blob: 9cc773483624e4cbb1592ddde74f9c8faa21ef87 [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
2 * Generic hugetlb support.
Nadia Yvette Chambers6d49e352012-12-06 10:39:54 +01003 * (C) Nadia Yvette Chambers, April 2004
Linus Torvalds1da177e2005-04-16 15:20:36 -07004 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07005#include <linux/list.h>
6#include <linux/init.h>
7#include <linux/module.h>
8#include <linux/mm.h>
Alexey Dobriyane1759c22008-10-15 23:50:22 +04009#include <linux/seq_file.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070010#include <linux/sysctl.h>
11#include <linux/highmem.h>
Andrea Arcangelicddb8a52008-07-28 15:46:29 -070012#include <linux/mmu_notifier.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070013#include <linux/nodemask.h>
David Gibson63551ae2005-06-21 17:14:44 -070014#include <linux/pagemap.h>
Christoph Lameter5da7ca82006-01-06 00:10:46 -080015#include <linux/mempolicy.h>
Gideon Israel Dsouza3b321232014-04-07 15:37:26 -070016#include <linux/compiler.h>
Christoph Lameteraea47ff2006-01-08 01:00:57 -080017#include <linux/cpuset.h>
David Gibson3935baa2006-03-22 00:08:53 -080018#include <linux/mutex.h>
Andi Kleenaa888a72008-07-23 21:27:47 -070019#include <linux/bootmem.h>
Nishanth Aravamudana3437872008-07-23 21:27:44 -070020#include <linux/sysfs.h>
Tejun Heo5a0e3ad2010-03-24 17:04:11 +090021#include <linux/slab.h>
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +090022#include <linux/rmap.h>
Naoya Horiguchifd6a03ed2010-05-28 09:29:21 +090023#include <linux/swap.h>
24#include <linux/swapops.h>
Naoya Horiguchic8721bb2013-09-11 14:22:09 -070025#include <linux/page-isolation.h>
Davidlohr Bueso8382d912014-04-03 14:47:31 -070026#include <linux/jhash.h>
Linus Torvaldsd6606682008-08-06 12:04:54 -070027
David Gibson63551ae2005-06-21 17:14:44 -070028#include <asm/page.h>
29#include <asm/pgtable.h>
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -070030#include <asm/tlb.h>
David Gibson63551ae2005-06-21 17:14:44 -070031
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -070032#include <linux/io.h>
David Gibson63551ae2005-06-21 17:14:44 -070033#include <linux/hugetlb.h>
Aneesh Kumar K.V9dd540e2012-07-31 16:42:15 -070034#include <linux/hugetlb_cgroup.h>
Lee Schermerhorn9a305232009-12-14 17:58:25 -080035#include <linux/node.h>
Nick Piggin7835e982006-03-22 00:08:40 -080036#include "internal.h"
Linus Torvalds1da177e2005-04-16 15:20:36 -070037
Andrey Ryabinin753162c2015-02-10 14:11:36 -080038int hugepages_treat_as_movable;
Andi Kleena5516432008-07-23 21:27:41 -070039
Aneesh Kumar K.Vc3f38a32012-07-31 16:42:10 -070040int hugetlb_max_hstate __read_mostly;
Andi Kleene5ff2152008-07-23 21:27:42 -070041unsigned int default_hstate_idx;
42struct hstate hstates[HUGE_MAX_HSTATE];
Naoya Horiguchi641844f2015-06-24 16:56:59 -070043/*
44 * Minimum page order among possible hugepage sizes, set to a proper value
45 * at boot time.
46 */
47static unsigned int minimum_order __read_mostly = UINT_MAX;
Andi Kleene5ff2152008-07-23 21:27:42 -070048
Jon Tollefson53ba51d2008-07-23 21:27:52 -070049__initdata LIST_HEAD(huge_boot_pages);
50
Andi Kleene5ff2152008-07-23 21:27:42 -070051/* for command line parsing */
52static struct hstate * __initdata parsed_hstate;
53static unsigned long __initdata default_hstate_max_huge_pages;
Nick Piggine11bfbf2008-07-23 21:27:52 -070054static unsigned long __initdata default_hstate_size;
Andi Kleene5ff2152008-07-23 21:27:42 -070055
David Gibson3935baa2006-03-22 00:08:53 -080056/*
Naoya Horiguchi31caf662013-09-11 14:21:59 -070057 * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages,
58 * free_huge_pages, and surplus_huge_pages.
David Gibson3935baa2006-03-22 00:08:53 -080059 */
Aneesh Kumar K.Vc3f38a32012-07-31 16:42:10 -070060DEFINE_SPINLOCK(hugetlb_lock);
Eric Paris0bd0f9f2005-11-21 21:32:28 -080061
Davidlohr Bueso8382d912014-04-03 14:47:31 -070062/*
63 * Serializes faults on the same logical page. This is used to
64 * prevent spurious OOMs when the hugepage pool is fully utilized.
65 */
66static int num_fault_mutexes;
Mike Kravetzc672c7f2015-09-08 15:01:35 -070067struct mutex *hugetlb_fault_mutex_table ____cacheline_aligned_in_smp;
Davidlohr Bueso8382d912014-04-03 14:47:31 -070068
Mike Kravetz7ca02d02015-04-15 16:13:42 -070069/* Forward declaration */
70static int hugetlb_acct_memory(struct hstate *h, long delta);
71
David Gibson90481622012-03-21 16:34:12 -070072static inline void unlock_or_release_subpool(struct hugepage_subpool *spool)
73{
74 bool free = (spool->count == 0) && (spool->used_hpages == 0);
75
76 spin_unlock(&spool->lock);
77
78 /* If no pages are used, and no other handles to the subpool
Mike Kravetz7ca02d02015-04-15 16:13:42 -070079 * remain, give up any reservations mased on minimum size and
80 * free the subpool */
81 if (free) {
82 if (spool->min_hpages != -1)
83 hugetlb_acct_memory(spool->hstate,
84 -spool->min_hpages);
David Gibson90481622012-03-21 16:34:12 -070085 kfree(spool);
Mike Kravetz7ca02d02015-04-15 16:13:42 -070086 }
David Gibson90481622012-03-21 16:34:12 -070087}
88
Mike Kravetz7ca02d02015-04-15 16:13:42 -070089struct hugepage_subpool *hugepage_new_subpool(struct hstate *h, long max_hpages,
90 long min_hpages)
David Gibson90481622012-03-21 16:34:12 -070091{
92 struct hugepage_subpool *spool;
93
Mike Kravetzc6a91822015-04-15 16:13:36 -070094 spool = kzalloc(sizeof(*spool), GFP_KERNEL);
David Gibson90481622012-03-21 16:34:12 -070095 if (!spool)
96 return NULL;
97
98 spin_lock_init(&spool->lock);
99 spool->count = 1;
Mike Kravetz7ca02d02015-04-15 16:13:42 -0700100 spool->max_hpages = max_hpages;
101 spool->hstate = h;
102 spool->min_hpages = min_hpages;
103
104 if (min_hpages != -1 && hugetlb_acct_memory(h, min_hpages)) {
105 kfree(spool);
106 return NULL;
107 }
108 spool->rsv_hpages = min_hpages;
David Gibson90481622012-03-21 16:34:12 -0700109
110 return spool;
111}
112
113void hugepage_put_subpool(struct hugepage_subpool *spool)
114{
115 spin_lock(&spool->lock);
116 BUG_ON(!spool->count);
117 spool->count--;
118 unlock_or_release_subpool(spool);
119}
120
Mike Kravetz1c5ecae2015-04-15 16:13:39 -0700121/*
122 * Subpool accounting for allocating and reserving pages.
123 * Return -ENOMEM if there are not enough resources to satisfy the
124 * the request. Otherwise, return the number of pages by which the
125 * global pools must be adjusted (upward). The returned value may
126 * only be different than the passed value (delta) in the case where
127 * a subpool minimum size must be manitained.
128 */
129static long hugepage_subpool_get_pages(struct hugepage_subpool *spool,
David Gibson90481622012-03-21 16:34:12 -0700130 long delta)
131{
Mike Kravetz1c5ecae2015-04-15 16:13:39 -0700132 long ret = delta;
David Gibson90481622012-03-21 16:34:12 -0700133
134 if (!spool)
Mike Kravetz1c5ecae2015-04-15 16:13:39 -0700135 return ret;
David Gibson90481622012-03-21 16:34:12 -0700136
137 spin_lock(&spool->lock);
David Gibson90481622012-03-21 16:34:12 -0700138
Mike Kravetz1c5ecae2015-04-15 16:13:39 -0700139 if (spool->max_hpages != -1) { /* maximum size accounting */
140 if ((spool->used_hpages + delta) <= spool->max_hpages)
141 spool->used_hpages += delta;
142 else {
143 ret = -ENOMEM;
144 goto unlock_ret;
145 }
146 }
147
148 if (spool->min_hpages != -1) { /* minimum size accounting */
149 if (delta > spool->rsv_hpages) {
150 /*
151 * Asking for more reserves than those already taken on
152 * behalf of subpool. Return difference.
153 */
154 ret = delta - spool->rsv_hpages;
155 spool->rsv_hpages = 0;
156 } else {
157 ret = 0; /* reserves already accounted for */
158 spool->rsv_hpages -= delta;
159 }
160 }
161
162unlock_ret:
163 spin_unlock(&spool->lock);
David Gibson90481622012-03-21 16:34:12 -0700164 return ret;
165}
166
Mike Kravetz1c5ecae2015-04-15 16:13:39 -0700167/*
168 * Subpool accounting for freeing and unreserving pages.
169 * Return the number of global page reservations that must be dropped.
170 * The return value may only be different than the passed value (delta)
171 * in the case where a subpool minimum size must be maintained.
172 */
173static long hugepage_subpool_put_pages(struct hugepage_subpool *spool,
David Gibson90481622012-03-21 16:34:12 -0700174 long delta)
175{
Mike Kravetz1c5ecae2015-04-15 16:13:39 -0700176 long ret = delta;
177
David Gibson90481622012-03-21 16:34:12 -0700178 if (!spool)
Mike Kravetz1c5ecae2015-04-15 16:13:39 -0700179 return delta;
David Gibson90481622012-03-21 16:34:12 -0700180
181 spin_lock(&spool->lock);
Mike Kravetz1c5ecae2015-04-15 16:13:39 -0700182
183 if (spool->max_hpages != -1) /* maximum size accounting */
184 spool->used_hpages -= delta;
185
186 if (spool->min_hpages != -1) { /* minimum size accounting */
187 if (spool->rsv_hpages + delta <= spool->min_hpages)
188 ret = 0;
189 else
190 ret = spool->rsv_hpages + delta - spool->min_hpages;
191
192 spool->rsv_hpages += delta;
193 if (spool->rsv_hpages > spool->min_hpages)
194 spool->rsv_hpages = spool->min_hpages;
195 }
196
197 /*
198 * If hugetlbfs_put_super couldn't free spool due to an outstanding
199 * quota reference, free it now.
200 */
David Gibson90481622012-03-21 16:34:12 -0700201 unlock_or_release_subpool(spool);
Mike Kravetz1c5ecae2015-04-15 16:13:39 -0700202
203 return ret;
David Gibson90481622012-03-21 16:34:12 -0700204}
205
206static inline struct hugepage_subpool *subpool_inode(struct inode *inode)
207{
208 return HUGETLBFS_SB(inode->i_sb)->spool;
209}
210
211static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma)
212{
Al Viro496ad9a2013-01-23 17:07:38 -0500213 return subpool_inode(file_inode(vma->vm_file));
David Gibson90481622012-03-21 16:34:12 -0700214}
215
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700216/*
Andy Whitcroft96822902008-07-23 21:27:29 -0700217 * Region tracking -- allows tracking of reservations and instantiated pages
218 * across the pages in a mapping.
Andy Whitcroft84afd992008-07-23 21:27:32 -0700219 *
Mike Kravetz1dd308a2015-06-24 16:57:52 -0700220 * The region data structures are embedded into a resv_map and protected
221 * by a resv_map's lock. The set of regions within the resv_map represent
222 * reservations for huge pages, or huge pages that have already been
223 * instantiated within the map. The from and to elements are huge page
224 * indicies into the associated mapping. from indicates the starting index
225 * of the region. to represents the first index past the end of the region.
226 *
227 * For example, a file region structure with from == 0 and to == 4 represents
228 * four huge pages in a mapping. It is important to note that the to element
229 * represents the first element past the end of the region. This is used in
230 * arithmetic as 4(to) - 0(from) = 4 huge pages in the region.
231 *
232 * Interval notation of the form [from, to) will be used to indicate that
233 * the endpoint from is inclusive and to is exclusive.
Andy Whitcroft96822902008-07-23 21:27:29 -0700234 */
235struct file_region {
236 struct list_head link;
237 long from;
238 long to;
239};
240
Mike Kravetz1dd308a2015-06-24 16:57:52 -0700241/*
242 * Add the huge page range represented by [f, t) to the reserve
Mike Kravetz5e911372015-09-08 15:01:28 -0700243 * map. In the normal case, existing regions will be expanded
244 * to accommodate the specified range. Sufficient regions should
245 * exist for expansion due to the previous call to region_chg
246 * with the same range. However, it is possible that region_del
247 * could have been called after region_chg and modifed the map
248 * in such a way that no region exists to be expanded. In this
249 * case, pull a region descriptor from the cache associated with
250 * the map and use that for the new range.
Mike Kravetzcf3ad202015-06-24 16:57:55 -0700251 *
252 * Return the number of new huge pages added to the map. This
253 * number is greater than or equal to zero.
Mike Kravetz1dd308a2015-06-24 16:57:52 -0700254 */
Joonsoo Kim1406ec92014-04-03 14:47:26 -0700255static long region_add(struct resv_map *resv, long f, long t)
Andy Whitcroft96822902008-07-23 21:27:29 -0700256{
Joonsoo Kim1406ec92014-04-03 14:47:26 -0700257 struct list_head *head = &resv->regions;
Andy Whitcroft96822902008-07-23 21:27:29 -0700258 struct file_region *rg, *nrg, *trg;
Mike Kravetzcf3ad202015-06-24 16:57:55 -0700259 long add = 0;
Andy Whitcroft96822902008-07-23 21:27:29 -0700260
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700261 spin_lock(&resv->lock);
Andy Whitcroft96822902008-07-23 21:27:29 -0700262 /* Locate the region we are either in or before. */
263 list_for_each_entry(rg, head, link)
264 if (f <= rg->to)
265 break;
266
Mike Kravetz5e911372015-09-08 15:01:28 -0700267 /*
268 * If no region exists which can be expanded to include the
269 * specified range, the list must have been modified by an
270 * interleving call to region_del(). Pull a region descriptor
271 * from the cache and use it for this range.
272 */
273 if (&rg->link == head || t < rg->from) {
274 VM_BUG_ON(resv->region_cache_count <= 0);
275
276 resv->region_cache_count--;
277 nrg = list_first_entry(&resv->region_cache, struct file_region,
278 link);
279 list_del(&nrg->link);
280
281 nrg->from = f;
282 nrg->to = t;
283 list_add(&nrg->link, rg->link.prev);
284
285 add += t - f;
286 goto out_locked;
287 }
288
Andy Whitcroft96822902008-07-23 21:27:29 -0700289 /* Round our left edge to the current segment if it encloses us. */
290 if (f > rg->from)
291 f = rg->from;
292
293 /* Check for and consume any regions we now overlap with. */
294 nrg = rg;
295 list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
296 if (&rg->link == head)
297 break;
298 if (rg->from > t)
299 break;
300
301 /* If this area reaches higher then extend our area to
302 * include it completely. If this is not the first area
303 * which we intend to reuse, free it. */
304 if (rg->to > t)
305 t = rg->to;
306 if (rg != nrg) {
Mike Kravetzcf3ad202015-06-24 16:57:55 -0700307 /* Decrement return value by the deleted range.
308 * Another range will span this area so that by
309 * end of routine add will be >= zero
310 */
311 add -= (rg->to - rg->from);
Andy Whitcroft96822902008-07-23 21:27:29 -0700312 list_del(&rg->link);
313 kfree(rg);
314 }
315 }
Mike Kravetzcf3ad202015-06-24 16:57:55 -0700316
317 add += (nrg->from - f); /* Added to beginning of region */
Andy Whitcroft96822902008-07-23 21:27:29 -0700318 nrg->from = f;
Mike Kravetzcf3ad202015-06-24 16:57:55 -0700319 add += t - nrg->to; /* Added to end of region */
Andy Whitcroft96822902008-07-23 21:27:29 -0700320 nrg->to = t;
Mike Kravetzcf3ad202015-06-24 16:57:55 -0700321
Mike Kravetz5e911372015-09-08 15:01:28 -0700322out_locked:
323 resv->adds_in_progress--;
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700324 spin_unlock(&resv->lock);
Mike Kravetzcf3ad202015-06-24 16:57:55 -0700325 VM_BUG_ON(add < 0);
326 return add;
Andy Whitcroft96822902008-07-23 21:27:29 -0700327}
328
Mike Kravetz1dd308a2015-06-24 16:57:52 -0700329/*
330 * Examine the existing reserve map and determine how many
331 * huge pages in the specified range [f, t) are NOT currently
332 * represented. This routine is called before a subsequent
333 * call to region_add that will actually modify the reserve
334 * map to add the specified range [f, t). region_chg does
335 * not change the number of huge pages represented by the
336 * map. However, if the existing regions in the map can not
337 * be expanded to represent the new range, a new file_region
338 * structure is added to the map as a placeholder. This is
339 * so that the subsequent region_add call will have all the
340 * regions it needs and will not fail.
341 *
Mike Kravetz5e911372015-09-08 15:01:28 -0700342 * Upon entry, region_chg will also examine the cache of region descriptors
343 * associated with the map. If there are not enough descriptors cached, one
344 * will be allocated for the in progress add operation.
345 *
346 * Returns the number of huge pages that need to be added to the existing
347 * reservation map for the range [f, t). This number is greater or equal to
348 * zero. -ENOMEM is returned if a new file_region structure or cache entry
349 * is needed and can not be allocated.
Mike Kravetz1dd308a2015-06-24 16:57:52 -0700350 */
Joonsoo Kim1406ec92014-04-03 14:47:26 -0700351static long region_chg(struct resv_map *resv, long f, long t)
Andy Whitcroft96822902008-07-23 21:27:29 -0700352{
Joonsoo Kim1406ec92014-04-03 14:47:26 -0700353 struct list_head *head = &resv->regions;
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700354 struct file_region *rg, *nrg = NULL;
Andy Whitcroft96822902008-07-23 21:27:29 -0700355 long chg = 0;
356
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700357retry:
358 spin_lock(&resv->lock);
Mike Kravetz5e911372015-09-08 15:01:28 -0700359retry_locked:
360 resv->adds_in_progress++;
361
362 /*
363 * Check for sufficient descriptors in the cache to accommodate
364 * the number of in progress add operations.
365 */
366 if (resv->adds_in_progress > resv->region_cache_count) {
367 struct file_region *trg;
368
369 VM_BUG_ON(resv->adds_in_progress - resv->region_cache_count > 1);
370 /* Must drop lock to allocate a new descriptor. */
371 resv->adds_in_progress--;
372 spin_unlock(&resv->lock);
373
374 trg = kmalloc(sizeof(*trg), GFP_KERNEL);
375 if (!trg)
376 return -ENOMEM;
377
378 spin_lock(&resv->lock);
379 list_add(&trg->link, &resv->region_cache);
380 resv->region_cache_count++;
381 goto retry_locked;
382 }
383
Andy Whitcroft96822902008-07-23 21:27:29 -0700384 /* Locate the region we are before or in. */
385 list_for_each_entry(rg, head, link)
386 if (f <= rg->to)
387 break;
388
389 /* If we are below the current region then a new region is required.
390 * Subtle, allocate a new region at the position but make it zero
391 * size such that we can guarantee to record the reservation. */
392 if (&rg->link == head || t < rg->from) {
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700393 if (!nrg) {
Mike Kravetz5e911372015-09-08 15:01:28 -0700394 resv->adds_in_progress--;
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700395 spin_unlock(&resv->lock);
396 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
397 if (!nrg)
398 return -ENOMEM;
Andy Whitcroft96822902008-07-23 21:27:29 -0700399
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700400 nrg->from = f;
401 nrg->to = f;
402 INIT_LIST_HEAD(&nrg->link);
403 goto retry;
404 }
405
406 list_add(&nrg->link, rg->link.prev);
407 chg = t - f;
408 goto out_nrg;
Andy Whitcroft96822902008-07-23 21:27:29 -0700409 }
410
411 /* Round our left edge to the current segment if it encloses us. */
412 if (f > rg->from)
413 f = rg->from;
414 chg = t - f;
415
416 /* Check for and consume any regions we now overlap with. */
417 list_for_each_entry(rg, rg->link.prev, link) {
418 if (&rg->link == head)
419 break;
420 if (rg->from > t)
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700421 goto out;
Andy Whitcroft96822902008-07-23 21:27:29 -0700422
Lucas De Marchi25985ed2011-03-30 22:57:33 -0300423 /* We overlap with this area, if it extends further than
Andy Whitcroft96822902008-07-23 21:27:29 -0700424 * us then we must extend ourselves. Account for its
425 * existing reservation. */
426 if (rg->to > t) {
427 chg += rg->to - t;
428 t = rg->to;
429 }
430 chg -= rg->to - rg->from;
431 }
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700432
433out:
434 spin_unlock(&resv->lock);
435 /* We already know we raced and no longer need the new region */
436 kfree(nrg);
437 return chg;
438out_nrg:
439 spin_unlock(&resv->lock);
Andy Whitcroft96822902008-07-23 21:27:29 -0700440 return chg;
441}
442
Mike Kravetz1dd308a2015-06-24 16:57:52 -0700443/*
Mike Kravetz5e911372015-09-08 15:01:28 -0700444 * Abort the in progress add operation. The adds_in_progress field
445 * of the resv_map keeps track of the operations in progress between
446 * calls to region_chg and region_add. Operations are sometimes
447 * aborted after the call to region_chg. In such cases, region_abort
448 * is called to decrement the adds_in_progress counter.
449 *
450 * NOTE: The range arguments [f, t) are not needed or used in this
451 * routine. They are kept to make reading the calling code easier as
452 * arguments will match the associated region_chg call.
453 */
454static void region_abort(struct resv_map *resv, long f, long t)
455{
456 spin_lock(&resv->lock);
457 VM_BUG_ON(!resv->region_cache_count);
458 resv->adds_in_progress--;
459 spin_unlock(&resv->lock);
460}
461
462/*
Mike Kravetzfeba16e2015-09-08 15:01:31 -0700463 * Delete the specified range [f, t) from the reserve map. If the
464 * t parameter is LONG_MAX, this indicates that ALL regions after f
465 * should be deleted. Locate the regions which intersect [f, t)
466 * and either trim, delete or split the existing regions.
467 *
468 * Returns the number of huge pages deleted from the reserve map.
469 * In the normal case, the return value is zero or more. In the
470 * case where a region must be split, a new region descriptor must
471 * be allocated. If the allocation fails, -ENOMEM will be returned.
472 * NOTE: If the parameter t == LONG_MAX, then we will never split
473 * a region and possibly return -ENOMEM. Callers specifying
474 * t == LONG_MAX do not need to check for -ENOMEM error.
Mike Kravetz1dd308a2015-06-24 16:57:52 -0700475 */
Mike Kravetzfeba16e2015-09-08 15:01:31 -0700476static long region_del(struct resv_map *resv, long f, long t)
Andy Whitcroft96822902008-07-23 21:27:29 -0700477{
Joonsoo Kim1406ec92014-04-03 14:47:26 -0700478 struct list_head *head = &resv->regions;
Andy Whitcroft96822902008-07-23 21:27:29 -0700479 struct file_region *rg, *trg;
Mike Kravetzfeba16e2015-09-08 15:01:31 -0700480 struct file_region *nrg = NULL;
481 long del = 0;
Andy Whitcroft96822902008-07-23 21:27:29 -0700482
Mike Kravetzfeba16e2015-09-08 15:01:31 -0700483retry:
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700484 spin_lock(&resv->lock);
Mike Kravetzfeba16e2015-09-08 15:01:31 -0700485 list_for_each_entry_safe(rg, trg, head, link) {
486 if (rg->to <= f)
487 continue;
488 if (rg->from >= t)
Andy Whitcroft96822902008-07-23 21:27:29 -0700489 break;
Andy Whitcroft96822902008-07-23 21:27:29 -0700490
Mike Kravetzfeba16e2015-09-08 15:01:31 -0700491 if (f > rg->from && t < rg->to) { /* Must split region */
492 /*
493 * Check for an entry in the cache before dropping
494 * lock and attempting allocation.
495 */
496 if (!nrg &&
497 resv->region_cache_count > resv->adds_in_progress) {
498 nrg = list_first_entry(&resv->region_cache,
499 struct file_region,
500 link);
501 list_del(&nrg->link);
502 resv->region_cache_count--;
503 }
504
505 if (!nrg) {
506 spin_unlock(&resv->lock);
507 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
508 if (!nrg)
509 return -ENOMEM;
510 goto retry;
511 }
512
513 del += t - f;
514
515 /* New entry for end of split region */
516 nrg->from = t;
517 nrg->to = rg->to;
518 INIT_LIST_HEAD(&nrg->link);
519
520 /* Original entry is trimmed */
521 rg->to = f;
522
523 list_add(&nrg->link, &rg->link);
524 nrg = NULL;
525 break;
526 }
527
528 if (f <= rg->from && t >= rg->to) { /* Remove entire region */
529 del += rg->to - rg->from;
530 list_del(&rg->link);
531 kfree(rg);
532 continue;
533 }
534
535 if (f <= rg->from) { /* Trim beginning of region */
536 del += t - rg->from;
537 rg->from = t;
538 } else { /* Trim end of region */
539 del += rg->to - f;
540 rg->to = f;
541 }
Andy Whitcroft96822902008-07-23 21:27:29 -0700542 }
543
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700544 spin_unlock(&resv->lock);
Mike Kravetzfeba16e2015-09-08 15:01:31 -0700545 kfree(nrg);
546 return del;
Andy Whitcroft96822902008-07-23 21:27:29 -0700547}
548
Mike Kravetz1dd308a2015-06-24 16:57:52 -0700549/*
Mike Kravetzb5cec282015-09-08 15:01:41 -0700550 * A rare out of memory error was encountered which prevented removal of
551 * the reserve map region for a page. The huge page itself was free'ed
552 * and removed from the page cache. This routine will adjust the subpool
553 * usage count, and the global reserve count if needed. By incrementing
554 * these counts, the reserve map entry which could not be deleted will
555 * appear as a "reserved" entry instead of simply dangling with incorrect
556 * counts.
557 */
558void hugetlb_fix_reserve_counts(struct inode *inode, bool restore_reserve)
559{
560 struct hugepage_subpool *spool = subpool_inode(inode);
561 long rsv_adjust;
562
563 rsv_adjust = hugepage_subpool_get_pages(spool, 1);
564 if (restore_reserve && rsv_adjust) {
565 struct hstate *h = hstate_inode(inode);
566
567 hugetlb_acct_memory(h, 1);
568 }
569}
570
571/*
Mike Kravetz1dd308a2015-06-24 16:57:52 -0700572 * Count and return the number of huge pages in the reserve map
573 * that intersect with the range [f, t).
