blob: 271e4432734c376baf0bf4b8953a38e391ac011c [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
2 * Generic hugetlb support.
Nadia Yvette Chambers6d49e352012-12-06 10:39:54 +01003 * (C) Nadia Yvette Chambers, April 2004
Linus Torvalds1da177e2005-04-16 15:20:36 -07004 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07005#include <linux/list.h>
6#include <linux/init.h>
7#include <linux/module.h>
8#include <linux/mm.h>
Alexey Dobriyane1759c22008-10-15 23:50:22 +04009#include <linux/seq_file.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070010#include <linux/sysctl.h>
11#include <linux/highmem.h>
Andrea Arcangelicddb8a52008-07-28 15:46:29 -070012#include <linux/mmu_notifier.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070013#include <linux/nodemask.h>
David Gibson63551ae2005-06-21 17:14:44 -070014#include <linux/pagemap.h>
Christoph Lameter5da7ca82006-01-06 00:10:46 -080015#include <linux/mempolicy.h>
Gideon Israel Dsouza3b321232014-04-07 15:37:26 -070016#include <linux/compiler.h>
Christoph Lameteraea47ff2006-01-08 01:00:57 -080017#include <linux/cpuset.h>
David Gibson3935baa2006-03-22 00:08:53 -080018#include <linux/mutex.h>
Andi Kleenaa888a72008-07-23 21:27:47 -070019#include <linux/bootmem.h>
Nishanth Aravamudana3437872008-07-23 21:27:44 -070020#include <linux/sysfs.h>
Tejun Heo5a0e3ad2010-03-24 17:04:11 +090021#include <linux/slab.h>
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +090022#include <linux/rmap.h>
Naoya Horiguchifd6a03e2010-05-28 09:29:21 +090023#include <linux/swap.h>
24#include <linux/swapops.h>
Naoya Horiguchic8721bb2013-09-11 14:22:09 -070025#include <linux/page-isolation.h>
Davidlohr Bueso8382d912014-04-03 14:47:31 -070026#include <linux/jhash.h>
Linus Torvaldsd6606682008-08-06 12:04:54 -070027
David Gibson63551ae2005-06-21 17:14:44 -070028#include <asm/page.h>
29#include <asm/pgtable.h>
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -070030#include <asm/tlb.h>
David Gibson63551ae2005-06-21 17:14:44 -070031
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -070032#include <linux/io.h>
David Gibson63551ae2005-06-21 17:14:44 -070033#include <linux/hugetlb.h>
Aneesh Kumar K.V9dd540e2012-07-31 16:42:15 -070034#include <linux/hugetlb_cgroup.h>
Lee Schermerhorn9a3052302009-12-14 17:58:25 -080035#include <linux/node.h>
Nick Piggin7835e982006-03-22 00:08:40 -080036#include "internal.h"
Linus Torvalds1da177e2005-04-16 15:20:36 -070037
Andrey Ryabinin753162c2015-02-10 14:11:36 -080038int hugepages_treat_as_movable;
Andi Kleena5516432008-07-23 21:27:41 -070039
Aneesh Kumar K.Vc3f38a32012-07-31 16:42:10 -070040int hugetlb_max_hstate __read_mostly;
Andi Kleene5ff2152008-07-23 21:27:42 -070041unsigned int default_hstate_idx;
42struct hstate hstates[HUGE_MAX_HSTATE];
43
Jon Tollefson53ba51d2008-07-23 21:27:52 -070044__initdata LIST_HEAD(huge_boot_pages);
45
Andi Kleene5ff2152008-07-23 21:27:42 -070046/* for command line parsing */
47static struct hstate * __initdata parsed_hstate;
48static unsigned long __initdata default_hstate_max_huge_pages;
Nick Piggine11bfbf2008-07-23 21:27:52 -070049static unsigned long __initdata default_hstate_size;
Andi Kleene5ff2152008-07-23 21:27:42 -070050
David Gibson3935baa2006-03-22 00:08:53 -080051/*
Naoya Horiguchi31caf662013-09-11 14:21:59 -070052 * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages,
53 * free_huge_pages, and surplus_huge_pages.
David Gibson3935baa2006-03-22 00:08:53 -080054 */
Aneesh Kumar K.Vc3f38a32012-07-31 16:42:10 -070055DEFINE_SPINLOCK(hugetlb_lock);
Eric Paris0bd0f9f2005-11-21 21:32:28 -080056
Davidlohr Bueso8382d912014-04-03 14:47:31 -070057/*
58 * Serializes faults on the same logical page. This is used to
59 * prevent spurious OOMs when the hugepage pool is fully utilized.
60 */
61static int num_fault_mutexes;
62static struct mutex *htlb_fault_mutex_table ____cacheline_aligned_in_smp;
63
Mike Kravetz7ca02d02015-04-15 16:13:42 -070064/* Forward declaration */
65static int hugetlb_acct_memory(struct hstate *h, long delta);
66
David Gibson90481622012-03-21 16:34:12 -070067static inline void unlock_or_release_subpool(struct hugepage_subpool *spool)
68{
69 bool free = (spool->count == 0) && (spool->used_hpages == 0);
70
71 spin_unlock(&spool->lock);
72
73 /* If no pages are used, and no other handles to the subpool
Mike Kravetz7ca02d02015-04-15 16:13:42 -070074 * remain, give up any reservations mased on minimum size and
75 * free the subpool */
76 if (free) {
77 if (spool->min_hpages != -1)
78 hugetlb_acct_memory(spool->hstate,
79 -spool->min_hpages);
David Gibson90481622012-03-21 16:34:12 -070080 kfree(spool);
Mike Kravetz7ca02d02015-04-15 16:13:42 -070081 }
David Gibson90481622012-03-21 16:34:12 -070082}
83
Mike Kravetz7ca02d02015-04-15 16:13:42 -070084struct hugepage_subpool *hugepage_new_subpool(struct hstate *h, long max_hpages,
85 long min_hpages)
David Gibson90481622012-03-21 16:34:12 -070086{
87 struct hugepage_subpool *spool;
88
Mike Kravetzc6a91822015-04-15 16:13:36 -070089 spool = kzalloc(sizeof(*spool), GFP_KERNEL);
David Gibson90481622012-03-21 16:34:12 -070090 if (!spool)
91 return NULL;
92
93 spin_lock_init(&spool->lock);
94 spool->count = 1;
Mike Kravetz7ca02d02015-04-15 16:13:42 -070095 spool->max_hpages = max_hpages;
96 spool->hstate = h;
97 spool->min_hpages = min_hpages;
98
99 if (min_hpages != -1 && hugetlb_acct_memory(h, min_hpages)) {
100 kfree(spool);
101 return NULL;
102 }
103 spool->rsv_hpages = min_hpages;
David Gibson90481622012-03-21 16:34:12 -0700104
105 return spool;
106}
107
108void hugepage_put_subpool(struct hugepage_subpool *spool)
109{
110 spin_lock(&spool->lock);
111 BUG_ON(!spool->count);
112 spool->count--;
113 unlock_or_release_subpool(spool);
114}
115
Mike Kravetz1c5ecae2015-04-15 16:13:39 -0700116/*
117 * Subpool accounting for allocating and reserving pages.
118 * Return -ENOMEM if there are not enough resources to satisfy the
119 * the request. Otherwise, return the number of pages by which the
120 * global pools must be adjusted (upward). The returned value may
121 * only be different than the passed value (delta) in the case where
122 * a subpool minimum size must be manitained.
123 */
124static long hugepage_subpool_get_pages(struct hugepage_subpool *spool,
David Gibson90481622012-03-21 16:34:12 -0700125 long delta)
126{
Mike Kravetz1c5ecae2015-04-15 16:13:39 -0700127 long ret = delta;
David Gibson90481622012-03-21 16:34:12 -0700128
129 if (!spool)
Mike Kravetz1c5ecae2015-04-15 16:13:39 -0700130 return ret;
David Gibson90481622012-03-21 16:34:12 -0700131
132 spin_lock(&spool->lock);
David Gibson90481622012-03-21 16:34:12 -0700133
Mike Kravetz1c5ecae2015-04-15 16:13:39 -0700134 if (spool->max_hpages != -1) { /* maximum size accounting */
135 if ((spool->used_hpages + delta) <= spool->max_hpages)
136 spool->used_hpages += delta;
137 else {
138 ret = -ENOMEM;
139 goto unlock_ret;
140 }
141 }
142
143 if (spool->min_hpages != -1) { /* minimum size accounting */
144 if (delta > spool->rsv_hpages) {
145 /*
146 * Asking for more reserves than those already taken on
147 * behalf of subpool. Return difference.
148 */
149 ret = delta - spool->rsv_hpages;
150 spool->rsv_hpages = 0;
151 } else {
152 ret = 0; /* reserves already accounted for */
153 spool->rsv_hpages -= delta;
154 }
155 }
156
157unlock_ret:
158 spin_unlock(&spool->lock);
David Gibson90481622012-03-21 16:34:12 -0700159 return ret;
160}
161
Mike Kravetz1c5ecae2015-04-15 16:13:39 -0700162/*
163 * Subpool accounting for freeing and unreserving pages.
164 * Return the number of global page reservations that must be dropped.
165 * The return value may only be different than the passed value (delta)
166 * in the case where a subpool minimum size must be maintained.
167 */
168static long hugepage_subpool_put_pages(struct hugepage_subpool *spool,
David Gibson90481622012-03-21 16:34:12 -0700169 long delta)
170{
Mike Kravetz1c5ecae2015-04-15 16:13:39 -0700171 long ret = delta;
172
David Gibson90481622012-03-21 16:34:12 -0700173 if (!spool)
Mike Kravetz1c5ecae2015-04-15 16:13:39 -0700174 return delta;
David Gibson90481622012-03-21 16:34:12 -0700175
176 spin_lock(&spool->lock);
Mike Kravetz1c5ecae2015-04-15 16:13:39 -0700177
178 if (spool->max_hpages != -1) /* maximum size accounting */
179 spool->used_hpages -= delta;
180
181 if (spool->min_hpages != -1) { /* minimum size accounting */
182 if (spool->rsv_hpages + delta <= spool->min_hpages)
183 ret = 0;
184 else
185 ret = spool->rsv_hpages + delta - spool->min_hpages;
186
187 spool->rsv_hpages += delta;
188 if (spool->rsv_hpages > spool->min_hpages)
189 spool->rsv_hpages = spool->min_hpages;
190 }
191
192 /*
193 * If hugetlbfs_put_super couldn't free spool due to an outstanding
194 * quota reference, free it now.
195 */
David Gibson90481622012-03-21 16:34:12 -0700196 unlock_or_release_subpool(spool);
Mike Kravetz1c5ecae2015-04-15 16:13:39 -0700197
198 return ret;
David Gibson90481622012-03-21 16:34:12 -0700199}
200
201static inline struct hugepage_subpool *subpool_inode(struct inode *inode)
202{
203 return HUGETLBFS_SB(inode->i_sb)->spool;
204}
205
206static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma)
207{
Al Viro496ad9a2013-01-23 17:07:38 -0500208 return subpool_inode(file_inode(vma->vm_file));
David Gibson90481622012-03-21 16:34:12 -0700209}
210
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700211/*
Andy Whitcroft96822902008-07-23 21:27:29 -0700212 * Region tracking -- allows tracking of reservations and instantiated pages
213 * across the pages in a mapping.
Andy Whitcroft84afd992008-07-23 21:27:32 -0700214 *
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700215 * The region data structures are embedded into a resv_map and
216 * protected by a resv_map's lock
Andy Whitcroft96822902008-07-23 21:27:29 -0700217 */
218struct file_region {
219 struct list_head link;
220 long from;
221 long to;
222};
223
Joonsoo Kim1406ec92014-04-03 14:47:26 -0700224static long region_add(struct resv_map *resv, long f, long t)
Andy Whitcroft96822902008-07-23 21:27:29 -0700225{
Joonsoo Kim1406ec92014-04-03 14:47:26 -0700226 struct list_head *head = &resv->regions;
Andy Whitcroft96822902008-07-23 21:27:29 -0700227 struct file_region *rg, *nrg, *trg;
228
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700229 spin_lock(&resv->lock);
Andy Whitcroft96822902008-07-23 21:27:29 -0700230 /* Locate the region we are either in or before. */
231 list_for_each_entry(rg, head, link)
232 if (f <= rg->to)
233 break;
234
235 /* Round our left edge to the current segment if it encloses us. */
236 if (f > rg->from)
237 f = rg->from;
238
239 /* Check for and consume any regions we now overlap with. */
240 nrg = rg;
241 list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
242 if (&rg->link == head)
243 break;
244 if (rg->from > t)
245 break;
246
247 /* If this area reaches higher then extend our area to
248 * include it completely. If this is not the first area
249 * which we intend to reuse, free it. */
250 if (rg->to > t)
251 t = rg->to;
252 if (rg != nrg) {
253 list_del(&rg->link);
254 kfree(rg);
255 }
256 }
257 nrg->from = f;
258 nrg->to = t;
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700259 spin_unlock(&resv->lock);
Andy Whitcroft96822902008-07-23 21:27:29 -0700260 return 0;
261}
262
Joonsoo Kim1406ec92014-04-03 14:47:26 -0700263static long region_chg(struct resv_map *resv, long f, long t)
Andy Whitcroft96822902008-07-23 21:27:29 -0700264{
Joonsoo Kim1406ec92014-04-03 14:47:26 -0700265 struct list_head *head = &resv->regions;
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700266 struct file_region *rg, *nrg = NULL;
Andy Whitcroft96822902008-07-23 21:27:29 -0700267 long chg = 0;
268
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700269retry:
270 spin_lock(&resv->lock);
Andy Whitcroft96822902008-07-23 21:27:29 -0700271 /* Locate the region we are before or in. */
272 list_for_each_entry(rg, head, link)
273 if (f <= rg->to)
274 break;
275
276 /* If we are below the current region then a new region is required.
277 * Subtle, allocate a new region at the position but make it zero
278 * size such that we can guarantee to record the reservation. */
279 if (&rg->link == head || t < rg->from) {
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700280 if (!nrg) {
281 spin_unlock(&resv->lock);
282 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
283 if (!nrg)
284 return -ENOMEM;
Andy Whitcroft96822902008-07-23 21:27:29 -0700285
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700286 nrg->from = f;
287 nrg->to = f;
288 INIT_LIST_HEAD(&nrg->link);
289 goto retry;
290 }
291
292 list_add(&nrg->link, rg->link.prev);
293 chg = t - f;
294 goto out_nrg;
Andy Whitcroft96822902008-07-23 21:27:29 -0700295 }
296
297 /* Round our left edge to the current segment if it encloses us. */
298 if (f > rg->from)
299 f = rg->from;
300 chg = t - f;
301
302 /* Check for and consume any regions we now overlap with. */
303 list_for_each_entry(rg, rg->link.prev, link) {
304 if (&rg->link == head)
305 break;
306 if (rg->from > t)
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700307 goto out;
Andy Whitcroft96822902008-07-23 21:27:29 -0700308
Lucas De Marchi25985ed2011-03-30 22:57:33 -0300309 /* We overlap with this area, if it extends further than
Andy Whitcroft96822902008-07-23 21:27:29 -0700310 * us then we must extend ourselves. Account for its
311 * existing reservation. */
312 if (rg->to > t) {
313 chg += rg->to - t;
314 t = rg->to;
315 }
316 chg -= rg->to - rg->from;
317 }
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700318
319out:
320 spin_unlock(&resv->lock);
321 /* We already know we raced and no longer need the new region */
322 kfree(nrg);
323 return chg;
324out_nrg:
325 spin_unlock(&resv->lock);
Andy Whitcroft96822902008-07-23 21:27:29 -0700326 return chg;
327}
328
Joonsoo Kim1406ec92014-04-03 14:47:26 -0700329static long region_truncate(struct resv_map *resv, long end)
Andy Whitcroft96822902008-07-23 21:27:29 -0700330{
Joonsoo Kim1406ec92014-04-03 14:47:26 -0700331 struct list_head *head = &resv->regions;
Andy Whitcroft96822902008-07-23 21:27:29 -0700332 struct file_region *rg, *trg;
333 long chg = 0;
334
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700335 spin_lock(&resv->lock);
Andy Whitcroft96822902008-07-23 21:27:29 -0700336 /* Locate the region we are either in or before. */
337 list_for_each_entry(rg, head, link)
338 if (end <= rg->to)
339 break;
340 if (&rg->link == head)
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700341 goto out;
Andy Whitcroft96822902008-07-23 21:27:29 -0700342
343 /* If we are in the middle of a region then adjust it. */
344 if (end > rg->from) {
345 chg = rg->to - end;
346 rg->to = end;
347 rg = list_entry(rg->link.next, typeof(*rg), link);
348 }
349
350 /* Drop any remaining regions. */
351 list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
352 if (&rg->link == head)
353 break;
354 chg += rg->to - rg->from;
355 list_del(&rg->link);
356 kfree(rg);
357 }
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700358
359out:
360 spin_unlock(&resv->lock);
Andy Whitcroft96822902008-07-23 21:27:29 -0700361 return chg;
362}
363
Joonsoo Kim1406ec92014-04-03 14:47:26 -0700364static long region_count(struct resv_map *resv, long f, long t)
Andy Whitcroft84afd992008-07-23 21:27:32 -0700365{
Joonsoo Kim1406ec92014-04-03 14:47:26 -0700366 struct list_head *head = &resv->regions;
Andy Whitcroft84afd992008-07-23 21:27:32 -0700367 struct file_region *rg;
368 long chg = 0;
369
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700370 spin_lock(&resv->lock);
Andy Whitcroft84afd992008-07-23 21:27:32 -0700371 /* Locate each segment we overlap with, and count that overlap. */
372 list_for_each_entry(rg, head, link) {
Wang Sheng-Huif2135a42012-05-29 15:06:17 -0700373 long seg_from;
374 long seg_to;
Andy Whitcroft84afd992008-07-23 21:27:32 -0700375
376 if (rg->to <= f)
377 continue;
378 if (rg->from >= t)
379 break;
380
381 seg_from = max(rg->from, f);
382 seg_to = min(rg->to, t);
383
384 chg += seg_to - seg_from;
385 }
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700386 spin_unlock(&resv->lock);
Andy Whitcroft84afd992008-07-23 21:27:32 -0700387
388 return chg;
389}
390
Andy Whitcroft96822902008-07-23 21:27:29 -0700391/*
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700392 * Convert the address within this vma to the page offset within
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700393 * the mapping, in pagecache page units; huge pages here.
394 */
Andi Kleena5516432008-07-23 21:27:41 -0700395static pgoff_t vma_hugecache_offset(struct hstate *h,
396 struct vm_area_struct *vma, unsigned long address)
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700397{
Andi Kleena5516432008-07-23 21:27:41 -0700398 return ((address - vma->vm_start) >> huge_page_shift(h)) +
399 (vma->vm_pgoff >> huge_page_order(h));
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700400}
401
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +0900402pgoff_t linear_hugepage_index(struct vm_area_struct *vma,
403 unsigned long address)
404{
405 return vma_hugecache_offset(hstate_vma(vma), vma, address);
406}
407
Andy Whitcroft84afd992008-07-23 21:27:32 -0700408/*
Mel Gorman08fba692009-01-06 14:38:53 -0800409 * Return the size of the pages allocated when backing a VMA. In the majority
410 * cases this will be same size as used by the page table entries.
411 */
412unsigned long vma_kernel_pagesize(struct vm_area_struct *vma)
413{
414 struct hstate *hstate;
415
416 if (!is_vm_hugetlb_page(vma))
417 return PAGE_SIZE;
418
419 hstate = hstate_vma(vma);
420
Wanpeng Li2415cf12013-07-03 15:02:43 -0700421 return 1UL << huge_page_shift(hstate);
Mel Gorman08fba692009-01-06 14:38:53 -0800422}
Joerg Roedelf340ca02009-06-19 15:16:22 +0200423EXPORT_SYMBOL_GPL(vma_kernel_pagesize);
Mel Gorman08fba692009-01-06 14:38:53 -0800424
425/*
Mel Gorman33402892009-01-06 14:38:54 -0800426 * Return the page size being used by the MMU to back a VMA. In the majority
427 * of cases, the page size used by the kernel matches the MMU size. On
428 * architectures where it differs, an architecture-specific version of this
429 * function is required.
430 */
431#ifndef vma_mmu_pagesize
432unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
433{
434 return vma_kernel_pagesize(vma);
435}
436#endif
437
438/*
Andy Whitcroft84afd992008-07-23 21:27:32 -0700439 * Flags for MAP_PRIVATE reservations. These are stored in the bottom
440 * bits of the reservation map pointer, which are always clear due to
441 * alignment.
442 */
443#define HPAGE_RESV_OWNER (1UL << 0)
444#define HPAGE_RESV_UNMAPPED (1UL << 1)
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700445#define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
Andy Whitcroft84afd992008-07-23 21:27:32 -0700446
Mel Gormana1e78772008-07-23 21:27:23 -0700447/*
448 * These helpers are used to track how many pages are reserved for
449 * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
450 * is guaranteed to have their future faults succeed.
451 *
452 * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
453 * the reserve counters are updated with the hugetlb_lock held. It is safe
454 * to reset the VMA at fork() time as it is not in use yet and there is no
455 * chance of the global counters getting corrupted as a result of the values.
Andy Whitcroft84afd992008-07-23 21:27:32 -0700456 *
457 * The private mapping reservation is represented in a subtly different
458 * manner to a shared mapping. A shared mapping has a region map associated
459 * with the underlying file, this region map represents the backing file
460 * pages which have ever had a reservation assigned which this persists even
461 * after the page is instantiated. A private mapping has a region map
462 * associated with the original mmap which is attached to all VMAs which
463 * reference it, this region map represents those offsets which have consumed
464 * reservation ie. where pages have been instantiated.
Mel Gormana1e78772008-07-23 21:27:23 -0700465 */
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700466static unsigned long get_vma_private_data(struct vm_area_struct *vma)
467{
468 return (unsigned long)vma->vm_private_data;
469}
470
471static void set_vma_private_data(struct vm_area_struct *vma,
472 unsigned long value)
473{
474 vma->vm_private_data = (void *)value;
475}
476
Joonsoo Kim9119a412014-04-03 14:47:25 -0700477struct resv_map *resv_map_alloc(void)
Andy Whitcroft84afd992008-07-23 21:27:32 -0700478{
479 struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
480 if (!resv_map)
481 return NULL;
482
483 kref_init(&resv_map->refs);
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700484 spin_lock_init(&resv_map->lock);
Andy Whitcroft84afd992008-07-23 21:27:32 -0700485 INIT_LIST_HEAD(&resv_map->regions);
486
487 return resv_map;
488}
489
Joonsoo Kim9119a412014-04-03 14:47:25 -0700490void resv_map_release(struct kref *ref)
Andy Whitcroft84afd992008-07-23 21:27:32 -0700491{
492 struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
493
494 /* Clear out any active regions before we release the map. */
Joonsoo Kim1406ec92014-04-03 14:47:26 -0700495 region_truncate(resv_map, 0);
Andy Whitcroft84afd992008-07-23 21:27:32 -0700496 kfree(resv_map);
497}
498
Joonsoo Kim4e35f482014-04-03 14:47:30 -0700499static inline struct resv_map *inode_resv_map(struct inode *inode)
500{
501 return inode->i_mapping->private_data;
502}
503
Andy Whitcroft84afd992008-07-23 21:27:32 -0700504static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
Mel Gormana1e78772008-07-23 21:27:23 -0700505{
Sasha Levin81d1b092014-10-09 15:28:10 -0700506 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
Joonsoo Kim4e35f482014-04-03 14:47:30 -0700507 if (vma->vm_flags & VM_MAYSHARE) {
508 struct address_space *mapping = vma->vm_file->f_mapping;
509 struct inode *inode = mapping->host;
510
511 return inode_resv_map(inode);
512
513 } else {
Andy Whitcroft84afd992008-07-23 21:27:32 -0700514 return (struct resv_map *)(get_vma_private_data(vma) &
515 ~HPAGE_RESV_MASK);
Joonsoo Kim4e35f482014-04-03 14:47:30 -0700516 }
Mel Gormana1e78772008-07-23 21:27:23 -0700517}
518
Andy Whitcroft84afd992008-07-23 21:27:32 -0700519static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
Mel Gormana1e78772008-07-23 21:27:23 -0700520{
Sasha Levin81d1b092014-10-09 15:28:10 -0700521 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
522 VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
Mel Gormana1e78772008-07-23 21:27:23 -0700523
Andy Whitcroft84afd992008-07-23 21:27:32 -0700524 set_vma_private_data(vma, (get_vma_private_data(vma) &
525 HPAGE_RESV_MASK) | (unsigned long)map);
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700526}
527
528static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
529{
Sasha Levin81d1b092014-10-09 15:28:10 -0700530 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
531 VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700532
533 set_vma_private_data(vma, get_vma_private_data(vma) | flags);
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700534}
535
536static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
537{
Sasha Levin81d1b092014-10-09 15:28:10 -0700538 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700539
540 return (get_vma_private_data(vma) & flag) != 0;
Mel Gormana1e78772008-07-23 21:27:23 -0700541}
542
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700543/* Reset counters to 0 and clear all HPAGE_RESV_* flags */
Mel Gormana1e78772008-07-23 21:27:23 -0700544void reset_vma_resv_huge_pages(struct vm_area_struct *vma)
545{
Sasha Levin81d1b092014-10-09 15:28:10 -0700546 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
Mel Gormanf83a2752009-05-28 14:34:40 -0700547 if (!(vma->vm_flags & VM_MAYSHARE))
Mel Gormana1e78772008-07-23 21:27:23 -0700548 vma->vm_private_data = (void *)0;
549}
550
551/* Returns true if the VMA has associated reserve pages */
Joonsoo Kimaf0ed732013-09-11 14:21:18 -0700552static int vma_has_reserves(struct vm_area_struct *vma, long chg)
Mel Gormana1e78772008-07-23 21:27:23 -0700553{
Joonsoo Kimaf0ed732013-09-11 14:21:18 -0700554 if (vma->vm_flags & VM_NORESERVE) {
555 /*
556 * This address is already reserved by other process(chg == 0),
557 * so, we should decrement reserved count. Without decrementing,
558 * reserve count remains after releasing inode, because this
559 * allocated page will go into page cache and is regarded as
560 * coming from reserved pool in releasing step. Currently, we
561 * don't have any other solution to deal with this situation
562 * properly, so add work-around here.
