blob: ef6963b577fd2920c1a4857f7b74b50e91c98b7a [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);
Mike Kravetzdbe409e2015-12-11 13:40:52 -0800375 if (!trg) {
376 kfree(nrg);
Mike Kravetz5e911372015-09-08 15:01:28 -0700377 return -ENOMEM;
Mike Kravetzdbe409e2015-12-11 13:40:52 -0800378 }
Mike Kravetz5e911372015-09-08 15:01:28 -0700379
380 spin_lock(&resv->lock);
381 list_add(&trg->link, &resv->region_cache);
382 resv->region_cache_count++;
383 goto retry_locked;
384 }
385
Andy Whitcroft96822902008-07-23 21:27:29 -0700386 /* Locate the region we are before or in. */
387 list_for_each_entry(rg, head, link)
388 if (f <= rg->to)
389 break;
390
391 /* If we are below the current region then a new region is required.
392 * Subtle, allocate a new region at the position but make it zero
393 * size such that we can guarantee to record the reservation. */
394 if (&rg->link == head || t < rg->from) {
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700395 if (!nrg) {
Mike Kravetz5e911372015-09-08 15:01:28 -0700396 resv->adds_in_progress--;
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700397 spin_unlock(&resv->lock);
398 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
399 if (!nrg)
400 return -ENOMEM;
Andy Whitcroft96822902008-07-23 21:27:29 -0700401
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700402 nrg->from = f;
403 nrg->to = f;
404 INIT_LIST_HEAD(&nrg->link);
405 goto retry;
406 }
407
408 list_add(&nrg->link, rg->link.prev);
409 chg = t - f;
410 goto out_nrg;
Andy Whitcroft96822902008-07-23 21:27:29 -0700411 }
412
413 /* Round our left edge to the current segment if it encloses us. */
414 if (f > rg->from)
415 f = rg->from;
416 chg = t - f;
417
418 /* Check for and consume any regions we now overlap with. */
419 list_for_each_entry(rg, rg->link.prev, link) {
420 if (&rg->link == head)
421 break;
422 if (rg->from > t)
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700423 goto out;
Andy Whitcroft96822902008-07-23 21:27:29 -0700424
Lucas De Marchi25985ed2011-03-30 22:57:33 -0300425 /* We overlap with this area, if it extends further than
Andy Whitcroft96822902008-07-23 21:27:29 -0700426 * us then we must extend ourselves. Account for its
427 * existing reservation. */
428 if (rg->to > t) {
429 chg += rg->to - t;
430 t = rg->to;
431 }
432 chg -= rg->to - rg->from;
433 }
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700434
435out:
436 spin_unlock(&resv->lock);
437 /* We already know we raced and no longer need the new region */
438 kfree(nrg);
439 return chg;
440out_nrg:
441 spin_unlock(&resv->lock);
Andy Whitcroft96822902008-07-23 21:27:29 -0700442 return chg;
443}
444
Mike Kravetz1dd308a2015-06-24 16:57:52 -0700445/*
Mike Kravetz5e911372015-09-08 15:01:28 -0700446 * Abort the in progress add operation. The adds_in_progress field
447 * of the resv_map keeps track of the operations in progress between
448 * calls to region_chg and region_add. Operations are sometimes
449 * aborted after the call to region_chg. In such cases, region_abort
450 * is called to decrement the adds_in_progress counter.
451 *
452 * NOTE: The range arguments [f, t) are not needed or used in this
453 * routine. They are kept to make reading the calling code easier as
454 * arguments will match the associated region_chg call.
455 */
456static void region_abort(struct resv_map *resv, long f, long t)
457{
458 spin_lock(&resv->lock);
459 VM_BUG_ON(!resv->region_cache_count);
460 resv->adds_in_progress--;
461 spin_unlock(&resv->lock);
462}
463
464/*
Mike Kravetzfeba16e2015-09-08 15:01:31 -0700465 * Delete the specified range [f, t) from the reserve map. If the
466 * t parameter is LONG_MAX, this indicates that ALL regions after f
467 * should be deleted. Locate the regions which intersect [f, t)
468 * and either trim, delete or split the existing regions.
469 *
470 * Returns the number of huge pages deleted from the reserve map.
471 * In the normal case, the return value is zero or more. In the
472 * case where a region must be split, a new region descriptor must
473 * be allocated. If the allocation fails, -ENOMEM will be returned.
474 * NOTE: If the parameter t == LONG_MAX, then we will never split
475 * a region and possibly return -ENOMEM. Callers specifying
476 * t == LONG_MAX do not need to check for -ENOMEM error.
Mike Kravetz1dd308a2015-06-24 16:57:52 -0700477 */
Mike Kravetzfeba16e2015-09-08 15:01:31 -0700478static long region_del(struct resv_map *resv, long f, long t)
Andy Whitcroft96822902008-07-23 21:27:29 -0700479{
Joonsoo Kim1406ec92014-04-03 14:47:26 -0700480 struct list_head *head = &resv->regions;
Andy Whitcroft96822902008-07-23 21:27:29 -0700481 struct file_region *rg, *trg;
Mike Kravetzfeba16e2015-09-08 15:01:31 -0700482 struct file_region *nrg = NULL;
483 long del = 0;
Andy Whitcroft96822902008-07-23 21:27:29 -0700484
Mike Kravetzfeba16e2015-09-08 15:01:31 -0700485retry:
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700486 spin_lock(&resv->lock);
Mike Kravetzfeba16e2015-09-08 15:01:31 -0700487 list_for_each_entry_safe(rg, trg, head, link) {
Mike Kravetzdbe409e2015-12-11 13:40:52 -0800488 /*
489 * Skip regions before the range to be deleted. file_region
490 * ranges are normally of the form [from, to). However, there
491 * may be a "placeholder" entry in the map which is of the form
492 * (from, to) with from == to. Check for placeholder entries
493 * at the beginning of the range to be deleted.
494 */
495 if (rg->to <= f && (rg->to != rg->from || rg->to != f))
Mike Kravetzfeba16e2015-09-08 15:01:31 -0700496 continue;
Mike Kravetzdbe409e2015-12-11 13:40:52 -0800497
Mike Kravetzfeba16e2015-09-08 15:01:31 -0700498 if (rg->from >= t)
Andy Whitcroft96822902008-07-23 21:27:29 -0700499 break;
Andy Whitcroft96822902008-07-23 21:27:29 -0700500
Mike Kravetzfeba16e2015-09-08 15:01:31 -0700501 if (f > rg->from && t < rg->to) { /* Must split region */
502 /*
503 * Check for an entry in the cache before dropping
504 * lock and attempting allocation.
505 */
506 if (!nrg &&
507 resv->region_cache_count > resv->adds_in_progress) {
508 nrg = list_first_entry(&resv->region_cache,
509 struct file_region,
510 link);
511 list_del(&nrg->link);
512 resv->region_cache_count--;
513 }
514
515 if (!nrg) {
516 spin_unlock(&resv->lock);
517 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
518 if (!nrg)
519 return -ENOMEM;
520 goto retry;
521 }
522
523 del += t - f;
524
525 /* New entry for end of split region */
526 nrg->from = t;
527 nrg->to = rg->to;
528 INIT_LIST_HEAD(&nrg->link);
529
530 /* Original entry is trimmed */
531 rg->to = f;
532
533 list_add(&nrg->link, &rg->link);
534 nrg = NULL;
535 break;
536 }
537
538 if (f <= rg->from && t >= rg->to) { /* Remove entire region */
539 del += rg->to - rg->from;
540 list_del(&rg->link);
541 kfree(rg);
542 continue;
543 }
544
545 if (f <= rg->from) { /* Trim beginning of region */
546 del += t - rg->from;
547 rg->from = t;
548 } else { /* Trim end of region */
549 del += rg->to - f;
550 rg->to = f;
551 }
Andy Whitcroft96822902008-07-23 21:27:29 -0700552 }
553
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700554 spin_unlock(&resv->lock);
Mike Kravetzfeba16e2015-09-08 15:01:31 -0700555 kfree(nrg);
556 return del;
Andy Whitcroft96822902008-07-23 21:27:29 -0700557}
558
Mike Kravetz1dd308a2015-06-24 16:57:52 -0700559/*
Mike Kravetzb5cec282015-09-08 15:01:41 -0700560 * A rare out of memory error was encountered which prevented removal of
561 * the reserve map region for a page. The huge page itself was free'ed
562 * and removed from the page cache. This routine will adjust the subpool
563 * usage count, and the global reserve count if needed. By incrementing
564 * these counts, the reserve map entry which could not be deleted will
565 * appear as a "reserved" entry instead of simply dangling with incorrect
566 * counts.
567 */
568void hugetlb_fix_reserve_counts(struct inode *inode, bool restore_reserve)
569{
570 struct hugepage_subpool *spool = subpool_inode(inode);
571 long rsv_adjust;
572
573 rsv_adjust = hugepage_subpool_get_pages(spool, 1);
574 if (restore_reserve && rsv_adjust) {
575 struct hstate *h = hstate_inode(inode);
576
577 hugetlb_acct_memory(h, 1);
578 }
579}
580
581/*
Mike Kravetz1dd308a2015-06-24 16:57:52 -0700582 * Count and return the number of huge pages in the reserve map
583 * that intersect with the range [f, t).
584 */
Joonsoo Kim1406ec92014-04-03 14:47:26 -0700585static long region_count(struct resv_map *resv, long f, long t)
Andy Whitcroft84afd992008-07-23 21:27:32 -0700586{
Joonsoo Kim1406ec92014-04-03 14:47:26 -0700587 struct list_head *head = &resv->regions;
Andy Whitcroft84afd992008-07-23 21:27:32 -0700588 struct file_region *rg;
589 long chg = 0;
590
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700591 spin_lock(&resv->lock);
Andy Whitcroft84afd992008-07-23 21:27:32 -0700592 /* Locate each segment we overlap with, and count that overlap. */
593 list_for_each_entry(rg, head, link) {
Wang Sheng-Huif2135a42012-05-29 15:06:17 -0700594 long seg_from;
595 long seg_to;
Andy Whitcroft84afd992008-07-23 21:27:32 -0700596
597 if (rg->to <= f)
598 continue;
599 if (rg->from >= t)
600 break;
601
602 seg_from = max(rg->from, f);
603 seg_to = min(rg->to, t);
604
605 chg += seg_to - seg_from;
606 }
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700607 spin_unlock(&resv->lock);
Andy Whitcroft84afd992008-07-23 21:27:32 -0700608
609 return chg;
610}
611
Andy Whitcroft96822902008-07-23 21:27:29 -0700612/*
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700613 * Convert the address within this vma to the page offset within
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700614 * the mapping, in pagecache page units; huge pages here.
615 */
Andi Kleena5516432008-07-23 21:27:41 -0700616static pgoff_t vma_hugecache_offset(struct hstate *h,
617 struct vm_area_struct *vma, unsigned long address)
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700618{
Andi Kleena5516432008-07-23 21:27:41 -0700619 return ((address - vma->vm_start) >> huge_page_shift(h)) +
620 (vma->vm_pgoff >> huge_page_order(h));
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700621}
622
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +0900623pgoff_t linear_hugepage_index(struct vm_area_struct *vma,
624 unsigned long address)
625{
626 return vma_hugecache_offset(hstate_vma(vma), vma, address);
627}
628
Andy Whitcroft84afd992008-07-23 21:27:32 -0700629/*
Mel Gorman08fba692009-01-06 14:38:53 -0800630 * Return the size of the pages allocated when backing a VMA. In the majority
631 * cases this will be same size as used by the page table entries.
632 */
633unsigned long vma_kernel_pagesize(struct vm_area_struct *vma)
634{
635 struct hstate *hstate;
636
637 if (!is_vm_hugetlb_page(vma))
638 return PAGE_SIZE;
639
640 hstate = hstate_vma(vma);
641
Wanpeng Li2415cf12013-07-03 15:02:43 -0700642 return 1UL << huge_page_shift(hstate);
Mel Gorman08fba692009-01-06 14:38:53 -0800643}
Joerg Roedelf340ca02009-06-19 15:16:22 +0200644EXPORT_SYMBOL_GPL(vma_kernel_pagesize);
Mel Gorman08fba692009-01-06 14:38:53 -0800645
646/*
Mel Gorman33402892009-01-06 14:38:54 -0800647 * Return the page size being used by the MMU to back a VMA. In the majority
648 * of cases, the page size used by the kernel matches the MMU size. On
649 * architectures where it differs, an architecture-specific version of this
650 * function is required.
651 */
652#ifndef vma_mmu_pagesize
653unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
654{
655 return vma_kernel_pagesize(vma);
656}
657#endif
658
659/*
Andy Whitcroft84afd992008-07-23 21:27:32 -0700660 * Flags for MAP_PRIVATE reservations. These are stored in the bottom
661 * bits of the reservation map pointer, which are always clear due to
662 * alignment.
663 */
664#define HPAGE_RESV_OWNER (1UL << 0)
665#define HPAGE_RESV_UNMAPPED (1UL << 1)
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700666#define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
Andy Whitcroft84afd992008-07-23 21:27:32 -0700667
Mel Gormana1e78772008-07-23 21:27:23 -0700668/*
669 * These helpers are used to track how many pages are reserved for
670 * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
671 * is guaranteed to have their future faults succeed.
672 *
673 * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
674 * the reserve counters are updated with the hugetlb_lock held. It is safe
675 * to reset the VMA at fork() time as it is not in use yet and there is no
676 * chance of the global counters getting corrupted as a result of the values.
Andy Whitcroft84afd992008-07-23 21:27:32 -0700677 *
678 * The private mapping reservation is represented in a subtly different
679 * manner to a shared mapping. A shared mapping has a region map associated
680 * with the underlying file, this region map represents the backing file
681 * pages which have ever had a reservation assigned which this persists even
682 * after the page is instantiated. A private mapping has a region map
683 * associated with the original mmap which is attached to all VMAs which
684 * reference it, this region map represents those offsets which have consumed
685 * reservation ie. where pages have been instantiated.
Mel Gormana1e78772008-07-23 21:27:23 -0700686 */
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700687static unsigned long get_vma_private_data(struct vm_area_struct *vma)
688{
689 return (unsigned long)vma->vm_private_data;
690}
691
692static void set_vma_private_data(struct vm_area_struct *vma,
693 unsigned long value)
694{
695 vma->vm_private_data = (void *)value;
696}
697
Joonsoo Kim9119a412014-04-03 14:47:25 -0700698struct resv_map *resv_map_alloc(void)
Andy Whitcroft84afd992008-07-23 21:27:32 -0700699{
700 struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
Mike Kravetz5e911372015-09-08 15:01:28 -0700701 struct file_region *rg = kmalloc(sizeof(*rg), GFP_KERNEL);
702
703 if (!resv_map || !rg) {
704 kfree(resv_map);
705 kfree(rg);
Andy Whitcroft84afd992008-07-23 21:27:32 -0700706 return NULL;
Mike Kravetz5e911372015-09-08 15:01:28 -0700707 }
Andy Whitcroft84afd992008-07-23 21:27:32 -0700708
709 kref_init(&resv_map->refs);
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700710 spin_lock_init(&resv_map->lock);
Andy Whitcroft84afd992008-07-23 21:27:32 -0700711 INIT_LIST_HEAD(&resv_map->regions);
712
Mike Kravetz5e911372015-09-08 15:01:28 -0700713 resv_map->adds_in_progress = 0;
714
715 INIT_LIST_HEAD(&resv_map->region_cache);
716 list_add(&rg->link, &resv_map->region_cache);
717 resv_map->region_cache_count = 1;
718
Andy Whitcroft84afd992008-07-23 21:27:32 -0700719 return resv_map;
720}
721
Joonsoo Kim9119a412014-04-03 14:47:25 -0700722void resv_map_release(struct kref *ref)
Andy Whitcroft84afd992008-07-23 21:27:32 -0700723{
724 struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
Mike Kravetz5e911372015-09-08 15:01:28 -0700725 struct list_head *head = &resv_map->region_cache;
726 struct file_region *rg, *trg;
Andy Whitcroft84afd992008-07-23 21:27:32 -0700727
728 /* Clear out any active regions before we release the map. */
Mike Kravetzfeba16e2015-09-08 15:01:31 -0700729 region_del(resv_map, 0, LONG_MAX);
Mike Kravetz5e911372015-09-08 15:01:28 -0700730
731 /* ... and any entries left in the cache */
732 list_for_each_entry_safe(rg, trg, head, link) {
733 list_del(&rg->link);
734 kfree(rg);
735 }
736
737 VM_BUG_ON(resv_map->adds_in_progress);
738
Andy Whitcroft84afd992008-07-23 21:27:32 -0700739 kfree(resv_map);
740}
741
Joonsoo Kim4e35f482014-04-03 14:47:30 -0700742static inline struct resv_map *inode_resv_map(struct inode *inode)
743{
744 return inode->i_mapping->private_data;
745}
746
Andy Whitcroft84afd992008-07-23 21:27:32 -0700747static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
Mel Gormana1e78772008-07-23 21:27:23 -0700748{
Sasha Levin81d1b092014-10-09 15:28:10 -0700749 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
Joonsoo Kim4e35f482014-04-03 14:47:30 -0700750 if (vma->vm_flags & VM_MAYSHARE) {
751 struct address_space *mapping = vma->vm_file->f_mapping;
752 struct inode *inode = mapping->host;
753
754 return inode_resv_map(inode);
755
756 } else {
Andy Whitcroft84afd992008-07-23 21:27:32 -0700757 return (struct resv_map *)(get_vma_private_data(vma) &
758 ~HPAGE_RESV_MASK);
Joonsoo Kim4e35f482014-04-03 14:47:30 -0700759 }
Mel Gormana1e78772008-07-23 21:27:23 -0700760}
761
Andy Whitcroft84afd992008-07-23 21:27:32 -0700762static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
Mel Gormana1e78772008-07-23 21:27:23 -0700763{
Sasha Levin81d1b092014-10-09 15:28:10 -0700764 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
765 VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
Mel Gormana1e78772008-07-23 21:27:23 -0700766
Andy Whitcroft84afd992008-07-23 21:27:32 -0700767 set_vma_private_data(vma, (get_vma_private_data(vma) &
768 HPAGE_RESV_MASK) | (unsigned long)map);
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700769}
770
771static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
772{
Sasha Levin81d1b092014-10-09 15:28:10 -0700773 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
774 VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700775
776 set_vma_private_data(vma, get_vma_private_data(vma) | flags);
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700777}
778
779static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
780{
Sasha Levin81d1b092014-10-09 15:28:10 -0700781 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700782
783 return (get_vma_private_data(vma) & flag) != 0;
Mel Gormana1e78772008-07-23 21:27:23 -0700784}
785
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700786/* Reset counters to 0 and clear all HPAGE_RESV_* flags */
Mel Gormana1e78772008-07-23 21:27:23 -0700787void reset_vma_resv_huge_pages(struct vm_area_struct *vma)
788{
Sasha Levin81d1b092014-10-09 15:28:10 -0700789 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
Mel Gormanf83a2752009-05-28 14:34:40 -0700790 if (!(vma->vm_flags & VM_MAYSHARE))
Mel Gormana1e78772008-07-23 21:27:23 -0700791 vma->vm_private_data = (void *)0;
792}
793
794/* Returns true if the VMA has associated reserve pages */
Nicholas Krause559ec2f2015-09-04 15:48:27 -0700795static bool vma_has_reserves(struct vm_area_struct *vma, long chg)
Mel Gormana1e78772008-07-23 21:27:23 -0700796{
Joonsoo Kimaf0ed732013-09-11 14:21:18 -0700797 if (vma->vm_flags & VM_NORESERVE) {
798 /*
799 * This address is already reserved by other process(chg == 0),
800 * so, we should decrement reserved count. Without decrementing,
801 * reserve count remains after releasing inode, because this
802 * allocated page will go into page cache and is regarded as
803 * coming from reserved pool in releasing step. Currently, we
804 * don't have any other solution to deal with this situation
805 * properly, so add work-around here.
806 */
807 if (vma->vm_flags & VM_MAYSHARE && chg == 0)
Nicholas Krause559ec2f2015-09-04 15:48:27 -0700808 return true;
Joonsoo Kimaf0ed732013-09-11 14:21:18 -0700809 else
Nicholas Krause559ec2f2015-09-04 15:48:27 -0700810 return false;
Joonsoo Kimaf0ed732013-09-11 14:21:18 -0700811 }
Joonsoo Kima63884e2013-09-11 14:21:07 -0700812
813 /* Shared mappings always use reserves */
Mike Kravetz1fb1b0e2015-09-08 15:01:44 -0700814 if (vma->vm_flags & VM_MAYSHARE) {
815 /*
816 * We know VM_NORESERVE is not set. Therefore, there SHOULD
817 * be a region map for all pages. The only situation where
818 * there is no region map is if a hole was punched via
819 * fallocate. In this case, there really are no reverves to
820 * use. This situation is indicated if chg != 0.
821 */
822 if (chg)
823 return false;
824 else
825 return true;
826 }
Joonsoo Kima63884e2013-09-11 14:21:07 -0700827
828 /*
829 * Only the process that called mmap() has reserves for
830 * private mappings.
831 */
Mel Gorman7f09ca52008-07-23 21:27:58 -0700832 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER))
Nicholas Krause559ec2f2015-09-04 15:48:27 -0700833 return true;
Joonsoo Kima63884e2013-09-11 14:21:07 -0700834
Nicholas Krause559ec2f2015-09-04 15:48:27 -0700835 return false;
Mel Gormana1e78772008-07-23 21:27:23 -0700836}
837
Andi Kleena5516432008-07-23 21:27:41 -0700838static void enqueue_huge_page(struct hstate *h, struct page *page)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700839{
840 int nid = page_to_nid(page);
Aneesh Kumar K.V0edaecf2012-07-31 16:42:07 -0700841 list_move(&page->lru, &h->hugepage_freelists[nid]);
Andi Kleena5516432008-07-23 21:27:41 -0700842 h->free_huge_pages++;
843 h->free_huge_pages_node[nid]++;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700844}
845
Naoya Horiguchibf50bab2010-09-08 10:19:33 +0900846static struct page *dequeue_huge_page_node(struct hstate *h, int nid)
847{
848 struct page *page;
849
Naoya Horiguchic8721bb2013-09-11 14:22:09 -0700850 list_for_each_entry(page, &h->hugepage_freelists[nid], lru)
851 if (!is_migrate_isolate_page(page))
852 break;
853 /*
854 * if 'non-isolated free hugepage' not found on the list,
855 * the allocation fails.
856 */
857 if (&h->hugepage_freelists[nid] == &page->lru)
Naoya Horiguchibf50bab2010-09-08 10:19:33 +0900858 return NULL;
Aneesh Kumar K.V0edaecf2012-07-31 16:42:07 -0700859 list_move(&page->lru, &h->hugepage_activelist);
Naoya Horiguchia9869b82010-09-08 10:19:37 +0900860 set_page_refcounted(page);
Naoya Horiguchibf50bab2010-09-08 10:19:33 +0900861 h->free_huge_pages--;
862 h->free_huge_pages_node[nid]--;
863 return page;
864}
865
Naoya Horiguchi86cdb462013-09-11 14:22:13 -0700866/* Movability of hugepages depends on migration support. */
867static inline gfp_t htlb_alloc_mask(struct hstate *h)
868{
Naoya Horiguchi100873d2014-06-04 16:10:56 -0700869 if (hugepages_treat_as_movable || hugepage_migration_supported(h))
Naoya Horiguchi86cdb462013-09-11 14:22:13 -0700870 return GFP_HIGHUSER_MOVABLE;
871 else
872 return GFP_HIGHUSER;
873}
874
Andi Kleena5516432008-07-23 21:27:41 -0700875static struct page *dequeue_huge_page_vma(struct hstate *h,
876 struct vm_area_struct *vma,
Joonsoo Kimaf0ed732013-09-11 14:21:18 -0700877 unsigned long address, int avoid_reserve,
878 long chg)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700879{
Konstantin Khlebnikovb1c12cb2012-04-25 16:01:46 -0700880 struct page *page = NULL;
Lee Schermerhorn480eccf2007-09-18 22:46:47 -0700881 struct mempolicy *mpol;
Mel Gorman19770b32008-04-28 02:12:18 -0700882 nodemask_t *nodemask;
Miao Xiec0ff7452010-05-24 14:32:08 -0700883 struct zonelist *zonelist;
Mel Gormandd1a2392008-04-28 02:12:17 -0700884 struct zone *zone;
885 struct zoneref *z;
Mel Gormancc9a6c82012-03-21 16:34:11 -0700886 unsigned int cpuset_mems_cookie;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700887
Mel Gormana1e78772008-07-23 21:27:23 -0700888 /*
889 * A child process with MAP_PRIVATE mappings created by their parent
890 * have no page reserves. This check ensures that reservations are
891 * not "stolen". The child may still get SIGKILLed
892 */
Joonsoo Kimaf0ed732013-09-11 14:21:18 -0700893 if (!vma_has_reserves(vma, chg) &&
Andi Kleena5516432008-07-23 21:27:41 -0700894 h->free_huge_pages - h->resv_huge_pages == 0)
Miao Xiec0ff7452010-05-24 14:32:08 -0700895 goto err;
Mel Gormana1e78772008-07-23 21:27:23 -0700896
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700897 /* If reserves cannot be used, ensure enough pages are in the pool */
Andi Kleena5516432008-07-23 21:27:41 -0700898 if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
Justin P. Mattock6eab04a2011-04-08 19:49:08 -0700899 goto err;
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700900
Joonsoo Kim9966c4b2013-09-11 14:20:50 -0700901retry_cpuset:
Mel Gormand26914d2014-04-03 14:47:24 -0700902 cpuset_mems_cookie = read_mems_allowed_begin();
Joonsoo Kim9966c4b2013-09-11 14:20:50 -0700903 zonelist = huge_zonelist(vma, address,
Naoya Horiguchi86cdb462013-09-11 14:22:13 -0700904 htlb_alloc_mask(h), &mpol, &nodemask);
Joonsoo Kim9966c4b2013-09-11 14:20:50 -0700905
Mel Gorman19770b32008-04-28 02:12:18 -0700906 for_each_zone_zonelist_nodemask(zone, z, zonelist,
907 MAX_NR_ZONES - 1, nodemask) {
Vladimir Davydov344736f2014-10-20 15:50:30 +0400908 if (cpuset_zone_allowed(zone, htlb_alloc_mask(h))) {
Naoya Horiguchibf50bab2010-09-08 10:19:33 +0900909 page = dequeue_huge_page_node(h, zone_to_nid(zone));
910 if (page) {
Joonsoo Kimaf0ed732013-09-11 14:21:18 -0700911 if (avoid_reserve)
912 break;
913 if (!vma_has_reserves(vma, chg))
914 break;
915
Joonsoo Kim07443a82013-09-11 14:21:58 -0700916 SetPagePrivate(page);
Joonsoo Kimaf0ed732013-09-11 14:21:18 -0700917 h->resv_huge_pages--;
Naoya Horiguchibf50bab2010-09-08 10:19:33 +0900918 break;
919 }
Andrew Morton3abf7af2007-07-19 01:49:08 -0700920 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700921 }
Mel Gormancc9a6c82012-03-21 16:34:11 -0700922
923 mpol_cond_put(mpol);
Mel Gormand26914d2014-04-03 14:47:24 -0700924 if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie)))
Mel Gormancc9a6c82012-03-21 16:34:11 -0700925 goto retry_cpuset;
926 return page;
927
Miao Xiec0ff7452010-05-24 14:32:08 -0700928err:
Mel Gormancc9a6c82012-03-21 16:34:11 -0700929 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700930}
931
Luiz Capitulino1cac6f22014-06-04 16:07:11 -0700932/*
933 * common helper functions for hstate_next_node_to_{alloc|free}.
