blob: 85032de5e20f88bdef36f8fa54b71ae826f53f13 [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
2 * Generic hugetlb support.
Nadia Yvette Chambers6d49e352012-12-06 10:39:54 +01003 * (C) Nadia Yvette Chambers, April 2004
Linus Torvalds1da177e2005-04-16 15:20:36 -07004 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07005#include <linux/list.h>
6#include <linux/init.h>
7#include <linux/module.h>
8#include <linux/mm.h>
Alexey Dobriyane1759c22008-10-15 23:50:22 +04009#include <linux/seq_file.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070010#include <linux/sysctl.h>
11#include <linux/highmem.h>
Andrea Arcangelicddb8a52008-07-28 15:46:29 -070012#include <linux/mmu_notifier.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070013#include <linux/nodemask.h>
David Gibson63551ae2005-06-21 17:14:44 -070014#include <linux/pagemap.h>
Christoph Lameter5da7ca82006-01-06 00:10:46 -080015#include <linux/mempolicy.h>
Gideon Israel Dsouza3b321232014-04-07 15:37:26 -070016#include <linux/compiler.h>
Christoph Lameteraea47ff2006-01-08 01:00:57 -080017#include <linux/cpuset.h>
David Gibson3935baa2006-03-22 00:08:53 -080018#include <linux/mutex.h>
Andi Kleenaa888a72008-07-23 21:27:47 -070019#include <linux/bootmem.h>
Nishanth Aravamudana3437872008-07-23 21:27:44 -070020#include <linux/sysfs.h>
Tejun Heo5a0e3ad2010-03-24 17:04:11 +090021#include <linux/slab.h>
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +090022#include <linux/rmap.h>
Naoya Horiguchifd6a03ed2010-05-28 09:29:21 +090023#include <linux/swap.h>
24#include <linux/swapops.h>
Naoya Horiguchic8721bb2013-09-11 14:22:09 -070025#include <linux/page-isolation.h>
Davidlohr Bueso8382d912014-04-03 14:47:31 -070026#include <linux/jhash.h>
Linus Torvaldsd6606682008-08-06 12:04:54 -070027
David Gibson63551ae2005-06-21 17:14:44 -070028#include <asm/page.h>
29#include <asm/pgtable.h>
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -070030#include <asm/tlb.h>
David Gibson63551ae2005-06-21 17:14:44 -070031
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -070032#include <linux/io.h>
David Gibson63551ae2005-06-21 17:14:44 -070033#include <linux/hugetlb.h>
Aneesh Kumar K.V9dd540e2012-07-31 16:42:15 -070034#include <linux/hugetlb_cgroup.h>
Lee Schermerhorn9a305232009-12-14 17:58:25 -080035#include <linux/node.h>
Nick Piggin7835e982006-03-22 00:08:40 -080036#include "internal.h"
Linus Torvalds1da177e2005-04-16 15:20:36 -070037
Mel Gorman396faf02007-07-17 04:03:13 -070038unsigned long 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
David Gibson90481622012-03-21 16:34:12 -070064static inline void unlock_or_release_subpool(struct hugepage_subpool *spool)
65{
66 bool free = (spool->count == 0) && (spool->used_hpages == 0);
67
68 spin_unlock(&spool->lock);
69
70 /* If no pages are used, and no other handles to the subpool
71 * remain, free the subpool the subpool remain */
72 if (free)
73 kfree(spool);
74}
75
76struct hugepage_subpool *hugepage_new_subpool(long nr_blocks)
77{
78 struct hugepage_subpool *spool;
79
80 spool = kmalloc(sizeof(*spool), GFP_KERNEL);
81 if (!spool)
82 return NULL;
83
84 spin_lock_init(&spool->lock);
85 spool->count = 1;
86 spool->max_hpages = nr_blocks;
87 spool->used_hpages = 0;
88
89 return spool;
90}
91
92void hugepage_put_subpool(struct hugepage_subpool *spool)
93{
94 spin_lock(&spool->lock);
95 BUG_ON(!spool->count);
96 spool->count--;
97 unlock_or_release_subpool(spool);
98}
99
100static int hugepage_subpool_get_pages(struct hugepage_subpool *spool,
101 long delta)
102{
103 int ret = 0;
104
105 if (!spool)
106 return 0;
107
108 spin_lock(&spool->lock);
109 if ((spool->used_hpages + delta) <= spool->max_hpages) {
110 spool->used_hpages += delta;
111 } else {
112 ret = -ENOMEM;
113 }
114 spin_unlock(&spool->lock);
115
116 return ret;
117}
118
119static void hugepage_subpool_put_pages(struct hugepage_subpool *spool,
120 long delta)
121{
122 if (!spool)
123 return;
124
125 spin_lock(&spool->lock);
126 spool->used_hpages -= delta;
127 /* If hugetlbfs_put_super couldn't free spool due to
128 * an outstanding quota reference, free it now. */
129 unlock_or_release_subpool(spool);
130}
131
132static inline struct hugepage_subpool *subpool_inode(struct inode *inode)
133{
134 return HUGETLBFS_SB(inode->i_sb)->spool;
135}
136
137static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma)
138{
Al Viro496ad9a2013-01-23 17:07:38 -0500139 return subpool_inode(file_inode(vma->vm_file));
David Gibson90481622012-03-21 16:34:12 -0700140}
141
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700142/*
Andy Whitcroft96822902008-07-23 21:27:29 -0700143 * Region tracking -- allows tracking of reservations and instantiated pages
144 * across the pages in a mapping.