563 */
564 if (vma->vm_flags & VM_MAYSHARE && chg == 0)
565 return 1;
566 else
567 return 0;
568 }
Joonsoo Kima63884e2013-09-11 14:21:07 -0700569
570 /* Shared mappings always use reserves */
Mel Gormanf83a2752009-05-28 14:34:40 -0700571 if (vma->vm_flags & VM_MAYSHARE)
Mel Gorman7f09ca52008-07-23 21:27:58 -0700572 return 1;
Joonsoo Kima63884e2013-09-11 14:21:07 -0700573
574 /*
575 * Only the process that called mmap() has reserves for
576 * private mappings.
577 */
Mel Gorman7f09ca52008-07-23 21:27:58 -0700578 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER))
579 return 1;
Joonsoo Kima63884e2013-09-11 14:21:07 -0700580
Mel Gorman7f09ca52008-07-23 21:27:58 -0700581 return 0;
Mel Gormana1e78772008-07-23 21:27:23 -0700582}
583
Andi Kleena5516432008-07-23 21:27:41 -0700584static void enqueue_huge_page(struct hstate *h, struct page *page)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700585{
586 int nid = page_to_nid(page);
Aneesh Kumar K.V0edaecf2012-07-31 16:42:07 -0700587 list_move(&page->lru, &h->hugepage_freelists[nid]);
Andi Kleena5516432008-07-23 21:27:41 -0700588 h->free_huge_pages++;
589 h->free_huge_pages_node[nid]++;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700590}
591
Naoya Horiguchibf50bab2010-09-08 10:19:33 +0900592static struct page *dequeue_huge_page_node(struct hstate *h, int nid)
593{
594 struct page *page;
595
Naoya Horiguchic8721bb2013-09-11 14:22:09 -0700596 list_for_each_entry(page, &h->hugepage_freelists[nid], lru)
597 if (!is_migrate_isolate_page(page))
598 break;
599 /*
600 * if 'non-isolated free hugepage' not found on the list,
601 * the allocation fails.
602 */
603 if (&h->hugepage_freelists[nid] == &page->lru)
Naoya Horiguchibf50bab2010-09-08 10:19:33 +0900604 return NULL;
Aneesh Kumar K.V0edaecf2012-07-31 16:42:07 -0700605 list_move(&page->lru, &h->hugepage_activelist);
Naoya Horiguchia9869b82010-09-08 10:19:37 +0900606 set_page_refcounted(page);
Naoya Horiguchibf50bab2010-09-08 10:19:33 +0900607 h->free_huge_pages--;
608 h->free_huge_pages_node[nid]--;
609 return page;
610}
611
Naoya Horiguchi86cdb462013-09-11 14:22:13 -0700612/* Movability of hugepages depends on migration support. */
613static inline gfp_t htlb_alloc_mask(struct hstate *h)
614{
Naoya Horiguchi100873d2014-06-04 16:10:56 -0700615 if (hugepages_treat_as_movable || hugepage_migration_supported(h))
Naoya Horiguchi86cdb462013-09-11 14:22:13 -0700616 return GFP_HIGHUSER_MOVABLE;
617 else
618 return GFP_HIGHUSER;
619}
620
Andi Kleena5516432008-07-23 21:27:41 -0700621static struct page *dequeue_huge_page_vma(struct hstate *h,
622 struct vm_area_struct *vma,
Joonsoo Kimaf0ed732013-09-11 14:21:18 -0700623 unsigned long address, int avoid_reserve,
624 long chg)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700625{
Konstantin Khlebnikovb1c12cb2012-04-25 16:01:46 -0700626 struct page *page = NULL;
Lee Schermerhorn480eccf2007-09-18 22:46:47 -0700627 struct mempolicy *mpol;
Mel Gorman19770b32008-04-28 02:12:18 -0700628 nodemask_t *nodemask;
Miao Xiec0ff7452010-05-24 14:32:08 -0700629 struct zonelist *zonelist;
Mel Gormandd1a2392008-04-28 02:12:17 -0700630 struct zone *zone;
631 struct zoneref *z;
Mel Gormancc9a6c82012-03-21 16:34:11 -0700632 unsigned int cpuset_mems_cookie;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700633
Mel Gormana1e78772008-07-23 21:27:23 -0700634 /*
635 * A child process with MAP_PRIVATE mappings created by their parent
636 * have no page reserves. This check ensures that reservations are
637 * not "stolen". The child may still get SIGKILLed
638 */
Joonsoo Kimaf0ed732013-09-11 14:21:18 -0700639 if (!vma_has_reserves(vma, chg) &&
Andi Kleena5516432008-07-23 21:27:41 -0700640 h->free_huge_pages - h->resv_huge_pages == 0)
Miao Xiec0ff7452010-05-24 14:32:08 -0700641 goto err;
Mel Gormana1e78772008-07-23 21:27:23 -0700642
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700643 /* If reserves cannot be used, ensure enough pages are in the pool */
Andi Kleena5516432008-07-23 21:27:41 -0700644 if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
Justin P. Mattock6eab04a2011-04-08 19:49:08 -0700645 goto err;
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700646
Joonsoo Kim9966c4b2013-09-11 14:20:50 -0700647retry_cpuset:
Mel Gormand26914d2014-04-03 14:47:24 -0700648 cpuset_mems_cookie = read_mems_allowed_begin();
Joonsoo Kim9966c4b2013-09-11 14:20:50 -0700649 zonelist = huge_zonelist(vma, address,
Naoya Horiguchi86cdb462013-09-11 14:22:13 -0700650 htlb_alloc_mask(h), &mpol, &nodemask);
Joonsoo Kim9966c4b2013-09-11 14:20:50 -0700651
Mel Gorman19770b32008-04-28 02:12:18 -0700652 for_each_zone_zonelist_nodemask(zone, z, zonelist,
653 MAX_NR_ZONES - 1, nodemask) {
Vladimir Davydov344736f2014-10-20 15:50:30 +0400654 if (cpuset_zone_allowed(zone, htlb_alloc_mask(h))) {
Naoya Horiguchibf50bab2010-09-08 10:19:33 +0900655 page = dequeue_huge_page_node(h, zone_to_nid(zone));
656 if (page) {
Joonsoo Kimaf0ed732013-09-11 14:21:18 -0700657 if (avoid_reserve)
658 break;
659 if (!vma_has_reserves(vma, chg))
660 break;
661
Joonsoo Kim07443a82013-09-11 14:21:58 -0700662 SetPagePrivate(page);
Joonsoo Kimaf0ed732013-09-11 14:21:18 -0700663 h->resv_huge_pages--;
Naoya Horiguchibf50bab2010-09-08 10:19:33 +0900664 break;
665 }
Andrew Morton3abf7af2007-07-19 01:49:08 -0700666 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700667 }
Mel Gormancc9a6c82012-03-21 16:34:11 -0700668
669 mpol_cond_put(mpol);
Mel Gormand26914d2014-04-03 14:47:24 -0700670 if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie)))
Mel Gormancc9a6c82012-03-21 16:34:11 -0700671 goto retry_cpuset;
672 return page;
673
Miao Xiec0ff7452010-05-24 14:32:08 -0700674err:
Mel Gormancc9a6c82012-03-21 16:34:11 -0700675 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700676}
677
Luiz Capitulino1cac6f22014-06-04 16:07:11 -0700678/*
679 * common helper functions for hstate_next_node_to_{alloc|free}.
680 * We may have allocated or freed a huge page based on a different
681 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
682 * be outside of *nodes_allowed. Ensure that we use an allowed
683 * node for alloc or free.
684 */
685static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
686{
687 nid = next_node(nid, *nodes_allowed);
688 if (nid == MAX_NUMNODES)
689 nid = first_node(*nodes_allowed);
690 VM_BUG_ON(nid >= MAX_NUMNODES);
691
692 return nid;
693}
694
695static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
696{
697 if (!node_isset(nid, *nodes_allowed))
698 nid = next_node_allowed(nid, nodes_allowed);
699 return nid;
700}
701
702/*
703 * returns the previously saved node ["this node"] from which to
704 * allocate a persistent huge page for the pool and advance the
705 * next node from which to allocate, handling wrap at end of node
706 * mask.
707 */
708static int hstate_next_node_to_alloc(struct hstate *h,
709 nodemask_t *nodes_allowed)
710{
711 int nid;
712
713 VM_BUG_ON(!nodes_allowed);
714
715 nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
716 h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
717
718 return nid;
719}
720
721/*
722 * helper for free_pool_huge_page() - return the previously saved
723 * node ["this node"] from which to free a huge page. Advance the
724 * next node id whether or not we find a free huge page to free so
725 * that the next attempt to free addresses the next node.
726 */
727static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
728{
729 int nid;
730
731 VM_BUG_ON(!nodes_allowed);
732
733 nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
734 h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
735
736 return nid;
737}
738
739#define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \
740 for (nr_nodes = nodes_weight(*mask); \
741 nr_nodes > 0 && \
742 ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \
743 nr_nodes--)
744
745#define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \
746 for (nr_nodes = nodes_weight(*mask); \
747 nr_nodes > 0 && \
748 ((node = hstate_next_node_to_free(hs, mask)) || 1); \
749 nr_nodes--)
750
Luiz Capitulino944d9fe2014-06-04 16:07:13 -0700751#if defined(CONFIG_CMA) && defined(CONFIG_X86_64)
752static void destroy_compound_gigantic_page(struct page *page,
753 unsigned long order)
754{
755 int i;
756 int nr_pages = 1 << order;
757 struct page *p = page + 1;
758
759 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
760 __ClearPageTail(p);
761 set_page_refcounted(p);
762 p->first_page = NULL;
763 }
764
765 set_compound_order(page, 0);
766 __ClearPageHead(page);
767}
768
769static void free_gigantic_page(struct page *page, unsigned order)
770{
771 free_contig_range(page_to_pfn(page), 1 << order);
772}
773
774static int __alloc_gigantic_page(unsigned long start_pfn,
775 unsigned long nr_pages)
776{
777 unsigned long end_pfn = start_pfn + nr_pages;
778 return alloc_contig_range(start_pfn, end_pfn, MIGRATE_MOVABLE);
779}
780
781static bool pfn_range_valid_gigantic(unsigned long start_pfn,
782 unsigned long nr_pages)
783{
784 unsigned long i, end_pfn = start_pfn + nr_pages;
785 struct page *page;
786
787 for (i = start_pfn; i < end_pfn; i++) {
788 if (!pfn_valid(i))
789 return false;
790
791 page = pfn_to_page(i);
792
793 if (PageReserved(page))
794 return false;
795
796 if (page_count(page) > 0)
797 return false;
798
799 if (PageHuge(page))
800 return false;
801 }
802
803 return true;
804}
805
806static bool zone_spans_last_pfn(const struct zone *zone,
807 unsigned long start_pfn, unsigned long nr_pages)
808{
809 unsigned long last_pfn = start_pfn + nr_pages - 1;
810 return zone_spans_pfn(zone, last_pfn);
811}
812
813static struct page *alloc_gigantic_page(int nid, unsigned order)
814{
815 unsigned long nr_pages = 1 << order;
816 unsigned long ret, pfn, flags;
817 struct zone *z;
818
819 z = NODE_DATA(nid)->node_zones;
820 for (; z - NODE_DATA(nid)->node_zones < MAX_NR_ZONES; z++) {
821 spin_lock_irqsave(&z->lock, flags);
822
823 pfn = ALIGN(z->zone_start_pfn, nr_pages);
824 while (zone_spans_last_pfn(z, pfn, nr_pages)) {
825 if (pfn_range_valid_gigantic(pfn, nr_pages)) {
826 /*
827 * We release the zone lock here because
828 * alloc_contig_range() will also lock the zone
829 * at some point. If there's an allocation
830 * spinning on this lock, it may win the race
831 * and cause alloc_contig_range() to fail...
832 */
833 spin_unlock_irqrestore(&z->lock, flags);
834 ret = __alloc_gigantic_page(pfn, nr_pages);
835 if (!ret)
836 return pfn_to_page(pfn);
837 spin_lock_irqsave(&z->lock, flags);
838 }
839 pfn += nr_pages;
840 }
841
842 spin_unlock_irqrestore(&z->lock, flags);
843 }
844
845 return NULL;
846}
847
848static void prep_new_huge_page(struct hstate *h, struct page *page, int nid);
849static void prep_compound_gigantic_page(struct page *page, unsigned long order);
850
851static struct page *alloc_fresh_gigantic_page_node(struct hstate *h, int nid)
852{
853 struct page *page;
854
855 page = alloc_gigantic_page(nid, huge_page_order(h));
856 if (page) {
857 prep_compound_gigantic_page(page, huge_page_order(h));
858 prep_new_huge_page(h, page, nid);
859 }
860
861 return page;
862}
863
864static int alloc_fresh_gigantic_page(struct hstate *h,
865 nodemask_t *nodes_allowed)
866{
867 struct page *page = NULL;
868 int nr_nodes, node;
869
870 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
871 page = alloc_fresh_gigantic_page_node(h, node);
872 if (page)
873 return 1;
874 }
875
876 return 0;
877}
878
879static inline bool gigantic_page_supported(void) { return true; }
880#else
881static inline bool gigantic_page_supported(void) { return false; }
882static inline void free_gigantic_page(struct page *page, unsigned order) { }
883static inline void destroy_compound_gigantic_page(struct page *page,
884 unsigned long order) { }
885static inline int alloc_fresh_gigantic_page(struct hstate *h,
886 nodemask_t *nodes_allowed) { return 0; }
887#endif
888
Andi Kleena5516432008-07-23 21:27:41 -0700889static void update_and_free_page(struct hstate *h, struct page *page)
Adam Litke6af2acb2007-10-16 01:26:16 -0700890{
891 int i;
Andi Kleena5516432008-07-23 21:27:41 -0700892
Luiz Capitulino944d9fe2014-06-04 16:07:13 -0700893 if (hstate_is_gigantic(h) && !gigantic_page_supported())
894 return;
Andy Whitcroft18229df2008-11-06 12:53:27 -0800895
Andi Kleena5516432008-07-23 21:27:41 -0700896 h->nr_huge_pages--;
897 h->nr_huge_pages_node[page_to_nid(page)]--;
898 for (i = 0; i < pages_per_huge_page(h); i++) {
Chris Forbes32f84522011-07-25 17:12:14 -0700899 page[i].flags &= ~(1 << PG_locked | 1 << PG_error |
900 1 << PG_referenced | 1 << PG_dirty |
Luiz Capitulinoa7407a22014-06-04 16:07:09 -0700901 1 << PG_active | 1 << PG_private |
902 1 << PG_writeback);
Adam Litke6af2acb2007-10-16 01:26:16 -0700903 }
Sasha Levin309381fea2014-01-23 15:52:54 -0800904 VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page);
Adam Litke6af2acb2007-10-16 01:26:16 -0700905 set_compound_page_dtor(page, NULL);
906 set_page_refcounted(page);
Luiz Capitulino944d9fe2014-06-04 16:07:13 -0700907 if (hstate_is_gigantic(h)) {
908 destroy_compound_gigantic_page(page, huge_page_order(h));
909 free_gigantic_page(page, huge_page_order(h));
910 } else {
911 arch_release_hugepage(page);
912 __free_pages(page, huge_page_order(h));
913 }
Adam Litke6af2acb2007-10-16 01:26:16 -0700914}
915
Andi Kleene5ff2152008-07-23 21:27:42 -0700916struct hstate *size_to_hstate(unsigned long size)
917{
918 struct hstate *h;
919
920 for_each_hstate(h) {
921 if (huge_page_size(h) == size)
922 return h;
923 }
924 return NULL;
925}
926
Naoya Horiguchibcc54222015-04-15 16:14:38 -0700927/*
928 * Test to determine whether the hugepage is "active/in-use" (i.e. being linked
929 * to hstate->hugepage_activelist.)
930 *
931 * This function can be called for tail pages, but never returns true for them.
932 */
933bool page_huge_active(struct page *page)
934{
935 VM_BUG_ON_PAGE(!PageHuge(page), page);
936 return PageHead(page) && PagePrivate(&page[1]);
937}
938
939/* never called for tail page */
940static void set_page_huge_active(struct page *page)
941{
942 VM_BUG_ON_PAGE(!PageHeadHuge(page), page);
943 SetPagePrivate(&page[1]);
944}
945
946static void clear_page_huge_active(struct page *page)
947{
948 VM_BUG_ON_PAGE(!PageHeadHuge(page), page);
949 ClearPagePrivate(&page[1]);
950}
951
Atsushi Kumagai8f1d26d2014-07-30 16:08:39 -0700952void free_huge_page(struct page *page)
David Gibson27a85ef2006-03-22 00:08:56 -0800953{
Andi Kleena5516432008-07-23 21:27:41 -0700954 /*
955 * Can't pass hstate in here because it is called from the
956 * compound page destructor.
957 */
Andi Kleene5ff2152008-07-23 21:27:42 -0700958 struct hstate *h = page_hstate(page);
Adam Litke7893d1d2007-10-16 01:26:18 -0700959 int nid = page_to_nid(page);
David Gibson90481622012-03-21 16:34:12 -0700960 struct hugepage_subpool *spool =
961 (struct hugepage_subpool *)page_private(page);
Joonsoo Kim07443a82013-09-11 14:21:58 -0700962 bool restore_reserve;
David Gibson27a85ef2006-03-22 00:08:56 -0800963
Andy Whitcrofte5df70a2008-02-23 15:23:32 -0800964 set_page_private(page, 0);
Mel Gorman23be7462010-04-23 13:17:56 -0400965 page->mapping = NULL;
Adam Litke7893d1d2007-10-16 01:26:18 -0700966 BUG_ON(page_count(page));
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +0900967 BUG_ON(page_mapcount(page));
Joonsoo Kim07443a82013-09-11 14:21:58 -0700968 restore_reserve = PagePrivate(page);
Joonsoo Kim16c794b2013-10-16 13:46:48 -0700969 ClearPagePrivate(page);
David Gibson27a85ef2006-03-22 00:08:56 -0800970
Mike Kravetz1c5ecae2015-04-15 16:13:39 -0700971 /*
972 * A return code of zero implies that the subpool will be under its
973 * minimum size if the reservation is not restored after page is free.
974 * Therefore, force restore_reserve operation.
975 */
976 if (hugepage_subpool_put_pages(spool, 1) == 0)
977 restore_reserve = true;
978
David Gibson27a85ef2006-03-22 00:08:56 -0800979 spin_lock(&hugetlb_lock);
Naoya Horiguchibcc54222015-04-15 16:14:38 -0700980 clear_page_huge_active(page);
Aneesh Kumar K.V6d76dcf2012-07-31 16:42:18 -0700981 hugetlb_cgroup_uncharge_page(hstate_index(h),
982 pages_per_huge_page(h), page);
Joonsoo Kim07443a82013-09-11 14:21:58 -0700983 if (restore_reserve)
984 h->resv_huge_pages++;
985
Luiz Capitulino944d9fe2014-06-04 16:07:13 -0700986 if (h->surplus_huge_pages_node[nid]) {
Aneesh Kumar K.V0edaecf2012-07-31 16:42:07 -0700987 /* remove the page from active list */
988 list_del(&page->lru);
Andi Kleena5516432008-07-23 21:27:41 -0700989 update_and_free_page(h, page);
990 h->surplus_huge_pages--;
991 h->surplus_huge_pages_node[nid]--;
Adam Litke7893d1d2007-10-16 01:26:18 -0700992 } else {
Will Deacon5d3a5512012-10-08 16:29:32 -0700993 arch_clear_hugepage_flags(page);
Andi Kleena5516432008-07-23 21:27:41 -0700994 enqueue_huge_page(h, page);
Adam Litke7893d1d2007-10-16 01:26:18 -0700995 }
David Gibson27a85ef2006-03-22 00:08:56 -0800996 spin_unlock(&hugetlb_lock);
997}
998
Andi Kleena5516432008-07-23 21:27:41 -0700999static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
Andi Kleenb7ba30c2008-07-23 21:27:40 -07001000{
Aneesh Kumar K.V0edaecf2012-07-31 16:42:07 -07001001 INIT_LIST_HEAD(&page->lru);
Andi Kleenb7ba30c2008-07-23 21:27:40 -07001002 set_compound_page_dtor(page, free_huge_page);
1003 spin_lock(&hugetlb_lock);
Aneesh Kumar K.V9dd540e2012-07-31 16:42:15 -07001004 set_hugetlb_cgroup(page, NULL);
Andi Kleena5516432008-07-23 21:27:41 -07001005 h->nr_huge_pages++;
1006 h->nr_huge_pages_node[nid]++;
Andi Kleenb7ba30c2008-07-23 21:27:40 -07001007 spin_unlock(&hugetlb_lock);
1008 put_page(page); /* free it into the hugepage allocator */
1009}
1010
Luiz Capitulino2906dd52014-06-04 16:07:06 -07001011static void prep_compound_gigantic_page(struct page *page, unsigned long order)
Wu Fengguang20a03072009-06-16 15:32:22 -07001012{
1013 int i;
1014 int nr_pages = 1 << order;
1015 struct page *p = page + 1;
1016
1017 /* we rely on prep_new_huge_page to set the destructor */
1018 set_compound_order(page, order);
1019 __SetPageHead(page);
Andrea Arcangelief5a22b2013-10-16 13:46:56 -07001020 __ClearPageReserved(page);
Wu Fengguang20a03072009-06-16 15:32:22 -07001021 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
Andrea Arcangelief5a22b2013-10-16 13:46:56 -07001022 /*
1023 * For gigantic hugepages allocated through bootmem at
1024 * boot, it's safer to be consistent with the not-gigantic
1025 * hugepages and clear the PG_reserved bit from all tail pages
1026 * too. Otherwse drivers using get_user_pages() to access tail
1027 * pages may get the reference counting wrong if they see
1028 * PG_reserved set on a tail page (despite the head page not
1029 * having PG_reserved set). Enforcing this consistency between
1030 * head and tail pages allows drivers to optimize away a check
1031 * on the head page when they need know if put_page() is needed
1032 * after get_user_pages().
1033 */
1034 __ClearPageReserved(p);
Youquan Song58a84aa2011-12-08 14:34:18 -08001035 set_page_count(p, 0);
Wu Fengguang20a03072009-06-16 15:32:22 -07001036 p->first_page = page;
David Rientjes44fc8052015-03-12 16:25:54 -07001037 /* Make sure p->first_page is always valid for PageTail() */
1038 smp_wmb();
1039 __SetPageTail(p);
Wu Fengguang20a03072009-06-16 15:32:22 -07001040 }
1041}
1042
Andrew Morton77959122012-10-08 16:34:11 -07001043/*
1044 * PageHuge() only returns true for hugetlbfs pages, but not for normal or
1045 * transparent huge pages. See the PageTransHuge() documentation for more
1046 * details.
1047 */
Wu Fengguang20a03072009-06-16 15:32:22 -07001048int PageHuge(struct page *page)
1049{
Wu Fengguang20a03072009-06-16 15:32:22 -07001050 if (!PageCompound(page))
1051 return 0;
1052
1053 page = compound_head(page);
Andrew Morton758f66a2014-01-21 15:48:57 -08001054 return get_compound_page_dtor(page) == free_huge_page;
Wu Fengguang20a03072009-06-16 15:32:22 -07001055}
Naoya Horiguchi43131e12010-05-28 09:29:22 +09001056EXPORT_SYMBOL_GPL(PageHuge);
1057
Andrea Arcangeli27c73ae2013-11-21 14:32:02 -08001058/*
1059 * PageHeadHuge() only returns true for hugetlbfs head page, but not for
1060 * normal or transparent huge pages.