574 */
Joonsoo Kim1406ec92014-04-03 14:47:26 -0700575static long region_count(struct resv_map *resv, long f, long t)
Andy Whitcroft84afd992008-07-23 21:27:32 -0700576{
Joonsoo Kim1406ec92014-04-03 14:47:26 -0700577 struct list_head *head = &resv->regions;
Andy Whitcroft84afd992008-07-23 21:27:32 -0700578 struct file_region *rg;
579 long chg = 0;
580
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700581 spin_lock(&resv->lock);
Andy Whitcroft84afd992008-07-23 21:27:32 -0700582 /* Locate each segment we overlap with, and count that overlap. */
583 list_for_each_entry(rg, head, link) {
Wang Sheng-Huif2135a42012-05-29 15:06:17 -0700584 long seg_from;
585 long seg_to;
Andy Whitcroft84afd992008-07-23 21:27:32 -0700586
587 if (rg->to <= f)
588 continue;
589 if (rg->from >= t)
590 break;
591
592 seg_from = max(rg->from, f);
593 seg_to = min(rg->to, t);
594
595 chg += seg_to - seg_from;
596 }
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700597 spin_unlock(&resv->lock);
Andy Whitcroft84afd992008-07-23 21:27:32 -0700598
599 return chg;
600}
601
Andy Whitcroft96822902008-07-23 21:27:29 -0700602/*
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700603 * Convert the address within this vma to the page offset within
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700604 * the mapping, in pagecache page units; huge pages here.
605 */
Andi Kleena5516432008-07-23 21:27:41 -0700606static pgoff_t vma_hugecache_offset(struct hstate *h,
607 struct vm_area_struct *vma, unsigned long address)
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700608{
Andi Kleena5516432008-07-23 21:27:41 -0700609 return ((address - vma->vm_start) >> huge_page_shift(h)) +
610 (vma->vm_pgoff >> huge_page_order(h));
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700611}
612
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +0900613pgoff_t linear_hugepage_index(struct vm_area_struct *vma,
614 unsigned long address)
615{
616 return vma_hugecache_offset(hstate_vma(vma), vma, address);
617}
618
Andy Whitcroft84afd992008-07-23 21:27:32 -0700619/*
Mel Gorman08fba692009-01-06 14:38:53 -0800620 * Return the size of the pages allocated when backing a VMA. In the majority
621 * cases this will be same size as used by the page table entries.
622 */
623unsigned long vma_kernel_pagesize(struct vm_area_struct *vma)
624{
625 struct hstate *hstate;
626
627 if (!is_vm_hugetlb_page(vma))
628 return PAGE_SIZE;
629
630 hstate = hstate_vma(vma);
631
Wanpeng Li2415cf12013-07-03 15:02:43 -0700632 return 1UL << huge_page_shift(hstate);
Mel Gorman08fba692009-01-06 14:38:53 -0800633}
Joerg Roedelf340ca02009-06-19 15:16:22 +0200634EXPORT_SYMBOL_GPL(vma_kernel_pagesize);
Mel Gorman08fba692009-01-06 14:38:53 -0800635
636/*
Mel Gorman33402892009-01-06 14:38:54 -0800637 * Return the page size being used by the MMU to back a VMA. In the majority
638 * of cases, the page size used by the kernel matches the MMU size. On
639 * architectures where it differs, an architecture-specific version of this
640 * function is required.
641 */
642#ifndef vma_mmu_pagesize
643unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
644{
645 return vma_kernel_pagesize(vma);
646}
647#endif
648
649/*
Andy Whitcroft84afd992008-07-23 21:27:32 -0700650 * Flags for MAP_PRIVATE reservations. These are stored in the bottom
651 * bits of the reservation map pointer, which are always clear due to
652 * alignment.
653 */
654#define HPAGE_RESV_OWNER (1UL << 0)
655#define HPAGE_RESV_UNMAPPED (1UL << 1)
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700656#define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
Andy Whitcroft84afd992008-07-23 21:27:32 -0700657
Mel Gormana1e78772008-07-23 21:27:23 -0700658/*
659 * These helpers are used to track how many pages are reserved for
660 * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
661 * is guaranteed to have their future faults succeed.
662 *
663 * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
664 * the reserve counters are updated with the hugetlb_lock held. It is safe
665 * to reset the VMA at fork() time as it is not in use yet and there is no
666 * chance of the global counters getting corrupted as a result of the values.
Andy Whitcroft84afd992008-07-23 21:27:32 -0700667 *
668 * The private mapping reservation is represented in a subtly different
669 * manner to a shared mapping. A shared mapping has a region map associated
670 * with the underlying file, this region map represents the backing file
671 * pages which have ever had a reservation assigned which this persists even
672 * after the page is instantiated. A private mapping has a region map
673 * associated with the original mmap which is attached to all VMAs which
674 * reference it, this region map represents those offsets which have consumed
675 * reservation ie. where pages have been instantiated.
Mel Gormana1e78772008-07-23 21:27:23 -0700676 */
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700677static unsigned long get_vma_private_data(struct vm_area_struct *vma)
678{
679 return (unsigned long)vma->vm_private_data;
680}
681
682static void set_vma_private_data(struct vm_area_struct *vma,
683 unsigned long value)
684{
685 vma->vm_private_data = (void *)value;
686}
687
Joonsoo Kim9119a412014-04-03 14:47:25 -0700688struct resv_map *resv_map_alloc(void)
Andy Whitcroft84afd992008-07-23 21:27:32 -0700689{
690 struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
Mike Kravetz5e911372015-09-08 15:01:28 -0700691 struct file_region *rg = kmalloc(sizeof(*rg), GFP_KERNEL);
692
693 if (!resv_map || !rg) {
694 kfree(resv_map);
695 kfree(rg);
Andy Whitcroft84afd992008-07-23 21:27:32 -0700696 return NULL;
Mike Kravetz5e911372015-09-08 15:01:28 -0700697 }
Andy Whitcroft84afd992008-07-23 21:27:32 -0700698
699 kref_init(&resv_map->refs);
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700700 spin_lock_init(&resv_map->lock);
Andy Whitcroft84afd992008-07-23 21:27:32 -0700701 INIT_LIST_HEAD(&resv_map->regions);
702
Mike Kravetz5e911372015-09-08 15:01:28 -0700703 resv_map->adds_in_progress = 0;
704
705 INIT_LIST_HEAD(&resv_map->region_cache);
706 list_add(&rg->link, &resv_map->region_cache);
707 resv_map->region_cache_count = 1;
708
Andy Whitcroft84afd992008-07-23 21:27:32 -0700709 return resv_map;
710}
711
Joonsoo Kim9119a412014-04-03 14:47:25 -0700712void resv_map_release(struct kref *ref)
Andy Whitcroft84afd992008-07-23 21:27:32 -0700713{
714 struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
Mike Kravetz5e911372015-09-08 15:01:28 -0700715 struct list_head *head = &resv_map->region_cache;
716 struct file_region *rg, *trg;
Andy Whitcroft84afd992008-07-23 21:27:32 -0700717
718 /* Clear out any active regions before we release the map. */
Mike Kravetzfeba16e2015-09-08 15:01:31 -0700719 region_del(resv_map, 0, LONG_MAX);
Mike Kravetz5e911372015-09-08 15:01:28 -0700720
721 /* ... and any entries left in the cache */
722 list_for_each_entry_safe(rg, trg, head, link) {
723 list_del(&rg->link);
724 kfree(rg);
725 }
726
727 VM_BUG_ON(resv_map->adds_in_progress);
728
Andy Whitcroft84afd992008-07-23 21:27:32 -0700729 kfree(resv_map);
730}
731
Joonsoo Kim4e35f482014-04-03 14:47:30 -0700732static inline struct resv_map *inode_resv_map(struct inode *inode)
733{
734 return inode->i_mapping->private_data;
735}
736
Andy Whitcroft84afd992008-07-23 21:27:32 -0700737static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
Mel Gormana1e78772008-07-23 21:27:23 -0700738{
Sasha Levin81d1b092014-10-09 15:28:10 -0700739 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
Joonsoo Kim4e35f482014-04-03 14:47:30 -0700740 if (vma->vm_flags & VM_MAYSHARE) {
741 struct address_space *mapping = vma->vm_file->f_mapping;
742 struct inode *inode = mapping->host;
743
744 return inode_resv_map(inode);
745
746 } else {
Andy Whitcroft84afd992008-07-23 21:27:32 -0700747 return (struct resv_map *)(get_vma_private_data(vma) &
748 ~HPAGE_RESV_MASK);
Joonsoo Kim4e35f482014-04-03 14:47:30 -0700749 }
Mel Gormana1e78772008-07-23 21:27:23 -0700750}
751
Andy Whitcroft84afd992008-07-23 21:27:32 -0700752static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
Mel Gormana1e78772008-07-23 21:27:23 -0700753{
Sasha Levin81d1b092014-10-09 15:28:10 -0700754 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
755 VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
Mel Gormana1e78772008-07-23 21:27:23 -0700756
Andy Whitcroft84afd992008-07-23 21:27:32 -0700757 set_vma_private_data(vma, (get_vma_private_data(vma) &
758 HPAGE_RESV_MASK) | (unsigned long)map);
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700759}
760
761static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
762{
Sasha Levin81d1b092014-10-09 15:28:10 -0700763 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
764 VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700765
766 set_vma_private_data(vma, get_vma_private_data(vma) | flags);
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700767}
768
769static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
770{
Sasha Levin81d1b092014-10-09 15:28:10 -0700771 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700772
773 return (get_vma_private_data(vma) & flag) != 0;
Mel Gormana1e78772008-07-23 21:27:23 -0700774}
775
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700776/* Reset counters to 0 and clear all HPAGE_RESV_* flags */
Mel Gormana1e78772008-07-23 21:27:23 -0700777void reset_vma_resv_huge_pages(struct vm_area_struct *vma)
778{
Sasha Levin81d1b092014-10-09 15:28:10 -0700779 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
Mel Gormanf83a2752009-05-28 14:34:40 -0700780 if (!(vma->vm_flags & VM_MAYSHARE))
Mel Gormana1e78772008-07-23 21:27:23 -0700781 vma->vm_private_data = (void *)0;
782}
783
784/* Returns true if the VMA has associated reserve pages */
Nicholas Krause559ec2f2015-09-04 15:48:27 -0700785static bool vma_has_reserves(struct vm_area_struct *vma, long chg)
Mel Gormana1e78772008-07-23 21:27:23 -0700786{
Joonsoo Kimaf0ed732013-09-11 14:21:18 -0700787 if (vma->vm_flags & VM_NORESERVE) {
788 /*
789 * This address is already reserved by other process(chg == 0),
790 * so, we should decrement reserved count. Without decrementing,
791 * reserve count remains after releasing inode, because this
792 * allocated page will go into page cache and is regarded as
793 * coming from reserved pool in releasing step. Currently, we
794 * don't have any other solution to deal with this situation
795 * properly, so add work-around here.
796 */
797 if (vma->vm_flags & VM_MAYSHARE && chg == 0)
Nicholas Krause559ec2f2015-09-04 15:48:27 -0700798 return true;
Joonsoo Kimaf0ed732013-09-11 14:21:18 -0700799 else
Nicholas Krause559ec2f2015-09-04 15:48:27 -0700800 return false;
Joonsoo Kimaf0ed732013-09-11 14:21:18 -0700801 }
Joonsoo Kima63884e2013-09-11 14:21:07 -0700802
803 /* Shared mappings always use reserves */
Mike Kravetz1fb1b0e2015-09-08 15:01:44 -0700804 if (vma->vm_flags & VM_MAYSHARE) {
805 /*
806 * We know VM_NORESERVE is not set. Therefore, there SHOULD
807 * be a region map for all pages. The only situation where
808 * there is no region map is if a hole was punched via
809 * fallocate. In this case, there really are no reverves to
810 * use. This situation is indicated if chg != 0.
811 */
812 if (chg)
813 return false;
814 else
815 return true;
816 }
Joonsoo Kima63884e2013-09-11 14:21:07 -0700817
818 /*
819 * Only the process that called mmap() has reserves for
820 * private mappings.
821 */
Mel Gorman7f09ca52008-07-23 21:27:58 -0700822 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER))
Nicholas Krause559ec2f2015-09-04 15:48:27 -0700823 return true;
Joonsoo Kima63884e2013-09-11 14:21:07 -0700824
Nicholas Krause559ec2f2015-09-04 15:48:27 -0700825 return false;
Mel Gormana1e78772008-07-23 21:27:23 -0700826}
827
Andi Kleena5516432008-07-23 21:27:41 -0700828static void enqueue_huge_page(struct hstate *h, struct page *page)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700829{
830 int nid = page_to_nid(page);
Aneesh Kumar K.V0edaecf2012-07-31 16:42:07 -0700831 list_move(&page->lru, &h->hugepage_freelists[nid]);
Andi Kleena5516432008-07-23 21:27:41 -0700832 h->free_huge_pages++;
833 h->free_huge_pages_node[nid]++;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700834}
835
Naoya Horiguchibf50bab2010-09-08 10:19:33 +0900836static struct page *dequeue_huge_page_node(struct hstate *h, int nid)
837{
838 struct page *page;
839
Naoya Horiguchic8721bb2013-09-11 14:22:09 -0700840 list_for_each_entry(page, &h->hugepage_freelists[nid], lru)
841 if (!is_migrate_isolate_page(page))
842 break;
843 /*
844 * if 'non-isolated free hugepage' not found on the list,
845 * the allocation fails.
846 */
847 if (&h->hugepage_freelists[nid] == &page->lru)
Naoya Horiguchibf50bab2010-09-08 10:19:33 +0900848 return NULL;
Aneesh Kumar K.V0edaecf2012-07-31 16:42:07 -0700849 list_move(&page->lru, &h->hugepage_activelist);
Naoya Horiguchia9869b82010-09-08 10:19:37 +0900850 set_page_refcounted(page);
Naoya Horiguchibf50bab2010-09-08 10:19:33 +0900851 h->free_huge_pages--;
852 h->free_huge_pages_node[nid]--;
853 return page;
854}
855
Naoya Horiguchi86cdb462013-09-11 14:22:13 -0700856/* Movability of hugepages depends on migration support. */
857static inline gfp_t htlb_alloc_mask(struct hstate *h)
858{
Naoya Horiguchi100873d2014-06-04 16:10:56 -0700859 if (hugepages_treat_as_movable || hugepage_migration_supported(h))
Naoya Horiguchi86cdb462013-09-11 14:22:13 -0700860 return GFP_HIGHUSER_MOVABLE;
861 else
862 return GFP_HIGHUSER;
863}
864
Andi Kleena5516432008-07-23 21:27:41 -0700865static struct page *dequeue_huge_page_vma(struct hstate *h,
866 struct vm_area_struct *vma,
Joonsoo Kimaf0ed732013-09-11 14:21:18 -0700867 unsigned long address, int avoid_reserve,
868 long chg)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700869{
Konstantin Khlebnikovb1c12cb2012-04-25 16:01:46 -0700870 struct page *page = NULL;
Lee Schermerhorn480eccf2007-09-18 22:46:47 -0700871 struct mempolicy *mpol;
Mel Gorman19770b32008-04-28 02:12:18 -0700872 nodemask_t *nodemask;
Miao Xiec0ff7452010-05-24 14:32:08 -0700873 struct zonelist *zonelist;
Mel Gormandd1a2392008-04-28 02:12:17 -0700874 struct zone *zone;
875 struct zoneref *z;
Mel Gormancc9a6c82012-03-21 16:34:11 -0700876 unsigned int cpuset_mems_cookie;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700877
Mel Gormana1e78772008-07-23 21:27:23 -0700878 /*
879 * A child process with MAP_PRIVATE mappings created by their parent
880 * have no page reserves. This check ensures that reservations are
881 * not "stolen". The child may still get SIGKILLed
882 */
Joonsoo Kimaf0ed732013-09-11 14:21:18 -0700883 if (!vma_has_reserves(vma, chg) &&
Andi Kleena5516432008-07-23 21:27:41 -0700884 h->free_huge_pages - h->resv_huge_pages == 0)
Miao Xiec0ff7452010-05-24 14:32:08 -0700885 goto err;
Mel Gormana1e78772008-07-23 21:27:23 -0700886
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700887 /* If reserves cannot be used, ensure enough pages are in the pool */
Andi Kleena5516432008-07-23 21:27:41 -0700888 if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
Justin P. Mattock6eab04a2011-04-08 19:49:08 -0700889 goto err;
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700890
Joonsoo Kim9966c4b2013-09-11 14:20:50 -0700891retry_cpuset:
Mel Gormand26914d2014-04-03 14:47:24 -0700892 cpuset_mems_cookie = read_mems_allowed_begin();
Joonsoo Kim9966c4b2013-09-11 14:20:50 -0700893 zonelist = huge_zonelist(vma, address,
Naoya Horiguchi86cdb462013-09-11 14:22:13 -0700894 htlb_alloc_mask(h), &mpol, &nodemask);
Joonsoo Kim9966c4b2013-09-11 14:20:50 -0700895
Mel Gorman19770b32008-04-28 02:12:18 -0700896 for_each_zone_zonelist_nodemask(zone, z, zonelist,
897 MAX_NR_ZONES - 1, nodemask) {
Vladimir Davydov344736f2014-10-20 15:50:30 +0400898 if (cpuset_zone_allowed(zone, htlb_alloc_mask(h))) {
Naoya Horiguchibf50bab2010-09-08 10:19:33 +0900899 page = dequeue_huge_page_node(h, zone_to_nid(zone));
900 if (page) {
Joonsoo Kimaf0ed732013-09-11 14:21:18 -0700901 if (avoid_reserve)
902 break;
903 if (!vma_has_reserves(vma, chg))
904 break;
905
Joonsoo Kim07443a82013-09-11 14:21:58 -0700906 SetPagePrivate(page);
Joonsoo Kimaf0ed732013-09-11 14:21:18 -0700907 h->resv_huge_pages--;
Naoya Horiguchibf50bab2010-09-08 10:19:33 +0900908 break;
909 }
Andrew Morton3abf7af2007-07-19 01:49:08 -0700910 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700911 }
Mel Gormancc9a6c82012-03-21 16:34:11 -0700912
913 mpol_cond_put(mpol);
Mel Gormand26914d2014-04-03 14:47:24 -0700914 if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie)))
Mel Gormancc9a6c82012-03-21 16:34:11 -0700915 goto retry_cpuset;
916 return page;
917
Miao Xiec0ff7452010-05-24 14:32:08 -0700918err:
Mel Gormancc9a6c82012-03-21 16:34:11 -0700919 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700920}
921
Luiz Capitulino1cac6f22014-06-04 16:07:11 -0700922/*
923 * common helper functions for hstate_next_node_to_{alloc|free}.
924 * We may have allocated or freed a huge page based on a different
925 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
926 * be outside of *nodes_allowed. Ensure that we use an allowed
927 * node for alloc or free.
928 */
929static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
930{
931 nid = next_node(nid, *nodes_allowed);
932 if (nid == MAX_NUMNODES)
933 nid = first_node(*nodes_allowed);
934 VM_BUG_ON(nid >= MAX_NUMNODES);
935
936 return nid;
937}
938
939static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
940{
941 if (!node_isset(nid, *nodes_allowed))
942 nid = next_node_allowed(nid, nodes_allowed);
943 return nid;
944}
945
946/*
947 * returns the previously saved node ["this node"] from which to
948 * allocate a persistent huge page for the pool and advance the
949 * next node from which to allocate, handling wrap at end of node
950 * mask.
951 */
952static int hstate_next_node_to_alloc(struct hstate *h,
953 nodemask_t *nodes_allowed)
954{
955 int nid;
956
957 VM_BUG_ON(!nodes_allowed);
958
959 nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
960 h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
961
962 return nid;
963}
964
965/*
966 * helper for free_pool_huge_page() - return the previously saved
967 * node ["this node"] from which to free a huge page. Advance the
968 * next node id whether or not we find a free huge page to free so
969 * that the next attempt to free addresses the next node.
970 */
971static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
972{
973 int nid;
974
975 VM_BUG_ON(!nodes_allowed);
976
977 nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
978 h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
979
980 return nid;
981}
982
983#define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \
984 for (nr_nodes = nodes_weight(*mask); \
985 nr_nodes > 0 && \
986 ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \
987 nr_nodes--)
988
989#define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \
990 for (nr_nodes = nodes_weight(*mask); \
991 nr_nodes > 0 && \
992 ((node = hstate_next_node_to_free(hs, mask)) || 1); \
993 nr_nodes--)
994
Luiz Capitulino944d9fe2014-06-04 16:07:13 -0700995#if defined(CONFIG_CMA) && defined(CONFIG_X86_64)
996static void destroy_compound_gigantic_page(struct page *page,
997 unsigned long order)
998{
999 int i;
1000 int nr_pages = 1 << order;
1001 struct page *p = page + 1;
1002
1003 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
1004 __ClearPageTail(p);
1005 set_page_refcounted(p);
1006 p->first_page = NULL;
1007 }
1008
1009 set_compound_order(page, 0);
1010 __ClearPageHead(page);
1011}
1012
1013static void free_gigantic_page(struct page *page, unsigned order)
1014{
1015 free_contig_range(page_to_pfn(page), 1 << order);
1016}
1017
1018static int __alloc_gigantic_page(unsigned long start_pfn,
1019 unsigned long nr_pages)
1020{
1021 unsigned long end_pfn = start_pfn + nr_pages;
1022 return alloc_contig_range(start_pfn, end_pfn, MIGRATE_MOVABLE);
1023}
1024
1025static bool pfn_range_valid_gigantic(unsigned long start_pfn,
1026 unsigned long nr_pages)
1027{
1028 unsigned long i, end_pfn = start_pfn + nr_pages;
1029 struct page *page;
1030
1031 for (i = start_pfn; i < end_pfn; i++) {
1032 if (!pfn_valid(i))
1033 return false;
1034
1035 page = pfn_to_page(i);
1036
1037 if (PageReserved(page))
1038 return false;
1039
1040 if (page_count(page) > 0)
1041 return false;
1042
1043 if (PageHuge(page))
1044 return false;
1045 }
1046
1047 return true;
1048}
1049
1050static bool zone_spans_last_pfn(const struct zone *zone,
1051 unsigned long start_pfn, unsigned long nr_pages)
1052{
1053 unsigned long last_pfn = start_pfn + nr_pages - 1;
1054 return zone_spans_pfn(zone, last_pfn);
1055}
1056
1057static struct page *alloc_gigantic_page(int nid, unsigned order)
1058{
1059 unsigned long nr_pages = 1 << order;
1060 unsigned long ret, pfn, flags;
1061 struct zone *z;
1062
1063 z = NODE_DATA(nid)->node_zones;
1064 for (; z - NODE_DATA(nid)->node_zones < MAX_NR_ZONES; z++) {
1065 spin_lock_irqsave(&z->lock, flags);
1066
1067 pfn = ALIGN(z->zone_start_pfn, nr_pages);
1068 while (zone_spans_last_pfn(z, pfn, nr_pages)) {
1069 if (pfn_range_valid_gigantic(pfn, nr_pages)) {
1070 /*
1071 * We release the zone lock here because
1072 * alloc_contig_range() will also lock the zone
1073 * at some point. If there's an allocation
1074 * spinning on this lock, it may win the race
1075 * and cause alloc_contig_range() to fail...
1076 */
1077 spin_unlock_irqrestore(&z->lock, flags);
1078 ret = __alloc_gigantic_page(pfn, nr_pages);
1079 if (!ret)
1080 return pfn_to_page(pfn);
1081 spin_lock_irqsave(&z->lock, flags);
1082 }
1083 pfn += nr_pages;
1084 }
1085
1086 spin_unlock_irqrestore(&z->lock, flags);
1087 }
1088
1089 return NULL;
1090}
1091
1092static void prep_new_huge_page(struct hstate *h, struct page *page, int nid);
1093static void prep_compound_gigantic_page(struct page *page, unsigned long order);
1094
1095static struct page *alloc_fresh_gigantic_page_node(struct hstate *h, int nid)
1096{
1097 struct page *page;
1098
1099 page = alloc_gigantic_page(nid, huge_page_order(h));
1100 if (page) {
1101 prep_compound_gigantic_page(page, huge_page_order(h));
1102 prep_new_huge_page(h, page, nid);
1103 }
1104
1105 return page;
1106}
1107
1108static int alloc_fresh_gigantic_page(struct hstate *h,
1109 nodemask_t *nodes_allowed)
1110{
1111 struct page *page = NULL;
1112 int nr_nodes, node;
1113
1114 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
1115 page = alloc_fresh_gigantic_page_node(h, node);
1116 if (page)
1117 return 1;
1118 }
1119
1120 return 0;
1121}
1122
1123static inline bool gigantic_page_supported(void) { return true; }
1124#else
1125static inline bool gigantic_page_supported(void) { return false; }
1126static inline void free_gigantic_page(struct page *page, unsigned order) { }
1127static inline void destroy_compound_gigantic_page(struct page *page,
1128 unsigned long order) { }
1129static inline int alloc_fresh_gigantic_page(struct hstate *h,
1130 nodemask_t *nodes_allowed) { return 0; }
1131#endif
1132
Andi Kleena5516432008-07-23 21:27:41 -07001133static void update_and_free_page(struct hstate *h, struct page *page)
Adam Litke6af2acb2007-10-16 01:26:16 -07001134{
1135 int i;
Andi Kleena5516432008-07-23 21:27:41 -07001136
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001137 if (hstate_is_gigantic(h) && !gigantic_page_supported())
1138 return;
Andy Whitcroft18229df2008-11-06 12:53:27 -08001139
Andi Kleena5516432008-07-23 21:27:41 -07001140 h->nr_huge_pages--;
1141 h->nr_huge_pages_node[page_to_nid(page)]--;
1142 for (i = 0; i < pages_per_huge_page(h); i++) {
Chris Forbes32f84522011-07-25 17:12:14 -07001143 page[i].flags &= ~(1 << PG_locked | 1 << PG_error |
1144 1 << PG_referenced | 1 << PG_dirty |
Luiz Capitulinoa7407a22014-06-04 16:07:09 -07001145 1 << PG_active | 1 << PG_private |
1146 1 << PG_writeback);
Adam Litke6af2acb2007-10-16 01:26:16 -07001147 }
Sasha Levin309381fea2014-01-23 15:52:54 -08001148 VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page);
Adam Litke6af2acb2007-10-16 01:26:16 -07001149 set_compound_page_dtor(page, NULL);
1150 set_page_refcounted(page);
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001151 if (hstate_is_gigantic(h)) {
1152 destroy_compound_gigantic_page(page, huge_page_order(h));
1153 free_gigantic_page(page, huge_page_order(h));
1154 } else {
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001155 __free_pages(page, huge_page_order(h));
1156 }
Adam Litke6af2acb2007-10-16 01:26:16 -07001157}
1158
Andi Kleene5ff2152008-07-23 21:27:42 -07001159struct hstate *size_to_hstate(unsigned long size)
1160{
1161 struct hstate *h;
1162
1163 for_each_hstate(h) {
1164 if (huge_page_size(h) == size)
1165 return h;
1166 }
1167 return NULL;
1168}
1169
Naoya Horiguchibcc54222015-04-15 16:14:38 -07001170/*
1171 * Test to determine whether the hugepage is "active/in-use" (i.e. being linked
1172 * to hstate->hugepage_activelist.)
1173 *
1174 * This function can be called for tail pages, but never returns true for them.
1175 */
1176bool page_huge_active(struct page *page)
1177{
1178 VM_BUG_ON_PAGE(!PageHuge(page), page);
1179 return PageHead(page) && PagePrivate(&page[1]);
1180}
1181
1182/* never called for tail page */
1183static void set_page_huge_active(struct page *page)
1184{
1185 VM_BUG_ON_PAGE(!PageHeadHuge(page), page);
1186 SetPagePrivate(&page[1]);
1187}
1188
1189static void clear_page_huge_active(struct page *page)
1190{
1191 VM_BUG_ON_PAGE(!PageHeadHuge(page), page);
1192 ClearPagePrivate(&page[1]);
1193}
1194
Atsushi Kumagai8f1d26d2014-07-30 16:08:39 -07001195void free_huge_page(struct page *page)
David Gibson27a85ef2006-03-22 00:08:56 -08001196{
Andi Kleena5516432008-07-23 21:27:41 -07001197 /*
1198 * Can't pass hstate in here because it is called from the
1199 * compound page destructor.