934 * We may have allocated or freed a huge page based on a different
935 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
936 * be outside of *nodes_allowed. Ensure that we use an allowed
937 * node for alloc or free.
938 */
939static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
940{
941 nid = next_node(nid, *nodes_allowed);
942 if (nid == MAX_NUMNODES)
943 nid = first_node(*nodes_allowed);
944 VM_BUG_ON(nid >= MAX_NUMNODES);
945
946 return nid;
947}
948
949static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
950{
951 if (!node_isset(nid, *nodes_allowed))
952 nid = next_node_allowed(nid, nodes_allowed);
953 return nid;
954}
955
956/*
957 * returns the previously saved node ["this node"] from which to
958 * allocate a persistent huge page for the pool and advance the
959 * next node from which to allocate, handling wrap at end of node
960 * mask.
961 */
962static int hstate_next_node_to_alloc(struct hstate *h,
963 nodemask_t *nodes_allowed)
964{
965 int nid;
966
967 VM_BUG_ON(!nodes_allowed);
968
969 nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
970 h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
971
972 return nid;
973}
974
975/*
976 * helper for free_pool_huge_page() - return the previously saved
977 * node ["this node"] from which to free a huge page. Advance the
978 * next node id whether or not we find a free huge page to free so
979 * that the next attempt to free addresses the next node.
980 */
981static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
982{
983 int nid;
984
985 VM_BUG_ON(!nodes_allowed);
986
987 nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
988 h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
989
990 return nid;
991}
992
993#define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \
994 for (nr_nodes = nodes_weight(*mask); \
995 nr_nodes > 0 && \
996 ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \
997 nr_nodes--)
998
999#define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \
1000 for (nr_nodes = nodes_weight(*mask); \
1001 nr_nodes > 0 && \
1002 ((node = hstate_next_node_to_free(hs, mask)) || 1); \
1003 nr_nodes--)
1004
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001005#if defined(CONFIG_CMA) && defined(CONFIG_X86_64)
1006static void destroy_compound_gigantic_page(struct page *page,
Kirill A. Shutemovd00181b2015-11-06 16:29:57 -08001007 unsigned int order)
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001008{
1009 int i;
1010 int nr_pages = 1 << order;
1011 struct page *p = page + 1;
1012
1013 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
Kirill A. Shutemov1d798ca2015-11-06 16:29:54 -08001014 clear_compound_head(p);
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001015 set_page_refcounted(p);
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001016 }
1017
1018 set_compound_order(page, 0);
1019 __ClearPageHead(page);
1020}
1021
Kirill A. Shutemovd00181b2015-11-06 16:29:57 -08001022static void free_gigantic_page(struct page *page, unsigned int order)
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001023{
1024 free_contig_range(page_to_pfn(page), 1 << order);
1025}
1026
1027static int __alloc_gigantic_page(unsigned long start_pfn,
1028 unsigned long nr_pages)
1029{
1030 unsigned long end_pfn = start_pfn + nr_pages;
1031 return alloc_contig_range(start_pfn, end_pfn, MIGRATE_MOVABLE);
1032}
1033
1034static bool pfn_range_valid_gigantic(unsigned long start_pfn,
1035 unsigned long nr_pages)
1036{
1037 unsigned long i, end_pfn = start_pfn + nr_pages;
1038 struct page *page;
1039
1040 for (i = start_pfn; i < end_pfn; i++) {
1041 if (!pfn_valid(i))
1042 return false;
1043
1044 page = pfn_to_page(i);
1045
1046 if (PageReserved(page))
1047 return false;
1048
1049 if (page_count(page) > 0)
1050 return false;
1051
1052 if (PageHuge(page))
1053 return false;
1054 }
1055
1056 return true;
1057}
1058
1059static bool zone_spans_last_pfn(const struct zone *zone,
1060 unsigned long start_pfn, unsigned long nr_pages)
1061{
1062 unsigned long last_pfn = start_pfn + nr_pages - 1;
1063 return zone_spans_pfn(zone, last_pfn);
1064}
1065
Kirill A. Shutemovd00181b2015-11-06 16:29:57 -08001066static struct page *alloc_gigantic_page(int nid, unsigned int order)
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001067{
1068 unsigned long nr_pages = 1 << order;
1069 unsigned long ret, pfn, flags;
1070 struct zone *z;
1071
1072 z = NODE_DATA(nid)->node_zones;
1073 for (; z - NODE_DATA(nid)->node_zones < MAX_NR_ZONES; z++) {
1074 spin_lock_irqsave(&z->lock, flags);
1075
1076 pfn = ALIGN(z->zone_start_pfn, nr_pages);
1077 while (zone_spans_last_pfn(z, pfn, nr_pages)) {
1078 if (pfn_range_valid_gigantic(pfn, nr_pages)) {
1079 /*
1080 * We release the zone lock here because
1081 * alloc_contig_range() will also lock the zone
1082 * at some point. If there's an allocation
1083 * spinning on this lock, it may win the race
1084 * and cause alloc_contig_range() to fail...
1085 */
1086 spin_unlock_irqrestore(&z->lock, flags);
1087 ret = __alloc_gigantic_page(pfn, nr_pages);
1088 if (!ret)
1089 return pfn_to_page(pfn);
1090 spin_lock_irqsave(&z->lock, flags);
1091 }
1092 pfn += nr_pages;
1093 }
1094
1095 spin_unlock_irqrestore(&z->lock, flags);
1096 }
1097
1098 return NULL;
1099}
1100
1101static void prep_new_huge_page(struct hstate *h, struct page *page, int nid);
Kirill A. Shutemovd00181b2015-11-06 16:29:57 -08001102static void prep_compound_gigantic_page(struct page *page, unsigned int order);
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001103
1104static struct page *alloc_fresh_gigantic_page_node(struct hstate *h, int nid)
1105{
1106 struct page *page;
1107
1108 page = alloc_gigantic_page(nid, huge_page_order(h));
1109 if (page) {
1110 prep_compound_gigantic_page(page, huge_page_order(h));
1111 prep_new_huge_page(h, page, nid);
1112 }
1113
1114 return page;
1115}
1116
1117static int alloc_fresh_gigantic_page(struct hstate *h,
1118 nodemask_t *nodes_allowed)
1119{
1120 struct page *page = NULL;
1121 int nr_nodes, node;
1122
1123 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
1124 page = alloc_fresh_gigantic_page_node(h, node);
1125 if (page)
1126 return 1;
1127 }
1128
1129 return 0;
1130}
1131
1132static inline bool gigantic_page_supported(void) { return true; }
1133#else
1134static inline bool gigantic_page_supported(void) { return false; }
Kirill A. Shutemovd00181b2015-11-06 16:29:57 -08001135static inline void free_gigantic_page(struct page *page, unsigned int order) { }
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001136static inline void destroy_compound_gigantic_page(struct page *page,
Kirill A. Shutemovd00181b2015-11-06 16:29:57 -08001137 unsigned int order) { }
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001138static inline int alloc_fresh_gigantic_page(struct hstate *h,
1139 nodemask_t *nodes_allowed) { return 0; }
1140#endif
1141
Andi Kleena5516432008-07-23 21:27:41 -07001142static void update_and_free_page(struct hstate *h, struct page *page)
Adam Litke6af2acb2007-10-16 01:26:16 -07001143{
1144 int i;
Andi Kleena5516432008-07-23 21:27:41 -07001145
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001146 if (hstate_is_gigantic(h) && !gigantic_page_supported())
1147 return;
Andy Whitcroft18229df2008-11-06 12:53:27 -08001148
Andi Kleena5516432008-07-23 21:27:41 -07001149 h->nr_huge_pages--;
1150 h->nr_huge_pages_node[page_to_nid(page)]--;
1151 for (i = 0; i < pages_per_huge_page(h); i++) {
Chris Forbes32f84522011-07-25 17:12:14 -07001152 page[i].flags &= ~(1 << PG_locked | 1 << PG_error |
1153 1 << PG_referenced | 1 << PG_dirty |
Luiz Capitulinoa7407a22014-06-04 16:07:09 -07001154 1 << PG_active | 1 << PG_private |
1155 1 << PG_writeback);
Adam Litke6af2acb2007-10-16 01:26:16 -07001156 }
Sasha Levin309381fea2014-01-23 15:52:54 -08001157 VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page);
Kirill A. Shutemovf1e61552015-11-06 16:29:50 -08001158 set_compound_page_dtor(page, NULL_COMPOUND_DTOR);
Adam Litke6af2acb2007-10-16 01:26:16 -07001159 set_page_refcounted(page);
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001160 if (hstate_is_gigantic(h)) {
1161 destroy_compound_gigantic_page(page, huge_page_order(h));
1162 free_gigantic_page(page, huge_page_order(h));
1163 } else {
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001164 __free_pages(page, huge_page_order(h));
1165 }
Adam Litke6af2acb2007-10-16 01:26:16 -07001166}
1167
Andi Kleene5ff2152008-07-23 21:27:42 -07001168struct hstate *size_to_hstate(unsigned long size)
1169{
1170 struct hstate *h;
1171
1172 for_each_hstate(h) {
1173 if (huge_page_size(h) == size)
1174 return h;
1175 }
1176 return NULL;
1177}
1178
Naoya Horiguchibcc54222015-04-15 16:14:38 -07001179/*
1180 * Test to determine whether the hugepage is "active/in-use" (i.e. being linked
1181 * to hstate->hugepage_activelist.)
1182 *
1183 * This function can be called for tail pages, but never returns true for them.
1184 */
1185bool page_huge_active(struct page *page)
1186{
1187 VM_BUG_ON_PAGE(!PageHuge(page), page);
1188 return PageHead(page) && PagePrivate(&page[1]);
1189}
1190
1191/* never called for tail page */
1192static void set_page_huge_active(struct page *page)
1193{
1194 VM_BUG_ON_PAGE(!PageHeadHuge(page), page);
1195 SetPagePrivate(&page[1]);
1196}
1197
1198static void clear_page_huge_active(struct page *page)
1199{
1200 VM_BUG_ON_PAGE(!PageHeadHuge(page), page);
1201 ClearPagePrivate(&page[1]);
1202}
1203
Atsushi Kumagai8f1d26d2014-07-30 16:08:39 -07001204void free_huge_page(struct page *page)
David Gibson27a85ef2006-03-22 00:08:56 -08001205{
Andi Kleena5516432008-07-23 21:27:41 -07001206 /*
1207 * Can't pass hstate in here because it is called from the
1208 * compound page destructor.
1209 */
Andi Kleene5ff2152008-07-23 21:27:42 -07001210 struct hstate *h = page_hstate(page);
Adam Litke7893d1d2007-10-16 01:26:18 -07001211 int nid = page_to_nid(page);
David Gibson90481622012-03-21 16:34:12 -07001212 struct hugepage_subpool *spool =
1213 (struct hugepage_subpool *)page_private(page);
Joonsoo Kim07443a82013-09-11 14:21:58 -07001214 bool restore_reserve;
David Gibson27a85ef2006-03-22 00:08:56 -08001215
Andy Whitcrofte5df70a2008-02-23 15:23:32 -08001216 set_page_private(page, 0);
Mel Gorman23be7462010-04-23 13:17:56 -04001217 page->mapping = NULL;
Adam Litke7893d1d2007-10-16 01:26:18 -07001218 BUG_ON(page_count(page));
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09001219 BUG_ON(page_mapcount(page));
Joonsoo Kim07443a82013-09-11 14:21:58 -07001220 restore_reserve = PagePrivate(page);
Joonsoo Kim16c794b2013-10-16 13:46:48 -07001221 ClearPagePrivate(page);
David Gibson27a85ef2006-03-22 00:08:56 -08001222
Mike Kravetz1c5ecae2015-04-15 16:13:39 -07001223 /*
1224 * A return code of zero implies that the subpool will be under its
1225 * minimum size if the reservation is not restored after page is free.
1226 * Therefore, force restore_reserve operation.
1227 */
1228 if (hugepage_subpool_put_pages(spool, 1) == 0)
1229 restore_reserve = true;
1230
David Gibson27a85ef2006-03-22 00:08:56 -08001231 spin_lock(&hugetlb_lock);
Naoya Horiguchibcc54222015-04-15 16:14:38 -07001232 clear_page_huge_active(page);
Aneesh Kumar K.V6d76dcf2012-07-31 16:42:18 -07001233 hugetlb_cgroup_uncharge_page(hstate_index(h),
1234 pages_per_huge_page(h), page);
Joonsoo Kim07443a82013-09-11 14:21:58 -07001235 if (restore_reserve)
1236 h->resv_huge_pages++;
1237
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001238 if (h->surplus_huge_pages_node[nid]) {
Aneesh Kumar K.V0edaecf2012-07-31 16:42:07 -07001239 /* remove the page from active list */
1240 list_del(&page->lru);
Andi Kleena5516432008-07-23 21:27:41 -07001241 update_and_free_page(h, page);
1242 h->surplus_huge_pages--;
1243 h->surplus_huge_pages_node[nid]--;
Adam Litke7893d1d2007-10-16 01:26:18 -07001244 } else {
Will Deacon5d3a5512012-10-08 16:29:32 -07001245 arch_clear_hugepage_flags(page);
Andi Kleena5516432008-07-23 21:27:41 -07001246 enqueue_huge_page(h, page);
Adam Litke7893d1d2007-10-16 01:26:18 -07001247 }
David Gibson27a85ef2006-03-22 00:08:56 -08001248 spin_unlock(&hugetlb_lock);
1249}
1250
Andi Kleena5516432008-07-23 21:27:41 -07001251static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
Andi Kleenb7ba30c2008-07-23 21:27:40 -07001252{
Aneesh Kumar K.V0edaecf2012-07-31 16:42:07 -07001253 INIT_LIST_HEAD(&page->lru);
Kirill A. Shutemovf1e61552015-11-06 16:29:50 -08001254 set_compound_page_dtor(page, HUGETLB_PAGE_DTOR);
Andi Kleenb7ba30c2008-07-23 21:27:40 -07001255 spin_lock(&hugetlb_lock);
Aneesh Kumar K.V9dd540e2012-07-31 16:42:15 -07001256 set_hugetlb_cgroup(page, NULL);
Andi Kleena5516432008-07-23 21:27:41 -07001257 h->nr_huge_pages++;
1258 h->nr_huge_pages_node[nid]++;
Andi Kleenb7ba30c2008-07-23 21:27:40 -07001259 spin_unlock(&hugetlb_lock);
1260 put_page(page); /* free it into the hugepage allocator */
1261}
1262
Kirill A. Shutemovd00181b2015-11-06 16:29:57 -08001263static void prep_compound_gigantic_page(struct page *page, unsigned int order)
Wu Fengguang20a03072009-06-16 15:32:22 -07001264{
1265 int i;
1266 int nr_pages = 1 << order;
1267 struct page *p = page + 1;
1268
1269 /* we rely on prep_new_huge_page to set the destructor */
1270 set_compound_order(page, order);
1271 __SetPageHead(page);
Andrea Arcangelief5a22b2013-10-16 13:46:56 -07001272 __ClearPageReserved(page);
Wu Fengguang20a03072009-06-16 15:32:22 -07001273 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
Andrea Arcangelief5a22b2013-10-16 13:46:56 -07001274 /*
1275 * For gigantic hugepages allocated through bootmem at
1276 * boot, it's safer to be consistent with the not-gigantic
1277 * hugepages and clear the PG_reserved bit from all tail pages
1278 * too. Otherwse drivers using get_user_pages() to access tail
1279 * pages may get the reference counting wrong if they see
1280 * PG_reserved set on a tail page (despite the head page not
1281 * having PG_reserved set). Enforcing this consistency between
1282 * head and tail pages allows drivers to optimize away a check
1283 * on the head page when they need know if put_page() is needed
1284 * after get_user_pages().
1285 */
1286 __ClearPageReserved(p);
Youquan Song58a84aa2011-12-08 14:34:18 -08001287 set_page_count(p, 0);
Kirill A. Shutemov1d798ca2015-11-06 16:29:54 -08001288 set_compound_head(p, page);
Wu Fengguang20a03072009-06-16 15:32:22 -07001289 }
1290}
1291
Andrew Morton77959122012-10-08 16:34:11 -07001292/*
1293 * PageHuge() only returns true for hugetlbfs pages, but not for normal or
1294 * transparent huge pages. See the PageTransHuge() documentation for more
1295 * details.
1296 */
Wu Fengguang20a03072009-06-16 15:32:22 -07001297int PageHuge(struct page *page)
1298{
Wu Fengguang20a03072009-06-16 15:32:22 -07001299 if (!PageCompound(page))
1300 return 0;
1301
1302 page = compound_head(page);
Kirill A. Shutemovf1e61552015-11-06 16:29:50 -08001303 return page[1].compound_dtor == HUGETLB_PAGE_DTOR;
Wu Fengguang20a03072009-06-16 15:32:22 -07001304}
Naoya Horiguchi43131e12010-05-28 09:29:22 +09001305EXPORT_SYMBOL_GPL(PageHuge);
1306
Andrea Arcangeli27c73ae2013-11-21 14:32:02 -08001307/*
1308 * PageHeadHuge() only returns true for hugetlbfs head page, but not for
1309 * normal or transparent huge pages.
1310 */
1311int PageHeadHuge(struct page *page_head)
1312{
Andrea Arcangeli27c73ae2013-11-21 14:32:02 -08001313 if (!PageHead(page_head))
1314 return 0;
1315
Andrew Morton758f66a2014-01-21 15:48:57 -08001316 return get_compound_page_dtor(page_head) == free_huge_page;
Andrea Arcangeli27c73ae2013-11-21 14:32:02 -08001317}
Andrea Arcangeli27c73ae2013-11-21 14:32:02 -08001318
Zhang Yi13d60f42013-06-25 21:19:31 +08001319pgoff_t __basepage_index(struct page *page)
1320{
1321 struct page *page_head = compound_head(page);
1322 pgoff_t index = page_index(page_head);
1323 unsigned long compound_idx;
1324
1325 if (!PageHuge(page_head))
1326 return page_index(page);
1327
1328 if (compound_order(page_head) >= MAX_ORDER)
1329 compound_idx = page_to_pfn(page) - page_to_pfn(page_head);
1330 else
1331 compound_idx = page - page_head;
1332
1333 return (index << compound_order(page_head)) + compound_idx;
1334}
1335
Andi Kleena5516432008-07-23 21:27:41 -07001336static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001337{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001338 struct page *page;
Joe Jinf96efd52007-07-15 23:38:12 -07001339
Vlastimil Babka96db8002015-09-08 15:03:50 -07001340 page = __alloc_pages_node(nid,
Naoya Horiguchi86cdb462013-09-11 14:22:13 -07001341 htlb_alloc_mask(h)|__GFP_COMP|__GFP_THISNODE|
Nishanth Aravamudan551883a2008-04-29 00:58:26 -07001342 __GFP_REPEAT|__GFP_NOWARN,
Andi Kleena5516432008-07-23 21:27:41 -07001343 huge_page_order(h));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001344 if (page) {
Andi Kleena5516432008-07-23 21:27:41 -07001345 prep_new_huge_page(h, page, nid);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001346 }
Nishanth Aravamudan63b46132007-10-16 01:26:24 -07001347
1348 return page;
1349}
1350
Joonsoo Kimb2261022013-09-11 14:21:00 -07001351static int alloc_fresh_huge_page(struct hstate *h, nodemask_t *nodes_allowed)
1352{
1353 struct page *page;
1354 int nr_nodes, node;
1355 int ret = 0;
1356
1357 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
1358 page = alloc_fresh_huge_page_node(h, node);
1359 if (page) {
1360 ret = 1;
1361 break;
1362 }
1363 }
1364
1365 if (ret)
1366 count_vm_event(HTLB_BUDDY_PGALLOC);
1367 else
1368 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
1369
1370 return ret;
1371}
1372
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001373/*
1374 * Free huge page from pool from next node to free.
1375 * Attempt to keep persistent huge pages more or less
1376 * balanced over allowed nodes.
1377 * Called with hugetlb_lock locked.
1378 */
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001379static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
1380 bool acct_surplus)
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001381{
Joonsoo Kimb2261022013-09-11 14:21:00 -07001382 int nr_nodes, node;
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001383 int ret = 0;
1384
Joonsoo Kimb2261022013-09-11 14:21:00 -07001385 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
Lee Schermerhorn685f3452009-09-21 17:01:23 -07001386 /*
1387 * If we're returning unused surplus pages, only examine
1388 * nodes with surplus pages.
1389 */
Joonsoo Kimb2261022013-09-11 14:21:00 -07001390 if ((!acct_surplus || h->surplus_huge_pages_node[node]) &&
1391 !list_empty(&h->hugepage_freelists[node])) {
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001392 struct page *page =
Joonsoo Kimb2261022013-09-11 14:21:00 -07001393 list_entry(h->hugepage_freelists[node].next,
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001394 struct page, lru);
1395 list_del(&page->lru);
1396 h->free_huge_pages--;
Joonsoo Kimb2261022013-09-11 14:21:00 -07001397 h->free_huge_pages_node[node]--;
Lee Schermerhorn685f3452009-09-21 17:01:23 -07001398 if (acct_surplus) {
1399 h->surplus_huge_pages--;
Joonsoo Kimb2261022013-09-11 14:21:00 -07001400 h->surplus_huge_pages_node[node]--;
Lee Schermerhorn685f3452009-09-21 17:01:23 -07001401 }
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001402 update_and_free_page(h, page);
1403 ret = 1;
Lee Schermerhorn9a76db02009-12-14 17:58:15 -08001404 break;
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001405 }
Joonsoo Kimb2261022013-09-11 14:21:00 -07001406 }
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001407
1408 return ret;
1409}
1410
Naoya Horiguchic8721bb2013-09-11 14:22:09 -07001411/*
1412 * Dissolve a given free hugepage into free buddy pages. This function does
1413 * nothing for in-use (including surplus) hugepages.
1414 */
1415static void dissolve_free_huge_page(struct page *page)
1416{
1417 spin_lock(&hugetlb_lock);
1418 if (PageHuge(page) && !page_count(page)) {
1419 struct hstate *h = page_hstate(page);
1420 int nid = page_to_nid(page);
1421 list_del(&page->lru);
1422 h->free_huge_pages--;
1423 h->free_huge_pages_node[nid]--;
1424 update_and_free_page(h, page);
1425 }
1426 spin_unlock(&hugetlb_lock);
1427}
1428
1429/*
1430 * Dissolve free hugepages in a given pfn range. Used by memory hotplug to
1431 * make specified memory blocks removable from the system.
1432 * Note that start_pfn should aligned with (minimum) hugepage size.
1433 */
1434void dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn)
1435{
Naoya Horiguchic8721bb2013-09-11 14:22:09 -07001436 unsigned long pfn;
Naoya Horiguchic8721bb2013-09-11 14:22:09 -07001437
Li Zhongd0177632014-08-06 16:07:56 -07001438 if (!hugepages_supported())
1439 return;
1440
Naoya Horiguchi641844f2015-06-24 16:56:59 -07001441 VM_BUG_ON(!IS_ALIGNED(start_pfn, 1 << minimum_order));
1442 for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << minimum_order)
Naoya Horiguchic8721bb2013-09-11 14:22:09 -07001443 dissolve_free_huge_page(pfn_to_page(pfn));
1444}
1445
Dave Hansen099730d2015-11-05 18:50:17 -08001446/*
1447 * There are 3 ways this can get called:
1448 * 1. With vma+addr: we use the VMA's memory policy
1449 * 2. With !vma, but nid=NUMA_NO_NODE: We try to allocate a huge
1450 * page from any node, and let the buddy allocator itself figure
1451 * it out.
1452 * 3. With !vma, but nid!=NUMA_NO_NODE. We allocate a huge page
1453 * strictly from 'nid'
1454 */
1455static struct page *__hugetlb_alloc_buddy_huge_page(struct hstate *h,
1456 struct vm_area_struct *vma, unsigned long addr, int nid)
1457{
1458 int order = huge_page_order(h);
1459 gfp_t gfp = htlb_alloc_mask(h)|__GFP_COMP|__GFP_REPEAT|__GFP_NOWARN;
1460 unsigned int cpuset_mems_cookie;
1461
1462 /*
1463 * We need a VMA to get a memory policy. If we do not
Dave Hansene0ec90e2015-11-05 18:50:20 -08001464 * have one, we use the 'nid' argument.
1465 *
1466 * The mempolicy stuff below has some non-inlined bits
1467 * and calls ->vm_ops. That makes it hard to optimize at
1468 * compile-time, even when NUMA is off and it does
1469 * nothing. This helps the compiler optimize it out.
