blob: 9801dc0250e209260f9ba6d9bf1f2c61ee71eccd [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
2 * Generic hugetlb support.
Nadia Yvette Chambers6d49e352012-12-06 10:39:54 +01003 * (C) Nadia Yvette Chambers, April 2004
Linus Torvalds1da177e2005-04-16 15:20:36 -07004 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07005#include <linux/list.h>
6#include <linux/init.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -07007#include <linux/mm.h>
Alexey Dobriyane1759c22008-10-15 23:50:22 +04008#include <linux/seq_file.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -07009#include <linux/sysctl.h>
10#include <linux/highmem.h>
Andrea Arcangelicddb8a52008-07-28 15:46:29 -070011#include <linux/mmu_notifier.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070012#include <linux/nodemask.h>
David Gibson63551ae2005-06-21 17:14:44 -070013#include <linux/pagemap.h>
Christoph Lameter5da7ca82006-01-06 00:10:46 -080014#include <linux/mempolicy.h>
Gideon Israel Dsouza3b321232014-04-07 15:37:26 -070015#include <linux/compiler.h>
Christoph Lameteraea47ff2006-01-08 01:00:57 -080016#include <linux/cpuset.h>
David Gibson3935baa2006-03-22 00:08:53 -080017#include <linux/mutex.h>
Andi Kleenaa888a72008-07-23 21:27:47 -070018#include <linux/bootmem.h>
Nishanth Aravamudana3437872008-07-23 21:27:44 -070019#include <linux/sysfs.h>
Tejun Heo5a0e3ad2010-03-24 17:04:11 +090020#include <linux/slab.h>
Mike Kravetz1e862842018-03-22 16:17:13 -070021#include <linux/mmdebug.h>
Ingo Molnar174cd4b2017-02-02 19:15:33 +010022#include <linux/sched/signal.h>
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +090023#include <linux/rmap.h>
Matthew Wilcoxc6247f72017-07-10 15:48:56 -070024#include <linux/string_helpers.h>
Naoya Horiguchifd6a03e2010-05-28 09:29:21 +090025#include <linux/swap.h>
26#include <linux/swapops.h>
Davidlohr Bueso8382d912014-04-03 14:47:31 -070027#include <linux/jhash.h>
Linus Torvaldsd6606682008-08-06 12:04:54 -070028
David Gibson63551ae2005-06-21 17:14:44 -070029#include <asm/page.h>
30#include <asm/pgtable.h>
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -070031#include <asm/tlb.h>
David Gibson63551ae2005-06-21 17:14:44 -070032
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -070033#include <linux/io.h>
David Gibson63551ae2005-06-21 17:14:44 -070034#include <linux/hugetlb.h>
Aneesh Kumar K.V9dd540e2012-07-31 16:42:15 -070035#include <linux/hugetlb_cgroup.h>
Lee Schermerhorn9a3052302009-12-14 17:58:25 -080036#include <linux/node.h>
Mike Kravetz1a1aad82017-02-22 15:43:01 -080037#include <linux/userfaultfd_k.h>
Nick Piggin7835e982006-03-22 00:08:40 -080038#include "internal.h"
Linus Torvalds1da177e2005-04-16 15:20:36 -070039
Andrey Ryabinin753162c2015-02-10 14:11:36 -080040int hugepages_treat_as_movable;
Andi Kleena5516432008-07-23 21:27:41 -070041
Aneesh Kumar K.Vc3f38a32012-07-31 16:42:10 -070042int hugetlb_max_hstate __read_mostly;
Andi Kleene5ff2152008-07-23 21:27:42 -070043unsigned int default_hstate_idx;
44struct hstate hstates[HUGE_MAX_HSTATE];
Naoya Horiguchi641844f2015-06-24 16:56:59 -070045/*
46 * Minimum page order among possible hugepage sizes, set to a proper value
47 * at boot time.
48 */
49static unsigned int minimum_order __read_mostly = UINT_MAX;
Andi Kleene5ff2152008-07-23 21:27:42 -070050
Jon Tollefson53ba51d2008-07-23 21:27:52 -070051__initdata LIST_HEAD(huge_boot_pages);
52
Andi Kleene5ff2152008-07-23 21:27:42 -070053/* for command line parsing */
54static struct hstate * __initdata parsed_hstate;
55static unsigned long __initdata default_hstate_max_huge_pages;
Nick Piggine11bfbf2008-07-23 21:27:52 -070056static unsigned long __initdata default_hstate_size;
Vaishali Thakkar9fee0212016-05-19 17:11:04 -070057static bool __initdata parsed_valid_hugepagesz = true;
Andi Kleene5ff2152008-07-23 21:27:42 -070058
David Gibson3935baa2006-03-22 00:08:53 -080059/*
Naoya Horiguchi31caf662013-09-11 14:21:59 -070060 * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages,
61 * free_huge_pages, and surplus_huge_pages.
David Gibson3935baa2006-03-22 00:08:53 -080062 */
Aneesh Kumar K.Vc3f38a32012-07-31 16:42:10 -070063DEFINE_SPINLOCK(hugetlb_lock);
Eric Paris0bd0f9f2005-11-21 21:32:28 -080064
Davidlohr Bueso8382d912014-04-03 14:47:31 -070065/*
66 * Serializes faults on the same logical page. This is used to
67 * prevent spurious OOMs when the hugepage pool is fully utilized.
68 */
69static int num_fault_mutexes;
Mike Kravetzc672c7f2015-09-08 15:01:35 -070070struct mutex *hugetlb_fault_mutex_table ____cacheline_aligned_in_smp;
Davidlohr Bueso8382d912014-04-03 14:47:31 -070071
Mike Kravetz7ca02d02015-04-15 16:13:42 -070072/* Forward declaration */
73static int hugetlb_acct_memory(struct hstate *h, long delta);
74
David Gibson90481622012-03-21 16:34:12 -070075static inline void unlock_or_release_subpool(struct hugepage_subpool *spool)
76{
77 bool free = (spool->count == 0) && (spool->used_hpages == 0);
78
79 spin_unlock(&spool->lock);
80
81 /* If no pages are used, and no other handles to the subpool
Mike Kravetz7ca02d02015-04-15 16:13:42 -070082 * remain, give up any reservations mased on minimum size and
83 * free the subpool */
84 if (free) {
85 if (spool->min_hpages != -1)
86 hugetlb_acct_memory(spool->hstate,
87 -spool->min_hpages);
David Gibson90481622012-03-21 16:34:12 -070088 kfree(spool);
Mike Kravetz7ca02d02015-04-15 16:13:42 -070089 }
David Gibson90481622012-03-21 16:34:12 -070090}
91
Mike Kravetz7ca02d02015-04-15 16:13:42 -070092struct hugepage_subpool *hugepage_new_subpool(struct hstate *h, long max_hpages,
93 long min_hpages)
David Gibson90481622012-03-21 16:34:12 -070094{
95 struct hugepage_subpool *spool;
96
Mike Kravetzc6a91822015-04-15 16:13:36 -070097 spool = kzalloc(sizeof(*spool), GFP_KERNEL);
David Gibson90481622012-03-21 16:34:12 -070098 if (!spool)
99 return NULL;
100
101 spin_lock_init(&spool->lock);
102 spool->count = 1;
Mike Kravetz7ca02d02015-04-15 16:13:42 -0700103 spool->max_hpages = max_hpages;
104 spool->hstate = h;
105 spool->min_hpages = min_hpages;
106
107 if (min_hpages != -1 && hugetlb_acct_memory(h, min_hpages)) {
108 kfree(spool);
109 return NULL;
110 }
111 spool->rsv_hpages = min_hpages;
David Gibson90481622012-03-21 16:34:12 -0700112
113 return spool;
114}
115
116void hugepage_put_subpool(struct hugepage_subpool *spool)
117{
118 spin_lock(&spool->lock);
119 BUG_ON(!spool->count);
120 spool->count--;
121 unlock_or_release_subpool(spool);
122}
123
Mike Kravetz1c5ecae2015-04-15 16:13:39 -0700124/*
125 * Subpool accounting for allocating and reserving pages.
126 * Return -ENOMEM if there are not enough resources to satisfy the
127 * the request. Otherwise, return the number of pages by which the
128 * global pools must be adjusted (upward). The returned value may
129 * only be different than the passed value (delta) in the case where
130 * a subpool minimum size must be manitained.
131 */
132static long hugepage_subpool_get_pages(struct hugepage_subpool *spool,
David Gibson90481622012-03-21 16:34:12 -0700133 long delta)
134{
Mike Kravetz1c5ecae2015-04-15 16:13:39 -0700135 long ret = delta;
David Gibson90481622012-03-21 16:34:12 -0700136
137 if (!spool)
Mike Kravetz1c5ecae2015-04-15 16:13:39 -0700138 return ret;
David Gibson90481622012-03-21 16:34:12 -0700139
140 spin_lock(&spool->lock);
David Gibson90481622012-03-21 16:34:12 -0700141
Mike Kravetz1c5ecae2015-04-15 16:13:39 -0700142 if (spool->max_hpages != -1) { /* maximum size accounting */
143 if ((spool->used_hpages + delta) <= spool->max_hpages)
144 spool->used_hpages += delta;
145 else {
146 ret = -ENOMEM;
147 goto unlock_ret;
148 }
149 }
150
Mike Kravetz09a95e22016-05-19 17:11:01 -0700151 /* minimum size accounting */
152 if (spool->min_hpages != -1 && spool->rsv_hpages) {
Mike Kravetz1c5ecae2015-04-15 16:13:39 -0700153 if (delta > spool->rsv_hpages) {
154 /*
155 * Asking for more reserves than those already taken on
156 * behalf of subpool. Return difference.
157 */
158 ret = delta - spool->rsv_hpages;
159 spool->rsv_hpages = 0;
160 } else {
161 ret = 0; /* reserves already accounted for */
162 spool->rsv_hpages -= delta;
163 }
164 }
165
166unlock_ret:
167 spin_unlock(&spool->lock);
David Gibson90481622012-03-21 16:34:12 -0700168 return ret;
169}
170
Mike Kravetz1c5ecae2015-04-15 16:13:39 -0700171/*
172 * Subpool accounting for freeing and unreserving pages.
173 * Return the number of global page reservations that must be dropped.
174 * The return value may only be different than the passed value (delta)
175 * in the case where a subpool minimum size must be maintained.
176 */
177static long hugepage_subpool_put_pages(struct hugepage_subpool *spool,
David Gibson90481622012-03-21 16:34:12 -0700178 long delta)
179{
Mike Kravetz1c5ecae2015-04-15 16:13:39 -0700180 long ret = delta;
181
David Gibson90481622012-03-21 16:34:12 -0700182 if (!spool)
Mike Kravetz1c5ecae2015-04-15 16:13:39 -0700183 return delta;
David Gibson90481622012-03-21 16:34:12 -0700184
185 spin_lock(&spool->lock);
Mike Kravetz1c5ecae2015-04-15 16:13:39 -0700186
187 if (spool->max_hpages != -1) /* maximum size accounting */
188 spool->used_hpages -= delta;
189
Mike Kravetz09a95e22016-05-19 17:11:01 -0700190 /* minimum size accounting */
191 if (spool->min_hpages != -1 && spool->used_hpages < spool->min_hpages) {
Mike Kravetz1c5ecae2015-04-15 16:13:39 -0700192 if (spool->rsv_hpages + delta <= spool->min_hpages)
193 ret = 0;
194 else
195 ret = spool->rsv_hpages + delta - spool->min_hpages;
196
197 spool->rsv_hpages += delta;
198 if (spool->rsv_hpages > spool->min_hpages)
199 spool->rsv_hpages = spool->min_hpages;
200 }
201
202 /*
203 * If hugetlbfs_put_super couldn't free spool due to an outstanding
204 * quota reference, free it now.
205 */
David Gibson90481622012-03-21 16:34:12 -0700206 unlock_or_release_subpool(spool);
Mike Kravetz1c5ecae2015-04-15 16:13:39 -0700207
208 return ret;
David Gibson90481622012-03-21 16:34:12 -0700209}
210
211static inline struct hugepage_subpool *subpool_inode(struct inode *inode)
212{
213 return HUGETLBFS_SB(inode->i_sb)->spool;
214}
215
216static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma)
217{
Al Viro496ad9a2013-01-23 17:07:38 -0500218 return subpool_inode(file_inode(vma->vm_file));
David Gibson90481622012-03-21 16:34:12 -0700219}
220
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700221/*
Andy Whitcroft96822902008-07-23 21:27:29 -0700222 * Region tracking -- allows tracking of reservations and instantiated pages
223 * across the pages in a mapping.
Andy Whitcroft84afd992008-07-23 21:27:32 -0700224 *
Mike Kravetz1dd308a2015-06-24 16:57:52 -0700225 * The region data structures are embedded into a resv_map and protected
226 * by a resv_map's lock. The set of regions within the resv_map represent
227 * reservations for huge pages, or huge pages that have already been
228 * instantiated within the map. The from and to elements are huge page
229 * indicies into the associated mapping. from indicates the starting index
230 * of the region. to represents the first index past the end of the region.
231 *
232 * For example, a file region structure with from == 0 and to == 4 represents
233 * four huge pages in a mapping. It is important to note that the to element
234 * represents the first element past the end of the region. This is used in
235 * arithmetic as 4(to) - 0(from) = 4 huge pages in the region.
236 *
237 * Interval notation of the form [from, to) will be used to indicate that
238 * the endpoint from is inclusive and to is exclusive.
Andy Whitcroft96822902008-07-23 21:27:29 -0700239 */
240struct file_region {
241 struct list_head link;
242 long from;
243 long to;
244};
245
Mike Kravetz1dd308a2015-06-24 16:57:52 -0700246/*
247 * Add the huge page range represented by [f, t) to the reserve
Mike Kravetz5e911372015-09-08 15:01:28 -0700248 * map. In the normal case, existing regions will be expanded
249 * to accommodate the specified range. Sufficient regions should
250 * exist for expansion due to the previous call to region_chg
251 * with the same range. However, it is possible that region_del
252 * could have been called after region_chg and modifed the map
253 * in such a way that no region exists to be expanded. In this
254 * case, pull a region descriptor from the cache associated with
255 * the map and use that for the new range.
Mike Kravetzcf3ad202015-06-24 16:57:55 -0700256 *
257 * Return the number of new huge pages added to the map. This
258 * number is greater than or equal to zero.
Mike Kravetz1dd308a2015-06-24 16:57:52 -0700259 */
Joonsoo Kim1406ec92014-04-03 14:47:26 -0700260static long region_add(struct resv_map *resv, long f, long t)
Andy Whitcroft96822902008-07-23 21:27:29 -0700261{
Joonsoo Kim1406ec92014-04-03 14:47:26 -0700262 struct list_head *head = &resv->regions;
Andy Whitcroft96822902008-07-23 21:27:29 -0700263 struct file_region *rg, *nrg, *trg;
Mike Kravetzcf3ad202015-06-24 16:57:55 -0700264 long add = 0;
Andy Whitcroft96822902008-07-23 21:27:29 -0700265
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700266 spin_lock(&resv->lock);
Andy Whitcroft96822902008-07-23 21:27:29 -0700267 /* Locate the region we are either in or before. */
268 list_for_each_entry(rg, head, link)
269 if (f <= rg->to)
270 break;
271
Mike Kravetz5e911372015-09-08 15:01:28 -0700272 /*
273 * If no region exists which can be expanded to include the
274 * specified range, the list must have been modified by an
275 * interleving call to region_del(). Pull a region descriptor
276 * from the cache and use it for this range.
277 */
278 if (&rg->link == head || t < rg->from) {
279 VM_BUG_ON(resv->region_cache_count <= 0);
280
281 resv->region_cache_count--;
282 nrg = list_first_entry(&resv->region_cache, struct file_region,
283 link);
284 list_del(&nrg->link);
285
286 nrg->from = f;
287 nrg->to = t;
288 list_add(&nrg->link, rg->link.prev);
289
290 add += t - f;
291 goto out_locked;
292 }
293
Andy Whitcroft96822902008-07-23 21:27:29 -0700294 /* Round our left edge to the current segment if it encloses us. */
295 if (f > rg->from)
296 f = rg->from;
297
298 /* Check for and consume any regions we now overlap with. */
299 nrg = rg;
300 list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
301 if (&rg->link == head)
302 break;
303 if (rg->from > t)
304 break;
305
306 /* If this area reaches higher then extend our area to
307 * include it completely. If this is not the first area
308 * which we intend to reuse, free it. */
309 if (rg->to > t)
310 t = rg->to;
311 if (rg != nrg) {
Mike Kravetzcf3ad202015-06-24 16:57:55 -0700312 /* Decrement return value by the deleted range.
313 * Another range will span this area so that by
314 * end of routine add will be >= zero
315 */
316 add -= (rg->to - rg->from);
Andy Whitcroft96822902008-07-23 21:27:29 -0700317 list_del(&rg->link);
318 kfree(rg);
319 }
320 }
Mike Kravetzcf3ad202015-06-24 16:57:55 -0700321
322 add += (nrg->from - f); /* Added to beginning of region */
Andy Whitcroft96822902008-07-23 21:27:29 -0700323 nrg->from = f;
Mike Kravetzcf3ad202015-06-24 16:57:55 -0700324 add += t - nrg->to; /* Added to end of region */
Andy Whitcroft96822902008-07-23 21:27:29 -0700325 nrg->to = t;
Mike Kravetzcf3ad202015-06-24 16:57:55 -0700326
Mike Kravetz5e911372015-09-08 15:01:28 -0700327out_locked:
328 resv->adds_in_progress--;
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700329 spin_unlock(&resv->lock);
Mike Kravetzcf3ad202015-06-24 16:57:55 -0700330 VM_BUG_ON(add < 0);
331 return add;
Andy Whitcroft96822902008-07-23 21:27:29 -0700332}
333
Mike Kravetz1dd308a2015-06-24 16:57:52 -0700334/*
335 * Examine the existing reserve map and determine how many
336 * huge pages in the specified range [f, t) are NOT currently
337 * represented. This routine is called before a subsequent
338 * call to region_add that will actually modify the reserve
339 * map to add the specified range [f, t). region_chg does
340 * not change the number of huge pages represented by the
341 * map. However, if the existing regions in the map can not
342 * be expanded to represent the new range, a new file_region
343 * structure is added to the map as a placeholder. This is
344 * so that the subsequent region_add call will have all the
345 * regions it needs and will not fail.
346 *
Mike Kravetz5e911372015-09-08 15:01:28 -0700347 * Upon entry, region_chg will also examine the cache of region descriptors
348 * associated with the map. If there are not enough descriptors cached, one
349 * will be allocated for the in progress add operation.
350 *
351 * Returns the number of huge pages that need to be added to the existing
352 * reservation map for the range [f, t). This number is greater or equal to
353 * zero. -ENOMEM is returned if a new file_region structure or cache entry
354 * is needed and can not be allocated.
Mike Kravetz1dd308a2015-06-24 16:57:52 -0700355 */
Joonsoo Kim1406ec92014-04-03 14:47:26 -0700356static long region_chg(struct resv_map *resv, long f, long t)
Andy Whitcroft96822902008-07-23 21:27:29 -0700357{
Joonsoo Kim1406ec92014-04-03 14:47:26 -0700358 struct list_head *head = &resv->regions;
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700359 struct file_region *rg, *nrg = NULL;
Andy Whitcroft96822902008-07-23 21:27:29 -0700360 long chg = 0;
361
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700362retry:
363 spin_lock(&resv->lock);
Mike Kravetz5e911372015-09-08 15:01:28 -0700364retry_locked:
365 resv->adds_in_progress++;
366
367 /*
368 * Check for sufficient descriptors in the cache to accommodate
369 * the number of in progress add operations.
370 */
371 if (resv->adds_in_progress > resv->region_cache_count) {
372 struct file_region *trg;
373
374 VM_BUG_ON(resv->adds_in_progress - resv->region_cache_count > 1);
375 /* Must drop lock to allocate a new descriptor. */
376 resv->adds_in_progress--;
377 spin_unlock(&resv->lock);
378
379 trg = kmalloc(sizeof(*trg), GFP_KERNEL);
Mike Kravetzdbe409e2015-12-11 13:40:52 -0800380 if (!trg) {
381 kfree(nrg);
Mike Kravetz5e911372015-09-08 15:01:28 -0700382 return -ENOMEM;
Mike Kravetzdbe409e2015-12-11 13:40:52 -0800383 }
Mike Kravetz5e911372015-09-08 15:01:28 -0700384
385 spin_lock(&resv->lock);
386 list_add(&trg->link, &resv->region_cache);
387 resv->region_cache_count++;
388 goto retry_locked;
389 }
390
Andy Whitcroft96822902008-07-23 21:27:29 -0700391 /* Locate the region we are before or in. */
392 list_for_each_entry(rg, head, link)
393 if (f <= rg->to)
394 break;
395
396 /* If we are below the current region then a new region is required.
397 * Subtle, allocate a new region at the position but make it zero
398 * size such that we can guarantee to record the reservation. */
399 if (&rg->link == head || t < rg->from) {
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700400 if (!nrg) {
Mike Kravetz5e911372015-09-08 15:01:28 -0700401 resv->adds_in_progress--;
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700402 spin_unlock(&resv->lock);
403 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
404 if (!nrg)
405 return -ENOMEM;
Andy Whitcroft96822902008-07-23 21:27:29 -0700406
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700407 nrg->from = f;
408 nrg->to = f;
409 INIT_LIST_HEAD(&nrg->link);
410 goto retry;
411 }
412
413 list_add(&nrg->link, rg->link.prev);
414 chg = t - f;
415 goto out_nrg;
Andy Whitcroft96822902008-07-23 21:27:29 -0700416 }
417
418 /* Round our left edge to the current segment if it encloses us. */
419 if (f > rg->from)
420 f = rg->from;
421 chg = t - f;
422
423 /* Check for and consume any regions we now overlap with. */
424 list_for_each_entry(rg, rg->link.prev, link) {
425 if (&rg->link == head)
426 break;
427 if (rg->from > t)
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700428 goto out;
Andy Whitcroft96822902008-07-23 21:27:29 -0700429
Lucas De Marchi25985ed2011-03-30 22:57:33 -0300430 /* We overlap with this area, if it extends further than
Andy Whitcroft96822902008-07-23 21:27:29 -0700431 * us then we must extend ourselves. Account for its
432 * existing reservation. */
433 if (rg->to > t) {
434 chg += rg->to - t;
435 t = rg->to;
436 }
437 chg -= rg->to - rg->from;
438 }
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700439
440out:
441 spin_unlock(&resv->lock);
442 /* We already know we raced and no longer need the new region */
443 kfree(nrg);
444 return chg;
445out_nrg:
446 spin_unlock(&resv->lock);
Andy Whitcroft96822902008-07-23 21:27:29 -0700447 return chg;
448}
449
Mike Kravetz1dd308a2015-06-24 16:57:52 -0700450/*
Mike Kravetz5e911372015-09-08 15:01:28 -0700451 * Abort the in progress add operation. The adds_in_progress field
452 * of the resv_map keeps track of the operations in progress between
453 * calls to region_chg and region_add. Operations are sometimes
454 * aborted after the call to region_chg. In such cases, region_abort
455 * is called to decrement the adds_in_progress counter.
456 *
457 * NOTE: The range arguments [f, t) are not needed or used in this
458 * routine. They are kept to make reading the calling code easier as
459 * arguments will match the associated region_chg call.
460 */
461static void region_abort(struct resv_map *resv, long f, long t)
462{
463 spin_lock(&resv->lock);
464 VM_BUG_ON(!resv->region_cache_count);
465 resv->adds_in_progress--;
466 spin_unlock(&resv->lock);
467}
468
469/*
Mike Kravetzfeba16e2015-09-08 15:01:31 -0700470 * Delete the specified range [f, t) from the reserve map. If the
471 * t parameter is LONG_MAX, this indicates that ALL regions after f
472 * should be deleted. Locate the regions which intersect [f, t)
473 * and either trim, delete or split the existing regions.
474 *
475 * Returns the number of huge pages deleted from the reserve map.
476 * In the normal case, the return value is zero or more. In the
477 * case where a region must be split, a new region descriptor must
478 * be allocated. If the allocation fails, -ENOMEM will be returned.
479 * NOTE: If the parameter t == LONG_MAX, then we will never split
480 * a region and possibly return -ENOMEM. Callers specifying
481 * t == LONG_MAX do not need to check for -ENOMEM error.
Mike Kravetz1dd308a2015-06-24 16:57:52 -0700482 */
Mike Kravetzfeba16e2015-09-08 15:01:31 -0700483static long region_del(struct resv_map *resv, long f, long t)
Andy Whitcroft96822902008-07-23 21:27:29 -0700484{
Joonsoo Kim1406ec92014-04-03 14:47:26 -0700485 struct list_head *head = &resv->regions;
Andy Whitcroft96822902008-07-23 21:27:29 -0700486 struct file_region *rg, *trg;
Mike Kravetzfeba16e2015-09-08 15:01:31 -0700487 struct file_region *nrg = NULL;
488 long del = 0;
Andy Whitcroft96822902008-07-23 21:27:29 -0700489
Mike Kravetzfeba16e2015-09-08 15:01:31 -0700490retry:
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700491 spin_lock(&resv->lock);
Mike Kravetzfeba16e2015-09-08 15:01:31 -0700492 list_for_each_entry_safe(rg, trg, head, link) {
Mike Kravetzdbe409e2015-12-11 13:40:52 -0800493 /*
494 * Skip regions before the range to be deleted. file_region
495 * ranges are normally of the form [from, to). However, there
496 * may be a "placeholder" entry in the map which is of the form
497 * (from, to) with from == to. Check for placeholder entries
498 * at the beginning of the range to be deleted.
499 */
500 if (rg->to <= f && (rg->to != rg->from || rg->to != f))
Mike Kravetzfeba16e2015-09-08 15:01:31 -0700501 continue;
Mike Kravetzdbe409e2015-12-11 13:40:52 -0800502
Mike Kravetzfeba16e2015-09-08 15:01:31 -0700503 if (rg->from >= t)
Andy Whitcroft96822902008-07-23 21:27:29 -0700504 break;
Andy Whitcroft96822902008-07-23 21:27:29 -0700505
Mike Kravetzfeba16e2015-09-08 15:01:31 -0700506 if (f > rg->from && t < rg->to) { /* Must split region */
507 /*
508 * Check for an entry in the cache before dropping
509 * lock and attempting allocation.
510 */
511 if (!nrg &&
512 resv->region_cache_count > resv->adds_in_progress) {
513 nrg = list_first_entry(&resv->region_cache,
514 struct file_region,
515 link);
516 list_del(&nrg->link);
517 resv->region_cache_count--;
518 }
519
520 if (!nrg) {
521 spin_unlock(&resv->lock);
522 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
523 if (!nrg)
524 return -ENOMEM;
525 goto retry;
526 }
527
528 del += t - f;
529
530 /* New entry for end of split region */
531 nrg->from = t;
532 nrg->to = rg->to;
533 INIT_LIST_HEAD(&nrg->link);
534
535 /* Original entry is trimmed */
536 rg->to = f;
537
538 list_add(&nrg->link, &rg->link);
539 nrg = NULL;
540 break;
541 }
542
543 if (f <= rg->from && t >= rg->to) { /* Remove entire region */
544 del += rg->to - rg->from;
545 list_del(&rg->link);
546 kfree(rg);
547 continue;
548 }
549
550 if (f <= rg->from) { /* Trim beginning of region */
551 del += t - rg->from;
552 rg->from = t;
553 } else { /* Trim end of region */
554 del += rg->to - f;
555 rg->to = f;
556 }
Andy Whitcroft96822902008-07-23 21:27:29 -0700557 }
558
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700559 spin_unlock(&resv->lock);
Mike Kravetzfeba16e2015-09-08 15:01:31 -0700560 kfree(nrg);
561 return del;
Andy Whitcroft96822902008-07-23 21:27:29 -0700562}
563
Mike Kravetz1dd308a2015-06-24 16:57:52 -0700564/*
Mike Kravetzb5cec282015-09-08 15:01:41 -0700565 * A rare out of memory error was encountered which prevented removal of
566 * the reserve map region for a page. The huge page itself was free'ed
567 * and removed from the page cache. This routine will adjust the subpool
568 * usage count, and the global reserve count if needed. By incrementing
569 * these counts, the reserve map entry which could not be deleted will
570 * appear as a "reserved" entry instead of simply dangling with incorrect
571 * counts.
572 */
zhong jiang72e29362016-10-07 17:02:01 -0700573void hugetlb_fix_reserve_counts(struct inode *inode)
Mike Kravetzb5cec282015-09-08 15:01:41 -0700574{
575 struct hugepage_subpool *spool = subpool_inode(inode);
576 long rsv_adjust;
577
578 rsv_adjust = hugepage_subpool_get_pages(spool, 1);
zhong jiang72e29362016-10-07 17:02:01 -0700579 if (rsv_adjust) {
Mike Kravetzb5cec282015-09-08 15:01:41 -0700580 struct hstate *h = hstate_inode(inode);
581
582 hugetlb_acct_memory(h, 1);
583 }
584}
585
586/*
Mike Kravetz1dd308a2015-06-24 16:57:52 -0700587 * Count and return the number of huge pages in the reserve map
588 * that intersect with the range [f, t).
589 */
Joonsoo Kim1406ec92014-04-03 14:47:26 -0700590static long region_count(struct resv_map *resv, long f, long t)
Andy Whitcroft84afd992008-07-23 21:27:32 -0700591{
Joonsoo Kim1406ec92014-04-03 14:47:26 -0700592 struct list_head *head = &resv->regions;
Andy Whitcroft84afd992008-07-23 21:27:32 -0700593 struct file_region *rg;
594 long chg = 0;
595
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700596 spin_lock(&resv->lock);
Andy Whitcroft84afd992008-07-23 21:27:32 -0700597 /* Locate each segment we overlap with, and count that overlap. */
598 list_for_each_entry(rg, head, link) {
Wang Sheng-Huif2135a42012-05-29 15:06:17 -0700599 long seg_from;
600 long seg_to;
Andy Whitcroft84afd992008-07-23 21:27:32 -0700601
602 if (rg->to <= f)
603 continue;
604 if (rg->from >= t)
605 break;
606
607 seg_from = max(rg->from, f);
608 seg_to = min(rg->to, t);
609
610 chg += seg_to - seg_from;
611 }
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700612 spin_unlock(&resv->lock);
Andy Whitcroft84afd992008-07-23 21:27:32 -0700613
614 return chg;
615}
616
Andy Whitcroft96822902008-07-23 21:27:29 -0700617/*
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700618 * Convert the address within this vma to the page offset within
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700619 * the mapping, in pagecache page units; huge pages here.
620 */
Andi Kleena5516432008-07-23 21:27:41 -0700621static pgoff_t vma_hugecache_offset(struct hstate *h,
622 struct vm_area_struct *vma, unsigned long address)
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700623{
Andi Kleena5516432008-07-23 21:27:41 -0700624 return ((address - vma->vm_start) >> huge_page_shift(h)) +
625 (vma->vm_pgoff >> huge_page_order(h));
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700626}
627
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +0900628pgoff_t linear_hugepage_index(struct vm_area_struct *vma,
629 unsigned long address)
630{
631 return vma_hugecache_offset(hstate_vma(vma), vma, address);
632}
Dan Williamsdee41072016-05-14 12:20:44 -0700633EXPORT_SYMBOL_GPL(linear_hugepage_index);
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +0900634
Andy Whitcroft84afd992008-07-23 21:27:32 -0700635/*
Mel Gorman08fba692009-01-06 14:38:53 -0800636 * Return the size of the pages allocated when backing a VMA. In the majority
637 * cases this will be same size as used by the page table entries.
