blob: b9aa1b0b38b0ecdb769cd33a8d77c0ffda2621bf [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
2 * Generic hugetlb support.
Nadia Yvette Chambers6d49e352012-12-06 10:39:54 +01003 * (C) Nadia Yvette Chambers, April 2004
Linus Torvalds1da177e2005-04-16 15:20:36 -07004 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07005#include <linux/list.h>
6#include <linux/init.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -07007#include <linux/mm.h>
Alexey Dobriyane1759c22008-10-15 23:50:22 +04008#include <linux/seq_file.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -07009#include <linux/sysctl.h>
10#include <linux/highmem.h>
Andrea Arcangelicddb8a52008-07-28 15:46:29 -070011#include <linux/mmu_notifier.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070012#include <linux/nodemask.h>
David Gibson63551ae2005-06-21 17:14:44 -070013#include <linux/pagemap.h>
Christoph Lameter5da7ca82006-01-06 00:10:46 -080014#include <linux/mempolicy.h>
Gideon Israel Dsouza3b321232014-04-07 15:37:26 -070015#include <linux/compiler.h>
Christoph Lameteraea47ff2006-01-08 01:00:57 -080016#include <linux/cpuset.h>
David Gibson3935baa2006-03-22 00:08:53 -080017#include <linux/mutex.h>
Andi Kleenaa888a72008-07-23 21:27:47 -070018#include <linux/bootmem.h>
Nishanth Aravamudana3437872008-07-23 21:27:44 -070019#include <linux/sysfs.h>
Tejun Heo5a0e3ad2010-03-24 17:04:11 +090020#include <linux/slab.h>
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +090021#include <linux/rmap.h>
Naoya Horiguchifd6a03ed2010-05-28 09:29:21 +090022#include <linux/swap.h>
23#include <linux/swapops.h>
Naoya Horiguchic8721bb2013-09-11 14:22:09 -070024#include <linux/page-isolation.h>
Davidlohr Bueso8382d912014-04-03 14:47:31 -070025#include <linux/jhash.h>
Linus Torvaldsd6606682008-08-06 12:04:54 -070026
David Gibson63551ae2005-06-21 17:14:44 -070027#include <asm/page.h>
28#include <asm/pgtable.h>
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -070029#include <asm/tlb.h>
David Gibson63551ae2005-06-21 17:14:44 -070030
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -070031#include <linux/io.h>
David Gibson63551ae2005-06-21 17:14:44 -070032#include <linux/hugetlb.h>
Aneesh Kumar K.V9dd540e2012-07-31 16:42:15 -070033#include <linux/hugetlb_cgroup.h>
Lee Schermerhorn9a305232009-12-14 17:58:25 -080034#include <linux/node.h>
Nick Piggin7835e982006-03-22 00:08:40 -080035#include "internal.h"
Linus Torvalds1da177e2005-04-16 15:20:36 -070036
Andrey Ryabinin753162c2015-02-10 14:11:36 -080037int hugepages_treat_as_movable;
Andi Kleena5516432008-07-23 21:27:41 -070038
Aneesh Kumar K.Vc3f38a32012-07-31 16:42:10 -070039int hugetlb_max_hstate __read_mostly;
Andi Kleene5ff2152008-07-23 21:27:42 -070040unsigned int default_hstate_idx;
41struct hstate hstates[HUGE_MAX_HSTATE];
Naoya Horiguchi641844f2015-06-24 16:56:59 -070042/*
43 * Minimum page order among possible hugepage sizes, set to a proper value
44 * at boot time.
45 */
46static unsigned int minimum_order __read_mostly = UINT_MAX;
Andi Kleene5ff2152008-07-23 21:27:42 -070047
Jon Tollefson53ba51d2008-07-23 21:27:52 -070048__initdata LIST_HEAD(huge_boot_pages);
49
Andi Kleene5ff2152008-07-23 21:27:42 -070050/* for command line parsing */
51static struct hstate * __initdata parsed_hstate;
52static unsigned long __initdata default_hstate_max_huge_pages;
Nick Piggine11bfbf2008-07-23 21:27:52 -070053static unsigned long __initdata default_hstate_size;
Vaishali Thakkar9fee0212016-05-19 17:11:04 -070054static bool __initdata parsed_valid_hugepagesz = true;
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
Mike Kravetz09a95e22016-05-19 17:11:01 -0700148 /* minimum size accounting */
149 if (spool->min_hpages != -1 && spool->rsv_hpages) {
Mike Kravetz1c5ecae2015-04-15 16:13:39 -0700150 if (delta > spool->rsv_hpages) {
151 /*
152 * Asking for more reserves than those already taken on
153 * behalf of subpool. Return difference.
154 */
155 ret = delta - spool->rsv_hpages;
156 spool->rsv_hpages = 0;
157 } else {
158 ret = 0; /* reserves already accounted for */
159 spool->rsv_hpages -= delta;
160 }
161 }
162
163unlock_ret:
164 spin_unlock(&spool->lock);
David Gibson90481622012-03-21 16:34:12 -0700165 return ret;
166}
167
Mike Kravetz1c5ecae2015-04-15 16:13:39 -0700168/*
169 * Subpool accounting for freeing and unreserving pages.
170 * Return the number of global page reservations that must be dropped.
171 * The return value may only be different than the passed value (delta)
172 * in the case where a subpool minimum size must be maintained.
173 */
174static long hugepage_subpool_put_pages(struct hugepage_subpool *spool,
David Gibson90481622012-03-21 16:34:12 -0700175 long delta)
176{
Mike Kravetz1c5ecae2015-04-15 16:13:39 -0700177 long ret = delta;
178
David Gibson90481622012-03-21 16:34:12 -0700179 if (!spool)
Mike Kravetz1c5ecae2015-04-15 16:13:39 -0700180 return delta;
David Gibson90481622012-03-21 16:34:12 -0700181
182 spin_lock(&spool->lock);
Mike Kravetz1c5ecae2015-04-15 16:13:39 -0700183
184 if (spool->max_hpages != -1) /* maximum size accounting */
185 spool->used_hpages -= delta;
186
Mike Kravetz09a95e22016-05-19 17:11:01 -0700187 /* minimum size accounting */
188 if (spool->min_hpages != -1 && spool->used_hpages < spool->min_hpages) {
Mike Kravetz1c5ecae2015-04-15 16:13:39 -0700189 if (spool->rsv_hpages + delta <= spool->min_hpages)
190 ret = 0;
191 else
192 ret = spool->rsv_hpages + delta - spool->min_hpages;
193
194 spool->rsv_hpages += delta;
195 if (spool->rsv_hpages > spool->min_hpages)
196 spool->rsv_hpages = spool->min_hpages;
197 }
198
199 /*
200 * If hugetlbfs_put_super couldn't free spool due to an outstanding
201 * quota reference, free it now.
202 */
David Gibson90481622012-03-21 16:34:12 -0700203 unlock_or_release_subpool(spool);
Mike Kravetz1c5ecae2015-04-15 16:13:39 -0700204
205 return ret;
David Gibson90481622012-03-21 16:34:12 -0700206}
207
208static inline struct hugepage_subpool *subpool_inode(struct inode *inode)
209{
210 return HUGETLBFS_SB(inode->i_sb)->spool;
211}
212
213static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma)
214{
Al Viro496ad9a2013-01-23 17:07:38 -0500215 return subpool_inode(file_inode(vma->vm_file));
David Gibson90481622012-03-21 16:34:12 -0700216}
217
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700218/*
Andy Whitcroft96822902008-07-23 21:27:29 -0700219 * Region tracking -- allows tracking of reservations and instantiated pages
220 * across the pages in a mapping.
Andy Whitcroft84afd992008-07-23 21:27:32 -0700221 *
Mike Kravetz1dd308a2015-06-24 16:57:52 -0700222 * The region data structures are embedded into a resv_map and protected
223 * by a resv_map's lock. The set of regions within the resv_map represent
224 * reservations for huge pages, or huge pages that have already been
225 * instantiated within the map. The from and to elements are huge page
226 * indicies into the associated mapping. from indicates the starting index
227 * of the region. to represents the first index past the end of the region.
228 *
229 * For example, a file region structure with from == 0 and to == 4 represents
230 * four huge pages in a mapping. It is important to note that the to element
231 * represents the first element past the end of the region. This is used in
232 * arithmetic as 4(to) - 0(from) = 4 huge pages in the region.
233 *
234 * Interval notation of the form [from, to) will be used to indicate that
235 * the endpoint from is inclusive and to is exclusive.
Andy Whitcroft96822902008-07-23 21:27:29 -0700236 */
237struct file_region {
238 struct list_head link;
239 long from;
240 long to;
241};
242
Mike Kravetz1dd308a2015-06-24 16:57:52 -0700243/*
244 * Add the huge page range represented by [f, t) to the reserve
Mike Kravetz5e911372015-09-08 15:01:28 -0700245 * map. In the normal case, existing regions will be expanded
246 * to accommodate the specified range. Sufficient regions should
247 * exist for expansion due to the previous call to region_chg
248 * with the same range. However, it is possible that region_del
249 * could have been called after region_chg and modifed the map
250 * in such a way that no region exists to be expanded. In this
251 * case, pull a region descriptor from the cache associated with
252 * the map and use that for the new range.
Mike Kravetzcf3ad202015-06-24 16:57:55 -0700253 *
254 * Return the number of new huge pages added to the map. This
255 * number is greater than or equal to zero.
Mike Kravetz1dd308a2015-06-24 16:57:52 -0700256 */
Joonsoo Kim1406ec92014-04-03 14:47:26 -0700257static long region_add(struct resv_map *resv, long f, long t)
Andy Whitcroft96822902008-07-23 21:27:29 -0700258{
Joonsoo Kim1406ec92014-04-03 14:47:26 -0700259 struct list_head *head = &resv->regions;
Andy Whitcroft96822902008-07-23 21:27:29 -0700260 struct file_region *rg, *nrg, *trg;
Mike Kravetzcf3ad202015-06-24 16:57:55 -0700261 long add = 0;
Andy Whitcroft96822902008-07-23 21:27:29 -0700262
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700263 spin_lock(&resv->lock);
Andy Whitcroft96822902008-07-23 21:27:29 -0700264 /* Locate the region we are either in or before. */
265 list_for_each_entry(rg, head, link)
266 if (f <= rg->to)
267 break;
268
Mike Kravetz5e911372015-09-08 15:01:28 -0700269 /*
270 * If no region exists which can be expanded to include the
271 * specified range, the list must have been modified by an
272 * interleving call to region_del(). Pull a region descriptor
273 * from the cache and use it for this range.
274 */
275 if (&rg->link == head || t < rg->from) {
276 VM_BUG_ON(resv->region_cache_count <= 0);
277
278 resv->region_cache_count--;
279 nrg = list_first_entry(&resv->region_cache, struct file_region,
280 link);
281 list_del(&nrg->link);
282
283 nrg->from = f;
284 nrg->to = t;
285 list_add(&nrg->link, rg->link.prev);
286
287 add += t - f;
288 goto out_locked;
289 }
290
Andy Whitcroft96822902008-07-23 21:27:29 -0700291 /* Round our left edge to the current segment if it encloses us. */
292 if (f > rg->from)
293 f = rg->from;
294
295 /* Check for and consume any regions we now overlap with. */
296 nrg = rg;
297 list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
298 if (&rg->link == head)
299 break;
300 if (rg->from > t)
301 break;
302
303 /* If this area reaches higher then extend our area to
304 * include it completely. If this is not the first area
305 * which we intend to reuse, free it. */
306 if (rg->to > t)
307 t = rg->to;
308 if (rg != nrg) {
Mike Kravetzcf3ad202015-06-24 16:57:55 -0700309 /* Decrement return value by the deleted range.
310 * Another range will span this area so that by
311 * end of routine add will be >= zero
312 */
313 add -= (rg->to - rg->from);
Andy Whitcroft96822902008-07-23 21:27:29 -0700314 list_del(&rg->link);
315 kfree(rg);
316 }
317 }
Mike Kravetzcf3ad202015-06-24 16:57:55 -0700318
319 add += (nrg->from - f); /* Added to beginning of region */
Andy Whitcroft96822902008-07-23 21:27:29 -0700320 nrg->from = f;
Mike Kravetzcf3ad202015-06-24 16:57:55 -0700321 add += t - nrg->to; /* Added to end of region */
Andy Whitcroft96822902008-07-23 21:27:29 -0700322 nrg->to = t;
Mike Kravetzcf3ad202015-06-24 16:57:55 -0700323
Mike Kravetz5e911372015-09-08 15:01:28 -0700324out_locked:
325 resv->adds_in_progress--;
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700326 spin_unlock(&resv->lock);
Mike Kravetzcf3ad202015-06-24 16:57:55 -0700327 VM_BUG_ON(add < 0);
328 return add;
Andy Whitcroft96822902008-07-23 21:27:29 -0700329}
330
Mike Kravetz1dd308a2015-06-24 16:57:52 -0700331/*
332 * Examine the existing reserve map and determine how many
333 * huge pages in the specified range [f, t) are NOT currently
334 * represented. This routine is called before a subsequent
335 * call to region_add that will actually modify the reserve
336 * map to add the specified range [f, t). region_chg does
337 * not change the number of huge pages represented by the
338 * map. However, if the existing regions in the map can not
339 * be expanded to represent the new range, a new file_region
340 * structure is added to the map as a placeholder. This is
341 * so that the subsequent region_add call will have all the
342 * regions it needs and will not fail.
343 *
Mike Kravetz5e911372015-09-08 15:01:28 -0700344 * Upon entry, region_chg will also examine the cache of region descriptors
345 * associated with the map. If there are not enough descriptors cached, one
346 * will be allocated for the in progress add operation.
347 *
348 * Returns the number of huge pages that need to be added to the existing
349 * reservation map for the range [f, t). This number is greater or equal to
350 * zero. -ENOMEM is returned if a new file_region structure or cache entry
351 * is needed and can not be allocated.
Mike Kravetz1dd308a2015-06-24 16:57:52 -0700352 */
Joonsoo Kim1406ec92014-04-03 14:47:26 -0700353static long region_chg(struct resv_map *resv, long f, long t)
Andy Whitcroft96822902008-07-23 21:27:29 -0700354{
Joonsoo Kim1406ec92014-04-03 14:47:26 -0700355 struct list_head *head = &resv->regions;
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700356 struct file_region *rg, *nrg = NULL;
Andy Whitcroft96822902008-07-23 21:27:29 -0700357 long chg = 0;
358
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700359retry:
360 spin_lock(&resv->lock);
Mike Kravetz5e911372015-09-08 15:01:28 -0700361retry_locked:
362 resv->adds_in_progress++;
363
364 /*
365 * Check for sufficient descriptors in the cache to accommodate
366 * the number of in progress add operations.
367 */
368 if (resv->adds_in_progress > resv->region_cache_count) {
369 struct file_region *trg;
370
371 VM_BUG_ON(resv->adds_in_progress - resv->region_cache_count > 1);
372 /* Must drop lock to allocate a new descriptor. */
373 resv->adds_in_progress--;
374 spin_unlock(&resv->lock);
375
376 trg = kmalloc(sizeof(*trg), GFP_KERNEL);
Mike Kravetzdbe409e2015-12-11 13:40:52 -0800377 if (!trg) {
378 kfree(nrg);
Mike Kravetz5e911372015-09-08 15:01:28 -0700379 return -ENOMEM;
Mike Kravetzdbe409e2015-12-11 13:40:52 -0800380 }
Mike Kravetz5e911372015-09-08 15:01:28 -0700381
382 spin_lock(&resv->lock);
383 list_add(&trg->link, &resv->region_cache);
384 resv->region_cache_count++;
385 goto retry_locked;
386 }
387
Andy Whitcroft96822902008-07-23 21:27:29 -0700388 /* Locate the region we are before or in. */
389 list_for_each_entry(rg, head, link)
390 if (f <= rg->to)
391 break;
392
393 /* If we are below the current region then a new region is required.
394 * Subtle, allocate a new region at the position but make it zero
395 * size such that we can guarantee to record the reservation. */
396 if (&rg->link == head || t < rg->from) {
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700397 if (!nrg) {
Mike Kravetz5e911372015-09-08 15:01:28 -0700398 resv->adds_in_progress--;
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700399 spin_unlock(&resv->lock);
400 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
401 if (!nrg)
402 return -ENOMEM;
Andy Whitcroft96822902008-07-23 21:27:29 -0700403
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700404 nrg->from = f;
405 nrg->to = f;
406 INIT_LIST_HEAD(&nrg->link);
407 goto retry;
408 }
409
410 list_add(&nrg->link, rg->link.prev);
411 chg = t - f;
412 goto out_nrg;
Andy Whitcroft96822902008-07-23 21:27:29 -0700413 }
414
415 /* Round our left edge to the current segment if it encloses us. */
416 if (f > rg->from)
417 f = rg->from;
418 chg = t - f;
419
420 /* Check for and consume any regions we now overlap with. */
421 list_for_each_entry(rg, rg->link.prev, link) {
422 if (&rg->link == head)
423 break;
424 if (rg->from > t)
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700425 goto out;
Andy Whitcroft96822902008-07-23 21:27:29 -0700426
Lucas De Marchi25985ed2011-03-30 22:57:33 -0300427 /* We overlap with this area, if it extends further than
Andy Whitcroft96822902008-07-23 21:27:29 -0700428 * us then we must extend ourselves. Account for its
429 * existing reservation. */
430 if (rg->to > t) {
431 chg += rg->to - t;
432 t = rg->to;
433 }
434 chg -= rg->to - rg->from;
435 }
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700436
437out:
438 spin_unlock(&resv->lock);
439 /* We already know we raced and no longer need the new region */
440 kfree(nrg);
441 return chg;
442out_nrg:
443 spin_unlock(&resv->lock);
Andy Whitcroft96822902008-07-23 21:27:29 -0700444 return chg;
445}
446
Mike Kravetz1dd308a2015-06-24 16:57:52 -0700447/*
Mike Kravetz5e911372015-09-08 15:01:28 -0700448 * Abort the in progress add operation. The adds_in_progress field
449 * of the resv_map keeps track of the operations in progress between
450 * calls to region_chg and region_add. Operations are sometimes
451 * aborted after the call to region_chg. In such cases, region_abort
452 * is called to decrement the adds_in_progress counter.
453 *
454 * NOTE: The range arguments [f, t) are not needed or used in this
455 * routine. They are kept to make reading the calling code easier as
456 * arguments will match the associated region_chg call.
457 */
458static void region_abort(struct resv_map *resv, long f, long t)
459{
460 spin_lock(&resv->lock);
461 VM_BUG_ON(!resv->region_cache_count);
462 resv->adds_in_progress--;
463 spin_unlock(&resv->lock);
464}
465
466/*
Mike Kravetzfeba16e2015-09-08 15:01:31 -0700467 * Delete the specified range [f, t) from the reserve map. If the
468 * t parameter is LONG_MAX, this indicates that ALL regions after f
469 * should be deleted. Locate the regions which intersect [f, t)
470 * and either trim, delete or split the existing regions.
471 *
472 * Returns the number of huge pages deleted from the reserve map.
473 * In the normal case, the return value is zero or more. In the
474 * case where a region must be split, a new region descriptor must
475 * be allocated. If the allocation fails, -ENOMEM will be returned.
476 * NOTE: If the parameter t == LONG_MAX, then we will never split
477 * a region and possibly return -ENOMEM. Callers specifying
478 * t == LONG_MAX do not need to check for -ENOMEM error.
Mike Kravetz1dd308a2015-06-24 16:57:52 -0700479 */
Mike Kravetzfeba16e2015-09-08 15:01:31 -0700480static long region_del(struct resv_map *resv, long f, long t)
Andy Whitcroft96822902008-07-23 21:27:29 -0700481{
Joonsoo Kim1406ec92014-04-03 14:47:26 -0700482 struct list_head *head = &resv->regions;
Andy Whitcroft96822902008-07-23 21:27:29 -0700483 struct file_region *rg, *trg;
Mike Kravetzfeba16e2015-09-08 15:01:31 -0700484 struct file_region *nrg = NULL;
485 long del = 0;
Andy Whitcroft96822902008-07-23 21:27:29 -0700486
Mike Kravetzfeba16e2015-09-08 15:01:31 -0700487retry:
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700488 spin_lock(&resv->lock);
Mike Kravetzfeba16e2015-09-08 15:01:31 -0700489 list_for_each_entry_safe(rg, trg, head, link) {
Mike Kravetzdbe409e2015-12-11 13:40:52 -0800490 /*
491 * Skip regions before the range to be deleted. file_region
492 * ranges are normally of the form [from, to). However, there
493 * may be a "placeholder" entry in the map which is of the form
494 * (from, to) with from == to. Check for placeholder entries
495 * at the beginning of the range to be deleted.
496 */
497 if (rg->to <= f && (rg->to != rg->from || rg->to != f))
Mike Kravetzfeba16e2015-09-08 15:01:31 -0700498 continue;
Mike Kravetzdbe409e2015-12-11 13:40:52 -0800499
Mike Kravetzfeba16e2015-09-08 15:01:31 -0700500 if (rg->from >= t)
Andy Whitcroft96822902008-07-23 21:27:29 -0700501 break;
Andy Whitcroft96822902008-07-23 21:27:29 -0700502
Mike Kravetzfeba16e2015-09-08 15:01:31 -0700503 if (f > rg->from && t < rg->to) { /* Must split region */
504 /*
505 * Check for an entry in the cache before dropping
506 * lock and attempting allocation.
507 */
508 if (!nrg &&
509 resv->region_cache_count > resv->adds_in_progress) {
510 nrg = list_first_entry(&resv->region_cache,
511 struct file_region,
512 link);
513 list_del(&nrg->link);
514 resv->region_cache_count--;
515 }
516
517 if (!nrg) {
518 spin_unlock(&resv->lock);
519 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
520 if (!nrg)
521 return -ENOMEM;
522 goto retry;
523 }
524
525 del += t - f;
526
527 /* New entry for end of split region */
528 nrg->from = t;
529 nrg->to = rg->to;
530 INIT_LIST_HEAD(&nrg->link);
531
532 /* Original entry is trimmed */
533 rg->to = f;
534
535 list_add(&nrg->link, &rg->link);
536 nrg = NULL;
537 break;
538 }
539
540 if (f <= rg->from && t >= rg->to) { /* Remove entire region */
541 del += rg->to - rg->from;
542 list_del(&rg->link);
543 kfree(rg);
544 continue;
545 }
546
547 if (f <= rg->from) { /* Trim beginning of region */
548 del += t - rg->from;
549 rg->from = t;
550 } else { /* Trim end of region */
551 del += rg->to - f;
552 rg->to = f;
553 }
Andy Whitcroft96822902008-07-23 21:27:29 -0700554 }
555
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700556 spin_unlock(&resv->lock);
Mike Kravetzfeba16e2015-09-08 15:01:31 -0700557 kfree(nrg);
558 return del;
Andy Whitcroft96822902008-07-23 21:27:29 -0700559}
560
Mike Kravetz1dd308a2015-06-24 16:57:52 -0700561/*
Mike Kravetzb5cec282015-09-08 15:01:41 -0700562 * A rare out of memory error was encountered which prevented removal of
563 * the reserve map region for a page. The huge page itself was free'ed
564 * and removed from the page cache. This routine will adjust the subpool
565 * usage count, and the global reserve count if needed. By incrementing
566 * these counts, the reserve map entry which could not be deleted will
567 * appear as a "reserved" entry instead of simply dangling with incorrect
568 * counts.
569 */
570void hugetlb_fix_reserve_counts(struct inode *inode, bool restore_reserve)
571{
572 struct hugepage_subpool *spool = subpool_inode(inode);
573 long rsv_adjust;
574
575 rsv_adjust = hugepage_subpool_get_pages(spool, 1);
576 if (restore_reserve && rsv_adjust) {
577 struct hstate *h = hstate_inode(inode);
578
579 hugetlb_acct_memory(h, 1);
580 }
581}
582
583/*
Mike Kravetz1dd308a2015-06-24 16:57:52 -0700584 * Count and return the number of huge pages in the reserve map
585 * that intersect with the range [f, t).
586 */
Joonsoo Kim1406ec92014-04-03 14:47:26 -0700587static long region_count(struct resv_map *resv, long f, long t)
Andy Whitcroft84afd992008-07-23 21:27:32 -0700588{
Joonsoo Kim1406ec92014-04-03 14:47:26 -0700589 struct list_head *head = &resv->regions;
Andy Whitcroft84afd992008-07-23 21:27:32 -0700590 struct file_region *rg;
591 long chg = 0;
592
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700593 spin_lock(&resv->lock);
Andy Whitcroft84afd992008-07-23 21:27:32 -0700594 /* Locate each segment we overlap with, and count that overlap. */
595 list_for_each_entry(rg, head, link) {
Wang Sheng-Huif2135a42012-05-29 15:06:17 -0700596 long seg_from;
597 long seg_to;
Andy Whitcroft84afd992008-07-23 21:27:32 -0700598
599 if (rg->to <= f)
600 continue;
601 if (rg->from >= t)
602 break;
603
604 seg_from = max(rg->from, f);
605 seg_to = min(rg->to, t);
606
607 chg += seg_to - seg_from;
608 }
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700609 spin_unlock(&resv->lock);
Andy Whitcroft84afd992008-07-23 21:27:32 -0700610
611 return chg;
612}
613
Andy Whitcroft96822902008-07-23 21:27:29 -0700614/*
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700615 * Convert the address within this vma to the page offset within
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700616 * the mapping, in pagecache page units; huge pages here.
617 */
Andi Kleena5516432008-07-23 21:27:41 -0700618static pgoff_t vma_hugecache_offset(struct hstate *h,
619 struct vm_area_struct *vma, unsigned long address)
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700620{
Andi Kleena5516432008-07-23 21:27:41 -0700621 return ((address - vma->vm_start) >> huge_page_shift(h)) +
622 (vma->vm_pgoff >> huge_page_order(h));
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700623}
624
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +0900625pgoff_t linear_hugepage_index(struct vm_area_struct *vma,
626 unsigned long address)
627{
628 return vma_hugecache_offset(hstate_vma(vma), vma, address);
629}
Dan Williamsdee41072016-05-14 12:20:44 -0700630EXPORT_SYMBOL_GPL(linear_hugepage_index);
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +0900631
Andy Whitcroft84afd992008-07-23 21:27:32 -0700632/*
Mel Gorman08fba692009-01-06 14:38:53 -0800633 * Return the size of the pages allocated when backing a VMA. In the majority
634 * cases this will be same size as used by the page table entries.
635 */
636unsigned long vma_kernel_pagesize(struct vm_area_struct *vma)
637{
638 struct hstate *hstate;
639
640 if (!is_vm_hugetlb_page(vma))
641 return PAGE_SIZE;
642
643 hstate = hstate_vma(vma);
644
Wanpeng Li2415cf12013-07-03 15:02:43 -0700645 return 1UL << huge_page_shift(hstate);
Mel Gorman08fba692009-01-06 14:38:53 -0800646}
Joerg Roedelf340ca02009-06-19 15:16:22 +0200647EXPORT_SYMBOL_GPL(vma_kernel_pagesize);
Mel Gorman08fba692009-01-06 14:38:53 -0800648
649/*
Mel Gorman33402892009-01-06 14:38:54 -0800650 * Return the page size being used by the MMU to back a VMA. In the majority
651 * of cases, the page size used by the kernel matches the MMU size. On
652 * architectures where it differs, an architecture-specific version of this
653 * function is required.
654 */
655#ifndef vma_mmu_pagesize
656unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
657{
658 return vma_kernel_pagesize(vma);
659}
660#endif
661
662/*
Andy Whitcroft84afd992008-07-23 21:27:32 -0700663 * Flags for MAP_PRIVATE reservations. These are stored in the bottom
664 * bits of the reservation map pointer, which are always clear due to
665 * alignment.
666 */
667#define HPAGE_RESV_OWNER (1UL << 0)
668#define HPAGE_RESV_UNMAPPED (1UL << 1)
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700669#define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
Andy Whitcroft84afd992008-07-23 21:27:32 -0700670
Mel Gormana1e78772008-07-23 21:27:23 -0700671/*
672 * These helpers are used to track how many pages are reserved for
673 * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
674 * is guaranteed to have their future faults succeed.
675 *
676 * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
677 * the reserve counters are updated with the hugetlb_lock held. It is safe
678 * to reset the VMA at fork() time as it is not in use yet and there is no
679 * chance of the global counters getting corrupted as a result of the values.