Dave Hansen099730d2015-11-05 18:50:17 -08001470 */
Dave Hansene0ec90e2015-11-05 18:50:20 -08001471 if (!IS_ENABLED(CONFIG_NUMA) || !vma) {
Dave Hansen099730d2015-11-05 18:50:17 -08001472 /*
1473 * If a specific node is requested, make sure to
1474 * get memory from there, but only when a node
1475 * is explicitly specified.
1476 */
1477 if (nid != NUMA_NO_NODE)
1478 gfp |= __GFP_THISNODE;
1479 /*
1480 * Make sure to call something that can handle
1481 * nid=NUMA_NO_NODE
1482 */
1483 return alloc_pages_node(nid, gfp, order);
1484 }
1485
1486 /*
1487 * OK, so we have a VMA. Fetch the mempolicy and try to
Dave Hansene0ec90e2015-11-05 18:50:20 -08001488 * allocate a huge page with it. We will only reach this
1489 * when CONFIG_NUMA=y.
Dave Hansen099730d2015-11-05 18:50:17 -08001490 */
1491 do {
1492 struct page *page;
1493 struct mempolicy *mpol;
1494 struct zonelist *zl;
1495 nodemask_t *nodemask;
1496
1497 cpuset_mems_cookie = read_mems_allowed_begin();
1498 zl = huge_zonelist(vma, addr, gfp, &mpol, &nodemask);
1499 mpol_cond_put(mpol);
1500 page = __alloc_pages_nodemask(gfp, order, zl, nodemask);
1501 if (page)
1502 return page;
1503 } while (read_mems_allowed_retry(cpuset_mems_cookie));
1504
1505 return NULL;
1506}
1507
1508/*
1509 * There are two ways to allocate a huge page:
1510 * 1. When you have a VMA and an address (like a fault)
1511 * 2. When you have no VMA (like when setting /proc/.../nr_hugepages)
1512 *
1513 * 'vma' and 'addr' are only for (1). 'nid' is always NUMA_NO_NODE in
1514 * this case which signifies that the allocation should be done with
1515 * respect for the VMA's memory policy.
1516 *
1517 * For (2), we ignore 'vma' and 'addr' and use 'nid' exclusively. This
1518 * implies that memory policies will not be taken in to account.
1519 */
1520static struct page *__alloc_buddy_huge_page(struct hstate *h,
1521 struct vm_area_struct *vma, unsigned long addr, int nid)
Adam Litke7893d1d2007-10-16 01:26:18 -07001522{
1523 struct page *page;
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001524 unsigned int r_nid;
Adam Litke7893d1d2007-10-16 01:26:18 -07001525
Luiz Capitulinobae7f4a2014-06-04 16:07:08 -07001526 if (hstate_is_gigantic(h))
Andi Kleenaa888a72008-07-23 21:27:47 -07001527 return NULL;
1528
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001529 /*
Dave Hansen099730d2015-11-05 18:50:17 -08001530 * Make sure that anyone specifying 'nid' is not also specifying a VMA.
1531 * This makes sure the caller is picking _one_ of the modes with which
1532 * we can call this function, not both.
1533 */
1534 if (vma || (addr != -1)) {
Dave Hansene0ec90e2015-11-05 18:50:20 -08001535 VM_WARN_ON_ONCE(addr == -1);
1536 VM_WARN_ON_ONCE(nid != NUMA_NO_NODE);
Dave Hansen099730d2015-11-05 18:50:17 -08001537 }
1538 /*
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001539 * Assume we will successfully allocate the surplus page to
1540 * prevent racing processes from causing the surplus to exceed
1541 * overcommit
1542 *
1543 * This however introduces a different race, where a process B
1544 * tries to grow the static hugepage pool while alloc_pages() is
1545 * called by process A. B will only examine the per-node
1546 * counters in determining if surplus huge pages can be
1547 * converted to normal huge pages in adjust_pool_surplus(). A
1548 * won't be able to increment the per-node counter, until the
1549 * lock is dropped by B, but B doesn't drop hugetlb_lock until
1550 * no more huge pages can be converted from surplus to normal
1551 * state (and doesn't try to convert again). Thus, we have a
1552 * case where a surplus huge page exists, the pool is grown, and
1553 * the surplus huge page still exists after, even though it
1554 * should just have been converted to a normal huge page. This
1555 * does not leak memory, though, as the hugepage will be freed
1556 * once it is out of use. It also does not allow the counters to
1557 * go out of whack in adjust_pool_surplus() as we don't modify
1558 * the node values until we've gotten the hugepage and only the
1559 * per-node value is checked there.
1560 */
1561 spin_lock(&hugetlb_lock);
Andi Kleena5516432008-07-23 21:27:41 -07001562 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001563 spin_unlock(&hugetlb_lock);
1564 return NULL;
1565 } else {
Andi Kleena5516432008-07-23 21:27:41 -07001566 h->nr_huge_pages++;
1567 h->surplus_huge_pages++;
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001568 }
1569 spin_unlock(&hugetlb_lock);
1570
Dave Hansen099730d2015-11-05 18:50:17 -08001571 page = __hugetlb_alloc_buddy_huge_page(h, vma, addr, nid);
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001572
1573 spin_lock(&hugetlb_lock);
Adam Litke7893d1d2007-10-16 01:26:18 -07001574 if (page) {
Aneesh Kumar K.V0edaecf2012-07-31 16:42:07 -07001575 INIT_LIST_HEAD(&page->lru);
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001576 r_nid = page_to_nid(page);
Kirill A. Shutemovf1e61552015-11-06 16:29:50 -08001577 set_compound_page_dtor(page, HUGETLB_PAGE_DTOR);
Aneesh Kumar K.V9dd540e2012-07-31 16:42:15 -07001578 set_hugetlb_cgroup(page, NULL);
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001579 /*
1580 * We incremented the global counters already
1581 */
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001582 h->nr_huge_pages_node[r_nid]++;
1583 h->surplus_huge_pages_node[r_nid]++;
Adam Litke3b116302008-04-28 02:13:06 -07001584 __count_vm_event(HTLB_BUDDY_PGALLOC);
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001585 } else {
Andi Kleena5516432008-07-23 21:27:41 -07001586 h->nr_huge_pages--;
1587 h->surplus_huge_pages--;
Adam Litke3b116302008-04-28 02:13:06 -07001588 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
Adam Litke7893d1d2007-10-16 01:26:18 -07001589 }
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001590 spin_unlock(&hugetlb_lock);
Adam Litke7893d1d2007-10-16 01:26:18 -07001591
1592 return page;
1593}
1594
Adam Litkee4e574b2007-10-16 01:26:19 -07001595/*
Dave Hansen099730d2015-11-05 18:50:17 -08001596 * Allocate a huge page from 'nid'. Note, 'nid' may be
1597 * NUMA_NO_NODE, which means that it may be allocated
1598 * anywhere.
1599 */
Dave Hansene0ec90e2015-11-05 18:50:20 -08001600static
Dave Hansen099730d2015-11-05 18:50:17 -08001601struct page *__alloc_buddy_huge_page_no_mpol(struct hstate *h, int nid)
1602{
1603 unsigned long addr = -1;
1604
1605 return __alloc_buddy_huge_page(h, NULL, addr, nid);
1606}
1607
1608/*
1609 * Use the VMA's mpolicy to allocate a huge page from the buddy.
1610 */
Dave Hansene0ec90e2015-11-05 18:50:20 -08001611static
Dave Hansen099730d2015-11-05 18:50:17 -08001612struct page *__alloc_buddy_huge_page_with_mpol(struct hstate *h,
1613 struct vm_area_struct *vma, unsigned long addr)
1614{
1615 return __alloc_buddy_huge_page(h, vma, addr, NUMA_NO_NODE);
1616}
1617
1618/*
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001619 * This allocation function is useful in the context where vma is irrelevant.
1620 * E.g. soft-offlining uses this function because it only cares physical
1621 * address of error page.
1622 */
1623struct page *alloc_huge_page_node(struct hstate *h, int nid)
1624{
Joonsoo Kim4ef91842013-09-11 14:21:51 -07001625 struct page *page = NULL;
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001626
1627 spin_lock(&hugetlb_lock);
Joonsoo Kim4ef91842013-09-11 14:21:51 -07001628 if (h->free_huge_pages - h->resv_huge_pages > 0)
1629 page = dequeue_huge_page_node(h, nid);
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001630 spin_unlock(&hugetlb_lock);
1631
Aneesh Kumar K.V94ae8ba2012-07-31 16:42:35 -07001632 if (!page)
Dave Hansen099730d2015-11-05 18:50:17 -08001633 page = __alloc_buddy_huge_page_no_mpol(h, nid);
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001634
1635 return page;
1636}
1637
1638/*
Lucas De Marchi25985ed2011-03-30 22:57:33 -03001639 * Increase the hugetlb pool such that it can accommodate a reservation
Adam Litkee4e574b2007-10-16 01:26:19 -07001640 * of size 'delta'.
1641 */
Andi Kleena5516432008-07-23 21:27:41 -07001642static int gather_surplus_pages(struct hstate *h, int delta)
Adam Litkee4e574b2007-10-16 01:26:19 -07001643{
1644 struct list_head surplus_list;
1645 struct page *page, *tmp;
1646 int ret, i;
1647 int needed, allocated;
Hillf Danton28073b02012-03-21 16:34:00 -07001648 bool alloc_ok = true;
Adam Litkee4e574b2007-10-16 01:26:19 -07001649
Andi Kleena5516432008-07-23 21:27:41 -07001650 needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
Adam Litkeac09b3a2008-03-04 14:29:38 -08001651 if (needed <= 0) {
Andi Kleena5516432008-07-23 21:27:41 -07001652 h->resv_huge_pages += delta;
Adam Litkee4e574b2007-10-16 01:26:19 -07001653 return 0;
Adam Litkeac09b3a2008-03-04 14:29:38 -08001654 }
Adam Litkee4e574b2007-10-16 01:26:19 -07001655
1656 allocated = 0;
1657 INIT_LIST_HEAD(&surplus_list);
1658
1659 ret = -ENOMEM;
1660retry:
1661 spin_unlock(&hugetlb_lock);
1662 for (i = 0; i < needed; i++) {
Dave Hansen099730d2015-11-05 18:50:17 -08001663 page = __alloc_buddy_huge_page_no_mpol(h, NUMA_NO_NODE);
Hillf Danton28073b02012-03-21 16:34:00 -07001664 if (!page) {
1665 alloc_ok = false;
1666 break;
1667 }
Adam Litkee4e574b2007-10-16 01:26:19 -07001668 list_add(&page->lru, &surplus_list);
1669 }
Hillf Danton28073b02012-03-21 16:34:00 -07001670 allocated += i;
Adam Litkee4e574b2007-10-16 01:26:19 -07001671
1672 /*
1673 * After retaking hugetlb_lock, we need to recalculate 'needed'
1674 * because either resv_huge_pages or free_huge_pages may have changed.
1675 */
1676 spin_lock(&hugetlb_lock);
Andi Kleena5516432008-07-23 21:27:41 -07001677 needed = (h->resv_huge_pages + delta) -
1678 (h->free_huge_pages + allocated);
Hillf Danton28073b02012-03-21 16:34:00 -07001679 if (needed > 0) {
1680 if (alloc_ok)
1681 goto retry;
1682 /*
1683 * We were not able to allocate enough pages to
1684 * satisfy the entire reservation so we free what
1685 * we've allocated so far.
1686 */
1687 goto free;
1688 }
Adam Litkee4e574b2007-10-16 01:26:19 -07001689 /*
1690 * The surplus_list now contains _at_least_ the number of extra pages
Lucas De Marchi25985ed2011-03-30 22:57:33 -03001691 * needed to accommodate the reservation. Add the appropriate number
Adam Litkee4e574b2007-10-16 01:26:19 -07001692 * of pages to the hugetlb pool and free the extras back to the buddy
Adam Litkeac09b3a2008-03-04 14:29:38 -08001693 * allocator. Commit the entire reservation here to prevent another
1694 * process from stealing the pages as they are added to the pool but
1695 * before they are reserved.
Adam Litkee4e574b2007-10-16 01:26:19 -07001696 */
1697 needed += allocated;
Andi Kleena5516432008-07-23 21:27:41 -07001698 h->resv_huge_pages += delta;
Adam Litkee4e574b2007-10-16 01:26:19 -07001699 ret = 0;
Naoya Horiguchia9869b82010-09-08 10:19:37 +09001700
Adam Litke19fc3f02008-04-28 02:12:20 -07001701 /* Free the needed pages to the hugetlb pool */
Adam Litkee4e574b2007-10-16 01:26:19 -07001702 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
Adam Litke19fc3f02008-04-28 02:12:20 -07001703 if ((--needed) < 0)
1704 break;
Naoya Horiguchia9869b82010-09-08 10:19:37 +09001705 /*
1706 * This page is now managed by the hugetlb allocator and has
1707 * no users -- drop the buddy allocator's reference.
1708 */
1709 put_page_testzero(page);
Sasha Levin309381fea2014-01-23 15:52:54 -08001710 VM_BUG_ON_PAGE(page_count(page), page);
Andi Kleena5516432008-07-23 21:27:41 -07001711 enqueue_huge_page(h, page);
Adam Litke19fc3f02008-04-28 02:12:20 -07001712 }
Hillf Danton28073b02012-03-21 16:34:00 -07001713free:
Hillf Dantonb0365c82011-12-28 15:57:16 -08001714 spin_unlock(&hugetlb_lock);
Adam Litke19fc3f02008-04-28 02:12:20 -07001715
1716 /* Free unnecessary surplus pages to the buddy allocator */
Joonsoo Kimc0d934b2013-09-11 14:21:02 -07001717 list_for_each_entry_safe(page, tmp, &surplus_list, lru)
1718 put_page(page);
Naoya Horiguchia9869b82010-09-08 10:19:37 +09001719 spin_lock(&hugetlb_lock);
Adam Litkee4e574b2007-10-16 01:26:19 -07001720
1721 return ret;
1722}
1723
1724/*
1725 * When releasing a hugetlb pool reservation, any surplus pages that were
1726 * allocated to satisfy the reservation must be explicitly freed if they were
1727 * never used.
Lee Schermerhorn685f3452009-09-21 17:01:23 -07001728 * Called with hugetlb_lock held.
Adam Litkee4e574b2007-10-16 01:26:19 -07001729 */
Andi Kleena5516432008-07-23 21:27:41 -07001730static void return_unused_surplus_pages(struct hstate *h,
1731 unsigned long unused_resv_pages)
Adam Litkee4e574b2007-10-16 01:26:19 -07001732{
Adam Litkee4e574b2007-10-16 01:26:19 -07001733 unsigned long nr_pages;
1734
Adam Litkeac09b3a2008-03-04 14:29:38 -08001735 /* Uncommit the reservation */
Andi Kleena5516432008-07-23 21:27:41 -07001736 h->resv_huge_pages -= unused_resv_pages;
Adam Litkeac09b3a2008-03-04 14:29:38 -08001737
Andi Kleenaa888a72008-07-23 21:27:47 -07001738 /* Cannot return gigantic pages currently */
Luiz Capitulinobae7f4a2014-06-04 16:07:08 -07001739 if (hstate_is_gigantic(h))
Andi Kleenaa888a72008-07-23 21:27:47 -07001740 return;
1741
Andi Kleena5516432008-07-23 21:27:41 -07001742 nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
Adam Litkee4e574b2007-10-16 01:26:19 -07001743
Lee Schermerhorn685f3452009-09-21 17:01:23 -07001744 /*
1745 * We want to release as many surplus pages as possible, spread
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -08001746 * evenly across all nodes with memory. Iterate across these nodes
1747 * until we can no longer free unreserved surplus pages. This occurs
1748 * when the nodes with surplus pages have no free pages.
1749 * free_pool_huge_page() will balance the the freed pages across the
1750 * on-line nodes with memory and will handle the hstate accounting.
Lee Schermerhorn685f3452009-09-21 17:01:23 -07001751 */
1752 while (nr_pages--) {
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08001753 if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1))
Lee Schermerhorn685f3452009-09-21 17:01:23 -07001754 break;
Mizuma, Masayoshi7848a4b2014-04-18 15:07:18 -07001755 cond_resched_lock(&hugetlb_lock);
Adam Litkee4e574b2007-10-16 01:26:19 -07001756 }
1757}
1758
Mike Kravetz5e911372015-09-08 15:01:28 -07001759
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001760/*
Mike Kravetzfeba16e2015-09-08 15:01:31 -07001761 * vma_needs_reservation, vma_commit_reservation and vma_end_reservation
Mike Kravetz5e911372015-09-08 15:01:28 -07001762 * are used by the huge page allocation routines to manage reservations.
Mike Kravetzcf3ad202015-06-24 16:57:55 -07001763 *
1764 * vma_needs_reservation is called to determine if the huge page at addr
1765 * within the vma has an associated reservation. If a reservation is
1766 * needed, the value 1 is returned. The caller is then responsible for
1767 * managing the global reservation and subpool usage counts. After
1768 * the huge page has been allocated, vma_commit_reservation is called
Mike Kravetzfeba16e2015-09-08 15:01:31 -07001769 * to add the page to the reservation map. If the page allocation fails,
1770 * the reservation must be ended instead of committed. vma_end_reservation
1771 * is called in such cases.
Mike Kravetzcf3ad202015-06-24 16:57:55 -07001772 *
1773 * In the normal case, vma_commit_reservation returns the same value
1774 * as the preceding vma_needs_reservation call. The only time this
1775 * is not the case is if a reserve map was changed between calls. It
1776 * is the responsibility of the caller to notice the difference and
1777 * take appropriate action.
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001778 */
Mike Kravetz5e911372015-09-08 15:01:28 -07001779enum vma_resv_mode {
1780 VMA_NEEDS_RESV,
1781 VMA_COMMIT_RESV,
Mike Kravetzfeba16e2015-09-08 15:01:31 -07001782 VMA_END_RESV,
Mike Kravetz5e911372015-09-08 15:01:28 -07001783};
Mike Kravetzcf3ad202015-06-24 16:57:55 -07001784static long __vma_reservation_common(struct hstate *h,
1785 struct vm_area_struct *vma, unsigned long addr,
Mike Kravetz5e911372015-09-08 15:01:28 -07001786 enum vma_resv_mode mode)
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001787{
Joonsoo Kim4e35f482014-04-03 14:47:30 -07001788 struct resv_map *resv;
1789 pgoff_t idx;
Mike Kravetzcf3ad202015-06-24 16:57:55 -07001790 long ret;
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001791
Joonsoo Kim4e35f482014-04-03 14:47:30 -07001792 resv = vma_resv_map(vma);
1793 if (!resv)
Andy Whitcroft84afd992008-07-23 21:27:32 -07001794 return 1;
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001795
Joonsoo Kim4e35f482014-04-03 14:47:30 -07001796 idx = vma_hugecache_offset(h, vma, addr);
Mike Kravetz5e911372015-09-08 15:01:28 -07001797 switch (mode) {
1798 case VMA_NEEDS_RESV:
Mike Kravetzcf3ad202015-06-24 16:57:55 -07001799 ret = region_chg(resv, idx, idx + 1);
Mike Kravetz5e911372015-09-08 15:01:28 -07001800 break;
1801 case VMA_COMMIT_RESV:
1802 ret = region_add(resv, idx, idx + 1);
1803 break;
Mike Kravetzfeba16e2015-09-08 15:01:31 -07001804 case VMA_END_RESV:
Mike Kravetz5e911372015-09-08 15:01:28 -07001805 region_abort(resv, idx, idx + 1);
1806 ret = 0;
1807 break;
1808 default:
1809 BUG();
1810 }
Andy Whitcroft84afd992008-07-23 21:27:32 -07001811
Joonsoo Kim4e35f482014-04-03 14:47:30 -07001812 if (vma->vm_flags & VM_MAYSHARE)
Mike Kravetzcf3ad202015-06-24 16:57:55 -07001813 return ret;
Joonsoo Kim4e35f482014-04-03 14:47:30 -07001814 else
Mike Kravetzcf3ad202015-06-24 16:57:55 -07001815 return ret < 0 ? ret : 0;
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001816}
Mike Kravetzcf3ad202015-06-24 16:57:55 -07001817
1818static long vma_needs_reservation(struct hstate *h,
Andi Kleena5516432008-07-23 21:27:41 -07001819 struct vm_area_struct *vma, unsigned long addr)
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001820{
Mike Kravetz5e911372015-09-08 15:01:28 -07001821 return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV);
Mike Kravetzcf3ad202015-06-24 16:57:55 -07001822}
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001823
Mike Kravetzcf3ad202015-06-24 16:57:55 -07001824static long vma_commit_reservation(struct hstate *h,
1825 struct vm_area_struct *vma, unsigned long addr)
1826{
Mike Kravetz5e911372015-09-08 15:01:28 -07001827 return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV);
1828}
1829
Mike Kravetzfeba16e2015-09-08 15:01:31 -07001830static void vma_end_reservation(struct hstate *h,
Mike Kravetz5e911372015-09-08 15:01:28 -07001831 struct vm_area_struct *vma, unsigned long addr)
1832{
Mike Kravetzfeba16e2015-09-08 15:01:31 -07001833 (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV);
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001834}
1835
Mike Kravetz70c35472015-09-08 15:01:54 -07001836struct page *alloc_huge_page(struct vm_area_struct *vma,
Mel Gorman04f2cbe2008-07-23 21:27:25 -07001837 unsigned long addr, int avoid_reserve)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001838{
David Gibson90481622012-03-21 16:34:12 -07001839 struct hugepage_subpool *spool = subpool_vma(vma);
Andi Kleena5516432008-07-23 21:27:41 -07001840 struct hstate *h = hstate_vma(vma);
Adam Litke348ea202007-11-14 16:59:37 -08001841 struct page *page;
Mike Kravetzd85f69b2015-09-08 15:01:47 -07001842 long map_chg, map_commit;
1843 long gbl_chg;
Aneesh Kumar K.V6d76dcf2012-07-31 16:42:18 -07001844 int ret, idx;
1845 struct hugetlb_cgroup *h_cg;
Adam Litke2fc39ce2007-11-14 16:59:39 -08001846
Aneesh Kumar K.V6d76dcf2012-07-31 16:42:18 -07001847 idx = hstate_index(h);
Mel Gormana1e78772008-07-23 21:27:23 -07001848 /*
Mike Kravetzd85f69b2015-09-08 15:01:47 -07001849 * Examine the region/reserve map to determine if the process
1850 * has a reservation for the page to be allocated. A return
1851 * code of zero indicates a reservation exists (no change).
Mel Gormana1e78772008-07-23 21:27:23 -07001852 */
Mike Kravetzd85f69b2015-09-08 15:01:47 -07001853 map_chg = gbl_chg = vma_needs_reservation(h, vma, addr);
1854 if (map_chg < 0)
Aneesh Kumar K.V76dcee72012-07-31 16:41:57 -07001855 return ERR_PTR(-ENOMEM);
Mike Kravetzd85f69b2015-09-08 15:01:47 -07001856
1857 /*
1858 * Processes that did not create the mapping will have no
1859 * reserves as indicated by the region/reserve map. Check
1860 * that the allocation will not exceed the subpool limit.
1861 * Allocations for MAP_NORESERVE mappings also need to be
1862 * checked against any subpool limit.
1863 */
1864 if (map_chg || avoid_reserve) {
1865 gbl_chg = hugepage_subpool_get_pages(spool, 1);
1866 if (gbl_chg < 0) {
Mike Kravetzfeba16e2015-09-08 15:01:31 -07001867 vma_end_reservation(h, vma, addr);
Aneesh Kumar K.V76dcee72012-07-31 16:41:57 -07001868 return ERR_PTR(-ENOSPC);
Mike Kravetz5e911372015-09-08 15:01:28 -07001869 }
Mel Gormana1e78772008-07-23 21:27:23 -07001870
Mike Kravetzd85f69b2015-09-08 15:01:47 -07001871 /*
1872 * Even though there was no reservation in the region/reserve
1873 * map, there could be reservations associated with the
1874 * subpool that can be used. This would be indicated if the
1875 * return value of hugepage_subpool_get_pages() is zero.
1876 * However, if avoid_reserve is specified we still avoid even
1877 * the subpool reservations.
1878 */
1879 if (avoid_reserve)
1880 gbl_chg = 1;
1881 }
1882
Aneesh Kumar K.V6d76dcf2012-07-31 16:42:18 -07001883 ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg);
Jianyu Zhan8f34af62014-06-04 16:10:36 -07001884 if (ret)
1885 goto out_subpool_put;
1886
Mel Gormana1e78772008-07-23 21:27:23 -07001887 spin_lock(&hugetlb_lock);
Mike Kravetzd85f69b2015-09-08 15:01:47 -07001888 /*
1889 * glb_chg is passed to indicate whether or not a page must be taken
1890 * from the global free pool (global change). gbl_chg == 0 indicates
1891 * a reservation exists for the allocation.
1892 */
1893 page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg);
Joonsoo Kim81a6fca2013-09-11 14:20:58 -07001894 if (!page) {
Aneesh Kumar K.V94ae8ba2012-07-31 16:42:35 -07001895 spin_unlock(&hugetlb_lock);
Dave Hansen099730d2015-11-05 18:50:17 -08001896 page = __alloc_buddy_huge_page_with_mpol(h, vma, addr);
Jianyu Zhan8f34af62014-06-04 16:10:36 -07001897 if (!page)
1898 goto out_uncharge_cgroup;
Naoya Horiguchia88c7692015-12-11 13:40:24 -08001899 if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) {
1900 SetPagePrivate(page);
1901 h->resv_huge_pages--;
1902 }
Aneesh Kumar K.V79dbb232012-07-31 16:42:32 -07001903 spin_lock(&hugetlb_lock);
1904 list_move(&page->lru, &h->hugepage_activelist);
Joonsoo Kim81a6fca2013-09-11 14:20:58 -07001905 /* Fall through */
Mel Gormana1e78772008-07-23 21:27:23 -07001906 }
Joonsoo Kim81a6fca2013-09-11 14:20:58 -07001907 hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page);
1908 spin_unlock(&hugetlb_lock);
Mel Gormana1e78772008-07-23 21:27:23 -07001909
David Gibson90481622012-03-21 16:34:12 -07001910 set_page_private(page, (unsigned long)spool);
Mel Gormana1e78772008-07-23 21:27:23 -07001911
Mike Kravetzd85f69b2015-09-08 15:01:47 -07001912 map_commit = vma_commit_reservation(h, vma, addr);
1913 if (unlikely(map_chg > map_commit)) {
Mike Kravetz33039672015-06-24 16:57:58 -07001914 /*
1915 * The page was added to the reservation map between
1916 * vma_needs_reservation and vma_commit_reservation.