1200 */
Andi Kleene5ff2152008-07-23 21:27:42 -07001201 struct hstate *h = page_hstate(page);
Adam Litke7893d1d2007-10-16 01:26:18 -07001202 int nid = page_to_nid(page);
David Gibson90481622012-03-21 16:34:12 -07001203 struct hugepage_subpool *spool =
1204 (struct hugepage_subpool *)page_private(page);
Joonsoo Kim07443a82013-09-11 14:21:58 -07001205 bool restore_reserve;
David Gibson27a85ef2006-03-22 00:08:56 -08001206
Andy Whitcrofte5df70a2008-02-23 15:23:32 -08001207 set_page_private(page, 0);
Mel Gorman23be7462010-04-23 13:17:56 -04001208 page->mapping = NULL;
Adam Litke7893d1d2007-10-16 01:26:18 -07001209 BUG_ON(page_count(page));
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09001210 BUG_ON(page_mapcount(page));
Joonsoo Kim07443a82013-09-11 14:21:58 -07001211 restore_reserve = PagePrivate(page);
Joonsoo Kim16c794b2013-10-16 13:46:48 -07001212 ClearPagePrivate(page);
David Gibson27a85ef2006-03-22 00:08:56 -08001213
Mike Kravetz1c5ecae2015-04-15 16:13:39 -07001214 /*
1215 * A return code of zero implies that the subpool will be under its
1216 * minimum size if the reservation is not restored after page is free.
1217 * Therefore, force restore_reserve operation.
1218 */
1219 if (hugepage_subpool_put_pages(spool, 1) == 0)
1220 restore_reserve = true;
1221
David Gibson27a85ef2006-03-22 00:08:56 -08001222 spin_lock(&hugetlb_lock);
Naoya Horiguchibcc54222015-04-15 16:14:38 -07001223 clear_page_huge_active(page);
Aneesh Kumar K.V6d76dcf2012-07-31 16:42:18 -07001224 hugetlb_cgroup_uncharge_page(hstate_index(h),
1225 pages_per_huge_page(h), page);
Joonsoo Kim07443a82013-09-11 14:21:58 -07001226 if (restore_reserve)
1227 h->resv_huge_pages++;
1228
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001229 if (h->surplus_huge_pages_node[nid]) {
Aneesh Kumar K.V0edaecf2012-07-31 16:42:07 -07001230 /* remove the page from active list */
1231 list_del(&page->lru);
Andi Kleena5516432008-07-23 21:27:41 -07001232 update_and_free_page(h, page);
1233 h->surplus_huge_pages--;
1234 h->surplus_huge_pages_node[nid]--;
Adam Litke7893d1d2007-10-16 01:26:18 -07001235 } else {
Will Deacon5d3a5512012-10-08 16:29:32 -07001236 arch_clear_hugepage_flags(page);
Andi Kleena5516432008-07-23 21:27:41 -07001237 enqueue_huge_page(h, page);
Adam Litke7893d1d2007-10-16 01:26:18 -07001238 }
David Gibson27a85ef2006-03-22 00:08:56 -08001239 spin_unlock(&hugetlb_lock);
1240}
1241
Andi Kleena5516432008-07-23 21:27:41 -07001242static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
Andi Kleenb7ba30c2008-07-23 21:27:40 -07001243{
Aneesh Kumar K.V0edaecf2012-07-31 16:42:07 -07001244 INIT_LIST_HEAD(&page->lru);
Andi Kleenb7ba30c2008-07-23 21:27:40 -07001245 set_compound_page_dtor(page, free_huge_page);
1246 spin_lock(&hugetlb_lock);
Aneesh Kumar K.V9dd540e2012-07-31 16:42:15 -07001247 set_hugetlb_cgroup(page, NULL);
Andi Kleena5516432008-07-23 21:27:41 -07001248 h->nr_huge_pages++;
1249 h->nr_huge_pages_node[nid]++;
Andi Kleenb7ba30c2008-07-23 21:27:40 -07001250 spin_unlock(&hugetlb_lock);
1251 put_page(page); /* free it into the hugepage allocator */
1252}
1253
Luiz Capitulino2906dd52014-06-04 16:07:06 -07001254static void prep_compound_gigantic_page(struct page *page, unsigned long order)
Wu Fengguang20a03072009-06-16 15:32:22 -07001255{
1256 int i;
1257 int nr_pages = 1 << order;
1258 struct page *p = page + 1;
1259
1260 /* we rely on prep_new_huge_page to set the destructor */
1261 set_compound_order(page, order);
1262 __SetPageHead(page);
Andrea Arcangelief5a22b2013-10-16 13:46:56 -07001263 __ClearPageReserved(page);
Wu Fengguang20a03072009-06-16 15:32:22 -07001264 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
Andrea Arcangelief5a22b2013-10-16 13:46:56 -07001265 /*
1266 * For gigantic hugepages allocated through bootmem at
1267 * boot, it's safer to be consistent with the not-gigantic
1268 * hugepages and clear the PG_reserved bit from all tail pages
1269 * too. Otherwse drivers using get_user_pages() to access tail
1270 * pages may get the reference counting wrong if they see
1271 * PG_reserved set on a tail page (despite the head page not
1272 * having PG_reserved set). Enforcing this consistency between
1273 * head and tail pages allows drivers to optimize away a check
1274 * on the head page when they need know if put_page() is needed
1275 * after get_user_pages().
1276 */
1277 __ClearPageReserved(p);
Youquan Song58a84aa2011-12-08 14:34:18 -08001278 set_page_count(p, 0);
Wu Fengguang20a03072009-06-16 15:32:22 -07001279 p->first_page = page;
David Rientjes44fc8052015-03-12 16:25:54 -07001280 /* Make sure p->first_page is always valid for PageTail() */
1281 smp_wmb();
1282 __SetPageTail(p);
Wu Fengguang20a03072009-06-16 15:32:22 -07001283 }
1284}
1285
Andrew Morton77959122012-10-08 16:34:11 -07001286/*
1287 * PageHuge() only returns true for hugetlbfs pages, but not for normal or
1288 * transparent huge pages. See the PageTransHuge() documentation for more
1289 * details.
1290 */
Wu Fengguang20a03072009-06-16 15:32:22 -07001291int PageHuge(struct page *page)
1292{
Wu Fengguang20a03072009-06-16 15:32:22 -07001293 if (!PageCompound(page))
1294 return 0;
1295
1296 page = compound_head(page);
Andrew Morton758f66a2014-01-21 15:48:57 -08001297 return get_compound_page_dtor(page) == free_huge_page;
Wu Fengguang20a03072009-06-16 15:32:22 -07001298}
Naoya Horiguchi43131e12010-05-28 09:29:22 +09001299EXPORT_SYMBOL_GPL(PageHuge);
1300
Andrea Arcangeli27c73ae2013-11-21 14:32:02 -08001301/*
1302 * PageHeadHuge() only returns true for hugetlbfs head page, but not for
1303 * normal or transparent huge pages.
1304 */
1305int PageHeadHuge(struct page *page_head)
1306{
Andrea Arcangeli27c73ae2013-11-21 14:32:02 -08001307 if (!PageHead(page_head))
1308 return 0;
1309
Andrew Morton758f66a2014-01-21 15:48:57 -08001310 return get_compound_page_dtor(page_head) == free_huge_page;
Andrea Arcangeli27c73ae2013-11-21 14:32:02 -08001311}
Andrea Arcangeli27c73ae2013-11-21 14:32:02 -08001312
Zhang Yi13d60f42013-06-25 21:19:31 +08001313pgoff_t __basepage_index(struct page *page)
1314{
1315 struct page *page_head = compound_head(page);
1316 pgoff_t index = page_index(page_head);
1317 unsigned long compound_idx;
1318
1319 if (!PageHuge(page_head))
1320 return page_index(page);
1321
1322 if (compound_order(page_head) >= MAX_ORDER)
1323 compound_idx = page_to_pfn(page) - page_to_pfn(page_head);
1324 else
1325 compound_idx = page - page_head;
1326
1327 return (index << compound_order(page_head)) + compound_idx;
1328}
1329
Andi Kleena5516432008-07-23 21:27:41 -07001330static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001331{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001332 struct page *page;
Joe Jinf96efd52007-07-15 23:38:12 -07001333
Vlastimil Babka96db8002015-09-08 15:03:50 -07001334 page = __alloc_pages_node(nid,
Naoya Horiguchi86cdb462013-09-11 14:22:13 -07001335 htlb_alloc_mask(h)|__GFP_COMP|__GFP_THISNODE|
Nishanth Aravamudan551883a2008-04-29 00:58:26 -07001336 __GFP_REPEAT|__GFP_NOWARN,
Andi Kleena5516432008-07-23 21:27:41 -07001337 huge_page_order(h));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001338 if (page) {
Andi Kleena5516432008-07-23 21:27:41 -07001339 prep_new_huge_page(h, page, nid);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001340 }
Nishanth Aravamudan63b46132007-10-16 01:26:24 -07001341
1342 return page;
1343}
1344
Joonsoo Kimb2261022013-09-11 14:21:00 -07001345static int alloc_fresh_huge_page(struct hstate *h, nodemask_t *nodes_allowed)
1346{
1347 struct page *page;
1348 int nr_nodes, node;
1349 int ret = 0;
1350
1351 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
1352 page = alloc_fresh_huge_page_node(h, node);
1353 if (page) {
1354 ret = 1;
1355 break;
1356 }
1357 }
1358
1359 if (ret)
1360 count_vm_event(HTLB_BUDDY_PGALLOC);
1361 else
1362 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
1363
1364 return ret;
1365}
1366
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001367/*
1368 * Free huge page from pool from next node to free.
1369 * Attempt to keep persistent huge pages more or less
1370 * balanced over allowed nodes.
1371 * Called with hugetlb_lock locked.
1372 */
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001373static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
1374 bool acct_surplus)
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001375{
Joonsoo Kimb2261022013-09-11 14:21:00 -07001376 int nr_nodes, node;
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001377 int ret = 0;
1378
Joonsoo Kimb2261022013-09-11 14:21:00 -07001379 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
Lee Schermerhorn685f3452009-09-21 17:01:23 -07001380 /*
1381 * If we're returning unused surplus pages, only examine
1382 * nodes with surplus pages.
1383 */
Joonsoo Kimb2261022013-09-11 14:21:00 -07001384 if ((!acct_surplus || h->surplus_huge_pages_node[node]) &&
1385 !list_empty(&h->hugepage_freelists[node])) {
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001386 struct page *page =
Joonsoo Kimb2261022013-09-11 14:21:00 -07001387 list_entry(h->hugepage_freelists[node].next,
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001388 struct page, lru);
1389 list_del(&page->lru);
1390 h->free_huge_pages--;
Joonsoo Kimb2261022013-09-11 14:21:00 -07001391 h->free_huge_pages_node[node]--;
Lee Schermerhorn685f3452009-09-21 17:01:23 -07001392 if (acct_surplus) {
1393 h->surplus_huge_pages--;
Joonsoo Kimb2261022013-09-11 14:21:00 -07001394 h->surplus_huge_pages_node[node]--;
Lee Schermerhorn685f3452009-09-21 17:01:23 -07001395 }
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001396 update_and_free_page(h, page);
1397 ret = 1;
Lee Schermerhorn9a76db02009-12-14 17:58:15 -08001398 break;
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001399 }
Joonsoo Kimb2261022013-09-11 14:21:00 -07001400 }
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001401
1402 return ret;
1403}
1404
Naoya Horiguchic8721bb2013-09-11 14:22:09 -07001405/*
1406 * Dissolve a given free hugepage into free buddy pages. This function does
1407 * nothing for in-use (including surplus) hugepages.
1408 */
1409static void dissolve_free_huge_page(struct page *page)
1410{
1411 spin_lock(&hugetlb_lock);
1412 if (PageHuge(page) && !page_count(page)) {
1413 struct hstate *h = page_hstate(page);
1414 int nid = page_to_nid(page);
1415 list_del(&page->lru);
1416 h->free_huge_pages--;
1417 h->free_huge_pages_node[nid]--;
1418 update_and_free_page(h, page);
1419 }
1420 spin_unlock(&hugetlb_lock);
1421}
1422
1423/*
1424 * Dissolve free hugepages in a given pfn range. Used by memory hotplug to
1425 * make specified memory blocks removable from the system.
1426 * Note that start_pfn should aligned with (minimum) hugepage size.
1427 */
1428void dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn)
1429{
Naoya Horiguchic8721bb2013-09-11 14:22:09 -07001430 unsigned long pfn;
Naoya Horiguchic8721bb2013-09-11 14:22:09 -07001431
Li Zhongd0177632014-08-06 16:07:56 -07001432 if (!hugepages_supported())
1433 return;
1434
Naoya Horiguchi641844f2015-06-24 16:56:59 -07001435 VM_BUG_ON(!IS_ALIGNED(start_pfn, 1 << minimum_order));
1436 for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << minimum_order)
Naoya Horiguchic8721bb2013-09-11 14:22:09 -07001437 dissolve_free_huge_page(pfn_to_page(pfn));
1438}
1439
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001440static struct page *alloc_buddy_huge_page(struct hstate *h, int nid)
Adam Litke7893d1d2007-10-16 01:26:18 -07001441{
1442 struct page *page;
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001443 unsigned int r_nid;
Adam Litke7893d1d2007-10-16 01:26:18 -07001444
Luiz Capitulinobae7f4a2014-06-04 16:07:08 -07001445 if (hstate_is_gigantic(h))
Andi Kleenaa888a72008-07-23 21:27:47 -07001446 return NULL;
1447
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001448 /*
1449 * Assume we will successfully allocate the surplus page to
1450 * prevent racing processes from causing the surplus to exceed
1451 * overcommit
1452 *
1453 * This however introduces a different race, where a process B
1454 * tries to grow the static hugepage pool while alloc_pages() is
1455 * called by process A. B will only examine the per-node
1456 * counters in determining if surplus huge pages can be
1457 * converted to normal huge pages in adjust_pool_surplus(). A
1458 * won't be able to increment the per-node counter, until the
1459 * lock is dropped by B, but B doesn't drop hugetlb_lock until
1460 * no more huge pages can be converted from surplus to normal
1461 * state (and doesn't try to convert again). Thus, we have a
1462 * case where a surplus huge page exists, the pool is grown, and
1463 * the surplus huge page still exists after, even though it
1464 * should just have been converted to a normal huge page. This
1465 * does not leak memory, though, as the hugepage will be freed
1466 * once it is out of use. It also does not allow the counters to
1467 * go out of whack in adjust_pool_surplus() as we don't modify
1468 * the node values until we've gotten the hugepage and only the
1469 * per-node value is checked there.
1470 */
1471 spin_lock(&hugetlb_lock);
Andi Kleena5516432008-07-23 21:27:41 -07001472 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001473 spin_unlock(&hugetlb_lock);
1474 return NULL;
1475 } else {
Andi Kleena5516432008-07-23 21:27:41 -07001476 h->nr_huge_pages++;
1477 h->surplus_huge_pages++;
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001478 }
1479 spin_unlock(&hugetlb_lock);
1480
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001481 if (nid == NUMA_NO_NODE)
Naoya Horiguchi86cdb462013-09-11 14:22:13 -07001482 page = alloc_pages(htlb_alloc_mask(h)|__GFP_COMP|
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001483 __GFP_REPEAT|__GFP_NOWARN,
1484 huge_page_order(h));
1485 else
Vlastimil Babka96db8002015-09-08 15:03:50 -07001486 page = __alloc_pages_node(nid,
Naoya Horiguchi86cdb462013-09-11 14:22:13 -07001487 htlb_alloc_mask(h)|__GFP_COMP|__GFP_THISNODE|
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001488 __GFP_REPEAT|__GFP_NOWARN, huge_page_order(h));
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001489
1490 spin_lock(&hugetlb_lock);
Adam Litke7893d1d2007-10-16 01:26:18 -07001491 if (page) {
Aneesh Kumar K.V0edaecf2012-07-31 16:42:07 -07001492 INIT_LIST_HEAD(&page->lru);
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001493 r_nid = page_to_nid(page);
Adam Litke7893d1d2007-10-16 01:26:18 -07001494 set_compound_page_dtor(page, free_huge_page);
Aneesh Kumar K.V9dd540e2012-07-31 16:42:15 -07001495 set_hugetlb_cgroup(page, NULL);
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001496 /*
1497 * We incremented the global counters already
1498 */
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001499 h->nr_huge_pages_node[r_nid]++;
1500 h->surplus_huge_pages_node[r_nid]++;
Adam Litke3b116302008-04-28 02:13:06 -07001501 __count_vm_event(HTLB_BUDDY_PGALLOC);
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001502 } else {
Andi Kleena5516432008-07-23 21:27:41 -07001503 h->nr_huge_pages--;
1504 h->surplus_huge_pages--;
Adam Litke3b116302008-04-28 02:13:06 -07001505 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
Adam Litke7893d1d2007-10-16 01:26:18 -07001506 }
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001507 spin_unlock(&hugetlb_lock);
Adam Litke7893d1d2007-10-16 01:26:18 -07001508
1509 return page;
1510}
1511
Adam Litkee4e574b2007-10-16 01:26:19 -07001512/*
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001513 * This allocation function is useful in the context where vma is irrelevant.
1514 * E.g. soft-offlining uses this function because it only cares physical
1515 * address of error page.
1516 */
1517struct page *alloc_huge_page_node(struct hstate *h, int nid)
1518{
Joonsoo Kim4ef91842013-09-11 14:21:51 -07001519 struct page *page = NULL;
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001520
1521 spin_lock(&hugetlb_lock);
Joonsoo Kim4ef91842013-09-11 14:21:51 -07001522 if (h->free_huge_pages - h->resv_huge_pages > 0)
1523 page = dequeue_huge_page_node(h, nid);
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001524 spin_unlock(&hugetlb_lock);
1525
Aneesh Kumar K.V94ae8ba2012-07-31 16:42:35 -07001526 if (!page)
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001527 page = alloc_buddy_huge_page(h, nid);
1528
1529 return page;
1530}
1531
1532/*
Lucas De Marchi25985ed2011-03-30 22:57:33 -03001533 * Increase the hugetlb pool such that it can accommodate a reservation
Adam Litkee4e574b2007-10-16 01:26:19 -07001534 * of size 'delta'.
1535 */
Andi Kleena5516432008-07-23 21:27:41 -07001536static int gather_surplus_pages(struct hstate *h, int delta)
Adam Litkee4e574b2007-10-16 01:26:19 -07001537{
1538 struct list_head surplus_list;
1539 struct page *page, *tmp;
1540 int ret, i;
1541 int needed, allocated;
Hillf Danton28073b02012-03-21 16:34:00 -07001542 bool alloc_ok = true;
Adam Litkee4e574b2007-10-16 01:26:19 -07001543
Andi Kleena5516432008-07-23 21:27:41 -07001544 needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
Adam Litkeac09b3a2008-03-04 14:29:38 -08001545 if (needed <= 0) {
Andi Kleena5516432008-07-23 21:27:41 -07001546 h->resv_huge_pages += delta;
Adam Litkee4e574b2007-10-16 01:26:19 -07001547 return 0;
Adam Litkeac09b3a2008-03-04 14:29:38 -08001548 }
Adam Litkee4e574b2007-10-16 01:26:19 -07001549
1550 allocated = 0;
1551 INIT_LIST_HEAD(&surplus_list);
1552
1553 ret = -ENOMEM;
1554retry:
1555 spin_unlock(&hugetlb_lock);
1556 for (i = 0; i < needed; i++) {
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001557 page = alloc_buddy_huge_page(h, NUMA_NO_NODE);
Hillf Danton28073b02012-03-21 16:34:00 -07001558 if (!page) {
1559 alloc_ok = false;
1560 break;
1561 }
Adam Litkee4e574b2007-10-16 01:26:19 -07001562 list_add(&page->lru, &surplus_list);
1563 }
Hillf Danton28073b02012-03-21 16:34:00 -07001564 allocated += i;
Adam Litkee4e574b2007-10-16 01:26:19 -07001565
1566 /*
1567 * After retaking hugetlb_lock, we need to recalculate 'needed'
1568 * because either resv_huge_pages or free_huge_pages may have changed.
1569 */
1570 spin_lock(&hugetlb_lock);
Andi Kleena5516432008-07-23 21:27:41 -07001571 needed = (h->resv_huge_pages + delta) -
1572 (h->free_huge_pages + allocated);
Hillf Danton28073b02012-03-21 16:34:00 -07001573 if (needed > 0) {
1574 if (alloc_ok)
1575 goto retry;
1576 /*
1577 * We were not able to allocate enough pages to
1578 * satisfy the entire reservation so we free what
1579 * we've allocated so far.
1580 */
1581 goto free;
1582 }
Adam Litkee4e574b2007-10-16 01:26:19 -07001583 /*
1584 * The surplus_list now contains _at_least_ the number of extra pages
Lucas De Marchi25985ed2011-03-30 22:57:33 -03001585 * needed to accommodate the reservation. Add the appropriate number
Adam Litkee4e574b2007-10-16 01:26:19 -07001586 * of pages to the hugetlb pool and free the extras back to the buddy
Adam Litkeac09b3a2008-03-04 14:29:38 -08001587 * allocator. Commit the entire reservation here to prevent another
1588 * process from stealing the pages as they are added to the pool but
1589 * before they are reserved.
Adam Litkee4e574b2007-10-16 01:26:19 -07001590 */
1591 needed += allocated;
Andi Kleena5516432008-07-23 21:27:41 -07001592 h->resv_huge_pages += delta;
Adam Litkee4e574b2007-10-16 01:26:19 -07001593 ret = 0;
Naoya Horiguchia9869b82010-09-08 10:19:37 +09001594
Adam Litke19fc3f02008-04-28 02:12:20 -07001595 /* Free the needed pages to the hugetlb pool */
Adam Litkee4e574b2007-10-16 01:26:19 -07001596 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
Adam Litke19fc3f02008-04-28 02:12:20 -07001597 if ((--needed) < 0)
1598 break;
Naoya Horiguchia9869b82010-09-08 10:19:37 +09001599 /*
1600 * This page is now managed by the hugetlb allocator and has
1601 * no users -- drop the buddy allocator's reference.
1602 */
1603 put_page_testzero(page);
Sasha Levin309381fea2014-01-23 15:52:54 -08001604 VM_BUG_ON_PAGE(page_count(page), page);
Andi Kleena5516432008-07-23 21:27:41 -07001605 enqueue_huge_page(h, page);
Adam Litke19fc3f02008-04-28 02:12:20 -07001606 }
Hillf Danton28073b02012-03-21 16:34:00 -07001607free:
Hillf Dantonb0365c82011-12-28 15:57:16 -08001608 spin_unlock(&hugetlb_lock);
Adam Litke19fc3f02008-04-28 02:12:20 -07001609
1610 /* Free unnecessary surplus pages to the buddy allocator */
Joonsoo Kimc0d934b2013-09-11 14:21:02 -07001611 list_for_each_entry_safe(page, tmp, &surplus_list, lru)
1612 put_page(page);
Naoya Horiguchia9869b82010-09-08 10:19:37 +09001613 spin_lock(&hugetlb_lock);
Adam Litkee4e574b2007-10-16 01:26:19 -07001614
1615 return ret;
1616}
1617
1618/*
1619 * When releasing a hugetlb pool reservation, any surplus pages that were
1620 * allocated to satisfy the reservation must be explicitly freed if they were
1621 * never used.
Lee Schermerhorn685f3452009-09-21 17:01:23 -07001622 * Called with hugetlb_lock held.
Adam Litkee4e574b2007-10-16 01:26:19 -07001623 */
Andi Kleena5516432008-07-23 21:27:41 -07001624static void return_unused_surplus_pages(struct hstate *h,
1625 unsigned long unused_resv_pages)
Adam Litkee4e574b2007-10-16 01:26:19 -07001626{
Adam Litkee4e574b2007-10-16 01:26:19 -07001627 unsigned long nr_pages;
1628
Adam Litkeac09b3a2008-03-04 14:29:38 -08001629 /* Uncommit the reservation */
Andi Kleena5516432008-07-23 21:27:41 -07001630 h->resv_huge_pages -= unused_resv_pages;
Adam Litkeac09b3a2008-03-04 14:29:38 -08001631
Andi Kleenaa888a72008-07-23 21:27:47 -07001632 /* Cannot return gigantic pages currently */
Luiz Capitulinobae7f4a2014-06-04 16:07:08 -07001633 if (hstate_is_gigantic(h))
Andi Kleenaa888a72008-07-23 21:27:47 -07001634 return;
1635
Andi Kleena5516432008-07-23 21:27:41 -07001636 nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
Adam Litkee4e574b2007-10-16 01:26:19 -07001637
Lee Schermerhorn685f3452009-09-21 17:01:23 -07001638 /*
1639 * We want to release as many surplus pages as possible, spread
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -08001640 * evenly across all nodes with memory. Iterate across these nodes
1641 * until we can no longer free unreserved surplus pages. This occurs
1642 * when the nodes with surplus pages have no free pages.
1643 * free_pool_huge_page() will balance the the freed pages across the
1644 * on-line nodes with memory and will handle the hstate accounting.
Lee Schermerhorn685f3452009-09-21 17:01:23 -07001645 */
1646 while (nr_pages--) {
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08001647 if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1))
Lee Schermerhorn685f3452009-09-21 17:01:23 -07001648 break;
Mizuma, Masayoshi7848a4b2014-04-18 15:07:18 -07001649 cond_resched_lock(&hugetlb_lock);
Adam Litkee4e574b2007-10-16 01:26:19 -07001650 }
1651}
1652
Mike Kravetz5e911372015-09-08 15:01:28 -07001653
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001654/*
Mike Kravetzfeba16e2015-09-08 15:01:31 -07001655 * vma_needs_reservation, vma_commit_reservation and vma_end_reservation
Mike Kravetz5e911372015-09-08 15:01:28 -07001656 * are used by the huge page allocation routines to manage reservations.
Mike Kravetzcf3ad202015-06-24 16:57:55 -07001657 *
1658 * vma_needs_reservation is called to determine if the huge page at addr
1659 * within the vma has an associated reservation. If a reservation is
1660 * needed, the value 1 is returned. The caller is then responsible for
1661 * managing the global reservation and subpool usage counts. After
1662 * the huge page has been allocated, vma_commit_reservation is called
Mike Kravetzfeba16e2015-09-08 15:01:31 -07001663 * to add the page to the reservation map. If the page allocation fails,
1664 * the reservation must be ended instead of committed. vma_end_reservation
1665 * is called in such cases.
Mike Kravetzcf3ad202015-06-24 16:57:55 -07001666 *
1667 * In the normal case, vma_commit_reservation returns the same value
1668 * as the preceding vma_needs_reservation call. The only time this
1669 * is not the case is if a reserve map was changed between calls. It
1670 * is the responsibility of the caller to notice the difference and
1671 * take appropriate action.