Andy Whitcroft84afd992008-07-23 21:27:32 -0700680 *
681 * The private mapping reservation is represented in a subtly different
682 * manner to a shared mapping. A shared mapping has a region map associated
683 * with the underlying file, this region map represents the backing file
684 * pages which have ever had a reservation assigned which this persists even
685 * after the page is instantiated. A private mapping has a region map
686 * associated with the original mmap which is attached to all VMAs which
687 * reference it, this region map represents those offsets which have consumed
688 * reservation ie. where pages have been instantiated.
Mel Gormana1e78772008-07-23 21:27:23 -0700689 */
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700690static unsigned long get_vma_private_data(struct vm_area_struct *vma)
691{
692 return (unsigned long)vma->vm_private_data;
693}
694
695static void set_vma_private_data(struct vm_area_struct *vma,
696 unsigned long value)
697{
698 vma->vm_private_data = (void *)value;
699}
700
Joonsoo Kim9119a412014-04-03 14:47:25 -0700701struct resv_map *resv_map_alloc(void)
Andy Whitcroft84afd992008-07-23 21:27:32 -0700702{
703 struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
Mike Kravetz5e911372015-09-08 15:01:28 -0700704 struct file_region *rg = kmalloc(sizeof(*rg), GFP_KERNEL);
705
706 if (!resv_map || !rg) {
707 kfree(resv_map);
708 kfree(rg);
Andy Whitcroft84afd992008-07-23 21:27:32 -0700709 return NULL;
Mike Kravetz5e911372015-09-08 15:01:28 -0700710 }
Andy Whitcroft84afd992008-07-23 21:27:32 -0700711
712 kref_init(&resv_map->refs);
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700713 spin_lock_init(&resv_map->lock);
Andy Whitcroft84afd992008-07-23 21:27:32 -0700714 INIT_LIST_HEAD(&resv_map->regions);
715
Mike Kravetz5e911372015-09-08 15:01:28 -0700716 resv_map->adds_in_progress = 0;
717
718 INIT_LIST_HEAD(&resv_map->region_cache);
719 list_add(&rg->link, &resv_map->region_cache);
720 resv_map->region_cache_count = 1;
721
Andy Whitcroft84afd992008-07-23 21:27:32 -0700722 return resv_map;
723}
724
Joonsoo Kim9119a412014-04-03 14:47:25 -0700725void resv_map_release(struct kref *ref)
Andy Whitcroft84afd992008-07-23 21:27:32 -0700726{
727 struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
Mike Kravetz5e911372015-09-08 15:01:28 -0700728 struct list_head *head = &resv_map->region_cache;
729 struct file_region *rg, *trg;
Andy Whitcroft84afd992008-07-23 21:27:32 -0700730
731 /* Clear out any active regions before we release the map. */
Mike Kravetzfeba16e2015-09-08 15:01:31 -0700732 region_del(resv_map, 0, LONG_MAX);
Mike Kravetz5e911372015-09-08 15:01:28 -0700733
734 /* ... and any entries left in the cache */
735 list_for_each_entry_safe(rg, trg, head, link) {
736 list_del(&rg->link);
737 kfree(rg);
738 }
739
740 VM_BUG_ON(resv_map->adds_in_progress);
741
Andy Whitcroft84afd992008-07-23 21:27:32 -0700742 kfree(resv_map);
743}
744
Joonsoo Kim4e35f482014-04-03 14:47:30 -0700745static inline struct resv_map *inode_resv_map(struct inode *inode)
746{
747 return inode->i_mapping->private_data;
748}
749
Andy Whitcroft84afd992008-07-23 21:27:32 -0700750static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
Mel Gormana1e78772008-07-23 21:27:23 -0700751{
Sasha Levin81d1b092014-10-09 15:28:10 -0700752 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
Joonsoo Kim4e35f482014-04-03 14:47:30 -0700753 if (vma->vm_flags & VM_MAYSHARE) {
754 struct address_space *mapping = vma->vm_file->f_mapping;
755 struct inode *inode = mapping->host;
756
757 return inode_resv_map(inode);
758
759 } else {
Andy Whitcroft84afd992008-07-23 21:27:32 -0700760 return (struct resv_map *)(get_vma_private_data(vma) &
761 ~HPAGE_RESV_MASK);
Joonsoo Kim4e35f482014-04-03 14:47:30 -0700762 }
Mel Gormana1e78772008-07-23 21:27:23 -0700763}
764
Andy Whitcroft84afd992008-07-23 21:27:32 -0700765static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
Mel Gormana1e78772008-07-23 21:27:23 -0700766{
Sasha Levin81d1b092014-10-09 15:28:10 -0700767 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
768 VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
Mel Gormana1e78772008-07-23 21:27:23 -0700769
Andy Whitcroft84afd992008-07-23 21:27:32 -0700770 set_vma_private_data(vma, (get_vma_private_data(vma) &
771 HPAGE_RESV_MASK) | (unsigned long)map);
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700772}
773
774static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
775{
Sasha Levin81d1b092014-10-09 15:28:10 -0700776 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
777 VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700778
779 set_vma_private_data(vma, get_vma_private_data(vma) | flags);
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700780}
781
782static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
783{
Sasha Levin81d1b092014-10-09 15:28:10 -0700784 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700785
786 return (get_vma_private_data(vma) & flag) != 0;
Mel Gormana1e78772008-07-23 21:27:23 -0700787}
788
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700789/* Reset counters to 0 and clear all HPAGE_RESV_* flags */
Mel Gormana1e78772008-07-23 21:27:23 -0700790void reset_vma_resv_huge_pages(struct vm_area_struct *vma)
791{
Sasha Levin81d1b092014-10-09 15:28:10 -0700792 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
Mel Gormanf83a2752009-05-28 14:34:40 -0700793 if (!(vma->vm_flags & VM_MAYSHARE))
Mel Gormana1e78772008-07-23 21:27:23 -0700794 vma->vm_private_data = (void *)0;
795}
796
797/* Returns true if the VMA has associated reserve pages */
Nicholas Krause559ec2f2015-09-04 15:48:27 -0700798static bool vma_has_reserves(struct vm_area_struct *vma, long chg)
Mel Gormana1e78772008-07-23 21:27:23 -0700799{
Joonsoo Kimaf0ed732013-09-11 14:21:18 -0700800 if (vma->vm_flags & VM_NORESERVE) {
801 /*
802 * This address is already reserved by other process(chg == 0),
803 * so, we should decrement reserved count. Without decrementing,
804 * reserve count remains after releasing inode, because this
805 * allocated page will go into page cache and is regarded as
806 * coming from reserved pool in releasing step. Currently, we
807 * don't have any other solution to deal with this situation
808 * properly, so add work-around here.
809 */
810 if (vma->vm_flags & VM_MAYSHARE && chg == 0)
Nicholas Krause559ec2f2015-09-04 15:48:27 -0700811 return true;
Joonsoo Kimaf0ed732013-09-11 14:21:18 -0700812 else
Nicholas Krause559ec2f2015-09-04 15:48:27 -0700813 return false;
Joonsoo Kimaf0ed732013-09-11 14:21:18 -0700814 }
Joonsoo Kima63884e2013-09-11 14:21:07 -0700815
816 /* Shared mappings always use reserves */
Mike Kravetz1fb1b0e2015-09-08 15:01:44 -0700817 if (vma->vm_flags & VM_MAYSHARE) {
818 /*
819 * We know VM_NORESERVE is not set. Therefore, there SHOULD
820 * be a region map for all pages. The only situation where
821 * there is no region map is if a hole was punched via
822 * fallocate. In this case, there really are no reverves to
823 * use. This situation is indicated if chg != 0.
824 */
825 if (chg)
826 return false;
827 else
828 return true;
829 }
Joonsoo Kima63884e2013-09-11 14:21:07 -0700830
831 /*
832 * Only the process that called mmap() has reserves for
833 * private mappings.
834 */
Mike Kravetz67961f92016-06-08 15:33:42 -0700835 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
836 /*
837 * Like the shared case above, a hole punch or truncate
838 * could have been performed on the private mapping.
839 * Examine the value of chg to determine if reserves
840 * actually exist or were previously consumed.
841 * Very Subtle - The value of chg comes from a previous
842 * call to vma_needs_reserves(). The reserve map for
843 * private mappings has different (opposite) semantics
844 * than that of shared mappings. vma_needs_reserves()
845 * has already taken this difference in semantics into
846 * account. Therefore, the meaning of chg is the same
847 * as in the shared case above. Code could easily be
848 * combined, but keeping it separate draws attention to
849 * subtle differences.
850 */
851 if (chg)
852 return false;
853 else
854 return true;
855 }
Joonsoo Kima63884e2013-09-11 14:21:07 -0700856
Nicholas Krause559ec2f2015-09-04 15:48:27 -0700857 return false;
Mel Gormana1e78772008-07-23 21:27:23 -0700858}
859
Andi Kleena5516432008-07-23 21:27:41 -0700860static void enqueue_huge_page(struct hstate *h, struct page *page)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700861{
862 int nid = page_to_nid(page);
Aneesh Kumar K.V0edaecf2012-07-31 16:42:07 -0700863 list_move(&page->lru, &h->hugepage_freelists[nid]);
Andi Kleena5516432008-07-23 21:27:41 -0700864 h->free_huge_pages++;
865 h->free_huge_pages_node[nid]++;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700866}
867
Naoya Horiguchibf50bab2010-09-08 10:19:33 +0900868static struct page *dequeue_huge_page_node(struct hstate *h, int nid)
869{
870 struct page *page;
871
Naoya Horiguchic8721bb2013-09-11 14:22:09 -0700872 list_for_each_entry(page, &h->hugepage_freelists[nid], lru)
873 if (!is_migrate_isolate_page(page))
874 break;
875 /*
876 * if 'non-isolated free hugepage' not found on the list,
877 * the allocation fails.
878 */
879 if (&h->hugepage_freelists[nid] == &page->lru)
Naoya Horiguchibf50bab2010-09-08 10:19:33 +0900880 return NULL;
Aneesh Kumar K.V0edaecf2012-07-31 16:42:07 -0700881 list_move(&page->lru, &h->hugepage_activelist);
Naoya Horiguchia9869b82010-09-08 10:19:37 +0900882 set_page_refcounted(page);
Naoya Horiguchibf50bab2010-09-08 10:19:33 +0900883 h->free_huge_pages--;
884 h->free_huge_pages_node[nid]--;
885 return page;
886}
887
Naoya Horiguchi86cdb462013-09-11 14:22:13 -0700888/* Movability of hugepages depends on migration support. */
889static inline gfp_t htlb_alloc_mask(struct hstate *h)
890{
Naoya Horiguchi100873d2014-06-04 16:10:56 -0700891 if (hugepages_treat_as_movable || hugepage_migration_supported(h))
Naoya Horiguchi86cdb462013-09-11 14:22:13 -0700892 return GFP_HIGHUSER_MOVABLE;
893 else
894 return GFP_HIGHUSER;
895}
896
Andi Kleena5516432008-07-23 21:27:41 -0700897static struct page *dequeue_huge_page_vma(struct hstate *h,
898 struct vm_area_struct *vma,
Joonsoo Kimaf0ed732013-09-11 14:21:18 -0700899 unsigned long address, int avoid_reserve,
900 long chg)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700901{
Konstantin Khlebnikovb1c12cb2012-04-25 16:01:46 -0700902 struct page *page = NULL;
Lee Schermerhorn480eccf2007-09-18 22:46:47 -0700903 struct mempolicy *mpol;
Mel Gorman19770b32008-04-28 02:12:18 -0700904 nodemask_t *nodemask;
Miao Xiec0ff7452010-05-24 14:32:08 -0700905 struct zonelist *zonelist;
Mel Gormandd1a2392008-04-28 02:12:17 -0700906 struct zone *zone;
907 struct zoneref *z;
Mel Gormancc9a6c82012-03-21 16:34:11 -0700908 unsigned int cpuset_mems_cookie;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700909
Mel Gormana1e78772008-07-23 21:27:23 -0700910 /*
911 * A child process with MAP_PRIVATE mappings created by their parent
912 * have no page reserves. This check ensures that reservations are
913 * not "stolen". The child may still get SIGKILLed
914 */
Joonsoo Kimaf0ed732013-09-11 14:21:18 -0700915 if (!vma_has_reserves(vma, chg) &&
Andi Kleena5516432008-07-23 21:27:41 -0700916 h->free_huge_pages - h->resv_huge_pages == 0)
Miao Xiec0ff7452010-05-24 14:32:08 -0700917 goto err;
Mel Gormana1e78772008-07-23 21:27:23 -0700918
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700919 /* If reserves cannot be used, ensure enough pages are in the pool */
Andi Kleena5516432008-07-23 21:27:41 -0700920 if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
Justin P. Mattock6eab04a2011-04-08 19:49:08 -0700921 goto err;
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700922
Joonsoo Kim9966c4b2013-09-11 14:20:50 -0700923retry_cpuset:
Mel Gormand26914d2014-04-03 14:47:24 -0700924 cpuset_mems_cookie = read_mems_allowed_begin();
Joonsoo Kim9966c4b2013-09-11 14:20:50 -0700925 zonelist = huge_zonelist(vma, address,
Naoya Horiguchi86cdb462013-09-11 14:22:13 -0700926 htlb_alloc_mask(h), &mpol, &nodemask);
Joonsoo Kim9966c4b2013-09-11 14:20:50 -0700927
Mel Gorman19770b32008-04-28 02:12:18 -0700928 for_each_zone_zonelist_nodemask(zone, z, zonelist,
929 MAX_NR_ZONES - 1, nodemask) {
Vladimir Davydov344736f2014-10-20 15:50:30 +0400930 if (cpuset_zone_allowed(zone, htlb_alloc_mask(h))) {
Naoya Horiguchibf50bab2010-09-08 10:19:33 +0900931 page = dequeue_huge_page_node(h, zone_to_nid(zone));
932 if (page) {
Joonsoo Kimaf0ed732013-09-11 14:21:18 -0700933 if (avoid_reserve)
934 break;
935 if (!vma_has_reserves(vma, chg))
936 break;
937
Joonsoo Kim07443a82013-09-11 14:21:58 -0700938 SetPagePrivate(page);
Joonsoo Kimaf0ed732013-09-11 14:21:18 -0700939 h->resv_huge_pages--;
Naoya Horiguchibf50bab2010-09-08 10:19:33 +0900940 break;
941 }
Andrew Morton3abf7af2007-07-19 01:49:08 -0700942 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700943 }
Mel Gormancc9a6c82012-03-21 16:34:11 -0700944
945 mpol_cond_put(mpol);
Mel Gormand26914d2014-04-03 14:47:24 -0700946 if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie)))
Mel Gormancc9a6c82012-03-21 16:34:11 -0700947 goto retry_cpuset;
948 return page;
949
Miao Xiec0ff7452010-05-24 14:32:08 -0700950err:
Mel Gormancc9a6c82012-03-21 16:34:11 -0700951 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700952}
953
Luiz Capitulino1cac6f22014-06-04 16:07:11 -0700954/*
955 * common helper functions for hstate_next_node_to_{alloc|free}.
956 * We may have allocated or freed a huge page based on a different
957 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
958 * be outside of *nodes_allowed. Ensure that we use an allowed
959 * node for alloc or free.
960 */
961static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
962{
Andrew Morton0edaf862016-05-19 17:10:58 -0700963 nid = next_node_in(nid, *nodes_allowed);
Luiz Capitulino1cac6f22014-06-04 16:07:11 -0700964 VM_BUG_ON(nid >= MAX_NUMNODES);
965
966 return nid;
967}
968
969static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
970{
971 if (!node_isset(nid, *nodes_allowed))
972 nid = next_node_allowed(nid, nodes_allowed);
973 return nid;
974}
975
976/*
977 * returns the previously saved node ["this node"] from which to
978 * allocate a persistent huge page for the pool and advance the
979 * next node from which to allocate, handling wrap at end of node
980 * mask.
981 */
982static int hstate_next_node_to_alloc(struct hstate *h,
983 nodemask_t *nodes_allowed)
984{
985 int nid;
986
987 VM_BUG_ON(!nodes_allowed);
988
989 nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
990 h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
991
992 return nid;
993}
994
995/*
996 * helper for free_pool_huge_page() - return the previously saved
997 * node ["this node"] from which to free a huge page. Advance the
998 * next node id whether or not we find a free huge page to free so
999 * that the next attempt to free addresses the next node.
1000 */
1001static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
1002{
1003 int nid;
1004
1005 VM_BUG_ON(!nodes_allowed);
1006
1007 nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
1008 h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
1009
1010 return nid;
1011}
1012
1013#define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \
1014 for (nr_nodes = nodes_weight(*mask); \
1015 nr_nodes > 0 && \
1016 ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \
1017 nr_nodes--)
1018
1019#define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \
1020 for (nr_nodes = nodes_weight(*mask); \
1021 nr_nodes > 0 && \
1022 ((node = hstate_next_node_to_free(hs, mask)) || 1); \
1023 nr_nodes--)
1024
Gerald Schaeferd08de8e2016-07-04 14:47:01 +02001025#if (defined(CONFIG_X86_64) || defined(CONFIG_S390)) && \
1026 ((defined(CONFIG_MEMORY_ISOLATION) && defined(CONFIG_COMPACTION)) || \
1027 defined(CONFIG_CMA))
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001028static void destroy_compound_gigantic_page(struct page *page,
Kirill A. Shutemovd00181b2015-11-06 16:29:57 -08001029 unsigned int order)
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001030{
1031 int i;
1032 int nr_pages = 1 << order;
1033 struct page *p = page + 1;
1034
Gerald Schaeferc8cc7082016-06-24 14:50:04 -07001035 atomic_set(compound_mapcount_ptr(page), 0);
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001036 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
Kirill A. Shutemov1d798ca2015-11-06 16:29:54 -08001037 clear_compound_head(p);
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001038 set_page_refcounted(p);
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001039 }
1040
1041 set_compound_order(page, 0);
1042 __ClearPageHead(page);
1043}
1044
Kirill A. Shutemovd00181b2015-11-06 16:29:57 -08001045static void free_gigantic_page(struct page *page, unsigned int order)
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001046{
1047 free_contig_range(page_to_pfn(page), 1 << order);
1048}
1049
1050static int __alloc_gigantic_page(unsigned long start_pfn,
1051 unsigned long nr_pages)
1052{
1053 unsigned long end_pfn = start_pfn + nr_pages;
1054 return alloc_contig_range(start_pfn, end_pfn, MIGRATE_MOVABLE);
1055}
1056
Joonsoo Kimf44b2dda82016-05-19 17:12:03 -07001057static bool pfn_range_valid_gigantic(struct zone *z,
1058 unsigned long start_pfn, unsigned long nr_pages)
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001059{
1060 unsigned long i, end_pfn = start_pfn + nr_pages;
1061 struct page *page;
1062
1063 for (i = start_pfn; i < end_pfn; i++) {
1064 if (!pfn_valid(i))
1065 return false;
1066
1067 page = pfn_to_page(i);
1068
Joonsoo Kimf44b2dda82016-05-19 17:12:03 -07001069 if (page_zone(page) != z)
1070 return false;
1071
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001072 if (PageReserved(page))
1073 return false;
1074
1075 if (page_count(page) > 0)
1076 return false;
1077
1078 if (PageHuge(page))
1079 return false;
1080 }
1081
1082 return true;
1083}
1084
1085static bool zone_spans_last_pfn(const struct zone *zone,
1086 unsigned long start_pfn, unsigned long nr_pages)
1087{
1088 unsigned long last_pfn = start_pfn + nr_pages - 1;
1089 return zone_spans_pfn(zone, last_pfn);
1090}
1091
Kirill A. Shutemovd00181b2015-11-06 16:29:57 -08001092static struct page *alloc_gigantic_page(int nid, unsigned int order)
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001093{
1094 unsigned long nr_pages = 1 << order;
1095 unsigned long ret, pfn, flags;
1096 struct zone *z;
1097
1098 z = NODE_DATA(nid)->node_zones;
1099 for (; z - NODE_DATA(nid)->node_zones < MAX_NR_ZONES; z++) {
1100 spin_lock_irqsave(&z->lock, flags);
1101
1102 pfn = ALIGN(z->zone_start_pfn, nr_pages);
1103 while (zone_spans_last_pfn(z, pfn, nr_pages)) {
Joonsoo Kimf44b2dda82016-05-19 17:12:03 -07001104 if (pfn_range_valid_gigantic(z, pfn, nr_pages)) {
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001105 /*
1106 * We release the zone lock here because
1107 * alloc_contig_range() will also lock the zone
1108 * at some point. If there's an allocation
1109 * spinning on this lock, it may win the race
1110 * and cause alloc_contig_range() to fail...
1111 */
1112 spin_unlock_irqrestore(&z->lock, flags);
1113 ret = __alloc_gigantic_page(pfn, nr_pages);
1114 if (!ret)
1115 return pfn_to_page(pfn);
1116 spin_lock_irqsave(&z->lock, flags);
1117 }
1118 pfn += nr_pages;
1119 }
1120
1121 spin_unlock_irqrestore(&z->lock, flags);
1122 }
1123
1124 return NULL;
1125}
1126
1127static void prep_new_huge_page(struct hstate *h, struct page *page, int nid);
Kirill A. Shutemovd00181b2015-11-06 16:29:57 -08001128static void prep_compound_gigantic_page(struct page *page, unsigned int order);
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001129
1130static struct page *alloc_fresh_gigantic_page_node(struct hstate *h, int nid)
1131{
1132 struct page *page;
1133
1134 page = alloc_gigantic_page(nid, huge_page_order(h));
1135 if (page) {
1136 prep_compound_gigantic_page(page, huge_page_order(h));
1137 prep_new_huge_page(h, page, nid);
1138 }
1139
1140 return page;
1141}
1142
1143static int alloc_fresh_gigantic_page(struct hstate *h,
1144 nodemask_t *nodes_allowed)
1145{
1146 struct page *page = NULL;
1147 int nr_nodes, node;
1148
1149 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
1150 page = alloc_fresh_gigantic_page_node(h, node);
1151 if (page)
1152 return 1;
1153 }
1154
1155 return 0;
1156}
1157
1158static inline bool gigantic_page_supported(void) { return true; }
1159#else
1160static inline bool gigantic_page_supported(void) { return false; }
Kirill A. Shutemovd00181b2015-11-06 16:29:57 -08001161static inline void free_gigantic_page(struct page *page, unsigned int order) { }
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001162static inline void destroy_compound_gigantic_page(struct page *page,
Kirill A. Shutemovd00181b2015-11-06 16:29:57 -08001163 unsigned int order) { }
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001164static inline int alloc_fresh_gigantic_page(struct hstate *h,
1165 nodemask_t *nodes_allowed) { return 0; }
1166#endif
1167
Andi Kleena5516432008-07-23 21:27:41 -07001168static void update_and_free_page(struct hstate *h, struct page *page)
Adam Litke6af2acb2007-10-16 01:26:16 -07001169{
1170 int i;
Andi Kleena5516432008-07-23 21:27:41 -07001171
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001172 if (hstate_is_gigantic(h) && !gigantic_page_supported())
1173 return;
Andy Whitcroft18229df2008-11-06 12:53:27 -08001174
Andi Kleena5516432008-07-23 21:27:41 -07001175 h->nr_huge_pages--;
1176 h->nr_huge_pages_node[page_to_nid(page)]--;
1177 for (i = 0; i < pages_per_huge_page(h); i++) {
Chris Forbes32f84522011-07-25 17:12:14 -07001178 page[i].flags &= ~(1 << PG_locked | 1 << PG_error |
1179 1 << PG_referenced | 1 << PG_dirty |
Luiz Capitulinoa7407a22014-06-04 16:07:09 -07001180 1 << PG_active | 1 << PG_private |
1181 1 << PG_writeback);
Adam Litke6af2acb2007-10-16 01:26:16 -07001182 }
Sasha Levin309381fea2014-01-23 15:52:54 -08001183 VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page);
Kirill A. Shutemovf1e61552015-11-06 16:29:50 -08001184 set_compound_page_dtor(page, NULL_COMPOUND_DTOR);
Adam Litke6af2acb2007-10-16 01:26:16 -07001185 set_page_refcounted(page);
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001186 if (hstate_is_gigantic(h)) {
1187 destroy_compound_gigantic_page(page, huge_page_order(h));
1188 free_gigantic_page(page, huge_page_order(h));
1189 } else {
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001190 __free_pages(page, huge_page_order(h));
1191 }
Adam Litke6af2acb2007-10-16 01:26:16 -07001192}
1193
Andi Kleene5ff2152008-07-23 21:27:42 -07001194struct hstate *size_to_hstate(unsigned long size)
1195{
1196 struct hstate *h;
1197
1198 for_each_hstate(h) {
1199 if (huge_page_size(h) == size)
1200 return h;
1201 }
1202 return NULL;
1203}
1204
Naoya Horiguchibcc54222015-04-15 16:14:38 -07001205/*
1206 * Test to determine whether the hugepage is "active/in-use" (i.e. being linked
1207 * to hstate->hugepage_activelist.)
1208 *
1209 * This function can be called for tail pages, but never returns true for them.
1210 */
1211bool page_huge_active(struct page *page)
1212{
1213 VM_BUG_ON_PAGE(!PageHuge(page), page);
1214 return PageHead(page) && PagePrivate(&page[1]);
1215}
1216
1217/* never called for tail page */
1218static void set_page_huge_active(struct page *page)
1219{
1220 VM_BUG_ON_PAGE(!PageHeadHuge(page), page);
1221 SetPagePrivate(&page[1]);
1222}
1223
1224static void clear_page_huge_active(struct page *page)
1225{
1226 VM_BUG_ON_PAGE(!PageHeadHuge(page), page);
1227 ClearPagePrivate(&page[1]);
1228}
1229
Atsushi Kumagai8f1d26d2014-07-30 16:08:39 -07001230void free_huge_page(struct page *page)
David Gibson27a85ef2006-03-22 00:08:56 -08001231{
Andi Kleena5516432008-07-23 21:27:41 -07001232 /*
1233 * Can't pass hstate in here because it is called from the
1234 * compound page destructor.
1235 */
Andi Kleene5ff2152008-07-23 21:27:42 -07001236 struct hstate *h = page_hstate(page);
Adam Litke7893d1d2007-10-16 01:26:18 -07001237 int nid = page_to_nid(page);
David Gibson90481622012-03-21 16:34:12 -07001238 struct hugepage_subpool *spool =
1239 (struct hugepage_subpool *)page_private(page);
Joonsoo Kim07443a82013-09-11 14:21:58 -07001240 bool restore_reserve;
David Gibson27a85ef2006-03-22 00:08:56 -08001241
Andy Whitcrofte5df70a2008-02-23 15:23:32 -08001242 set_page_private(page, 0);
Mel Gorman23be7462010-04-23 13:17:56 -04001243 page->mapping = NULL;
Mike Kravetzb4330af2016-02-05 15:36:38 -08001244 VM_BUG_ON_PAGE(page_count(page), page);
1245 VM_BUG_ON_PAGE(page_mapcount(page), page);
Joonsoo Kim07443a82013-09-11 14:21:58 -07001246 restore_reserve = PagePrivate(page);
Joonsoo Kim16c794b2013-10-16 13:46:48 -07001247 ClearPagePrivate(page);
David Gibson27a85ef2006-03-22 00:08:56 -08001248
Mike Kravetz1c5ecae2015-04-15 16:13:39 -07001249 /*
1250 * A return code of zero implies that the subpool will be under its
1251 * minimum size if the reservation is not restored after page is free.
1252 * Therefore, force restore_reserve operation.