1061 */
1062int PageHeadHuge(struct page *page_head)
1063{
Andrea Arcangeli27c73ae2013-11-21 14:32:02 -08001064 if (!PageHead(page_head))
1065 return 0;
1066
Andrew Morton758f66a2014-01-21 15:48:57 -08001067 return get_compound_page_dtor(page_head) == free_huge_page;
Andrea Arcangeli27c73ae2013-11-21 14:32:02 -08001068}
Andrea Arcangeli27c73ae2013-11-21 14:32:02 -08001069
Zhang Yi13d60f42013-06-25 21:19:31 +08001070pgoff_t __basepage_index(struct page *page)
1071{
1072 struct page *page_head = compound_head(page);
1073 pgoff_t index = page_index(page_head);
1074 unsigned long compound_idx;
1075
1076 if (!PageHuge(page_head))
1077 return page_index(page);
1078
1079 if (compound_order(page_head) >= MAX_ORDER)
1080 compound_idx = page_to_pfn(page) - page_to_pfn(page_head);
1081 else
1082 compound_idx = page - page_head;
1083
1084 return (index << compound_order(page_head)) + compound_idx;
1085}
1086
Andi Kleena5516432008-07-23 21:27:41 -07001087static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001088{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001089 struct page *page;
Joe Jinf96efd52007-07-15 23:38:12 -07001090
Mel Gorman6484eb32009-06-16 15:31:54 -07001091 page = alloc_pages_exact_node(nid,
Naoya Horiguchi86cdb462013-09-11 14:22:13 -07001092 htlb_alloc_mask(h)|__GFP_COMP|__GFP_THISNODE|
Nishanth Aravamudan551883a2008-04-29 00:58:26 -07001093 __GFP_REPEAT|__GFP_NOWARN,
Andi Kleena5516432008-07-23 21:27:41 -07001094 huge_page_order(h));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001095 if (page) {
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07001096 if (arch_prepare_hugepage(page)) {
Gerald Schaefercaff3a22008-08-12 15:08:38 -07001097 __free_pages(page, huge_page_order(h));
Harvey Harrison7b8ee842008-04-28 14:13:19 -07001098 return NULL;
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07001099 }
Andi Kleena5516432008-07-23 21:27:41 -07001100 prep_new_huge_page(h, page, nid);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001101 }
Nishanth Aravamudan63b46132007-10-16 01:26:24 -07001102
1103 return page;
1104}
1105
Joonsoo Kimb2261022013-09-11 14:21:00 -07001106static int alloc_fresh_huge_page(struct hstate *h, nodemask_t *nodes_allowed)
1107{
1108 struct page *page;
1109 int nr_nodes, node;
1110 int ret = 0;
1111
1112 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
1113 page = alloc_fresh_huge_page_node(h, node);
1114 if (page) {
1115 ret = 1;
1116 break;
1117 }
1118 }
1119
1120 if (ret)
1121 count_vm_event(HTLB_BUDDY_PGALLOC);
1122 else
1123 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
1124
1125 return ret;
1126}
1127
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001128/*
1129 * Free huge page from pool from next node to free.
1130 * Attempt to keep persistent huge pages more or less
1131 * balanced over allowed nodes.
1132 * Called with hugetlb_lock locked.
1133 */
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001134static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
1135 bool acct_surplus)
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001136{
Joonsoo Kimb2261022013-09-11 14:21:00 -07001137 int nr_nodes, node;
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001138 int ret = 0;
1139
Joonsoo Kimb2261022013-09-11 14:21:00 -07001140 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
Lee Schermerhorn685f3452009-09-21 17:01:23 -07001141 /*
1142 * If we're returning unused surplus pages, only examine
1143 * nodes with surplus pages.
1144 */
Joonsoo Kimb2261022013-09-11 14:21:00 -07001145 if ((!acct_surplus || h->surplus_huge_pages_node[node]) &&
1146 !list_empty(&h->hugepage_freelists[node])) {
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001147 struct page *page =
Joonsoo Kimb2261022013-09-11 14:21:00 -07001148 list_entry(h->hugepage_freelists[node].next,
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001149 struct page, lru);
1150 list_del(&page->lru);
1151 h->free_huge_pages--;
Joonsoo Kimb2261022013-09-11 14:21:00 -07001152 h->free_huge_pages_node[node]--;
Lee Schermerhorn685f3452009-09-21 17:01:23 -07001153 if (acct_surplus) {
1154 h->surplus_huge_pages--;
Joonsoo Kimb2261022013-09-11 14:21:00 -07001155 h->surplus_huge_pages_node[node]--;
Lee Schermerhorn685f3452009-09-21 17:01:23 -07001156 }
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001157 update_and_free_page(h, page);
1158 ret = 1;
Lee Schermerhorn9a76db02009-12-14 17:58:15 -08001159 break;
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001160 }
Joonsoo Kimb2261022013-09-11 14:21:00 -07001161 }
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001162
1163 return ret;
1164}
1165
Naoya Horiguchic8721bb2013-09-11 14:22:09 -07001166/*
1167 * Dissolve a given free hugepage into free buddy pages. This function does
1168 * nothing for in-use (including surplus) hugepages.
1169 */
1170static void dissolve_free_huge_page(struct page *page)
1171{
1172 spin_lock(&hugetlb_lock);
1173 if (PageHuge(page) && !page_count(page)) {
1174 struct hstate *h = page_hstate(page);
1175 int nid = page_to_nid(page);
1176 list_del(&page->lru);
1177 h->free_huge_pages--;
1178 h->free_huge_pages_node[nid]--;
1179 update_and_free_page(h, page);
1180 }
1181 spin_unlock(&hugetlb_lock);
1182}
1183
1184/*
1185 * Dissolve free hugepages in a given pfn range. Used by memory hotplug to
1186 * make specified memory blocks removable from the system.
1187 * Note that start_pfn should aligned with (minimum) hugepage size.
1188 */
1189void dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn)
1190{
1191 unsigned int order = 8 * sizeof(void *);
1192 unsigned long pfn;
1193 struct hstate *h;
1194
Li Zhongd0177632014-08-06 16:07:56 -07001195 if (!hugepages_supported())
1196 return;
1197
Naoya Horiguchic8721bb2013-09-11 14:22:09 -07001198 /* Set scan step to minimum hugepage size */
1199 for_each_hstate(h)
1200 if (order > huge_page_order(h))
1201 order = huge_page_order(h);
1202 VM_BUG_ON(!IS_ALIGNED(start_pfn, 1 << order));
1203 for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << order)
1204 dissolve_free_huge_page(pfn_to_page(pfn));
1205}
1206
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001207static struct page *alloc_buddy_huge_page(struct hstate *h, int nid)
Adam Litke7893d1d2007-10-16 01:26:18 -07001208{
1209 struct page *page;
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001210 unsigned int r_nid;
Adam Litke7893d1d2007-10-16 01:26:18 -07001211
Luiz Capitulinobae7f4a2014-06-04 16:07:08 -07001212 if (hstate_is_gigantic(h))
Andi Kleenaa888a72008-07-23 21:27:47 -07001213 return NULL;
1214
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001215 /*
1216 * Assume we will successfully allocate the surplus page to
1217 * prevent racing processes from causing the surplus to exceed
1218 * overcommit
1219 *
1220 * This however introduces a different race, where a process B
1221 * tries to grow the static hugepage pool while alloc_pages() is
1222 * called by process A. B will only examine the per-node
1223 * counters in determining if surplus huge pages can be
1224 * converted to normal huge pages in adjust_pool_surplus(). A
1225 * won't be able to increment the per-node counter, until the
1226 * lock is dropped by B, but B doesn't drop hugetlb_lock until
1227 * no more huge pages can be converted from surplus to normal
1228 * state (and doesn't try to convert again). Thus, we have a
1229 * case where a surplus huge page exists, the pool is grown, and
1230 * the surplus huge page still exists after, even though it
1231 * should just have been converted to a normal huge page. This
1232 * does not leak memory, though, as the hugepage will be freed
1233 * once it is out of use. It also does not allow the counters to
1234 * go out of whack in adjust_pool_surplus() as we don't modify
1235 * the node values until we've gotten the hugepage and only the
1236 * per-node value is checked there.
1237 */
1238 spin_lock(&hugetlb_lock);
Andi Kleena5516432008-07-23 21:27:41 -07001239 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001240 spin_unlock(&hugetlb_lock);
1241 return NULL;
1242 } else {
Andi Kleena5516432008-07-23 21:27:41 -07001243 h->nr_huge_pages++;
1244 h->surplus_huge_pages++;
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001245 }
1246 spin_unlock(&hugetlb_lock);
1247
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001248 if (nid == NUMA_NO_NODE)
Naoya Horiguchi86cdb462013-09-11 14:22:13 -07001249 page = alloc_pages(htlb_alloc_mask(h)|__GFP_COMP|
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001250 __GFP_REPEAT|__GFP_NOWARN,
1251 huge_page_order(h));
1252 else
1253 page = alloc_pages_exact_node(nid,
Naoya Horiguchi86cdb462013-09-11 14:22:13 -07001254 htlb_alloc_mask(h)|__GFP_COMP|__GFP_THISNODE|
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001255 __GFP_REPEAT|__GFP_NOWARN, huge_page_order(h));
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001256
Gerald Schaefercaff3a22008-08-12 15:08:38 -07001257 if (page && arch_prepare_hugepage(page)) {
1258 __free_pages(page, huge_page_order(h));
Hillf Dantonea5768c2012-01-10 15:08:30 -08001259 page = NULL;
Gerald Schaefercaff3a22008-08-12 15:08:38 -07001260 }
1261
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001262 spin_lock(&hugetlb_lock);
Adam Litke7893d1d2007-10-16 01:26:18 -07001263 if (page) {
Aneesh Kumar K.V0edaecf2012-07-31 16:42:07 -07001264 INIT_LIST_HEAD(&page->lru);
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001265 r_nid = page_to_nid(page);
Adam Litke7893d1d2007-10-16 01:26:18 -07001266 set_compound_page_dtor(page, free_huge_page);
Aneesh Kumar K.V9dd540e2012-07-31 16:42:15 -07001267 set_hugetlb_cgroup(page, NULL);
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001268 /*
1269 * We incremented the global counters already
1270 */
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001271 h->nr_huge_pages_node[r_nid]++;
1272 h->surplus_huge_pages_node[r_nid]++;
Adam Litke3b116302008-04-28 02:13:06 -07001273 __count_vm_event(HTLB_BUDDY_PGALLOC);
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001274 } else {
Andi Kleena5516432008-07-23 21:27:41 -07001275 h->nr_huge_pages--;
1276 h->surplus_huge_pages--;
Adam Litke3b116302008-04-28 02:13:06 -07001277 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
Adam Litke7893d1d2007-10-16 01:26:18 -07001278 }
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001279 spin_unlock(&hugetlb_lock);
Adam Litke7893d1d2007-10-16 01:26:18 -07001280
1281 return page;
1282}
1283
Adam Litkee4e574b2007-10-16 01:26:19 -07001284/*
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001285 * This allocation function is useful in the context where vma is irrelevant.
1286 * E.g. soft-offlining uses this function because it only cares physical
1287 * address of error page.
1288 */
1289struct page *alloc_huge_page_node(struct hstate *h, int nid)
1290{
Joonsoo Kim4ef91842013-09-11 14:21:51 -07001291 struct page *page = NULL;
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001292
1293 spin_lock(&hugetlb_lock);
Joonsoo Kim4ef91842013-09-11 14:21:51 -07001294 if (h->free_huge_pages - h->resv_huge_pages > 0)
1295 page = dequeue_huge_page_node(h, nid);
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001296 spin_unlock(&hugetlb_lock);
1297
Aneesh Kumar K.V94ae8ba2012-07-31 16:42:35 -07001298 if (!page)
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001299 page = alloc_buddy_huge_page(h, nid);
1300
1301 return page;
1302}
1303
1304/*
Lucas De Marchi25985ed2011-03-30 22:57:33 -03001305 * Increase the hugetlb pool such that it can accommodate a reservation
Adam Litkee4e574b2007-10-16 01:26:19 -07001306 * of size 'delta'.
1307 */
Andi Kleena5516432008-07-23 21:27:41 -07001308static int gather_surplus_pages(struct hstate *h, int delta)
Adam Litkee4e574b2007-10-16 01:26:19 -07001309{
1310 struct list_head surplus_list;
1311 struct page *page, *tmp;
1312 int ret, i;
1313 int needed, allocated;
Hillf Danton28073b02012-03-21 16:34:00 -07001314 bool alloc_ok = true;
Adam Litkee4e574b2007-10-16 01:26:19 -07001315
Andi Kleena5516432008-07-23 21:27:41 -07001316 needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
Adam Litkeac09b3a2008-03-04 14:29:38 -08001317 if (needed <= 0) {
Andi Kleena5516432008-07-23 21:27:41 -07001318 h->resv_huge_pages += delta;
Adam Litkee4e574b2007-10-16 01:26:19 -07001319 return 0;
Adam Litkeac09b3a2008-03-04 14:29:38 -08001320 }
Adam Litkee4e574b2007-10-16 01:26:19 -07001321
1322 allocated = 0;
1323 INIT_LIST_HEAD(&surplus_list);
1324
1325 ret = -ENOMEM;
1326retry:
1327 spin_unlock(&hugetlb_lock);
1328 for (i = 0; i < needed; i++) {
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001329 page = alloc_buddy_huge_page(h, NUMA_NO_NODE);
Hillf Danton28073b02012-03-21 16:34:00 -07001330 if (!page) {
1331 alloc_ok = false;
1332 break;
1333 }
Adam Litkee4e574b2007-10-16 01:26:19 -07001334 list_add(&page->lru, &surplus_list);
1335 }
Hillf Danton28073b02012-03-21 16:34:00 -07001336 allocated += i;
Adam Litkee4e574b2007-10-16 01:26:19 -07001337
1338 /*
1339 * After retaking hugetlb_lock, we need to recalculate 'needed'
1340 * because either resv_huge_pages or free_huge_pages may have changed.
1341 */
1342 spin_lock(&hugetlb_lock);
Andi Kleena5516432008-07-23 21:27:41 -07001343 needed = (h->resv_huge_pages + delta) -
1344 (h->free_huge_pages + allocated);
Hillf Danton28073b02012-03-21 16:34:00 -07001345 if (needed > 0) {
1346 if (alloc_ok)
1347 goto retry;
1348 /*
1349 * We were not able to allocate enough pages to
1350 * satisfy the entire reservation so we free what
1351 * we've allocated so far.
1352 */
1353 goto free;
1354 }
Adam Litkee4e574b2007-10-16 01:26:19 -07001355 /*
1356 * The surplus_list now contains _at_least_ the number of extra pages
Lucas De Marchi25985ed2011-03-30 22:57:33 -03001357 * needed to accommodate the reservation. Add the appropriate number
Adam Litkee4e574b2007-10-16 01:26:19 -07001358 * of pages to the hugetlb pool and free the extras back to the buddy
Adam Litkeac09b3a2008-03-04 14:29:38 -08001359 * allocator. Commit the entire reservation here to prevent another
1360 * process from stealing the pages as they are added to the pool but
1361 * before they are reserved.
Adam Litkee4e574b2007-10-16 01:26:19 -07001362 */
1363 needed += allocated;
Andi Kleena5516432008-07-23 21:27:41 -07001364 h->resv_huge_pages += delta;
Adam Litkee4e574b2007-10-16 01:26:19 -07001365 ret = 0;
Naoya Horiguchia9869b82010-09-08 10:19:37 +09001366
Adam Litke19fc3f02008-04-28 02:12:20 -07001367 /* Free the needed pages to the hugetlb pool */
Adam Litkee4e574b2007-10-16 01:26:19 -07001368 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
Adam Litke19fc3f02008-04-28 02:12:20 -07001369 if ((--needed) < 0)
1370 break;
Naoya Horiguchia9869b82010-09-08 10:19:37 +09001371 /*
1372 * This page is now managed by the hugetlb allocator and has
1373 * no users -- drop the buddy allocator's reference.
1374 */
1375 put_page_testzero(page);
Sasha Levin309381fea2014-01-23 15:52:54 -08001376 VM_BUG_ON_PAGE(page_count(page), page);
Andi Kleena5516432008-07-23 21:27:41 -07001377 enqueue_huge_page(h, page);
Adam Litke19fc3f02008-04-28 02:12:20 -07001378 }
Hillf Danton28073b02012-03-21 16:34:00 -07001379free:
Hillf Dantonb0365c82011-12-28 15:57:16 -08001380 spin_unlock(&hugetlb_lock);
Adam Litke19fc3f02008-04-28 02:12:20 -07001381
1382 /* Free unnecessary surplus pages to the buddy allocator */
Joonsoo Kimc0d934b2013-09-11 14:21:02 -07001383 list_for_each_entry_safe(page, tmp, &surplus_list, lru)
1384 put_page(page);
Naoya Horiguchia9869b82010-09-08 10:19:37 +09001385 spin_lock(&hugetlb_lock);
Adam Litkee4e574b2007-10-16 01:26:19 -07001386
1387 return ret;
1388}
1389
1390/*
1391 * When releasing a hugetlb pool reservation, any surplus pages that were
1392 * allocated to satisfy the reservation must be explicitly freed if they were
1393 * never used.
Lee Schermerhorn685f3452009-09-21 17:01:23 -07001394 * Called with hugetlb_lock held.
Adam Litkee4e574b2007-10-16 01:26:19 -07001395 */
Andi Kleena5516432008-07-23 21:27:41 -07001396static void return_unused_surplus_pages(struct hstate *h,
1397 unsigned long unused_resv_pages)
Adam Litkee4e574b2007-10-16 01:26:19 -07001398{
Adam Litkee4e574b2007-10-16 01:26:19 -07001399 unsigned long nr_pages;
1400
Adam Litkeac09b3a2008-03-04 14:29:38 -08001401 /* Uncommit the reservation */
Andi Kleena5516432008-07-23 21:27:41 -07001402 h->resv_huge_pages -= unused_resv_pages;
Adam Litkeac09b3a2008-03-04 14:29:38 -08001403
Andi Kleenaa888a72008-07-23 21:27:47 -07001404 /* Cannot return gigantic pages currently */
Luiz Capitulinobae7f4a2014-06-04 16:07:08 -07001405 if (hstate_is_gigantic(h))
Andi Kleenaa888a72008-07-23 21:27:47 -07001406 return;
1407
Andi Kleena5516432008-07-23 21:27:41 -07001408 nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
Adam Litkee4e574b2007-10-16 01:26:19 -07001409
Lee Schermerhorn685f3452009-09-21 17:01:23 -07001410 /*
1411 * We want to release as many surplus pages as possible, spread
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -08001412 * evenly across all nodes with memory. Iterate across these nodes
1413 * until we can no longer free unreserved surplus pages. This occurs
1414 * when the nodes with surplus pages have no free pages.
1415 * free_pool_huge_page() will balance the the freed pages across the
1416 * on-line nodes with memory and will handle the hstate accounting.
Lee Schermerhorn685f3452009-09-21 17:01:23 -07001417 */
1418 while (nr_pages--) {
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08001419 if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1))
Lee Schermerhorn685f3452009-09-21 17:01:23 -07001420 break;
Mizuma, Masayoshi7848a4b2014-04-18 15:07:18 -07001421 cond_resched_lock(&hugetlb_lock);
Adam Litkee4e574b2007-10-16 01:26:19 -07001422 }
1423}
1424
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001425/*
1426 * Determine if the huge page at addr within the vma has an associated
1427 * reservation. Where it does not we will need to logically increase
David Gibson90481622012-03-21 16:34:12 -07001428 * reservation and actually increase subpool usage before an allocation
1429 * can occur. Where any new reservation would be required the
1430 * reservation change is prepared, but not committed. Once the page
1431 * has been allocated from the subpool and instantiated the change should
1432 * be committed via vma_commit_reservation. No action is required on
1433 * failure.
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001434 */
Roel Kluine2f17d92009-03-31 15:23:15 -07001435static long vma_needs_reservation(struct hstate *h,
Andi Kleena5516432008-07-23 21:27:41 -07001436 struct vm_area_struct *vma, unsigned long addr)
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001437{
Joonsoo Kim4e35f482014-04-03 14:47:30 -07001438 struct resv_map *resv;
1439 pgoff_t idx;
1440 long chg;
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001441
Joonsoo Kim4e35f482014-04-03 14:47:30 -07001442 resv = vma_resv_map(vma);
1443 if (!resv)
Andy Whitcroft84afd992008-07-23 21:27:32 -07001444 return 1;
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001445
Joonsoo Kim4e35f482014-04-03 14:47:30 -07001446 idx = vma_hugecache_offset(h, vma, addr);
1447 chg = region_chg(resv, idx, idx + 1);
Andy Whitcroft84afd992008-07-23 21:27:32 -07001448
Joonsoo Kim4e35f482014-04-03 14:47:30 -07001449 if (vma->vm_flags & VM_MAYSHARE)
1450 return chg;
1451 else
1452 return chg < 0 ? chg : 0;
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001453}
Andi Kleena5516432008-07-23 21:27:41 -07001454static void vma_commit_reservation(struct hstate *h,
1455 struct vm_area_struct *vma, unsigned long addr)
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001456{
Joonsoo Kim4e35f482014-04-03 14:47:30 -07001457 struct resv_map *resv;
1458 pgoff_t idx;
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001459
Joonsoo Kim4e35f482014-04-03 14:47:30 -07001460 resv = vma_resv_map(vma);
1461 if (!resv)
1462 return;
Joonsoo Kim9119a412014-04-03 14:47:25 -07001463
Joonsoo Kim4e35f482014-04-03 14:47:30 -07001464 idx = vma_hugecache_offset(h, vma, addr);
1465 region_add(resv, idx, idx + 1);
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001466}
1467
David Gibson27a85ef2006-03-22 00:08:56 -08001468static struct page *alloc_huge_page(struct vm_area_struct *vma,
Mel Gorman04f2cbe2008-07-23 21:27:25 -07001469 unsigned long addr, int avoid_reserve)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001470{
David Gibson90481622012-03-21 16:34:12 -07001471 struct hugepage_subpool *spool = subpool_vma(vma);
Andi Kleena5516432008-07-23 21:27:41 -07001472 struct hstate *h = hstate_vma(vma);
Adam Litke348ea202007-11-14 16:59:37 -08001473 struct page *page;
Roel Kluine2f17d92009-03-31 15:23:15 -07001474 long chg;
Aneesh Kumar K.V6d76dcf2012-07-31 16:42:18 -07001475 int ret, idx;
1476 struct hugetlb_cgroup *h_cg;
Adam Litke2fc39ce2007-11-14 16:59:39 -08001477
Aneesh Kumar K.V6d76dcf2012-07-31 16:42:18 -07001478 idx = hstate_index(h);
Mel Gormana1e78772008-07-23 21:27:23 -07001479 /*
David Gibson90481622012-03-21 16:34:12 -07001480 * Processes that did not create the mapping will have no
1481 * reserves and will not have accounted against subpool
1482 * limit. Check that the subpool limit can be made before
1483 * satisfying the allocation MAP_NORESERVE mappings may also
1484 * need pages and subpool limit allocated allocated if no reserve
1485 * mapping overlaps.
Mel Gormana1e78772008-07-23 21:27:23 -07001486 */
Andi Kleena5516432008-07-23 21:27:41 -07001487 chg = vma_needs_reservation(h, vma, addr);
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001488 if (chg < 0)
Aneesh Kumar K.V76dcee72012-07-31 16:41:57 -07001489 return ERR_PTR(-ENOMEM);
Joonsoo Kim8bb3f122013-09-11 14:21:54 -07001490 if (chg || avoid_reserve)
Mike Kravetz1c5ecae2015-04-15 16:13:39 -07001491 if (hugepage_subpool_get_pages(spool, 1) < 0)
Aneesh Kumar K.V76dcee72012-07-31 16:41:57 -07001492 return ERR_PTR(-ENOSPC);
Mel Gormana1e78772008-07-23 21:27:23 -07001493
Aneesh Kumar K.V6d76dcf2012-07-31 16:42:18 -07001494 ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg);
Jianyu Zhan8f34af62014-06-04 16:10:36 -07001495 if (ret)
1496 goto out_subpool_put;
1497
Mel Gormana1e78772008-07-23 21:27:23 -07001498 spin_lock(&hugetlb_lock);
Joonsoo Kimaf0ed732013-09-11 14:21:18 -07001499 page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, chg);
Joonsoo Kim81a6fca2013-09-11 14:20:58 -07001500 if (!page) {
Aneesh Kumar K.V94ae8ba2012-07-31 16:42:35 -07001501 spin_unlock(&hugetlb_lock);
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001502 page = alloc_buddy_huge_page(h, NUMA_NO_NODE);
Jianyu Zhan8f34af62014-06-04 16:10:36 -07001503 if (!page)
1504 goto out_uncharge_cgroup;
1505
Aneesh Kumar K.V79dbb232012-07-31 16:42:32 -07001506 spin_lock(&hugetlb_lock);
1507 list_move(&page->lru, &h->hugepage_activelist);
Joonsoo Kim81a6fca2013-09-11 14:20:58 -07001508 /* Fall through */
Mel Gormana1e78772008-07-23 21:27:23 -07001509 }
Joonsoo Kim81a6fca2013-09-11 14:20:58 -07001510 hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page);
1511 spin_unlock(&hugetlb_lock);
Mel Gormana1e78772008-07-23 21:27:23 -07001512
David Gibson90481622012-03-21 16:34:12 -07001513 set_page_private(page, (unsigned long)spool);
Mel Gormana1e78772008-07-23 21:27:23 -07001514
Andi Kleena5516432008-07-23 21:27:41 -07001515 vma_commit_reservation(h, vma, addr);
Adam Litke90d8b7e2007-11-14 16:59:42 -08001516 return page;
Jianyu Zhan8f34af62014-06-04 16:10:36 -07001517
1518out_uncharge_cgroup:
1519 hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg);
1520out_subpool_put:
1521 if (chg || avoid_reserve)
1522 hugepage_subpool_put_pages(spool, 1);
1523 return ERR_PTR(-ENOSPC);
David Gibsonb45b5bd2006-03-22 00:08:55 -08001524}
1525
Naoya Horiguchi74060e42013-09-11 14:22:06 -07001526/*
1527 * alloc_huge_page()'s wrapper which simply returns the page if allocation
1528 * succeeds, otherwise NULL. This function is called from new_vma_page(),
1529 * where no ERR_VALUE is expected to be returned.
