blob: 7bd390797092c3a4cf0765e3541765e7ef05a264 [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 */
zhong jiang72e29362016-10-07 17:02:01 -0700570void hugetlb_fix_reserve_counts(struct inode *inode)
Mike Kravetzb5cec282015-09-08 15:01:41 -0700571{
572 struct hugepage_subpool *spool = subpool_inode(inode);
573 long rsv_adjust;
574
575 rsv_adjust = hugepage_subpool_get_pages(spool, 1);
zhong jiang72e29362016-10-07 17:02:01 -0700576 if (rsv_adjust) {
Mike Kravetzb5cec282015-09-08 15:01:41 -0700577 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
Yisheng Xie461a7182016-10-07 17:01:46 -07001025#if defined(CONFIG_ARCH_HAS_GIGANTIC_PAGE) && \
Gerald Schaeferd08de8e2016-07-04 14:47:01 +02001026 ((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
Gerald Schaefer082d5b62016-10-07 17:01:10 -07001440 * nothing for in-use (including surplus) hugepages. Returns -EBUSY if the
1441 * number of free hugepages would be reduced below the number of reserved
1442 * hugepages.
Naoya Horiguchic8721bb2013-09-11 14:22:09 -07001443 */
Gerald Schaefer082d5b62016-10-07 17:01:10 -07001444static int dissolve_free_huge_page(struct page *page)
Naoya Horiguchic8721bb2013-09-11 14:22:09 -07001445{
Gerald Schaefer082d5b62016-10-07 17:01:10 -07001446 int rc = 0;
1447
Naoya Horiguchic8721bb2013-09-11 14:22:09 -07001448 spin_lock(&hugetlb_lock);
1449 if (PageHuge(page) && !page_count(page)) {
Gerald Schaefer2247bb32016-10-07 17:01:07 -07001450 struct page *head = compound_head(page);
1451 struct hstate *h = page_hstate(head);
1452 int nid = page_to_nid(head);
Gerald Schaefer082d5b62016-10-07 17:01:10 -07001453 if (h->free_huge_pages - h->resv_huge_pages == 0) {
1454 rc = -EBUSY;
1455 goto out;
1456 }
Gerald Schaefer2247bb32016-10-07 17:01:07 -07001457 list_del(&head->lru);
Naoya Horiguchic8721bb2013-09-11 14:22:09 -07001458 h->free_huge_pages--;
1459 h->free_huge_pages_node[nid]--;
zhong jiangc1470b32016-08-11 15:32:55 -07001460 h->max_huge_pages--;
Gerald Schaefer2247bb32016-10-07 17:01:07 -07001461 update_and_free_page(h, head);
Naoya Horiguchic8721bb2013-09-11 14:22:09 -07001462 }
Gerald Schaefer082d5b62016-10-07 17:01:10 -07001463out:
Naoya Horiguchic8721bb2013-09-11 14:22:09 -07001464 spin_unlock(&hugetlb_lock);
Gerald Schaefer082d5b62016-10-07 17:01:10 -07001465 return rc;
Naoya Horiguchic8721bb2013-09-11 14:22:09 -07001466}
1467
1468/*
1469 * Dissolve free hugepages in a given pfn range. Used by memory hotplug to
1470 * make specified memory blocks removable from the system.
Gerald Schaefer2247bb32016-10-07 17:01:07 -07001471 * Note that this will dissolve a free gigantic hugepage completely, if any
1472 * part of it lies within the given range.
Gerald Schaefer082d5b62016-10-07 17:01:10 -07001473 * Also note that if dissolve_free_huge_page() returns with an error, all
1474 * free hugepages that were dissolved before that error are lost.
Naoya Horiguchic8721bb2013-09-11 14:22:09 -07001475 */
Gerald Schaefer082d5b62016-10-07 17:01:10 -07001476int dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn)
Naoya Horiguchic8721bb2013-09-11 14:22:09 -07001477{
Naoya Horiguchic8721bb2013-09-11 14:22:09 -07001478 unsigned long pfn;
Gerald Schaefereb03aa02016-10-07 17:01:13 -07001479 struct page *page;
Gerald Schaefer082d5b62016-10-07 17:01:10 -07001480 int rc = 0;
Naoya Horiguchic8721bb2013-09-11 14:22:09 -07001481
Li Zhongd0177632014-08-06 16:07:56 -07001482 if (!hugepages_supported())
Gerald Schaefer082d5b62016-10-07 17:01:10 -07001483 return rc;
Li Zhongd0177632014-08-06 16:07:56 -07001484
Gerald Schaefereb03aa02016-10-07 17:01:13 -07001485 for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << minimum_order) {
1486 page = pfn_to_page(pfn);
1487 if (PageHuge(page) && !page_count(page)) {
1488 rc = dissolve_free_huge_page(page);
1489 if (rc)
1490 break;
1491 }
1492 }
Gerald Schaefer082d5b62016-10-07 17:01:10 -07001493
1494 return rc;
Naoya Horiguchic8721bb2013-09-11 14:22:09 -07001495}
1496
Dave Hansen099730d2015-11-05 18:50:17 -08001497/*
1498 * There are 3 ways this can get called:
1499 * 1. With vma+addr: we use the VMA's memory policy
1500 * 2. With !vma, but nid=NUMA_NO_NODE: We try to allocate a huge
1501 * page from any node, and let the buddy allocator itself figure
1502 * it out.
1503 * 3. With !vma, but nid!=NUMA_NO_NODE. We allocate a huge page
1504 * strictly from 'nid'
1505 */
1506static struct page *__hugetlb_alloc_buddy_huge_page(struct hstate *h,
1507 struct vm_area_struct *vma, unsigned long addr, int nid)
1508{
1509 int order = huge_page_order(h);
1510 gfp_t gfp = htlb_alloc_mask(h)|__GFP_COMP|__GFP_REPEAT|__GFP_NOWARN;
1511 unsigned int cpuset_mems_cookie;
1512
1513 /*
1514 * We need a VMA to get a memory policy. If we do not
Dave Hansene0ec90e2015-11-05 18:50:20 -08001515 * have one, we use the 'nid' argument.
1516 *
1517 * The mempolicy stuff below has some non-inlined bits
1518 * and calls ->vm_ops. That makes it hard to optimize at
1519 * compile-time, even when NUMA is off and it does
1520 * nothing. This helps the compiler optimize it out.
Dave Hansen099730d2015-11-05 18:50:17 -08001521 */
Dave Hansene0ec90e2015-11-05 18:50:20 -08001522 if (!IS_ENABLED(CONFIG_NUMA) || !vma) {
Dave Hansen099730d2015-11-05 18:50:17 -08001523 /*
1524 * If a specific node is requested, make sure to
1525 * get memory from there, but only when a node
1526 * is explicitly specified.
1527 */
1528 if (nid != NUMA_NO_NODE)
1529 gfp |= __GFP_THISNODE;
1530 /*
1531 * Make sure to call something that can handle
1532 * nid=NUMA_NO_NODE
1533 */
1534 return alloc_pages_node(nid, gfp, order);
1535 }
1536
1537 /*
1538 * OK, so we have a VMA. Fetch the mempolicy and try to
Dave Hansene0ec90e2015-11-05 18:50:20 -08001539 * allocate a huge page with it. We will only reach this
1540 * when CONFIG_NUMA=y.
Dave Hansen099730d2015-11-05 18:50:17 -08001541 */
1542 do {
1543 struct page *page;
1544 struct mempolicy *mpol;
1545 struct zonelist *zl;
1546 nodemask_t *nodemask;
1547
1548 cpuset_mems_cookie = read_mems_allowed_begin();
1549 zl = huge_zonelist(vma, addr, gfp, &mpol, &nodemask);
1550 mpol_cond_put(mpol);
1551 page = __alloc_pages_nodemask(gfp, order, zl, nodemask);
1552 if (page)
1553 return page;
1554 } while (read_mems_allowed_retry(cpuset_mems_cookie));
1555
1556 return NULL;
1557}
1558
1559/*
1560 * There are two ways to allocate a huge page:
1561 * 1. When you have a VMA and an address (like a fault)
1562 * 2. When you have no VMA (like when setting /proc/.../nr_hugepages)
1563 *
1564 * 'vma' and 'addr' are only for (1). 'nid' is always NUMA_NO_NODE in
1565 * this case which signifies that the allocation should be done with
1566 * respect for the VMA's memory policy.
1567 *
1568 * For (2), we ignore 'vma' and 'addr' and use 'nid' exclusively. This
1569 * implies that memory policies will not be taken in to account.
1570 */
1571static struct page *__alloc_buddy_huge_page(struct hstate *h,
1572 struct vm_area_struct *vma, unsigned long addr, int nid)
Adam Litke7893d1d2007-10-16 01:26:18 -07001573{
1574 struct page *page;
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001575 unsigned int r_nid;
Adam Litke7893d1d2007-10-16 01:26:18 -07001576
Luiz Capitulinobae7f4a2014-06-04 16:07:08 -07001577 if (hstate_is_gigantic(h))
Andi Kleenaa888a72008-07-23 21:27:47 -07001578 return NULL;
1579
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001580 /*
Dave Hansen099730d2015-11-05 18:50:17 -08001581 * Make sure that anyone specifying 'nid' is not also specifying a VMA.
1582 * This makes sure the caller is picking _one_ of the modes with which
1583 * we can call this function, not both.
1584 */
1585 if (vma || (addr != -1)) {
Dave Hansene0ec90e2015-11-05 18:50:20 -08001586 VM_WARN_ON_ONCE(addr == -1);
1587 VM_WARN_ON_ONCE(nid != NUMA_NO_NODE);
Dave Hansen099730d2015-11-05 18:50:17 -08001588 }
1589 /*
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001590 * Assume we will successfully allocate the surplus page to
1591 * prevent racing processes from causing the surplus to exceed
1592 * overcommit
1593 *
1594 * This however introduces a different race, where a process B
1595 * tries to grow the static hugepage pool while alloc_pages() is
1596 * called by process A. B will only examine the per-node
1597 * counters in determining if surplus huge pages can be
1598 * converted to normal huge pages in adjust_pool_surplus(). A
1599 * won't be able to increment the per-node counter, until the
1600 * lock is dropped by B, but B doesn't drop hugetlb_lock until
1601 * no more huge pages can be converted from surplus to normal
1602 * state (and doesn't try to convert again). Thus, we have a
1603 * case where a surplus huge page exists, the pool is grown, and
1604 * the surplus huge page still exists after, even though it
1605 * should just have been converted to a normal huge page. This
1606 * does not leak memory, though, as the hugepage will be freed
1607 * once it is out of use. It also does not allow the counters to
1608 * go out of whack in adjust_pool_surplus() as we don't modify
1609 * the node values until we've gotten the hugepage and only the
1610 * per-node value is checked there.
1611 */
1612 spin_lock(&hugetlb_lock);
Andi Kleena5516432008-07-23 21:27:41 -07001613 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001614 spin_unlock(&hugetlb_lock);
1615 return NULL;
1616 } else {
Andi Kleena5516432008-07-23 21:27:41 -07001617 h->nr_huge_pages++;
1618 h->surplus_huge_pages++;
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001619 }
1620 spin_unlock(&hugetlb_lock);
1621
Dave Hansen099730d2015-11-05 18:50:17 -08001622 page = __hugetlb_alloc_buddy_huge_page(h, vma, addr, nid);
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001623
1624 spin_lock(&hugetlb_lock);
Adam Litke7893d1d2007-10-16 01:26:18 -07001625 if (page) {
Aneesh Kumar K.V0edaecf2012-07-31 16:42:07 -07001626 INIT_LIST_HEAD(&page->lru);
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001627 r_nid = page_to_nid(page);
Kirill A. Shutemovf1e61552015-11-06 16:29:50 -08001628 set_compound_page_dtor(page, HUGETLB_PAGE_DTOR);
Aneesh Kumar K.V9dd540e2012-07-31 16:42:15 -07001629 set_hugetlb_cgroup(page, NULL);
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001630 /*
1631 * We incremented the global counters already
1632 */
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001633 h->nr_huge_pages_node[r_nid]++;
1634 h->surplus_huge_pages_node[r_nid]++;
Adam Litke3b116302008-04-28 02:13:06 -07001635 __count_vm_event(HTLB_BUDDY_PGALLOC);
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001636 } else {
Andi Kleena5516432008-07-23 21:27:41 -07001637 h->nr_huge_pages--;
1638 h->surplus_huge_pages--;
Adam Litke3b116302008-04-28 02:13:06 -07001639 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
Adam Litke7893d1d2007-10-16 01:26:18 -07001640 }
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001641 spin_unlock(&hugetlb_lock);
Adam Litke7893d1d2007-10-16 01:26:18 -07001642
1643 return page;
1644}
1645
Adam Litkee4e574b2007-10-16 01:26:19 -07001646/*
Dave Hansen099730d2015-11-05 18:50:17 -08001647 * Allocate a huge page from 'nid'. Note, 'nid' may be
1648 * NUMA_NO_NODE, which means that it may be allocated
1649 * anywhere.
1650 */
Dave Hansene0ec90e2015-11-05 18:50:20 -08001651static
Dave Hansen099730d2015-11-05 18:50:17 -08001652struct page *__alloc_buddy_huge_page_no_mpol(struct hstate *h, int nid)
1653{
1654 unsigned long addr = -1;
1655
1656 return __alloc_buddy_huge_page(h, NULL, addr, nid);
1657}
1658
1659/*
1660 * Use the VMA's mpolicy to allocate a huge page from the buddy.
1661 */
Dave Hansene0ec90e2015-11-05 18:50:20 -08001662static
Dave Hansen099730d2015-11-05 18:50:17 -08001663struct page *__alloc_buddy_huge_page_with_mpol(struct hstate *h,
1664 struct vm_area_struct *vma, unsigned long addr)
1665{
1666 return __alloc_buddy_huge_page(h, vma, addr, NUMA_NO_NODE);
1667}
1668
1669/*
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001670 * This allocation function is useful in the context where vma is irrelevant.
1671 * E.g. soft-offlining uses this function because it only cares physical
1672 * address of error page.
1673 */
1674struct page *alloc_huge_page_node(struct hstate *h, int nid)
1675{
Joonsoo Kim4ef91842013-09-11 14:21:51 -07001676 struct page *page = NULL;
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001677
1678 spin_lock(&hugetlb_lock);
Joonsoo Kim4ef91842013-09-11 14:21:51 -07001679 if (h->free_huge_pages - h->resv_huge_pages > 0)
1680 page = dequeue_huge_page_node(h, nid);
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001681 spin_unlock(&hugetlb_lock);
1682
Aneesh Kumar K.V94ae8ba2012-07-31 16:42:35 -07001683 if (!page)
Dave Hansen099730d2015-11-05 18:50:17 -08001684 page = __alloc_buddy_huge_page_no_mpol(h, nid);
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001685
1686 return page;
1687}
1688
1689/*
Lucas De Marchi25985ed2011-03-30 22:57:33 -03001690 * Increase the hugetlb pool such that it can accommodate a reservation
Adam Litkee4e574b2007-10-16 01:26:19 -07001691 * of size 'delta'.
1692 */
Andi Kleena5516432008-07-23 21:27:41 -07001693static int gather_surplus_pages(struct hstate *h, int delta)
Adam Litkee4e574b2007-10-16 01:26:19 -07001694{
1695 struct list_head surplus_list;
1696 struct page *page, *tmp;
1697 int ret, i;
1698 int needed, allocated;
Hillf Danton28073b02012-03-21 16:34:00 -07001699 bool alloc_ok = true;
Adam Litkee4e574b2007-10-16 01:26:19 -07001700
Andi Kleena5516432008-07-23 21:27:41 -07001701 needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
Adam Litkeac09b3a2008-03-04 14:29:38 -08001702 if (needed <= 0) {
Andi Kleena5516432008-07-23 21:27:41 -07001703 h->resv_huge_pages += delta;
Adam Litkee4e574b2007-10-16 01:26:19 -07001704 return 0;
Adam Litkeac09b3a2008-03-04 14:29:38 -08001705 }
Adam Litkee4e574b2007-10-16 01:26:19 -07001706
1707 allocated = 0;
1708 INIT_LIST_HEAD(&surplus_list);
1709
1710 ret = -ENOMEM;
1711retry:
1712 spin_unlock(&hugetlb_lock);
1713 for (i = 0; i < needed; i++) {
Dave Hansen099730d2015-11-05 18:50:17 -08001714 page = __alloc_buddy_huge_page_no_mpol(h, NUMA_NO_NODE);
Hillf Danton28073b02012-03-21 16:34:00 -07001715 if (!page) {
1716 alloc_ok = false;
1717 break;
1718 }
Adam Litkee4e574b2007-10-16 01:26:19 -07001719 list_add(&page->lru, &surplus_list);
1720 }
Hillf Danton28073b02012-03-21 16:34:00 -07001721 allocated += i;
Adam Litkee4e574b2007-10-16 01:26:19 -07001722
1723 /*
1724 * After retaking hugetlb_lock, we need to recalculate 'needed'
1725 * because either resv_huge_pages or free_huge_pages may have changed.
1726 */
1727 spin_lock(&hugetlb_lock);
Andi Kleena5516432008-07-23 21:27:41 -07001728 needed = (h->resv_huge_pages + delta) -
1729 (h->free_huge_pages + allocated);
Hillf Danton28073b02012-03-21 16:34:00 -07001730 if (needed > 0) {
1731 if (alloc_ok)
1732 goto retry;
1733 /*
1734 * We were not able to allocate enough pages to
1735 * satisfy the entire reservation so we free what
1736 * we've allocated so far.
1737 */
1738 goto free;
1739 }
Adam Litkee4e574b2007-10-16 01:26:19 -07001740 /*
1741 * The surplus_list now contains _at_least_ the number of extra pages
Lucas De Marchi25985ed2011-03-30 22:57:33 -03001742 * needed to accommodate the reservation. Add the appropriate number
Adam Litkee4e574b2007-10-16 01:26:19 -07001743 * of pages to the hugetlb pool and free the extras back to the buddy
Adam Litkeac09b3a2008-03-04 14:29:38 -08001744 * allocator. Commit the entire reservation here to prevent another
1745 * process from stealing the pages as they are added to the pool but
1746 * before they are reserved.
Adam Litkee4e574b2007-10-16 01:26:19 -07001747 */
1748 needed += allocated;
Andi Kleena5516432008-07-23 21:27:41 -07001749 h->resv_huge_pages += delta;
Adam Litkee4e574b2007-10-16 01:26:19 -07001750 ret = 0;
Naoya Horiguchia9869b82010-09-08 10:19:37 +09001751
Adam Litke19fc3f02008-04-28 02:12:20 -07001752 /* Free the needed pages to the hugetlb pool */
Adam Litkee4e574b2007-10-16 01:26:19 -07001753 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
Adam Litke19fc3f02008-04-28 02:12:20 -07001754 if ((--needed) < 0)
1755 break;
Naoya Horiguchia9869b82010-09-08 10:19:37 +09001756 /*
1757 * This page is now managed by the hugetlb allocator and has
1758 * no users -- drop the buddy allocator's reference.
1759 */
1760 put_page_testzero(page);
Sasha Levin309381fea2014-01-23 15:52:54 -08001761 VM_BUG_ON_PAGE(page_count(page), page);
Andi Kleena5516432008-07-23 21:27:41 -07001762 enqueue_huge_page(h, page);
Adam Litke19fc3f02008-04-28 02:12:20 -07001763 }
Hillf Danton28073b02012-03-21 16:34:00 -07001764free:
Hillf Dantonb0365c82011-12-28 15:57:16 -08001765 spin_unlock(&hugetlb_lock);
Adam Litke19fc3f02008-04-28 02:12:20 -07001766
1767 /* Free unnecessary surplus pages to the buddy allocator */
Joonsoo Kimc0d934b2013-09-11 14:21:02 -07001768 list_for_each_entry_safe(page, tmp, &surplus_list, lru)
1769 put_page(page);
Naoya Horiguchia9869b82010-09-08 10:19:37 +09001770 spin_lock(&hugetlb_lock);
Adam Litkee4e574b2007-10-16 01:26:19 -07001771
1772 return ret;
1773}
1774
1775/*
Mike Kravetz1e26cec2017-01-10 16:58:27 -08001776 * This routine has two main purposes:
1777 * 1) Decrement the reservation count (resv_huge_pages) by the value passed
1778 * in unused_resv_pages. This corresponds to the prior adjustments made
1779 * to the associated reservation map.
1780 * 2) Free any unused surplus pages that may have been allocated to satisfy
1781 * the reservation. As many as unused_resv_pages may be freed.
1782 *
1783 * Called with hugetlb_lock held. However, the lock could be dropped (and
1784 * reacquired) during calls to cond_resched_lock. Whenever dropping the lock,
1785 * we must make sure nobody else can claim pages we are in the process of
1786 * freeing. Do this by ensuring resv_huge_page always is greater than the
1787 * number of huge pages we plan to free when dropping the lock.
Adam Litkee4e574b2007-10-16 01:26:19 -07001788 */
Andi Kleena5516432008-07-23 21:27:41 -07001789static void return_unused_surplus_pages(struct hstate *h,
1790 unsigned long unused_resv_pages)
Adam Litkee4e574b2007-10-16 01:26:19 -07001791{
Adam Litkee4e574b2007-10-16 01:26:19 -07001792 unsigned long nr_pages;
1793
Andi Kleenaa888a72008-07-23 21:27:47 -07001794 /* Cannot return gigantic pages currently */
Luiz Capitulinobae7f4a2014-06-04 16:07:08 -07001795 if (hstate_is_gigantic(h))
Mike Kravetz1e26cec2017-01-10 16:58:27 -08001796 goto out;
Andi Kleenaa888a72008-07-23 21:27:47 -07001797
Mike Kravetz1e26cec2017-01-10 16:58:27 -08001798 /*
1799 * Part (or even all) of the reservation could have been backed
1800 * by pre-allocated pages. Only free surplus pages.