1253 */
1254 if (hugepage_subpool_put_pages(spool, 1) == 0)
1255 restore_reserve = true;
1256
David Gibson27a85ef2006-03-22 00:08:56 -08001257 spin_lock(&hugetlb_lock);
Naoya Horiguchibcc54222015-04-15 16:14:38 -07001258 clear_page_huge_active(page);
Aneesh Kumar K.V6d76dcf2012-07-31 16:42:18 -07001259 hugetlb_cgroup_uncharge_page(hstate_index(h),
1260 pages_per_huge_page(h), page);
Joonsoo Kim07443a82013-09-11 14:21:58 -07001261 if (restore_reserve)
1262 h->resv_huge_pages++;
1263
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001264 if (h->surplus_huge_pages_node[nid]) {
Aneesh Kumar K.V0edaecf2012-07-31 16:42:07 -07001265 /* remove the page from active list */
1266 list_del(&page->lru);
Andi Kleena5516432008-07-23 21:27:41 -07001267 update_and_free_page(h, page);
1268 h->surplus_huge_pages--;
1269 h->surplus_huge_pages_node[nid]--;
Adam Litke7893d1d2007-10-16 01:26:18 -07001270 } else {
Will Deacon5d3a5512012-10-08 16:29:32 -07001271 arch_clear_hugepage_flags(page);
Andi Kleena5516432008-07-23 21:27:41 -07001272 enqueue_huge_page(h, page);
Adam Litke7893d1d2007-10-16 01:26:18 -07001273 }
David Gibson27a85ef2006-03-22 00:08:56 -08001274 spin_unlock(&hugetlb_lock);
1275}
1276
Andi Kleena5516432008-07-23 21:27:41 -07001277static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
Andi Kleenb7ba30c2008-07-23 21:27:40 -07001278{
Aneesh Kumar K.V0edaecf2012-07-31 16:42:07 -07001279 INIT_LIST_HEAD(&page->lru);
Kirill A. Shutemovf1e61552015-11-06 16:29:50 -08001280 set_compound_page_dtor(page, HUGETLB_PAGE_DTOR);
Andi Kleenb7ba30c2008-07-23 21:27:40 -07001281 spin_lock(&hugetlb_lock);
Aneesh Kumar K.V9dd540e2012-07-31 16:42:15 -07001282 set_hugetlb_cgroup(page, NULL);
Andi Kleena5516432008-07-23 21:27:41 -07001283 h->nr_huge_pages++;
1284 h->nr_huge_pages_node[nid]++;
Andi Kleenb7ba30c2008-07-23 21:27:40 -07001285 spin_unlock(&hugetlb_lock);
1286 put_page(page); /* free it into the hugepage allocator */
1287}
1288
Kirill A. Shutemovd00181b2015-11-06 16:29:57 -08001289static void prep_compound_gigantic_page(struct page *page, unsigned int order)
Wu Fengguang20a03072009-06-16 15:32:22 -07001290{
1291 int i;
1292 int nr_pages = 1 << order;
1293 struct page *p = page + 1;
1294
1295 /* we rely on prep_new_huge_page to set the destructor */
1296 set_compound_order(page, order);
Andrea Arcangelief5a22b2013-10-16 13:46:56 -07001297 __ClearPageReserved(page);
Kirill A. Shutemovde09d312016-01-15 16:51:42 -08001298 __SetPageHead(page);
Wu Fengguang20a03072009-06-16 15:32:22 -07001299 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
Andrea Arcangelief5a22b2013-10-16 13:46:56 -07001300 /*
1301 * For gigantic hugepages allocated through bootmem at
1302 * boot, it's safer to be consistent with the not-gigantic
1303 * hugepages and clear the PG_reserved bit from all tail pages
1304 * too. Otherwse drivers using get_user_pages() to access tail
1305 * pages may get the reference counting wrong if they see
1306 * PG_reserved set on a tail page (despite the head page not
1307 * having PG_reserved set). Enforcing this consistency between
1308 * head and tail pages allows drivers to optimize away a check
1309 * on the head page when they need know if put_page() is needed
1310 * after get_user_pages().
1311 */
1312 __ClearPageReserved(p);
Youquan Song58a84aa2011-12-08 14:34:18 -08001313 set_page_count(p, 0);
Kirill A. Shutemov1d798ca2015-11-06 16:29:54 -08001314 set_compound_head(p, page);
Wu Fengguang20a03072009-06-16 15:32:22 -07001315 }
Mike Kravetzb4330af2016-02-05 15:36:38 -08001316 atomic_set(compound_mapcount_ptr(page), -1);
Wu Fengguang20a03072009-06-16 15:32:22 -07001317}
1318
Andrew Morton77959122012-10-08 16:34:11 -07001319/*
1320 * PageHuge() only returns true for hugetlbfs pages, but not for normal or
1321 * transparent huge pages. See the PageTransHuge() documentation for more
1322 * details.
1323 */
Wu Fengguang20a03072009-06-16 15:32:22 -07001324int PageHuge(struct page *page)
1325{
Wu Fengguang20a03072009-06-16 15:32:22 -07001326 if (!PageCompound(page))
1327 return 0;
1328
1329 page = compound_head(page);
Kirill A. Shutemovf1e61552015-11-06 16:29:50 -08001330 return page[1].compound_dtor == HUGETLB_PAGE_DTOR;
Wu Fengguang20a03072009-06-16 15:32:22 -07001331}
Naoya Horiguchi43131e12010-05-28 09:29:22 +09001332EXPORT_SYMBOL_GPL(PageHuge);
1333
Andrea Arcangeli27c73ae2013-11-21 14:32:02 -08001334/*
1335 * PageHeadHuge() only returns true for hugetlbfs head page, but not for
1336 * normal or transparent huge pages.
1337 */
1338int PageHeadHuge(struct page *page_head)
1339{
Andrea Arcangeli27c73ae2013-11-21 14:32:02 -08001340 if (!PageHead(page_head))
1341 return 0;
1342
Andrew Morton758f66a2014-01-21 15:48:57 -08001343 return get_compound_page_dtor(page_head) == free_huge_page;
Andrea Arcangeli27c73ae2013-11-21 14:32:02 -08001344}
Andrea Arcangeli27c73ae2013-11-21 14:32:02 -08001345
Zhang Yi13d60f42013-06-25 21:19:31 +08001346pgoff_t __basepage_index(struct page *page)
1347{
1348 struct page *page_head = compound_head(page);
1349 pgoff_t index = page_index(page_head);
1350 unsigned long compound_idx;
1351
1352 if (!PageHuge(page_head))
1353 return page_index(page);
1354
1355 if (compound_order(page_head) >= MAX_ORDER)
1356 compound_idx = page_to_pfn(page) - page_to_pfn(page_head);
1357 else
1358 compound_idx = page - page_head;
1359
1360 return (index << compound_order(page_head)) + compound_idx;
1361}
1362
Andi Kleena5516432008-07-23 21:27:41 -07001363static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001364{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001365 struct page *page;
Joe Jinf96efd52007-07-15 23:38:12 -07001366
Vlastimil Babka96db8002015-09-08 15:03:50 -07001367 page = __alloc_pages_node(nid,
Naoya Horiguchi86cdb462013-09-11 14:22:13 -07001368 htlb_alloc_mask(h)|__GFP_COMP|__GFP_THISNODE|
Nishanth Aravamudan551883a2008-04-29 00:58:26 -07001369 __GFP_REPEAT|__GFP_NOWARN,
Andi Kleena5516432008-07-23 21:27:41 -07001370 huge_page_order(h));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001371 if (page) {
Andi Kleena5516432008-07-23 21:27:41 -07001372 prep_new_huge_page(h, page, nid);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001373 }
Nishanth Aravamudan63b46132007-10-16 01:26:24 -07001374
1375 return page;
1376}
1377
Joonsoo Kimb2261022013-09-11 14:21:00 -07001378static int alloc_fresh_huge_page(struct hstate *h, nodemask_t *nodes_allowed)
1379{
1380 struct page *page;
1381 int nr_nodes, node;
1382 int ret = 0;
1383
1384 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
1385 page = alloc_fresh_huge_page_node(h, node);
1386 if (page) {
1387 ret = 1;
1388 break;
1389 }
1390 }
1391
1392 if (ret)
1393 count_vm_event(HTLB_BUDDY_PGALLOC);
1394 else
1395 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
1396
1397 return ret;
1398}
1399
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001400/*
1401 * Free huge page from pool from next node to free.
1402 * Attempt to keep persistent huge pages more or less
1403 * balanced over allowed nodes.
1404 * Called with hugetlb_lock locked.
1405 */
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001406static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
1407 bool acct_surplus)
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001408{
Joonsoo Kimb2261022013-09-11 14:21:00 -07001409 int nr_nodes, node;
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001410 int ret = 0;
1411
Joonsoo Kimb2261022013-09-11 14:21:00 -07001412 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
Lee Schermerhorn685f3452009-09-21 17:01:23 -07001413 /*
1414 * If we're returning unused surplus pages, only examine
1415 * nodes with surplus pages.
1416 */
Joonsoo Kimb2261022013-09-11 14:21:00 -07001417 if ((!acct_surplus || h->surplus_huge_pages_node[node]) &&
1418 !list_empty(&h->hugepage_freelists[node])) {
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001419 struct page *page =
Joonsoo Kimb2261022013-09-11 14:21:00 -07001420 list_entry(h->hugepage_freelists[node].next,
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001421 struct page, lru);
1422 list_del(&page->lru);
1423 h->free_huge_pages--;
Joonsoo Kimb2261022013-09-11 14:21:00 -07001424 h->free_huge_pages_node[node]--;
Lee Schermerhorn685f3452009-09-21 17:01:23 -07001425 if (acct_surplus) {
1426 h->surplus_huge_pages--;
Joonsoo Kimb2261022013-09-11 14:21:00 -07001427 h->surplus_huge_pages_node[node]--;
Lee Schermerhorn685f3452009-09-21 17:01:23 -07001428 }
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001429 update_and_free_page(h, page);
1430 ret = 1;
Lee Schermerhorn9a76db02009-12-14 17:58:15 -08001431 break;
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001432 }
Joonsoo Kimb2261022013-09-11 14:21:00 -07001433 }
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001434
1435 return ret;
1436}
1437
Naoya Horiguchic8721bb2013-09-11 14:22:09 -07001438/*
1439 * Dissolve a given free hugepage into free buddy pages. This function does
1440 * nothing for in-use (including surplus) hugepages.
1441 */
1442static void dissolve_free_huge_page(struct page *page)
1443{
1444 spin_lock(&hugetlb_lock);
1445 if (PageHuge(page) && !page_count(page)) {
1446 struct hstate *h = page_hstate(page);
1447 int nid = page_to_nid(page);
1448 list_del(&page->lru);
1449 h->free_huge_pages--;
1450 h->free_huge_pages_node[nid]--;
1451 update_and_free_page(h, page);
1452 }
1453 spin_unlock(&hugetlb_lock);
1454}
1455
1456/*
1457 * Dissolve free hugepages in a given pfn range. Used by memory hotplug to
1458 * make specified memory blocks removable from the system.
1459 * Note that start_pfn should aligned with (minimum) hugepage size.
1460 */
1461void dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn)
1462{
Naoya Horiguchic8721bb2013-09-11 14:22:09 -07001463 unsigned long pfn;
Naoya Horiguchic8721bb2013-09-11 14:22:09 -07001464
Li Zhongd0177632014-08-06 16:07:56 -07001465 if (!hugepages_supported())
1466 return;
1467
Naoya Horiguchi641844f2015-06-24 16:56:59 -07001468 VM_BUG_ON(!IS_ALIGNED(start_pfn, 1 << minimum_order));
1469 for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << minimum_order)
Naoya Horiguchic8721bb2013-09-11 14:22:09 -07001470 dissolve_free_huge_page(pfn_to_page(pfn));
1471}
1472
Dave Hansen099730d2015-11-05 18:50:17 -08001473/*
1474 * There are 3 ways this can get called:
1475 * 1. With vma+addr: we use the VMA's memory policy
1476 * 2. With !vma, but nid=NUMA_NO_NODE: We try to allocate a huge
1477 * page from any node, and let the buddy allocator itself figure
1478 * it out.
1479 * 3. With !vma, but nid!=NUMA_NO_NODE. We allocate a huge page
1480 * strictly from 'nid'
1481 */
1482static struct page *__hugetlb_alloc_buddy_huge_page(struct hstate *h,
1483 struct vm_area_struct *vma, unsigned long addr, int nid)
1484{
1485 int order = huge_page_order(h);
1486 gfp_t gfp = htlb_alloc_mask(h)|__GFP_COMP|__GFP_REPEAT|__GFP_NOWARN;
1487 unsigned int cpuset_mems_cookie;
1488
1489 /*
1490 * We need a VMA to get a memory policy. If we do not
Dave Hansene0ec90e2015-11-05 18:50:20 -08001491 * have one, we use the 'nid' argument.
1492 *
1493 * The mempolicy stuff below has some non-inlined bits
1494 * and calls ->vm_ops. That makes it hard to optimize at
1495 * compile-time, even when NUMA is off and it does
1496 * nothing. This helps the compiler optimize it out.
Dave Hansen099730d2015-11-05 18:50:17 -08001497 */
Dave Hansene0ec90e2015-11-05 18:50:20 -08001498 if (!IS_ENABLED(CONFIG_NUMA) || !vma) {
Dave Hansen099730d2015-11-05 18:50:17 -08001499 /*
1500 * If a specific node is requested, make sure to
1501 * get memory from there, but only when a node
1502 * is explicitly specified.
1503 */
1504 if (nid != NUMA_NO_NODE)
1505 gfp |= __GFP_THISNODE;
1506 /*
1507 * Make sure to call something that can handle
1508 * nid=NUMA_NO_NODE
1509 */
1510 return alloc_pages_node(nid, gfp, order);
1511 }
1512
1513 /*
1514 * OK, so we have a VMA. Fetch the mempolicy and try to
Dave Hansene0ec90e2015-11-05 18:50:20 -08001515 * allocate a huge page with it. We will only reach this
1516 * when CONFIG_NUMA=y.
Dave Hansen099730d2015-11-05 18:50:17 -08001517 */
1518 do {
1519 struct page *page;
1520 struct mempolicy *mpol;
1521 struct zonelist *zl;
1522 nodemask_t *nodemask;
1523
1524 cpuset_mems_cookie = read_mems_allowed_begin();
1525 zl = huge_zonelist(vma, addr, gfp, &mpol, &nodemask);
1526 mpol_cond_put(mpol);
1527 page = __alloc_pages_nodemask(gfp, order, zl, nodemask);
1528 if (page)
1529 return page;
1530 } while (read_mems_allowed_retry(cpuset_mems_cookie));
1531
1532 return NULL;
1533}
1534
1535/*
1536 * There are two ways to allocate a huge page:
1537 * 1. When you have a VMA and an address (like a fault)
1538 * 2. When you have no VMA (like when setting /proc/.../nr_hugepages)
1539 *
1540 * 'vma' and 'addr' are only for (1). 'nid' is always NUMA_NO_NODE in
1541 * this case which signifies that the allocation should be done with
1542 * respect for the VMA's memory policy.
1543 *
1544 * For (2), we ignore 'vma' and 'addr' and use 'nid' exclusively. This
1545 * implies that memory policies will not be taken in to account.
1546 */
1547static struct page *__alloc_buddy_huge_page(struct hstate *h,
1548 struct vm_area_struct *vma, unsigned long addr, int nid)
Adam Litke7893d1d2007-10-16 01:26:18 -07001549{
1550 struct page *page;
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001551 unsigned int r_nid;
Adam Litke7893d1d2007-10-16 01:26:18 -07001552
Luiz Capitulinobae7f4a2014-06-04 16:07:08 -07001553 if (hstate_is_gigantic(h))
Andi Kleenaa888a72008-07-23 21:27:47 -07001554 return NULL;
1555
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001556 /*
Dave Hansen099730d2015-11-05 18:50:17 -08001557 * Make sure that anyone specifying 'nid' is not also specifying a VMA.
1558 * This makes sure the caller is picking _one_ of the modes with which
1559 * we can call this function, not both.
1560 */
1561 if (vma || (addr != -1)) {
Dave Hansene0ec90e2015-11-05 18:50:20 -08001562 VM_WARN_ON_ONCE(addr == -1);
1563 VM_WARN_ON_ONCE(nid != NUMA_NO_NODE);
Dave Hansen099730d2015-11-05 18:50:17 -08001564 }
1565 /*
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001566 * Assume we will successfully allocate the surplus page to
1567 * prevent racing processes from causing the surplus to exceed
1568 * overcommit
1569 *
1570 * This however introduces a different race, where a process B
1571 * tries to grow the static hugepage pool while alloc_pages() is
1572 * called by process A. B will only examine the per-node
1573 * counters in determining if surplus huge pages can be
1574 * converted to normal huge pages in adjust_pool_surplus(). A
1575 * won't be able to increment the per-node counter, until the
1576 * lock is dropped by B, but B doesn't drop hugetlb_lock until
1577 * no more huge pages can be converted from surplus to normal
1578 * state (and doesn't try to convert again). Thus, we have a
1579 * case where a surplus huge page exists, the pool is grown, and
1580 * the surplus huge page still exists after, even though it
1581 * should just have been converted to a normal huge page. This
1582 * does not leak memory, though, as the hugepage will be freed
1583 * once it is out of use. It also does not allow the counters to
1584 * go out of whack in adjust_pool_surplus() as we don't modify
1585 * the node values until we've gotten the hugepage and only the
1586 * per-node value is checked there.
1587 */
1588 spin_lock(&hugetlb_lock);
Andi Kleena5516432008-07-23 21:27:41 -07001589 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001590 spin_unlock(&hugetlb_lock);
1591 return NULL;
1592 } else {
Andi Kleena5516432008-07-23 21:27:41 -07001593 h->nr_huge_pages++;
1594 h->surplus_huge_pages++;
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001595 }
1596 spin_unlock(&hugetlb_lock);
1597
Dave Hansen099730d2015-11-05 18:50:17 -08001598 page = __hugetlb_alloc_buddy_huge_page(h, vma, addr, nid);
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001599
1600 spin_lock(&hugetlb_lock);
Adam Litke7893d1d2007-10-16 01:26:18 -07001601 if (page) {
Aneesh Kumar K.V0edaecf2012-07-31 16:42:07 -07001602 INIT_LIST_HEAD(&page->lru);
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001603 r_nid = page_to_nid(page);
Kirill A. Shutemovf1e61552015-11-06 16:29:50 -08001604 set_compound_page_dtor(page, HUGETLB_PAGE_DTOR);
Aneesh Kumar K.V9dd540e2012-07-31 16:42:15 -07001605 set_hugetlb_cgroup(page, NULL);
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001606 /*
1607 * We incremented the global counters already
1608 */
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001609 h->nr_huge_pages_node[r_nid]++;
1610 h->surplus_huge_pages_node[r_nid]++;
Adam Litke3b116302008-04-28 02:13:06 -07001611 __count_vm_event(HTLB_BUDDY_PGALLOC);
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001612 } else {
Andi Kleena5516432008-07-23 21:27:41 -07001613 h->nr_huge_pages--;
1614 h->surplus_huge_pages--;
Adam Litke3b116302008-04-28 02:13:06 -07001615 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
Adam Litke7893d1d2007-10-16 01:26:18 -07001616 }
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001617 spin_unlock(&hugetlb_lock);
Adam Litke7893d1d2007-10-16 01:26:18 -07001618
1619 return page;
1620}
1621
Adam Litkee4e574b2007-10-16 01:26:19 -07001622/*
Dave Hansen099730d2015-11-05 18:50:17 -08001623 * Allocate a huge page from 'nid'. Note, 'nid' may be
1624 * NUMA_NO_NODE, which means that it may be allocated
1625 * anywhere.
1626 */
Dave Hansene0ec90e2015-11-05 18:50:20 -08001627static
Dave Hansen099730d2015-11-05 18:50:17 -08001628struct page *__alloc_buddy_huge_page_no_mpol(struct hstate *h, int nid)
1629{
1630 unsigned long addr = -1;
1631
1632 return __alloc_buddy_huge_page(h, NULL, addr, nid);
1633}
1634
1635/*
1636 * Use the VMA's mpolicy to allocate a huge page from the buddy.
1637 */
Dave Hansene0ec90e2015-11-05 18:50:20 -08001638static
Dave Hansen099730d2015-11-05 18:50:17 -08001639struct page *__alloc_buddy_huge_page_with_mpol(struct hstate *h,
1640 struct vm_area_struct *vma, unsigned long addr)
1641{
1642 return __alloc_buddy_huge_page(h, vma, addr, NUMA_NO_NODE);
1643}
1644
1645/*
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001646 * This allocation function is useful in the context where vma is irrelevant.
1647 * E.g. soft-offlining uses this function because it only cares physical
1648 * address of error page.
1649 */
1650struct page *alloc_huge_page_node(struct hstate *h, int nid)
1651{
Joonsoo Kim4ef91842013-09-11 14:21:51 -07001652 struct page *page = NULL;
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001653
1654 spin_lock(&hugetlb_lock);
Joonsoo Kim4ef91842013-09-11 14:21:51 -07001655 if (h->free_huge_pages - h->resv_huge_pages > 0)
1656 page = dequeue_huge_page_node(h, nid);
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001657 spin_unlock(&hugetlb_lock);
1658
Aneesh Kumar K.V94ae8ba2012-07-31 16:42:35 -07001659 if (!page)
Dave Hansen099730d2015-11-05 18:50:17 -08001660 page = __alloc_buddy_huge_page_no_mpol(h, nid);
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001661
1662 return page;
1663}
1664
1665/*
Lucas De Marchi25985ed2011-03-30 22:57:33 -03001666 * Increase the hugetlb pool such that it can accommodate a reservation
Adam Litkee4e574b2007-10-16 01:26:19 -07001667 * of size 'delta'.
1668 */
Andi Kleena5516432008-07-23 21:27:41 -07001669static int gather_surplus_pages(struct hstate *h, int delta)
Adam Litkee4e574b2007-10-16 01:26:19 -07001670{
1671 struct list_head surplus_list;
1672 struct page *page, *tmp;
1673 int ret, i;
1674 int needed, allocated;
Hillf Danton28073b02012-03-21 16:34:00 -07001675 bool alloc_ok = true;
Adam Litkee4e574b2007-10-16 01:26:19 -07001676
Andi Kleena5516432008-07-23 21:27:41 -07001677 needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
Adam Litkeac09b3a2008-03-04 14:29:38 -08001678 if (needed <= 0) {
Andi Kleena5516432008-07-23 21:27:41 -07001679 h->resv_huge_pages += delta;
Adam Litkee4e574b2007-10-16 01:26:19 -07001680 return 0;
Adam Litkeac09b3a2008-03-04 14:29:38 -08001681 }
Adam Litkee4e574b2007-10-16 01:26:19 -07001682
1683 allocated = 0;
1684 INIT_LIST_HEAD(&surplus_list);
1685
1686 ret = -ENOMEM;
1687retry:
1688 spin_unlock(&hugetlb_lock);
1689 for (i = 0; i < needed; i++) {
Dave Hansen099730d2015-11-05 18:50:17 -08001690 page = __alloc_buddy_huge_page_no_mpol(h, NUMA_NO_NODE);
Hillf Danton28073b02012-03-21 16:34:00 -07001691 if (!page) {
1692 alloc_ok = false;
1693 break;
1694 }
Adam Litkee4e574b2007-10-16 01:26:19 -07001695 list_add(&page->lru, &surplus_list);
1696 }
Hillf Danton28073b02012-03-21 16:34:00 -07001697 allocated += i;
Adam Litkee4e574b2007-10-16 01:26:19 -07001698
1699 /*
1700 * After retaking hugetlb_lock, we need to recalculate 'needed'
1701 * because either resv_huge_pages or free_huge_pages may have changed.
1702 */
1703 spin_lock(&hugetlb_lock);
Andi Kleena5516432008-07-23 21:27:41 -07001704 needed = (h->resv_huge_pages + delta) -
1705 (h->free_huge_pages + allocated);
Hillf Danton28073b02012-03-21 16:34:00 -07001706 if (needed > 0) {
1707 if (alloc_ok)
1708 goto retry;
1709 /*
1710 * We were not able to allocate enough pages to
1711 * satisfy the entire reservation so we free what
1712 * we've allocated so far.
1713 */
1714 goto free;
1715 }
Adam Litkee4e574b2007-10-16 01:26:19 -07001716 /*
1717 * The surplus_list now contains _at_least_ the number of extra pages
Lucas De Marchi25985ed2011-03-30 22:57:33 -03001718 * needed to accommodate the reservation. Add the appropriate number
Adam Litkee4e574b2007-10-16 01:26:19 -07001719 * of pages to the hugetlb pool and free the extras back to the buddy
Adam Litkeac09b3a2008-03-04 14:29:38 -08001720 * allocator. Commit the entire reservation here to prevent another
1721 * process from stealing the pages as they are added to the pool but
1722 * before they are reserved.
Adam Litkee4e574b2007-10-16 01:26:19 -07001723 */
1724 needed += allocated;
Andi Kleena5516432008-07-23 21:27:41 -07001725 h->resv_huge_pages += delta;
Adam Litkee4e574b2007-10-16 01:26:19 -07001726 ret = 0;
Naoya Horiguchia9869b82010-09-08 10:19:37 +09001727
Adam Litke19fc3f02008-04-28 02:12:20 -07001728 /* Free the needed pages to the hugetlb pool */
Adam Litkee4e574b2007-10-16 01:26:19 -07001729 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
Adam Litke19fc3f02008-04-28 02:12:20 -07001730 if ((--needed) < 0)
1731 break;
Naoya Horiguchia9869b82010-09-08 10:19:37 +09001732 /*
1733 * This page is now managed by the hugetlb allocator and has
1734 * no users -- drop the buddy allocator's reference.
1735 */
1736 put_page_testzero(page);
Sasha Levin309381fea2014-01-23 15:52:54 -08001737 VM_BUG_ON_PAGE(page_count(page), page);
Andi Kleena5516432008-07-23 21:27:41 -07001738 enqueue_huge_page(h, page);
Adam Litke19fc3f02008-04-28 02:12:20 -07001739 }
Hillf Danton28073b02012-03-21 16:34:00 -07001740free:
Hillf Dantonb0365c82011-12-28 15:57:16 -08001741 spin_unlock(&hugetlb_lock);
Adam Litke19fc3f02008-04-28 02:12:20 -07001742
1743 /* Free unnecessary surplus pages to the buddy allocator */
Joonsoo Kimc0d934b2013-09-11 14:21:02 -07001744 list_for_each_entry_safe(page, tmp, &surplus_list, lru)
1745 put_page(page);
Naoya Horiguchia9869b82010-09-08 10:19:37 +09001746 spin_lock(&hugetlb_lock);
Adam Litkee4e574b2007-10-16 01:26:19 -07001747
1748 return ret;
1749}
1750
1751/*
1752 * When releasing a hugetlb pool reservation, any surplus pages that were
1753 * allocated to satisfy the reservation must be explicitly freed if they were
1754 * never used.
Lee Schermerhorn685f3452009-09-21 17:01:23 -07001755 * Called with hugetlb_lock held.
Adam Litkee4e574b2007-10-16 01:26:19 -07001756 */
Andi Kleena5516432008-07-23 21:27:41 -07001757static void return_unused_surplus_pages(struct hstate *h,
1758 unsigned long unused_resv_pages)
Adam Litkee4e574b2007-10-16 01:26:19 -07001759{
Adam Litkee4e574b2007-10-16 01:26:19 -07001760 unsigned long nr_pages;
1761
Adam Litkeac09b3a2008-03-04 14:29:38 -08001762 /* Uncommit the reservation */
Andi Kleena5516432008-07-23 21:27:41 -07001763 h->resv_huge_pages -= unused_resv_pages;
Adam Litkeac09b3a2008-03-04 14:29:38 -08001764
Andi Kleenaa888a72008-07-23 21:27:47 -07001765 /* Cannot return gigantic pages currently */
Luiz Capitulinobae7f4a2014-06-04 16:07:08 -07001766 if (hstate_is_gigantic(h))
Andi Kleenaa888a72008-07-23 21:27:47 -07001767 return;
1768
Andi Kleena5516432008-07-23 21:27:41 -07001769 nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
Adam Litkee4e574b2007-10-16 01:26:19 -07001770
Lee Schermerhorn685f3452009-09-21 17:01:23 -07001771 /*
1772 * We want to release as many surplus pages as possible, spread
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -08001773 * evenly across all nodes with memory. Iterate across these nodes
1774 * until we can no longer free unreserved surplus pages. This occurs
1775 * when the nodes with surplus pages have no free pages.
1776 * free_pool_huge_page() will balance the the freed pages across the
1777 * on-line nodes with memory and will handle the hstate accounting.
Lee Schermerhorn685f3452009-09-21 17:01:23 -07001778 */
1779 while (nr_pages--) {
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08001780 if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1))
Lee Schermerhorn685f3452009-09-21 17:01:23 -07001781 break;
Mizuma, Masayoshi7848a4b2014-04-18 15:07:18 -07001782 cond_resched_lock(&hugetlb_lock);
Adam Litkee4e574b2007-10-16 01:26:19 -07001783 }
1784}
1785
Mike Kravetz5e911372015-09-08 15:01:28 -07001786
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001787/*
Mike Kravetzfeba16e2015-09-08 15:01:31 -07001788 * vma_needs_reservation, vma_commit_reservation and vma_end_reservation
Mike Kravetz5e911372015-09-08 15:01:28 -07001789 * are used by the huge page allocation routines to manage reservations.
