15:32:22 -07002106
Joonsoo Kimb2261022013-09-11 14:21:00 -07002107 if (delta < 0) {
2108 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
2109 if (h->surplus_huge_pages_node[node])
2110 goto found;
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07002111 }
Joonsoo Kimb2261022013-09-11 14:21:00 -07002112 } else {
2113 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
2114 if (h->surplus_huge_pages_node[node] <
2115 h->nr_huge_pages_node[node])
2116 goto found;
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07002117 }
Joonsoo Kimb2261022013-09-11 14:21:00 -07002118 }
2119 return 0;
Wu Fengguang20a03072009-06-16 15:32:22 -07002120
Joonsoo Kimb2261022013-09-11 14:21:00 -07002121found:
2122 h->surplus_huge_pages += delta;
2123 h->surplus_huge_pages_node[node] += delta;
2124 return 1;
Wu Fengguang20a03072009-06-16 15:32:22 -07002125}
2126
Andi Kleena5516432008-07-23 21:27:41 -07002127#define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08002128static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count,
2129 nodemask_t *nodes_allowed)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002130{
Adam Litke7893d1d2007-10-16 01:26:18 -07002131 unsigned long min_count, ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002132
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07002133 if (hstate_is_gigantic(h) && !gigantic_page_supported())
Andi Kleenaa888a72008-07-23 21:27:47 -07002134 return h->max_huge_pages;
2135
Adam Litke7893d1d2007-10-16 01:26:18 -07002136 /*
2137 * Increase the pool size
2138 * First take pages out of surplus state. Then make up the
2139 * remaining difference by allocating fresh huge pages.
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08002140 *
Naoya Horiguchid15c7c02015-11-10 14:45:11 -08002141 * We might race with __alloc_buddy_huge_page() here and be unable
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08002142 * to convert a surplus huge page to a normal huge page. That is
2143 * not critical, though, it just means the overall size of the
2144 * pool might be one hugepage larger than it needs to be, but
2145 * within all the constraints specified by the sysctls.
Adam Litke7893d1d2007-10-16 01:26:18 -07002146 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002147 spin_lock(&hugetlb_lock);
Andi Kleena5516432008-07-23 21:27:41 -07002148 while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08002149 if (!adjust_pool_surplus(h, nodes_allowed, -1))
Adam Litke7893d1d2007-10-16 01:26:18 -07002150 break;
2151 }
2152
Andi Kleena5516432008-07-23 21:27:41 -07002153 while (count > persistent_huge_pages(h)) {
Adam Litke7893d1d2007-10-16 01:26:18 -07002154 /*
2155 * If this allocation races such that we no longer need the
2156 * page, free_huge_page will handle it by freeing the page
2157 * and reducing the surplus.
2158 */
2159 spin_unlock(&hugetlb_lock);
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07002160 if (hstate_is_gigantic(h))
2161 ret = alloc_fresh_gigantic_page(h, nodes_allowed);
2162 else
2163 ret = alloc_fresh_huge_page(h, nodes_allowed);
Adam Litke7893d1d2007-10-16 01:26:18 -07002164 spin_lock(&hugetlb_lock);
2165 if (!ret)
2166 goto out;
2167
Mel Gorman536240f22009-12-14 17:59:56 -08002168 /* Bail for signals. Probably ctrl-c from user */
2169 if (signal_pending(current))
2170 goto out;
Adam Litke7893d1d2007-10-16 01:26:18 -07002171 }
Adam Litke7893d1d2007-10-16 01:26:18 -07002172
2173 /*
2174 * Decrease the pool size
2175 * First return free pages to the buddy allocator (being careful
2176 * to keep enough around to satisfy reservations). Then place
2177 * pages into surplus state as needed so the pool will shrink
2178 * to the desired size as pages become free.
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08002179 *
2180 * By placing pages into the surplus state independent of the
2181 * overcommit value, we are allowing the surplus pool size to
2182 * exceed overcommit. There are few sane options here. Since
Naoya Horiguchid15c7c02015-11-10 14:45:11 -08002183 * __alloc_buddy_huge_page() is checking the global counter,
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08002184 * though, we'll note that we're not allowed to exceed surplus
2185 * and won't grow the pool anywhere else. Not until one of the
2186 * sysctls are changed, or the surplus pages go out of use.
Adam Litke7893d1d2007-10-16 01:26:18 -07002187 */
Andi Kleena5516432008-07-23 21:27:41 -07002188 min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
Adam Litke6b0c8802007-10-16 01:26:23 -07002189 min_count = max(count, min_count);
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08002190 try_to_free_low(h, min_count, nodes_allowed);
Andi Kleena5516432008-07-23 21:27:41 -07002191 while (min_count < persistent_huge_pages(h)) {
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08002192 if (!free_pool_huge_page(h, nodes_allowed, 0))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002193 break;
Mizuma, Masayoshi55f67142014-04-07 15:37:54 -07002194 cond_resched_lock(&hugetlb_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002195 }
Andi Kleena5516432008-07-23 21:27:41 -07002196 while (count < persistent_huge_pages(h)) {
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08002197 if (!adjust_pool_surplus(h, nodes_allowed, 1))
Adam Litke7893d1d2007-10-16 01:26:18 -07002198 break;
2199 }
2200out:
Andi Kleena5516432008-07-23 21:27:41 -07002201 ret = persistent_huge_pages(h);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002202 spin_unlock(&hugetlb_lock);
Adam Litke7893d1d2007-10-16 01:26:18 -07002203 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002204}
2205
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002206#define HSTATE_ATTR_RO(_name) \
2207 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
2208
2209#define HSTATE_ATTR(_name) \
2210 static struct kobj_attribute _name##_attr = \
2211 __ATTR(_name, 0644, _name##_show, _name##_store)
2212
2213static struct kobject *hugepages_kobj;
2214static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
2215
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002216static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
2217
2218static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002219{
2220 int i;
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002221
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002222 for (i = 0; i < HUGE_MAX_HSTATE; i++)
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002223 if (hstate_kobjs[i] == kobj) {
2224 if (nidp)
2225 *nidp = NUMA_NO_NODE;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002226 return &hstates[i];
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002227 }
2228
2229 return kobj_to_node_hstate(kobj, nidp);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002230}
2231
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002232static ssize_t nr_hugepages_show_common(struct kobject *kobj,
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002233 struct kobj_attribute *attr, char *buf)
2234{
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002235 struct hstate *h;
2236 unsigned long nr_huge_pages;
2237 int nid;
2238
2239 h = kobj_to_hstate(kobj, &nid);
2240 if (nid == NUMA_NO_NODE)
2241 nr_huge_pages = h->nr_huge_pages;
2242 else
2243 nr_huge_pages = h->nr_huge_pages_node[nid];
2244
2245 return sprintf(buf, "%lu\n", nr_huge_pages);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002246}
Eric B Munsonadbe8722011-01-13 15:47:27 -08002247
David Rientjes238d3c12014-08-06 16:06:51 -07002248static ssize_t __nr_hugepages_store_common(bool obey_mempolicy,
2249 struct hstate *h, int nid,
2250 unsigned long count, size_t len)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002251{
2252 int err;
David Rientjesbad44b52009-12-14 17:58:38 -08002253 NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002254
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07002255 if (hstate_is_gigantic(h) && !gigantic_page_supported()) {
Eric B Munsonadbe8722011-01-13 15:47:27 -08002256 err = -EINVAL;
2257 goto out;
2258 }
2259
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002260 if (nid == NUMA_NO_NODE) {
2261 /*
2262 * global hstate attribute
2263 */
2264 if (!(obey_mempolicy &&
2265 init_nodemask_of_mempolicy(nodes_allowed))) {
2266 NODEMASK_FREE(nodes_allowed);
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08002267 nodes_allowed = &node_states[N_MEMORY];
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002268 }
2269 } else if (nodes_allowed) {
2270 /*
2271 * per node hstate attribute: adjust count to global,
2272 * but restrict alloc/free to the specified node.
2273 */
2274 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
2275 init_nodemask_of_node(nodes_allowed, nid);
2276 } else
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08002277 nodes_allowed = &node_states[N_MEMORY];
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002278
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002279 h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002280
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08002281 if (nodes_allowed != &node_states[N_MEMORY])
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002282 NODEMASK_FREE(nodes_allowed);
2283
2284 return len;
Eric B Munsonadbe8722011-01-13 15:47:27 -08002285out:
2286 NODEMASK_FREE(nodes_allowed);
2287 return err;
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002288}
2289
David Rientjes238d3c12014-08-06 16:06:51 -07002290static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
2291 struct kobject *kobj, const char *buf,
2292 size_t len)
2293{
2294 struct hstate *h;
2295 unsigned long count;
2296 int nid;
2297 int err;
2298
2299 err = kstrtoul(buf, 10, &count);
2300 if (err)
2301 return err;
2302
2303 h = kobj_to_hstate(kobj, &nid);
2304 return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len);
2305}
2306
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002307static ssize_t nr_hugepages_show(struct kobject *kobj,
2308 struct kobj_attribute *attr, char *buf)
2309{
2310 return nr_hugepages_show_common(kobj, attr, buf);
2311}
2312
2313static ssize_t nr_hugepages_store(struct kobject *kobj,
2314 struct kobj_attribute *attr, const char *buf, size_t len)
2315{
David Rientjes238d3c12014-08-06 16:06:51 -07002316 return nr_hugepages_store_common(false, kobj, buf, len);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002317}
2318HSTATE_ATTR(nr_hugepages);
2319
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002320#ifdef CONFIG_NUMA
2321
2322/*
2323 * hstate attribute for optionally mempolicy-based constraint on persistent
2324 * huge page alloc/free.
2325 */
2326static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
2327 struct kobj_attribute *attr, char *buf)
2328{
2329 return nr_hugepages_show_common(kobj, attr, buf);
2330}
2331
2332static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
2333 struct kobj_attribute *attr, const char *buf, size_t len)
2334{
David Rientjes238d3c12014-08-06 16:06:51 -07002335 return nr_hugepages_store_common(true, kobj, buf, len);
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002336}
2337HSTATE_ATTR(nr_hugepages_mempolicy);
2338#endif
2339
2340
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002341static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
2342 struct kobj_attribute *attr, char *buf)
2343{
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002344 struct hstate *h = kobj_to_hstate(kobj, NULL);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002345 return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
2346}
Eric B Munsonadbe8722011-01-13 15:47:27 -08002347
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002348static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
2349 struct kobj_attribute *attr, const char *buf, size_t count)
2350{
2351 int err;
2352 unsigned long input;
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002353 struct hstate *h = kobj_to_hstate(kobj, NULL);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002354
Luiz Capitulinobae7f4a2014-06-04 16:07:08 -07002355 if (hstate_is_gigantic(h))
Eric B Munsonadbe8722011-01-13 15:47:27 -08002356 return -EINVAL;
2357
Jingoo Han3dbb95f2013-09-11 14:20:25 -07002358 err = kstrtoul(buf, 10, &input);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002359 if (err)
Eric B Munson73ae31e2011-01-13 15:47:28 -08002360 return err;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002361
2362 spin_lock(&hugetlb_lock);
2363 h->nr_overcommit_huge_pages = input;
2364 spin_unlock(&hugetlb_lock);
2365
2366 return count;
2367}
2368HSTATE_ATTR(nr_overcommit_hugepages);
2369
2370static ssize_t free_hugepages_show(struct kobject *kobj,
2371 struct kobj_attribute *attr, char *buf)
2372{
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002373 struct hstate *h;
2374 unsigned long free_huge_pages;
2375 int nid;
2376
2377 h = kobj_to_hstate(kobj, &nid);
2378 if (nid == NUMA_NO_NODE)
2379 free_huge_pages = h->free_huge_pages;
2380 else
2381 free_huge_pages = h->free_huge_pages_node[nid];
2382
2383 return sprintf(buf, "%lu\n", free_huge_pages);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002384}
2385HSTATE_ATTR_RO(free_hugepages);
2386
2387static ssize_t resv_hugepages_show(struct kobject *kobj,
2388 struct kobj_attribute *attr, char *buf)
2389{
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002390 struct hstate *h = kobj_to_hstate(kobj, NULL);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002391 return sprintf(buf, "%lu\n", h->resv_huge_pages);
2392}
2393HSTATE_ATTR_RO(resv_hugepages);
2394
2395static ssize_t surplus_hugepages_show(struct kobject *kobj,
2396 struct kobj_attribute *attr, char *buf)
2397{
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002398 struct hstate *h;
2399 unsigned long surplus_huge_pages;
2400 int nid;
2401
2402 h = kobj_to_hstate(kobj, &nid);
2403 if (nid == NUMA_NO_NODE)
2404 surplus_huge_pages = h->surplus_huge_pages;
2405 else
2406 surplus_huge_pages = h->surplus_huge_pages_node[nid];
2407
2408 return sprintf(buf, "%lu\n", surplus_huge_pages);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002409}
2410HSTATE_ATTR_RO(surplus_hugepages);
2411
2412static struct attribute *hstate_attrs[] = {
2413 &nr_hugepages_attr.attr,
2414 &nr_overcommit_hugepages_attr.attr,
2415 &free_hugepages_attr.attr,
2416 &resv_hugepages_attr.attr,
2417 &surplus_hugepages_attr.attr,
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002418#ifdef CONFIG_NUMA
2419 &nr_hugepages_mempolicy_attr.attr,
2420#endif
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002421 NULL,
2422};
2423
2424static struct attribute_group hstate_attr_group = {
2425 .attrs = hstate_attrs,
2426};
2427
Jeff Mahoney094e9532010-02-02 13:44:14 -08002428static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
2429 struct kobject **hstate_kobjs,
2430 struct attribute_group *hstate_attr_group)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002431{
2432 int retval;
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07002433 int hi = hstate_index(h);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002434
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002435 hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
2436 if (!hstate_kobjs[hi])
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002437 return -ENOMEM;
2438
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002439 retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002440 if (retval)
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002441 kobject_put(hstate_kobjs[hi]);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002442
2443 return retval;
2444}
2445
2446static void __init hugetlb_sysfs_init(void)
2447{
2448 struct hstate *h;
2449 int err;
2450
2451 hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
2452 if (!hugepages_kobj)
2453 return;
2454
2455 for_each_hstate(h) {
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002456 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
2457 hstate_kobjs, &hstate_attr_group);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002458 if (err)
Andrew Mortonffb22af2013-02-22 16:32:08 -08002459 pr_err("Hugetlb: Unable to add hstate %s", h->name);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002460 }
2461}
2462
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002463#ifdef CONFIG_NUMA
2464
2465/*
2466 * node_hstate/s - associate per node hstate attributes, via their kobjects,
Kay Sievers10fbcf42011-12-21 14:48:43 -08002467 * with node devices in node_devices[] using a parallel array. The array
2468 * index of a node device or _hstate == node id.
2469 * This is here to avoid any static dependency of the node device driver, in
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002470 * the base kernel, on the hugetlb module.