1801 */
Andi Kleena5516432008-07-23 21:27:41 -07001802 nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
Adam Litkee4e574b2007-10-16 01:26:19 -07001803
Lee Schermerhorn685f3452009-09-21 17:01:23 -07001804 /*
1805 * We want to release as many surplus pages as possible, spread
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -08001806 * evenly across all nodes with memory. Iterate across these nodes
1807 * until we can no longer free unreserved surplus pages. This occurs
1808 * when the nodes with surplus pages have no free pages.
1809 * free_pool_huge_page() will balance the the freed pages across the
1810 * on-line nodes with memory and will handle the hstate accounting.
Mike Kravetz1e26cec2017-01-10 16:58:27 -08001811 *
1812 * Note that we decrement resv_huge_pages as we free the pages. If
1813 * we drop the lock, resv_huge_pages will still be sufficiently large
1814 * to cover subsequent pages we may free.
Lee Schermerhorn685f3452009-09-21 17:01:23 -07001815 */
1816 while (nr_pages--) {
Mike Kravetz1e26cec2017-01-10 16:58:27 -08001817 h->resv_huge_pages--;
1818 unused_resv_pages--;
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08001819 if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1))
Mike Kravetz1e26cec2017-01-10 16:58:27 -08001820 goto out;
Mizuma, Masayoshi7848a4b2014-04-18 15:07:18 -07001821 cond_resched_lock(&hugetlb_lock);
Adam Litkee4e574b2007-10-16 01:26:19 -07001822 }
Mike Kravetz1e26cec2017-01-10 16:58:27 -08001823
1824out:
1825 /* Fully uncommit the reservation */
1826 h->resv_huge_pages -= unused_resv_pages;
Adam Litkee4e574b2007-10-16 01:26:19 -07001827}
1828
Mike Kravetz5e911372015-09-08 15:01:28 -07001829
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001830/*
Mike Kravetzfeba16e2015-09-08 15:01:31 -07001831 * vma_needs_reservation, vma_commit_reservation and vma_end_reservation
Mike Kravetz5e911372015-09-08 15:01:28 -07001832 * are used by the huge page allocation routines to manage reservations.
Mike Kravetzcf3ad202015-06-24 16:57:55 -07001833 *
1834 * vma_needs_reservation is called to determine if the huge page at addr
1835 * within the vma has an associated reservation. If a reservation is
1836 * needed, the value 1 is returned. The caller is then responsible for
1837 * managing the global reservation and subpool usage counts. After
1838 * the huge page has been allocated, vma_commit_reservation is called
Mike Kravetzfeba16e2015-09-08 15:01:31 -07001839 * to add the page to the reservation map. If the page allocation fails,
1840 * the reservation must be ended instead of committed. vma_end_reservation
1841 * is called in such cases.
Mike Kravetzcf3ad202015-06-24 16:57:55 -07001842 *
1843 * In the normal case, vma_commit_reservation returns the same value
1844 * as the preceding vma_needs_reservation call. The only time this
1845 * is not the case is if a reserve map was changed between calls. It
1846 * is the responsibility of the caller to notice the difference and
1847 * take appropriate action.
Mike Kravetz96b96a92016-11-10 10:46:32 -08001848 *
1849 * vma_add_reservation is used in error paths where a reservation must
1850 * be restored when a newly allocated huge page must be freed. It is
1851 * to be called after calling vma_needs_reservation to determine if a
1852 * reservation exists.
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001853 */
Mike Kravetz5e911372015-09-08 15:01:28 -07001854enum vma_resv_mode {
1855 VMA_NEEDS_RESV,
1856 VMA_COMMIT_RESV,
Mike Kravetzfeba16e2015-09-08 15:01:31 -07001857 VMA_END_RESV,
Mike Kravetz96b96a92016-11-10 10:46:32 -08001858 VMA_ADD_RESV,
Mike Kravetz5e911372015-09-08 15:01:28 -07001859};
Mike Kravetzcf3ad202015-06-24 16:57:55 -07001860static long __vma_reservation_common(struct hstate *h,
1861 struct vm_area_struct *vma, unsigned long addr,
Mike Kravetz5e911372015-09-08 15:01:28 -07001862 enum vma_resv_mode mode)
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001863{
Joonsoo Kim4e35f482014-04-03 14:47:30 -07001864 struct resv_map *resv;
1865 pgoff_t idx;
Mike Kravetzcf3ad202015-06-24 16:57:55 -07001866 long ret;
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001867
Joonsoo Kim4e35f482014-04-03 14:47:30 -07001868 resv = vma_resv_map(vma);
1869 if (!resv)
Andy Whitcroft84afd992008-07-23 21:27:32 -07001870 return 1;
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001871
Joonsoo Kim4e35f482014-04-03 14:47:30 -07001872 idx = vma_hugecache_offset(h, vma, addr);
Mike Kravetz5e911372015-09-08 15:01:28 -07001873 switch (mode) {
1874 case VMA_NEEDS_RESV:
Mike Kravetzcf3ad202015-06-24 16:57:55 -07001875 ret = region_chg(resv, idx, idx + 1);
Mike Kravetz5e911372015-09-08 15:01:28 -07001876 break;
1877 case VMA_COMMIT_RESV:
1878 ret = region_add(resv, idx, idx + 1);
1879 break;
Mike Kravetzfeba16e2015-09-08 15:01:31 -07001880 case VMA_END_RESV:
Mike Kravetz5e911372015-09-08 15:01:28 -07001881 region_abort(resv, idx, idx + 1);
1882 ret = 0;
1883 break;
Mike Kravetz96b96a92016-11-10 10:46:32 -08001884 case VMA_ADD_RESV:
1885 if (vma->vm_flags & VM_MAYSHARE)
1886 ret = region_add(resv, idx, idx + 1);
1887 else {
1888 region_abort(resv, idx, idx + 1);
1889 ret = region_del(resv, idx, idx + 1);
1890 }
1891 break;
Mike Kravetz5e911372015-09-08 15:01:28 -07001892 default:
1893 BUG();
1894 }
Andy Whitcroft84afd992008-07-23 21:27:32 -07001895
Joonsoo Kim4e35f482014-04-03 14:47:30 -07001896 if (vma->vm_flags & VM_MAYSHARE)
Mike Kravetzcf3ad202015-06-24 16:57:55 -07001897 return ret;
Mike Kravetz67961f92016-06-08 15:33:42 -07001898 else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && ret >= 0) {
1899 /*
1900 * In most cases, reserves always exist for private mappings.
1901 * However, a file associated with mapping could have been
1902 * hole punched or truncated after reserves were consumed.
1903 * As subsequent fault on such a range will not use reserves.
1904 * Subtle - The reserve map for private mappings has the
1905 * opposite meaning than that of shared mappings. If NO
1906 * entry is in the reserve map, it means a reservation exists.
1907 * If an entry exists in the reserve map, it means the
1908 * reservation has already been consumed. As a result, the
1909 * return value of this routine is the opposite of the
1910 * value returned from reserve map manipulation routines above.
1911 */
1912 if (ret)
1913 return 0;
1914 else
1915 return 1;
1916 }
Joonsoo Kim4e35f482014-04-03 14:47:30 -07001917 else
Mike Kravetzcf3ad202015-06-24 16:57:55 -07001918 return ret < 0 ? ret : 0;
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001919}
Mike Kravetzcf3ad202015-06-24 16:57:55 -07001920
1921static long vma_needs_reservation(struct hstate *h,
Andi Kleena5516432008-07-23 21:27:41 -07001922 struct vm_area_struct *vma, unsigned long addr)
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001923{
Mike Kravetz5e911372015-09-08 15:01:28 -07001924 return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV);
Mike Kravetzcf3ad202015-06-24 16:57:55 -07001925}
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001926
Mike Kravetzcf3ad202015-06-24 16:57:55 -07001927static long vma_commit_reservation(struct hstate *h,
1928 struct vm_area_struct *vma, unsigned long addr)
1929{
Mike Kravetz5e911372015-09-08 15:01:28 -07001930 return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV);
1931}
1932
Mike Kravetzfeba16e2015-09-08 15:01:31 -07001933static void vma_end_reservation(struct hstate *h,
Mike Kravetz5e911372015-09-08 15:01:28 -07001934 struct vm_area_struct *vma, unsigned long addr)
1935{
Mike Kravetzfeba16e2015-09-08 15:01:31 -07001936 (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV);
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001937}
1938
Mike Kravetz96b96a92016-11-10 10:46:32 -08001939static long vma_add_reservation(struct hstate *h,
1940 struct vm_area_struct *vma, unsigned long addr)
1941{
1942 return __vma_reservation_common(h, vma, addr, VMA_ADD_RESV);
1943}
1944
1945/*
1946 * This routine is called to restore a reservation on error paths. In the
1947 * specific error paths, a huge page was allocated (via alloc_huge_page)
1948 * and is about to be freed. If a reservation for the page existed,
1949 * alloc_huge_page would have consumed the reservation and set PagePrivate
1950 * in the newly allocated page. When the page is freed via free_huge_page,
1951 * the global reservation count will be incremented if PagePrivate is set.
1952 * However, free_huge_page can not adjust the reserve map. Adjust the
1953 * reserve map here to be consistent with global reserve count adjustments
1954 * to be made by free_huge_page.
1955 */
1956static void restore_reserve_on_error(struct hstate *h,
1957 struct vm_area_struct *vma, unsigned long address,
1958 struct page *page)
1959{
1960 if (unlikely(PagePrivate(page))) {
1961 long rc = vma_needs_reservation(h, vma, address);
1962
1963 if (unlikely(rc < 0)) {
1964 /*
1965 * Rare out of memory condition in reserve map
1966 * manipulation. Clear PagePrivate so that
1967 * global reserve count will not be incremented
1968 * by free_huge_page. This will make it appear
1969 * as though the reservation for this page was
1970 * consumed. This may prevent the task from
1971 * faulting in the page at a later time. This
1972 * is better than inconsistent global huge page
1973 * accounting of reserve counts.
1974 */
1975 ClearPagePrivate(page);
1976 } else if (rc) {
1977 rc = vma_add_reservation(h, vma, address);
1978 if (unlikely(rc < 0))
1979 /*
1980 * See above comment about rare out of
1981 * memory condition.
1982 */
1983 ClearPagePrivate(page);
1984 } else
1985 vma_end_reservation(h, vma, address);
1986 }
1987}
1988
Mike Kravetz70c35472015-09-08 15:01:54 -07001989struct page *alloc_huge_page(struct vm_area_struct *vma,
Mel Gorman04f2cbe2008-07-23 21:27:25 -07001990 unsigned long addr, int avoid_reserve)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001991{
David Gibson90481622012-03-21 16:34:12 -07001992 struct hugepage_subpool *spool = subpool_vma(vma);
Andi Kleena5516432008-07-23 21:27:41 -07001993 struct hstate *h = hstate_vma(vma);
Adam Litke348ea202007-11-14 16:59:37 -08001994 struct page *page;
Mike Kravetzd85f69b2015-09-08 15:01:47 -07001995 long map_chg, map_commit;
1996 long gbl_chg;
Aneesh Kumar K.V6d76dcf2012-07-31 16:42:18 -07001997 int ret, idx;
1998 struct hugetlb_cgroup *h_cg;
Adam Litke2fc39ce2007-11-14 16:59:39 -08001999
Aneesh Kumar K.V6d76dcf2012-07-31 16:42:18 -07002000 idx = hstate_index(h);
Mel Gormana1e78772008-07-23 21:27:23 -07002001 /*
Mike Kravetzd85f69b2015-09-08 15:01:47 -07002002 * Examine the region/reserve map to determine if the process
2003 * has a reservation for the page to be allocated. A return
2004 * code of zero indicates a reservation exists (no change).
Mel Gormana1e78772008-07-23 21:27:23 -07002005 */
Mike Kravetzd85f69b2015-09-08 15:01:47 -07002006 map_chg = gbl_chg = vma_needs_reservation(h, vma, addr);
2007 if (map_chg < 0)
Aneesh Kumar K.V76dcee72012-07-31 16:41:57 -07002008 return ERR_PTR(-ENOMEM);
Mike Kravetzd85f69b2015-09-08 15:01:47 -07002009
2010 /*
2011 * Processes that did not create the mapping will have no
2012 * reserves as indicated by the region/reserve map. Check
2013 * that the allocation will not exceed the subpool limit.
2014 * Allocations for MAP_NORESERVE mappings also need to be
2015 * checked against any subpool limit.
2016 */
2017 if (map_chg || avoid_reserve) {
2018 gbl_chg = hugepage_subpool_get_pages(spool, 1);
2019 if (gbl_chg < 0) {
Mike Kravetzfeba16e2015-09-08 15:01:31 -07002020 vma_end_reservation(h, vma, addr);
Aneesh Kumar K.V76dcee72012-07-31 16:41:57 -07002021 return ERR_PTR(-ENOSPC);
Mike Kravetz5e911372015-09-08 15:01:28 -07002022 }
Mel Gormana1e78772008-07-23 21:27:23 -07002023
Mike Kravetzd85f69b2015-09-08 15:01:47 -07002024 /*
2025 * Even though there was no reservation in the region/reserve
2026 * map, there could be reservations associated with the
2027 * subpool that can be used. This would be indicated if the
2028 * return value of hugepage_subpool_get_pages() is zero.
2029 * However, if avoid_reserve is specified we still avoid even
2030 * the subpool reservations.
2031 */
2032 if (avoid_reserve)
2033 gbl_chg = 1;
2034 }
2035
Aneesh Kumar K.V6d76dcf2012-07-31 16:42:18 -07002036 ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg);
Jianyu Zhan8f34af62014-06-04 16:10:36 -07002037 if (ret)
2038 goto out_subpool_put;
2039
Mel Gormana1e78772008-07-23 21:27:23 -07002040 spin_lock(&hugetlb_lock);
Mike Kravetzd85f69b2015-09-08 15:01:47 -07002041 /*
2042 * glb_chg is passed to indicate whether or not a page must be taken
2043 * from the global free pool (global change). gbl_chg == 0 indicates
2044 * a reservation exists for the allocation.
2045 */
2046 page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg);
Joonsoo Kim81a6fca2013-09-11 14:20:58 -07002047 if (!page) {
Aneesh Kumar K.V94ae8ba2012-07-31 16:42:35 -07002048 spin_unlock(&hugetlb_lock);
Dave Hansen099730d2015-11-05 18:50:17 -08002049 page = __alloc_buddy_huge_page_with_mpol(h, vma, addr);
Jianyu Zhan8f34af62014-06-04 16:10:36 -07002050 if (!page)
2051 goto out_uncharge_cgroup;
Naoya Horiguchia88c7692015-12-11 13:40:24 -08002052 if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) {
2053 SetPagePrivate(page);
2054 h->resv_huge_pages--;
2055 }
Aneesh Kumar K.V79dbb232012-07-31 16:42:32 -07002056 spin_lock(&hugetlb_lock);
2057 list_move(&page->lru, &h->hugepage_activelist);
Joonsoo Kim81a6fca2013-09-11 14:20:58 -07002058 /* Fall through */
Mel Gormana1e78772008-07-23 21:27:23 -07002059 }
Joonsoo Kim81a6fca2013-09-11 14:20:58 -07002060 hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page);
2061 spin_unlock(&hugetlb_lock);
Mel Gormana1e78772008-07-23 21:27:23 -07002062
David Gibson90481622012-03-21 16:34:12 -07002063 set_page_private(page, (unsigned long)spool);
Mel Gormana1e78772008-07-23 21:27:23 -07002064
Mike Kravetzd85f69b2015-09-08 15:01:47 -07002065 map_commit = vma_commit_reservation(h, vma, addr);
2066 if (unlikely(map_chg > map_commit)) {
Mike Kravetz33039672015-06-24 16:57:58 -07002067 /*
2068 * The page was added to the reservation map between
2069 * vma_needs_reservation and vma_commit_reservation.
2070 * This indicates a race with hugetlb_reserve_pages.
2071 * Adjust for the subpool count incremented above AND
2072 * in hugetlb_reserve_pages for the same page. Also,
2073 * the reservation count added in hugetlb_reserve_pages
2074 * no longer applies.
2075 */
2076 long rsv_adjust;
2077
2078 rsv_adjust = hugepage_subpool_put_pages(spool, 1);
2079 hugetlb_acct_memory(h, -rsv_adjust);
2080 }
Adam Litke90d8b7e2007-11-14 16:59:42 -08002081 return page;
Jianyu Zhan8f34af62014-06-04 16:10:36 -07002082
2083out_uncharge_cgroup:
2084 hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg);
2085out_subpool_put:
Mike Kravetzd85f69b2015-09-08 15:01:47 -07002086 if (map_chg || avoid_reserve)
Jianyu Zhan8f34af62014-06-04 16:10:36 -07002087 hugepage_subpool_put_pages(spool, 1);
Mike Kravetzfeba16e2015-09-08 15:01:31 -07002088 vma_end_reservation(h, vma, addr);
Jianyu Zhan8f34af62014-06-04 16:10:36 -07002089 return ERR_PTR(-ENOSPC);
David Gibsonb45b5bd2006-03-22 00:08:55 -08002090}
2091
Naoya Horiguchi74060e42013-09-11 14:22:06 -07002092/*
2093 * alloc_huge_page()'s wrapper which simply returns the page if allocation
2094 * succeeds, otherwise NULL. This function is called from new_vma_page(),
2095 * where no ERR_VALUE is expected to be returned.
2096 */
2097struct page *alloc_huge_page_noerr(struct vm_area_struct *vma,
2098 unsigned long addr, int avoid_reserve)
2099{
2100 struct page *page = alloc_huge_page(vma, addr, avoid_reserve);
2101 if (IS_ERR(page))
2102 page = NULL;
2103 return page;
2104}
2105
Cyrill Gorcunov91f47662009-01-06 14:40:33 -08002106int __weak alloc_bootmem_huge_page(struct hstate *h)
Andi Kleenaa888a72008-07-23 21:27:47 -07002107{
2108 struct huge_bootmem_page *m;
Joonsoo Kimb2261022013-09-11 14:21:00 -07002109 int nr_nodes, node;
Andi Kleenaa888a72008-07-23 21:27:47 -07002110
Joonsoo Kimb2261022013-09-11 14:21:00 -07002111 for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) {
Andi Kleenaa888a72008-07-23 21:27:47 -07002112 void *addr;
2113
Grygorii Strashko8b89a112014-01-21 15:50:36 -08002114 addr = memblock_virt_alloc_try_nid_nopanic(
2115 huge_page_size(h), huge_page_size(h),
2116 0, BOOTMEM_ALLOC_ACCESSIBLE, node);
Andi Kleenaa888a72008-07-23 21:27:47 -07002117 if (addr) {
2118 /*
2119 * Use the beginning of the huge page to store the
2120 * huge_bootmem_page struct (until gather_bootmem
2121 * puts them into the mem_map).
2122 */
2123 m = addr;
Cyrill Gorcunov91f47662009-01-06 14:40:33 -08002124 goto found;
Andi Kleenaa888a72008-07-23 21:27:47 -07002125 }
Andi Kleenaa888a72008-07-23 21:27:47 -07002126 }
2127 return 0;
2128
2129found:
Luiz Capitulinodf994ea2014-12-12 16:55:21 -08002130 BUG_ON(!IS_ALIGNED(virt_to_phys(m), huge_page_size(h)));
Andi Kleenaa888a72008-07-23 21:27:47 -07002131 /* Put them into a private list first because mem_map is not up yet */
2132 list_add(&m->list, &huge_boot_pages);
2133 m->hstate = h;
2134 return 1;
2135}
2136
Kirill A. Shutemovd00181b2015-11-06 16:29:57 -08002137static void __init prep_compound_huge_page(struct page *page,
2138 unsigned int order)
Andy Whitcroft18229df2008-11-06 12:53:27 -08002139{
2140 if (unlikely(order > (MAX_ORDER - 1)))
2141 prep_compound_gigantic_page(page, order);
2142 else
2143 prep_compound_page(page, order);
2144}
2145
Andi Kleenaa888a72008-07-23 21:27:47 -07002146/* Put bootmem huge pages into the standard lists after mem_map is up */
2147static void __init gather_bootmem_prealloc(void)
2148{
2149 struct huge_bootmem_page *m;
2150
2151 list_for_each_entry(m, &huge_boot_pages, list) {
Andi Kleenaa888a72008-07-23 21:27:47 -07002152 struct hstate *h = m->hstate;
Becky Bruceee8f2482011-07-25 17:11:50 -07002153 struct page *page;
2154
2155#ifdef CONFIG_HIGHMEM
2156 page = pfn_to_page(m->phys >> PAGE_SHIFT);
Grygorii Strashko8b89a112014-01-21 15:50:36 -08002157 memblock_free_late(__pa(m),
2158 sizeof(struct huge_bootmem_page));
Becky Bruceee8f2482011-07-25 17:11:50 -07002159#else
2160 page = virt_to_page(m);
2161#endif
Andi Kleenaa888a72008-07-23 21:27:47 -07002162 WARN_ON(page_count(page) != 1);
Andy Whitcroft18229df2008-11-06 12:53:27 -08002163 prep_compound_huge_page(page, h->order);
Andrea Arcangelief5a22b2013-10-16 13:46:56 -07002164 WARN_ON(PageReserved(page));
Andi Kleenaa888a72008-07-23 21:27:47 -07002165 prep_new_huge_page(h, page, page_to_nid(page));
Rafael Aquinib0320c72011-06-15 15:08:39 -07002166 /*
2167 * If we had gigantic hugepages allocated at boot time, we need
2168 * to restore the 'stolen' pages to totalram_pages in order to
2169 * fix confusing memory reports from free(1) and another
2170 * side-effects, like CommitLimit going negative.