Mike Kravetzcf3ad202015-06-24 16:57:55 -07001790 *
1791 * vma_needs_reservation is called to determine if the huge page at addr
1792 * within the vma has an associated reservation. If a reservation is
1793 * needed, the value 1 is returned. The caller is then responsible for
1794 * managing the global reservation and subpool usage counts. After
1795 * the huge page has been allocated, vma_commit_reservation is called
Mike Kravetzfeba16e2015-09-08 15:01:31 -07001796 * to add the page to the reservation map. If the page allocation fails,
1797 * the reservation must be ended instead of committed. vma_end_reservation
1798 * is called in such cases.
Mike Kravetzcf3ad202015-06-24 16:57:55 -07001799 *
1800 * In the normal case, vma_commit_reservation returns the same value
1801 * as the preceding vma_needs_reservation call. The only time this
1802 * is not the case is if a reserve map was changed between calls. It
1803 * is the responsibility of the caller to notice the difference and
1804 * take appropriate action.
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001805 */
Mike Kravetz5e911372015-09-08 15:01:28 -07001806enum vma_resv_mode {
1807 VMA_NEEDS_RESV,
1808 VMA_COMMIT_RESV,
Mike Kravetzfeba16e2015-09-08 15:01:31 -07001809 VMA_END_RESV,
Mike Kravetz5e911372015-09-08 15:01:28 -07001810};
Mike Kravetzcf3ad202015-06-24 16:57:55 -07001811static long __vma_reservation_common(struct hstate *h,
1812 struct vm_area_struct *vma, unsigned long addr,
Mike Kravetz5e911372015-09-08 15:01:28 -07001813 enum vma_resv_mode mode)
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001814{
Joonsoo Kim4e35f482014-04-03 14:47:30 -07001815 struct resv_map *resv;
1816 pgoff_t idx;
Mike Kravetzcf3ad202015-06-24 16:57:55 -07001817 long ret;
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001818
Joonsoo Kim4e35f482014-04-03 14:47:30 -07001819 resv = vma_resv_map(vma);
1820 if (!resv)
Andy Whitcroft84afd992008-07-23 21:27:32 -07001821 return 1;
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001822
Joonsoo Kim4e35f482014-04-03 14:47:30 -07001823 idx = vma_hugecache_offset(h, vma, addr);
Mike Kravetz5e911372015-09-08 15:01:28 -07001824 switch (mode) {
1825 case VMA_NEEDS_RESV:
Mike Kravetzcf3ad202015-06-24 16:57:55 -07001826 ret = region_chg(resv, idx, idx + 1);
Mike Kravetz5e911372015-09-08 15:01:28 -07001827 break;
1828 case VMA_COMMIT_RESV:
1829 ret = region_add(resv, idx, idx + 1);
1830 break;
Mike Kravetzfeba16e2015-09-08 15:01:31 -07001831 case VMA_END_RESV:
Mike Kravetz5e911372015-09-08 15:01:28 -07001832 region_abort(resv, idx, idx + 1);
1833 ret = 0;
1834 break;
1835 default:
1836 BUG();
1837 }
Andy Whitcroft84afd992008-07-23 21:27:32 -07001838
Joonsoo Kim4e35f482014-04-03 14:47:30 -07001839 if (vma->vm_flags & VM_MAYSHARE)
Mike Kravetzcf3ad202015-06-24 16:57:55 -07001840 return ret;
Mike Kravetz67961f92016-06-08 15:33:42 -07001841 else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && ret >= 0) {
1842 /*
1843 * In most cases, reserves always exist for private mappings.
1844 * However, a file associated with mapping could have been
1845 * hole punched or truncated after reserves were consumed.
1846 * As subsequent fault on such a range will not use reserves.
1847 * Subtle - The reserve map for private mappings has the
1848 * opposite meaning than that of shared mappings. If NO
1849 * entry is in the reserve map, it means a reservation exists.
1850 * If an entry exists in the reserve map, it means the
1851 * reservation has already been consumed. As a result, the
1852 * return value of this routine is the opposite of the
1853 * value returned from reserve map manipulation routines above.
1854 */
1855 if (ret)
1856 return 0;
1857 else
1858 return 1;
1859 }
Joonsoo Kim4e35f482014-04-03 14:47:30 -07001860 else
Mike Kravetzcf3ad202015-06-24 16:57:55 -07001861 return ret < 0 ? ret : 0;
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001862}
Mike Kravetzcf3ad202015-06-24 16:57:55 -07001863
1864static long vma_needs_reservation(struct hstate *h,
Andi Kleena5516432008-07-23 21:27:41 -07001865 struct vm_area_struct *vma, unsigned long addr)
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001866{
Mike Kravetz5e911372015-09-08 15:01:28 -07001867 return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV);
Mike Kravetzcf3ad202015-06-24 16:57:55 -07001868}
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001869
Mike Kravetzcf3ad202015-06-24 16:57:55 -07001870static long vma_commit_reservation(struct hstate *h,
1871 struct vm_area_struct *vma, unsigned long addr)
1872{
Mike Kravetz5e911372015-09-08 15:01:28 -07001873 return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV);
1874}
1875
Mike Kravetzfeba16e2015-09-08 15:01:31 -07001876static void vma_end_reservation(struct hstate *h,
Mike Kravetz5e911372015-09-08 15:01:28 -07001877 struct vm_area_struct *vma, unsigned long addr)
1878{
Mike Kravetzfeba16e2015-09-08 15:01:31 -07001879 (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV);
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001880}
1881
Mike Kravetz70c35472015-09-08 15:01:54 -07001882struct page *alloc_huge_page(struct vm_area_struct *vma,
Mel Gorman04f2cbe2008-07-23 21:27:25 -07001883 unsigned long addr, int avoid_reserve)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001884{
David Gibson90481622012-03-21 16:34:12 -07001885 struct hugepage_subpool *spool = subpool_vma(vma);
Andi Kleena5516432008-07-23 21:27:41 -07001886 struct hstate *h = hstate_vma(vma);
Adam Litke348ea202007-11-14 16:59:37 -08001887 struct page *page;
Mike Kravetzd85f69b2015-09-08 15:01:47 -07001888 long map_chg, map_commit;
1889 long gbl_chg;
Aneesh Kumar K.V6d76dcf2012-07-31 16:42:18 -07001890 int ret, idx;
1891 struct hugetlb_cgroup *h_cg;
Adam Litke2fc39ce2007-11-14 16:59:39 -08001892
Aneesh Kumar K.V6d76dcf2012-07-31 16:42:18 -07001893 idx = hstate_index(h);
Mel Gormana1e78772008-07-23 21:27:23 -07001894 /*
Mike Kravetzd85f69b2015-09-08 15:01:47 -07001895 * Examine the region/reserve map to determine if the process
1896 * has a reservation for the page to be allocated. A return
1897 * code of zero indicates a reservation exists (no change).
Mel Gormana1e78772008-07-23 21:27:23 -07001898 */
Mike Kravetzd85f69b2015-09-08 15:01:47 -07001899 map_chg = gbl_chg = vma_needs_reservation(h, vma, addr);
1900 if (map_chg < 0)
Aneesh Kumar K.V76dcee72012-07-31 16:41:57 -07001901 return ERR_PTR(-ENOMEM);
Mike Kravetzd85f69b2015-09-08 15:01:47 -07001902
1903 /*
1904 * Processes that did not create the mapping will have no
1905 * reserves as indicated by the region/reserve map. Check
1906 * that the allocation will not exceed the subpool limit.
1907 * Allocations for MAP_NORESERVE mappings also need to be
1908 * checked against any subpool limit.
1909 */
1910 if (map_chg || avoid_reserve) {
1911 gbl_chg = hugepage_subpool_get_pages(spool, 1);
1912 if (gbl_chg < 0) {
Mike Kravetzfeba16e2015-09-08 15:01:31 -07001913 vma_end_reservation(h, vma, addr);
Aneesh Kumar K.V76dcee72012-07-31 16:41:57 -07001914 return ERR_PTR(-ENOSPC);
Mike Kravetz5e911372015-09-08 15:01:28 -07001915 }
Mel Gormana1e78772008-07-23 21:27:23 -07001916
Mike Kravetzd85f69b2015-09-08 15:01:47 -07001917 /*
1918 * Even though there was no reservation in the region/reserve
1919 * map, there could be reservations associated with the
1920 * subpool that can be used. This would be indicated if the
1921 * return value of hugepage_subpool_get_pages() is zero.
1922 * However, if avoid_reserve is specified we still avoid even
1923 * the subpool reservations.
1924 */
1925 if (avoid_reserve)
1926 gbl_chg = 1;
1927 }
1928
Aneesh Kumar K.V6d76dcf2012-07-31 16:42:18 -07001929 ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg);
Jianyu Zhan8f34af62014-06-04 16:10:36 -07001930 if (ret)
1931 goto out_subpool_put;
1932
Mel Gormana1e78772008-07-23 21:27:23 -07001933 spin_lock(&hugetlb_lock);
Mike Kravetzd85f69b2015-09-08 15:01:47 -07001934 /*
1935 * glb_chg is passed to indicate whether or not a page must be taken
1936 * from the global free pool (global change). gbl_chg == 0 indicates
1937 * a reservation exists for the allocation.
1938 */
1939 page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg);
Joonsoo Kim81a6fca2013-09-11 14:20:58 -07001940 if (!page) {
Aneesh Kumar K.V94ae8ba2012-07-31 16:42:35 -07001941 spin_unlock(&hugetlb_lock);
Dave Hansen099730d2015-11-05 18:50:17 -08001942 page = __alloc_buddy_huge_page_with_mpol(h, vma, addr);
Jianyu Zhan8f34af62014-06-04 16:10:36 -07001943 if (!page)
1944 goto out_uncharge_cgroup;
Naoya Horiguchia88c7692015-12-11 13:40:24 -08001945 if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) {
1946 SetPagePrivate(page);
1947 h->resv_huge_pages--;
1948 }
Aneesh Kumar K.V79dbb232012-07-31 16:42:32 -07001949 spin_lock(&hugetlb_lock);
1950 list_move(&page->lru, &h->hugepage_activelist);
Joonsoo Kim81a6fca2013-09-11 14:20:58 -07001951 /* Fall through */
Mel Gormana1e78772008-07-23 21:27:23 -07001952 }
Joonsoo Kim81a6fca2013-09-11 14:20:58 -07001953 hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page);
1954 spin_unlock(&hugetlb_lock);
Mel Gormana1e78772008-07-23 21:27:23 -07001955
David Gibson90481622012-03-21 16:34:12 -07001956 set_page_private(page, (unsigned long)spool);
Mel Gormana1e78772008-07-23 21:27:23 -07001957
Mike Kravetzd85f69b2015-09-08 15:01:47 -07001958 map_commit = vma_commit_reservation(h, vma, addr);
1959 if (unlikely(map_chg > map_commit)) {
Mike Kravetz33039672015-06-24 16:57:58 -07001960 /*
1961 * The page was added to the reservation map between
1962 * vma_needs_reservation and vma_commit_reservation.
1963 * This indicates a race with hugetlb_reserve_pages.
1964 * Adjust for the subpool count incremented above AND
1965 * in hugetlb_reserve_pages for the same page. Also,
1966 * the reservation count added in hugetlb_reserve_pages
1967 * no longer applies.
1968 */
1969 long rsv_adjust;
1970
1971 rsv_adjust = hugepage_subpool_put_pages(spool, 1);
1972 hugetlb_acct_memory(h, -rsv_adjust);
1973 }
Adam Litke90d8b7e2007-11-14 16:59:42 -08001974 return page;
Jianyu Zhan8f34af62014-06-04 16:10:36 -07001975
1976out_uncharge_cgroup:
1977 hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg);
1978out_subpool_put:
Mike Kravetzd85f69b2015-09-08 15:01:47 -07001979 if (map_chg || avoid_reserve)
Jianyu Zhan8f34af62014-06-04 16:10:36 -07001980 hugepage_subpool_put_pages(spool, 1);
Mike Kravetzfeba16e2015-09-08 15:01:31 -07001981 vma_end_reservation(h, vma, addr);
Jianyu Zhan8f34af62014-06-04 16:10:36 -07001982 return ERR_PTR(-ENOSPC);
David Gibsonb45b5bd2006-03-22 00:08:55 -08001983}
1984
Naoya Horiguchi74060e42013-09-11 14:22:06 -07001985/*
1986 * alloc_huge_page()'s wrapper which simply returns the page if allocation
1987 * succeeds, otherwise NULL. This function is called from new_vma_page(),
1988 * where no ERR_VALUE is expected to be returned.
1989 */
1990struct page *alloc_huge_page_noerr(struct vm_area_struct *vma,
1991 unsigned long addr, int avoid_reserve)
1992{
1993 struct page *page = alloc_huge_page(vma, addr, avoid_reserve);
1994 if (IS_ERR(page))
1995 page = NULL;
1996 return page;
1997}
1998
Cyrill Gorcunov91f47662009-01-06 14:40:33 -08001999int __weak alloc_bootmem_huge_page(struct hstate *h)
Andi Kleenaa888a72008-07-23 21:27:47 -07002000{
2001 struct huge_bootmem_page *m;
Joonsoo Kimb2261022013-09-11 14:21:00 -07002002 int nr_nodes, node;
Andi Kleenaa888a72008-07-23 21:27:47 -07002003
Joonsoo Kimb2261022013-09-11 14:21:00 -07002004 for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) {
Andi Kleenaa888a72008-07-23 21:27:47 -07002005 void *addr;
2006
Grygorii Strashko8b89a112014-01-21 15:50:36 -08002007 addr = memblock_virt_alloc_try_nid_nopanic(
2008 huge_page_size(h), huge_page_size(h),
2009 0, BOOTMEM_ALLOC_ACCESSIBLE, node);
Andi Kleenaa888a72008-07-23 21:27:47 -07002010 if (addr) {
2011 /*
2012 * Use the beginning of the huge page to store the
2013 * huge_bootmem_page struct (until gather_bootmem
2014 * puts them into the mem_map).
2015 */
2016 m = addr;
Cyrill Gorcunov91f47662009-01-06 14:40:33 -08002017 goto found;
Andi Kleenaa888a72008-07-23 21:27:47 -07002018 }
Andi Kleenaa888a72008-07-23 21:27:47 -07002019 }
2020 return 0;
2021
2022found:
Luiz Capitulinodf994ea2014-12-12 16:55:21 -08002023 BUG_ON(!IS_ALIGNED(virt_to_phys(m), huge_page_size(h)));
Andi Kleenaa888a72008-07-23 21:27:47 -07002024 /* Put them into a private list first because mem_map is not up yet */
2025 list_add(&m->list, &huge_boot_pages);
2026 m->hstate = h;
2027 return 1;
2028}
2029
Kirill A. Shutemovd00181b2015-11-06 16:29:57 -08002030static void __init prep_compound_huge_page(struct page *page,
2031 unsigned int order)
Andy Whitcroft18229df2008-11-06 12:53:27 -08002032{
2033 if (unlikely(order > (MAX_ORDER - 1)))
2034 prep_compound_gigantic_page(page, order);
2035 else
2036 prep_compound_page(page, order);
2037}
2038
Andi Kleenaa888a72008-07-23 21:27:47 -07002039/* Put bootmem huge pages into the standard lists after mem_map is up */
2040static void __init gather_bootmem_prealloc(void)
2041{
2042 struct huge_bootmem_page *m;
2043
2044 list_for_each_entry(m, &huge_boot_pages, list) {
Andi Kleenaa888a72008-07-23 21:27:47 -07002045 struct hstate *h = m->hstate;
Becky Bruceee8f2482011-07-25 17:11:50 -07002046 struct page *page;
2047
2048#ifdef CONFIG_HIGHMEM
2049 page = pfn_to_page(m->phys >> PAGE_SHIFT);
Grygorii Strashko8b89a112014-01-21 15:50:36 -08002050 memblock_free_late(__pa(m),
2051 sizeof(struct huge_bootmem_page));
Becky Bruceee8f2482011-07-25 17:11:50 -07002052#else
2053 page = virt_to_page(m);
2054#endif
Andi Kleenaa888a72008-07-23 21:27:47 -07002055 WARN_ON(page_count(page) != 1);
Andy Whitcroft18229df2008-11-06 12:53:27 -08002056 prep_compound_huge_page(page, h->order);
Andrea Arcangelief5a22b2013-10-16 13:46:56 -07002057 WARN_ON(PageReserved(page));
Andi Kleenaa888a72008-07-23 21:27:47 -07002058 prep_new_huge_page(h, page, page_to_nid(page));
Rafael Aquinib0320c72011-06-15 15:08:39 -07002059 /*
2060 * If we had gigantic hugepages allocated at boot time, we need
2061 * to restore the 'stolen' pages to totalram_pages in order to
2062 * fix confusing memory reports from free(1) and another
2063 * side-effects, like CommitLimit going negative.
2064 */
Luiz Capitulinobae7f4a2014-06-04 16:07:08 -07002065 if (hstate_is_gigantic(h))
Jiang Liu3dcc0572013-07-03 15:03:21 -07002066 adjust_managed_page_count(page, 1 << h->order);
Andi Kleenaa888a72008-07-23 21:27:47 -07002067 }
2068}
2069
Andi Kleen8faa8b02008-07-23 21:27:48 -07002070static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002071{
2072 unsigned long i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002073
Andi Kleene5ff2152008-07-23 21:27:42 -07002074 for (i = 0; i < h->max_huge_pages; ++i) {
Luiz Capitulinobae7f4a2014-06-04 16:07:08 -07002075 if (hstate_is_gigantic(h)) {
Andi Kleenaa888a72008-07-23 21:27:47 -07002076 if (!alloc_bootmem_huge_page(h))
2077 break;
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -08002078 } else if (!alloc_fresh_huge_page(h,
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08002079 &node_states[N_MEMORY]))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002080 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002081 }
Andi Kleen8faa8b02008-07-23 21:27:48 -07002082 h->max_huge_pages = i;
Andi Kleene5ff2152008-07-23 21:27:42 -07002083}
2084
2085static void __init hugetlb_init_hstates(void)
2086{
2087 struct hstate *h;
2088
2089 for_each_hstate(h) {
Naoya Horiguchi641844f2015-06-24 16:56:59 -07002090 if (minimum_order > huge_page_order(h))
2091 minimum_order = huge_page_order(h);
2092
Andi Kleen8faa8b02008-07-23 21:27:48 -07002093 /* oversize hugepages were init'ed in early boot */
Luiz Capitulinobae7f4a2014-06-04 16:07:08 -07002094 if (!hstate_is_gigantic(h))
Andi Kleen8faa8b02008-07-23 21:27:48 -07002095 hugetlb_hstate_alloc_pages(h);
Andi Kleene5ff2152008-07-23 21:27:42 -07002096 }
Naoya Horiguchi641844f2015-06-24 16:56:59 -07002097 VM_BUG_ON(minimum_order == UINT_MAX);
Andi Kleene5ff2152008-07-23 21:27:42 -07002098}
2099
Andi Kleen4abd32d2008-07-23 21:27:49 -07002100static char * __init memfmt(char *buf, unsigned long n)
2101{
2102 if (n >= (1UL << 30))
2103 sprintf(buf, "%lu GB", n >> 30);
2104 else if (n >= (1UL << 20))
2105 sprintf(buf, "%lu MB", n >> 20);
2106 else
2107 sprintf(buf, "%lu KB", n >> 10);
2108 return buf;
2109}
2110
Andi Kleene5ff2152008-07-23 21:27:42 -07002111static void __init report_hugepages(void)
2112{
2113 struct hstate *h;
2114
2115 for_each_hstate(h) {
Andi Kleen4abd32d2008-07-23 21:27:49 -07002116 char buf[32];
Andrew Mortonffb22af2013-02-22 16:32:08 -08002117 pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n",
Andi Kleen4abd32d2008-07-23 21:27:49 -07002118 memfmt(buf, huge_page_size(h)),
2119 h->free_huge_pages);
Andi Kleene5ff2152008-07-23 21:27:42 -07002120 }
2121}
2122
Linus Torvalds1da177e2005-04-16 15:20:36 -07002123#ifdef CONFIG_HIGHMEM
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08002124static void try_to_free_low(struct hstate *h, unsigned long count,
2125 nodemask_t *nodes_allowed)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002126{
Christoph Lameter4415cc82006-09-25 23:31:55 -07002127 int i;
2128
Luiz Capitulinobae7f4a2014-06-04 16:07:08 -07002129 if (hstate_is_gigantic(h))
Andi Kleenaa888a72008-07-23 21:27:47 -07002130 return;
2131
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08002132 for_each_node_mask(i, *nodes_allowed) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002133 struct page *page, *next;
Andi Kleena5516432008-07-23 21:27:41 -07002134 struct list_head *freel = &h->hugepage_freelists[i];
2135 list_for_each_entry_safe(page, next, freel, lru) {
2136 if (count >= h->nr_huge_pages)
Adam Litke6b0c8802007-10-16 01:26:23 -07002137 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002138 if (PageHighMem(page))
2139 continue;
2140 list_del(&page->lru);
Andi Kleene5ff2152008-07-23 21:27:42 -07002141 update_and_free_page(h, page);
Andi Kleena5516432008-07-23 21:27:41 -07002142 h->free_huge_pages--;
2143 h->free_huge_pages_node[page_to_nid(page)]--;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002144 }
2145 }
2146}
2147#else
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08002148static inline void try_to_free_low(struct hstate *h, unsigned long count,
2149 nodemask_t *nodes_allowed)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002150{
2151}
2152#endif
2153
Wu Fengguang20a03072009-06-16 15:32:22 -07002154/*
2155 * Increment or decrement surplus_huge_pages. Keep node-specific counters
2156 * balanced by operating on them in a round-robin fashion.
2157 * Returns 1 if an adjustment was made.
2158 */
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08002159static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
2160 int delta)
Wu Fengguang20a03072009-06-16 15:32:22 -07002161{
Joonsoo Kimb2261022013-09-11 14:21:00 -07002162 int nr_nodes, node;
Wu Fengguang20a03072009-06-16 15:32:22 -07002163
2164 VM_BUG_ON(delta != -1 && delta != 1);
Wu Fengguang20a03072009-06-16 15:32:22 -07002165
Joonsoo Kimb2261022013-09-11 14:21:00 -07002166 if (delta < 0) {
2167 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
2168 if (h->surplus_huge_pages_node[node])
2169 goto found;
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07002170 }
Joonsoo Kimb2261022013-09-11 14:21:00 -07002171 } else {
2172 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
2173 if (h->surplus_huge_pages_node[node] <
2174 h->nr_huge_pages_node[node])
2175 goto found;
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07002176 }
Joonsoo Kimb2261022013-09-11 14:21:00 -07002177 }
2178 return 0;
Wu Fengguang20a03072009-06-16 15:32:22 -07002179
Joonsoo Kimb2261022013-09-11 14:21:00 -07002180found:
2181 h->surplus_huge_pages += delta;
2182 h->surplus_huge_pages_node[node] += delta;
2183 return 1;
Wu Fengguang20a03072009-06-16 15:32:22 -07002184}
2185
Andi Kleena5516432008-07-23 21:27:41 -07002186#define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08002187static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count,
2188 nodemask_t *nodes_allowed)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002189{
Adam Litke7893d1d2007-10-16 01:26:18 -07002190 unsigned long min_count, ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002191
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07002192 if (hstate_is_gigantic(h) && !gigantic_page_supported())
Andi Kleenaa888a72008-07-23 21:27:47 -07002193 return h->max_huge_pages;
2194
Adam Litke7893d1d2007-10-16 01:26:18 -07002195 /*
2196 * Increase the pool size
2197 * First take pages out of surplus state. Then make up the
2198 * remaining difference by allocating fresh huge pages.
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08002199 *
Naoya Horiguchid15c7c02015-11-10 14:45:11 -08002200 * We might race with __alloc_buddy_huge_page() here and be unable
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08002201 * to convert a surplus huge page to a normal huge page. That is
2202 * not critical, though, it just means the overall size of the
2203 * pool might be one hugepage larger than it needs to be, but
2204 * within all the constraints specified by the sysctls.
Adam Litke7893d1d2007-10-16 01:26:18 -07002205 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002206 spin_lock(&hugetlb_lock);
Andi Kleena5516432008-07-23 21:27:41 -07002207 while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08002208 if (!adjust_pool_surplus(h, nodes_allowed, -1))
Adam Litke7893d1d2007-10-16 01:26:18 -07002209 break;
2210 }
2211
Andi Kleena5516432008-07-23 21:27:41 -07002212 while (count > persistent_huge_pages(h)) {
Adam Litke7893d1d2007-10-16 01:26:18 -07002213 /*
2214 * If this allocation races such that we no longer need the
2215 * page, free_huge_page will handle it by freeing the page
2216 * and reducing the surplus.
2217 */
2218 spin_unlock(&hugetlb_lock);
Jia He649920c2016-08-02 14:02:31 -07002219
2220 /* yield cpu to avoid soft lockup */
2221 cond_resched();
2222
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07002223 if (hstate_is_gigantic(h))
2224 ret = alloc_fresh_gigantic_page(h, nodes_allowed);
2225 else
2226 ret = alloc_fresh_huge_page(h, nodes_allowed);
Adam Litke7893d1d2007-10-16 01:26:18 -07002227 spin_lock(&hugetlb_lock);
2228 if (!ret)
2229 goto out;
2230
Mel Gorman536240f22009-12-14 17:59:56 -08002231 /* Bail for signals. Probably ctrl-c from user */
2232 if (signal_pending(current))
2233 goto out;
Adam Litke7893d1d2007-10-16 01:26:18 -07002234 }
Adam Litke7893d1d2007-10-16 01:26:18 -07002235
2236 /*
2237 * Decrease the pool size
2238 * First return free pages to the buddy allocator (being careful
2239 * to keep enough around to satisfy reservations). Then place
2240 * pages into surplus state as needed so the pool will shrink
2241 * to the desired size as pages become free.
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08002242 *
2243 * By placing pages into the surplus state independent of the
2244 * overcommit value, we are allowing the surplus pool size to
2245 * exceed overcommit. There are few sane options here. Since
Naoya Horiguchid15c7c02015-11-10 14:45:11 -08002246 * __alloc_buddy_huge_page() is checking the global counter,
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08002247 * though, we'll note that we're not allowed to exceed surplus
2248 * and won't grow the pool anywhere else. Not until one of the
2249 * sysctls are changed, or the surplus pages go out of use.
Adam Litke7893d1d2007-10-16 01:26:18 -07002250 */
Andi Kleena5516432008-07-23 21:27:41 -07002251 min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
Adam Litke6b0c8802007-10-16 01:26:23 -07002252 min_count = max(count, min_count);
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08002253 try_to_free_low(h, min_count, nodes_allowed);
Andi Kleena5516432008-07-23 21:27:41 -07002254 while (min_count < persistent_huge_pages(h)) {
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08002255 if (!free_pool_huge_page(h, nodes_allowed, 0))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002256 break;
Mizuma, Masayoshi55f67142014-04-07 15:37:54 -07002257 cond_resched_lock(&hugetlb_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002258 }
Andi Kleena5516432008-07-23 21:27:41 -07002259 while (count < persistent_huge_pages(h)) {
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08002260 if (!adjust_pool_surplus(h, nodes_allowed, 1))
Adam Litke7893d1d2007-10-16 01:26:18 -07002261 break;
2262 }
2263out:
Andi Kleena5516432008-07-23 21:27:41 -07002264 ret = persistent_huge_pages(h);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002265 spin_unlock(&hugetlb_lock);
Adam Litke7893d1d2007-10-16 01:26:18 -07002266 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002267}
2268
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002269#define HSTATE_ATTR_RO(_name) \
2270 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
2271
2272#define HSTATE_ATTR(_name) \
2273 static struct kobj_attribute _name##_attr = \
2274 __ATTR(_name, 0644, _name##_show, _name##_store)
2275
2276static struct kobject *hugepages_kobj;
2277static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
2278
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002279static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
2280
2281static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002282{
2283 int i;
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002284
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002285 for (i = 0; i < HUGE_MAX_HSTATE; i++)
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002286 if (hstate_kobjs[i] == kobj) {
2287 if (nidp)
2288 *nidp = NUMA_NO_NODE;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002289 return &hstates[i];
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002290 }
2291
2292 return kobj_to_node_hstate(kobj, nidp);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002293}
2294
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002295static ssize_t nr_hugepages_show_common(struct kobject *kobj,
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002296 struct kobj_attribute *attr, char *buf)
2297{
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002298 struct hstate *h;
2299 unsigned long nr_huge_pages;
2300 int nid;
2301
2302 h = kobj_to_hstate(kobj, &nid);
2303 if (nid == NUMA_NO_NODE)
2304 nr_huge_pages = h->nr_huge_pages;
2305 else
2306 nr_huge_pages = h->nr_huge_pages_node[nid];
2307
2308 return sprintf(buf, "%lu\n", nr_huge_pages);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002309}
Eric B Munsonadbe8722011-01-13 15:47:27 -08002310
David Rientjes238d3c12014-08-06 16:06:51 -07002311static ssize_t __nr_hugepages_store_common(bool obey_mempolicy,
2312 struct hstate *h, int nid,
2313 unsigned long count, size_t len)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002314{
2315 int err;
David Rientjesbad44b52009-12-14 17:58:38 -08002316 NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002317
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07002318 if (hstate_is_gigantic(h) && !gigantic_page_supported()) {
Eric B Munsonadbe8722011-01-13 15:47:27 -08002319 err = -EINVAL;
2320 goto out;
2321 }
2322
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002323 if (nid == NUMA_NO_NODE) {
2324 /*
2325 * global hstate attribute
2326 */
2327 if (!(obey_mempolicy &&
2328 init_nodemask_of_mempolicy(nodes_allowed))) {
2329 NODEMASK_FREE(nodes_allowed);
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08002330 nodes_allowed = &node_states[N_MEMORY];
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002331 }
2332 } else if (nodes_allowed) {
2333 /*
2334 * per node hstate attribute: adjust count to global,
2335 * but restrict alloc/free to the specified node.