2471 */
2472struct node_hstate {
2473 struct kobject *hugepages_kobj;
2474 struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
2475};
Alexander Kuleshovb4e289a2015-11-05 18:50:14 -08002476static struct node_hstate node_hstates[MAX_NUMNODES];
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002477
2478/*
Kay Sievers10fbcf42011-12-21 14:48:43 -08002479 * A subset of global hstate attributes for node devices
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002480 */
2481static struct attribute *per_node_hstate_attrs[] = {
2482 &nr_hugepages_attr.attr,
2483 &free_hugepages_attr.attr,
2484 &surplus_hugepages_attr.attr,
2485 NULL,
2486};
2487
2488static struct attribute_group per_node_hstate_attr_group = {
2489 .attrs = per_node_hstate_attrs,
2490};
2491
2492/*
Kay Sievers10fbcf42011-12-21 14:48:43 -08002493 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002494 * Returns node id via non-NULL nidp.
2495 */
2496static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
2497{
2498 int nid;
2499
2500 for (nid = 0; nid < nr_node_ids; nid++) {
2501 struct node_hstate *nhs = &node_hstates[nid];
2502 int i;
2503 for (i = 0; i < HUGE_MAX_HSTATE; i++)
2504 if (nhs->hstate_kobjs[i] == kobj) {
2505 if (nidp)
2506 *nidp = nid;
2507 return &hstates[i];
2508 }
2509 }
2510
2511 BUG();
2512 return NULL;
2513}
2514
2515/*
Kay Sievers10fbcf42011-12-21 14:48:43 -08002516 * Unregister hstate attributes from a single node device.
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002517 * No-op if no hstate attributes attached.
2518 */
Claudiu Ghioc3cd8b442013-03-04 12:46:15 +02002519static void hugetlb_unregister_node(struct node *node)
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002520{
2521 struct hstate *h;
Kay Sievers10fbcf42011-12-21 14:48:43 -08002522 struct node_hstate *nhs = &node_hstates[node->dev.id];
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002523
2524 if (!nhs->hugepages_kobj)
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -08002525 return; /* no hstate attributes */
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002526
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07002527 for_each_hstate(h) {
2528 int idx = hstate_index(h);
2529 if (nhs->hstate_kobjs[idx]) {
2530 kobject_put(nhs->hstate_kobjs[idx]);
2531 nhs->hstate_kobjs[idx] = NULL;
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002532 }
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07002533 }
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002534
2535 kobject_put(nhs->hugepages_kobj);
2536 nhs->hugepages_kobj = NULL;
2537}
2538
2539/*
Kay Sievers10fbcf42011-12-21 14:48:43 -08002540 * hugetlb module exit: unregister hstate attributes from node devices
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002541 * that have them.
2542 */
2543static void hugetlb_unregister_all_nodes(void)
2544{
2545 int nid;
2546
2547 /*
Kay Sievers10fbcf42011-12-21 14:48:43 -08002548 * disable node device registrations.
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002549 */
2550 register_hugetlbfs_with_node(NULL, NULL);
2551
2552 /*
2553 * remove hstate attributes from any nodes that have them.
2554 */
2555 for (nid = 0; nid < nr_node_ids; nid++)
Wen Congyang87327942012-12-11 16:00:56 -08002556 hugetlb_unregister_node(node_devices[nid]);
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002557}
2558
2559/*
Kay Sievers10fbcf42011-12-21 14:48:43 -08002560 * Register hstate attributes for a single node device.
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002561 * No-op if attributes already registered.
2562 */
Claudiu Ghioc3cd8b442013-03-04 12:46:15 +02002563static void hugetlb_register_node(struct node *node)
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002564{
2565 struct hstate *h;
Kay Sievers10fbcf42011-12-21 14:48:43 -08002566 struct node_hstate *nhs = &node_hstates[node->dev.id];
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002567 int err;
2568
2569 if (nhs->hugepages_kobj)
2570 return; /* already allocated */
2571
2572 nhs->hugepages_kobj = kobject_create_and_add("hugepages",
Kay Sievers10fbcf42011-12-21 14:48:43 -08002573 &node->dev.kobj);
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002574 if (!nhs->hugepages_kobj)
2575 return;
2576
2577 for_each_hstate(h) {
2578 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
2579 nhs->hstate_kobjs,
2580 &per_node_hstate_attr_group);
2581 if (err) {
Andrew Mortonffb22af2013-02-22 16:32:08 -08002582 pr_err("Hugetlb: Unable to add hstate %s for node %d\n",
2583 h->name, node->dev.id);
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002584 hugetlb_unregister_node(node);
2585 break;
2586 }
2587 }
2588}
2589
2590/*
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -08002591 * hugetlb init time: register hstate attributes for all registered node
Kay Sievers10fbcf42011-12-21 14:48:43 -08002592 * devices of nodes that have memory. All on-line nodes should have
2593 * registered their associated device by this time.
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002594 */
Luiz Capitulino7d9ca002014-12-12 16:55:24 -08002595static void __init hugetlb_register_all_nodes(void)
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002596{
2597 int nid;
2598
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08002599 for_each_node_state(nid, N_MEMORY) {
Wen Congyang87327942012-12-11 16:00:56 -08002600 struct node *node = node_devices[nid];
Kay Sievers10fbcf42011-12-21 14:48:43 -08002601 if (node->dev.id == nid)
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002602 hugetlb_register_node(node);
2603 }
2604
2605 /*
Kay Sievers10fbcf42011-12-21 14:48:43 -08002606 * Let the node device driver know we're here so it can
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002607 * [un]register hstate attributes on node hotplug.
2608 */
2609 register_hugetlbfs_with_node(hugetlb_register_node,
2610 hugetlb_unregister_node);
2611}
2612#else /* !CONFIG_NUMA */
2613
2614static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
2615{
2616 BUG();
2617 if (nidp)
2618 *nidp = -1;
2619 return NULL;
2620}
2621
2622static void hugetlb_unregister_all_nodes(void) { }
2623
2624static void hugetlb_register_all_nodes(void) { }
2625
2626#endif
2627
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002628static void __exit hugetlb_exit(void)
2629{
2630 struct hstate *h;
2631
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002632 hugetlb_unregister_all_nodes();
2633
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002634 for_each_hstate(h) {
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07002635 kobject_put(hstate_kobjs[hstate_index(h)]);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002636 }
2637
2638 kobject_put(hugepages_kobj);
Mike Kravetzc672c7f2015-09-08 15:01:35 -07002639 kfree(hugetlb_fault_mutex_table);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002640}
2641module_exit(hugetlb_exit);
2642
2643static int __init hugetlb_init(void)
2644{
Davidlohr Bueso8382d912014-04-03 14:47:31 -07002645 int i;
2646
Nishanth Aravamudan457c1b22014-05-06 12:50:00 -07002647 if (!hugepages_supported())
Benjamin Herrenschmidt0ef89d22008-07-31 00:07:30 -07002648 return 0;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002649
Nick Piggine11bfbf2008-07-23 21:27:52 -07002650 if (!size_to_hstate(default_hstate_size)) {
2651 default_hstate_size = HPAGE_SIZE;
2652 if (!size_to_hstate(default_hstate_size))
2653 hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002654 }
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07002655 default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size));
Nick Piggine11bfbf2008-07-23 21:27:52 -07002656 if (default_hstate_max_huge_pages)
2657 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002658
2659 hugetlb_init_hstates();
Andi Kleenaa888a72008-07-23 21:27:47 -07002660 gather_bootmem_prealloc();
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002661 report_hugepages();
2662
2663 hugetlb_sysfs_init();
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002664 hugetlb_register_all_nodes();
Jianguo Wu7179e7b2012-12-18 14:23:19 -08002665 hugetlb_cgroup_file_init();
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002666
Davidlohr Bueso8382d912014-04-03 14:47:31 -07002667#ifdef CONFIG_SMP
2668 num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus());
2669#else
2670 num_fault_mutexes = 1;
2671#endif
Mike Kravetzc672c7f2015-09-08 15:01:35 -07002672 hugetlb_fault_mutex_table =
Davidlohr Bueso8382d912014-04-03 14:47:31 -07002673 kmalloc(sizeof(struct mutex) * num_fault_mutexes, GFP_KERNEL);
Mike Kravetzc672c7f2015-09-08 15:01:35 -07002674 BUG_ON(!hugetlb_fault_mutex_table);
Davidlohr Bueso8382d912014-04-03 14:47:31 -07002675
2676 for (i = 0; i < num_fault_mutexes; i++)
Mike Kravetzc672c7f2015-09-08 15:01:35 -07002677 mutex_init(&hugetlb_fault_mutex_table[i]);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002678 return 0;
2679}
2680module_init(hugetlb_init);
2681
2682/* Should be called on processing a hugepagesz=... option */
Kirill A. Shutemovd00181b2015-11-06 16:29:57 -08002683void __init hugetlb_add_hstate(unsigned int order)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002684{
2685 struct hstate *h;
Andi Kleen8faa8b02008-07-23 21:27:48 -07002686 unsigned long i;
2687
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002688 if (size_to_hstate(PAGE_SIZE << order)) {
Andrew Mortonffb22af2013-02-22 16:32:08 -08002689 pr_warning("hugepagesz= specified twice, ignoring\n");
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002690 return;
2691 }
Aneesh Kumar K.V47d38342012-07-31 16:41:54 -07002692 BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002693 BUG_ON(order == 0);
Aneesh Kumar K.V47d38342012-07-31 16:41:54 -07002694 h = &hstates[hugetlb_max_hstate++];
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002695 h->order = order;
2696 h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
Andi Kleen8faa8b02008-07-23 21:27:48 -07002697 h->nr_huge_pages = 0;
2698 h->free_huge_pages = 0;
2699 for (i = 0; i < MAX_NUMNODES; ++i)
2700 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
Aneesh Kumar K.V0edaecf2012-07-31 16:42:07 -07002701 INIT_LIST_HEAD(&h->hugepage_activelist);
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08002702 h->next_nid_to_alloc = first_node(node_states[N_MEMORY]);
2703 h->next_nid_to_free = first_node(node_states[N_MEMORY]);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002704 snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
2705 huge_page_size(h)/1024);
Andi Kleen8faa8b02008-07-23 21:27:48 -07002706
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002707 parsed_hstate = h;
2708}
2709
Nick Piggine11bfbf2008-07-23 21:27:52 -07002710static int __init hugetlb_nrpages_setup(char *s)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002711{
2712 unsigned long *mhp;
Andi Kleen8faa8b02008-07-23 21:27:48 -07002713 static unsigned long *last_mhp;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002714
2715 /*
Aneesh Kumar K.V47d38342012-07-31 16:41:54 -07002716 * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet,
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002717 * so this hugepages= parameter goes to the "default hstate".
2718 */
Aneesh Kumar K.V47d38342012-07-31 16:41:54 -07002719 if (!hugetlb_max_hstate)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002720 mhp = &default_hstate_max_huge_pages;
2721 else
2722 mhp = &parsed_hstate->max_huge_pages;
2723
Andi Kleen8faa8b02008-07-23 21:27:48 -07002724 if (mhp == last_mhp) {
Andrew Mortonffb22af2013-02-22 16:32:08 -08002725 pr_warning("hugepages= specified twice without "
2726 "interleaving hugepagesz=, ignoring\n");
Andi Kleen8faa8b02008-07-23 21:27:48 -07002727 return 1;
2728 }
2729
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002730 if (sscanf(s, "%lu", mhp) <= 0)
2731 *mhp = 0;
2732
Andi Kleen8faa8b02008-07-23 21:27:48 -07002733 /*
2734 * Global state is always initialized later in hugetlb_init.
2735 * But we need to allocate >= MAX_ORDER hstates here early to still
2736 * use the bootmem allocator.