2171 */
Luiz Capitulinobae7f4a2014-06-04 16:07:08 -07002172 if (hstate_is_gigantic(h))
Jiang Liu3dcc0572013-07-03 15:03:21 -07002173 adjust_managed_page_count(page, 1 << h->order);
Cannon Matthews433c1832018-07-03 17:02:43 -07002174 cond_resched();
Andi Kleenaa888a72008-07-23 21:27:47 -07002175 }
2176}
2177
Andi Kleen8faa8b02008-07-23 21:27:48 -07002178static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002179{
2180 unsigned long i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002181
Andi Kleene5ff2152008-07-23 21:27:42 -07002182 for (i = 0; i < h->max_huge_pages; ++i) {
Luiz Capitulinobae7f4a2014-06-04 16:07:08 -07002183 if (hstate_is_gigantic(h)) {
Andi Kleenaa888a72008-07-23 21:27:47 -07002184 if (!alloc_bootmem_huge_page(h))
2185 break;
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -08002186 } else if (!alloc_fresh_huge_page(h,
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08002187 &node_states[N_MEMORY]))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002188 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002189 }
Andi Kleen8faa8b02008-07-23 21:27:48 -07002190 h->max_huge_pages = i;
Andi Kleene5ff2152008-07-23 21:27:42 -07002191}
2192
2193static void __init hugetlb_init_hstates(void)
2194{
2195 struct hstate *h;
2196
2197 for_each_hstate(h) {
Naoya Horiguchi641844f2015-06-24 16:56:59 -07002198 if (minimum_order > huge_page_order(h))
2199 minimum_order = huge_page_order(h);
2200
Andi Kleen8faa8b02008-07-23 21:27:48 -07002201 /* oversize hugepages were init'ed in early boot */
Luiz Capitulinobae7f4a2014-06-04 16:07:08 -07002202 if (!hstate_is_gigantic(h))
Andi Kleen8faa8b02008-07-23 21:27:48 -07002203 hugetlb_hstate_alloc_pages(h);
Andi Kleene5ff2152008-07-23 21:27:42 -07002204 }
Naoya Horiguchi641844f2015-06-24 16:56:59 -07002205 VM_BUG_ON(minimum_order == UINT_MAX);
Andi Kleene5ff2152008-07-23 21:27:42 -07002206}
2207
Andi Kleen4abd32d2008-07-23 21:27:49 -07002208static char * __init memfmt(char *buf, unsigned long n)
2209{
2210 if (n >= (1UL << 30))
2211 sprintf(buf, "%lu GB", n >> 30);
2212 else if (n >= (1UL << 20))
2213 sprintf(buf, "%lu MB", n >> 20);
2214 else
2215 sprintf(buf, "%lu KB", n >> 10);
2216 return buf;
2217}
2218
Andi Kleene5ff2152008-07-23 21:27:42 -07002219static void __init report_hugepages(void)
2220{
2221 struct hstate *h;
2222
2223 for_each_hstate(h) {
Andi Kleen4abd32d2008-07-23 21:27:49 -07002224 char buf[32];
Andrew Mortonffb22af2013-02-22 16:32:08 -08002225 pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n",
Andi Kleen4abd32d2008-07-23 21:27:49 -07002226 memfmt(buf, huge_page_size(h)),
2227 h->free_huge_pages);
Andi Kleene5ff2152008-07-23 21:27:42 -07002228 }
2229}
2230
Linus Torvalds1da177e2005-04-16 15:20:36 -07002231#ifdef CONFIG_HIGHMEM
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08002232static void try_to_free_low(struct hstate *h, unsigned long count,
2233 nodemask_t *nodes_allowed)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002234{
Christoph Lameter4415cc82006-09-25 23:31:55 -07002235 int i;
2236
Luiz Capitulinobae7f4a2014-06-04 16:07:08 -07002237 if (hstate_is_gigantic(h))
Andi Kleenaa888a72008-07-23 21:27:47 -07002238 return;
2239
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08002240 for_each_node_mask(i, *nodes_allowed) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002241 struct page *page, *next;
Andi Kleena5516432008-07-23 21:27:41 -07002242 struct list_head *freel = &h->hugepage_freelists[i];
2243 list_for_each_entry_safe(page, next, freel, lru) {
2244 if (count >= h->nr_huge_pages)
Adam Litke6b0c8802007-10-16 01:26:23 -07002245 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002246 if (PageHighMem(page))
2247 continue;
2248 list_del(&page->lru);
Andi Kleene5ff2152008-07-23 21:27:42 -07002249 update_and_free_page(h, page);
Andi Kleena5516432008-07-23 21:27:41 -07002250 h->free_huge_pages--;
2251 h->free_huge_pages_node[page_to_nid(page)]--;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002252 }
2253 }
2254}
2255#else
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08002256static inline void try_to_free_low(struct hstate *h, unsigned long count,
2257 nodemask_t *nodes_allowed)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002258{
2259}
2260#endif
2261
Wu Fengguang20a03072009-06-16 15:32:22 -07002262/*
2263 * Increment or decrement surplus_huge_pages. Keep node-specific counters
2264 * balanced by operating on them in a round-robin fashion.
2265 * Returns 1 if an adjustment was made.
2266 */
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08002267static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
2268 int delta)
Wu Fengguang20a03072009-06-16 15:32:22 -07002269{
Joonsoo Kimb2261022013-09-11 14:21:00 -07002270 int nr_nodes, node;
Wu Fengguang20a03072009-06-16 15:32:22 -07002271
2272 VM_BUG_ON(delta != -1 && delta != 1);
Wu Fengguang20a03072009-06-16 15:32:22 -07002273
Joonsoo Kimb2261022013-09-11 14:21:00 -07002274 if (delta < 0) {
2275 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
2276 if (h->surplus_huge_pages_node[node])
2277 goto found;
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07002278 }
Joonsoo Kimb2261022013-09-11 14:21:00 -07002279 } else {
2280 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
2281 if (h->surplus_huge_pages_node[node] <
2282 h->nr_huge_pages_node[node])
2283 goto found;
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07002284 }
Joonsoo Kimb2261022013-09-11 14:21:00 -07002285 }
2286 return 0;
Wu Fengguang20a03072009-06-16 15:32:22 -07002287
Joonsoo Kimb2261022013-09-11 14:21:00 -07002288found:
2289 h->surplus_huge_pages += delta;
2290 h->surplus_huge_pages_node[node] += delta;
2291 return 1;
Wu Fengguang20a03072009-06-16 15:32:22 -07002292}
2293
Andi Kleena5516432008-07-23 21:27:41 -07002294#define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08002295static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count,
2296 nodemask_t *nodes_allowed)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002297{
Adam Litke7893d1d2007-10-16 01:26:18 -07002298 unsigned long min_count, ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002299
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07002300 if (hstate_is_gigantic(h) && !gigantic_page_supported())
Andi Kleenaa888a72008-07-23 21:27:47 -07002301 return h->max_huge_pages;
2302
Adam Litke7893d1d2007-10-16 01:26:18 -07002303 /*
2304 * Increase the pool size
2305 * First take pages out of surplus state. Then make up the
2306 * remaining difference by allocating fresh huge pages.
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08002307 *
Naoya Horiguchid15c7c02015-11-10 14:45:11 -08002308 * We might race with __alloc_buddy_huge_page() here and be unable
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08002309 * to convert a surplus huge page to a normal huge page. That is
2310 * not critical, though, it just means the overall size of the
2311 * pool might be one hugepage larger than it needs to be, but
2312 * within all the constraints specified by the sysctls.
Adam Litke7893d1d2007-10-16 01:26:18 -07002313 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002314 spin_lock(&hugetlb_lock);
Andi Kleena5516432008-07-23 21:27:41 -07002315 while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08002316 if (!adjust_pool_surplus(h, nodes_allowed, -1))
Adam Litke7893d1d2007-10-16 01:26:18 -07002317 break;
2318 }
2319
Andi Kleena5516432008-07-23 21:27:41 -07002320 while (count > persistent_huge_pages(h)) {
Adam Litke7893d1d2007-10-16 01:26:18 -07002321 /*
2322 * If this allocation races such that we no longer need the
2323 * page, free_huge_page will handle it by freeing the page
2324 * and reducing the surplus.
2325 */
2326 spin_unlock(&hugetlb_lock);
Jia He649920c2016-08-02 14:02:31 -07002327
2328 /* yield cpu to avoid soft lockup */
2329 cond_resched();
2330
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07002331 if (hstate_is_gigantic(h))
2332 ret = alloc_fresh_gigantic_page(h, nodes_allowed);
2333 else
2334 ret = alloc_fresh_huge_page(h, nodes_allowed);
Adam Litke7893d1d2007-10-16 01:26:18 -07002335 spin_lock(&hugetlb_lock);
2336 if (!ret)
2337 goto out;
2338
Mel Gorman536240f22009-12-14 17:59:56 -08002339 /* Bail for signals. Probably ctrl-c from user */
2340 if (signal_pending(current))
2341 goto out;
Adam Litke7893d1d2007-10-16 01:26:18 -07002342 }
Adam Litke7893d1d2007-10-16 01:26:18 -07002343
2344 /*
2345 * Decrease the pool size
2346 * First return free pages to the buddy allocator (being careful
2347 * to keep enough around to satisfy reservations). Then place
2348 * pages into surplus state as needed so the pool will shrink
2349 * to the desired size as pages become free.
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08002350 *
2351 * By placing pages into the surplus state independent of the
2352 * overcommit value, we are allowing the surplus pool size to
2353 * exceed overcommit. There are few sane options here. Since
Naoya Horiguchid15c7c02015-11-10 14:45:11 -08002354 * __alloc_buddy_huge_page() is checking the global counter,
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08002355 * though, we'll note that we're not allowed to exceed surplus
2356 * and won't grow the pool anywhere else. Not until one of the
2357 * sysctls are changed, or the surplus pages go out of use.
Adam Litke7893d1d2007-10-16 01:26:18 -07002358 */
Andi Kleena5516432008-07-23 21:27:41 -07002359 min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
Adam Litke6b0c8802007-10-16 01:26:23 -07002360 min_count = max(count, min_count);
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08002361 try_to_free_low(h, min_count, nodes_allowed);
Andi Kleena5516432008-07-23 21:27:41 -07002362 while (min_count < persistent_huge_pages(h)) {
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08002363 if (!free_pool_huge_page(h, nodes_allowed, 0))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002364 break;
Mizuma, Masayoshi55f67142014-04-07 15:37:54 -07002365 cond_resched_lock(&hugetlb_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002366 }
Andi Kleena5516432008-07-23 21:27:41 -07002367 while (count < persistent_huge_pages(h)) {
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08002368 if (!adjust_pool_surplus(h, nodes_allowed, 1))
Adam Litke7893d1d2007-10-16 01:26:18 -07002369 break;
2370 }
2371out:
Andi Kleena5516432008-07-23 21:27:41 -07002372 ret = persistent_huge_pages(h);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002373 spin_unlock(&hugetlb_lock);
Adam Litke7893d1d2007-10-16 01:26:18 -07002374 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002375}
2376
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002377#define HSTATE_ATTR_RO(_name) \
2378 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
2379
2380#define HSTATE_ATTR(_name) \
2381 static struct kobj_attribute _name##_attr = \
2382 __ATTR(_name, 0644, _name##_show, _name##_store)
2383
2384static struct kobject *hugepages_kobj;
2385static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
2386
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002387static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
2388
2389static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002390{
2391 int i;
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002392
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002393 for (i = 0; i < HUGE_MAX_HSTATE; i++)
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002394 if (hstate_kobjs[i] == kobj) {
2395 if (nidp)
2396 *nidp = NUMA_NO_NODE;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002397 return &hstates[i];
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002398 }
2399
2400 return kobj_to_node_hstate(kobj, nidp);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002401}
2402
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002403static ssize_t nr_hugepages_show_common(struct kobject *kobj,
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002404 struct kobj_attribute *attr, char *buf)
2405{
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002406 struct hstate *h;
2407 unsigned long nr_huge_pages;
2408 int nid;
2409
2410 h = kobj_to_hstate(kobj, &nid);
2411 if (nid == NUMA_NO_NODE)
2412 nr_huge_pages = h->nr_huge_pages;
2413 else
2414 nr_huge_pages = h->nr_huge_pages_node[nid];
2415
2416 return sprintf(buf, "%lu\n", nr_huge_pages);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002417}
Eric B Munsonadbe8722011-01-13 15:47:27 -08002418
David Rientjes238d3c12014-08-06 16:06:51 -07002419static ssize_t __nr_hugepages_store_common(bool obey_mempolicy,
2420 struct hstate *h, int nid,
2421 unsigned long count, size_t len)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002422{
2423 int err;
David Rientjesbad44b52009-12-14 17:58:38 -08002424 NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002425
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07002426 if (hstate_is_gigantic(h) && !gigantic_page_supported()) {
Eric B Munsonadbe8722011-01-13 15:47:27 -08002427 err = -EINVAL;
2428 goto out;
2429 }
2430
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002431 if (nid == NUMA_NO_NODE) {
2432 /*
2433 * global hstate attribute
2434 */
2435 if (!(obey_mempolicy &&
2436 init_nodemask_of_mempolicy(nodes_allowed))) {
2437 NODEMASK_FREE(nodes_allowed);
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08002438 nodes_allowed = &node_states[N_MEMORY];
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002439 }
2440 } else if (nodes_allowed) {
2441 /*
2442 * per node hstate attribute: adjust count to global,
2443 * but restrict alloc/free to the specified node.
2444 */
2445 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
2446 init_nodemask_of_node(nodes_allowed, nid);
2447 } else
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08002448 nodes_allowed = &node_states[N_MEMORY];
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002449
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002450 h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002451
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08002452 if (nodes_allowed != &node_states[N_MEMORY])
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002453 NODEMASK_FREE(nodes_allowed);
2454
2455 return len;
Eric B Munsonadbe8722011-01-13 15:47:27 -08002456out:
2457 NODEMASK_FREE(nodes_allowed);
2458 return err;
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002459}
2460
David Rientjes238d3c12014-08-06 16:06:51 -07002461static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
2462 struct kobject *kobj, const char *buf,
2463 size_t len)
2464{
2465 struct hstate *h;
2466 unsigned long count;
2467 int nid;
2468 int err;
2469
2470 err = kstrtoul(buf, 10, &count);
2471 if (err)
2472 return err;
2473
2474 h = kobj_to_hstate(kobj, &nid);
2475 return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len);
2476}
2477
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002478static ssize_t nr_hugepages_show(struct kobject *kobj,
2479 struct kobj_attribute *attr, char *buf)
2480{
2481 return nr_hugepages_show_common(kobj, attr, buf);
2482}
2483
2484static ssize_t nr_hugepages_store(struct kobject *kobj,
2485 struct kobj_attribute *attr, const char *buf, size_t len)
2486{
David Rientjes238d3c12014-08-06 16:06:51 -07002487 return nr_hugepages_store_common(false, kobj, buf, len);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002488}
2489HSTATE_ATTR(nr_hugepages);
2490
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002491#ifdef CONFIG_NUMA
2492
2493/*
2494 * hstate attribute for optionally mempolicy-based constraint on persistent
2495 * huge page alloc/free.
2496 */
2497static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
2498 struct kobj_attribute *attr, char *buf)
2499{
2500 return nr_hugepages_show_common(kobj, attr, buf);
2501}
2502
2503static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
2504 struct kobj_attribute *attr, const char *buf, size_t len)
2505{
David Rientjes238d3c12014-08-06 16:06:51 -07002506 return nr_hugepages_store_common(true, kobj, buf, len);
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002507}
2508HSTATE_ATTR(nr_hugepages_mempolicy);
2509#endif
2510
2511
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002512static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
2513 struct kobj_attribute *attr, char *buf)
2514{
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002515 struct hstate *h = kobj_to_hstate(kobj, NULL);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002516 return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
2517}
Eric B Munsonadbe8722011-01-13 15:47:27 -08002518
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002519static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
2520 struct kobj_attribute *attr, const char *buf, size_t count)
2521{
2522 int err;
2523 unsigned long input;
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002524 struct hstate *h = kobj_to_hstate(kobj, NULL);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002525
Luiz Capitulinobae7f4a2014-06-04 16:07:08 -07002526 if (hstate_is_gigantic(h))
Eric B Munsonadbe8722011-01-13 15:47:27 -08002527 return -EINVAL;
2528
Jingoo Han3dbb95f2013-09-11 14:20:25 -07002529 err = kstrtoul(buf, 10, &input);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002530 if (err)
Eric B Munson73ae31e2011-01-13 15:47:28 -08002531 return err;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002532
2533 spin_lock(&hugetlb_lock);
2534 h->nr_overcommit_huge_pages = input;
2535 spin_unlock(&hugetlb_lock);
2536
2537 return count;
2538}
2539HSTATE_ATTR(nr_overcommit_hugepages);
2540
2541static ssize_t free_hugepages_show(struct kobject *kobj,
2542 struct kobj_attribute *attr, char *buf)
2543{
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002544 struct hstate *h;
2545 unsigned long free_huge_pages;
2546 int nid;
2547
2548 h = kobj_to_hstate(kobj, &nid);
2549 if (nid == NUMA_NO_NODE)
2550 free_huge_pages = h->free_huge_pages;
2551 else
2552 free_huge_pages = h->free_huge_pages_node[nid];
2553
2554 return sprintf(buf, "%lu\n", free_huge_pages);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002555}
2556HSTATE_ATTR_RO(free_hugepages);
2557
2558static ssize_t resv_hugepages_show(struct kobject *kobj,
2559 struct kobj_attribute *attr, char *buf)
2560{
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002561 struct hstate *h = kobj_to_hstate(kobj, NULL);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002562 return sprintf(buf, "%lu\n", h->resv_huge_pages);
2563}
2564HSTATE_ATTR_RO(resv_hugepages);
2565
2566static ssize_t surplus_hugepages_show(struct kobject *kobj,
2567 struct kobj_attribute *attr, char *buf)
2568{
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002569 struct hstate *h;
2570 unsigned long surplus_huge_pages;
2571 int nid;
2572
2573 h = kobj_to_hstate(kobj, &nid);
2574 if (nid == NUMA_NO_NODE)
2575 surplus_huge_pages = h->surplus_huge_pages;
2576 else
2577 surplus_huge_pages = h->surplus_huge_pages_node[nid];
2578
2579 return sprintf(buf, "%lu\n", surplus_huge_pages);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002580}
2581HSTATE_ATTR_RO(surplus_hugepages);
2582
2583static struct attribute *hstate_attrs[] = {
2584 &nr_hugepages_attr.attr,
2585 &nr_overcommit_hugepages_attr.attr,
2586 &free_hugepages_attr.attr,
2587 &resv_hugepages_attr.attr,
2588 &surplus_hugepages_attr.attr,
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002589#ifdef CONFIG_NUMA
2590 &nr_hugepages_mempolicy_attr.attr,
2591#endif
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002592 NULL,
2593};
2594
2595static struct attribute_group hstate_attr_group = {
2596 .attrs = hstate_attrs,
2597};
2598
Jeff Mahoney094e9532010-02-02 13:44:14 -08002599static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
2600 struct kobject **hstate_kobjs,
2601 struct attribute_group *hstate_attr_group)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002602{
2603 int retval;
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07002604 int hi = hstate_index(h);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002605
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002606 hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
2607 if (!hstate_kobjs[hi])
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002608 return -ENOMEM;
2609
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002610 retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002611 if (retval)
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002612 kobject_put(hstate_kobjs[hi]);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002613
2614 return retval;
2615}
2616
2617static void __init hugetlb_sysfs_init(void)
2618{
2619 struct hstate *h;
2620 int err;
2621
2622 hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
2623 if (!hugepages_kobj)
2624 return;
2625
2626 for_each_hstate(h) {
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002627 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
2628 hstate_kobjs, &hstate_attr_group);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002629 if (err)
Andrew Mortonffb22af2013-02-22 16:32:08 -08002630 pr_err("Hugetlb: Unable to add hstate %s", h->name);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002631 }
2632}
2633
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002634#ifdef CONFIG_NUMA
2635
2636/*
2637 * node_hstate/s - associate per node hstate attributes, via their kobjects,
Kay Sievers10fbcf42011-12-21 14:48:43 -08002638 * with node devices in node_devices[] using a parallel array. The array
2639 * index of a node device or _hstate == node id.
2640 * This is here to avoid any static dependency of the node device driver, in
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002641 * the base kernel, on the hugetlb module.
2642 */
2643struct node_hstate {
2644 struct kobject *hugepages_kobj;
2645 struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
2646};
Alexander Kuleshovb4e289a2015-11-05 18:50:14 -08002647static struct node_hstate node_hstates[MAX_NUMNODES];
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002648
2649/*
Kay Sievers10fbcf42011-12-21 14:48:43 -08002650 * A subset of global hstate attributes for node devices
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002651 */
2652static struct attribute *per_node_hstate_attrs[] = {
2653 &nr_hugepages_attr.attr,
2654 &free_hugepages_attr.attr,
2655 &surplus_hugepages_attr.attr,
2656 NULL,
2657};
2658
2659static struct attribute_group per_node_hstate_attr_group = {
2660 .attrs = per_node_hstate_attrs,
2661};
2662
2663/*
Kay Sievers10fbcf42011-12-21 14:48:43 -08002664 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002665 * Returns node id via non-NULL nidp.