2336 */
2337 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
2338 init_nodemask_of_node(nodes_allowed, nid);
2339 } else
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08002340 nodes_allowed = &node_states[N_MEMORY];
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002341
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002342 h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002343
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08002344 if (nodes_allowed != &node_states[N_MEMORY])
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002345 NODEMASK_FREE(nodes_allowed);
2346
2347 return len;
Eric B Munsonadbe8722011-01-13 15:47:27 -08002348out:
2349 NODEMASK_FREE(nodes_allowed);
2350 return err;
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002351}
2352
David Rientjes238d3c12014-08-06 16:06:51 -07002353static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
2354 struct kobject *kobj, const char *buf,
2355 size_t len)
2356{
2357 struct hstate *h;
2358 unsigned long count;
2359 int nid;
2360 int err;
2361
2362 err = kstrtoul(buf, 10, &count);
2363 if (err)
2364 return err;
2365
2366 h = kobj_to_hstate(kobj, &nid);
2367 return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len);
2368}
2369
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002370static ssize_t nr_hugepages_show(struct kobject *kobj,
2371 struct kobj_attribute *attr, char *buf)
2372{
2373 return nr_hugepages_show_common(kobj, attr, buf);
2374}
2375
2376static ssize_t nr_hugepages_store(struct kobject *kobj,
2377 struct kobj_attribute *attr, const char *buf, size_t len)
2378{
David Rientjes238d3c12014-08-06 16:06:51 -07002379 return nr_hugepages_store_common(false, kobj, buf, len);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002380}
2381HSTATE_ATTR(nr_hugepages);
2382
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002383#ifdef CONFIG_NUMA
2384
2385/*
2386 * hstate attribute for optionally mempolicy-based constraint on persistent
2387 * huge page alloc/free.
2388 */
2389static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
2390 struct kobj_attribute *attr, char *buf)
2391{
2392 return nr_hugepages_show_common(kobj, attr, buf);
2393}
2394
2395static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
2396 struct kobj_attribute *attr, const char *buf, size_t len)
2397{
David Rientjes238d3c12014-08-06 16:06:51 -07002398 return nr_hugepages_store_common(true, kobj, buf, len);
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002399}
2400HSTATE_ATTR(nr_hugepages_mempolicy);
2401#endif
2402
2403
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002404static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
2405 struct kobj_attribute *attr, char *buf)
2406{
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002407 struct hstate *h = kobj_to_hstate(kobj, NULL);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002408 return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
2409}
Eric B Munsonadbe8722011-01-13 15:47:27 -08002410
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002411static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
2412 struct kobj_attribute *attr, const char *buf, size_t count)
2413{
2414 int err;
2415 unsigned long input;
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002416 struct hstate *h = kobj_to_hstate(kobj, NULL);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002417
Luiz Capitulinobae7f4a2014-06-04 16:07:08 -07002418 if (hstate_is_gigantic(h))
Eric B Munsonadbe8722011-01-13 15:47:27 -08002419 return -EINVAL;
2420
Jingoo Han3dbb95f2013-09-11 14:20:25 -07002421 err = kstrtoul(buf, 10, &input);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002422 if (err)
Eric B Munson73ae31e2011-01-13 15:47:28 -08002423 return err;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002424
2425 spin_lock(&hugetlb_lock);
2426 h->nr_overcommit_huge_pages = input;
2427 spin_unlock(&hugetlb_lock);
2428
2429 return count;
2430}
2431HSTATE_ATTR(nr_overcommit_hugepages);
2432
2433static ssize_t free_hugepages_show(struct kobject *kobj,
2434 struct kobj_attribute *attr, char *buf)
2435{
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002436 struct hstate *h;
2437 unsigned long free_huge_pages;
2438 int nid;
2439
2440 h = kobj_to_hstate(kobj, &nid);
2441 if (nid == NUMA_NO_NODE)
2442 free_huge_pages = h->free_huge_pages;
2443 else
2444 free_huge_pages = h->free_huge_pages_node[nid];
2445
2446 return sprintf(buf, "%lu\n", free_huge_pages);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002447}
2448HSTATE_ATTR_RO(free_hugepages);
2449
2450static ssize_t resv_hugepages_show(struct kobject *kobj,
2451 struct kobj_attribute *attr, char *buf)
2452{
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002453 struct hstate *h = kobj_to_hstate(kobj, NULL);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002454 return sprintf(buf, "%lu\n", h->resv_huge_pages);
2455}
2456HSTATE_ATTR_RO(resv_hugepages);
2457
2458static ssize_t surplus_hugepages_show(struct kobject *kobj,
2459 struct kobj_attribute *attr, char *buf)
2460{
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002461 struct hstate *h;
2462 unsigned long surplus_huge_pages;
2463 int nid;
2464
2465 h = kobj_to_hstate(kobj, &nid);
2466 if (nid == NUMA_NO_NODE)
2467 surplus_huge_pages = h->surplus_huge_pages;
2468 else
2469 surplus_huge_pages = h->surplus_huge_pages_node[nid];
2470
2471 return sprintf(buf, "%lu\n", surplus_huge_pages);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002472}
2473HSTATE_ATTR_RO(surplus_hugepages);
2474
2475static struct attribute *hstate_attrs[] = {
2476 &nr_hugepages_attr.attr,
2477 &nr_overcommit_hugepages_attr.attr,
2478 &free_hugepages_attr.attr,
2479 &resv_hugepages_attr.attr,
2480 &surplus_hugepages_attr.attr,
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002481#ifdef CONFIG_NUMA
2482 &nr_hugepages_mempolicy_attr.attr,
2483#endif
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002484 NULL,
2485};
2486
2487static struct attribute_group hstate_attr_group = {
2488 .attrs = hstate_attrs,
2489};
2490
Jeff Mahoney094e9532010-02-02 13:44:14 -08002491static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
2492 struct kobject **hstate_kobjs,
2493 struct attribute_group *hstate_attr_group)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002494{
2495 int retval;
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07002496 int hi = hstate_index(h);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002497
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002498 hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
2499 if (!hstate_kobjs[hi])
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002500 return -ENOMEM;
2501
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002502 retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002503 if (retval)
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002504 kobject_put(hstate_kobjs[hi]);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002505
2506 return retval;
2507}
2508
2509static void __init hugetlb_sysfs_init(void)
2510{
2511 struct hstate *h;
2512 int err;
2513
2514 hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
2515 if (!hugepages_kobj)
2516 return;
2517
2518 for_each_hstate(h) {
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002519 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
2520 hstate_kobjs, &hstate_attr_group);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002521 if (err)
Andrew Mortonffb22af2013-02-22 16:32:08 -08002522 pr_err("Hugetlb: Unable to add hstate %s", h->name);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002523 }
2524}
2525
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002526#ifdef CONFIG_NUMA
2527
2528/*
2529 * node_hstate/s - associate per node hstate attributes, via their kobjects,
Kay Sievers10fbcf42011-12-21 14:48:43 -08002530 * with node devices in node_devices[] using a parallel array. The array
2531 * index of a node device or _hstate == node id.
2532 * This is here to avoid any static dependency of the node device driver, in
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002533 * the base kernel, on the hugetlb module.
2534 */
2535struct node_hstate {
2536 struct kobject *hugepages_kobj;
2537 struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
2538};
Alexander Kuleshovb4e289a2015-11-05 18:50:14 -08002539static struct node_hstate node_hstates[MAX_NUMNODES];
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002540
2541/*
Kay Sievers10fbcf42011-12-21 14:48:43 -08002542 * A subset of global hstate attributes for node devices
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002543 */
2544static struct attribute *per_node_hstate_attrs[] = {
2545 &nr_hugepages_attr.attr,
2546 &free_hugepages_attr.attr,
2547 &surplus_hugepages_attr.attr,
2548 NULL,
2549};
2550
2551static struct attribute_group per_node_hstate_attr_group = {
2552 .attrs = per_node_hstate_attrs,
2553};
2554
2555/*
Kay Sievers10fbcf42011-12-21 14:48:43 -08002556 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002557 * Returns node id via non-NULL nidp.
2558 */
2559static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
2560{
2561 int nid;
2562
2563 for (nid = 0; nid < nr_node_ids; nid++) {
2564 struct node_hstate *nhs = &node_hstates[nid];
2565 int i;
2566 for (i = 0; i < HUGE_MAX_HSTATE; i++)
2567 if (nhs->hstate_kobjs[i] == kobj) {
2568 if (nidp)
2569 *nidp = nid;
2570 return &hstates[i];
2571 }
2572 }
2573
2574 BUG();
2575 return NULL;
2576}
2577
2578/*
Kay Sievers10fbcf42011-12-21 14:48:43 -08002579 * Unregister hstate attributes from a single node device.
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002580 * No-op if no hstate attributes attached.
2581 */
Claudiu Ghioc3cd8b442013-03-04 12:46:15 +02002582static void hugetlb_unregister_node(struct node *node)
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002583{
2584 struct hstate *h;
Kay Sievers10fbcf42011-12-21 14:48:43 -08002585 struct node_hstate *nhs = &node_hstates[node->dev.id];
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002586
2587 if (!nhs->hugepages_kobj)
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -08002588 return; /* no hstate attributes */
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002589
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07002590 for_each_hstate(h) {
2591 int idx = hstate_index(h);
2592 if (nhs->hstate_kobjs[idx]) {
2593 kobject_put(nhs->hstate_kobjs[idx]);
2594 nhs->hstate_kobjs[idx] = NULL;
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002595 }
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07002596 }
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002597
2598 kobject_put(nhs->hugepages_kobj);
2599 nhs->hugepages_kobj = NULL;
2600}
2601
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002602
2603/*
Kay Sievers10fbcf42011-12-21 14:48:43 -08002604 * Register hstate attributes for a single node device.
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002605 * No-op if attributes already registered.
2606 */
Claudiu Ghioc3cd8b442013-03-04 12:46:15 +02002607static void hugetlb_register_node(struct node *node)
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002608{
2609 struct hstate *h;
Kay Sievers10fbcf42011-12-21 14:48:43 -08002610 struct node_hstate *nhs = &node_hstates[node->dev.id];
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002611 int err;
2612
2613 if (nhs->hugepages_kobj)
2614 return; /* already allocated */
2615
2616 nhs->hugepages_kobj = kobject_create_and_add("hugepages",
Kay Sievers10fbcf42011-12-21 14:48:43 -08002617 &node->dev.kobj);
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002618 if (!nhs->hugepages_kobj)
2619 return;
2620
2621 for_each_hstate(h) {
2622 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
2623 nhs->hstate_kobjs,
2624 &per_node_hstate_attr_group);
2625 if (err) {
Andrew Mortonffb22af2013-02-22 16:32:08 -08002626 pr_err("Hugetlb: Unable to add hstate %s for node %d\n",
2627 h->name, node->dev.id);
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002628 hugetlb_unregister_node(node);
2629 break;
2630 }
2631 }
2632}
2633
2634/*
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -08002635 * hugetlb init time: register hstate attributes for all registered node
Kay Sievers10fbcf42011-12-21 14:48:43 -08002636 * devices of nodes that have memory. All on-line nodes should have
2637 * registered their associated device by this time.
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002638 */
Luiz Capitulino7d9ca002014-12-12 16:55:24 -08002639static void __init hugetlb_register_all_nodes(void)
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002640{
2641 int nid;
2642
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08002643 for_each_node_state(nid, N_MEMORY) {
Wen Congyang87327942012-12-11 16:00:56 -08002644 struct node *node = node_devices[nid];
Kay Sievers10fbcf42011-12-21 14:48:43 -08002645 if (node->dev.id == nid)
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002646 hugetlb_register_node(node);
2647 }
2648
2649 /*
Kay Sievers10fbcf42011-12-21 14:48:43 -08002650 * Let the node device driver know we're here so it can
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002651 * [un]register hstate attributes on node hotplug.
2652 */
2653 register_hugetlbfs_with_node(hugetlb_register_node,
2654 hugetlb_unregister_node);
2655}
2656#else /* !CONFIG_NUMA */
2657
2658static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
2659{
2660 BUG();
2661 if (nidp)
2662 *nidp = -1;
2663 return NULL;
2664}
2665
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002666static void hugetlb_register_all_nodes(void) { }
2667
2668#endif
2669
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002670static int __init hugetlb_init(void)
2671{
Davidlohr Bueso8382d912014-04-03 14:47:31 -07002672 int i;
2673
Nishanth Aravamudan457c1b22014-05-06 12:50:00 -07002674 if (!hugepages_supported())
Benjamin Herrenschmidt0ef89d22008-07-31 00:07:30 -07002675 return 0;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002676
Nick Piggine11bfbf2008-07-23 21:27:52 -07002677 if (!size_to_hstate(default_hstate_size)) {
2678 default_hstate_size = HPAGE_SIZE;
2679 if (!size_to_hstate(default_hstate_size))
2680 hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002681 }
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07002682 default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size));
Vaishali Thakkarf8b74812016-02-17 13:11:26 -08002683 if (default_hstate_max_huge_pages) {
2684 if (!default_hstate.max_huge_pages)
2685 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
2686 }
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002687
2688 hugetlb_init_hstates();
Andi Kleenaa888a72008-07-23 21:27:47 -07002689 gather_bootmem_prealloc();
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002690 report_hugepages();
2691
2692 hugetlb_sysfs_init();
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002693 hugetlb_register_all_nodes();
Jianguo Wu7179e7b2012-12-18 14:23:19 -08002694 hugetlb_cgroup_file_init();
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002695
Davidlohr Bueso8382d912014-04-03 14:47:31 -07002696#ifdef CONFIG_SMP
2697 num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus());
2698#else
2699 num_fault_mutexes = 1;
2700#endif
Mike Kravetzc672c7f2015-09-08 15:01:35 -07002701 hugetlb_fault_mutex_table =
Davidlohr Bueso8382d912014-04-03 14:47:31 -07002702 kmalloc(sizeof(struct mutex) * num_fault_mutexes, GFP_KERNEL);
Mike Kravetzc672c7f2015-09-08 15:01:35 -07002703 BUG_ON(!hugetlb_fault_mutex_table);
Davidlohr Bueso8382d912014-04-03 14:47:31 -07002704
2705 for (i = 0; i < num_fault_mutexes; i++)
Mike Kravetzc672c7f2015-09-08 15:01:35 -07002706 mutex_init(&hugetlb_fault_mutex_table[i]);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002707 return 0;
2708}
Paul Gortmaker3e89e1c2016-01-14 15:21:52 -08002709subsys_initcall(hugetlb_init);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002710
2711/* Should be called on processing a hugepagesz=... option */
Vaishali Thakkar9fee0212016-05-19 17:11:04 -07002712void __init hugetlb_bad_size(void)
2713{
2714 parsed_valid_hugepagesz = false;
2715}
2716
Kirill A. Shutemovd00181b2015-11-06 16:29:57 -08002717void __init hugetlb_add_hstate(unsigned int order)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002718{
2719 struct hstate *h;
Andi Kleen8faa8b02008-07-23 21:27:48 -07002720 unsigned long i;
2721
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002722 if (size_to_hstate(PAGE_SIZE << order)) {
Joe Perches598d8092016-03-17 14:19:44 -07002723 pr_warn("hugepagesz= specified twice, ignoring\n");
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002724 return;
2725 }
Aneesh Kumar K.V47d38342012-07-31 16:41:54 -07002726 BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002727 BUG_ON(order == 0);
Aneesh Kumar K.V47d38342012-07-31 16:41:54 -07002728 h = &hstates[hugetlb_max_hstate++];
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002729 h->order = order;
2730 h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
Andi Kleen8faa8b02008-07-23 21:27:48 -07002731 h->nr_huge_pages = 0;
2732 h->free_huge_pages = 0;
2733 for (i = 0; i < MAX_NUMNODES; ++i)
2734 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
Aneesh Kumar K.V0edaecf2012-07-31 16:42:07 -07002735 INIT_LIST_HEAD(&h->hugepage_activelist);
Andrew Morton54f18d32016-05-19 17:11:40 -07002736 h->next_nid_to_alloc = first_memory_node;
2737 h->next_nid_to_free = first_memory_node;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002738 snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
2739 huge_page_size(h)/1024);
Andi Kleen8faa8b02008-07-23 21:27:48 -07002740
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002741 parsed_hstate = h;
2742}
2743
Nick Piggine11bfbf2008-07-23 21:27:52 -07002744static int __init hugetlb_nrpages_setup(char *s)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002745{
2746 unsigned long *mhp;
Andi Kleen8faa8b02008-07-23 21:27:48 -07002747 static unsigned long *last_mhp;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002748
Vaishali Thakkar9fee0212016-05-19 17:11:04 -07002749 if (!parsed_valid_hugepagesz) {
2750 pr_warn("hugepages = %s preceded by "
2751 "an unsupported hugepagesz, ignoring\n", s);
2752 parsed_valid_hugepagesz = true;
2753 return 1;
2754 }
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002755 /*
Aneesh Kumar K.V47d38342012-07-31 16:41:54 -07002756 * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet,
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002757 * so this hugepages= parameter goes to the "default hstate".
2758 */
Vaishali Thakkar9fee0212016-05-19 17:11:04 -07002759 else if (!hugetlb_max_hstate)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002760 mhp = &default_hstate_max_huge_pages;
2761 else
2762 mhp = &parsed_hstate->max_huge_pages;
2763
Andi Kleen8faa8b02008-07-23 21:27:48 -07002764 if (mhp == last_mhp) {
Joe Perches598d8092016-03-17 14:19:44 -07002765 pr_warn("hugepages= specified twice without interleaving hugepagesz=, ignoring\n");
Andi Kleen8faa8b02008-07-23 21:27:48 -07002766 return 1;
2767 }
2768
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002769 if (sscanf(s, "%lu", mhp) <= 0)
2770 *mhp = 0;
2771
Andi Kleen8faa8b02008-07-23 21:27:48 -07002772 /*
2773 * Global state is always initialized later in hugetlb_init.
2774 * But we need to allocate >= MAX_ORDER hstates here early to still
2775 * use the bootmem allocator.
2776 */
Aneesh Kumar K.V47d38342012-07-31 16:41:54 -07002777 if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER)
Andi Kleen8faa8b02008-07-23 21:27:48 -07002778 hugetlb_hstate_alloc_pages(parsed_hstate);
2779
2780 last_mhp = mhp;
2781
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002782 return 1;
2783}
Nick Piggine11bfbf2008-07-23 21:27:52 -07002784__setup("hugepages=", hugetlb_nrpages_setup);
2785
2786static int __init hugetlb_default_setup(char *s)
2787{
2788 default_hstate_size = memparse(s, &s);
2789 return 1;
2790}
2791__setup("default_hugepagesz=", hugetlb_default_setup);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002792
Nishanth Aravamudan8a213462008-07-25 19:44:37 -07002793static unsigned int cpuset_mems_nr(unsigned int *array)
2794{
2795 int node;
2796 unsigned int nr = 0;
2797
2798 for_each_node_mask(node, cpuset_current_mems_allowed)
2799 nr += array[node];
2800
2801 return nr;
2802}
2803
2804#ifdef CONFIG_SYSCTL
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002805static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
2806 struct ctl_table *table, int write,
2807 void __user *buffer, size_t *length, loff_t *ppos)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002808{
Andi Kleene5ff2152008-07-23 21:27:42 -07002809 struct hstate *h = &default_hstate;
David Rientjes238d3c12014-08-06 16:06:51 -07002810 unsigned long tmp = h->max_huge_pages;
Michal Hocko08d4a242011-01-13 15:47:26 -08002811 int ret;
Andi Kleene5ff2152008-07-23 21:27:42 -07002812
Nishanth Aravamudan457c1b22014-05-06 12:50:00 -07002813 if (!hugepages_supported())
Jan Stancek86613622016-03-09 14:08:35 -08002814 return -EOPNOTSUPP;
Nishanth Aravamudan457c1b22014-05-06 12:50:00 -07002815
Andi Kleene5ff2152008-07-23 21:27:42 -07002816 table->data = &tmp;
2817 table->maxlen = sizeof(unsigned long);
Michal Hocko08d4a242011-01-13 15:47:26 -08002818 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2819 if (ret)
2820 goto out;
Andi Kleene5ff2152008-07-23 21:27:42 -07002821
David Rientjes238d3c12014-08-06 16:06:51 -07002822 if (write)
2823 ret = __nr_hugepages_store_common(obey_mempolicy, h,
2824 NUMA_NO_NODE, tmp, *length);
Michal Hocko08d4a242011-01-13 15:47:26 -08002825out:
2826 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002827}
Mel Gorman396faf02007-07-17 04:03:13 -07002828
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002829int hugetlb_sysctl_handler(struct ctl_table *table, int write,
2830 void __user *buffer, size_t *length, loff_t *ppos)
2831{
2832
2833 return hugetlb_sysctl_handler_common(false, table, write,
2834 buffer, length, ppos);
2835}
2836
2837#ifdef CONFIG_NUMA
2838int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
2839 void __user *buffer, size_t *length, loff_t *ppos)
2840{
2841 return hugetlb_sysctl_handler_common(true, table, write,
2842 buffer, length, ppos);
2843}
2844#endif /* CONFIG_NUMA */
2845
Nishanth Aravamudana3d0c6a2008-02-08 04:18:18 -08002846int hugetlb_overcommit_handler(struct ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07002847 void __user *buffer,
Nishanth Aravamudana3d0c6a2008-02-08 04:18:18 -08002848 size_t *length, loff_t *ppos)
2849{
Andi Kleena5516432008-07-23 21:27:41 -07002850 struct hstate *h = &default_hstate;
Andi Kleene5ff2152008-07-23 21:27:42 -07002851 unsigned long tmp;
Michal Hocko08d4a242011-01-13 15:47:26 -08002852 int ret;
Andi Kleene5ff2152008-07-23 21:27:42 -07002853
Nishanth Aravamudan457c1b22014-05-06 12:50:00 -07002854 if (!hugepages_supported())
Jan Stancek86613622016-03-09 14:08:35 -08002855 return -EOPNOTSUPP;
Nishanth Aravamudan457c1b22014-05-06 12:50:00 -07002856
Petr Holasekc033a932011-03-22 16:33:05 -07002857 tmp = h->nr_overcommit_huge_pages;
Andi Kleene5ff2152008-07-23 21:27:42 -07002858
Luiz Capitulinobae7f4a2014-06-04 16:07:08 -07002859 if (write && hstate_is_gigantic(h))
Eric B Munsonadbe8722011-01-13 15:47:27 -08002860 return -EINVAL;
2861
Andi Kleene5ff2152008-07-23 21:27:42 -07002862 table->data = &tmp;
2863 table->maxlen = sizeof(unsigned long);
Michal Hocko08d4a242011-01-13 15:47:26 -08002864 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2865 if (ret)
2866 goto out;
Andi Kleene5ff2152008-07-23 21:27:42 -07002867
2868 if (write) {
2869 spin_lock(&hugetlb_lock);
2870 h->nr_overcommit_huge_pages = tmp;
2871 spin_unlock(&hugetlb_lock);
2872 }
Michal Hocko08d4a242011-01-13 15:47:26 -08002873out:
2874 return ret;
Nishanth Aravamudana3d0c6a2008-02-08 04:18:18 -08002875}
2876
Linus Torvalds1da177e2005-04-16 15:20:36 -07002877#endif /* CONFIG_SYSCTL */
2878
Alexey Dobriyane1759c22008-10-15 23:50:22 +04002879void hugetlb_report_meminfo(struct seq_file *m)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002880{
Andi Kleena5516432008-07-23 21:27:41 -07002881 struct hstate *h = &default_hstate;
Nishanth Aravamudan457c1b22014-05-06 12:50:00 -07002882 if (!hugepages_supported())
2883 return;
Alexey Dobriyane1759c22008-10-15 23:50:22 +04002884 seq_printf(m,
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002885 "HugePages_Total: %5lu\n"
2886 "HugePages_Free: %5lu\n"
2887 "HugePages_Rsvd: %5lu\n"
2888 "HugePages_Surp: %5lu\n"
2889 "Hugepagesize: %8lu kB\n",
Andi Kleena5516432008-07-23 21:27:41 -07002890 h->nr_huge_pages,
2891 h->free_huge_pages,
2892 h->resv_huge_pages,
2893 h->surplus_huge_pages,
2894 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002895}
2896
2897int hugetlb_report_node_meminfo(int nid, char *buf)
2898{
Andi Kleena5516432008-07-23 21:27:41 -07002899 struct hstate *h = &default_hstate;
Nishanth Aravamudan457c1b22014-05-06 12:50:00 -07002900 if (!hugepages_supported())
2901 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002902 return sprintf(buf,
2903 "Node %d HugePages_Total: %5u\n"
Nishanth Aravamudana1de0912008-03-26 14:37:53 -07002904 "Node %d HugePages_Free: %5u\n"
2905 "Node %d HugePages_Surp: %5u\n",
Andi Kleena5516432008-07-23 21:27:41 -07002906 nid, h->nr_huge_pages_node[nid],
2907 nid, h->free_huge_pages_node[nid],
2908 nid, h->surplus_huge_pages_node[nid]);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002909}
2910
David Rientjes949f7ec2013-04-29 15:07:48 -07002911void hugetlb_show_meminfo(void)
2912{
2913 struct hstate *h;
2914 int nid;
2915
Nishanth Aravamudan457c1b22014-05-06 12:50:00 -07002916 if (!hugepages_supported())
2917 return;
2918
David Rientjes949f7ec2013-04-29 15:07:48 -07002919 for_each_node_state(nid, N_MEMORY)
2920 for_each_hstate(h)
2921 pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
2922 nid,
2923 h->nr_huge_pages_node[nid],
2924 h->free_huge_pages_node[nid],
2925 h->surplus_huge_pages_node[nid],
2926 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
2927}
2928
Naoya Horiguchi5d317b22015-11-05 18:47:14 -08002929void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm)
2930{
2931 seq_printf(m, "HugetlbPages:\t%8lu kB\n",
2932 atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10));
2933}
2934
Linus Torvalds1da177e2005-04-16 15:20:36 -07002935/* Return the number pages of memory we physically have, in PAGE_SIZE units. */
2936unsigned long hugetlb_total_pages(void)
2937{
Wanpeng Lid0028582013-03-22 15:04:40 -07002938 struct hstate *h;
2939 unsigned long nr_total_pages = 0;
2940
2941 for_each_hstate(h)
2942 nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h);
2943 return nr_total_pages;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002944}
Linus Torvalds1da177e2005-04-16 15:20:36 -07002945
Andi Kleena5516432008-07-23 21:27:41 -07002946static int hugetlb_acct_memory(struct hstate *h, long delta)
Mel Gormanfc1b8a72008-07-23 21:27:22 -07002947{
2948 int ret = -ENOMEM;
2949
2950 spin_lock(&hugetlb_lock);
2951 /*
2952 * When cpuset is configured, it breaks the strict hugetlb page
2953 * reservation as the accounting is done on a global variable. Such
2954 * reservation is completely rubbish in the presence of cpuset because
2955 * the reservation is not checked against page availability for the
2956 * current cpuset. Application can still potentially OOM'ed by kernel
2957 * with lack of free htlb page in cpuset that the task is in.