Andy Whitcroft84afd992008-07-23 21:27:32 -0700145 *
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700146 * The region data structures are embedded into a resv_map and
147 * protected by a resv_map's lock
Andy Whitcroft96822902008-07-23 21:27:29 -0700148 */
149struct file_region {
150 struct list_head link;
151 long from;
152 long to;
153};
154
Joonsoo Kim1406ec92014-04-03 14:47:26 -0700155static long region_add(struct resv_map *resv, long f, long t)
Andy Whitcroft96822902008-07-23 21:27:29 -0700156{
Joonsoo Kim1406ec92014-04-03 14:47:26 -0700157 struct list_head *head = &resv->regions;
Andy Whitcroft96822902008-07-23 21:27:29 -0700158 struct file_region *rg, *nrg, *trg;
159
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700160 spin_lock(&resv->lock);
Andy Whitcroft96822902008-07-23 21:27:29 -0700161 /* Locate the region we are either in or before. */
162 list_for_each_entry(rg, head, link)
163 if (f <= rg->to)
164 break;
165
166 /* Round our left edge to the current segment if it encloses us. */
167 if (f > rg->from)
168 f = rg->from;
169
170 /* Check for and consume any regions we now overlap with. */
171 nrg = rg;
172 list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
173 if (&rg->link == head)
174 break;
175 if (rg->from > t)
176 break;
177
178 /* If this area reaches higher then extend our area to
179 * include it completely. If this is not the first area
180 * which we intend to reuse, free it. */
181 if (rg->to > t)
182 t = rg->to;
183 if (rg != nrg) {
184 list_del(&rg->link);
185 kfree(rg);
186 }
187 }
188 nrg->from = f;
189 nrg->to = t;
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700190 spin_unlock(&resv->lock);
Andy Whitcroft96822902008-07-23 21:27:29 -0700191 return 0;
192}
193
Joonsoo Kim1406ec92014-04-03 14:47:26 -0700194static long region_chg(struct resv_map *resv, long f, long t)
Andy Whitcroft96822902008-07-23 21:27:29 -0700195{
Joonsoo Kim1406ec92014-04-03 14:47:26 -0700196 struct list_head *head = &resv->regions;
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700197 struct file_region *rg, *nrg = NULL;
Andy Whitcroft96822902008-07-23 21:27:29 -0700198 long chg = 0;
199
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700200retry:
201 spin_lock(&resv->lock);
Andy Whitcroft96822902008-07-23 21:27:29 -0700202 /* Locate the region we are before or in. */
203 list_for_each_entry(rg, head, link)
204 if (f <= rg->to)
205 break;
206
207 /* If we are below the current region then a new region is required.
208 * Subtle, allocate a new region at the position but make it zero
209 * size such that we can guarantee to record the reservation. */
210 if (&rg->link == head || t < rg->from) {
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700211 if (!nrg) {
212 spin_unlock(&resv->lock);
213 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
214 if (!nrg)
215 return -ENOMEM;
Andy Whitcroft96822902008-07-23 21:27:29 -0700216
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700217 nrg->from = f;
218 nrg->to = f;
219 INIT_LIST_HEAD(&nrg->link);
220 goto retry;
221 }
222
223 list_add(&nrg->link, rg->link.prev);
224 chg = t - f;
225 goto out_nrg;
Andy Whitcroft96822902008-07-23 21:27:29 -0700226 }
227
228 /* Round our left edge to the current segment if it encloses us. */
229 if (f > rg->from)
230 f = rg->from;
231 chg = t - f;
232
233 /* Check for and consume any regions we now overlap with. */
234 list_for_each_entry(rg, rg->link.prev, link) {
235 if (&rg->link == head)
236 break;
237 if (rg->from > t)
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700238 goto out;
Andy Whitcroft96822902008-07-23 21:27:29 -0700239
Lucas De Marchi25985ed2011-03-30 22:57:33 -0300240 /* We overlap with this area, if it extends further than
Andy Whitcroft96822902008-07-23 21:27:29 -0700241 * us then we must extend ourselves. Account for its
242 * existing reservation. */
243 if (rg->to > t) {
244 chg += rg->to - t;
245 t = rg->to;
246 }
247 chg -= rg->to - rg->from;
248 }
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700249
250out:
251 spin_unlock(&resv->lock);
252 /* We already know we raced and no longer need the new region */
253 kfree(nrg);
254 return chg;
255out_nrg:
256 spin_unlock(&resv->lock);
Andy Whitcroft96822902008-07-23 21:27:29 -0700257 return chg;
258}
259
Joonsoo Kim1406ec92014-04-03 14:47:26 -0700260static long region_truncate(struct resv_map *resv, long end)
Andy Whitcroft96822902008-07-23 21:27:29 -0700261{