2666 */
2667static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
2668{
2669 int nid;
2670
2671 for (nid = 0; nid < nr_node_ids; nid++) {
2672 struct node_hstate *nhs = &node_hstates[nid];
2673 int i;
2674 for (i = 0; i < HUGE_MAX_HSTATE; i++)
2675 if (nhs->hstate_kobjs[i] == kobj) {
2676 if (nidp)
2677 *nidp = nid;
2678 return &hstates[i];
2679 }
2680 }
2681
2682 BUG();
2683 return NULL;
2684}
2685
2686/*
Kay Sievers10fbcf42011-12-21 14:48:43 -08002687 * Unregister hstate attributes from a single node device.
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002688 * No-op if no hstate attributes attached.
2689 */
Claudiu Ghioc3cd8b442013-03-04 12:46:15 +02002690static void hugetlb_unregister_node(struct node *node)
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002691{
2692 struct hstate *h;
Kay Sievers10fbcf42011-12-21 14:48:43 -08002693 struct node_hstate *nhs = &node_hstates[node->dev.id];
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002694
2695 if (!nhs->hugepages_kobj)
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -08002696 return; /* no hstate attributes */
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002697
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07002698 for_each_hstate(h) {
2699 int idx = hstate_index(h);
2700 if (nhs->hstate_kobjs[idx]) {
2701 kobject_put(nhs->hstate_kobjs[idx]);
2702 nhs->hstate_kobjs[idx] = NULL;
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002703 }
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07002704 }
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002705
2706 kobject_put(nhs->hugepages_kobj);
2707 nhs->hugepages_kobj = NULL;
2708}
2709
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002710
2711/*
Kay Sievers10fbcf42011-12-21 14:48:43 -08002712 * Register hstate attributes for a single node device.
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002713 * No-op if attributes already registered.
2714 */
Claudiu Ghioc3cd8b442013-03-04 12:46:15 +02002715static void hugetlb_register_node(struct node *node)
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002716{
2717 struct hstate *h;
Kay Sievers10fbcf42011-12-21 14:48:43 -08002718 struct node_hstate *nhs = &node_hstates[node->dev.id];
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002719 int err;
2720
2721 if (nhs->hugepages_kobj)
2722 return; /* already allocated */
2723
2724 nhs->hugepages_kobj = kobject_create_and_add("hugepages",
Kay Sievers10fbcf42011-12-21 14:48:43 -08002725 &node->dev.kobj);
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002726 if (!nhs->hugepages_kobj)
2727 return;
2728
2729 for_each_hstate(h) {
2730 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
2731 nhs->hstate_kobjs,
2732 &per_node_hstate_attr_group);
2733 if (err) {
Andrew Mortonffb22af2013-02-22 16:32:08 -08002734 pr_err("Hugetlb: Unable to add hstate %s for node %d\n",
2735 h->name, node->dev.id);
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002736 hugetlb_unregister_node(node);
2737 break;
2738 }
2739 }
2740}
2741
2742/*
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -08002743 * hugetlb init time: register hstate attributes for all registered node
Kay Sievers10fbcf42011-12-21 14:48:43 -08002744 * devices of nodes that have memory. All on-line nodes should have
2745 * registered their associated device by this time.
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002746 */
Luiz Capitulino7d9ca002014-12-12 16:55:24 -08002747static void __init hugetlb_register_all_nodes(void)
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002748{
2749 int nid;
2750
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08002751 for_each_node_state(nid, N_MEMORY) {
Wen Congyang87327942012-12-11 16:00:56 -08002752 struct node *node = node_devices[nid];
Kay Sievers10fbcf42011-12-21 14:48:43 -08002753 if (node->dev.id == nid)
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002754 hugetlb_register_node(node);
2755 }
2756
2757 /*
Kay Sievers10fbcf42011-12-21 14:48:43 -08002758 * Let the node device driver know we're here so it can
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002759 * [un]register hstate attributes on node hotplug.
2760 */
2761 register_hugetlbfs_with_node(hugetlb_register_node,
2762 hugetlb_unregister_node);
2763}
2764#else /* !CONFIG_NUMA */
2765
2766static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
2767{
2768 BUG();
2769 if (nidp)
2770 *nidp = -1;
2771 return NULL;
2772}
2773
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002774static void hugetlb_register_all_nodes(void) { }
2775
2776#endif
2777
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002778static int __init hugetlb_init(void)
2779{
Davidlohr Bueso8382d912014-04-03 14:47:31 -07002780 int i;
2781
Nishanth Aravamudan457c1b22014-05-06 12:50:00 -07002782 if (!hugepages_supported())
Benjamin Herrenschmidt0ef89d22008-07-31 00:07:30 -07002783 return 0;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002784
Nick Piggine11bfbf2008-07-23 21:27:52 -07002785 if (!size_to_hstate(default_hstate_size)) {
2786 default_hstate_size = HPAGE_SIZE;
2787 if (!size_to_hstate(default_hstate_size))
2788 hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002789 }
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07002790 default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size));
Vaishali Thakkarf8b74812016-02-17 13:11:26 -08002791 if (default_hstate_max_huge_pages) {
2792 if (!default_hstate.max_huge_pages)
2793 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
2794 }
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002795
2796 hugetlb_init_hstates();
Andi Kleenaa888a72008-07-23 21:27:47 -07002797 gather_bootmem_prealloc();
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002798 report_hugepages();
2799
2800 hugetlb_sysfs_init();
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002801 hugetlb_register_all_nodes();
Jianguo Wu7179e7b2012-12-18 14:23:19 -08002802 hugetlb_cgroup_file_init();
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002803
Davidlohr Bueso8382d912014-04-03 14:47:31 -07002804#ifdef CONFIG_SMP
2805 num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus());
2806#else
2807 num_fault_mutexes = 1;
2808#endif
Mike Kravetzc672c7f2015-09-08 15:01:35 -07002809 hugetlb_fault_mutex_table =
Davidlohr Bueso8382d912014-04-03 14:47:31 -07002810 kmalloc(sizeof(struct mutex) * num_fault_mutexes, GFP_KERNEL);
Mike Kravetzc672c7f2015-09-08 15:01:35 -07002811 BUG_ON(!hugetlb_fault_mutex_table);
Davidlohr Bueso8382d912014-04-03 14:47:31 -07002812
2813 for (i = 0; i < num_fault_mutexes; i++)
Mike Kravetzc672c7f2015-09-08 15:01:35 -07002814 mutex_init(&hugetlb_fault_mutex_table[i]);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002815 return 0;
2816}
Paul Gortmaker3e89e1c2016-01-14 15:21:52 -08002817subsys_initcall(hugetlb_init);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002818
2819/* Should be called on processing a hugepagesz=... option */
Vaishali Thakkar9fee0212016-05-19 17:11:04 -07002820void __init hugetlb_bad_size(void)
2821{
2822 parsed_valid_hugepagesz = false;
2823}
2824
Kirill A. Shutemovd00181b2015-11-06 16:29:57 -08002825void __init hugetlb_add_hstate(unsigned int order)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002826{
2827 struct hstate *h;
Andi Kleen8faa8b02008-07-23 21:27:48 -07002828 unsigned long i;
2829
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002830 if (size_to_hstate(PAGE_SIZE << order)) {
Joe Perches598d8092016-03-17 14:19:44 -07002831 pr_warn("hugepagesz= specified twice, ignoring\n");
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002832 return;
2833 }
Aneesh Kumar K.V47d38342012-07-31 16:41:54 -07002834 BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002835 BUG_ON(order == 0);
Aneesh Kumar K.V47d38342012-07-31 16:41:54 -07002836 h = &hstates[hugetlb_max_hstate++];
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002837 h->order = order;
2838 h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
Andi Kleen8faa8b02008-07-23 21:27:48 -07002839 h->nr_huge_pages = 0;
2840 h->free_huge_pages = 0;
2841 for (i = 0; i < MAX_NUMNODES; ++i)
2842 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
Aneesh Kumar K.V0edaecf2012-07-31 16:42:07 -07002843 INIT_LIST_HEAD(&h->hugepage_activelist);
Andrew Morton54f18d32016-05-19 17:11:40 -07002844 h->next_nid_to_alloc = first_memory_node;
2845 h->next_nid_to_free = first_memory_node;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002846 snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
2847 huge_page_size(h)/1024);
Andi Kleen8faa8b02008-07-23 21:27:48 -07002848
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002849 parsed_hstate = h;
2850}
2851
Nick Piggine11bfbf2008-07-23 21:27:52 -07002852static int __init hugetlb_nrpages_setup(char *s)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002853{
2854 unsigned long *mhp;
Andi Kleen8faa8b02008-07-23 21:27:48 -07002855 static unsigned long *last_mhp;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002856
Vaishali Thakkar9fee0212016-05-19 17:11:04 -07002857 if (!parsed_valid_hugepagesz) {
2858 pr_warn("hugepages = %s preceded by "
2859 "an unsupported hugepagesz, ignoring\n", s);
2860 parsed_valid_hugepagesz = true;
2861 return 1;
2862 }
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002863 /*
Aneesh Kumar K.V47d38342012-07-31 16:41:54 -07002864 * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet,
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002865 * so this hugepages= parameter goes to the "default hstate".
2866 */
Vaishali Thakkar9fee0212016-05-19 17:11:04 -07002867 else if (!hugetlb_max_hstate)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002868 mhp = &default_hstate_max_huge_pages;
2869 else
2870 mhp = &parsed_hstate->max_huge_pages;
2871
Andi Kleen8faa8b02008-07-23 21:27:48 -07002872 if (mhp == last_mhp) {
Joe Perches598d8092016-03-17 14:19:44 -07002873 pr_warn("hugepages= specified twice without interleaving hugepagesz=, ignoring\n");
Andi Kleen8faa8b02008-07-23 21:27:48 -07002874 return 1;
2875 }
2876
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002877 if (sscanf(s, "%lu", mhp) <= 0)
2878 *mhp = 0;
2879
Andi Kleen8faa8b02008-07-23 21:27:48 -07002880 /*
2881 * Global state is always initialized later in hugetlb_init.
2882 * But we need to allocate >= MAX_ORDER hstates here early to still
2883 * use the bootmem allocator.
2884 */
Aneesh Kumar K.V47d38342012-07-31 16:41:54 -07002885 if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER)
Andi Kleen8faa8b02008-07-23 21:27:48 -07002886 hugetlb_hstate_alloc_pages(parsed_hstate);
2887
2888 last_mhp = mhp;
2889
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002890 return 1;
2891}
Nick Piggine11bfbf2008-07-23 21:27:52 -07002892__setup("hugepages=", hugetlb_nrpages_setup);
2893
2894static int __init hugetlb_default_setup(char *s)
2895{
2896 default_hstate_size = memparse(s, &s);
2897 return 1;
2898}
2899__setup("default_hugepagesz=", hugetlb_default_setup);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002900
Nishanth Aravamudan8a213462008-07-25 19:44:37 -07002901static unsigned int cpuset_mems_nr(unsigned int *array)
2902{
2903 int node;
2904 unsigned int nr = 0;
2905
2906 for_each_node_mask(node, cpuset_current_mems_allowed)
2907 nr += array[node];
2908
2909 return nr;
2910}
2911
2912#ifdef CONFIG_SYSCTL
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002913static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
2914 struct ctl_table *table, int write,
2915 void __user *buffer, size_t *length, loff_t *ppos)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002916{
Andi Kleene5ff2152008-07-23 21:27:42 -07002917 struct hstate *h = &default_hstate;
David Rientjes238d3c12014-08-06 16:06:51 -07002918 unsigned long tmp = h->max_huge_pages;
Michal Hocko08d4a242011-01-13 15:47:26 -08002919 int ret;
Andi Kleene5ff2152008-07-23 21:27:42 -07002920
Nishanth Aravamudan457c1b22014-05-06 12:50:00 -07002921 if (!hugepages_supported())
Jan Stancek86613622016-03-09 14:08:35 -08002922 return -EOPNOTSUPP;
Nishanth Aravamudan457c1b22014-05-06 12:50:00 -07002923
Andi Kleene5ff2152008-07-23 21:27:42 -07002924 table->data = &tmp;
2925 table->maxlen = sizeof(unsigned long);
Michal Hocko08d4a242011-01-13 15:47:26 -08002926 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2927 if (ret)
2928 goto out;
Andi Kleene5ff2152008-07-23 21:27:42 -07002929
David Rientjes238d3c12014-08-06 16:06:51 -07002930 if (write)
2931 ret = __nr_hugepages_store_common(obey_mempolicy, h,
2932 NUMA_NO_NODE, tmp, *length);
Michal Hocko08d4a242011-01-13 15:47:26 -08002933out:
2934 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002935}
Mel Gorman396faf02007-07-17 04:03:13 -07002936
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002937int hugetlb_sysctl_handler(struct ctl_table *table, int write,
2938 void __user *buffer, size_t *length, loff_t *ppos)
2939{
2940
2941 return hugetlb_sysctl_handler_common(false, table, write,
2942 buffer, length, ppos);
2943}
2944
2945#ifdef CONFIG_NUMA
2946int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
2947 void __user *buffer, size_t *length, loff_t *ppos)
2948{
2949 return hugetlb_sysctl_handler_common(true, table, write,
2950 buffer, length, ppos);
2951}
2952#endif /* CONFIG_NUMA */
2953
Nishanth Aravamudana3d0c6a2008-02-08 04:18:18 -08002954int hugetlb_overcommit_handler(struct ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07002955 void __user *buffer,
Nishanth Aravamudana3d0c6a2008-02-08 04:18:18 -08002956 size_t *length, loff_t *ppos)
2957{
Andi Kleena5516432008-07-23 21:27:41 -07002958 struct hstate *h = &default_hstate;
Andi Kleene5ff2152008-07-23 21:27:42 -07002959 unsigned long tmp;
Michal Hocko08d4a242011-01-13 15:47:26 -08002960 int ret;
Andi Kleene5ff2152008-07-23 21:27:42 -07002961
Nishanth Aravamudan457c1b22014-05-06 12:50:00 -07002962 if (!hugepages_supported())
Jan Stancek86613622016-03-09 14:08:35 -08002963 return -EOPNOTSUPP;
Nishanth Aravamudan457c1b22014-05-06 12:50:00 -07002964
Petr Holasekc033a932011-03-22 16:33:05 -07002965 tmp = h->nr_overcommit_huge_pages;
Andi Kleene5ff2152008-07-23 21:27:42 -07002966
Luiz Capitulinobae7f4a2014-06-04 16:07:08 -07002967 if (write && hstate_is_gigantic(h))
Eric B Munsonadbe8722011-01-13 15:47:27 -08002968 return -EINVAL;
2969
Andi Kleene5ff2152008-07-23 21:27:42 -07002970 table->data = &tmp;
2971 table->maxlen = sizeof(unsigned long);
Michal Hocko08d4a242011-01-13 15:47:26 -08002972 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2973 if (ret)
2974 goto out;
Andi Kleene5ff2152008-07-23 21:27:42 -07002975
2976 if (write) {
2977 spin_lock(&hugetlb_lock);
2978 h->nr_overcommit_huge_pages = tmp;
2979 spin_unlock(&hugetlb_lock);
2980 }
Michal Hocko08d4a242011-01-13 15:47:26 -08002981out:
2982 return ret;
Nishanth Aravamudana3d0c6a2008-02-08 04:18:18 -08002983}
2984
Linus Torvalds1da177e2005-04-16 15:20:36 -07002985#endif /* CONFIG_SYSCTL */
2986
Alexey Dobriyane1759c22008-10-15 23:50:22 +04002987void hugetlb_report_meminfo(struct seq_file *m)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002988{
Andi Kleena5516432008-07-23 21:27:41 -07002989 struct hstate *h = &default_hstate;
Nishanth Aravamudan457c1b22014-05-06 12:50:00 -07002990 if (!hugepages_supported())
2991 return;
Alexey Dobriyane1759c22008-10-15 23:50:22 +04002992 seq_printf(m,
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002993 "HugePages_Total: %5lu\n"
2994 "HugePages_Free: %5lu\n"
2995 "HugePages_Rsvd: %5lu\n"
2996 "HugePages_Surp: %5lu\n"
2997 "Hugepagesize: %8lu kB\n",
Andi Kleena5516432008-07-23 21:27:41 -07002998 h->nr_huge_pages,
2999 h->free_huge_pages,
3000 h->resv_huge_pages,
3001 h->surplus_huge_pages,
3002 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003003}
3004
3005int hugetlb_report_node_meminfo(int nid, char *buf)
3006{
Andi Kleena5516432008-07-23 21:27:41 -07003007 struct hstate *h = &default_hstate;
Nishanth Aravamudan457c1b22014-05-06 12:50:00 -07003008 if (!hugepages_supported())
3009 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003010 return sprintf(buf,
3011 "Node %d HugePages_Total: %5u\n"
Nishanth Aravamudana1de0912008-03-26 14:37:53 -07003012 "Node %d HugePages_Free: %5u\n"
3013 "Node %d HugePages_Surp: %5u\n",
Andi Kleena5516432008-07-23 21:27:41 -07003014 nid, h->nr_huge_pages_node[nid],
3015 nid, h->free_huge_pages_node[nid],
3016 nid, h->surplus_huge_pages_node[nid]);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003017}
3018
David Rientjes949f7ec2013-04-29 15:07:48 -07003019void hugetlb_show_meminfo(void)
3020{
3021 struct hstate *h;
3022 int nid;
3023
Nishanth Aravamudan457c1b22014-05-06 12:50:00 -07003024 if (!hugepages_supported())
3025 return;
3026
David Rientjes949f7ec2013-04-29 15:07:48 -07003027 for_each_node_state(nid, N_MEMORY)
3028 for_each_hstate(h)
3029 pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
3030 nid,
3031 h->nr_huge_pages_node[nid],
3032 h->free_huge_pages_node[nid],
3033 h->surplus_huge_pages_node[nid],
3034 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
3035}
3036
Naoya Horiguchi5d317b22015-11-05 18:47:14 -08003037void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm)
3038{
3039 seq_printf(m, "HugetlbPages:\t%8lu kB\n",
3040 atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10));
3041}
3042
Linus Torvalds1da177e2005-04-16 15:20:36 -07003043/* Return the number pages of memory we physically have, in PAGE_SIZE units. */
3044unsigned long hugetlb_total_pages(void)
3045{
Wanpeng Lid0028582013-03-22 15:04:40 -07003046 struct hstate *h;
3047 unsigned long nr_total_pages = 0;
3048
3049 for_each_hstate(h)
3050 nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h);
3051 return nr_total_pages;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003052}
Linus Torvalds1da177e2005-04-16 15:20:36 -07003053
Andi Kleena5516432008-07-23 21:27:41 -07003054static int hugetlb_acct_memory(struct hstate *h, long delta)
Mel Gormanfc1b8a72008-07-23 21:27:22 -07003055{
3056 int ret = -ENOMEM;
3057
3058 spin_lock(&hugetlb_lock);
3059 /*
3060 * When cpuset is configured, it breaks the strict hugetlb page
3061 * reservation as the accounting is done on a global variable. Such
3062 * reservation is completely rubbish in the presence of cpuset because
3063 * the reservation is not checked against page availability for the
3064 * current cpuset. Application can still potentially OOM'ed by kernel
3065 * with lack of free htlb page in cpuset that the task is in.
3066 * Attempt to enforce strict accounting with cpuset is almost
3067 * impossible (or too ugly) because cpuset is too fluid that
3068 * task or memory node can be dynamically moved between cpusets.
3069 *
3070 * The change of semantics for shared hugetlb mapping with cpuset is
3071 * undesirable. However, in order to preserve some of the semantics,
3072 * we fall back to check against current free page availability as
3073 * a best attempt and hopefully to minimize the impact of changing
3074 * semantics that cpuset has.
3075 */
3076 if (delta > 0) {
Andi Kleena5516432008-07-23 21:27:41 -07003077 if (gather_surplus_pages(h, delta) < 0)
Mel Gormanfc1b8a72008-07-23 21:27:22 -07003078 goto out;
3079
Andi Kleena5516432008-07-23 21:27:41 -07003080 if (delta > cpuset_mems_nr(h->free_huge_pages_node)) {
3081 return_unused_surplus_pages(h, delta);
Mel Gormanfc1b8a72008-07-23 21:27:22 -07003082 goto out;
3083 }
3084 }
3085
3086 ret = 0;
3087 if (delta < 0)
Andi Kleena5516432008-07-23 21:27:41 -07003088 return_unused_surplus_pages(h, (unsigned long) -delta);
Mel Gormanfc1b8a72008-07-23 21:27:22 -07003089
3090out:
3091 spin_unlock(&hugetlb_lock);
3092 return ret;
3093}
3094
Andy Whitcroft84afd992008-07-23 21:27:32 -07003095static void hugetlb_vm_op_open(struct vm_area_struct *vma)
3096{
Joonsoo Kimf522c3a2013-09-11 14:21:53 -07003097 struct resv_map *resv = vma_resv_map(vma);
Andy Whitcroft84afd992008-07-23 21:27:32 -07003098
3099 /*
3100 * This new VMA should share its siblings reservation map if present.
3101 * The VMA will only ever have a valid reservation map pointer where
3102 * it is being copied for another still existing VMA. As that VMA
Lucas De Marchi25985ed2011-03-30 22:57:33 -03003103 * has a reference to the reservation map it cannot disappear until
Andy Whitcroft84afd992008-07-23 21:27:32 -07003104 * after this open call completes. It is therefore safe to take a
3105 * new reference here without additional locking.