2958 * Attempt to enforce strict accounting with cpuset is almost
2959 * impossible (or too ugly) because cpuset is too fluid that
2960 * task or memory node can be dynamically moved between cpusets.
2961 *
2962 * The change of semantics for shared hugetlb mapping with cpuset is
2963 * undesirable. However, in order to preserve some of the semantics,
2964 * we fall back to check against current free page availability as
2965 * a best attempt and hopefully to minimize the impact of changing
2966 * semantics that cpuset has.
2967 */
2968 if (delta > 0) {
Andi Kleena5516432008-07-23 21:27:41 -07002969 if (gather_surplus_pages(h, delta) < 0)
Mel Gormanfc1b8a72008-07-23 21:27:22 -07002970 goto out;
2971
Andi Kleena5516432008-07-23 21:27:41 -07002972 if (delta > cpuset_mems_nr(h->free_huge_pages_node)) {
2973 return_unused_surplus_pages(h, delta);
Mel Gormanfc1b8a72008-07-23 21:27:22 -07002974 goto out;
2975 }
2976 }
2977
2978 ret = 0;
2979 if (delta < 0)
Andi Kleena5516432008-07-23 21:27:41 -07002980 return_unused_surplus_pages(h, (unsigned long) -delta);
Mel Gormanfc1b8a72008-07-23 21:27:22 -07002981
2982out:
2983 spin_unlock(&hugetlb_lock);
2984 return ret;
2985}
2986
Andy Whitcroft84afd992008-07-23 21:27:32 -07002987static void hugetlb_vm_op_open(struct vm_area_struct *vma)
2988{
Joonsoo Kimf522c3a2013-09-11 14:21:53 -07002989 struct resv_map *resv = vma_resv_map(vma);
Andy Whitcroft84afd992008-07-23 21:27:32 -07002990
2991 /*
2992 * This new VMA should share its siblings reservation map if present.
2993 * The VMA will only ever have a valid reservation map pointer where
2994 * it is being copied for another still existing VMA. As that VMA
Lucas De Marchi25985ed2011-03-30 22:57:33 -03002995 * has a reference to the reservation map it cannot disappear until
Andy Whitcroft84afd992008-07-23 21:27:32 -07002996 * after this open call completes. It is therefore safe to take a
2997 * new reference here without additional locking.
2998 */
Joonsoo Kim4e35f482014-04-03 14:47:30 -07002999 if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
Joonsoo Kimf522c3a2013-09-11 14:21:53 -07003000 kref_get(&resv->refs);
Andy Whitcroft84afd992008-07-23 21:27:32 -07003001}
3002
Mel Gormana1e78772008-07-23 21:27:23 -07003003static void hugetlb_vm_op_close(struct vm_area_struct *vma)
3004{
Andi Kleena5516432008-07-23 21:27:41 -07003005 struct hstate *h = hstate_vma(vma);
Joonsoo Kimf522c3a2013-09-11 14:21:53 -07003006 struct resv_map *resv = vma_resv_map(vma);
David Gibson90481622012-03-21 16:34:12 -07003007 struct hugepage_subpool *spool = subpool_vma(vma);
Joonsoo Kim4e35f482014-04-03 14:47:30 -07003008 unsigned long reserve, start, end;
Mike Kravetz1c5ecae2015-04-15 16:13:39 -07003009 long gbl_reserve;
Andy Whitcroft84afd992008-07-23 21:27:32 -07003010
Joonsoo Kim4e35f482014-04-03 14:47:30 -07003011 if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER))
3012 return;
Andy Whitcroft84afd992008-07-23 21:27:32 -07003013
Joonsoo Kim4e35f482014-04-03 14:47:30 -07003014 start = vma_hugecache_offset(h, vma, vma->vm_start);
3015 end = vma_hugecache_offset(h, vma, vma->vm_end);
Andy Whitcroft84afd992008-07-23 21:27:32 -07003016
Joonsoo Kim4e35f482014-04-03 14:47:30 -07003017 reserve = (end - start) - region_count(resv, start, end);
Andy Whitcroft84afd992008-07-23 21:27:32 -07003018
Joonsoo Kim4e35f482014-04-03 14:47:30 -07003019 kref_put(&resv->refs, resv_map_release);
3020
3021 if (reserve) {
Mike Kravetz1c5ecae2015-04-15 16:13:39 -07003022 /*
3023 * Decrement reserve counts. The global reserve count may be
3024 * adjusted if the subpool has a minimum size.
3025 */
3026 gbl_reserve = hugepage_subpool_put_pages(spool, reserve);
3027 hugetlb_acct_memory(h, -gbl_reserve);
Andy Whitcroft84afd992008-07-23 21:27:32 -07003028 }
Mel Gormana1e78772008-07-23 21:27:23 -07003029}
3030
Linus Torvalds1da177e2005-04-16 15:20:36 -07003031/*
3032 * We cannot handle pagefaults against hugetlb pages at all. They cause
3033 * handle_mm_fault() to try to instantiate regular-sized pages in the
3034 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
3035 * this far.
3036 */
Nick Piggind0217ac2007-07-19 01:47:03 -07003037static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003038{
3039 BUG();
Nick Piggind0217ac2007-07-19 01:47:03 -07003040 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003041}
3042
Alexey Dobriyanf0f37e22009-09-27 22:29:37 +04003043const struct vm_operations_struct hugetlb_vm_ops = {
Nick Piggind0217ac2007-07-19 01:47:03 -07003044 .fault = hugetlb_vm_op_fault,
Andy Whitcroft84afd992008-07-23 21:27:32 -07003045 .open = hugetlb_vm_op_open,
Mel Gormana1e78772008-07-23 21:27:23 -07003046 .close = hugetlb_vm_op_close,
Linus Torvalds1da177e2005-04-16 15:20:36 -07003047};
3048
David Gibson1e8f8892006-01-06 00:10:44 -08003049static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
3050 int writable)
David Gibson63551ae2005-06-21 17:14:44 -07003051{
3052 pte_t entry;
3053
David Gibson1e8f8892006-01-06 00:10:44 -08003054 if (writable) {
Gerald Schaefer106c9922013-04-29 15:07:23 -07003055 entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page,
3056 vma->vm_page_prot)));
David Gibson63551ae2005-06-21 17:14:44 -07003057 } else {
Gerald Schaefer106c9922013-04-29 15:07:23 -07003058 entry = huge_pte_wrprotect(mk_huge_pte(page,
3059 vma->vm_page_prot));
David Gibson63551ae2005-06-21 17:14:44 -07003060 }
3061 entry = pte_mkyoung(entry);
3062 entry = pte_mkhuge(entry);
Chris Metcalfd9ed9fa2012-04-01 14:01:34 -04003063 entry = arch_make_huge_pte(entry, vma, page, writable);
David Gibson63551ae2005-06-21 17:14:44 -07003064
3065 return entry;
3066}
3067
David Gibson1e8f8892006-01-06 00:10:44 -08003068static void set_huge_ptep_writable(struct vm_area_struct *vma,
3069 unsigned long address, pte_t *ptep)
3070{
3071 pte_t entry;
3072
Gerald Schaefer106c9922013-04-29 15:07:23 -07003073 entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep)));
Chris Forbes32f84522011-07-25 17:12:14 -07003074 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
Russell King4b3073e2009-12-18 16:40:18 +00003075 update_mmu_cache(vma, address, ptep);
David Gibson1e8f8892006-01-06 00:10:44 -08003076}
3077
Naoya Horiguchi4a705fe2014-06-23 13:22:03 -07003078static int is_hugetlb_entry_migration(pte_t pte)
3079{
3080 swp_entry_t swp;
3081
3082 if (huge_pte_none(pte) || pte_present(pte))
3083 return 0;
3084 swp = pte_to_swp_entry(pte);
3085 if (non_swap_entry(swp) && is_migration_entry(swp))
3086 return 1;
3087 else
3088 return 0;
3089}
3090
3091static int is_hugetlb_entry_hwpoisoned(pte_t pte)
3092{
3093 swp_entry_t swp;
3094
3095 if (huge_pte_none(pte) || pte_present(pte))
3096 return 0;
3097 swp = pte_to_swp_entry(pte);
3098 if (non_swap_entry(swp) && is_hwpoison_entry(swp))
3099 return 1;
3100 else
3101 return 0;
3102}
David Gibson1e8f8892006-01-06 00:10:44 -08003103
David Gibson63551ae2005-06-21 17:14:44 -07003104int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
3105 struct vm_area_struct *vma)
3106{
3107 pte_t *src_pte, *dst_pte, entry;
3108 struct page *ptepage;
Hugh Dickins1c598272005-10-19 21:23:43 -07003109 unsigned long addr;
David Gibson1e8f8892006-01-06 00:10:44 -08003110 int cow;
Andi Kleena5516432008-07-23 21:27:41 -07003111 struct hstate *h = hstate_vma(vma);
3112 unsigned long sz = huge_page_size(h);
Andreas Sandberge8569dd2014-01-21 15:49:09 -08003113 unsigned long mmun_start; /* For mmu_notifiers */
3114 unsigned long mmun_end; /* For mmu_notifiers */
3115 int ret = 0;
David Gibson1e8f8892006-01-06 00:10:44 -08003116
3117 cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
David Gibson63551ae2005-06-21 17:14:44 -07003118
Andreas Sandberge8569dd2014-01-21 15:49:09 -08003119 mmun_start = vma->vm_start;
3120 mmun_end = vma->vm_end;
3121 if (cow)
3122 mmu_notifier_invalidate_range_start(src, mmun_start, mmun_end);
3123
Andi Kleena5516432008-07-23 21:27:41 -07003124 for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003125 spinlock_t *src_ptl, *dst_ptl;
Hugh Dickinsc74df322005-10-29 18:16:23 -07003126 src_pte = huge_pte_offset(src, addr);
3127 if (!src_pte)
3128 continue;
Andi Kleena5516432008-07-23 21:27:41 -07003129 dst_pte = huge_pte_alloc(dst, addr, sz);
Andreas Sandberge8569dd2014-01-21 15:49:09 -08003130 if (!dst_pte) {
3131 ret = -ENOMEM;
3132 break;
3133 }
Larry Woodmanc5c99422008-01-24 05:49:25 -08003134
3135 /* If the pagetables are shared don't copy or take references */
3136 if (dst_pte == src_pte)
3137 continue;
3138
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003139 dst_ptl = huge_pte_lock(h, dst, dst_pte);
3140 src_ptl = huge_pte_lockptr(h, src, src_pte);
3141 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
Naoya Horiguchi4a705fe2014-06-23 13:22:03 -07003142 entry = huge_ptep_get(src_pte);
3143 if (huge_pte_none(entry)) { /* skip none entry */
3144 ;
3145 } else if (unlikely(is_hugetlb_entry_migration(entry) ||
3146 is_hugetlb_entry_hwpoisoned(entry))) {
3147 swp_entry_t swp_entry = pte_to_swp_entry(entry);
3148
3149 if (is_write_migration_entry(swp_entry) && cow) {
3150 /*
3151 * COW mappings require pages in both
3152 * parent and child to be set to read.
3153 */
3154 make_migration_entry_read(&swp_entry);
3155 entry = swp_entry_to_pte(swp_entry);
3156 set_huge_pte_at(src, addr, src_pte, entry);
3157 }
3158 set_huge_pte_at(dst, addr, dst_pte, entry);
3159 } else {
Joerg Roedel34ee6452014-11-13 13:46:09 +11003160 if (cow) {
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07003161 huge_ptep_set_wrprotect(src, addr, src_pte);
Joerg Roedel34ee6452014-11-13 13:46:09 +11003162 mmu_notifier_invalidate_range(src, mmun_start,
3163 mmun_end);
3164 }
Naoya Horiguchi0253d632014-07-23 14:00:19 -07003165 entry = huge_ptep_get(src_pte);
Hugh Dickins1c598272005-10-19 21:23:43 -07003166 ptepage = pte_page(entry);
3167 get_page(ptepage);
Kirill A. Shutemov53f92632016-01-15 16:53:42 -08003168 page_dup_rmap(ptepage, true);
Hugh Dickins1c598272005-10-19 21:23:43 -07003169 set_huge_pte_at(dst, addr, dst_pte, entry);
Naoya Horiguchi5d317b22015-11-05 18:47:14 -08003170 hugetlb_count_add(pages_per_huge_page(h), dst);
Hugh Dickins1c598272005-10-19 21:23:43 -07003171 }
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003172 spin_unlock(src_ptl);
3173 spin_unlock(dst_ptl);
David Gibson63551ae2005-06-21 17:14:44 -07003174 }
David Gibson63551ae2005-06-21 17:14:44 -07003175
Andreas Sandberge8569dd2014-01-21 15:49:09 -08003176 if (cow)
3177 mmu_notifier_invalidate_range_end(src, mmun_start, mmun_end);
3178
3179 return ret;
David Gibson63551ae2005-06-21 17:14:44 -07003180}
3181
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07003182void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
3183 unsigned long start, unsigned long end,
3184 struct page *ref_page)
David Gibson63551ae2005-06-21 17:14:44 -07003185{
3186 struct mm_struct *mm = vma->vm_mm;
3187 unsigned long address;
David Gibsonc7546f82005-08-05 11:59:35 -07003188 pte_t *ptep;
David Gibson63551ae2005-06-21 17:14:44 -07003189 pte_t pte;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003190 spinlock_t *ptl;
David Gibson63551ae2005-06-21 17:14:44 -07003191 struct page *page;
Andi Kleena5516432008-07-23 21:27:41 -07003192 struct hstate *h = hstate_vma(vma);
3193 unsigned long sz = huge_page_size(h);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07003194 const unsigned long mmun_start = start; /* For mmu_notifiers */
3195 const unsigned long mmun_end = end; /* For mmu_notifiers */
Andi Kleena5516432008-07-23 21:27:41 -07003196
David Gibson63551ae2005-06-21 17:14:44 -07003197 WARN_ON(!is_vm_hugetlb_page(vma));
Andi Kleena5516432008-07-23 21:27:41 -07003198 BUG_ON(start & ~huge_page_mask(h));
3199 BUG_ON(end & ~huge_page_mask(h));
David Gibson63551ae2005-06-21 17:14:44 -07003200
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07003201 tlb_start_vma(tlb, vma);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07003202 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
Hillf Danton569f48b2014-12-10 15:44:41 -08003203 address = start;
Hillf Danton569f48b2014-12-10 15:44:41 -08003204 for (; address < end; address += sz) {
David Gibsonc7546f82005-08-05 11:59:35 -07003205 ptep = huge_pte_offset(mm, address);
Adam Litke4c887262005-10-29 18:16:46 -07003206 if (!ptep)
David Gibsonc7546f82005-08-05 11:59:35 -07003207 continue;
3208
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003209 ptl = huge_pte_lock(h, mm, ptep);
Aneesh Kumar K.V31d49da2016-07-26 15:24:06 -07003210 if (huge_pmd_unshare(mm, &address, ptep)) {
3211 spin_unlock(ptl);
3212 continue;
3213 }
Chen, Kenneth W39dde652006-12-06 20:32:03 -08003214
Hillf Danton66293262012-03-23 15:01:48 -07003215 pte = huge_ptep_get(ptep);
Aneesh Kumar K.V31d49da2016-07-26 15:24:06 -07003216 if (huge_pte_none(pte)) {
3217 spin_unlock(ptl);
3218 continue;
3219 }
Hillf Danton66293262012-03-23 15:01:48 -07003220
3221 /*
Naoya Horiguchi9fbc1f62015-02-11 15:25:32 -08003222 * Migrating hugepage or HWPoisoned hugepage is already
3223 * unmapped and its refcount is dropped, so just clear pte here.
Hillf Danton66293262012-03-23 15:01:48 -07003224 */
Naoya Horiguchi9fbc1f62015-02-11 15:25:32 -08003225 if (unlikely(!pte_present(pte))) {
Gerald Schaefer106c9922013-04-29 15:07:23 -07003226 huge_pte_clear(mm, address, ptep);
Aneesh Kumar K.V31d49da2016-07-26 15:24:06 -07003227 spin_unlock(ptl);
3228 continue;
Naoya Horiguchi8c4894c2012-12-12 13:52:28 -08003229 }
Hillf Danton66293262012-03-23 15:01:48 -07003230
3231 page = pte_page(pte);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003232 /*
3233 * If a reference page is supplied, it is because a specific
3234 * page is being unmapped, not a range. Ensure the page we
3235 * are about to unmap is the actual page of interest.
3236 */
3237 if (ref_page) {
Aneesh Kumar K.V31d49da2016-07-26 15:24:06 -07003238 if (page != ref_page) {
3239 spin_unlock(ptl);
3240 continue;
3241 }
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003242 /*
3243 * Mark the VMA as having unmapped its page so that
3244 * future faults in this VMA will fail rather than
3245 * looking like data was lost
3246 */
3247 set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
3248 }
3249
David Gibsonc7546f82005-08-05 11:59:35 -07003250 pte = huge_ptep_get_and_clear(mm, address, ptep);
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07003251 tlb_remove_tlb_entry(tlb, ptep, address);
Gerald Schaefer106c9922013-04-29 15:07:23 -07003252 if (huge_pte_dirty(pte))
Ken Chen6649a382007-02-08 14:20:27 -08003253 set_page_dirty(page);
Hillf Danton9e811302012-03-21 16:34:03 -07003254
Naoya Horiguchi5d317b22015-11-05 18:47:14 -08003255 hugetlb_count_sub(pages_per_huge_page(h), mm);
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -08003256 page_remove_rmap(page, true);
Aneesh Kumar K.V31d49da2016-07-26 15:24:06 -07003257
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003258 spin_unlock(ptl);
Aneesh Kumar K.Ve77b0852016-07-26 15:24:12 -07003259 tlb_remove_page_size(tlb, page, huge_page_size(h));
Aneesh Kumar K.V31d49da2016-07-26 15:24:06 -07003260 /*
3261 * Bail out after unmapping reference page if supplied
3262 */
3263 if (ref_page)
3264 break;
Chen, Kenneth Wfe1668a2006-10-04 02:15:24 -07003265 }
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07003266 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07003267 tlb_end_vma(tlb, vma);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003268}
David Gibson63551ae2005-06-21 17:14:44 -07003269
Mel Gormand8333522012-07-31 16:46:20 -07003270void __unmap_hugepage_range_final(struct mmu_gather *tlb,
3271 struct vm_area_struct *vma, unsigned long start,
3272 unsigned long end, struct page *ref_page)
3273{
3274 __unmap_hugepage_range(tlb, vma, start, end, ref_page);
3275
3276 /*
3277 * Clear this flag so that x86's huge_pmd_share page_table_shareable
3278 * test will fail on a vma being torn down, and not grab a page table
3279 * on its way out. We're lucky that the flag has such an appropriate
3280 * name, and can in fact be safely cleared here. We could clear it
3281 * before the __unmap_hugepage_range above, but all that's necessary
Davidlohr Buesoc8c06ef2014-12-12 16:54:24 -08003282 * is to clear it before releasing the i_mmap_rwsem. This works
Mel Gormand8333522012-07-31 16:46:20 -07003283 * because in the context this is called, the VMA is about to be
Davidlohr Buesoc8c06ef2014-12-12 16:54:24 -08003284 * destroyed and the i_mmap_rwsem is held.
Mel Gormand8333522012-07-31 16:46:20 -07003285 */
3286 vma->vm_flags &= ~VM_MAYSHARE;
3287}
3288
Chen, Kenneth W502717f2006-10-11 01:20:46 -07003289void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003290 unsigned long end, struct page *ref_page)
Chen, Kenneth W502717f2006-10-11 01:20:46 -07003291{
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07003292 struct mm_struct *mm;
3293 struct mmu_gather tlb;
3294
3295 mm = vma->vm_mm;
3296
Linus Torvalds2b047252013-08-15 11:42:25 -07003297 tlb_gather_mmu(&tlb, mm, start, end);
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07003298 __unmap_hugepage_range(&tlb, vma, start, end, ref_page);
3299 tlb_finish_mmu(&tlb, start, end);
Chen, Kenneth W502717f2006-10-11 01:20:46 -07003300}
3301
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003302/*
3303 * This is called when the original mapper is failing to COW a MAP_PRIVATE
3304 * mappping it owns the reserve page for. The intention is to unmap the page
3305 * from other VMAs and let the children be SIGKILLed if they are faulting the
3306 * same region.
3307 */
Davidlohr Bueso2f4612a2014-08-06 16:06:45 -07003308static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
3309 struct page *page, unsigned long address)
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003310{
Adam Litke75266742008-11-12 13:24:56 -08003311 struct hstate *h = hstate_vma(vma);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003312 struct vm_area_struct *iter_vma;
3313 struct address_space *mapping;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003314 pgoff_t pgoff;
3315
3316 /*
3317 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
3318 * from page cache lookup which is in HPAGE_SIZE units.
3319 */
Adam Litke75266742008-11-12 13:24:56 -08003320 address = address & huge_page_mask(h);
Michal Hocko36e4f202012-10-08 16:33:31 -07003321 pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) +
3322 vma->vm_pgoff;
Al Viro93c76a32015-12-04 23:45:44 -05003323 mapping = vma->vm_file->f_mapping;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003324
Mel Gorman4eb2b1d2009-12-14 17:59:53 -08003325 /*
3326 * Take the mapping lock for the duration of the table walk. As
3327 * this mapping should be shared between all the VMAs,
3328 * __unmap_hugepage_range() is called as the lock is already held
3329 */
Davidlohr Bueso83cde9e2014-12-12 16:54:21 -08003330 i_mmap_lock_write(mapping);
Michel Lespinasse6b2dbba2012-10-08 16:31:25 -07003331 vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) {
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003332 /* Do not unmap the current VMA */
3333 if (iter_vma == vma)
3334 continue;
3335
3336 /*
Mel Gorman2f84a892015-10-01 15:36:57 -07003337 * Shared VMAs have their own reserves and do not affect
3338 * MAP_PRIVATE accounting but it is possible that a shared
3339 * VMA is using the same page so check and skip such VMAs.
3340 */
3341 if (iter_vma->vm_flags & VM_MAYSHARE)
3342 continue;
3343
3344 /*
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003345 * Unmap the page from other VMAs without their own reserves.
3346 * They get marked to be SIGKILLed if they fault in these
3347 * areas. This is because a future no-page fault on this VMA
3348 * could insert a zeroed page instead of the data existing
3349 * from the time of fork. This would look like data corruption
3350 */
3351 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07003352 unmap_hugepage_range(iter_vma, address,
3353 address + huge_page_size(h), page);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003354 }
Davidlohr Bueso83cde9e2014-12-12 16:54:21 -08003355 i_mmap_unlock_write(mapping);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003356}
3357
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09003358/*
3359 * Hugetlb_cow() should be called with page lock of the original hugepage held.
Michal Hockoef009b22012-01-10 15:07:21 -08003360 * Called with hugetlb_instantiation_mutex held and pte_page locked so we
3361 * cannot race with other handlers or page migration.
3362 * Keep the pte_same checks anyway to make transition from the mutex easier.
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09003363 */
David Gibson1e8f8892006-01-06 00:10:44 -08003364static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003365 unsigned long address, pte_t *ptep, pte_t pte,
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003366 struct page *pagecache_page, spinlock_t *ptl)
David Gibson1e8f8892006-01-06 00:10:44 -08003367{
Andi Kleena5516432008-07-23 21:27:41 -07003368 struct hstate *h = hstate_vma(vma);
David Gibson1e8f8892006-01-06 00:10:44 -08003369 struct page *old_page, *new_page;
Davidlohr Buesoad4404a2014-08-06 16:06:47 -07003370 int ret = 0, outside_reserve = 0;
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07003371 unsigned long mmun_start; /* For mmu_notifiers */
3372 unsigned long mmun_end; /* For mmu_notifiers */
David Gibson1e8f8892006-01-06 00:10:44 -08003373
3374 old_page = pte_page(pte);
3375
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003376retry_avoidcopy:
David Gibson1e8f8892006-01-06 00:10:44 -08003377 /* If no-one else is actually using this page, avoid the copy
3378 * and just make the page writable */
Joonsoo Kim37a21402013-09-11 14:21:04 -07003379 if (page_mapcount(old_page) == 1 && PageAnon(old_page)) {
Hugh Dickins5a499732016-07-14 12:07:38 -07003380 page_move_anon_rmap(old_page, vma);
David Gibson1e8f8892006-01-06 00:10:44 -08003381 set_huge_ptep_writable(vma, address, ptep);
Nick Piggin83c54072007-07-19 01:47:05 -07003382 return 0;
David Gibson1e8f8892006-01-06 00:10:44 -08003383 }
3384
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003385 /*
3386 * If the process that created a MAP_PRIVATE mapping is about to
3387 * perform a COW due to a shared page count, attempt to satisfy
3388 * the allocation without using the existing reserves. The pagecache
3389 * page is used to determine if the reserve at this address was
3390 * consumed or not. If reserves were used, a partial faulted mapping
3391 * at the time of fork() could consume its reserves on COW instead
3392 * of the full address range.
3393 */
Joonsoo Kim5944d012013-09-11 14:21:55 -07003394 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003395 old_page != pagecache_page)
3396 outside_reserve = 1;
3397
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03003398 get_page(old_page);
Larry Woodmanb76c8cf2009-12-14 17:59:37 -08003399
Davidlohr Buesoad4404a2014-08-06 16:06:47 -07003400 /*
3401 * Drop page table lock as buddy allocator may be called. It will
3402 * be acquired again before returning to the caller, as expected.
3403 */
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003404 spin_unlock(ptl);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003405 new_page = alloc_huge_page(vma, address, outside_reserve);
David Gibson1e8f8892006-01-06 00:10:44 -08003406
Adam Litke2fc39ce2007-11-14 16:59:39 -08003407 if (IS_ERR(new_page)) {
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003408 /*
3409 * If a process owning a MAP_PRIVATE mapping fails to COW,
3410 * it is due to references held by a child and an insufficient
3411 * huge page pool. To guarantee the original mappers
3412 * reliability, unmap the page from child processes. The child
3413 * may get SIGKILLed if it later faults.
3414 */
3415 if (outside_reserve) {
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03003416 put_page(old_page);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003417 BUG_ON(huge_pte_none(pte));
Davidlohr Bueso2f4612a2014-08-06 16:06:45 -07003418 unmap_ref_private(mm, vma, old_page, address);
3419 BUG_ON(huge_pte_none(pte));
3420 spin_lock(ptl);
3421 ptep = huge_pte_offset(mm, address & huge_page_mask(h));
3422 if (likely(ptep &&
3423 pte_same(huge_ptep_get(ptep), pte)))
3424 goto retry_avoidcopy;
3425 /*
3426 * race occurs while re-acquiring page table
3427 * lock, and our job is done.
3428 */
3429 return 0;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003430 }
3431
Davidlohr Buesoad4404a2014-08-06 16:06:47 -07003432 ret = (PTR_ERR(new_page) == -ENOMEM) ?
3433 VM_FAULT_OOM : VM_FAULT_SIGBUS;
3434 goto out_release_old;
David Gibson1e8f8892006-01-06 00:10:44 -08003435 }
3436
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09003437 /*
3438 * When the original hugepage is shared one, it does not have
3439 * anon_vma prepared.