1917 * This indicates a race with hugetlb_reserve_pages.
1918 * Adjust for the subpool count incremented above AND
1919 * in hugetlb_reserve_pages for the same page. Also,
1920 * the reservation count added in hugetlb_reserve_pages
1921 * no longer applies.
1922 */
1923 long rsv_adjust;
1924
1925 rsv_adjust = hugepage_subpool_put_pages(spool, 1);
1926 hugetlb_acct_memory(h, -rsv_adjust);
1927 }
Adam Litke90d8b7e2007-11-14 16:59:42 -08001928 return page;
Jianyu Zhan8f34af62014-06-04 16:10:36 -07001929
1930out_uncharge_cgroup:
1931 hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg);
1932out_subpool_put:
Mike Kravetzd85f69b2015-09-08 15:01:47 -07001933 if (map_chg || avoid_reserve)
Jianyu Zhan8f34af62014-06-04 16:10:36 -07001934 hugepage_subpool_put_pages(spool, 1);
Mike Kravetzfeba16e2015-09-08 15:01:31 -07001935 vma_end_reservation(h, vma, addr);
Jianyu Zhan8f34af62014-06-04 16:10:36 -07001936 return ERR_PTR(-ENOSPC);
David Gibsonb45b5bd2006-03-22 00:08:55 -08001937}
1938
Naoya Horiguchi74060e42013-09-11 14:22:06 -07001939/*
1940 * alloc_huge_page()'s wrapper which simply returns the page if allocation
1941 * succeeds, otherwise NULL. This function is called from new_vma_page(),
1942 * where no ERR_VALUE is expected to be returned.
1943 */
1944struct page *alloc_huge_page_noerr(struct vm_area_struct *vma,
1945 unsigned long addr, int avoid_reserve)
1946{
1947 struct page *page = alloc_huge_page(vma, addr, avoid_reserve);
1948 if (IS_ERR(page))
1949 page = NULL;
1950 return page;
1951}
1952
Cyrill Gorcunov91f47662009-01-06 14:40:33 -08001953int __weak alloc_bootmem_huge_page(struct hstate *h)
Andi Kleenaa888a72008-07-23 21:27:47 -07001954{
1955 struct huge_bootmem_page *m;
Joonsoo Kimb2261022013-09-11 14:21:00 -07001956 int nr_nodes, node;
Andi Kleenaa888a72008-07-23 21:27:47 -07001957
Joonsoo Kimb2261022013-09-11 14:21:00 -07001958 for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) {
Andi Kleenaa888a72008-07-23 21:27:47 -07001959 void *addr;
1960
Grygorii Strashko8b89a112014-01-21 15:50:36 -08001961 addr = memblock_virt_alloc_try_nid_nopanic(
1962 huge_page_size(h), huge_page_size(h),
1963 0, BOOTMEM_ALLOC_ACCESSIBLE, node);
Andi Kleenaa888a72008-07-23 21:27:47 -07001964 if (addr) {
1965 /*
1966 * Use the beginning of the huge page to store the
1967 * huge_bootmem_page struct (until gather_bootmem
1968 * puts them into the mem_map).
1969 */
1970 m = addr;
Cyrill Gorcunov91f47662009-01-06 14:40:33 -08001971 goto found;
Andi Kleenaa888a72008-07-23 21:27:47 -07001972 }
Andi Kleenaa888a72008-07-23 21:27:47 -07001973 }
1974 return 0;
1975
1976found:
Luiz Capitulinodf994ea2014-12-12 16:55:21 -08001977 BUG_ON(!IS_ALIGNED(virt_to_phys(m), huge_page_size(h)));
Andi Kleenaa888a72008-07-23 21:27:47 -07001978 /* Put them into a private list first because mem_map is not up yet */
1979 list_add(&m->list, &huge_boot_pages);
1980 m->hstate = h;
1981 return 1;
1982}
1983
Kirill A. Shutemovd00181b2015-11-06 16:29:57 -08001984static void __init prep_compound_huge_page(struct page *page,
1985 unsigned int order)
Andy Whitcroft18229df2008-11-06 12:53:27 -08001986{
1987 if (unlikely(order > (MAX_ORDER - 1)))
1988 prep_compound_gigantic_page(page, order);
1989 else
1990 prep_compound_page(page, order);
1991}
1992
Andi Kleenaa888a72008-07-23 21:27:47 -07001993/* Put bootmem huge pages into the standard lists after mem_map is up */
1994static void __init gather_bootmem_prealloc(void)
1995{
1996 struct huge_bootmem_page *m;
1997
1998 list_for_each_entry(m, &huge_boot_pages, list) {
Andi Kleenaa888a72008-07-23 21:27:47 -07001999 struct hstate *h = m->hstate;
Becky Bruceee8f2482011-07-25 17:11:50 -07002000 struct page *page;
2001
2002#ifdef CONFIG_HIGHMEM
2003 page = pfn_to_page(m->phys >> PAGE_SHIFT);
Grygorii Strashko8b89a112014-01-21 15:50:36 -08002004 memblock_free_late(__pa(m),
2005 sizeof(struct huge_bootmem_page));
Becky Bruceee8f2482011-07-25 17:11:50 -07002006#else
2007 page = virt_to_page(m);
2008#endif
Andi Kleenaa888a72008-07-23 21:27:47 -07002009 WARN_ON(page_count(page) != 1);
Andy Whitcroft18229df2008-11-06 12:53:27 -08002010 prep_compound_huge_page(page, h->order);
Andrea Arcangelief5a22b2013-10-16 13:46:56 -07002011 WARN_ON(PageReserved(page));
Andi Kleenaa888a72008-07-23 21:27:47 -07002012 prep_new_huge_page(h, page, page_to_nid(page));
Rafael Aquinib0320c72011-06-15 15:08:39 -07002013 /*
2014 * If we had gigantic hugepages allocated at boot time, we need
2015 * to restore the 'stolen' pages to totalram_pages in order to
2016 * fix confusing memory reports from free(1) and another
2017 * side-effects, like CommitLimit going negative.
2018 */
Luiz Capitulinobae7f4a2014-06-04 16:07:08 -07002019 if (hstate_is_gigantic(h))
Jiang Liu3dcc0572013-07-03 15:03:21 -07002020 adjust_managed_page_count(page, 1 << h->order);
Andi Kleenaa888a72008-07-23 21:27:47 -07002021 }
2022}
2023
Andi Kleen8faa8b02008-07-23 21:27:48 -07002024static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002025{
2026 unsigned long i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002027
Andi Kleene5ff2152008-07-23 21:27:42 -07002028 for (i = 0; i < h->max_huge_pages; ++i) {
Luiz Capitulinobae7f4a2014-06-04 16:07:08 -07002029 if (hstate_is_gigantic(h)) {
Andi Kleenaa888a72008-07-23 21:27:47 -07002030 if (!alloc_bootmem_huge_page(h))
2031 break;
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -08002032 } else if (!alloc_fresh_huge_page(h,
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08002033 &node_states[N_MEMORY]))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002034 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002035 }
Andi Kleen8faa8b02008-07-23 21:27:48 -07002036 h->max_huge_pages = i;
Andi Kleene5ff2152008-07-23 21:27:42 -07002037}
2038
2039static void __init hugetlb_init_hstates(void)
2040{
2041 struct hstate *h;
2042
2043 for_each_hstate(h) {
Naoya Horiguchi641844f2015-06-24 16:56:59 -07002044 if (minimum_order > huge_page_order(h))
2045 minimum_order = huge_page_order(h);
2046
Andi Kleen8faa8b02008-07-23 21:27:48 -07002047 /* oversize hugepages were init'ed in early boot */
Luiz Capitulinobae7f4a2014-06-04 16:07:08 -07002048 if (!hstate_is_gigantic(h))
Andi Kleen8faa8b02008-07-23 21:27:48 -07002049 hugetlb_hstate_alloc_pages(h);
Andi Kleene5ff2152008-07-23 21:27:42 -07002050 }
Naoya Horiguchi641844f2015-06-24 16:56:59 -07002051 VM_BUG_ON(minimum_order == UINT_MAX);
Andi Kleene5ff2152008-07-23 21:27:42 -07002052}
2053
Andi Kleen4abd32d2008-07-23 21:27:49 -07002054static char * __init memfmt(char *buf, unsigned long n)
2055{
2056 if (n >= (1UL << 30))
2057 sprintf(buf, "%lu GB", n >> 30);
2058 else if (n >= (1UL << 20))
2059 sprintf(buf, "%lu MB", n >> 20);
2060 else
2061 sprintf(buf, "%lu KB", n >> 10);
2062 return buf;
2063}
2064
Andi Kleene5ff2152008-07-23 21:27:42 -07002065static void __init report_hugepages(void)
2066{
2067 struct hstate *h;
2068
2069 for_each_hstate(h) {
Andi Kleen4abd32d2008-07-23 21:27:49 -07002070 char buf[32];
Andrew Mortonffb22af2013-02-22 16:32:08 -08002071 pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n",
Andi Kleen4abd32d2008-07-23 21:27:49 -07002072 memfmt(buf, huge_page_size(h)),
2073 h->free_huge_pages);
Andi Kleene5ff2152008-07-23 21:27:42 -07002074 }
2075}
2076
Linus Torvalds1da177e2005-04-16 15:20:36 -07002077#ifdef CONFIG_HIGHMEM
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08002078static void try_to_free_low(struct hstate *h, unsigned long count,
2079 nodemask_t *nodes_allowed)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002080{
Christoph Lameter4415cc82006-09-25 23:31:55 -07002081 int i;
2082
Luiz Capitulinobae7f4a2014-06-04 16:07:08 -07002083 if (hstate_is_gigantic(h))
Andi Kleenaa888a72008-07-23 21:27:47 -07002084 return;
2085
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08002086 for_each_node_mask(i, *nodes_allowed) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002087 struct page *page, *next;
Andi Kleena5516432008-07-23 21:27:41 -07002088 struct list_head *freel = &h->hugepage_freelists[i];
2089 list_for_each_entry_safe(page, next, freel, lru) {
2090 if (count >= h->nr_huge_pages)
Adam Litke6b0c8802007-10-16 01:26:23 -07002091 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002092 if (PageHighMem(page))
2093 continue;
2094 list_del(&page->lru);
Andi Kleene5ff2152008-07-23 21:27:42 -07002095 update_and_free_page(h, page);
Andi Kleena5516432008-07-23 21:27:41 -07002096 h->free_huge_pages--;
2097 h->free_huge_pages_node[page_to_nid(page)]--;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002098 }
2099 }
2100}
2101#else
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08002102static inline void try_to_free_low(struct hstate *h, unsigned long count,
2103 nodemask_t *nodes_allowed)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002104{
2105}
2106#endif
2107
Wu Fengguang20a03072009-06-16 15:32:22 -07002108/*
2109 * Increment or decrement surplus_huge_pages. Keep node-specific counters
2110 * balanced by operating on them in a round-robin fashion.
2111 * Returns 1 if an adjustment was made.
2112 */
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08002113static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
2114 int delta)
Wu Fengguang20a03072009-06-16 15:32:22 -07002115{
Joonsoo Kimb2261022013-09-11 14:21:00 -07002116 int nr_nodes, node;
Wu Fengguang20a03072009-06-16 15:32:22 -07002117
2118 VM_BUG_ON(delta != -1 && delta != 1);
Wu Fengguang20a03072009-06-16 15:32:22 -07002119
Joonsoo Kimb2261022013-09-11 14:21:00 -07002120 if (delta < 0) {
2121 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
2122 if (h->surplus_huge_pages_node[node])
2123 goto found;
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07002124 }
Joonsoo Kimb2261022013-09-11 14:21:00 -07002125 } else {
2126 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
2127 if (h->surplus_huge_pages_node[node] <
2128 h->nr_huge_pages_node[node])
2129 goto found;
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07002130 }
Joonsoo Kimb2261022013-09-11 14:21:00 -07002131 }
2132 return 0;
Wu Fengguang20a03072009-06-16 15:32:22 -07002133
Joonsoo Kimb2261022013-09-11 14:21:00 -07002134found:
2135 h->surplus_huge_pages += delta;
2136 h->surplus_huge_pages_node[node] += delta;
2137 return 1;
Wu Fengguang20a03072009-06-16 15:32:22 -07002138}
2139
Andi Kleena5516432008-07-23 21:27:41 -07002140#define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08002141static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count,
2142 nodemask_t *nodes_allowed)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002143{
Adam Litke7893d1d2007-10-16 01:26:18 -07002144 unsigned long min_count, ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002145
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07002146 if (hstate_is_gigantic(h) && !gigantic_page_supported())
Andi Kleenaa888a72008-07-23 21:27:47 -07002147 return h->max_huge_pages;
2148
Adam Litke7893d1d2007-10-16 01:26:18 -07002149 /*
2150 * Increase the pool size
2151 * First take pages out of surplus state. Then make up the
2152 * remaining difference by allocating fresh huge pages.
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08002153 *
Naoya Horiguchid15c7c02015-11-10 14:45:11 -08002154 * We might race with __alloc_buddy_huge_page() here and be unable
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08002155 * to convert a surplus huge page to a normal huge page. That is
2156 * not critical, though, it just means the overall size of the
2157 * pool might be one hugepage larger than it needs to be, but
2158 * within all the constraints specified by the sysctls.
Adam Litke7893d1d2007-10-16 01:26:18 -07002159 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002160 spin_lock(&hugetlb_lock);
Andi Kleena5516432008-07-23 21:27:41 -07002161 while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08002162 if (!adjust_pool_surplus(h, nodes_allowed, -1))
Adam Litke7893d1d2007-10-16 01:26:18 -07002163 break;
2164 }
2165
Andi Kleena5516432008-07-23 21:27:41 -07002166 while (count > persistent_huge_pages(h)) {
Adam Litke7893d1d2007-10-16 01:26:18 -07002167 /*
2168 * If this allocation races such that we no longer need the
2169 * page, free_huge_page will handle it by freeing the page
2170 * and reducing the surplus.
2171 */
2172 spin_unlock(&hugetlb_lock);
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07002173 if (hstate_is_gigantic(h))
2174 ret = alloc_fresh_gigantic_page(h, nodes_allowed);
2175 else
2176 ret = alloc_fresh_huge_page(h, nodes_allowed);
Adam Litke7893d1d2007-10-16 01:26:18 -07002177 spin_lock(&hugetlb_lock);
2178 if (!ret)
2179 goto out;
2180
Mel Gorman536240f22009-12-14 17:59:56 -08002181 /* Bail for signals. Probably ctrl-c from user */
2182 if (signal_pending(current))
2183 goto out;
Adam Litke7893d1d2007-10-16 01:26:18 -07002184 }
Adam Litke7893d1d2007-10-16 01:26:18 -07002185
2186 /*
2187 * Decrease the pool size
2188 * First return free pages to the buddy allocator (being careful
2189 * to keep enough around to satisfy reservations). Then place
2190 * pages into surplus state as needed so the pool will shrink
2191 * to the desired size as pages become free.
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08002192 *
2193 * By placing pages into the surplus state independent of the
2194 * overcommit value, we are allowing the surplus pool size to
2195 * exceed overcommit. There are few sane options here. Since
Naoya Horiguchid15c7c02015-11-10 14:45:11 -08002196 * __alloc_buddy_huge_page() is checking the global counter,
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08002197 * though, we'll note that we're not allowed to exceed surplus
2198 * and won't grow the pool anywhere else. Not until one of the
2199 * sysctls are changed, or the surplus pages go out of use.
Adam Litke7893d1d2007-10-16 01:26:18 -07002200 */
Andi Kleena5516432008-07-23 21:27:41 -07002201 min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
Adam Litke6b0c8802007-10-16 01:26:23 -07002202 min_count = max(count, min_count);
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08002203 try_to_free_low(h, min_count, nodes_allowed);
Andi Kleena5516432008-07-23 21:27:41 -07002204 while (min_count < persistent_huge_pages(h)) {
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08002205 if (!free_pool_huge_page(h, nodes_allowed, 0))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002206 break;
Mizuma, Masayoshi55f67142014-04-07 15:37:54 -07002207 cond_resched_lock(&hugetlb_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002208 }
Andi Kleena5516432008-07-23 21:27:41 -07002209 while (count < persistent_huge_pages(h)) {
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08002210 if (!adjust_pool_surplus(h, nodes_allowed, 1))
Adam Litke7893d1d2007-10-16 01:26:18 -07002211 break;
2212 }
2213out:
Andi Kleena5516432008-07-23 21:27:41 -07002214 ret = persistent_huge_pages(h);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002215 spin_unlock(&hugetlb_lock);
Adam Litke7893d1d2007-10-16 01:26:18 -07002216 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002217}
2218
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002219#define HSTATE_ATTR_RO(_name) \
2220 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
2221
2222#define HSTATE_ATTR(_name) \
2223 static struct kobj_attribute _name##_attr = \
2224 __ATTR(_name, 0644, _name##_show, _name##_store)
2225
2226static struct kobject *hugepages_kobj;
2227static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
2228
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002229static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
2230
2231static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002232{
2233 int i;
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002234
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002235 for (i = 0; i < HUGE_MAX_HSTATE; i++)
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002236 if (hstate_kobjs[i] == kobj) {
2237 if (nidp)
2238 *nidp = NUMA_NO_NODE;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002239 return &hstates[i];
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002240 }
2241
2242 return kobj_to_node_hstate(kobj, nidp);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002243}
2244
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002245static ssize_t nr_hugepages_show_common(struct kobject *kobj,
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002246 struct kobj_attribute *attr, char *buf)
2247{
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002248 struct hstate *h;
2249 unsigned long nr_huge_pages;
2250 int nid;
2251
2252 h = kobj_to_hstate(kobj, &nid);
2253 if (nid == NUMA_NO_NODE)
2254 nr_huge_pages = h->nr_huge_pages;
2255 else
2256 nr_huge_pages = h->nr_huge_pages_node[nid];
2257
2258 return sprintf(buf, "%lu\n", nr_huge_pages);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002259}
Eric B Munsonadbe8722011-01-13 15:47:27 -08002260
David Rientjes238d3c12014-08-06 16:06:51 -07002261static ssize_t __nr_hugepages_store_common(bool obey_mempolicy,
2262 struct hstate *h, int nid,
2263 unsigned long count, size_t len)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002264{
2265 int err;
David Rientjesbad44b52009-12-14 17:58:38 -08002266 NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002267
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07002268 if (hstate_is_gigantic(h) && !gigantic_page_supported()) {
Eric B Munsonadbe8722011-01-13 15:47:27 -08002269 err = -EINVAL;
2270 goto out;
2271 }
2272
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002273 if (nid == NUMA_NO_NODE) {
2274 /*
2275 * global hstate attribute
2276 */
2277 if (!(obey_mempolicy &&
2278 init_nodemask_of_mempolicy(nodes_allowed))) {
2279 NODEMASK_FREE(nodes_allowed);
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08002280 nodes_allowed = &node_states[N_MEMORY];
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002281 }
2282 } else if (nodes_allowed) {
2283 /*
2284 * per node hstate attribute: adjust count to global,
2285 * but restrict alloc/free to the specified node.
2286 */
2287 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
2288 init_nodemask_of_node(nodes_allowed, nid);
2289 } else
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08002290 nodes_allowed = &node_states[N_MEMORY];
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002291
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002292 h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002293
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08002294 if (nodes_allowed != &node_states[N_MEMORY])
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002295 NODEMASK_FREE(nodes_allowed);
2296
2297 return len;
Eric B Munsonadbe8722011-01-13 15:47:27 -08002298out:
2299 NODEMASK_FREE(nodes_allowed);
2300 return err;
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002301}
2302
David Rientjes238d3c12014-08-06 16:06:51 -07002303static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
2304 struct kobject *kobj, const char *buf,
2305 size_t len)
2306{
2307 struct hstate *h;
2308 unsigned long count;
2309 int nid;
2310 int err;
2311
2312 err = kstrtoul(buf, 10, &count);
2313 if (err)
2314 return err;
2315
2316 h = kobj_to_hstate(kobj, &nid);
2317 return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len);
2318}
2319
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002320static ssize_t nr_hugepages_show(struct kobject *kobj,
2321 struct kobj_attribute *attr, char *buf)
2322{
2323 return nr_hugepages_show_common(kobj, attr, buf);
2324}
2325
2326static ssize_t nr_hugepages_store(struct kobject *kobj,
2327 struct kobj_attribute *attr, const char *buf, size_t len)
2328{
David Rientjes238d3c12014-08-06 16:06:51 -07002329 return nr_hugepages_store_common(false, kobj, buf, len);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002330}
2331HSTATE_ATTR(nr_hugepages);
2332
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002333#ifdef CONFIG_NUMA
2334
2335/*
2336 * hstate attribute for optionally mempolicy-based constraint on persistent
2337 * huge page alloc/free.
2338 */
2339static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
2340 struct kobj_attribute *attr, char *buf)
2341{
2342 return nr_hugepages_show_common(kobj, attr, buf);
2343}
2344
2345static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
2346 struct kobj_attribute *attr, const char *buf, size_t len)
2347{
David Rientjes238d3c12014-08-06 16:06:51 -07002348 return nr_hugepages_store_common(true, kobj, buf, len);
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002349}
2350HSTATE_ATTR(nr_hugepages_mempolicy);
2351#endif
2352
2353
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002354static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
2355 struct kobj_attribute *attr, char *buf)
2356{
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002357 struct hstate *h = kobj_to_hstate(kobj, NULL);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002358 return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
2359}
Eric B Munsonadbe8722011-01-13 15:47:27 -08002360
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002361static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
2362 struct kobj_attribute *attr, const char *buf, size_t count)
2363{
2364 int err;
2365 unsigned long input;
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002366 struct hstate *h = kobj_to_hstate(kobj, NULL);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002367
Luiz Capitulinobae7f4a2014-06-04 16:07:08 -07002368 if (hstate_is_gigantic(h))
Eric B Munsonadbe8722011-01-13 15:47:27 -08002369 return -EINVAL;
2370
Jingoo Han3dbb95f2013-09-11 14:20:25 -07002371 err = kstrtoul(buf, 10, &input);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002372 if (err)
Eric B Munson73ae31e2011-01-13 15:47:28 -08002373 return err;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002374
2375 spin_lock(&hugetlb_lock);
2376 h->nr_overcommit_huge_pages = input;
2377 spin_unlock(&hugetlb_lock);
2378
2379 return count;
2380}
2381HSTATE_ATTR(nr_overcommit_hugepages);
2382
2383static ssize_t free_hugepages_show(struct kobject *kobj,
2384 struct kobj_attribute *attr, char *buf)
2385{
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002386 struct hstate *h;
2387 unsigned long free_huge_pages;
2388 int nid;
2389
2390 h = kobj_to_hstate(kobj, &nid);
2391 if (nid == NUMA_NO_NODE)
2392 free_huge_pages = h->free_huge_pages;
2393 else
2394 free_huge_pages = h->free_huge_pages_node[nid];
2395
2396 return sprintf(buf, "%lu\n", free_huge_pages);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002397}
2398HSTATE_ATTR_RO(free_hugepages);
2399
2400static ssize_t resv_hugepages_show(struct kobject *kobj,
2401 struct kobj_attribute *attr, char *buf)
2402{
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002403 struct hstate *h = kobj_to_hstate(kobj, NULL);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002404 return sprintf(buf, "%lu\n", h->resv_huge_pages);
2405}
2406HSTATE_ATTR_RO(resv_hugepages);
2407
2408static ssize_t surplus_hugepages_show(struct kobject *kobj,
2409 struct kobj_attribute *attr, char *buf)
2410{
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002411 struct hstate *h;
2412 unsigned long surplus_huge_pages;
2413 int nid;
2414
2415 h = kobj_to_hstate(kobj, &nid);
2416 if (nid == NUMA_NO_NODE)
2417 surplus_huge_pages = h->surplus_huge_pages;
2418 else
2419 surplus_huge_pages = h->surplus_huge_pages_node[nid];
2420
2421 return sprintf(buf, "%lu\n", surplus_huge_pages);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002422}
2423HSTATE_ATTR_RO(surplus_hugepages);
2424
2425static struct attribute *hstate_attrs[] = {
2426 &nr_hugepages_attr.attr,
2427 &nr_overcommit_hugepages_attr.attr,
2428 &free_hugepages_attr.attr,
2429 &resv_hugepages_attr.attr,
2430 &surplus_hugepages_attr.attr,
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002431#ifdef CONFIG_NUMA
2432 &nr_hugepages_mempolicy_attr.attr,
2433#endif
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002434 NULL,
2435};
2436
2437static struct attribute_group hstate_attr_group = {
2438 .attrs = hstate_attrs,
2439};
2440
Jeff Mahoney094e9532010-02-02 13:44:14 -08002441static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
2442 struct kobject **hstate_kobjs,
2443 struct attribute_group *hstate_attr_group)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002444{
2445 int retval;
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07002446 int hi = hstate_index(h);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002447
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002448 hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
2449 if (!hstate_kobjs[hi])
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002450 return -ENOMEM;
2451
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002452 retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002453 if (retval)
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002454 kobject_put(hstate_kobjs[hi]);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002455
2456 return retval;
2457}
2458
2459static void __init hugetlb_sysfs_init(void)
2460{
2461 struct hstate *h;
2462 int err;
2463
2464 hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
2465 if (!hugepages_kobj)
2466 return;
2467
2468 for_each_hstate(h) {
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002469 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
2470 hstate_kobjs, &hstate_attr_group);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002471 if (err)
Andrew Mortonffb22af2013-02-22 16:32:08 -08002472 pr_err("Hugetlb: Unable to add hstate %s", h->name);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002473 }
2474}
2475
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002476#ifdef CONFIG_NUMA
2477
2478/*
2479 * node_hstate/s - associate per node hstate attributes, via their kobjects,
Kay Sievers10fbcf42011-12-21 14:48:43 -08002480 * with node devices in node_devices[] using a parallel array. The array
2481 * index of a node device or _hstate == node id.
2482 * This is here to avoid any static dependency of the node device driver, in
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002483 * the base kernel, on the hugetlb module.