638 */
639unsigned long vma_kernel_pagesize(struct vm_area_struct *vma)
640{
641 struct hstate *hstate;
642
643 if (!is_vm_hugetlb_page(vma))
644 return PAGE_SIZE;
645
646 hstate = hstate_vma(vma);
647
Wanpeng Li2415cf12013-07-03 15:02:43 -0700648 return 1UL << huge_page_shift(hstate);
Mel Gorman08fba692009-01-06 14:38:53 -0800649}
Joerg Roedelf340ca02009-06-19 15:16:22 +0200650EXPORT_SYMBOL_GPL(vma_kernel_pagesize);
Mel Gorman08fba692009-01-06 14:38:53 -0800651
652/*
Mel Gorman33402892009-01-06 14:38:54 -0800653 * Return the page size being used by the MMU to back a VMA. In the majority
654 * of cases, the page size used by the kernel matches the MMU size. On
655 * architectures where it differs, an architecture-specific version of this
656 * function is required.
657 */
658#ifndef vma_mmu_pagesize
659unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
660{
661 return vma_kernel_pagesize(vma);
662}
663#endif
664
665/*
Andy Whitcroft84afd992008-07-23 21:27:32 -0700666 * Flags for MAP_PRIVATE reservations. These are stored in the bottom
667 * bits of the reservation map pointer, which are always clear due to
668 * alignment.
669 */
670#define HPAGE_RESV_OWNER (1UL << 0)
671#define HPAGE_RESV_UNMAPPED (1UL << 1)
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700672#define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
Andy Whitcroft84afd992008-07-23 21:27:32 -0700673
Mel Gormana1e78772008-07-23 21:27:23 -0700674/*
675 * These helpers are used to track how many pages are reserved for
676 * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
677 * is guaranteed to have their future faults succeed.
678 *
679 * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
680 * the reserve counters are updated with the hugetlb_lock held. It is safe
681 * to reset the VMA at fork() time as it is not in use yet and there is no
682 * chance of the global counters getting corrupted as a result of the values.
Andy Whitcroft84afd992008-07-23 21:27:32 -0700683 *
684 * The private mapping reservation is represented in a subtly different
685 * manner to a shared mapping. A shared mapping has a region map associated
686 * with the underlying file, this region map represents the backing file
687 * pages which have ever had a reservation assigned which this persists even
688 * after the page is instantiated. A private mapping has a region map
689 * associated with the original mmap which is attached to all VMAs which
690 * reference it, this region map represents those offsets which have consumed
691 * reservation ie. where pages have been instantiated.
Mel Gormana1e78772008-07-23 21:27:23 -0700692 */
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700693static unsigned long get_vma_private_data(struct vm_area_struct *vma)
694{
695 return (unsigned long)vma->vm_private_data;
696}
697
698static void set_vma_private_data(struct vm_area_struct *vma,
699 unsigned long value)
700{
701 vma->vm_private_data = (void *)value;
702}
703
Joonsoo Kim9119a412014-04-03 14:47:25 -0700704struct resv_map *resv_map_alloc(void)
Andy Whitcroft84afd992008-07-23 21:27:32 -0700705{
706 struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
Mike Kravetz5e911372015-09-08 15:01:28 -0700707 struct file_region *rg = kmalloc(sizeof(*rg), GFP_KERNEL);
708
709 if (!resv_map || !rg) {
710 kfree(resv_map);
711 kfree(rg);
Andy Whitcroft84afd992008-07-23 21:27:32 -0700712 return NULL;
Mike Kravetz5e911372015-09-08 15:01:28 -0700713 }
Andy Whitcroft84afd992008-07-23 21:27:32 -0700714
715 kref_init(&resv_map->refs);
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700716 spin_lock_init(&resv_map->lock);
Andy Whitcroft84afd992008-07-23 21:27:32 -0700717 INIT_LIST_HEAD(&resv_map->regions);
718
Mike Kravetz5e911372015-09-08 15:01:28 -0700719 resv_map->adds_in_progress = 0;
720
721 INIT_LIST_HEAD(&resv_map->region_cache);
722 list_add(&rg->link, &resv_map->region_cache);
723 resv_map->region_cache_count = 1;
724
Andy Whitcroft84afd992008-07-23 21:27:32 -0700725 return resv_map;
726}
727
Joonsoo Kim9119a412014-04-03 14:47:25 -0700728void resv_map_release(struct kref *ref)
Andy Whitcroft84afd992008-07-23 21:27:32 -0700729{
730 struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
Mike Kravetz5e911372015-09-08 15:01:28 -0700731 struct list_head *head = &resv_map->region_cache;
732 struct file_region *rg, *trg;
Andy Whitcroft84afd992008-07-23 21:27:32 -0700733
734 /* Clear out any active regions before we release the map. */
Mike Kravetzfeba16e2015-09-08 15:01:31 -0700735 region_del(resv_map, 0, LONG_MAX);
Mike Kravetz5e911372015-09-08 15:01:28 -0700736
737 /* ... and any entries left in the cache */
738 list_for_each_entry_safe(rg, trg, head, link) {
739 list_del(&rg->link);
740 kfree(rg);
741 }
742
743 VM_BUG_ON(resv_map->adds_in_progress);
744
Andy Whitcroft84afd992008-07-23 21:27:32 -0700745 kfree(resv_map);
746}
747
Joonsoo Kim4e35f482014-04-03 14:47:30 -0700748static inline struct resv_map *inode_resv_map(struct inode *inode)
749{
750 return inode->i_mapping->private_data;
751}
752
Andy Whitcroft84afd992008-07-23 21:27:32 -0700753static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
Mel Gormana1e78772008-07-23 21:27:23 -0700754{
Sasha Levin81d1b092014-10-09 15:28:10 -0700755 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
Joonsoo Kim4e35f482014-04-03 14:47:30 -0700756 if (vma->vm_flags & VM_MAYSHARE) {
757 struct address_space *mapping = vma->vm_file->f_mapping;
758 struct inode *inode = mapping->host;
759
760 return inode_resv_map(inode);
761
762 } else {
Andy Whitcroft84afd992008-07-23 21:27:32 -0700763 return (struct resv_map *)(get_vma_private_data(vma) &
764 ~HPAGE_RESV_MASK);
Joonsoo Kim4e35f482014-04-03 14:47:30 -0700765 }
Mel Gormana1e78772008-07-23 21:27:23 -0700766}
767
Andy Whitcroft84afd992008-07-23 21:27:32 -0700768static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
Mel Gormana1e78772008-07-23 21:27:23 -0700769{
Sasha Levin81d1b092014-10-09 15:28:10 -0700770 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
771 VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
Mel Gormana1e78772008-07-23 21:27:23 -0700772
Andy Whitcroft84afd992008-07-23 21:27:32 -0700773 set_vma_private_data(vma, (get_vma_private_data(vma) &
774 HPAGE_RESV_MASK) | (unsigned long)map);
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700775}
776
777static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
778{
Sasha Levin81d1b092014-10-09 15:28:10 -0700779 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
780 VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700781
782 set_vma_private_data(vma, get_vma_private_data(vma) | flags);
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700783}
784
785static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
786{
Sasha Levin81d1b092014-10-09 15:28:10 -0700787 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700788
789 return (get_vma_private_data(vma) & flag) != 0;
Mel Gormana1e78772008-07-23 21:27:23 -0700790}
791
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700792/* Reset counters to 0 and clear all HPAGE_RESV_* flags */
Mel Gormana1e78772008-07-23 21:27:23 -0700793void reset_vma_resv_huge_pages(struct vm_area_struct *vma)
794{
Sasha Levin81d1b092014-10-09 15:28:10 -0700795 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
Mel Gormanf83a2752009-05-28 14:34:40 -0700796 if (!(vma->vm_flags & VM_MAYSHARE))
Mel Gormana1e78772008-07-23 21:27:23 -0700797 vma->vm_private_data = (void *)0;
798}
799
800/* Returns true if the VMA has associated reserve pages */
Nicholas Krause559ec2f2015-09-04 15:48:27 -0700801static bool vma_has_reserves(struct vm_area_struct *vma, long chg)
Mel Gormana1e78772008-07-23 21:27:23 -0700802{
Joonsoo Kimaf0ed732013-09-11 14:21:18 -0700803 if (vma->vm_flags & VM_NORESERVE) {
804 /*
805 * This address is already reserved by other process(chg == 0),
806 * so, we should decrement reserved count. Without decrementing,
807 * reserve count remains after releasing inode, because this
808 * allocated page will go into page cache and is regarded as
809 * coming from reserved pool in releasing step. Currently, we
810 * don't have any other solution to deal with this situation
811 * properly, so add work-around here.
812 */
813 if (vma->vm_flags & VM_MAYSHARE && chg == 0)
Nicholas Krause559ec2f2015-09-04 15:48:27 -0700814 return true;
Joonsoo Kimaf0ed732013-09-11 14:21:18 -0700815 else
Nicholas Krause559ec2f2015-09-04 15:48:27 -0700816 return false;
Joonsoo Kimaf0ed732013-09-11 14:21:18 -0700817 }
Joonsoo Kima63884e2013-09-11 14:21:07 -0700818
819 /* Shared mappings always use reserves */
Mike Kravetz1fb1b0e2015-09-08 15:01:44 -0700820 if (vma->vm_flags & VM_MAYSHARE) {
821 /*
822 * We know VM_NORESERVE is not set. Therefore, there SHOULD
823 * be a region map for all pages. The only situation where
824 * there is no region map is if a hole was punched via
825 * fallocate. In this case, there really are no reverves to
826 * use. This situation is indicated if chg != 0.
827 */
828 if (chg)
829 return false;
830 else
831 return true;
832 }
Joonsoo Kima63884e2013-09-11 14:21:07 -0700833
834 /*
835 * Only the process that called mmap() has reserves for
836 * private mappings.
837 */
Mike Kravetz67961f92016-06-08 15:33:42 -0700838 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
839 /*
840 * Like the shared case above, a hole punch or truncate
841 * could have been performed on the private mapping.
842 * Examine the value of chg to determine if reserves
843 * actually exist or were previously consumed.
844 * Very Subtle - The value of chg comes from a previous
845 * call to vma_needs_reserves(). The reserve map for
846 * private mappings has different (opposite) semantics
847 * than that of shared mappings. vma_needs_reserves()
848 * has already taken this difference in semantics into
849 * account. Therefore, the meaning of chg is the same
850 * as in the shared case above. Code could easily be
851 * combined, but keeping it separate draws attention to
852 * subtle differences.
853 */
854 if (chg)
855 return false;
856 else
857 return true;
858 }
Joonsoo Kima63884e2013-09-11 14:21:07 -0700859
Nicholas Krause559ec2f2015-09-04 15:48:27 -0700860 return false;
Mel Gormana1e78772008-07-23 21:27:23 -0700861}
862
Andi Kleena5516432008-07-23 21:27:41 -0700863static void enqueue_huge_page(struct hstate *h, struct page *page)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700864{
865 int nid = page_to_nid(page);
Aneesh Kumar K.V0edaecf2012-07-31 16:42:07 -0700866 list_move(&page->lru, &h->hugepage_freelists[nid]);
Andi Kleena5516432008-07-23 21:27:41 -0700867 h->free_huge_pages++;
868 h->free_huge_pages_node[nid]++;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700869}
870
Anshuman Khandual94310cb2017-07-06 15:38:38 -0700871static struct page *dequeue_huge_page_node_exact(struct hstate *h, int nid)
Naoya Horiguchibf50bab2010-09-08 10:19:33 +0900872{
873 struct page *page;
874
Naoya Horiguchic8721bb2013-09-11 14:22:09 -0700875 list_for_each_entry(page, &h->hugepage_freelists[nid], lru)
Naoya Horiguchi243abd52017-07-10 15:47:32 -0700876 if (!PageHWPoison(page))
Naoya Horiguchic8721bb2013-09-11 14:22:09 -0700877 break;
878 /*
879 * if 'non-isolated free hugepage' not found on the list,
880 * the allocation fails.
881 */
882 if (&h->hugepage_freelists[nid] == &page->lru)
Naoya Horiguchibf50bab2010-09-08 10:19:33 +0900883 return NULL;
Aneesh Kumar K.V0edaecf2012-07-31 16:42:07 -0700884 list_move(&page->lru, &h->hugepage_activelist);
Naoya Horiguchia9869b82010-09-08 10:19:37 +0900885 set_page_refcounted(page);
Naoya Horiguchibf50bab2010-09-08 10:19:33 +0900886 h->free_huge_pages--;
887 h->free_huge_pages_node[nid]--;
888 return page;
889}
890
Michal Hocko3e59fcb2017-07-10 15:49:11 -0700891static struct page *dequeue_huge_page_nodemask(struct hstate *h, gfp_t gfp_mask, int nid,
892 nodemask_t *nmask)
Anshuman Khandual94310cb2017-07-06 15:38:38 -0700893{
Michal Hocko3e59fcb2017-07-10 15:49:11 -0700894 unsigned int cpuset_mems_cookie;
895 struct zonelist *zonelist;
896 struct zone *zone;
897 struct zoneref *z;
898 int node = -1;
Anshuman Khandual94310cb2017-07-06 15:38:38 -0700899
Michal Hocko3e59fcb2017-07-10 15:49:11 -0700900 zonelist = node_zonelist(nid, gfp_mask);
Anshuman Khandual94310cb2017-07-06 15:38:38 -0700901
Michal Hocko3e59fcb2017-07-10 15:49:11 -0700902retry_cpuset:
903 cpuset_mems_cookie = read_mems_allowed_begin();
904 for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), nmask) {
905 struct page *page;
906
907 if (!cpuset_zone_allowed(zone, gfp_mask))
908 continue;
909 /*
910 * no need to ask again on the same node. Pool is node rather than
911 * zone aware
912 */
913 if (zone_to_nid(zone) == node)
914 continue;
915 node = zone_to_nid(zone);
916
Anshuman Khandual94310cb2017-07-06 15:38:38 -0700917 page = dequeue_huge_page_node_exact(h, node);
918 if (page)
919 return page;
920 }
Michal Hocko3e59fcb2017-07-10 15:49:11 -0700921 if (unlikely(read_mems_allowed_retry(cpuset_mems_cookie)))
922 goto retry_cpuset;
923
Anshuman Khandual94310cb2017-07-06 15:38:38 -0700924 return NULL;
925}
926
Naoya Horiguchi86cdb462013-09-11 14:22:13 -0700927/* Movability of hugepages depends on migration support. */
928static inline gfp_t htlb_alloc_mask(struct hstate *h)
929{
Naoya Horiguchi100873d2014-06-04 16:10:56 -0700930 if (hugepages_treat_as_movable || hugepage_migration_supported(h))
Naoya Horiguchi86cdb462013-09-11 14:22:13 -0700931 return GFP_HIGHUSER_MOVABLE;
932 else
933 return GFP_HIGHUSER;
934}
935
Andi Kleena5516432008-07-23 21:27:41 -0700936static struct page *dequeue_huge_page_vma(struct hstate *h,
937 struct vm_area_struct *vma,
Joonsoo Kimaf0ed732013-09-11 14:21:18 -0700938 unsigned long address, int avoid_reserve,
939 long chg)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700940{
Michal Hocko3e59fcb2017-07-10 15:49:11 -0700941 struct page *page;
Lee Schermerhorn480eccf2007-09-18 22:46:47 -0700942 struct mempolicy *mpol;
Vlastimil Babka04ec6262017-07-06 15:40:03 -0700943 gfp_t gfp_mask;
Michal Hocko3e59fcb2017-07-10 15:49:11 -0700944 nodemask_t *nodemask;
Vlastimil Babka04ec6262017-07-06 15:40:03 -0700945 int nid;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700946
Mel Gormana1e78772008-07-23 21:27:23 -0700947 /*
948 * A child process with MAP_PRIVATE mappings created by their parent
949 * have no page reserves. This check ensures that reservations are
950 * not "stolen". The child may still get SIGKILLed
951 */
Joonsoo Kimaf0ed732013-09-11 14:21:18 -0700952 if (!vma_has_reserves(vma, chg) &&
Andi Kleena5516432008-07-23 21:27:41 -0700953 h->free_huge_pages - h->resv_huge_pages == 0)
Miao Xiec0ff7452010-05-24 14:32:08 -0700954 goto err;
Mel Gormana1e78772008-07-23 21:27:23 -0700955
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700956 /* If reserves cannot be used, ensure enough pages are in the pool */
Andi Kleena5516432008-07-23 21:27:41 -0700957 if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
Justin P. Mattock6eab04a2011-04-08 19:49:08 -0700958 goto err;
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700959
Vlastimil Babka04ec6262017-07-06 15:40:03 -0700960 gfp_mask = htlb_alloc_mask(h);
961 nid = huge_node(vma, address, gfp_mask, &mpol, &nodemask);
Michal Hocko3e59fcb2017-07-10 15:49:11 -0700962 page = dequeue_huge_page_nodemask(h, gfp_mask, nid, nodemask);
963 if (page && !avoid_reserve && vma_has_reserves(vma, chg)) {
964 SetPagePrivate(page);
965 h->resv_huge_pages--;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700966 }
Mel Gormancc9a6c82012-03-21 16:34:11 -0700967
968 mpol_cond_put(mpol);
Mel Gormancc9a6c82012-03-21 16:34:11 -0700969 return page;
970
Miao Xiec0ff7452010-05-24 14:32:08 -0700971err:
Mel Gormancc9a6c82012-03-21 16:34:11 -0700972 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700973}
974
Luiz Capitulino1cac6f22014-06-04 16:07:11 -0700975/*
976 * common helper functions for hstate_next_node_to_{alloc|free}.
977 * We may have allocated or freed a huge page based on a different
978 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
979 * be outside of *nodes_allowed. Ensure that we use an allowed
980 * node for alloc or free.
981 */
982static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
983{
Andrew Morton0edaf862016-05-19 17:10:58 -0700984 nid = next_node_in(nid, *nodes_allowed);
Luiz Capitulino1cac6f22014-06-04 16:07:11 -0700985 VM_BUG_ON(nid >= MAX_NUMNODES);
986
987 return nid;
988}
989
990static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
991{
992 if (!node_isset(nid, *nodes_allowed))
993 nid = next_node_allowed(nid, nodes_allowed);
994 return nid;
995}
996
997/*
998 * returns the previously saved node ["this node"] from which to
999 * allocate a persistent huge page for the pool and advance the
1000 * next node from which to allocate, handling wrap at end of node
1001 * mask.
1002 */
1003static int hstate_next_node_to_alloc(struct hstate *h,
1004 nodemask_t *nodes_allowed)
1005{
1006 int nid;
1007
1008 VM_BUG_ON(!nodes_allowed);
1009
1010 nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
1011 h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
1012
1013 return nid;
1014}
1015
1016/*
1017 * helper for free_pool_huge_page() - return the previously saved
1018 * node ["this node"] from which to free a huge page. Advance the
1019 * next node id whether or not we find a free huge page to free so
1020 * that the next attempt to free addresses the next node.
1021 */
1022static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
1023{
1024 int nid;
1025
1026 VM_BUG_ON(!nodes_allowed);
1027
1028 nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
1029 h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
1030
1031 return nid;
1032}
1033
1034#define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \
1035 for (nr_nodes = nodes_weight(*mask); \
1036 nr_nodes > 0 && \
1037 ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \
1038 nr_nodes--)
1039
1040#define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \
1041 for (nr_nodes = nodes_weight(*mask); \
1042 nr_nodes > 0 && \
1043 ((node = hstate_next_node_to_free(hs, mask)) || 1); \
1044 nr_nodes--)
1045
Aneesh Kumar K.Ve1073d12017-07-06 15:39:17 -07001046#ifdef CONFIG_ARCH_HAS_GIGANTIC_PAGE
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001047static void destroy_compound_gigantic_page(struct page *page,
Kirill A. Shutemovd00181b92015-11-06 16:29:57 -08001048 unsigned int order)
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001049{
1050 int i;
1051 int nr_pages = 1 << order;
1052 struct page *p = page + 1;
1053
Gerald Schaeferc8cc7082016-06-24 14:50:04 -07001054 atomic_set(compound_mapcount_ptr(page), 0);
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001055 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
Kirill A. Shutemov1d798ca2015-11-06 16:29:54 -08001056 clear_compound_head(p);
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001057 set_page_refcounted(p);
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001058 }
1059
1060 set_compound_order(page, 0);
1061 __ClearPageHead(page);
1062}
1063
Kirill A. Shutemovd00181b92015-11-06 16:29:57 -08001064static void free_gigantic_page(struct page *page, unsigned int order)
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001065{
1066 free_contig_range(page_to_pfn(page), 1 << order);
1067}
1068
1069static int __alloc_gigantic_page(unsigned long start_pfn,
Michal Hocko79b63f12017-09-06 16:24:03 -07001070 unsigned long nr_pages, gfp_t gfp_mask)
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001071{
1072 unsigned long end_pfn = start_pfn + nr_pages;
Lucas Stachca96b622017-02-24 14:58:37 -08001073 return alloc_contig_range(start_pfn, end_pfn, MIGRATE_MOVABLE,
Michal Hocko79b63f12017-09-06 16:24:03 -07001074 gfp_mask);
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001075}
1076
Joonsoo Kimf44b2dda82016-05-19 17:12:03 -07001077static bool pfn_range_valid_gigantic(struct zone *z,
1078 unsigned long start_pfn, unsigned long nr_pages)
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001079{
1080 unsigned long i, end_pfn = start_pfn + nr_pages;
1081 struct page *page;
1082
1083 for (i = start_pfn; i < end_pfn; i++) {
1084 if (!pfn_valid(i))
1085 return false;
1086
1087 page = pfn_to_page(i);
1088
Joonsoo Kimf44b2dda82016-05-19 17:12:03 -07001089 if (page_zone(page) != z)
1090 return false;
1091
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001092 if (PageReserved(page))
1093 return false;
1094
1095 if (page_count(page) > 0)
1096 return false;
1097
1098 if (PageHuge(page))
1099 return false;
1100 }
1101
1102 return true;
1103}
1104
1105static bool zone_spans_last_pfn(const struct zone *zone,
1106 unsigned long start_pfn, unsigned long nr_pages)
1107{
1108 unsigned long last_pfn = start_pfn + nr_pages - 1;
1109 return zone_spans_pfn(zone, last_pfn);
1110}
1111
Michal Hocko79b63f12017-09-06 16:24:03 -07001112static struct page *alloc_gigantic_page(int nid, struct hstate *h)
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001113{
Michal Hocko79b63f12017-09-06 16:24:03 -07001114 unsigned int order = huge_page_order(h);
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001115 unsigned long nr_pages = 1 << order;
1116 unsigned long ret, pfn, flags;
Michal Hocko79b63f12017-09-06 16:24:03 -07001117 struct zonelist *zonelist;
1118 struct zone *zone;
1119 struct zoneref *z;
1120 gfp_t gfp_mask;
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001121
Michal Hocko79b63f12017-09-06 16:24:03 -07001122 gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE;
1123 zonelist = node_zonelist(nid, gfp_mask);
1124 for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), NULL) {
1125 spin_lock_irqsave(&zone->lock, flags);
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001126
Michal Hocko79b63f12017-09-06 16:24:03 -07001127 pfn = ALIGN(zone->zone_start_pfn, nr_pages);
1128 while (zone_spans_last_pfn(zone, pfn, nr_pages)) {
1129 if (pfn_range_valid_gigantic(zone, pfn, nr_pages)) {
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001130 /*
1131 * We release the zone lock here because
1132 * alloc_contig_range() will also lock the zone
1133 * at some point. If there's an allocation
1134 * spinning on this lock, it may win the race
1135 * and cause alloc_contig_range() to fail...
1136 */
Michal Hocko79b63f12017-09-06 16:24:03 -07001137 spin_unlock_irqrestore(&zone->lock, flags);
1138 ret = __alloc_gigantic_page(pfn, nr_pages, gfp_mask);
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001139 if (!ret)
1140 return pfn_to_page(pfn);
Michal Hocko79b63f12017-09-06 16:24:03 -07001141 spin_lock_irqsave(&zone->lock, flags);
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001142 }
1143 pfn += nr_pages;
1144 }
1145
Michal Hocko79b63f12017-09-06 16:24:03 -07001146 spin_unlock_irqrestore(&zone->lock, flags);
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001147 }
1148
1149 return NULL;
1150}
1151
1152static void prep_new_huge_page(struct hstate *h, struct page *page, int nid);
Kirill A. Shutemovd00181b92015-11-06 16:29:57 -08001153static void prep_compound_gigantic_page(struct page *page, unsigned int order);
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001154
1155static struct page *alloc_fresh_gigantic_page_node(struct hstate *h, int nid)
1156{
1157 struct page *page;
1158
Michal Hocko79b63f12017-09-06 16:24:03 -07001159 page = alloc_gigantic_page(nid, h);
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001160 if (page) {
1161 prep_compound_gigantic_page(page, huge_page_order(h));
1162 prep_new_huge_page(h, page, nid);
1163 }
1164
1165 return page;
1166}
1167
1168static int alloc_fresh_gigantic_page(struct hstate *h,
1169 nodemask_t *nodes_allowed)
1170{
1171 struct page *page = NULL;
1172 int nr_nodes, node;
1173
1174 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
1175 page = alloc_fresh_gigantic_page_node(h, node);
1176 if (page)
1177 return 1;
1178 }
1179
1180 return 0;
1181}
1182
Aneesh Kumar K.Ve1073d12017-07-06 15:39:17 -07001183#else /* !CONFIG_ARCH_HAS_GIGANTIC_PAGE */
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001184static inline bool gigantic_page_supported(void) { return false; }
Kirill A. Shutemovd00181b92015-11-06 16:29:57 -08001185static inline void free_gigantic_page(struct page *page, unsigned int order) { }
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001186static inline void destroy_compound_gigantic_page(struct page *page,
Kirill A. Shutemovd00181b92015-11-06 16:29:57 -08001187 unsigned int order) { }
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001188static inline int alloc_fresh_gigantic_page(struct hstate *h,
1189 nodemask_t *nodes_allowed) { return 0; }
1190#endif
1191
Andi Kleena5516432008-07-23 21:27:41 -07001192static void update_and_free_page(struct hstate *h, struct page *page)
Adam Litke6af2acb2007-10-16 01:26:16 -07001193{
1194 int i;
Andi Kleena5516432008-07-23 21:27:41 -07001195
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001196 if (hstate_is_gigantic(h) && !gigantic_page_supported())
1197 return;
Andy Whitcroft18229df2008-11-06 12:53:27 -08001198
Andi Kleena5516432008-07-23 21:27:41 -07001199 h->nr_huge_pages--;
1200 h->nr_huge_pages_node[page_to_nid(page)]--;
1201 for (i = 0; i < pages_per_huge_page(h); i++) {
Chris Forbes32f84522011-07-25 17:12:14 -07001202 page[i].flags &= ~(1 << PG_locked | 1 << PG_error |
1203 1 << PG_referenced | 1 << PG_dirty |
Luiz Capitulinoa7407a22014-06-04 16:07:09 -07001204 1 << PG_active | 1 << PG_private |
1205 1 << PG_writeback);
Adam Litke6af2acb2007-10-16 01:26:16 -07001206 }
Sasha Levin309381fea2014-01-23 15:52:54 -08001207 VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page);
Kirill A. Shutemovf1e61552015-11-06 16:29:50 -08001208 set_compound_page_dtor(page, NULL_COMPOUND_DTOR);
Adam Litke6af2acb2007-10-16 01:26:16 -07001209 set_page_refcounted(page);
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001210 if (hstate_is_gigantic(h)) {
1211 destroy_compound_gigantic_page(page, huge_page_order(h));
1212 free_gigantic_page(page, huge_page_order(h));
1213 } else {
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001214 __free_pages(page, huge_page_order(h));
1215 }
Adam Litke6af2acb2007-10-16 01:26:16 -07001216}
1217
Andi Kleene5ff2152008-07-23 21:27:42 -07001218struct hstate *size_to_hstate(unsigned long size)
1219{
1220 struct hstate *h;
1221
1222 for_each_hstate(h) {
1223 if (huge_page_size(h) == size)
1224 return h;
1225 }
1226 return NULL;
1227}
1228
Naoya Horiguchibcc54222015-04-15 16:14:38 -07001229/*
1230 * Test to determine whether the hugepage is "active/in-use" (i.e. being linked
1231 * to hstate->hugepage_activelist.)
1232 *
1233 * This function can be called for tail pages, but never returns true for them.
1234 */
1235bool page_huge_active(struct page *page)
1236{
1237 VM_BUG_ON_PAGE(!PageHuge(page), page);
1238 return PageHead(page) && PagePrivate(&page[1]);
1239}
1240
1241/* never called for tail page */
1242static void set_page_huge_active(struct page *page)
1243{
1244 VM_BUG_ON_PAGE(!PageHeadHuge(page), page);
1245 SetPagePrivate(&page[1]);
1246}
1247
1248static void clear_page_huge_active(struct page *page)
1249{
1250 VM_BUG_ON_PAGE(!PageHeadHuge(page), page);
1251 ClearPagePrivate(&page[1]);
1252}
1253
Atsushi Kumagai8f1d26d2014-07-30 16:08:39 -07001254void free_huge_page(struct page *page)
David Gibson27a85ef2006-03-22 00:08:56 -08001255{
Andi Kleena5516432008-07-23 21:27:41 -07001256 /*
1257 * Can't pass hstate in here because it is called from the
1258 * compound page destructor.
1259 */
Andi Kleene5ff2152008-07-23 21:27:42 -07001260 struct hstate *h = page_hstate(page);
Adam Litke7893d1d2007-10-16 01:26:18 -07001261 int nid = page_to_nid(page);
David Gibson90481622012-03-21 16:34:12 -07001262 struct hugepage_subpool *spool =
1263 (struct hugepage_subpool *)page_private(page);
Joonsoo Kim07443a82013-09-11 14:21:58 -07001264 bool restore_reserve;
David Gibson27a85ef2006-03-22 00:08:56 -08001265
Andy Whitcrofte5df70a2008-02-23 15:23:32 -08001266 set_page_private(page, 0);
Mel Gorman23be7462010-04-23 13:17:56 -04001267 page->mapping = NULL;
Mike Kravetzb4330af2016-02-05 15:36:38 -08001268 VM_BUG_ON_PAGE(page_count(page), page);
1269 VM_BUG_ON_PAGE(page_mapcount(page), page);
Joonsoo Kim07443a82013-09-11 14:21:58 -07001270 restore_reserve = PagePrivate(page);
Joonsoo Kim16c794b2013-10-16 13:46:48 -07001271 ClearPagePrivate(page);
David Gibson27a85ef2006-03-22 00:08:56 -08001272
Mike Kravetz1c5ecae2015-04-15 16:13:39 -07001273 /*
1274 * A return code of zero implies that the subpool will be under its
1275 * minimum size if the reservation is not restored after page is free.
1276 * Therefore, force restore_reserve operation.