3440 */
Dean Nelson44e2aa92010-10-26 14:22:08 -07003441 if (unlikely(anon_vma_prepare(vma))) {
Davidlohr Buesoad4404a2014-08-06 16:06:47 -07003442 ret = VM_FAULT_OOM;
3443 goto out_release_all;
Dean Nelson44e2aa92010-10-26 14:22:08 -07003444 }
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09003445
Andrea Arcangeli47ad8472011-01-13 15:46:47 -08003446 copy_user_huge_page(new_page, old_page, address, vma,
3447 pages_per_huge_page(h));
Nick Piggin0ed361d2008-02-04 22:29:34 -08003448 __SetPageUptodate(new_page);
Naoya Horiguchibcc54222015-04-15 16:14:38 -07003449 set_page_huge_active(new_page);
David Gibson1e8f8892006-01-06 00:10:44 -08003450
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07003451 mmun_start = address & huge_page_mask(h);
3452 mmun_end = mmun_start + huge_page_size(h);
3453 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
Davidlohr Buesoad4404a2014-08-06 16:06:47 -07003454
Larry Woodmanb76c8cf2009-12-14 17:59:37 -08003455 /*
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003456 * Retake the page table lock to check for racing updates
Larry Woodmanb76c8cf2009-12-14 17:59:37 -08003457 * before the page tables are altered
3458 */
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003459 spin_lock(ptl);
Andi Kleena5516432008-07-23 21:27:41 -07003460 ptep = huge_pte_offset(mm, address & huge_page_mask(h));
Naoya Horiguchia9af0c52014-04-07 15:36:54 -07003461 if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) {
Joonsoo Kim07443a82013-09-11 14:21:58 -07003462 ClearPagePrivate(new_page);
3463
David Gibson1e8f8892006-01-06 00:10:44 -08003464 /* Break COW */
Gerald Schaefer8fe627e2008-04-28 02:13:28 -07003465 huge_ptep_clear_flush(vma, address, ptep);
Joerg Roedel34ee6452014-11-13 13:46:09 +11003466 mmu_notifier_invalidate_range(mm, mmun_start, mmun_end);
David Gibson1e8f8892006-01-06 00:10:44 -08003467 set_huge_pte_at(mm, address, ptep,
3468 make_huge_pte(vma, new_page, 1));
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -08003469 page_remove_rmap(old_page, true);
Naoya Horiguchicd67f0d2010-09-10 13:23:04 +09003470 hugepage_add_new_anon_rmap(new_page, vma, address);
David Gibson1e8f8892006-01-06 00:10:44 -08003471 /* Make the old page be freed below */
3472 new_page = old_page;
3473 }
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003474 spin_unlock(ptl);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07003475 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
Davidlohr Buesoad4404a2014-08-06 16:06:47 -07003476out_release_all:
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03003477 put_page(new_page);
Davidlohr Buesoad4404a2014-08-06 16:06:47 -07003478out_release_old:
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03003479 put_page(old_page);
Joonsoo Kim83120342013-09-11 14:21:57 -07003480
Davidlohr Buesoad4404a2014-08-06 16:06:47 -07003481 spin_lock(ptl); /* Caller expects lock to be held */
3482 return ret;
David Gibson1e8f8892006-01-06 00:10:44 -08003483}
3484
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003485/* Return the pagecache page at a given address within a VMA */
Andi Kleena5516432008-07-23 21:27:41 -07003486static struct page *hugetlbfs_pagecache_page(struct hstate *h,
3487 struct vm_area_struct *vma, unsigned long address)
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003488{
3489 struct address_space *mapping;
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -07003490 pgoff_t idx;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003491
3492 mapping = vma->vm_file->f_mapping;
Andi Kleena5516432008-07-23 21:27:41 -07003493 idx = vma_hugecache_offset(h, vma, address);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003494
3495 return find_lock_page(mapping, idx);
3496}
3497
Hugh Dickins3ae77f42009-09-21 17:03:33 -07003498/*
3499 * Return whether there is a pagecache page to back given address within VMA.
3500 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
3501 */
3502static bool hugetlbfs_pagecache_present(struct hstate *h,
Hugh Dickins2a15efc2009-09-21 17:03:27 -07003503 struct vm_area_struct *vma, unsigned long address)
3504{
3505 struct address_space *mapping;
3506 pgoff_t idx;
3507 struct page *page;
3508
3509 mapping = vma->vm_file->f_mapping;
3510 idx = vma_hugecache_offset(h, vma, address);
3511
3512 page = find_get_page(mapping, idx);
3513 if (page)
3514 put_page(page);
3515 return page != NULL;
3516}
3517
Mike Kravetzab76ad52015-09-08 15:01:50 -07003518int huge_add_to_page_cache(struct page *page, struct address_space *mapping,
3519 pgoff_t idx)
3520{
3521 struct inode *inode = mapping->host;
3522 struct hstate *h = hstate_inode(inode);
3523 int err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
3524
3525 if (err)
3526 return err;
3527 ClearPagePrivate(page);
3528
3529 spin_lock(&inode->i_lock);
3530 inode->i_blocks += blocks_per_huge_page(h);
3531 spin_unlock(&inode->i_lock);
3532 return 0;
3533}
3534
Robert P. J. Daya1ed3dd2007-07-17 04:03:33 -07003535static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
Davidlohr Bueso8382d912014-04-03 14:47:31 -07003536 struct address_space *mapping, pgoff_t idx,
3537 unsigned long address, pte_t *ptep, unsigned int flags)
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01003538{
Andi Kleena5516432008-07-23 21:27:41 -07003539 struct hstate *h = hstate_vma(vma);
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01003540 int ret = VM_FAULT_SIGBUS;
Hillf Danton409eb8c2012-01-20 14:34:13 -08003541 int anon_rmap = 0;
Adam Litke4c887262005-10-29 18:16:46 -07003542 unsigned long size;
Adam Litke4c887262005-10-29 18:16:46 -07003543 struct page *page;
David Gibson1e8f8892006-01-06 00:10:44 -08003544 pte_t new_pte;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003545 spinlock_t *ptl;
Adam Litke4c887262005-10-29 18:16:46 -07003546
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003547 /*
3548 * Currently, we are forced to kill the process in the event the
3549 * original mapper has unmapped pages from the child due to a failed
Lucas De Marchi25985ed2011-03-30 22:57:33 -03003550 * COW. Warn that such a situation has occurred as it may not be obvious
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003551 */
3552 if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
Geoffrey Thomas910154d2016-03-09 14:08:04 -08003553 pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n",
Andrew Mortonffb22af2013-02-22 16:32:08 -08003554 current->pid);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003555 return ret;
3556 }
3557
Adam Litke4c887262005-10-29 18:16:46 -07003558 /*
3559 * Use page lock to guard against racing truncation
3560 * before we get page_table_lock.
3561 */
Christoph Lameter6bda6662006-01-06 00:10:49 -08003562retry:
3563 page = find_lock_page(mapping, idx);
3564 if (!page) {
Andi Kleena5516432008-07-23 21:27:41 -07003565 size = i_size_read(mapping->host) >> huge_page_shift(h);
Hugh Dickinsebed4bf2006-10-28 10:38:43 -07003566 if (idx >= size)
3567 goto out;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003568 page = alloc_huge_page(vma, address, 0);
Adam Litke2fc39ce2007-11-14 16:59:39 -08003569 if (IS_ERR(page)) {
Aneesh Kumar K.V76dcee72012-07-31 16:41:57 -07003570 ret = PTR_ERR(page);
3571 if (ret == -ENOMEM)
3572 ret = VM_FAULT_OOM;
3573 else
3574 ret = VM_FAULT_SIGBUS;
Christoph Lameter6bda6662006-01-06 00:10:49 -08003575 goto out;
3576 }
Andrea Arcangeli47ad8472011-01-13 15:46:47 -08003577 clear_huge_page(page, address, pages_per_huge_page(h));
Nick Piggin0ed361d2008-02-04 22:29:34 -08003578 __SetPageUptodate(page);
Naoya Horiguchibcc54222015-04-15 16:14:38 -07003579 set_page_huge_active(page);
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01003580
Mel Gormanf83a2752009-05-28 14:34:40 -07003581 if (vma->vm_flags & VM_MAYSHARE) {
Mike Kravetzab76ad52015-09-08 15:01:50 -07003582 int err = huge_add_to_page_cache(page, mapping, idx);
Christoph Lameter6bda6662006-01-06 00:10:49 -08003583 if (err) {
3584 put_page(page);
Christoph Lameter6bda6662006-01-06 00:10:49 -08003585 if (err == -EEXIST)
3586 goto retry;
3587 goto out;
3588 }
Mel Gorman23be7462010-04-23 13:17:56 -04003589 } else {
Christoph Lameter6bda6662006-01-06 00:10:49 -08003590 lock_page(page);
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09003591 if (unlikely(anon_vma_prepare(vma))) {
3592 ret = VM_FAULT_OOM;
3593 goto backout_unlocked;
3594 }
Hillf Danton409eb8c2012-01-20 14:34:13 -08003595 anon_rmap = 1;
Mel Gorman23be7462010-04-23 13:17:56 -04003596 }
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09003597 } else {
Naoya Horiguchi998b4382010-09-08 10:19:32 +09003598 /*
3599 * If memory error occurs between mmap() and fault, some process
3600 * don't have hwpoisoned swap entry for errored virtual address.
3601 * So we need to block hugepage fault by PG_hwpoison bit check.
3602 */
3603 if (unlikely(PageHWPoison(page))) {
Chris Forbes32f84522011-07-25 17:12:14 -07003604 ret = VM_FAULT_HWPOISON |
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07003605 VM_FAULT_SET_HINDEX(hstate_index(h));
Naoya Horiguchi998b4382010-09-08 10:19:32 +09003606 goto backout_unlocked;
3607 }
Christoph Lameter6bda6662006-01-06 00:10:49 -08003608 }
David Gibson1e8f8892006-01-06 00:10:44 -08003609
Andy Whitcroft57303d82008-08-12 15:08:47 -07003610 /*
3611 * If we are going to COW a private mapping later, we examine the
3612 * pending reservations for this page now. This will ensure that
3613 * any allocations necessary to record that reservation occur outside
3614 * the spinlock.
3615 */
Mike Kravetz5e911372015-09-08 15:01:28 -07003616 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
Andy Whitcroft2b267362008-08-12 15:08:49 -07003617 if (vma_needs_reservation(h, vma, address) < 0) {
3618 ret = VM_FAULT_OOM;
3619 goto backout_unlocked;
3620 }
Mike Kravetz5e911372015-09-08 15:01:28 -07003621 /* Just decrements count, does not deallocate */
Mike Kravetzfeba16e2015-09-08 15:01:31 -07003622 vma_end_reservation(h, vma, address);
Mike Kravetz5e911372015-09-08 15:01:28 -07003623 }
Andy Whitcroft57303d82008-08-12 15:08:47 -07003624
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003625 ptl = huge_pte_lockptr(h, mm, ptep);
3626 spin_lock(ptl);
Andi Kleena5516432008-07-23 21:27:41 -07003627 size = i_size_read(mapping->host) >> huge_page_shift(h);
Adam Litke4c887262005-10-29 18:16:46 -07003628 if (idx >= size)
3629 goto backout;
3630
Nick Piggin83c54072007-07-19 01:47:05 -07003631 ret = 0;
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07003632 if (!huge_pte_none(huge_ptep_get(ptep)))
Adam Litke4c887262005-10-29 18:16:46 -07003633 goto backout;
3634
Joonsoo Kim07443a82013-09-11 14:21:58 -07003635 if (anon_rmap) {
3636 ClearPagePrivate(page);
Hillf Danton409eb8c2012-01-20 14:34:13 -08003637 hugepage_add_new_anon_rmap(page, vma, address);
Choi Gi-yongac714902014-04-07 15:37:36 -07003638 } else
Kirill A. Shutemov53f92632016-01-15 16:53:42 -08003639 page_dup_rmap(page, true);
David Gibson1e8f8892006-01-06 00:10:44 -08003640 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
3641 && (vma->vm_flags & VM_SHARED)));
3642 set_huge_pte_at(mm, address, ptep, new_pte);
3643
Naoya Horiguchi5d317b22015-11-05 18:47:14 -08003644 hugetlb_count_add(pages_per_huge_page(h), mm);
Hugh Dickins788c7df2009-06-23 13:49:05 +01003645 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
David Gibson1e8f8892006-01-06 00:10:44 -08003646 /* Optimization, do the COW without a second fault */
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003647 ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page, ptl);
David Gibson1e8f8892006-01-06 00:10:44 -08003648 }
3649
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003650 spin_unlock(ptl);
Adam Litke4c887262005-10-29 18:16:46 -07003651 unlock_page(page);
3652out:
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01003653 return ret;
Adam Litke4c887262005-10-29 18:16:46 -07003654
3655backout:
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003656 spin_unlock(ptl);
Andy Whitcroft2b267362008-08-12 15:08:49 -07003657backout_unlocked:
Adam Litke4c887262005-10-29 18:16:46 -07003658 unlock_page(page);
3659 put_page(page);
3660 goto out;
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01003661}
3662
Davidlohr Bueso8382d912014-04-03 14:47:31 -07003663#ifdef CONFIG_SMP
Mike Kravetzc672c7f2015-09-08 15:01:35 -07003664u32 hugetlb_fault_mutex_hash(struct hstate *h, struct mm_struct *mm,
Davidlohr Bueso8382d912014-04-03 14:47:31 -07003665 struct vm_area_struct *vma,
3666 struct address_space *mapping,
3667 pgoff_t idx, unsigned long address)
3668{
3669 unsigned long key[2];
3670 u32 hash;
3671
3672 if (vma->vm_flags & VM_SHARED) {
3673 key[0] = (unsigned long) mapping;
3674 key[1] = idx;
3675 } else {
3676 key[0] = (unsigned long) mm;
3677 key[1] = address >> huge_page_shift(h);
3678 }
3679
3680 hash = jhash2((u32 *)&key, sizeof(key)/sizeof(u32), 0);
3681
3682 return hash & (num_fault_mutexes - 1);
3683}
3684#else
3685/*
3686 * For uniprocesor systems we always use a single mutex, so just
3687 * return 0 and avoid the hashing overhead.
3688 */
Mike Kravetzc672c7f2015-09-08 15:01:35 -07003689u32 hugetlb_fault_mutex_hash(struct hstate *h, struct mm_struct *mm,
Davidlohr Bueso8382d912014-04-03 14:47:31 -07003690 struct vm_area_struct *vma,
3691 struct address_space *mapping,
3692 pgoff_t idx, unsigned long address)
3693{
3694 return 0;
3695}
3696#endif
3697
Adam Litke86e52162006-01-06 00:10:43 -08003698int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
Hugh Dickins788c7df2009-06-23 13:49:05 +01003699 unsigned long address, unsigned int flags)
Adam Litke86e52162006-01-06 00:10:43 -08003700{
Davidlohr Bueso8382d912014-04-03 14:47:31 -07003701 pte_t *ptep, entry;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003702 spinlock_t *ptl;
David Gibson1e8f8892006-01-06 00:10:44 -08003703 int ret;
Davidlohr Bueso8382d912014-04-03 14:47:31 -07003704 u32 hash;
3705 pgoff_t idx;
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09003706 struct page *page = NULL;
Andy Whitcroft57303d82008-08-12 15:08:47 -07003707 struct page *pagecache_page = NULL;
Andi Kleena5516432008-07-23 21:27:41 -07003708 struct hstate *h = hstate_vma(vma);
Davidlohr Bueso8382d912014-04-03 14:47:31 -07003709 struct address_space *mapping;
Naoya Horiguchi0f792cf2015-02-11 15:25:25 -08003710 int need_wait_lock = 0;
Adam Litke86e52162006-01-06 00:10:43 -08003711
KAMEZAWA Hiroyuki1e16a532012-01-10 15:07:22 -08003712 address &= huge_page_mask(h);
3713
Naoya Horiguchifd6a03ed2010-05-28 09:29:21 +09003714 ptep = huge_pte_offset(mm, address);
3715 if (ptep) {
3716 entry = huge_ptep_get(ptep);
Naoya Horiguchi290408d2010-09-08 10:19:35 +09003717 if (unlikely(is_hugetlb_entry_migration(entry))) {
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003718 migration_entry_wait_huge(vma, mm, ptep);
Naoya Horiguchi290408d2010-09-08 10:19:35 +09003719 return 0;
3720 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
Chris Forbes32f84522011-07-25 17:12:14 -07003721 return VM_FAULT_HWPOISON_LARGE |
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07003722 VM_FAULT_SET_HINDEX(hstate_index(h));
Naoya Horiguchi0d777df2015-12-11 13:40:49 -08003723 } else {
3724 ptep = huge_pte_alloc(mm, address, huge_page_size(h));
3725 if (!ptep)
3726 return VM_FAULT_OOM;
Naoya Horiguchifd6a03ed2010-05-28 09:29:21 +09003727 }
3728
Davidlohr Bueso8382d912014-04-03 14:47:31 -07003729 mapping = vma->vm_file->f_mapping;
3730 idx = vma_hugecache_offset(h, vma, address);
3731
David Gibson3935baa2006-03-22 00:08:53 -08003732 /*
3733 * Serialize hugepage allocation and instantiation, so that we don't
3734 * get spurious allocation failures if two CPUs race to instantiate
3735 * the same page in the page cache.
3736 */
Mike Kravetzc672c7f2015-09-08 15:01:35 -07003737 hash = hugetlb_fault_mutex_hash(h, mm, vma, mapping, idx, address);
3738 mutex_lock(&hugetlb_fault_mutex_table[hash]);
Davidlohr Bueso8382d912014-04-03 14:47:31 -07003739
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07003740 entry = huge_ptep_get(ptep);
3741 if (huge_pte_none(entry)) {
Davidlohr Bueso8382d912014-04-03 14:47:31 -07003742 ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags);
David Gibsonb4d1d992008-10-15 22:01:11 -07003743 goto out_mutex;
David Gibson3935baa2006-03-22 00:08:53 -08003744 }
Adam Litke86e52162006-01-06 00:10:43 -08003745
Nick Piggin83c54072007-07-19 01:47:05 -07003746 ret = 0;
David Gibson1e8f8892006-01-06 00:10:44 -08003747
Andy Whitcroft57303d82008-08-12 15:08:47 -07003748 /*
Naoya Horiguchi0f792cf2015-02-11 15:25:25 -08003749 * entry could be a migration/hwpoison entry at this point, so this
3750 * check prevents the kernel from going below assuming that we have
3751 * a active hugepage in pagecache. This goto expects the 2nd page fault,
3752 * and is_hugetlb_entry_(migration|hwpoisoned) check will properly
3753 * handle it.
3754 */
3755 if (!pte_present(entry))
3756 goto out_mutex;
3757
3758 /*
Andy Whitcroft57303d82008-08-12 15:08:47 -07003759 * If we are going to COW the mapping later, we examine the pending
3760 * reservations for this page now. This will ensure that any
3761 * allocations necessary to record that reservation occur outside the
3762 * spinlock. For private mappings, we also lookup the pagecache
3763 * page now as it is used to determine if a reservation has been
3764 * consumed.
3765 */
Gerald Schaefer106c9922013-04-29 15:07:23 -07003766 if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) {
Andy Whitcroft2b267362008-08-12 15:08:49 -07003767 if (vma_needs_reservation(h, vma, address) < 0) {
3768 ret = VM_FAULT_OOM;
David Gibsonb4d1d992008-10-15 22:01:11 -07003769 goto out_mutex;
Andy Whitcroft2b267362008-08-12 15:08:49 -07003770 }
Mike Kravetz5e911372015-09-08 15:01:28 -07003771 /* Just decrements count, does not deallocate */
Mike Kravetzfeba16e2015-09-08 15:01:31 -07003772 vma_end_reservation(h, vma, address);
Andy Whitcroft57303d82008-08-12 15:08:47 -07003773
Mel Gormanf83a2752009-05-28 14:34:40 -07003774 if (!(vma->vm_flags & VM_MAYSHARE))
Andy Whitcroft57303d82008-08-12 15:08:47 -07003775 pagecache_page = hugetlbfs_pagecache_page(h,
3776 vma, address);
3777 }
3778
Naoya Horiguchi0f792cf2015-02-11 15:25:25 -08003779 ptl = huge_pte_lock(h, mm, ptep);
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09003780
David Gibson1e8f8892006-01-06 00:10:44 -08003781 /* Check for a racing update before calling hugetlb_cow */
David Gibsonb4d1d992008-10-15 22:01:11 -07003782 if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003783 goto out_ptl;
David Gibsonb4d1d992008-10-15 22:01:11 -07003784
Naoya Horiguchi0f792cf2015-02-11 15:25:25 -08003785 /*
3786 * hugetlb_cow() requires page locks of pte_page(entry) and
3787 * pagecache_page, so here we need take the former one
3788 * when page != pagecache_page or !pagecache_page.
3789 */
3790 page = pte_page(entry);
3791 if (page != pagecache_page)
3792 if (!trylock_page(page)) {
3793 need_wait_lock = 1;
3794 goto out_ptl;
3795 }
3796
3797 get_page(page);
David Gibsonb4d1d992008-10-15 22:01:11 -07003798
Hugh Dickins788c7df2009-06-23 13:49:05 +01003799 if (flags & FAULT_FLAG_WRITE) {
Gerald Schaefer106c9922013-04-29 15:07:23 -07003800 if (!huge_pte_write(entry)) {
Andy Whitcroft57303d82008-08-12 15:08:47 -07003801 ret = hugetlb_cow(mm, vma, address, ptep, entry,
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003802 pagecache_page, ptl);
Naoya Horiguchi0f792cf2015-02-11 15:25:25 -08003803 goto out_put_page;
David Gibsonb4d1d992008-10-15 22:01:11 -07003804 }
Gerald Schaefer106c9922013-04-29 15:07:23 -07003805 entry = huge_pte_mkdirty(entry);
David Gibsonb4d1d992008-10-15 22:01:11 -07003806 }
3807 entry = pte_mkyoung(entry);
Hugh Dickins788c7df2009-06-23 13:49:05 +01003808 if (huge_ptep_set_access_flags(vma, address, ptep, entry,
3809 flags & FAULT_FLAG_WRITE))
Russell King4b3073e2009-12-18 16:40:18 +00003810 update_mmu_cache(vma, address, ptep);
Naoya Horiguchi0f792cf2015-02-11 15:25:25 -08003811out_put_page:
3812 if (page != pagecache_page)
3813 unlock_page(page);
3814 put_page(page);
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003815out_ptl:
3816 spin_unlock(ptl);
Andy Whitcroft57303d82008-08-12 15:08:47 -07003817
3818 if (pagecache_page) {
3819 unlock_page(pagecache_page);
3820 put_page(pagecache_page);
3821 }
David Gibsonb4d1d992008-10-15 22:01:11 -07003822out_mutex:
Mike Kravetzc672c7f2015-09-08 15:01:35 -07003823 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
Naoya Horiguchi0f792cf2015-02-11 15:25:25 -08003824 /*
3825 * Generally it's safe to hold refcount during waiting page lock. But
3826 * here we just wait to defer the next page fault to avoid busy loop and
3827 * the page is not used after unlocked before returning from the current
3828 * page fault. So we are safe from accessing freed page, even if we wait
3829 * here without taking refcount.
3830 */
3831 if (need_wait_lock)
3832 wait_on_page_locked(page);
David Gibson1e8f8892006-01-06 00:10:44 -08003833 return ret;
Adam Litke86e52162006-01-06 00:10:43 -08003834}
3835
Michel Lespinasse28a35712013-02-22 16:35:55 -08003836long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
3837 struct page **pages, struct vm_area_struct **vmas,
3838 unsigned long *position, unsigned long *nr_pages,
3839 long i, unsigned int flags)
David Gibson63551ae2005-06-21 17:14:44 -07003840{
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08003841 unsigned long pfn_offset;
3842 unsigned long vaddr = *position;
Michel Lespinasse28a35712013-02-22 16:35:55 -08003843 unsigned long remainder = *nr_pages;
Andi Kleena5516432008-07-23 21:27:41 -07003844 struct hstate *h = hstate_vma(vma);
David Gibson63551ae2005-06-21 17:14:44 -07003845
David Gibson63551ae2005-06-21 17:14:44 -07003846 while (vaddr < vma->vm_end && remainder) {
Adam Litke4c887262005-10-29 18:16:46 -07003847 pte_t *pte;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003848 spinlock_t *ptl = NULL;
Hugh Dickins2a15efc2009-09-21 17:03:27 -07003849 int absent;
Adam Litke4c887262005-10-29 18:16:46 -07003850 struct page *page;
3851
3852 /*
David Rientjes02057962015-04-14 15:48:24 -07003853 * If we have a pending SIGKILL, don't keep faulting pages and
3854 * potentially allocating memory.
3855 */
3856 if (unlikely(fatal_signal_pending(current))) {
3857 remainder = 0;
3858 break;
3859 }
3860
3861 /*
Adam Litke4c887262005-10-29 18:16:46 -07003862 * Some archs (sparc64, sh*) have multiple pte_ts to
Hugh Dickins2a15efc2009-09-21 17:03:27 -07003863 * each hugepage. We have to make sure we get the
Adam Litke4c887262005-10-29 18:16:46 -07003864 * first, for the page indexing below to work.
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003865 *
3866 * Note that page table lock is not held when pte is null.
Adam Litke4c887262005-10-29 18:16:46 -07003867 */
Andi Kleena5516432008-07-23 21:27:41 -07003868 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h));
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003869 if (pte)
3870 ptl = huge_pte_lock(h, mm, pte);
Hugh Dickins2a15efc2009-09-21 17:03:27 -07003871 absent = !pte || huge_pte_none(huge_ptep_get(pte));
Adam Litke4c887262005-10-29 18:16:46 -07003872
Hugh Dickins2a15efc2009-09-21 17:03:27 -07003873 /*
3874 * When coredumping, it suits get_dump_page if we just return
Hugh Dickins3ae77f42009-09-21 17:03:33 -07003875 * an error where there's an empty slot with no huge pagecache
3876 * to back it. This way, we avoid allocating a hugepage, and
3877 * the sparse dumpfile avoids allocating disk blocks, but its
3878 * huge holes still show up with zeroes where they need to be.
Hugh Dickins2a15efc2009-09-21 17:03:27 -07003879 */
Hugh Dickins3ae77f42009-09-21 17:03:33 -07003880 if (absent && (flags & FOLL_DUMP) &&
3881 !hugetlbfs_pagecache_present(h, vma, vaddr)) {
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003882 if (pte)
3883 spin_unlock(ptl);
Hugh Dickins2a15efc2009-09-21 17:03:27 -07003884 remainder = 0;
3885 break;
3886 }
3887
Naoya Horiguchi9cc3a5b2013-04-17 15:58:30 -07003888 /*
3889 * We need call hugetlb_fault for both hugepages under migration
3890 * (in which case hugetlb_fault waits for the migration,) and
3891 * hwpoisoned hugepages (in which case we need to prevent the
3892 * caller from accessing to them.) In order to do this, we use
3893 * here is_swap_pte instead of is_hugetlb_entry_migration and
3894 * is_hugetlb_entry_hwpoisoned. This is because it simply covers
3895 * both cases, and because we can't follow correct pages
3896 * directly from any kind of swap entries.
3897 */
3898 if (absent || is_swap_pte(huge_ptep_get(pte)) ||
Gerald Schaefer106c9922013-04-29 15:07:23 -07003899 ((flags & FOLL_WRITE) &&
3900 !huge_pte_write(huge_ptep_get(pte)))) {
Adam Litke4c887262005-10-29 18:16:46 -07003901 int ret;
3902
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003903 if (pte)
3904 spin_unlock(ptl);
Hugh Dickins2a15efc2009-09-21 17:03:27 -07003905 ret = hugetlb_fault(mm, vma, vaddr,
3906 (flags & FOLL_WRITE) ? FAULT_FLAG_WRITE : 0);
Adam Litkea89182c2007-08-22 14:01:51 -07003907 if (!(ret & VM_FAULT_ERROR))
Adam Litke4c887262005-10-29 18:16:46 -07003908 continue;
3909
3910 remainder = 0;
Adam Litke4c887262005-10-29 18:16:46 -07003911 break;
3912 }
David Gibson63551ae2005-06-21 17:14:44 -07003913
Andi Kleena5516432008-07-23 21:27:41 -07003914 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07003915 page = pte_page(huge_ptep_get(pte));
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08003916same_page:
Chen, Kenneth Wd6692182006-03-31 02:29:57 -08003917 if (pages) {
Hugh Dickins2a15efc2009-09-21 17:03:27 -07003918 pages[i] = mem_map_offset(page, pfn_offset);
Kirill A. Shutemovddc58f22016-01-15 16:52:56 -08003919 get_page(pages[i]);
Chen, Kenneth Wd6692182006-03-31 02:29:57 -08003920 }
David Gibson63551ae2005-06-21 17:14:44 -07003921
3922 if (vmas)
3923 vmas[i] = vma;
3924
3925 vaddr += PAGE_SIZE;
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08003926 ++pfn_offset;
David Gibson63551ae2005-06-21 17:14:44 -07003927 --remainder;
3928 ++i;
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08003929 if (vaddr < vma->vm_end && remainder &&
Andi Kleena5516432008-07-23 21:27:41 -07003930 pfn_offset < pages_per_huge_page(h)) {
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08003931 /*
3932 * We use pfn_offset to avoid touching the pageframes
3933 * of this compound page.