2484 */
2485struct node_hstate {
2486 struct kobject *hugepages_kobj;
2487 struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
2488};
Alexander Kuleshovb4e289a2015-11-05 18:50:14 -08002489static struct node_hstate node_hstates[MAX_NUMNODES];
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002490
2491/*
Kay Sievers10fbcf42011-12-21 14:48:43 -08002492 * A subset of global hstate attributes for node devices
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002493 */
2494static struct attribute *per_node_hstate_attrs[] = {
2495 &nr_hugepages_attr.attr,
2496 &free_hugepages_attr.attr,
2497 &surplus_hugepages_attr.attr,
2498 NULL,
2499};
2500
2501static struct attribute_group per_node_hstate_attr_group = {
2502 .attrs = per_node_hstate_attrs,
2503};
2504
2505/*
Kay Sievers10fbcf42011-12-21 14:48:43 -08002506 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002507 * Returns node id via non-NULL nidp.
2508 */
2509static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
2510{
2511 int nid;
2512
2513 for (nid = 0; nid < nr_node_ids; nid++) {
2514 struct node_hstate *nhs = &node_hstates[nid];
2515 int i;
2516 for (i = 0; i < HUGE_MAX_HSTATE; i++)
2517 if (nhs->hstate_kobjs[i] == kobj) {
2518 if (nidp)
2519 *nidp = nid;
2520 return &hstates[i];
2521 }
2522 }
2523
2524 BUG();
2525 return NULL;
2526}
2527
2528/*
Kay Sievers10fbcf42011-12-21 14:48:43 -08002529 * Unregister hstate attributes from a single node device.
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002530 * No-op if no hstate attributes attached.
2531 */
Claudiu Ghioc3cd8b442013-03-04 12:46:15 +02002532static void hugetlb_unregister_node(struct node *node)
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002533{
2534 struct hstate *h;
Kay Sievers10fbcf42011-12-21 14:48:43 -08002535 struct node_hstate *nhs = &node_hstates[node->dev.id];
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002536
2537 if (!nhs->hugepages_kobj)
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -08002538 return; /* no hstate attributes */
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002539
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07002540 for_each_hstate(h) {
2541 int idx = hstate_index(h);
2542 if (nhs->hstate_kobjs[idx]) {
2543 kobject_put(nhs->hstate_kobjs[idx]);
2544 nhs->hstate_kobjs[idx] = NULL;
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002545 }
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07002546 }
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002547
2548 kobject_put(nhs->hugepages_kobj);
2549 nhs->hugepages_kobj = NULL;
2550}
2551
2552/*
Kay Sievers10fbcf42011-12-21 14:48:43 -08002553 * hugetlb module exit: unregister hstate attributes from node devices
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002554 * that have them.
2555 */
2556static void hugetlb_unregister_all_nodes(void)
2557{
2558 int nid;
2559
2560 /*
Kay Sievers10fbcf42011-12-21 14:48:43 -08002561 * disable node device registrations.
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002562 */
2563 register_hugetlbfs_with_node(NULL, NULL);
2564
2565 /*
2566 * remove hstate attributes from any nodes that have them.
2567 */
2568 for (nid = 0; nid < nr_node_ids; nid++)
Wen Congyang87327942012-12-11 16:00:56 -08002569 hugetlb_unregister_node(node_devices[nid]);
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002570}
2571
2572/*
Kay Sievers10fbcf42011-12-21 14:48:43 -08002573 * Register hstate attributes for a single node device.
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002574 * No-op if attributes already registered.
2575 */
Claudiu Ghioc3cd8b442013-03-04 12:46:15 +02002576static void hugetlb_register_node(struct node *node)
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002577{
2578 struct hstate *h;
Kay Sievers10fbcf42011-12-21 14:48:43 -08002579 struct node_hstate *nhs = &node_hstates[node->dev.id];
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002580 int err;
2581
2582 if (nhs->hugepages_kobj)
2583 return; /* already allocated */
2584
2585 nhs->hugepages_kobj = kobject_create_and_add("hugepages",
Kay Sievers10fbcf42011-12-21 14:48:43 -08002586 &node->dev.kobj);
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002587 if (!nhs->hugepages_kobj)
2588 return;
2589
2590 for_each_hstate(h) {
2591 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
2592 nhs->hstate_kobjs,
2593 &per_node_hstate_attr_group);
2594 if (err) {
Andrew Mortonffb22af2013-02-22 16:32:08 -08002595 pr_err("Hugetlb: Unable to add hstate %s for node %d\n",
2596 h->name, node->dev.id);
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002597 hugetlb_unregister_node(node);
2598 break;
2599 }
2600 }
2601}
2602
2603/*
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -08002604 * hugetlb init time: register hstate attributes for all registered node
Kay Sievers10fbcf42011-12-21 14:48:43 -08002605 * devices of nodes that have memory. All on-line nodes should have
2606 * registered their associated device by this time.
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002607 */
Luiz Capitulino7d9ca002014-12-12 16:55:24 -08002608static void __init hugetlb_register_all_nodes(void)
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002609{
2610 int nid;
2611
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08002612 for_each_node_state(nid, N_MEMORY) {
Wen Congyang87327942012-12-11 16:00:56 -08002613 struct node *node = node_devices[nid];
Kay Sievers10fbcf42011-12-21 14:48:43 -08002614 if (node->dev.id == nid)
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002615 hugetlb_register_node(node);
2616 }
2617
2618 /*
Kay Sievers10fbcf42011-12-21 14:48:43 -08002619 * Let the node device driver know we're here so it can
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002620 * [un]register hstate attributes on node hotplug.
2621 */
2622 register_hugetlbfs_with_node(hugetlb_register_node,
2623 hugetlb_unregister_node);
2624}
2625#else /* !CONFIG_NUMA */
2626
2627static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
2628{
2629 BUG();
2630 if (nidp)
2631 *nidp = -1;
2632 return NULL;
2633}
2634
2635static void hugetlb_unregister_all_nodes(void) { }
2636
2637static void hugetlb_register_all_nodes(void) { }
2638
2639#endif
2640
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002641static void __exit hugetlb_exit(void)
2642{
2643 struct hstate *h;
2644
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002645 hugetlb_unregister_all_nodes();
2646
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002647 for_each_hstate(h) {
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07002648 kobject_put(hstate_kobjs[hstate_index(h)]);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002649 }
2650
2651 kobject_put(hugepages_kobj);
Mike Kravetzc672c7f2015-09-08 15:01:35 -07002652 kfree(hugetlb_fault_mutex_table);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002653}
2654module_exit(hugetlb_exit);
2655
2656static int __init hugetlb_init(void)
2657{
Davidlohr Bueso8382d912014-04-03 14:47:31 -07002658 int i;
2659
Nishanth Aravamudan457c1b22014-05-06 12:50:00 -07002660 if (!hugepages_supported())
Benjamin Herrenschmidt0ef89d22008-07-31 00:07:30 -07002661 return 0;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002662
Nick Piggine11bfbf2008-07-23 21:27:52 -07002663 if (!size_to_hstate(default_hstate_size)) {
2664 default_hstate_size = HPAGE_SIZE;
2665 if (!size_to_hstate(default_hstate_size))
2666 hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002667 }
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07002668 default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size));
Nick Piggine11bfbf2008-07-23 21:27:52 -07002669 if (default_hstate_max_huge_pages)
2670 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002671
2672 hugetlb_init_hstates();
Andi Kleenaa888a72008-07-23 21:27:47 -07002673 gather_bootmem_prealloc();
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002674 report_hugepages();
2675
2676 hugetlb_sysfs_init();
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002677 hugetlb_register_all_nodes();
Jianguo Wu7179e7b2012-12-18 14:23:19 -08002678 hugetlb_cgroup_file_init();
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002679
Davidlohr Bueso8382d912014-04-03 14:47:31 -07002680#ifdef CONFIG_SMP
2681 num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus());
2682#else
2683 num_fault_mutexes = 1;
2684#endif
Mike Kravetzc672c7f2015-09-08 15:01:35 -07002685 hugetlb_fault_mutex_table =
Davidlohr Bueso8382d912014-04-03 14:47:31 -07002686 kmalloc(sizeof(struct mutex) * num_fault_mutexes, GFP_KERNEL);
Mike Kravetzc672c7f2015-09-08 15:01:35 -07002687 BUG_ON(!hugetlb_fault_mutex_table);
Davidlohr Bueso8382d912014-04-03 14:47:31 -07002688
2689 for (i = 0; i < num_fault_mutexes; i++)
Mike Kravetzc672c7f2015-09-08 15:01:35 -07002690 mutex_init(&hugetlb_fault_mutex_table[i]);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002691 return 0;
2692}
2693module_init(hugetlb_init);
2694
2695/* Should be called on processing a hugepagesz=... option */
Kirill A. Shutemovd00181b2015-11-06 16:29:57 -08002696void __init hugetlb_add_hstate(unsigned int order)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002697{
2698 struct hstate *h;
Andi Kleen8faa8b02008-07-23 21:27:48 -07002699 unsigned long i;
2700
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002701 if (size_to_hstate(PAGE_SIZE << order)) {
Andrew Mortonffb22af2013-02-22 16:32:08 -08002702 pr_warning("hugepagesz= specified twice, ignoring\n");
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002703 return;
2704 }
Aneesh Kumar K.V47d38342012-07-31 16:41:54 -07002705 BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002706 BUG_ON(order == 0);
Aneesh Kumar K.V47d38342012-07-31 16:41:54 -07002707 h = &hstates[hugetlb_max_hstate++];
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002708 h->order = order;
2709 h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
Andi Kleen8faa8b02008-07-23 21:27:48 -07002710 h->nr_huge_pages = 0;
2711 h->free_huge_pages = 0;
2712 for (i = 0; i < MAX_NUMNODES; ++i)
2713 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
Aneesh Kumar K.V0edaecf2012-07-31 16:42:07 -07002714 INIT_LIST_HEAD(&h->hugepage_activelist);
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08002715 h->next_nid_to_alloc = first_node(node_states[N_MEMORY]);
2716 h->next_nid_to_free = first_node(node_states[N_MEMORY]);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002717 snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
2718 huge_page_size(h)/1024);
Andi Kleen8faa8b02008-07-23 21:27:48 -07002719
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002720 parsed_hstate = h;
2721}
2722
Nick Piggine11bfbf2008-07-23 21:27:52 -07002723static int __init hugetlb_nrpages_setup(char *s)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002724{
2725 unsigned long *mhp;
Andi Kleen8faa8b02008-07-23 21:27:48 -07002726 static unsigned long *last_mhp;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002727
2728 /*
Aneesh Kumar K.V47d38342012-07-31 16:41:54 -07002729 * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet,
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002730 * so this hugepages= parameter goes to the "default hstate".
2731 */
Aneesh Kumar K.V47d38342012-07-31 16:41:54 -07002732 if (!hugetlb_max_hstate)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002733 mhp = &default_hstate_max_huge_pages;
2734 else
2735 mhp = &parsed_hstate->max_huge_pages;
2736
Andi Kleen8faa8b02008-07-23 21:27:48 -07002737 if (mhp == last_mhp) {
Andrew Mortonffb22af2013-02-22 16:32:08 -08002738 pr_warning("hugepages= specified twice without "
2739 "interleaving hugepagesz=, ignoring\n");
Andi Kleen8faa8b02008-07-23 21:27:48 -07002740 return 1;
2741 }
2742
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002743 if (sscanf(s, "%lu", mhp) <= 0)
2744 *mhp = 0;
2745
Andi Kleen8faa8b02008-07-23 21:27:48 -07002746 /*
2747 * Global state is always initialized later in hugetlb_init.
2748 * But we need to allocate >= MAX_ORDER hstates here early to still
2749 * use the bootmem allocator.
2750 */
Aneesh Kumar K.V47d38342012-07-31 16:41:54 -07002751 if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER)
Andi Kleen8faa8b02008-07-23 21:27:48 -07002752 hugetlb_hstate_alloc_pages(parsed_hstate);
2753
2754 last_mhp = mhp;
2755
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002756 return 1;
2757}
Nick Piggine11bfbf2008-07-23 21:27:52 -07002758__setup("hugepages=", hugetlb_nrpages_setup);
2759
2760static int __init hugetlb_default_setup(char *s)
2761{
2762 default_hstate_size = memparse(s, &s);
2763 return 1;
2764}
2765__setup("default_hugepagesz=", hugetlb_default_setup);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002766
Nishanth Aravamudan8a213462008-07-25 19:44:37 -07002767static unsigned int cpuset_mems_nr(unsigned int *array)
2768{
2769 int node;
2770 unsigned int nr = 0;
2771
2772 for_each_node_mask(node, cpuset_current_mems_allowed)
2773 nr += array[node];
2774
2775 return nr;
2776}
2777
2778#ifdef CONFIG_SYSCTL
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002779static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
2780 struct ctl_table *table, int write,
2781 void __user *buffer, size_t *length, loff_t *ppos)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002782{
Andi Kleene5ff2152008-07-23 21:27:42 -07002783 struct hstate *h = &default_hstate;
David Rientjes238d3c12014-08-06 16:06:51 -07002784 unsigned long tmp = h->max_huge_pages;
Michal Hocko08d4a242011-01-13 15:47:26 -08002785 int ret;
Andi Kleene5ff2152008-07-23 21:27:42 -07002786
Nishanth Aravamudan457c1b22014-05-06 12:50:00 -07002787 if (!hugepages_supported())
2788 return -ENOTSUPP;
2789
Andi Kleene5ff2152008-07-23 21:27:42 -07002790 table->data = &tmp;
2791 table->maxlen = sizeof(unsigned long);
Michal Hocko08d4a242011-01-13 15:47:26 -08002792 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2793 if (ret)
2794 goto out;
Andi Kleene5ff2152008-07-23 21:27:42 -07002795
David Rientjes238d3c12014-08-06 16:06:51 -07002796 if (write)
2797 ret = __nr_hugepages_store_common(obey_mempolicy, h,
2798 NUMA_NO_NODE, tmp, *length);
Michal Hocko08d4a242011-01-13 15:47:26 -08002799out:
2800 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002801}
Mel Gorman396faf02007-07-17 04:03:13 -07002802
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002803int hugetlb_sysctl_handler(struct ctl_table *table, int write,
2804 void __user *buffer, size_t *length, loff_t *ppos)
2805{
2806
2807 return hugetlb_sysctl_handler_common(false, table, write,
2808 buffer, length, ppos);
2809}
2810
2811#ifdef CONFIG_NUMA
2812int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
2813 void __user *buffer, size_t *length, loff_t *ppos)
2814{
2815 return hugetlb_sysctl_handler_common(true, table, write,
2816 buffer, length, ppos);
2817}
2818#endif /* CONFIG_NUMA */
2819
Nishanth Aravamudana3d0c6a2008-02-08 04:18:18 -08002820int hugetlb_overcommit_handler(struct ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07002821 void __user *buffer,
Nishanth Aravamudana3d0c6a2008-02-08 04:18:18 -08002822 size_t *length, loff_t *ppos)
2823{
Andi Kleena5516432008-07-23 21:27:41 -07002824 struct hstate *h = &default_hstate;
Andi Kleene5ff2152008-07-23 21:27:42 -07002825 unsigned long tmp;
Michal Hocko08d4a242011-01-13 15:47:26 -08002826 int ret;
Andi Kleene5ff2152008-07-23 21:27:42 -07002827
Nishanth Aravamudan457c1b22014-05-06 12:50:00 -07002828 if (!hugepages_supported())
2829 return -ENOTSUPP;
2830
Petr Holasekc033a932011-03-22 16:33:05 -07002831 tmp = h->nr_overcommit_huge_pages;
Andi Kleene5ff2152008-07-23 21:27:42 -07002832
Luiz Capitulinobae7f4a2014-06-04 16:07:08 -07002833 if (write && hstate_is_gigantic(h))
Eric B Munsonadbe8722011-01-13 15:47:27 -08002834 return -EINVAL;
2835
Andi Kleene5ff2152008-07-23 21:27:42 -07002836 table->data = &tmp;
2837 table->maxlen = sizeof(unsigned long);
Michal Hocko08d4a242011-01-13 15:47:26 -08002838 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2839 if (ret)
2840 goto out;
Andi Kleene5ff2152008-07-23 21:27:42 -07002841
2842 if (write) {
2843 spin_lock(&hugetlb_lock);
2844 h->nr_overcommit_huge_pages = tmp;
2845 spin_unlock(&hugetlb_lock);
2846 }
Michal Hocko08d4a242011-01-13 15:47:26 -08002847out:
2848 return ret;
Nishanth Aravamudana3d0c6a2008-02-08 04:18:18 -08002849}
2850
Linus Torvalds1da177e2005-04-16 15:20:36 -07002851#endif /* CONFIG_SYSCTL */
2852
Alexey Dobriyane1759c22008-10-15 23:50:22 +04002853void hugetlb_report_meminfo(struct seq_file *m)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002854{
Andi Kleena5516432008-07-23 21:27:41 -07002855 struct hstate *h = &default_hstate;
Nishanth Aravamudan457c1b22014-05-06 12:50:00 -07002856 if (!hugepages_supported())
2857 return;
Alexey Dobriyane1759c22008-10-15 23:50:22 +04002858 seq_printf(m,
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002859 "HugePages_Total: %5lu\n"
2860 "HugePages_Free: %5lu\n"
2861 "HugePages_Rsvd: %5lu\n"
2862 "HugePages_Surp: %5lu\n"
2863 "Hugepagesize: %8lu kB\n",
Andi Kleena5516432008-07-23 21:27:41 -07002864 h->nr_huge_pages,
2865 h->free_huge_pages,
2866 h->resv_huge_pages,
2867 h->surplus_huge_pages,
2868 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002869}
2870
2871int hugetlb_report_node_meminfo(int nid, char *buf)
2872{
Andi Kleena5516432008-07-23 21:27:41 -07002873 struct hstate *h = &default_hstate;
Nishanth Aravamudan457c1b22014-05-06 12:50:00 -07002874 if (!hugepages_supported())
2875 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002876 return sprintf(buf,
2877 "Node %d HugePages_Total: %5u\n"
Nishanth Aravamudana1de0912008-03-26 14:37:53 -07002878 "Node %d HugePages_Free: %5u\n"
2879 "Node %d HugePages_Surp: %5u\n",
Andi Kleena5516432008-07-23 21:27:41 -07002880 nid, h->nr_huge_pages_node[nid],
2881 nid, h->free_huge_pages_node[nid],
2882 nid, h->surplus_huge_pages_node[nid]);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002883}
2884
David Rientjes949f7ec2013-04-29 15:07:48 -07002885void hugetlb_show_meminfo(void)
2886{
2887 struct hstate *h;
2888 int nid;
2889
Nishanth Aravamudan457c1b22014-05-06 12:50:00 -07002890 if (!hugepages_supported())
2891 return;
2892
David Rientjes949f7ec2013-04-29 15:07:48 -07002893 for_each_node_state(nid, N_MEMORY)
2894 for_each_hstate(h)
2895 pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
2896 nid,
2897 h->nr_huge_pages_node[nid],
2898 h->free_huge_pages_node[nid],
2899 h->surplus_huge_pages_node[nid],
2900 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
2901}
2902
Naoya Horiguchi5d317b22015-11-05 18:47:14 -08002903void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm)
2904{
2905 seq_printf(m, "HugetlbPages:\t%8lu kB\n",
2906 atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10));
2907}
2908
Linus Torvalds1da177e2005-04-16 15:20:36 -07002909/* Return the number pages of memory we physically have, in PAGE_SIZE units. */
2910unsigned long hugetlb_total_pages(void)
2911{
Wanpeng Lid0028582013-03-22 15:04:40 -07002912 struct hstate *h;
2913 unsigned long nr_total_pages = 0;
2914
2915 for_each_hstate(h)
2916 nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h);
2917 return nr_total_pages;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002918}
Linus Torvalds1da177e2005-04-16 15:20:36 -07002919
Andi Kleena5516432008-07-23 21:27:41 -07002920static int hugetlb_acct_memory(struct hstate *h, long delta)
Mel Gormanfc1b8a72008-07-23 21:27:22 -07002921{
2922 int ret = -ENOMEM;
2923
2924 spin_lock(&hugetlb_lock);
2925 /*
2926 * When cpuset is configured, it breaks the strict hugetlb page
2927 * reservation as the accounting is done on a global variable. Such
2928 * reservation is completely rubbish in the presence of cpuset because
2929 * the reservation is not checked against page availability for the
2930 * current cpuset. Application can still potentially OOM'ed by kernel
2931 * with lack of free htlb page in cpuset that the task is in.
2932 * Attempt to enforce strict accounting with cpuset is almost
2933 * impossible (or too ugly) because cpuset is too fluid that
2934 * task or memory node can be dynamically moved between cpusets.
2935 *
2936 * The change of semantics for shared hugetlb mapping with cpuset is
2937 * undesirable. However, in order to preserve some of the semantics,
2938 * we fall back to check against current free page availability as
2939 * a best attempt and hopefully to minimize the impact of changing
2940 * semantics that cpuset has.
2941 */
2942 if (delta > 0) {
Andi Kleena5516432008-07-23 21:27:41 -07002943 if (gather_surplus_pages(h, delta) < 0)
Mel Gormanfc1b8a72008-07-23 21:27:22 -07002944 goto out;
2945
Andi Kleena5516432008-07-23 21:27:41 -07002946 if (delta > cpuset_mems_nr(h->free_huge_pages_node)) {
2947 return_unused_surplus_pages(h, delta);
Mel Gormanfc1b8a72008-07-23 21:27:22 -07002948 goto out;
2949 }
2950 }
2951
2952 ret = 0;
2953 if (delta < 0)
Andi Kleena5516432008-07-23 21:27:41 -07002954 return_unused_surplus_pages(h, (unsigned long) -delta);
Mel Gormanfc1b8a72008-07-23 21:27:22 -07002955
2956out:
2957 spin_unlock(&hugetlb_lock);
2958 return ret;
2959}
2960
Andy Whitcroft84afd992008-07-23 21:27:32 -07002961static void hugetlb_vm_op_open(struct vm_area_struct *vma)
2962{
Joonsoo Kimf522c3a2013-09-11 14:21:53 -07002963 struct resv_map *resv = vma_resv_map(vma);
Andy Whitcroft84afd992008-07-23 21:27:32 -07002964
2965 /*
2966 * This new VMA should share its siblings reservation map if present.
2967 * The VMA will only ever have a valid reservation map pointer where
2968 * it is being copied for another still existing VMA. As that VMA
Lucas De Marchi25985ed2011-03-30 22:57:33 -03002969 * has a reference to the reservation map it cannot disappear until
Andy Whitcroft84afd992008-07-23 21:27:32 -07002970 * after this open call completes. It is therefore safe to take a
2971 * new reference here without additional locking.
2972 */
Joonsoo Kim4e35f482014-04-03 14:47:30 -07002973 if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
Joonsoo Kimf522c3a2013-09-11 14:21:53 -07002974 kref_get(&resv->refs);
Andy Whitcroft84afd992008-07-23 21:27:32 -07002975}
2976
Mel Gormana1e78772008-07-23 21:27:23 -07002977static void hugetlb_vm_op_close(struct vm_area_struct *vma)
2978{
Andi Kleena5516432008-07-23 21:27:41 -07002979 struct hstate *h = hstate_vma(vma);
Joonsoo Kimf522c3a2013-09-11 14:21:53 -07002980 struct resv_map *resv = vma_resv_map(vma);
David Gibson90481622012-03-21 16:34:12 -07002981 struct hugepage_subpool *spool = subpool_vma(vma);
Joonsoo Kim4e35f482014-04-03 14:47:30 -07002982 unsigned long reserve, start, end;
Mike Kravetz1c5ecae2015-04-15 16:13:39 -07002983 long gbl_reserve;
Andy Whitcroft84afd992008-07-23 21:27:32 -07002984
Joonsoo Kim4e35f482014-04-03 14:47:30 -07002985 if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER))
2986 return;
Andy Whitcroft84afd992008-07-23 21:27:32 -07002987
Joonsoo Kim4e35f482014-04-03 14:47:30 -07002988 start = vma_hugecache_offset(h, vma, vma->vm_start);
2989 end = vma_hugecache_offset(h, vma, vma->vm_end);
Andy Whitcroft84afd992008-07-23 21:27:32 -07002990
Joonsoo Kim4e35f482014-04-03 14:47:30 -07002991 reserve = (end - start) - region_count(resv, start, end);
Andy Whitcroft84afd992008-07-23 21:27:32 -07002992
Joonsoo Kim4e35f482014-04-03 14:47:30 -07002993 kref_put(&resv->refs, resv_map_release);
2994
2995 if (reserve) {
Mike Kravetz1c5ecae2015-04-15 16:13:39 -07002996 /*
2997 * Decrement reserve counts. The global reserve count may be
2998 * adjusted if the subpool has a minimum size.
2999 */
3000 gbl_reserve = hugepage_subpool_put_pages(spool, reserve);
3001 hugetlb_acct_memory(h, -gbl_reserve);
Andy Whitcroft84afd992008-07-23 21:27:32 -07003002 }
Mel Gormana1e78772008-07-23 21:27:23 -07003003}
3004
Linus Torvalds1da177e2005-04-16 15:20:36 -07003005/*
3006 * We cannot handle pagefaults against hugetlb pages at all. They cause
3007 * handle_mm_fault() to try to instantiate regular-sized pages in the
3008 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
3009 * this far.