2737 */
Aneesh Kumar K.V47d38342012-07-31 16:41:54 -07002738 if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER)
Andi Kleen8faa8b02008-07-23 21:27:48 -07002739 hugetlb_hstate_alloc_pages(parsed_hstate);
2740
2741 last_mhp = mhp;
2742
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002743 return 1;
2744}
Nick Piggine11bfbf2008-07-23 21:27:52 -07002745__setup("hugepages=", hugetlb_nrpages_setup);
2746
2747static int __init hugetlb_default_setup(char *s)
2748{
2749 default_hstate_size = memparse(s, &s);
2750 return 1;
2751}
2752__setup("default_hugepagesz=", hugetlb_default_setup);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002753
Nishanth Aravamudan8a213462008-07-25 19:44:37 -07002754static unsigned int cpuset_mems_nr(unsigned int *array)
2755{
2756 int node;
2757 unsigned int nr = 0;
2758
2759 for_each_node_mask(node, cpuset_current_mems_allowed)
2760 nr += array[node];
2761
2762 return nr;
2763}
2764
2765#ifdef CONFIG_SYSCTL
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002766static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
2767 struct ctl_table *table, int write,
2768 void __user *buffer, size_t *length, loff_t *ppos)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002769{
Andi Kleene5ff2152008-07-23 21:27:42 -07002770 struct hstate *h = &default_hstate;
David Rientjes238d3c12014-08-06 16:06:51 -07002771 unsigned long tmp = h->max_huge_pages;
Michal Hocko08d4a242011-01-13 15:47:26 -08002772 int ret;
Andi Kleene5ff2152008-07-23 21:27:42 -07002773
Nishanth Aravamudan457c1b22014-05-06 12:50:00 -07002774 if (!hugepages_supported())
2775 return -ENOTSUPP;
2776
Andi Kleene5ff2152008-07-23 21:27:42 -07002777 table->data = &tmp;
2778 table->maxlen = sizeof(unsigned long);
Michal Hocko08d4a242011-01-13 15:47:26 -08002779 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2780 if (ret)
2781 goto out;
Andi Kleene5ff2152008-07-23 21:27:42 -07002782
David Rientjes238d3c12014-08-06 16:06:51 -07002783 if (write)
2784 ret = __nr_hugepages_store_common(obey_mempolicy, h,
2785 NUMA_NO_NODE, tmp, *length);
Michal Hocko08d4a242011-01-13 15:47:26 -08002786out:
2787 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002788}
Mel Gorman396faf02007-07-17 04:03:13 -07002789
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002790int hugetlb_sysctl_handler(struct ctl_table *table, int write,
2791 void __user *buffer, size_t *length, loff_t *ppos)
2792{
2793
2794 return hugetlb_sysctl_handler_common(false, table, write,
2795 buffer, length, ppos);
2796}
2797
2798#ifdef CONFIG_NUMA
2799int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
2800 void __user *buffer, size_t *length, loff_t *ppos)
2801{
2802 return hugetlb_sysctl_handler_common(true, table, write,
2803 buffer, length, ppos);
2804}
2805#endif /* CONFIG_NUMA */
2806
Nishanth Aravamudana3d0c6a2008-02-08 04:18:18 -08002807int hugetlb_overcommit_handler(struct ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07002808 void __user *buffer,
Nishanth Aravamudana3d0c6a2008-02-08 04:18:18 -08002809 size_t *length, loff_t *ppos)
2810{
Andi Kleena5516432008-07-23 21:27:41 -07002811 struct hstate *h = &default_hstate;
Andi Kleene5ff2152008-07-23 21:27:42 -07002812 unsigned long tmp;
Michal Hocko08d4a242011-01-13 15:47:26 -08002813 int ret;
Andi Kleene5ff2152008-07-23 21:27:42 -07002814
Nishanth Aravamudan457c1b22014-05-06 12:50:00 -07002815 if (!hugepages_supported())
2816 return -ENOTSUPP;
2817
Petr Holasekc033a932011-03-22 16:33:05 -07002818 tmp = h->nr_overcommit_huge_pages;
Andi Kleene5ff2152008-07-23 21:27:42 -07002819
Luiz Capitulinobae7f4a2014-06-04 16:07:08 -07002820 if (write && hstate_is_gigantic(h))
Eric B Munsonadbe8722011-01-13 15:47:27 -08002821 return -EINVAL;
2822
Andi Kleene5ff2152008-07-23 21:27:42 -07002823 table->data = &tmp;
2824 table->maxlen = sizeof(unsigned long);
Michal Hocko08d4a242011-01-13 15:47:26 -08002825 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2826 if (ret)
2827 goto out;
Andi Kleene5ff2152008-07-23 21:27:42 -07002828
2829 if (write) {
2830 spin_lock(&hugetlb_lock);
2831 h->nr_overcommit_huge_pages = tmp;
2832 spin_unlock(&hugetlb_lock);
2833 }
Michal Hocko08d4a242011-01-13 15:47:26 -08002834out:
2835 return ret;
Nishanth Aravamudana3d0c6a2008-02-08 04:18:18 -08002836}
2837
Linus Torvalds1da177e2005-04-16 15:20:36 -07002838#endif /* CONFIG_SYSCTL */
2839
Alexey Dobriyane1759c22008-10-15 23:50:22 +04002840void hugetlb_report_meminfo(struct seq_file *m)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002841{
Andi Kleena5516432008-07-23 21:27:41 -07002842 struct hstate *h = &default_hstate;
Nishanth Aravamudan457c1b22014-05-06 12:50:00 -07002843 if (!hugepages_supported())
2844 return;
Alexey Dobriyane1759c22008-10-15 23:50:22 +04002845 seq_printf(m,
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002846 "HugePages_Total: %5lu\n"
2847 "HugePages_Free: %5lu\n"
2848 "HugePages_Rsvd: %5lu\n"
2849 "HugePages_Surp: %5lu\n"
2850 "Hugepagesize: %8lu kB\n",
Andi Kleena5516432008-07-23 21:27:41 -07002851 h->nr_huge_pages,
2852 h->free_huge_pages,
2853 h->resv_huge_pages,
2854 h->surplus_huge_pages,
2855 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002856}
2857
2858int hugetlb_report_node_meminfo(int nid, char *buf)
2859{
Andi Kleena5516432008-07-23 21:27:41 -07002860 struct hstate *h = &default_hstate;
Nishanth Aravamudan457c1b22014-05-06 12:50:00 -07002861 if (!hugepages_supported())
2862 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002863 return sprintf(buf,
2864 "Node %d HugePages_Total: %5u\n"
Nishanth Aravamudana1de0912008-03-26 14:37:53 -07002865 "Node %d HugePages_Free: %5u\n"
2866 "Node %d HugePages_Surp: %5u\n",
Andi Kleena5516432008-07-23 21:27:41 -07002867 nid, h->nr_huge_pages_node[nid],
2868 nid, h->free_huge_pages_node[nid],
2869 nid, h->surplus_huge_pages_node[nid]);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002870}
2871
David Rientjes949f7ec2013-04-29 15:07:48 -07002872void hugetlb_show_meminfo(void)
2873{
2874 struct hstate *h;
2875 int nid;
2876
Nishanth Aravamudan457c1b22014-05-06 12:50:00 -07002877 if (!hugepages_supported())
2878 return;
2879
David Rientjes949f7ec2013-04-29 15:07:48 -07002880 for_each_node_state(nid, N_MEMORY)
2881 for_each_hstate(h)
2882 pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
2883 nid,
2884 h->nr_huge_pages_node[nid],
2885 h->free_huge_pages_node[nid],
2886 h->surplus_huge_pages_node[nid],
2887 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
2888}
2889
Naoya Horiguchi5d317b22015-11-05 18:47:14 -08002890void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm)
2891{
2892 seq_printf(m, "HugetlbPages:\t%8lu kB\n",
2893 atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10));
2894}
2895
Linus Torvalds1da177e2005-04-16 15:20:36 -07002896/* Return the number pages of memory we physically have, in PAGE_SIZE units. */
2897unsigned long hugetlb_total_pages(void)
2898{
Wanpeng Lid0028582013-03-22 15:04:40 -07002899 struct hstate *h;
2900 unsigned long nr_total_pages = 0;
2901
2902 for_each_hstate(h)
2903 nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h);
2904 return nr_total_pages;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002905}
Linus Torvalds1da177e2005-04-16 15:20:36 -07002906
Andi Kleena5516432008-07-23 21:27:41 -07002907static int hugetlb_acct_memory(struct hstate *h, long delta)
Mel Gormanfc1b8a72008-07-23 21:27:22 -07002908{
2909 int ret = -ENOMEM;
2910
2911 spin_lock(&hugetlb_lock);
2912 /*
2913 * When cpuset is configured, it breaks the strict hugetlb page
2914 * reservation as the accounting is done on a global variable. Such
2915 * reservation is completely rubbish in the presence of cpuset because
2916 * the reservation is not checked against page availability for the
2917 * current cpuset. Application can still potentially OOM'ed by kernel
2918 * with lack of free htlb page in cpuset that the task is in.
2919 * Attempt to enforce strict accounting with cpuset is almost
2920 * impossible (or too ugly) because cpuset is too fluid that
2921 * task or memory node can be dynamically moved between cpusets.
2922 *
2923 * The change of semantics for shared hugetlb mapping with cpuset is
2924 * undesirable. However, in order to preserve some of the semantics,
2925 * we fall back to check against current free page availability as
2926 * a best attempt and hopefully to minimize the impact of changing
2927 * semantics that cpuset has.
2928 */
2929 if (delta > 0) {
Andi Kleena5516432008-07-23 21:27:41 -07002930 if (gather_surplus_pages(h, delta) < 0)
Mel Gormanfc1b8a72008-07-23 21:27:22 -07002931 goto out;
2932
Andi Kleena5516432008-07-23 21:27:41 -07002933 if (delta > cpuset_mems_nr(h->free_huge_pages_node)) {
2934 return_unused_surplus_pages(h, delta);
Mel Gormanfc1b8a72008-07-23 21:27:22 -07002935 goto out;
2936 }
2937 }
2938
2939 ret = 0;
2940 if (delta < 0)
Andi Kleena5516432008-07-23 21:27:41 -07002941 return_unused_surplus_pages(h, (unsigned long) -delta);
Mel Gormanfc1b8a72008-07-23 21:27:22 -07002942
2943out:
2944 spin_unlock(&hugetlb_lock);
2945 return ret;
2946}
2947
Andy Whitcroft84afd992008-07-23 21:27:32 -07002948static void hugetlb_vm_op_open(struct vm_area_struct *vma)
2949{
Joonsoo Kimf522c3a2013-09-11 14:21:53 -07002950 struct resv_map *resv = vma_resv_map(vma);
Andy Whitcroft84afd992008-07-23 21:27:32 -07002951
2952 /*
2953 * This new VMA should share its siblings reservation map if present.
2954 * The VMA will only ever have a valid reservation map pointer where
2955 * it is being copied for another still existing VMA. As that VMA
Lucas De Marchi25985ed2011-03-30 22:57:33 -03002956 * has a reference to the reservation map it cannot disappear until
Andy Whitcroft84afd992008-07-23 21:27:32 -07002957 * after this open call completes. It is therefore safe to take a
2958 * new reference here without additional locking.
2959 */
Joonsoo Kim4e35f482014-04-03 14:47:30 -07002960 if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
Joonsoo Kimf522c3a2013-09-11 14:21:53 -07002961 kref_get(&resv->refs);
Andy Whitcroft84afd992008-07-23 21:27:32 -07002962}
2963
Mel Gormana1e78772008-07-23 21:27:23 -07002964static void hugetlb_vm_op_close(struct vm_area_struct *vma)
2965{
Andi Kleena5516432008-07-23 21:27:41 -07002966 struct hstate *h = hstate_vma(vma);
Joonsoo Kimf522c3a2013-09-11 14:21:53 -07002967 struct resv_map *resv = vma_resv_map(vma);
David Gibson90481622012-03-21 16:34:12 -07002968 struct hugepage_subpool *spool = subpool_vma(vma);
Joonsoo Kim4e35f482014-04-03 14:47:30 -07002969 unsigned long reserve, start, end;
Mike Kravetz1c5ecae2015-04-15 16:13:39 -07002970 long gbl_reserve;
Andy Whitcroft84afd992008-07-23 21:27:32 -07002971
Joonsoo Kim4e35f482014-04-03 14:47:30 -07002972 if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER))
2973 return;
Andy Whitcroft84afd992008-07-23 21:27:32 -07002974
Joonsoo Kim4e35f482014-04-03 14:47:30 -07002975 start = vma_hugecache_offset(h, vma, vma->vm_start);
2976 end = vma_hugecache_offset(h, vma, vma->vm_end);
Andy Whitcroft84afd992008-07-23 21:27:32 -07002977
Joonsoo Kim4e35f482014-04-03 14:47:30 -07002978 reserve = (end - start) - region_count(resv, start, end);
Andy Whitcroft84afd992008-07-23 21:27:32 -07002979
Joonsoo Kim4e35f482014-04-03 14:47:30 -07002980 kref_put(&resv->refs, resv_map_release);
2981
2982 if (reserve) {
Mike Kravetz1c5ecae2015-04-15 16:13:39 -07002983 /*
2984 * Decrement reserve counts. The global reserve count may be
2985 * adjusted if the subpool has a minimum size.
2986 */
2987 gbl_reserve = hugepage_subpool_put_pages(spool, reserve);
2988 hugetlb_acct_memory(h, -gbl_reserve);
Andy Whitcroft84afd992008-07-23 21:27:32 -07002989 }
Mel Gormana1e78772008-07-23 21:27:23 -07002990}
2991
Linus Torvalds1da177e2005-04-16 15:20:36 -07002992/*
2993 * We cannot handle pagefaults against hugetlb pages at all. They cause
2994 * handle_mm_fault() to try to instantiate regular-sized pages in the
2995 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
2996 * this far.
2997 */
Nick Piggind0217ac2007-07-19 01:47:03 -07002998static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002999{
3000 BUG();
Nick Piggind0217ac2007-07-19 01:47:03 -07003001 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003002}
3003
Alexey Dobriyanf0f37e22009-09-27 22:29:37 +04003004const struct vm_operations_struct hugetlb_vm_ops = {
Nick Piggind0217ac2007-07-19 01:47:03 -07003005 .fault = hugetlb_vm_op_fault,
Andy Whitcroft84afd992008-07-23 21:27:32 -07003006 .open = hugetlb_vm_op_open,
Mel Gormana1e78772008-07-23 21:27:23 -07003007 .close = hugetlb_vm_op_close,
Linus Torvalds1da177e2005-04-16 15:20:36 -07003008};
3009
David Gibson1e8f8892006-01-06 00:10:44 -08003010static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
3011 int writable)
David Gibson63551ae2005-06-21 17:14:44 -07003012{
3013 pte_t entry;
3014
David Gibson1e8f8892006-01-06 00:10:44 -08003015 if (writable) {
Gerald Schaefer106c9922013-04-29 15:07:23 -07003016 entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page,
3017 vma->vm_page_prot)));
David Gibson63551ae2005-06-21 17:14:44 -07003018 } else {
Gerald Schaefer106c9922013-04-29 15:07:23 -07003019 entry = huge_pte_wrprotect(mk_huge_pte(page,
3020 vma->vm_page_prot));
David Gibson63551ae2005-06-21 17:14:44 -07003021 }
3022 entry = pte_mkyoung(entry);
3023 entry = pte_mkhuge(entry);
Chris Metcalfd9ed9fa2012-04-01 14:01:34 -04003024 entry = arch_make_huge_pte(entry, vma, page, writable);
David Gibson63551ae2005-06-21 17:14:44 -07003025
3026 return entry;
3027}
3028
David Gibson1e8f8892006-01-06 00:10:44 -08003029static void set_huge_ptep_writable(struct vm_area_struct *vma,
3030 unsigned long address, pte_t *ptep)
3031{
3032 pte_t entry;
3033
Gerald Schaefer106c9922013-04-29 15:07:23 -07003034 entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep)));
Chris Forbes32f84522011-07-25 17:12:14 -07003035 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
Russell King4b3073e2009-12-18 16:40:18 +00003036 update_mmu_cache(vma, address, ptep);
David Gibson1e8f8892006-01-06 00:10:44 -08003037}
3038
Naoya Horiguchi4a705fe2014-06-23 13:22:03 -07003039static int is_hugetlb_entry_migration(pte_t pte)
3040{
3041 swp_entry_t swp;
3042
3043 if (huge_pte_none(pte) || pte_present(pte))
3044 return 0;
3045 swp = pte_to_swp_entry(pte);
3046 if (non_swap_entry(swp) && is_migration_entry(swp))
3047 return 1;
3048 else
3049 return 0;
3050}
3051
3052static int is_hugetlb_entry_hwpoisoned(pte_t pte)
3053{
3054 swp_entry_t swp;
3055
3056 if (huge_pte_none(pte) || pte_present(pte))
3057 return 0;
3058 swp = pte_to_swp_entry(pte);
3059 if (non_swap_entry(swp) && is_hwpoison_entry(swp))
3060 return 1;
3061 else
3062 return 0;
3063}
David Gibson1e8f8892006-01-06 00:10:44 -08003064
David Gibson63551ae2005-06-21 17:14:44 -07003065int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
3066 struct vm_area_struct *vma)
3067{
3068 pte_t *src_pte, *dst_pte, entry;
3069 struct page *ptepage;
Hugh Dickins1c598272005-10-19 21:23:43 -07003070 unsigned long addr;
David Gibson1e8f8892006-01-06 00:10:44 -08003071 int cow;
Andi Kleena5516432008-07-23 21:27:41 -07003072 struct hstate *h = hstate_vma(vma);
3073 unsigned long sz = huge_page_size(h);
Andreas Sandberge8569dd2014-01-21 15:49:09 -08003074 unsigned long mmun_start; /* For mmu_notifiers */
3075 unsigned long mmun_end; /* For mmu_notifiers */
3076 int ret = 0;
David Gibson1e8f8892006-01-06 00:10:44 -08003077
3078 cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
David Gibson63551ae2005-06-21 17:14:44 -07003079
Andreas Sandberge8569dd2014-01-21 15:49:09 -08003080 mmun_start = vma->vm_start;
3081 mmun_end = vma->vm_end;
3082 if (cow)
3083 mmu_notifier_invalidate_range_start(src, mmun_start, mmun_end);
3084
Andi Kleena5516432008-07-23 21:27:41 -07003085 for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003086 spinlock_t *src_ptl, *dst_ptl;
Hugh Dickinsc74df322005-10-29 18:16:23 -07003087 src_pte = huge_pte_offset(src, addr);
3088 if (!src_pte)
3089 continue;
Andi Kleena5516432008-07-23 21:27:41 -07003090 dst_pte = huge_pte_alloc(dst, addr, sz);
Andreas Sandberge8569dd2014-01-21 15:49:09 -08003091 if (!dst_pte) {
3092 ret = -ENOMEM;
3093 break;
3094 }
Larry Woodmanc5c99422008-01-24 05:49:25 -08003095
3096 /* If the pagetables are shared don't copy or take references */
3097 if (dst_pte == src_pte)
3098 continue;
3099
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003100 dst_ptl = huge_pte_lock(h, dst, dst_pte);
3101 src_ptl = huge_pte_lockptr(h, src, src_pte);
3102 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
Naoya Horiguchi4a705fe2014-06-23 13:22:03 -07003103 entry = huge_ptep_get(src_pte);
3104 if (huge_pte_none(entry)) { /* skip none entry */
3105 ;
3106 } else if (unlikely(is_hugetlb_entry_migration(entry) ||
3107 is_hugetlb_entry_hwpoisoned(entry))) {
3108 swp_entry_t swp_entry = pte_to_swp_entry(entry);
3109
3110 if (is_write_migration_entry(swp_entry) && cow) {
3111 /*
3112 * COW mappings require pages in both
3113 * parent and child to be set to read.