3106 */
Joonsoo Kim4e35f482014-04-03 14:47:30 -07003107 if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
Joonsoo Kimf522c3a2013-09-11 14:21:53 -07003108 kref_get(&resv->refs);
Andy Whitcroft84afd992008-07-23 21:27:32 -07003109}
3110
Mel Gormana1e78772008-07-23 21:27:23 -07003111static void hugetlb_vm_op_close(struct vm_area_struct *vma)
3112{
Andi Kleena5516432008-07-23 21:27:41 -07003113 struct hstate *h = hstate_vma(vma);
Joonsoo Kimf522c3a2013-09-11 14:21:53 -07003114 struct resv_map *resv = vma_resv_map(vma);
David Gibson90481622012-03-21 16:34:12 -07003115 struct hugepage_subpool *spool = subpool_vma(vma);
Joonsoo Kim4e35f482014-04-03 14:47:30 -07003116 unsigned long reserve, start, end;
Mike Kravetz1c5ecae2015-04-15 16:13:39 -07003117 long gbl_reserve;
Andy Whitcroft84afd992008-07-23 21:27:32 -07003118
Joonsoo Kim4e35f482014-04-03 14:47:30 -07003119 if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER))
3120 return;
Andy Whitcroft84afd992008-07-23 21:27:32 -07003121
Joonsoo Kim4e35f482014-04-03 14:47:30 -07003122 start = vma_hugecache_offset(h, vma, vma->vm_start);
3123 end = vma_hugecache_offset(h, vma, vma->vm_end);
Andy Whitcroft84afd992008-07-23 21:27:32 -07003124
Joonsoo Kim4e35f482014-04-03 14:47:30 -07003125 reserve = (end - start) - region_count(resv, start, end);
Andy Whitcroft84afd992008-07-23 21:27:32 -07003126
Joonsoo Kim4e35f482014-04-03 14:47:30 -07003127 kref_put(&resv->refs, resv_map_release);
3128
3129 if (reserve) {
Mike Kravetz1c5ecae2015-04-15 16:13:39 -07003130 /*
3131 * Decrement reserve counts. The global reserve count may be
3132 * adjusted if the subpool has a minimum size.
3133 */
3134 gbl_reserve = hugepage_subpool_put_pages(spool, reserve);
3135 hugetlb_acct_memory(h, -gbl_reserve);
Andy Whitcroft84afd992008-07-23 21:27:32 -07003136 }
Mel Gormana1e78772008-07-23 21:27:23 -07003137}
3138
Dan Williamscebe1392017-11-29 16:10:28 -08003139static int hugetlb_vm_op_split(struct vm_area_struct *vma, unsigned long addr)
3140{
3141 if (addr & ~(huge_page_mask(hstate_vma(vma))))
3142 return -EINVAL;
3143 return 0;
3144}
3145
Linus Torvalds1da177e2005-04-16 15:20:36 -07003146/*
3147 * We cannot handle pagefaults against hugetlb pages at all. They cause
3148 * handle_mm_fault() to try to instantiate regular-sized pages in the
3149 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
3150 * this far.
3151 */
Nick Piggind0217ac2007-07-19 01:47:03 -07003152static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003153{
3154 BUG();
Nick Piggind0217ac2007-07-19 01:47:03 -07003155 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003156}
3157
Alexey Dobriyanf0f37e22009-09-27 22:29:37 +04003158const struct vm_operations_struct hugetlb_vm_ops = {
Nick Piggind0217ac2007-07-19 01:47:03 -07003159 .fault = hugetlb_vm_op_fault,
Andy Whitcroft84afd992008-07-23 21:27:32 -07003160 .open = hugetlb_vm_op_open,
Mel Gormana1e78772008-07-23 21:27:23 -07003161 .close = hugetlb_vm_op_close,
Dan Williamscebe1392017-11-29 16:10:28 -08003162 .split = hugetlb_vm_op_split,
Linus Torvalds1da177e2005-04-16 15:20:36 -07003163};
3164
David Gibson1e8f8892006-01-06 00:10:44 -08003165static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
3166 int writable)
David Gibson63551ae2005-06-21 17:14:44 -07003167{
3168 pte_t entry;
3169
David Gibson1e8f8892006-01-06 00:10:44 -08003170 if (writable) {
Gerald Schaefer106c9922013-04-29 15:07:23 -07003171 entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page,
3172 vma->vm_page_prot)));
David Gibson63551ae2005-06-21 17:14:44 -07003173 } else {
Gerald Schaefer106c9922013-04-29 15:07:23 -07003174 entry = huge_pte_wrprotect(mk_huge_pte(page,
3175 vma->vm_page_prot));
David Gibson63551ae2005-06-21 17:14:44 -07003176 }
3177 entry = pte_mkyoung(entry);
3178 entry = pte_mkhuge(entry);
Chris Metcalfd9ed9fa2012-04-01 14:01:34 -04003179 entry = arch_make_huge_pte(entry, vma, page, writable);
David Gibson63551ae2005-06-21 17:14:44 -07003180
3181 return entry;
3182}
3183
David Gibson1e8f8892006-01-06 00:10:44 -08003184static void set_huge_ptep_writable(struct vm_area_struct *vma,
3185 unsigned long address, pte_t *ptep)
3186{
3187 pte_t entry;
3188
Gerald Schaefer106c9922013-04-29 15:07:23 -07003189 entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep)));
Chris Forbes32f84522011-07-25 17:12:14 -07003190 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
Russell King4b3073e2009-12-18 16:40:18 +00003191 update_mmu_cache(vma, address, ptep);
David Gibson1e8f8892006-01-06 00:10:44 -08003192}
3193
Naoya Horiguchi4a705fe2014-06-23 13:22:03 -07003194static int is_hugetlb_entry_migration(pte_t pte)
3195{
3196 swp_entry_t swp;
3197
3198 if (huge_pte_none(pte) || pte_present(pte))
3199 return 0;
3200 swp = pte_to_swp_entry(pte);
3201 if (non_swap_entry(swp) && is_migration_entry(swp))
3202 return 1;
3203 else
3204 return 0;
3205}
3206
3207static int is_hugetlb_entry_hwpoisoned(pte_t pte)
3208{
3209 swp_entry_t swp;
3210
3211 if (huge_pte_none(pte) || pte_present(pte))
3212 return 0;
3213 swp = pte_to_swp_entry(pte);
3214 if (non_swap_entry(swp) && is_hwpoison_entry(swp))
3215 return 1;
3216 else
3217 return 0;
3218}
David Gibson1e8f8892006-01-06 00:10:44 -08003219
David Gibson63551ae2005-06-21 17:14:44 -07003220int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
3221 struct vm_area_struct *vma)
3222{
Mike Kravetzf8d4c942018-11-16 15:08:04 -08003223 pte_t *src_pte, *dst_pte, entry, dst_entry;
David Gibson63551ae2005-06-21 17:14:44 -07003224 struct page *ptepage;
Hugh Dickins1c598272005-10-19 21:23:43 -07003225 unsigned long addr;
David Gibson1e8f8892006-01-06 00:10:44 -08003226 int cow;
Andi Kleena5516432008-07-23 21:27:41 -07003227 struct hstate *h = hstate_vma(vma);
3228 unsigned long sz = huge_page_size(h);
Andreas Sandberge8569dd2014-01-21 15:49:09 -08003229 unsigned long mmun_start; /* For mmu_notifiers */
3230 unsigned long mmun_end; /* For mmu_notifiers */
3231 int ret = 0;
David Gibson1e8f8892006-01-06 00:10:44 -08003232
3233 cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
David Gibson63551ae2005-06-21 17:14:44 -07003234
Andreas Sandberge8569dd2014-01-21 15:49:09 -08003235 mmun_start = vma->vm_start;
3236 mmun_end = vma->vm_end;
3237 if (cow)
3238 mmu_notifier_invalidate_range_start(src, mmun_start, mmun_end);
3239
Andi Kleena5516432008-07-23 21:27:41 -07003240 for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003241 spinlock_t *src_ptl, *dst_ptl;
Hugh Dickinsc74df322005-10-29 18:16:23 -07003242 src_pte = huge_pte_offset(src, addr);
3243 if (!src_pte)
3244 continue;
Andi Kleena5516432008-07-23 21:27:41 -07003245 dst_pte = huge_pte_alloc(dst, addr, sz);
Andreas Sandberge8569dd2014-01-21 15:49:09 -08003246 if (!dst_pte) {
3247 ret = -ENOMEM;
3248 break;
3249 }
Larry Woodmanc5c99422008-01-24 05:49:25 -08003250
Mike Kravetzf8d4c942018-11-16 15:08:04 -08003251 /*
3252 * If the pagetables are shared don't copy or take references.
3253 * dst_pte == src_pte is the common case of src/dest sharing.
3254 *
3255 * However, src could have 'unshared' and dst shares with
3256 * another vma. If dst_pte !none, this implies sharing.
3257 * Check here before taking page table lock, and once again
3258 * after taking the lock below.
3259 */
3260 dst_entry = huge_ptep_get(dst_pte);
3261 if ((dst_pte == src_pte) || !huge_pte_none(dst_entry))
Larry Woodmanc5c99422008-01-24 05:49:25 -08003262 continue;
3263
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003264 dst_ptl = huge_pte_lock(h, dst, dst_pte);
3265 src_ptl = huge_pte_lockptr(h, src, src_pte);
3266 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
Naoya Horiguchi4a705fe2014-06-23 13:22:03 -07003267 entry = huge_ptep_get(src_pte);
Mike Kravetzf8d4c942018-11-16 15:08:04 -08003268 dst_entry = huge_ptep_get(dst_pte);
3269 if (huge_pte_none(entry) || !huge_pte_none(dst_entry)) {
3270 /*
3271 * Skip if src entry none. Also, skip in the
3272 * unlikely case dst entry !none as this implies
3273 * sharing with another vma.
3274 */
Naoya Horiguchi4a705fe2014-06-23 13:22:03 -07003275 ;
3276 } else if (unlikely(is_hugetlb_entry_migration(entry) ||
3277 is_hugetlb_entry_hwpoisoned(entry))) {
3278 swp_entry_t swp_entry = pte_to_swp_entry(entry);
3279
3280 if (is_write_migration_entry(swp_entry) && cow) {
3281 /*
3282 * COW mappings require pages in both
3283 * parent and child to be set to read.
3284 */
3285 make_migration_entry_read(&swp_entry);
3286 entry = swp_entry_to_pte(swp_entry);
3287 set_huge_pte_at(src, addr, src_pte, entry);
3288 }
3289 set_huge_pte_at(dst, addr, dst_pte, entry);
3290 } else {
Joerg Roedel34ee6452014-11-13 13:46:09 +11003291 if (cow) {
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07003292 huge_ptep_set_wrprotect(src, addr, src_pte);
Joerg Roedel34ee6452014-11-13 13:46:09 +11003293 mmu_notifier_invalidate_range(src, mmun_start,
3294 mmun_end);
3295 }
Naoya Horiguchi0253d632014-07-23 14:00:19 -07003296 entry = huge_ptep_get(src_pte);
Hugh Dickins1c598272005-10-19 21:23:43 -07003297 ptepage = pte_page(entry);
3298 get_page(ptepage);
Kirill A. Shutemov53f92632016-01-15 16:53:42 -08003299 page_dup_rmap(ptepage, true);
Hugh Dickins1c598272005-10-19 21:23:43 -07003300 set_huge_pte_at(dst, addr, dst_pte, entry);
Naoya Horiguchi5d317b22015-11-05 18:47:14 -08003301 hugetlb_count_add(pages_per_huge_page(h), dst);
Hugh Dickins1c598272005-10-19 21:23:43 -07003302 }
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003303 spin_unlock(src_ptl);
3304 spin_unlock(dst_ptl);
David Gibson63551ae2005-06-21 17:14:44 -07003305 }
David Gibson63551ae2005-06-21 17:14:44 -07003306
Andreas Sandberge8569dd2014-01-21 15:49:09 -08003307 if (cow)
3308 mmu_notifier_invalidate_range_end(src, mmun_start, mmun_end);
3309
3310 return ret;
David Gibson63551ae2005-06-21 17:14:44 -07003311}
3312
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07003313void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
3314 unsigned long start, unsigned long end,
3315 struct page *ref_page)
David Gibson63551ae2005-06-21 17:14:44 -07003316{
3317 struct mm_struct *mm = vma->vm_mm;
3318 unsigned long address;
David Gibsonc7546f82005-08-05 11:59:35 -07003319 pte_t *ptep;
David Gibson63551ae2005-06-21 17:14:44 -07003320 pte_t pte;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003321 spinlock_t *ptl;
David Gibson63551ae2005-06-21 17:14:44 -07003322 struct page *page;
Andi Kleena5516432008-07-23 21:27:41 -07003323 struct hstate *h = hstate_vma(vma);
3324 unsigned long sz = huge_page_size(h);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07003325 const unsigned long mmun_start = start; /* For mmu_notifiers */
3326 const unsigned long mmun_end = end; /* For mmu_notifiers */
Andi Kleena5516432008-07-23 21:27:41 -07003327
David Gibson63551ae2005-06-21 17:14:44 -07003328 WARN_ON(!is_vm_hugetlb_page(vma));
Andi Kleena5516432008-07-23 21:27:41 -07003329 BUG_ON(start & ~huge_page_mask(h));
3330 BUG_ON(end & ~huge_page_mask(h));
David Gibson63551ae2005-06-21 17:14:44 -07003331
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07003332 tlb_start_vma(tlb, vma);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07003333 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
Hillf Danton569f48b2014-12-10 15:44:41 -08003334 address = start;
Hillf Danton569f48b2014-12-10 15:44:41 -08003335 for (; address < end; address += sz) {
David Gibsonc7546f82005-08-05 11:59:35 -07003336 ptep = huge_pte_offset(mm, address);
Adam Litke4c887262005-10-29 18:16:46 -07003337 if (!ptep)
David Gibsonc7546f82005-08-05 11:59:35 -07003338 continue;
3339
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003340 ptl = huge_pte_lock(h, mm, ptep);
Aneesh Kumar K.V31d49da2016-07-26 15:24:06 -07003341 if (huge_pmd_unshare(mm, &address, ptep)) {
3342 spin_unlock(ptl);
3343 continue;
3344 }
Chen, Kenneth W39dde652006-12-06 20:32:03 -08003345
Hillf Danton66293262012-03-23 15:01:48 -07003346 pte = huge_ptep_get(ptep);
Aneesh Kumar K.V31d49da2016-07-26 15:24:06 -07003347 if (huge_pte_none(pte)) {
3348 spin_unlock(ptl);
3349 continue;
3350 }
Hillf Danton66293262012-03-23 15:01:48 -07003351
3352 /*
Naoya Horiguchi9fbc1f62015-02-11 15:25:32 -08003353 * Migrating hugepage or HWPoisoned hugepage is already
3354 * unmapped and its refcount is dropped, so just clear pte here.
Hillf Danton66293262012-03-23 15:01:48 -07003355 */
Naoya Horiguchi9fbc1f62015-02-11 15:25:32 -08003356 if (unlikely(!pte_present(pte))) {
Gerald Schaefer106c9922013-04-29 15:07:23 -07003357 huge_pte_clear(mm, address, ptep);
Aneesh Kumar K.V31d49da2016-07-26 15:24:06 -07003358 spin_unlock(ptl);
3359 continue;
Naoya Horiguchi8c4894c2012-12-12 13:52:28 -08003360 }
Hillf Danton66293262012-03-23 15:01:48 -07003361
3362 page = pte_page(pte);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003363 /*
3364 * If a reference page is supplied, it is because a specific
3365 * page is being unmapped, not a range. Ensure the page we
3366 * are about to unmap is the actual page of interest.
3367 */
3368 if (ref_page) {
Aneesh Kumar K.V31d49da2016-07-26 15:24:06 -07003369 if (page != ref_page) {
3370 spin_unlock(ptl);
3371 continue;
3372 }
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003373 /*
3374 * Mark the VMA as having unmapped its page so that
3375 * future faults in this VMA will fail rather than
3376 * looking like data was lost
3377 */
3378 set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
3379 }
3380
David Gibsonc7546f82005-08-05 11:59:35 -07003381 pte = huge_ptep_get_and_clear(mm, address, ptep);
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07003382 tlb_remove_tlb_entry(tlb, ptep, address);
Gerald Schaefer106c9922013-04-29 15:07:23 -07003383 if (huge_pte_dirty(pte))
Ken Chen6649a382007-02-08 14:20:27 -08003384 set_page_dirty(page);
Hillf Danton9e811302012-03-21 16:34:03 -07003385
Naoya Horiguchi5d317b22015-11-05 18:47:14 -08003386 hugetlb_count_sub(pages_per_huge_page(h), mm);
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -08003387 page_remove_rmap(page, true);
Aneesh Kumar K.V31d49da2016-07-26 15:24:06 -07003388
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003389 spin_unlock(ptl);
Aneesh Kumar K.Ve77b0852016-07-26 15:24:12 -07003390 tlb_remove_page_size(tlb, page, huge_page_size(h));
Aneesh Kumar K.V31d49da2016-07-26 15:24:06 -07003391 /*
3392 * Bail out after unmapping reference page if supplied
3393 */
3394 if (ref_page)
3395 break;
Chen, Kenneth Wfe1668a2006-10-04 02:15:24 -07003396 }
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07003397 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07003398 tlb_end_vma(tlb, vma);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003399}
David Gibson63551ae2005-06-21 17:14:44 -07003400
Mel Gormand8333522012-07-31 16:46:20 -07003401void __unmap_hugepage_range_final(struct mmu_gather *tlb,
3402 struct vm_area_struct *vma, unsigned long start,
3403 unsigned long end, struct page *ref_page)
3404{
3405 __unmap_hugepage_range(tlb, vma, start, end, ref_page);
3406
3407 /*
3408 * Clear this flag so that x86's huge_pmd_share page_table_shareable
3409 * test will fail on a vma being torn down, and not grab a page table
3410 * on its way out. We're lucky that the flag has such an appropriate
3411 * name, and can in fact be safely cleared here. We could clear it
3412 * before the __unmap_hugepage_range above, but all that's necessary
Davidlohr Buesoc8c06ef2014-12-12 16:54:24 -08003413 * is to clear it before releasing the i_mmap_rwsem. This works
Mel Gormand8333522012-07-31 16:46:20 -07003414 * because in the context this is called, the VMA is about to be
Davidlohr Buesoc8c06ef2014-12-12 16:54:24 -08003415 * destroyed and the i_mmap_rwsem is held.
Mel Gormand8333522012-07-31 16:46:20 -07003416 */
3417 vma->vm_flags &= ~VM_MAYSHARE;
3418}
3419
Chen, Kenneth W502717f2006-10-11 01:20:46 -07003420void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003421 unsigned long end, struct page *ref_page)
Chen, Kenneth W502717f2006-10-11 01:20:46 -07003422{
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07003423 struct mm_struct *mm;
3424 struct mmu_gather tlb;
3425
3426 mm = vma->vm_mm;
3427
Linus Torvalds2b047252013-08-15 11:42:25 -07003428 tlb_gather_mmu(&tlb, mm, start, end);
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07003429 __unmap_hugepage_range(&tlb, vma, start, end, ref_page);
3430 tlb_finish_mmu(&tlb, start, end);
Chen, Kenneth W502717f2006-10-11 01:20:46 -07003431}
3432
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003433/*
3434 * This is called when the original mapper is failing to COW a MAP_PRIVATE
3435 * mappping it owns the reserve page for. The intention is to unmap the page
3436 * from other VMAs and let the children be SIGKILLed if they are faulting the
3437 * same region.
3438 */
Davidlohr Bueso2f4612a2014-08-06 16:06:45 -07003439static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
3440 struct page *page, unsigned long address)
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003441{
Adam Litke75266742008-11-12 13:24:56 -08003442 struct hstate *h = hstate_vma(vma);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003443 struct vm_area_struct *iter_vma;
3444 struct address_space *mapping;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003445 pgoff_t pgoff;
3446
3447 /*
3448 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
3449 * from page cache lookup which is in HPAGE_SIZE units.
3450 */
Adam Litke75266742008-11-12 13:24:56 -08003451 address = address & huge_page_mask(h);
Michal Hocko36e4f202012-10-08 16:33:31 -07003452 pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) +
3453 vma->vm_pgoff;
Al Viro93c76a32015-12-04 23:45:44 -05003454 mapping = vma->vm_file->f_mapping;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003455
Mel Gorman4eb2b1d2009-12-14 17:59:53 -08003456 /*
3457 * Take the mapping lock for the duration of the table walk. As
3458 * this mapping should be shared between all the VMAs,
3459 * __unmap_hugepage_range() is called as the lock is already held
3460 */
Davidlohr Bueso83cde9e2014-12-12 16:54:21 -08003461 i_mmap_lock_write(mapping);
Michel Lespinasse6b2dbba2012-10-08 16:31:25 -07003462 vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) {
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003463 /* Do not unmap the current VMA */
3464 if (iter_vma == vma)
3465 continue;
3466
3467 /*
Mel Gorman2f84a892015-10-01 15:36:57 -07003468 * Shared VMAs have their own reserves and do not affect
3469 * MAP_PRIVATE accounting but it is possible that a shared
3470 * VMA is using the same page so check and skip such VMAs.
3471 */
3472 if (iter_vma->vm_flags & VM_MAYSHARE)
3473 continue;
3474
3475 /*
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003476 * Unmap the page from other VMAs without their own reserves.