1530 */
1531struct page *alloc_huge_page_noerr(struct vm_area_struct *vma,
1532 unsigned long addr, int avoid_reserve)
1533{
1534 struct page *page = alloc_huge_page(vma, addr, avoid_reserve);
1535 if (IS_ERR(page))
1536 page = NULL;
1537 return page;
1538}
1539
Cyrill Gorcunov91f47662009-01-06 14:40:33 -08001540int __weak alloc_bootmem_huge_page(struct hstate *h)
Andi Kleenaa888a72008-07-23 21:27:47 -07001541{
1542 struct huge_bootmem_page *m;
Joonsoo Kimb2261022013-09-11 14:21:00 -07001543 int nr_nodes, node;
Andi Kleenaa888a72008-07-23 21:27:47 -07001544
Joonsoo Kimb2261022013-09-11 14:21:00 -07001545 for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) {
Andi Kleenaa888a72008-07-23 21:27:47 -07001546 void *addr;
1547
Grygorii Strashko8b89a112014-01-21 15:50:36 -08001548 addr = memblock_virt_alloc_try_nid_nopanic(
1549 huge_page_size(h), huge_page_size(h),
1550 0, BOOTMEM_ALLOC_ACCESSIBLE, node);
Andi Kleenaa888a72008-07-23 21:27:47 -07001551 if (addr) {
1552 /*
1553 * Use the beginning of the huge page to store the
1554 * huge_bootmem_page struct (until gather_bootmem
1555 * puts them into the mem_map).
1556 */
1557 m = addr;
Cyrill Gorcunov91f47662009-01-06 14:40:33 -08001558 goto found;
Andi Kleenaa888a72008-07-23 21:27:47 -07001559 }
Andi Kleenaa888a72008-07-23 21:27:47 -07001560 }
1561 return 0;
1562
1563found:
Luiz Capitulinodf994ea2014-12-12 16:55:21 -08001564 BUG_ON(!IS_ALIGNED(virt_to_phys(m), huge_page_size(h)));
Andi Kleenaa888a72008-07-23 21:27:47 -07001565 /* Put them into a private list first because mem_map is not up yet */
1566 list_add(&m->list, &huge_boot_pages);
1567 m->hstate = h;
1568 return 1;
1569}
1570
David Rientjesf412c972014-04-03 14:47:59 -07001571static void __init prep_compound_huge_page(struct page *page, int order)
Andy Whitcroft18229df2008-11-06 12:53:27 -08001572{
1573 if (unlikely(order > (MAX_ORDER - 1)))
1574 prep_compound_gigantic_page(page, order);
1575 else
1576 prep_compound_page(page, order);
1577}
1578
Andi Kleenaa888a72008-07-23 21:27:47 -07001579/* Put bootmem huge pages into the standard lists after mem_map is up */
1580static void __init gather_bootmem_prealloc(void)
1581{
1582 struct huge_bootmem_page *m;
1583
1584 list_for_each_entry(m, &huge_boot_pages, list) {
Andi Kleenaa888a72008-07-23 21:27:47 -07001585 struct hstate *h = m->hstate;
Becky Bruceee8f2482011-07-25 17:11:50 -07001586 struct page *page;
1587
1588#ifdef CONFIG_HIGHMEM
1589 page = pfn_to_page(m->phys >> PAGE_SHIFT);
Grygorii Strashko8b89a112014-01-21 15:50:36 -08001590 memblock_free_late(__pa(m),
1591 sizeof(struct huge_bootmem_page));
Becky Bruceee8f2482011-07-25 17:11:50 -07001592#else
1593 page = virt_to_page(m);
1594#endif
Andi Kleenaa888a72008-07-23 21:27:47 -07001595 WARN_ON(page_count(page) != 1);
Andy Whitcroft18229df2008-11-06 12:53:27 -08001596 prep_compound_huge_page(page, h->order);
Andrea Arcangelief5a22b2013-10-16 13:46:56 -07001597 WARN_ON(PageReserved(page));
Andi Kleenaa888a72008-07-23 21:27:47 -07001598 prep_new_huge_page(h, page, page_to_nid(page));
Rafael Aquinib0320c72011-06-15 15:08:39 -07001599 /*
1600 * If we had gigantic hugepages allocated at boot time, we need
1601 * to restore the 'stolen' pages to totalram_pages in order to
1602 * fix confusing memory reports from free(1) and another
1603 * side-effects, like CommitLimit going negative.
1604 */
Luiz Capitulinobae7f4a2014-06-04 16:07:08 -07001605 if (hstate_is_gigantic(h))
Jiang Liu3dcc0572013-07-03 15:03:21 -07001606 adjust_managed_page_count(page, 1 << h->order);
Andi Kleenaa888a72008-07-23 21:27:47 -07001607 }
1608}
1609
Andi Kleen8faa8b02008-07-23 21:27:48 -07001610static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001611{
1612 unsigned long i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001613
Andi Kleene5ff2152008-07-23 21:27:42 -07001614 for (i = 0; i < h->max_huge_pages; ++i) {
Luiz Capitulinobae7f4a2014-06-04 16:07:08 -07001615 if (hstate_is_gigantic(h)) {
Andi Kleenaa888a72008-07-23 21:27:47 -07001616 if (!alloc_bootmem_huge_page(h))
1617 break;
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -08001618 } else if (!alloc_fresh_huge_page(h,
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08001619 &node_states[N_MEMORY]))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001620 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001621 }
Andi Kleen8faa8b02008-07-23 21:27:48 -07001622 h->max_huge_pages = i;
Andi Kleene5ff2152008-07-23 21:27:42 -07001623}
1624
1625static void __init hugetlb_init_hstates(void)
1626{
1627 struct hstate *h;
1628
1629 for_each_hstate(h) {
Andi Kleen8faa8b02008-07-23 21:27:48 -07001630 /* oversize hugepages were init'ed in early boot */
Luiz Capitulinobae7f4a2014-06-04 16:07:08 -07001631 if (!hstate_is_gigantic(h))
Andi Kleen8faa8b02008-07-23 21:27:48 -07001632 hugetlb_hstate_alloc_pages(h);
Andi Kleene5ff2152008-07-23 21:27:42 -07001633 }
1634}
1635
Andi Kleen4abd32d2008-07-23 21:27:49 -07001636static char * __init memfmt(char *buf, unsigned long n)
1637{
1638 if (n >= (1UL << 30))
1639 sprintf(buf, "%lu GB", n >> 30);
1640 else if (n >= (1UL << 20))
1641 sprintf(buf, "%lu MB", n >> 20);
1642 else
1643 sprintf(buf, "%lu KB", n >> 10);
1644 return buf;
1645}
1646
Andi Kleene5ff2152008-07-23 21:27:42 -07001647static void __init report_hugepages(void)
1648{
1649 struct hstate *h;
1650
1651 for_each_hstate(h) {
Andi Kleen4abd32d2008-07-23 21:27:49 -07001652 char buf[32];
Andrew Mortonffb22af2013-02-22 16:32:08 -08001653 pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n",
Andi Kleen4abd32d2008-07-23 21:27:49 -07001654 memfmt(buf, huge_page_size(h)),
1655 h->free_huge_pages);
Andi Kleene5ff2152008-07-23 21:27:42 -07001656 }
1657}
1658
Linus Torvalds1da177e2005-04-16 15:20:36 -07001659#ifdef CONFIG_HIGHMEM
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001660static void try_to_free_low(struct hstate *h, unsigned long count,
1661 nodemask_t *nodes_allowed)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001662{
Christoph Lameter4415cc82006-09-25 23:31:55 -07001663 int i;
1664
Luiz Capitulinobae7f4a2014-06-04 16:07:08 -07001665 if (hstate_is_gigantic(h))
Andi Kleenaa888a72008-07-23 21:27:47 -07001666 return;
1667
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001668 for_each_node_mask(i, *nodes_allowed) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001669 struct page *page, *next;
Andi Kleena5516432008-07-23 21:27:41 -07001670 struct list_head *freel = &h->hugepage_freelists[i];
1671 list_for_each_entry_safe(page, next, freel, lru) {
1672 if (count >= h->nr_huge_pages)
Adam Litke6b0c8802007-10-16 01:26:23 -07001673 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001674 if (PageHighMem(page))
1675 continue;
1676 list_del(&page->lru);
Andi Kleene5ff2152008-07-23 21:27:42 -07001677 update_and_free_page(h, page);
Andi Kleena5516432008-07-23 21:27:41 -07001678 h->free_huge_pages--;
1679 h->free_huge_pages_node[page_to_nid(page)]--;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001680 }
1681 }
1682}
1683#else
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001684static inline void try_to_free_low(struct hstate *h, unsigned long count,
1685 nodemask_t *nodes_allowed)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001686{
1687}
1688#endif
1689
Wu Fengguang20a03072009-06-16 15:32:22 -07001690/*
1691 * Increment or decrement surplus_huge_pages. Keep node-specific counters
1692 * balanced by operating on them in a round-robin fashion.
1693 * Returns 1 if an adjustment was made.
1694 */
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001695static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
1696 int delta)
Wu Fengguang20a03072009-06-16 15:32:22 -07001697{
Joonsoo Kimb2261022013-09-11 14:21:00 -07001698 int nr_nodes, node;
Wu Fengguang20a03072009-06-16 15:32:22 -07001699
1700 VM_BUG_ON(delta != -1 && delta != 1);
Wu Fengguang20a03072009-06-16 15:32:22 -07001701
Joonsoo Kimb2261022013-09-11 14:21:00 -07001702 if (delta < 0) {
1703 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
1704 if (h->surplus_huge_pages_node[node])
1705 goto found;
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001706 }
Joonsoo Kimb2261022013-09-11 14:21:00 -07001707 } else {
1708 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
1709 if (h->surplus_huge_pages_node[node] <
1710 h->nr_huge_pages_node[node])
1711 goto found;
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001712 }
Joonsoo Kimb2261022013-09-11 14:21:00 -07001713 }
1714 return 0;
Wu Fengguang20a03072009-06-16 15:32:22 -07001715
Joonsoo Kimb2261022013-09-11 14:21:00 -07001716found:
1717 h->surplus_huge_pages += delta;
1718 h->surplus_huge_pages_node[node] += delta;
1719 return 1;
Wu Fengguang20a03072009-06-16 15:32:22 -07001720}
1721
Andi Kleena5516432008-07-23 21:27:41 -07001722#define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001723static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count,
1724 nodemask_t *nodes_allowed)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001725{
Adam Litke7893d1d2007-10-16 01:26:18 -07001726 unsigned long min_count, ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001727
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001728 if (hstate_is_gigantic(h) && !gigantic_page_supported())
Andi Kleenaa888a72008-07-23 21:27:47 -07001729 return h->max_huge_pages;
1730
Adam Litke7893d1d2007-10-16 01:26:18 -07001731 /*
1732 * Increase the pool size
1733 * First take pages out of surplus state. Then make up the
1734 * remaining difference by allocating fresh huge pages.
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001735 *
1736 * We might race with alloc_buddy_huge_page() here and be unable
1737 * to convert a surplus huge page to a normal huge page. That is
1738 * not critical, though, it just means the overall size of the
1739 * pool might be one hugepage larger than it needs to be, but
1740 * within all the constraints specified by the sysctls.
Adam Litke7893d1d2007-10-16 01:26:18 -07001741 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001742 spin_lock(&hugetlb_lock);
Andi Kleena5516432008-07-23 21:27:41 -07001743 while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001744 if (!adjust_pool_surplus(h, nodes_allowed, -1))
Adam Litke7893d1d2007-10-16 01:26:18 -07001745 break;
1746 }
1747
Andi Kleena5516432008-07-23 21:27:41 -07001748 while (count > persistent_huge_pages(h)) {
Adam Litke7893d1d2007-10-16 01:26:18 -07001749 /*
1750 * If this allocation races such that we no longer need the
1751 * page, free_huge_page will handle it by freeing the page
1752 * and reducing the surplus.
1753 */
1754 spin_unlock(&hugetlb_lock);
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001755 if (hstate_is_gigantic(h))
1756 ret = alloc_fresh_gigantic_page(h, nodes_allowed);
1757 else
1758 ret = alloc_fresh_huge_page(h, nodes_allowed);
Adam Litke7893d1d2007-10-16 01:26:18 -07001759 spin_lock(&hugetlb_lock);
1760 if (!ret)
1761 goto out;
1762
Mel Gorman536240f22009-12-14 17:59:56 -08001763 /* Bail for signals. Probably ctrl-c from user */
1764 if (signal_pending(current))
1765 goto out;
Adam Litke7893d1d2007-10-16 01:26:18 -07001766 }
Adam Litke7893d1d2007-10-16 01:26:18 -07001767
1768 /*
1769 * Decrease the pool size
1770 * First return free pages to the buddy allocator (being careful
1771 * to keep enough around to satisfy reservations). Then place
1772 * pages into surplus state as needed so the pool will shrink
1773 * to the desired size as pages become free.
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001774 *
1775 * By placing pages into the surplus state independent of the
1776 * overcommit value, we are allowing the surplus pool size to
1777 * exceed overcommit. There are few sane options here. Since
1778 * alloc_buddy_huge_page() is checking the global counter,
1779 * though, we'll note that we're not allowed to exceed surplus
1780 * and won't grow the pool anywhere else. Not until one of the
1781 * sysctls are changed, or the surplus pages go out of use.
Adam Litke7893d1d2007-10-16 01:26:18 -07001782 */
Andi Kleena5516432008-07-23 21:27:41 -07001783 min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
Adam Litke6b0c8802007-10-16 01:26:23 -07001784 min_count = max(count, min_count);
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001785 try_to_free_low(h, min_count, nodes_allowed);
Andi Kleena5516432008-07-23 21:27:41 -07001786 while (min_count < persistent_huge_pages(h)) {
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001787 if (!free_pool_huge_page(h, nodes_allowed, 0))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001788 break;
Mizuma, Masayoshi55f67142014-04-07 15:37:54 -07001789 cond_resched_lock(&hugetlb_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001790 }
Andi Kleena5516432008-07-23 21:27:41 -07001791 while (count < persistent_huge_pages(h)) {
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001792 if (!adjust_pool_surplus(h, nodes_allowed, 1))
Adam Litke7893d1d2007-10-16 01:26:18 -07001793 break;
1794 }
1795out:
Andi Kleena5516432008-07-23 21:27:41 -07001796 ret = persistent_huge_pages(h);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001797 spin_unlock(&hugetlb_lock);
Adam Litke7893d1d2007-10-16 01:26:18 -07001798 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001799}
1800
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001801#define HSTATE_ATTR_RO(_name) \
1802 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
1803
1804#define HSTATE_ATTR(_name) \
1805 static struct kobj_attribute _name##_attr = \
1806 __ATTR(_name, 0644, _name##_show, _name##_store)
1807
1808static struct kobject *hugepages_kobj;
1809static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
1810
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001811static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
1812
1813static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001814{
1815 int i;
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001816
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001817 for (i = 0; i < HUGE_MAX_HSTATE; i++)
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001818 if (hstate_kobjs[i] == kobj) {
1819 if (nidp)
1820 *nidp = NUMA_NO_NODE;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001821 return &hstates[i];
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001822 }
1823
1824 return kobj_to_node_hstate(kobj, nidp);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001825}
1826
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001827static ssize_t nr_hugepages_show_common(struct kobject *kobj,
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001828 struct kobj_attribute *attr, char *buf)
1829{
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001830 struct hstate *h;
1831 unsigned long nr_huge_pages;
1832 int nid;
1833
1834 h = kobj_to_hstate(kobj, &nid);
1835 if (nid == NUMA_NO_NODE)
1836 nr_huge_pages = h->nr_huge_pages;
1837 else
1838 nr_huge_pages = h->nr_huge_pages_node[nid];
1839
1840 return sprintf(buf, "%lu\n", nr_huge_pages);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001841}
Eric B Munsonadbe8722011-01-13 15:47:27 -08001842
David Rientjes238d3c12014-08-06 16:06:51 -07001843static ssize_t __nr_hugepages_store_common(bool obey_mempolicy,
1844 struct hstate *h, int nid,
1845 unsigned long count, size_t len)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001846{
1847 int err;
David Rientjesbad44b52009-12-14 17:58:38 -08001848 NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001849
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001850 if (hstate_is_gigantic(h) && !gigantic_page_supported()) {
Eric B Munsonadbe8722011-01-13 15:47:27 -08001851 err = -EINVAL;
1852 goto out;
1853 }
1854
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001855 if (nid == NUMA_NO_NODE) {
1856 /*
1857 * global hstate attribute
1858 */
1859 if (!(obey_mempolicy &&
1860 init_nodemask_of_mempolicy(nodes_allowed))) {
1861 NODEMASK_FREE(nodes_allowed);
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08001862 nodes_allowed = &node_states[N_MEMORY];
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001863 }
1864 } else if (nodes_allowed) {
1865 /*
1866 * per node hstate attribute: adjust count to global,
1867 * but restrict alloc/free to the specified node.
1868 */
1869 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
1870 init_nodemask_of_node(nodes_allowed, nid);
1871 } else
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08001872 nodes_allowed = &node_states[N_MEMORY];
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001873
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001874 h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001875
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08001876 if (nodes_allowed != &node_states[N_MEMORY])
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001877 NODEMASK_FREE(nodes_allowed);
1878
1879 return len;
Eric B Munsonadbe8722011-01-13 15:47:27 -08001880out:
1881 NODEMASK_FREE(nodes_allowed);
1882 return err;
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001883}
1884
David Rientjes238d3c12014-08-06 16:06:51 -07001885static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
1886 struct kobject *kobj, const char *buf,
1887 size_t len)
1888{
1889 struct hstate *h;
1890 unsigned long count;
1891 int nid;
1892 int err;
1893
1894 err = kstrtoul(buf, 10, &count);
1895 if (err)
1896 return err;
1897
1898 h = kobj_to_hstate(kobj, &nid);
1899 return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len);
1900}
1901
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001902static ssize_t nr_hugepages_show(struct kobject *kobj,
1903 struct kobj_attribute *attr, char *buf)
1904{
1905 return nr_hugepages_show_common(kobj, attr, buf);
1906}
1907
1908static ssize_t nr_hugepages_store(struct kobject *kobj,
1909 struct kobj_attribute *attr, const char *buf, size_t len)
1910{
David Rientjes238d3c12014-08-06 16:06:51 -07001911 return nr_hugepages_store_common(false, kobj, buf, len);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001912}
1913HSTATE_ATTR(nr_hugepages);
1914
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001915#ifdef CONFIG_NUMA
1916
1917/*
1918 * hstate attribute for optionally mempolicy-based constraint on persistent
1919 * huge page alloc/free.
1920 */
1921static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
1922 struct kobj_attribute *attr, char *buf)
1923{
1924 return nr_hugepages_show_common(kobj, attr, buf);
1925}
1926
1927static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
1928 struct kobj_attribute *attr, const char *buf, size_t len)
1929{
David Rientjes238d3c12014-08-06 16:06:51 -07001930 return nr_hugepages_store_common(true, kobj, buf, len);
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001931}
1932HSTATE_ATTR(nr_hugepages_mempolicy);
1933#endif
1934
1935
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001936static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
1937 struct kobj_attribute *attr, char *buf)
1938{
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001939 struct hstate *h = kobj_to_hstate(kobj, NULL);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001940 return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
1941}
Eric B Munsonadbe8722011-01-13 15:47:27 -08001942
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001943static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
1944 struct kobj_attribute *attr, const char *buf, size_t count)
1945{
1946 int err;
1947 unsigned long input;
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001948 struct hstate *h = kobj_to_hstate(kobj, NULL);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001949
Luiz Capitulinobae7f4a2014-06-04 16:07:08 -07001950 if (hstate_is_gigantic(h))
Eric B Munsonadbe8722011-01-13 15:47:27 -08001951 return -EINVAL;
1952
Jingoo Han3dbb95f2013-09-11 14:20:25 -07001953 err = kstrtoul(buf, 10, &input);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001954 if (err)
Eric B Munson73ae31e2011-01-13 15:47:28 -08001955 return err;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001956
1957 spin_lock(&hugetlb_lock);
1958 h->nr_overcommit_huge_pages = input;
1959 spin_unlock(&hugetlb_lock);
1960
1961 return count;
1962}
1963HSTATE_ATTR(nr_overcommit_hugepages);
1964
1965static ssize_t free_hugepages_show(struct kobject *kobj,
1966 struct kobj_attribute *attr, char *buf)
1967{
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001968 struct hstate *h;
1969 unsigned long free_huge_pages;
1970 int nid;
1971
1972 h = kobj_to_hstate(kobj, &nid);
1973 if (nid == NUMA_NO_NODE)
1974 free_huge_pages = h->free_huge_pages;
1975 else
1976 free_huge_pages = h->free_huge_pages_node[nid];
1977
1978 return sprintf(buf, "%lu\n", free_huge_pages);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001979}
1980HSTATE_ATTR_RO(free_hugepages);
1981
1982static ssize_t resv_hugepages_show(struct kobject *kobj,
1983 struct kobj_attribute *attr, char *buf)
1984{
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001985 struct hstate *h = kobj_to_hstate(kobj, NULL);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001986 return sprintf(buf, "%lu\n", h->resv_huge_pages);
1987}
1988HSTATE_ATTR_RO(resv_hugepages);
1989
1990static ssize_t surplus_hugepages_show(struct kobject *kobj,
1991 struct kobj_attribute *attr, char *buf)
1992{
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001993 struct hstate *h;
1994 unsigned long surplus_huge_pages;
1995 int nid;
1996
1997 h = kobj_to_hstate(kobj, &nid);
1998 if (nid == NUMA_NO_NODE)
1999 surplus_huge_pages = h->surplus_huge_pages;
2000 else
2001 surplus_huge_pages = h->surplus_huge_pages_node[nid];
2002
2003 return sprintf(buf, "%lu\n", surplus_huge_pages);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002004}
2005HSTATE_ATTR_RO(surplus_hugepages);
2006
2007static struct attribute *hstate_attrs[] = {
2008 &nr_hugepages_attr.attr,
2009 &nr_overcommit_hugepages_attr.attr,
2010 &free_hugepages_attr.attr,
2011 &resv_hugepages_attr.attr,
2012 &surplus_hugepages_attr.attr,
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002013#ifdef CONFIG_NUMA
2014 &nr_hugepages_mempolicy_attr.attr,
2015#endif
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002016 NULL,
2017};
2018
2019static struct attribute_group hstate_attr_group = {
2020 .attrs = hstate_attrs,
2021};
2022
Jeff Mahoney094e9532010-02-02 13:44:14 -08002023static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
2024 struct kobject **hstate_kobjs,
2025 struct attribute_group *hstate_attr_group)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002026{
2027 int retval;
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07002028 int hi = hstate_index(h);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002029
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002030 hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
2031 if (!hstate_kobjs[hi])
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002032 return -ENOMEM;
2033
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002034 retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002035 if (retval)
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002036 kobject_put(hstate_kobjs[hi]);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002037
2038 return retval;
2039}
2040
2041static void __init hugetlb_sysfs_init(void)
2042{
2043 struct hstate *h;
2044 int err;
2045
2046 hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
2047 if (!hugepages_kobj)
2048 return;
2049
2050 for_each_hstate(h) {
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002051 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
2052 hstate_kobjs, &hstate_attr_group);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002053 if (err)
Andrew Mortonffb22af2013-02-22 16:32:08 -08002054 pr_err("Hugetlb: Unable to add hstate %s", h->name);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002055 }
2056}
2057
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002058#ifdef CONFIG_NUMA
2059
2060/*
2061 * node_hstate/s - associate per node hstate attributes, via their kobjects,
Kay Sievers10fbcf42011-12-21 14:48:43 -08002062 * with node devices in node_devices[] using a parallel array. The array
2063 * index of a node device or _hstate == node id.
2064 * This is here to avoid any static dependency of the node device driver, in
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002065 * the base kernel, on the hugetlb module.