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001672 */
Mike Kravetz5e911372015-09-08 15:01:28 -07001673enum vma_resv_mode {
1674 VMA_NEEDS_RESV,
1675 VMA_COMMIT_RESV,
Mike Kravetzfeba16e2015-09-08 15:01:31 -07001676 VMA_END_RESV,
Mike Kravetz5e911372015-09-08 15:01:28 -07001677};
Mike Kravetzcf3ad202015-06-24 16:57:55 -07001678static long __vma_reservation_common(struct hstate *h,
1679 struct vm_area_struct *vma, unsigned long addr,
Mike Kravetz5e911372015-09-08 15:01:28 -07001680 enum vma_resv_mode mode)
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001681{
Joonsoo Kim4e35f482014-04-03 14:47:30 -07001682 struct resv_map *resv;
1683 pgoff_t idx;
Mike Kravetzcf3ad202015-06-24 16:57:55 -07001684 long ret;
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001685
Joonsoo Kim4e35f482014-04-03 14:47:30 -07001686 resv = vma_resv_map(vma);
1687 if (!resv)
Andy Whitcroft84afd992008-07-23 21:27:32 -07001688 return 1;
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001689
Joonsoo Kim4e35f482014-04-03 14:47:30 -07001690 idx = vma_hugecache_offset(h, vma, addr);
Mike Kravetz5e911372015-09-08 15:01:28 -07001691 switch (mode) {
1692 case VMA_NEEDS_RESV:
Mike Kravetzcf3ad202015-06-24 16:57:55 -07001693 ret = region_chg(resv, idx, idx + 1);
Mike Kravetz5e911372015-09-08 15:01:28 -07001694 break;
1695 case VMA_COMMIT_RESV:
1696 ret = region_add(resv, idx, idx + 1);
1697 break;
Mike Kravetzfeba16e2015-09-08 15:01:31 -07001698 case VMA_END_RESV:
Mike Kravetz5e911372015-09-08 15:01:28 -07001699 region_abort(resv, idx, idx + 1);
1700 ret = 0;
1701 break;
1702 default:
1703 BUG();
1704 }
Andy Whitcroft84afd992008-07-23 21:27:32 -07001705
Joonsoo Kim4e35f482014-04-03 14:47:30 -07001706 if (vma->vm_flags & VM_MAYSHARE)
Mike Kravetzcf3ad202015-06-24 16:57:55 -07001707 return ret;
Joonsoo Kim4e35f482014-04-03 14:47:30 -07001708 else
Mike Kravetzcf3ad202015-06-24 16:57:55 -07001709 return ret < 0 ? ret : 0;
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001710}
Mike Kravetzcf3ad202015-06-24 16:57:55 -07001711
1712static long vma_needs_reservation(struct hstate *h,
Andi Kleena5516432008-07-23 21:27:41 -07001713 struct vm_area_struct *vma, unsigned long addr)
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001714{
Mike Kravetz5e911372015-09-08 15:01:28 -07001715 return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV);
Mike Kravetzcf3ad202015-06-24 16:57:55 -07001716}
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001717
Mike Kravetzcf3ad202015-06-24 16:57:55 -07001718static long vma_commit_reservation(struct hstate *h,
1719 struct vm_area_struct *vma, unsigned long addr)
1720{
Mike Kravetz5e911372015-09-08 15:01:28 -07001721 return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV);
1722}
1723
Mike Kravetzfeba16e2015-09-08 15:01:31 -07001724static void vma_end_reservation(struct hstate *h,
Mike Kravetz5e911372015-09-08 15:01:28 -07001725 struct vm_area_struct *vma, unsigned long addr)
1726{
Mike Kravetzfeba16e2015-09-08 15:01:31 -07001727 (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV);
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001728}
1729
Mike Kravetz70c35472015-09-08 15:01:54 -07001730struct page *alloc_huge_page(struct vm_area_struct *vma,
Mel Gorman04f2cbe2008-07-23 21:27:25 -07001731 unsigned long addr, int avoid_reserve)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001732{
David Gibson90481622012-03-21 16:34:12 -07001733 struct hugepage_subpool *spool = subpool_vma(vma);
Andi Kleena5516432008-07-23 21:27:41 -07001734 struct hstate *h = hstate_vma(vma);
Adam Litke348ea202007-11-14 16:59:37 -08001735 struct page *page;
Mike Kravetzd85f69b2015-09-08 15:01:47 -07001736 long map_chg, map_commit;
1737 long gbl_chg;
Aneesh Kumar K.V6d76dcf2012-07-31 16:42:18 -07001738 int ret, idx;
1739 struct hugetlb_cgroup *h_cg;
Adam Litke2fc39ce2007-11-14 16:59:39 -08001740
Aneesh Kumar K.V6d76dcf2012-07-31 16:42:18 -07001741 idx = hstate_index(h);
Mel Gormana1e78772008-07-23 21:27:23 -07001742 /*
Mike Kravetzd85f69b2015-09-08 15:01:47 -07001743 * Examine the region/reserve map to determine if the process
1744 * has a reservation for the page to be allocated. A return
1745 * code of zero indicates a reservation exists (no change).
Mel Gormana1e78772008-07-23 21:27:23 -07001746 */
Mike Kravetzd85f69b2015-09-08 15:01:47 -07001747 map_chg = gbl_chg = vma_needs_reservation(h, vma, addr);
1748 if (map_chg < 0)
Aneesh Kumar K.V76dcee72012-07-31 16:41:57 -07001749 return ERR_PTR(-ENOMEM);
Mike Kravetzd85f69b2015-09-08 15:01:47 -07001750
1751 /*
1752 * Processes that did not create the mapping will have no
1753 * reserves as indicated by the region/reserve map. Check
1754 * that the allocation will not exceed the subpool limit.
1755 * Allocations for MAP_NORESERVE mappings also need to be
1756 * checked against any subpool limit.
1757 */
1758 if (map_chg || avoid_reserve) {
1759 gbl_chg = hugepage_subpool_get_pages(spool, 1);
1760 if (gbl_chg < 0) {
Mike Kravetzfeba16e2015-09-08 15:01:31 -07001761 vma_end_reservation(h, vma, addr);
Aneesh Kumar K.V76dcee72012-07-31 16:41:57 -07001762 return ERR_PTR(-ENOSPC);
Mike Kravetz5e911372015-09-08 15:01:28 -07001763 }
Mel Gormana1e78772008-07-23 21:27:23 -07001764
Mike Kravetzd85f69b2015-09-08 15:01:47 -07001765 /*
1766 * Even though there was no reservation in the region/reserve
1767 * map, there could be reservations associated with the
1768 * subpool that can be used. This would be indicated if the
1769 * return value of hugepage_subpool_get_pages() is zero.
1770 * However, if avoid_reserve is specified we still avoid even
1771 * the subpool reservations.
1772 */
1773 if (avoid_reserve)
1774 gbl_chg = 1;
1775 }
1776
Aneesh Kumar K.V6d76dcf2012-07-31 16:42:18 -07001777 ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg);
Jianyu Zhan8f34af62014-06-04 16:10:36 -07001778 if (ret)
1779 goto out_subpool_put;
1780
Mel Gormana1e78772008-07-23 21:27:23 -07001781 spin_lock(&hugetlb_lock);
Mike Kravetzd85f69b2015-09-08 15:01:47 -07001782 /*
1783 * glb_chg is passed to indicate whether or not a page must be taken
1784 * from the global free pool (global change). gbl_chg == 0 indicates
1785 * a reservation exists for the allocation.
1786 */
1787 page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg);
Joonsoo Kim81a6fca2013-09-11 14:20:58 -07001788 if (!page) {
Aneesh Kumar K.V94ae8ba2012-07-31 16:42:35 -07001789 spin_unlock(&hugetlb_lock);
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001790 page = alloc_buddy_huge_page(h, NUMA_NO_NODE);
Jianyu Zhan8f34af62014-06-04 16:10:36 -07001791 if (!page)
1792 goto out_uncharge_cgroup;
1793
Aneesh Kumar K.V79dbb232012-07-31 16:42:32 -07001794 spin_lock(&hugetlb_lock);
1795 list_move(&page->lru, &h->hugepage_activelist);
Joonsoo Kim81a6fca2013-09-11 14:20:58 -07001796 /* Fall through */
Mel Gormana1e78772008-07-23 21:27:23 -07001797 }
Joonsoo Kim81a6fca2013-09-11 14:20:58 -07001798 hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page);
1799 spin_unlock(&hugetlb_lock);
Mel Gormana1e78772008-07-23 21:27:23 -07001800
David Gibson90481622012-03-21 16:34:12 -07001801 set_page_private(page, (unsigned long)spool);
Mel Gormana1e78772008-07-23 21:27:23 -07001802
Mike Kravetzd85f69b2015-09-08 15:01:47 -07001803 map_commit = vma_commit_reservation(h, vma, addr);
1804 if (unlikely(map_chg > map_commit)) {
Mike Kravetz33039672015-06-24 16:57:58 -07001805 /*
1806 * The page was added to the reservation map between
1807 * vma_needs_reservation and vma_commit_reservation.
1808 * This indicates a race with hugetlb_reserve_pages.
1809 * Adjust for the subpool count incremented above AND
1810 * in hugetlb_reserve_pages for the same page. Also,
1811 * the reservation count added in hugetlb_reserve_pages
1812 * no longer applies.
1813 */
1814 long rsv_adjust;
1815
1816 rsv_adjust = hugepage_subpool_put_pages(spool, 1);
1817 hugetlb_acct_memory(h, -rsv_adjust);
1818 }
Adam Litke90d8b7e2007-11-14 16:59:42 -08001819 return page;
Jianyu Zhan8f34af62014-06-04 16:10:36 -07001820
1821out_uncharge_cgroup:
1822 hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg);
1823out_subpool_put:
Mike Kravetzd85f69b2015-09-08 15:01:47 -07001824 if (map_chg || avoid_reserve)
Jianyu Zhan8f34af62014-06-04 16:10:36 -07001825 hugepage_subpool_put_pages(spool, 1);
Mike Kravetzfeba16e2015-09-08 15:01:31 -07001826 vma_end_reservation(h, vma, addr);
Jianyu Zhan8f34af62014-06-04 16:10:36 -07001827 return ERR_PTR(-ENOSPC);
David Gibsonb45b5bd2006-03-22 00:08:55 -08001828}
1829
Naoya Horiguchi74060e42013-09-11 14:22:06 -07001830/*
1831 * alloc_huge_page()'s wrapper which simply returns the page if allocation
1832 * succeeds, otherwise NULL. This function is called from new_vma_page(),
1833 * where no ERR_VALUE is expected to be returned.
1834 */
1835struct page *alloc_huge_page_noerr(struct vm_area_struct *vma,
1836 unsigned long addr, int avoid_reserve)
1837{
1838 struct page *page = alloc_huge_page(vma, addr, avoid_reserve);
1839 if (IS_ERR(page))
1840 page = NULL;
1841 return page;
1842}
1843
Cyrill Gorcunov91f47662009-01-06 14:40:33 -08001844int __weak alloc_bootmem_huge_page(struct hstate *h)
Andi Kleenaa888a72008-07-23 21:27:47 -07001845{
1846 struct huge_bootmem_page *m;
Joonsoo Kimb2261022013-09-11 14:21:00 -07001847 int nr_nodes, node;
Andi Kleenaa888a72008-07-23 21:27:47 -07001848
Joonsoo Kimb2261022013-09-11 14:21:00 -07001849 for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) {
Andi Kleenaa888a72008-07-23 21:27:47 -07001850 void *addr;
1851
Grygorii Strashko8b89a112014-01-21 15:50:36 -08001852 addr = memblock_virt_alloc_try_nid_nopanic(
1853 huge_page_size(h), huge_page_size(h),
1854 0, BOOTMEM_ALLOC_ACCESSIBLE, node);
Andi Kleenaa888a72008-07-23 21:27:47 -07001855 if (addr) {
1856 /*
1857 * Use the beginning of the huge page to store the
1858 * huge_bootmem_page struct (until gather_bootmem
1859 * puts them into the mem_map).
1860 */
1861 m = addr;
Cyrill Gorcunov91f47662009-01-06 14:40:33 -08001862 goto found;
Andi Kleenaa888a72008-07-23 21:27:47 -07001863 }
Andi Kleenaa888a72008-07-23 21:27:47 -07001864 }
1865 return 0;
1866
1867found:
Luiz Capitulinodf994ea2014-12-12 16:55:21 -08001868 BUG_ON(!IS_ALIGNED(virt_to_phys(m), huge_page_size(h)));
Andi Kleenaa888a72008-07-23 21:27:47 -07001869 /* Put them into a private list first because mem_map is not up yet */
1870 list_add(&m->list, &huge_boot_pages);
1871 m->hstate = h;
1872 return 1;
1873}
1874
David Rientjesf412c972014-04-03 14:47:59 -07001875static void __init prep_compound_huge_page(struct page *page, int order)
Andy Whitcroft18229df2008-11-06 12:53:27 -08001876{
1877 if (unlikely(order > (MAX_ORDER - 1)))
1878 prep_compound_gigantic_page(page, order);
1879 else
1880 prep_compound_page(page, order);
1881}
1882
Andi Kleenaa888a72008-07-23 21:27:47 -07001883/* Put bootmem huge pages into the standard lists after mem_map is up */
1884static void __init gather_bootmem_prealloc(void)
1885{
1886 struct huge_bootmem_page *m;
1887
1888 list_for_each_entry(m, &huge_boot_pages, list) {
Andi Kleenaa888a72008-07-23 21:27:47 -07001889 struct hstate *h = m->hstate;
Becky Bruceee8f2482011-07-25 17:11:50 -07001890 struct page *page;
1891
1892#ifdef CONFIG_HIGHMEM
1893 page = pfn_to_page(m->phys >> PAGE_SHIFT);
Grygorii Strashko8b89a112014-01-21 15:50:36 -08001894 memblock_free_late(__pa(m),
1895 sizeof(struct huge_bootmem_page));
Becky Bruceee8f2482011-07-25 17:11:50 -07001896#else
1897 page = virt_to_page(m);
1898#endif
Andi Kleenaa888a72008-07-23 21:27:47 -07001899 WARN_ON(page_count(page) != 1);
Andy Whitcroft18229df2008-11-06 12:53:27 -08001900 prep_compound_huge_page(page, h->order);
Andrea Arcangelief5a22b2013-10-16 13:46:56 -07001901 WARN_ON(PageReserved(page));
Andi Kleenaa888a72008-07-23 21:27:47 -07001902 prep_new_huge_page(h, page, page_to_nid(page));
Rafael Aquinib0320c72011-06-15 15:08:39 -07001903 /*
1904 * If we had gigantic hugepages allocated at boot time, we need
1905 * to restore the 'stolen' pages to totalram_pages in order to
1906 * fix confusing memory reports from free(1) and another
1907 * side-effects, like CommitLimit going negative.
1908 */
Luiz Capitulinobae7f4a2014-06-04 16:07:08 -07001909 if (hstate_is_gigantic(h))
Jiang Liu3dcc0572013-07-03 15:03:21 -07001910 adjust_managed_page_count(page, 1 << h->order);
Andi Kleenaa888a72008-07-23 21:27:47 -07001911 }
1912}
1913
Andi Kleen8faa8b02008-07-23 21:27:48 -07001914static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001915{
1916 unsigned long i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001917
Andi Kleene5ff2152008-07-23 21:27:42 -07001918 for (i = 0; i < h->max_huge_pages; ++i) {
Luiz Capitulinobae7f4a2014-06-04 16:07:08 -07001919 if (hstate_is_gigantic(h)) {
Andi Kleenaa888a72008-07-23 21:27:47 -07001920 if (!alloc_bootmem_huge_page(h))
1921 break;
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -08001922 } else if (!alloc_fresh_huge_page(h,
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08001923 &node_states[N_MEMORY]))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001924 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001925 }
Andi Kleen8faa8b02008-07-23 21:27:48 -07001926 h->max_huge_pages = i;
Andi Kleene5ff2152008-07-23 21:27:42 -07001927}
1928
1929static void __init hugetlb_init_hstates(void)
1930{
1931 struct hstate *h;
1932
1933 for_each_hstate(h) {
Naoya Horiguchi641844f2015-06-24 16:56:59 -07001934 if (minimum_order > huge_page_order(h))
1935 minimum_order = huge_page_order(h);
1936
Andi Kleen8faa8b02008-07-23 21:27:48 -07001937 /* oversize hugepages were init'ed in early boot */
Luiz Capitulinobae7f4a2014-06-04 16:07:08 -07001938 if (!hstate_is_gigantic(h))
Andi Kleen8faa8b02008-07-23 21:27:48 -07001939 hugetlb_hstate_alloc_pages(h);
Andi Kleene5ff2152008-07-23 21:27:42 -07001940 }
Naoya Horiguchi641844f2015-06-24 16:56:59 -07001941 VM_BUG_ON(minimum_order == UINT_MAX);
Andi Kleene5ff2152008-07-23 21:27:42 -07001942}
1943
Andi Kleen4abd32d2008-07-23 21:27:49 -07001944static char * __init memfmt(char *buf, unsigned long n)
1945{
1946 if (n >= (1UL << 30))
1947 sprintf(buf, "%lu GB", n >> 30);
1948 else if (n >= (1UL << 20))
1949 sprintf(buf, "%lu MB", n >> 20);
1950 else
1951 sprintf(buf, "%lu KB", n >> 10);
1952 return buf;
1953}
1954
Andi Kleene5ff2152008-07-23 21:27:42 -07001955static void __init report_hugepages(void)
1956{
1957 struct hstate *h;
1958
1959 for_each_hstate(h) {
Andi Kleen4abd32d2008-07-23 21:27:49 -07001960 char buf[32];
Andrew Mortonffb22af2013-02-22 16:32:08 -08001961 pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n",
Andi Kleen4abd32d2008-07-23 21:27:49 -07001962 memfmt(buf, huge_page_size(h)),
1963 h->free_huge_pages);
Andi Kleene5ff2152008-07-23 21:27:42 -07001964 }
1965}
1966
Linus Torvalds1da177e2005-04-16 15:20:36 -07001967#ifdef CONFIG_HIGHMEM
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001968static void try_to_free_low(struct hstate *h, unsigned long count,
1969 nodemask_t *nodes_allowed)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001970{
Christoph Lameter4415cc82006-09-25 23:31:55 -07001971 int i;
1972
Luiz Capitulinobae7f4a2014-06-04 16:07:08 -07001973 if (hstate_is_gigantic(h))
Andi Kleenaa888a72008-07-23 21:27:47 -07001974 return;
1975
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001976 for_each_node_mask(i, *nodes_allowed) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001977 struct page *page, *next;
Andi Kleena5516432008-07-23 21:27:41 -07001978 struct list_head *freel = &h->hugepage_freelists[i];
1979 list_for_each_entry_safe(page, next, freel, lru) {
1980 if (count >= h->nr_huge_pages)
Adam Litke6b0c8802007-10-16 01:26:23 -07001981 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001982 if (PageHighMem(page))
1983 continue;
1984 list_del(&page->lru);
Andi Kleene5ff2152008-07-23 21:27:42 -07001985 update_and_free_page(h, page);
Andi Kleena5516432008-07-23 21:27:41 -07001986 h->free_huge_pages--;
1987 h->free_huge_pages_node[page_to_nid(page)]--;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001988 }
1989 }
1990}
1991#else
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001992static inline void try_to_free_low(struct hstate *h, unsigned long count,
1993 nodemask_t *nodes_allowed)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001994{
1995}
1996#endif
1997
Wu Fengguang20a03072009-06-16 15:32:22 -07001998/*
1999 * Increment or decrement surplus_huge_pages. Keep node-specific counters
2000 * balanced by operating on them in a round-robin fashion.
2001 * Returns 1 if an adjustment was made.
2002 */
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08002003static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
2004 int delta)
Wu Fengguang20a03072009-06-16 15:32:22 -07002005{
Joonsoo Kimb2261022013-09-11 14:21:00 -07002006 int nr_nodes, node;
Wu Fengguang20a03072009-06-16 15:32:22 -07002007
2008 VM_BUG_ON(delta != -1 && delta != 1);
Wu Fengguang20a03072009-06-16 15:32:22 -07002009
Joonsoo Kimb2261022013-09-11 14:21:00 -07002010 if (delta < 0) {
2011 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
2012 if (h->surplus_huge_pages_node[node])
2013 goto found;
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07002014 }
Joonsoo Kimb2261022013-09-11 14:21:00 -07002015 } else {
2016 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
2017 if (h->surplus_huge_pages_node[node] <
2018 h->nr_huge_pages_node[node])
2019 goto found;
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07002020 }
Joonsoo Kimb2261022013-09-11 14:21:00 -07002021 }
2022 return 0;
Wu Fengguang20a03072009-06-16 15:32:22 -07002023
Joonsoo Kimb2261022013-09-11 14:21:00 -07002024found:
2025 h->surplus_huge_pages += delta;
2026 h->surplus_huge_pages_node[node] += delta;
2027 return 1;
Wu Fengguang20a03072009-06-16 15:32:22 -07002028}
2029
Andi Kleena5516432008-07-23 21:27:41 -07002030#define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08002031static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count,
2032 nodemask_t *nodes_allowed)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002033{
Adam Litke7893d1d2007-10-16 01:26:18 -07002034 unsigned long min_count, ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002035
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07002036 if (hstate_is_gigantic(h) && !gigantic_page_supported())
Andi Kleenaa888a72008-07-23 21:27:47 -07002037 return h->max_huge_pages;
2038
Adam Litke7893d1d2007-10-16 01:26:18 -07002039 /*
2040 * Increase the pool size
2041 * First take pages out of surplus state. Then make up the
2042 * remaining difference by allocating fresh huge pages.
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08002043 *
2044 * We might race with alloc_buddy_huge_page() here and be unable
2045 * to convert a surplus huge page to a normal huge page. That is
2046 * not critical, though, it just means the overall size of the
2047 * pool might be one hugepage larger than it needs to be, but
2048 * within all the constraints specified by the sysctls.
Adam Litke7893d1d2007-10-16 01:26:18 -07002049 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002050 spin_lock(&hugetlb_lock);
Andi Kleena5516432008-07-23 21:27:41 -07002051 while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08002052 if (!adjust_pool_surplus(h, nodes_allowed, -1))
Adam Litke7893d1d2007-10-16 01:26:18 -07002053 break;
2054 }
2055
Andi Kleena5516432008-07-23 21:27:41 -07002056 while (count > persistent_huge_pages(h)) {
Adam Litke7893d1d2007-10-16 01:26:18 -07002057 /*
2058 * If this allocation races such that we no longer need the
2059 * page, free_huge_page will handle it by freeing the page
2060 * and reducing the surplus.
2061 */
2062 spin_unlock(&hugetlb_lock);
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07002063 if (hstate_is_gigantic(h))
2064 ret = alloc_fresh_gigantic_page(h, nodes_allowed);
2065 else
2066 ret = alloc_fresh_huge_page(h, nodes_allowed);
Adam Litke7893d1d2007-10-16 01:26:18 -07002067 spin_lock(&hugetlb_lock);
2068 if (!ret)
2069 goto out;
2070
Mel Gorman536240f22009-12-14 17:59:56 -08002071 /* Bail for signals. Probably ctrl-c from user */
2072 if (signal_pending(current))
2073 goto out;
Adam Litke7893d1d2007-10-16 01:26:18 -07002074 }
Adam Litke7893d1d2007-10-16 01:26:18 -07002075
2076 /*
2077 * Decrease the pool size
2078 * First return free pages to the buddy allocator (being careful
2079 * to keep enough around to satisfy reservations). Then place
2080 * pages into surplus state as needed so the pool will shrink
2081 * to the desired size as pages become free.
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08002082 *
2083 * By placing pages into the surplus state independent of the
2084 * overcommit value, we are allowing the surplus pool size to
2085 * exceed overcommit. There are few sane options here. Since
2086 * alloc_buddy_huge_page() is checking the global counter,
2087 * though, we'll note that we're not allowed to exceed surplus
2088 * and won't grow the pool anywhere else. Not until one of the
2089 * sysctls are changed, or the surplus pages go out of use.
Adam Litke7893d1d2007-10-16 01:26:18 -07002090 */
Andi Kleena5516432008-07-23 21:27:41 -07002091 min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
Adam Litke6b0c8802007-10-16 01:26:23 -07002092 min_count = max(count, min_count);
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08002093 try_to_free_low(h, min_count, nodes_allowed);
Andi Kleena5516432008-07-23 21:27:41 -07002094 while (min_count < persistent_huge_pages(h)) {
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08002095 if (!free_pool_huge_page(h, nodes_allowed, 0))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002096 break;
Mizuma, Masayoshi55f67142014-04-07 15:37:54 -07002097 cond_resched_lock(&hugetlb_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002098 }
Andi Kleena5516432008-07-23 21:27:41 -07002099 while (count < persistent_huge_pages(h)) {
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08002100 if (!adjust_pool_surplus(h, nodes_allowed, 1))
Adam Litke7893d1d2007-10-16 01:26:18 -07002101 break;
2102 }
2103out:
Andi Kleena5516432008-07-23 21:27:41 -07002104 ret = persistent_huge_pages(h);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002105 spin_unlock(&hugetlb_lock);
Adam Litke7893d1d2007-10-16 01:26:18 -07002106 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002107}
2108
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002109#define HSTATE_ATTR_RO(_name) \
2110 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
2111
2112#define HSTATE_ATTR(_name) \
2113 static struct kobj_attribute _name##_attr = \
2114 __ATTR(_name, 0644, _name##_show, _name##_store)
2115
2116static struct kobject *hugepages_kobj;
2117static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
2118
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002119static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
2120
2121static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002122{
2123 int i;
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002124
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002125 for (i = 0; i < HUGE_MAX_HSTATE; i++)
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002126 if (hstate_kobjs[i] == kobj) {
2127 if (nidp)
2128 *nidp = NUMA_NO_NODE;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002129 return &hstates[i];
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002130 }
2131
2132 return kobj_to_node_hstate(kobj, nidp);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002133}
2134
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002135static ssize_t nr_hugepages_show_common(struct kobject *kobj,
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002136 struct kobj_attribute *attr, char *buf)
2137{
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002138 struct hstate *h;
2139 unsigned long nr_huge_pages;
2140 int nid;
2141
2142 h = kobj_to_hstate(kobj, &nid);
2143 if (nid == NUMA_NO_NODE)
2144 nr_huge_pages = h->nr_huge_pages;
2145 else
2146 nr_huge_pages = h->nr_huge_pages_node[nid];
2147
2148 return sprintf(buf, "%lu\n", nr_huge_pages);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002149}
Eric B Munsonadbe8722011-01-13 15:47:27 -08002150
David Rientjes238d3c12014-08-06 16:06:51 -07002151static ssize_t __nr_hugepages_store_common(bool obey_mempolicy,
2152 struct hstate *h, int nid,
2153 unsigned long count, size_t len)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002154{
2155 int err;
David Rientjesbad44b52009-12-14 17:58:38 -08002156 NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002157
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07002158 if (hstate_is_gigantic(h) && !gigantic_page_supported()) {
Eric B Munsonadbe8722011-01-13 15:47:27 -08002159 err = -EINVAL;
2160 goto out;
2161 }
2162
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002163 if (nid == NUMA_NO_NODE) {
2164 /*
2165 * global hstate attribute
2166 */
2167 if (!(obey_mempolicy &&
2168 init_nodemask_of_mempolicy(nodes_allowed))) {
2169 NODEMASK_FREE(nodes_allowed);
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08002170 nodes_allowed = &node_states[N_MEMORY];
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002171 }
2172 } else if (nodes_allowed) {
2173 /*
2174 * per node hstate attribute: adjust count to global,
2175 * but restrict alloc/free to the specified node.