1277 */
1278 if (hugepage_subpool_put_pages(spool, 1) == 0)
1279 restore_reserve = true;
1280
David Gibson27a85ef2006-03-22 00:08:56 -08001281 spin_lock(&hugetlb_lock);
Naoya Horiguchibcc54222015-04-15 16:14:38 -07001282 clear_page_huge_active(page);
Aneesh Kumar K.V6d76dcf2012-07-31 16:42:18 -07001283 hugetlb_cgroup_uncharge_page(hstate_index(h),
1284 pages_per_huge_page(h), page);
Joonsoo Kim07443a82013-09-11 14:21:58 -07001285 if (restore_reserve)
1286 h->resv_huge_pages++;
1287
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001288 if (h->surplus_huge_pages_node[nid]) {
Aneesh Kumar K.V0edaecf2012-07-31 16:42:07 -07001289 /* remove the page from active list */
1290 list_del(&page->lru);
Andi Kleena5516432008-07-23 21:27:41 -07001291 update_and_free_page(h, page);
1292 h->surplus_huge_pages--;
1293 h->surplus_huge_pages_node[nid]--;
Adam Litke7893d1d2007-10-16 01:26:18 -07001294 } else {
Will Deacon5d3a5512012-10-08 16:29:32 -07001295 arch_clear_hugepage_flags(page);
Andi Kleena5516432008-07-23 21:27:41 -07001296 enqueue_huge_page(h, page);
Adam Litke7893d1d2007-10-16 01:26:18 -07001297 }
David Gibson27a85ef2006-03-22 00:08:56 -08001298 spin_unlock(&hugetlb_lock);
1299}
1300
Andi Kleena5516432008-07-23 21:27:41 -07001301static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
Andi Kleenb7ba30c2008-07-23 21:27:40 -07001302{
Aneesh Kumar K.V0edaecf2012-07-31 16:42:07 -07001303 INIT_LIST_HEAD(&page->lru);
Kirill A. Shutemovf1e61552015-11-06 16:29:50 -08001304 set_compound_page_dtor(page, HUGETLB_PAGE_DTOR);
Andi Kleenb7ba30c2008-07-23 21:27:40 -07001305 spin_lock(&hugetlb_lock);
Aneesh Kumar K.V9dd540e2012-07-31 16:42:15 -07001306 set_hugetlb_cgroup(page, NULL);
Andi Kleena5516432008-07-23 21:27:41 -07001307 h->nr_huge_pages++;
1308 h->nr_huge_pages_node[nid]++;
Andi Kleenb7ba30c2008-07-23 21:27:40 -07001309 spin_unlock(&hugetlb_lock);
1310 put_page(page); /* free it into the hugepage allocator */
1311}
1312
Kirill A. Shutemovd00181b92015-11-06 16:29:57 -08001313static void prep_compound_gigantic_page(struct page *page, unsigned int order)
Wu Fengguang20a03072009-06-16 15:32:22 -07001314{
1315 int i;
1316 int nr_pages = 1 << order;
1317 struct page *p = page + 1;
1318
1319 /* we rely on prep_new_huge_page to set the destructor */
1320 set_compound_order(page, order);
Andrea Arcangelief5a22b2013-10-16 13:46:56 -07001321 __ClearPageReserved(page);
Kirill A. Shutemovde09d312016-01-15 16:51:42 -08001322 __SetPageHead(page);
Wu Fengguang20a03072009-06-16 15:32:22 -07001323 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
Andrea Arcangelief5a22b2013-10-16 13:46:56 -07001324 /*
1325 * For gigantic hugepages allocated through bootmem at
1326 * boot, it's safer to be consistent with the not-gigantic
1327 * hugepages and clear the PG_reserved bit from all tail pages
1328 * too. Otherwse drivers using get_user_pages() to access tail
1329 * pages may get the reference counting wrong if they see
1330 * PG_reserved set on a tail page (despite the head page not
1331 * having PG_reserved set). Enforcing this consistency between
1332 * head and tail pages allows drivers to optimize away a check
1333 * on the head page when they need know if put_page() is needed
1334 * after get_user_pages().
1335 */
1336 __ClearPageReserved(p);
Youquan Song58a84aa2011-12-08 14:34:18 -08001337 set_page_count(p, 0);
Kirill A. Shutemov1d798ca2015-11-06 16:29:54 -08001338 set_compound_head(p, page);
Wu Fengguang20a03072009-06-16 15:32:22 -07001339 }
Mike Kravetzb4330af2016-02-05 15:36:38 -08001340 atomic_set(compound_mapcount_ptr(page), -1);
Wu Fengguang20a03072009-06-16 15:32:22 -07001341}
1342
Andrew Morton77959122012-10-08 16:34:11 -07001343/*
1344 * PageHuge() only returns true for hugetlbfs pages, but not for normal or
1345 * transparent huge pages. See the PageTransHuge() documentation for more
1346 * details.
1347 */
Wu Fengguang20a03072009-06-16 15:32:22 -07001348int PageHuge(struct page *page)
1349{
Wu Fengguang20a03072009-06-16 15:32:22 -07001350 if (!PageCompound(page))
1351 return 0;
1352
1353 page = compound_head(page);
Kirill A. Shutemovf1e61552015-11-06 16:29:50 -08001354 return page[1].compound_dtor == HUGETLB_PAGE_DTOR;
Wu Fengguang20a03072009-06-16 15:32:22 -07001355}
Naoya Horiguchi43131e12010-05-28 09:29:22 +09001356EXPORT_SYMBOL_GPL(PageHuge);
1357
Andrea Arcangeli27c73ae2013-11-21 14:32:02 -08001358/*
1359 * PageHeadHuge() only returns true for hugetlbfs head page, but not for
1360 * normal or transparent huge pages.
1361 */
1362int PageHeadHuge(struct page *page_head)
1363{
Andrea Arcangeli27c73ae2013-11-21 14:32:02 -08001364 if (!PageHead(page_head))
1365 return 0;
1366
Andrew Morton758f66a2014-01-21 15:48:57 -08001367 return get_compound_page_dtor(page_head) == free_huge_page;
Andrea Arcangeli27c73ae2013-11-21 14:32:02 -08001368}
Andrea Arcangeli27c73ae2013-11-21 14:32:02 -08001369
Zhang Yi13d60f42013-06-25 21:19:31 +08001370pgoff_t __basepage_index(struct page *page)
1371{
1372 struct page *page_head = compound_head(page);
1373 pgoff_t index = page_index(page_head);
1374 unsigned long compound_idx;
1375
1376 if (!PageHuge(page_head))
1377 return page_index(page);
1378
1379 if (compound_order(page_head) >= MAX_ORDER)
1380 compound_idx = page_to_pfn(page) - page_to_pfn(page_head);
1381 else
1382 compound_idx = page - page_head;
1383
1384 return (index << compound_order(page_head)) + compound_idx;
1385}
1386
Andi Kleena5516432008-07-23 21:27:41 -07001387static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001388{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001389 struct page *page;
Joe Jinf96efd52007-07-15 23:38:12 -07001390
Vlastimil Babka96db8002015-09-08 15:03:50 -07001391 page = __alloc_pages_node(nid,
Naoya Horiguchi86cdb462013-09-11 14:22:13 -07001392 htlb_alloc_mask(h)|__GFP_COMP|__GFP_THISNODE|
Michal Hockodcda9b02017-07-12 14:36:45 -07001393 __GFP_RETRY_MAYFAIL|__GFP_NOWARN,
Andi Kleena5516432008-07-23 21:27:41 -07001394 huge_page_order(h));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001395 if (page) {
Andi Kleena5516432008-07-23 21:27:41 -07001396 prep_new_huge_page(h, page, nid);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001397 }
Nishanth Aravamudan63b46132007-10-16 01:26:24 -07001398
1399 return page;
1400}
1401
Joonsoo Kimb2261022013-09-11 14:21:00 -07001402static int alloc_fresh_huge_page(struct hstate *h, nodemask_t *nodes_allowed)
1403{
1404 struct page *page;
1405 int nr_nodes, node;
1406 int ret = 0;
1407
1408 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
1409 page = alloc_fresh_huge_page_node(h, node);
1410 if (page) {
1411 ret = 1;
1412 break;
1413 }
1414 }
1415
1416 if (ret)
1417 count_vm_event(HTLB_BUDDY_PGALLOC);
1418 else
1419 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
1420
1421 return ret;
1422}
1423
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001424/*
1425 * Free huge page from pool from next node to free.
1426 * Attempt to keep persistent huge pages more or less
1427 * balanced over allowed nodes.
1428 * Called with hugetlb_lock locked.
1429 */
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001430static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
1431 bool acct_surplus)
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001432{
Joonsoo Kimb2261022013-09-11 14:21:00 -07001433 int nr_nodes, node;
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001434 int ret = 0;
1435
Joonsoo Kimb2261022013-09-11 14:21:00 -07001436 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
Lee Schermerhorn685f3452009-09-21 17:01:23 -07001437 /*
1438 * If we're returning unused surplus pages, only examine
1439 * nodes with surplus pages.
1440 */
Joonsoo Kimb2261022013-09-11 14:21:00 -07001441 if ((!acct_surplus || h->surplus_huge_pages_node[node]) &&
1442 !list_empty(&h->hugepage_freelists[node])) {
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001443 struct page *page =
Joonsoo Kimb2261022013-09-11 14:21:00 -07001444 list_entry(h->hugepage_freelists[node].next,
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001445 struct page, lru);
1446 list_del(&page->lru);
1447 h->free_huge_pages--;
Joonsoo Kimb2261022013-09-11 14:21:00 -07001448 h->free_huge_pages_node[node]--;
Lee Schermerhorn685f3452009-09-21 17:01:23 -07001449 if (acct_surplus) {
1450 h->surplus_huge_pages--;
Joonsoo Kimb2261022013-09-11 14:21:00 -07001451 h->surplus_huge_pages_node[node]--;
Lee Schermerhorn685f3452009-09-21 17:01:23 -07001452 }
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001453 update_and_free_page(h, page);
1454 ret = 1;
Lee Schermerhorn9a76db02009-12-14 17:58:15 -08001455 break;
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001456 }
Joonsoo Kimb2261022013-09-11 14:21:00 -07001457 }
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001458
1459 return ret;
1460}
1461
Naoya Horiguchic8721bb2013-09-11 14:22:09 -07001462/*
1463 * Dissolve a given free hugepage into free buddy pages. This function does
Gerald Schaefer082d5b62016-10-07 17:01:10 -07001464 * nothing for in-use (including surplus) hugepages. Returns -EBUSY if the
1465 * number of free hugepages would be reduced below the number of reserved
1466 * hugepages.
Naoya Horiguchic8721bb2013-09-11 14:22:09 -07001467 */
Anshuman Khandualc3114a82017-07-10 15:47:41 -07001468int dissolve_free_huge_page(struct page *page)
Naoya Horiguchic8721bb2013-09-11 14:22:09 -07001469{
Gerald Schaefer082d5b62016-10-07 17:01:10 -07001470 int rc = 0;
1471
Naoya Horiguchic8721bb2013-09-11 14:22:09 -07001472 spin_lock(&hugetlb_lock);
1473 if (PageHuge(page) && !page_count(page)) {
Gerald Schaefer2247bb32016-10-07 17:01:07 -07001474 struct page *head = compound_head(page);
1475 struct hstate *h = page_hstate(head);
1476 int nid = page_to_nid(head);
Gerald Schaefer082d5b62016-10-07 17:01:10 -07001477 if (h->free_huge_pages - h->resv_huge_pages == 0) {
1478 rc = -EBUSY;
1479 goto out;
1480 }
Anshuman Khandualc3114a82017-07-10 15:47:41 -07001481 /*
1482 * Move PageHWPoison flag from head page to the raw error page,
1483 * which makes any subpages rather than the error page reusable.
1484 */
1485 if (PageHWPoison(head) && page != head) {
1486 SetPageHWPoison(page);
1487 ClearPageHWPoison(head);
1488 }
Gerald Schaefer2247bb32016-10-07 17:01:07 -07001489 list_del(&head->lru);
Naoya Horiguchic8721bb2013-09-11 14:22:09 -07001490 h->free_huge_pages--;
1491 h->free_huge_pages_node[nid]--;
zhong jiangc1470b32016-08-11 15:32:55 -07001492 h->max_huge_pages--;
Gerald Schaefer2247bb32016-10-07 17:01:07 -07001493 update_and_free_page(h, head);
Naoya Horiguchic8721bb2013-09-11 14:22:09 -07001494 }
Gerald Schaefer082d5b62016-10-07 17:01:10 -07001495out:
Naoya Horiguchic8721bb2013-09-11 14:22:09 -07001496 spin_unlock(&hugetlb_lock);
Gerald Schaefer082d5b62016-10-07 17:01:10 -07001497 return rc;
Naoya Horiguchic8721bb2013-09-11 14:22:09 -07001498}
1499
1500/*
1501 * Dissolve free hugepages in a given pfn range. Used by memory hotplug to
1502 * make specified memory blocks removable from the system.
Gerald Schaefer2247bb32016-10-07 17:01:07 -07001503 * Note that this will dissolve a free gigantic hugepage completely, if any
1504 * part of it lies within the given range.
Gerald Schaefer082d5b62016-10-07 17:01:10 -07001505 * Also note that if dissolve_free_huge_page() returns with an error, all
1506 * free hugepages that were dissolved before that error are lost.
Naoya Horiguchic8721bb2013-09-11 14:22:09 -07001507 */
Gerald Schaefer082d5b62016-10-07 17:01:10 -07001508int dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn)
Naoya Horiguchic8721bb2013-09-11 14:22:09 -07001509{
Naoya Horiguchic8721bb2013-09-11 14:22:09 -07001510 unsigned long pfn;
Gerald Schaefereb03aa02016-10-07 17:01:13 -07001511 struct page *page;
Gerald Schaefer082d5b62016-10-07 17:01:10 -07001512 int rc = 0;
Naoya Horiguchic8721bb2013-09-11 14:22:09 -07001513
Li Zhongd0177632014-08-06 16:07:56 -07001514 if (!hugepages_supported())
Gerald Schaefer082d5b62016-10-07 17:01:10 -07001515 return rc;
Li Zhongd0177632014-08-06 16:07:56 -07001516
Gerald Schaefereb03aa02016-10-07 17:01:13 -07001517 for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << minimum_order) {
1518 page = pfn_to_page(pfn);
1519 if (PageHuge(page) && !page_count(page)) {
1520 rc = dissolve_free_huge_page(page);
1521 if (rc)
1522 break;
1523 }
1524 }
Gerald Schaefer082d5b62016-10-07 17:01:10 -07001525
1526 return rc;
Naoya Horiguchic8721bb2013-09-11 14:22:09 -07001527}
1528
Dave Hansen099730d2015-11-05 18:50:17 -08001529static struct page *__hugetlb_alloc_buddy_huge_page(struct hstate *h,
Michal Hockoaaf14e42017-07-10 15:49:08 -07001530 gfp_t gfp_mask, int nid, nodemask_t *nmask)
Dave Hansen099730d2015-11-05 18:50:17 -08001531{
1532 int order = huge_page_order(h);
Dave Hansen099730d2015-11-05 18:50:17 -08001533
Michal Hockodcda9b02017-07-12 14:36:45 -07001534 gfp_mask |= __GFP_COMP|__GFP_RETRY_MAYFAIL|__GFP_NOWARN;
Michal Hockoaaf14e42017-07-10 15:49:08 -07001535 if (nid == NUMA_NO_NODE)
1536 nid = numa_mem_id();
1537 return __alloc_pages_nodemask(gfp_mask, order, nid, nmask);
Dave Hansen099730d2015-11-05 18:50:17 -08001538}
1539
Michal Hockoaaf14e42017-07-10 15:49:08 -07001540static struct page *__alloc_buddy_huge_page(struct hstate *h, gfp_t gfp_mask,
1541 int nid, nodemask_t *nmask)
Adam Litke7893d1d2007-10-16 01:26:18 -07001542{
1543 struct page *page;
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001544 unsigned int r_nid;
Adam Litke7893d1d2007-10-16 01:26:18 -07001545
Luiz Capitulinobae7f4a2014-06-04 16:07:08 -07001546 if (hstate_is_gigantic(h))
Andi Kleenaa888a72008-07-23 21:27:47 -07001547 return NULL;
1548
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001549 /*
1550 * Assume we will successfully allocate the surplus page to
1551 * prevent racing processes from causing the surplus to exceed
1552 * overcommit
1553 *
1554 * This however introduces a different race, where a process B
1555 * tries to grow the static hugepage pool while alloc_pages() is
1556 * called by process A. B will only examine the per-node
1557 * counters in determining if surplus huge pages can be
1558 * converted to normal huge pages in adjust_pool_surplus(). A
1559 * won't be able to increment the per-node counter, until the
1560 * lock is dropped by B, but B doesn't drop hugetlb_lock until
1561 * no more huge pages can be converted from surplus to normal
1562 * state (and doesn't try to convert again). Thus, we have a
1563 * case where a surplus huge page exists, the pool is grown, and
1564 * the surplus huge page still exists after, even though it
1565 * should just have been converted to a normal huge page. This
1566 * does not leak memory, though, as the hugepage will be freed
1567 * once it is out of use. It also does not allow the counters to
1568 * go out of whack in adjust_pool_surplus() as we don't modify
1569 * the node values until we've gotten the hugepage and only the
1570 * per-node value is checked there.
1571 */
1572 spin_lock(&hugetlb_lock);
Andi Kleena5516432008-07-23 21:27:41 -07001573 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001574 spin_unlock(&hugetlb_lock);
1575 return NULL;
1576 } else {
Andi Kleena5516432008-07-23 21:27:41 -07001577 h->nr_huge_pages++;
1578 h->surplus_huge_pages++;
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001579 }
1580 spin_unlock(&hugetlb_lock);
1581
Michal Hockoaaf14e42017-07-10 15:49:08 -07001582 page = __hugetlb_alloc_buddy_huge_page(h, gfp_mask, nid, nmask);
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001583
1584 spin_lock(&hugetlb_lock);
Adam Litke7893d1d2007-10-16 01:26:18 -07001585 if (page) {
Aneesh Kumar K.V0edaecf2012-07-31 16:42:07 -07001586 INIT_LIST_HEAD(&page->lru);
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001587 r_nid = page_to_nid(page);
Kirill A. Shutemovf1e61552015-11-06 16:29:50 -08001588 set_compound_page_dtor(page, HUGETLB_PAGE_DTOR);
Aneesh Kumar K.V9dd540e2012-07-31 16:42:15 -07001589 set_hugetlb_cgroup(page, NULL);
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001590 /*
1591 * We incremented the global counters already
1592 */
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001593 h->nr_huge_pages_node[r_nid]++;
1594 h->surplus_huge_pages_node[r_nid]++;
Adam Litke3b116302008-04-28 02:13:06 -07001595 __count_vm_event(HTLB_BUDDY_PGALLOC);
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001596 } else {
Andi Kleena5516432008-07-23 21:27:41 -07001597 h->nr_huge_pages--;
1598 h->surplus_huge_pages--;
Adam Litke3b116302008-04-28 02:13:06 -07001599 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
Adam Litke7893d1d2007-10-16 01:26:18 -07001600 }
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001601 spin_unlock(&hugetlb_lock);
Adam Litke7893d1d2007-10-16 01:26:18 -07001602
1603 return page;
1604}
1605
Adam Litkee4e574b2007-10-16 01:26:19 -07001606/*
Dave Hansen099730d2015-11-05 18:50:17 -08001607 * Use the VMA's mpolicy to allocate a huge page from the buddy.
1608 */
Dave Hansene0ec90e2015-11-05 18:50:20 -08001609static
Dave Hansen099730d2015-11-05 18:50:17 -08001610struct page *__alloc_buddy_huge_page_with_mpol(struct hstate *h,
1611 struct vm_area_struct *vma, unsigned long addr)
1612{
Michal Hockoaaf14e42017-07-10 15:49:08 -07001613 struct page *page;
1614 struct mempolicy *mpol;
1615 gfp_t gfp_mask = htlb_alloc_mask(h);
1616 int nid;
1617 nodemask_t *nodemask;
1618
1619 nid = huge_node(vma, addr, gfp_mask, &mpol, &nodemask);
1620 page = __alloc_buddy_huge_page(h, gfp_mask, nid, nodemask);
1621 mpol_cond_put(mpol);
1622
1623 return page;
Dave Hansen099730d2015-11-05 18:50:17 -08001624}
1625
1626/*
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001627 * This allocation function is useful in the context where vma is irrelevant.
1628 * E.g. soft-offlining uses this function because it only cares physical
1629 * address of error page.
1630 */
1631struct page *alloc_huge_page_node(struct hstate *h, int nid)
1632{
Michal Hockoaaf14e42017-07-10 15:49:08 -07001633 gfp_t gfp_mask = htlb_alloc_mask(h);
Joonsoo Kim4ef91842013-09-11 14:21:51 -07001634 struct page *page = NULL;
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001635
Michal Hockoaaf14e42017-07-10 15:49:08 -07001636 if (nid != NUMA_NO_NODE)
1637 gfp_mask |= __GFP_THISNODE;
1638
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001639 spin_lock(&hugetlb_lock);
Joonsoo Kim4ef91842013-09-11 14:21:51 -07001640 if (h->free_huge_pages - h->resv_huge_pages > 0)
Michal Hocko3e59fcb2017-07-10 15:49:11 -07001641 page = dequeue_huge_page_nodemask(h, gfp_mask, nid, NULL);
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001642 spin_unlock(&hugetlb_lock);
1643
Aneesh Kumar K.V94ae8ba2012-07-31 16:42:35 -07001644 if (!page)
Michal Hockoaaf14e42017-07-10 15:49:08 -07001645 page = __alloc_buddy_huge_page(h, gfp_mask, nid, NULL);
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001646
1647 return page;
1648}
1649
Michal Hocko3e59fcb2017-07-10 15:49:11 -07001650
1651struct page *alloc_huge_page_nodemask(struct hstate *h, int preferred_nid,
1652 nodemask_t *nmask)
Michal Hocko4db9b2e2017-07-10 15:48:44 -07001653{
Michal Hockoaaf14e42017-07-10 15:49:08 -07001654 gfp_t gfp_mask = htlb_alloc_mask(h);
Michal Hocko4db9b2e2017-07-10 15:48:44 -07001655
1656 spin_lock(&hugetlb_lock);
1657 if (h->free_huge_pages - h->resv_huge_pages > 0) {
Michal Hocko3e59fcb2017-07-10 15:49:11 -07001658 struct page *page;
1659
1660 page = dequeue_huge_page_nodemask(h, gfp_mask, preferred_nid, nmask);
1661 if (page) {
1662 spin_unlock(&hugetlb_lock);
1663 return page;
Michal Hocko4db9b2e2017-07-10 15:48:44 -07001664 }
1665 }
1666 spin_unlock(&hugetlb_lock);
Michal Hocko4db9b2e2017-07-10 15:48:44 -07001667
1668 /* No reservations, try to overcommit */
Michal Hocko3e59fcb2017-07-10 15:49:11 -07001669
1670 return __alloc_buddy_huge_page(h, gfp_mask, preferred_nid, nmask);
Michal Hocko4db9b2e2017-07-10 15:48:44 -07001671}
1672
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001673/*
Lucas De Marchi25985ed2011-03-30 22:57:33 -03001674 * Increase the hugetlb pool such that it can accommodate a reservation
Adam Litkee4e574b2007-10-16 01:26:19 -07001675 * of size 'delta'.
1676 */
Andi Kleena5516432008-07-23 21:27:41 -07001677static int gather_surplus_pages(struct hstate *h, int delta)
Adam Litkee4e574b2007-10-16 01:26:19 -07001678{
1679 struct list_head surplus_list;
1680 struct page *page, *tmp;
1681 int ret, i;
1682 int needed, allocated;
Hillf Danton28073b02012-03-21 16:34:00 -07001683 bool alloc_ok = true;
Adam Litkee4e574b2007-10-16 01:26:19 -07001684
Andi Kleena5516432008-07-23 21:27:41 -07001685 needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
Adam Litkeac09b3a2008-03-04 14:29:38 -08001686 if (needed <= 0) {
Andi Kleena5516432008-07-23 21:27:41 -07001687 h->resv_huge_pages += delta;
Adam Litkee4e574b2007-10-16 01:26:19 -07001688 return 0;
Adam Litkeac09b3a2008-03-04 14:29:38 -08001689 }
Adam Litkee4e574b2007-10-16 01:26:19 -07001690
1691 allocated = 0;
1692 INIT_LIST_HEAD(&surplus_list);
1693
1694 ret = -ENOMEM;
1695retry:
1696 spin_unlock(&hugetlb_lock);
1697 for (i = 0; i < needed; i++) {
Michal Hockoaaf14e42017-07-10 15:49:08 -07001698 page = __alloc_buddy_huge_page(h, htlb_alloc_mask(h),
1699 NUMA_NO_NODE, NULL);
Hillf Danton28073b02012-03-21 16:34:00 -07001700 if (!page) {
1701 alloc_ok = false;
1702 break;
1703 }
Adam Litkee4e574b2007-10-16 01:26:19 -07001704 list_add(&page->lru, &surplus_list);
David Rientjes69ed7792017-07-10 15:48:50 -07001705 cond_resched();
Adam Litkee4e574b2007-10-16 01:26:19 -07001706 }
Hillf Danton28073b02012-03-21 16:34:00 -07001707 allocated += i;
Adam Litkee4e574b2007-10-16 01:26:19 -07001708
1709 /*
1710 * After retaking hugetlb_lock, we need to recalculate 'needed'
1711 * because either resv_huge_pages or free_huge_pages may have changed.
1712 */
1713 spin_lock(&hugetlb_lock);
Andi Kleena5516432008-07-23 21:27:41 -07001714 needed = (h->resv_huge_pages + delta) -
1715 (h->free_huge_pages + allocated);
Hillf Danton28073b02012-03-21 16:34:00 -07001716 if (needed > 0) {
1717 if (alloc_ok)
1718 goto retry;
1719 /*
1720 * We were not able to allocate enough pages to
1721 * satisfy the entire reservation so we free what
1722 * we've allocated so far.
1723 */
1724 goto free;
1725 }
Adam Litkee4e574b2007-10-16 01:26:19 -07001726 /*
1727 * The surplus_list now contains _at_least_ the number of extra pages
Lucas De Marchi25985ed2011-03-30 22:57:33 -03001728 * needed to accommodate the reservation. Add the appropriate number
Adam Litkee4e574b2007-10-16 01:26:19 -07001729 * of pages to the hugetlb pool and free the extras back to the buddy
Adam Litkeac09b3a2008-03-04 14:29:38 -08001730 * allocator. Commit the entire reservation here to prevent another
1731 * process from stealing the pages as they are added to the pool but
1732 * before they are reserved.
Adam Litkee4e574b2007-10-16 01:26:19 -07001733 */
1734 needed += allocated;
Andi Kleena5516432008-07-23 21:27:41 -07001735 h->resv_huge_pages += delta;
Adam Litkee4e574b2007-10-16 01:26:19 -07001736 ret = 0;
Naoya Horiguchia9869b82010-09-08 10:19:37 +09001737
Adam Litke19fc3f02008-04-28 02:12:20 -07001738 /* Free the needed pages to the hugetlb pool */
Adam Litkee4e574b2007-10-16 01:26:19 -07001739 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
Adam Litke19fc3f02008-04-28 02:12:20 -07001740 if ((--needed) < 0)
1741 break;
Naoya Horiguchia9869b82010-09-08 10:19:37 +09001742 /*
1743 * This page is now managed by the hugetlb allocator and has
1744 * no users -- drop the buddy allocator's reference.
1745 */
1746 put_page_testzero(page);
Sasha Levin309381fea2014-01-23 15:52:54 -08001747 VM_BUG_ON_PAGE(page_count(page), page);
Andi Kleena5516432008-07-23 21:27:41 -07001748 enqueue_huge_page(h, page);
Adam Litke19fc3f02008-04-28 02:12:20 -07001749 }
Hillf Danton28073b02012-03-21 16:34:00 -07001750free:
Hillf Dantonb0365c82011-12-28 15:57:16 -08001751 spin_unlock(&hugetlb_lock);
Adam Litke19fc3f02008-04-28 02:12:20 -07001752
1753 /* Free unnecessary surplus pages to the buddy allocator */
Joonsoo Kimc0d934b2013-09-11 14:21:02 -07001754 list_for_each_entry_safe(page, tmp, &surplus_list, lru)
1755 put_page(page);
Naoya Horiguchia9869b82010-09-08 10:19:37 +09001756 spin_lock(&hugetlb_lock);
Adam Litkee4e574b2007-10-16 01:26:19 -07001757
1758 return ret;
1759}
1760
1761/*
Mike Kravetze5bbc8a2017-01-10 16:58:27 -08001762 * This routine has two main purposes:
1763 * 1) Decrement the reservation count (resv_huge_pages) by the value passed
1764 * in unused_resv_pages. This corresponds to the prior adjustments made
1765 * to the associated reservation map.
1766 * 2) Free any unused surplus pages that may have been allocated to satisfy
1767 * the reservation. As many as unused_resv_pages may be freed.
1768 *
1769 * Called with hugetlb_lock held. However, the lock could be dropped (and
1770 * reacquired) during calls to cond_resched_lock. Whenever dropping the lock,
1771 * we must make sure nobody else can claim pages we are in the process of
1772 * freeing. Do this by ensuring resv_huge_page always is greater than the
1773 * number of huge pages we plan to free when dropping the lock.
Adam Litkee4e574b2007-10-16 01:26:19 -07001774 */
Andi Kleena5516432008-07-23 21:27:41 -07001775static void return_unused_surplus_pages(struct hstate *h,
1776 unsigned long unused_resv_pages)
Adam Litkee4e574b2007-10-16 01:26:19 -07001777{
Adam Litkee4e574b2007-10-16 01:26:19 -07001778 unsigned long nr_pages;
1779
Andi Kleenaa888a72008-07-23 21:27:47 -07001780 /* Cannot return gigantic pages currently */
Luiz Capitulinobae7f4a2014-06-04 16:07:08 -07001781 if (hstate_is_gigantic(h))
Mike Kravetze5bbc8a2017-01-10 16:58:27 -08001782 goto out;
Andi Kleenaa888a72008-07-23 21:27:47 -07001783
Mike Kravetze5bbc8a2017-01-10 16:58:27 -08001784 /*
1785 * Part (or even all) of the reservation could have been backed
1786 * by pre-allocated pages. Only free surplus pages.
1787 */
Andi Kleena5516432008-07-23 21:27:41 -07001788 nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
Adam Litkee4e574b2007-10-16 01:26:19 -07001789
Lee Schermerhorn685f3452009-09-21 17:01:23 -07001790 /*
1791 * We want to release as many surplus pages as possible, spread
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -08001792 * evenly across all nodes with memory. Iterate across these nodes
1793 * until we can no longer free unreserved surplus pages. This occurs
1794 * when the nodes with surplus pages have no free pages.
1795 * free_pool_huge_page() will balance the the freed pages across the
1796 * on-line nodes with memory and will handle the hstate accounting.
Mike Kravetze5bbc8a2017-01-10 16:58:27 -08001797 *
1798 * Note that we decrement resv_huge_pages as we free the pages. If
1799 * we drop the lock, resv_huge_pages will still be sufficiently large
1800 * to cover subsequent pages we may free.
Lee Schermerhorn685f3452009-09-21 17:01:23 -07001801 */
1802 while (nr_pages--) {
Mike Kravetze5bbc8a2017-01-10 16:58:27 -08001803 h->resv_huge_pages--;
1804 unused_resv_pages--;
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08001805 if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1))
Mike Kravetze5bbc8a2017-01-10 16:58:27 -08001806 goto out;
Mizuma, Masayoshi7848a4b2014-04-18 15:07:18 -07001807 cond_resched_lock(&hugetlb_lock);
Adam Litkee4e574b2007-10-16 01:26:19 -07001808 }
Mike Kravetze5bbc8a2017-01-10 16:58:27 -08001809
1810out:
1811 /* Fully uncommit the reservation */
1812 h->resv_huge_pages -= unused_resv_pages;
Adam Litkee4e574b2007-10-16 01:26:19 -07001813}
1814
Mike Kravetz5e911372015-09-08 15:01:28 -07001815
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001816/*
Mike Kravetzfeba16e2015-09-08 15:01:31 -07001817 * vma_needs_reservation, vma_commit_reservation and vma_end_reservation
Mike Kravetz5e911372015-09-08 15:01:28 -07001818 * are used by the huge page allocation routines to manage reservations.
Mike Kravetzcf3ad202015-06-24 16:57:55 -07001819 *
1820 * vma_needs_reservation is called to determine if the huge page at addr
1821 * within the vma has an associated reservation. If a reservation is
1822 * needed, the value 1 is returned. The caller is then responsible for
1823 * managing the global reservation and subpool usage counts. After
1824 * the huge page has been allocated, vma_commit_reservation is called
Mike Kravetzfeba16e2015-09-08 15:01:31 -07001825 * to add the page to the reservation map. If the page allocation fails,
1826 * the reservation must be ended instead of committed. vma_end_reservation
1827 * is called in such cases.
Mike Kravetzcf3ad202015-06-24 16:57:55 -07001828 *
1829 * In the normal case, vma_commit_reservation returns the same value
1830 * as the preceding vma_needs_reservation call. The only time this
1831 * is not the case is if a reserve map was changed between calls. It
1832 * is the responsibility of the caller to notice the difference and
1833 * take appropriate action.
Mike Kravetz96b96a92016-11-10 10:46:32 -08001834 *
1835 * vma_add_reservation is used in error paths where a reservation must
1836 * be restored when a newly allocated huge page must be freed. It is
1837 * to be called after calling vma_needs_reservation to determine if a
1838 * reservation exists.