2066 */
2067struct node_hstate {
2068 struct kobject *hugepages_kobj;
2069 struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
2070};
2071struct node_hstate node_hstates[MAX_NUMNODES];
2072
2073/*
Kay Sievers10fbcf42011-12-21 14:48:43 -08002074 * A subset of global hstate attributes for node devices
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002075 */
2076static struct attribute *per_node_hstate_attrs[] = {
2077 &nr_hugepages_attr.attr,
2078 &free_hugepages_attr.attr,
2079 &surplus_hugepages_attr.attr,
2080 NULL,
2081};
2082
2083static struct attribute_group per_node_hstate_attr_group = {
2084 .attrs = per_node_hstate_attrs,
2085};
2086
2087/*
Kay Sievers10fbcf42011-12-21 14:48:43 -08002088 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002089 * Returns node id via non-NULL nidp.
2090 */
2091static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
2092{
2093 int nid;
2094
2095 for (nid = 0; nid < nr_node_ids; nid++) {
2096 struct node_hstate *nhs = &node_hstates[nid];
2097 int i;
2098 for (i = 0; i < HUGE_MAX_HSTATE; i++)
2099 if (nhs->hstate_kobjs[i] == kobj) {
2100 if (nidp)
2101 *nidp = nid;
2102 return &hstates[i];
2103 }
2104 }
2105
2106 BUG();
2107 return NULL;
2108}
2109
2110/*
Kay Sievers10fbcf42011-12-21 14:48:43 -08002111 * Unregister hstate attributes from a single node device.
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002112 * No-op if no hstate attributes attached.
2113 */
Claudiu Ghioc3cd8b442013-03-04 12:46:15 +02002114static void hugetlb_unregister_node(struct node *node)
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002115{
2116 struct hstate *h;
Kay Sievers10fbcf42011-12-21 14:48:43 -08002117 struct node_hstate *nhs = &node_hstates[node->dev.id];
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002118
2119 if (!nhs->hugepages_kobj)
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -08002120 return; /* no hstate attributes */
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002121
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07002122 for_each_hstate(h) {
2123 int idx = hstate_index(h);
2124 if (nhs->hstate_kobjs[idx]) {
2125 kobject_put(nhs->hstate_kobjs[idx]);
2126 nhs->hstate_kobjs[idx] = NULL;
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002127 }
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07002128 }
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002129
2130 kobject_put(nhs->hugepages_kobj);
2131 nhs->hugepages_kobj = NULL;
2132}
2133
2134/*
Kay Sievers10fbcf42011-12-21 14:48:43 -08002135 * hugetlb module exit: unregister hstate attributes from node devices
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002136 * that have them.
2137 */
2138static void hugetlb_unregister_all_nodes(void)
2139{
2140 int nid;
2141
2142 /*
Kay Sievers10fbcf42011-12-21 14:48:43 -08002143 * disable node device registrations.
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002144 */
2145 register_hugetlbfs_with_node(NULL, NULL);
2146
2147 /*
2148 * remove hstate attributes from any nodes that have them.
2149 */
2150 for (nid = 0; nid < nr_node_ids; nid++)
Wen Congyang87327942012-12-11 16:00:56 -08002151 hugetlb_unregister_node(node_devices[nid]);
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002152}
2153
2154/*
Kay Sievers10fbcf42011-12-21 14:48:43 -08002155 * Register hstate attributes for a single node device.
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002156 * No-op if attributes already registered.
2157 */
Claudiu Ghioc3cd8b442013-03-04 12:46:15 +02002158static void hugetlb_register_node(struct node *node)
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002159{
2160 struct hstate *h;
Kay Sievers10fbcf42011-12-21 14:48:43 -08002161 struct node_hstate *nhs = &node_hstates[node->dev.id];
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002162 int err;
2163
2164 if (nhs->hugepages_kobj)
2165 return; /* already allocated */
2166
2167 nhs->hugepages_kobj = kobject_create_and_add("hugepages",
Kay Sievers10fbcf42011-12-21 14:48:43 -08002168 &node->dev.kobj);
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002169 if (!nhs->hugepages_kobj)
2170 return;
2171
2172 for_each_hstate(h) {
2173 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
2174 nhs->hstate_kobjs,
2175 &per_node_hstate_attr_group);
2176 if (err) {
Andrew Mortonffb22af2013-02-22 16:32:08 -08002177 pr_err("Hugetlb: Unable to add hstate %s for node %d\n",
2178 h->name, node->dev.id);
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002179 hugetlb_unregister_node(node);
2180 break;
2181 }
2182 }
2183}
2184
2185/*
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -08002186 * hugetlb init time: register hstate attributes for all registered node
Kay Sievers10fbcf42011-12-21 14:48:43 -08002187 * devices of nodes that have memory. All on-line nodes should have
2188 * registered their associated device by this time.
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002189 */
Luiz Capitulino7d9ca002014-12-12 16:55:24 -08002190static void __init hugetlb_register_all_nodes(void)
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002191{
2192 int nid;
2193
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08002194 for_each_node_state(nid, N_MEMORY) {
Wen Congyang87327942012-12-11 16:00:56 -08002195 struct node *node = node_devices[nid];
Kay Sievers10fbcf42011-12-21 14:48:43 -08002196 if (node->dev.id == nid)
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002197 hugetlb_register_node(node);
2198 }
2199
2200 /*
Kay Sievers10fbcf42011-12-21 14:48:43 -08002201 * Let the node device driver know we're here so it can
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002202 * [un]register hstate attributes on node hotplug.
2203 */
2204 register_hugetlbfs_with_node(hugetlb_register_node,
2205 hugetlb_unregister_node);
2206}
2207#else /* !CONFIG_NUMA */
2208
2209static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
2210{
2211 BUG();
2212 if (nidp)
2213 *nidp = -1;
2214 return NULL;
2215}
2216
2217static void hugetlb_unregister_all_nodes(void) { }
2218
2219static void hugetlb_register_all_nodes(void) { }
2220
2221#endif
2222
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002223static void __exit hugetlb_exit(void)
2224{
2225 struct hstate *h;
2226
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002227 hugetlb_unregister_all_nodes();
2228
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002229 for_each_hstate(h) {
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07002230 kobject_put(hstate_kobjs[hstate_index(h)]);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002231 }
2232
2233 kobject_put(hugepages_kobj);
Davidlohr Bueso8382d912014-04-03 14:47:31 -07002234 kfree(htlb_fault_mutex_table);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002235}
2236module_exit(hugetlb_exit);
2237
2238static int __init hugetlb_init(void)
2239{
Davidlohr Bueso8382d912014-04-03 14:47:31 -07002240 int i;
2241
Nishanth Aravamudan457c1b22014-05-06 12:50:00 -07002242 if (!hugepages_supported())
Benjamin Herrenschmidt0ef89d22008-07-31 00:07:30 -07002243 return 0;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002244
Nick Piggine11bfbf2008-07-23 21:27:52 -07002245 if (!size_to_hstate(default_hstate_size)) {
2246 default_hstate_size = HPAGE_SIZE;
2247 if (!size_to_hstate(default_hstate_size))
2248 hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002249 }
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07002250 default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size));
Nick Piggine11bfbf2008-07-23 21:27:52 -07002251 if (default_hstate_max_huge_pages)
2252 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002253
2254 hugetlb_init_hstates();
Andi Kleenaa888a72008-07-23 21:27:47 -07002255 gather_bootmem_prealloc();
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002256 report_hugepages();
2257
2258 hugetlb_sysfs_init();
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002259 hugetlb_register_all_nodes();
Jianguo Wu7179e7b2012-12-18 14:23:19 -08002260 hugetlb_cgroup_file_init();
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002261
Davidlohr Bueso8382d912014-04-03 14:47:31 -07002262#ifdef CONFIG_SMP
2263 num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus());
2264#else
2265 num_fault_mutexes = 1;
2266#endif
2267 htlb_fault_mutex_table =
2268 kmalloc(sizeof(struct mutex) * num_fault_mutexes, GFP_KERNEL);
2269 BUG_ON(!htlb_fault_mutex_table);
2270
2271 for (i = 0; i < num_fault_mutexes; i++)
2272 mutex_init(&htlb_fault_mutex_table[i]);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002273 return 0;
2274}
2275module_init(hugetlb_init);
2276
2277/* Should be called on processing a hugepagesz=... option */
2278void __init hugetlb_add_hstate(unsigned order)
2279{
2280 struct hstate *h;
Andi Kleen8faa8b02008-07-23 21:27:48 -07002281 unsigned long i;
2282
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002283 if (size_to_hstate(PAGE_SIZE << order)) {
Andrew Mortonffb22af2013-02-22 16:32:08 -08002284 pr_warning("hugepagesz= specified twice, ignoring\n");
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002285 return;
2286 }
Aneesh Kumar K.V47d38342012-07-31 16:41:54 -07002287 BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002288 BUG_ON(order == 0);
Aneesh Kumar K.V47d38342012-07-31 16:41:54 -07002289 h = &hstates[hugetlb_max_hstate++];
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002290 h->order = order;
2291 h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
Andi Kleen8faa8b02008-07-23 21:27:48 -07002292 h->nr_huge_pages = 0;
2293 h->free_huge_pages = 0;
2294 for (i = 0; i < MAX_NUMNODES; ++i)
2295 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
Aneesh Kumar K.V0edaecf2012-07-31 16:42:07 -07002296 INIT_LIST_HEAD(&h->hugepage_activelist);
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08002297 h->next_nid_to_alloc = first_node(node_states[N_MEMORY]);
2298 h->next_nid_to_free = first_node(node_states[N_MEMORY]);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002299 snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
2300 huge_page_size(h)/1024);
Andi Kleen8faa8b02008-07-23 21:27:48 -07002301
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002302 parsed_hstate = h;
2303}
2304
Nick Piggine11bfbf2008-07-23 21:27:52 -07002305static int __init hugetlb_nrpages_setup(char *s)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002306{
2307 unsigned long *mhp;
Andi Kleen8faa8b02008-07-23 21:27:48 -07002308 static unsigned long *last_mhp;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002309
2310 /*
Aneesh Kumar K.V47d38342012-07-31 16:41:54 -07002311 * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet,
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002312 * so this hugepages= parameter goes to the "default hstate".
2313 */
Aneesh Kumar K.V47d38342012-07-31 16:41:54 -07002314 if (!hugetlb_max_hstate)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002315 mhp = &default_hstate_max_huge_pages;
2316 else
2317 mhp = &parsed_hstate->max_huge_pages;
2318
Andi Kleen8faa8b02008-07-23 21:27:48 -07002319 if (mhp == last_mhp) {
Andrew Mortonffb22af2013-02-22 16:32:08 -08002320 pr_warning("hugepages= specified twice without "
2321 "interleaving hugepagesz=, ignoring\n");
Andi Kleen8faa8b02008-07-23 21:27:48 -07002322 return 1;
2323 }
2324
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002325 if (sscanf(s, "%lu", mhp) <= 0)
2326 *mhp = 0;
2327
Andi Kleen8faa8b02008-07-23 21:27:48 -07002328 /*
2329 * Global state is always initialized later in hugetlb_init.
2330 * But we need to allocate >= MAX_ORDER hstates here early to still
2331 * use the bootmem allocator.
2332 */
Aneesh Kumar K.V47d38342012-07-31 16:41:54 -07002333 if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER)
Andi Kleen8faa8b02008-07-23 21:27:48 -07002334 hugetlb_hstate_alloc_pages(parsed_hstate);
2335
2336 last_mhp = mhp;
2337
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002338 return 1;
2339}
Nick Piggine11bfbf2008-07-23 21:27:52 -07002340__setup("hugepages=", hugetlb_nrpages_setup);
2341
2342static int __init hugetlb_default_setup(char *s)
2343{
2344 default_hstate_size = memparse(s, &s);
2345 return 1;
2346}
2347__setup("default_hugepagesz=", hugetlb_default_setup);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002348
Nishanth Aravamudan8a213462008-07-25 19:44:37 -07002349static unsigned int cpuset_mems_nr(unsigned int *array)
2350{
2351 int node;
2352 unsigned int nr = 0;
2353
2354 for_each_node_mask(node, cpuset_current_mems_allowed)
2355 nr += array[node];
2356
2357 return nr;
2358}
2359
2360#ifdef CONFIG_SYSCTL
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002361static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
2362 struct ctl_table *table, int write,
2363 void __user *buffer, size_t *length, loff_t *ppos)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002364{
Andi Kleene5ff2152008-07-23 21:27:42 -07002365 struct hstate *h = &default_hstate;
David Rientjes238d3c12014-08-06 16:06:51 -07002366 unsigned long tmp = h->max_huge_pages;
Michal Hocko08d4a242011-01-13 15:47:26 -08002367 int ret;
Andi Kleene5ff2152008-07-23 21:27:42 -07002368
Nishanth Aravamudan457c1b22014-05-06 12:50:00 -07002369 if (!hugepages_supported())
2370 return -ENOTSUPP;
2371
Andi Kleene5ff2152008-07-23 21:27:42 -07002372 table->data = &tmp;
2373 table->maxlen = sizeof(unsigned long);
Michal Hocko08d4a242011-01-13 15:47:26 -08002374 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2375 if (ret)
2376 goto out;
Andi Kleene5ff2152008-07-23 21:27:42 -07002377
David Rientjes238d3c12014-08-06 16:06:51 -07002378 if (write)
2379 ret = __nr_hugepages_store_common(obey_mempolicy, h,
2380 NUMA_NO_NODE, tmp, *length);
Michal Hocko08d4a242011-01-13 15:47:26 -08002381out:
2382 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002383}
Mel Gorman396faf02007-07-17 04:03:13 -07002384
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002385int hugetlb_sysctl_handler(struct ctl_table *table, int write,
2386 void __user *buffer, size_t *length, loff_t *ppos)
2387{
2388
2389 return hugetlb_sysctl_handler_common(false, table, write,
2390 buffer, length, ppos);
2391}
2392
2393#ifdef CONFIG_NUMA
2394int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
2395 void __user *buffer, size_t *length, loff_t *ppos)
2396{
2397 return hugetlb_sysctl_handler_common(true, table, write,
2398 buffer, length, ppos);
2399}
2400#endif /* CONFIG_NUMA */
2401
Nishanth Aravamudana3d0c6a2008-02-08 04:18:18 -08002402int hugetlb_overcommit_handler(struct ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07002403 void __user *buffer,
Nishanth Aravamudana3d0c6a2008-02-08 04:18:18 -08002404 size_t *length, loff_t *ppos)
2405{
Andi Kleena5516432008-07-23 21:27:41 -07002406 struct hstate *h = &default_hstate;
Andi Kleene5ff2152008-07-23 21:27:42 -07002407 unsigned long tmp;
Michal Hocko08d4a242011-01-13 15:47:26 -08002408 int ret;
Andi Kleene5ff2152008-07-23 21:27:42 -07002409
Nishanth Aravamudan457c1b22014-05-06 12:50:00 -07002410 if (!hugepages_supported())
2411 return -ENOTSUPP;
2412
Petr Holasekc033a932011-03-22 16:33:05 -07002413 tmp = h->nr_overcommit_huge_pages;
Andi Kleene5ff2152008-07-23 21:27:42 -07002414
Luiz Capitulinobae7f4a2014-06-04 16:07:08 -07002415 if (write && hstate_is_gigantic(h))
Eric B Munsonadbe8722011-01-13 15:47:27 -08002416 return -EINVAL;
2417
Andi Kleene5ff2152008-07-23 21:27:42 -07002418 table->data = &tmp;
2419 table->maxlen = sizeof(unsigned long);
Michal Hocko08d4a242011-01-13 15:47:26 -08002420 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2421 if (ret)
2422 goto out;
Andi Kleene5ff2152008-07-23 21:27:42 -07002423
2424 if (write) {
2425 spin_lock(&hugetlb_lock);
2426 h->nr_overcommit_huge_pages = tmp;
2427 spin_unlock(&hugetlb_lock);
2428 }
Michal Hocko08d4a242011-01-13 15:47:26 -08002429out:
2430 return ret;
Nishanth Aravamudana3d0c6a2008-02-08 04:18:18 -08002431}
2432
Linus Torvalds1da177e2005-04-16 15:20:36 -07002433#endif /* CONFIG_SYSCTL */
2434
Alexey Dobriyane1759c22008-10-15 23:50:22 +04002435void hugetlb_report_meminfo(struct seq_file *m)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002436{
Andi Kleena5516432008-07-23 21:27:41 -07002437 struct hstate *h = &default_hstate;
Nishanth Aravamudan457c1b22014-05-06 12:50:00 -07002438 if (!hugepages_supported())
2439 return;
Alexey Dobriyane1759c22008-10-15 23:50:22 +04002440 seq_printf(m,
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002441 "HugePages_Total: %5lu\n"
2442 "HugePages_Free: %5lu\n"
2443 "HugePages_Rsvd: %5lu\n"
2444 "HugePages_Surp: %5lu\n"
2445 "Hugepagesize: %8lu kB\n",
Andi Kleena5516432008-07-23 21:27:41 -07002446 h->nr_huge_pages,
2447 h->free_huge_pages,
2448 h->resv_huge_pages,
2449 h->surplus_huge_pages,
2450 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002451}
2452
2453int hugetlb_report_node_meminfo(int nid, char *buf)
2454{
Andi Kleena5516432008-07-23 21:27:41 -07002455 struct hstate *h = &default_hstate;
Nishanth Aravamudan457c1b22014-05-06 12:50:00 -07002456 if (!hugepages_supported())
2457 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002458 return sprintf(buf,
2459 "Node %d HugePages_Total: %5u\n"
Nishanth Aravamudana1de0912008-03-26 14:37:53 -07002460 "Node %d HugePages_Free: %5u\n"
2461 "Node %d HugePages_Surp: %5u\n",
Andi Kleena5516432008-07-23 21:27:41 -07002462 nid, h->nr_huge_pages_node[nid],
2463 nid, h->free_huge_pages_node[nid],
2464 nid, h->surplus_huge_pages_node[nid]);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002465}
2466
David Rientjes949f7ec2013-04-29 15:07:48 -07002467void hugetlb_show_meminfo(void)
2468{
2469 struct hstate *h;
2470 int nid;
2471
Nishanth Aravamudan457c1b22014-05-06 12:50:00 -07002472 if (!hugepages_supported())
2473 return;
2474
David Rientjes949f7ec2013-04-29 15:07:48 -07002475 for_each_node_state(nid, N_MEMORY)
2476 for_each_hstate(h)
2477 pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
2478 nid,
2479 h->nr_huge_pages_node[nid],
2480 h->free_huge_pages_node[nid],
2481 h->surplus_huge_pages_node[nid],
2482 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
2483}
2484
Linus Torvalds1da177e2005-04-16 15:20:36 -07002485/* Return the number pages of memory we physically have, in PAGE_SIZE units. */
2486unsigned long hugetlb_total_pages(void)
2487{
Wanpeng Lid0028582013-03-22 15:04:40 -07002488 struct hstate *h;
2489 unsigned long nr_total_pages = 0;
2490
2491 for_each_hstate(h)
2492 nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h);
2493 return nr_total_pages;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002494}
Linus Torvalds1da177e2005-04-16 15:20:36 -07002495
Andi Kleena5516432008-07-23 21:27:41 -07002496static int hugetlb_acct_memory(struct hstate *h, long delta)
Mel Gormanfc1b8a72008-07-23 21:27:22 -07002497{
2498 int ret = -ENOMEM;
2499
2500 spin_lock(&hugetlb_lock);
2501 /*
2502 * When cpuset is configured, it breaks the strict hugetlb page
2503 * reservation as the accounting is done on a global variable. Such
2504 * reservation is completely rubbish in the presence of cpuset because
2505 * the reservation is not checked against page availability for the
2506 * current cpuset. Application can still potentially OOM'ed by kernel
2507 * with lack of free htlb page in cpuset that the task is in.
2508 * Attempt to enforce strict accounting with cpuset is almost
2509 * impossible (or too ugly) because cpuset is too fluid that
2510 * task or memory node can be dynamically moved between cpusets.
2511 *
2512 * The change of semantics for shared hugetlb mapping with cpuset is
2513 * undesirable. However, in order to preserve some of the semantics,
2514 * we fall back to check against current free page availability as
2515 * a best attempt and hopefully to minimize the impact of changing
2516 * semantics that cpuset has.
2517 */
2518 if (delta > 0) {
Andi Kleena5516432008-07-23 21:27:41 -07002519 if (gather_surplus_pages(h, delta) < 0)
Mel Gormanfc1b8a72008-07-23 21:27:22 -07002520 goto out;
2521
Andi Kleena5516432008-07-23 21:27:41 -07002522 if (delta > cpuset_mems_nr(h->free_huge_pages_node)) {
2523 return_unused_surplus_pages(h, delta);
Mel Gormanfc1b8a72008-07-23 21:27:22 -07002524 goto out;
2525 }
2526 }
2527
2528 ret = 0;
2529 if (delta < 0)
Andi Kleena5516432008-07-23 21:27:41 -07002530 return_unused_surplus_pages(h, (unsigned long) -delta);
Mel Gormanfc1b8a72008-07-23 21:27:22 -07002531
2532out:
2533 spin_unlock(&hugetlb_lock);
2534 return ret;
2535}
2536
Andy Whitcroft84afd992008-07-23 21:27:32 -07002537static void hugetlb_vm_op_open(struct vm_area_struct *vma)
2538{
Joonsoo Kimf522c3a2013-09-11 14:21:53 -07002539 struct resv_map *resv = vma_resv_map(vma);
Andy Whitcroft84afd992008-07-23 21:27:32 -07002540
2541 /*
2542 * This new VMA should share its siblings reservation map if present.
2543 * The VMA will only ever have a valid reservation map pointer where
2544 * it is being copied for another still existing VMA. As that VMA
Lucas De Marchi25985ed2011-03-30 22:57:33 -03002545 * has a reference to the reservation map it cannot disappear until
Andy Whitcroft84afd992008-07-23 21:27:32 -07002546 * after this open call completes. It is therefore safe to take a
2547 * new reference here without additional locking.
2548 */
Joonsoo Kim4e35f482014-04-03 14:47:30 -07002549 if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
Joonsoo Kimf522c3a2013-09-11 14:21:53 -07002550 kref_get(&resv->refs);
Andy Whitcroft84afd992008-07-23 21:27:32 -07002551}
2552
Mel Gormana1e78772008-07-23 21:27:23 -07002553static void hugetlb_vm_op_close(struct vm_area_struct *vma)
2554{
Andi Kleena5516432008-07-23 21:27:41 -07002555 struct hstate *h = hstate_vma(vma);
Joonsoo Kimf522c3a2013-09-11 14:21:53 -07002556 struct resv_map *resv = vma_resv_map(vma);
David Gibson90481622012-03-21 16:34:12 -07002557 struct hugepage_subpool *spool = subpool_vma(vma);
Joonsoo Kim4e35f482014-04-03 14:47:30 -07002558 unsigned long reserve, start, end;
Mike Kravetz1c5ecae2015-04-15 16:13:39 -07002559 long gbl_reserve;
Andy Whitcroft84afd992008-07-23 21:27:32 -07002560
Joonsoo Kim4e35f482014-04-03 14:47:30 -07002561 if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER))
2562 return;
Andy Whitcroft84afd992008-07-23 21:27:32 -07002563
Joonsoo Kim4e35f482014-04-03 14:47:30 -07002564 start = vma_hugecache_offset(h, vma, vma->vm_start);
2565 end = vma_hugecache_offset(h, vma, vma->vm_end);
Andy Whitcroft84afd992008-07-23 21:27:32 -07002566
Joonsoo Kim4e35f482014-04-03 14:47:30 -07002567 reserve = (end - start) - region_count(resv, start, end);
Andy Whitcroft84afd992008-07-23 21:27:32 -07002568
Joonsoo Kim4e35f482014-04-03 14:47:30 -07002569 kref_put(&resv->refs, resv_map_release);
2570
2571 if (reserve) {
Mike Kravetz1c5ecae2015-04-15 16:13:39 -07002572 /*
2573 * Decrement reserve counts. The global reserve count may be
2574 * adjusted if the subpool has a minimum size.
2575 */
2576 gbl_reserve = hugepage_subpool_put_pages(spool, reserve);
2577 hugetlb_acct_memory(h, -gbl_reserve);
Andy Whitcroft84afd992008-07-23 21:27:32 -07002578 }
Mel Gormana1e78772008-07-23 21:27:23 -07002579}
2580
Linus Torvalds1da177e2005-04-16 15:20:36 -07002581/*
2582 * We cannot handle pagefaults against hugetlb pages at all. They cause
2583 * handle_mm_fault() to try to instantiate regular-sized pages in the
2584 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
2585 * this far.