2176 */
2177 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
2178 init_nodemask_of_node(nodes_allowed, nid);
2179 } else
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08002180 nodes_allowed = &node_states[N_MEMORY];
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002181
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002182 h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002183
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08002184 if (nodes_allowed != &node_states[N_MEMORY])
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002185 NODEMASK_FREE(nodes_allowed);
2186
2187 return len;
Eric B Munsonadbe8722011-01-13 15:47:27 -08002188out:
2189 NODEMASK_FREE(nodes_allowed);
2190 return err;
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002191}
2192
David Rientjes238d3c12014-08-06 16:06:51 -07002193static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
2194 struct kobject *kobj, const char *buf,
2195 size_t len)
2196{
2197 struct hstate *h;
2198 unsigned long count;
2199 int nid;
2200 int err;
2201
2202 err = kstrtoul(buf, 10, &count);
2203 if (err)
2204 return err;
2205
2206 h = kobj_to_hstate(kobj, &nid);
2207 return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len);
2208}
2209
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002210static ssize_t nr_hugepages_show(struct kobject *kobj,
2211 struct kobj_attribute *attr, char *buf)
2212{
2213 return nr_hugepages_show_common(kobj, attr, buf);
2214}
2215
2216static ssize_t nr_hugepages_store(struct kobject *kobj,
2217 struct kobj_attribute *attr, const char *buf, size_t len)
2218{
David Rientjes238d3c12014-08-06 16:06:51 -07002219 return nr_hugepages_store_common(false, kobj, buf, len);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002220}
2221HSTATE_ATTR(nr_hugepages);
2222
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002223#ifdef CONFIG_NUMA
2224
2225/*
2226 * hstate attribute for optionally mempolicy-based constraint on persistent
2227 * huge page alloc/free.
2228 */
2229static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
2230 struct kobj_attribute *attr, char *buf)
2231{
2232 return nr_hugepages_show_common(kobj, attr, buf);
2233}
2234
2235static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
2236 struct kobj_attribute *attr, const char *buf, size_t len)
2237{
David Rientjes238d3c12014-08-06 16:06:51 -07002238 return nr_hugepages_store_common(true, kobj, buf, len);
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002239}
2240HSTATE_ATTR(nr_hugepages_mempolicy);
2241#endif
2242
2243
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002244static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
2245 struct kobj_attribute *attr, char *buf)
2246{
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002247 struct hstate *h = kobj_to_hstate(kobj, NULL);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002248 return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
2249}
Eric B Munsonadbe8722011-01-13 15:47:27 -08002250
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002251static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
2252 struct kobj_attribute *attr, const char *buf, size_t count)
2253{
2254 int err;
2255 unsigned long input;
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002256 struct hstate *h = kobj_to_hstate(kobj, NULL);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002257
Luiz Capitulinobae7f4a2014-06-04 16:07:08 -07002258 if (hstate_is_gigantic(h))
Eric B Munsonadbe8722011-01-13 15:47:27 -08002259 return -EINVAL;
2260
Jingoo Han3dbb95f2013-09-11 14:20:25 -07002261 err = kstrtoul(buf, 10, &input);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002262 if (err)
Eric B Munson73ae31e2011-01-13 15:47:28 -08002263 return err;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002264
2265 spin_lock(&hugetlb_lock);
2266 h->nr_overcommit_huge_pages = input;
2267 spin_unlock(&hugetlb_lock);
2268
2269 return count;
2270}
2271HSTATE_ATTR(nr_overcommit_hugepages);
2272
2273static ssize_t free_hugepages_show(struct kobject *kobj,
2274 struct kobj_attribute *attr, char *buf)
2275{
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002276 struct hstate *h;
2277 unsigned long free_huge_pages;
2278 int nid;
2279
2280 h = kobj_to_hstate(kobj, &nid);
2281 if (nid == NUMA_NO_NODE)
2282 free_huge_pages = h->free_huge_pages;
2283 else
2284 free_huge_pages = h->free_huge_pages_node[nid];
2285
2286 return sprintf(buf, "%lu\n", free_huge_pages);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002287}
2288HSTATE_ATTR_RO(free_hugepages);
2289
2290static ssize_t resv_hugepages_show(struct kobject *kobj,
2291 struct kobj_attribute *attr, char *buf)
2292{
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002293 struct hstate *h = kobj_to_hstate(kobj, NULL);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002294 return sprintf(buf, "%lu\n", h->resv_huge_pages);
2295}
2296HSTATE_ATTR_RO(resv_hugepages);
2297
2298static ssize_t surplus_hugepages_show(struct kobject *kobj,
2299 struct kobj_attribute *attr, char *buf)
2300{
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002301 struct hstate *h;
2302 unsigned long surplus_huge_pages;
2303 int nid;
2304
2305 h = kobj_to_hstate(kobj, &nid);
2306 if (nid == NUMA_NO_NODE)
2307 surplus_huge_pages = h->surplus_huge_pages;
2308 else
2309 surplus_huge_pages = h->surplus_huge_pages_node[nid];
2310
2311 return sprintf(buf, "%lu\n", surplus_huge_pages);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002312}
2313HSTATE_ATTR_RO(surplus_hugepages);
2314
2315static struct attribute *hstate_attrs[] = {
2316 &nr_hugepages_attr.attr,
2317 &nr_overcommit_hugepages_attr.attr,
2318 &free_hugepages_attr.attr,
2319 &resv_hugepages_attr.attr,
2320 &surplus_hugepages_attr.attr,
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002321#ifdef CONFIG_NUMA
2322 &nr_hugepages_mempolicy_attr.attr,
2323#endif
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002324 NULL,
2325};
2326
2327static struct attribute_group hstate_attr_group = {
2328 .attrs = hstate_attrs,
2329};
2330
Jeff Mahoney094e9532010-02-02 13:44:14 -08002331static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
2332 struct kobject **hstate_kobjs,
2333 struct attribute_group *hstate_attr_group)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002334{
2335 int retval;
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07002336 int hi = hstate_index(h);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002337
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002338 hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
2339 if (!hstate_kobjs[hi])
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002340 return -ENOMEM;
2341
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002342 retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002343 if (retval)
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002344 kobject_put(hstate_kobjs[hi]);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002345
2346 return retval;
2347}
2348
2349static void __init hugetlb_sysfs_init(void)
2350{
2351 struct hstate *h;
2352 int err;
2353
2354 hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
2355 if (!hugepages_kobj)
2356 return;
2357
2358 for_each_hstate(h) {
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002359 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
2360 hstate_kobjs, &hstate_attr_group);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002361 if (err)
Andrew Mortonffb22af2013-02-22 16:32:08 -08002362 pr_err("Hugetlb: Unable to add hstate %s", h->name);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002363 }
2364}
2365
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002366#ifdef CONFIG_NUMA
2367
2368/*
2369 * node_hstate/s - associate per node hstate attributes, via their kobjects,
Kay Sievers10fbcf42011-12-21 14:48:43 -08002370 * with node devices in node_devices[] using a parallel array. The array
2371 * index of a node device or _hstate == node id.
2372 * This is here to avoid any static dependency of the node device driver, in
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002373 * the base kernel, on the hugetlb module.
2374 */
2375struct node_hstate {
2376 struct kobject *hugepages_kobj;
2377 struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
2378};
2379struct node_hstate node_hstates[MAX_NUMNODES];
2380
2381/*
Kay Sievers10fbcf42011-12-21 14:48:43 -08002382 * A subset of global hstate attributes for node devices
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002383 */
2384static struct attribute *per_node_hstate_attrs[] = {
2385 &nr_hugepages_attr.attr,
2386 &free_hugepages_attr.attr,
2387 &surplus_hugepages_attr.attr,
2388 NULL,
2389};
2390
2391static struct attribute_group per_node_hstate_attr_group = {
2392 .attrs = per_node_hstate_attrs,
2393};
2394
2395/*
Kay Sievers10fbcf42011-12-21 14:48:43 -08002396 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002397 * Returns node id via non-NULL nidp.
2398 */
2399static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
2400{
2401 int nid;
2402
2403 for (nid = 0; nid < nr_node_ids; nid++) {
2404 struct node_hstate *nhs = &node_hstates[nid];
2405 int i;
2406 for (i = 0; i < HUGE_MAX_HSTATE; i++)
2407 if (nhs->hstate_kobjs[i] == kobj) {
2408 if (nidp)
2409 *nidp = nid;
2410 return &hstates[i];
2411 }
2412 }
2413
2414 BUG();
2415 return NULL;
2416}
2417
2418/*
Kay Sievers10fbcf42011-12-21 14:48:43 -08002419 * Unregister hstate attributes from a single node device.
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002420 * No-op if no hstate attributes attached.
2421 */
Claudiu Ghioc3cd8b442013-03-04 12:46:15 +02002422static void hugetlb_unregister_node(struct node *node)
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002423{
2424 struct hstate *h;
Kay Sievers10fbcf42011-12-21 14:48:43 -08002425 struct node_hstate *nhs = &node_hstates[node->dev.id];
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002426
2427 if (!nhs->hugepages_kobj)
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -08002428 return; /* no hstate attributes */
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002429
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07002430 for_each_hstate(h) {
2431 int idx = hstate_index(h);
2432 if (nhs->hstate_kobjs[idx]) {
2433 kobject_put(nhs->hstate_kobjs[idx]);
2434 nhs->hstate_kobjs[idx] = NULL;
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002435 }
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07002436 }
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002437
2438 kobject_put(nhs->hugepages_kobj);
2439 nhs->hugepages_kobj = NULL;
2440}
2441
2442/*
Kay Sievers10fbcf42011-12-21 14:48:43 -08002443 * hugetlb module exit: unregister hstate attributes from node devices
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002444 * that have them.
2445 */
2446static void hugetlb_unregister_all_nodes(void)
2447{
2448 int nid;
2449
2450 /*
Kay Sievers10fbcf42011-12-21 14:48:43 -08002451 * disable node device registrations.
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002452 */
2453 register_hugetlbfs_with_node(NULL, NULL);
2454
2455 /*
2456 * remove hstate attributes from any nodes that have them.
2457 */
2458 for (nid = 0; nid < nr_node_ids; nid++)
Wen Congyang87327942012-12-11 16:00:56 -08002459 hugetlb_unregister_node(node_devices[nid]);
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002460}
2461
2462/*
Kay Sievers10fbcf42011-12-21 14:48:43 -08002463 * Register hstate attributes for a single node device.
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002464 * No-op if attributes already registered.
2465 */
Claudiu Ghioc3cd8b442013-03-04 12:46:15 +02002466static void hugetlb_register_node(struct node *node)
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002467{
2468 struct hstate *h;
Kay Sievers10fbcf42011-12-21 14:48:43 -08002469 struct node_hstate *nhs = &node_hstates[node->dev.id];
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002470 int err;
2471
2472 if (nhs->hugepages_kobj)
2473 return; /* already allocated */
2474
2475 nhs->hugepages_kobj = kobject_create_and_add("hugepages",
Kay Sievers10fbcf42011-12-21 14:48:43 -08002476 &node->dev.kobj);
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002477 if (!nhs->hugepages_kobj)
2478 return;
2479
2480 for_each_hstate(h) {
2481 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
2482 nhs->hstate_kobjs,
2483 &per_node_hstate_attr_group);
2484 if (err) {
Andrew Mortonffb22af2013-02-22 16:32:08 -08002485 pr_err("Hugetlb: Unable to add hstate %s for node %d\n",
2486 h->name, node->dev.id);
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002487 hugetlb_unregister_node(node);
2488 break;
2489 }
2490 }
2491}
2492
2493/*
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -08002494 * hugetlb init time: register hstate attributes for all registered node
Kay Sievers10fbcf42011-12-21 14:48:43 -08002495 * devices of nodes that have memory. All on-line nodes should have
2496 * registered their associated device by this time.
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002497 */
Luiz Capitulino7d9ca002014-12-12 16:55:24 -08002498static void __init hugetlb_register_all_nodes(void)
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002499{
2500 int nid;
2501
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08002502 for_each_node_state(nid, N_MEMORY) {
Wen Congyang87327942012-12-11 16:00:56 -08002503 struct node *node = node_devices[nid];
Kay Sievers10fbcf42011-12-21 14:48:43 -08002504 if (node->dev.id == nid)
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002505 hugetlb_register_node(node);
2506 }
2507
2508 /*
Kay Sievers10fbcf42011-12-21 14:48:43 -08002509 * Let the node device driver know we're here so it can
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002510 * [un]register hstate attributes on node hotplug.
2511 */
2512 register_hugetlbfs_with_node(hugetlb_register_node,
2513 hugetlb_unregister_node);
2514}
2515#else /* !CONFIG_NUMA */
2516
2517static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
2518{
2519 BUG();
2520 if (nidp)
2521 *nidp = -1;
2522 return NULL;
2523}
2524
2525static void hugetlb_unregister_all_nodes(void) { }
2526
2527static void hugetlb_register_all_nodes(void) { }
2528
2529#endif
2530
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002531static void __exit hugetlb_exit(void)
2532{
2533 struct hstate *h;
2534
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002535 hugetlb_unregister_all_nodes();
2536
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002537 for_each_hstate(h) {
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07002538 kobject_put(hstate_kobjs[hstate_index(h)]);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002539 }
2540
2541 kobject_put(hugepages_kobj);
Mike Kravetzc672c7f2015-09-08 15:01:35 -07002542 kfree(hugetlb_fault_mutex_table);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002543}
2544module_exit(hugetlb_exit);
2545
2546static int __init hugetlb_init(void)
2547{
Davidlohr Bueso8382d912014-04-03 14:47:31 -07002548 int i;
2549
Nishanth Aravamudan457c1b22014-05-06 12:50:00 -07002550 if (!hugepages_supported())
Benjamin Herrenschmidt0ef89d22008-07-31 00:07:30 -07002551 return 0;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002552
Nick Piggine11bfbf2008-07-23 21:27:52 -07002553 if (!size_to_hstate(default_hstate_size)) {
2554 default_hstate_size = HPAGE_SIZE;
2555 if (!size_to_hstate(default_hstate_size))
2556 hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002557 }
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07002558 default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size));
Nick Piggine11bfbf2008-07-23 21:27:52 -07002559 if (default_hstate_max_huge_pages)
2560 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002561
2562 hugetlb_init_hstates();
Andi Kleenaa888a72008-07-23 21:27:47 -07002563 gather_bootmem_prealloc();
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002564 report_hugepages();
2565
2566 hugetlb_sysfs_init();
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002567 hugetlb_register_all_nodes();
Jianguo Wu7179e7b2012-12-18 14:23:19 -08002568 hugetlb_cgroup_file_init();
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002569
Davidlohr Bueso8382d912014-04-03 14:47:31 -07002570#ifdef CONFIG_SMP
2571 num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus());
2572#else
2573 num_fault_mutexes = 1;
2574#endif
Mike Kravetzc672c7f2015-09-08 15:01:35 -07002575 hugetlb_fault_mutex_table =
Davidlohr Bueso8382d912014-04-03 14:47:31 -07002576 kmalloc(sizeof(struct mutex) * num_fault_mutexes, GFP_KERNEL);
Mike Kravetzc672c7f2015-09-08 15:01:35 -07002577 BUG_ON(!hugetlb_fault_mutex_table);
Davidlohr Bueso8382d912014-04-03 14:47:31 -07002578
2579 for (i = 0; i < num_fault_mutexes; i++)
Mike Kravetzc672c7f2015-09-08 15:01:35 -07002580 mutex_init(&hugetlb_fault_mutex_table[i]);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002581 return 0;
2582}
2583module_init(hugetlb_init);
2584
2585/* Should be called on processing a hugepagesz=... option */
2586void __init hugetlb_add_hstate(unsigned order)
2587{
2588 struct hstate *h;
Andi Kleen8faa8b02008-07-23 21:27:48 -07002589 unsigned long i;
2590
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002591 if (size_to_hstate(PAGE_SIZE << order)) {
Andrew Mortonffb22af2013-02-22 16:32:08 -08002592 pr_warning("hugepagesz= specified twice, ignoring\n");
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002593 return;
2594 }
Aneesh Kumar K.V47d38342012-07-31 16:41:54 -07002595 BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002596 BUG_ON(order == 0);
Aneesh Kumar K.V47d38342012-07-31 16:41:54 -07002597 h = &hstates[hugetlb_max_hstate++];
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002598 h->order = order;
2599 h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
Andi Kleen8faa8b02008-07-23 21:27:48 -07002600 h->nr_huge_pages = 0;
2601 h->free_huge_pages = 0;
2602 for (i = 0; i < MAX_NUMNODES; ++i)
2603 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
Aneesh Kumar K.V0edaecf2012-07-31 16:42:07 -07002604 INIT_LIST_HEAD(&h->hugepage_activelist);
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08002605 h->next_nid_to_alloc = first_node(node_states[N_MEMORY]);
2606 h->next_nid_to_free = first_node(node_states[N_MEMORY]);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002607 snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
2608 huge_page_size(h)/1024);
Andi Kleen8faa8b02008-07-23 21:27:48 -07002609
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002610 parsed_hstate = h;
2611}
2612
Nick Piggine11bfbf2008-07-23 21:27:52 -07002613static int __init hugetlb_nrpages_setup(char *s)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002614{
2615 unsigned long *mhp;
Andi Kleen8faa8b02008-07-23 21:27:48 -07002616 static unsigned long *last_mhp;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002617
2618 /*
Aneesh Kumar K.V47d38342012-07-31 16:41:54 -07002619 * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet,
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002620 * so this hugepages= parameter goes to the "default hstate".
2621 */
Aneesh Kumar K.V47d38342012-07-31 16:41:54 -07002622 if (!hugetlb_max_hstate)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002623 mhp = &default_hstate_max_huge_pages;
2624 else
2625 mhp = &parsed_hstate->max_huge_pages;
2626
Andi Kleen8faa8b02008-07-23 21:27:48 -07002627 if (mhp == last_mhp) {
Andrew Mortonffb22af2013-02-22 16:32:08 -08002628 pr_warning("hugepages= specified twice without "
2629 "interleaving hugepagesz=, ignoring\n");
Andi Kleen8faa8b02008-07-23 21:27:48 -07002630 return 1;
2631 }
2632
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002633 if (sscanf(s, "%lu", mhp) <= 0)
2634 *mhp = 0;
2635
Andi Kleen8faa8b02008-07-23 21:27:48 -07002636 /*
2637 * Global state is always initialized later in hugetlb_init.
2638 * But we need to allocate >= MAX_ORDER hstates here early to still
2639 * use the bootmem allocator.
2640 */
Aneesh Kumar K.V47d38342012-07-31 16:41:54 -07002641 if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER)
Andi Kleen8faa8b02008-07-23 21:27:48 -07002642 hugetlb_hstate_alloc_pages(parsed_hstate);
2643
2644 last_mhp = mhp;
2645
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002646 return 1;
2647}
Nick Piggine11bfbf2008-07-23 21:27:52 -07002648__setup("hugepages=", hugetlb_nrpages_setup);
2649
2650static int __init hugetlb_default_setup(char *s)
2651{
2652 default_hstate_size = memparse(s, &s);
2653 return 1;
2654}
2655__setup("default_hugepagesz=", hugetlb_default_setup);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002656
Nishanth Aravamudan8a213462008-07-25 19:44:37 -07002657static unsigned int cpuset_mems_nr(unsigned int *array)
2658{
2659 int node;
2660 unsigned int nr = 0;
2661
2662 for_each_node_mask(node, cpuset_current_mems_allowed)
2663 nr += array[node];
2664
2665 return nr;
2666}
2667
2668#ifdef CONFIG_SYSCTL
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002669static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
2670 struct ctl_table *table, int write,
2671 void __user *buffer, size_t *length, loff_t *ppos)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002672{
Andi Kleene5ff2152008-07-23 21:27:42 -07002673 struct hstate *h = &default_hstate;
David Rientjes238d3c12014-08-06 16:06:51 -07002674 unsigned long tmp = h->max_huge_pages;
Michal Hocko08d4a242011-01-13 15:47:26 -08002675 int ret;
Andi Kleene5ff2152008-07-23 21:27:42 -07002676
Nishanth Aravamudan457c1b22014-05-06 12:50:00 -07002677 if (!hugepages_supported())
2678 return -ENOTSUPP;
2679
Andi Kleene5ff2152008-07-23 21:27:42 -07002680 table->data = &tmp;
2681 table->maxlen = sizeof(unsigned long);
Michal Hocko08d4a242011-01-13 15:47:26 -08002682 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2683 if (ret)
2684 goto out;
Andi Kleene5ff2152008-07-23 21:27:42 -07002685
David Rientjes238d3c12014-08-06 16:06:51 -07002686 if (write)
2687 ret = __nr_hugepages_store_common(obey_mempolicy, h,
2688 NUMA_NO_NODE, tmp, *length);
Michal Hocko08d4a242011-01-13 15:47:26 -08002689out:
2690 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002691}
Mel Gorman396faf02007-07-17 04:03:13 -07002692
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002693int hugetlb_sysctl_handler(struct ctl_table *table, int write,
2694 void __user *buffer, size_t *length, loff_t *ppos)
2695{
2696
2697 return hugetlb_sysctl_handler_common(false, table, write,
2698 buffer, length, ppos);
2699}
2700
2701#ifdef CONFIG_NUMA
2702int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
2703 void __user *buffer, size_t *length, loff_t *ppos)
2704{
2705 return hugetlb_sysctl_handler_common(true, table, write,
2706 buffer, length, ppos);
2707}
2708#endif /* CONFIG_NUMA */
2709
Nishanth Aravamudana3d0c6a2008-02-08 04:18:18 -08002710int hugetlb_overcommit_handler(struct ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07002711 void __user *buffer,
Nishanth Aravamudana3d0c6a2008-02-08 04:18:18 -08002712 size_t *length, loff_t *ppos)
2713{
Andi Kleena5516432008-07-23 21:27:41 -07002714 struct hstate *h = &default_hstate;
Andi Kleene5ff2152008-07-23 21:27:42 -07002715 unsigned long tmp;
Michal Hocko08d4a242011-01-13 15:47:26 -08002716 int ret;
Andi Kleene5ff2152008-07-23 21:27:42 -07002717
Nishanth Aravamudan457c1b22014-05-06 12:50:00 -07002718 if (!hugepages_supported())
2719 return -ENOTSUPP;
2720
Petr Holasekc033a932011-03-22 16:33:05 -07002721 tmp = h->nr_overcommit_huge_pages;
Andi Kleene5ff2152008-07-23 21:27:42 -07002722
Luiz Capitulinobae7f4a2014-06-04 16:07:08 -07002723 if (write && hstate_is_gigantic(h))
Eric B Munsonadbe8722011-01-13 15:47:27 -08002724 return -EINVAL;
2725
Andi Kleene5ff2152008-07-23 21:27:42 -07002726 table->data = &tmp;
2727 table->maxlen = sizeof(unsigned long);
Michal Hocko08d4a242011-01-13 15:47:26 -08002728 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2729 if (ret)
2730 goto out;
Andi Kleene5ff2152008-07-23 21:27:42 -07002731
2732 if (write) {
2733 spin_lock(&hugetlb_lock);
2734 h->nr_overcommit_huge_pages = tmp;
2735 spin_unlock(&hugetlb_lock);
2736 }
Michal Hocko08d4a242011-01-13 15:47:26 -08002737out:
2738 return ret;
Nishanth Aravamudana3d0c6a2008-02-08 04:18:18 -08002739}
2740
Linus Torvalds1da177e2005-04-16 15:20:36 -07002741#endif /* CONFIG_SYSCTL */
2742
Alexey Dobriyane1759c22008-10-15 23:50:22 +04002743void hugetlb_report_meminfo(struct seq_file *m)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002744{
Andi Kleena5516432008-07-23 21:27:41 -07002745 struct hstate *h = &default_hstate;
Nishanth Aravamudan457c1b22014-05-06 12:50:00 -07002746 if (!hugepages_supported())
2747 return;
Alexey Dobriyane1759c22008-10-15 23:50:22 +04002748 seq_printf(m,
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002749 "HugePages_Total: %5lu\n"
2750 "HugePages_Free: %5lu\n"
2751 "HugePages_Rsvd: %5lu\n"
2752 "HugePages_Surp: %5lu\n"
2753 "Hugepagesize: %8lu kB\n",
Andi Kleena5516432008-07-23 21:27:41 -07002754 h->nr_huge_pages,
2755 h->free_huge_pages,
2756 h->resv_huge_pages,
2757 h->surplus_huge_pages,
2758 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002759}
2760
2761int hugetlb_report_node_meminfo(int nid, char *buf)
2762{
Andi Kleena5516432008-07-23 21:27:41 -07002763 struct hstate *h = &default_hstate;
Nishanth Aravamudan457c1b22014-05-06 12:50:00 -07002764 if (!hugepages_supported())
2765 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002766 return sprintf(buf,
2767 "Node %d HugePages_Total: %5u\n"
Nishanth Aravamudana1de0912008-03-26 14:37:53 -07002768 "Node %d HugePages_Free: %5u\n"
2769 "Node %d HugePages_Surp: %5u\n",
Andi Kleena5516432008-07-23 21:27:41 -07002770 nid, h->nr_huge_pages_node[nid],
2771 nid, h->free_huge_pages_node[nid],
2772 nid, h->surplus_huge_pages_node[nid]);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002773}
2774
David Rientjes949f7ec2013-04-29 15:07:48 -07002775void hugetlb_show_meminfo(void)
2776{
2777 struct hstate *h;
2778 int nid;
2779
Nishanth Aravamudan457c1b22014-05-06 12:50:00 -07002780 if (!hugepages_supported())
2781 return;
2782
David Rientjes949f7ec2013-04-29 15:07:48 -07002783 for_each_node_state(nid, N_MEMORY)
2784 for_each_hstate(h)
2785 pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
2786 nid,
2787 h->nr_huge_pages_node[nid],
2788 h->free_huge_pages_node[nid],
2789 h->surplus_huge_pages_node[nid],
2790 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
2791}
2792
Linus Torvalds1da177e2005-04-16 15:20:36 -07002793/* Return the number pages of memory we physically have, in PAGE_SIZE units. */
2794unsigned long hugetlb_total_pages(void)
2795{
Wanpeng Lid0028582013-03-22 15:04:40 -07002796 struct hstate *h;
2797 unsigned long nr_total_pages = 0;
2798
2799 for_each_hstate(h)
2800 nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h);
2801 return nr_total_pages;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002802}
Linus Torvalds1da177e2005-04-16 15:20:36 -07002803
Andi Kleena5516432008-07-23 21:27:41 -07002804static int hugetlb_acct_memory(struct hstate *h, long delta)
Mel Gormanfc1b8a72008-07-23 21:27:22 -07002805{
2806 int ret = -ENOMEM;
2807
2808 spin_lock(&hugetlb_lock);
2809 /*
2810 * When cpuset is configured, it breaks the strict hugetlb page
2811 * reservation as the accounting is done on a global variable. Such
2812 * reservation is completely rubbish in the presence of cpuset because
2813 * the reservation is not checked against page availability for the
2814 * current cpuset. Application can still potentially OOM'ed by kernel
2815 * with lack of free htlb page in cpuset that the task is in.
2816 * Attempt to enforce strict accounting with cpuset is almost
2817 * impossible (or too ugly) because cpuset is too fluid that
2818 * task or memory node can be dynamically moved between cpusets.
2819 *
2820 * The change of semantics for shared hugetlb mapping with cpuset is
2821 * undesirable. However, in order to preserve some of the semantics,
2822 * we fall back to check against current free page availability as
2823 * a best attempt and hopefully to minimize the impact of changing
2824 * semantics that cpuset has.