3010 */
Nick Piggind0217ac2007-07-19 01:47:03 -07003011static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003012{
3013 BUG();
Nick Piggind0217ac2007-07-19 01:47:03 -07003014 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003015}
3016
Alexey Dobriyanf0f37e22009-09-27 22:29:37 +04003017const struct vm_operations_struct hugetlb_vm_ops = {
Nick Piggind0217ac2007-07-19 01:47:03 -07003018 .fault = hugetlb_vm_op_fault,
Andy Whitcroft84afd992008-07-23 21:27:32 -07003019 .open = hugetlb_vm_op_open,
Mel Gormana1e78772008-07-23 21:27:23 -07003020 .close = hugetlb_vm_op_close,
Linus Torvalds1da177e2005-04-16 15:20:36 -07003021};
3022
David Gibson1e8f8892006-01-06 00:10:44 -08003023static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
3024 int writable)
David Gibson63551ae2005-06-21 17:14:44 -07003025{
3026 pte_t entry;
3027
David Gibson1e8f8892006-01-06 00:10:44 -08003028 if (writable) {
Gerald Schaefer106c9922013-04-29 15:07:23 -07003029 entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page,
3030 vma->vm_page_prot)));
David Gibson63551ae2005-06-21 17:14:44 -07003031 } else {
Gerald Schaefer106c9922013-04-29 15:07:23 -07003032 entry = huge_pte_wrprotect(mk_huge_pte(page,
3033 vma->vm_page_prot));
David Gibson63551ae2005-06-21 17:14:44 -07003034 }
3035 entry = pte_mkyoung(entry);
3036 entry = pte_mkhuge(entry);
Chris Metcalfd9ed9fa2012-04-01 14:01:34 -04003037 entry = arch_make_huge_pte(entry, vma, page, writable);
David Gibson63551ae2005-06-21 17:14:44 -07003038
3039 return entry;
3040}
3041
David Gibson1e8f8892006-01-06 00:10:44 -08003042static void set_huge_ptep_writable(struct vm_area_struct *vma,
3043 unsigned long address, pte_t *ptep)
3044{
3045 pte_t entry;
3046
Gerald Schaefer106c9922013-04-29 15:07:23 -07003047 entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep)));
Chris Forbes32f84522011-07-25 17:12:14 -07003048 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
Russell King4b3073e2009-12-18 16:40:18 +00003049 update_mmu_cache(vma, address, ptep);
David Gibson1e8f8892006-01-06 00:10:44 -08003050}
3051
Naoya Horiguchi4a705fe2014-06-23 13:22:03 -07003052static int is_hugetlb_entry_migration(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_migration_entry(swp))
3060 return 1;
3061 else
3062 return 0;
3063}
3064
3065static int is_hugetlb_entry_hwpoisoned(pte_t pte)
3066{
3067 swp_entry_t swp;
3068
3069 if (huge_pte_none(pte) || pte_present(pte))
3070 return 0;
3071 swp = pte_to_swp_entry(pte);
3072 if (non_swap_entry(swp) && is_hwpoison_entry(swp))
3073 return 1;
3074 else
3075 return 0;
3076}
David Gibson1e8f8892006-01-06 00:10:44 -08003077
David Gibson63551ae2005-06-21 17:14:44 -07003078int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
3079 struct vm_area_struct *vma)
3080{
3081 pte_t *src_pte, *dst_pte, entry;
3082 struct page *ptepage;
Hugh Dickins1c598272005-10-19 21:23:43 -07003083 unsigned long addr;
David Gibson1e8f8892006-01-06 00:10:44 -08003084 int cow;
Andi Kleena5516432008-07-23 21:27:41 -07003085 struct hstate *h = hstate_vma(vma);
3086 unsigned long sz = huge_page_size(h);
Andreas Sandberge8569dd2014-01-21 15:49:09 -08003087 unsigned long mmun_start; /* For mmu_notifiers */
3088 unsigned long mmun_end; /* For mmu_notifiers */
3089 int ret = 0;
David Gibson1e8f8892006-01-06 00:10:44 -08003090
3091 cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
David Gibson63551ae2005-06-21 17:14:44 -07003092
Andreas Sandberge8569dd2014-01-21 15:49:09 -08003093 mmun_start = vma->vm_start;
3094 mmun_end = vma->vm_end;
3095 if (cow)
3096 mmu_notifier_invalidate_range_start(src, mmun_start, mmun_end);
3097
Andi Kleena5516432008-07-23 21:27:41 -07003098 for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003099 spinlock_t *src_ptl, *dst_ptl;
Hugh Dickinsc74df322005-10-29 18:16:23 -07003100 src_pte = huge_pte_offset(src, addr);
3101 if (!src_pte)
3102 continue;
Andi Kleena5516432008-07-23 21:27:41 -07003103 dst_pte = huge_pte_alloc(dst, addr, sz);
Andreas Sandberge8569dd2014-01-21 15:49:09 -08003104 if (!dst_pte) {
3105 ret = -ENOMEM;
3106 break;
3107 }
Larry Woodmanc5c99422008-01-24 05:49:25 -08003108
3109 /* If the pagetables are shared don't copy or take references */
3110 if (dst_pte == src_pte)
3111 continue;
3112
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003113 dst_ptl = huge_pte_lock(h, dst, dst_pte);
3114 src_ptl = huge_pte_lockptr(h, src, src_pte);
3115 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
Naoya Horiguchi4a705fe2014-06-23 13:22:03 -07003116 entry = huge_ptep_get(src_pte);
3117 if (huge_pte_none(entry)) { /* skip none entry */
3118 ;
3119 } else if (unlikely(is_hugetlb_entry_migration(entry) ||
3120 is_hugetlb_entry_hwpoisoned(entry))) {
3121 swp_entry_t swp_entry = pte_to_swp_entry(entry);
3122
3123 if (is_write_migration_entry(swp_entry) && cow) {
3124 /*
3125 * COW mappings require pages in both
3126 * parent and child to be set to read.
3127 */
3128 make_migration_entry_read(&swp_entry);
3129 entry = swp_entry_to_pte(swp_entry);
3130 set_huge_pte_at(src, addr, src_pte, entry);
3131 }
3132 set_huge_pte_at(dst, addr, dst_pte, entry);
3133 } else {
Joerg Roedel34ee6452014-11-13 13:46:09 +11003134 if (cow) {
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07003135 huge_ptep_set_wrprotect(src, addr, src_pte);
Joerg Roedel34ee6452014-11-13 13:46:09 +11003136 mmu_notifier_invalidate_range(src, mmun_start,
3137 mmun_end);
3138 }
Naoya Horiguchi0253d632014-07-23 14:00:19 -07003139 entry = huge_ptep_get(src_pte);
Hugh Dickins1c598272005-10-19 21:23:43 -07003140 ptepage = pte_page(entry);
3141 get_page(ptepage);
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09003142 page_dup_rmap(ptepage);
Hugh Dickins1c598272005-10-19 21:23:43 -07003143 set_huge_pte_at(dst, addr, dst_pte, entry);
Naoya Horiguchi5d317b22015-11-05 18:47:14 -08003144 hugetlb_count_add(pages_per_huge_page(h), dst);
Hugh Dickins1c598272005-10-19 21:23:43 -07003145 }
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003146 spin_unlock(src_ptl);
3147 spin_unlock(dst_ptl);
David Gibson63551ae2005-06-21 17:14:44 -07003148 }
David Gibson63551ae2005-06-21 17:14:44 -07003149
Andreas Sandberge8569dd2014-01-21 15:49:09 -08003150 if (cow)
3151 mmu_notifier_invalidate_range_end(src, mmun_start, mmun_end);
3152
3153 return ret;
David Gibson63551ae2005-06-21 17:14:44 -07003154}
3155
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07003156void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
3157 unsigned long start, unsigned long end,
3158 struct page *ref_page)
David Gibson63551ae2005-06-21 17:14:44 -07003159{
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07003160 int force_flush = 0;
David Gibson63551ae2005-06-21 17:14:44 -07003161 struct mm_struct *mm = vma->vm_mm;
3162 unsigned long address;
David Gibsonc7546f82005-08-05 11:59:35 -07003163 pte_t *ptep;
David Gibson63551ae2005-06-21 17:14:44 -07003164 pte_t pte;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003165 spinlock_t *ptl;
David Gibson63551ae2005-06-21 17:14:44 -07003166 struct page *page;
Andi Kleena5516432008-07-23 21:27:41 -07003167 struct hstate *h = hstate_vma(vma);
3168 unsigned long sz = huge_page_size(h);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07003169 const unsigned long mmun_start = start; /* For mmu_notifiers */
3170 const unsigned long mmun_end = end; /* For mmu_notifiers */
Andi Kleena5516432008-07-23 21:27:41 -07003171
David Gibson63551ae2005-06-21 17:14:44 -07003172 WARN_ON(!is_vm_hugetlb_page(vma));
Andi Kleena5516432008-07-23 21:27:41 -07003173 BUG_ON(start & ~huge_page_mask(h));
3174 BUG_ON(end & ~huge_page_mask(h));
David Gibson63551ae2005-06-21 17:14:44 -07003175
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07003176 tlb_start_vma(tlb, vma);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07003177 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
Hillf Danton569f48b2014-12-10 15:44:41 -08003178 address = start;
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07003179again:
Hillf Danton569f48b2014-12-10 15:44:41 -08003180 for (; address < end; address += sz) {
David Gibsonc7546f82005-08-05 11:59:35 -07003181 ptep = huge_pte_offset(mm, address);
Adam Litke4c887262005-10-29 18:16:46 -07003182 if (!ptep)
David Gibsonc7546f82005-08-05 11:59:35 -07003183 continue;
3184
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003185 ptl = huge_pte_lock(h, mm, ptep);
Chen, Kenneth W39dde652006-12-06 20:32:03 -08003186 if (huge_pmd_unshare(mm, &address, ptep))
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003187 goto unlock;
Chen, Kenneth W39dde652006-12-06 20:32:03 -08003188
Hillf Danton66293262012-03-23 15:01:48 -07003189 pte = huge_ptep_get(ptep);
3190 if (huge_pte_none(pte))
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003191 goto unlock;
Hillf Danton66293262012-03-23 15:01:48 -07003192
3193 /*
Naoya Horiguchi9fbc1f62015-02-11 15:25:32 -08003194 * Migrating hugepage or HWPoisoned hugepage is already
3195 * unmapped and its refcount is dropped, so just clear pte here.
Hillf Danton66293262012-03-23 15:01:48 -07003196 */
Naoya Horiguchi9fbc1f62015-02-11 15:25:32 -08003197 if (unlikely(!pte_present(pte))) {
Gerald Schaefer106c9922013-04-29 15:07:23 -07003198 huge_pte_clear(mm, address, ptep);
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003199 goto unlock;
Naoya Horiguchi8c4894c2012-12-12 13:52:28 -08003200 }
Hillf Danton66293262012-03-23 15:01:48 -07003201
3202 page = pte_page(pte);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003203 /*
3204 * If a reference page is supplied, it is because a specific
3205 * page is being unmapped, not a range. Ensure the page we
3206 * are about to unmap is the actual page of interest.
3207 */
3208 if (ref_page) {
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003209 if (page != ref_page)
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003210 goto unlock;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003211
3212 /*
3213 * Mark the VMA as having unmapped its page so that
3214 * future faults in this VMA will fail rather than
3215 * looking like data was lost
3216 */
3217 set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
3218 }
3219
David Gibsonc7546f82005-08-05 11:59:35 -07003220 pte = huge_ptep_get_and_clear(mm, address, ptep);
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07003221 tlb_remove_tlb_entry(tlb, ptep, address);
Gerald Schaefer106c9922013-04-29 15:07:23 -07003222 if (huge_pte_dirty(pte))
Ken Chen6649a382007-02-08 14:20:27 -08003223 set_page_dirty(page);
Hillf Danton9e811302012-03-21 16:34:03 -07003224
Naoya Horiguchi5d317b22015-11-05 18:47:14 -08003225 hugetlb_count_sub(pages_per_huge_page(h), mm);
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07003226 page_remove_rmap(page);
3227 force_flush = !__tlb_remove_page(tlb, page);
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003228 if (force_flush) {
Hillf Danton569f48b2014-12-10 15:44:41 -08003229 address += sz;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003230 spin_unlock(ptl);
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07003231 break;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003232 }
Hillf Danton9e811302012-03-21 16:34:03 -07003233 /* Bail out after unmapping reference page if supplied */
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003234 if (ref_page) {
3235 spin_unlock(ptl);
Hillf Danton9e811302012-03-21 16:34:03 -07003236 break;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003237 }
3238unlock:
3239 spin_unlock(ptl);
David Gibson63551ae2005-06-21 17:14:44 -07003240 }
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07003241 /*
3242 * mmu_gather ran out of room to batch pages, we break out of
3243 * the PTE lock to avoid doing the potential expensive TLB invalidate
3244 * and page-free while holding it.
3245 */
3246 if (force_flush) {
3247 force_flush = 0;
3248 tlb_flush_mmu(tlb);
3249 if (address < end && !ref_page)
3250 goto again;
Chen, Kenneth Wfe1668a2006-10-04 02:15:24 -07003251 }
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07003252 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07003253 tlb_end_vma(tlb, vma);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003254}
David Gibson63551ae2005-06-21 17:14:44 -07003255
Mel Gormand8333522012-07-31 16:46:20 -07003256void __unmap_hugepage_range_final(struct mmu_gather *tlb,
3257 struct vm_area_struct *vma, unsigned long start,
3258 unsigned long end, struct page *ref_page)
3259{
3260 __unmap_hugepage_range(tlb, vma, start, end, ref_page);
3261
3262 /*
3263 * Clear this flag so that x86's huge_pmd_share page_table_shareable
3264 * test will fail on a vma being torn down, and not grab a page table
3265 * on its way out. We're lucky that the flag has such an appropriate
3266 * name, and can in fact be safely cleared here. We could clear it
3267 * before the __unmap_hugepage_range above, but all that's necessary
Davidlohr Buesoc8c06ef2014-12-12 16:54:24 -08003268 * is to clear it before releasing the i_mmap_rwsem. This works
Mel Gormand8333522012-07-31 16:46:20 -07003269 * because in the context this is called, the VMA is about to be
Davidlohr Buesoc8c06ef2014-12-12 16:54:24 -08003270 * destroyed and the i_mmap_rwsem is held.
Mel Gormand8333522012-07-31 16:46:20 -07003271 */
3272 vma->vm_flags &= ~VM_MAYSHARE;
3273}
3274
Chen, Kenneth W502717f2006-10-11 01:20:46 -07003275void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003276 unsigned long end, struct page *ref_page)
Chen, Kenneth W502717f2006-10-11 01:20:46 -07003277{
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07003278 struct mm_struct *mm;
3279 struct mmu_gather tlb;
3280
3281 mm = vma->vm_mm;
3282
Linus Torvalds2b047252013-08-15 11:42:25 -07003283 tlb_gather_mmu(&tlb, mm, start, end);
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07003284 __unmap_hugepage_range(&tlb, vma, start, end, ref_page);
3285 tlb_finish_mmu(&tlb, start, end);
Chen, Kenneth W502717f2006-10-11 01:20:46 -07003286}
3287
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003288/*
3289 * This is called when the original mapper is failing to COW a MAP_PRIVATE
3290 * mappping it owns the reserve page for. The intention is to unmap the page
3291 * from other VMAs and let the children be SIGKILLed if they are faulting the
3292 * same region.
3293 */
Davidlohr Bueso2f4612a2014-08-06 16:06:45 -07003294static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
3295 struct page *page, unsigned long address)
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003296{
Adam Litke75266742008-11-12 13:24:56 -08003297 struct hstate *h = hstate_vma(vma);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003298 struct vm_area_struct *iter_vma;
3299 struct address_space *mapping;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003300 pgoff_t pgoff;
3301
3302 /*
3303 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
3304 * from page cache lookup which is in HPAGE_SIZE units.
3305 */
Adam Litke75266742008-11-12 13:24:56 -08003306 address = address & huge_page_mask(h);
Michal Hocko36e4f202012-10-08 16:33:31 -07003307 pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) +
3308 vma->vm_pgoff;
Al Viro496ad9a2013-01-23 17:07:38 -05003309 mapping = file_inode(vma->vm_file)->i_mapping;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003310
Mel Gorman4eb2b1d2009-12-14 17:59:53 -08003311 /*
3312 * Take the mapping lock for the duration of the table walk. As
3313 * this mapping should be shared between all the VMAs,
3314 * __unmap_hugepage_range() is called as the lock is already held
3315 */
Davidlohr Bueso83cde9e2014-12-12 16:54:21 -08003316 i_mmap_lock_write(mapping);
Michel Lespinasse6b2dbba2012-10-08 16:31:25 -07003317 vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) {
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003318 /* Do not unmap the current VMA */
3319 if (iter_vma == vma)
3320 continue;
3321
3322 /*
Mel Gorman2f84a892015-10-01 15:36:57 -07003323 * Shared VMAs have their own reserves and do not affect
3324 * MAP_PRIVATE accounting but it is possible that a shared
3325 * VMA is using the same page so check and skip such VMAs.
3326 */
3327 if (iter_vma->vm_flags & VM_MAYSHARE)
3328 continue;
3329
3330 /*
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003331 * Unmap the page from other VMAs without their own reserves.
3332 * They get marked to be SIGKILLed if they fault in these
3333 * areas. This is because a future no-page fault on this VMA
3334 * could insert a zeroed page instead of the data existing
3335 * from the time of fork. This would look like data corruption
3336 */
3337 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07003338 unmap_hugepage_range(iter_vma, address,
3339 address + huge_page_size(h), page);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003340 }
Davidlohr Bueso83cde9e2014-12-12 16:54:21 -08003341 i_mmap_unlock_write(mapping);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003342}
3343
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09003344/*
3345 * Hugetlb_cow() should be called with page lock of the original hugepage held.
Michal Hockoef009b22012-01-10 15:07:21 -08003346 * Called with hugetlb_instantiation_mutex held and pte_page locked so we
3347 * cannot race with other handlers or page migration.
3348 * Keep the pte_same checks anyway to make transition from the mutex easier.
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09003349 */
David Gibson1e8f8892006-01-06 00:10:44 -08003350static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003351 unsigned long address, pte_t *ptep, pte_t pte,
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003352 struct page *pagecache_page, spinlock_t *ptl)
David Gibson1e8f8892006-01-06 00:10:44 -08003353{
Andi Kleena5516432008-07-23 21:27:41 -07003354 struct hstate *h = hstate_vma(vma);
David Gibson1e8f8892006-01-06 00:10:44 -08003355 struct page *old_page, *new_page;
Davidlohr Buesoad4404a2014-08-06 16:06:47 -07003356 int ret = 0, outside_reserve = 0;
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07003357 unsigned long mmun_start; /* For mmu_notifiers */
3358 unsigned long mmun_end; /* For mmu_notifiers */
David Gibson1e8f8892006-01-06 00:10:44 -08003359
3360 old_page = pte_page(pte);
3361
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003362retry_avoidcopy:
David Gibson1e8f8892006-01-06 00:10:44 -08003363 /* If no-one else is actually using this page, avoid the copy
3364 * and just make the page writable */
Joonsoo Kim37a21402013-09-11 14:21:04 -07003365 if (page_mapcount(old_page) == 1 && PageAnon(old_page)) {
3366 page_move_anon_rmap(old_page, vma, address);
David Gibson1e8f8892006-01-06 00:10:44 -08003367 set_huge_ptep_writable(vma, address, ptep);
Nick Piggin83c54072007-07-19 01:47:05 -07003368 return 0;
David Gibson1e8f8892006-01-06 00:10:44 -08003369 }
3370
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003371 /*
3372 * If the process that created a MAP_PRIVATE mapping is about to
3373 * perform a COW due to a shared page count, attempt to satisfy
3374 * the allocation without using the existing reserves. The pagecache
3375 * page is used to determine if the reserve at this address was
3376 * consumed or not. If reserves were used, a partial faulted mapping
3377 * at the time of fork() could consume its reserves on COW instead
3378 * of the full address range.
3379 */
Joonsoo Kim5944d012013-09-11 14:21:55 -07003380 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003381 old_page != pagecache_page)
3382 outside_reserve = 1;
3383
David Gibson1e8f8892006-01-06 00:10:44 -08003384 page_cache_get(old_page);
Larry Woodmanb76c8cf2009-12-14 17:59:37 -08003385
Davidlohr Buesoad4404a2014-08-06 16:06:47 -07003386 /*
3387 * Drop page table lock as buddy allocator may be called. It will
3388 * be acquired again before returning to the caller, as expected.
3389 */
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003390 spin_unlock(ptl);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003391 new_page = alloc_huge_page(vma, address, outside_reserve);
David Gibson1e8f8892006-01-06 00:10:44 -08003392
Adam Litke2fc39ce2007-11-14 16:59:39 -08003393 if (IS_ERR(new_page)) {
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003394 /*
3395 * If a process owning a MAP_PRIVATE mapping fails to COW,
3396 * it is due to references held by a child and an insufficient
3397 * huge page pool. To guarantee the original mappers
3398 * reliability, unmap the page from child processes. The child
3399 * may get SIGKILLed if it later faults.
3400 */
3401 if (outside_reserve) {
Davidlohr Buesoad4404a2014-08-06 16:06:47 -07003402 page_cache_release(old_page);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003403 BUG_ON(huge_pte_none(pte));
Davidlohr Bueso2f4612a2014-08-06 16:06:45 -07003404 unmap_ref_private(mm, vma, old_page, address);
3405 BUG_ON(huge_pte_none(pte));
3406 spin_lock(ptl);
3407 ptep = huge_pte_offset(mm, address & huge_page_mask(h));
3408 if (likely(ptep &&
3409 pte_same(huge_ptep_get(ptep), pte)))
3410 goto retry_avoidcopy;
3411 /*
3412 * race occurs while re-acquiring page table
3413 * lock, and our job is done.
3414 */
3415 return 0;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003416 }
3417
Davidlohr Buesoad4404a2014-08-06 16:06:47 -07003418 ret = (PTR_ERR(new_page) == -ENOMEM) ?
3419 VM_FAULT_OOM : VM_FAULT_SIGBUS;
3420 goto out_release_old;
David Gibson1e8f8892006-01-06 00:10:44 -08003421 }
3422
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09003423 /*
3424 * When the original hugepage is shared one, it does not have
3425 * anon_vma prepared.
3426 */
Dean Nelson44e2aa92010-10-26 14:22:08 -07003427 if (unlikely(anon_vma_prepare(vma))) {
Davidlohr Buesoad4404a2014-08-06 16:06:47 -07003428 ret = VM_FAULT_OOM;
3429 goto out_release_all;
Dean Nelson44e2aa92010-10-26 14:22:08 -07003430 }
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09003431
Andrea Arcangeli47ad8472011-01-13 15:46:47 -08003432 copy_user_huge_page(new_page, old_page, address, vma,
3433 pages_per_huge_page(h));
Nick Piggin0ed361d2008-02-04 22:29:34 -08003434 __SetPageUptodate(new_page);
Naoya Horiguchibcc54222015-04-15 16:14:38 -07003435 set_page_huge_active(new_page);
David Gibson1e8f8892006-01-06 00:10:44 -08003436
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07003437 mmun_start = address & huge_page_mask(h);
3438 mmun_end = mmun_start + huge_page_size(h);
3439 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
Davidlohr Buesoad4404a2014-08-06 16:06:47 -07003440
Larry Woodmanb76c8cf2009-12-14 17:59:37 -08003441 /*
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003442 * Retake the page table lock to check for racing updates
Larry Woodmanb76c8cf2009-12-14 17:59:37 -08003443 * before the page tables are altered
3444 */
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003445 spin_lock(ptl);
Andi Kleena5516432008-07-23 21:27:41 -07003446 ptep = huge_pte_offset(mm, address & huge_page_mask(h));
Naoya Horiguchia9af0c52014-04-07 15:36:54 -07003447 if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) {
Joonsoo Kim07443a82013-09-11 14:21:58 -07003448 ClearPagePrivate(new_page);
3449
David Gibson1e8f8892006-01-06 00:10:44 -08003450 /* Break COW */
Gerald Schaefer8fe627e2008-04-28 02:13:28 -07003451 huge_ptep_clear_flush(vma, address, ptep);
Joerg Roedel34ee6452014-11-13 13:46:09 +11003452 mmu_notifier_invalidate_range(mm, mmun_start, mmun_end);
David Gibson1e8f8892006-01-06 00:10:44 -08003453 set_huge_pte_at(mm, address, ptep,
3454 make_huge_pte(vma, new_page, 1));
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09003455 page_remove_rmap(old_page);
Naoya Horiguchicd67f0d2010-09-10 13:23:04 +09003456 hugepage_add_new_anon_rmap(new_page, vma, address);
David Gibson1e8f8892006-01-06 00:10:44 -08003457 /* Make the old page be freed below */
3458 new_page = old_page;
3459 }
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003460 spin_unlock(ptl);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07003461 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
Davidlohr Buesoad4404a2014-08-06 16:06:47 -07003462out_release_all:
David Gibson1e8f8892006-01-06 00:10:44 -08003463 page_cache_release(new_page);
Davidlohr Buesoad4404a2014-08-06 16:06:47 -07003464out_release_old:
David Gibson1e8f8892006-01-06 00:10:44 -08003465 page_cache_release(old_page);
Joonsoo Kim83120342013-09-11 14:21:57 -07003466
Davidlohr Buesoad4404a2014-08-06 16:06:47 -07003467 spin_lock(ptl); /* Caller expects lock to be held */
3468 return ret;
David Gibson1e8f8892006-01-06 00:10:44 -08003469}
3470
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003471/* Return the pagecache page at a given address within a VMA */
Andi Kleena5516432008-07-23 21:27:41 -07003472static struct page *hugetlbfs_pagecache_page(struct hstate *h,
3473 struct vm_area_struct *vma, unsigned long address)
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003474{
3475 struct address_space *mapping;
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -07003476 pgoff_t idx;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003477
3478 mapping = vma->vm_file->f_mapping;
Andi Kleena5516432008-07-23 21:27:41 -07003479 idx = vma_hugecache_offset(h, vma, address);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003480
3481 return find_lock_page(mapping, idx);
3482}
3483
Hugh Dickins3ae77f42009-09-21 17:03:33 -07003484/*
3485 * Return whether there is a pagecache page to back given address within VMA.