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001839 */
Mike Kravetz5e911372015-09-08 15:01:28 -07001840enum vma_resv_mode {
1841 VMA_NEEDS_RESV,
1842 VMA_COMMIT_RESV,
Mike Kravetzfeba16e2015-09-08 15:01:31 -07001843 VMA_END_RESV,
Mike Kravetz96b96a92016-11-10 10:46:32 -08001844 VMA_ADD_RESV,
Mike Kravetz5e911372015-09-08 15:01:28 -07001845};
Mike Kravetzcf3ad202015-06-24 16:57:55 -07001846static long __vma_reservation_common(struct hstate *h,
1847 struct vm_area_struct *vma, unsigned long addr,
Mike Kravetz5e911372015-09-08 15:01:28 -07001848 enum vma_resv_mode mode)
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001849{
Joonsoo Kim4e35f482014-04-03 14:47:30 -07001850 struct resv_map *resv;
1851 pgoff_t idx;
Mike Kravetzcf3ad202015-06-24 16:57:55 -07001852 long ret;
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001853
Joonsoo Kim4e35f482014-04-03 14:47:30 -07001854 resv = vma_resv_map(vma);
1855 if (!resv)
Andy Whitcroft84afd992008-07-23 21:27:32 -07001856 return 1;
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001857
Joonsoo Kim4e35f482014-04-03 14:47:30 -07001858 idx = vma_hugecache_offset(h, vma, addr);
Mike Kravetz5e911372015-09-08 15:01:28 -07001859 switch (mode) {
1860 case VMA_NEEDS_RESV:
Mike Kravetzcf3ad202015-06-24 16:57:55 -07001861 ret = region_chg(resv, idx, idx + 1);
Mike Kravetz5e911372015-09-08 15:01:28 -07001862 break;
1863 case VMA_COMMIT_RESV:
1864 ret = region_add(resv, idx, idx + 1);
1865 break;
Mike Kravetzfeba16e2015-09-08 15:01:31 -07001866 case VMA_END_RESV:
Mike Kravetz5e911372015-09-08 15:01:28 -07001867 region_abort(resv, idx, idx + 1);
1868 ret = 0;
1869 break;
Mike Kravetz96b96a92016-11-10 10:46:32 -08001870 case VMA_ADD_RESV:
1871 if (vma->vm_flags & VM_MAYSHARE)
1872 ret = region_add(resv, idx, idx + 1);
1873 else {
1874 region_abort(resv, idx, idx + 1);
1875 ret = region_del(resv, idx, idx + 1);
1876 }
1877 break;
Mike Kravetz5e911372015-09-08 15:01:28 -07001878 default:
1879 BUG();
1880 }
Andy Whitcroft84afd992008-07-23 21:27:32 -07001881
Joonsoo Kim4e35f482014-04-03 14:47:30 -07001882 if (vma->vm_flags & VM_MAYSHARE)
Mike Kravetzcf3ad202015-06-24 16:57:55 -07001883 return ret;
Mike Kravetz67961f92016-06-08 15:33:42 -07001884 else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && ret >= 0) {
1885 /*
1886 * In most cases, reserves always exist for private mappings.
1887 * However, a file associated with mapping could have been
1888 * hole punched or truncated after reserves were consumed.
1889 * As subsequent fault on such a range will not use reserves.
1890 * Subtle - The reserve map for private mappings has the
1891 * opposite meaning than that of shared mappings. If NO
1892 * entry is in the reserve map, it means a reservation exists.
1893 * If an entry exists in the reserve map, it means the
1894 * reservation has already been consumed. As a result, the
1895 * return value of this routine is the opposite of the
1896 * value returned from reserve map manipulation routines above.
1897 */
1898 if (ret)
1899 return 0;
1900 else
1901 return 1;
1902 }
Joonsoo Kim4e35f482014-04-03 14:47:30 -07001903 else
Mike Kravetzcf3ad202015-06-24 16:57:55 -07001904 return ret < 0 ? ret : 0;
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001905}
Mike Kravetzcf3ad202015-06-24 16:57:55 -07001906
1907static long vma_needs_reservation(struct hstate *h,
Andi Kleena5516432008-07-23 21:27:41 -07001908 struct vm_area_struct *vma, unsigned long addr)
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001909{
Mike Kravetz5e911372015-09-08 15:01:28 -07001910 return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV);
Mike Kravetzcf3ad202015-06-24 16:57:55 -07001911}
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001912
Mike Kravetzcf3ad202015-06-24 16:57:55 -07001913static long vma_commit_reservation(struct hstate *h,
1914 struct vm_area_struct *vma, unsigned long addr)
1915{
Mike Kravetz5e911372015-09-08 15:01:28 -07001916 return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV);
1917}
1918
Mike Kravetzfeba16e2015-09-08 15:01:31 -07001919static void vma_end_reservation(struct hstate *h,
Mike Kravetz5e911372015-09-08 15:01:28 -07001920 struct vm_area_struct *vma, unsigned long addr)
1921{
Mike Kravetzfeba16e2015-09-08 15:01:31 -07001922 (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV);
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001923}
1924
Mike Kravetz96b96a92016-11-10 10:46:32 -08001925static long vma_add_reservation(struct hstate *h,
1926 struct vm_area_struct *vma, unsigned long addr)
1927{
1928 return __vma_reservation_common(h, vma, addr, VMA_ADD_RESV);
1929}
1930
1931/*
1932 * This routine is called to restore a reservation on error paths. In the
1933 * specific error paths, a huge page was allocated (via alloc_huge_page)
1934 * and is about to be freed. If a reservation for the page existed,
1935 * alloc_huge_page would have consumed the reservation and set PagePrivate
1936 * in the newly allocated page. When the page is freed via free_huge_page,
1937 * the global reservation count will be incremented if PagePrivate is set.
1938 * However, free_huge_page can not adjust the reserve map. Adjust the
1939 * reserve map here to be consistent with global reserve count adjustments
1940 * to be made by free_huge_page.
1941 */
1942static void restore_reserve_on_error(struct hstate *h,
1943 struct vm_area_struct *vma, unsigned long address,
1944 struct page *page)
1945{
1946 if (unlikely(PagePrivate(page))) {
1947 long rc = vma_needs_reservation(h, vma, address);
1948
1949 if (unlikely(rc < 0)) {
1950 /*
1951 * Rare out of memory condition in reserve map
1952 * manipulation. Clear PagePrivate so that
1953 * global reserve count will not be incremented
1954 * by free_huge_page. This will make it appear
1955 * as though the reservation for this page was
1956 * consumed. This may prevent the task from
1957 * faulting in the page at a later time. This
1958 * is better than inconsistent global huge page
1959 * accounting of reserve counts.
1960 */
1961 ClearPagePrivate(page);
1962 } else if (rc) {
1963 rc = vma_add_reservation(h, vma, address);
1964 if (unlikely(rc < 0))
1965 /*
1966 * See above comment about rare out of
1967 * memory condition.
1968 */
1969 ClearPagePrivate(page);
1970 } else
1971 vma_end_reservation(h, vma, address);
1972 }
1973}
1974
Mike Kravetz70c35472015-09-08 15:01:54 -07001975struct page *alloc_huge_page(struct vm_area_struct *vma,
Mel Gorman04f2cbe2008-07-23 21:27:25 -07001976 unsigned long addr, int avoid_reserve)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001977{
David Gibson90481622012-03-21 16:34:12 -07001978 struct hugepage_subpool *spool = subpool_vma(vma);
Andi Kleena5516432008-07-23 21:27:41 -07001979 struct hstate *h = hstate_vma(vma);
Adam Litke348ea202007-11-14 16:59:37 -08001980 struct page *page;
Mike Kravetzd85f69b2015-09-08 15:01:47 -07001981 long map_chg, map_commit;
1982 long gbl_chg;
Aneesh Kumar K.V6d76dcf2012-07-31 16:42:18 -07001983 int ret, idx;
1984 struct hugetlb_cgroup *h_cg;
Adam Litke2fc39ce2007-11-14 16:59:39 -08001985
Aneesh Kumar K.V6d76dcf2012-07-31 16:42:18 -07001986 idx = hstate_index(h);
Mel Gormana1e78772008-07-23 21:27:23 -07001987 /*
Mike Kravetzd85f69b2015-09-08 15:01:47 -07001988 * Examine the region/reserve map to determine if the process
1989 * has a reservation for the page to be allocated. A return
1990 * code of zero indicates a reservation exists (no change).
Mel Gormana1e78772008-07-23 21:27:23 -07001991 */
Mike Kravetzd85f69b2015-09-08 15:01:47 -07001992 map_chg = gbl_chg = vma_needs_reservation(h, vma, addr);
1993 if (map_chg < 0)
Aneesh Kumar K.V76dcee72012-07-31 16:41:57 -07001994 return ERR_PTR(-ENOMEM);
Mike Kravetzd85f69b2015-09-08 15:01:47 -07001995
1996 /*
1997 * Processes that did not create the mapping will have no
1998 * reserves as indicated by the region/reserve map. Check
1999 * that the allocation will not exceed the subpool limit.
2000 * Allocations for MAP_NORESERVE mappings also need to be
2001 * checked against any subpool limit.
2002 */
2003 if (map_chg || avoid_reserve) {
2004 gbl_chg = hugepage_subpool_get_pages(spool, 1);
2005 if (gbl_chg < 0) {
Mike Kravetzfeba16e2015-09-08 15:01:31 -07002006 vma_end_reservation(h, vma, addr);
Aneesh Kumar K.V76dcee72012-07-31 16:41:57 -07002007 return ERR_PTR(-ENOSPC);
Mike Kravetz5e911372015-09-08 15:01:28 -07002008 }
Mel Gormana1e78772008-07-23 21:27:23 -07002009
Mike Kravetzd85f69b2015-09-08 15:01:47 -07002010 /*
2011 * Even though there was no reservation in the region/reserve
2012 * map, there could be reservations associated with the
2013 * subpool that can be used. This would be indicated if the
2014 * return value of hugepage_subpool_get_pages() is zero.
2015 * However, if avoid_reserve is specified we still avoid even
2016 * the subpool reservations.
2017 */
2018 if (avoid_reserve)
2019 gbl_chg = 1;
2020 }
2021
Aneesh Kumar K.V6d76dcf2012-07-31 16:42:18 -07002022 ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg);
Jianyu Zhan8f34af62014-06-04 16:10:36 -07002023 if (ret)
2024 goto out_subpool_put;
2025
Mel Gormana1e78772008-07-23 21:27:23 -07002026 spin_lock(&hugetlb_lock);
Mike Kravetzd85f69b2015-09-08 15:01:47 -07002027 /*
2028 * glb_chg is passed to indicate whether or not a page must be taken
2029 * from the global free pool (global change). gbl_chg == 0 indicates
2030 * a reservation exists for the allocation.
2031 */
2032 page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg);
Joonsoo Kim81a6fca2013-09-11 14:20:58 -07002033 if (!page) {
Aneesh Kumar K.V94ae8ba2012-07-31 16:42:35 -07002034 spin_unlock(&hugetlb_lock);
Dave Hansen099730d2015-11-05 18:50:17 -08002035 page = __alloc_buddy_huge_page_with_mpol(h, vma, addr);
Jianyu Zhan8f34af62014-06-04 16:10:36 -07002036 if (!page)
2037 goto out_uncharge_cgroup;
Naoya Horiguchia88c7692015-12-11 13:40:24 -08002038 if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) {
2039 SetPagePrivate(page);
2040 h->resv_huge_pages--;
2041 }
Aneesh Kumar K.V79dbb232012-07-31 16:42:32 -07002042 spin_lock(&hugetlb_lock);
2043 list_move(&page->lru, &h->hugepage_activelist);
Joonsoo Kim81a6fca2013-09-11 14:20:58 -07002044 /* Fall through */
Mel Gormana1e78772008-07-23 21:27:23 -07002045 }
Joonsoo Kim81a6fca2013-09-11 14:20:58 -07002046 hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page);
2047 spin_unlock(&hugetlb_lock);
Mel Gormana1e78772008-07-23 21:27:23 -07002048
David Gibson90481622012-03-21 16:34:12 -07002049 set_page_private(page, (unsigned long)spool);
Mel Gormana1e78772008-07-23 21:27:23 -07002050
Mike Kravetzd85f69b2015-09-08 15:01:47 -07002051 map_commit = vma_commit_reservation(h, vma, addr);
2052 if (unlikely(map_chg > map_commit)) {
Mike Kravetz33039672015-06-24 16:57:58 -07002053 /*
2054 * The page was added to the reservation map between
2055 * vma_needs_reservation and vma_commit_reservation.
2056 * This indicates a race with hugetlb_reserve_pages.
2057 * Adjust for the subpool count incremented above AND
2058 * in hugetlb_reserve_pages for the same page. Also,
2059 * the reservation count added in hugetlb_reserve_pages
2060 * no longer applies.
2061 */
2062 long rsv_adjust;
2063
2064 rsv_adjust = hugepage_subpool_put_pages(spool, 1);
2065 hugetlb_acct_memory(h, -rsv_adjust);
2066 }
Adam Litke90d8b7e2007-11-14 16:59:42 -08002067 return page;
Jianyu Zhan8f34af62014-06-04 16:10:36 -07002068
2069out_uncharge_cgroup:
2070 hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg);
2071out_subpool_put:
Mike Kravetzd85f69b2015-09-08 15:01:47 -07002072 if (map_chg || avoid_reserve)
Jianyu Zhan8f34af62014-06-04 16:10:36 -07002073 hugepage_subpool_put_pages(spool, 1);
Mike Kravetzfeba16e2015-09-08 15:01:31 -07002074 vma_end_reservation(h, vma, addr);
Jianyu Zhan8f34af62014-06-04 16:10:36 -07002075 return ERR_PTR(-ENOSPC);
David Gibsonb45b5bd2006-03-22 00:08:55 -08002076}
2077
Naoya Horiguchi74060e42013-09-11 14:22:06 -07002078/*
2079 * alloc_huge_page()'s wrapper which simply returns the page if allocation
2080 * succeeds, otherwise NULL. This function is called from new_vma_page(),
2081 * where no ERR_VALUE is expected to be returned.
2082 */
2083struct page *alloc_huge_page_noerr(struct vm_area_struct *vma,
2084 unsigned long addr, int avoid_reserve)
2085{
2086 struct page *page = alloc_huge_page(vma, addr, avoid_reserve);
2087 if (IS_ERR(page))
2088 page = NULL;
2089 return page;
2090}
2091
Aneesh Kumar K.Ve24a1302017-07-28 10:31:25 +05302092int alloc_bootmem_huge_page(struct hstate *h)
2093 __attribute__ ((weak, alias("__alloc_bootmem_huge_page")));
2094int __alloc_bootmem_huge_page(struct hstate *h)
Andi Kleenaa888a72008-07-23 21:27:47 -07002095{
2096 struct huge_bootmem_page *m;
Joonsoo Kimb2261022013-09-11 14:21:00 -07002097 int nr_nodes, node;
Andi Kleenaa888a72008-07-23 21:27:47 -07002098
Joonsoo Kimb2261022013-09-11 14:21:00 -07002099 for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) {
Andi Kleenaa888a72008-07-23 21:27:47 -07002100 void *addr;
2101
Grygorii Strashko8b89a112014-01-21 15:50:36 -08002102 addr = memblock_virt_alloc_try_nid_nopanic(
2103 huge_page_size(h), huge_page_size(h),
2104 0, BOOTMEM_ALLOC_ACCESSIBLE, node);
Andi Kleenaa888a72008-07-23 21:27:47 -07002105 if (addr) {
2106 /*
2107 * Use the beginning of the huge page to store the
2108 * huge_bootmem_page struct (until gather_bootmem
2109 * puts them into the mem_map).
2110 */
2111 m = addr;
Cyrill Gorcunov91f47662009-01-06 14:40:33 -08002112 goto found;
Andi Kleenaa888a72008-07-23 21:27:47 -07002113 }
Andi Kleenaa888a72008-07-23 21:27:47 -07002114 }
2115 return 0;
2116
2117found:
Luiz Capitulinodf994ea2014-12-12 16:55:21 -08002118 BUG_ON(!IS_ALIGNED(virt_to_phys(m), huge_page_size(h)));
Andi Kleenaa888a72008-07-23 21:27:47 -07002119 /* Put them into a private list first because mem_map is not up yet */
2120 list_add(&m->list, &huge_boot_pages);
2121 m->hstate = h;
2122 return 1;
2123}
2124
Kirill A. Shutemovd00181b92015-11-06 16:29:57 -08002125static void __init prep_compound_huge_page(struct page *page,
2126 unsigned int order)
Andy Whitcroft18229df2008-11-06 12:53:27 -08002127{
2128 if (unlikely(order > (MAX_ORDER - 1)))
2129 prep_compound_gigantic_page(page, order);
2130 else
2131 prep_compound_page(page, order);
2132}
2133
Andi Kleenaa888a72008-07-23 21:27:47 -07002134/* Put bootmem huge pages into the standard lists after mem_map is up */
2135static void __init gather_bootmem_prealloc(void)
2136{
2137 struct huge_bootmem_page *m;
2138
2139 list_for_each_entry(m, &huge_boot_pages, list) {
Andi Kleenaa888a72008-07-23 21:27:47 -07002140 struct hstate *h = m->hstate;
Becky Bruceee8f2482011-07-25 17:11:50 -07002141 struct page *page;
2142
2143#ifdef CONFIG_HIGHMEM
2144 page = pfn_to_page(m->phys >> PAGE_SHIFT);
Grygorii Strashko8b89a112014-01-21 15:50:36 -08002145 memblock_free_late(__pa(m),
2146 sizeof(struct huge_bootmem_page));
Becky Bruceee8f2482011-07-25 17:11:50 -07002147#else
2148 page = virt_to_page(m);
2149#endif
Andi Kleenaa888a72008-07-23 21:27:47 -07002150 WARN_ON(page_count(page) != 1);
Andy Whitcroft18229df2008-11-06 12:53:27 -08002151 prep_compound_huge_page(page, h->order);
Andrea Arcangelief5a22b2013-10-16 13:46:56 -07002152 WARN_ON(PageReserved(page));
Andi Kleenaa888a72008-07-23 21:27:47 -07002153 prep_new_huge_page(h, page, page_to_nid(page));
Rafael Aquinib0320c72011-06-15 15:08:39 -07002154 /*
2155 * If we had gigantic hugepages allocated at boot time, we need
2156 * to restore the 'stolen' pages to totalram_pages in order to
2157 * fix confusing memory reports from free(1) and another
2158 * side-effects, like CommitLimit going negative.
2159 */
Luiz Capitulinobae7f4a2014-06-04 16:07:08 -07002160 if (hstate_is_gigantic(h))
Jiang Liu3dcc0572013-07-03 15:03:21 -07002161 adjust_managed_page_count(page, 1 << h->order);
Cannon Matthews48b019a2018-07-03 17:02:43 -07002162 cond_resched();
Andi Kleenaa888a72008-07-23 21:27:47 -07002163 }
2164}
2165
Andi Kleen8faa8b02008-07-23 21:27:48 -07002166static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002167{
2168 unsigned long i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002169
Andi Kleene5ff2152008-07-23 21:27:42 -07002170 for (i = 0; i < h->max_huge_pages; ++i) {
Luiz Capitulinobae7f4a2014-06-04 16:07:08 -07002171 if (hstate_is_gigantic(h)) {
Andi Kleenaa888a72008-07-23 21:27:47 -07002172 if (!alloc_bootmem_huge_page(h))
2173 break;
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -08002174 } else if (!alloc_fresh_huge_page(h,
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08002175 &node_states[N_MEMORY]))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002176 break;
David Rientjes69ed7792017-07-10 15:48:50 -07002177 cond_resched();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002178 }
Liam R. Howlettd715cf82017-07-10 15:48:15 -07002179 if (i < h->max_huge_pages) {
2180 char buf[32];
2181
Matthew Wilcoxc6247f72017-07-10 15:48:56 -07002182 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
Liam R. Howlettd715cf82017-07-10 15:48:15 -07002183 pr_warn("HugeTLB: allocating %lu of page size %s failed. Only allocated %lu hugepages.\n",
2184 h->max_huge_pages, buf, i);
2185 h->max_huge_pages = i;
2186 }
Andi Kleene5ff2152008-07-23 21:27:42 -07002187}
2188
2189static void __init hugetlb_init_hstates(void)
2190{
2191 struct hstate *h;
2192
2193 for_each_hstate(h) {
Naoya Horiguchi641844f2015-06-24 16:56:59 -07002194 if (minimum_order > huge_page_order(h))
2195 minimum_order = huge_page_order(h);
2196
Andi Kleen8faa8b02008-07-23 21:27:48 -07002197 /* oversize hugepages were init'ed in early boot */
Luiz Capitulinobae7f4a2014-06-04 16:07:08 -07002198 if (!hstate_is_gigantic(h))
Andi Kleen8faa8b02008-07-23 21:27:48 -07002199 hugetlb_hstate_alloc_pages(h);
Andi Kleene5ff2152008-07-23 21:27:42 -07002200 }
Naoya Horiguchi641844f2015-06-24 16:56:59 -07002201 VM_BUG_ON(minimum_order == UINT_MAX);
Andi Kleene5ff2152008-07-23 21:27:42 -07002202}
2203
2204static void __init report_hugepages(void)
2205{
2206 struct hstate *h;
2207
2208 for_each_hstate(h) {
Andi Kleen4abd32d2008-07-23 21:27:49 -07002209 char buf[32];
Matthew Wilcoxc6247f72017-07-10 15:48:56 -07002210
2211 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
Andrew Mortonffb22af2013-02-22 16:32:08 -08002212 pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n",
Matthew Wilcoxc6247f72017-07-10 15:48:56 -07002213 buf, h->free_huge_pages);
Andi Kleene5ff2152008-07-23 21:27:42 -07002214 }
2215}
2216
Linus Torvalds1da177e2005-04-16 15:20:36 -07002217#ifdef CONFIG_HIGHMEM
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08002218static void try_to_free_low(struct hstate *h, unsigned long count,
2219 nodemask_t *nodes_allowed)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002220{
Christoph Lameter4415cc82006-09-25 23:31:55 -07002221 int i;
2222
Luiz Capitulinobae7f4a2014-06-04 16:07:08 -07002223 if (hstate_is_gigantic(h))
Andi Kleenaa888a72008-07-23 21:27:47 -07002224 return;
2225
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08002226 for_each_node_mask(i, *nodes_allowed) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002227 struct page *page, *next;
Andi Kleena5516432008-07-23 21:27:41 -07002228 struct list_head *freel = &h->hugepage_freelists[i];
2229 list_for_each_entry_safe(page, next, freel, lru) {
2230 if (count >= h->nr_huge_pages)
Adam Litke6b0c8802007-10-16 01:26:23 -07002231 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002232 if (PageHighMem(page))
2233 continue;
2234 list_del(&page->lru);
Andi Kleene5ff2152008-07-23 21:27:42 -07002235 update_and_free_page(h, page);
Andi Kleena5516432008-07-23 21:27:41 -07002236 h->free_huge_pages--;
2237 h->free_huge_pages_node[page_to_nid(page)]--;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002238 }
2239 }
2240}
2241#else
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08002242static inline void try_to_free_low(struct hstate *h, unsigned long count,
2243 nodemask_t *nodes_allowed)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002244{
2245}
2246#endif
2247
Wu Fengguang20a03072009-06-16 15:32:22 -07002248/*
2249 * Increment or decrement surplus_huge_pages. Keep node-specific counters
2250 * balanced by operating on them in a round-robin fashion.
2251 * Returns 1 if an adjustment was made.
2252 */
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08002253static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
2254 int delta)
Wu Fengguang20a03072009-06-16 15:32:22 -07002255{
Joonsoo Kimb2261022013-09-11 14:21:00 -07002256 int nr_nodes, node;
Wu Fengguang20a03072009-06-16 15:32:22 -07002257
2258 VM_BUG_ON(delta != -1 && delta != 1);
Wu Fengguang20a03072009-06-16 15:32:22 -07002259
Joonsoo Kimb2261022013-09-11 14:21:00 -07002260 if (delta < 0) {
2261 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
2262 if (h->surplus_huge_pages_node[node])
2263 goto found;
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07002264 }
Joonsoo Kimb2261022013-09-11 14:21:00 -07002265 } else {
2266 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
2267 if (h->surplus_huge_pages_node[node] <
2268 h->nr_huge_pages_node[node])
2269 goto found;
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07002270 }
Joonsoo Kimb2261022013-09-11 14:21:00 -07002271 }
2272 return 0;
Wu Fengguang20a03072009-06-16 15:32:22 -07002273
Joonsoo Kimb2261022013-09-11 14:21:00 -07002274found:
2275 h->surplus_huge_pages += delta;
2276 h->surplus_huge_pages_node[node] += delta;
2277 return 1;
Wu Fengguang20a03072009-06-16 15:32:22 -07002278}
2279
Andi Kleena5516432008-07-23 21:27:41 -07002280#define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08002281static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count,
2282 nodemask_t *nodes_allowed)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002283{
Adam Litke7893d1d2007-10-16 01:26:18 -07002284 unsigned long min_count, ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002285
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07002286 if (hstate_is_gigantic(h) && !gigantic_page_supported())
Andi Kleenaa888a72008-07-23 21:27:47 -07002287 return h->max_huge_pages;
2288
Adam Litke7893d1d2007-10-16 01:26:18 -07002289 /*
2290 * Increase the pool size
2291 * First take pages out of surplus state. Then make up the
2292 * remaining difference by allocating fresh huge pages.
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08002293 *
Naoya Horiguchid15c7c02015-11-10 14:45:11 -08002294 * We might race with __alloc_buddy_huge_page() here and be unable
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08002295 * to convert a surplus huge page to a normal huge page. That is
2296 * not critical, though, it just means the overall size of the
2297 * pool might be one hugepage larger than it needs to be, but
2298 * within all the constraints specified by the sysctls.
Adam Litke7893d1d2007-10-16 01:26:18 -07002299 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002300 spin_lock(&hugetlb_lock);
Andi Kleena5516432008-07-23 21:27:41 -07002301 while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08002302 if (!adjust_pool_surplus(h, nodes_allowed, -1))
Adam Litke7893d1d2007-10-16 01:26:18 -07002303 break;
2304 }
2305
Andi Kleena5516432008-07-23 21:27:41 -07002306 while (count > persistent_huge_pages(h)) {
Adam Litke7893d1d2007-10-16 01:26:18 -07002307 /*
2308 * If this allocation races such that we no longer need the
2309 * page, free_huge_page will handle it by freeing the page
2310 * and reducing the surplus.
2311 */
2312 spin_unlock(&hugetlb_lock);
Jia He649920c2016-08-02 14:02:31 -07002313
2314 /* yield cpu to avoid soft lockup */
2315 cond_resched();
2316
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07002317 if (hstate_is_gigantic(h))
2318 ret = alloc_fresh_gigantic_page(h, nodes_allowed);
2319 else
2320 ret = alloc_fresh_huge_page(h, nodes_allowed);
Adam Litke7893d1d2007-10-16 01:26:18 -07002321 spin_lock(&hugetlb_lock);
2322 if (!ret)
2323 goto out;
2324
Mel Gorman536240f22009-12-14 17:59:56 -08002325 /* Bail for signals. Probably ctrl-c from user */
2326 if (signal_pending(current))
2327 goto out;
Adam Litke7893d1d2007-10-16 01:26:18 -07002328 }
Adam Litke7893d1d2007-10-16 01:26:18 -07002329
2330 /*
2331 * Decrease the pool size
2332 * First return free pages to the buddy allocator (being careful
2333 * to keep enough around to satisfy reservations). Then place
2334 * pages into surplus state as needed so the pool will shrink
2335 * to the desired size as pages become free.
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08002336 *
2337 * By placing pages into the surplus state independent of the
2338 * overcommit value, we are allowing the surplus pool size to
2339 * exceed overcommit. There are few sane options here. Since
Naoya Horiguchid15c7c02015-11-10 14:45:11 -08002340 * __alloc_buddy_huge_page() is checking the global counter,
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08002341 * though, we'll note that we're not allowed to exceed surplus
2342 * and won't grow the pool anywhere else. Not until one of the
2343 * sysctls are changed, or the surplus pages go out of use.
Adam Litke7893d1d2007-10-16 01:26:18 -07002344 */
Andi Kleena5516432008-07-23 21:27:41 -07002345 min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
Adam Litke6b0c8802007-10-16 01:26:23 -07002346 min_count = max(count, min_count);
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08002347 try_to_free_low(h, min_count, nodes_allowed);
Andi Kleena5516432008-07-23 21:27:41 -07002348 while (min_count < persistent_huge_pages(h)) {
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08002349 if (!free_pool_huge_page(h, nodes_allowed, 0))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002350 break;
Mizuma, Masayoshi55f67142014-04-07 15:37:54 -07002351 cond_resched_lock(&hugetlb_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002352 }
Andi Kleena5516432008-07-23 21:27:41 -07002353 while (count < persistent_huge_pages(h)) {
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08002354 if (!adjust_pool_surplus(h, nodes_allowed, 1))
Adam Litke7893d1d2007-10-16 01:26:18 -07002355 break;
2356 }
2357out:
Andi Kleena5516432008-07-23 21:27:41 -07002358 ret = persistent_huge_pages(h);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002359 spin_unlock(&hugetlb_lock);
Adam Litke7893d1d2007-10-16 01:26:18 -07002360 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002361}
2362
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002363#define HSTATE_ATTR_RO(_name) \
2364 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
2365
2366#define HSTATE_ATTR(_name) \
2367 static struct kobj_attribute _name##_attr = \
2368 __ATTR(_name, 0644, _name##_show, _name##_store)
2369
2370static struct kobject *hugepages_kobj;
2371static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
2372
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002373static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
2374
2375static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002376{
2377 int i;
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002378
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002379 for (i = 0; i < HUGE_MAX_HSTATE; i++)
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002380 if (hstate_kobjs[i] == kobj) {
2381 if (nidp)
2382 *nidp = NUMA_NO_NODE;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002383 return &hstates[i];
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002384 }
2385
2386 return kobj_to_node_hstate(kobj, nidp);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002387}
2388
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002389static ssize_t nr_hugepages_show_common(struct kobject *kobj,
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002390 struct kobj_attribute *attr, char *buf)
2391{
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002392 struct hstate *h;
2393 unsigned long nr_huge_pages;
2394 int nid;
2395
2396 h = kobj_to_hstate(kobj, &nid);
2397 if (nid == NUMA_NO_NODE)
2398 nr_huge_pages = h->nr_huge_pages;
2399 else
2400 nr_huge_pages = h->nr_huge_pages_node[nid];
2401
2402 return sprintf(buf, "%lu\n", nr_huge_pages);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002403}
Eric B Munsonadbe8722011-01-13 15:47:27 -08002404
David Rientjes238d3c12014-08-06 16:06:51 -07002405static ssize_t __nr_hugepages_store_common(bool obey_mempolicy,
2406 struct hstate *h, int nid,
2407 unsigned long count, size_t len)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002408{
2409 int err;
David Rientjesbad44b52009-12-14 17:58:38 -08002410 NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002411
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07002412 if (hstate_is_gigantic(h) && !gigantic_page_supported()) {
Eric B Munsonadbe8722011-01-13 15:47:27 -08002413 err = -EINVAL;
2414 goto out;
2415 }
2416
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002417 if (nid == NUMA_NO_NODE) {
2418 /*
2419 * global hstate attribute
2420 */
2421 if (!(obey_mempolicy &&
2422 init_nodemask_of_mempolicy(nodes_allowed))) {
2423 NODEMASK_FREE(nodes_allowed);
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08002424 nodes_allowed = &node_states[N_MEMORY];
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002425 }
2426 } else if (nodes_allowed) {
2427 /*
2428 * per node hstate attribute: adjust count to global,
2429 * but restrict alloc/free to the specified node.