2825 */
2826 if (delta > 0) {
Andi Kleena5516432008-07-23 21:27:41 -07002827 if (gather_surplus_pages(h, delta) < 0)
Mel Gormanfc1b8a72008-07-23 21:27:22 -07002828 goto out;
2829
Andi Kleena5516432008-07-23 21:27:41 -07002830 if (delta > cpuset_mems_nr(h->free_huge_pages_node)) {
2831 return_unused_surplus_pages(h, delta);
Mel Gormanfc1b8a72008-07-23 21:27:22 -07002832 goto out;
2833 }
2834 }
2835
2836 ret = 0;
2837 if (delta < 0)
Andi Kleena5516432008-07-23 21:27:41 -07002838 return_unused_surplus_pages(h, (unsigned long) -delta);
Mel Gormanfc1b8a72008-07-23 21:27:22 -07002839
2840out:
2841 spin_unlock(&hugetlb_lock);
2842 return ret;
2843}
2844
Andy Whitcroft84afd992008-07-23 21:27:32 -07002845static void hugetlb_vm_op_open(struct vm_area_struct *vma)
2846{
Joonsoo Kimf522c3a2013-09-11 14:21:53 -07002847 struct resv_map *resv = vma_resv_map(vma);
Andy Whitcroft84afd992008-07-23 21:27:32 -07002848
2849 /*
2850 * This new VMA should share its siblings reservation map if present.
2851 * The VMA will only ever have a valid reservation map pointer where
2852 * it is being copied for another still existing VMA. As that VMA
Lucas De Marchi25985ed2011-03-30 22:57:33 -03002853 * has a reference to the reservation map it cannot disappear until
Andy Whitcroft84afd992008-07-23 21:27:32 -07002854 * after this open call completes. It is therefore safe to take a
2855 * new reference here without additional locking.
2856 */
Joonsoo Kim4e35f482014-04-03 14:47:30 -07002857 if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
Joonsoo Kimf522c3a2013-09-11 14:21:53 -07002858 kref_get(&resv->refs);
Andy Whitcroft84afd992008-07-23 21:27:32 -07002859}
2860
Mel Gormana1e78772008-07-23 21:27:23 -07002861static void hugetlb_vm_op_close(struct vm_area_struct *vma)
2862{
Andi Kleena5516432008-07-23 21:27:41 -07002863 struct hstate *h = hstate_vma(vma);
Joonsoo Kimf522c3a2013-09-11 14:21:53 -07002864 struct resv_map *resv = vma_resv_map(vma);
David Gibson90481622012-03-21 16:34:12 -07002865 struct hugepage_subpool *spool = subpool_vma(vma);
Joonsoo Kim4e35f482014-04-03 14:47:30 -07002866 unsigned long reserve, start, end;
Mike Kravetz1c5ecae2015-04-15 16:13:39 -07002867 long gbl_reserve;
Andy Whitcroft84afd992008-07-23 21:27:32 -07002868
Joonsoo Kim4e35f482014-04-03 14:47:30 -07002869 if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER))
2870 return;
Andy Whitcroft84afd992008-07-23 21:27:32 -07002871
Joonsoo Kim4e35f482014-04-03 14:47:30 -07002872 start = vma_hugecache_offset(h, vma, vma->vm_start);
2873 end = vma_hugecache_offset(h, vma, vma->vm_end);
Andy Whitcroft84afd992008-07-23 21:27:32 -07002874
Joonsoo Kim4e35f482014-04-03 14:47:30 -07002875 reserve = (end - start) - region_count(resv, start, end);
Andy Whitcroft84afd992008-07-23 21:27:32 -07002876
Joonsoo Kim4e35f482014-04-03 14:47:30 -07002877 kref_put(&resv->refs, resv_map_release);
2878
2879 if (reserve) {
Mike Kravetz1c5ecae2015-04-15 16:13:39 -07002880 /*
2881 * Decrement reserve counts. The global reserve count may be
2882 * adjusted if the subpool has a minimum size.
2883 */
2884 gbl_reserve = hugepage_subpool_put_pages(spool, reserve);
2885 hugetlb_acct_memory(h, -gbl_reserve);
Andy Whitcroft84afd992008-07-23 21:27:32 -07002886 }
Mel Gormana1e78772008-07-23 21:27:23 -07002887}
2888
Linus Torvalds1da177e2005-04-16 15:20:36 -07002889/*
2890 * We cannot handle pagefaults against hugetlb pages at all. They cause
2891 * handle_mm_fault() to try to instantiate regular-sized pages in the
2892 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
2893 * this far.
2894 */
Nick Piggind0217ac2007-07-19 01:47:03 -07002895static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002896{
2897 BUG();
Nick Piggind0217ac2007-07-19 01:47:03 -07002898 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002899}
2900
Alexey Dobriyanf0f37e22009-09-27 22:29:37 +04002901const struct vm_operations_struct hugetlb_vm_ops = {
Nick Piggind0217ac2007-07-19 01:47:03 -07002902 .fault = hugetlb_vm_op_fault,
Andy Whitcroft84afd992008-07-23 21:27:32 -07002903 .open = hugetlb_vm_op_open,
Mel Gormana1e78772008-07-23 21:27:23 -07002904 .close = hugetlb_vm_op_close,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002905};
2906
David Gibson1e8f8892006-01-06 00:10:44 -08002907static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
2908 int writable)
David Gibson63551ae2005-06-21 17:14:44 -07002909{
2910 pte_t entry;
2911
David Gibson1e8f8892006-01-06 00:10:44 -08002912 if (writable) {
Gerald Schaefer106c9922013-04-29 15:07:23 -07002913 entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page,
2914 vma->vm_page_prot)));
David Gibson63551ae2005-06-21 17:14:44 -07002915 } else {
Gerald Schaefer106c9922013-04-29 15:07:23 -07002916 entry = huge_pte_wrprotect(mk_huge_pte(page,
2917 vma->vm_page_prot));
David Gibson63551ae2005-06-21 17:14:44 -07002918 }
2919 entry = pte_mkyoung(entry);
2920 entry = pte_mkhuge(entry);
Chris Metcalfd9ed9fa2012-04-01 14:01:34 -04002921 entry = arch_make_huge_pte(entry, vma, page, writable);
David Gibson63551ae2005-06-21 17:14:44 -07002922
2923 return entry;
2924}
2925
David Gibson1e8f8892006-01-06 00:10:44 -08002926static void set_huge_ptep_writable(struct vm_area_struct *vma,
2927 unsigned long address, pte_t *ptep)
2928{
2929 pte_t entry;
2930
Gerald Schaefer106c9922013-04-29 15:07:23 -07002931 entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep)));
Chris Forbes32f84522011-07-25 17:12:14 -07002932 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
Russell King4b3073e2009-12-18 16:40:18 +00002933 update_mmu_cache(vma, address, ptep);
David Gibson1e8f8892006-01-06 00:10:44 -08002934}
2935
Naoya Horiguchi4a705fe2014-06-23 13:22:03 -07002936static int is_hugetlb_entry_migration(pte_t pte)
2937{
2938 swp_entry_t swp;
2939
2940 if (huge_pte_none(pte) || pte_present(pte))
2941 return 0;
2942 swp = pte_to_swp_entry(pte);
2943 if (non_swap_entry(swp) && is_migration_entry(swp))
2944 return 1;
2945 else
2946 return 0;
2947}
2948
2949static int is_hugetlb_entry_hwpoisoned(pte_t pte)
2950{
2951 swp_entry_t swp;
2952
2953 if (huge_pte_none(pte) || pte_present(pte))
2954 return 0;
2955 swp = pte_to_swp_entry(pte);
2956 if (non_swap_entry(swp) && is_hwpoison_entry(swp))
2957 return 1;
2958 else
2959 return 0;
2960}
David Gibson1e8f8892006-01-06 00:10:44 -08002961
David Gibson63551ae2005-06-21 17:14:44 -07002962int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
2963 struct vm_area_struct *vma)
2964{
2965 pte_t *src_pte, *dst_pte, entry;
2966 struct page *ptepage;
Hugh Dickins1c598272005-10-19 21:23:43 -07002967 unsigned long addr;
David Gibson1e8f8892006-01-06 00:10:44 -08002968 int cow;
Andi Kleena5516432008-07-23 21:27:41 -07002969 struct hstate *h = hstate_vma(vma);
2970 unsigned long sz = huge_page_size(h);
Andreas Sandberge8569dd2014-01-21 15:49:09 -08002971 unsigned long mmun_start; /* For mmu_notifiers */
2972 unsigned long mmun_end; /* For mmu_notifiers */
2973 int ret = 0;
David Gibson1e8f8892006-01-06 00:10:44 -08002974
2975 cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
David Gibson63551ae2005-06-21 17:14:44 -07002976
Andreas Sandberge8569dd2014-01-21 15:49:09 -08002977 mmun_start = vma->vm_start;
2978 mmun_end = vma->vm_end;
2979 if (cow)
2980 mmu_notifier_invalidate_range_start(src, mmun_start, mmun_end);
2981
Andi Kleena5516432008-07-23 21:27:41 -07002982 for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002983 spinlock_t *src_ptl, *dst_ptl;
Hugh Dickinsc74df322005-10-29 18:16:23 -07002984 src_pte = huge_pte_offset(src, addr);
2985 if (!src_pte)
2986 continue;
Andi Kleena5516432008-07-23 21:27:41 -07002987 dst_pte = huge_pte_alloc(dst, addr, sz);
Andreas Sandberge8569dd2014-01-21 15:49:09 -08002988 if (!dst_pte) {
2989 ret = -ENOMEM;
2990 break;
2991 }
Larry Woodmanc5c99422008-01-24 05:49:25 -08002992
2993 /* If the pagetables are shared don't copy or take references */
2994 if (dst_pte == src_pte)
2995 continue;
2996
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002997 dst_ptl = huge_pte_lock(h, dst, dst_pte);
2998 src_ptl = huge_pte_lockptr(h, src, src_pte);
2999 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
Naoya Horiguchi4a705fe2014-06-23 13:22:03 -07003000 entry = huge_ptep_get(src_pte);
3001 if (huge_pte_none(entry)) { /* skip none entry */
3002 ;
3003 } else if (unlikely(is_hugetlb_entry_migration(entry) ||
3004 is_hugetlb_entry_hwpoisoned(entry))) {
3005 swp_entry_t swp_entry = pte_to_swp_entry(entry);
3006
3007 if (is_write_migration_entry(swp_entry) && cow) {
3008 /*
3009 * COW mappings require pages in both
3010 * parent and child to be set to read.
3011 */
3012 make_migration_entry_read(&swp_entry);
3013 entry = swp_entry_to_pte(swp_entry);
3014 set_huge_pte_at(src, addr, src_pte, entry);
3015 }
3016 set_huge_pte_at(dst, addr, dst_pte, entry);
3017 } else {
Joerg Roedel34ee6452014-11-13 13:46:09 +11003018 if (cow) {
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07003019 huge_ptep_set_wrprotect(src, addr, src_pte);
Joerg Roedel34ee6452014-11-13 13:46:09 +11003020 mmu_notifier_invalidate_range(src, mmun_start,
3021 mmun_end);
3022 }
Naoya Horiguchi0253d632014-07-23 14:00:19 -07003023 entry = huge_ptep_get(src_pte);
Hugh Dickins1c598272005-10-19 21:23:43 -07003024 ptepage = pte_page(entry);
3025 get_page(ptepage);
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09003026 page_dup_rmap(ptepage);
Hugh Dickins1c598272005-10-19 21:23:43 -07003027 set_huge_pte_at(dst, addr, dst_pte, entry);
3028 }
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003029 spin_unlock(src_ptl);
3030 spin_unlock(dst_ptl);
David Gibson63551ae2005-06-21 17:14:44 -07003031 }
David Gibson63551ae2005-06-21 17:14:44 -07003032
Andreas Sandberge8569dd2014-01-21 15:49:09 -08003033 if (cow)
3034 mmu_notifier_invalidate_range_end(src, mmun_start, mmun_end);
3035
3036 return ret;
David Gibson63551ae2005-06-21 17:14:44 -07003037}
3038
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07003039void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
3040 unsigned long start, unsigned long end,
3041 struct page *ref_page)
David Gibson63551ae2005-06-21 17:14:44 -07003042{
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07003043 int force_flush = 0;
David Gibson63551ae2005-06-21 17:14:44 -07003044 struct mm_struct *mm = vma->vm_mm;
3045 unsigned long address;
David Gibsonc7546f82005-08-05 11:59:35 -07003046 pte_t *ptep;
David Gibson63551ae2005-06-21 17:14:44 -07003047 pte_t pte;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003048 spinlock_t *ptl;
David Gibson63551ae2005-06-21 17:14:44 -07003049 struct page *page;
Andi Kleena5516432008-07-23 21:27:41 -07003050 struct hstate *h = hstate_vma(vma);
3051 unsigned long sz = huge_page_size(h);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07003052 const unsigned long mmun_start = start; /* For mmu_notifiers */
3053 const unsigned long mmun_end = end; /* For mmu_notifiers */
Andi Kleena5516432008-07-23 21:27:41 -07003054
David Gibson63551ae2005-06-21 17:14:44 -07003055 WARN_ON(!is_vm_hugetlb_page(vma));
Andi Kleena5516432008-07-23 21:27:41 -07003056 BUG_ON(start & ~huge_page_mask(h));
3057 BUG_ON(end & ~huge_page_mask(h));
David Gibson63551ae2005-06-21 17:14:44 -07003058
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07003059 tlb_start_vma(tlb, vma);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07003060 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
Hillf Danton569f48b2014-12-10 15:44:41 -08003061 address = start;
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07003062again:
Hillf Danton569f48b2014-12-10 15:44:41 -08003063 for (; address < end; address += sz) {
David Gibsonc7546f82005-08-05 11:59:35 -07003064 ptep = huge_pte_offset(mm, address);
Adam Litke4c887262005-10-29 18:16:46 -07003065 if (!ptep)
David Gibsonc7546f82005-08-05 11:59:35 -07003066 continue;
3067
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003068 ptl = huge_pte_lock(h, mm, ptep);
Chen, Kenneth W39dde652006-12-06 20:32:03 -08003069 if (huge_pmd_unshare(mm, &address, ptep))
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003070 goto unlock;
Chen, Kenneth W39dde652006-12-06 20:32:03 -08003071
Hillf Danton66293262012-03-23 15:01:48 -07003072 pte = huge_ptep_get(ptep);
3073 if (huge_pte_none(pte))
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003074 goto unlock;
Hillf Danton66293262012-03-23 15:01:48 -07003075
3076 /*
Naoya Horiguchi9fbc1f62015-02-11 15:25:32 -08003077 * Migrating hugepage or HWPoisoned hugepage is already
3078 * unmapped and its refcount is dropped, so just clear pte here.
Hillf Danton66293262012-03-23 15:01:48 -07003079 */
Naoya Horiguchi9fbc1f62015-02-11 15:25:32 -08003080 if (unlikely(!pte_present(pte))) {
Gerald Schaefer106c9922013-04-29 15:07:23 -07003081 huge_pte_clear(mm, address, ptep);
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003082 goto unlock;
Naoya Horiguchi8c4894c2012-12-12 13:52:28 -08003083 }
Hillf Danton66293262012-03-23 15:01:48 -07003084
3085 page = pte_page(pte);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003086 /*
3087 * If a reference page is supplied, it is because a specific
3088 * page is being unmapped, not a range. Ensure the page we
3089 * are about to unmap is the actual page of interest.
3090 */
3091 if (ref_page) {
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003092 if (page != ref_page)
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003093 goto unlock;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003094
3095 /*
3096 * Mark the VMA as having unmapped its page so that
3097 * future faults in this VMA will fail rather than
3098 * looking like data was lost
3099 */
3100 set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
3101 }
3102
David Gibsonc7546f82005-08-05 11:59:35 -07003103 pte = huge_ptep_get_and_clear(mm, address, ptep);
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07003104 tlb_remove_tlb_entry(tlb, ptep, address);
Gerald Schaefer106c9922013-04-29 15:07:23 -07003105 if (huge_pte_dirty(pte))
Ken Chen6649a382007-02-08 14:20:27 -08003106 set_page_dirty(page);
Hillf Danton9e811302012-03-21 16:34:03 -07003107
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07003108 page_remove_rmap(page);
3109 force_flush = !__tlb_remove_page(tlb, page);
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003110 if (force_flush) {
Hillf Danton569f48b2014-12-10 15:44:41 -08003111 address += sz;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003112 spin_unlock(ptl);
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07003113 break;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003114 }
Hillf Danton9e811302012-03-21 16:34:03 -07003115 /* Bail out after unmapping reference page if supplied */
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003116 if (ref_page) {
3117 spin_unlock(ptl);
Hillf Danton9e811302012-03-21 16:34:03 -07003118 break;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003119 }
3120unlock:
3121 spin_unlock(ptl);
David Gibson63551ae2005-06-21 17:14:44 -07003122 }
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07003123 /*
3124 * mmu_gather ran out of room to batch pages, we break out of
3125 * the PTE lock to avoid doing the potential expensive TLB invalidate
3126 * and page-free while holding it.
3127 */
3128 if (force_flush) {
3129 force_flush = 0;
3130 tlb_flush_mmu(tlb);
3131 if (address < end && !ref_page)
3132 goto again;
Chen, Kenneth Wfe1668a2006-10-04 02:15:24 -07003133 }
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07003134 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07003135 tlb_end_vma(tlb, vma);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003136}
David Gibson63551ae2005-06-21 17:14:44 -07003137
Mel Gormand8333522012-07-31 16:46:20 -07003138void __unmap_hugepage_range_final(struct mmu_gather *tlb,
3139 struct vm_area_struct *vma, unsigned long start,
3140 unsigned long end, struct page *ref_page)
3141{
3142 __unmap_hugepage_range(tlb, vma, start, end, ref_page);
3143
3144 /*
3145 * Clear this flag so that x86's huge_pmd_share page_table_shareable
3146 * test will fail on a vma being torn down, and not grab a page table
3147 * on its way out. We're lucky that the flag has such an appropriate
3148 * name, and can in fact be safely cleared here. We could clear it
3149 * before the __unmap_hugepage_range above, but all that's necessary
Davidlohr Buesoc8c06ef2014-12-12 16:54:24 -08003150 * is to clear it before releasing the i_mmap_rwsem. This works
Mel Gormand8333522012-07-31 16:46:20 -07003151 * because in the context this is called, the VMA is about to be
Davidlohr Buesoc8c06ef2014-12-12 16:54:24 -08003152 * destroyed and the i_mmap_rwsem is held.
Mel Gormand8333522012-07-31 16:46:20 -07003153 */
3154 vma->vm_flags &= ~VM_MAYSHARE;
3155}
3156
Chen, Kenneth W502717f2006-10-11 01:20:46 -07003157void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003158 unsigned long end, struct page *ref_page)
Chen, Kenneth W502717f2006-10-11 01:20:46 -07003159{
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07003160 struct mm_struct *mm;
3161 struct mmu_gather tlb;
3162
3163 mm = vma->vm_mm;
3164
Linus Torvalds2b047252013-08-15 11:42:25 -07003165 tlb_gather_mmu(&tlb, mm, start, end);
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07003166 __unmap_hugepage_range(&tlb, vma, start, end, ref_page);
3167 tlb_finish_mmu(&tlb, start, end);
Chen, Kenneth W502717f2006-10-11 01:20:46 -07003168}
3169
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003170/*
3171 * This is called when the original mapper is failing to COW a MAP_PRIVATE
3172 * mappping it owns the reserve page for. The intention is to unmap the page
3173 * from other VMAs and let the children be SIGKILLed if they are faulting the
3174 * same region.
3175 */
Davidlohr Bueso2f4612a2014-08-06 16:06:45 -07003176static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
3177 struct page *page, unsigned long address)
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003178{
Adam Litke75266742008-11-12 13:24:56 -08003179 struct hstate *h = hstate_vma(vma);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003180 struct vm_area_struct *iter_vma;
3181 struct address_space *mapping;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003182 pgoff_t pgoff;
3183
3184 /*
3185 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
3186 * from page cache lookup which is in HPAGE_SIZE units.
3187 */
Adam Litke75266742008-11-12 13:24:56 -08003188 address = address & huge_page_mask(h);
Michal Hocko36e4f202012-10-08 16:33:31 -07003189 pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) +
3190 vma->vm_pgoff;
Al Viro496ad9a2013-01-23 17:07:38 -05003191 mapping = file_inode(vma->vm_file)->i_mapping;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003192
Mel Gorman4eb2b1d2009-12-14 17:59:53 -08003193 /*
3194 * Take the mapping lock for the duration of the table walk. As
3195 * this mapping should be shared between all the VMAs,
3196 * __unmap_hugepage_range() is called as the lock is already held
3197 */
Davidlohr Bueso83cde9e2014-12-12 16:54:21 -08003198 i_mmap_lock_write(mapping);
Michel Lespinasse6b2dbba2012-10-08 16:31:25 -07003199 vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) {
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003200 /* Do not unmap the current VMA */
3201 if (iter_vma == vma)
3202 continue;
3203
3204 /*
Mel Gorman2f84a892015-10-01 15:36:57 -07003205 * Shared VMAs have their own reserves and do not affect
3206 * MAP_PRIVATE accounting but it is possible that a shared
3207 * VMA is using the same page so check and skip such VMAs.
3208 */
3209 if (iter_vma->vm_flags & VM_MAYSHARE)
3210 continue;
3211
3212 /*
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003213 * Unmap the page from other VMAs without their own reserves.
3214 * They get marked to be SIGKILLed if they fault in these
3215 * areas. This is because a future no-page fault on this VMA
3216 * could insert a zeroed page instead of the data existing
3217 * from the time of fork. This would look like data corruption
3218 */
3219 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07003220 unmap_hugepage_range(iter_vma, address,
3221 address + huge_page_size(h), page);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003222 }
Davidlohr Bueso83cde9e2014-12-12 16:54:21 -08003223 i_mmap_unlock_write(mapping);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003224}
3225
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09003226/*
3227 * Hugetlb_cow() should be called with page lock of the original hugepage held.
Michal Hockoef009b22012-01-10 15:07:21 -08003228 * Called with hugetlb_instantiation_mutex held and pte_page locked so we
3229 * cannot race with other handlers or page migration.
3230 * Keep the pte_same checks anyway to make transition from the mutex easier.
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09003231 */
David Gibson1e8f8892006-01-06 00:10:44 -08003232static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003233 unsigned long address, pte_t *ptep, pte_t pte,
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003234 struct page *pagecache_page, spinlock_t *ptl)
David Gibson1e8f8892006-01-06 00:10:44 -08003235{
Andi Kleena5516432008-07-23 21:27:41 -07003236 struct hstate *h = hstate_vma(vma);
David Gibson1e8f8892006-01-06 00:10:44 -08003237 struct page *old_page, *new_page;
Davidlohr Buesoad4404a2014-08-06 16:06:47 -07003238 int ret = 0, outside_reserve = 0;
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07003239 unsigned long mmun_start; /* For mmu_notifiers */
3240 unsigned long mmun_end; /* For mmu_notifiers */
David Gibson1e8f8892006-01-06 00:10:44 -08003241
3242 old_page = pte_page(pte);
3243
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003244retry_avoidcopy:
David Gibson1e8f8892006-01-06 00:10:44 -08003245 /* If no-one else is actually using this page, avoid the copy
3246 * and just make the page writable */
Joonsoo Kim37a21402013-09-11 14:21:04 -07003247 if (page_mapcount(old_page) == 1 && PageAnon(old_page)) {
3248 page_move_anon_rmap(old_page, vma, address);
David Gibson1e8f8892006-01-06 00:10:44 -08003249 set_huge_ptep_writable(vma, address, ptep);
Nick Piggin83c54072007-07-19 01:47:05 -07003250 return 0;
David Gibson1e8f8892006-01-06 00:10:44 -08003251 }
3252
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003253 /*
3254 * If the process that created a MAP_PRIVATE mapping is about to
3255 * perform a COW due to a shared page count, attempt to satisfy
3256 * the allocation without using the existing reserves. The pagecache
3257 * page is used to determine if the reserve at this address was
3258 * consumed or not. If reserves were used, a partial faulted mapping
3259 * at the time of fork() could consume its reserves on COW instead
3260 * of the full address range.
3261 */
Joonsoo Kim5944d012013-09-11 14:21:55 -07003262 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003263 old_page != pagecache_page)
3264 outside_reserve = 1;
3265
David Gibson1e8f8892006-01-06 00:10:44 -08003266 page_cache_get(old_page);
Larry Woodmanb76c8cf2009-12-14 17:59:37 -08003267
Davidlohr Buesoad4404a2014-08-06 16:06:47 -07003268 /*
3269 * Drop page table lock as buddy allocator may be called. It will
3270 * be acquired again before returning to the caller, as expected.
3271 */
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003272 spin_unlock(ptl);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003273 new_page = alloc_huge_page(vma, address, outside_reserve);
David Gibson1e8f8892006-01-06 00:10:44 -08003274
Adam Litke2fc39ce2007-11-14 16:59:39 -08003275 if (IS_ERR(new_page)) {
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003276 /*
3277 * If a process owning a MAP_PRIVATE mapping fails to COW,
3278 * it is due to references held by a child and an insufficient
3279 * huge page pool. To guarantee the original mappers
3280 * reliability, unmap the page from child processes. The child
3281 * may get SIGKILLed if it later faults.
3282 */
3283 if (outside_reserve) {
Davidlohr Buesoad4404a2014-08-06 16:06:47 -07003284 page_cache_release(old_page);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003285 BUG_ON(huge_pte_none(pte));
Davidlohr Bueso2f4612a2014-08-06 16:06:45 -07003286 unmap_ref_private(mm, vma, old_page, address);
3287 BUG_ON(huge_pte_none(pte));
3288 spin_lock(ptl);
3289 ptep = huge_pte_offset(mm, address & huge_page_mask(h));
3290 if (likely(ptep &&
3291 pte_same(huge_ptep_get(ptep), pte)))
3292 goto retry_avoidcopy;
3293 /*
3294 * race occurs while re-acquiring page table
3295 * lock, and our job is done.
3296 */
3297 return 0;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003298 }
3299
Davidlohr Buesoad4404a2014-08-06 16:06:47 -07003300 ret = (PTR_ERR(new_page) == -ENOMEM) ?
3301 VM_FAULT_OOM : VM_FAULT_SIGBUS;
3302 goto out_release_old;
David Gibson1e8f8892006-01-06 00:10:44 -08003303 }
3304
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09003305 /*
3306 * When the original hugepage is shared one, it does not have
3307 * anon_vma prepared.