2586 */
Nick Piggind0217ac2007-07-19 01:47:03 -07002587static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002588{
2589 BUG();
Nick Piggind0217ac2007-07-19 01:47:03 -07002590 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002591}
2592
Alexey Dobriyanf0f37e22009-09-27 22:29:37 +04002593const struct vm_operations_struct hugetlb_vm_ops = {
Nick Piggind0217ac2007-07-19 01:47:03 -07002594 .fault = hugetlb_vm_op_fault,
Andy Whitcroft84afd992008-07-23 21:27:32 -07002595 .open = hugetlb_vm_op_open,
Mel Gormana1e78772008-07-23 21:27:23 -07002596 .close = hugetlb_vm_op_close,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002597};
2598
David Gibson1e8f8892006-01-06 00:10:44 -08002599static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
2600 int writable)
David Gibson63551ae2005-06-21 17:14:44 -07002601{
2602 pte_t entry;
2603
David Gibson1e8f8892006-01-06 00:10:44 -08002604 if (writable) {
Gerald Schaefer106c9922013-04-29 15:07:23 -07002605 entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page,
2606 vma->vm_page_prot)));
David Gibson63551ae2005-06-21 17:14:44 -07002607 } else {
Gerald Schaefer106c9922013-04-29 15:07:23 -07002608 entry = huge_pte_wrprotect(mk_huge_pte(page,
2609 vma->vm_page_prot));
David Gibson63551ae2005-06-21 17:14:44 -07002610 }
2611 entry = pte_mkyoung(entry);
2612 entry = pte_mkhuge(entry);
Chris Metcalfd9ed9fa2012-04-01 14:01:34 -04002613 entry = arch_make_huge_pte(entry, vma, page, writable);
David Gibson63551ae2005-06-21 17:14:44 -07002614
2615 return entry;
2616}
2617
David Gibson1e8f8892006-01-06 00:10:44 -08002618static void set_huge_ptep_writable(struct vm_area_struct *vma,
2619 unsigned long address, pte_t *ptep)
2620{
2621 pte_t entry;
2622
Gerald Schaefer106c9922013-04-29 15:07:23 -07002623 entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep)));
Chris Forbes32f84522011-07-25 17:12:14 -07002624 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
Russell King4b3073e2009-12-18 16:40:18 +00002625 update_mmu_cache(vma, address, ptep);
David Gibson1e8f8892006-01-06 00:10:44 -08002626}
2627
Naoya Horiguchi4a705fe2014-06-23 13:22:03 -07002628static int is_hugetlb_entry_migration(pte_t pte)
2629{
2630 swp_entry_t swp;
2631
2632 if (huge_pte_none(pte) || pte_present(pte))
2633 return 0;
2634 swp = pte_to_swp_entry(pte);
2635 if (non_swap_entry(swp) && is_migration_entry(swp))
2636 return 1;
2637 else
2638 return 0;
2639}
2640
2641static int is_hugetlb_entry_hwpoisoned(pte_t pte)
2642{
2643 swp_entry_t swp;
2644
2645 if (huge_pte_none(pte) || pte_present(pte))
2646 return 0;
2647 swp = pte_to_swp_entry(pte);
2648 if (non_swap_entry(swp) && is_hwpoison_entry(swp))
2649 return 1;
2650 else
2651 return 0;
2652}
David Gibson1e8f8892006-01-06 00:10:44 -08002653
David Gibson63551ae2005-06-21 17:14:44 -07002654int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
2655 struct vm_area_struct *vma)
2656{
2657 pte_t *src_pte, *dst_pte, entry;
2658 struct page *ptepage;
Hugh Dickins1c598272005-10-19 21:23:43 -07002659 unsigned long addr;
David Gibson1e8f8892006-01-06 00:10:44 -08002660 int cow;
Andi Kleena5516432008-07-23 21:27:41 -07002661 struct hstate *h = hstate_vma(vma);
2662 unsigned long sz = huge_page_size(h);
Andreas Sandberge8569dd2014-01-21 15:49:09 -08002663 unsigned long mmun_start; /* For mmu_notifiers */
2664 unsigned long mmun_end; /* For mmu_notifiers */
2665 int ret = 0;
David Gibson1e8f8892006-01-06 00:10:44 -08002666
2667 cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
David Gibson63551ae2005-06-21 17:14:44 -07002668
Andreas Sandberge8569dd2014-01-21 15:49:09 -08002669 mmun_start = vma->vm_start;
2670 mmun_end = vma->vm_end;
2671 if (cow)
2672 mmu_notifier_invalidate_range_start(src, mmun_start, mmun_end);
2673
Andi Kleena5516432008-07-23 21:27:41 -07002674 for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002675 spinlock_t *src_ptl, *dst_ptl;
Hugh Dickinsc74df322005-10-29 18:16:23 -07002676 src_pte = huge_pte_offset(src, addr);
2677 if (!src_pte)
2678 continue;
Andi Kleena5516432008-07-23 21:27:41 -07002679 dst_pte = huge_pte_alloc(dst, addr, sz);
Andreas Sandberge8569dd2014-01-21 15:49:09 -08002680 if (!dst_pte) {
2681 ret = -ENOMEM;
2682 break;
2683 }
Larry Woodmanc5c99422008-01-24 05:49:25 -08002684
2685 /* If the pagetables are shared don't copy or take references */
2686 if (dst_pte == src_pte)
2687 continue;
2688
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002689 dst_ptl = huge_pte_lock(h, dst, dst_pte);
2690 src_ptl = huge_pte_lockptr(h, src, src_pte);
2691 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
Naoya Horiguchi4a705fe2014-06-23 13:22:03 -07002692 entry = huge_ptep_get(src_pte);
2693 if (huge_pte_none(entry)) { /* skip none entry */
2694 ;
2695 } else if (unlikely(is_hugetlb_entry_migration(entry) ||
2696 is_hugetlb_entry_hwpoisoned(entry))) {
2697 swp_entry_t swp_entry = pte_to_swp_entry(entry);
2698
2699 if (is_write_migration_entry(swp_entry) && cow) {
2700 /*
2701 * COW mappings require pages in both
2702 * parent and child to be set to read.
2703 */
2704 make_migration_entry_read(&swp_entry);
2705 entry = swp_entry_to_pte(swp_entry);
2706 set_huge_pte_at(src, addr, src_pte, entry);
2707 }
2708 set_huge_pte_at(dst, addr, dst_pte, entry);
2709 } else {
Joerg Roedel34ee6452014-11-13 13:46:09 +11002710 if (cow) {
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07002711 huge_ptep_set_wrprotect(src, addr, src_pte);
Joerg Roedel34ee6452014-11-13 13:46:09 +11002712 mmu_notifier_invalidate_range(src, mmun_start,
2713 mmun_end);
2714 }
Naoya Horiguchi0253d632014-07-23 14:00:19 -07002715 entry = huge_ptep_get(src_pte);
Hugh Dickins1c598272005-10-19 21:23:43 -07002716 ptepage = pte_page(entry);
2717 get_page(ptepage);
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002718 page_dup_rmap(ptepage);
Hugh Dickins1c598272005-10-19 21:23:43 -07002719 set_huge_pte_at(dst, addr, dst_pte, entry);
2720 }
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002721 spin_unlock(src_ptl);
2722 spin_unlock(dst_ptl);
David Gibson63551ae2005-06-21 17:14:44 -07002723 }
David Gibson63551ae2005-06-21 17:14:44 -07002724
Andreas Sandberge8569dd2014-01-21 15:49:09 -08002725 if (cow)
2726 mmu_notifier_invalidate_range_end(src, mmun_start, mmun_end);
2727
2728 return ret;
David Gibson63551ae2005-06-21 17:14:44 -07002729}
2730
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07002731void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
2732 unsigned long start, unsigned long end,
2733 struct page *ref_page)
David Gibson63551ae2005-06-21 17:14:44 -07002734{
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07002735 int force_flush = 0;
David Gibson63551ae2005-06-21 17:14:44 -07002736 struct mm_struct *mm = vma->vm_mm;
2737 unsigned long address;
David Gibsonc7546f82005-08-05 11:59:35 -07002738 pte_t *ptep;
David Gibson63551ae2005-06-21 17:14:44 -07002739 pte_t pte;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002740 spinlock_t *ptl;
David Gibson63551ae2005-06-21 17:14:44 -07002741 struct page *page;
Andi Kleena5516432008-07-23 21:27:41 -07002742 struct hstate *h = hstate_vma(vma);
2743 unsigned long sz = huge_page_size(h);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07002744 const unsigned long mmun_start = start; /* For mmu_notifiers */
2745 const unsigned long mmun_end = end; /* For mmu_notifiers */
Andi Kleena5516432008-07-23 21:27:41 -07002746
David Gibson63551ae2005-06-21 17:14:44 -07002747 WARN_ON(!is_vm_hugetlb_page(vma));
Andi Kleena5516432008-07-23 21:27:41 -07002748 BUG_ON(start & ~huge_page_mask(h));
2749 BUG_ON(end & ~huge_page_mask(h));
David Gibson63551ae2005-06-21 17:14:44 -07002750
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07002751 tlb_start_vma(tlb, vma);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07002752 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
Hillf Danton569f48b82014-12-10 15:44:41 -08002753 address = start;
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07002754again:
Hillf Danton569f48b82014-12-10 15:44:41 -08002755 for (; address < end; address += sz) {
David Gibsonc7546f82005-08-05 11:59:35 -07002756 ptep = huge_pte_offset(mm, address);
Adam Litke4c887262005-10-29 18:16:46 -07002757 if (!ptep)
David Gibsonc7546f82005-08-05 11:59:35 -07002758 continue;
2759
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002760 ptl = huge_pte_lock(h, mm, ptep);
Chen, Kenneth W39dde652006-12-06 20:32:03 -08002761 if (huge_pmd_unshare(mm, &address, ptep))
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002762 goto unlock;
Chen, Kenneth W39dde652006-12-06 20:32:03 -08002763
Hillf Danton66293262012-03-23 15:01:48 -07002764 pte = huge_ptep_get(ptep);
2765 if (huge_pte_none(pte))
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002766 goto unlock;
Hillf Danton66293262012-03-23 15:01:48 -07002767
2768 /*
Naoya Horiguchi9fbc1f62015-02-11 15:25:32 -08002769 * Migrating hugepage or HWPoisoned hugepage is already
2770 * unmapped and its refcount is dropped, so just clear pte here.
Hillf Danton66293262012-03-23 15:01:48 -07002771 */
Naoya Horiguchi9fbc1f62015-02-11 15:25:32 -08002772 if (unlikely(!pte_present(pte))) {
Gerald Schaefer106c9922013-04-29 15:07:23 -07002773 huge_pte_clear(mm, address, ptep);
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002774 goto unlock;
Naoya Horiguchi8c4894c2012-12-12 13:52:28 -08002775 }
Hillf Danton66293262012-03-23 15:01:48 -07002776
2777 page = pte_page(pte);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002778 /*
2779 * If a reference page is supplied, it is because a specific
2780 * page is being unmapped, not a range. Ensure the page we
2781 * are about to unmap is the actual page of interest.
2782 */
2783 if (ref_page) {
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002784 if (page != ref_page)
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002785 goto unlock;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002786
2787 /*
2788 * Mark the VMA as having unmapped its page so that
2789 * future faults in this VMA will fail rather than
2790 * looking like data was lost
2791 */
2792 set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
2793 }
2794
David Gibsonc7546f82005-08-05 11:59:35 -07002795 pte = huge_ptep_get_and_clear(mm, address, ptep);
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07002796 tlb_remove_tlb_entry(tlb, ptep, address);
Gerald Schaefer106c9922013-04-29 15:07:23 -07002797 if (huge_pte_dirty(pte))
Ken Chen6649a382007-02-08 14:20:27 -08002798 set_page_dirty(page);
Hillf Danton9e811302012-03-21 16:34:03 -07002799
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07002800 page_remove_rmap(page);
2801 force_flush = !__tlb_remove_page(tlb, page);
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002802 if (force_flush) {
Hillf Danton569f48b82014-12-10 15:44:41 -08002803 address += sz;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002804 spin_unlock(ptl);
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07002805 break;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002806 }
Hillf Danton9e811302012-03-21 16:34:03 -07002807 /* Bail out after unmapping reference page if supplied */
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002808 if (ref_page) {
2809 spin_unlock(ptl);
Hillf Danton9e811302012-03-21 16:34:03 -07002810 break;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002811 }
2812unlock:
2813 spin_unlock(ptl);
David Gibson63551ae2005-06-21 17:14:44 -07002814 }
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07002815 /*
2816 * mmu_gather ran out of room to batch pages, we break out of
2817 * the PTE lock to avoid doing the potential expensive TLB invalidate
2818 * and page-free while holding it.
2819 */
2820 if (force_flush) {
2821 force_flush = 0;
2822 tlb_flush_mmu(tlb);
2823 if (address < end && !ref_page)
2824 goto again;
Chen, Kenneth Wfe1668a2006-10-04 02:15:24 -07002825 }
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07002826 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07002827 tlb_end_vma(tlb, vma);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002828}
David Gibson63551ae2005-06-21 17:14:44 -07002829
Mel Gormand8333522012-07-31 16:46:20 -07002830void __unmap_hugepage_range_final(struct mmu_gather *tlb,
2831 struct vm_area_struct *vma, unsigned long start,
2832 unsigned long end, struct page *ref_page)
2833{
2834 __unmap_hugepage_range(tlb, vma, start, end, ref_page);
2835
2836 /*
2837 * Clear this flag so that x86's huge_pmd_share page_table_shareable
2838 * test will fail on a vma being torn down, and not grab a page table
2839 * on its way out. We're lucky that the flag has such an appropriate
2840 * name, and can in fact be safely cleared here. We could clear it
2841 * before the __unmap_hugepage_range above, but all that's necessary
Davidlohr Buesoc8c06ef2014-12-12 16:54:24 -08002842 * is to clear it before releasing the i_mmap_rwsem. This works
Mel Gormand8333522012-07-31 16:46:20 -07002843 * because in the context this is called, the VMA is about to be
Davidlohr Buesoc8c06ef2014-12-12 16:54:24 -08002844 * destroyed and the i_mmap_rwsem is held.
Mel Gormand8333522012-07-31 16:46:20 -07002845 */
2846 vma->vm_flags &= ~VM_MAYSHARE;
2847}
2848
Chen, Kenneth W502717f2006-10-11 01:20:46 -07002849void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002850 unsigned long end, struct page *ref_page)
Chen, Kenneth W502717f2006-10-11 01:20:46 -07002851{
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07002852 struct mm_struct *mm;
2853 struct mmu_gather tlb;
2854
2855 mm = vma->vm_mm;
2856
Linus Torvalds2b047252013-08-15 11:42:25 -07002857 tlb_gather_mmu(&tlb, mm, start, end);
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07002858 __unmap_hugepage_range(&tlb, vma, start, end, ref_page);
2859 tlb_finish_mmu(&tlb, start, end);
Chen, Kenneth W502717f2006-10-11 01:20:46 -07002860}
2861
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002862/*
2863 * This is called when the original mapper is failing to COW a MAP_PRIVATE
2864 * mappping it owns the reserve page for. The intention is to unmap the page
2865 * from other VMAs and let the children be SIGKILLed if they are faulting the
2866 * same region.
2867 */
Davidlohr Bueso2f4612a2014-08-06 16:06:45 -07002868static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
2869 struct page *page, unsigned long address)
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002870{
Adam Litke75266742008-11-12 13:24:56 -08002871 struct hstate *h = hstate_vma(vma);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002872 struct vm_area_struct *iter_vma;
2873 struct address_space *mapping;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002874 pgoff_t pgoff;
2875
2876 /*
2877 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
2878 * from page cache lookup which is in HPAGE_SIZE units.
2879 */
Adam Litke75266742008-11-12 13:24:56 -08002880 address = address & huge_page_mask(h);
Michal Hocko36e4f202012-10-08 16:33:31 -07002881 pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) +
2882 vma->vm_pgoff;
Al Viro496ad9a2013-01-23 17:07:38 -05002883 mapping = file_inode(vma->vm_file)->i_mapping;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002884
Mel Gorman4eb2b1d2009-12-14 17:59:53 -08002885 /*
2886 * Take the mapping lock for the duration of the table walk. As
2887 * this mapping should be shared between all the VMAs,
2888 * __unmap_hugepage_range() is called as the lock is already held
2889 */
Davidlohr Bueso83cde9e2014-12-12 16:54:21 -08002890 i_mmap_lock_write(mapping);
Michel Lespinasse6b2dbba2012-10-08 16:31:25 -07002891 vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) {
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002892 /* Do not unmap the current VMA */
2893 if (iter_vma == vma)
2894 continue;
2895
2896 /*
2897 * Unmap the page from other VMAs without their own reserves.
2898 * They get marked to be SIGKILLed if they fault in these
2899 * areas. This is because a future no-page fault on this VMA
2900 * could insert a zeroed page instead of the data existing
2901 * from the time of fork. This would look like data corruption
2902 */
2903 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07002904 unmap_hugepage_range(iter_vma, address,
2905 address + huge_page_size(h), page);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002906 }
Davidlohr Bueso83cde9e2014-12-12 16:54:21 -08002907 i_mmap_unlock_write(mapping);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002908}
2909
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002910/*
2911 * Hugetlb_cow() should be called with page lock of the original hugepage held.
Michal Hockoef009b22012-01-10 15:07:21 -08002912 * Called with hugetlb_instantiation_mutex held and pte_page locked so we
2913 * cannot race with other handlers or page migration.
2914 * Keep the pte_same checks anyway to make transition from the mutex easier.
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002915 */
David Gibson1e8f8892006-01-06 00:10:44 -08002916static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002917 unsigned long address, pte_t *ptep, pte_t pte,
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002918 struct page *pagecache_page, spinlock_t *ptl)
David Gibson1e8f8892006-01-06 00:10:44 -08002919{
Andi Kleena5516432008-07-23 21:27:41 -07002920 struct hstate *h = hstate_vma(vma);
David Gibson1e8f8892006-01-06 00:10:44 -08002921 struct page *old_page, *new_page;
Davidlohr Buesoad4404a2014-08-06 16:06:47 -07002922 int ret = 0, outside_reserve = 0;
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07002923 unsigned long mmun_start; /* For mmu_notifiers */
2924 unsigned long mmun_end; /* For mmu_notifiers */
David Gibson1e8f8892006-01-06 00:10:44 -08002925
2926 old_page = pte_page(pte);
2927
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002928retry_avoidcopy:
David Gibson1e8f8892006-01-06 00:10:44 -08002929 /* If no-one else is actually using this page, avoid the copy
2930 * and just make the page writable */
Joonsoo Kim37a21402013-09-11 14:21:04 -07002931 if (page_mapcount(old_page) == 1 && PageAnon(old_page)) {
2932 page_move_anon_rmap(old_page, vma, address);
David Gibson1e8f8892006-01-06 00:10:44 -08002933 set_huge_ptep_writable(vma, address, ptep);
Nick Piggin83c54072007-07-19 01:47:05 -07002934 return 0;
David Gibson1e8f8892006-01-06 00:10:44 -08002935 }
2936
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002937 /*
2938 * If the process that created a MAP_PRIVATE mapping is about to
2939 * perform a COW due to a shared page count, attempt to satisfy
2940 * the allocation without using the existing reserves. The pagecache
2941 * page is used to determine if the reserve at this address was
2942 * consumed or not. If reserves were used, a partial faulted mapping
2943 * at the time of fork() could consume its reserves on COW instead
2944 * of the full address range.
2945 */
Joonsoo Kim5944d012013-09-11 14:21:55 -07002946 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002947 old_page != pagecache_page)
2948 outside_reserve = 1;
2949
David Gibson1e8f8892006-01-06 00:10:44 -08002950 page_cache_get(old_page);
Larry Woodmanb76c8cf2009-12-14 17:59:37 -08002951
Davidlohr Buesoad4404a2014-08-06 16:06:47 -07002952 /*
2953 * Drop page table lock as buddy allocator may be called. It will
2954 * be acquired again before returning to the caller, as expected.
2955 */
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002956 spin_unlock(ptl);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002957 new_page = alloc_huge_page(vma, address, outside_reserve);
David Gibson1e8f8892006-01-06 00:10:44 -08002958
Adam Litke2fc39ce2007-11-14 16:59:39 -08002959 if (IS_ERR(new_page)) {
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002960 /*
2961 * If a process owning a MAP_PRIVATE mapping fails to COW,
2962 * it is due to references held by a child and an insufficient
2963 * huge page pool. To guarantee the original mappers
2964 * reliability, unmap the page from child processes. The child
2965 * may get SIGKILLed if it later faults.
2966 */
2967 if (outside_reserve) {
Davidlohr Buesoad4404a2014-08-06 16:06:47 -07002968 page_cache_release(old_page);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002969 BUG_ON(huge_pte_none(pte));
Davidlohr Bueso2f4612a2014-08-06 16:06:45 -07002970 unmap_ref_private(mm, vma, old_page, address);
2971 BUG_ON(huge_pte_none(pte));
2972 spin_lock(ptl);
2973 ptep = huge_pte_offset(mm, address & huge_page_mask(h));
2974 if (likely(ptep &&
2975 pte_same(huge_ptep_get(ptep), pte)))
2976 goto retry_avoidcopy;
2977 /*
2978 * race occurs while re-acquiring page table
2979 * lock, and our job is done.
2980 */
2981 return 0;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002982 }
2983
Davidlohr Buesoad4404a2014-08-06 16:06:47 -07002984 ret = (PTR_ERR(new_page) == -ENOMEM) ?
2985 VM_FAULT_OOM : VM_FAULT_SIGBUS;
2986 goto out_release_old;
David Gibson1e8f8892006-01-06 00:10:44 -08002987 }
2988
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002989 /*
2990 * When the original hugepage is shared one, it does not have
2991 * anon_vma prepared.
2992 */
Dean Nelson44e2aa92010-10-26 14:22:08 -07002993 if (unlikely(anon_vma_prepare(vma))) {
Davidlohr Buesoad4404a2014-08-06 16:06:47 -07002994 ret = VM_FAULT_OOM;
2995 goto out_release_all;
Dean Nelson44e2aa92010-10-26 14:22:08 -07002996 }
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002997
Andrea Arcangeli47ad8472011-01-13 15:46:47 -08002998 copy_user_huge_page(new_page, old_page, address, vma,
2999 pages_per_huge_page(h));
Nick Piggin0ed361d2008-02-04 22:29:34 -08003000 __SetPageUptodate(new_page);
Naoya Horiguchibcc54222015-04-15 16:14:38 -07003001 set_page_huge_active(new_page);
David Gibson1e8f8892006-01-06 00:10:44 -08003002
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07003003 mmun_start = address & huge_page_mask(h);
3004 mmun_end = mmun_start + huge_page_size(h);
3005 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
Davidlohr Buesoad4404a2014-08-06 16:06:47 -07003006
Larry Woodmanb76c8cf2009-12-14 17:59:37 -08003007 /*
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003008 * Retake the page table lock to check for racing updates
Larry Woodmanb76c8cf2009-12-14 17:59:37 -08003009 * before the page tables are altered
3010 */
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003011 spin_lock(ptl);
Andi Kleena5516432008-07-23 21:27:41 -07003012 ptep = huge_pte_offset(mm, address & huge_page_mask(h));
Naoya Horiguchia9af0c52014-04-07 15:36:54 -07003013 if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) {
Joonsoo Kim07443a82013-09-11 14:21:58 -07003014 ClearPagePrivate(new_page);
3015
David Gibson1e8f8892006-01-06 00:10:44 -08003016 /* Break COW */
Gerald Schaefer8fe627e2008-04-28 02:13:28 -07003017 huge_ptep_clear_flush(vma, address, ptep);
Joerg Roedel34ee6452014-11-13 13:46:09 +11003018 mmu_notifier_invalidate_range(mm, mmun_start, mmun_end);
David Gibson1e8f8892006-01-06 00:10:44 -08003019 set_huge_pte_at(mm, address, ptep,
3020 make_huge_pte(vma, new_page, 1));
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09003021 page_remove_rmap(old_page);
Naoya Horiguchicd67f0d2010-09-10 13:23:04 +09003022 hugepage_add_new_anon_rmap(new_page, vma, address);
David Gibson1e8f8892006-01-06 00:10:44 -08003023 /* Make the old page be freed below */
3024 new_page = old_page;
3025 }
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003026 spin_unlock(ptl);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07003027 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
Davidlohr Buesoad4404a2014-08-06 16:06:47 -07003028out_release_all:
David Gibson1e8f8892006-01-06 00:10:44 -08003029 page_cache_release(new_page);
Davidlohr Buesoad4404a2014-08-06 16:06:47 -07003030out_release_old:
David Gibson1e8f8892006-01-06 00:10:44 -08003031 page_cache_release(old_page);
Joonsoo Kim83120342013-09-11 14:21:57 -07003032
Davidlohr Buesoad4404a2014-08-06 16:06:47 -07003033 spin_lock(ptl); /* Caller expects lock to be held */
3034 return ret;
David Gibson1e8f8892006-01-06 00:10:44 -08003035}
3036
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003037/* Return the pagecache page at a given address within a VMA */
Andi Kleena5516432008-07-23 21:27:41 -07003038static struct page *hugetlbfs_pagecache_page(struct hstate *h,
3039 struct vm_area_struct *vma, unsigned long address)
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003040{
3041 struct address_space *mapping;
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -07003042 pgoff_t idx;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003043
3044 mapping = vma->vm_file->f_mapping;
Andi Kleena5516432008-07-23 21:27:41 -07003045 idx = vma_hugecache_offset(h, vma, address);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003046
3047 return find_lock_page(mapping, idx);
3048}
3049
Hugh Dickins3ae77f42009-09-21 17:03:33 -07003050/*
3051 * Return whether there is a pagecache page to back given address within VMA.