2430 */
2431 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
2432 init_nodemask_of_node(nodes_allowed, nid);
2433 } else
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08002434 nodes_allowed = &node_states[N_MEMORY];
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002435
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002436 h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002437
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08002438 if (nodes_allowed != &node_states[N_MEMORY])
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002439 NODEMASK_FREE(nodes_allowed);
2440
2441 return len;
Eric B Munsonadbe8722011-01-13 15:47:27 -08002442out:
2443 NODEMASK_FREE(nodes_allowed);
2444 return err;
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002445}
2446
David Rientjes238d3c12014-08-06 16:06:51 -07002447static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
2448 struct kobject *kobj, const char *buf,
2449 size_t len)
2450{
2451 struct hstate *h;
2452 unsigned long count;
2453 int nid;
2454 int err;
2455
2456 err = kstrtoul(buf, 10, &count);
2457 if (err)
2458 return err;
2459
2460 h = kobj_to_hstate(kobj, &nid);
2461 return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len);
2462}
2463
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002464static ssize_t nr_hugepages_show(struct kobject *kobj,
2465 struct kobj_attribute *attr, char *buf)
2466{
2467 return nr_hugepages_show_common(kobj, attr, buf);
2468}
2469
2470static ssize_t nr_hugepages_store(struct kobject *kobj,
2471 struct kobj_attribute *attr, const char *buf, size_t len)
2472{
David Rientjes238d3c12014-08-06 16:06:51 -07002473 return nr_hugepages_store_common(false, kobj, buf, len);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002474}
2475HSTATE_ATTR(nr_hugepages);
2476
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002477#ifdef CONFIG_NUMA
2478
2479/*
2480 * hstate attribute for optionally mempolicy-based constraint on persistent
2481 * huge page alloc/free.
2482 */
2483static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
2484 struct kobj_attribute *attr, char *buf)
2485{
2486 return nr_hugepages_show_common(kobj, attr, buf);
2487}
2488
2489static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
2490 struct kobj_attribute *attr, const char *buf, size_t len)
2491{
David Rientjes238d3c12014-08-06 16:06:51 -07002492 return nr_hugepages_store_common(true, kobj, buf, len);
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002493}
2494HSTATE_ATTR(nr_hugepages_mempolicy);
2495#endif
2496
2497
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002498static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
2499 struct kobj_attribute *attr, char *buf)
2500{
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002501 struct hstate *h = kobj_to_hstate(kobj, NULL);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002502 return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
2503}
Eric B Munsonadbe8722011-01-13 15:47:27 -08002504
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002505static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
2506 struct kobj_attribute *attr, const char *buf, size_t count)
2507{
2508 int err;
2509 unsigned long input;
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002510 struct hstate *h = kobj_to_hstate(kobj, NULL);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002511
Luiz Capitulinobae7f4a2014-06-04 16:07:08 -07002512 if (hstate_is_gigantic(h))
Eric B Munsonadbe8722011-01-13 15:47:27 -08002513 return -EINVAL;
2514
Jingoo Han3dbb95f2013-09-11 14:20:25 -07002515 err = kstrtoul(buf, 10, &input);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002516 if (err)
Eric B Munson73ae31e2011-01-13 15:47:28 -08002517 return err;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002518
2519 spin_lock(&hugetlb_lock);
2520 h->nr_overcommit_huge_pages = input;
2521 spin_unlock(&hugetlb_lock);
2522
2523 return count;
2524}
2525HSTATE_ATTR(nr_overcommit_hugepages);
2526
2527static ssize_t free_hugepages_show(struct kobject *kobj,
2528 struct kobj_attribute *attr, char *buf)
2529{
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002530 struct hstate *h;
2531 unsigned long free_huge_pages;
2532 int nid;
2533
2534 h = kobj_to_hstate(kobj, &nid);
2535 if (nid == NUMA_NO_NODE)
2536 free_huge_pages = h->free_huge_pages;
2537 else
2538 free_huge_pages = h->free_huge_pages_node[nid];
2539
2540 return sprintf(buf, "%lu\n", free_huge_pages);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002541}
2542HSTATE_ATTR_RO(free_hugepages);
2543
2544static ssize_t resv_hugepages_show(struct kobject *kobj,
2545 struct kobj_attribute *attr, char *buf)
2546{
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002547 struct hstate *h = kobj_to_hstate(kobj, NULL);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002548 return sprintf(buf, "%lu\n", h->resv_huge_pages);
2549}
2550HSTATE_ATTR_RO(resv_hugepages);
2551
2552static ssize_t surplus_hugepages_show(struct kobject *kobj,
2553 struct kobj_attribute *attr, char *buf)
2554{
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002555 struct hstate *h;
2556 unsigned long surplus_huge_pages;
2557 int nid;
2558
2559 h = kobj_to_hstate(kobj, &nid);
2560 if (nid == NUMA_NO_NODE)
2561 surplus_huge_pages = h->surplus_huge_pages;
2562 else
2563 surplus_huge_pages = h->surplus_huge_pages_node[nid];
2564
2565 return sprintf(buf, "%lu\n", surplus_huge_pages);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002566}
2567HSTATE_ATTR_RO(surplus_hugepages);
2568
2569static struct attribute *hstate_attrs[] = {
2570 &nr_hugepages_attr.attr,
2571 &nr_overcommit_hugepages_attr.attr,
2572 &free_hugepages_attr.attr,
2573 &resv_hugepages_attr.attr,
2574 &surplus_hugepages_attr.attr,
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002575#ifdef CONFIG_NUMA
2576 &nr_hugepages_mempolicy_attr.attr,
2577#endif
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002578 NULL,
2579};
2580
Arvind Yadav67e5ed92017-09-06 16:22:06 -07002581static const struct attribute_group hstate_attr_group = {
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002582 .attrs = hstate_attrs,
2583};
2584
Jeff Mahoney094e9532010-02-02 13:44:14 -08002585static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
2586 struct kobject **hstate_kobjs,
Arvind Yadav67e5ed92017-09-06 16:22:06 -07002587 const struct attribute_group *hstate_attr_group)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002588{
2589 int retval;
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07002590 int hi = hstate_index(h);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002591
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002592 hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
2593 if (!hstate_kobjs[hi])
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002594 return -ENOMEM;
2595
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002596 retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002597 if (retval)
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002598 kobject_put(hstate_kobjs[hi]);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002599
2600 return retval;
2601}
2602
2603static void __init hugetlb_sysfs_init(void)
2604{
2605 struct hstate *h;
2606 int err;
2607
2608 hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
2609 if (!hugepages_kobj)
2610 return;
2611
2612 for_each_hstate(h) {
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002613 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
2614 hstate_kobjs, &hstate_attr_group);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002615 if (err)
Andrew Mortonffb22af2013-02-22 16:32:08 -08002616 pr_err("Hugetlb: Unable to add hstate %s", h->name);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002617 }
2618}
2619
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002620#ifdef CONFIG_NUMA
2621
2622/*
2623 * node_hstate/s - associate per node hstate attributes, via their kobjects,
Kay Sievers10fbcf42011-12-21 14:48:43 -08002624 * with node devices in node_devices[] using a parallel array. The array
2625 * index of a node device or _hstate == node id.
2626 * This is here to avoid any static dependency of the node device driver, in
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002627 * the base kernel, on the hugetlb module.
2628 */
2629struct node_hstate {
2630 struct kobject *hugepages_kobj;
2631 struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
2632};
Alexander Kuleshovb4e289a2015-11-05 18:50:14 -08002633static struct node_hstate node_hstates[MAX_NUMNODES];
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002634
2635/*
Kay Sievers10fbcf42011-12-21 14:48:43 -08002636 * A subset of global hstate attributes for node devices
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002637 */
2638static struct attribute *per_node_hstate_attrs[] = {
2639 &nr_hugepages_attr.attr,
2640 &free_hugepages_attr.attr,
2641 &surplus_hugepages_attr.attr,
2642 NULL,
2643};
2644
Arvind Yadav67e5ed92017-09-06 16:22:06 -07002645static const struct attribute_group per_node_hstate_attr_group = {
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002646 .attrs = per_node_hstate_attrs,
2647};
2648
2649/*
Kay Sievers10fbcf42011-12-21 14:48:43 -08002650 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002651 * Returns node id via non-NULL nidp.
2652 */
2653static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
2654{
2655 int nid;
2656
2657 for (nid = 0; nid < nr_node_ids; nid++) {
2658 struct node_hstate *nhs = &node_hstates[nid];
2659 int i;
2660 for (i = 0; i < HUGE_MAX_HSTATE; i++)
2661 if (nhs->hstate_kobjs[i] == kobj) {
2662 if (nidp)
2663 *nidp = nid;
2664 return &hstates[i];
2665 }
2666 }
2667
2668 BUG();
2669 return NULL;
2670}
2671
2672/*
Kay Sievers10fbcf42011-12-21 14:48:43 -08002673 * Unregister hstate attributes from a single node device.
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002674 * No-op if no hstate attributes attached.
2675 */
Claudiu Ghioc3cd8b442013-03-04 12:46:15 +02002676static void hugetlb_unregister_node(struct node *node)
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002677{
2678 struct hstate *h;
Kay Sievers10fbcf42011-12-21 14:48:43 -08002679 struct node_hstate *nhs = &node_hstates[node->dev.id];
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002680
2681 if (!nhs->hugepages_kobj)
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -08002682 return; /* no hstate attributes */
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002683
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07002684 for_each_hstate(h) {
2685 int idx = hstate_index(h);
2686 if (nhs->hstate_kobjs[idx]) {
2687 kobject_put(nhs->hstate_kobjs[idx]);
2688 nhs->hstate_kobjs[idx] = NULL;
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002689 }
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07002690 }
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002691
2692 kobject_put(nhs->hugepages_kobj);
2693 nhs->hugepages_kobj = NULL;
2694}
2695
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002696
2697/*
Kay Sievers10fbcf42011-12-21 14:48:43 -08002698 * Register hstate attributes for a single node device.
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002699 * No-op if attributes already registered.
2700 */
Claudiu Ghioc3cd8b442013-03-04 12:46:15 +02002701static void hugetlb_register_node(struct node *node)
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002702{
2703 struct hstate *h;
Kay Sievers10fbcf42011-12-21 14:48:43 -08002704 struct node_hstate *nhs = &node_hstates[node->dev.id];
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002705 int err;
2706
2707 if (nhs->hugepages_kobj)
2708 return; /* already allocated */
2709
2710 nhs->hugepages_kobj = kobject_create_and_add("hugepages",
Kay Sievers10fbcf42011-12-21 14:48:43 -08002711 &node->dev.kobj);
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002712 if (!nhs->hugepages_kobj)
2713 return;
2714
2715 for_each_hstate(h) {
2716 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
2717 nhs->hstate_kobjs,
2718 &per_node_hstate_attr_group);
2719 if (err) {
Andrew Mortonffb22af2013-02-22 16:32:08 -08002720 pr_err("Hugetlb: Unable to add hstate %s for node %d\n",
2721 h->name, node->dev.id);
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002722 hugetlb_unregister_node(node);
2723 break;
2724 }
2725 }
2726}
2727
2728/*
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -08002729 * hugetlb init time: register hstate attributes for all registered node
Kay Sievers10fbcf42011-12-21 14:48:43 -08002730 * devices of nodes that have memory. All on-line nodes should have
2731 * registered their associated device by this time.
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002732 */
Luiz Capitulino7d9ca002014-12-12 16:55:24 -08002733static void __init hugetlb_register_all_nodes(void)
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002734{
2735 int nid;
2736
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08002737 for_each_node_state(nid, N_MEMORY) {
Wen Congyang87327942012-12-11 16:00:56 -08002738 struct node *node = node_devices[nid];
Kay Sievers10fbcf42011-12-21 14:48:43 -08002739 if (node->dev.id == nid)
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002740 hugetlb_register_node(node);
2741 }
2742
2743 /*
Kay Sievers10fbcf42011-12-21 14:48:43 -08002744 * Let the node device driver know we're here so it can
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002745 * [un]register hstate attributes on node hotplug.
2746 */
2747 register_hugetlbfs_with_node(hugetlb_register_node,
2748 hugetlb_unregister_node);
2749}
2750#else /* !CONFIG_NUMA */
2751
2752static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
2753{
2754 BUG();
2755 if (nidp)
2756 *nidp = -1;
2757 return NULL;
2758}
2759
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002760static void hugetlb_register_all_nodes(void) { }
2761
2762#endif
2763
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002764static int __init hugetlb_init(void)
2765{
Davidlohr Bueso8382d912014-04-03 14:47:31 -07002766 int i;
2767
Nishanth Aravamudan457c1b22014-05-06 12:50:00 -07002768 if (!hugepages_supported())
Benjamin Herrenschmidt0ef89d22008-07-31 00:07:30 -07002769 return 0;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002770
Nick Piggine11bfbf2008-07-23 21:27:52 -07002771 if (!size_to_hstate(default_hstate_size)) {
Liam R. Howlettd715cf82017-07-10 15:48:15 -07002772 if (default_hstate_size != 0) {
2773 pr_err("HugeTLB: unsupported default_hugepagesz %lu. Reverting to %lu\n",
2774 default_hstate_size, HPAGE_SIZE);
2775 }
2776
Nick Piggine11bfbf2008-07-23 21:27:52 -07002777 default_hstate_size = HPAGE_SIZE;
2778 if (!size_to_hstate(default_hstate_size))
2779 hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002780 }
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07002781 default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size));
Vaishali Thakkarf8b74812016-02-17 13:11:26 -08002782 if (default_hstate_max_huge_pages) {
2783 if (!default_hstate.max_huge_pages)
2784 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
2785 }
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002786
2787 hugetlb_init_hstates();
Andi Kleenaa888a72008-07-23 21:27:47 -07002788 gather_bootmem_prealloc();
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002789 report_hugepages();
2790
2791 hugetlb_sysfs_init();
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002792 hugetlb_register_all_nodes();
Jianguo Wu7179e7b2012-12-18 14:23:19 -08002793 hugetlb_cgroup_file_init();
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002794
Davidlohr Bueso8382d912014-04-03 14:47:31 -07002795#ifdef CONFIG_SMP
2796 num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus());
2797#else
2798 num_fault_mutexes = 1;
2799#endif
Mike Kravetzc672c7f2015-09-08 15:01:35 -07002800 hugetlb_fault_mutex_table =
Davidlohr Bueso8382d912014-04-03 14:47:31 -07002801 kmalloc(sizeof(struct mutex) * num_fault_mutexes, GFP_KERNEL);
Mike Kravetzc672c7f2015-09-08 15:01:35 -07002802 BUG_ON(!hugetlb_fault_mutex_table);
Davidlohr Bueso8382d912014-04-03 14:47:31 -07002803
2804 for (i = 0; i < num_fault_mutexes; i++)
Mike Kravetzc672c7f2015-09-08 15:01:35 -07002805 mutex_init(&hugetlb_fault_mutex_table[i]);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002806 return 0;
2807}
Paul Gortmaker3e89e1c2016-01-14 15:21:52 -08002808subsys_initcall(hugetlb_init);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002809
2810/* Should be called on processing a hugepagesz=... option */
Vaishali Thakkar9fee0212016-05-19 17:11:04 -07002811void __init hugetlb_bad_size(void)
2812{
2813 parsed_valid_hugepagesz = false;
2814}
2815
Kirill A. Shutemovd00181b92015-11-06 16:29:57 -08002816void __init hugetlb_add_hstate(unsigned int order)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002817{
2818 struct hstate *h;
Andi Kleen8faa8b02008-07-23 21:27:48 -07002819 unsigned long i;
2820
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002821 if (size_to_hstate(PAGE_SIZE << order)) {
Joe Perches598d8092016-03-17 14:19:44 -07002822 pr_warn("hugepagesz= specified twice, ignoring\n");
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002823 return;
2824 }
Aneesh Kumar K.V47d38342012-07-31 16:41:54 -07002825 BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002826 BUG_ON(order == 0);
Aneesh Kumar K.V47d38342012-07-31 16:41:54 -07002827 h = &hstates[hugetlb_max_hstate++];
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002828 h->order = order;
2829 h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
Andi Kleen8faa8b02008-07-23 21:27:48 -07002830 h->nr_huge_pages = 0;
2831 h->free_huge_pages = 0;
2832 for (i = 0; i < MAX_NUMNODES; ++i)
2833 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
Aneesh Kumar K.V0edaecf2012-07-31 16:42:07 -07002834 INIT_LIST_HEAD(&h->hugepage_activelist);
Andrew Morton54f18d32016-05-19 17:11:40 -07002835 h->next_nid_to_alloc = first_memory_node;
2836 h->next_nid_to_free = first_memory_node;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002837 snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
2838 huge_page_size(h)/1024);
Andi Kleen8faa8b02008-07-23 21:27:48 -07002839
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002840 parsed_hstate = h;
2841}
2842
Nick Piggine11bfbf2008-07-23 21:27:52 -07002843static int __init hugetlb_nrpages_setup(char *s)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002844{
2845 unsigned long *mhp;
Andi Kleen8faa8b02008-07-23 21:27:48 -07002846 static unsigned long *last_mhp;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002847
Vaishali Thakkar9fee0212016-05-19 17:11:04 -07002848 if (!parsed_valid_hugepagesz) {
2849 pr_warn("hugepages = %s preceded by "
2850 "an unsupported hugepagesz, ignoring\n", s);
2851 parsed_valid_hugepagesz = true;
2852 return 1;
2853 }
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002854 /*
Aneesh Kumar K.V47d38342012-07-31 16:41:54 -07002855 * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet,
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002856 * so this hugepages= parameter goes to the "default hstate".
2857 */
Vaishali Thakkar9fee0212016-05-19 17:11:04 -07002858 else if (!hugetlb_max_hstate)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002859 mhp = &default_hstate_max_huge_pages;
2860 else
2861 mhp = &parsed_hstate->max_huge_pages;
2862
Andi Kleen8faa8b02008-07-23 21:27:48 -07002863 if (mhp == last_mhp) {
Joe Perches598d8092016-03-17 14:19:44 -07002864 pr_warn("hugepages= specified twice without interleaving hugepagesz=, ignoring\n");
Andi Kleen8faa8b02008-07-23 21:27:48 -07002865 return 1;
2866 }
2867
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002868 if (sscanf(s, "%lu", mhp) <= 0)
2869 *mhp = 0;
2870
Andi Kleen8faa8b02008-07-23 21:27:48 -07002871 /*
2872 * Global state is always initialized later in hugetlb_init.
2873 * But we need to allocate >= MAX_ORDER hstates here early to still
2874 * use the bootmem allocator.
2875 */
Aneesh Kumar K.V47d38342012-07-31 16:41:54 -07002876 if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER)
Andi Kleen8faa8b02008-07-23 21:27:48 -07002877 hugetlb_hstate_alloc_pages(parsed_hstate);
2878
2879 last_mhp = mhp;
2880
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002881 return 1;
2882}
Nick Piggine11bfbf2008-07-23 21:27:52 -07002883__setup("hugepages=", hugetlb_nrpages_setup);
2884
2885static int __init hugetlb_default_setup(char *s)
2886{
2887 default_hstate_size = memparse(s, &s);
2888 return 1;
2889}
2890__setup("default_hugepagesz=", hugetlb_default_setup);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002891
Nishanth Aravamudan8a213462008-07-25 19:44:37 -07002892static unsigned int cpuset_mems_nr(unsigned int *array)
2893{
2894 int node;
2895 unsigned int nr = 0;
2896
2897 for_each_node_mask(node, cpuset_current_mems_allowed)
2898 nr += array[node];
2899
2900 return nr;
2901}
2902
2903#ifdef CONFIG_SYSCTL
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002904static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
2905 struct ctl_table *table, int write,
2906 void __user *buffer, size_t *length, loff_t *ppos)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002907{
Andi Kleene5ff2152008-07-23 21:27:42 -07002908 struct hstate *h = &default_hstate;
David Rientjes238d3c12014-08-06 16:06:51 -07002909 unsigned long tmp = h->max_huge_pages;
Michal Hocko08d4a242011-01-13 15:47:26 -08002910 int ret;
Andi Kleene5ff2152008-07-23 21:27:42 -07002911
Nishanth Aravamudan457c1b22014-05-06 12:50:00 -07002912 if (!hugepages_supported())
Jan Stancek86613622016-03-09 14:08:35 -08002913 return -EOPNOTSUPP;
Nishanth Aravamudan457c1b22014-05-06 12:50:00 -07002914
Andi Kleene5ff2152008-07-23 21:27:42 -07002915 table->data = &tmp;
2916 table->maxlen = sizeof(unsigned long);
Michal Hocko08d4a242011-01-13 15:47:26 -08002917 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2918 if (ret)
2919 goto out;
Andi Kleene5ff2152008-07-23 21:27:42 -07002920
David Rientjes238d3c12014-08-06 16:06:51 -07002921 if (write)
2922 ret = __nr_hugepages_store_common(obey_mempolicy, h,
2923 NUMA_NO_NODE, tmp, *length);
Michal Hocko08d4a242011-01-13 15:47:26 -08002924out:
2925 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002926}
Mel Gorman396faf02007-07-17 04:03:13 -07002927
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002928int hugetlb_sysctl_handler(struct ctl_table *table, int write,
2929 void __user *buffer, size_t *length, loff_t *ppos)
2930{
2931
2932 return hugetlb_sysctl_handler_common(false, table, write,
2933 buffer, length, ppos);
2934}
2935
2936#ifdef CONFIG_NUMA
2937int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
2938 void __user *buffer, size_t *length, loff_t *ppos)
2939{
2940 return hugetlb_sysctl_handler_common(true, table, write,
2941 buffer, length, ppos);
2942}
2943#endif /* CONFIG_NUMA */
2944
Nishanth Aravamudana3d0c6a2008-02-08 04:18:18 -08002945int hugetlb_overcommit_handler(struct ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07002946 void __user *buffer,
Nishanth Aravamudana3d0c6a2008-02-08 04:18:18 -08002947 size_t *length, loff_t *ppos)
2948{
Andi Kleena5516432008-07-23 21:27:41 -07002949 struct hstate *h = &default_hstate;
Andi Kleene5ff2152008-07-23 21:27:42 -07002950 unsigned long tmp;
Michal Hocko08d4a242011-01-13 15:47:26 -08002951 int ret;
Andi Kleene5ff2152008-07-23 21:27:42 -07002952
Nishanth Aravamudan457c1b22014-05-06 12:50:00 -07002953 if (!hugepages_supported())
Jan Stancek86613622016-03-09 14:08:35 -08002954 return -EOPNOTSUPP;
Nishanth Aravamudan457c1b22014-05-06 12:50:00 -07002955
Petr Holasekc033a932011-03-22 16:33:05 -07002956 tmp = h->nr_overcommit_huge_pages;
Andi Kleene5ff2152008-07-23 21:27:42 -07002957
Luiz Capitulinobae7f4a2014-06-04 16:07:08 -07002958 if (write && hstate_is_gigantic(h))
Eric B Munsonadbe8722011-01-13 15:47:27 -08002959 return -EINVAL;
2960
Andi Kleene5ff2152008-07-23 21:27:42 -07002961 table->data = &tmp;
2962 table->maxlen = sizeof(unsigned long);
Michal Hocko08d4a242011-01-13 15:47:26 -08002963 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2964 if (ret)
2965 goto out;
Andi Kleene5ff2152008-07-23 21:27:42 -07002966
2967 if (write) {
2968 spin_lock(&hugetlb_lock);
2969 h->nr_overcommit_huge_pages = tmp;
2970 spin_unlock(&hugetlb_lock);
2971 }
Michal Hocko08d4a242011-01-13 15:47:26 -08002972out:
2973 return ret;
Nishanth Aravamudana3d0c6a2008-02-08 04:18:18 -08002974}
2975
Linus Torvalds1da177e2005-04-16 15:20:36 -07002976#endif /* CONFIG_SYSCTL */
2977
Alexey Dobriyane1759c22008-10-15 23:50:22 +04002978void hugetlb_report_meminfo(struct seq_file *m)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002979{
Andi Kleena5516432008-07-23 21:27:41 -07002980 struct hstate *h = &default_hstate;
Nishanth Aravamudan457c1b22014-05-06 12:50:00 -07002981 if (!hugepages_supported())
2982 return;
Alexey Dobriyane1759c22008-10-15 23:50:22 +04002983 seq_printf(m,
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002984 "HugePages_Total: %5lu\n"
2985 "HugePages_Free: %5lu\n"
2986 "HugePages_Rsvd: %5lu\n"
2987 "HugePages_Surp: %5lu\n"
2988 "Hugepagesize: %8lu kB\n",
Andi Kleena5516432008-07-23 21:27:41 -07002989 h->nr_huge_pages,
2990 h->free_huge_pages,
2991 h->resv_huge_pages,
2992 h->surplus_huge_pages,
2993 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002994}
2995
2996int hugetlb_report_node_meminfo(int nid, char *buf)
2997{
Andi Kleena5516432008-07-23 21:27:41 -07002998 struct hstate *h = &default_hstate;
Nishanth Aravamudan457c1b22014-05-06 12:50:00 -07002999 if (!hugepages_supported())
3000 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003001 return sprintf(buf,
3002 "Node %d HugePages_Total: %5u\n"
Nishanth Aravamudana1de0912008-03-26 14:37:53 -07003003 "Node %d HugePages_Free: %5u\n"
3004 "Node %d HugePages_Surp: %5u\n",
Andi Kleena5516432008-07-23 21:27:41 -07003005 nid, h->nr_huge_pages_node[nid],
3006 nid, h->free_huge_pages_node[nid],
3007 nid, h->surplus_huge_pages_node[nid]);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003008}
3009
David Rientjes949f7ec2013-04-29 15:07:48 -07003010void hugetlb_show_meminfo(void)
3011{
3012 struct hstate *h;
3013 int nid;
3014
Nishanth Aravamudan457c1b22014-05-06 12:50:00 -07003015 if (!hugepages_supported())
3016 return;
3017
David Rientjes949f7ec2013-04-29 15:07:48 -07003018 for_each_node_state(nid, N_MEMORY)
3019 for_each_hstate(h)
3020 pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
3021 nid,
3022 h->nr_huge_pages_node[nid],
3023 h->free_huge_pages_node[nid],
3024 h->surplus_huge_pages_node[nid],
3025 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
3026}
3027
Naoya Horiguchi5d317b22015-11-05 18:47:14 -08003028void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm)
3029{
3030 seq_printf(m, "HugetlbPages:\t%8lu kB\n",
3031 atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10));
3032}
3033
Linus Torvalds1da177e2005-04-16 15:20:36 -07003034/* Return the number pages of memory we physically have, in PAGE_SIZE units. */
3035unsigned long hugetlb_total_pages(void)
3036{
Wanpeng Lid0028582013-03-22 15:04:40 -07003037 struct hstate *h;
3038 unsigned long nr_total_pages = 0;
3039
3040 for_each_hstate(h)
3041 nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h);
3042 return nr_total_pages;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003043}
Linus Torvalds1da177e2005-04-16 15:20:36 -07003044
Andi Kleena5516432008-07-23 21:27:41 -07003045static int hugetlb_acct_memory(struct hstate *h, long delta)
Mel Gormanfc1b8a72008-07-23 21:27:22 -07003046{
3047 int ret = -ENOMEM;
3048
3049 spin_lock(&hugetlb_lock);
3050 /*
3051 * When cpuset is configured, it breaks the strict hugetlb page
3052 * reservation as the accounting is done on a global variable. Such
3053 * reservation is completely rubbish in the presence of cpuset because
3054 * the reservation is not checked against page availability for the
3055 * current cpuset. Application can still potentially OOM'ed by kernel
3056 * with lack of free htlb page in cpuset that the task is in.
3057 * Attempt to enforce strict accounting with cpuset is almost
3058 * impossible (or too ugly) because cpuset is too fluid that
3059 * task or memory node can be dynamically moved between cpusets.
3060 *
3061 * The change of semantics for shared hugetlb mapping with cpuset is
3062 * undesirable. However, in order to preserve some of the semantics,
3063 * we fall back to check against current free page availability as
3064 * a best attempt and hopefully to minimize the impact of changing
3065 * semantics that cpuset has.
3066 */
3067 if (delta > 0) {
Andi Kleena5516432008-07-23 21:27:41 -07003068 if (gather_surplus_pages(h, delta) < 0)
Mel Gormanfc1b8a72008-07-23 21:27:22 -07003069 goto out;
3070
Andi Kleena5516432008-07-23 21:27:41 -07003071 if (delta > cpuset_mems_nr(h->free_huge_pages_node)) {
3072 return_unused_surplus_pages(h, delta);
Mel Gormanfc1b8a72008-07-23 21:27:22 -07003073 goto out;
3074 }
3075 }
3076
3077 ret = 0;
3078 if (delta < 0)
Andi Kleena5516432008-07-23 21:27:41 -07003079 return_unused_surplus_pages(h, (unsigned long) -delta);
Mel Gormanfc1b8a72008-07-23 21:27:22 -07003080
3081out:
3082 spin_unlock(&hugetlb_lock);
3083 return ret;
3084}
3085
Andy Whitcroft84afd992008-07-23 21:27:32 -07003086static void hugetlb_vm_op_open(struct vm_area_struct *vma)
3087{
Joonsoo Kimf522c3a2013-09-11 14:21:53 -07003088 struct resv_map *resv = vma_resv_map(vma);
Andy Whitcroft84afd992008-07-23 21:27:32 -07003089
3090 /*
3091 * This new VMA should share its siblings reservation map if present.
3092 * The VMA will only ever have a valid reservation map pointer where
3093 * it is being copied for another still existing VMA. As that VMA
Lucas De Marchi25985ed2011-03-30 22:57:33 -03003094 * has a reference to the reservation map it cannot disappear until
Andy Whitcroft84afd992008-07-23 21:27:32 -07003095 * after this open call completes. It is therefore safe to take a
3096 * new reference here without additional locking.
3097 */
Joonsoo Kim4e35f482014-04-03 14:47:30 -07003098 if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
Joonsoo Kimf522c3a2013-09-11 14:21:53 -07003099 kref_get(&resv->refs);
Andy Whitcroft84afd992008-07-23 21:27:32 -07003100}
3101
Mel Gormana1e78772008-07-23 21:27:23 -07003102static void hugetlb_vm_op_close(struct vm_area_struct *vma)
3103{
Andi Kleena5516432008-07-23 21:27:41 -07003104 struct hstate *h = hstate_vma(vma);
Joonsoo Kimf522c3a2013-09-11 14:21:53 -07003105 struct resv_map *resv = vma_resv_map(vma);
David Gibson90481622012-03-21 16:34:12 -07003106 struct hugepage_subpool *spool = subpool_vma(vma);
Joonsoo Kim4e35f482014-04-03 14:47:30 -07003107 unsigned long reserve, start, end;
Mike Kravetz1c5ecae2015-04-15 16:13:39 -07003108 long gbl_reserve;
Andy Whitcroft84afd992008-07-23 21:27:32 -07003109
Joonsoo Kim4e35f482014-04-03 14:47:30 -07003110 if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER))
3111 return;
Andy Whitcroft84afd992008-07-23 21:27:32 -07003112
Joonsoo Kim4e35f482014-04-03 14:47:30 -07003113 start = vma_hugecache_offset(h, vma, vma->vm_start);
3114 end = vma_hugecache_offset(h, vma, vma->vm_end);
Andy Whitcroft84afd992008-07-23 21:27:32 -07003115
Joonsoo Kim4e35f482014-04-03 14:47:30 -07003116 reserve = (end - start) - region_count(resv, start, end);
Andy Whitcroft84afd992008-07-23 21:27:32 -07003117
Joonsoo Kim4e35f482014-04-03 14:47:30 -07003118 kref_put(&resv->refs, resv_map_release);
3119
3120 if (reserve) {
Mike Kravetz1c5ecae2015-04-15 16:13:39 -07003121 /*
3122 * Decrement reserve counts. The global reserve count may be
3123 * adjusted if the subpool has a minimum size.
3124 */
3125 gbl_reserve = hugepage_subpool_put_pages(spool, reserve);
3126 hugetlb_acct_memory(h, -gbl_reserve);
Andy Whitcroft84afd992008-07-23 21:27:32 -07003127 }
Mel Gormana1e78772008-07-23 21:27:23 -07003128}
3129
Dan Williamsc6c78a12017-11-29 16:10:28 -08003130static int hugetlb_vm_op_split(struct vm_area_struct *vma, unsigned long addr)
3131{
3132 if (addr & ~(huge_page_mask(hstate_vma(vma))))
3133 return -EINVAL;
3134 return 0;
3135}
3136
Linus Torvalds1da177e2005-04-16 15:20:36 -07003137/*
3138 * We cannot handle pagefaults against hugetlb pages at all. They cause
3139 * handle_mm_fault() to try to instantiate regular-sized pages in the
3140 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
3141 * this far.