3308 */
Dean Nelson44e2aa92010-10-26 14:22:08 -07003309 if (unlikely(anon_vma_prepare(vma))) {
Davidlohr Buesoad4404a2014-08-06 16:06:47 -07003310 ret = VM_FAULT_OOM;
3311 goto out_release_all;
Dean Nelson44e2aa92010-10-26 14:22:08 -07003312 }
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09003313
Andrea Arcangeli47ad8472011-01-13 15:46:47 -08003314 copy_user_huge_page(new_page, old_page, address, vma,
3315 pages_per_huge_page(h));
Nick Piggin0ed361d2008-02-04 22:29:34 -08003316 __SetPageUptodate(new_page);
Naoya Horiguchibcc54222015-04-15 16:14:38 -07003317 set_page_huge_active(new_page);
David Gibson1e8f8892006-01-06 00:10:44 -08003318
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07003319 mmun_start = address & huge_page_mask(h);
3320 mmun_end = mmun_start + huge_page_size(h);
3321 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
Davidlohr Buesoad4404a2014-08-06 16:06:47 -07003322
Larry Woodmanb76c8cf2009-12-14 17:59:37 -08003323 /*
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003324 * Retake the page table lock to check for racing updates
Larry Woodmanb76c8cf2009-12-14 17:59:37 -08003325 * before the page tables are altered
3326 */
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003327 spin_lock(ptl);
Andi Kleena5516432008-07-23 21:27:41 -07003328 ptep = huge_pte_offset(mm, address & huge_page_mask(h));
Naoya Horiguchia9af0c52014-04-07 15:36:54 -07003329 if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) {
Joonsoo Kim07443a82013-09-11 14:21:58 -07003330 ClearPagePrivate(new_page);
3331
David Gibson1e8f8892006-01-06 00:10:44 -08003332 /* Break COW */
Gerald Schaefer8fe627e2008-04-28 02:13:28 -07003333 huge_ptep_clear_flush(vma, address, ptep);
Joerg Roedel34ee6452014-11-13 13:46:09 +11003334 mmu_notifier_invalidate_range(mm, mmun_start, mmun_end);
David Gibson1e8f8892006-01-06 00:10:44 -08003335 set_huge_pte_at(mm, address, ptep,
3336 make_huge_pte(vma, new_page, 1));
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09003337 page_remove_rmap(old_page);
Naoya Horiguchicd67f0d2010-09-10 13:23:04 +09003338 hugepage_add_new_anon_rmap(new_page, vma, address);
David Gibson1e8f8892006-01-06 00:10:44 -08003339 /* Make the old page be freed below */
3340 new_page = old_page;
3341 }
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003342 spin_unlock(ptl);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07003343 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
Davidlohr Buesoad4404a2014-08-06 16:06:47 -07003344out_release_all:
David Gibson1e8f8892006-01-06 00:10:44 -08003345 page_cache_release(new_page);
Davidlohr Buesoad4404a2014-08-06 16:06:47 -07003346out_release_old:
David Gibson1e8f8892006-01-06 00:10:44 -08003347 page_cache_release(old_page);
Joonsoo Kim83120342013-09-11 14:21:57 -07003348
Davidlohr Buesoad4404a2014-08-06 16:06:47 -07003349 spin_lock(ptl); /* Caller expects lock to be held */
3350 return ret;
David Gibson1e8f8892006-01-06 00:10:44 -08003351}
3352
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003353/* Return the pagecache page at a given address within a VMA */
Andi Kleena5516432008-07-23 21:27:41 -07003354static struct page *hugetlbfs_pagecache_page(struct hstate *h,
3355 struct vm_area_struct *vma, unsigned long address)
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003356{
3357 struct address_space *mapping;
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -07003358 pgoff_t idx;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003359
3360 mapping = vma->vm_file->f_mapping;
Andi Kleena5516432008-07-23 21:27:41 -07003361 idx = vma_hugecache_offset(h, vma, address);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003362
3363 return find_lock_page(mapping, idx);
3364}
3365
Hugh Dickins3ae77f42009-09-21 17:03:33 -07003366/*
3367 * Return whether there is a pagecache page to back given address within VMA.
3368 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
3369 */
3370static bool hugetlbfs_pagecache_present(struct hstate *h,
Hugh Dickins2a15efc2009-09-21 17:03:27 -07003371 struct vm_area_struct *vma, unsigned long address)
3372{
3373 struct address_space *mapping;
3374 pgoff_t idx;
3375 struct page *page;
3376
3377 mapping = vma->vm_file->f_mapping;
3378 idx = vma_hugecache_offset(h, vma, address);
3379
3380 page = find_get_page(mapping, idx);
3381 if (page)
3382 put_page(page);
3383 return page != NULL;
3384}
3385
Mike Kravetzab76ad52015-09-08 15:01:50 -07003386int huge_add_to_page_cache(struct page *page, struct address_space *mapping,
3387 pgoff_t idx)
3388{
3389 struct inode *inode = mapping->host;
3390 struct hstate *h = hstate_inode(inode);
3391 int err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
3392
3393 if (err)
3394 return err;
3395 ClearPagePrivate(page);
3396
3397 spin_lock(&inode->i_lock);
3398 inode->i_blocks += blocks_per_huge_page(h);
3399 spin_unlock(&inode->i_lock);
3400 return 0;
3401}
3402
Robert P. J. Daya1ed3dd2007-07-17 04:03:33 -07003403static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
Davidlohr Bueso8382d912014-04-03 14:47:31 -07003404 struct address_space *mapping, pgoff_t idx,
3405 unsigned long address, pte_t *ptep, unsigned int flags)
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01003406{
Andi Kleena5516432008-07-23 21:27:41 -07003407 struct hstate *h = hstate_vma(vma);
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01003408 int ret = VM_FAULT_SIGBUS;
Hillf Danton409eb8c2012-01-20 14:34:13 -08003409 int anon_rmap = 0;
Adam Litke4c887262005-10-29 18:16:46 -07003410 unsigned long size;
Adam Litke4c887262005-10-29 18:16:46 -07003411 struct page *page;
David Gibson1e8f8892006-01-06 00:10:44 -08003412 pte_t new_pte;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003413 spinlock_t *ptl;
Adam Litke4c887262005-10-29 18:16:46 -07003414
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003415 /*
3416 * Currently, we are forced to kill the process in the event the
3417 * original mapper has unmapped pages from the child due to a failed
Lucas De Marchi25985ed2011-03-30 22:57:33 -03003418 * COW. Warn that such a situation has occurred as it may not be obvious
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003419 */
3420 if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
Andrew Mortonffb22af2013-02-22 16:32:08 -08003421 pr_warning("PID %d killed due to inadequate hugepage pool\n",
3422 current->pid);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003423 return ret;
3424 }
3425
Adam Litke4c887262005-10-29 18:16:46 -07003426 /*
3427 * Use page lock to guard against racing truncation
3428 * before we get page_table_lock.
3429 */
Christoph Lameter6bda6662006-01-06 00:10:49 -08003430retry:
3431 page = find_lock_page(mapping, idx);
3432 if (!page) {
Andi Kleena5516432008-07-23 21:27:41 -07003433 size = i_size_read(mapping->host) >> huge_page_shift(h);
Hugh Dickinsebed4bf2006-10-28 10:38:43 -07003434 if (idx >= size)
3435 goto out;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003436 page = alloc_huge_page(vma, address, 0);
Adam Litke2fc39ce2007-11-14 16:59:39 -08003437 if (IS_ERR(page)) {
Aneesh Kumar K.V76dcee72012-07-31 16:41:57 -07003438 ret = PTR_ERR(page);
3439 if (ret == -ENOMEM)
3440 ret = VM_FAULT_OOM;
3441 else
3442 ret = VM_FAULT_SIGBUS;
Christoph Lameter6bda6662006-01-06 00:10:49 -08003443 goto out;
3444 }
Andrea Arcangeli47ad8472011-01-13 15:46:47 -08003445 clear_huge_page(page, address, pages_per_huge_page(h));
Nick Piggin0ed361d2008-02-04 22:29:34 -08003446 __SetPageUptodate(page);
Naoya Horiguchibcc54222015-04-15 16:14:38 -07003447 set_page_huge_active(page);
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01003448
Mel Gormanf83a2752009-05-28 14:34:40 -07003449 if (vma->vm_flags & VM_MAYSHARE) {
Mike Kravetzab76ad52015-09-08 15:01:50 -07003450 int err = huge_add_to_page_cache(page, mapping, idx);
Christoph Lameter6bda6662006-01-06 00:10:49 -08003451 if (err) {
3452 put_page(page);
Christoph Lameter6bda6662006-01-06 00:10:49 -08003453 if (err == -EEXIST)
3454 goto retry;
3455 goto out;
3456 }
Mel Gorman23be7462010-04-23 13:17:56 -04003457 } else {
Christoph Lameter6bda6662006-01-06 00:10:49 -08003458 lock_page(page);
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09003459 if (unlikely(anon_vma_prepare(vma))) {
3460 ret = VM_FAULT_OOM;
3461 goto backout_unlocked;
3462 }
Hillf Danton409eb8c2012-01-20 14:34:13 -08003463 anon_rmap = 1;
Mel Gorman23be7462010-04-23 13:17:56 -04003464 }
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09003465 } else {
Naoya Horiguchi998b4382010-09-08 10:19:32 +09003466 /*
3467 * If memory error occurs between mmap() and fault, some process
3468 * don't have hwpoisoned swap entry for errored virtual address.
3469 * So we need to block hugepage fault by PG_hwpoison bit check.
3470 */
3471 if (unlikely(PageHWPoison(page))) {
Chris Forbes32f84522011-07-25 17:12:14 -07003472 ret = VM_FAULT_HWPOISON |
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07003473 VM_FAULT_SET_HINDEX(hstate_index(h));
Naoya Horiguchi998b4382010-09-08 10:19:32 +09003474 goto backout_unlocked;
3475 }
Christoph Lameter6bda6662006-01-06 00:10:49 -08003476 }
David Gibson1e8f8892006-01-06 00:10:44 -08003477
Andy Whitcroft57303d82008-08-12 15:08:47 -07003478 /*
3479 * If we are going to COW a private mapping later, we examine the
3480 * pending reservations for this page now. This will ensure that
3481 * any allocations necessary to record that reservation occur outside
3482 * the spinlock.
3483 */
Mike Kravetz5e911372015-09-08 15:01:28 -07003484 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
Andy Whitcroft2b267362008-08-12 15:08:49 -07003485 if (vma_needs_reservation(h, vma, address) < 0) {
3486 ret = VM_FAULT_OOM;
3487 goto backout_unlocked;
3488 }
Mike Kravetz5e911372015-09-08 15:01:28 -07003489 /* Just decrements count, does not deallocate */
Mike Kravetzfeba16e2015-09-08 15:01:31 -07003490 vma_end_reservation(h, vma, address);
Mike Kravetz5e911372015-09-08 15:01:28 -07003491 }
Andy Whitcroft57303d82008-08-12 15:08:47 -07003492
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003493 ptl = huge_pte_lockptr(h, mm, ptep);
3494 spin_lock(ptl);
Andi Kleena5516432008-07-23 21:27:41 -07003495 size = i_size_read(mapping->host) >> huge_page_shift(h);
Adam Litke4c887262005-10-29 18:16:46 -07003496 if (idx >= size)
3497 goto backout;
3498
Nick Piggin83c54072007-07-19 01:47:05 -07003499 ret = 0;
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07003500 if (!huge_pte_none(huge_ptep_get(ptep)))
Adam Litke4c887262005-10-29 18:16:46 -07003501 goto backout;
3502
Joonsoo Kim07443a82013-09-11 14:21:58 -07003503 if (anon_rmap) {
3504 ClearPagePrivate(page);
Hillf Danton409eb8c2012-01-20 14:34:13 -08003505 hugepage_add_new_anon_rmap(page, vma, address);
Choi Gi-yongac714902014-04-07 15:37:36 -07003506 } else
Hillf Danton409eb8c2012-01-20 14:34:13 -08003507 page_dup_rmap(page);
David Gibson1e8f8892006-01-06 00:10:44 -08003508 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
3509 && (vma->vm_flags & VM_SHARED)));
3510 set_huge_pte_at(mm, address, ptep, new_pte);
3511
Hugh Dickins788c7df2009-06-23 13:49:05 +01003512 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
David Gibson1e8f8892006-01-06 00:10:44 -08003513 /* Optimization, do the COW without a second fault */
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003514 ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page, ptl);
David Gibson1e8f8892006-01-06 00:10:44 -08003515 }
3516
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003517 spin_unlock(ptl);
Adam Litke4c887262005-10-29 18:16:46 -07003518 unlock_page(page);
3519out:
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01003520 return ret;
Adam Litke4c887262005-10-29 18:16:46 -07003521
3522backout:
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003523 spin_unlock(ptl);
Andy Whitcroft2b267362008-08-12 15:08:49 -07003524backout_unlocked:
Adam Litke4c887262005-10-29 18:16:46 -07003525 unlock_page(page);
3526 put_page(page);
3527 goto out;
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01003528}
3529
Davidlohr Bueso8382d912014-04-03 14:47:31 -07003530#ifdef CONFIG_SMP
Mike Kravetzc672c7f2015-09-08 15:01:35 -07003531u32 hugetlb_fault_mutex_hash(struct hstate *h, struct mm_struct *mm,
Davidlohr Bueso8382d912014-04-03 14:47:31 -07003532 struct vm_area_struct *vma,
3533 struct address_space *mapping,
3534 pgoff_t idx, unsigned long address)
3535{
3536 unsigned long key[2];
3537 u32 hash;
3538
3539 if (vma->vm_flags & VM_SHARED) {
3540 key[0] = (unsigned long) mapping;
3541 key[1] = idx;
3542 } else {
3543 key[0] = (unsigned long) mm;
3544 key[1] = address >> huge_page_shift(h);
3545 }
3546
3547 hash = jhash2((u32 *)&key, sizeof(key)/sizeof(u32), 0);
3548
3549 return hash & (num_fault_mutexes - 1);
3550}
3551#else
3552/*
3553 * For uniprocesor systems we always use a single mutex, so just
3554 * return 0 and avoid the hashing overhead.
3555 */
Mike Kravetzc672c7f2015-09-08 15:01:35 -07003556u32 hugetlb_fault_mutex_hash(struct hstate *h, struct mm_struct *mm,
Davidlohr Bueso8382d912014-04-03 14:47:31 -07003557 struct vm_area_struct *vma,
3558 struct address_space *mapping,
3559 pgoff_t idx, unsigned long address)
3560{
3561 return 0;
3562}
3563#endif
3564
Adam Litke86e52162006-01-06 00:10:43 -08003565int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
Hugh Dickins788c7df2009-06-23 13:49:05 +01003566 unsigned long address, unsigned int flags)
Adam Litke86e52162006-01-06 00:10:43 -08003567{
Davidlohr Bueso8382d912014-04-03 14:47:31 -07003568 pte_t *ptep, entry;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003569 spinlock_t *ptl;
David Gibson1e8f8892006-01-06 00:10:44 -08003570 int ret;
Davidlohr Bueso8382d912014-04-03 14:47:31 -07003571 u32 hash;
3572 pgoff_t idx;
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09003573 struct page *page = NULL;
Andy Whitcroft57303d82008-08-12 15:08:47 -07003574 struct page *pagecache_page = NULL;
Andi Kleena5516432008-07-23 21:27:41 -07003575 struct hstate *h = hstate_vma(vma);
Davidlohr Bueso8382d912014-04-03 14:47:31 -07003576 struct address_space *mapping;
Naoya Horiguchi0f792cf2015-02-11 15:25:25 -08003577 int need_wait_lock = 0;
Adam Litke86e52162006-01-06 00:10:43 -08003578
KAMEZAWA Hiroyuki1e16a532012-01-10 15:07:22 -08003579 address &= huge_page_mask(h);
3580
Naoya Horiguchifd6a03ed2010-05-28 09:29:21 +09003581 ptep = huge_pte_offset(mm, address);
3582 if (ptep) {
3583 entry = huge_ptep_get(ptep);
Naoya Horiguchi290408d2010-09-08 10:19:35 +09003584 if (unlikely(is_hugetlb_entry_migration(entry))) {
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003585 migration_entry_wait_huge(vma, mm, ptep);
Naoya Horiguchi290408d2010-09-08 10:19:35 +09003586 return 0;
3587 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
Chris Forbes32f84522011-07-25 17:12:14 -07003588 return VM_FAULT_HWPOISON_LARGE |
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07003589 VM_FAULT_SET_HINDEX(hstate_index(h));
Naoya Horiguchifd6a03ed2010-05-28 09:29:21 +09003590 }
3591
Andi Kleena5516432008-07-23 21:27:41 -07003592 ptep = huge_pte_alloc(mm, address, huge_page_size(h));
Adam Litke86e52162006-01-06 00:10:43 -08003593 if (!ptep)
3594 return VM_FAULT_OOM;
3595
Davidlohr Bueso8382d912014-04-03 14:47:31 -07003596 mapping = vma->vm_file->f_mapping;
3597 idx = vma_hugecache_offset(h, vma, address);
3598
David Gibson3935baa2006-03-22 00:08:53 -08003599 /*
3600 * Serialize hugepage allocation and instantiation, so that we don't
3601 * get spurious allocation failures if two CPUs race to instantiate
3602 * the same page in the page cache.
3603 */
Mike Kravetzc672c7f2015-09-08 15:01:35 -07003604 hash = hugetlb_fault_mutex_hash(h, mm, vma, mapping, idx, address);
3605 mutex_lock(&hugetlb_fault_mutex_table[hash]);
Davidlohr Bueso8382d912014-04-03 14:47:31 -07003606
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07003607 entry = huge_ptep_get(ptep);
3608 if (huge_pte_none(entry)) {
Davidlohr Bueso8382d912014-04-03 14:47:31 -07003609 ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags);
David Gibsonb4d1d992008-10-15 22:01:11 -07003610 goto out_mutex;
David Gibson3935baa2006-03-22 00:08:53 -08003611 }
Adam Litke86e52162006-01-06 00:10:43 -08003612
Nick Piggin83c54072007-07-19 01:47:05 -07003613 ret = 0;
David Gibson1e8f8892006-01-06 00:10:44 -08003614
Andy Whitcroft57303d82008-08-12 15:08:47 -07003615 /*
Naoya Horiguchi0f792cf2015-02-11 15:25:25 -08003616 * entry could be a migration/hwpoison entry at this point, so this
3617 * check prevents the kernel from going below assuming that we have
3618 * a active hugepage in pagecache. This goto expects the 2nd page fault,
3619 * and is_hugetlb_entry_(migration|hwpoisoned) check will properly
3620 * handle it.
3621 */
3622 if (!pte_present(entry))
3623 goto out_mutex;
3624
3625 /*
Andy Whitcroft57303d82008-08-12 15:08:47 -07003626 * If we are going to COW the mapping later, we examine the pending
3627 * reservations for this page now. This will ensure that any
3628 * allocations necessary to record that reservation occur outside the
3629 * spinlock. For private mappings, we also lookup the pagecache
3630 * page now as it is used to determine if a reservation has been
3631 * consumed.
3632 */
Gerald Schaefer106c9922013-04-29 15:07:23 -07003633 if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) {
Andy Whitcroft2b267362008-08-12 15:08:49 -07003634 if (vma_needs_reservation(h, vma, address) < 0) {
3635 ret = VM_FAULT_OOM;
David Gibsonb4d1d992008-10-15 22:01:11 -07003636 goto out_mutex;
Andy Whitcroft2b267362008-08-12 15:08:49 -07003637 }
Mike Kravetz5e911372015-09-08 15:01:28 -07003638 /* Just decrements count, does not deallocate */
Mike Kravetzfeba16e2015-09-08 15:01:31 -07003639 vma_end_reservation(h, vma, address);
Andy Whitcroft57303d82008-08-12 15:08:47 -07003640
Mel Gormanf83a2752009-05-28 14:34:40 -07003641 if (!(vma->vm_flags & VM_MAYSHARE))
Andy Whitcroft57303d82008-08-12 15:08:47 -07003642 pagecache_page = hugetlbfs_pagecache_page(h,
3643 vma, address);
3644 }
3645
Naoya Horiguchi0f792cf2015-02-11 15:25:25 -08003646 ptl = huge_pte_lock(h, mm, ptep);
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09003647
David Gibson1e8f8892006-01-06 00:10:44 -08003648 /* Check for a racing update before calling hugetlb_cow */
David Gibsonb4d1d992008-10-15 22:01:11 -07003649 if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003650 goto out_ptl;
David Gibsonb4d1d992008-10-15 22:01:11 -07003651
Naoya Horiguchi0f792cf2015-02-11 15:25:25 -08003652 /*
3653 * hugetlb_cow() requires page locks of pte_page(entry) and
3654 * pagecache_page, so here we need take the former one
3655 * when page != pagecache_page or !pagecache_page.
3656 */
3657 page = pte_page(entry);
3658 if (page != pagecache_page)
3659 if (!trylock_page(page)) {
3660 need_wait_lock = 1;
3661 goto out_ptl;
3662 }
3663
3664 get_page(page);
David Gibsonb4d1d992008-10-15 22:01:11 -07003665
Hugh Dickins788c7df2009-06-23 13:49:05 +01003666 if (flags & FAULT_FLAG_WRITE) {
Gerald Schaefer106c9922013-04-29 15:07:23 -07003667 if (!huge_pte_write(entry)) {
Andy Whitcroft57303d82008-08-12 15:08:47 -07003668 ret = hugetlb_cow(mm, vma, address, ptep, entry,
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003669 pagecache_page, ptl);
Naoya Horiguchi0f792cf2015-02-11 15:25:25 -08003670 goto out_put_page;
David Gibsonb4d1d992008-10-15 22:01:11 -07003671 }
Gerald Schaefer106c9922013-04-29 15:07:23 -07003672 entry = huge_pte_mkdirty(entry);
David Gibsonb4d1d992008-10-15 22:01:11 -07003673 }
3674 entry = pte_mkyoung(entry);
Hugh Dickins788c7df2009-06-23 13:49:05 +01003675 if (huge_ptep_set_access_flags(vma, address, ptep, entry,
3676 flags & FAULT_FLAG_WRITE))
Russell King4b3073e2009-12-18 16:40:18 +00003677 update_mmu_cache(vma, address, ptep);
Naoya Horiguchi0f792cf2015-02-11 15:25:25 -08003678out_put_page:
3679 if (page != pagecache_page)
3680 unlock_page(page);
3681 put_page(page);
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003682out_ptl:
3683 spin_unlock(ptl);
Andy Whitcroft57303d82008-08-12 15:08:47 -07003684
3685 if (pagecache_page) {
3686 unlock_page(pagecache_page);
3687 put_page(pagecache_page);
3688 }
David Gibsonb4d1d992008-10-15 22:01:11 -07003689out_mutex:
Mike Kravetzc672c7f2015-09-08 15:01:35 -07003690 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
Naoya Horiguchi0f792cf2015-02-11 15:25:25 -08003691 /*
3692 * Generally it's safe to hold refcount during waiting page lock. But
3693 * here we just wait to defer the next page fault to avoid busy loop and
3694 * the page is not used after unlocked before returning from the current
3695 * page fault. So we are safe from accessing freed page, even if we wait
3696 * here without taking refcount.
3697 */
3698 if (need_wait_lock)
3699 wait_on_page_locked(page);
David Gibson1e8f8892006-01-06 00:10:44 -08003700 return ret;
Adam Litke86e52162006-01-06 00:10:43 -08003701}
3702
Michel Lespinasse28a35712013-02-22 16:35:55 -08003703long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
3704 struct page **pages, struct vm_area_struct **vmas,
3705 unsigned long *position, unsigned long *nr_pages,
3706 long i, unsigned int flags)
David Gibson63551ae2005-06-21 17:14:44 -07003707{
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08003708 unsigned long pfn_offset;
3709 unsigned long vaddr = *position;
Michel Lespinasse28a35712013-02-22 16:35:55 -08003710 unsigned long remainder = *nr_pages;
Andi Kleena5516432008-07-23 21:27:41 -07003711 struct hstate *h = hstate_vma(vma);
David Gibson63551ae2005-06-21 17:14:44 -07003712
David Gibson63551ae2005-06-21 17:14:44 -07003713 while (vaddr < vma->vm_end && remainder) {
Adam Litke4c887262005-10-29 18:16:46 -07003714 pte_t *pte;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003715 spinlock_t *ptl = NULL;
Hugh Dickins2a15efc2009-09-21 17:03:27 -07003716 int absent;
Adam Litke4c887262005-10-29 18:16:46 -07003717 struct page *page;
3718
3719 /*
David Rientjes02057962015-04-14 15:48:24 -07003720 * If we have a pending SIGKILL, don't keep faulting pages and
3721 * potentially allocating memory.
3722 */
3723 if (unlikely(fatal_signal_pending(current))) {
3724 remainder = 0;
3725 break;
3726 }
3727
3728 /*
Adam Litke4c887262005-10-29 18:16:46 -07003729 * Some archs (sparc64, sh*) have multiple pte_ts to
Hugh Dickins2a15efc2009-09-21 17:03:27 -07003730 * each hugepage. We have to make sure we get the
Adam Litke4c887262005-10-29 18:16:46 -07003731 * first, for the page indexing below to work.
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003732 *
3733 * Note that page table lock is not held when pte is null.
Adam Litke4c887262005-10-29 18:16:46 -07003734 */
Andi Kleena5516432008-07-23 21:27:41 -07003735 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h));
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003736 if (pte)
3737 ptl = huge_pte_lock(h, mm, pte);
Hugh Dickins2a15efc2009-09-21 17:03:27 -07003738 absent = !pte || huge_pte_none(huge_ptep_get(pte));
Adam Litke4c887262005-10-29 18:16:46 -07003739
Hugh Dickins2a15efc2009-09-21 17:03:27 -07003740 /*
3741 * When coredumping, it suits get_dump_page if we just return
Hugh Dickins3ae77f42009-09-21 17:03:33 -07003742 * an error where there's an empty slot with no huge pagecache
3743 * to back it. This way, we avoid allocating a hugepage, and
3744 * the sparse dumpfile avoids allocating disk blocks, but its
3745 * huge holes still show up with zeroes where they need to be.
Hugh Dickins2a15efc2009-09-21 17:03:27 -07003746 */
Hugh Dickins3ae77f42009-09-21 17:03:33 -07003747 if (absent && (flags & FOLL_DUMP) &&
3748 !hugetlbfs_pagecache_present(h, vma, vaddr)) {
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003749 if (pte)
3750 spin_unlock(ptl);
Hugh Dickins2a15efc2009-09-21 17:03:27 -07003751 remainder = 0;
3752 break;
3753 }
3754
Naoya Horiguchi9cc3a5b2013-04-17 15:58:30 -07003755 /*
3756 * We need call hugetlb_fault for both hugepages under migration
3757 * (in which case hugetlb_fault waits for the migration,) and
3758 * hwpoisoned hugepages (in which case we need to prevent the
3759 * caller from accessing to them.) In order to do this, we use
3760 * here is_swap_pte instead of is_hugetlb_entry_migration and
3761 * is_hugetlb_entry_hwpoisoned. This is because it simply covers
3762 * both cases, and because we can't follow correct pages
3763 * directly from any kind of swap entries.
3764 */
3765 if (absent || is_swap_pte(huge_ptep_get(pte)) ||
Gerald Schaefer106c9922013-04-29 15:07:23 -07003766 ((flags & FOLL_WRITE) &&
3767 !huge_pte_write(huge_ptep_get(pte)))) {
Adam Litke4c887262005-10-29 18:16:46 -07003768 int ret;
3769
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003770 if (pte)
3771 spin_unlock(ptl);
Hugh Dickins2a15efc2009-09-21 17:03:27 -07003772 ret = hugetlb_fault(mm, vma, vaddr,
3773 (flags & FOLL_WRITE) ? FAULT_FLAG_WRITE : 0);
Adam Litkea89182c2007-08-22 14:01:51 -07003774 if (!(ret & VM_FAULT_ERROR))
Adam Litke4c887262005-10-29 18:16:46 -07003775 continue;
3776
3777 remainder = 0;
Adam Litke4c887262005-10-29 18:16:46 -07003778 break;
3779 }
David Gibson63551ae2005-06-21 17:14:44 -07003780
Andi Kleena5516432008-07-23 21:27:41 -07003781 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07003782 page = pte_page(huge_ptep_get(pte));
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08003783same_page:
Chen, Kenneth Wd6692182006-03-31 02:29:57 -08003784 if (pages) {
Hugh Dickins2a15efc2009-09-21 17:03:27 -07003785 pages[i] = mem_map_offset(page, pfn_offset);
Andrea Arcangelia0368d42014-01-21 15:48:49 -08003786 get_page_foll(pages[i]);
Chen, Kenneth Wd6692182006-03-31 02:29:57 -08003787 }
David Gibson63551ae2005-06-21 17:14:44 -07003788
3789 if (vmas)
3790 vmas[i] = vma;
3791
3792 vaddr += PAGE_SIZE;
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08003793 ++pfn_offset;
David Gibson63551ae2005-06-21 17:14:44 -07003794 --remainder;
3795 ++i;
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08003796 if (vaddr < vma->vm_end && remainder &&
Andi Kleena5516432008-07-23 21:27:41 -07003797 pfn_offset < pages_per_huge_page(h)) {
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08003798 /*
3799 * We use pfn_offset to avoid touching the pageframes
3800 * of this compound page.