3052 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
3053 */
3054static bool hugetlbfs_pagecache_present(struct hstate *h,
Hugh Dickins2a15efc2009-09-21 17:03:27 -07003055 struct vm_area_struct *vma, unsigned long address)
3056{
3057 struct address_space *mapping;
3058 pgoff_t idx;
3059 struct page *page;
3060
3061 mapping = vma->vm_file->f_mapping;
3062 idx = vma_hugecache_offset(h, vma, address);
3063
3064 page = find_get_page(mapping, idx);
3065 if (page)
3066 put_page(page);
3067 return page != NULL;
3068}
3069
Robert P. J. Daya1ed3dd2007-07-17 04:03:33 -07003070static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
Davidlohr Bueso8382d912014-04-03 14:47:31 -07003071 struct address_space *mapping, pgoff_t idx,
3072 unsigned long address, pte_t *ptep, unsigned int flags)
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01003073{
Andi Kleena5516432008-07-23 21:27:41 -07003074 struct hstate *h = hstate_vma(vma);
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01003075 int ret = VM_FAULT_SIGBUS;
Hillf Danton409eb8c2012-01-20 14:34:13 -08003076 int anon_rmap = 0;
Adam Litke4c887262005-10-29 18:16:46 -07003077 unsigned long size;
Adam Litke4c887262005-10-29 18:16:46 -07003078 struct page *page;
David Gibson1e8f8892006-01-06 00:10:44 -08003079 pte_t new_pte;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003080 spinlock_t *ptl;
Adam Litke4c887262005-10-29 18:16:46 -07003081
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003082 /*
3083 * Currently, we are forced to kill the process in the event the
3084 * original mapper has unmapped pages from the child due to a failed
Lucas De Marchi25985ed2011-03-30 22:57:33 -03003085 * COW. Warn that such a situation has occurred as it may not be obvious
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003086 */
3087 if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
Andrew Mortonffb22af2013-02-22 16:32:08 -08003088 pr_warning("PID %d killed due to inadequate hugepage pool\n",
3089 current->pid);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003090 return ret;
3091 }
3092
Adam Litke4c887262005-10-29 18:16:46 -07003093 /*
3094 * Use page lock to guard against racing truncation
3095 * before we get page_table_lock.
3096 */
Christoph Lameter6bda6662006-01-06 00:10:49 -08003097retry:
3098 page = find_lock_page(mapping, idx);
3099 if (!page) {
Andi Kleena5516432008-07-23 21:27:41 -07003100 size = i_size_read(mapping->host) >> huge_page_shift(h);
Hugh Dickinsebed4bf2006-10-28 10:38:43 -07003101 if (idx >= size)
3102 goto out;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003103 page = alloc_huge_page(vma, address, 0);
Adam Litke2fc39ce2007-11-14 16:59:39 -08003104 if (IS_ERR(page)) {
Aneesh Kumar K.V76dcee72012-07-31 16:41:57 -07003105 ret = PTR_ERR(page);
3106 if (ret == -ENOMEM)
3107 ret = VM_FAULT_OOM;
3108 else
3109 ret = VM_FAULT_SIGBUS;
Christoph Lameter6bda6662006-01-06 00:10:49 -08003110 goto out;
3111 }
Andrea Arcangeli47ad8472011-01-13 15:46:47 -08003112 clear_huge_page(page, address, pages_per_huge_page(h));
Nick Piggin0ed361d2008-02-04 22:29:34 -08003113 __SetPageUptodate(page);
Naoya Horiguchibcc54222015-04-15 16:14:38 -07003114 set_page_huge_active(page);
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01003115
Mel Gormanf83a2752009-05-28 14:34:40 -07003116 if (vma->vm_flags & VM_MAYSHARE) {
Christoph Lameter6bda6662006-01-06 00:10:49 -08003117 int err;
Ken Chen45c682a2007-11-14 16:59:44 -08003118 struct inode *inode = mapping->host;
Christoph Lameter6bda6662006-01-06 00:10:49 -08003119
3120 err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
3121 if (err) {
3122 put_page(page);
Christoph Lameter6bda6662006-01-06 00:10:49 -08003123 if (err == -EEXIST)
3124 goto retry;
3125 goto out;
3126 }
Joonsoo Kim07443a82013-09-11 14:21:58 -07003127 ClearPagePrivate(page);
Ken Chen45c682a2007-11-14 16:59:44 -08003128
3129 spin_lock(&inode->i_lock);
Andi Kleena5516432008-07-23 21:27:41 -07003130 inode->i_blocks += blocks_per_huge_page(h);
Ken Chen45c682a2007-11-14 16:59:44 -08003131 spin_unlock(&inode->i_lock);
Mel Gorman23be7462010-04-23 13:17:56 -04003132 } else {
Christoph Lameter6bda6662006-01-06 00:10:49 -08003133 lock_page(page);
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09003134 if (unlikely(anon_vma_prepare(vma))) {
3135 ret = VM_FAULT_OOM;
3136 goto backout_unlocked;
3137 }
Hillf Danton409eb8c2012-01-20 14:34:13 -08003138 anon_rmap = 1;
Mel Gorman23be7462010-04-23 13:17:56 -04003139 }
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09003140 } else {
Naoya Horiguchi998b4382010-09-08 10:19:32 +09003141 /*
3142 * If memory error occurs between mmap() and fault, some process
3143 * don't have hwpoisoned swap entry for errored virtual address.
3144 * So we need to block hugepage fault by PG_hwpoison bit check.
3145 */
3146 if (unlikely(PageHWPoison(page))) {
Chris Forbes32f84522011-07-25 17:12:14 -07003147 ret = VM_FAULT_HWPOISON |
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07003148 VM_FAULT_SET_HINDEX(hstate_index(h));
Naoya Horiguchi998b4382010-09-08 10:19:32 +09003149 goto backout_unlocked;
3150 }
Christoph Lameter6bda6662006-01-06 00:10:49 -08003151 }
David Gibson1e8f8892006-01-06 00:10:44 -08003152
Andy Whitcroft57303d82008-08-12 15:08:47 -07003153 /*
3154 * If we are going to COW a private mapping later, we examine the
3155 * pending reservations for this page now. This will ensure that
3156 * any allocations necessary to record that reservation occur outside
3157 * the spinlock.
3158 */
Hugh Dickins788c7df2009-06-23 13:49:05 +01003159 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED))
Andy Whitcroft2b267362008-08-12 15:08:49 -07003160 if (vma_needs_reservation(h, vma, address) < 0) {
3161 ret = VM_FAULT_OOM;
3162 goto backout_unlocked;
3163 }
Andy Whitcroft57303d82008-08-12 15:08:47 -07003164
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003165 ptl = huge_pte_lockptr(h, mm, ptep);
3166 spin_lock(ptl);
Andi Kleena5516432008-07-23 21:27:41 -07003167 size = i_size_read(mapping->host) >> huge_page_shift(h);
Adam Litke4c887262005-10-29 18:16:46 -07003168 if (idx >= size)
3169 goto backout;
3170
Nick Piggin83c54072007-07-19 01:47:05 -07003171 ret = 0;
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07003172 if (!huge_pte_none(huge_ptep_get(ptep)))
Adam Litke4c887262005-10-29 18:16:46 -07003173 goto backout;
3174
Joonsoo Kim07443a82013-09-11 14:21:58 -07003175 if (anon_rmap) {
3176 ClearPagePrivate(page);
Hillf Danton409eb8c2012-01-20 14:34:13 -08003177 hugepage_add_new_anon_rmap(page, vma, address);
Choi Gi-yongac714902014-04-07 15:37:36 -07003178 } else
Hillf Danton409eb8c2012-01-20 14:34:13 -08003179 page_dup_rmap(page);
David Gibson1e8f8892006-01-06 00:10:44 -08003180 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
3181 && (vma->vm_flags & VM_SHARED)));
3182 set_huge_pte_at(mm, address, ptep, new_pte);
3183
Hugh Dickins788c7df2009-06-23 13:49:05 +01003184 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
David Gibson1e8f8892006-01-06 00:10:44 -08003185 /* Optimization, do the COW without a second fault */
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003186 ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page, ptl);
David Gibson1e8f8892006-01-06 00:10:44 -08003187 }
3188
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003189 spin_unlock(ptl);
Adam Litke4c887262005-10-29 18:16:46 -07003190 unlock_page(page);
3191out:
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01003192 return ret;
Adam Litke4c887262005-10-29 18:16:46 -07003193
3194backout:
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003195 spin_unlock(ptl);
Andy Whitcroft2b267362008-08-12 15:08:49 -07003196backout_unlocked:
Adam Litke4c887262005-10-29 18:16:46 -07003197 unlock_page(page);
3198 put_page(page);
3199 goto out;
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01003200}
3201
Davidlohr Bueso8382d912014-04-03 14:47:31 -07003202#ifdef CONFIG_SMP
3203static u32 fault_mutex_hash(struct hstate *h, struct mm_struct *mm,
3204 struct vm_area_struct *vma,
3205 struct address_space *mapping,
3206 pgoff_t idx, unsigned long address)
3207{
3208 unsigned long key[2];
3209 u32 hash;
3210
3211 if (vma->vm_flags & VM_SHARED) {
3212 key[0] = (unsigned long) mapping;
3213 key[1] = idx;
3214 } else {
3215 key[0] = (unsigned long) mm;
3216 key[1] = address >> huge_page_shift(h);
3217 }
3218
3219 hash = jhash2((u32 *)&key, sizeof(key)/sizeof(u32), 0);
3220
3221 return hash & (num_fault_mutexes - 1);
3222}
3223#else
3224/*
3225 * For uniprocesor systems we always use a single mutex, so just
3226 * return 0 and avoid the hashing overhead.
3227 */
3228static u32 fault_mutex_hash(struct hstate *h, struct mm_struct *mm,
3229 struct vm_area_struct *vma,
3230 struct address_space *mapping,
3231 pgoff_t idx, unsigned long address)
3232{
3233 return 0;
3234}
3235#endif
3236
Adam Litke86e52162006-01-06 00:10:43 -08003237int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
Hugh Dickins788c7df2009-06-23 13:49:05 +01003238 unsigned long address, unsigned int flags)
Adam Litke86e52162006-01-06 00:10:43 -08003239{
Davidlohr Bueso8382d912014-04-03 14:47:31 -07003240 pte_t *ptep, entry;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003241 spinlock_t *ptl;
David Gibson1e8f8892006-01-06 00:10:44 -08003242 int ret;
Davidlohr Bueso8382d912014-04-03 14:47:31 -07003243 u32 hash;
3244 pgoff_t idx;
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09003245 struct page *page = NULL;
Andy Whitcroft57303d82008-08-12 15:08:47 -07003246 struct page *pagecache_page = NULL;
Andi Kleena5516432008-07-23 21:27:41 -07003247 struct hstate *h = hstate_vma(vma);
Davidlohr Bueso8382d912014-04-03 14:47:31 -07003248 struct address_space *mapping;
Naoya Horiguchi0f792cf2015-02-11 15:25:25 -08003249 int need_wait_lock = 0;
Adam Litke86e52162006-01-06 00:10:43 -08003250
KAMEZAWA Hiroyuki1e16a532012-01-10 15:07:22 -08003251 address &= huge_page_mask(h);
3252
Naoya Horiguchifd6a03e2010-05-28 09:29:21 +09003253 ptep = huge_pte_offset(mm, address);
3254 if (ptep) {
3255 entry = huge_ptep_get(ptep);
Naoya Horiguchi290408d2010-09-08 10:19:35 +09003256 if (unlikely(is_hugetlb_entry_migration(entry))) {
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003257 migration_entry_wait_huge(vma, mm, ptep);
Naoya Horiguchi290408d2010-09-08 10:19:35 +09003258 return 0;
3259 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
Chris Forbes32f84522011-07-25 17:12:14 -07003260 return VM_FAULT_HWPOISON_LARGE |
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07003261 VM_FAULT_SET_HINDEX(hstate_index(h));
Naoya Horiguchifd6a03e2010-05-28 09:29:21 +09003262 }
3263
Andi Kleena5516432008-07-23 21:27:41 -07003264 ptep = huge_pte_alloc(mm, address, huge_page_size(h));
Adam Litke86e52162006-01-06 00:10:43 -08003265 if (!ptep)
3266 return VM_FAULT_OOM;
3267
Davidlohr Bueso8382d912014-04-03 14:47:31 -07003268 mapping = vma->vm_file->f_mapping;
3269 idx = vma_hugecache_offset(h, vma, address);
3270
David Gibson3935baa2006-03-22 00:08:53 -08003271 /*
3272 * Serialize hugepage allocation and instantiation, so that we don't
3273 * get spurious allocation failures if two CPUs race to instantiate
3274 * the same page in the page cache.
3275 */
Davidlohr Bueso8382d912014-04-03 14:47:31 -07003276 hash = fault_mutex_hash(h, mm, vma, mapping, idx, address);
3277 mutex_lock(&htlb_fault_mutex_table[hash]);
3278
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07003279 entry = huge_ptep_get(ptep);
3280 if (huge_pte_none(entry)) {
Davidlohr Bueso8382d912014-04-03 14:47:31 -07003281 ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags);
David Gibsonb4d1d992008-10-15 22:01:11 -07003282 goto out_mutex;
David Gibson3935baa2006-03-22 00:08:53 -08003283 }
Adam Litke86e52162006-01-06 00:10:43 -08003284
Nick Piggin83c54072007-07-19 01:47:05 -07003285 ret = 0;
David Gibson1e8f8892006-01-06 00:10:44 -08003286
Andy Whitcroft57303d82008-08-12 15:08:47 -07003287 /*
Naoya Horiguchi0f792cf2015-02-11 15:25:25 -08003288 * entry could be a migration/hwpoison entry at this point, so this
3289 * check prevents the kernel from going below assuming that we have
3290 * a active hugepage in pagecache. This goto expects the 2nd page fault,
3291 * and is_hugetlb_entry_(migration|hwpoisoned) check will properly
3292 * handle it.
3293 */
3294 if (!pte_present(entry))
3295 goto out_mutex;
3296
3297 /*
Andy Whitcroft57303d82008-08-12 15:08:47 -07003298 * If we are going to COW the mapping later, we examine the pending
3299 * reservations for this page now. This will ensure that any
3300 * allocations necessary to record that reservation occur outside the
3301 * spinlock. For private mappings, we also lookup the pagecache
3302 * page now as it is used to determine if a reservation has been
3303 * consumed.
3304 */
Gerald Schaefer106c9922013-04-29 15:07:23 -07003305 if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) {
Andy Whitcroft2b267362008-08-12 15:08:49 -07003306 if (vma_needs_reservation(h, vma, address) < 0) {
3307 ret = VM_FAULT_OOM;
David Gibsonb4d1d992008-10-15 22:01:11 -07003308 goto out_mutex;
Andy Whitcroft2b267362008-08-12 15:08:49 -07003309 }
Andy Whitcroft57303d82008-08-12 15:08:47 -07003310
Mel Gormanf83a2752009-05-28 14:34:40 -07003311 if (!(vma->vm_flags & VM_MAYSHARE))
Andy Whitcroft57303d82008-08-12 15:08:47 -07003312 pagecache_page = hugetlbfs_pagecache_page(h,
3313 vma, address);
3314 }
3315
Naoya Horiguchi0f792cf2015-02-11 15:25:25 -08003316 ptl = huge_pte_lock(h, mm, ptep);
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09003317
David Gibson1e8f8892006-01-06 00:10:44 -08003318 /* Check for a racing update before calling hugetlb_cow */
David Gibsonb4d1d992008-10-15 22:01:11 -07003319 if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003320 goto out_ptl;
David Gibsonb4d1d992008-10-15 22:01:11 -07003321
Naoya Horiguchi0f792cf2015-02-11 15:25:25 -08003322 /*
3323 * hugetlb_cow() requires page locks of pte_page(entry) and
3324 * pagecache_page, so here we need take the former one
3325 * when page != pagecache_page or !pagecache_page.
3326 */
3327 page = pte_page(entry);
3328 if (page != pagecache_page)
3329 if (!trylock_page(page)) {
3330 need_wait_lock = 1;
3331 goto out_ptl;
3332 }
3333
3334 get_page(page);
David Gibsonb4d1d992008-10-15 22:01:11 -07003335
Hugh Dickins788c7df2009-06-23 13:49:05 +01003336 if (flags & FAULT_FLAG_WRITE) {
Gerald Schaefer106c9922013-04-29 15:07:23 -07003337 if (!huge_pte_write(entry)) {
Andy Whitcroft57303d82008-08-12 15:08:47 -07003338 ret = hugetlb_cow(mm, vma, address, ptep, entry,
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003339 pagecache_page, ptl);
Naoya Horiguchi0f792cf2015-02-11 15:25:25 -08003340 goto out_put_page;
David Gibsonb4d1d992008-10-15 22:01:11 -07003341 }
Gerald Schaefer106c9922013-04-29 15:07:23 -07003342 entry = huge_pte_mkdirty(entry);
David Gibsonb4d1d992008-10-15 22:01:11 -07003343 }
3344 entry = pte_mkyoung(entry);
Hugh Dickins788c7df2009-06-23 13:49:05 +01003345 if (huge_ptep_set_access_flags(vma, address, ptep, entry,
3346 flags & FAULT_FLAG_WRITE))
Russell King4b3073e2009-12-18 16:40:18 +00003347 update_mmu_cache(vma, address, ptep);
Naoya Horiguchi0f792cf2015-02-11 15:25:25 -08003348out_put_page:
3349 if (page != pagecache_page)
3350 unlock_page(page);
3351 put_page(page);
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003352out_ptl:
3353 spin_unlock(ptl);
Andy Whitcroft57303d82008-08-12 15:08:47 -07003354
3355 if (pagecache_page) {
3356 unlock_page(pagecache_page);
3357 put_page(pagecache_page);
3358 }
David Gibsonb4d1d992008-10-15 22:01:11 -07003359out_mutex:
Davidlohr Bueso8382d912014-04-03 14:47:31 -07003360 mutex_unlock(&htlb_fault_mutex_table[hash]);
Naoya Horiguchi0f792cf2015-02-11 15:25:25 -08003361 /*
3362 * Generally it's safe to hold refcount during waiting page lock. But
3363 * here we just wait to defer the next page fault to avoid busy loop and
3364 * the page is not used after unlocked before returning from the current
3365 * page fault. So we are safe from accessing freed page, even if we wait
3366 * here without taking refcount.
3367 */
3368 if (need_wait_lock)
3369 wait_on_page_locked(page);
David Gibson1e8f8892006-01-06 00:10:44 -08003370 return ret;
Adam Litke86e52162006-01-06 00:10:43 -08003371}
3372
Michel Lespinasse28a35712013-02-22 16:35:55 -08003373long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
3374 struct page **pages, struct vm_area_struct **vmas,
3375 unsigned long *position, unsigned long *nr_pages,
3376 long i, unsigned int flags)
David Gibson63551ae2005-06-21 17:14:44 -07003377{
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08003378 unsigned long pfn_offset;
3379 unsigned long vaddr = *position;
Michel Lespinasse28a35712013-02-22 16:35:55 -08003380 unsigned long remainder = *nr_pages;
Andi Kleena5516432008-07-23 21:27:41 -07003381 struct hstate *h = hstate_vma(vma);
David Gibson63551ae2005-06-21 17:14:44 -07003382
David Gibson63551ae2005-06-21 17:14:44 -07003383 while (vaddr < vma->vm_end && remainder) {
Adam Litke4c887262005-10-29 18:16:46 -07003384 pte_t *pte;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003385 spinlock_t *ptl = NULL;
Hugh Dickins2a15efc2009-09-21 17:03:27 -07003386 int absent;
Adam Litke4c887262005-10-29 18:16:46 -07003387 struct page *page;
3388
3389 /*
David Rientjes02057962015-04-14 15:48:24 -07003390 * If we have a pending SIGKILL, don't keep faulting pages and
3391 * potentially allocating memory.
3392 */
3393 if (unlikely(fatal_signal_pending(current))) {
3394 remainder = 0;
3395 break;
3396 }
3397
3398 /*
Adam Litke4c887262005-10-29 18:16:46 -07003399 * Some archs (sparc64, sh*) have multiple pte_ts to
Hugh Dickins2a15efc2009-09-21 17:03:27 -07003400 * each hugepage. We have to make sure we get the
Adam Litke4c887262005-10-29 18:16:46 -07003401 * first, for the page indexing below to work.
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003402 *
3403 * Note that page table lock is not held when pte is null.
Adam Litke4c887262005-10-29 18:16:46 -07003404 */
Andi Kleena5516432008-07-23 21:27:41 -07003405 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h));
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003406 if (pte)
3407 ptl = huge_pte_lock(h, mm, pte);
Hugh Dickins2a15efc2009-09-21 17:03:27 -07003408 absent = !pte || huge_pte_none(huge_ptep_get(pte));
Adam Litke4c887262005-10-29 18:16:46 -07003409
Hugh Dickins2a15efc2009-09-21 17:03:27 -07003410 /*
3411 * When coredumping, it suits get_dump_page if we just return
Hugh Dickins3ae77f42009-09-21 17:03:33 -07003412 * an error where there's an empty slot with no huge pagecache
3413 * to back it. This way, we avoid allocating a hugepage, and
3414 * the sparse dumpfile avoids allocating disk blocks, but its
3415 * huge holes still show up with zeroes where they need to be.
Hugh Dickins2a15efc2009-09-21 17:03:27 -07003416 */
Hugh Dickins3ae77f42009-09-21 17:03:33 -07003417 if (absent && (flags & FOLL_DUMP) &&
3418 !hugetlbfs_pagecache_present(h, vma, vaddr)) {
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003419 if (pte)
3420 spin_unlock(ptl);
Hugh Dickins2a15efc2009-09-21 17:03:27 -07003421 remainder = 0;
3422 break;
3423 }
3424
Naoya Horiguchi9cc3a5b2013-04-17 15:58:30 -07003425 /*
3426 * We need call hugetlb_fault for both hugepages under migration
3427 * (in which case hugetlb_fault waits for the migration,) and
3428 * hwpoisoned hugepages (in which case we need to prevent the
3429 * caller from accessing to them.) In order to do this, we use
3430 * here is_swap_pte instead of is_hugetlb_entry_migration and
3431 * is_hugetlb_entry_hwpoisoned. This is because it simply covers
3432 * both cases, and because we can't follow correct pages
3433 * directly from any kind of swap entries.
3434 */
3435 if (absent || is_swap_pte(huge_ptep_get(pte)) ||
Gerald Schaefer106c9922013-04-29 15:07:23 -07003436 ((flags & FOLL_WRITE) &&
3437 !huge_pte_write(huge_ptep_get(pte)))) {
Adam Litke4c887262005-10-29 18:16:46 -07003438 int ret;
3439
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003440 if (pte)
3441 spin_unlock(ptl);
Hugh Dickins2a15efc2009-09-21 17:03:27 -07003442 ret = hugetlb_fault(mm, vma, vaddr,
3443 (flags & FOLL_WRITE) ? FAULT_FLAG_WRITE : 0);
Adam Litkea89182c2007-08-22 14:01:51 -07003444 if (!(ret & VM_FAULT_ERROR))
Adam Litke4c887262005-10-29 18:16:46 -07003445 continue;
3446
3447 remainder = 0;
Adam Litke4c887262005-10-29 18:16:46 -07003448 break;
3449 }
David Gibson63551ae2005-06-21 17:14:44 -07003450
Andi Kleena5516432008-07-23 21:27:41 -07003451 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07003452 page = pte_page(huge_ptep_get(pte));
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08003453same_page:
Chen, Kenneth Wd6692182006-03-31 02:29:57 -08003454 if (pages) {
Hugh Dickins2a15efc2009-09-21 17:03:27 -07003455 pages[i] = mem_map_offset(page, pfn_offset);
Andrea Arcangelia0368d42014-01-21 15:48:49 -08003456 get_page_foll(pages[i]);
Chen, Kenneth Wd6692182006-03-31 02:29:57 -08003457 }
David Gibson63551ae2005-06-21 17:14:44 -07003458
3459 if (vmas)
3460 vmas[i] = vma;
3461
3462 vaddr += PAGE_SIZE;
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08003463 ++pfn_offset;
David Gibson63551ae2005-06-21 17:14:44 -07003464 --remainder;
3465 ++i;
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08003466 if (vaddr < vma->vm_end && remainder &&
Andi Kleena5516432008-07-23 21:27:41 -07003467 pfn_offset < pages_per_huge_page(h)) {
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08003468 /*
3469 * We use pfn_offset to avoid touching the pageframes
3470 * of this compound page.