3114 */
3115 make_migration_entry_read(&swp_entry);
3116 entry = swp_entry_to_pte(swp_entry);
3117 set_huge_pte_at(src, addr, src_pte, entry);
3118 }
3119 set_huge_pte_at(dst, addr, dst_pte, entry);
3120 } else {
Joerg Roedel34ee6452014-11-13 13:46:09 +11003121 if (cow) {
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07003122 huge_ptep_set_wrprotect(src, addr, src_pte);
Joerg Roedel34ee6452014-11-13 13:46:09 +11003123 mmu_notifier_invalidate_range(src, mmun_start,
3124 mmun_end);
3125 }
Naoya Horiguchi0253d632014-07-23 14:00:19 -07003126 entry = huge_ptep_get(src_pte);
Hugh Dickins1c598272005-10-19 21:23:43 -07003127 ptepage = pte_page(entry);
3128 get_page(ptepage);
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09003129 page_dup_rmap(ptepage);
Hugh Dickins1c598272005-10-19 21:23:43 -07003130 set_huge_pte_at(dst, addr, dst_pte, entry);
Naoya Horiguchi5d317b22015-11-05 18:47:14 -08003131 hugetlb_count_add(pages_per_huge_page(h), dst);
Hugh Dickins1c598272005-10-19 21:23:43 -07003132 }
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003133 spin_unlock(src_ptl);
3134 spin_unlock(dst_ptl);
David Gibson63551ae2005-06-21 17:14:44 -07003135 }
David Gibson63551ae2005-06-21 17:14:44 -07003136
Andreas Sandberge8569dd2014-01-21 15:49:09 -08003137 if (cow)
3138 mmu_notifier_invalidate_range_end(src, mmun_start, mmun_end);
3139
3140 return ret;
David Gibson63551ae2005-06-21 17:14:44 -07003141}
3142
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07003143void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
3144 unsigned long start, unsigned long end,
3145 struct page *ref_page)
David Gibson63551ae2005-06-21 17:14:44 -07003146{
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07003147 int force_flush = 0;
David Gibson63551ae2005-06-21 17:14:44 -07003148 struct mm_struct *mm = vma->vm_mm;
3149 unsigned long address;
David Gibsonc7546f82005-08-05 11:59:35 -07003150 pte_t *ptep;
David Gibson63551ae2005-06-21 17:14:44 -07003151 pte_t pte;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003152 spinlock_t *ptl;
David Gibson63551ae2005-06-21 17:14:44 -07003153 struct page *page;
Andi Kleena5516432008-07-23 21:27:41 -07003154 struct hstate *h = hstate_vma(vma);
3155 unsigned long sz = huge_page_size(h);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07003156 const unsigned long mmun_start = start; /* For mmu_notifiers */
3157 const unsigned long mmun_end = end; /* For mmu_notifiers */
Andi Kleena5516432008-07-23 21:27:41 -07003158
David Gibson63551ae2005-06-21 17:14:44 -07003159 WARN_ON(!is_vm_hugetlb_page(vma));
Andi Kleena5516432008-07-23 21:27:41 -07003160 BUG_ON(start & ~huge_page_mask(h));
3161 BUG_ON(end & ~huge_page_mask(h));
David Gibson63551ae2005-06-21 17:14:44 -07003162
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07003163 tlb_start_vma(tlb, vma);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07003164 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
Hillf Danton569f48b2014-12-10 15:44:41 -08003165 address = start;
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07003166again:
Hillf Danton569f48b2014-12-10 15:44:41 -08003167 for (; address < end; address += sz) {
David Gibsonc7546f82005-08-05 11:59:35 -07003168 ptep = huge_pte_offset(mm, address);
Adam Litke4c887262005-10-29 18:16:46 -07003169 if (!ptep)
David Gibsonc7546f82005-08-05 11:59:35 -07003170 continue;
3171
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003172 ptl = huge_pte_lock(h, mm, ptep);
Chen, Kenneth W39dde652006-12-06 20:32:03 -08003173 if (huge_pmd_unshare(mm, &address, ptep))
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003174 goto unlock;
Chen, Kenneth W39dde652006-12-06 20:32:03 -08003175
Hillf Danton66293262012-03-23 15:01:48 -07003176 pte = huge_ptep_get(ptep);
3177 if (huge_pte_none(pte))
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003178 goto unlock;
Hillf Danton66293262012-03-23 15:01:48 -07003179
3180 /*
Naoya Horiguchi9fbc1f62015-02-11 15:25:32 -08003181 * Migrating hugepage or HWPoisoned hugepage is already
3182 * unmapped and its refcount is dropped, so just clear pte here.
Hillf Danton66293262012-03-23 15:01:48 -07003183 */
Naoya Horiguchi9fbc1f62015-02-11 15:25:32 -08003184 if (unlikely(!pte_present(pte))) {
Gerald Schaefer106c9922013-04-29 15:07:23 -07003185 huge_pte_clear(mm, address, ptep);
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003186 goto unlock;
Naoya Horiguchi8c4894c2012-12-12 13:52:28 -08003187 }
Hillf Danton66293262012-03-23 15:01:48 -07003188
3189 page = pte_page(pte);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003190 /*
3191 * If a reference page is supplied, it is because a specific
3192 * page is being unmapped, not a range. Ensure the page we
3193 * are about to unmap is the actual page of interest.
3194 */
3195 if (ref_page) {
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003196 if (page != ref_page)
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003197 goto unlock;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003198
3199 /*
3200 * Mark the VMA as having unmapped its page so that
3201 * future faults in this VMA will fail rather than
3202 * looking like data was lost
3203 */
3204 set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
3205 }
3206
David Gibsonc7546f82005-08-05 11:59:35 -07003207 pte = huge_ptep_get_and_clear(mm, address, ptep);
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07003208 tlb_remove_tlb_entry(tlb, ptep, address);
Gerald Schaefer106c9922013-04-29 15:07:23 -07003209 if (huge_pte_dirty(pte))
Ken Chen6649a382007-02-08 14:20:27 -08003210 set_page_dirty(page);
Hillf Danton9e811302012-03-21 16:34:03 -07003211
Naoya Horiguchi5d317b22015-11-05 18:47:14 -08003212 hugetlb_count_sub(pages_per_huge_page(h), mm);
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07003213 page_remove_rmap(page);
3214 force_flush = !__tlb_remove_page(tlb, page);
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003215 if (force_flush) {
Hillf Danton569f48b2014-12-10 15:44:41 -08003216 address += sz;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003217 spin_unlock(ptl);
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07003218 break;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003219 }
Hillf Danton9e811302012-03-21 16:34:03 -07003220 /* Bail out after unmapping reference page if supplied */
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003221 if (ref_page) {
3222 spin_unlock(ptl);
Hillf Danton9e811302012-03-21 16:34:03 -07003223 break;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003224 }
3225unlock:
3226 spin_unlock(ptl);
David Gibson63551ae2005-06-21 17:14:44 -07003227 }
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07003228 /*
3229 * mmu_gather ran out of room to batch pages, we break out of
3230 * the PTE lock to avoid doing the potential expensive TLB invalidate
3231 * and page-free while holding it.
3232 */
3233 if (force_flush) {
3234 force_flush = 0;
3235 tlb_flush_mmu(tlb);
3236 if (address < end && !ref_page)
3237 goto again;
Chen, Kenneth Wfe1668a2006-10-04 02:15:24 -07003238 }
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07003239 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07003240 tlb_end_vma(tlb, vma);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003241}
David Gibson63551ae2005-06-21 17:14:44 -07003242
Mel Gormand8333522012-07-31 16:46:20 -07003243void __unmap_hugepage_range_final(struct mmu_gather *tlb,
3244 struct vm_area_struct *vma, unsigned long start,
3245 unsigned long end, struct page *ref_page)
3246{
3247 __unmap_hugepage_range(tlb, vma, start, end, ref_page);
3248
3249 /*
3250 * Clear this flag so that x86's huge_pmd_share page_table_shareable
3251 * test will fail on a vma being torn down, and not grab a page table
3252 * on its way out. We're lucky that the flag has such an appropriate
3253 * name, and can in fact be safely cleared here. We could clear it
3254 * before the __unmap_hugepage_range above, but all that's necessary
Davidlohr Buesoc8c06ef2014-12-12 16:54:24 -08003255 * is to clear it before releasing the i_mmap_rwsem. This works
Mel Gormand8333522012-07-31 16:46:20 -07003256 * because in the context this is called, the VMA is about to be
Davidlohr Buesoc8c06ef2014-12-12 16:54:24 -08003257 * destroyed and the i_mmap_rwsem is held.
Mel Gormand8333522012-07-31 16:46:20 -07003258 */
3259 vma->vm_flags &= ~VM_MAYSHARE;
3260}
3261
Chen, Kenneth W502717f2006-10-11 01:20:46 -07003262void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003263 unsigned long end, struct page *ref_page)
Chen, Kenneth W502717f2006-10-11 01:20:46 -07003264{
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07003265 struct mm_struct *mm;
3266 struct mmu_gather tlb;
3267
3268 mm = vma->vm_mm;
3269
Linus Torvalds2b047252013-08-15 11:42:25 -07003270 tlb_gather_mmu(&tlb, mm, start, end);
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07003271 __unmap_hugepage_range(&tlb, vma, start, end, ref_page);
3272 tlb_finish_mmu(&tlb, start, end);
Chen, Kenneth W502717f2006-10-11 01:20:46 -07003273}
3274
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003275/*
3276 * This is called when the original mapper is failing to COW a MAP_PRIVATE
3277 * mappping it owns the reserve page for. The intention is to unmap the page
3278 * from other VMAs and let the children be SIGKILLed if they are faulting the
3279 * same region.
3280 */
Davidlohr Bueso2f4612a2014-08-06 16:06:45 -07003281static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
3282 struct page *page, unsigned long address)
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003283{
Adam Litke75266742008-11-12 13:24:56 -08003284 struct hstate *h = hstate_vma(vma);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003285 struct vm_area_struct *iter_vma;
3286 struct address_space *mapping;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003287 pgoff_t pgoff;
3288
3289 /*
3290 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
3291 * from page cache lookup which is in HPAGE_SIZE units.
3292 */
Adam Litke75266742008-11-12 13:24:56 -08003293 address = address & huge_page_mask(h);
Michal Hocko36e4f202012-10-08 16:33:31 -07003294 pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) +
3295 vma->vm_pgoff;
Al Viro496ad9a2013-01-23 17:07:38 -05003296 mapping = file_inode(vma->vm_file)->i_mapping;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003297
Mel Gorman4eb2b1d2009-12-14 17:59:53 -08003298 /*
3299 * Take the mapping lock for the duration of the table walk. As
3300 * this mapping should be shared between all the VMAs,
3301 * __unmap_hugepage_range() is called as the lock is already held
3302 */
Davidlohr Bueso83cde9e2014-12-12 16:54:21 -08003303 i_mmap_lock_write(mapping);
Michel Lespinasse6b2dbba2012-10-08 16:31:25 -07003304 vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) {
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003305 /* Do not unmap the current VMA */
3306 if (iter_vma == vma)
3307 continue;
3308
3309 /*
Mel Gorman2f84a892015-10-01 15:36:57 -07003310 * Shared VMAs have their own reserves and do not affect
3311 * MAP_PRIVATE accounting but it is possible that a shared
3312 * VMA is using the same page so check and skip such VMAs.
3313 */
3314 if (iter_vma->vm_flags & VM_MAYSHARE)
3315 continue;
3316
3317 /*
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003318 * Unmap the page from other VMAs without their own reserves.
3319 * They get marked to be SIGKILLed if they fault in these
3320 * areas. This is because a future no-page fault on this VMA
3321 * could insert a zeroed page instead of the data existing
3322 * from the time of fork. This would look like data corruption
3323 */
3324 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07003325 unmap_hugepage_range(iter_vma, address,
3326 address + huge_page_size(h), page);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003327 }
Davidlohr Bueso83cde9e2014-12-12 16:54:21 -08003328 i_mmap_unlock_write(mapping);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003329}
3330
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09003331/*
3332 * Hugetlb_cow() should be called with page lock of the original hugepage held.
Michal Hockoef009b22012-01-10 15:07:21 -08003333 * Called with hugetlb_instantiation_mutex held and pte_page locked so we
3334 * cannot race with other handlers or page migration.
3335 * Keep the pte_same checks anyway to make transition from the mutex easier.
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09003336 */
David Gibson1e8f8892006-01-06 00:10:44 -08003337static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003338 unsigned long address, pte_t *ptep, pte_t pte,
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003339 struct page *pagecache_page, spinlock_t *ptl)
David Gibson1e8f8892006-01-06 00:10:44 -08003340{
Andi Kleena5516432008-07-23 21:27:41 -07003341 struct hstate *h = hstate_vma(vma);
David Gibson1e8f8892006-01-06 00:10:44 -08003342 struct page *old_page, *new_page;
Davidlohr Buesoad4404a2014-08-06 16:06:47 -07003343 int ret = 0, outside_reserve = 0;
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07003344 unsigned long mmun_start; /* For mmu_notifiers */
3345 unsigned long mmun_end; /* For mmu_notifiers */
David Gibson1e8f8892006-01-06 00:10:44 -08003346
3347 old_page = pte_page(pte);
3348
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003349retry_avoidcopy:
David Gibson1e8f8892006-01-06 00:10:44 -08003350 /* If no-one else is actually using this page, avoid the copy
3351 * and just make the page writable */
Joonsoo Kim37a21402013-09-11 14:21:04 -07003352 if (page_mapcount(old_page) == 1 && PageAnon(old_page)) {
3353 page_move_anon_rmap(old_page, vma, address);
David Gibson1e8f8892006-01-06 00:10:44 -08003354 set_huge_ptep_writable(vma, address, ptep);
Nick Piggin83c54072007-07-19 01:47:05 -07003355 return 0;
David Gibson1e8f8892006-01-06 00:10:44 -08003356 }
3357
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003358 /*
3359 * If the process that created a MAP_PRIVATE mapping is about to
3360 * perform a COW due to a shared page count, attempt to satisfy
3361 * the allocation without using the existing reserves. The pagecache
3362 * page is used to determine if the reserve at this address was
3363 * consumed or not. If reserves were used, a partial faulted mapping
3364 * at the time of fork() could consume its reserves on COW instead
3365 * of the full address range.