3801 */
3802 goto same_page;
3803 }
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003804 spin_unlock(ptl);
David Gibson63551ae2005-06-21 17:14:44 -07003805 }
Michel Lespinasse28a35712013-02-22 16:35:55 -08003806 *nr_pages = remainder;
David Gibson63551ae2005-06-21 17:14:44 -07003807 *position = vaddr;
3808
Hugh Dickins2a15efc2009-09-21 17:03:27 -07003809 return i ? i : -EFAULT;
David Gibson63551ae2005-06-21 17:14:44 -07003810}
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003811
Peter Zijlstra7da4d642012-11-19 03:14:23 +01003812unsigned long hugetlb_change_protection(struct vm_area_struct *vma,
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003813 unsigned long address, unsigned long end, pgprot_t newprot)
3814{
3815 struct mm_struct *mm = vma->vm_mm;
3816 unsigned long start = address;
3817 pte_t *ptep;
3818 pte_t pte;
Andi Kleena5516432008-07-23 21:27:41 -07003819 struct hstate *h = hstate_vma(vma);
Peter Zijlstra7da4d642012-11-19 03:14:23 +01003820 unsigned long pages = 0;
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003821
3822 BUG_ON(address >= end);
3823 flush_cache_range(vma, address, end);
3824
Rik van Riela5338092014-04-07 15:36:57 -07003825 mmu_notifier_invalidate_range_start(mm, start, end);
Davidlohr Bueso83cde9e2014-12-12 16:54:21 -08003826 i_mmap_lock_write(vma->vm_file->f_mapping);
Andi Kleena5516432008-07-23 21:27:41 -07003827 for (; address < end; address += huge_page_size(h)) {
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003828 spinlock_t *ptl;
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003829 ptep = huge_pte_offset(mm, address);
3830 if (!ptep)
3831 continue;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003832 ptl = huge_pte_lock(h, mm, ptep);
Peter Zijlstra7da4d642012-11-19 03:14:23 +01003833 if (huge_pmd_unshare(mm, &address, ptep)) {
3834 pages++;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003835 spin_unlock(ptl);
Chen, Kenneth W39dde652006-12-06 20:32:03 -08003836 continue;
Peter Zijlstra7da4d642012-11-19 03:14:23 +01003837 }
Naoya Horiguchia8bda282015-02-11 15:25:28 -08003838 pte = huge_ptep_get(ptep);
3839 if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) {
3840 spin_unlock(ptl);
3841 continue;
3842 }
3843 if (unlikely(is_hugetlb_entry_migration(pte))) {
3844 swp_entry_t entry = pte_to_swp_entry(pte);
3845
3846 if (is_write_migration_entry(entry)) {
3847 pte_t newpte;
3848
3849 make_migration_entry_read(&entry);
3850 newpte = swp_entry_to_pte(entry);
3851 set_huge_pte_at(mm, address, ptep, newpte);
3852 pages++;
3853 }
3854 spin_unlock(ptl);
3855 continue;
3856 }
3857 if (!huge_pte_none(pte)) {
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003858 pte = huge_ptep_get_and_clear(mm, address, ptep);
Gerald Schaefer106c9922013-04-29 15:07:23 -07003859 pte = pte_mkhuge(huge_pte_modify(pte, newprot));
Tony Lube7517d2013-02-04 14:28:46 -08003860 pte = arch_make_huge_pte(pte, vma, NULL, 0);
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003861 set_huge_pte_at(mm, address, ptep, pte);
Peter Zijlstra7da4d642012-11-19 03:14:23 +01003862 pages++;
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003863 }
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003864 spin_unlock(ptl);
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003865 }
Mel Gormand8333522012-07-31 16:46:20 -07003866 /*
Davidlohr Buesoc8c06ef2014-12-12 16:54:24 -08003867 * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare
Mel Gormand8333522012-07-31 16:46:20 -07003868 * may have cleared our pud entry and done put_page on the page table:
Davidlohr Buesoc8c06ef2014-12-12 16:54:24 -08003869 * once we release i_mmap_rwsem, another task can do the final put_page
Mel Gormand8333522012-07-31 16:46:20 -07003870 * and that page table be reused and filled with junk.
3871 */
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003872 flush_tlb_range(vma, start, end);
Joerg Roedel34ee6452014-11-13 13:46:09 +11003873 mmu_notifier_invalidate_range(mm, start, end);
Davidlohr Bueso83cde9e2014-12-12 16:54:21 -08003874 i_mmap_unlock_write(vma->vm_file->f_mapping);
Rik van Riela5338092014-04-07 15:36:57 -07003875 mmu_notifier_invalidate_range_end(mm, start, end);
Peter Zijlstra7da4d642012-11-19 03:14:23 +01003876
3877 return pages << h->order;
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003878}
3879
Mel Gormana1e78772008-07-23 21:27:23 -07003880int hugetlb_reserve_pages(struct inode *inode,
3881 long from, long to,
Mel Gorman5a6fe122009-02-10 14:02:27 +00003882 struct vm_area_struct *vma,
KOSAKI Motohiroca16d142011-05-26 19:16:19 +09003883 vm_flags_t vm_flags)
Adam Litkee4e574b2007-10-16 01:26:19 -07003884{
Mel Gorman17c9d122009-02-11 16:34:16 +00003885 long ret, chg;
Andi Kleena5516432008-07-23 21:27:41 -07003886 struct hstate *h = hstate_inode(inode);
David Gibson90481622012-03-21 16:34:12 -07003887 struct hugepage_subpool *spool = subpool_inode(inode);
Joonsoo Kim9119a412014-04-03 14:47:25 -07003888 struct resv_map *resv_map;
Mike Kravetz1c5ecae2015-04-15 16:13:39 -07003889 long gbl_reserve;
Adam Litkee4e574b2007-10-16 01:26:19 -07003890
Mel Gormana1e78772008-07-23 21:27:23 -07003891 /*
Mel Gorman17c9d122009-02-11 16:34:16 +00003892 * Only apply hugepage reservation if asked. At fault time, an
3893 * attempt will be made for VM_NORESERVE to allocate a page
David Gibson90481622012-03-21 16:34:12 -07003894 * without using reserves
Mel Gorman17c9d122009-02-11 16:34:16 +00003895 */
KOSAKI Motohiroca16d142011-05-26 19:16:19 +09003896 if (vm_flags & VM_NORESERVE)
Mel Gorman17c9d122009-02-11 16:34:16 +00003897 return 0;
3898
3899 /*
Mel Gormana1e78772008-07-23 21:27:23 -07003900 * Shared mappings base their reservation on the number of pages that
3901 * are already allocated on behalf of the file. Private mappings need
3902 * to reserve the full area even if read-only as mprotect() may be
3903 * called to make the mapping read-write. Assume !vma is a shm mapping
3904 */
Joonsoo Kim9119a412014-04-03 14:47:25 -07003905 if (!vma || vma->vm_flags & VM_MAYSHARE) {
Joonsoo Kim4e35f482014-04-03 14:47:30 -07003906 resv_map = inode_resv_map(inode);
Joonsoo Kim9119a412014-04-03 14:47:25 -07003907
Joonsoo Kim1406ec92014-04-03 14:47:26 -07003908 chg = region_chg(resv_map, from, to);
Joonsoo Kim9119a412014-04-03 14:47:25 -07003909
3910 } else {
3911 resv_map = resv_map_alloc();
Mel Gorman5a6fe122009-02-10 14:02:27 +00003912 if (!resv_map)
3913 return -ENOMEM;
3914
Mel Gorman17c9d122009-02-11 16:34:16 +00003915 chg = to - from;
3916
Mel Gorman5a6fe122009-02-10 14:02:27 +00003917 set_vma_resv_map(vma, resv_map);
3918 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
3919 }
3920
Dave Hansenc50ac052012-05-29 15:06:46 -07003921 if (chg < 0) {
3922 ret = chg;
3923 goto out_err;
3924 }
Mel Gorman17c9d122009-02-11 16:34:16 +00003925
Mike Kravetz1c5ecae2015-04-15 16:13:39 -07003926 /*
3927 * There must be enough pages in the subpool for the mapping. If
3928 * the subpool has a minimum size, there may be some global
3929 * reservations already in place (gbl_reserve).
3930 */
3931 gbl_reserve = hugepage_subpool_get_pages(spool, chg);
3932 if (gbl_reserve < 0) {
Dave Hansenc50ac052012-05-29 15:06:46 -07003933 ret = -ENOSPC;
3934 goto out_err;
3935 }
Mel Gorman17c9d122009-02-11 16:34:16 +00003936
3937 /*
3938 * Check enough hugepages are available for the reservation.
David Gibson90481622012-03-21 16:34:12 -07003939 * Hand the pages back to the subpool if there are not
Mel Gorman17c9d122009-02-11 16:34:16 +00003940 */
Mike Kravetz1c5ecae2015-04-15 16:13:39 -07003941 ret = hugetlb_acct_memory(h, gbl_reserve);
Mel Gorman17c9d122009-02-11 16:34:16 +00003942 if (ret < 0) {
Mike Kravetz1c5ecae2015-04-15 16:13:39 -07003943 /* put back original number of pages, chg */
3944 (void)hugepage_subpool_put_pages(spool, chg);
Dave Hansenc50ac052012-05-29 15:06:46 -07003945 goto out_err;
Mel Gorman17c9d122009-02-11 16:34:16 +00003946 }
3947
3948 /*
3949 * Account for the reservations made. Shared mappings record regions
3950 * that have reservations as they are shared by multiple VMAs.
3951 * When the last VMA disappears, the region map says how much
3952 * the reservation was and the page cache tells how much of
3953 * the reservation was consumed. Private mappings are per-VMA and
3954 * only the consumed reservations are tracked. When the VMA
3955 * disappears, the original reservation is the VMA size and the
3956 * consumed reservations are stored in the map. Hence, nothing
3957 * else has to be done for private mappings here
3958 */
Mike Kravetz33039672015-06-24 16:57:58 -07003959 if (!vma || vma->vm_flags & VM_MAYSHARE) {
3960 long add = region_add(resv_map, from, to);
3961
3962 if (unlikely(chg > add)) {
3963 /*
3964 * pages in this range were added to the reserve
3965 * map between region_chg and region_add. This
3966 * indicates a race with alloc_huge_page. Adjust
3967 * the subpool and reserve counts modified above
3968 * based on the difference.
3969 */
3970 long rsv_adjust;
3971
3972 rsv_adjust = hugepage_subpool_put_pages(spool,
3973 chg - add);
3974 hugetlb_acct_memory(h, -rsv_adjust);
3975 }
3976 }
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -07003977 return 0;
Dave Hansenc50ac052012-05-29 15:06:46 -07003978out_err:
Mike Kravetz5e911372015-09-08 15:01:28 -07003979 if (!vma || vma->vm_flags & VM_MAYSHARE)
3980 region_abort(resv_map, from, to);
Joonsoo Kimf031dd22014-04-03 14:47:28 -07003981 if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
3982 kref_put(&resv_map->refs, resv_map_release);
Dave Hansenc50ac052012-05-29 15:06:46 -07003983 return ret;
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -07003984}
3985
Mike Kravetzb5cec282015-09-08 15:01:41 -07003986long hugetlb_unreserve_pages(struct inode *inode, long start, long end,
3987 long freed)
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -07003988{
Andi Kleena5516432008-07-23 21:27:41 -07003989 struct hstate *h = hstate_inode(inode);
Joonsoo Kim4e35f482014-04-03 14:47:30 -07003990 struct resv_map *resv_map = inode_resv_map(inode);
Joonsoo Kim9119a412014-04-03 14:47:25 -07003991 long chg = 0;
David Gibson90481622012-03-21 16:34:12 -07003992 struct hugepage_subpool *spool = subpool_inode(inode);
Mike Kravetz1c5ecae2015-04-15 16:13:39 -07003993 long gbl_reserve;
Ken Chen45c682a2007-11-14 16:59:44 -08003994
Mike Kravetzb5cec282015-09-08 15:01:41 -07003995 if (resv_map) {
3996 chg = region_del(resv_map, start, end);
3997 /*
3998 * region_del() can fail in the rare case where a region
3999 * must be split and another region descriptor can not be
4000 * allocated. If end == LONG_MAX, it will not fail.
4001 */
4002 if (chg < 0)
4003 return chg;
4004 }
4005
Ken Chen45c682a2007-11-14 16:59:44 -08004006 spin_lock(&inode->i_lock);
Eric Sandeene4c6f8b2009-07-29 15:02:16 -07004007 inode->i_blocks -= (blocks_per_huge_page(h) * freed);
Ken Chen45c682a2007-11-14 16:59:44 -08004008 spin_unlock(&inode->i_lock);
4009
Mike Kravetz1c5ecae2015-04-15 16:13:39 -07004010 /*
4011 * If the subpool has a minimum size, the number of global
4012 * reservations to be released may be adjusted.
4013 */
4014 gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed));
4015 hugetlb_acct_memory(h, -gbl_reserve);
Mike Kravetzb5cec282015-09-08 15:01:41 -07004016
4017 return 0;
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -07004018}
Naoya Horiguchi93f70f92010-05-28 09:29:20 +09004019
Steve Capper3212b532013-04-23 12:35:02 +01004020#ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
4021static unsigned long page_table_shareable(struct vm_area_struct *svma,
4022 struct vm_area_struct *vma,
4023 unsigned long addr, pgoff_t idx)
4024{
4025 unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) +
4026 svma->vm_start;
4027 unsigned long sbase = saddr & PUD_MASK;
4028 unsigned long s_end = sbase + PUD_SIZE;
4029
4030 /* Allow segments to share if only one is marked locked */
4031 unsigned long vm_flags = vma->vm_flags & ~VM_LOCKED;
4032 unsigned long svm_flags = svma->vm_flags & ~VM_LOCKED;
4033
4034 /*
4035 * match the virtual addresses, permission and the alignment of the
4036 * page table page.
4037 */
4038 if (pmd_index(addr) != pmd_index(saddr) ||
4039 vm_flags != svm_flags ||
4040 sbase < svma->vm_start || svma->vm_end < s_end)
4041 return 0;
4042
4043 return saddr;
4044}
4045
Nicholas Krause31aafb42015-09-04 15:47:58 -07004046static bool vma_shareable(struct vm_area_struct *vma, unsigned long addr)
Steve Capper3212b532013-04-23 12:35:02 +01004047{
4048 unsigned long base = addr & PUD_MASK;
4049 unsigned long end = base + PUD_SIZE;
4050
4051 /*
4052 * check on proper vm_flags and page table alignment
4053 */
4054 if (vma->vm_flags & VM_MAYSHARE &&
4055 vma->vm_start <= base && end <= vma->vm_end)
Nicholas Krause31aafb42015-09-04 15:47:58 -07004056 return true;
4057 return false;
Steve Capper3212b532013-04-23 12:35:02 +01004058}
4059
4060/*
4061 * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
4062 * and returns the corresponding pte. While this is not necessary for the
4063 * !shared pmd case because we can allocate the pmd later as well, it makes the
4064 * code much cleaner. pmd allocation is essential for the shared case because
Davidlohr Buesoc8c06ef2014-12-12 16:54:24 -08004065 * pud has to be populated inside the same i_mmap_rwsem section - otherwise
Steve Capper3212b532013-04-23 12:35:02 +01004066 * racing tasks could either miss the sharing (see huge_pte_offset) or select a
4067 * bad pmd for sharing.
4068 */
4069pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
4070{
4071 struct vm_area_struct *vma = find_vma(mm, addr);
4072 struct address_space *mapping = vma->vm_file->f_mapping;
4073 pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) +
4074 vma->vm_pgoff;
4075 struct vm_area_struct *svma;
4076 unsigned long saddr;
4077 pte_t *spte = NULL;
4078 pte_t *pte;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08004079 spinlock_t *ptl;
Steve Capper3212b532013-04-23 12:35:02 +01004080
4081 if (!vma_shareable(vma, addr))
4082 return (pte_t *)pmd_alloc(mm, pud, addr);
4083
Davidlohr Bueso83cde9e2014-12-12 16:54:21 -08004084 i_mmap_lock_write(mapping);
Steve Capper3212b532013-04-23 12:35:02 +01004085 vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) {
4086 if (svma == vma)
4087 continue;
4088
4089 saddr = page_table_shareable(svma, vma, addr, idx);
4090 if (saddr) {
4091 spte = huge_pte_offset(svma->vm_mm, saddr);
4092 if (spte) {
Kirill A. Shutemovdc6c9a32015-02-11 15:26:50 -08004093 mm_inc_nr_pmds(mm);
Steve Capper3212b532013-04-23 12:35:02 +01004094 get_page(virt_to_page(spte));
4095 break;
4096 }
4097 }
4098 }
4099
4100 if (!spte)
4101 goto out;
4102
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08004103 ptl = huge_pte_lockptr(hstate_vma(vma), mm, spte);
4104 spin_lock(ptl);
Kirill A. Shutemovdc6c9a32015-02-11 15:26:50 -08004105 if (pud_none(*pud)) {
Steve Capper3212b532013-04-23 12:35:02 +01004106 pud_populate(mm, pud,
4107 (pmd_t *)((unsigned long)spte & PAGE_MASK));
Kirill A. Shutemovdc6c9a32015-02-11 15:26:50 -08004108 } else {
Steve Capper3212b532013-04-23 12:35:02 +01004109 put_page(virt_to_page(spte));
Kirill A. Shutemovdc6c9a32015-02-11 15:26:50 -08004110 mm_inc_nr_pmds(mm);
4111 }
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08004112 spin_unlock(ptl);
Steve Capper3212b532013-04-23 12:35:02 +01004113out:
4114 pte = (pte_t *)pmd_alloc(mm, pud, addr);
Davidlohr Bueso83cde9e2014-12-12 16:54:21 -08004115 i_mmap_unlock_write(mapping);
Steve Capper3212b532013-04-23 12:35:02 +01004116 return pte;
4117}
4118
4119/*
4120 * unmap huge page backed by shared pte.
4121 *
4122 * Hugetlb pte page is ref counted at the time of mapping. If pte is shared
4123 * indicated by page_count > 1, unmap is achieved by clearing pud and
4124 * decrementing the ref count. If count == 1, the pte page is not shared.
4125 *
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08004126 * called with page table lock held.
Steve Capper3212b532013-04-23 12:35:02 +01004127 *
4128 * returns: 1 successfully unmapped a shared pte page
4129 * 0 the underlying pte page is not shared, or it is the last user
4130 */
4131int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
4132{
4133 pgd_t *pgd = pgd_offset(mm, *addr);
4134 pud_t *pud = pud_offset(pgd, *addr);
4135
4136 BUG_ON(page_count(virt_to_page(ptep)) == 0);
4137 if (page_count(virt_to_page(ptep)) == 1)
4138 return 0;
4139
4140 pud_clear(pud);
4141 put_page(virt_to_page(ptep));
Kirill A. Shutemovdc6c9a32015-02-11 15:26:50 -08004142 mm_dec_nr_pmds(mm);
Steve Capper3212b532013-04-23 12:35:02 +01004143 *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE;
4144 return 1;
4145}
Steve Capper9e5fc742013-04-30 08:02:03 +01004146#define want_pmd_share() (1)
4147#else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
4148pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
4149{
4150 return NULL;
4151}
Zhang Zhene81f2d22015-06-24 16:56:13 -07004152
4153int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
4154{
4155 return 0;
4156}
Steve Capper9e5fc742013-04-30 08:02:03 +01004157#define want_pmd_share() (0)
Steve Capper3212b532013-04-23 12:35:02 +01004158#endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
4159
Steve Capper9e5fc742013-04-30 08:02:03 +01004160#ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
4161pte_t *huge_pte_alloc(struct mm_struct *mm,
4162 unsigned long addr, unsigned long sz)
4163{
4164 pgd_t *pgd;
4165 pud_t *pud;
4166 pte_t *pte = NULL;
4167
4168 pgd = pgd_offset(mm, addr);
4169 pud = pud_alloc(mm, pgd, addr);
4170 if (pud) {
4171 if (sz == PUD_SIZE) {
4172 pte = (pte_t *)pud;
4173 } else {
4174 BUG_ON(sz != PMD_SIZE);
4175 if (want_pmd_share() && pud_none(*pud))
4176 pte = huge_pmd_share(mm, addr, pud);
4177 else
4178 pte = (pte_t *)pmd_alloc(mm, pud, addr);
4179 }
4180 }
4181 BUG_ON(pte && !pte_none(*pte) && !pte_huge(*pte));
4182
4183 return pte;
4184}
4185
4186pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr)
4187{
4188 pgd_t *pgd;
4189 pud_t *pud;
4190 pmd_t *pmd = NULL;
4191
4192 pgd = pgd_offset(mm, addr);
4193 if (pgd_present(*pgd)) {
4194 pud = pud_offset(pgd, addr);
4195 if (pud_present(*pud)) {
4196 if (pud_huge(*pud))
4197 return (pte_t *)pud;
4198 pmd = pmd_offset(pud, addr);
4199 }
4200 }
4201 return (pte_t *) pmd;
4202}
4203
Naoya Horiguchi61f77ed2015-02-11 15:25:15 -08004204#endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
4205
4206/*
4207 * These functions are overwritable if your architecture needs its own
4208 * behavior.
4209 */
4210struct page * __weak
4211follow_huge_addr(struct mm_struct *mm, unsigned long address,
4212 int write)
4213{
4214 return ERR_PTR(-EINVAL);
4215}
4216
4217struct page * __weak
Steve Capper9e5fc742013-04-30 08:02:03 +01004218follow_huge_pmd(struct mm_struct *mm, unsigned long address,
Naoya Horiguchie66f17f2015-02-11 15:25:22 -08004219 pmd_t *pmd, int flags)
Steve Capper9e5fc742013-04-30 08:02:03 +01004220{
Naoya Horiguchie66f17f2015-02-11 15:25:22 -08004221 struct page *page = NULL;
4222 spinlock_t *ptl;
4223retry:
4224 ptl = pmd_lockptr(mm, pmd);
4225 spin_lock(ptl);
4226 /*
4227 * make sure that the address range covered by this pmd is not
4228 * unmapped from other threads.
4229 */
4230 if (!pmd_huge(*pmd))
4231 goto out;
4232 if (pmd_present(*pmd)) {
Gerald Schaefer97534122015-04-14 15:42:30 -07004233 page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT);
Naoya Horiguchie66f17f2015-02-11 15:25:22 -08004234 if (flags & FOLL_GET)
4235 get_page(page);
4236 } else {
4237 if (is_hugetlb_entry_migration(huge_ptep_get((pte_t *)pmd))) {
4238 spin_unlock(ptl);
4239 __migration_entry_wait(mm, (pte_t *)pmd, ptl);
4240 goto retry;
4241 }
4242 /*
4243 * hwpoisoned entry is treated as no_page_table in
4244 * follow_page_mask().
4245 */
4246 }
4247out:
4248 spin_unlock(ptl);
Steve Capper9e5fc742013-04-30 08:02:03 +01004249 return page;
4250}
4251
Naoya Horiguchi61f77ed2015-02-11 15:25:15 -08004252struct page * __weak
Steve Capper9e5fc742013-04-30 08:02:03 +01004253follow_huge_pud(struct mm_struct *mm, unsigned long address,
Naoya Horiguchie66f17f2015-02-11 15:25:22 -08004254 pud_t *pud, int flags)
Steve Capper9e5fc742013-04-30 08:02:03 +01004255{
Naoya Horiguchie66f17f2015-02-11 15:25:22 -08004256 if (flags & FOLL_GET)
4257 return NULL;
Steve Capper9e5fc742013-04-30 08:02:03 +01004258
Naoya Horiguchie66f17f2015-02-11 15:25:22 -08004259 return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT);
Steve Capper9e5fc742013-04-30 08:02:03 +01004260}
4261
Andi Kleend5bd9102010-09-27 09:00:12 +02004262#ifdef CONFIG_MEMORY_FAILURE
4263
Naoya Horiguchi93f70f92010-05-28 09:29:20 +09004264/*
4265 * This function is called from memory failure code.
4266 * Assume the caller holds page lock of the head page.
4267 */
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +09004268int dequeue_hwpoisoned_huge_page(struct page *hpage)
Naoya Horiguchi93f70f92010-05-28 09:29:20 +09004269{
4270 struct hstate *h = page_hstate(hpage);
4271 int nid = page_to_nid(hpage);
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +09004272 int ret = -EBUSY;
Naoya Horiguchi93f70f92010-05-28 09:29:20 +09004273
4274 spin_lock(&hugetlb_lock);
Naoya Horiguchi7e1f0492015-04-15 16:14:41 -07004275 /*
4276 * Just checking !page_huge_active is not enough, because that could be
4277 * an isolated/hwpoisoned hugepage (which have >0 refcount).
4278 */
4279 if (!page_huge_active(hpage) && !page_count(hpage)) {
Naoya Horiguchi56f2fb12012-12-12 13:52:33 -08004280 /*
4281 * Hwpoisoned hugepage isn't linked to activelist or freelist,
4282 * but dangling hpage->lru can trigger list-debug warnings
4283 * (this happens when we call unpoison_memory() on it),
4284 * so let it point to itself with list_del_init().
4285 */
4286 list_del_init(&hpage->lru);
Naoya Horiguchi8c6c2ec2010-09-08 10:19:38 +09004287 set_page_refcounted(hpage);
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +09004288 h->free_huge_pages--;
4289 h->free_huge_pages_node[nid]--;
4290 ret = 0;
4291 }
Naoya Horiguchi93f70f92010-05-28 09:29:20 +09004292 spin_unlock(&hugetlb_lock);
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +09004293 return ret;
Naoya Horiguchi93f70f92010-05-28 09:29:20 +09004294}
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +09004295#endif
Naoya Horiguchi31caf662013-09-11 14:21:59 -07004296
4297bool isolate_huge_page(struct page *page, struct list_head *list)
4298{
Naoya Horiguchibcc54222015-04-15 16:14:38 -07004299 bool ret = true;
4300
Sasha Levin309381fea2014-01-23 15:52:54 -08004301 VM_BUG_ON_PAGE(!PageHead(page), page);
Naoya Horiguchi31caf662013-09-11 14:21:59 -07004302 spin_lock(&hugetlb_lock);
Naoya Horiguchibcc54222015-04-15 16:14:38 -07004303 if (!page_huge_active(page) || !get_page_unless_zero(page)) {
4304 ret = false;
4305 goto unlock;
4306 }
4307 clear_page_huge_active(page);
Naoya Horiguchi31caf662013-09-11 14:21:59 -07004308 list_move_tail(&page->lru, list);
Naoya Horiguchibcc54222015-04-15 16:14:38 -07004309unlock:
Naoya Horiguchi31caf662013-09-11 14:21:59 -07004310 spin_unlock(&hugetlb_lock);
Naoya Horiguchibcc54222015-04-15 16:14:38 -07004311 return ret;
Naoya Horiguchi31caf662013-09-11 14:21:59 -07004312}
4313
4314void putback_active_hugepage(struct page *page)
4315{
Sasha Levin309381fea2014-01-23 15:52:54 -08004316 VM_BUG_ON_PAGE(!PageHead(page), page);
Naoya Horiguchi31caf662013-09-11 14:21:59 -07004317 spin_lock(&hugetlb_lock);
Naoya Horiguchibcc54222015-04-15 16:14:38 -07004318 set_page_huge_active(page);
Naoya Horiguchi31caf662013-09-11 14:21:59 -07004319 list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist);
4320 spin_unlock(&hugetlb_lock);
4321 put_page(page);
4322}