3142 */
Dave Jiang11bac802017-02-24 14:56:41 -08003143static int hugetlb_vm_op_fault(struct vm_fault *vmf)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003144{
3145 BUG();
Nick Piggind0217ac2007-07-19 01:47:03 -07003146 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003147}
3148
Alexey Dobriyanf0f37e22009-09-27 22:29:37 +04003149const struct vm_operations_struct hugetlb_vm_ops = {
Nick Piggind0217ac2007-07-19 01:47:03 -07003150 .fault = hugetlb_vm_op_fault,
Andy Whitcroft84afd992008-07-23 21:27:32 -07003151 .open = hugetlb_vm_op_open,
Mel Gormana1e78772008-07-23 21:27:23 -07003152 .close = hugetlb_vm_op_close,
Dan Williamsc6c78a12017-11-29 16:10:28 -08003153 .split = hugetlb_vm_op_split,
Linus Torvalds1da177e2005-04-16 15:20:36 -07003154};
3155
David Gibson1e8f8892006-01-06 00:10:44 -08003156static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
3157 int writable)
David Gibson63551ae2005-06-21 17:14:44 -07003158{
3159 pte_t entry;
3160
David Gibson1e8f8892006-01-06 00:10:44 -08003161 if (writable) {
Gerald Schaefer106c9922013-04-29 15:07:23 -07003162 entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page,
3163 vma->vm_page_prot)));
David Gibson63551ae2005-06-21 17:14:44 -07003164 } else {
Gerald Schaefer106c9922013-04-29 15:07:23 -07003165 entry = huge_pte_wrprotect(mk_huge_pte(page,
3166 vma->vm_page_prot));
David Gibson63551ae2005-06-21 17:14:44 -07003167 }
3168 entry = pte_mkyoung(entry);
3169 entry = pte_mkhuge(entry);
Chris Metcalfd9ed9fa2012-04-01 14:01:34 -04003170 entry = arch_make_huge_pte(entry, vma, page, writable);
David Gibson63551ae2005-06-21 17:14:44 -07003171
3172 return entry;
3173}
3174
David Gibson1e8f8892006-01-06 00:10:44 -08003175static void set_huge_ptep_writable(struct vm_area_struct *vma,
3176 unsigned long address, pte_t *ptep)
3177{
3178 pte_t entry;
3179
Gerald Schaefer106c9922013-04-29 15:07:23 -07003180 entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep)));
Chris Forbes32f84522011-07-25 17:12:14 -07003181 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
Russell King4b3073e2009-12-18 16:40:18 +00003182 update_mmu_cache(vma, address, ptep);
David Gibson1e8f8892006-01-06 00:10:44 -08003183}
3184
Aneesh Kumar K.Vd5ed7442017-07-06 15:38:47 -07003185bool is_hugetlb_entry_migration(pte_t pte)
Naoya Horiguchi4a705fe2014-06-23 13:22:03 -07003186{
3187 swp_entry_t swp;
3188
3189 if (huge_pte_none(pte) || pte_present(pte))
Aneesh Kumar K.Vd5ed7442017-07-06 15:38:47 -07003190 return false;
Naoya Horiguchi4a705fe2014-06-23 13:22:03 -07003191 swp = pte_to_swp_entry(pte);
3192 if (non_swap_entry(swp) && is_migration_entry(swp))
Aneesh Kumar K.Vd5ed7442017-07-06 15:38:47 -07003193 return true;
Naoya Horiguchi4a705fe2014-06-23 13:22:03 -07003194 else
Aneesh Kumar K.Vd5ed7442017-07-06 15:38:47 -07003195 return false;
Naoya Horiguchi4a705fe2014-06-23 13:22:03 -07003196}
3197
3198static int is_hugetlb_entry_hwpoisoned(pte_t pte)
3199{
3200 swp_entry_t swp;
3201
3202 if (huge_pte_none(pte) || pte_present(pte))
3203 return 0;
3204 swp = pte_to_swp_entry(pte);
3205 if (non_swap_entry(swp) && is_hwpoison_entry(swp))
3206 return 1;
3207 else
3208 return 0;
3209}
David Gibson1e8f8892006-01-06 00:10:44 -08003210
David Gibson63551ae2005-06-21 17:14:44 -07003211int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
3212 struct vm_area_struct *vma)
3213{
3214 pte_t *src_pte, *dst_pte, entry;
3215 struct page *ptepage;
Hugh Dickins1c598272005-10-19 21:23:43 -07003216 unsigned long addr;
David Gibson1e8f8892006-01-06 00:10:44 -08003217 int cow;
Andi Kleena5516432008-07-23 21:27:41 -07003218 struct hstate *h = hstate_vma(vma);
3219 unsigned long sz = huge_page_size(h);
Andreas Sandberge8569dd2014-01-21 15:49:09 -08003220 unsigned long mmun_start; /* For mmu_notifiers */
3221 unsigned long mmun_end; /* For mmu_notifiers */
3222 int ret = 0;
David Gibson1e8f8892006-01-06 00:10:44 -08003223
3224 cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
David Gibson63551ae2005-06-21 17:14:44 -07003225
Andreas Sandberge8569dd2014-01-21 15:49:09 -08003226 mmun_start = vma->vm_start;
3227 mmun_end = vma->vm_end;
3228 if (cow)
3229 mmu_notifier_invalidate_range_start(src, mmun_start, mmun_end);
3230
Andi Kleena5516432008-07-23 21:27:41 -07003231 for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003232 spinlock_t *src_ptl, *dst_ptl;
Punit Agrawal7868a202017-07-06 15:39:42 -07003233 src_pte = huge_pte_offset(src, addr, sz);
Hugh Dickinsc74df322005-10-29 18:16:23 -07003234 if (!src_pte)
3235 continue;
Andi Kleena5516432008-07-23 21:27:41 -07003236 dst_pte = huge_pte_alloc(dst, addr, sz);
Andreas Sandberge8569dd2014-01-21 15:49:09 -08003237 if (!dst_pte) {
3238 ret = -ENOMEM;
3239 break;
3240 }
Larry Woodmanc5c99422008-01-24 05:49:25 -08003241
3242 /* If the pagetables are shared don't copy or take references */
3243 if (dst_pte == src_pte)
3244 continue;
3245
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003246 dst_ptl = huge_pte_lock(h, dst, dst_pte);
3247 src_ptl = huge_pte_lockptr(h, src, src_pte);
3248 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
Naoya Horiguchi4a705fe2014-06-23 13:22:03 -07003249 entry = huge_ptep_get(src_pte);
3250 if (huge_pte_none(entry)) { /* skip none entry */
3251 ;
3252 } else if (unlikely(is_hugetlb_entry_migration(entry) ||
3253 is_hugetlb_entry_hwpoisoned(entry))) {
3254 swp_entry_t swp_entry = pte_to_swp_entry(entry);
3255
3256 if (is_write_migration_entry(swp_entry) && cow) {
3257 /*
3258 * COW mappings require pages in both
3259 * parent and child to be set to read.
3260 */
3261 make_migration_entry_read(&swp_entry);
3262 entry = swp_entry_to_pte(swp_entry);
Punit Agrawale5251fd2017-07-06 15:39:50 -07003263 set_huge_swap_pte_at(src, addr, src_pte,
3264 entry, sz);
Naoya Horiguchi4a705fe2014-06-23 13:22:03 -07003265 }
Punit Agrawale5251fd2017-07-06 15:39:50 -07003266 set_huge_swap_pte_at(dst, addr, dst_pte, entry, sz);
Naoya Horiguchi4a705fe2014-06-23 13:22:03 -07003267 } else {
Joerg Roedel34ee6452014-11-13 13:46:09 +11003268 if (cow) {
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07003269 huge_ptep_set_wrprotect(src, addr, src_pte);
Joerg Roedel34ee6452014-11-13 13:46:09 +11003270 mmu_notifier_invalidate_range(src, mmun_start,
3271 mmun_end);
3272 }
Naoya Horiguchi0253d632014-07-23 14:00:19 -07003273 entry = huge_ptep_get(src_pte);
Hugh Dickins1c598272005-10-19 21:23:43 -07003274 ptepage = pte_page(entry);
3275 get_page(ptepage);
Kirill A. Shutemov53f92632016-01-15 16:53:42 -08003276 page_dup_rmap(ptepage, true);
Hugh Dickins1c598272005-10-19 21:23:43 -07003277 set_huge_pte_at(dst, addr, dst_pte, entry);
Naoya Horiguchi5d317b22015-11-05 18:47:14 -08003278 hugetlb_count_add(pages_per_huge_page(h), dst);
Hugh Dickins1c598272005-10-19 21:23:43 -07003279 }
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003280 spin_unlock(src_ptl);
3281 spin_unlock(dst_ptl);
David Gibson63551ae2005-06-21 17:14:44 -07003282 }
David Gibson63551ae2005-06-21 17:14:44 -07003283
Andreas Sandberge8569dd2014-01-21 15:49:09 -08003284 if (cow)
3285 mmu_notifier_invalidate_range_end(src, mmun_start, mmun_end);
3286
3287 return ret;
David Gibson63551ae2005-06-21 17:14:44 -07003288}
3289
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07003290void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
3291 unsigned long start, unsigned long end,
3292 struct page *ref_page)
David Gibson63551ae2005-06-21 17:14:44 -07003293{
3294 struct mm_struct *mm = vma->vm_mm;
3295 unsigned long address;
David Gibsonc7546f82005-08-05 11:59:35 -07003296 pte_t *ptep;
David Gibson63551ae2005-06-21 17:14:44 -07003297 pte_t pte;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003298 spinlock_t *ptl;
David Gibson63551ae2005-06-21 17:14:44 -07003299 struct page *page;
Andi Kleena5516432008-07-23 21:27:41 -07003300 struct hstate *h = hstate_vma(vma);
3301 unsigned long sz = huge_page_size(h);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07003302 const unsigned long mmun_start = start; /* For mmu_notifiers */
3303 const unsigned long mmun_end = end; /* For mmu_notifiers */
Andi Kleena5516432008-07-23 21:27:41 -07003304
David Gibson63551ae2005-06-21 17:14:44 -07003305 WARN_ON(!is_vm_hugetlb_page(vma));
Andi Kleena5516432008-07-23 21:27:41 -07003306 BUG_ON(start & ~huge_page_mask(h));
3307 BUG_ON(end & ~huge_page_mask(h));
David Gibson63551ae2005-06-21 17:14:44 -07003308
Aneesh Kumar K.V07e32662016-12-12 16:42:40 -08003309 /*
3310 * This is a hugetlb vma, all the pte entries should point
3311 * to huge page.
3312 */
3313 tlb_remove_check_page_size_change(tlb, sz);
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07003314 tlb_start_vma(tlb, vma);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07003315 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
Hillf Danton569f48b82014-12-10 15:44:41 -08003316 address = start;
Hillf Danton569f48b82014-12-10 15:44:41 -08003317 for (; address < end; address += sz) {
Punit Agrawal7868a202017-07-06 15:39:42 -07003318 ptep = huge_pte_offset(mm, address, sz);
Adam Litke4c887262005-10-29 18:16:46 -07003319 if (!ptep)
David Gibsonc7546f82005-08-05 11:59:35 -07003320 continue;
3321
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003322 ptl = huge_pte_lock(h, mm, ptep);
Aneesh Kumar K.V31d49da2016-07-26 15:24:06 -07003323 if (huge_pmd_unshare(mm, &address, ptep)) {
3324 spin_unlock(ptl);
3325 continue;
3326 }
Chen, Kenneth W39dde652006-12-06 20:32:03 -08003327
Hillf Danton66293262012-03-23 15:01:48 -07003328 pte = huge_ptep_get(ptep);
Aneesh Kumar K.V31d49da2016-07-26 15:24:06 -07003329 if (huge_pte_none(pte)) {
3330 spin_unlock(ptl);
3331 continue;
3332 }
Hillf Danton66293262012-03-23 15:01:48 -07003333
3334 /*
Naoya Horiguchi9fbc1f62015-02-11 15:25:32 -08003335 * Migrating hugepage or HWPoisoned hugepage is already
3336 * unmapped and its refcount is dropped, so just clear pte here.
Hillf Danton66293262012-03-23 15:01:48 -07003337 */
Naoya Horiguchi9fbc1f62015-02-11 15:25:32 -08003338 if (unlikely(!pte_present(pte))) {
Punit Agrawal9386fac2017-07-06 15:39:46 -07003339 huge_pte_clear(mm, address, ptep, sz);
Aneesh Kumar K.V31d49da2016-07-26 15:24:06 -07003340 spin_unlock(ptl);
3341 continue;
Naoya Horiguchi8c4894c2012-12-12 13:52:28 -08003342 }
Hillf Danton66293262012-03-23 15:01:48 -07003343
3344 page = pte_page(pte);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003345 /*
3346 * If a reference page is supplied, it is because a specific
3347 * page is being unmapped, not a range. Ensure the page we
3348 * are about to unmap is the actual page of interest.
3349 */
3350 if (ref_page) {
Aneesh Kumar K.V31d49da2016-07-26 15:24:06 -07003351 if (page != ref_page) {
3352 spin_unlock(ptl);
3353 continue;
3354 }
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003355 /*
3356 * Mark the VMA as having unmapped its page so that
3357 * future faults in this VMA will fail rather than
3358 * looking like data was lost
3359 */
3360 set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
3361 }
3362
David Gibsonc7546f82005-08-05 11:59:35 -07003363 pte = huge_ptep_get_and_clear(mm, address, ptep);
Aneesh Kumar K.Vb528e4b2016-12-12 16:42:37 -08003364 tlb_remove_huge_tlb_entry(h, tlb, ptep, address);
Gerald Schaefer106c9922013-04-29 15:07:23 -07003365 if (huge_pte_dirty(pte))
Ken Chen6649a382007-02-08 14:20:27 -08003366 set_page_dirty(page);
Hillf Danton9e811302012-03-21 16:34:03 -07003367
Naoya Horiguchi5d317b22015-11-05 18:47:14 -08003368 hugetlb_count_sub(pages_per_huge_page(h), mm);
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -08003369 page_remove_rmap(page, true);
Aneesh Kumar K.V31d49da2016-07-26 15:24:06 -07003370
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003371 spin_unlock(ptl);
Aneesh Kumar K.Ve77b0852016-07-26 15:24:12 -07003372 tlb_remove_page_size(tlb, page, huge_page_size(h));
Aneesh Kumar K.V31d49da2016-07-26 15:24:06 -07003373 /*
3374 * Bail out after unmapping reference page if supplied
3375 */
3376 if (ref_page)
3377 break;
Chen, Kenneth Wfe1668a2006-10-04 02:15:24 -07003378 }
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07003379 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07003380 tlb_end_vma(tlb, vma);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003381}
David Gibson63551ae2005-06-21 17:14:44 -07003382
Mel Gormand8333522012-07-31 16:46:20 -07003383void __unmap_hugepage_range_final(struct mmu_gather *tlb,
3384 struct vm_area_struct *vma, unsigned long start,
3385 unsigned long end, struct page *ref_page)
3386{
3387 __unmap_hugepage_range(tlb, vma, start, end, ref_page);
3388
3389 /*
3390 * Clear this flag so that x86's huge_pmd_share page_table_shareable
3391 * test will fail on a vma being torn down, and not grab a page table
3392 * on its way out. We're lucky that the flag has such an appropriate
3393 * name, and can in fact be safely cleared here. We could clear it
3394 * before the __unmap_hugepage_range above, but all that's necessary
Davidlohr Buesoc8c06ef2014-12-12 16:54:24 -08003395 * is to clear it before releasing the i_mmap_rwsem. This works
Mel Gormand8333522012-07-31 16:46:20 -07003396 * because in the context this is called, the VMA is about to be
Davidlohr Buesoc8c06ef2014-12-12 16:54:24 -08003397 * destroyed and the i_mmap_rwsem is held.
Mel Gormand8333522012-07-31 16:46:20 -07003398 */
3399 vma->vm_flags &= ~VM_MAYSHARE;
3400}
3401
Chen, Kenneth W502717f2006-10-11 01:20:46 -07003402void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003403 unsigned long end, struct page *ref_page)
Chen, Kenneth W502717f2006-10-11 01:20:46 -07003404{
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07003405 struct mm_struct *mm;
3406 struct mmu_gather tlb;
3407
3408 mm = vma->vm_mm;
3409
Linus Torvalds2b047252013-08-15 11:42:25 -07003410 tlb_gather_mmu(&tlb, mm, start, end);
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07003411 __unmap_hugepage_range(&tlb, vma, start, end, ref_page);
3412 tlb_finish_mmu(&tlb, start, end);
Chen, Kenneth W502717f2006-10-11 01:20:46 -07003413}
3414
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003415/*
3416 * This is called when the original mapper is failing to COW a MAP_PRIVATE
3417 * mappping it owns the reserve page for. The intention is to unmap the page
3418 * from other VMAs and let the children be SIGKILLed if they are faulting the
3419 * same region.
3420 */
Davidlohr Bueso2f4612a2014-08-06 16:06:45 -07003421static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
3422 struct page *page, unsigned long address)
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003423{
Adam Litke75266742008-11-12 13:24:56 -08003424 struct hstate *h = hstate_vma(vma);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003425 struct vm_area_struct *iter_vma;
3426 struct address_space *mapping;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003427 pgoff_t pgoff;
3428
3429 /*
3430 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
3431 * from page cache lookup which is in HPAGE_SIZE units.
3432 */
Adam Litke75266742008-11-12 13:24:56 -08003433 address = address & huge_page_mask(h);
Michal Hocko36e4f202012-10-08 16:33:31 -07003434 pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) +
3435 vma->vm_pgoff;
Al Viro93c76a32015-12-04 23:45:44 -05003436 mapping = vma->vm_file->f_mapping;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003437
Mel Gorman4eb2b1d2009-12-14 17:59:53 -08003438 /*
3439 * Take the mapping lock for the duration of the table walk. As
3440 * this mapping should be shared between all the VMAs,
3441 * __unmap_hugepage_range() is called as the lock is already held
3442 */
Davidlohr Bueso83cde9e2014-12-12 16:54:21 -08003443 i_mmap_lock_write(mapping);
Michel Lespinasse6b2dbba2012-10-08 16:31:25 -07003444 vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) {
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003445 /* Do not unmap the current VMA */
3446 if (iter_vma == vma)
3447 continue;
3448
3449 /*
Mel Gorman2f84a892015-10-01 15:36:57 -07003450 * Shared VMAs have their own reserves and do not affect
3451 * MAP_PRIVATE accounting but it is possible that a shared
3452 * VMA is using the same page so check and skip such VMAs.
3453 */
3454 if (iter_vma->vm_flags & VM_MAYSHARE)
3455 continue;
3456
3457 /*
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003458 * Unmap the page from other VMAs without their own reserves.
3459 * They get marked to be SIGKILLed if they fault in these
3460 * areas. This is because a future no-page fault on this VMA
3461 * could insert a zeroed page instead of the data existing
3462 * from the time of fork. This would look like data corruption
3463 */
3464 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07003465 unmap_hugepage_range(iter_vma, address,
3466 address + huge_page_size(h), page);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003467 }
Davidlohr Bueso83cde9e2014-12-12 16:54:21 -08003468 i_mmap_unlock_write(mapping);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003469}
3470
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09003471/*
3472 * Hugetlb_cow() should be called with page lock of the original hugepage held.
Michal Hockoef009b22012-01-10 15:07:21 -08003473 * Called with hugetlb_instantiation_mutex held and pte_page locked so we
3474 * cannot race with other handlers or page migration.
3475 * Keep the pte_same checks anyway to make transition from the mutex easier.
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09003476 */
David Gibson1e8f8892006-01-06 00:10:44 -08003477static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
Aneesh Kumar K.V3999f522016-12-12 16:41:56 -08003478 unsigned long address, pte_t *ptep,
3479 struct page *pagecache_page, spinlock_t *ptl)
David Gibson1e8f8892006-01-06 00:10:44 -08003480{
Aneesh Kumar K.V3999f522016-12-12 16:41:56 -08003481 pte_t pte;
Andi Kleena5516432008-07-23 21:27:41 -07003482 struct hstate *h = hstate_vma(vma);
David Gibson1e8f8892006-01-06 00:10:44 -08003483 struct page *old_page, *new_page;
Davidlohr Buesoad4404a2014-08-06 16:06:47 -07003484 int ret = 0, outside_reserve = 0;
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07003485 unsigned long mmun_start; /* For mmu_notifiers */
3486 unsigned long mmun_end; /* For mmu_notifiers */
David Gibson1e8f8892006-01-06 00:10:44 -08003487
Aneesh Kumar K.V3999f522016-12-12 16:41:56 -08003488 pte = huge_ptep_get(ptep);
David Gibson1e8f8892006-01-06 00:10:44 -08003489 old_page = pte_page(pte);
3490
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003491retry_avoidcopy:
David Gibson1e8f8892006-01-06 00:10:44 -08003492 /* If no-one else is actually using this page, avoid the copy
3493 * and just make the page writable */
Joonsoo Kim37a21402013-09-11 14:21:04 -07003494 if (page_mapcount(old_page) == 1 && PageAnon(old_page)) {
Hugh Dickins5a499732016-07-14 12:07:38 -07003495 page_move_anon_rmap(old_page, vma);
David Gibson1e8f8892006-01-06 00:10:44 -08003496 set_huge_ptep_writable(vma, address, ptep);
Nick Piggin83c54072007-07-19 01:47:05 -07003497 return 0;
David Gibson1e8f8892006-01-06 00:10:44 -08003498 }
3499
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003500 /*
3501 * If the process that created a MAP_PRIVATE mapping is about to
3502 * perform a COW due to a shared page count, attempt to satisfy
3503 * the allocation without using the existing reserves. The pagecache
3504 * page is used to determine if the reserve at this address was
3505 * consumed or not. If reserves were used, a partial faulted mapping
3506 * at the time of fork() could consume its reserves on COW instead
3507 * of the full address range.
3508 */
Joonsoo Kim5944d012013-09-11 14:21:55 -07003509 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003510 old_page != pagecache_page)
3511 outside_reserve = 1;
3512
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03003513 get_page(old_page);
Larry Woodmanb76c8cf2009-12-14 17:59:37 -08003514
Davidlohr Buesoad4404a2014-08-06 16:06:47 -07003515 /*
3516 * Drop page table lock as buddy allocator may be called. It will
3517 * be acquired again before returning to the caller, as expected.
3518 */
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003519 spin_unlock(ptl);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003520 new_page = alloc_huge_page(vma, address, outside_reserve);
David Gibson1e8f8892006-01-06 00:10:44 -08003521
Adam Litke2fc39ce2007-11-14 16:59:39 -08003522 if (IS_ERR(new_page)) {
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003523 /*
3524 * If a process owning a MAP_PRIVATE mapping fails to COW,
3525 * it is due to references held by a child and an insufficient
3526 * huge page pool. To guarantee the original mappers
3527 * reliability, unmap the page from child processes. The child
3528 * may get SIGKILLed if it later faults.
3529 */
3530 if (outside_reserve) {
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03003531 put_page(old_page);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003532 BUG_ON(huge_pte_none(pte));
Davidlohr Bueso2f4612a2014-08-06 16:06:45 -07003533 unmap_ref_private(mm, vma, old_page, address);
3534 BUG_ON(huge_pte_none(pte));
3535 spin_lock(ptl);
Punit Agrawal7868a202017-07-06 15:39:42 -07003536 ptep = huge_pte_offset(mm, address & huge_page_mask(h),
3537 huge_page_size(h));
Davidlohr Bueso2f4612a2014-08-06 16:06:45 -07003538 if (likely(ptep &&
3539 pte_same(huge_ptep_get(ptep), pte)))
3540 goto retry_avoidcopy;
3541 /*
3542 * race occurs while re-acquiring page table
3543 * lock, and our job is done.
3544 */
3545 return 0;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003546 }
3547
Davidlohr Buesoad4404a2014-08-06 16:06:47 -07003548 ret = (PTR_ERR(new_page) == -ENOMEM) ?
3549 VM_FAULT_OOM : VM_FAULT_SIGBUS;
3550 goto out_release_old;
David Gibson1e8f8892006-01-06 00:10:44 -08003551 }
3552
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09003553 /*
3554 * When the original hugepage is shared one, it does not have
3555 * anon_vma prepared.
3556 */
Dean Nelson44e2aa92010-10-26 14:22:08 -07003557 if (unlikely(anon_vma_prepare(vma))) {
Davidlohr Buesoad4404a2014-08-06 16:06:47 -07003558 ret = VM_FAULT_OOM;
3559 goto out_release_all;
Dean Nelson44e2aa92010-10-26 14:22:08 -07003560 }
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09003561
Andrea Arcangeli47ad8472011-01-13 15:46:47 -08003562 copy_user_huge_page(new_page, old_page, address, vma,
3563 pages_per_huge_page(h));
Nick Piggin0ed361d2008-02-04 22:29:34 -08003564 __SetPageUptodate(new_page);
Naoya Horiguchibcc54222015-04-15 16:14:38 -07003565 set_page_huge_active(new_page);
David Gibson1e8f8892006-01-06 00:10:44 -08003566
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07003567 mmun_start = address & huge_page_mask(h);
3568 mmun_end = mmun_start + huge_page_size(h);
3569 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
Davidlohr Buesoad4404a2014-08-06 16:06:47 -07003570
Larry Woodmanb76c8cf2009-12-14 17:59:37 -08003571 /*
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003572 * Retake the page table lock to check for racing updates
Larry Woodmanb76c8cf2009-12-14 17:59:37 -08003573 * before the page tables are altered
3574 */
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003575 spin_lock(ptl);
Punit Agrawal7868a202017-07-06 15:39:42 -07003576 ptep = huge_pte_offset(mm, address & huge_page_mask(h),
3577 huge_page_size(h));
Naoya Horiguchia9af0c52014-04-07 15:36:54 -07003578 if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) {
Joonsoo Kim07443a82013-09-11 14:21:58 -07003579 ClearPagePrivate(new_page);
3580
David Gibson1e8f8892006-01-06 00:10:44 -08003581 /* Break COW */
Gerald Schaefer8fe627e2008-04-28 02:13:28 -07003582 huge_ptep_clear_flush(vma, address, ptep);
Joerg Roedel34ee6452014-11-13 13:46:09 +11003583 mmu_notifier_invalidate_range(mm, mmun_start, mmun_end);
David Gibson1e8f8892006-01-06 00:10:44 -08003584 set_huge_pte_at(mm, address, ptep,
3585 make_huge_pte(vma, new_page, 1));
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -08003586 page_remove_rmap(old_page, true);
Naoya Horiguchicd67f0d2010-09-10 13:23:04 +09003587 hugepage_add_new_anon_rmap(new_page, vma, address);
David Gibson1e8f8892006-01-06 00:10:44 -08003588 /* Make the old page be freed below */
3589 new_page = old_page;
3590 }
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003591 spin_unlock(ptl);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07003592 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
Davidlohr Buesoad4404a2014-08-06 16:06:47 -07003593out_release_all:
Mike Kravetz96b96a92016-11-10 10:46:32 -08003594 restore_reserve_on_error(h, vma, address, new_page);
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03003595 put_page(new_page);
Davidlohr Buesoad4404a2014-08-06 16:06:47 -07003596out_release_old:
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03003597 put_page(old_page);
Joonsoo Kim83120342013-09-11 14:21:57 -07003598
Davidlohr Buesoad4404a2014-08-06 16:06:47 -07003599 spin_lock(ptl); /* Caller expects lock to be held */
3600 return ret;
David Gibson1e8f8892006-01-06 00:10:44 -08003601}
3602
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003603/* Return the pagecache page at a given address within a VMA */
Andi Kleena5516432008-07-23 21:27:41 -07003604static struct page *hugetlbfs_pagecache_page(struct hstate *h,
3605 struct vm_area_struct *vma, unsigned long address)
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003606{
3607 struct address_space *mapping;
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -07003608 pgoff_t idx;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003609
3610 mapping = vma->vm_file->f_mapping;
Andi Kleena5516432008-07-23 21:27:41 -07003611 idx = vma_hugecache_offset(h, vma, address);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003612
3613 return find_lock_page(mapping, idx);
3614}
3615
Hugh Dickins3ae77f42009-09-21 17:03:33 -07003616/*
3617 * Return whether there is a pagecache page to back given address within VMA.
3618 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
3619 */
3620static bool hugetlbfs_pagecache_present(struct hstate *h,
Hugh Dickins2a15efc2009-09-21 17:03:27 -07003621 struct vm_area_struct *vma, unsigned long address)
3622{
3623 struct address_space *mapping;
3624 pgoff_t idx;
3625 struct page *page;
3626
3627 mapping = vma->vm_file->f_mapping;
3628 idx = vma_hugecache_offset(h, vma, address);
3629
3630 page = find_get_page(mapping, idx);
3631 if (page)
3632 put_page(page);
3633 return page != NULL;
3634}
3635
Mike Kravetzab76ad52015-09-08 15:01:50 -07003636int huge_add_to_page_cache(struct page *page, struct address_space *mapping,
3637 pgoff_t idx)
3638{
3639 struct inode *inode = mapping->host;
3640 struct hstate *h = hstate_inode(inode);
3641 int err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
3642
3643 if (err)
3644 return err;
3645 ClearPagePrivate(page);
3646
3647 spin_lock(&inode->i_lock);
3648 inode->i_blocks += blocks_per_huge_page(h);
3649 spin_unlock(&inode->i_lock);
3650 return 0;
3651}
3652
Robert P. J. Daya1ed3dd2007-07-17 04:03:33 -07003653static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
Davidlohr Bueso8382d912014-04-03 14:47:31 -07003654 struct address_space *mapping, pgoff_t idx,
3655 unsigned long address, pte_t *ptep, unsigned int flags)
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01003656{
Andi Kleena5516432008-07-23 21:27:41 -07003657 struct hstate *h = hstate_vma(vma);
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01003658 int ret = VM_FAULT_SIGBUS;
Hillf Danton409eb8c2012-01-20 14:34:13 -08003659 int anon_rmap = 0;
Adam Litke4c887262005-10-29 18:16:46 -07003660 unsigned long size;
Adam Litke4c887262005-10-29 18:16:46 -07003661 struct page *page;
David Gibson1e8f8892006-01-06 00:10:44 -08003662 pte_t new_pte;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003663 spinlock_t *ptl;
Adam Litke4c887262005-10-29 18:16:46 -07003664
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003665 /*
3666 * Currently, we are forced to kill the process in the event the
3667 * original mapper has unmapped pages from the child due to a failed
Lucas De Marchi25985ed2011-03-30 22:57:33 -03003668 * COW. Warn that such a situation has occurred as it may not be obvious
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003669 */
3670 if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
Geoffrey Thomas910154d2016-03-09 14:08:04 -08003671 pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n",
Andrew Mortonffb22af2013-02-22 16:32:08 -08003672 current->pid);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003673 return ret;
3674 }
3675
Adam Litke4c887262005-10-29 18:16:46 -07003676 /*
3677 * Use page lock to guard against racing truncation
3678 * before we get page_table_lock.
3679 */
Christoph Lameter6bda6662006-01-06 00:10:49 -08003680retry:
3681 page = find_lock_page(mapping, idx);
3682 if (!page) {
Andi Kleena5516432008-07-23 21:27:41 -07003683 size = i_size_read(mapping->host) >> huge_page_shift(h);
Hugh Dickinsebed4bf2006-10-28 10:38:43 -07003684 if (idx >= size)
3685 goto out;
Mike Kravetz1a1aad82017-02-22 15:43:01 -08003686
3687 /*
3688 * Check for page in userfault range
3689 */
3690 if (userfaultfd_missing(vma)) {
3691 u32 hash;
3692 struct vm_fault vmf = {
3693 .vma = vma,
3694 .address = address,
3695 .flags = flags,
3696 /*
3697 * Hard to debug if it ends up being
3698 * used by a callee that assumes
3699 * something about the other
3700 * uninitialized fields... same as in
3701 * memory.c
3702 */
3703 };
3704
3705 /*
3706 * hugetlb_fault_mutex must be dropped before
3707 * handling userfault. Reacquire after handling
3708 * fault to make calling code simpler.