3477 * They get marked to be SIGKILLed if they fault in these
3478 * areas. This is because a future no-page fault on this VMA
3479 * could insert a zeroed page instead of the data existing
3480 * from the time of fork. This would look like data corruption
3481 */
3482 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07003483 unmap_hugepage_range(iter_vma, address,
3484 address + huge_page_size(h), page);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003485 }
Davidlohr Bueso83cde9e2014-12-12 16:54:21 -08003486 i_mmap_unlock_write(mapping);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003487}
3488
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09003489/*
3490 * Hugetlb_cow() should be called with page lock of the original hugepage held.
Michal Hockoef009b22012-01-10 15:07:21 -08003491 * Called with hugetlb_instantiation_mutex held and pte_page locked so we
3492 * cannot race with other handlers or page migration.
3493 * Keep the pte_same checks anyway to make transition from the mutex easier.
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09003494 */
David Gibson1e8f8892006-01-06 00:10:44 -08003495static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
Aneesh Kumar K.V66c67702016-12-12 16:41:56 -08003496 unsigned long address, pte_t *ptep,
3497 struct page *pagecache_page, spinlock_t *ptl)
David Gibson1e8f8892006-01-06 00:10:44 -08003498{
Aneesh Kumar K.V66c67702016-12-12 16:41:56 -08003499 pte_t pte;
Andi Kleena5516432008-07-23 21:27:41 -07003500 struct hstate *h = hstate_vma(vma);
David Gibson1e8f8892006-01-06 00:10:44 -08003501 struct page *old_page, *new_page;
Davidlohr Buesoad4404a2014-08-06 16:06:47 -07003502 int ret = 0, outside_reserve = 0;
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07003503 unsigned long mmun_start; /* For mmu_notifiers */
3504 unsigned long mmun_end; /* For mmu_notifiers */
David Gibson1e8f8892006-01-06 00:10:44 -08003505
Aneesh Kumar K.V66c67702016-12-12 16:41:56 -08003506 pte = huge_ptep_get(ptep);
David Gibson1e8f8892006-01-06 00:10:44 -08003507 old_page = pte_page(pte);
3508
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003509retry_avoidcopy:
David Gibson1e8f8892006-01-06 00:10:44 -08003510 /* If no-one else is actually using this page, avoid the copy
3511 * and just make the page writable */
Joonsoo Kim37a21402013-09-11 14:21:04 -07003512 if (page_mapcount(old_page) == 1 && PageAnon(old_page)) {
Hugh Dickins5a499732016-07-14 12:07:38 -07003513 page_move_anon_rmap(old_page, vma);
David Gibson1e8f8892006-01-06 00:10:44 -08003514 set_huge_ptep_writable(vma, address, ptep);
Nick Piggin83c54072007-07-19 01:47:05 -07003515 return 0;
David Gibson1e8f8892006-01-06 00:10:44 -08003516 }
3517
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003518 /*
3519 * If the process that created a MAP_PRIVATE mapping is about to
3520 * perform a COW due to a shared page count, attempt to satisfy
3521 * the allocation without using the existing reserves. The pagecache
3522 * page is used to determine if the reserve at this address was
3523 * consumed or not. If reserves were used, a partial faulted mapping
3524 * at the time of fork() could consume its reserves on COW instead
3525 * of the full address range.
3526 */
Joonsoo Kim5944d012013-09-11 14:21:55 -07003527 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003528 old_page != pagecache_page)
3529 outside_reserve = 1;
3530
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03003531 get_page(old_page);
Larry Woodmanb76c8cf2009-12-14 17:59:37 -08003532
Davidlohr Buesoad4404a2014-08-06 16:06:47 -07003533 /*
3534 * Drop page table lock as buddy allocator may be called. It will
3535 * be acquired again before returning to the caller, as expected.
3536 */
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003537 spin_unlock(ptl);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003538 new_page = alloc_huge_page(vma, address, outside_reserve);
David Gibson1e8f8892006-01-06 00:10:44 -08003539
Adam Litke2fc39ce2007-11-14 16:59:39 -08003540 if (IS_ERR(new_page)) {
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003541 /*
3542 * If a process owning a MAP_PRIVATE mapping fails to COW,
3543 * it is due to references held by a child and an insufficient
3544 * huge page pool. To guarantee the original mappers
3545 * reliability, unmap the page from child processes. The child
3546 * may get SIGKILLed if it later faults.
3547 */
3548 if (outside_reserve) {
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03003549 put_page(old_page);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003550 BUG_ON(huge_pte_none(pte));
Davidlohr Bueso2f4612a2014-08-06 16:06:45 -07003551 unmap_ref_private(mm, vma, old_page, address);
3552 BUG_ON(huge_pte_none(pte));
3553 spin_lock(ptl);
3554 ptep = huge_pte_offset(mm, address & huge_page_mask(h));
3555 if (likely(ptep &&
3556 pte_same(huge_ptep_get(ptep), pte)))
3557 goto retry_avoidcopy;
3558 /*
3559 * race occurs while re-acquiring page table
3560 * lock, and our job is done.
3561 */
3562 return 0;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003563 }
3564
Davidlohr Buesoad4404a2014-08-06 16:06:47 -07003565 ret = (PTR_ERR(new_page) == -ENOMEM) ?
3566 VM_FAULT_OOM : VM_FAULT_SIGBUS;
3567 goto out_release_old;
David Gibson1e8f8892006-01-06 00:10:44 -08003568 }
3569
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09003570 /*
3571 * When the original hugepage is shared one, it does not have
3572 * anon_vma prepared.
3573 */
Dean Nelson44e2aa92010-10-26 14:22:08 -07003574 if (unlikely(anon_vma_prepare(vma))) {
Davidlohr Buesoad4404a2014-08-06 16:06:47 -07003575 ret = VM_FAULT_OOM;
3576 goto out_release_all;
Dean Nelson44e2aa92010-10-26 14:22:08 -07003577 }
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09003578
Andrea Arcangeli47ad8472011-01-13 15:46:47 -08003579 copy_user_huge_page(new_page, old_page, address, vma,
3580 pages_per_huge_page(h));
Nick Piggin0ed361d2008-02-04 22:29:34 -08003581 __SetPageUptodate(new_page);
Naoya Horiguchibcc54222015-04-15 16:14:38 -07003582 set_page_huge_active(new_page);
David Gibson1e8f8892006-01-06 00:10:44 -08003583
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07003584 mmun_start = address & huge_page_mask(h);
3585 mmun_end = mmun_start + huge_page_size(h);
3586 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
Davidlohr Buesoad4404a2014-08-06 16:06:47 -07003587
Larry Woodmanb76c8cf2009-12-14 17:59:37 -08003588 /*
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003589 * Retake the page table lock to check for racing updates
Larry Woodmanb76c8cf2009-12-14 17:59:37 -08003590 * before the page tables are altered
3591 */
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003592 spin_lock(ptl);
Andi Kleena5516432008-07-23 21:27:41 -07003593 ptep = huge_pte_offset(mm, address & huge_page_mask(h));
Naoya Horiguchia9af0c52014-04-07 15:36:54 -07003594 if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) {
Joonsoo Kim07443a82013-09-11 14:21:58 -07003595 ClearPagePrivate(new_page);
3596
David Gibson1e8f8892006-01-06 00:10:44 -08003597 /* Break COW */
Gerald Schaefer8fe627e2008-04-28 02:13:28 -07003598 huge_ptep_clear_flush(vma, address, ptep);
Joerg Roedel34ee6452014-11-13 13:46:09 +11003599 mmu_notifier_invalidate_range(mm, mmun_start, mmun_end);
David Gibson1e8f8892006-01-06 00:10:44 -08003600 set_huge_pte_at(mm, address, ptep,
3601 make_huge_pte(vma, new_page, 1));
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -08003602 page_remove_rmap(old_page, true);
Naoya Horiguchicd67f0d2010-09-10 13:23:04 +09003603 hugepage_add_new_anon_rmap(new_page, vma, address);
David Gibson1e8f8892006-01-06 00:10:44 -08003604 /* Make the old page be freed below */
3605 new_page = old_page;
3606 }
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003607 spin_unlock(ptl);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07003608 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
Davidlohr Buesoad4404a2014-08-06 16:06:47 -07003609out_release_all:
Mike Kravetz96b96a92016-11-10 10:46:32 -08003610 restore_reserve_on_error(h, vma, address, new_page);
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03003611 put_page(new_page);
Davidlohr Buesoad4404a2014-08-06 16:06:47 -07003612out_release_old:
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03003613 put_page(old_page);
Joonsoo Kim83120342013-09-11 14:21:57 -07003614
Davidlohr Buesoad4404a2014-08-06 16:06:47 -07003615 spin_lock(ptl); /* Caller expects lock to be held */
3616 return ret;
David Gibson1e8f8892006-01-06 00:10:44 -08003617}
3618
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003619/* Return the pagecache page at a given address within a VMA */
Andi Kleena5516432008-07-23 21:27:41 -07003620static struct page *hugetlbfs_pagecache_page(struct hstate *h,
3621 struct vm_area_struct *vma, unsigned long address)
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003622{
3623 struct address_space *mapping;
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -07003624 pgoff_t idx;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003625
3626 mapping = vma->vm_file->f_mapping;
Andi Kleena5516432008-07-23 21:27:41 -07003627 idx = vma_hugecache_offset(h, vma, address);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003628
3629 return find_lock_page(mapping, idx);
3630}
3631
Hugh Dickins3ae77f42009-09-21 17:03:33 -07003632/*
3633 * Return whether there is a pagecache page to back given address within VMA.
3634 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
3635 */
3636static bool hugetlbfs_pagecache_present(struct hstate *h,
Hugh Dickins2a15efc2009-09-21 17:03:27 -07003637 struct vm_area_struct *vma, unsigned long address)
3638{
3639 struct address_space *mapping;
3640 pgoff_t idx;
3641 struct page *page;
3642
3643 mapping = vma->vm_file->f_mapping;
3644 idx = vma_hugecache_offset(h, vma, address);
3645
3646 page = find_get_page(mapping, idx);
3647 if (page)
3648 put_page(page);
3649 return page != NULL;
3650}
3651
Mike Kravetzab76ad52015-09-08 15:01:50 -07003652int huge_add_to_page_cache(struct page *page, struct address_space *mapping,
3653 pgoff_t idx)
3654{
3655 struct inode *inode = mapping->host;
3656 struct hstate *h = hstate_inode(inode);
3657 int err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
3658
3659 if (err)
3660 return err;
3661 ClearPagePrivate(page);
3662
Mike Kravetzcbf05aa2018-10-26 15:10:58 -07003663 /*
3664 * set page dirty so that it will not be removed from cache/file
3665 * by non-hugetlbfs specific code paths.
3666 */
3667 set_page_dirty(page);
3668
Mike Kravetzab76ad52015-09-08 15:01:50 -07003669 spin_lock(&inode->i_lock);
3670 inode->i_blocks += blocks_per_huge_page(h);
3671 spin_unlock(&inode->i_lock);
3672 return 0;
3673}
3674
Robert P. J. Daya1ed3dd2007-07-17 04:03:33 -07003675static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
Davidlohr Bueso8382d912014-04-03 14:47:31 -07003676 struct address_space *mapping, pgoff_t idx,
3677 unsigned long address, pte_t *ptep, unsigned int flags)
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01003678{
Andi Kleena5516432008-07-23 21:27:41 -07003679 struct hstate *h = hstate_vma(vma);
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01003680 int ret = VM_FAULT_SIGBUS;
Hillf Danton409eb8c2012-01-20 14:34:13 -08003681 int anon_rmap = 0;
Adam Litke4c887262005-10-29 18:16:46 -07003682 unsigned long size;
Adam Litke4c887262005-10-29 18:16:46 -07003683 struct page *page;
David Gibson1e8f8892006-01-06 00:10:44 -08003684 pte_t new_pte;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003685 spinlock_t *ptl;
Adam Litke4c887262005-10-29 18:16:46 -07003686
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003687 /*
3688 * Currently, we are forced to kill the process in the event the
3689 * original mapper has unmapped pages from the child due to a failed
Lucas De Marchi25985ed2011-03-30 22:57:33 -03003690 * COW. Warn that such a situation has occurred as it may not be obvious
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003691 */
3692 if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
Geoffrey Thomas910154d2016-03-09 14:08:04 -08003693 pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n",
Andrew Mortonffb22af2013-02-22 16:32:08 -08003694 current->pid);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003695 return ret;
3696 }
3697
Adam Litke4c887262005-10-29 18:16:46 -07003698 /*
3699 * Use page lock to guard against racing truncation
3700 * before we get page_table_lock.
3701 */
Christoph Lameter6bda6662006-01-06 00:10:49 -08003702retry:
3703 page = find_lock_page(mapping, idx);
3704 if (!page) {
Andi Kleena5516432008-07-23 21:27:41 -07003705 size = i_size_read(mapping->host) >> huge_page_shift(h);
Hugh Dickinsebed4bf2006-10-28 10:38:43 -07003706 if (idx >= size)
3707 goto out;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003708 page = alloc_huge_page(vma, address, 0);
Adam Litke2fc39ce2007-11-14 16:59:39 -08003709 if (IS_ERR(page)) {
Aneesh Kumar K.V76dcee72012-07-31 16:41:57 -07003710 ret = PTR_ERR(page);
3711 if (ret == -ENOMEM)
3712 ret = VM_FAULT_OOM;
3713 else
3714 ret = VM_FAULT_SIGBUS;
Christoph Lameter6bda6662006-01-06 00:10:49 -08003715 goto out;
3716 }
Andrea Arcangeli47ad8472011-01-13 15:46:47 -08003717 clear_huge_page(page, address, pages_per_huge_page(h));
Nick Piggin0ed361d2008-02-04 22:29:34 -08003718 __SetPageUptodate(page);
Naoya Horiguchibcc54222015-04-15 16:14:38 -07003719 set_page_huge_active(page);
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01003720
Mel Gormanf83a2752009-05-28 14:34:40 -07003721 if (vma->vm_flags & VM_MAYSHARE) {
Mike Kravetzab76ad52015-09-08 15:01:50 -07003722 int err = huge_add_to_page_cache(page, mapping, idx);
Christoph Lameter6bda6662006-01-06 00:10:49 -08003723 if (err) {
3724 put_page(page);
Christoph Lameter6bda6662006-01-06 00:10:49 -08003725 if (err == -EEXIST)
3726 goto retry;
3727 goto out;
3728 }
Mel Gorman23be7462010-04-23 13:17:56 -04003729 } else {
Christoph Lameter6bda6662006-01-06 00:10:49 -08003730 lock_page(page);
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09003731 if (unlikely(anon_vma_prepare(vma))) {
3732 ret = VM_FAULT_OOM;
3733 goto backout_unlocked;
3734 }
Hillf Danton409eb8c2012-01-20 14:34:13 -08003735 anon_rmap = 1;
Mel Gorman23be7462010-04-23 13:17:56 -04003736 }
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09003737 } else {
Naoya Horiguchi998b4382010-09-08 10:19:32 +09003738 /*
3739 * If memory error occurs between mmap() and fault, some process
3740 * don't have hwpoisoned swap entry for errored virtual address.
3741 * So we need to block hugepage fault by PG_hwpoison bit check.
3742 */
3743 if (unlikely(PageHWPoison(page))) {
Chris Forbes32f84522011-07-25 17:12:14 -07003744 ret = VM_FAULT_HWPOISON |
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07003745 VM_FAULT_SET_HINDEX(hstate_index(h));
Naoya Horiguchi998b4382010-09-08 10:19:32 +09003746 goto backout_unlocked;
3747 }
Christoph Lameter6bda6662006-01-06 00:10:49 -08003748 }
David Gibson1e8f8892006-01-06 00:10:44 -08003749
Andy Whitcroft57303d82008-08-12 15:08:47 -07003750 /*
3751 * If we are going to COW a private mapping later, we examine the
3752 * pending reservations for this page now. This will ensure that
3753 * any allocations necessary to record that reservation occur outside
3754 * the spinlock.
3755 */
Mike Kravetz5e911372015-09-08 15:01:28 -07003756 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
Andy Whitcroft2b267362008-08-12 15:08:49 -07003757 if (vma_needs_reservation(h, vma, address) < 0) {
3758 ret = VM_FAULT_OOM;
3759 goto backout_unlocked;
3760 }
Mike Kravetz5e911372015-09-08 15:01:28 -07003761 /* Just decrements count, does not deallocate */
Mike Kravetzfeba16e2015-09-08 15:01:31 -07003762 vma_end_reservation(h, vma, address);
Mike Kravetz5e911372015-09-08 15:01:28 -07003763 }
Andy Whitcroft57303d82008-08-12 15:08:47 -07003764
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003765 ptl = huge_pte_lockptr(h, mm, ptep);
3766 spin_lock(ptl);
Andi Kleena5516432008-07-23 21:27:41 -07003767 size = i_size_read(mapping->host) >> huge_page_shift(h);
Adam Litke4c887262005-10-29 18:16:46 -07003768 if (idx >= size)
3769 goto backout;
3770
Nick Piggin83c54072007-07-19 01:47:05 -07003771 ret = 0;
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07003772 if (!huge_pte_none(huge_ptep_get(ptep)))
Adam Litke4c887262005-10-29 18:16:46 -07003773 goto backout;
3774
Joonsoo Kim07443a82013-09-11 14:21:58 -07003775 if (anon_rmap) {
3776 ClearPagePrivate(page);
Hillf Danton409eb8c2012-01-20 14:34:13 -08003777 hugepage_add_new_anon_rmap(page, vma, address);
Choi Gi-yongac714902014-04-07 15:37:36 -07003778 } else
Kirill A. Shutemov53f92632016-01-15 16:53:42 -08003779 page_dup_rmap(page, true);
David Gibson1e8f8892006-01-06 00:10:44 -08003780 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
3781 && (vma->vm_flags & VM_SHARED)));
3782 set_huge_pte_at(mm, address, ptep, new_pte);
3783
Naoya Horiguchi5d317b22015-11-05 18:47:14 -08003784 hugetlb_count_add(pages_per_huge_page(h), mm);
Hugh Dickins788c7df2009-06-23 13:49:05 +01003785 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
David Gibson1e8f8892006-01-06 00:10:44 -08003786 /* Optimization, do the COW without a second fault */
Aneesh Kumar K.V66c67702016-12-12 16:41:56 -08003787 ret = hugetlb_cow(mm, vma, address, ptep, page, ptl);
David Gibson1e8f8892006-01-06 00:10:44 -08003788 }
3789
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003790 spin_unlock(ptl);
Adam Litke4c887262005-10-29 18:16:46 -07003791 unlock_page(page);
3792out:
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01003793 return ret;
Adam Litke4c887262005-10-29 18:16:46 -07003794
3795backout:
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003796 spin_unlock(ptl);
Andy Whitcroft2b267362008-08-12 15:08:49 -07003797backout_unlocked:
Adam Litke4c887262005-10-29 18:16:46 -07003798 unlock_page(page);
Mike Kravetz96b96a92016-11-10 10:46:32 -08003799 restore_reserve_on_error(h, vma, address, page);
Adam Litke4c887262005-10-29 18:16:46 -07003800 put_page(page);
3801 goto out;
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01003802}
3803
Davidlohr Bueso8382d912014-04-03 14:47:31 -07003804#ifdef CONFIG_SMP
Mike Kravetzc672c7f2015-09-08 15:01:35 -07003805u32 hugetlb_fault_mutex_hash(struct hstate *h, struct mm_struct *mm,
Davidlohr Bueso8382d912014-04-03 14:47:31 -07003806 struct vm_area_struct *vma,
3807 struct address_space *mapping,
3808 pgoff_t idx, unsigned long address)
3809{
3810 unsigned long key[2];
3811 u32 hash;
3812
3813 if (vma->vm_flags & VM_SHARED) {
3814 key[0] = (unsigned long) mapping;
3815 key[1] = idx;
3816 } else {
3817 key[0] = (unsigned long) mm;
3818 key[1] = address >> huge_page_shift(h);
3819 }
3820
3821 hash = jhash2((u32 *)&key, sizeof(key)/sizeof(u32), 0);
3822
3823 return hash & (num_fault_mutexes - 1);
3824}
3825#else
3826/*
3827 * For uniprocesor systems we always use a single mutex, so just
3828 * return 0 and avoid the hashing overhead.