3366 */
Joonsoo Kim5944d012013-09-11 14:21:55 -07003367 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003368 old_page != pagecache_page)
3369 outside_reserve = 1;
3370
David Gibson1e8f8892006-01-06 00:10:44 -08003371 page_cache_get(old_page);
Larry Woodmanb76c8cf2009-12-14 17:59:37 -08003372
Davidlohr Buesoad4404a2014-08-06 16:06:47 -07003373 /*
3374 * Drop page table lock as buddy allocator may be called. It will
3375 * be acquired again before returning to the caller, as expected.
3376 */
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003377 spin_unlock(ptl);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003378 new_page = alloc_huge_page(vma, address, outside_reserve);
David Gibson1e8f8892006-01-06 00:10:44 -08003379
Adam Litke2fc39ce2007-11-14 16:59:39 -08003380 if (IS_ERR(new_page)) {
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003381 /*
3382 * If a process owning a MAP_PRIVATE mapping fails to COW,
3383 * it is due to references held by a child and an insufficient
3384 * huge page pool. To guarantee the original mappers
3385 * reliability, unmap the page from child processes. The child
3386 * may get SIGKILLed if it later faults.
3387 */
3388 if (outside_reserve) {
Davidlohr Buesoad4404a2014-08-06 16:06:47 -07003389 page_cache_release(old_page);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003390 BUG_ON(huge_pte_none(pte));
Davidlohr Bueso2f4612a2014-08-06 16:06:45 -07003391 unmap_ref_private(mm, vma, old_page, address);
3392 BUG_ON(huge_pte_none(pte));
3393 spin_lock(ptl);
3394 ptep = huge_pte_offset(mm, address & huge_page_mask(h));
3395 if (likely(ptep &&
3396 pte_same(huge_ptep_get(ptep), pte)))
3397 goto retry_avoidcopy;
3398 /*
3399 * race occurs while re-acquiring page table
3400 * lock, and our job is done.
3401 */
3402 return 0;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003403 }
3404
Davidlohr Buesoad4404a2014-08-06 16:06:47 -07003405 ret = (PTR_ERR(new_page) == -ENOMEM) ?
3406 VM_FAULT_OOM : VM_FAULT_SIGBUS;
3407 goto out_release_old;
David Gibson1e8f8892006-01-06 00:10:44 -08003408 }
3409
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09003410 /*
3411 * When the original hugepage is shared one, it does not have
3412 * anon_vma prepared.
3413 */
Dean Nelson44e2aa92010-10-26 14:22:08 -07003414 if (unlikely(anon_vma_prepare(vma))) {
Davidlohr Buesoad4404a2014-08-06 16:06:47 -07003415 ret = VM_FAULT_OOM;
3416 goto out_release_all;
Dean Nelson44e2aa92010-10-26 14:22:08 -07003417 }
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09003418
Andrea Arcangeli47ad8472011-01-13 15:46:47 -08003419 copy_user_huge_page(new_page, old_page, address, vma,
3420 pages_per_huge_page(h));
Nick Piggin0ed361d2008-02-04 22:29:34 -08003421 __SetPageUptodate(new_page);
Naoya Horiguchibcc54222015-04-15 16:14:38 -07003422 set_page_huge_active(new_page);
David Gibson1e8f8892006-01-06 00:10:44 -08003423
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07003424 mmun_start = address & huge_page_mask(h);
3425 mmun_end = mmun_start + huge_page_size(h);
3426 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
Davidlohr Buesoad4404a2014-08-06 16:06:47 -07003427
Larry Woodmanb76c8cf2009-12-14 17:59:37 -08003428 /*
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003429 * Retake the page table lock to check for racing updates
Larry Woodmanb76c8cf2009-12-14 17:59:37 -08003430 * before the page tables are altered
3431 */
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003432 spin_lock(ptl);
Andi Kleena5516432008-07-23 21:27:41 -07003433 ptep = huge_pte_offset(mm, address & huge_page_mask(h));
Naoya Horiguchia9af0c52014-04-07 15:36:54 -07003434 if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) {
Joonsoo Kim07443a82013-09-11 14:21:58 -07003435 ClearPagePrivate(new_page);
3436
David Gibson1e8f8892006-01-06 00:10:44 -08003437 /* Break COW */
Gerald Schaefer8fe627e2008-04-28 02:13:28 -07003438 huge_ptep_clear_flush(vma, address, ptep);
Joerg Roedel34ee6452014-11-13 13:46:09 +11003439 mmu_notifier_invalidate_range(mm, mmun_start, mmun_end);
David Gibson1e8f8892006-01-06 00:10:44 -08003440 set_huge_pte_at(mm, address, ptep,
3441 make_huge_pte(vma, new_page, 1));
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09003442 page_remove_rmap(old_page);
Naoya Horiguchicd67f0d2010-09-10 13:23:04 +09003443 hugepage_add_new_anon_rmap(new_page, vma, address);
David Gibson1e8f8892006-01-06 00:10:44 -08003444 /* Make the old page be freed below */
3445 new_page = old_page;
3446 }
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003447 spin_unlock(ptl);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07003448 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
Davidlohr Buesoad4404a2014-08-06 16:06:47 -07003449out_release_all:
David Gibson1e8f8892006-01-06 00:10:44 -08003450 page_cache_release(new_page);
Davidlohr Buesoad4404a2014-08-06 16:06:47 -07003451out_release_old:
David Gibson1e8f8892006-01-06 00:10:44 -08003452 page_cache_release(old_page);
Joonsoo Kim83120342013-09-11 14:21:57 -07003453
Davidlohr Buesoad4404a2014-08-06 16:06:47 -07003454 spin_lock(ptl); /* Caller expects lock to be held */
3455 return ret;
David Gibson1e8f8892006-01-06 00:10:44 -08003456}
3457
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003458/* Return the pagecache page at a given address within a VMA */
Andi Kleena5516432008-07-23 21:27:41 -07003459static struct page *hugetlbfs_pagecache_page(struct hstate *h,
3460 struct vm_area_struct *vma, unsigned long address)
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003461{
3462 struct address_space *mapping;
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -07003463 pgoff_t idx;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003464
3465 mapping = vma->vm_file->f_mapping;
Andi Kleena5516432008-07-23 21:27:41 -07003466 idx = vma_hugecache_offset(h, vma, address);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003467
3468 return find_lock_page(mapping, idx);
3469}
3470
Hugh Dickins3ae77f42009-09-21 17:03:33 -07003471/*
3472 * Return whether there is a pagecache page to back given address within VMA.
3473 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
3474 */
3475static bool hugetlbfs_pagecache_present(struct hstate *h,
Hugh Dickins2a15efc2009-09-21 17:03:27 -07003476 struct vm_area_struct *vma, unsigned long address)
3477{
3478 struct address_space *mapping;
3479 pgoff_t idx;
3480 struct page *page;
3481
3482 mapping = vma->vm_file->f_mapping;
3483 idx = vma_hugecache_offset(h, vma, address);
3484
3485 page = find_get_page(mapping, idx);
3486 if (page)
3487 put_page(page);
3488 return page != NULL;
3489}
3490
Mike Kravetzab76ad52015-09-08 15:01:50 -07003491int huge_add_to_page_cache(struct page *page, struct address_space *mapping,
3492 pgoff_t idx)
3493{
3494 struct inode *inode = mapping->host;
3495 struct hstate *h = hstate_inode(inode);
3496 int err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
3497
3498 if (err)
3499 return err;
3500 ClearPagePrivate(page);
3501
3502 spin_lock(&inode->i_lock);
3503 inode->i_blocks += blocks_per_huge_page(h);
3504 spin_unlock(&inode->i_lock);
3505 return 0;
3506}
3507
Robert P. J. Daya1ed3dd2007-07-17 04:03:33 -07003508static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
Davidlohr Bueso8382d912014-04-03 14:47:31 -07003509 struct address_space *mapping, pgoff_t idx,
3510 unsigned long address, pte_t *ptep, unsigned int flags)
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01003511{
Andi Kleena5516432008-07-23 21:27:41 -07003512 struct hstate *h = hstate_vma(vma);
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01003513 int ret = VM_FAULT_SIGBUS;
Hillf Danton409eb8c2012-01-20 14:34:13 -08003514 int anon_rmap = 0;
Adam Litke4c887262005-10-29 18:16:46 -07003515 unsigned long size;
Adam Litke4c887262005-10-29 18:16:46 -07003516 struct page *page;
David Gibson1e8f8892006-01-06 00:10:44 -08003517 pte_t new_pte;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003518 spinlock_t *ptl;
Adam Litke4c887262005-10-29 18:16:46 -07003519
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003520 /*
3521 * Currently, we are forced to kill the process in the event the
3522 * original mapper has unmapped pages from the child due to a failed
Lucas De Marchi25985ed2011-03-30 22:57:33 -03003523 * COW. Warn that such a situation has occurred as it may not be obvious
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003524 */
3525 if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
Andrew Mortonffb22af2013-02-22 16:32:08 -08003526 pr_warning("PID %d killed due to inadequate hugepage pool\n",
3527 current->pid);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003528 return ret;
3529 }
3530
Adam Litke4c887262005-10-29 18:16:46 -07003531 /*
3532 * Use page lock to guard against racing truncation
3533 * before we get page_table_lock.
3534 */
Christoph Lameter6bda6662006-01-06 00:10:49 -08003535retry:
3536 page = find_lock_page(mapping, idx);
3537 if (!page) {
Andi Kleena5516432008-07-23 21:27:41 -07003538 size = i_size_read(mapping->host) >> huge_page_shift(h);
Hugh Dickinsebed4bf2006-10-28 10:38:43 -07003539 if (idx >= size)
3540 goto out;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003541 page = alloc_huge_page(vma, address, 0);
Adam Litke2fc39ce2007-11-14 16:59:39 -08003542 if (IS_ERR(page)) {
Aneesh Kumar K.V76dcee72012-07-31 16:41:57 -07003543 ret = PTR_ERR(page);
3544 if (ret == -ENOMEM)
3545 ret = VM_FAULT_OOM;
3546 else
3547 ret = VM_FAULT_SIGBUS;
Christoph Lameter6bda6662006-01-06 00:10:49 -08003548 goto out;
3549 }
Andrea Arcangeli47ad8472011-01-13 15:46:47 -08003550 clear_huge_page(page, address, pages_per_huge_page(h));
Nick Piggin0ed361d2008-02-04 22:29:34 -08003551 __SetPageUptodate(page);
Naoya Horiguchibcc54222015-04-15 16:14:38 -07003552 set_page_huge_active(page);
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01003553
Mel Gormanf83a2752009-05-28 14:34:40 -07003554 if (vma->vm_flags & VM_MAYSHARE) {
Mike Kravetzab76ad52015-09-08 15:01:50 -07003555 int err = huge_add_to_page_cache(page, mapping, idx);
Christoph Lameter6bda6662006-01-06 00:10:49 -08003556 if (err) {
3557 put_page(page);
Christoph Lameter6bda6662006-01-06 00:10:49 -08003558 if (err == -EEXIST)
3559 goto retry;
3560 goto out;
3561 }
Mel Gorman23be7462010-04-23 13:17:56 -04003562 } else {
Christoph Lameter6bda6662006-01-06 00:10:49 -08003563 lock_page(page);
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09003564 if (unlikely(anon_vma_prepare(vma))) {
3565 ret = VM_FAULT_OOM;
3566 goto backout_unlocked;
3567 }
Hillf Danton409eb8c2012-01-20 14:34:13 -08003568 anon_rmap = 1;
Mel Gorman23be7462010-04-23 13:17:56 -04003569 }
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09003570 } else {
Naoya Horiguchi998b4382010-09-08 10:19:32 +09003571 /*
3572 * If memory error occurs between mmap() and fault, some process
3573 * don't have hwpoisoned swap entry for errored virtual address.
3574 * So we need to block hugepage fault by PG_hwpoison bit check.
3575 */
3576 if (unlikely(PageHWPoison(page))) {
Chris Forbes32f84522011-07-25 17:12:14 -07003577 ret = VM_FAULT_HWPOISON |
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07003578 VM_FAULT_SET_HINDEX(hstate_index(h));
Naoya Horiguchi998b4382010-09-08 10:19:32 +09003579 goto backout_unlocked;
3580 }
Christoph Lameter6bda6662006-01-06 00:10:49 -08003581 }
David Gibson1e8f8892006-01-06 00:10:44 -08003582
Andy Whitcroft57303d82008-08-12 15:08:47 -07003583 /*
3584 * If we are going to COW a private mapping later, we examine the
3585 * pending reservations for this page now. This will ensure that
3586 * any allocations necessary to record that reservation occur outside
3587 * the spinlock.
3588 */
Mike Kravetz5e911372015-09-08 15:01:28 -07003589 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
Andy Whitcroft2b267362008-08-12 15:08:49 -07003590 if (vma_needs_reservation(h, vma, address) < 0) {
3591 ret = VM_FAULT_OOM;
3592 goto backout_unlocked;
3593 }
Mike Kravetz5e911372015-09-08 15:01:28 -07003594 /* Just decrements count, does not deallocate */
Mike Kravetzfeba16e2015-09-08 15:01:31 -07003595 vma_end_reservation(h, vma, address);
Mike Kravetz5e911372015-09-08 15:01:28 -07003596 }
Andy Whitcroft57303d82008-08-12 15:08:47 -07003597
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003598 ptl = huge_pte_lockptr(h, mm, ptep);
3599 spin_lock(ptl);
Andi Kleena5516432008-07-23 21:27:41 -07003600 size = i_size_read(mapping->host) >> huge_page_shift(h);
Adam Litke4c887262005-10-29 18:16:46 -07003601 if (idx >= size)
3602 goto backout;
3603
Nick Piggin83c54072007-07-19 01:47:05 -07003604 ret = 0;
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07003605 if (!huge_pte_none(huge_ptep_get(ptep)))
Adam Litke4c887262005-10-29 18:16:46 -07003606 goto backout;
3607
Joonsoo Kim07443a82013-09-11 14:21:58 -07003608 if (anon_rmap) {
3609 ClearPagePrivate(page);
Hillf Danton409eb8c2012-01-20 14:34:13 -08003610 hugepage_add_new_anon_rmap(page, vma, address);
Choi Gi-yongac714902014-04-07 15:37:36 -07003611 } else
Hillf Danton409eb8c2012-01-20 14:34:13 -08003612 page_dup_rmap(page);
David Gibson1e8f8892006-01-06 00:10:44 -08003613 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
3614 && (vma->vm_flags & VM_SHARED)));
3615 set_huge_pte_at(mm, address, ptep, new_pte);
3616
Naoya Horiguchi5d317b22015-11-05 18:47:14 -08003617 hugetlb_count_add(pages_per_huge_page(h), mm);
Hugh Dickins788c7df2009-06-23 13:49:05 +01003618 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
David Gibson1e8f8892006-01-06 00:10:44 -08003619 /* Optimization, do the COW without a second fault */
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003620 ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page, ptl);
David Gibson1e8f8892006-01-06 00:10:44 -08003621 }
3622
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003623 spin_unlock(ptl);
Adam Litke4c887262005-10-29 18:16:46 -07003624 unlock_page(page);
3625out:
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01003626 return ret;
Adam Litke4c887262005-10-29 18:16:46 -07003627
3628backout:
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003629 spin_unlock(ptl);
Andy Whitcroft2b267362008-08-12 15:08:49 -07003630backout_unlocked:
Adam Litke4c887262005-10-29 18:16:46 -07003631 unlock_page(page);
3632 put_page(page);
3633 goto out;
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01003634}
3635
Davidlohr Bueso8382d912014-04-03 14:47:31 -07003636#ifdef CONFIG_SMP
Mike Kravetzc672c7f2015-09-08 15:01:35 -07003637u32 hugetlb_fault_mutex_hash(struct hstate *h, struct mm_struct *mm,
Davidlohr Bueso8382d912014-04-03 14:47:31 -07003638 struct vm_area_struct *vma,
3639 struct address_space *mapping,
3640 pgoff_t idx, unsigned long address)
3641{
3642 unsigned long key[2];
3643 u32 hash;
3644
3645 if (vma->vm_flags & VM_SHARED) {
3646 key[0] = (unsigned long) mapping;
3647 key[1] = idx;
3648 } else {
3649 key[0] = (unsigned long) mm;
3650 key[1] = address >> huge_page_shift(h);
3651 }
3652
3653 hash = jhash2((u32 *)&key, sizeof(key)/sizeof(u32), 0);
3654
3655 return hash & (num_fault_mutexes - 1);
3656}
3657#else
3658/*
3659 * For uniprocesor systems we always use a single mutex, so just
3660 * return 0 and avoid the hashing overhead.