3934 */
3935 goto same_page;
3936 }
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003937 spin_unlock(ptl);
David Gibson63551ae2005-06-21 17:14:44 -07003938 }
Michel Lespinasse28a35712013-02-22 16:35:55 -08003939 *nr_pages = remainder;
David Gibson63551ae2005-06-21 17:14:44 -07003940 *position = vaddr;
3941
Hugh Dickins2a15efc2009-09-21 17:03:27 -07003942 return i ? i : -EFAULT;
David Gibson63551ae2005-06-21 17:14:44 -07003943}
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003944
Aneesh Kumar K.V5491ae72016-07-13 15:06:43 +05303945#ifndef __HAVE_ARCH_FLUSH_HUGETLB_TLB_RANGE
3946/*
3947 * ARCHes with special requirements for evicting HUGETLB backing TLB entries can
3948 * implement this.
3949 */
3950#define flush_hugetlb_tlb_range(vma, addr, end) flush_tlb_range(vma, addr, end)
3951#endif
3952
Peter Zijlstra7da4d642012-11-19 03:14:23 +01003953unsigned long hugetlb_change_protection(struct vm_area_struct *vma,
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003954 unsigned long address, unsigned long end, pgprot_t newprot)
3955{
3956 struct mm_struct *mm = vma->vm_mm;
3957 unsigned long start = address;
3958 pte_t *ptep;
3959 pte_t pte;
Andi Kleena5516432008-07-23 21:27:41 -07003960 struct hstate *h = hstate_vma(vma);
Peter Zijlstra7da4d642012-11-19 03:14:23 +01003961 unsigned long pages = 0;
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003962
3963 BUG_ON(address >= end);
3964 flush_cache_range(vma, address, end);
3965
Rik van Riela5338092014-04-07 15:36:57 -07003966 mmu_notifier_invalidate_range_start(mm, start, end);
Davidlohr Bueso83cde9e2014-12-12 16:54:21 -08003967 i_mmap_lock_write(vma->vm_file->f_mapping);
Andi Kleena5516432008-07-23 21:27:41 -07003968 for (; address < end; address += huge_page_size(h)) {
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003969 spinlock_t *ptl;
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003970 ptep = huge_pte_offset(mm, address);
3971 if (!ptep)
3972 continue;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003973 ptl = huge_pte_lock(h, mm, ptep);
Peter Zijlstra7da4d642012-11-19 03:14:23 +01003974 if (huge_pmd_unshare(mm, &address, ptep)) {
3975 pages++;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003976 spin_unlock(ptl);
Chen, Kenneth W39dde652006-12-06 20:32:03 -08003977 continue;
Peter Zijlstra7da4d642012-11-19 03:14:23 +01003978 }
Naoya Horiguchia8bda282015-02-11 15:25:28 -08003979 pte = huge_ptep_get(ptep);
3980 if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) {
3981 spin_unlock(ptl);
3982 continue;
3983 }
3984 if (unlikely(is_hugetlb_entry_migration(pte))) {
3985 swp_entry_t entry = pte_to_swp_entry(pte);
3986
3987 if (is_write_migration_entry(entry)) {
3988 pte_t newpte;
3989
3990 make_migration_entry_read(&entry);
3991 newpte = swp_entry_to_pte(entry);
3992 set_huge_pte_at(mm, address, ptep, newpte);
3993 pages++;
3994 }
3995 spin_unlock(ptl);
3996 continue;
3997 }
3998 if (!huge_pte_none(pte)) {
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003999 pte = huge_ptep_get_and_clear(mm, address, ptep);
Gerald Schaefer106c9922013-04-29 15:07:23 -07004000 pte = pte_mkhuge(huge_pte_modify(pte, newprot));
Tony Lube7517d2013-02-04 14:28:46 -08004001 pte = arch_make_huge_pte(pte, vma, NULL, 0);
Zhang, Yanmin8f860592006-03-22 00:08:50 -08004002 set_huge_pte_at(mm, address, ptep, pte);
Peter Zijlstra7da4d642012-11-19 03:14:23 +01004003 pages++;
Zhang, Yanmin8f860592006-03-22 00:08:50 -08004004 }
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08004005 spin_unlock(ptl);
Zhang, Yanmin8f860592006-03-22 00:08:50 -08004006 }
Mel Gormand8333522012-07-31 16:46:20 -07004007 /*
Davidlohr Buesoc8c06ef2014-12-12 16:54:24 -08004008 * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare
Mel Gormand8333522012-07-31 16:46:20 -07004009 * may have cleared our pud entry and done put_page on the page table:
Davidlohr Buesoc8c06ef2014-12-12 16:54:24 -08004010 * once we release i_mmap_rwsem, another task can do the final put_page
Mel Gormand8333522012-07-31 16:46:20 -07004011 * and that page table be reused and filled with junk.
4012 */
Aneesh Kumar K.V5491ae72016-07-13 15:06:43 +05304013 flush_hugetlb_tlb_range(vma, start, end);
Joerg Roedel34ee6452014-11-13 13:46:09 +11004014 mmu_notifier_invalidate_range(mm, start, end);
Davidlohr Bueso83cde9e2014-12-12 16:54:21 -08004015 i_mmap_unlock_write(vma->vm_file->f_mapping);
Rik van Riela5338092014-04-07 15:36:57 -07004016 mmu_notifier_invalidate_range_end(mm, start, end);
Peter Zijlstra7da4d642012-11-19 03:14:23 +01004017
4018 return pages << h->order;
Zhang, Yanmin8f860592006-03-22 00:08:50 -08004019}
4020
Mel Gormana1e78772008-07-23 21:27:23 -07004021int hugetlb_reserve_pages(struct inode *inode,
4022 long from, long to,
Mel Gorman5a6fe122009-02-10 14:02:27 +00004023 struct vm_area_struct *vma,
KOSAKI Motohiroca16d142011-05-26 19:16:19 +09004024 vm_flags_t vm_flags)
Adam Litkee4e574b2007-10-16 01:26:19 -07004025{
Mel Gorman17c9d122009-02-11 16:34:16 +00004026 long ret, chg;
Andi Kleena5516432008-07-23 21:27:41 -07004027 struct hstate *h = hstate_inode(inode);
David Gibson90481622012-03-21 16:34:12 -07004028 struct hugepage_subpool *spool = subpool_inode(inode);
Joonsoo Kim9119a412014-04-03 14:47:25 -07004029 struct resv_map *resv_map;
Mike Kravetz1c5ecae2015-04-15 16:13:39 -07004030 long gbl_reserve;
Adam Litkee4e574b2007-10-16 01:26:19 -07004031
Mel Gormana1e78772008-07-23 21:27:23 -07004032 /*
Mel Gorman17c9d122009-02-11 16:34:16 +00004033 * Only apply hugepage reservation if asked. At fault time, an
4034 * attempt will be made for VM_NORESERVE to allocate a page
David Gibson90481622012-03-21 16:34:12 -07004035 * without using reserves
Mel Gorman17c9d122009-02-11 16:34:16 +00004036 */
KOSAKI Motohiroca16d142011-05-26 19:16:19 +09004037 if (vm_flags & VM_NORESERVE)
Mel Gorman17c9d122009-02-11 16:34:16 +00004038 return 0;
4039
4040 /*
Mel Gormana1e78772008-07-23 21:27:23 -07004041 * Shared mappings base their reservation on the number of pages that
4042 * are already allocated on behalf of the file. Private mappings need
4043 * to reserve the full area even if read-only as mprotect() may be
4044 * called to make the mapping read-write. Assume !vma is a shm mapping
4045 */
Joonsoo Kim9119a412014-04-03 14:47:25 -07004046 if (!vma || vma->vm_flags & VM_MAYSHARE) {
Joonsoo Kim4e35f482014-04-03 14:47:30 -07004047 resv_map = inode_resv_map(inode);
Joonsoo Kim9119a412014-04-03 14:47:25 -07004048
Joonsoo Kim1406ec92014-04-03 14:47:26 -07004049 chg = region_chg(resv_map, from, to);
Joonsoo Kim9119a412014-04-03 14:47:25 -07004050
4051 } else {
4052 resv_map = resv_map_alloc();
Mel Gorman5a6fe122009-02-10 14:02:27 +00004053 if (!resv_map)
4054 return -ENOMEM;
4055
Mel Gorman17c9d122009-02-11 16:34:16 +00004056 chg = to - from;
4057
Mel Gorman5a6fe122009-02-10 14:02:27 +00004058 set_vma_resv_map(vma, resv_map);
4059 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
4060 }
4061
Dave Hansenc50ac052012-05-29 15:06:46 -07004062 if (chg < 0) {
4063 ret = chg;
4064 goto out_err;
4065 }
Mel Gorman17c9d122009-02-11 16:34:16 +00004066
Mike Kravetz1c5ecae2015-04-15 16:13:39 -07004067 /*
4068 * There must be enough pages in the subpool for the mapping. If
4069 * the subpool has a minimum size, there may be some global
4070 * reservations already in place (gbl_reserve).
4071 */
4072 gbl_reserve = hugepage_subpool_get_pages(spool, chg);
4073 if (gbl_reserve < 0) {
Dave Hansenc50ac052012-05-29 15:06:46 -07004074 ret = -ENOSPC;
4075 goto out_err;
4076 }
Mel Gorman17c9d122009-02-11 16:34:16 +00004077
4078 /*
4079 * Check enough hugepages are available for the reservation.
David Gibson90481622012-03-21 16:34:12 -07004080 * Hand the pages back to the subpool if there are not
Mel Gorman17c9d122009-02-11 16:34:16 +00004081 */
Mike Kravetz1c5ecae2015-04-15 16:13:39 -07004082 ret = hugetlb_acct_memory(h, gbl_reserve);
Mel Gorman17c9d122009-02-11 16:34:16 +00004083 if (ret < 0) {
Mike Kravetz1c5ecae2015-04-15 16:13:39 -07004084 /* put back original number of pages, chg */
4085 (void)hugepage_subpool_put_pages(spool, chg);
Dave Hansenc50ac052012-05-29 15:06:46 -07004086 goto out_err;
Mel Gorman17c9d122009-02-11 16:34:16 +00004087 }
4088
4089 /*
4090 * Account for the reservations made. Shared mappings record regions
4091 * that have reservations as they are shared by multiple VMAs.
4092 * When the last VMA disappears, the region map says how much
4093 * the reservation was and the page cache tells how much of
4094 * the reservation was consumed. Private mappings are per-VMA and
4095 * only the consumed reservations are tracked. When the VMA
4096 * disappears, the original reservation is the VMA size and the
4097 * consumed reservations are stored in the map. Hence, nothing
4098 * else has to be done for private mappings here
4099 */
Mike Kravetz33039672015-06-24 16:57:58 -07004100 if (!vma || vma->vm_flags & VM_MAYSHARE) {
4101 long add = region_add(resv_map, from, to);
4102
4103 if (unlikely(chg > add)) {
4104 /*
4105 * pages in this range were added to the reserve
4106 * map between region_chg and region_add. This
4107 * indicates a race with alloc_huge_page. Adjust
4108 * the subpool and reserve counts modified above
4109 * based on the difference.
4110 */
4111 long rsv_adjust;
4112
4113 rsv_adjust = hugepage_subpool_put_pages(spool,
4114 chg - add);
4115 hugetlb_acct_memory(h, -rsv_adjust);
4116 }
4117 }
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -07004118 return 0;
Dave Hansenc50ac052012-05-29 15:06:46 -07004119out_err:
Mike Kravetz5e911372015-09-08 15:01:28 -07004120 if (!vma || vma->vm_flags & VM_MAYSHARE)
4121 region_abort(resv_map, from, to);
Joonsoo Kimf031dd22014-04-03 14:47:28 -07004122 if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
4123 kref_put(&resv_map->refs, resv_map_release);
Dave Hansenc50ac052012-05-29 15:06:46 -07004124 return ret;
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -07004125}
4126
Mike Kravetzb5cec282015-09-08 15:01:41 -07004127long hugetlb_unreserve_pages(struct inode *inode, long start, long end,
4128 long freed)
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -07004129{
Andi Kleena5516432008-07-23 21:27:41 -07004130 struct hstate *h = hstate_inode(inode);
Joonsoo Kim4e35f482014-04-03 14:47:30 -07004131 struct resv_map *resv_map = inode_resv_map(inode);
Joonsoo Kim9119a412014-04-03 14:47:25 -07004132 long chg = 0;
David Gibson90481622012-03-21 16:34:12 -07004133 struct hugepage_subpool *spool = subpool_inode(inode);
Mike Kravetz1c5ecae2015-04-15 16:13:39 -07004134 long gbl_reserve;
Ken Chen45c682a2007-11-14 16:59:44 -08004135
Mike Kravetzb5cec282015-09-08 15:01:41 -07004136 if (resv_map) {
4137 chg = region_del(resv_map, start, end);
4138 /*
4139 * region_del() can fail in the rare case where a region
4140 * must be split and another region descriptor can not be
4141 * allocated. If end == LONG_MAX, it will not fail.
4142 */
4143 if (chg < 0)
4144 return chg;
4145 }
4146
Ken Chen45c682a2007-11-14 16:59:44 -08004147 spin_lock(&inode->i_lock);
Eric Sandeene4c6f8b2009-07-29 15:02:16 -07004148 inode->i_blocks -= (blocks_per_huge_page(h) * freed);
Ken Chen45c682a2007-11-14 16:59:44 -08004149 spin_unlock(&inode->i_lock);
4150
Mike Kravetz1c5ecae2015-04-15 16:13:39 -07004151 /*
4152 * If the subpool has a minimum size, the number of global
4153 * reservations to be released may be adjusted.
4154 */
4155 gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed));
4156 hugetlb_acct_memory(h, -gbl_reserve);
Mike Kravetzb5cec282015-09-08 15:01:41 -07004157
4158 return 0;
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -07004159}
Naoya Horiguchi93f70f92010-05-28 09:29:20 +09004160
Steve Capper3212b532013-04-23 12:35:02 +01004161#ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
4162static unsigned long page_table_shareable(struct vm_area_struct *svma,
4163 struct vm_area_struct *vma,
4164 unsigned long addr, pgoff_t idx)
4165{
4166 unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) +
4167 svma->vm_start;
4168 unsigned long sbase = saddr & PUD_MASK;
4169 unsigned long s_end = sbase + PUD_SIZE;
4170
4171 /* Allow segments to share if only one is marked locked */
Eric B Munsonde60f5f2015-11-05 18:51:36 -08004172 unsigned long vm_flags = vma->vm_flags & VM_LOCKED_CLEAR_MASK;
4173 unsigned long svm_flags = svma->vm_flags & VM_LOCKED_CLEAR_MASK;
Steve Capper3212b532013-04-23 12:35:02 +01004174
4175 /*
4176 * match the virtual addresses, permission and the alignment of the
4177 * page table page.
4178 */
4179 if (pmd_index(addr) != pmd_index(saddr) ||
4180 vm_flags != svm_flags ||
4181 sbase < svma->vm_start || svma->vm_end < s_end)
4182 return 0;
4183
4184 return saddr;
4185}
4186
Nicholas Krause31aafb42015-09-04 15:47:58 -07004187static bool vma_shareable(struct vm_area_struct *vma, unsigned long addr)
Steve Capper3212b532013-04-23 12:35:02 +01004188{
4189 unsigned long base = addr & PUD_MASK;
4190 unsigned long end = base + PUD_SIZE;
4191
4192 /*
4193 * check on proper vm_flags and page table alignment
4194 */
4195 if (vma->vm_flags & VM_MAYSHARE &&
4196 vma->vm_start <= base && end <= vma->vm_end)
Nicholas Krause31aafb42015-09-04 15:47:58 -07004197 return true;
4198 return false;
Steve Capper3212b532013-04-23 12:35:02 +01004199}
4200
4201/*
4202 * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
4203 * and returns the corresponding pte. While this is not necessary for the
4204 * !shared pmd case because we can allocate the pmd later as well, it makes the
4205 * code much cleaner. pmd allocation is essential for the shared case because
Davidlohr Buesoc8c06ef2014-12-12 16:54:24 -08004206 * pud has to be populated inside the same i_mmap_rwsem section - otherwise
Steve Capper3212b532013-04-23 12:35:02 +01004207 * racing tasks could either miss the sharing (see huge_pte_offset) or select a
4208 * bad pmd for sharing.
4209 */
4210pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
4211{
4212 struct vm_area_struct *vma = find_vma(mm, addr);
4213 struct address_space *mapping = vma->vm_file->f_mapping;
4214 pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) +
4215 vma->vm_pgoff;
4216 struct vm_area_struct *svma;
4217 unsigned long saddr;
4218 pte_t *spte = NULL;
4219 pte_t *pte;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08004220 spinlock_t *ptl;
Steve Capper3212b532013-04-23 12:35:02 +01004221
4222 if (!vma_shareable(vma, addr))
4223 return (pte_t *)pmd_alloc(mm, pud, addr);
4224
Davidlohr Bueso83cde9e2014-12-12 16:54:21 -08004225 i_mmap_lock_write(mapping);
Steve Capper3212b532013-04-23 12:35:02 +01004226 vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) {
4227 if (svma == vma)
4228 continue;
4229
4230 saddr = page_table_shareable(svma, vma, addr, idx);
4231 if (saddr) {
4232 spte = huge_pte_offset(svma->vm_mm, saddr);
4233 if (spte) {
4234 get_page(virt_to_page(spte));
4235 break;
4236 }
4237 }
4238 }
4239
4240 if (!spte)
4241 goto out;
4242
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08004243 ptl = huge_pte_lockptr(hstate_vma(vma), mm, spte);
4244 spin_lock(ptl);
Kirill A. Shutemovdc6c9a32015-02-11 15:26:50 -08004245 if (pud_none(*pud)) {
Steve Capper3212b532013-04-23 12:35:02 +01004246 pud_populate(mm, pud,
4247 (pmd_t *)((unsigned long)spte & PAGE_MASK));
Kirill A. Shutemovc17b1f42016-06-24 14:49:51 -07004248 mm_inc_nr_pmds(mm);
Kirill A. Shutemovdc6c9a32015-02-11 15:26:50 -08004249 } else {
Steve Capper3212b532013-04-23 12:35:02 +01004250 put_page(virt_to_page(spte));
Kirill A. Shutemovdc6c9a32015-02-11 15:26:50 -08004251 }
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08004252 spin_unlock(ptl);
Steve Capper3212b532013-04-23 12:35:02 +01004253out:
4254 pte = (pte_t *)pmd_alloc(mm, pud, addr);
Davidlohr Bueso83cde9e2014-12-12 16:54:21 -08004255 i_mmap_unlock_write(mapping);
Steve Capper3212b532013-04-23 12:35:02 +01004256 return pte;
4257}
4258
4259/*
4260 * unmap huge page backed by shared pte.
4261 *
4262 * Hugetlb pte page is ref counted at the time of mapping. If pte is shared
4263 * indicated by page_count > 1, unmap is achieved by clearing pud and
4264 * decrementing the ref count. If count == 1, the pte page is not shared.
4265 *
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08004266 * called with page table lock held.
Steve Capper3212b532013-04-23 12:35:02 +01004267 *
4268 * returns: 1 successfully unmapped a shared pte page
4269 * 0 the underlying pte page is not shared, or it is the last user
4270 */
4271int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
4272{
4273 pgd_t *pgd = pgd_offset(mm, *addr);
4274 pud_t *pud = pud_offset(pgd, *addr);
4275
4276 BUG_ON(page_count(virt_to_page(ptep)) == 0);
4277 if (page_count(virt_to_page(ptep)) == 1)
4278 return 0;
4279
4280 pud_clear(pud);
4281 put_page(virt_to_page(ptep));
Kirill A. Shutemovdc6c9a32015-02-11 15:26:50 -08004282 mm_dec_nr_pmds(mm);
Steve Capper3212b532013-04-23 12:35:02 +01004283 *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE;
4284 return 1;
4285}
Steve Capper9e5fc742013-04-30 08:02:03 +01004286#define want_pmd_share() (1)
4287#else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
4288pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
4289{
4290 return NULL;
4291}
Zhang Zhene81f2d22015-06-24 16:56:13 -07004292
4293int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
4294{
4295 return 0;
4296}
Steve Capper9e5fc742013-04-30 08:02:03 +01004297#define want_pmd_share() (0)
Steve Capper3212b532013-04-23 12:35:02 +01004298#endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
4299
Steve Capper9e5fc742013-04-30 08:02:03 +01004300#ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
4301pte_t *huge_pte_alloc(struct mm_struct *mm,
4302 unsigned long addr, unsigned long sz)
4303{
4304 pgd_t *pgd;
4305 pud_t *pud;
4306 pte_t *pte = NULL;
4307
4308 pgd = pgd_offset(mm, addr);
4309 pud = pud_alloc(mm, pgd, addr);
4310 if (pud) {
4311 if (sz == PUD_SIZE) {
4312 pte = (pte_t *)pud;
4313 } else {
4314 BUG_ON(sz != PMD_SIZE);
4315 if (want_pmd_share() && pud_none(*pud))
4316 pte = huge_pmd_share(mm, addr, pud);
4317 else
4318 pte = (pte_t *)pmd_alloc(mm, pud, addr);
4319 }
4320 }
Michal Hocko4e666312016-08-02 14:02:34 -07004321 BUG_ON(pte && pte_present(*pte) && !pte_huge(*pte));
Steve Capper9e5fc742013-04-30 08:02:03 +01004322
4323 return pte;
4324}
4325
4326pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr)
4327{
4328 pgd_t *pgd;
4329 pud_t *pud;
4330 pmd_t *pmd = NULL;
4331
4332 pgd = pgd_offset(mm, addr);
4333 if (pgd_present(*pgd)) {
4334 pud = pud_offset(pgd, addr);
4335 if (pud_present(*pud)) {
4336 if (pud_huge(*pud))
4337 return (pte_t *)pud;
4338 pmd = pmd_offset(pud, addr);
4339 }
4340 }
4341 return (pte_t *) pmd;
4342}
4343
Naoya Horiguchi61f77ed2015-02-11 15:25:15 -08004344#endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
4345
4346/*
4347 * These functions are overwritable if your architecture needs its own
4348 * behavior.
4349 */
4350struct page * __weak
4351follow_huge_addr(struct mm_struct *mm, unsigned long address,
4352 int write)
4353{
4354 return ERR_PTR(-EINVAL);
4355}
4356
4357struct page * __weak
Steve Capper9e5fc742013-04-30 08:02:03 +01004358follow_huge_pmd(struct mm_struct *mm, unsigned long address,
Naoya Horiguchie66f17f2015-02-11 15:25:22 -08004359 pmd_t *pmd, int flags)
Steve Capper9e5fc742013-04-30 08:02:03 +01004360{
Naoya Horiguchie66f17f2015-02-11 15:25:22 -08004361 struct page *page = NULL;
4362 spinlock_t *ptl;
4363retry:
4364 ptl = pmd_lockptr(mm, pmd);
4365 spin_lock(ptl);
4366 /*
4367 * make sure that the address range covered by this pmd is not
4368 * unmapped from other threads.
4369 */
4370 if (!pmd_huge(*pmd))
4371 goto out;
4372 if (pmd_present(*pmd)) {
Gerald Schaefer97534122015-04-14 15:42:30 -07004373 page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT);
Naoya Horiguchie66f17f2015-02-11 15:25:22 -08004374 if (flags & FOLL_GET)
4375 get_page(page);
4376 } else {
4377 if (is_hugetlb_entry_migration(huge_ptep_get((pte_t *)pmd))) {
4378 spin_unlock(ptl);
4379 __migration_entry_wait(mm, (pte_t *)pmd, ptl);
4380 goto retry;
4381 }
4382 /*
4383 * hwpoisoned entry is treated as no_page_table in
4384 * follow_page_mask().
4385 */
4386 }
4387out:
4388 spin_unlock(ptl);
Steve Capper9e5fc742013-04-30 08:02:03 +01004389 return page;
4390}
4391
Naoya Horiguchi61f77ed2015-02-11 15:25:15 -08004392struct page * __weak
Steve Capper9e5fc742013-04-30 08:02:03 +01004393follow_huge_pud(struct mm_struct *mm, unsigned long address,
Naoya Horiguchie66f17f2015-02-11 15:25:22 -08004394 pud_t *pud, int flags)
Steve Capper9e5fc742013-04-30 08:02:03 +01004395{
Naoya Horiguchie66f17f2015-02-11 15:25:22 -08004396 if (flags & FOLL_GET)
4397 return NULL;
Steve Capper9e5fc742013-04-30 08:02:03 +01004398
Naoya Horiguchie66f17f2015-02-11 15:25:22 -08004399 return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT);
Steve Capper9e5fc742013-04-30 08:02:03 +01004400}
4401
Andi Kleend5bd9102010-09-27 09:00:12 +02004402#ifdef CONFIG_MEMORY_FAILURE
4403
Naoya Horiguchi93f70f92010-05-28 09:29:20 +09004404/*
4405 * This function is called from memory failure code.
Naoya Horiguchi93f70f92010-05-28 09:29:20 +09004406 */
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +09004407int dequeue_hwpoisoned_huge_page(struct page *hpage)
Naoya Horiguchi93f70f92010-05-28 09:29:20 +09004408{
4409 struct hstate *h = page_hstate(hpage);
4410 int nid = page_to_nid(hpage);
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +09004411 int ret = -EBUSY;
Naoya Horiguchi93f70f92010-05-28 09:29:20 +09004412
4413 spin_lock(&hugetlb_lock);
Naoya Horiguchi7e1f0492015-04-15 16:14:41 -07004414 /*
4415 * Just checking !page_huge_active is not enough, because that could be
4416 * an isolated/hwpoisoned hugepage (which have >0 refcount).
4417 */
4418 if (!page_huge_active(hpage) && !page_count(hpage)) {
Naoya Horiguchi56f2fb12012-12-12 13:52:33 -08004419 /*
4420 * Hwpoisoned hugepage isn't linked to activelist or freelist,
4421 * but dangling hpage->lru can trigger list-debug warnings
4422 * (this happens when we call unpoison_memory() on it),
4423 * so let it point to itself with list_del_init().
4424 */
4425 list_del_init(&hpage->lru);
Naoya Horiguchi8c6c2ec2010-09-08 10:19:38 +09004426 set_page_refcounted(hpage);
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +09004427 h->free_huge_pages--;
4428 h->free_huge_pages_node[nid]--;
4429 ret = 0;
4430 }
Naoya Horiguchi93f70f92010-05-28 09:29:20 +09004431 spin_unlock(&hugetlb_lock);
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +09004432 return ret;
Naoya Horiguchi93f70f92010-05-28 09:29:20 +09004433}
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +09004434#endif
Naoya Horiguchi31caf662013-09-11 14:21:59 -07004435
4436bool isolate_huge_page(struct page *page, struct list_head *list)
4437{
Naoya Horiguchibcc54222015-04-15 16:14:38 -07004438 bool ret = true;
4439
Sasha Levin309381fea2014-01-23 15:52:54 -08004440 VM_BUG_ON_PAGE(!PageHead(page), page);
Naoya Horiguchi31caf662013-09-11 14:21:59 -07004441 spin_lock(&hugetlb_lock);
Naoya Horiguchibcc54222015-04-15 16:14:38 -07004442 if (!page_huge_active(page) || !get_page_unless_zero(page)) {
4443 ret = false;
4444 goto unlock;
4445 }
4446 clear_page_huge_active(page);
Naoya Horiguchi31caf662013-09-11 14:21:59 -07004447 list_move_tail(&page->lru, list);
Naoya Horiguchibcc54222015-04-15 16:14:38 -07004448unlock:
Naoya Horiguchi31caf662013-09-11 14:21:59 -07004449 spin_unlock(&hugetlb_lock);
Naoya Horiguchibcc54222015-04-15 16:14:38 -07004450 return ret;
Naoya Horiguchi31caf662013-09-11 14:21:59 -07004451}
4452
4453void putback_active_hugepage(struct page *page)
4454{
Sasha Levin309381fea2014-01-23 15:52:54 -08004455 VM_BUG_ON_PAGE(!PageHead(page), page);
Naoya Horiguchi31caf662013-09-11 14:21:59 -07004456 spin_lock(&hugetlb_lock);
Naoya Horiguchibcc54222015-04-15 16:14:38 -07004457 set_page_huge_active(page);
Naoya Horiguchi31caf662013-09-11 14:21:59 -07004458 list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist);
4459 spin_unlock(&hugetlb_lock);
4460 put_page(page);
4461}