3709 */
3710 hash = hugetlb_fault_mutex_hash(h, mm, vma, mapping,
3711 idx, address);
3712 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
3713 ret = handle_userfault(&vmf, VM_UFFD_MISSING);
3714 mutex_lock(&hugetlb_fault_mutex_table[hash]);
3715 goto out;
3716 }
3717
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003718 page = alloc_huge_page(vma, address, 0);
Adam Litke2fc39ce2007-11-14 16:59:39 -08003719 if (IS_ERR(page)) {
Aneesh Kumar K.V76dcee72012-07-31 16:41:57 -07003720 ret = PTR_ERR(page);
3721 if (ret == -ENOMEM)
3722 ret = VM_FAULT_OOM;
3723 else
3724 ret = VM_FAULT_SIGBUS;
Christoph Lameter6bda6662006-01-06 00:10:49 -08003725 goto out;
3726 }
Andrea Arcangeli47ad8472011-01-13 15:46:47 -08003727 clear_huge_page(page, address, pages_per_huge_page(h));
Nick Piggin0ed361d2008-02-04 22:29:34 -08003728 __SetPageUptodate(page);
Naoya Horiguchibcc54222015-04-15 16:14:38 -07003729 set_page_huge_active(page);
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01003730
Mel Gormanf83a2752009-05-28 14:34:40 -07003731 if (vma->vm_flags & VM_MAYSHARE) {
Mike Kravetzab76ad52015-09-08 15:01:50 -07003732 int err = huge_add_to_page_cache(page, mapping, idx);
Christoph Lameter6bda6662006-01-06 00:10:49 -08003733 if (err) {
3734 put_page(page);
Christoph Lameter6bda6662006-01-06 00:10:49 -08003735 if (err == -EEXIST)
3736 goto retry;
3737 goto out;
3738 }
Mel Gorman23be7462010-04-23 13:17:56 -04003739 } else {
Christoph Lameter6bda6662006-01-06 00:10:49 -08003740 lock_page(page);
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09003741 if (unlikely(anon_vma_prepare(vma))) {
3742 ret = VM_FAULT_OOM;
3743 goto backout_unlocked;
3744 }
Hillf Danton409eb8c2012-01-20 14:34:13 -08003745 anon_rmap = 1;
Mel Gorman23be7462010-04-23 13:17:56 -04003746 }
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09003747 } else {
Naoya Horiguchi998b4382010-09-08 10:19:32 +09003748 /*
3749 * If memory error occurs between mmap() and fault, some process
3750 * don't have hwpoisoned swap entry for errored virtual address.
3751 * So we need to block hugepage fault by PG_hwpoison bit check.
3752 */
3753 if (unlikely(PageHWPoison(page))) {
Chris Forbes32f84522011-07-25 17:12:14 -07003754 ret = VM_FAULT_HWPOISON |
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07003755 VM_FAULT_SET_HINDEX(hstate_index(h));
Naoya Horiguchi998b4382010-09-08 10:19:32 +09003756 goto backout_unlocked;
3757 }
Christoph Lameter6bda6662006-01-06 00:10:49 -08003758 }
David Gibson1e8f8892006-01-06 00:10:44 -08003759
Andy Whitcroft57303d82008-08-12 15:08:47 -07003760 /*
3761 * If we are going to COW a private mapping later, we examine the
3762 * pending reservations for this page now. This will ensure that
3763 * any allocations necessary to record that reservation occur outside
3764 * the spinlock.
3765 */
Mike Kravetz5e911372015-09-08 15:01:28 -07003766 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
Andy Whitcroft2b267362008-08-12 15:08:49 -07003767 if (vma_needs_reservation(h, vma, address) < 0) {
3768 ret = VM_FAULT_OOM;
3769 goto backout_unlocked;
3770 }
Mike Kravetz5e911372015-09-08 15:01:28 -07003771 /* Just decrements count, does not deallocate */
Mike Kravetzfeba16e2015-09-08 15:01:31 -07003772 vma_end_reservation(h, vma, address);
Mike Kravetz5e911372015-09-08 15:01:28 -07003773 }
Andy Whitcroft57303d82008-08-12 15:08:47 -07003774
Aneesh Kumar K.V8bea8052016-12-12 16:41:59 -08003775 ptl = huge_pte_lock(h, mm, ptep);
Andi Kleena5516432008-07-23 21:27:41 -07003776 size = i_size_read(mapping->host) >> huge_page_shift(h);
Adam Litke4c887262005-10-29 18:16:46 -07003777 if (idx >= size)
3778 goto backout;
3779
Nick Piggin83c54072007-07-19 01:47:05 -07003780 ret = 0;
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07003781 if (!huge_pte_none(huge_ptep_get(ptep)))
Adam Litke4c887262005-10-29 18:16:46 -07003782 goto backout;
3783
Joonsoo Kim07443a82013-09-11 14:21:58 -07003784 if (anon_rmap) {
3785 ClearPagePrivate(page);
Hillf Danton409eb8c2012-01-20 14:34:13 -08003786 hugepage_add_new_anon_rmap(page, vma, address);
Choi Gi-yongac714902014-04-07 15:37:36 -07003787 } else
Kirill A. Shutemov53f92632016-01-15 16:53:42 -08003788 page_dup_rmap(page, true);
David Gibson1e8f8892006-01-06 00:10:44 -08003789 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
3790 && (vma->vm_flags & VM_SHARED)));
3791 set_huge_pte_at(mm, address, ptep, new_pte);
3792
Naoya Horiguchi5d317b22015-11-05 18:47:14 -08003793 hugetlb_count_add(pages_per_huge_page(h), mm);
Hugh Dickins788c7df2009-06-23 13:49:05 +01003794 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
David Gibson1e8f8892006-01-06 00:10:44 -08003795 /* Optimization, do the COW without a second fault */
Aneesh Kumar K.V3999f522016-12-12 16:41:56 -08003796 ret = hugetlb_cow(mm, vma, address, ptep, page, ptl);
David Gibson1e8f8892006-01-06 00:10:44 -08003797 }
3798
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003799 spin_unlock(ptl);
Adam Litke4c887262005-10-29 18:16:46 -07003800 unlock_page(page);
3801out:
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01003802 return ret;
Adam Litke4c887262005-10-29 18:16:46 -07003803
3804backout:
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003805 spin_unlock(ptl);
Andy Whitcroft2b267362008-08-12 15:08:49 -07003806backout_unlocked:
Adam Litke4c887262005-10-29 18:16:46 -07003807 unlock_page(page);
Mike Kravetz96b96a92016-11-10 10:46:32 -08003808 restore_reserve_on_error(h, vma, address, page);
Adam Litke4c887262005-10-29 18:16:46 -07003809 put_page(page);
3810 goto out;
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01003811}
3812
Davidlohr Bueso8382d912014-04-03 14:47:31 -07003813#ifdef CONFIG_SMP
Mike Kravetzc672c7f2015-09-08 15:01:35 -07003814u32 hugetlb_fault_mutex_hash(struct hstate *h, struct mm_struct *mm,
Davidlohr Bueso8382d912014-04-03 14:47:31 -07003815 struct vm_area_struct *vma,
3816 struct address_space *mapping,
3817 pgoff_t idx, unsigned long address)
3818{
3819 unsigned long key[2];
3820 u32 hash;
3821
3822 if (vma->vm_flags & VM_SHARED) {
3823 key[0] = (unsigned long) mapping;
3824 key[1] = idx;
3825 } else {
3826 key[0] = (unsigned long) mm;
3827 key[1] = address >> huge_page_shift(h);
3828 }
3829
3830 hash = jhash2((u32 *)&key, sizeof(key)/sizeof(u32), 0);
3831
3832 return hash & (num_fault_mutexes - 1);
3833}
3834#else
3835/*
3836 * For uniprocesor systems we always use a single mutex, so just
3837 * return 0 and avoid the hashing overhead.
3838 */
Mike Kravetzc672c7f2015-09-08 15:01:35 -07003839u32 hugetlb_fault_mutex_hash(struct hstate *h, struct mm_struct *mm,
Davidlohr Bueso8382d912014-04-03 14:47:31 -07003840 struct vm_area_struct *vma,
3841 struct address_space *mapping,
3842 pgoff_t idx, unsigned long address)
3843{
3844 return 0;
3845}
3846#endif
3847
Adam Litke86e52162006-01-06 00:10:43 -08003848int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
Hugh Dickins788c7df2009-06-23 13:49:05 +01003849 unsigned long address, unsigned int flags)
Adam Litke86e52162006-01-06 00:10:43 -08003850{
Davidlohr Bueso8382d912014-04-03 14:47:31 -07003851 pte_t *ptep, entry;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003852 spinlock_t *ptl;
David Gibson1e8f8892006-01-06 00:10:44 -08003853 int ret;
Davidlohr Bueso8382d912014-04-03 14:47:31 -07003854 u32 hash;
3855 pgoff_t idx;
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09003856 struct page *page = NULL;
Andy Whitcroft57303d82008-08-12 15:08:47 -07003857 struct page *pagecache_page = NULL;
Andi Kleena5516432008-07-23 21:27:41 -07003858 struct hstate *h = hstate_vma(vma);
Davidlohr Bueso8382d912014-04-03 14:47:31 -07003859 struct address_space *mapping;
Naoya Horiguchi0f792cf2015-02-11 15:25:25 -08003860 int need_wait_lock = 0;
Adam Litke86e52162006-01-06 00:10:43 -08003861
KAMEZAWA Hiroyuki1e16a5392012-01-10 15:07:22 -08003862 address &= huge_page_mask(h);
3863
Punit Agrawal7868a202017-07-06 15:39:42 -07003864 ptep = huge_pte_offset(mm, address, huge_page_size(h));
Naoya Horiguchifd6a03e2010-05-28 09:29:21 +09003865 if (ptep) {
3866 entry = huge_ptep_get(ptep);
Naoya Horiguchi290408d2010-09-08 10:19:35 +09003867 if (unlikely(is_hugetlb_entry_migration(entry))) {
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003868 migration_entry_wait_huge(vma, mm, ptep);
Naoya Horiguchi290408d2010-09-08 10:19:35 +09003869 return 0;
3870 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
Chris Forbes32f84522011-07-25 17:12:14 -07003871 return VM_FAULT_HWPOISON_LARGE |
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07003872 VM_FAULT_SET_HINDEX(hstate_index(h));
Naoya Horiguchi0d777df2015-12-11 13:40:49 -08003873 } else {
3874 ptep = huge_pte_alloc(mm, address, huge_page_size(h));
3875 if (!ptep)
3876 return VM_FAULT_OOM;
Naoya Horiguchifd6a03e2010-05-28 09:29:21 +09003877 }
3878
Davidlohr Bueso8382d912014-04-03 14:47:31 -07003879 mapping = vma->vm_file->f_mapping;
3880 idx = vma_hugecache_offset(h, vma, address);
3881
David Gibson3935baa2006-03-22 00:08:53 -08003882 /*
3883 * Serialize hugepage allocation and instantiation, so that we don't
3884 * get spurious allocation failures if two CPUs race to instantiate
3885 * the same page in the page cache.
3886 */
Mike Kravetzc672c7f2015-09-08 15:01:35 -07003887 hash = hugetlb_fault_mutex_hash(h, mm, vma, mapping, idx, address);
3888 mutex_lock(&hugetlb_fault_mutex_table[hash]);
Davidlohr Bueso8382d912014-04-03 14:47:31 -07003889
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07003890 entry = huge_ptep_get(ptep);
3891 if (huge_pte_none(entry)) {
Davidlohr Bueso8382d912014-04-03 14:47:31 -07003892 ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags);
David Gibsonb4d1d992008-10-15 22:01:11 -07003893 goto out_mutex;
David Gibson3935baa2006-03-22 00:08:53 -08003894 }
Adam Litke86e52162006-01-06 00:10:43 -08003895
Nick Piggin83c54072007-07-19 01:47:05 -07003896 ret = 0;
David Gibson1e8f8892006-01-06 00:10:44 -08003897
Andy Whitcroft57303d82008-08-12 15:08:47 -07003898 /*
Naoya Horiguchi0f792cf2015-02-11 15:25:25 -08003899 * entry could be a migration/hwpoison entry at this point, so this
3900 * check prevents the kernel from going below assuming that we have
3901 * a active hugepage in pagecache. This goto expects the 2nd page fault,
3902 * and is_hugetlb_entry_(migration|hwpoisoned) check will properly
3903 * handle it.
3904 */
3905 if (!pte_present(entry))
3906 goto out_mutex;
3907
3908 /*
Andy Whitcroft57303d82008-08-12 15:08:47 -07003909 * If we are going to COW the mapping later, we examine the pending
3910 * reservations for this page now. This will ensure that any
3911 * allocations necessary to record that reservation occur outside the
3912 * spinlock. For private mappings, we also lookup the pagecache
3913 * page now as it is used to determine if a reservation has been
3914 * consumed.
3915 */
Gerald Schaefer106c9922013-04-29 15:07:23 -07003916 if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) {
Andy Whitcroft2b267362008-08-12 15:08:49 -07003917 if (vma_needs_reservation(h, vma, address) < 0) {
3918 ret = VM_FAULT_OOM;
David Gibsonb4d1d992008-10-15 22:01:11 -07003919 goto out_mutex;
Andy Whitcroft2b267362008-08-12 15:08:49 -07003920 }
Mike Kravetz5e911372015-09-08 15:01:28 -07003921 /* Just decrements count, does not deallocate */
Mike Kravetzfeba16e2015-09-08 15:01:31 -07003922 vma_end_reservation(h, vma, address);
Andy Whitcroft57303d82008-08-12 15:08:47 -07003923
Mel Gormanf83a2752009-05-28 14:34:40 -07003924 if (!(vma->vm_flags & VM_MAYSHARE))
Andy Whitcroft57303d82008-08-12 15:08:47 -07003925 pagecache_page = hugetlbfs_pagecache_page(h,
3926 vma, address);
3927 }
3928
Naoya Horiguchi0f792cf2015-02-11 15:25:25 -08003929 ptl = huge_pte_lock(h, mm, ptep);
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09003930
David Gibson1e8f8892006-01-06 00:10:44 -08003931 /* Check for a racing update before calling hugetlb_cow */
David Gibsonb4d1d992008-10-15 22:01:11 -07003932 if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003933 goto out_ptl;
David Gibsonb4d1d992008-10-15 22:01:11 -07003934
Naoya Horiguchi0f792cf2015-02-11 15:25:25 -08003935 /*
3936 * hugetlb_cow() requires page locks of pte_page(entry) and
3937 * pagecache_page, so here we need take the former one
3938 * when page != pagecache_page or !pagecache_page.
3939 */
3940 page = pte_page(entry);
3941 if (page != pagecache_page)
3942 if (!trylock_page(page)) {
3943 need_wait_lock = 1;
3944 goto out_ptl;
3945 }
3946
3947 get_page(page);
David Gibsonb4d1d992008-10-15 22:01:11 -07003948
Hugh Dickins788c7df2009-06-23 13:49:05 +01003949 if (flags & FAULT_FLAG_WRITE) {
Gerald Schaefer106c9922013-04-29 15:07:23 -07003950 if (!huge_pte_write(entry)) {
Aneesh Kumar K.V3999f522016-12-12 16:41:56 -08003951 ret = hugetlb_cow(mm, vma, address, ptep,
3952 pagecache_page, ptl);
Naoya Horiguchi0f792cf2015-02-11 15:25:25 -08003953 goto out_put_page;
David Gibsonb4d1d992008-10-15 22:01:11 -07003954 }
Gerald Schaefer106c9922013-04-29 15:07:23 -07003955 entry = huge_pte_mkdirty(entry);
David Gibsonb4d1d992008-10-15 22:01:11 -07003956 }
3957 entry = pte_mkyoung(entry);
Hugh Dickins788c7df2009-06-23 13:49:05 +01003958 if (huge_ptep_set_access_flags(vma, address, ptep, entry,
3959 flags & FAULT_FLAG_WRITE))
Russell King4b3073e2009-12-18 16:40:18 +00003960 update_mmu_cache(vma, address, ptep);
Naoya Horiguchi0f792cf2015-02-11 15:25:25 -08003961out_put_page:
3962 if (page != pagecache_page)
3963 unlock_page(page);
3964 put_page(page);
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003965out_ptl:
3966 spin_unlock(ptl);
Andy Whitcroft57303d82008-08-12 15:08:47 -07003967
3968 if (pagecache_page) {
3969 unlock_page(pagecache_page);
3970 put_page(pagecache_page);
3971 }
David Gibsonb4d1d992008-10-15 22:01:11 -07003972out_mutex:
Mike Kravetzc672c7f2015-09-08 15:01:35 -07003973 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
Naoya Horiguchi0f792cf2015-02-11 15:25:25 -08003974 /*
3975 * Generally it's safe to hold refcount during waiting page lock. But
3976 * here we just wait to defer the next page fault to avoid busy loop and
3977 * the page is not used after unlocked before returning from the current
3978 * page fault. So we are safe from accessing freed page, even if we wait
3979 * here without taking refcount.
3980 */
3981 if (need_wait_lock)
3982 wait_on_page_locked(page);
David Gibson1e8f8892006-01-06 00:10:44 -08003983 return ret;
Adam Litke86e52162006-01-06 00:10:43 -08003984}
3985
Mike Kravetz8fb5deb2017-02-22 15:42:52 -08003986/*
3987 * Used by userfaultfd UFFDIO_COPY. Based on mcopy_atomic_pte with
3988 * modifications for huge pages.
3989 */
3990int hugetlb_mcopy_atomic_pte(struct mm_struct *dst_mm,
3991 pte_t *dst_pte,
3992 struct vm_area_struct *dst_vma,
3993 unsigned long dst_addr,
3994 unsigned long src_addr,
3995 struct page **pagep)
3996{
Andrea Arcangeli1e3921472017-11-02 15:59:29 -07003997 struct address_space *mapping;
3998 pgoff_t idx;
3999 unsigned long size;
Mike Kravetz1c9e8de2017-02-22 15:43:43 -08004000 int vm_shared = dst_vma->vm_flags & VM_SHARED;
Mike Kravetz8fb5deb2017-02-22 15:42:52 -08004001 struct hstate *h = hstate_vma(dst_vma);
4002 pte_t _dst_pte;
4003 spinlock_t *ptl;
4004 int ret;
4005 struct page *page;
4006
4007 if (!*pagep) {
4008 ret = -ENOMEM;
4009 page = alloc_huge_page(dst_vma, dst_addr, 0);
4010 if (IS_ERR(page))
4011 goto out;
4012
4013 ret = copy_huge_page_from_user(page,
4014 (const void __user *) src_addr,
Mike Kravetz810a56b2017-02-22 15:42:58 -08004015 pages_per_huge_page(h), false);
Mike Kravetz8fb5deb2017-02-22 15:42:52 -08004016
4017 /* fallback to copy_from_user outside mmap_sem */
4018 if (unlikely(ret)) {
4019 ret = -EFAULT;
4020 *pagep = page;
4021 /* don't free the page */
4022 goto out;
4023 }
4024 } else {
4025 page = *pagep;
4026 *pagep = NULL;
4027 }
4028
4029 /*
4030 * The memory barrier inside __SetPageUptodate makes sure that
4031 * preceding stores to the page contents become visible before
4032 * the set_pte_at() write.
4033 */
4034 __SetPageUptodate(page);
4035 set_page_huge_active(page);
4036
Andrea Arcangeli1e3921472017-11-02 15:59:29 -07004037 mapping = dst_vma->vm_file->f_mapping;
4038 idx = vma_hugecache_offset(h, dst_vma, dst_addr);
4039
Mike Kravetz1c9e8de2017-02-22 15:43:43 -08004040 /*
4041 * If shared, add to page cache
4042 */
4043 if (vm_shared) {
Andrea Arcangeli1e3921472017-11-02 15:59:29 -07004044 size = i_size_read(mapping->host) >> huge_page_shift(h);
4045 ret = -EFAULT;
4046 if (idx >= size)
4047 goto out_release_nounlock;
Mike Kravetz1c9e8de2017-02-22 15:43:43 -08004048
Andrea Arcangeli1e3921472017-11-02 15:59:29 -07004049 /*
4050 * Serialization between remove_inode_hugepages() and
4051 * huge_add_to_page_cache() below happens through the
4052 * hugetlb_fault_mutex_table that here must be hold by
4053 * the caller.
4054 */
Mike Kravetz1c9e8de2017-02-22 15:43:43 -08004055 ret = huge_add_to_page_cache(page, mapping, idx);
4056 if (ret)
4057 goto out_release_nounlock;
4058 }
4059
Mike Kravetz8fb5deb2017-02-22 15:42:52 -08004060 ptl = huge_pte_lockptr(h, dst_mm, dst_pte);
4061 spin_lock(ptl);
4062
Andrea Arcangeli1e3921472017-11-02 15:59:29 -07004063 /*
4064 * Recheck the i_size after holding PT lock to make sure not
4065 * to leave any page mapped (as page_mapped()) beyond the end
4066 * of the i_size (remove_inode_hugepages() is strict about
4067 * enforcing that). If we bail out here, we'll also leave a
4068 * page in the radix tree in the vm_shared case beyond the end
4069 * of the i_size, but remove_inode_hugepages() will take care
4070 * of it as soon as we drop the hugetlb_fault_mutex_table.
4071 */
4072 size = i_size_read(mapping->host) >> huge_page_shift(h);
4073 ret = -EFAULT;
4074 if (idx >= size)
4075 goto out_release_unlock;
4076
Mike Kravetz8fb5deb2017-02-22 15:42:52 -08004077 ret = -EEXIST;
4078 if (!huge_pte_none(huge_ptep_get(dst_pte)))
4079 goto out_release_unlock;
4080
Mike Kravetz1c9e8de2017-02-22 15:43:43 -08004081 if (vm_shared) {
4082 page_dup_rmap(page, true);
4083 } else {
4084 ClearPagePrivate(page);
4085 hugepage_add_new_anon_rmap(page, dst_vma, dst_addr);
4086 }
Mike Kravetz8fb5deb2017-02-22 15:42:52 -08004087
4088 _dst_pte = make_huge_pte(dst_vma, page, dst_vma->vm_flags & VM_WRITE);
4089 if (dst_vma->vm_flags & VM_WRITE)
4090 _dst_pte = huge_pte_mkdirty(_dst_pte);
4091 _dst_pte = pte_mkyoung(_dst_pte);
4092
4093 set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte);
4094
4095 (void)huge_ptep_set_access_flags(dst_vma, dst_addr, dst_pte, _dst_pte,
4096 dst_vma->vm_flags & VM_WRITE);
4097 hugetlb_count_add(pages_per_huge_page(h), dst_mm);
4098
4099 /* No need to invalidate - it was non-present before */
4100 update_mmu_cache(dst_vma, dst_addr, dst_pte);
4101
4102 spin_unlock(ptl);
Mike Kravetz1c9e8de2017-02-22 15:43:43 -08004103 if (vm_shared)
4104 unlock_page(page);
Mike Kravetz8fb5deb2017-02-22 15:42:52 -08004105 ret = 0;
4106out:
4107 return ret;
4108out_release_unlock:
4109 spin_unlock(ptl);
Mike Kravetz1c9e8de2017-02-22 15:43:43 -08004110 if (vm_shared)
4111 unlock_page(page);
Andrea Arcangeli5af10df2017-08-10 15:23:38 -07004112out_release_nounlock:
Mike Kravetz8fb5deb2017-02-22 15:42:52 -08004113 put_page(page);
4114 goto out;
4115}
4116
Michel Lespinasse28a35712013-02-22 16:35:55 -08004117long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
4118 struct page **pages, struct vm_area_struct **vmas,
4119 unsigned long *position, unsigned long *nr_pages,
Andrea Arcangeli87ffc112017-02-22 15:43:13 -08004120 long i, unsigned int flags, int *nonblocking)
David Gibson63551ae2005-06-21 17:14:44 -07004121{
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08004122 unsigned long pfn_offset;
4123 unsigned long vaddr = *position;
Michel Lespinasse28a35712013-02-22 16:35:55 -08004124 unsigned long remainder = *nr_pages;
Andi Kleena5516432008-07-23 21:27:41 -07004125 struct hstate *h = hstate_vma(vma);
Daniel Jordan2be7cfe2017-08-02 13:31:47 -07004126 int err = -EFAULT;
David Gibson63551ae2005-06-21 17:14:44 -07004127
David Gibson63551ae2005-06-21 17:14:44 -07004128 while (vaddr < vma->vm_end && remainder) {
Adam Litke4c887262005-10-29 18:16:46 -07004129 pte_t *pte;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08004130 spinlock_t *ptl = NULL;
Hugh Dickins2a15efc2009-09-21 17:03:27 -07004131 int absent;
Adam Litke4c887262005-10-29 18:16:46 -07004132 struct page *page;
4133
4134 /*
David Rientjes02057962015-04-14 15:48:24 -07004135 * If we have a pending SIGKILL, don't keep faulting pages and
4136 * potentially allocating memory.
4137 */
4138 if (unlikely(fatal_signal_pending(current))) {
4139 remainder = 0;
4140 break;
4141 }
4142
4143 /*
Adam Litke4c887262005-10-29 18:16:46 -07004144 * Some archs (sparc64, sh*) have multiple pte_ts to
Hugh Dickins2a15efc2009-09-21 17:03:27 -07004145 * each hugepage. We have to make sure we get the
Adam Litke4c887262005-10-29 18:16:46 -07004146 * first, for the page indexing below to work.
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08004147 *
4148 * Note that page table lock is not held when pte is null.
Adam Litke4c887262005-10-29 18:16:46 -07004149 */
Punit Agrawal7868a202017-07-06 15:39:42 -07004150 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h),
4151 huge_page_size(h));
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08004152 if (pte)
4153 ptl = huge_pte_lock(h, mm, pte);
Hugh Dickins2a15efc2009-09-21 17:03:27 -07004154 absent = !pte || huge_pte_none(huge_ptep_get(pte));
Adam Litke4c887262005-10-29 18:16:46 -07004155
Hugh Dickins2a15efc2009-09-21 17:03:27 -07004156 /*
4157 * When coredumping, it suits get_dump_page if we just return
Hugh Dickins3ae77f42009-09-21 17:03:33 -07004158 * an error where there's an empty slot with no huge pagecache
4159 * to back it. This way, we avoid allocating a hugepage, and
4160 * the sparse dumpfile avoids allocating disk blocks, but its
4161 * huge holes still show up with zeroes where they need to be.
Hugh Dickins2a15efc2009-09-21 17:03:27 -07004162 */
Hugh Dickins3ae77f42009-09-21 17:03:33 -07004163 if (absent && (flags & FOLL_DUMP) &&
4164 !hugetlbfs_pagecache_present(h, vma, vaddr)) {
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08004165 if (pte)
4166 spin_unlock(ptl);
Hugh Dickins2a15efc2009-09-21 17:03:27 -07004167 remainder = 0;
4168 break;
4169 }
4170
Naoya Horiguchi9cc3a5b2013-04-17 15:58:30 -07004171 /*
4172 * We need call hugetlb_fault for both hugepages under migration
4173 * (in which case hugetlb_fault waits for the migration,) and
4174 * hwpoisoned hugepages (in which case we need to prevent the
4175 * caller from accessing to them.) In order to do this, we use
4176 * here is_swap_pte instead of is_hugetlb_entry_migration and
4177 * is_hugetlb_entry_hwpoisoned. This is because it simply covers
4178 * both cases, and because we can't follow correct pages
4179 * directly from any kind of swap entries.
4180 */
4181 if (absent || is_swap_pte(huge_ptep_get(pte)) ||
Gerald Schaefer106c9922013-04-29 15:07:23 -07004182 ((flags & FOLL_WRITE) &&
4183 !huge_pte_write(huge_ptep_get(pte)))) {
Adam Litke4c887262005-10-29 18:16:46 -07004184 int ret;
Andrea Arcangeli87ffc112017-02-22 15:43:13 -08004185 unsigned int fault_flags = 0;
Adam Litke4c887262005-10-29 18:16:46 -07004186
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08004187 if (pte)
4188 spin_unlock(ptl);
Andrea Arcangeli87ffc112017-02-22 15:43:13 -08004189 if (flags & FOLL_WRITE)
4190 fault_flags |= FAULT_FLAG_WRITE;
4191 if (nonblocking)
4192 fault_flags |= FAULT_FLAG_ALLOW_RETRY;
4193 if (flags & FOLL_NOWAIT)
4194 fault_flags |= FAULT_FLAG_ALLOW_RETRY |
4195 FAULT_FLAG_RETRY_NOWAIT;
4196 if (flags & FOLL_TRIED) {
4197 VM_WARN_ON_ONCE(fault_flags &
4198 FAULT_FLAG_ALLOW_RETRY);
4199 fault_flags |= FAULT_FLAG_TRIED;
4200 }
4201 ret = hugetlb_fault(mm, vma, vaddr, fault_flags);
4202 if (ret & VM_FAULT_ERROR) {
Daniel Jordan2be7cfe2017-08-02 13:31:47 -07004203 err = vm_fault_to_errno(ret, flags);
Andrea Arcangeli87ffc112017-02-22 15:43:13 -08004204 remainder = 0;
4205 break;
4206 }
4207 if (ret & VM_FAULT_RETRY) {
4208 if (nonblocking)
4209 *nonblocking = 0;
4210 *nr_pages = 0;
4211 /*
4212 * VM_FAULT_RETRY must not return an
4213 * error, it will return zero
4214 * instead.
4215 *
4216 * No need to update "position" as the
4217 * caller will not check it after
4218 * *nr_pages is set to 0.
4219 */
4220 return i;
4221 }
4222 continue;
Adam Litke4c887262005-10-29 18:16:46 -07004223 }
David Gibson63551ae2005-06-21 17:14:44 -07004224
Andi Kleena5516432008-07-23 21:27:41 -07004225 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07004226 page = pte_page(huge_ptep_get(pte));
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08004227same_page:
Chen, Kenneth Wd6692182006-03-31 02:29:57 -08004228 if (pages) {
Hugh Dickins2a15efc2009-09-21 17:03:27 -07004229 pages[i] = mem_map_offset(page, pfn_offset);
Kirill A. Shutemovddc58f22016-01-15 16:52:56 -08004230 get_page(pages[i]);
Chen, Kenneth Wd6692182006-03-31 02:29:57 -08004231 }
David Gibson63551ae2005-06-21 17:14:44 -07004232
4233 if (vmas)
4234 vmas[i] = vma;
4235
4236 vaddr += PAGE_SIZE;
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08004237 ++pfn_offset;
David Gibson63551ae2005-06-21 17:14:44 -07004238 --remainder;
4239 ++i;
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08004240 if (vaddr < vma->vm_end && remainder &&
Andi Kleena5516432008-07-23 21:27:41 -07004241 pfn_offset < pages_per_huge_page(h)) {
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08004242 /*
4243 * We use pfn_offset to avoid touching the pageframes
4244 * of this compound page.
4245 */
4246 goto same_page;
4247 }
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08004248 spin_unlock(ptl);
David Gibson63551ae2005-06-21 17:14:44 -07004249 }
Michel Lespinasse28a35712013-02-22 16:35:55 -08004250 *nr_pages = remainder;
Andrea Arcangeli87ffc112017-02-22 15:43:13 -08004251 /*
4252 * setting position is actually required only if remainder is
4253 * not zero but it's faster not to add a "if (remainder)"
4254 * branch.
4255 */
David Gibson63551ae2005-06-21 17:14:44 -07004256 *position = vaddr;
4257
Daniel Jordan2be7cfe2017-08-02 13:31:47 -07004258 return i ? i : err;
David Gibson63551ae2005-06-21 17:14:44 -07004259}
Zhang, Yanmin8f860592006-03-22 00:08:50 -08004260
Aneesh Kumar K.V5491ae72016-07-13 15:06:43 +05304261#ifndef __HAVE_ARCH_FLUSH_HUGETLB_TLB_RANGE
4262/*
4263 * ARCHes with special requirements for evicting HUGETLB backing TLB entries can
4264 * implement this.