3661 */
Mike Kravetzc672c7f2015-09-08 15:01:35 -07003662u32 hugetlb_fault_mutex_hash(struct hstate *h, struct mm_struct *mm,
Davidlohr Bueso8382d912014-04-03 14:47:31 -07003663 struct vm_area_struct *vma,
3664 struct address_space *mapping,
3665 pgoff_t idx, unsigned long address)
3666{
3667 return 0;
3668}
3669#endif
3670
Adam Litke86e52162006-01-06 00:10:43 -08003671int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
Hugh Dickins788c7df2009-06-23 13:49:05 +01003672 unsigned long address, unsigned int flags)
Adam Litke86e52162006-01-06 00:10:43 -08003673{
Davidlohr Bueso8382d912014-04-03 14:47:31 -07003674 pte_t *ptep, entry;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003675 spinlock_t *ptl;
David Gibson1e8f8892006-01-06 00:10:44 -08003676 int ret;
Davidlohr Bueso8382d912014-04-03 14:47:31 -07003677 u32 hash;
3678 pgoff_t idx;
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09003679 struct page *page = NULL;
Andy Whitcroft57303d82008-08-12 15:08:47 -07003680 struct page *pagecache_page = NULL;
Andi Kleena5516432008-07-23 21:27:41 -07003681 struct hstate *h = hstate_vma(vma);
Davidlohr Bueso8382d912014-04-03 14:47:31 -07003682 struct address_space *mapping;
Naoya Horiguchi0f792cf2015-02-11 15:25:25 -08003683 int need_wait_lock = 0;
Adam Litke86e52162006-01-06 00:10:43 -08003684
KAMEZAWA Hiroyuki1e16a532012-01-10 15:07:22 -08003685 address &= huge_page_mask(h);
3686
Naoya Horiguchifd6a03ed2010-05-28 09:29:21 +09003687 ptep = huge_pte_offset(mm, address);
3688 if (ptep) {
3689 entry = huge_ptep_get(ptep);
Naoya Horiguchi290408d2010-09-08 10:19:35 +09003690 if (unlikely(is_hugetlb_entry_migration(entry))) {
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003691 migration_entry_wait_huge(vma, mm, ptep);
Naoya Horiguchi290408d2010-09-08 10:19:35 +09003692 return 0;
3693 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
Chris Forbes32f84522011-07-25 17:12:14 -07003694 return VM_FAULT_HWPOISON_LARGE |
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07003695 VM_FAULT_SET_HINDEX(hstate_index(h));
Naoya Horiguchifd6a03ed2010-05-28 09:29:21 +09003696 }
3697
Andi Kleena5516432008-07-23 21:27:41 -07003698 ptep = huge_pte_alloc(mm, address, huge_page_size(h));
Adam Litke86e52162006-01-06 00:10:43 -08003699 if (!ptep)
3700 return VM_FAULT_OOM;
3701
Davidlohr Bueso8382d912014-04-03 14:47:31 -07003702 mapping = vma->vm_file->f_mapping;
3703 idx = vma_hugecache_offset(h, vma, address);
3704
David Gibson3935baa2006-03-22 00:08:53 -08003705 /*
3706 * Serialize hugepage allocation and instantiation, so that we don't
3707 * get spurious allocation failures if two CPUs race to instantiate
3708 * the same page in the page cache.
3709 */
Mike Kravetzc672c7f2015-09-08 15:01:35 -07003710 hash = hugetlb_fault_mutex_hash(h, mm, vma, mapping, idx, address);
3711 mutex_lock(&hugetlb_fault_mutex_table[hash]);
Davidlohr Bueso8382d912014-04-03 14:47:31 -07003712
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07003713 entry = huge_ptep_get(ptep);
3714 if (huge_pte_none(entry)) {
Davidlohr Bueso8382d912014-04-03 14:47:31 -07003715 ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags);
David Gibsonb4d1d992008-10-15 22:01:11 -07003716 goto out_mutex;
David Gibson3935baa2006-03-22 00:08:53 -08003717 }
Adam Litke86e52162006-01-06 00:10:43 -08003718
Nick Piggin83c54072007-07-19 01:47:05 -07003719 ret = 0;
David Gibson1e8f8892006-01-06 00:10:44 -08003720
Andy Whitcroft57303d82008-08-12 15:08:47 -07003721 /*
Naoya Horiguchi0f792cf2015-02-11 15:25:25 -08003722 * entry could be a migration/hwpoison entry at this point, so this
3723 * check prevents the kernel from going below assuming that we have
3724 * a active hugepage in pagecache. This goto expects the 2nd page fault,
3725 * and is_hugetlb_entry_(migration|hwpoisoned) check will properly
3726 * handle it.
3727 */
3728 if (!pte_present(entry))
3729 goto out_mutex;
3730
3731 /*
Andy Whitcroft57303d82008-08-12 15:08:47 -07003732 * If we are going to COW the mapping later, we examine the pending
3733 * reservations for this page now. This will ensure that any
3734 * allocations necessary to record that reservation occur outside the
3735 * spinlock. For private mappings, we also lookup the pagecache
3736 * page now as it is used to determine if a reservation has been
3737 * consumed.
3738 */
Gerald Schaefer106c9922013-04-29 15:07:23 -07003739 if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) {
Andy Whitcroft2b267362008-08-12 15:08:49 -07003740 if (vma_needs_reservation(h, vma, address) < 0) {
3741 ret = VM_FAULT_OOM;
David Gibsonb4d1d992008-10-15 22:01:11 -07003742 goto out_mutex;
Andy Whitcroft2b267362008-08-12 15:08:49 -07003743 }
Mike Kravetz5e911372015-09-08 15:01:28 -07003744 /* Just decrements count, does not deallocate */
Mike Kravetzfeba16e2015-09-08 15:01:31 -07003745 vma_end_reservation(h, vma, address);
Andy Whitcroft57303d82008-08-12 15:08:47 -07003746
Mel Gormanf83a2752009-05-28 14:34:40 -07003747 if (!(vma->vm_flags & VM_MAYSHARE))
Andy Whitcroft57303d82008-08-12 15:08:47 -07003748 pagecache_page = hugetlbfs_pagecache_page(h,
3749 vma, address);
3750 }
3751
Naoya Horiguchi0f792cf2015-02-11 15:25:25 -08003752 ptl = huge_pte_lock(h, mm, ptep);
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09003753
David Gibson1e8f8892006-01-06 00:10:44 -08003754 /* Check for a racing update before calling hugetlb_cow */
David Gibsonb4d1d992008-10-15 22:01:11 -07003755 if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003756 goto out_ptl;
David Gibsonb4d1d992008-10-15 22:01:11 -07003757
Naoya Horiguchi0f792cf2015-02-11 15:25:25 -08003758 /*
3759 * hugetlb_cow() requires page locks of pte_page(entry) and
3760 * pagecache_page, so here we need take the former one
3761 * when page != pagecache_page or !pagecache_page.
3762 */
3763 page = pte_page(entry);
3764 if (page != pagecache_page)
3765 if (!trylock_page(page)) {
3766 need_wait_lock = 1;
3767 goto out_ptl;
3768 }
3769
3770 get_page(page);
David Gibsonb4d1d992008-10-15 22:01:11 -07003771
Hugh Dickins788c7df2009-06-23 13:49:05 +01003772 if (flags & FAULT_FLAG_WRITE) {
Gerald Schaefer106c9922013-04-29 15:07:23 -07003773 if (!huge_pte_write(entry)) {
Andy Whitcroft57303d82008-08-12 15:08:47 -07003774 ret = hugetlb_cow(mm, vma, address, ptep, entry,
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003775 pagecache_page, ptl);
Naoya Horiguchi0f792cf2015-02-11 15:25:25 -08003776 goto out_put_page;
David Gibsonb4d1d992008-10-15 22:01:11 -07003777 }
Gerald Schaefer106c9922013-04-29 15:07:23 -07003778 entry = huge_pte_mkdirty(entry);
David Gibsonb4d1d992008-10-15 22:01:11 -07003779 }
3780 entry = pte_mkyoung(entry);
Hugh Dickins788c7df2009-06-23 13:49:05 +01003781 if (huge_ptep_set_access_flags(vma, address, ptep, entry,
3782 flags & FAULT_FLAG_WRITE))
Russell King4b3073e2009-12-18 16:40:18 +00003783 update_mmu_cache(vma, address, ptep);
Naoya Horiguchi0f792cf2015-02-11 15:25:25 -08003784out_put_page:
3785 if (page != pagecache_page)
3786 unlock_page(page);
3787 put_page(page);
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003788out_ptl:
3789 spin_unlock(ptl);
Andy Whitcroft57303d82008-08-12 15:08:47 -07003790
3791 if (pagecache_page) {
3792 unlock_page(pagecache_page);
3793 put_page(pagecache_page);
3794 }
David Gibsonb4d1d992008-10-15 22:01:11 -07003795out_mutex:
Mike Kravetzc672c7f2015-09-08 15:01:35 -07003796 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
Naoya Horiguchi0f792cf2015-02-11 15:25:25 -08003797 /*
3798 * Generally it's safe to hold refcount during waiting page lock. But
3799 * here we just wait to defer the next page fault to avoid busy loop and
3800 * the page is not used after unlocked before returning from the current
3801 * page fault. So we are safe from accessing freed page, even if we wait
3802 * here without taking refcount.
3803 */
3804 if (need_wait_lock)
3805 wait_on_page_locked(page);
David Gibson1e8f8892006-01-06 00:10:44 -08003806 return ret;
Adam Litke86e52162006-01-06 00:10:43 -08003807}
3808
Michel Lespinasse28a35712013-02-22 16:35:55 -08003809long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
3810 struct page **pages, struct vm_area_struct **vmas,
3811 unsigned long *position, unsigned long *nr_pages,
3812 long i, unsigned int flags)
David Gibson63551ae2005-06-21 17:14:44 -07003813{
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08003814 unsigned long pfn_offset;
3815 unsigned long vaddr = *position;
Michel Lespinasse28a35712013-02-22 16:35:55 -08003816 unsigned long remainder = *nr_pages;
Andi Kleena5516432008-07-23 21:27:41 -07003817 struct hstate *h = hstate_vma(vma);
David Gibson63551ae2005-06-21 17:14:44 -07003818
David Gibson63551ae2005-06-21 17:14:44 -07003819 while (vaddr < vma->vm_end && remainder) {
Adam Litke4c887262005-10-29 18:16:46 -07003820 pte_t *pte;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003821 spinlock_t *ptl = NULL;
Hugh Dickins2a15efc2009-09-21 17:03:27 -07003822 int absent;
Adam Litke4c887262005-10-29 18:16:46 -07003823 struct page *page;
3824
3825 /*
David Rientjes02057962015-04-14 15:48:24 -07003826 * If we have a pending SIGKILL, don't keep faulting pages and
3827 * potentially allocating memory.
3828 */
3829 if (unlikely(fatal_signal_pending(current))) {
3830 remainder = 0;
3831 break;
3832 }
3833
3834 /*
Adam Litke4c887262005-10-29 18:16:46 -07003835 * Some archs (sparc64, sh*) have multiple pte_ts to
Hugh Dickins2a15efc2009-09-21 17:03:27 -07003836 * each hugepage. We have to make sure we get the
Adam Litke4c887262005-10-29 18:16:46 -07003837 * first, for the page indexing below to work.
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003838 *
3839 * Note that page table lock is not held when pte is null.
Adam Litke4c887262005-10-29 18:16:46 -07003840 */
Andi Kleena5516432008-07-23 21:27:41 -07003841 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h));
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003842 if (pte)
3843 ptl = huge_pte_lock(h, mm, pte);
Hugh Dickins2a15efc2009-09-21 17:03:27 -07003844 absent = !pte || huge_pte_none(huge_ptep_get(pte));
Adam Litke4c887262005-10-29 18:16:46 -07003845
Hugh Dickins2a15efc2009-09-21 17:03:27 -07003846 /*
3847 * When coredumping, it suits get_dump_page if we just return
Hugh Dickins3ae77f42009-09-21 17:03:33 -07003848 * an error where there's an empty slot with no huge pagecache
3849 * to back it. This way, we avoid allocating a hugepage, and
3850 * the sparse dumpfile avoids allocating disk blocks, but its
3851 * huge holes still show up with zeroes where they need to be.
Hugh Dickins2a15efc2009-09-21 17:03:27 -07003852 */
Hugh Dickins3ae77f42009-09-21 17:03:33 -07003853 if (absent && (flags & FOLL_DUMP) &&
3854 !hugetlbfs_pagecache_present(h, vma, vaddr)) {
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003855 if (pte)
3856 spin_unlock(ptl);
Hugh Dickins2a15efc2009-09-21 17:03:27 -07003857 remainder = 0;
3858 break;
3859 }
3860
Naoya Horiguchi9cc3a5b2013-04-17 15:58:30 -07003861 /*
3862 * We need call hugetlb_fault for both hugepages under migration
3863 * (in which case hugetlb_fault waits for the migration,) and
3864 * hwpoisoned hugepages (in which case we need to prevent the
3865 * caller from accessing to them.) In order to do this, we use
3866 * here is_swap_pte instead of is_hugetlb_entry_migration and
3867 * is_hugetlb_entry_hwpoisoned. This is because it simply covers
3868 * both cases, and because we can't follow correct pages
3869 * directly from any kind of swap entries.