3486 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
3487 */
3488static bool hugetlbfs_pagecache_present(struct hstate *h,
Hugh Dickins2a15efc2009-09-21 17:03:27 -07003489 struct vm_area_struct *vma, unsigned long address)
3490{
3491 struct address_space *mapping;
3492 pgoff_t idx;
3493 struct page *page;
3494
3495 mapping = vma->vm_file->f_mapping;
3496 idx = vma_hugecache_offset(h, vma, address);
3497
3498 page = find_get_page(mapping, idx);
3499 if (page)
3500 put_page(page);
3501 return page != NULL;
3502}
3503
Mike Kravetzab76ad52015-09-08 15:01:50 -07003504int huge_add_to_page_cache(struct page *page, struct address_space *mapping,
3505 pgoff_t idx)
3506{
3507 struct inode *inode = mapping->host;
3508 struct hstate *h = hstate_inode(inode);
3509 int err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
3510
3511 if (err)
3512 return err;
3513 ClearPagePrivate(page);
3514
3515 spin_lock(&inode->i_lock);
3516 inode->i_blocks += blocks_per_huge_page(h);
3517 spin_unlock(&inode->i_lock);
3518 return 0;
3519}
3520
Robert P. J. Daya1ed3dd2007-07-17 04:03:33 -07003521static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
Davidlohr Bueso8382d912014-04-03 14:47:31 -07003522 struct address_space *mapping, pgoff_t idx,
3523 unsigned long address, pte_t *ptep, unsigned int flags)
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01003524{
Andi Kleena5516432008-07-23 21:27:41 -07003525 struct hstate *h = hstate_vma(vma);
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01003526 int ret = VM_FAULT_SIGBUS;
Hillf Danton409eb8c2012-01-20 14:34:13 -08003527 int anon_rmap = 0;
Adam Litke4c887262005-10-29 18:16:46 -07003528 unsigned long size;
Adam Litke4c887262005-10-29 18:16:46 -07003529 struct page *page;
David Gibson1e8f8892006-01-06 00:10:44 -08003530 pte_t new_pte;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003531 spinlock_t *ptl;
Adam Litke4c887262005-10-29 18:16:46 -07003532
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003533 /*
3534 * Currently, we are forced to kill the process in the event the
3535 * original mapper has unmapped pages from the child due to a failed
Lucas De Marchi25985ed2011-03-30 22:57:33 -03003536 * COW. Warn that such a situation has occurred as it may not be obvious
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003537 */
3538 if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
Andrew Mortonffb22af2013-02-22 16:32:08 -08003539 pr_warning("PID %d killed due to inadequate hugepage pool\n",
3540 current->pid);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003541 return ret;
3542 }
3543
Adam Litke4c887262005-10-29 18:16:46 -07003544 /*
3545 * Use page lock to guard against racing truncation
3546 * before we get page_table_lock.
3547 */
Christoph Lameter6bda6662006-01-06 00:10:49 -08003548retry:
3549 page = find_lock_page(mapping, idx);
3550 if (!page) {
Andi Kleena5516432008-07-23 21:27:41 -07003551 size = i_size_read(mapping->host) >> huge_page_shift(h);
Hugh Dickinsebed4bf2006-10-28 10:38:43 -07003552 if (idx >= size)
3553 goto out;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003554 page = alloc_huge_page(vma, address, 0);
Adam Litke2fc39ce2007-11-14 16:59:39 -08003555 if (IS_ERR(page)) {
Aneesh Kumar K.V76dcee72012-07-31 16:41:57 -07003556 ret = PTR_ERR(page);
3557 if (ret == -ENOMEM)
3558 ret = VM_FAULT_OOM;
3559 else
3560 ret = VM_FAULT_SIGBUS;
Christoph Lameter6bda6662006-01-06 00:10:49 -08003561 goto out;
3562 }
Andrea Arcangeli47ad8472011-01-13 15:46:47 -08003563 clear_huge_page(page, address, pages_per_huge_page(h));
Nick Piggin0ed361d2008-02-04 22:29:34 -08003564 __SetPageUptodate(page);
Naoya Horiguchibcc54222015-04-15 16:14:38 -07003565 set_page_huge_active(page);
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01003566
Mel Gormanf83a2752009-05-28 14:34:40 -07003567 if (vma->vm_flags & VM_MAYSHARE) {
Mike Kravetzab76ad52015-09-08 15:01:50 -07003568 int err = huge_add_to_page_cache(page, mapping, idx);
Christoph Lameter6bda6662006-01-06 00:10:49 -08003569 if (err) {
3570 put_page(page);
Christoph Lameter6bda6662006-01-06 00:10:49 -08003571 if (err == -EEXIST)
3572 goto retry;
3573 goto out;
3574 }
Mel Gorman23be7462010-04-23 13:17:56 -04003575 } else {
Christoph Lameter6bda6662006-01-06 00:10:49 -08003576 lock_page(page);
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09003577 if (unlikely(anon_vma_prepare(vma))) {
3578 ret = VM_FAULT_OOM;
3579 goto backout_unlocked;
3580 }
Hillf Danton409eb8c2012-01-20 14:34:13 -08003581 anon_rmap = 1;
Mel Gorman23be7462010-04-23 13:17:56 -04003582 }
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09003583 } else {
Naoya Horiguchi998b4382010-09-08 10:19:32 +09003584 /*
3585 * If memory error occurs between mmap() and fault, some process
3586 * don't have hwpoisoned swap entry for errored virtual address.
3587 * So we need to block hugepage fault by PG_hwpoison bit check.
3588 */
3589 if (unlikely(PageHWPoison(page))) {
Chris Forbes32f84522011-07-25 17:12:14 -07003590 ret = VM_FAULT_HWPOISON |
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07003591 VM_FAULT_SET_HINDEX(hstate_index(h));
Naoya Horiguchi998b4382010-09-08 10:19:32 +09003592 goto backout_unlocked;
3593 }
Christoph Lameter6bda6662006-01-06 00:10:49 -08003594 }
David Gibson1e8f8892006-01-06 00:10:44 -08003595
Andy Whitcroft57303d82008-08-12 15:08:47 -07003596 /*
3597 * If we are going to COW a private mapping later, we examine the
3598 * pending reservations for this page now. This will ensure that
3599 * any allocations necessary to record that reservation occur outside
3600 * the spinlock.
3601 */
Mike Kravetz5e911372015-09-08 15:01:28 -07003602 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
Andy Whitcroft2b267362008-08-12 15:08:49 -07003603 if (vma_needs_reservation(h, vma, address) < 0) {
3604 ret = VM_FAULT_OOM;
3605 goto backout_unlocked;
3606 }
Mike Kravetz5e911372015-09-08 15:01:28 -07003607 /* Just decrements count, does not deallocate */
Mike Kravetzfeba16e2015-09-08 15:01:31 -07003608 vma_end_reservation(h, vma, address);
Mike Kravetz5e911372015-09-08 15:01:28 -07003609 }
Andy Whitcroft57303d82008-08-12 15:08:47 -07003610
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003611 ptl = huge_pte_lockptr(h, mm, ptep);
3612 spin_lock(ptl);
Andi Kleena5516432008-07-23 21:27:41 -07003613 size = i_size_read(mapping->host) >> huge_page_shift(h);
Adam Litke4c887262005-10-29 18:16:46 -07003614 if (idx >= size)
3615 goto backout;
3616
Nick Piggin83c54072007-07-19 01:47:05 -07003617 ret = 0;
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07003618 if (!huge_pte_none(huge_ptep_get(ptep)))
Adam Litke4c887262005-10-29 18:16:46 -07003619 goto backout;
3620
Joonsoo Kim07443a82013-09-11 14:21:58 -07003621 if (anon_rmap) {
3622 ClearPagePrivate(page);
Hillf Danton409eb8c2012-01-20 14:34:13 -08003623 hugepage_add_new_anon_rmap(page, vma, address);
Choi Gi-yongac714902014-04-07 15:37:36 -07003624 } else
Hillf Danton409eb8c2012-01-20 14:34:13 -08003625 page_dup_rmap(page);
David Gibson1e8f8892006-01-06 00:10:44 -08003626 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
3627 && (vma->vm_flags & VM_SHARED)));
3628 set_huge_pte_at(mm, address, ptep, new_pte);
3629
Naoya Horiguchi5d317b22015-11-05 18:47:14 -08003630 hugetlb_count_add(pages_per_huge_page(h), mm);
Hugh Dickins788c7df2009-06-23 13:49:05 +01003631 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
David Gibson1e8f8892006-01-06 00:10:44 -08003632 /* Optimization, do the COW without a second fault */
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003633 ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page, ptl);
David Gibson1e8f8892006-01-06 00:10:44 -08003634 }
3635
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003636 spin_unlock(ptl);
Adam Litke4c887262005-10-29 18:16:46 -07003637 unlock_page(page);
3638out:
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01003639 return ret;
Adam Litke4c887262005-10-29 18:16:46 -07003640
3641backout:
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003642 spin_unlock(ptl);
Andy Whitcroft2b267362008-08-12 15:08:49 -07003643backout_unlocked:
Adam Litke4c887262005-10-29 18:16:46 -07003644 unlock_page(page);
3645 put_page(page);
3646 goto out;
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01003647}
3648
Davidlohr Bueso8382d912014-04-03 14:47:31 -07003649#ifdef CONFIG_SMP
Mike Kravetzc672c7f2015-09-08 15:01:35 -07003650u32 hugetlb_fault_mutex_hash(struct hstate *h, struct mm_struct *mm,
Davidlohr Bueso8382d912014-04-03 14:47:31 -07003651 struct vm_area_struct *vma,
3652 struct address_space *mapping,
3653 pgoff_t idx, unsigned long address)
3654{
3655 unsigned long key[2];
3656 u32 hash;
3657
3658 if (vma->vm_flags & VM_SHARED) {
3659 key[0] = (unsigned long) mapping;
3660 key[1] = idx;
3661 } else {
3662 key[0] = (unsigned long) mm;
3663 key[1] = address >> huge_page_shift(h);
3664 }
3665
3666 hash = jhash2((u32 *)&key, sizeof(key)/sizeof(u32), 0);
3667
3668 return hash & (num_fault_mutexes - 1);
3669}
3670#else
3671/*
3672 * For uniprocesor systems we always use a single mutex, so just
3673 * return 0 and avoid the hashing overhead.
3674 */
Mike Kravetzc672c7f2015-09-08 15:01:35 -07003675u32 hugetlb_fault_mutex_hash(struct hstate *h, struct mm_struct *mm,
Davidlohr Bueso8382d912014-04-03 14:47:31 -07003676 struct vm_area_struct *vma,
3677 struct address_space *mapping,
3678 pgoff_t idx, unsigned long address)
3679{
3680 return 0;
3681}
3682#endif
3683
Adam Litke86e52162006-01-06 00:10:43 -08003684int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
Hugh Dickins788c7df2009-06-23 13:49:05 +01003685 unsigned long address, unsigned int flags)
Adam Litke86e52162006-01-06 00:10:43 -08003686{
Davidlohr Bueso8382d912014-04-03 14:47:31 -07003687 pte_t *ptep, entry;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003688 spinlock_t *ptl;
David Gibson1e8f8892006-01-06 00:10:44 -08003689 int ret;
Davidlohr Bueso8382d912014-04-03 14:47:31 -07003690 u32 hash;
3691 pgoff_t idx;
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09003692 struct page *page = NULL;
Andy Whitcroft57303d82008-08-12 15:08:47 -07003693 struct page *pagecache_page = NULL;
Andi Kleena5516432008-07-23 21:27:41 -07003694 struct hstate *h = hstate_vma(vma);
Davidlohr Bueso8382d912014-04-03 14:47:31 -07003695 struct address_space *mapping;
Naoya Horiguchi0f792cf2015-02-11 15:25:25 -08003696 int need_wait_lock = 0;
Adam Litke86e52162006-01-06 00:10:43 -08003697
KAMEZAWA Hiroyuki1e16a532012-01-10 15:07:22 -08003698 address &= huge_page_mask(h);
3699
Naoya Horiguchifd6a03ed2010-05-28 09:29:21 +09003700 ptep = huge_pte_offset(mm, address);
3701 if (ptep) {
3702 entry = huge_ptep_get(ptep);
Naoya Horiguchi290408d2010-09-08 10:19:35 +09003703 if (unlikely(is_hugetlb_entry_migration(entry))) {
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003704 migration_entry_wait_huge(vma, mm, ptep);
Naoya Horiguchi290408d2010-09-08 10:19:35 +09003705 return 0;
3706 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
Chris Forbes32f84522011-07-25 17:12:14 -07003707 return VM_FAULT_HWPOISON_LARGE |
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07003708 VM_FAULT_SET_HINDEX(hstate_index(h));
Naoya Horiguchi0d777df2015-12-11 13:40:49 -08003709 } else {
3710 ptep = huge_pte_alloc(mm, address, huge_page_size(h));
3711 if (!ptep)
3712 return VM_FAULT_OOM;
Naoya Horiguchifd6a03ed2010-05-28 09:29:21 +09003713 }
3714
Davidlohr Bueso8382d912014-04-03 14:47:31 -07003715 mapping = vma->vm_file->f_mapping;
3716 idx = vma_hugecache_offset(h, vma, address);
3717
David Gibson3935baa2006-03-22 00:08:53 -08003718 /*
3719 * Serialize hugepage allocation and instantiation, so that we don't
3720 * get spurious allocation failures if two CPUs race to instantiate
3721 * the same page in the page cache.
3722 */
Mike Kravetzc672c7f2015-09-08 15:01:35 -07003723 hash = hugetlb_fault_mutex_hash(h, mm, vma, mapping, idx, address);
3724 mutex_lock(&hugetlb_fault_mutex_table[hash]);
Davidlohr Bueso8382d912014-04-03 14:47:31 -07003725
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07003726 entry = huge_ptep_get(ptep);
3727 if (huge_pte_none(entry)) {
Davidlohr Bueso8382d912014-04-03 14:47:31 -07003728 ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags);
David Gibsonb4d1d992008-10-15 22:01:11 -07003729 goto out_mutex;
David Gibson3935baa2006-03-22 00:08:53 -08003730 }
Adam Litke86e52162006-01-06 00:10:43 -08003731
Nick Piggin83c54072007-07-19 01:47:05 -07003732 ret = 0;
David Gibson1e8f8892006-01-06 00:10:44 -08003733
Andy Whitcroft57303d82008-08-12 15:08:47 -07003734 /*
Naoya Horiguchi0f792cf2015-02-11 15:25:25 -08003735 * entry could be a migration/hwpoison entry at this point, so this
3736 * check prevents the kernel from going below assuming that we have
3737 * a active hugepage in pagecache. This goto expects the 2nd page fault,
3738 * and is_hugetlb_entry_(migration|hwpoisoned) check will properly
3739 * handle it.
3740 */
3741 if (!pte_present(entry))
3742 goto out_mutex;
3743
3744 /*
Andy Whitcroft57303d82008-08-12 15:08:47 -07003745 * If we are going to COW the mapping later, we examine the pending
3746 * reservations for this page now. This will ensure that any
3747 * allocations necessary to record that reservation occur outside the
3748 * spinlock. For private mappings, we also lookup the pagecache
3749 * page now as it is used to determine if a reservation has been
3750 * consumed.
3751 */
Gerald Schaefer106c9922013-04-29 15:07:23 -07003752 if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) {
Andy Whitcroft2b267362008-08-12 15:08:49 -07003753 if (vma_needs_reservation(h, vma, address) < 0) {
3754 ret = VM_FAULT_OOM;
David Gibsonb4d1d992008-10-15 22:01:11 -07003755 goto out_mutex;
Andy Whitcroft2b267362008-08-12 15:08:49 -07003756 }
Mike Kravetz5e911372015-09-08 15:01:28 -07003757 /* Just decrements count, does not deallocate */
Mike Kravetzfeba16e2015-09-08 15:01:31 -07003758 vma_end_reservation(h, vma, address);
Andy Whitcroft57303d82008-08-12 15:08:47 -07003759
Mel Gormanf83a2752009-05-28 14:34:40 -07003760 if (!(vma->vm_flags & VM_MAYSHARE))
Andy Whitcroft57303d82008-08-12 15:08:47 -07003761 pagecache_page = hugetlbfs_pagecache_page(h,
3762 vma, address);
3763 }
3764
Naoya Horiguchi0f792cf2015-02-11 15:25:25 -08003765 ptl = huge_pte_lock(h, mm, ptep);
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09003766
David Gibson1e8f8892006-01-06 00:10:44 -08003767 /* Check for a racing update before calling hugetlb_cow */
David Gibsonb4d1d992008-10-15 22:01:11 -07003768 if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003769 goto out_ptl;
David Gibsonb4d1d992008-10-15 22:01:11 -07003770
Naoya Horiguchi0f792cf2015-02-11 15:25:25 -08003771 /*
3772 * hugetlb_cow() requires page locks of pte_page(entry) and
3773 * pagecache_page, so here we need take the former one
3774 * when page != pagecache_page or !pagecache_page.
3775 */
3776 page = pte_page(entry);
3777 if (page != pagecache_page)
3778 if (!trylock_page(page)) {
3779 need_wait_lock = 1;
3780 goto out_ptl;
3781 }
3782
3783 get_page(page);
David Gibsonb4d1d992008-10-15 22:01:11 -07003784
Hugh Dickins788c7df2009-06-23 13:49:05 +01003785 if (flags & FAULT_FLAG_WRITE) {
Gerald Schaefer106c9922013-04-29 15:07:23 -07003786 if (!huge_pte_write(entry)) {
Andy Whitcroft57303d82008-08-12 15:08:47 -07003787 ret = hugetlb_cow(mm, vma, address, ptep, entry,
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003788 pagecache_page, ptl);
Naoya Horiguchi0f792cf2015-02-11 15:25:25 -08003789 goto out_put_page;
David Gibsonb4d1d992008-10-15 22:01:11 -07003790 }
Gerald Schaefer106c9922013-04-29 15:07:23 -07003791 entry = huge_pte_mkdirty(entry);
David Gibsonb4d1d992008-10-15 22:01:11 -07003792 }
3793 entry = pte_mkyoung(entry);
Hugh Dickins788c7df2009-06-23 13:49:05 +01003794 if (huge_ptep_set_access_flags(vma, address, ptep, entry,
3795 flags & FAULT_FLAG_WRITE))
Russell King4b3073e2009-12-18 16:40:18 +00003796 update_mmu_cache(vma, address, ptep);
Naoya Horiguchi0f792cf2015-02-11 15:25:25 -08003797out_put_page:
3798 if (page != pagecache_page)
3799 unlock_page(page);
3800 put_page(page);
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003801out_ptl:
3802 spin_unlock(ptl);
Andy Whitcroft57303d82008-08-12 15:08:47 -07003803
3804 if (pagecache_page) {
3805 unlock_page(pagecache_page);
3806 put_page(pagecache_page);
3807 }
David Gibsonb4d1d992008-10-15 22:01:11 -07003808out_mutex:
Mike Kravetzc672c7f2015-09-08 15:01:35 -07003809 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
Naoya Horiguchi0f792cf2015-02-11 15:25:25 -08003810 /*
3811 * Generally it's safe to hold refcount during waiting page lock. But
3812 * here we just wait to defer the next page fault to avoid busy loop and
3813 * the page is not used after unlocked before returning from the current
3814 * page fault. So we are safe from accessing freed page, even if we wait
3815 * here without taking refcount.
3816 */
3817 if (need_wait_lock)
3818 wait_on_page_locked(page);
David Gibson1e8f8892006-01-06 00:10:44 -08003819 return ret;
Adam Litke86e52162006-01-06 00:10:43 -08003820}
3821
Michel Lespinasse28a35712013-02-22 16:35:55 -08003822long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
3823 struct page **pages, struct vm_area_struct **vmas,
3824 unsigned long *position, unsigned long *nr_pages,
3825 long i, unsigned int flags)
David Gibson63551ae2005-06-21 17:14:44 -07003826{
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08003827 unsigned long pfn_offset;
3828 unsigned long vaddr = *position;
Michel Lespinasse28a35712013-02-22 16:35:55 -08003829 unsigned long remainder = *nr_pages;
Andi Kleena5516432008-07-23 21:27:41 -07003830 struct hstate *h = hstate_vma(vma);
David Gibson63551ae2005-06-21 17:14:44 -07003831
David Gibson63551ae2005-06-21 17:14:44 -07003832 while (vaddr < vma->vm_end && remainder) {
Adam Litke4c887262005-10-29 18:16:46 -07003833 pte_t *pte;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003834 spinlock_t *ptl = NULL;
Hugh Dickins2a15efc2009-09-21 17:03:27 -07003835 int absent;
Adam Litke4c887262005-10-29 18:16:46 -07003836 struct page *page;
3837
3838 /*
David Rientjes02057962015-04-14 15:48:24 -07003839 * If we have a pending SIGKILL, don't keep faulting pages and
3840 * potentially allocating memory.
3841 */
3842 if (unlikely(fatal_signal_pending(current))) {
3843 remainder = 0;
3844 break;
3845 }
3846
3847 /*
Adam Litke4c887262005-10-29 18:16:46 -07003848 * Some archs (sparc64, sh*) have multiple pte_ts to
Hugh Dickins2a15efc2009-09-21 17:03:27 -07003849 * each hugepage. We have to make sure we get the
Adam Litke4c887262005-10-29 18:16:46 -07003850 * first, for the page indexing below to work.
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003851 *
3852 * Note that page table lock is not held when pte is null.
Adam Litke4c887262005-10-29 18:16:46 -07003853 */
Andi Kleena5516432008-07-23 21:27:41 -07003854 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h));
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003855 if (pte)
3856 ptl = huge_pte_lock(h, mm, pte);
Hugh Dickins2a15efc2009-09-21 17:03:27 -07003857 absent = !pte || huge_pte_none(huge_ptep_get(pte));
Adam Litke4c887262005-10-29 18:16:46 -07003858
Hugh Dickins2a15efc2009-09-21 17:03:27 -07003859 /*
3860 * When coredumping, it suits get_dump_page if we just return
Hugh Dickins3ae77f42009-09-21 17:03:33 -07003861 * an error where there's an empty slot with no huge pagecache
3862 * to back it. This way, we avoid allocating a hugepage, and
3863 * the sparse dumpfile avoids allocating disk blocks, but its
3864 * huge holes still show up with zeroes where they need to be.
Hugh Dickins2a15efc2009-09-21 17:03:27 -07003865 */
Hugh Dickins3ae77f42009-09-21 17:03:33 -07003866 if (absent && (flags & FOLL_DUMP) &&
3867 !hugetlbfs_pagecache_present(h, vma, vaddr)) {
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003868 if (pte)
3869 spin_unlock(ptl);
Hugh Dickins2a15efc2009-09-21 17:03:27 -07003870 remainder = 0;
3871 break;
3872 }
3873
Naoya Horiguchi9cc3a5b2013-04-17 15:58:30 -07003874 /*
3875 * We need call hugetlb_fault for both hugepages under migration
3876 * (in which case hugetlb_fault waits for the migration,) and
3877 * hwpoisoned hugepages (in which case we need to prevent the
3878 * caller from accessing to them.) In order to do this, we use
3879 * here is_swap_pte instead of is_hugetlb_entry_migration and
3880 * is_hugetlb_entry_hwpoisoned. This is because it simply covers
3881 * both cases, and because we can't follow correct pages
3882 * directly from any kind of swap entries.
3883 */
3884 if (absent || is_swap_pte(huge_ptep_get(pte)) ||
Gerald Schaefer106c9922013-04-29 15:07:23 -07003885 ((flags & FOLL_WRITE) &&
3886 !huge_pte_write(huge_ptep_get(pte)))) {
Adam Litke4c887262005-10-29 18:16:46 -07003887 int ret;
3888
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003889 if (pte)
3890 spin_unlock(ptl);
Hugh Dickins2a15efc2009-09-21 17:03:27 -07003891 ret = hugetlb_fault(mm, vma, vaddr,
3892 (flags & FOLL_WRITE) ? FAULT_FLAG_WRITE : 0);
Adam Litkea89182c2007-08-22 14:01:51 -07003893 if (!(ret & VM_FAULT_ERROR))
Adam Litke4c887262005-10-29 18:16:46 -07003894 continue;
3895
3896 remainder = 0;
Adam Litke4c887262005-10-29 18:16:46 -07003897 break;
3898 }
David Gibson63551ae2005-06-21 17:14:44 -07003899
Andi Kleena5516432008-07-23 21:27:41 -07003900 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07003901 page = pte_page(huge_ptep_get(pte));
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08003902same_page:
Chen, Kenneth Wd6692182006-03-31 02:29:57 -08003903 if (pages) {
Hugh Dickins2a15efc2009-09-21 17:03:27 -07003904 pages[i] = mem_map_offset(page, pfn_offset);
Andrea Arcangelia0368d42014-01-21 15:48:49 -08003905 get_page_foll(pages[i]);
Chen, Kenneth Wd6692182006-03-31 02:29:57 -08003906 }
David Gibson63551ae2005-06-21 17:14:44 -07003907
3908 if (vmas)
3909 vmas[i] = vma;
3910
3911 vaddr += PAGE_SIZE;
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08003912 ++pfn_offset;
David Gibson63551ae2005-06-21 17:14:44 -07003913 --remainder;
3914 ++i;
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08003915 if (vaddr < vma->vm_end && remainder &&
Andi Kleena5516432008-07-23 21:27:41 -07003916 pfn_offset < pages_per_huge_page(h)) {
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08003917 /*
3918 * We use pfn_offset to avoid touching the pageframes
3919 * of this compound page.