4265 */
4266#define flush_hugetlb_tlb_range(vma, addr, end) flush_tlb_range(vma, addr, end)
4267#endif
4268
Peter Zijlstra7da4d642012-11-19 03:14:23 +01004269unsigned long hugetlb_change_protection(struct vm_area_struct *vma,
Zhang, Yanmin8f860592006-03-22 00:08:50 -08004270 unsigned long address, unsigned long end, pgprot_t newprot)
4271{
4272 struct mm_struct *mm = vma->vm_mm;
4273 unsigned long start = address;
4274 pte_t *ptep;
4275 pte_t pte;
Andi Kleena5516432008-07-23 21:27:41 -07004276 struct hstate *h = hstate_vma(vma);
Peter Zijlstra7da4d642012-11-19 03:14:23 +01004277 unsigned long pages = 0;
Zhang, Yanmin8f860592006-03-22 00:08:50 -08004278
4279 BUG_ON(address >= end);
4280 flush_cache_range(vma, address, end);
4281
Rik van Riela5338092014-04-07 15:36:57 -07004282 mmu_notifier_invalidate_range_start(mm, start, end);
Davidlohr Bueso83cde9e2014-12-12 16:54:21 -08004283 i_mmap_lock_write(vma->vm_file->f_mapping);
Andi Kleena5516432008-07-23 21:27:41 -07004284 for (; address < end; address += huge_page_size(h)) {
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08004285 spinlock_t *ptl;
Punit Agrawal7868a202017-07-06 15:39:42 -07004286 ptep = huge_pte_offset(mm, address, huge_page_size(h));
Zhang, Yanmin8f860592006-03-22 00:08:50 -08004287 if (!ptep)
4288 continue;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08004289 ptl = huge_pte_lock(h, mm, ptep);
Peter Zijlstra7da4d642012-11-19 03:14:23 +01004290 if (huge_pmd_unshare(mm, &address, ptep)) {
4291 pages++;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08004292 spin_unlock(ptl);
Chen, Kenneth W39dde652006-12-06 20:32:03 -08004293 continue;
Peter Zijlstra7da4d642012-11-19 03:14:23 +01004294 }
Naoya Horiguchia8bda282015-02-11 15:25:28 -08004295 pte = huge_ptep_get(ptep);
4296 if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) {
4297 spin_unlock(ptl);
4298 continue;
4299 }
4300 if (unlikely(is_hugetlb_entry_migration(pte))) {
4301 swp_entry_t entry = pte_to_swp_entry(pte);
4302
4303 if (is_write_migration_entry(entry)) {
4304 pte_t newpte;
4305
4306 make_migration_entry_read(&entry);
4307 newpte = swp_entry_to_pte(entry);
Punit Agrawale5251fd2017-07-06 15:39:50 -07004308 set_huge_swap_pte_at(mm, address, ptep,
4309 newpte, huge_page_size(h));
Naoya Horiguchia8bda282015-02-11 15:25:28 -08004310 pages++;
4311 }
4312 spin_unlock(ptl);
4313 continue;
4314 }
4315 if (!huge_pte_none(pte)) {
Zhang, Yanmin8f860592006-03-22 00:08:50 -08004316 pte = huge_ptep_get_and_clear(mm, address, ptep);
Gerald Schaefer106c9922013-04-29 15:07:23 -07004317 pte = pte_mkhuge(huge_pte_modify(pte, newprot));
Tony Lube7517d2013-02-04 14:28:46 -08004318 pte = arch_make_huge_pte(pte, vma, NULL, 0);
Zhang, Yanmin8f860592006-03-22 00:08:50 -08004319 set_huge_pte_at(mm, address, ptep, pte);
Peter Zijlstra7da4d642012-11-19 03:14:23 +01004320 pages++;
Zhang, Yanmin8f860592006-03-22 00:08:50 -08004321 }
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08004322 spin_unlock(ptl);
Zhang, Yanmin8f860592006-03-22 00:08:50 -08004323 }
Mel Gormand8333522012-07-31 16:46:20 -07004324 /*
Davidlohr Buesoc8c06ef2014-12-12 16:54:24 -08004325 * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare
Mel Gormand8333522012-07-31 16:46:20 -07004326 * may have cleared our pud entry and done put_page on the page table:
Davidlohr Buesoc8c06ef2014-12-12 16:54:24 -08004327 * once we release i_mmap_rwsem, another task can do the final put_page
Mel Gormand8333522012-07-31 16:46:20 -07004328 * and that page table be reused and filled with junk.
4329 */
Aneesh Kumar K.V5491ae72016-07-13 15:06:43 +05304330 flush_hugetlb_tlb_range(vma, start, end);
Joerg Roedel34ee6452014-11-13 13:46:09 +11004331 mmu_notifier_invalidate_range(mm, start, end);
Davidlohr Bueso83cde9e2014-12-12 16:54:21 -08004332 i_mmap_unlock_write(vma->vm_file->f_mapping);
Rik van Riela5338092014-04-07 15:36:57 -07004333 mmu_notifier_invalidate_range_end(mm, start, end);
Peter Zijlstra7da4d642012-11-19 03:14:23 +01004334
4335 return pages << h->order;
Zhang, Yanmin8f860592006-03-22 00:08:50 -08004336}
4337
Mel Gormana1e78772008-07-23 21:27:23 -07004338int hugetlb_reserve_pages(struct inode *inode,
4339 long from, long to,
Mel Gorman5a6fe122009-02-10 14:02:27 +00004340 struct vm_area_struct *vma,
KOSAKI Motohiroca16d142011-05-26 19:16:19 +09004341 vm_flags_t vm_flags)
Adam Litkee4e574b2007-10-16 01:26:19 -07004342{
Mel Gorman17c9d122009-02-11 16:34:16 +00004343 long ret, chg;
Andi Kleena5516432008-07-23 21:27:41 -07004344 struct hstate *h = hstate_inode(inode);
David Gibson90481622012-03-21 16:34:12 -07004345 struct hugepage_subpool *spool = subpool_inode(inode);
Joonsoo Kim9119a412014-04-03 14:47:25 -07004346 struct resv_map *resv_map;
Mike Kravetz1c5ecae2015-04-15 16:13:39 -07004347 long gbl_reserve;
Adam Litkee4e574b2007-10-16 01:26:19 -07004348
Mike Kravetz1e862842018-03-22 16:17:13 -07004349 /* This should never happen */
4350 if (from > to) {
4351 VM_WARN(1, "%s called with a negative range\n", __func__);
4352 return -EINVAL;
4353 }
4354
Mel Gormana1e78772008-07-23 21:27:23 -07004355 /*
Mel Gorman17c9d122009-02-11 16:34:16 +00004356 * Only apply hugepage reservation if asked. At fault time, an
4357 * attempt will be made for VM_NORESERVE to allocate a page
David Gibson90481622012-03-21 16:34:12 -07004358 * without using reserves
Mel Gorman17c9d122009-02-11 16:34:16 +00004359 */
KOSAKI Motohiroca16d142011-05-26 19:16:19 +09004360 if (vm_flags & VM_NORESERVE)
Mel Gorman17c9d122009-02-11 16:34:16 +00004361 return 0;
4362
4363 /*
Mel Gormana1e78772008-07-23 21:27:23 -07004364 * Shared mappings base their reservation on the number of pages that
4365 * are already allocated on behalf of the file. Private mappings need
4366 * to reserve the full area even if read-only as mprotect() may be
4367 * called to make the mapping read-write. Assume !vma is a shm mapping
4368 */
Joonsoo Kim9119a412014-04-03 14:47:25 -07004369 if (!vma || vma->vm_flags & VM_MAYSHARE) {
Joonsoo Kim4e35f482014-04-03 14:47:30 -07004370 resv_map = inode_resv_map(inode);
Joonsoo Kim9119a412014-04-03 14:47:25 -07004371
Joonsoo Kim1406ec92014-04-03 14:47:26 -07004372 chg = region_chg(resv_map, from, to);
Joonsoo Kim9119a412014-04-03 14:47:25 -07004373
4374 } else {
4375 resv_map = resv_map_alloc();
Mel Gorman5a6fe122009-02-10 14:02:27 +00004376 if (!resv_map)
4377 return -ENOMEM;
4378
Mel Gorman17c9d122009-02-11 16:34:16 +00004379 chg = to - from;
4380
Mel Gorman5a6fe122009-02-10 14:02:27 +00004381 set_vma_resv_map(vma, resv_map);
4382 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
4383 }
4384
Dave Hansenc50ac052012-05-29 15:06:46 -07004385 if (chg < 0) {
4386 ret = chg;
4387 goto out_err;
4388 }
Mel Gorman17c9d122009-02-11 16:34:16 +00004389
Mike Kravetz1c5ecae2015-04-15 16:13:39 -07004390 /*
4391 * There must be enough pages in the subpool for the mapping. If
4392 * the subpool has a minimum size, there may be some global
4393 * reservations already in place (gbl_reserve).
4394 */
4395 gbl_reserve = hugepage_subpool_get_pages(spool, chg);
4396 if (gbl_reserve < 0) {
Dave Hansenc50ac052012-05-29 15:06:46 -07004397 ret = -ENOSPC;
4398 goto out_err;
4399 }
Mel Gorman17c9d122009-02-11 16:34:16 +00004400
4401 /*
4402 * Check enough hugepages are available for the reservation.
David Gibson90481622012-03-21 16:34:12 -07004403 * Hand the pages back to the subpool if there are not
Mel Gorman17c9d122009-02-11 16:34:16 +00004404 */
Mike Kravetz1c5ecae2015-04-15 16:13:39 -07004405 ret = hugetlb_acct_memory(h, gbl_reserve);
Mel Gorman17c9d122009-02-11 16:34:16 +00004406 if (ret < 0) {
Mike Kravetz1c5ecae2015-04-15 16:13:39 -07004407 /* put back original number of pages, chg */
4408 (void)hugepage_subpool_put_pages(spool, chg);
Dave Hansenc50ac052012-05-29 15:06:46 -07004409 goto out_err;
Mel Gorman17c9d122009-02-11 16:34:16 +00004410 }
4411
4412 /*
4413 * Account for the reservations made. Shared mappings record regions
4414 * that have reservations as they are shared by multiple VMAs.
4415 * When the last VMA disappears, the region map says how much
4416 * the reservation was and the page cache tells how much of
4417 * the reservation was consumed. Private mappings are per-VMA and
4418 * only the consumed reservations are tracked. When the VMA
4419 * disappears, the original reservation is the VMA size and the
4420 * consumed reservations are stored in the map. Hence, nothing
4421 * else has to be done for private mappings here
4422 */
Mike Kravetz33039672015-06-24 16:57:58 -07004423 if (!vma || vma->vm_flags & VM_MAYSHARE) {
4424 long add = region_add(resv_map, from, to);
4425
4426 if (unlikely(chg > add)) {
4427 /*
4428 * pages in this range were added to the reserve
4429 * map between region_chg and region_add. This
4430 * indicates a race with alloc_huge_page. Adjust
4431 * the subpool and reserve counts modified above
4432 * based on the difference.
4433 */
4434 long rsv_adjust;
4435
4436 rsv_adjust = hugepage_subpool_put_pages(spool,
4437 chg - add);
4438 hugetlb_acct_memory(h, -rsv_adjust);
4439 }
4440 }
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -07004441 return 0;
Dave Hansenc50ac052012-05-29 15:06:46 -07004442out_err:
Mike Kravetz5e911372015-09-08 15:01:28 -07004443 if (!vma || vma->vm_flags & VM_MAYSHARE)
Mike Kravetzff8c0c52017-03-31 15:12:07 -07004444 /* Don't call region_abort if region_chg failed */
4445 if (chg >= 0)
4446 region_abort(resv_map, from, to);
Joonsoo Kimf031dd22014-04-03 14:47:28 -07004447 if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
4448 kref_put(&resv_map->refs, resv_map_release);
Dave Hansenc50ac052012-05-29 15:06:46 -07004449 return ret;
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -07004450}
4451
Mike Kravetzb5cec282015-09-08 15:01:41 -07004452long hugetlb_unreserve_pages(struct inode *inode, long start, long end,
4453 long freed)
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -07004454{
Andi Kleena5516432008-07-23 21:27:41 -07004455 struct hstate *h = hstate_inode(inode);
Joonsoo Kim4e35f482014-04-03 14:47:30 -07004456 struct resv_map *resv_map = inode_resv_map(inode);
Joonsoo Kim9119a412014-04-03 14:47:25 -07004457 long chg = 0;
David Gibson90481622012-03-21 16:34:12 -07004458 struct hugepage_subpool *spool = subpool_inode(inode);
Mike Kravetz1c5ecae2015-04-15 16:13:39 -07004459 long gbl_reserve;
Ken Chen45c682a2007-11-14 16:59:44 -08004460
Mike Kravetzb5cec282015-09-08 15:01:41 -07004461 if (resv_map) {
4462 chg = region_del(resv_map, start, end);
4463 /*
4464 * region_del() can fail in the rare case where a region
4465 * must be split and another region descriptor can not be
4466 * allocated. If end == LONG_MAX, it will not fail.
4467 */
4468 if (chg < 0)
4469 return chg;
4470 }
4471
Ken Chen45c682a2007-11-14 16:59:44 -08004472 spin_lock(&inode->i_lock);
Eric Sandeene4c6f8b2009-07-29 15:02:16 -07004473 inode->i_blocks -= (blocks_per_huge_page(h) * freed);
Ken Chen45c682a2007-11-14 16:59:44 -08004474 spin_unlock(&inode->i_lock);
4475
Mike Kravetz1c5ecae2015-04-15 16:13:39 -07004476 /*
4477 * If the subpool has a minimum size, the number of global
4478 * reservations to be released may be adjusted.
4479 */
4480 gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed));
4481 hugetlb_acct_memory(h, -gbl_reserve);
Mike Kravetzb5cec282015-09-08 15:01:41 -07004482
4483 return 0;
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -07004484}
Naoya Horiguchi93f70f92010-05-28 09:29:20 +09004485
Steve Capper3212b532013-04-23 12:35:02 +01004486#ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
4487static unsigned long page_table_shareable(struct vm_area_struct *svma,
4488 struct vm_area_struct *vma,
4489 unsigned long addr, pgoff_t idx)
4490{
4491 unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) +
4492 svma->vm_start;
4493 unsigned long sbase = saddr & PUD_MASK;
4494 unsigned long s_end = sbase + PUD_SIZE;
4495
4496 /* Allow segments to share if only one is marked locked */
Eric B Munsonde60f5f2015-11-05 18:51:36 -08004497 unsigned long vm_flags = vma->vm_flags & VM_LOCKED_CLEAR_MASK;
4498 unsigned long svm_flags = svma->vm_flags & VM_LOCKED_CLEAR_MASK;
Steve Capper3212b532013-04-23 12:35:02 +01004499
4500 /*
4501 * match the virtual addresses, permission and the alignment of the
4502 * page table page.
4503 */
4504 if (pmd_index(addr) != pmd_index(saddr) ||
4505 vm_flags != svm_flags ||
4506 sbase < svma->vm_start || svma->vm_end < s_end)
4507 return 0;
4508
4509 return saddr;
4510}
4511
Nicholas Krause31aafb42015-09-04 15:47:58 -07004512static bool vma_shareable(struct vm_area_struct *vma, unsigned long addr)
Steve Capper3212b532013-04-23 12:35:02 +01004513{
4514 unsigned long base = addr & PUD_MASK;
4515 unsigned long end = base + PUD_SIZE;
4516
4517 /*
4518 * check on proper vm_flags and page table alignment
4519 */
Mike Kravetz5f4f5b12018-10-05 15:51:29 -07004520 if (vma->vm_flags & VM_MAYSHARE && range_in_vma(vma, base, end))
Nicholas Krause31aafb42015-09-04 15:47:58 -07004521 return true;
4522 return false;
Steve Capper3212b532013-04-23 12:35:02 +01004523}
4524
4525/*
Mike Kravetz5f4f5b12018-10-05 15:51:29 -07004526 * Determine if start,end range within vma could be mapped by shared pmd.
4527 * If yes, adjust start and end to cover range associated with possible
4528 * shared pmd mappings.
4529 */
4530void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma,
4531 unsigned long *start, unsigned long *end)
4532{
4533 unsigned long check_addr = *start;
4534
4535 if (!(vma->vm_flags & VM_MAYSHARE))
4536 return;
4537
4538 for (check_addr = *start; check_addr < *end; check_addr += PUD_SIZE) {
4539 unsigned long a_start = check_addr & PUD_MASK;
4540 unsigned long a_end = a_start + PUD_SIZE;
4541
4542 /*
4543 * If sharing is possible, adjust start/end if necessary.
4544 */
4545 if (range_in_vma(vma, a_start, a_end)) {
4546 if (a_start < *start)
4547 *start = a_start;
4548 if (a_end > *end)
4549 *end = a_end;
4550 }
4551 }
4552}
4553
4554/*
Steve Capper3212b532013-04-23 12:35:02 +01004555 * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
4556 * and returns the corresponding pte. While this is not necessary for the
4557 * !shared pmd case because we can allocate the pmd later as well, it makes the
4558 * code much cleaner. pmd allocation is essential for the shared case because
Davidlohr Buesoc8c06ef2014-12-12 16:54:24 -08004559 * pud has to be populated inside the same i_mmap_rwsem section - otherwise
Steve Capper3212b532013-04-23 12:35:02 +01004560 * racing tasks could either miss the sharing (see huge_pte_offset) or select a
4561 * bad pmd for sharing.
4562 */
4563pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
4564{
4565 struct vm_area_struct *vma = find_vma(mm, addr);
4566 struct address_space *mapping = vma->vm_file->f_mapping;
4567 pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) +
4568 vma->vm_pgoff;
4569 struct vm_area_struct *svma;
4570 unsigned long saddr;
4571 pte_t *spte = NULL;
4572 pte_t *pte;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08004573 spinlock_t *ptl;
Steve Capper3212b532013-04-23 12:35:02 +01004574
4575 if (!vma_shareable(vma, addr))
4576 return (pte_t *)pmd_alloc(mm, pud, addr);
4577
Davidlohr Bueso83cde9e2014-12-12 16:54:21 -08004578 i_mmap_lock_write(mapping);
Steve Capper3212b532013-04-23 12:35:02 +01004579 vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) {
4580 if (svma == vma)
4581 continue;
4582
4583 saddr = page_table_shareable(svma, vma, addr, idx);
4584 if (saddr) {
Punit Agrawal7868a202017-07-06 15:39:42 -07004585 spte = huge_pte_offset(svma->vm_mm, saddr,
4586 vma_mmu_pagesize(svma));
Steve Capper3212b532013-04-23 12:35:02 +01004587 if (spte) {
4588 get_page(virt_to_page(spte));
4589 break;
4590 }
4591 }
4592 }
4593
4594 if (!spte)
4595 goto out;
4596
Aneesh Kumar K.V8bea8052016-12-12 16:41:59 -08004597 ptl = huge_pte_lock(hstate_vma(vma), mm, spte);
Kirill A. Shutemovdc6c9a32015-02-11 15:26:50 -08004598 if (pud_none(*pud)) {
Steve Capper3212b532013-04-23 12:35:02 +01004599 pud_populate(mm, pud,
4600 (pmd_t *)((unsigned long)spte & PAGE_MASK));
Kirill A. Shutemovc17b1f42016-06-24 14:49:51 -07004601 mm_inc_nr_pmds(mm);
Kirill A. Shutemovdc6c9a32015-02-11 15:26:50 -08004602 } else {
Steve Capper3212b532013-04-23 12:35:02 +01004603 put_page(virt_to_page(spte));
Kirill A. Shutemovdc6c9a32015-02-11 15:26:50 -08004604 }
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08004605 spin_unlock(ptl);
Steve Capper3212b532013-04-23 12:35:02 +01004606out:
4607 pte = (pte_t *)pmd_alloc(mm, pud, addr);
Davidlohr Bueso83cde9e2014-12-12 16:54:21 -08004608 i_mmap_unlock_write(mapping);
Steve Capper3212b532013-04-23 12:35:02 +01004609 return pte;
4610}
4611
4612/*
4613 * unmap huge page backed by shared pte.
4614 *
4615 * Hugetlb pte page is ref counted at the time of mapping. If pte is shared
4616 * indicated by page_count > 1, unmap is achieved by clearing pud and
4617 * decrementing the ref count. If count == 1, the pte page is not shared.
4618 *
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08004619 * called with page table lock held.
Steve Capper3212b532013-04-23 12:35:02 +01004620 *
4621 * returns: 1 successfully unmapped a shared pte page
4622 * 0 the underlying pte page is not shared, or it is the last user
4623 */
4624int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
4625{
4626 pgd_t *pgd = pgd_offset(mm, *addr);
Kirill A. Shutemovc2febaf2017-03-09 17:24:07 +03004627 p4d_t *p4d = p4d_offset(pgd, *addr);
4628 pud_t *pud = pud_offset(p4d, *addr);
Steve Capper3212b532013-04-23 12:35:02 +01004629
4630 BUG_ON(page_count(virt_to_page(ptep)) == 0);
4631 if (page_count(virt_to_page(ptep)) == 1)
4632 return 0;
4633
4634 pud_clear(pud);
4635 put_page(virt_to_page(ptep));
Kirill A. Shutemovdc6c9a32015-02-11 15:26:50 -08004636 mm_dec_nr_pmds(mm);
Steve Capper3212b532013-04-23 12:35:02 +01004637 *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE;
4638 return 1;
4639}
Steve Capper9e5fc742013-04-30 08:02:03 +01004640#define want_pmd_share() (1)
4641#else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
4642pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
4643{
4644 return NULL;
4645}
Zhang Zhene81f2d22015-06-24 16:56:13 -07004646
4647int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
4648{
4649 return 0;
4650}
Mike Kravetz5f4f5b12018-10-05 15:51:29 -07004651
4652void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma,
4653 unsigned long *start, unsigned long *end)
4654{
4655}
Steve Capper9e5fc742013-04-30 08:02:03 +01004656#define want_pmd_share() (0)
Steve Capper3212b532013-04-23 12:35:02 +01004657#endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
4658
Steve Capper9e5fc742013-04-30 08:02:03 +01004659#ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
4660pte_t *huge_pte_alloc(struct mm_struct *mm,
4661 unsigned long addr, unsigned long sz)
4662{
4663 pgd_t *pgd;
Kirill A. Shutemovc2febaf2017-03-09 17:24:07 +03004664 p4d_t *p4d;
Steve Capper9e5fc742013-04-30 08:02:03 +01004665 pud_t *pud;
4666 pte_t *pte = NULL;
4667
4668 pgd = pgd_offset(mm, addr);
Kirill A. Shutemovbe75ad82017-11-29 16:11:30 -08004669 p4d = p4d_alloc(mm, pgd, addr);
4670 if (!p4d)
4671 return NULL;
Kirill A. Shutemovc2febaf2017-03-09 17:24:07 +03004672 pud = pud_alloc(mm, p4d, addr);
Steve Capper9e5fc742013-04-30 08:02:03 +01004673 if (pud) {
4674 if (sz == PUD_SIZE) {
4675 pte = (pte_t *)pud;
4676 } else {
4677 BUG_ON(sz != PMD_SIZE);
4678 if (want_pmd_share() && pud_none(*pud))
4679 pte = huge_pmd_share(mm, addr, pud);
4680 else
4681 pte = (pte_t *)pmd_alloc(mm, pud, addr);
4682 }
4683 }
Michal Hocko4e666312016-08-02 14:02:34 -07004684 BUG_ON(pte && pte_present(*pte) && !pte_huge(*pte));
Steve Capper9e5fc742013-04-30 08:02:03 +01004685
4686 return pte;
4687}
4688
Punit Agrawal9b19df22017-09-06 16:21:01 -07004689/*
4690 * huge_pte_offset() - Walk the page table to resolve the hugepage
4691 * entry at address @addr
4692 *
4693 * Return: Pointer to page table or swap entry (PUD or PMD) for
4694 * address @addr, or NULL if a p*d_none() entry is encountered and the
4695 * size @sz doesn't match the hugepage size at this level of the page
4696 * table.
4697 */
Punit Agrawal7868a202017-07-06 15:39:42 -07004698pte_t *huge_pte_offset(struct mm_struct *mm,
4699 unsigned long addr, unsigned long sz)
Steve Capper9e5fc742013-04-30 08:02:03 +01004700{
4701 pgd_t *pgd;
Kirill A. Shutemovc2febaf2017-03-09 17:24:07 +03004702 p4d_t *p4d;
Steve Capper9e5fc742013-04-30 08:02:03 +01004703 pud_t *pud;
Kirill A. Shutemovc2febaf2017-03-09 17:24:07 +03004704 pmd_t *pmd;
Steve Capper9e5fc742013-04-30 08:02:03 +01004705
4706 pgd = pgd_offset(mm, addr);
Kirill A. Shutemovc2febaf2017-03-09 17:24:07 +03004707 if (!pgd_present(*pgd))
4708 return NULL;
4709 p4d = p4d_offset(pgd, addr);
4710 if (!p4d_present(*p4d))
4711 return NULL;
Punit Agrawal9b19df22017-09-06 16:21:01 -07004712
Kirill A. Shutemovc2febaf2017-03-09 17:24:07 +03004713 pud = pud_offset(p4d, addr);
Punit Agrawal9b19df22017-09-06 16:21:01 -07004714 if (sz != PUD_SIZE && pud_none(*pud))
Kirill A. Shutemovc2febaf2017-03-09 17:24:07 +03004715 return NULL;
Punit Agrawal9b19df22017-09-06 16:21:01 -07004716 /* hugepage or swap? */
4717 if (pud_huge(*pud) || !pud_present(*pud))
Kirill A. Shutemovc2febaf2017-03-09 17:24:07 +03004718 return (pte_t *)pud;
Punit Agrawal9b19df22017-09-06 16:21:01 -07004719
Kirill A. Shutemovc2febaf2017-03-09 17:24:07 +03004720 pmd = pmd_offset(pud, addr);
Punit Agrawal9b19df22017-09-06 16:21:01 -07004721 if (sz != PMD_SIZE && pmd_none(*pmd))
4722 return NULL;
4723 /* hugepage or swap? */
4724 if (pmd_huge(*pmd) || !pmd_present(*pmd))
4725 return (pte_t *)pmd;
4726
4727 return NULL;
Steve Capper9e5fc742013-04-30 08:02:03 +01004728}
4729
Naoya Horiguchi61f77ed2015-02-11 15:25:15 -08004730#endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
4731
4732/*
4733 * These functions are overwritable if your architecture needs its own
4734 * behavior.
4735 */
4736struct page * __weak
4737follow_huge_addr(struct mm_struct *mm, unsigned long address,
4738 int write)
4739{
4740 return ERR_PTR(-EINVAL);
4741}
4742
4743struct page * __weak
Aneesh Kumar K.V4dc71452017-07-06 15:38:56 -07004744follow_huge_pd(struct vm_area_struct *vma,
4745 unsigned long address, hugepd_t hpd, int flags, int pdshift)
4746{
4747 WARN(1, "hugepd follow called with no support for hugepage directory format\n");
4748 return NULL;
4749}
4750
4751struct page * __weak
Steve Capper9e5fc742013-04-30 08:02:03 +01004752follow_huge_pmd(struct mm_struct *mm, unsigned long address,
Naoya Horiguchie66f17f2015-02-11 15:25:22 -08004753 pmd_t *pmd, int flags)
Steve Capper9e5fc742013-04-30 08:02:03 +01004754{
Naoya Horiguchie66f17f2015-02-11 15:25:22 -08004755 struct page *page = NULL;
4756 spinlock_t *ptl;
Naoya Horiguchic9d398f2017-03-31 15:11:55 -07004757 pte_t pte;
Naoya Horiguchie66f17f2015-02-11 15:25:22 -08004758retry:
4759 ptl = pmd_lockptr(mm, pmd);
4760 spin_lock(ptl);
4761 /*
4762 * make sure that the address range covered by this pmd is not
4763 * unmapped from other threads.
4764 */
4765 if (!pmd_huge(*pmd))
4766 goto out;
Naoya Horiguchic9d398f2017-03-31 15:11:55 -07004767 pte = huge_ptep_get((pte_t *)pmd);
4768 if (pte_present(pte)) {
Gerald Schaefer97534122015-04-14 15:42:30 -07004769 page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT);
Naoya Horiguchie66f17f2015-02-11 15:25:22 -08004770 if (flags & FOLL_GET)
4771 get_page(page);
4772 } else {
Naoya Horiguchic9d398f2017-03-31 15:11:55 -07004773 if (is_hugetlb_entry_migration(pte)) {
Naoya Horiguchie66f17f2015-02-11 15:25:22 -08004774 spin_unlock(ptl);
4775 __migration_entry_wait(mm, (pte_t *)pmd, ptl);
4776 goto retry;
4777 }
4778 /*
4779 * hwpoisoned entry is treated as no_page_table in
4780 * follow_page_mask().
4781 */
4782 }
4783out:
4784 spin_unlock(ptl);
Steve Capper9e5fc742013-04-30 08:02:03 +01004785 return page;
4786}
4787
Naoya Horiguchi61f77ed2015-02-11 15:25:15 -08004788struct page * __weak
Steve Capper9e5fc742013-04-30 08:02:03 +01004789follow_huge_pud(struct mm_struct *mm, unsigned long address,
Naoya Horiguchie66f17f2015-02-11 15:25:22 -08004790 pud_t *pud, int flags)
Steve Capper9e5fc742013-04-30 08:02:03 +01004791{
Naoya Horiguchie66f17f2015-02-11 15:25:22 -08004792 if (flags & FOLL_GET)
4793 return NULL;
Steve Capper9e5fc742013-04-30 08:02:03 +01004794
Naoya Horiguchie66f17f2015-02-11 15:25:22 -08004795 return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT);
Steve Capper9e5fc742013-04-30 08:02:03 +01004796}
4797
Anshuman Khandualfaaa5b62017-07-06 15:38:50 -07004798struct page * __weak
4799follow_huge_pgd(struct mm_struct *mm, unsigned long address, pgd_t *pgd, int flags)
4800{
4801 if (flags & FOLL_GET)
4802 return NULL;
4803
4804 return pte_page(*(pte_t *)pgd) + ((address & ~PGDIR_MASK) >> PAGE_SHIFT);
4805}
4806
Naoya Horiguchi31caf662013-09-11 14:21:59 -07004807bool isolate_huge_page(struct page *page, struct list_head *list)
4808{
Naoya Horiguchibcc54222015-04-15 16:14:38 -07004809 bool ret = true;
4810
Sasha Levin309381fea2014-01-23 15:52:54 -08004811 VM_BUG_ON_PAGE(!PageHead(page), page);
Naoya Horiguchi31caf662013-09-11 14:21:59 -07004812 spin_lock(&hugetlb_lock);
Naoya Horiguchibcc54222015-04-15 16:14:38 -07004813 if (!page_huge_active(page) || !get_page_unless_zero(page)) {
4814 ret = false;
4815 goto unlock;
4816 }
4817 clear_page_huge_active(page);
Naoya Horiguchi31caf662013-09-11 14:21:59 -07004818 list_move_tail(&page->lru, list);
Naoya Horiguchibcc54222015-04-15 16:14:38 -07004819unlock:
Naoya Horiguchi31caf662013-09-11 14:21:59 -07004820 spin_unlock(&hugetlb_lock);
Naoya Horiguchibcc54222015-04-15 16:14:38 -07004821 return ret;
Naoya Horiguchi31caf662013-09-11 14:21:59 -07004822}
4823
4824void putback_active_hugepage(struct page *page)
4825{
Sasha Levin309381fea2014-01-23 15:52:54 -08004826 VM_BUG_ON_PAGE(!PageHead(page), page);
Naoya Horiguchi31caf662013-09-11 14:21:59 -07004827 spin_lock(&hugetlb_lock);
Naoya Horiguchibcc54222015-04-15 16:14:38 -07004828 set_page_huge_active(page);
Naoya Horiguchi31caf662013-09-11 14:21:59 -07004829 list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist);
4830 spin_unlock(&hugetlb_lock);
4831 put_page(page);
4832}