3870 */
3871 if (absent || is_swap_pte(huge_ptep_get(pte)) ||
Gerald Schaefer106c9922013-04-29 15:07:23 -07003872 ((flags & FOLL_WRITE) &&
3873 !huge_pte_write(huge_ptep_get(pte)))) {
Adam Litke4c887262005-10-29 18:16:46 -07003874 int ret;
3875
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003876 if (pte)
3877 spin_unlock(ptl);
Hugh Dickins2a15efc2009-09-21 17:03:27 -07003878 ret = hugetlb_fault(mm, vma, vaddr,
3879 (flags & FOLL_WRITE) ? FAULT_FLAG_WRITE : 0);
Adam Litkea89182c2007-08-22 14:01:51 -07003880 if (!(ret & VM_FAULT_ERROR))
Adam Litke4c887262005-10-29 18:16:46 -07003881 continue;
3882
3883 remainder = 0;
Adam Litke4c887262005-10-29 18:16:46 -07003884 break;
3885 }
David Gibson63551ae2005-06-21 17:14:44 -07003886
Andi Kleena5516432008-07-23 21:27:41 -07003887 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07003888 page = pte_page(huge_ptep_get(pte));
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08003889same_page:
Chen, Kenneth Wd6692182006-03-31 02:29:57 -08003890 if (pages) {
Hugh Dickins2a15efc2009-09-21 17:03:27 -07003891 pages[i] = mem_map_offset(page, pfn_offset);
Andrea Arcangelia0368d42014-01-21 15:48:49 -08003892 get_page_foll(pages[i]);
Chen, Kenneth Wd6692182006-03-31 02:29:57 -08003893 }
David Gibson63551ae2005-06-21 17:14:44 -07003894
3895 if (vmas)
3896 vmas[i] = vma;
3897
3898 vaddr += PAGE_SIZE;
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08003899 ++pfn_offset;
David Gibson63551ae2005-06-21 17:14:44 -07003900 --remainder;
3901 ++i;
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08003902 if (vaddr < vma->vm_end && remainder &&
Andi Kleena5516432008-07-23 21:27:41 -07003903 pfn_offset < pages_per_huge_page(h)) {
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08003904 /*
3905 * We use pfn_offset to avoid touching the pageframes
3906 * of this compound page.
3907 */
3908 goto same_page;
3909 }
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003910 spin_unlock(ptl);
David Gibson63551ae2005-06-21 17:14:44 -07003911 }
Michel Lespinasse28a35712013-02-22 16:35:55 -08003912 *nr_pages = remainder;
David Gibson63551ae2005-06-21 17:14:44 -07003913 *position = vaddr;
3914
Hugh Dickins2a15efc2009-09-21 17:03:27 -07003915 return i ? i : -EFAULT;
David Gibson63551ae2005-06-21 17:14:44 -07003916}
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003917
Peter Zijlstra7da4d642012-11-19 03:14:23 +01003918unsigned long hugetlb_change_protection(struct vm_area_struct *vma,
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003919 unsigned long address, unsigned long end, pgprot_t newprot)
3920{
3921 struct mm_struct *mm = vma->vm_mm;
3922 unsigned long start = address;
3923 pte_t *ptep;
3924 pte_t pte;
Andi Kleena5516432008-07-23 21:27:41 -07003925 struct hstate *h = hstate_vma(vma);
Peter Zijlstra7da4d642012-11-19 03:14:23 +01003926 unsigned long pages = 0;
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003927
3928 BUG_ON(address >= end);
3929 flush_cache_range(vma, address, end);
3930
Rik van Riela5338092014-04-07 15:36:57 -07003931 mmu_notifier_invalidate_range_start(mm, start, end);
Davidlohr Bueso83cde9e2014-12-12 16:54:21 -08003932 i_mmap_lock_write(vma->vm_file->f_mapping);
Andi Kleena5516432008-07-23 21:27:41 -07003933 for (; address < end; address += huge_page_size(h)) {
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003934 spinlock_t *ptl;
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003935 ptep = huge_pte_offset(mm, address);
3936 if (!ptep)
3937 continue;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003938 ptl = huge_pte_lock(h, mm, ptep);
Peter Zijlstra7da4d642012-11-19 03:14:23 +01003939 if (huge_pmd_unshare(mm, &address, ptep)) {
3940 pages++;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003941 spin_unlock(ptl);
Chen, Kenneth W39dde652006-12-06 20:32:03 -08003942 continue;
Peter Zijlstra7da4d642012-11-19 03:14:23 +01003943 }
Naoya Horiguchia8bda282015-02-11 15:25:28 -08003944 pte = huge_ptep_get(ptep);
3945 if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) {
3946 spin_unlock(ptl);
3947 continue;
3948 }
3949 if (unlikely(is_hugetlb_entry_migration(pte))) {
3950 swp_entry_t entry = pte_to_swp_entry(pte);
3951
3952 if (is_write_migration_entry(entry)) {
3953 pte_t newpte;
3954
3955 make_migration_entry_read(&entry);
3956 newpte = swp_entry_to_pte(entry);
3957 set_huge_pte_at(mm, address, ptep, newpte);
3958 pages++;
3959 }
3960 spin_unlock(ptl);
3961 continue;
3962 }
3963 if (!huge_pte_none(pte)) {
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003964 pte = huge_ptep_get_and_clear(mm, address, ptep);
Gerald Schaefer106c9922013-04-29 15:07:23 -07003965 pte = pte_mkhuge(huge_pte_modify(pte, newprot));
Tony Lube7517d2013-02-04 14:28:46 -08003966 pte = arch_make_huge_pte(pte, vma, NULL, 0);
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003967 set_huge_pte_at(mm, address, ptep, pte);
Peter Zijlstra7da4d642012-11-19 03:14:23 +01003968 pages++;
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003969 }
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003970 spin_unlock(ptl);
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003971 }
Mel Gormand8333522012-07-31 16:46:20 -07003972 /*
Davidlohr Buesoc8c06ef2014-12-12 16:54:24 -08003973 * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare
Mel Gormand8333522012-07-31 16:46:20 -07003974 * may have cleared our pud entry and done put_page on the page table:
Davidlohr Buesoc8c06ef2014-12-12 16:54:24 -08003975 * once we release i_mmap_rwsem, another task can do the final put_page
Mel Gormand8333522012-07-31 16:46:20 -07003976 * and that page table be reused and filled with junk.
3977 */
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003978 flush_tlb_range(vma, start, end);
Joerg Roedel34ee6452014-11-13 13:46:09 +11003979 mmu_notifier_invalidate_range(mm, start, end);
Davidlohr Bueso83cde9e2014-12-12 16:54:21 -08003980 i_mmap_unlock_write(vma->vm_file->f_mapping);
Rik van Riela5338092014-04-07 15:36:57 -07003981 mmu_notifier_invalidate_range_end(mm, start, end);
Peter Zijlstra7da4d642012-11-19 03:14:23 +01003982
3983 return pages << h->order;
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003984}
3985
Mel Gormana1e78772008-07-23 21:27:23 -07003986int hugetlb_reserve_pages(struct inode *inode,
3987 long from, long to,
Mel Gorman5a6fe122009-02-10 14:02:27 +00003988 struct vm_area_struct *vma,
KOSAKI Motohiroca16d142011-05-26 19:16:19 +09003989 vm_flags_t vm_flags)
Adam Litkee4e574b2007-10-16 01:26:19 -07003990{
Mel Gorman17c9d122009-02-11 16:34:16 +00003991 long ret, chg;
Andi Kleena5516432008-07-23 21:27:41 -07003992 struct hstate *h = hstate_inode(inode);
David Gibson90481622012-03-21 16:34:12 -07003993 struct hugepage_subpool *spool = subpool_inode(inode);
Joonsoo Kim9119a412014-04-03 14:47:25 -07003994 struct resv_map *resv_map;
Mike Kravetz1c5ecae2015-04-15 16:13:39 -07003995 long gbl_reserve;
Adam Litkee4e574b2007-10-16 01:26:19 -07003996
Mel Gormana1e78772008-07-23 21:27:23 -07003997 /*
Mel Gorman17c9d122009-02-11 16:34:16 +00003998 * Only apply hugepage reservation if asked. At fault time, an
3999 * attempt will be made for VM_NORESERVE to allocate a page
David Gibson90481622012-03-21 16:34:12 -07004000 * without using reserves
Mel Gorman17c9d122009-02-11 16:34:16 +00004001 */
KOSAKI Motohiroca16d142011-05-26 19:16:19 +09004002 if (vm_flags & VM_NORESERVE)
Mel Gorman17c9d122009-02-11 16:34:16 +00004003 return 0;
4004
4005 /*
Mel Gormana1e78772008-07-23 21:27:23 -07004006 * Shared mappings base their reservation on the number of pages that
4007 * are already allocated on behalf of the file. Private mappings need
4008 * to reserve the full area even if read-only as mprotect() may be
4009 * called to make the mapping read-write. Assume !vma is a shm mapping
4010 */
Joonsoo Kim9119a412014-04-03 14:47:25 -07004011 if (!vma || vma->vm_flags & VM_MAYSHARE) {
Joonsoo Kim4e35f482014-04-03 14:47:30 -07004012 resv_map = inode_resv_map(inode);
Joonsoo Kim9119a412014-04-03 14:47:25 -07004013
Joonsoo Kim1406ec92014-04-03 14:47:26 -07004014 chg = region_chg(resv_map, from, to);
Joonsoo Kim9119a412014-04-03 14:47:25 -07004015
4016 } else {
4017 resv_map = resv_map_alloc();
Mel Gorman5a6fe122009-02-10 14:02:27 +00004018 if (!resv_map)
4019 return -ENOMEM;
4020
Mel Gorman17c9d122009-02-11 16:34:16 +00004021 chg = to - from;
4022
Mel Gorman5a6fe122009-02-10 14:02:27 +00004023 set_vma_resv_map(vma, resv_map);
4024 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
4025 }
4026
Dave Hansenc50ac052012-05-29 15:06:46 -07004027 if (chg < 0) {
4028 ret = chg;
4029 goto out_err;
4030 }
Mel Gorman17c9d122009-02-11 16:34:16 +00004031
Mike Kravetz1c5ecae2015-04-15 16:13:39 -07004032 /*
4033 * There must be enough pages in the subpool for the mapping. If
4034 * the subpool has a minimum size, there may be some global
4035 * reservations already in place (gbl_reserve).
4036 */
4037 gbl_reserve = hugepage_subpool_get_pages(spool, chg);
4038 if (gbl_reserve < 0) {
Dave Hansenc50ac052012-05-29 15:06:46 -07004039 ret = -ENOSPC;
4040 goto out_err;
4041 }
Mel Gorman17c9d122009-02-11 16:34:16 +00004042
4043 /*
4044 * Check enough hugepages are available for the reservation.
David Gibson90481622012-03-21 16:34:12 -07004045 * Hand the pages back to the subpool if there are not
Mel Gorman17c9d122009-02-11 16:34:16 +00004046 */
Mike Kravetz1c5ecae2015-04-15 16:13:39 -07004047 ret = hugetlb_acct_memory(h, gbl_reserve);
Mel Gorman17c9d122009-02-11 16:34:16 +00004048 if (ret < 0) {
Mike Kravetz1c5ecae2015-04-15 16:13:39 -07004049 /* put back original number of pages, chg */
4050 (void)hugepage_subpool_put_pages(spool, chg);
Dave Hansenc50ac052012-05-29 15:06:46 -07004051 goto out_err;
Mel Gorman17c9d122009-02-11 16:34:16 +00004052 }
4053
4054 /*
4055 * Account for the reservations made. Shared mappings record regions
4056 * that have reservations as they are shared by multiple VMAs.
4057 * When the last VMA disappears, the region map says how much
4058 * the reservation was and the page cache tells how much of
4059 * the reservation was consumed. Private mappings are per-VMA and
4060 * only the consumed reservations are tracked. When the VMA
4061 * disappears, the original reservation is the VMA size and the
4062 * consumed reservations are stored in the map. Hence, nothing
4063 * else has to be done for private mappings here
4064 */
Mike Kravetz33039672015-06-24 16:57:58 -07004065 if (!vma || vma->vm_flags & VM_MAYSHARE) {
4066 long add = region_add(resv_map, from, to);
4067
4068 if (unlikely(chg > add)) {
4069 /*
4070 * pages in this range were added to the reserve
4071 * map between region_chg and region_add. This
4072 * indicates a race with alloc_huge_page. Adjust
4073 * the subpool and reserve counts modified above
4074 * based on the difference.
4075 */
4076 long rsv_adjust;
4077
4078 rsv_adjust = hugepage_subpool_put_pages(spool,
4079 chg - add);
4080 hugetlb_acct_memory(h, -rsv_adjust);
4081 }
4082 }
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -07004083 return 0;
Dave Hansenc50ac052012-05-29 15:06:46 -07004084out_err:
Mike Kravetz5e911372015-09-08 15:01:28 -07004085 if (!vma || vma->vm_flags & VM_MAYSHARE)
4086 region_abort(resv_map, from, to);
Joonsoo Kimf031dd22014-04-03 14:47:28 -07004087 if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
4088 kref_put(&resv_map->refs, resv_map_release);
Dave Hansenc50ac052012-05-29 15:06:46 -07004089 return ret;
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -07004090}
4091
Mike Kravetzb5cec282015-09-08 15:01:41 -07004092long hugetlb_unreserve_pages(struct inode *inode, long start, long end,
4093 long freed)
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -07004094{
Andi Kleena5516432008-07-23 21:27:41 -07004095 struct hstate *h = hstate_inode(inode);
Joonsoo Kim4e35f482014-04-03 14:47:30 -07004096 struct resv_map *resv_map = inode_resv_map(inode);
Joonsoo Kim9119a412014-04-03 14:47:25 -07004097 long chg = 0;
David Gibson90481622012-03-21 16:34:12 -07004098 struct hugepage_subpool *spool = subpool_inode(inode);
Mike Kravetz1c5ecae2015-04-15 16:13:39 -07004099 long gbl_reserve;
Ken Chen45c682a2007-11-14 16:59:44 -08004100
Mike Kravetzb5cec282015-09-08 15:01:41 -07004101 if (resv_map) {
4102 chg = region_del(resv_map, start, end);
4103 /*
4104 * region_del() can fail in the rare case where a region
4105 * must be split and another region descriptor can not be
4106 * allocated. If end == LONG_MAX, it will not fail.
4107 */
4108 if (chg < 0)
4109 return chg;
4110 }
4111
Ken Chen45c682a2007-11-14 16:59:44 -08004112 spin_lock(&inode->i_lock);
Eric Sandeene4c6f8b2009-07-29 15:02:16 -07004113 inode->i_blocks -= (blocks_per_huge_page(h) * freed);
Ken Chen45c682a2007-11-14 16:59:44 -08004114 spin_unlock(&inode->i_lock);
4115
Mike Kravetz1c5ecae2015-04-15 16:13:39 -07004116 /*
4117 * If the subpool has a minimum size, the number of global
4118 * reservations to be released may be adjusted.