3829 */
Mike Kravetzc672c7f2015-09-08 15:01:35 -07003830u32 hugetlb_fault_mutex_hash(struct hstate *h, struct mm_struct *mm,
Davidlohr Bueso8382d912014-04-03 14:47:31 -07003831 struct vm_area_struct *vma,
3832 struct address_space *mapping,
3833 pgoff_t idx, unsigned long address)
3834{
3835 return 0;
3836}
3837#endif
3838
Adam Litke86e52162006-01-06 00:10:43 -08003839int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
Hugh Dickins788c7df2009-06-23 13:49:05 +01003840 unsigned long address, unsigned int flags)
Adam Litke86e52162006-01-06 00:10:43 -08003841{
Davidlohr Bueso8382d912014-04-03 14:47:31 -07003842 pte_t *ptep, entry;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003843 spinlock_t *ptl;
David Gibson1e8f8892006-01-06 00:10:44 -08003844 int ret;
Davidlohr Bueso8382d912014-04-03 14:47:31 -07003845 u32 hash;
3846 pgoff_t idx;
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09003847 struct page *page = NULL;
Andy Whitcroft57303d82008-08-12 15:08:47 -07003848 struct page *pagecache_page = NULL;
Andi Kleena5516432008-07-23 21:27:41 -07003849 struct hstate *h = hstate_vma(vma);
Davidlohr Bueso8382d912014-04-03 14:47:31 -07003850 struct address_space *mapping;
Naoya Horiguchi0f792cf2015-02-11 15:25:25 -08003851 int need_wait_lock = 0;
Adam Litke86e52162006-01-06 00:10:43 -08003852
KAMEZAWA Hiroyuki1e16a532012-01-10 15:07:22 -08003853 address &= huge_page_mask(h);
3854
Naoya Horiguchifd6a03ed2010-05-28 09:29:21 +09003855 ptep = huge_pte_offset(mm, address);
3856 if (ptep) {
3857 entry = huge_ptep_get(ptep);
Naoya Horiguchi290408d2010-09-08 10:19:35 +09003858 if (unlikely(is_hugetlb_entry_migration(entry))) {
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003859 migration_entry_wait_huge(vma, mm, ptep);
Naoya Horiguchi290408d2010-09-08 10:19:35 +09003860 return 0;
3861 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
Chris Forbes32f84522011-07-25 17:12:14 -07003862 return VM_FAULT_HWPOISON_LARGE |
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07003863 VM_FAULT_SET_HINDEX(hstate_index(h));
Naoya Horiguchi0d777df2015-12-11 13:40:49 -08003864 } else {
3865 ptep = huge_pte_alloc(mm, address, huge_page_size(h));
3866 if (!ptep)
3867 return VM_FAULT_OOM;
Naoya Horiguchifd6a03ed2010-05-28 09:29:21 +09003868 }
3869
Davidlohr Bueso8382d912014-04-03 14:47:31 -07003870 mapping = vma->vm_file->f_mapping;
3871 idx = vma_hugecache_offset(h, vma, address);
3872
David Gibson3935baa2006-03-22 00:08:53 -08003873 /*
3874 * Serialize hugepage allocation and instantiation, so that we don't
3875 * get spurious allocation failures if two CPUs race to instantiate
3876 * the same page in the page cache.
3877 */
Mike Kravetzc672c7f2015-09-08 15:01:35 -07003878 hash = hugetlb_fault_mutex_hash(h, mm, vma, mapping, idx, address);
3879 mutex_lock(&hugetlb_fault_mutex_table[hash]);
Davidlohr Bueso8382d912014-04-03 14:47:31 -07003880
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07003881 entry = huge_ptep_get(ptep);
3882 if (huge_pte_none(entry)) {
Davidlohr Bueso8382d912014-04-03 14:47:31 -07003883 ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags);
David Gibsonb4d1d992008-10-15 22:01:11 -07003884 goto out_mutex;
David Gibson3935baa2006-03-22 00:08:53 -08003885 }
Adam Litke86e52162006-01-06 00:10:43 -08003886
Nick Piggin83c54072007-07-19 01:47:05 -07003887 ret = 0;
David Gibson1e8f8892006-01-06 00:10:44 -08003888
Andy Whitcroft57303d82008-08-12 15:08:47 -07003889 /*
Naoya Horiguchi0f792cf2015-02-11 15:25:25 -08003890 * entry could be a migration/hwpoison entry at this point, so this
3891 * check prevents the kernel from going below assuming that we have
3892 * a active hugepage in pagecache. This goto expects the 2nd page fault,
3893 * and is_hugetlb_entry_(migration|hwpoisoned) check will properly
3894 * handle it.
3895 */
3896 if (!pte_present(entry))
3897 goto out_mutex;
3898
3899 /*
Andy Whitcroft57303d82008-08-12 15:08:47 -07003900 * If we are going to COW the mapping later, we examine the pending
3901 * reservations for this page now. This will ensure that any
3902 * allocations necessary to record that reservation occur outside the
3903 * spinlock. For private mappings, we also lookup the pagecache
3904 * page now as it is used to determine if a reservation has been
3905 * consumed.
3906 */
Gerald Schaefer106c9922013-04-29 15:07:23 -07003907 if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) {
Andy Whitcroft2b267362008-08-12 15:08:49 -07003908 if (vma_needs_reservation(h, vma, address) < 0) {
3909 ret = VM_FAULT_OOM;
David Gibsonb4d1d992008-10-15 22:01:11 -07003910 goto out_mutex;
Andy Whitcroft2b267362008-08-12 15:08:49 -07003911 }
Mike Kravetz5e911372015-09-08 15:01:28 -07003912 /* Just decrements count, does not deallocate */
Mike Kravetzfeba16e2015-09-08 15:01:31 -07003913 vma_end_reservation(h, vma, address);
Andy Whitcroft57303d82008-08-12 15:08:47 -07003914
Mel Gormanf83a2752009-05-28 14:34:40 -07003915 if (!(vma->vm_flags & VM_MAYSHARE))
Andy Whitcroft57303d82008-08-12 15:08:47 -07003916 pagecache_page = hugetlbfs_pagecache_page(h,
3917 vma, address);
3918 }
3919
Naoya Horiguchi0f792cf2015-02-11 15:25:25 -08003920 ptl = huge_pte_lock(h, mm, ptep);
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09003921
David Gibson1e8f8892006-01-06 00:10:44 -08003922 /* Check for a racing update before calling hugetlb_cow */
David Gibsonb4d1d992008-10-15 22:01:11 -07003923 if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003924 goto out_ptl;
David Gibsonb4d1d992008-10-15 22:01:11 -07003925
Naoya Horiguchi0f792cf2015-02-11 15:25:25 -08003926 /*
3927 * hugetlb_cow() requires page locks of pte_page(entry) and
3928 * pagecache_page, so here we need take the former one
3929 * when page != pagecache_page or !pagecache_page.
3930 */
3931 page = pte_page(entry);
3932 if (page != pagecache_page)
3933 if (!trylock_page(page)) {
3934 need_wait_lock = 1;
3935 goto out_ptl;
3936 }
3937
3938 get_page(page);
David Gibsonb4d1d992008-10-15 22:01:11 -07003939
Hugh Dickins788c7df2009-06-23 13:49:05 +01003940 if (flags & FAULT_FLAG_WRITE) {
Gerald Schaefer106c9922013-04-29 15:07:23 -07003941 if (!huge_pte_write(entry)) {
Aneesh Kumar K.V66c67702016-12-12 16:41:56 -08003942 ret = hugetlb_cow(mm, vma, address, ptep,
3943 pagecache_page, ptl);
Naoya Horiguchi0f792cf2015-02-11 15:25:25 -08003944 goto out_put_page;
David Gibsonb4d1d992008-10-15 22:01:11 -07003945 }
Gerald Schaefer106c9922013-04-29 15:07:23 -07003946 entry = huge_pte_mkdirty(entry);
David Gibsonb4d1d992008-10-15 22:01:11 -07003947 }
3948 entry = pte_mkyoung(entry);
Hugh Dickins788c7df2009-06-23 13:49:05 +01003949 if (huge_ptep_set_access_flags(vma, address, ptep, entry,
3950 flags & FAULT_FLAG_WRITE))
Russell King4b3073e2009-12-18 16:40:18 +00003951 update_mmu_cache(vma, address, ptep);
Naoya Horiguchi0f792cf2015-02-11 15:25:25 -08003952out_put_page:
3953 if (page != pagecache_page)
3954 unlock_page(page);
3955 put_page(page);
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003956out_ptl:
3957 spin_unlock(ptl);
Andy Whitcroft57303d82008-08-12 15:08:47 -07003958
3959 if (pagecache_page) {
3960 unlock_page(pagecache_page);
3961 put_page(pagecache_page);
3962 }
David Gibsonb4d1d992008-10-15 22:01:11 -07003963out_mutex:
Mike Kravetzc672c7f2015-09-08 15:01:35 -07003964 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
Naoya Horiguchi0f792cf2015-02-11 15:25:25 -08003965 /*
3966 * Generally it's safe to hold refcount during waiting page lock. But
3967 * here we just wait to defer the next page fault to avoid busy loop and
3968 * the page is not used after unlocked before returning from the current
3969 * page fault. So we are safe from accessing freed page, even if we wait
3970 * here without taking refcount.
3971 */
3972 if (need_wait_lock)
3973 wait_on_page_locked(page);
David Gibson1e8f8892006-01-06 00:10:44 -08003974 return ret;
Adam Litke86e52162006-01-06 00:10:43 -08003975}
3976
Michel Lespinasse28a35712013-02-22 16:35:55 -08003977long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
3978 struct page **pages, struct vm_area_struct **vmas,
3979 unsigned long *position, unsigned long *nr_pages,
3980 long i, unsigned int flags)
David Gibson63551ae2005-06-21 17:14:44 -07003981{
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08003982 unsigned long pfn_offset;
3983 unsigned long vaddr = *position;
Michel Lespinasse28a35712013-02-22 16:35:55 -08003984 unsigned long remainder = *nr_pages;
Andi Kleena5516432008-07-23 21:27:41 -07003985 struct hstate *h = hstate_vma(vma);
David Gibson63551ae2005-06-21 17:14:44 -07003986
David Gibson63551ae2005-06-21 17:14:44 -07003987 while (vaddr < vma->vm_end && remainder) {
Adam Litke4c887262005-10-29 18:16:46 -07003988 pte_t *pte;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003989 spinlock_t *ptl = NULL;
Hugh Dickins2a15efc2009-09-21 17:03:27 -07003990 int absent;
Adam Litke4c887262005-10-29 18:16:46 -07003991 struct page *page;
3992
3993 /*
David Rientjes02057962015-04-14 15:48:24 -07003994 * If we have a pending SIGKILL, don't keep faulting pages and
3995 * potentially allocating memory.
3996 */
3997 if (unlikely(fatal_signal_pending(current))) {
3998 remainder = 0;
3999 break;
4000 }
4001
4002 /*
Adam Litke4c887262005-10-29 18:16:46 -07004003 * Some archs (sparc64, sh*) have multiple pte_ts to
Hugh Dickins2a15efc2009-09-21 17:03:27 -07004004 * each hugepage. We have to make sure we get the
Adam Litke4c887262005-10-29 18:16:46 -07004005 * first, for the page indexing below to work.
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08004006 *
4007 * Note that page table lock is not held when pte is null.
Adam Litke4c887262005-10-29 18:16:46 -07004008 */
Andi Kleena5516432008-07-23 21:27:41 -07004009 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h));
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08004010 if (pte)
4011 ptl = huge_pte_lock(h, mm, pte);
Hugh Dickins2a15efc2009-09-21 17:03:27 -07004012 absent = !pte || huge_pte_none(huge_ptep_get(pte));
Adam Litke4c887262005-10-29 18:16:46 -07004013
Hugh Dickins2a15efc2009-09-21 17:03:27 -07004014 /*
4015 * When coredumping, it suits get_dump_page if we just return
Hugh Dickins3ae77f42009-09-21 17:03:33 -07004016 * an error where there's an empty slot with no huge pagecache
4017 * to back it. This way, we avoid allocating a hugepage, and
4018 * the sparse dumpfile avoids allocating disk blocks, but its
4019 * huge holes still show up with zeroes where they need to be.
Hugh Dickins2a15efc2009-09-21 17:03:27 -07004020 */
Hugh Dickins3ae77f42009-09-21 17:03:33 -07004021 if (absent && (flags & FOLL_DUMP) &&
4022 !hugetlbfs_pagecache_present(h, vma, vaddr)) {
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08004023 if (pte)
4024 spin_unlock(ptl);
Hugh Dickins2a15efc2009-09-21 17:03:27 -07004025 remainder = 0;
4026 break;
4027 }
4028
Naoya Horiguchi9cc3a5b2013-04-17 15:58:30 -07004029 /*
4030 * We need call hugetlb_fault for both hugepages under migration
4031 * (in which case hugetlb_fault waits for the migration,) and
4032 * hwpoisoned hugepages (in which case we need to prevent the
4033 * caller from accessing to them.) In order to do this, we use
4034 * here is_swap_pte instead of is_hugetlb_entry_migration and
4035 * is_hugetlb_entry_hwpoisoned. This is because it simply covers
4036 * both cases, and because we can't follow correct pages
4037 * directly from any kind of swap entries.
4038 */
4039 if (absent || is_swap_pte(huge_ptep_get(pte)) ||
Gerald Schaefer106c9922013-04-29 15:07:23 -07004040 ((flags & FOLL_WRITE) &&
4041 !huge_pte_write(huge_ptep_get(pte)))) {
Adam Litke4c887262005-10-29 18:16:46 -07004042 int ret;
4043
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08004044 if (pte)
4045 spin_unlock(ptl);
Hugh Dickins2a15efc2009-09-21 17:03:27 -07004046 ret = hugetlb_fault(mm, vma, vaddr,
4047 (flags & FOLL_WRITE) ? FAULT_FLAG_WRITE : 0);
Adam Litkea89182c2007-08-22 14:01:51 -07004048 if (!(ret & VM_FAULT_ERROR))
Adam Litke4c887262005-10-29 18:16:46 -07004049 continue;
4050
4051 remainder = 0;
Adam Litke4c887262005-10-29 18:16:46 -07004052 break;
4053 }
David Gibson63551ae2005-06-21 17:14:44 -07004054
Andi Kleena5516432008-07-23 21:27:41 -07004055 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07004056 page = pte_page(huge_ptep_get(pte));
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08004057same_page:
Chen, Kenneth Wd6692182006-03-31 02:29:57 -08004058 if (pages) {
Hugh Dickins2a15efc2009-09-21 17:03:27 -07004059 pages[i] = mem_map_offset(page, pfn_offset);
Kirill A. Shutemovddc58f22016-01-15 16:52:56 -08004060 get_page(pages[i]);
Chen, Kenneth Wd6692182006-03-31 02:29:57 -08004061 }
David Gibson63551ae2005-06-21 17:14:44 -07004062
4063 if (vmas)
4064 vmas[i] = vma;
4065
4066 vaddr += PAGE_SIZE;
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08004067 ++pfn_offset;
David Gibson63551ae2005-06-21 17:14:44 -07004068 --remainder;
4069 ++i;
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08004070 if (vaddr < vma->vm_end && remainder &&
Andi Kleena5516432008-07-23 21:27:41 -07004071 pfn_offset < pages_per_huge_page(h)) {
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08004072 /*
4073 * We use pfn_offset to avoid touching the pageframes
4074 * of this compound page.
4075 */
4076 goto same_page;
4077 }
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08004078 spin_unlock(ptl);
David Gibson63551ae2005-06-21 17:14:44 -07004079 }
Michel Lespinasse28a35712013-02-22 16:35:55 -08004080 *nr_pages = remainder;
David Gibson63551ae2005-06-21 17:14:44 -07004081 *position = vaddr;
4082
Hugh Dickins2a15efc2009-09-21 17:03:27 -07004083 return i ? i : -EFAULT;
David Gibson63551ae2005-06-21 17:14:44 -07004084}
Zhang, Yanmin8f860592006-03-22 00:08:50 -08004085
Aneesh Kumar K.V5491ae72016-07-13 15:06:43 +05304086#ifndef __HAVE_ARCH_FLUSH_HUGETLB_TLB_RANGE
4087/*
4088 * ARCHes with special requirements for evicting HUGETLB backing TLB entries can
4089 * implement this.
4090 */
4091#define flush_hugetlb_tlb_range(vma, addr, end) flush_tlb_range(vma, addr, end)
4092#endif
4093
Peter Zijlstra7da4d642012-11-19 03:14:23 +01004094unsigned long hugetlb_change_protection(struct vm_area_struct *vma,
Zhang, Yanmin8f860592006-03-22 00:08:50 -08004095 unsigned long address, unsigned long end, pgprot_t newprot)
4096{
4097 struct mm_struct *mm = vma->vm_mm;
4098 unsigned long start = address;
4099 pte_t *ptep;
4100 pte_t pte;
Andi Kleena5516432008-07-23 21:27:41 -07004101 struct hstate *h = hstate_vma(vma);
Peter Zijlstra7da4d642012-11-19 03:14:23 +01004102 unsigned long pages = 0;
Zhang, Yanmin8f860592006-03-22 00:08:50 -08004103
4104 BUG_ON(address >= end);
4105 flush_cache_range(vma, address, end);
4106
Rik van Riela5338092014-04-07 15:36:57 -07004107 mmu_notifier_invalidate_range_start(mm, start, end);
Davidlohr Bueso83cde9e2014-12-12 16:54:21 -08004108 i_mmap_lock_write(vma->vm_file->f_mapping);
Andi Kleena5516432008-07-23 21:27:41 -07004109 for (; address < end; address += huge_page_size(h)) {
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08004110 spinlock_t *ptl;
Zhang, Yanmin8f860592006-03-22 00:08:50 -08004111 ptep = huge_pte_offset(mm, address);
4112 if (!ptep)
4113 continue;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08004114 ptl = huge_pte_lock(h, mm, ptep);
Peter Zijlstra7da4d642012-11-19 03:14:23 +01004115 if (huge_pmd_unshare(mm, &address, ptep)) {
4116 pages++;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08004117 spin_unlock(ptl);
Chen, Kenneth W39dde652006-12-06 20:32:03 -08004118 continue;
Peter Zijlstra7da4d642012-11-19 03:14:23 +01004119 }
Naoya Horiguchia8bda282015-02-11 15:25:28 -08004120 pte = huge_ptep_get(ptep);
4121 if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) {
4122 spin_unlock(ptl);
4123 continue;
4124 }
4125 if (unlikely(is_hugetlb_entry_migration(pte))) {
4126 swp_entry_t entry = pte_to_swp_entry(pte);
4127
4128 if (is_write_migration_entry(entry)) {
4129 pte_t newpte;
4130
4131 make_migration_entry_read(&entry);
4132 newpte = swp_entry_to_pte(entry);
4133 set_huge_pte_at(mm, address, ptep, newpte);
4134 pages++;
4135 }
4136 spin_unlock(ptl);
4137 continue;
4138 }
4139 if (!huge_pte_none(pte)) {
Zhang, Yanmin8f860592006-03-22 00:08:50 -08004140 pte = huge_ptep_get_and_clear(mm, address, ptep);
Gerald Schaefer106c9922013-04-29 15:07:23 -07004141 pte = pte_mkhuge(huge_pte_modify(pte, newprot));
Tony Lube7517d2013-02-04 14:28:46 -08004142 pte = arch_make_huge_pte(pte, vma, NULL, 0);
Zhang, Yanmin8f860592006-03-22 00:08:50 -08004143 set_huge_pte_at(mm, address, ptep, pte);
Peter Zijlstra7da4d642012-11-19 03:14:23 +01004144 pages++;
Zhang, Yanmin8f860592006-03-22 00:08:50 -08004145 }
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08004146 spin_unlock(ptl);
Zhang, Yanmin8f860592006-03-22 00:08:50 -08004147 }
Mel Gormand8333522012-07-31 16:46:20 -07004148 /*
Davidlohr Buesoc8c06ef2014-12-12 16:54:24 -08004149 * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare
Mel Gormand8333522012-07-31 16:46:20 -07004150 * may have cleared our pud entry and done put_page on the page table:
Davidlohr Buesoc8c06ef2014-12-12 16:54:24 -08004151 * once we release i_mmap_rwsem, another task can do the final put_page
Mel Gormand8333522012-07-31 16:46:20 -07004152 * and that page table be reused and filled with junk.
4153 */
Aneesh Kumar K.V5491ae72016-07-13 15:06:43 +05304154 flush_hugetlb_tlb_range(vma, start, end);
Joerg Roedel34ee6452014-11-13 13:46:09 +11004155 mmu_notifier_invalidate_range(mm, start, end);
Davidlohr Bueso83cde9e2014-12-12 16:54:21 -08004156 i_mmap_unlock_write(vma->vm_file->f_mapping);
Rik van Riela5338092014-04-07 15:36:57 -07004157 mmu_notifier_invalidate_range_end(mm, start, end);
Peter Zijlstra7da4d642012-11-19 03:14:23 +01004158
4159 return pages << h->order;
Zhang, Yanmin8f860592006-03-22 00:08:50 -08004160}
4161
Mel Gormana1e78772008-07-23 21:27:23 -07004162int hugetlb_reserve_pages(struct inode *inode,
4163 long from, long to,
Mel Gorman5a6fe122009-02-10 14:02:27 +00004164 struct vm_area_struct *vma,
KOSAKI Motohiroca16d142011-05-26 19:16:19 +09004165 vm_flags_t vm_flags)
Adam Litkee4e574b2007-10-16 01:26:19 -07004166{
Mel Gorman17c9d122009-02-11 16:34:16 +00004167 long ret, chg;
Andi Kleena5516432008-07-23 21:27:41 -07004168 struct hstate *h = hstate_inode(inode);
David Gibson90481622012-03-21 16:34:12 -07004169 struct hugepage_subpool *spool = subpool_inode(inode);
Joonsoo Kim9119a412014-04-03 14:47:25 -07004170 struct resv_map *resv_map;
Mike Kravetz1c5ecae2015-04-15 16:13:39 -07004171 long gbl_reserve;
Adam Litkee4e574b2007-10-16 01:26:19 -07004172
Mel Gormana1e78772008-07-23 21:27:23 -07004173 /*
Mel Gorman17c9d122009-02-11 16:34:16 +00004174 * Only apply hugepage reservation if asked. At fault time, an
4175 * attempt will be made for VM_NORESERVE to allocate a page
David Gibson90481622012-03-21 16:34:12 -07004176 * without using reserves
Mel Gorman17c9d122009-02-11 16:34:16 +00004177 */
KOSAKI Motohiroca16d142011-05-26 19:16:19 +09004178 if (vm_flags & VM_NORESERVE)
Mel Gorman17c9d122009-02-11 16:34:16 +00004179 return 0;
4180
4181 /*
Mel Gormana1e78772008-07-23 21:27:23 -07004182 * Shared mappings base their reservation on the number of pages that
4183 * are already allocated on behalf of the file. Private mappings need
4184 * to reserve the full area even if read-only as mprotect() may be
4185 * called to make the mapping read-write. Assume !vma is a shm mapping
4186 */
Joonsoo Kim9119a412014-04-03 14:47:25 -07004187 if (!vma || vma->vm_flags & VM_MAYSHARE) {
Joonsoo Kim4e35f482014-04-03 14:47:30 -07004188 resv_map = inode_resv_map(inode);
Joonsoo Kim9119a412014-04-03 14:47:25 -07004189
Joonsoo Kim1406ec92014-04-03 14:47:26 -07004190 chg = region_chg(resv_map, from, to);
Joonsoo Kim9119a412014-04-03 14:47:25 -07004191
4192 } else {
4193 resv_map = resv_map_alloc();
Mel Gorman5a6fe122009-02-10 14:02:27 +00004194 if (!resv_map)
4195 return -ENOMEM;
4196
Mel Gorman17c9d122009-02-11 16:34:16 +00004197 chg = to - from;
4198
Mel Gorman5a6fe122009-02-10 14:02:27 +00004199 set_vma_resv_map(vma, resv_map);
4200 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
4201 }
4202
Dave Hansenc50ac052012-05-29 15:06:46 -07004203 if (chg < 0) {
4204 ret = chg;
4205 goto out_err;
4206 }
Mel Gorman17c9d122009-02-11 16:34:16 +00004207
Mike Kravetz1c5ecae2015-04-15 16:13:39 -07004208 /*
4209 * There must be enough pages in the subpool for the mapping. If
4210 * the subpool has a minimum size, there may be some global
4211 * reservations already in place (gbl_reserve).