3471 */
3472 goto same_page;
3473 }
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003474 spin_unlock(ptl);
David Gibson63551ae2005-06-21 17:14:44 -07003475 }
Michel Lespinasse28a35712013-02-22 16:35:55 -08003476 *nr_pages = remainder;
David Gibson63551ae2005-06-21 17:14:44 -07003477 *position = vaddr;
3478
Hugh Dickins2a15efc2009-09-21 17:03:27 -07003479 return i ? i : -EFAULT;
David Gibson63551ae2005-06-21 17:14:44 -07003480}
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003481
Peter Zijlstra7da4d642012-11-19 03:14:23 +01003482unsigned long hugetlb_change_protection(struct vm_area_struct *vma,
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003483 unsigned long address, unsigned long end, pgprot_t newprot)
3484{
3485 struct mm_struct *mm = vma->vm_mm;
3486 unsigned long start = address;
3487 pte_t *ptep;
3488 pte_t pte;
Andi Kleena5516432008-07-23 21:27:41 -07003489 struct hstate *h = hstate_vma(vma);
Peter Zijlstra7da4d642012-11-19 03:14:23 +01003490 unsigned long pages = 0;
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003491
3492 BUG_ON(address >= end);
3493 flush_cache_range(vma, address, end);
3494
Rik van Riela5338092014-04-07 15:36:57 -07003495 mmu_notifier_invalidate_range_start(mm, start, end);
Davidlohr Bueso83cde9e2014-12-12 16:54:21 -08003496 i_mmap_lock_write(vma->vm_file->f_mapping);
Andi Kleena5516432008-07-23 21:27:41 -07003497 for (; address < end; address += huge_page_size(h)) {
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003498 spinlock_t *ptl;
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003499 ptep = huge_pte_offset(mm, address);
3500 if (!ptep)
3501 continue;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003502 ptl = huge_pte_lock(h, mm, ptep);
Peter Zijlstra7da4d642012-11-19 03:14:23 +01003503 if (huge_pmd_unshare(mm, &address, ptep)) {
3504 pages++;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003505 spin_unlock(ptl);
Chen, Kenneth W39dde652006-12-06 20:32:03 -08003506 continue;
Peter Zijlstra7da4d642012-11-19 03:14:23 +01003507 }
Naoya Horiguchia8bda282015-02-11 15:25:28 -08003508 pte = huge_ptep_get(ptep);
3509 if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) {
3510 spin_unlock(ptl);
3511 continue;
3512 }
3513 if (unlikely(is_hugetlb_entry_migration(pte))) {
3514 swp_entry_t entry = pte_to_swp_entry(pte);
3515
3516 if (is_write_migration_entry(entry)) {
3517 pte_t newpte;
3518
3519 make_migration_entry_read(&entry);
3520 newpte = swp_entry_to_pte(entry);
3521 set_huge_pte_at(mm, address, ptep, newpte);
3522 pages++;
3523 }
3524 spin_unlock(ptl);
3525 continue;
3526 }
3527 if (!huge_pte_none(pte)) {
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003528 pte = huge_ptep_get_and_clear(mm, address, ptep);
Gerald Schaefer106c9922013-04-29 15:07:23 -07003529 pte = pte_mkhuge(huge_pte_modify(pte, newprot));
Tony Lube7517d2013-02-04 14:28:46 -08003530 pte = arch_make_huge_pte(pte, vma, NULL, 0);
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003531 set_huge_pte_at(mm, address, ptep, pte);
Peter Zijlstra7da4d642012-11-19 03:14:23 +01003532 pages++;
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003533 }
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003534 spin_unlock(ptl);
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003535 }
Mel Gormand8333522012-07-31 16:46:20 -07003536 /*
Davidlohr Buesoc8c06ef2014-12-12 16:54:24 -08003537 * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare
Mel Gormand8333522012-07-31 16:46:20 -07003538 * may have cleared our pud entry and done put_page on the page table:
Davidlohr Buesoc8c06ef2014-12-12 16:54:24 -08003539 * once we release i_mmap_rwsem, another task can do the final put_page
Mel Gormand8333522012-07-31 16:46:20 -07003540 * and that page table be reused and filled with junk.
3541 */
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003542 flush_tlb_range(vma, start, end);
Joerg Roedel34ee6452014-11-13 13:46:09 +11003543 mmu_notifier_invalidate_range(mm, start, end);
Davidlohr Bueso83cde9e2014-12-12 16:54:21 -08003544 i_mmap_unlock_write(vma->vm_file->f_mapping);
Rik van Riela5338092014-04-07 15:36:57 -07003545 mmu_notifier_invalidate_range_end(mm, start, end);
Peter Zijlstra7da4d642012-11-19 03:14:23 +01003546
3547 return pages << h->order;
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003548}
3549
Mel Gormana1e78772008-07-23 21:27:23 -07003550int hugetlb_reserve_pages(struct inode *inode,
3551 long from, long to,
Mel Gorman5a6fe122009-02-10 14:02:27 +00003552 struct vm_area_struct *vma,
KOSAKI Motohiroca16d142011-05-26 19:16:19 +09003553 vm_flags_t vm_flags)
Adam Litkee4e574b2007-10-16 01:26:19 -07003554{
Mel Gorman17c9d122009-02-11 16:34:16 +00003555 long ret, chg;
Andi Kleena5516432008-07-23 21:27:41 -07003556 struct hstate *h = hstate_inode(inode);
David Gibson90481622012-03-21 16:34:12 -07003557 struct hugepage_subpool *spool = subpool_inode(inode);
Joonsoo Kim9119a412014-04-03 14:47:25 -07003558 struct resv_map *resv_map;
Mike Kravetz1c5ecae2015-04-15 16:13:39 -07003559 long gbl_reserve;
Adam Litkee4e574b2007-10-16 01:26:19 -07003560
Mel Gormana1e78772008-07-23 21:27:23 -07003561 /*
Mel Gorman17c9d122009-02-11 16:34:16 +00003562 * Only apply hugepage reservation if asked. At fault time, an
3563 * attempt will be made for VM_NORESERVE to allocate a page
David Gibson90481622012-03-21 16:34:12 -07003564 * without using reserves
Mel Gorman17c9d122009-02-11 16:34:16 +00003565 */
KOSAKI Motohiroca16d142011-05-26 19:16:19 +09003566 if (vm_flags & VM_NORESERVE)
Mel Gorman17c9d122009-02-11 16:34:16 +00003567 return 0;
3568
3569 /*
Mel Gormana1e78772008-07-23 21:27:23 -07003570 * Shared mappings base their reservation on the number of pages that
3571 * are already allocated on behalf of the file. Private mappings need
3572 * to reserve the full area even if read-only as mprotect() may be
3573 * called to make the mapping read-write. Assume !vma is a shm mapping
3574 */
Joonsoo Kim9119a412014-04-03 14:47:25 -07003575 if (!vma || vma->vm_flags & VM_MAYSHARE) {
Joonsoo Kim4e35f482014-04-03 14:47:30 -07003576 resv_map = inode_resv_map(inode);
Joonsoo Kim9119a412014-04-03 14:47:25 -07003577
Joonsoo Kim1406ec92014-04-03 14:47:26 -07003578 chg = region_chg(resv_map, from, to);
Joonsoo Kim9119a412014-04-03 14:47:25 -07003579
3580 } else {
3581 resv_map = resv_map_alloc();
Mel Gorman5a6fe122009-02-10 14:02:27 +00003582 if (!resv_map)
3583 return -ENOMEM;
3584
Mel Gorman17c9d122009-02-11 16:34:16 +00003585 chg = to - from;
3586
Mel Gorman5a6fe122009-02-10 14:02:27 +00003587 set_vma_resv_map(vma, resv_map);
3588 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
3589 }
3590
Dave Hansenc50ac052012-05-29 15:06:46 -07003591 if (chg < 0) {
3592 ret = chg;
3593 goto out_err;
3594 }
Mel Gorman17c9d122009-02-11 16:34:16 +00003595
Mike Kravetz1c5ecae2015-04-15 16:13:39 -07003596 /*
3597 * There must be enough pages in the subpool for the mapping. If
3598 * the subpool has a minimum size, there may be some global
3599 * reservations already in place (gbl_reserve).
3600 */
3601 gbl_reserve = hugepage_subpool_get_pages(spool, chg);
3602 if (gbl_reserve < 0) {
Dave Hansenc50ac052012-05-29 15:06:46 -07003603 ret = -ENOSPC;
3604 goto out_err;
3605 }
Mel Gorman17c9d122009-02-11 16:34:16 +00003606
3607 /*
3608 * Check enough hugepages are available for the reservation.
David Gibson90481622012-03-21 16:34:12 -07003609 * Hand the pages back to the subpool if there are not
Mel Gorman17c9d122009-02-11 16:34:16 +00003610 */
Mike Kravetz1c5ecae2015-04-15 16:13:39 -07003611 ret = hugetlb_acct_memory(h, gbl_reserve);
Mel Gorman17c9d122009-02-11 16:34:16 +00003612 if (ret < 0) {
Mike Kravetz1c5ecae2015-04-15 16:13:39 -07003613 /* put back original number of pages, chg */
3614 (void)hugepage_subpool_put_pages(spool, chg);
Dave Hansenc50ac052012-05-29 15:06:46 -07003615 goto out_err;
Mel Gorman17c9d122009-02-11 16:34:16 +00003616 }
3617
3618 /*
3619 * Account for the reservations made. Shared mappings record regions
3620 * that have reservations as they are shared by multiple VMAs.
3621 * When the last VMA disappears, the region map says how much
3622 * the reservation was and the page cache tells how much of
3623 * the reservation was consumed. Private mappings are per-VMA and
3624 * only the consumed reservations are tracked. When the VMA
3625 * disappears, the original reservation is the VMA size and the
3626 * consumed reservations are stored in the map. Hence, nothing
3627 * else has to be done for private mappings here
3628 */
Mel Gormanf83a2752009-05-28 14:34:40 -07003629 if (!vma || vma->vm_flags & VM_MAYSHARE)
Joonsoo Kim1406ec92014-04-03 14:47:26 -07003630 region_add(resv_map, from, to);
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -07003631 return 0;
Dave Hansenc50ac052012-05-29 15:06:46 -07003632out_err:
Joonsoo Kimf031dd22014-04-03 14:47:28 -07003633 if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
3634 kref_put(&resv_map->refs, resv_map_release);
Dave Hansenc50ac052012-05-29 15:06:46 -07003635 return ret;
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -07003636}
3637
3638void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed)
3639{
Andi Kleena5516432008-07-23 21:27:41 -07003640 struct hstate *h = hstate_inode(inode);
Joonsoo Kim4e35f482014-04-03 14:47:30 -07003641 struct resv_map *resv_map = inode_resv_map(inode);
Joonsoo Kim9119a412014-04-03 14:47:25 -07003642 long chg = 0;
David Gibson90481622012-03-21 16:34:12 -07003643 struct hugepage_subpool *spool = subpool_inode(inode);
Mike Kravetz1c5ecae2015-04-15 16:13:39 -07003644 long gbl_reserve;
Ken Chen45c682a2007-11-14 16:59:44 -08003645
Joonsoo Kim9119a412014-04-03 14:47:25 -07003646 if (resv_map)
Joonsoo Kim1406ec92014-04-03 14:47:26 -07003647 chg = region_truncate(resv_map, offset);
Ken Chen45c682a2007-11-14 16:59:44 -08003648 spin_lock(&inode->i_lock);
Eric Sandeene4c6f8b2009-07-29 15:02:16 -07003649 inode->i_blocks -= (blocks_per_huge_page(h) * freed);
Ken Chen45c682a2007-11-14 16:59:44 -08003650 spin_unlock(&inode->i_lock);
3651
Mike Kravetz1c5ecae2015-04-15 16:13:39 -07003652 /*
3653 * If the subpool has a minimum size, the number of global
3654 * reservations to be released may be adjusted.
3655 */
3656 gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed));
3657 hugetlb_acct_memory(h, -gbl_reserve);
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -07003658}
Naoya Horiguchi93f70f92010-05-28 09:29:20 +09003659
Steve Capper3212b532013-04-23 12:35:02 +01003660#ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
3661static unsigned long page_table_shareable(struct vm_area_struct *svma,
3662 struct vm_area_struct *vma,
3663 unsigned long addr, pgoff_t idx)
3664{
3665 unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) +
3666 svma->vm_start;
3667 unsigned long sbase = saddr & PUD_MASK;
3668 unsigned long s_end = sbase + PUD_SIZE;
3669
3670 /* Allow segments to share if only one is marked locked */
3671 unsigned long vm_flags = vma->vm_flags & ~VM_LOCKED;
3672 unsigned long svm_flags = svma->vm_flags & ~VM_LOCKED;
3673
3674 /*
3675 * match the virtual addresses, permission and the alignment of the
3676 * page table page.
3677 */
3678 if (pmd_index(addr) != pmd_index(saddr) ||
3679 vm_flags != svm_flags ||
3680 sbase < svma->vm_start || svma->vm_end < s_end)
3681 return 0;
3682
3683 return saddr;
3684}
3685
3686static int vma_shareable(struct vm_area_struct *vma, unsigned long addr)
3687{
3688 unsigned long base = addr & PUD_MASK;
3689 unsigned long end = base + PUD_SIZE;
3690
3691 /*
3692 * check on proper vm_flags and page table alignment
3693 */
3694 if (vma->vm_flags & VM_MAYSHARE &&
3695 vma->vm_start <= base && end <= vma->vm_end)
3696 return 1;
3697 return 0;
3698}
3699
3700/*
3701 * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
3702 * and returns the corresponding pte. While this is not necessary for the
3703 * !shared pmd case because we can allocate the pmd later as well, it makes the
3704 * code much cleaner. pmd allocation is essential for the shared case because
Davidlohr Buesoc8c06ef2014-12-12 16:54:24 -08003705 * pud has to be populated inside the same i_mmap_rwsem section - otherwise
Steve Capper3212b532013-04-23 12:35:02 +01003706 * racing tasks could either miss the sharing (see huge_pte_offset) or select a
3707 * bad pmd for sharing.
3708 */
3709pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
3710{
3711 struct vm_area_struct *vma = find_vma(mm, addr);
3712 struct address_space *mapping = vma->vm_file->f_mapping;
3713 pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) +
3714 vma->vm_pgoff;
3715 struct vm_area_struct *svma;
3716 unsigned long saddr;
3717 pte_t *spte = NULL;
3718 pte_t *pte;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003719 spinlock_t *ptl;
Steve Capper3212b532013-04-23 12:35:02 +01003720
3721 if (!vma_shareable(vma, addr))
3722 return (pte_t *)pmd_alloc(mm, pud, addr);
3723
Davidlohr Bueso83cde9e2014-12-12 16:54:21 -08003724 i_mmap_lock_write(mapping);
Steve Capper3212b532013-04-23 12:35:02 +01003725 vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) {
3726 if (svma == vma)
3727 continue;
3728
3729 saddr = page_table_shareable(svma, vma, addr, idx);
3730 if (saddr) {
3731 spte = huge_pte_offset(svma->vm_mm, saddr);
3732 if (spte) {
Kirill A. Shutemovdc6c9a32015-02-11 15:26:50 -08003733 mm_inc_nr_pmds(mm);
Steve Capper3212b532013-04-23 12:35:02 +01003734 get_page(virt_to_page(spte));
3735 break;
3736 }
3737 }
3738 }
3739
3740 if (!spte)
3741 goto out;
3742
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003743 ptl = huge_pte_lockptr(hstate_vma(vma), mm, spte);
3744 spin_lock(ptl);
Kirill A. Shutemovdc6c9a32015-02-11 15:26:50 -08003745 if (pud_none(*pud)) {
Steve Capper3212b532013-04-23 12:35:02 +01003746 pud_populate(mm, pud,
3747 (pmd_t *)((unsigned long)spte & PAGE_MASK));
Kirill A. Shutemovdc6c9a32015-02-11 15:26:50 -08003748 } else {
Steve Capper3212b532013-04-23 12:35:02 +01003749 put_page(virt_to_page(spte));
Kirill A. Shutemovdc6c9a32015-02-11 15:26:50 -08003750 mm_inc_nr_pmds(mm);
3751 }
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003752 spin_unlock(ptl);
Steve Capper3212b532013-04-23 12:35:02 +01003753out:
3754 pte = (pte_t *)pmd_alloc(mm, pud, addr);
Davidlohr Bueso83cde9e2014-12-12 16:54:21 -08003755 i_mmap_unlock_write(mapping);
Steve Capper3212b532013-04-23 12:35:02 +01003756 return pte;
3757}
3758
3759/*
3760 * unmap huge page backed by shared pte.
3761 *
3762 * Hugetlb pte page is ref counted at the time of mapping. If pte is shared
3763 * indicated by page_count > 1, unmap is achieved by clearing pud and
3764 * decrementing the ref count. If count == 1, the pte page is not shared.
3765 *
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003766 * called with page table lock held.
Steve Capper3212b532013-04-23 12:35:02 +01003767 *
3768 * returns: 1 successfully unmapped a shared pte page
3769 * 0 the underlying pte page is not shared, or it is the last user
3770 */
3771int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
3772{
3773 pgd_t *pgd = pgd_offset(mm, *addr);
3774 pud_t *pud = pud_offset(pgd, *addr);
3775
3776 BUG_ON(page_count(virt_to_page(ptep)) == 0);
3777 if (page_count(virt_to_page(ptep)) == 1)
3778 return 0;
3779
3780 pud_clear(pud);
3781 put_page(virt_to_page(ptep));
Kirill A. Shutemovdc6c9a32015-02-11 15:26:50 -08003782 mm_dec_nr_pmds(mm);
Steve Capper3212b532013-04-23 12:35:02 +01003783 *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE;
3784 return 1;
3785}
Steve Capper9e5fc742013-04-30 08:02:03 +01003786#define want_pmd_share() (1)
3787#else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
3788pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
3789{
3790 return NULL;
3791}
3792#define want_pmd_share() (0)
Steve Capper3212b532013-04-23 12:35:02 +01003793#endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
3794
Steve Capper9e5fc742013-04-30 08:02:03 +01003795#ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
3796pte_t *huge_pte_alloc(struct mm_struct *mm,
3797 unsigned long addr, unsigned long sz)
3798{
3799 pgd_t *pgd;
3800 pud_t *pud;
3801 pte_t *pte = NULL;
3802
3803 pgd = pgd_offset(mm, addr);
3804 pud = pud_alloc(mm, pgd, addr);
3805 if (pud) {
3806 if (sz == PUD_SIZE) {
3807 pte = (pte_t *)pud;
3808 } else {
3809 BUG_ON(sz != PMD_SIZE);
3810 if (want_pmd_share() && pud_none(*pud))
3811 pte = huge_pmd_share(mm, addr, pud);
3812 else
3813 pte = (pte_t *)pmd_alloc(mm, pud, addr);
3814 }
3815 }
3816 BUG_ON(pte && !pte_none(*pte) && !pte_huge(*pte));
3817
3818 return pte;
3819}
3820
3821pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr)
3822{
3823 pgd_t *pgd;
3824 pud_t *pud;
3825 pmd_t *pmd = NULL;
3826
3827 pgd = pgd_offset(mm, addr);
3828 if (pgd_present(*pgd)) {
3829 pud = pud_offset(pgd, addr);
3830 if (pud_present(*pud)) {
3831 if (pud_huge(*pud))
3832 return (pte_t *)pud;
3833 pmd = pmd_offset(pud, addr);
3834 }
3835 }
3836 return (pte_t *) pmd;
3837}
3838
Naoya Horiguchi61f77ed2015-02-11 15:25:15 -08003839#endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
3840
3841/*
3842 * These functions are overwritable if your architecture needs its own
3843 * behavior.
3844 */
3845struct page * __weak
3846follow_huge_addr(struct mm_struct *mm, unsigned long address,
3847 int write)
3848{
3849 return ERR_PTR(-EINVAL);
3850}
3851
3852struct page * __weak
Steve Capper9e5fc742013-04-30 08:02:03 +01003853follow_huge_pmd(struct mm_struct *mm, unsigned long address,
Naoya Horiguchie66f17f2015-02-11 15:25:22 -08003854 pmd_t *pmd, int flags)
Steve Capper9e5fc742013-04-30 08:02:03 +01003855{
Naoya Horiguchie66f17f2015-02-11 15:25:22 -08003856 struct page *page = NULL;
3857 spinlock_t *ptl;
3858retry:
3859 ptl = pmd_lockptr(mm, pmd);
3860 spin_lock(ptl);
3861 /*
3862 * make sure that the address range covered by this pmd is not
3863 * unmapped from other threads.
3864 */
3865 if (!pmd_huge(*pmd))
3866 goto out;
3867 if (pmd_present(*pmd)) {
Gerald Schaefer97534122015-04-14 15:42:30 -07003868 page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT);
Naoya Horiguchie66f17f2015-02-11 15:25:22 -08003869 if (flags & FOLL_GET)
3870 get_page(page);
3871 } else {
3872 if (is_hugetlb_entry_migration(huge_ptep_get((pte_t *)pmd))) {
3873 spin_unlock(ptl);
3874 __migration_entry_wait(mm, (pte_t *)pmd, ptl);
3875 goto retry;
3876 }
3877 /*
3878 * hwpoisoned entry is treated as no_page_table in
3879 * follow_page_mask().
3880 */
3881 }
3882out:
3883 spin_unlock(ptl);
Steve Capper9e5fc742013-04-30 08:02:03 +01003884 return page;
3885}
3886
Naoya Horiguchi61f77ed2015-02-11 15:25:15 -08003887struct page * __weak
Steve Capper9e5fc742013-04-30 08:02:03 +01003888follow_huge_pud(struct mm_struct *mm, unsigned long address,
Naoya Horiguchie66f17f2015-02-11 15:25:22 -08003889 pud_t *pud, int flags)
Steve Capper9e5fc742013-04-30 08:02:03 +01003890{
Naoya Horiguchie66f17f2015-02-11 15:25:22 -08003891 if (flags & FOLL_GET)
3892 return NULL;
Steve Capper9e5fc742013-04-30 08:02:03 +01003893
Naoya Horiguchie66f17f2015-02-11 15:25:22 -08003894 return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT);
Steve Capper9e5fc742013-04-30 08:02:03 +01003895}
3896
Andi Kleend5bd9102010-09-27 09:00:12 +02003897#ifdef CONFIG_MEMORY_FAILURE
3898
Naoya Horiguchi93f70f92010-05-28 09:29:20 +09003899/*
3900 * This function is called from memory failure code.
3901 * Assume the caller holds page lock of the head page.
3902 */
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +09003903int dequeue_hwpoisoned_huge_page(struct page *hpage)
Naoya Horiguchi93f70f92010-05-28 09:29:20 +09003904{
3905 struct hstate *h = page_hstate(hpage);
3906 int nid = page_to_nid(hpage);
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +09003907 int ret = -EBUSY;
Naoya Horiguchi93f70f92010-05-28 09:29:20 +09003908
3909 spin_lock(&hugetlb_lock);
Naoya Horiguchi7e1f0492015-04-15 16:14:41 -07003910 /*
3911 * Just checking !page_huge_active is not enough, because that could be
3912 * an isolated/hwpoisoned hugepage (which have >0 refcount).
3913 */
3914 if (!page_huge_active(hpage) && !page_count(hpage)) {
Naoya Horiguchi56f2fb12012-12-12 13:52:33 -08003915 /*
3916 * Hwpoisoned hugepage isn't linked to activelist or freelist,
3917 * but dangling hpage->lru can trigger list-debug warnings
3918 * (this happens when we call unpoison_memory() on it),
3919 * so let it point to itself with list_del_init().
3920 */
3921 list_del_init(&hpage->lru);
Naoya Horiguchi8c6c2ec2010-09-08 10:19:38 +09003922 set_page_refcounted(hpage);
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +09003923 h->free_huge_pages--;
3924 h->free_huge_pages_node[nid]--;
3925 ret = 0;
3926 }
Naoya Horiguchi93f70f92010-05-28 09:29:20 +09003927 spin_unlock(&hugetlb_lock);
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +09003928 return ret;
Naoya Horiguchi93f70f92010-05-28 09:29:20 +09003929}
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +09003930#endif
Naoya Horiguchi31caf662013-09-11 14:21:59 -07003931
3932bool isolate_huge_page(struct page *page, struct list_head *list)
3933{
Naoya Horiguchibcc54222015-04-15 16:14:38 -07003934 bool ret = true;
3935
Sasha Levin309381fea2014-01-23 15:52:54 -08003936 VM_BUG_ON_PAGE(!PageHead(page), page);
Naoya Horiguchi31caf662013-09-11 14:21:59 -07003937 spin_lock(&hugetlb_lock);
Naoya Horiguchibcc54222015-04-15 16:14:38 -07003938 if (!page_huge_active(page) || !get_page_unless_zero(page)) {
3939 ret = false;
3940 goto unlock;
3941 }
3942 clear_page_huge_active(page);
Naoya Horiguchi31caf662013-09-11 14:21:59 -07003943 list_move_tail(&page->lru, list);
Naoya Horiguchibcc54222015-04-15 16:14:38 -07003944unlock:
Naoya Horiguchi31caf662013-09-11 14:21:59 -07003945 spin_unlock(&hugetlb_lock);
Naoya Horiguchibcc54222015-04-15 16:14:38 -07003946 return ret;
Naoya Horiguchi31caf662013-09-11 14:21:59 -07003947}
3948
3949void putback_active_hugepage(struct page *page)
3950{
Sasha Levin309381fea2014-01-23 15:52:54 -08003951 VM_BUG_ON_PAGE(!PageHead(page), page);
Naoya Horiguchi31caf662013-09-11 14:21:59 -07003952 spin_lock(&hugetlb_lock);
Naoya Horiguchibcc54222015-04-15 16:14:38 -07003953 set_page_huge_active(page);
Naoya Horiguchi31caf662013-09-11 14:21:59 -07003954 list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist);
3955 spin_unlock(&hugetlb_lock);
3956 put_page(page);
3957}