4119 */
4120 gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed));
4121 hugetlb_acct_memory(h, -gbl_reserve);
Mike Kravetzb5cec282015-09-08 15:01:41 -07004122
4123 return 0;
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -07004124}
Naoya Horiguchi93f70f92010-05-28 09:29:20 +09004125
Steve Capper3212b532013-04-23 12:35:02 +01004126#ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
4127static unsigned long page_table_shareable(struct vm_area_struct *svma,
4128 struct vm_area_struct *vma,
4129 unsigned long addr, pgoff_t idx)
4130{
4131 unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) +
4132 svma->vm_start;
4133 unsigned long sbase = saddr & PUD_MASK;
4134 unsigned long s_end = sbase + PUD_SIZE;
4135
4136 /* Allow segments to share if only one is marked locked */
Eric B Munsonde60f5f2015-11-05 18:51:36 -08004137 unsigned long vm_flags = vma->vm_flags & VM_LOCKED_CLEAR_MASK;
4138 unsigned long svm_flags = svma->vm_flags & VM_LOCKED_CLEAR_MASK;
Steve Capper3212b532013-04-23 12:35:02 +01004139
4140 /*
4141 * match the virtual addresses, permission and the alignment of the
4142 * page table page.
4143 */
4144 if (pmd_index(addr) != pmd_index(saddr) ||
4145 vm_flags != svm_flags ||
4146 sbase < svma->vm_start || svma->vm_end < s_end)
4147 return 0;
4148
4149 return saddr;
4150}
4151
Nicholas Krause31aafb42015-09-04 15:47:58 -07004152static bool vma_shareable(struct vm_area_struct *vma, unsigned long addr)
Steve Capper3212b532013-04-23 12:35:02 +01004153{
4154 unsigned long base = addr & PUD_MASK;
4155 unsigned long end = base + PUD_SIZE;
4156
4157 /*
4158 * check on proper vm_flags and page table alignment
4159 */
4160 if (vma->vm_flags & VM_MAYSHARE &&
4161 vma->vm_start <= base && end <= vma->vm_end)
Nicholas Krause31aafb42015-09-04 15:47:58 -07004162 return true;
4163 return false;
Steve Capper3212b532013-04-23 12:35:02 +01004164}
4165
4166/*
4167 * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
4168 * and returns the corresponding pte. While this is not necessary for the
4169 * !shared pmd case because we can allocate the pmd later as well, it makes the
4170 * code much cleaner. pmd allocation is essential for the shared case because
Davidlohr Buesoc8c06ef2014-12-12 16:54:24 -08004171 * pud has to be populated inside the same i_mmap_rwsem section - otherwise
Steve Capper3212b532013-04-23 12:35:02 +01004172 * racing tasks could either miss the sharing (see huge_pte_offset) or select a
4173 * bad pmd for sharing.
4174 */
4175pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
4176{
4177 struct vm_area_struct *vma = find_vma(mm, addr);
4178 struct address_space *mapping = vma->vm_file->f_mapping;
4179 pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) +
4180 vma->vm_pgoff;
4181 struct vm_area_struct *svma;
4182 unsigned long saddr;
4183 pte_t *spte = NULL;
4184 pte_t *pte;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08004185 spinlock_t *ptl;
Steve Capper3212b532013-04-23 12:35:02 +01004186
4187 if (!vma_shareable(vma, addr))
4188 return (pte_t *)pmd_alloc(mm, pud, addr);
4189
Davidlohr Bueso83cde9e2014-12-12 16:54:21 -08004190 i_mmap_lock_write(mapping);
Steve Capper3212b532013-04-23 12:35:02 +01004191 vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) {
4192 if (svma == vma)
4193 continue;
4194
4195 saddr = page_table_shareable(svma, vma, addr, idx);
4196 if (saddr) {
4197 spte = huge_pte_offset(svma->vm_mm, saddr);
4198 if (spte) {
Kirill A. Shutemovdc6c9a32015-02-11 15:26:50 -08004199 mm_inc_nr_pmds(mm);
Steve Capper3212b532013-04-23 12:35:02 +01004200 get_page(virt_to_page(spte));
4201 break;
4202 }
4203 }
4204 }
4205
4206 if (!spte)
4207 goto out;
4208
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08004209 ptl = huge_pte_lockptr(hstate_vma(vma), mm, spte);
4210 spin_lock(ptl);
Kirill A. Shutemovdc6c9a32015-02-11 15:26:50 -08004211 if (pud_none(*pud)) {
Steve Capper3212b532013-04-23 12:35:02 +01004212 pud_populate(mm, pud,
4213 (pmd_t *)((unsigned long)spte & PAGE_MASK));
Kirill A. Shutemovdc6c9a32015-02-11 15:26:50 -08004214 } else {
Steve Capper3212b532013-04-23 12:35:02 +01004215 put_page(virt_to_page(spte));
Kirill A. Shutemovdc6c9a32015-02-11 15:26:50 -08004216 mm_inc_nr_pmds(mm);
4217 }
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08004218 spin_unlock(ptl);
Steve Capper3212b532013-04-23 12:35:02 +01004219out:
4220 pte = (pte_t *)pmd_alloc(mm, pud, addr);
Davidlohr Bueso83cde9e2014-12-12 16:54:21 -08004221 i_mmap_unlock_write(mapping);
Steve Capper3212b532013-04-23 12:35:02 +01004222 return pte;
4223}
4224
4225/*
4226 * unmap huge page backed by shared pte.
4227 *
4228 * Hugetlb pte page is ref counted at the time of mapping. If pte is shared
4229 * indicated by page_count > 1, unmap is achieved by clearing pud and
4230 * decrementing the ref count. If count == 1, the pte page is not shared.
4231 *
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08004232 * called with page table lock held.
Steve Capper3212b532013-04-23 12:35:02 +01004233 *
4234 * returns: 1 successfully unmapped a shared pte page
4235 * 0 the underlying pte page is not shared, or it is the last user
4236 */
4237int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
4238{
4239 pgd_t *pgd = pgd_offset(mm, *addr);
4240 pud_t *pud = pud_offset(pgd, *addr);
4241
4242 BUG_ON(page_count(virt_to_page(ptep)) == 0);
4243 if (page_count(virt_to_page(ptep)) == 1)
4244 return 0;
4245
4246 pud_clear(pud);
4247 put_page(virt_to_page(ptep));
Kirill A. Shutemovdc6c9a32015-02-11 15:26:50 -08004248 mm_dec_nr_pmds(mm);
Steve Capper3212b532013-04-23 12:35:02 +01004249 *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE;
4250 return 1;
4251}
Steve Capper9e5fc742013-04-30 08:02:03 +01004252#define want_pmd_share() (1)
4253#else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
4254pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
4255{
4256 return NULL;
4257}
Zhang Zhene81f2d22015-06-24 16:56:13 -07004258
4259int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
4260{
4261 return 0;
4262}
Steve Capper9e5fc742013-04-30 08:02:03 +01004263#define want_pmd_share() (0)
Steve Capper3212b532013-04-23 12:35:02 +01004264#endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
4265
Steve Capper9e5fc742013-04-30 08:02:03 +01004266#ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
4267pte_t *huge_pte_alloc(struct mm_struct *mm,
4268 unsigned long addr, unsigned long sz)
4269{
4270 pgd_t *pgd;
4271 pud_t *pud;
4272 pte_t *pte = NULL;
4273
4274 pgd = pgd_offset(mm, addr);
4275 pud = pud_alloc(mm, pgd, addr);
4276 if (pud) {
4277 if (sz == PUD_SIZE) {
4278 pte = (pte_t *)pud;
4279 } else {
4280 BUG_ON(sz != PMD_SIZE);
4281 if (want_pmd_share() && pud_none(*pud))
4282 pte = huge_pmd_share(mm, addr, pud);
4283 else
4284 pte = (pte_t *)pmd_alloc(mm, pud, addr);
4285 }
4286 }
4287 BUG_ON(pte && !pte_none(*pte) && !pte_huge(*pte));
4288
4289 return pte;
4290}
4291
4292pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr)
4293{
4294 pgd_t *pgd;
4295 pud_t *pud;
4296 pmd_t *pmd = NULL;
4297
4298 pgd = pgd_offset(mm, addr);
4299 if (pgd_present(*pgd)) {
4300 pud = pud_offset(pgd, addr);
4301 if (pud_present(*pud)) {
4302 if (pud_huge(*pud))
4303 return (pte_t *)pud;
4304 pmd = pmd_offset(pud, addr);
4305 }
4306 }
4307 return (pte_t *) pmd;
4308}
4309
Naoya Horiguchi61f77ed2015-02-11 15:25:15 -08004310#endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
4311
4312/*
4313 * These functions are overwritable if your architecture needs its own
4314 * behavior.
4315 */
4316struct page * __weak
4317follow_huge_addr(struct mm_struct *mm, unsigned long address,
4318 int write)
4319{
4320 return ERR_PTR(-EINVAL);
4321}
4322
4323struct page * __weak
Steve Capper9e5fc742013-04-30 08:02:03 +01004324follow_huge_pmd(struct mm_struct *mm, unsigned long address,
Naoya Horiguchie66f17f2015-02-11 15:25:22 -08004325 pmd_t *pmd, int flags)
Steve Capper9e5fc742013-04-30 08:02:03 +01004326{
Naoya Horiguchie66f17f2015-02-11 15:25:22 -08004327 struct page *page = NULL;
4328 spinlock_t *ptl;
4329retry:
4330 ptl = pmd_lockptr(mm, pmd);
4331 spin_lock(ptl);
4332 /*
4333 * make sure that the address range covered by this pmd is not
4334 * unmapped from other threads.
4335 */
4336 if (!pmd_huge(*pmd))
4337 goto out;
4338 if (pmd_present(*pmd)) {
Gerald Schaefer97534122015-04-14 15:42:30 -07004339 page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT);
Naoya Horiguchie66f17f2015-02-11 15:25:22 -08004340 if (flags & FOLL_GET)
4341 get_page(page);
4342 } else {
4343 if (is_hugetlb_entry_migration(huge_ptep_get((pte_t *)pmd))) {
4344 spin_unlock(ptl);
4345 __migration_entry_wait(mm, (pte_t *)pmd, ptl);
4346 goto retry;
4347 }
4348 /*
4349 * hwpoisoned entry is treated as no_page_table in
4350 * follow_page_mask().
4351 */
4352 }
4353out:
4354 spin_unlock(ptl);
Steve Capper9e5fc742013-04-30 08:02:03 +01004355 return page;
4356}
4357
Naoya Horiguchi61f77ed2015-02-11 15:25:15 -08004358struct page * __weak
Steve Capper9e5fc742013-04-30 08:02:03 +01004359follow_huge_pud(struct mm_struct *mm, unsigned long address,
Naoya Horiguchie66f17f2015-02-11 15:25:22 -08004360 pud_t *pud, int flags)
Steve Capper9e5fc742013-04-30 08:02:03 +01004361{
Naoya Horiguchie66f17f2015-02-11 15:25:22 -08004362 if (flags & FOLL_GET)
4363 return NULL;
Steve Capper9e5fc742013-04-30 08:02:03 +01004364
Naoya Horiguchie66f17f2015-02-11 15:25:22 -08004365 return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT);
Steve Capper9e5fc742013-04-30 08:02:03 +01004366}
4367
Andi Kleend5bd9102010-09-27 09:00:12 +02004368#ifdef CONFIG_MEMORY_FAILURE
4369
Naoya Horiguchi93f70f92010-05-28 09:29:20 +09004370/*
4371 * This function is called from memory failure code.
4372 * Assume the caller holds page lock of the head page.
4373 */
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +09004374int dequeue_hwpoisoned_huge_page(struct page *hpage)
Naoya Horiguchi93f70f92010-05-28 09:29:20 +09004375{
4376 struct hstate *h = page_hstate(hpage);
4377 int nid = page_to_nid(hpage);
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +09004378 int ret = -EBUSY;
Naoya Horiguchi93f70f92010-05-28 09:29:20 +09004379
4380 spin_lock(&hugetlb_lock);
Naoya Horiguchi7e1f0492015-04-15 16:14:41 -07004381 /*
4382 * Just checking !page_huge_active is not enough, because that could be
4383 * an isolated/hwpoisoned hugepage (which have >0 refcount).
4384 */
4385 if (!page_huge_active(hpage) && !page_count(hpage)) {
Naoya Horiguchi56f2fb12012-12-12 13:52:33 -08004386 /*
4387 * Hwpoisoned hugepage isn't linked to activelist or freelist,
4388 * but dangling hpage->lru can trigger list-debug warnings
4389 * (this happens when we call unpoison_memory() on it),
4390 * so let it point to itself with list_del_init().
4391 */
4392 list_del_init(&hpage->lru);
Naoya Horiguchi8c6c2ec2010-09-08 10:19:38 +09004393 set_page_refcounted(hpage);
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +09004394 h->free_huge_pages--;
4395 h->free_huge_pages_node[nid]--;
4396 ret = 0;
4397 }
Naoya Horiguchi93f70f92010-05-28 09:29:20 +09004398 spin_unlock(&hugetlb_lock);
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +09004399 return ret;
Naoya Horiguchi93f70f92010-05-28 09:29:20 +09004400}
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +09004401#endif
Naoya Horiguchi31caf662013-09-11 14:21:59 -07004402
4403bool isolate_huge_page(struct page *page, struct list_head *list)
4404{
Naoya Horiguchibcc54222015-04-15 16:14:38 -07004405 bool ret = true;
4406
Sasha Levin309381fea2014-01-23 15:52:54 -08004407 VM_BUG_ON_PAGE(!PageHead(page), page);
Naoya Horiguchi31caf662013-09-11 14:21:59 -07004408 spin_lock(&hugetlb_lock);
Naoya Horiguchibcc54222015-04-15 16:14:38 -07004409 if (!page_huge_active(page) || !get_page_unless_zero(page)) {
4410 ret = false;
4411 goto unlock;
4412 }
4413 clear_page_huge_active(page);
Naoya Horiguchi31caf662013-09-11 14:21:59 -07004414 list_move_tail(&page->lru, list);
Naoya Horiguchibcc54222015-04-15 16:14:38 -07004415unlock:
Naoya Horiguchi31caf662013-09-11 14:21:59 -07004416 spin_unlock(&hugetlb_lock);
Naoya Horiguchibcc54222015-04-15 16:14:38 -07004417 return ret;
Naoya Horiguchi31caf662013-09-11 14:21:59 -07004418}
4419
4420void putback_active_hugepage(struct page *page)
4421{
Sasha Levin309381fea2014-01-23 15:52:54 -08004422 VM_BUG_ON_PAGE(!PageHead(page), page);
Naoya Horiguchi31caf662013-09-11 14:21:59 -07004423 spin_lock(&hugetlb_lock);
Naoya Horiguchibcc54222015-04-15 16:14:38 -07004424 set_page_huge_active(page);
Naoya Horiguchi31caf662013-09-11 14:21:59 -07004425 list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist);
4426 spin_unlock(&hugetlb_lock);
4427 put_page(page);
4428}