4212 */
4213 gbl_reserve = hugepage_subpool_get_pages(spool, chg);
4214 if (gbl_reserve < 0) {
Dave Hansenc50ac052012-05-29 15:06:46 -07004215 ret = -ENOSPC;
4216 goto out_err;
4217 }
Mel Gorman17c9d122009-02-11 16:34:16 +00004218
4219 /*
4220 * Check enough hugepages are available for the reservation.
David Gibson90481622012-03-21 16:34:12 -07004221 * Hand the pages back to the subpool if there are not
Mel Gorman17c9d122009-02-11 16:34:16 +00004222 */
Mike Kravetz1c5ecae2015-04-15 16:13:39 -07004223 ret = hugetlb_acct_memory(h, gbl_reserve);
Mel Gorman17c9d122009-02-11 16:34:16 +00004224 if (ret < 0) {
Mike Kravetz1c5ecae2015-04-15 16:13:39 -07004225 /* put back original number of pages, chg */
4226 (void)hugepage_subpool_put_pages(spool, chg);
Dave Hansenc50ac052012-05-29 15:06:46 -07004227 goto out_err;
Mel Gorman17c9d122009-02-11 16:34:16 +00004228 }
4229
4230 /*
4231 * Account for the reservations made. Shared mappings record regions
4232 * that have reservations as they are shared by multiple VMAs.
4233 * When the last VMA disappears, the region map says how much
4234 * the reservation was and the page cache tells how much of
4235 * the reservation was consumed. Private mappings are per-VMA and
4236 * only the consumed reservations are tracked. When the VMA
4237 * disappears, the original reservation is the VMA size and the
4238 * consumed reservations are stored in the map. Hence, nothing
4239 * else has to be done for private mappings here
4240 */
Mike Kravetz33039672015-06-24 16:57:58 -07004241 if (!vma || vma->vm_flags & VM_MAYSHARE) {
4242 long add = region_add(resv_map, from, to);
4243
4244 if (unlikely(chg > add)) {
4245 /*
4246 * pages in this range were added to the reserve
4247 * map between region_chg and region_add. This
4248 * indicates a race with alloc_huge_page. Adjust
4249 * the subpool and reserve counts modified above
4250 * based on the difference.
4251 */
4252 long rsv_adjust;
4253
4254 rsv_adjust = hugepage_subpool_put_pages(spool,
4255 chg - add);
4256 hugetlb_acct_memory(h, -rsv_adjust);
4257 }
4258 }
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -07004259 return 0;
Dave Hansenc50ac052012-05-29 15:06:46 -07004260out_err:
Mike Kravetz5e911372015-09-08 15:01:28 -07004261 if (!vma || vma->vm_flags & VM_MAYSHARE)
4262 region_abort(resv_map, from, to);
Joonsoo Kimf031dd22014-04-03 14:47:28 -07004263 if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
4264 kref_put(&resv_map->refs, resv_map_release);
Dave Hansenc50ac052012-05-29 15:06:46 -07004265 return ret;
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -07004266}
4267
Mike Kravetzb5cec282015-09-08 15:01:41 -07004268long hugetlb_unreserve_pages(struct inode *inode, long start, long end,
4269 long freed)
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -07004270{
Andi Kleena5516432008-07-23 21:27:41 -07004271 struct hstate *h = hstate_inode(inode);
Joonsoo Kim4e35f482014-04-03 14:47:30 -07004272 struct resv_map *resv_map = inode_resv_map(inode);
Joonsoo Kim9119a412014-04-03 14:47:25 -07004273 long chg = 0;
David Gibson90481622012-03-21 16:34:12 -07004274 struct hugepage_subpool *spool = subpool_inode(inode);
Mike Kravetz1c5ecae2015-04-15 16:13:39 -07004275 long gbl_reserve;
Ken Chen45c682a2007-11-14 16:59:44 -08004276
Mike Kravetzb5cec282015-09-08 15:01:41 -07004277 if (resv_map) {
4278 chg = region_del(resv_map, start, end);
4279 /*
4280 * region_del() can fail in the rare case where a region
4281 * must be split and another region descriptor can not be
4282 * allocated. If end == LONG_MAX, it will not fail.
4283 */
4284 if (chg < 0)
4285 return chg;
4286 }
4287
Ken Chen45c682a2007-11-14 16:59:44 -08004288 spin_lock(&inode->i_lock);
Eric Sandeene4c6f8b2009-07-29 15:02:16 -07004289 inode->i_blocks -= (blocks_per_huge_page(h) * freed);
Ken Chen45c682a2007-11-14 16:59:44 -08004290 spin_unlock(&inode->i_lock);
4291
Mike Kravetz1c5ecae2015-04-15 16:13:39 -07004292 /*
4293 * If the subpool has a minimum size, the number of global
4294 * reservations to be released may be adjusted.
4295 */
4296 gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed));
4297 hugetlb_acct_memory(h, -gbl_reserve);
Mike Kravetzb5cec282015-09-08 15:01:41 -07004298
4299 return 0;
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -07004300}
Naoya Horiguchi93f70f92010-05-28 09:29:20 +09004301
Steve Capper3212b532013-04-23 12:35:02 +01004302#ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
4303static unsigned long page_table_shareable(struct vm_area_struct *svma,
4304 struct vm_area_struct *vma,
4305 unsigned long addr, pgoff_t idx)
4306{
4307 unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) +
4308 svma->vm_start;
4309 unsigned long sbase = saddr & PUD_MASK;
4310 unsigned long s_end = sbase + PUD_SIZE;
4311
4312 /* Allow segments to share if only one is marked locked */
Eric B Munsonde60f5f2015-11-05 18:51:36 -08004313 unsigned long vm_flags = vma->vm_flags & VM_LOCKED_CLEAR_MASK;
4314 unsigned long svm_flags = svma->vm_flags & VM_LOCKED_CLEAR_MASK;
Steve Capper3212b532013-04-23 12:35:02 +01004315
4316 /*
4317 * match the virtual addresses, permission and the alignment of the
4318 * page table page.
4319 */
4320 if (pmd_index(addr) != pmd_index(saddr) ||
4321 vm_flags != svm_flags ||
4322 sbase < svma->vm_start || svma->vm_end < s_end)
4323 return 0;
4324
4325 return saddr;
4326}
4327
Nicholas Krause31aafb42015-09-04 15:47:58 -07004328static bool vma_shareable(struct vm_area_struct *vma, unsigned long addr)
Steve Capper3212b532013-04-23 12:35:02 +01004329{
4330 unsigned long base = addr & PUD_MASK;
4331 unsigned long end = base + PUD_SIZE;
4332
4333 /*
4334 * check on proper vm_flags and page table alignment
4335 */
4336 if (vma->vm_flags & VM_MAYSHARE &&
4337 vma->vm_start <= base && end <= vma->vm_end)
Nicholas Krause31aafb42015-09-04 15:47:58 -07004338 return true;
4339 return false;
Steve Capper3212b532013-04-23 12:35:02 +01004340}
4341
4342/*
4343 * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
4344 * and returns the corresponding pte. While this is not necessary for the
4345 * !shared pmd case because we can allocate the pmd later as well, it makes the
4346 * code much cleaner. pmd allocation is essential for the shared case because
Davidlohr Buesoc8c06ef2014-12-12 16:54:24 -08004347 * pud has to be populated inside the same i_mmap_rwsem section - otherwise
Steve Capper3212b532013-04-23 12:35:02 +01004348 * racing tasks could either miss the sharing (see huge_pte_offset) or select a
4349 * bad pmd for sharing.
4350 */
4351pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
4352{
4353 struct vm_area_struct *vma = find_vma(mm, addr);
4354 struct address_space *mapping = vma->vm_file->f_mapping;
4355 pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) +
4356 vma->vm_pgoff;
4357 struct vm_area_struct *svma;
4358 unsigned long saddr;
4359 pte_t *spte = NULL;
4360 pte_t *pte;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08004361 spinlock_t *ptl;
Steve Capper3212b532013-04-23 12:35:02 +01004362
4363 if (!vma_shareable(vma, addr))
4364 return (pte_t *)pmd_alloc(mm, pud, addr);
4365
Davidlohr Bueso83cde9e2014-12-12 16:54:21 -08004366 i_mmap_lock_write(mapping);
Steve Capper3212b532013-04-23 12:35:02 +01004367 vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) {
4368 if (svma == vma)
4369 continue;
4370
4371 saddr = page_table_shareable(svma, vma, addr, idx);
4372 if (saddr) {
4373 spte = huge_pte_offset(svma->vm_mm, saddr);
4374 if (spte) {
4375 get_page(virt_to_page(spte));
4376 break;
4377 }
4378 }
4379 }
4380
4381 if (!spte)
4382 goto out;
4383
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08004384 ptl = huge_pte_lockptr(hstate_vma(vma), mm, spte);
4385 spin_lock(ptl);
Kirill A. Shutemovdc6c9a32015-02-11 15:26:50 -08004386 if (pud_none(*pud)) {
Steve Capper3212b532013-04-23 12:35:02 +01004387 pud_populate(mm, pud,
4388 (pmd_t *)((unsigned long)spte & PAGE_MASK));
Kirill A. Shutemovc17b1f42016-06-24 14:49:51 -07004389 mm_inc_nr_pmds(mm);
Kirill A. Shutemovdc6c9a32015-02-11 15:26:50 -08004390 } else {
Steve Capper3212b532013-04-23 12:35:02 +01004391 put_page(virt_to_page(spte));
Kirill A. Shutemovdc6c9a32015-02-11 15:26:50 -08004392 }
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08004393 spin_unlock(ptl);
Steve Capper3212b532013-04-23 12:35:02 +01004394out:
4395 pte = (pte_t *)pmd_alloc(mm, pud, addr);
Davidlohr Bueso83cde9e2014-12-12 16:54:21 -08004396 i_mmap_unlock_write(mapping);
Steve Capper3212b532013-04-23 12:35:02 +01004397 return pte;
4398}
4399
4400/*
4401 * unmap huge page backed by shared pte.
4402 *
4403 * Hugetlb pte page is ref counted at the time of mapping. If pte is shared
4404 * indicated by page_count > 1, unmap is achieved by clearing pud and
4405 * decrementing the ref count. If count == 1, the pte page is not shared.
4406 *
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08004407 * called with page table lock held.
Steve Capper3212b532013-04-23 12:35:02 +01004408 *
4409 * returns: 1 successfully unmapped a shared pte page
4410 * 0 the underlying pte page is not shared, or it is the last user
4411 */
4412int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
4413{
4414 pgd_t *pgd = pgd_offset(mm, *addr);
4415 pud_t *pud = pud_offset(pgd, *addr);
4416
4417 BUG_ON(page_count(virt_to_page(ptep)) == 0);
4418 if (page_count(virt_to_page(ptep)) == 1)
4419 return 0;
4420
4421 pud_clear(pud);
4422 put_page(virt_to_page(ptep));
Kirill A. Shutemovdc6c9a32015-02-11 15:26:50 -08004423 mm_dec_nr_pmds(mm);
Steve Capper3212b532013-04-23 12:35:02 +01004424 *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE;
4425 return 1;
4426}
Steve Capper9e5fc742013-04-30 08:02:03 +01004427#define want_pmd_share() (1)
4428#else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
4429pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
4430{
4431 return NULL;
4432}
Zhang Zhene81f2d22015-06-24 16:56:13 -07004433
4434int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
4435{
4436 return 0;
4437}
Steve Capper9e5fc742013-04-30 08:02:03 +01004438#define want_pmd_share() (0)
Steve Capper3212b532013-04-23 12:35:02 +01004439#endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
4440
Steve Capper9e5fc742013-04-30 08:02:03 +01004441#ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
4442pte_t *huge_pte_alloc(struct mm_struct *mm,
4443 unsigned long addr, unsigned long sz)
4444{
4445 pgd_t *pgd;
4446 pud_t *pud;
4447 pte_t *pte = NULL;
4448
4449 pgd = pgd_offset(mm, addr);
4450 pud = pud_alloc(mm, pgd, addr);
4451 if (pud) {
4452 if (sz == PUD_SIZE) {
4453 pte = (pte_t *)pud;
4454 } else {
4455 BUG_ON(sz != PMD_SIZE);
4456 if (want_pmd_share() && pud_none(*pud))
4457 pte = huge_pmd_share(mm, addr, pud);
4458 else
4459 pte = (pte_t *)pmd_alloc(mm, pud, addr);
4460 }
4461 }
Michal Hocko4e666312016-08-02 14:02:34 -07004462 BUG_ON(pte && pte_present(*pte) && !pte_huge(*pte));
Steve Capper9e5fc742013-04-30 08:02:03 +01004463
4464 return pte;
4465}
4466
4467pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr)
4468{
4469 pgd_t *pgd;
4470 pud_t *pud;
4471 pmd_t *pmd = NULL;
4472
4473 pgd = pgd_offset(mm, addr);
4474 if (pgd_present(*pgd)) {
4475 pud = pud_offset(pgd, addr);
4476 if (pud_present(*pud)) {
4477 if (pud_huge(*pud))
4478 return (pte_t *)pud;
4479 pmd = pmd_offset(pud, addr);
4480 }
4481 }
4482 return (pte_t *) pmd;
4483}
4484
Naoya Horiguchi61f77ed2015-02-11 15:25:15 -08004485#endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
4486
4487/*
4488 * These functions are overwritable if your architecture needs its own
4489 * behavior.
4490 */
4491struct page * __weak
4492follow_huge_addr(struct mm_struct *mm, unsigned long address,
4493 int write)
4494{
4495 return ERR_PTR(-EINVAL);
4496}
4497
4498struct page * __weak
Steve Capper9e5fc742013-04-30 08:02:03 +01004499follow_huge_pmd(struct mm_struct *mm, unsigned long address,
Naoya Horiguchie66f17f2015-02-11 15:25:22 -08004500 pmd_t *pmd, int flags)
Steve Capper9e5fc742013-04-30 08:02:03 +01004501{
Naoya Horiguchie66f17f2015-02-11 15:25:22 -08004502 struct page *page = NULL;
4503 spinlock_t *ptl;
Naoya Horiguchi40c5b992017-03-31 15:11:55 -07004504 pte_t pte;
Naoya Horiguchie66f17f2015-02-11 15:25:22 -08004505retry:
4506 ptl = pmd_lockptr(mm, pmd);
4507 spin_lock(ptl);
4508 /*
4509 * make sure that the address range covered by this pmd is not
4510 * unmapped from other threads.
4511 */
4512 if (!pmd_huge(*pmd))
4513 goto out;
Naoya Horiguchi40c5b992017-03-31 15:11:55 -07004514 pte = huge_ptep_get((pte_t *)pmd);
4515 if (pte_present(pte)) {
Gerald Schaefer97534122015-04-14 15:42:30 -07004516 page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT);
Naoya Horiguchie66f17f2015-02-11 15:25:22 -08004517 if (flags & FOLL_GET)
4518 get_page(page);
4519 } else {
Naoya Horiguchi40c5b992017-03-31 15:11:55 -07004520 if (is_hugetlb_entry_migration(pte)) {
Naoya Horiguchie66f17f2015-02-11 15:25:22 -08004521 spin_unlock(ptl);
4522 __migration_entry_wait(mm, (pte_t *)pmd, ptl);
4523 goto retry;
4524 }
4525 /*
4526 * hwpoisoned entry is treated as no_page_table in
4527 * follow_page_mask().
4528 */
4529 }
4530out:
4531 spin_unlock(ptl);
Steve Capper9e5fc742013-04-30 08:02:03 +01004532 return page;
4533}
4534
Naoya Horiguchi61f77ed2015-02-11 15:25:15 -08004535struct page * __weak
Steve Capper9e5fc742013-04-30 08:02:03 +01004536follow_huge_pud(struct mm_struct *mm, unsigned long address,
Naoya Horiguchie66f17f2015-02-11 15:25:22 -08004537 pud_t *pud, int flags)
Steve Capper9e5fc742013-04-30 08:02:03 +01004538{
Naoya Horiguchie66f17f2015-02-11 15:25:22 -08004539 if (flags & FOLL_GET)
4540 return NULL;
Steve Capper9e5fc742013-04-30 08:02:03 +01004541
Naoya Horiguchie66f17f2015-02-11 15:25:22 -08004542 return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT);
Steve Capper9e5fc742013-04-30 08:02:03 +01004543}
4544
Andi Kleend5bd9102010-09-27 09:00:12 +02004545#ifdef CONFIG_MEMORY_FAILURE
4546
Naoya Horiguchi93f70f92010-05-28 09:29:20 +09004547/*
4548 * This function is called from memory failure code.
Naoya Horiguchi93f70f92010-05-28 09:29:20 +09004549 */
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +09004550int dequeue_hwpoisoned_huge_page(struct page *hpage)
Naoya Horiguchi93f70f92010-05-28 09:29:20 +09004551{
4552 struct hstate *h = page_hstate(hpage);
4553 int nid = page_to_nid(hpage);
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +09004554 int ret = -EBUSY;
Naoya Horiguchi93f70f92010-05-28 09:29:20 +09004555
4556 spin_lock(&hugetlb_lock);
Naoya Horiguchi7e1f0492015-04-15 16:14:41 -07004557 /*
4558 * Just checking !page_huge_active is not enough, because that could be
4559 * an isolated/hwpoisoned hugepage (which have >0 refcount).
4560 */
4561 if (!page_huge_active(hpage) && !page_count(hpage)) {
Naoya Horiguchi56f2fb12012-12-12 13:52:33 -08004562 /*
4563 * Hwpoisoned hugepage isn't linked to activelist or freelist,
4564 * but dangling hpage->lru can trigger list-debug warnings
4565 * (this happens when we call unpoison_memory() on it),
4566 * so let it point to itself with list_del_init().
4567 */
4568 list_del_init(&hpage->lru);
Naoya Horiguchi8c6c2ec2010-09-08 10:19:38 +09004569 set_page_refcounted(hpage);
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +09004570 h->free_huge_pages--;
4571 h->free_huge_pages_node[nid]--;
4572 ret = 0;
4573 }
Naoya Horiguchi93f70f92010-05-28 09:29:20 +09004574 spin_unlock(&hugetlb_lock);
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +09004575 return ret;
Naoya Horiguchi93f70f92010-05-28 09:29:20 +09004576}
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +09004577#endif
Naoya Horiguchi31caf662013-09-11 14:21:59 -07004578
4579bool isolate_huge_page(struct page *page, struct list_head *list)
4580{
Naoya Horiguchibcc54222015-04-15 16:14:38 -07004581 bool ret = true;
4582
Sasha Levin309381fea2014-01-23 15:52:54 -08004583 VM_BUG_ON_PAGE(!PageHead(page), page);
Naoya Horiguchi31caf662013-09-11 14:21:59 -07004584 spin_lock(&hugetlb_lock);
Naoya Horiguchibcc54222015-04-15 16:14:38 -07004585 if (!page_huge_active(page) || !get_page_unless_zero(page)) {
4586 ret = false;
4587 goto unlock;
4588 }
4589 clear_page_huge_active(page);
Naoya Horiguchi31caf662013-09-11 14:21:59 -07004590 list_move_tail(&page->lru, list);
Naoya Horiguchibcc54222015-04-15 16:14:38 -07004591unlock:
Naoya Horiguchi31caf662013-09-11 14:21:59 -07004592 spin_unlock(&hugetlb_lock);
Naoya Horiguchibcc54222015-04-15 16:14:38 -07004593 return ret;
Naoya Horiguchi31caf662013-09-11 14:21:59 -07004594}
4595
4596void putback_active_hugepage(struct page *page)
4597{
Sasha Levin309381fea2014-01-23 15:52:54 -08004598 VM_BUG_ON_PAGE(!PageHead(page), page);
Naoya Horiguchi31caf662013-09-11 14:21:59 -07004599 spin_lock(&hugetlb_lock);
Naoya Horiguchibcc54222015-04-15 16:14:38 -07004600 set_page_huge_active(page);
Naoya Horiguchi31caf662013-09-11 14:21:59 -07004601 list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist);
4602 spin_unlock(&hugetlb_lock);
4603 put_page(page);
4604}