3920 */
3921 goto same_page;
3922 }
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003923 spin_unlock(ptl);
David Gibson63551ae2005-06-21 17:14:44 -07003924 }
Michel Lespinasse28a35712013-02-22 16:35:55 -08003925 *nr_pages = remainder;
David Gibson63551ae2005-06-21 17:14:44 -07003926 *position = vaddr;
3927
Hugh Dickins2a15efc2009-09-21 17:03:27 -07003928 return i ? i : -EFAULT;
David Gibson63551ae2005-06-21 17:14:44 -07003929}
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003930
Peter Zijlstra7da4d642012-11-19 03:14:23 +01003931unsigned long hugetlb_change_protection(struct vm_area_struct *vma,
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003932 unsigned long address, unsigned long end, pgprot_t newprot)
3933{
3934 struct mm_struct *mm = vma->vm_mm;
3935 unsigned long start = address;
3936 pte_t *ptep;
3937 pte_t pte;
Andi Kleena5516432008-07-23 21:27:41 -07003938 struct hstate *h = hstate_vma(vma);
Peter Zijlstra7da4d642012-11-19 03:14:23 +01003939 unsigned long pages = 0;
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003940
3941 BUG_ON(address >= end);
3942 flush_cache_range(vma, address, end);
3943
Rik van Riela5338092014-04-07 15:36:57 -07003944 mmu_notifier_invalidate_range_start(mm, start, end);
Davidlohr Bueso83cde9e2014-12-12 16:54:21 -08003945 i_mmap_lock_write(vma->vm_file->f_mapping);
Andi Kleena5516432008-07-23 21:27:41 -07003946 for (; address < end; address += huge_page_size(h)) {
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003947 spinlock_t *ptl;
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003948 ptep = huge_pte_offset(mm, address);
3949 if (!ptep)
3950 continue;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003951 ptl = huge_pte_lock(h, mm, ptep);
Peter Zijlstra7da4d642012-11-19 03:14:23 +01003952 if (huge_pmd_unshare(mm, &address, ptep)) {
3953 pages++;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003954 spin_unlock(ptl);
Chen, Kenneth W39dde652006-12-06 20:32:03 -08003955 continue;
Peter Zijlstra7da4d642012-11-19 03:14:23 +01003956 }
Naoya Horiguchia8bda282015-02-11 15:25:28 -08003957 pte = huge_ptep_get(ptep);
3958 if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) {
3959 spin_unlock(ptl);
3960 continue;
3961 }
3962 if (unlikely(is_hugetlb_entry_migration(pte))) {
3963 swp_entry_t entry = pte_to_swp_entry(pte);
3964
3965 if (is_write_migration_entry(entry)) {
3966 pte_t newpte;
3967
3968 make_migration_entry_read(&entry);
3969 newpte = swp_entry_to_pte(entry);
3970 set_huge_pte_at(mm, address, ptep, newpte);
3971 pages++;
3972 }
3973 spin_unlock(ptl);
3974 continue;
3975 }
3976 if (!huge_pte_none(pte)) {
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003977 pte = huge_ptep_get_and_clear(mm, address, ptep);
Gerald Schaefer106c9922013-04-29 15:07:23 -07003978 pte = pte_mkhuge(huge_pte_modify(pte, newprot));
Tony Lube7517d2013-02-04 14:28:46 -08003979 pte = arch_make_huge_pte(pte, vma, NULL, 0);
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003980 set_huge_pte_at(mm, address, ptep, pte);
Peter Zijlstra7da4d642012-11-19 03:14:23 +01003981 pages++;
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003982 }
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003983 spin_unlock(ptl);
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003984 }
Mel Gormand8333522012-07-31 16:46:20 -07003985 /*
Davidlohr Buesoc8c06ef2014-12-12 16:54:24 -08003986 * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare
Mel Gormand8333522012-07-31 16:46:20 -07003987 * may have cleared our pud entry and done put_page on the page table:
Davidlohr Buesoc8c06ef2014-12-12 16:54:24 -08003988 * once we release i_mmap_rwsem, another task can do the final put_page
Mel Gormand8333522012-07-31 16:46:20 -07003989 * and that page table be reused and filled with junk.
3990 */
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003991 flush_tlb_range(vma, start, end);
Joerg Roedel34ee6452014-11-13 13:46:09 +11003992 mmu_notifier_invalidate_range(mm, start, end);
Davidlohr Bueso83cde9e2014-12-12 16:54:21 -08003993 i_mmap_unlock_write(vma->vm_file->f_mapping);
Rik van Riela5338092014-04-07 15:36:57 -07003994 mmu_notifier_invalidate_range_end(mm, start, end);
Peter Zijlstra7da4d642012-11-19 03:14:23 +01003995
3996 return pages << h->order;
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003997}
3998
Mel Gormana1e78772008-07-23 21:27:23 -07003999int hugetlb_reserve_pages(struct inode *inode,
4000 long from, long to,
Mel Gorman5a6fe122009-02-10 14:02:27 +00004001 struct vm_area_struct *vma,
KOSAKI Motohiroca16d142011-05-26 19:16:19 +09004002 vm_flags_t vm_flags)
Adam Litkee4e574b2007-10-16 01:26:19 -07004003{
Mel Gorman17c9d122009-02-11 16:34:16 +00004004 long ret, chg;
Andi Kleena5516432008-07-23 21:27:41 -07004005 struct hstate *h = hstate_inode(inode);
David Gibson90481622012-03-21 16:34:12 -07004006 struct hugepage_subpool *spool = subpool_inode(inode);
Joonsoo Kim9119a412014-04-03 14:47:25 -07004007 struct resv_map *resv_map;
Mike Kravetz1c5ecae2015-04-15 16:13:39 -07004008 long gbl_reserve;
Adam Litkee4e574b2007-10-16 01:26:19 -07004009
Mel Gormana1e78772008-07-23 21:27:23 -07004010 /*
Mel Gorman17c9d122009-02-11 16:34:16 +00004011 * Only apply hugepage reservation if asked. At fault time, an
4012 * attempt will be made for VM_NORESERVE to allocate a page
David Gibson90481622012-03-21 16:34:12 -07004013 * without using reserves
Mel Gorman17c9d122009-02-11 16:34:16 +00004014 */
KOSAKI Motohiroca16d142011-05-26 19:16:19 +09004015 if (vm_flags & VM_NORESERVE)
Mel Gorman17c9d122009-02-11 16:34:16 +00004016 return 0;
4017
4018 /*
Mel Gormana1e78772008-07-23 21:27:23 -07004019 * Shared mappings base their reservation on the number of pages that
4020 * are already allocated on behalf of the file. Private mappings need
4021 * to reserve the full area even if read-only as mprotect() may be
4022 * called to make the mapping read-write. Assume !vma is a shm mapping
4023 */
Joonsoo Kim9119a412014-04-03 14:47:25 -07004024 if (!vma || vma->vm_flags & VM_MAYSHARE) {
Joonsoo Kim4e35f482014-04-03 14:47:30 -07004025 resv_map = inode_resv_map(inode);
Joonsoo Kim9119a412014-04-03 14:47:25 -07004026
Joonsoo Kim1406ec92014-04-03 14:47:26 -07004027 chg = region_chg(resv_map, from, to);
Joonsoo Kim9119a412014-04-03 14:47:25 -07004028
4029 } else {
4030 resv_map = resv_map_alloc();
Mel Gorman5a6fe122009-02-10 14:02:27 +00004031 if (!resv_map)
4032 return -ENOMEM;
4033
Mel Gorman17c9d122009-02-11 16:34:16 +00004034 chg = to - from;
4035
Mel Gorman5a6fe122009-02-10 14:02:27 +00004036 set_vma_resv_map(vma, resv_map);
4037 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
4038 }
4039
Dave Hansenc50ac052012-05-29 15:06:46 -07004040 if (chg < 0) {
4041 ret = chg;
4042 goto out_err;
4043 }
Mel Gorman17c9d122009-02-11 16:34:16 +00004044
Mike Kravetz1c5ecae2015-04-15 16:13:39 -07004045 /*
4046 * There must be enough pages in the subpool for the mapping. If
4047 * the subpool has a minimum size, there may be some global
4048 * reservations already in place (gbl_reserve).
4049 */
4050 gbl_reserve = hugepage_subpool_get_pages(spool, chg);
4051 if (gbl_reserve < 0) {
Dave Hansenc50ac052012-05-29 15:06:46 -07004052 ret = -ENOSPC;
4053 goto out_err;
4054 }
Mel Gorman17c9d122009-02-11 16:34:16 +00004055
4056 /*
4057 * Check enough hugepages are available for the reservation.
David Gibson90481622012-03-21 16:34:12 -07004058 * Hand the pages back to the subpool if there are not
Mel Gorman17c9d122009-02-11 16:34:16 +00004059 */
Mike Kravetz1c5ecae2015-04-15 16:13:39 -07004060 ret = hugetlb_acct_memory(h, gbl_reserve);
Mel Gorman17c9d122009-02-11 16:34:16 +00004061 if (ret < 0) {
Mike Kravetz1c5ecae2015-04-15 16:13:39 -07004062 /* put back original number of pages, chg */
4063 (void)hugepage_subpool_put_pages(spool, chg);
Dave Hansenc50ac052012-05-29 15:06:46 -07004064 goto out_err;
Mel Gorman17c9d122009-02-11 16:34:16 +00004065 }
4066
4067 /*
4068 * Account for the reservations made. Shared mappings record regions
4069 * that have reservations as they are shared by multiple VMAs.
4070 * When the last VMA disappears, the region map says how much
4071 * the reservation was and the page cache tells how much of
4072 * the reservation was consumed. Private mappings are per-VMA and
4073 * only the consumed reservations are tracked. When the VMA
4074 * disappears, the original reservation is the VMA size and the
4075 * consumed reservations are stored in the map. Hence, nothing
4076 * else has to be done for private mappings here
4077 */
Mike Kravetz33039672015-06-24 16:57:58 -07004078 if (!vma || vma->vm_flags & VM_MAYSHARE) {
4079 long add = region_add(resv_map, from, to);
4080
4081 if (unlikely(chg > add)) {
4082 /*
4083 * pages in this range were added to the reserve
4084 * map between region_chg and region_add. This
4085 * indicates a race with alloc_huge_page. Adjust
4086 * the subpool and reserve counts modified above
4087 * based on the difference.
4088 */
4089 long rsv_adjust;
4090
4091 rsv_adjust = hugepage_subpool_put_pages(spool,
4092 chg - add);
4093 hugetlb_acct_memory(h, -rsv_adjust);
4094 }
4095 }
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -07004096 return 0;
Dave Hansenc50ac052012-05-29 15:06:46 -07004097out_err:
Mike Kravetz5e911372015-09-08 15:01:28 -07004098 if (!vma || vma->vm_flags & VM_MAYSHARE)
4099 region_abort(resv_map, from, to);
Joonsoo Kimf031dd22014-04-03 14:47:28 -07004100 if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
4101 kref_put(&resv_map->refs, resv_map_release);
Dave Hansenc50ac052012-05-29 15:06:46 -07004102 return ret;
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -07004103}
4104
Mike Kravetzb5cec282015-09-08 15:01:41 -07004105long hugetlb_unreserve_pages(struct inode *inode, long start, long end,
4106 long freed)
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -07004107{
Andi Kleena5516432008-07-23 21:27:41 -07004108 struct hstate *h = hstate_inode(inode);
Joonsoo Kim4e35f482014-04-03 14:47:30 -07004109 struct resv_map *resv_map = inode_resv_map(inode);
Joonsoo Kim9119a412014-04-03 14:47:25 -07004110 long chg = 0;
David Gibson90481622012-03-21 16:34:12 -07004111 struct hugepage_subpool *spool = subpool_inode(inode);
Mike Kravetz1c5ecae2015-04-15 16:13:39 -07004112 long gbl_reserve;
Ken Chen45c682a2007-11-14 16:59:44 -08004113
Mike Kravetzb5cec282015-09-08 15:01:41 -07004114 if (resv_map) {
4115 chg = region_del(resv_map, start, end);
4116 /*
4117 * region_del() can fail in the rare case where a region
4118 * must be split and another region descriptor can not be
4119 * allocated. If end == LONG_MAX, it will not fail.
4120 */
4121 if (chg < 0)
4122 return chg;
4123 }
4124
Ken Chen45c682a2007-11-14 16:59:44 -08004125 spin_lock(&inode->i_lock);
Eric Sandeene4c6f8b2009-07-29 15:02:16 -07004126 inode->i_blocks -= (blocks_per_huge_page(h) * freed);
Ken Chen45c682a2007-11-14 16:59:44 -08004127 spin_unlock(&inode->i_lock);
4128
Mike Kravetz1c5ecae2015-04-15 16:13:39 -07004129 /*
4130 * If the subpool has a minimum size, the number of global
4131 * reservations to be released may be adjusted.
4132 */
4133 gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed));
4134 hugetlb_acct_memory(h, -gbl_reserve);
Mike Kravetzb5cec282015-09-08 15:01:41 -07004135
4136 return 0;
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -07004137}
Naoya Horiguchi93f70f92010-05-28 09:29:20 +09004138
Steve Capper3212b532013-04-23 12:35:02 +01004139#ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
4140static unsigned long page_table_shareable(struct vm_area_struct *svma,
4141 struct vm_area_struct *vma,
4142 unsigned long addr, pgoff_t idx)
4143{
4144 unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) +
4145 svma->vm_start;
4146 unsigned long sbase = saddr & PUD_MASK;
4147 unsigned long s_end = sbase + PUD_SIZE;
4148
4149 /* Allow segments to share if only one is marked locked */
Eric B Munsonde60f5f2015-11-05 18:51:36 -08004150 unsigned long vm_flags = vma->vm_flags & VM_LOCKED_CLEAR_MASK;
4151 unsigned long svm_flags = svma->vm_flags & VM_LOCKED_CLEAR_MASK;
Steve Capper3212b532013-04-23 12:35:02 +01004152
4153 /*
4154 * match the virtual addresses, permission and the alignment of the
4155 * page table page.
4156 */
4157 if (pmd_index(addr) != pmd_index(saddr) ||
4158 vm_flags != svm_flags ||
4159 sbase < svma->vm_start || svma->vm_end < s_end)
4160 return 0;
4161
4162 return saddr;
4163}
4164
Nicholas Krause31aafb42015-09-04 15:47:58 -07004165static bool vma_shareable(struct vm_area_struct *vma, unsigned long addr)
Steve Capper3212b532013-04-23 12:35:02 +01004166{
4167 unsigned long base = addr & PUD_MASK;
4168 unsigned long end = base + PUD_SIZE;
4169
4170 /*
4171 * check on proper vm_flags and page table alignment
4172 */
4173 if (vma->vm_flags & VM_MAYSHARE &&
4174 vma->vm_start <= base && end <= vma->vm_end)
Nicholas Krause31aafb42015-09-04 15:47:58 -07004175 return true;
4176 return false;
Steve Capper3212b532013-04-23 12:35:02 +01004177}
4178
4179/*
4180 * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
4181 * and returns the corresponding pte. While this is not necessary for the
4182 * !shared pmd case because we can allocate the pmd later as well, it makes the
4183 * code much cleaner. pmd allocation is essential for the shared case because
Davidlohr Buesoc8c06ef2014-12-12 16:54:24 -08004184 * pud has to be populated inside the same i_mmap_rwsem section - otherwise
Steve Capper3212b532013-04-23 12:35:02 +01004185 * racing tasks could either miss the sharing (see huge_pte_offset) or select a
4186 * bad pmd for sharing.
4187 */
4188pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
4189{
4190 struct vm_area_struct *vma = find_vma(mm, addr);
4191 struct address_space *mapping = vma->vm_file->f_mapping;
4192 pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) +
4193 vma->vm_pgoff;
4194 struct vm_area_struct *svma;
4195 unsigned long saddr;
4196 pte_t *spte = NULL;
4197 pte_t *pte;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08004198 spinlock_t *ptl;
Steve Capper3212b532013-04-23 12:35:02 +01004199
4200 if (!vma_shareable(vma, addr))
4201 return (pte_t *)pmd_alloc(mm, pud, addr);
4202
Davidlohr Bueso83cde9e2014-12-12 16:54:21 -08004203 i_mmap_lock_write(mapping);
Steve Capper3212b532013-04-23 12:35:02 +01004204 vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) {
4205 if (svma == vma)
4206 continue;
4207
4208 saddr = page_table_shareable(svma, vma, addr, idx);
4209 if (saddr) {
4210 spte = huge_pte_offset(svma->vm_mm, saddr);
4211 if (spte) {
Kirill A. Shutemovdc6c9a32015-02-11 15:26:50 -08004212 mm_inc_nr_pmds(mm);
Steve Capper3212b532013-04-23 12:35:02 +01004213 get_page(virt_to_page(spte));
4214 break;
4215 }
4216 }
4217 }
4218
4219 if (!spte)
4220 goto out;
4221
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08004222 ptl = huge_pte_lockptr(hstate_vma(vma), mm, spte);
4223 spin_lock(ptl);
Kirill A. Shutemovdc6c9a32015-02-11 15:26:50 -08004224 if (pud_none(*pud)) {
Steve Capper3212b532013-04-23 12:35:02 +01004225 pud_populate(mm, pud,
4226 (pmd_t *)((unsigned long)spte & PAGE_MASK));
Kirill A. Shutemovdc6c9a32015-02-11 15:26:50 -08004227 } else {
Steve Capper3212b532013-04-23 12:35:02 +01004228 put_page(virt_to_page(spte));
Kirill A. Shutemovdc6c9a32015-02-11 15:26:50 -08004229 mm_inc_nr_pmds(mm);
4230 }
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08004231 spin_unlock(ptl);
Steve Capper3212b532013-04-23 12:35:02 +01004232out:
4233 pte = (pte_t *)pmd_alloc(mm, pud, addr);
Davidlohr Bueso83cde9e2014-12-12 16:54:21 -08004234 i_mmap_unlock_write(mapping);
Steve Capper3212b532013-04-23 12:35:02 +01004235 return pte;
4236}
4237
4238/*
4239 * unmap huge page backed by shared pte.
4240 *
4241 * Hugetlb pte page is ref counted at the time of mapping. If pte is shared
4242 * indicated by page_count > 1, unmap is achieved by clearing pud and
4243 * decrementing the ref count. If count == 1, the pte page is not shared.
4244 *
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08004245 * called with page table lock held.
Steve Capper3212b532013-04-23 12:35:02 +01004246 *
4247 * returns: 1 successfully unmapped a shared pte page
4248 * 0 the underlying pte page is not shared, or it is the last user
4249 */
4250int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
4251{
4252 pgd_t *pgd = pgd_offset(mm, *addr);
4253 pud_t *pud = pud_offset(pgd, *addr);
4254
4255 BUG_ON(page_count(virt_to_page(ptep)) == 0);
4256 if (page_count(virt_to_page(ptep)) == 1)
4257 return 0;
4258
4259 pud_clear(pud);
4260 put_page(virt_to_page(ptep));
Kirill A. Shutemovdc6c9a32015-02-11 15:26:50 -08004261 mm_dec_nr_pmds(mm);
Steve Capper3212b532013-04-23 12:35:02 +01004262 *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE;
4263 return 1;
4264}
Steve Capper9e5fc742013-04-30 08:02:03 +01004265#define want_pmd_share() (1)
4266#else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
4267pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
4268{
4269 return NULL;
4270}
Zhang Zhene81f2d22015-06-24 16:56:13 -07004271
4272int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
4273{
4274 return 0;
4275}
Steve Capper9e5fc742013-04-30 08:02:03 +01004276#define want_pmd_share() (0)
Steve Capper3212b532013-04-23 12:35:02 +01004277#endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
4278
Steve Capper9e5fc742013-04-30 08:02:03 +01004279#ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
4280pte_t *huge_pte_alloc(struct mm_struct *mm,
4281 unsigned long addr, unsigned long sz)
4282{
4283 pgd_t *pgd;
4284 pud_t *pud;
4285 pte_t *pte = NULL;
4286
4287 pgd = pgd_offset(mm, addr);
4288 pud = pud_alloc(mm, pgd, addr);
4289 if (pud) {
4290 if (sz == PUD_SIZE) {
4291 pte = (pte_t *)pud;
4292 } else {
4293 BUG_ON(sz != PMD_SIZE);
4294 if (want_pmd_share() && pud_none(*pud))
4295 pte = huge_pmd_share(mm, addr, pud);
4296 else
4297 pte = (pte_t *)pmd_alloc(mm, pud, addr);
4298 }
4299 }
4300 BUG_ON(pte && !pte_none(*pte) && !pte_huge(*pte));
4301
4302 return pte;
4303}
4304
4305pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr)
4306{
4307 pgd_t *pgd;
4308 pud_t *pud;
4309 pmd_t *pmd = NULL;
4310
4311 pgd = pgd_offset(mm, addr);
4312 if (pgd_present(*pgd)) {
4313 pud = pud_offset(pgd, addr);
4314 if (pud_present(*pud)) {
4315 if (pud_huge(*pud))
4316 return (pte_t *)pud;
4317 pmd = pmd_offset(pud, addr);
4318 }
4319 }
4320 return (pte_t *) pmd;
4321}
4322
Naoya Horiguchi61f77ed2015-02-11 15:25:15 -08004323#endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
4324
4325/*
4326 * These functions are overwritable if your architecture needs its own
4327 * behavior.
4328 */
4329struct page * __weak
4330follow_huge_addr(struct mm_struct *mm, unsigned long address,
4331 int write)
4332{
4333 return ERR_PTR(-EINVAL);
4334}
4335
4336struct page * __weak
Steve Capper9e5fc742013-04-30 08:02:03 +01004337follow_huge_pmd(struct mm_struct *mm, unsigned long address,
Naoya Horiguchie66f17f2015-02-11 15:25:22 -08004338 pmd_t *pmd, int flags)
Steve Capper9e5fc742013-04-30 08:02:03 +01004339{
Naoya Horiguchie66f17f2015-02-11 15:25:22 -08004340 struct page *page = NULL;
4341 spinlock_t *ptl;
4342retry:
4343 ptl = pmd_lockptr(mm, pmd);
4344 spin_lock(ptl);
4345 /*
4346 * make sure that the address range covered by this pmd is not
4347 * unmapped from other threads.
4348 */
4349 if (!pmd_huge(*pmd))
4350 goto out;
4351 if (pmd_present(*pmd)) {
Gerald Schaefer97534122015-04-14 15:42:30 -07004352 page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT);
Naoya Horiguchie66f17f2015-02-11 15:25:22 -08004353 if (flags & FOLL_GET)
4354 get_page(page);
4355 } else {
4356 if (is_hugetlb_entry_migration(huge_ptep_get((pte_t *)pmd))) {
4357 spin_unlock(ptl);
4358 __migration_entry_wait(mm, (pte_t *)pmd, ptl);
4359 goto retry;
4360 }
4361 /*
4362 * hwpoisoned entry is treated as no_page_table in
4363 * follow_page_mask().
4364 */
4365 }
4366out:
4367 spin_unlock(ptl);
Steve Capper9e5fc742013-04-30 08:02:03 +01004368 return page;
4369}
4370
Naoya Horiguchi61f77ed2015-02-11 15:25:15 -08004371struct page * __weak
Steve Capper9e5fc742013-04-30 08:02:03 +01004372follow_huge_pud(struct mm_struct *mm, unsigned long address,
Naoya Horiguchie66f17f2015-02-11 15:25:22 -08004373 pud_t *pud, int flags)
Steve Capper9e5fc742013-04-30 08:02:03 +01004374{
Naoya Horiguchie66f17f2015-02-11 15:25:22 -08004375 if (flags & FOLL_GET)
4376 return NULL;
Steve Capper9e5fc742013-04-30 08:02:03 +01004377
Naoya Horiguchie66f17f2015-02-11 15:25:22 -08004378 return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT);
Steve Capper9e5fc742013-04-30 08:02:03 +01004379}
4380
Andi Kleend5bd9102010-09-27 09:00:12 +02004381#ifdef CONFIG_MEMORY_FAILURE
4382
Naoya Horiguchi93f70f92010-05-28 09:29:20 +09004383/*
4384 * This function is called from memory failure code.
4385 * Assume the caller holds page lock of the head page.
4386 */
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +09004387int dequeue_hwpoisoned_huge_page(struct page *hpage)
Naoya Horiguchi93f70f92010-05-28 09:29:20 +09004388{
4389 struct hstate *h = page_hstate(hpage);
4390 int nid = page_to_nid(hpage);
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +09004391 int ret = -EBUSY;
Naoya Horiguchi93f70f92010-05-28 09:29:20 +09004392
4393 spin_lock(&hugetlb_lock);
Naoya Horiguchi7e1f0492015-04-15 16:14:41 -07004394 /*
4395 * Just checking !page_huge_active is not enough, because that could be
4396 * an isolated/hwpoisoned hugepage (which have >0 refcount).
4397 */
4398 if (!page_huge_active(hpage) && !page_count(hpage)) {
Naoya Horiguchi56f2fb12012-12-12 13:52:33 -08004399 /*
4400 * Hwpoisoned hugepage isn't linked to activelist or freelist,
4401 * but dangling hpage->lru can trigger list-debug warnings
4402 * (this happens when we call unpoison_memory() on it),
4403 * so let it point to itself with list_del_init().
4404 */
4405 list_del_init(&hpage->lru);
Naoya Horiguchi8c6c2ec2010-09-08 10:19:38 +09004406 set_page_refcounted(hpage);
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +09004407 h->free_huge_pages--;
4408 h->free_huge_pages_node[nid]--;
4409 ret = 0;
4410 }
Naoya Horiguchi93f70f92010-05-28 09:29:20 +09004411 spin_unlock(&hugetlb_lock);
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +09004412 return ret;
Naoya Horiguchi93f70f92010-05-28 09:29:20 +09004413}
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +09004414#endif
Naoya Horiguchi31caf662013-09-11 14:21:59 -07004415
4416bool isolate_huge_page(struct page *page, struct list_head *list)
4417{
Naoya Horiguchibcc54222015-04-15 16:14:38 -07004418 bool ret = true;
4419
Sasha Levin309381fea2014-01-23 15:52:54 -08004420 VM_BUG_ON_PAGE(!PageHead(page), page);
Naoya Horiguchi31caf662013-09-11 14:21:59 -07004421 spin_lock(&hugetlb_lock);
Naoya Horiguchibcc54222015-04-15 16:14:38 -07004422 if (!page_huge_active(page) || !get_page_unless_zero(page)) {
4423 ret = false;
4424 goto unlock;
4425 }
4426 clear_page_huge_active(page);
Naoya Horiguchi31caf662013-09-11 14:21:59 -07004427 list_move_tail(&page->lru, list);
Naoya Horiguchibcc54222015-04-15 16:14:38 -07004428unlock:
Naoya Horiguchi31caf662013-09-11 14:21:59 -07004429 spin_unlock(&hugetlb_lock);
Naoya Horiguchibcc54222015-04-15 16:14:38 -07004430 return ret;
Naoya Horiguchi31caf662013-09-11 14:21:59 -07004431}
4432
4433void putback_active_hugepage(struct page *page)
4434{
Sasha Levin309381fea2014-01-23 15:52:54 -08004435 VM_BUG_ON_PAGE(!PageHead(page), page);
Naoya Horiguchi31caf662013-09-11 14:21:59 -07004436 spin_lock(&hugetlb_lock);
Naoya Horiguchibcc54222015-04-15 16:14:38 -07004437 set_page_huge_active(page);
Naoya Horiguchi31caf662013-09-11 14:21:59 -07004438 list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist);
4439 spin_unlock(&hugetlb_lock);
4440 put_page(page);
4441}