Joonsoo Kim1406ec92014-04-03 14:47:26 -0700262 struct list_head *head = &resv->regions;
Andy Whitcroft96822902008-07-23 21:27:29 -0700263 struct file_region *rg, *trg;
264 long chg = 0;
265
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700266 spin_lock(&resv->lock);
Andy Whitcroft96822902008-07-23 21:27:29 -0700267 /* Locate the region we are either in or before. */
268 list_for_each_entry(rg, head, link)
269 if (end <= rg->to)
270 break;
271 if (&rg->link == head)
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700272 goto out;
Andy Whitcroft96822902008-07-23 21:27:29 -0700273
274 /* If we are in the middle of a region then adjust it. */
275 if (end > rg->from) {
276 chg = rg->to - end;
277 rg->to = end;
278 rg = list_entry(rg->link.next, typeof(*rg), link);
279 }
280
281 /* Drop any remaining regions. */
282 list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
283 if (&rg->link == head)
284 break;
285 chg += rg->to - rg->from;
286 list_del(&rg->link);
287 kfree(rg);
288 }
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700289
290out:
291 spin_unlock(&resv->lock);
Andy Whitcroft96822902008-07-23 21:27:29 -0700292 return chg;
293}
294
Joonsoo Kim1406ec92014-04-03 14:47:26 -0700295static long region_count(struct resv_map *resv, long f, long t)
Andy Whitcroft84afd992008-07-23 21:27:32 -0700296{
Joonsoo Kim1406ec92014-04-03 14:47:26 -0700297 struct list_head *head = &resv->regions;
Andy Whitcroft84afd992008-07-23 21:27:32 -0700298 struct file_region *rg;
299 long chg = 0;
300
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700301 spin_lock(&resv->lock);
Andy Whitcroft84afd992008-07-23 21:27:32 -0700302 /* Locate each segment we overlap with, and count that overlap. */
303 list_for_each_entry(rg, head, link) {
Wang Sheng-Huif2135a42012-05-29 15:06:17 -0700304 long seg_from;
305 long seg_to;
Andy Whitcroft84afd992008-07-23 21:27:32 -0700306
307 if (rg->to <= f)
308 continue;
309 if (rg->from >= t)
310 break;
311
312 seg_from = max(rg->from, f);
313 seg_to = min(rg->to, t);
314
315 chg += seg_to - seg_from;
316 }
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700317 spin_unlock(&resv->lock);
Andy Whitcroft84afd992008-07-23 21:27:32 -0700318
319 return chg;
320}
321
Andy Whitcroft96822902008-07-23 21:27:29 -0700322/*
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700323 * Convert the address within this vma to the page offset within
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700324 * the mapping, in pagecache page units; huge pages here.
325 */
Andi Kleena5516432008-07-23 21:27:41 -0700326static pgoff_t vma_hugecache_offset(struct hstate *h,
327 struct vm_area_struct *vma, unsigned long address)
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700328{
Andi Kleena5516432008-07-23 21:27:41 -0700329 return ((address - vma->vm_start) >> huge_page_shift(h)) +
330 (vma->vm_pgoff >> huge_page_order(h));
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700331}
332
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +0900333pgoff_t linear_hugepage_index(struct vm_area_struct *vma,
334 unsigned long address)
335{
336 return vma_hugecache_offset(hstate_vma(vma), vma, address);
337}
338
Andy Whitcroft84afd992008-07-23 21:27:32 -0700339/*
Mel Gorman08fba692009-01-06 14:38:53 -0800340 * Return the size of the pages allocated when backing a VMA. In the majority
341 * cases this will be same size as used by the page table entries.
342 */
343unsigned long vma_kernel_pagesize(struct vm_area_struct *vma)
344{
345 struct hstate *hstate;
346
347 if (!is_vm_hugetlb_page(vma))
348 return PAGE_SIZE;
349
350 hstate = hstate_vma(vma);
351
Wanpeng Li2415cf12013-07-03 15:02:43 -0700352 return 1UL << huge_page_shift(hstate);
Mel Gorman08fba692009-01-06 14:38:53 -0800353}
Joerg Roedelf340ca02009-06-19 15:16:22 +0200354EXPORT_SYMBOL_GPL(vma_kernel_pagesize);
Mel Gorman08fba692009-01-06 14:38:53 -0800355
356/*
Mel Gorman33402892009-01-06 14:38:54 -0800357 * Return the page size being used by the MMU to back a VMA. In the majority
358 * of cases, the page size used by the kernel matches the MMU size. On
359 * architectures where it differs, an architecture-specific version of this
360 * function is required.
361 */
362#ifndef vma_mmu_pagesize
363unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
364{
365 return vma_kernel_pagesize(vma);
366}
367#endif
368
369/*
Andy Whitcroft84afd992008-07-23 21:27:32 -0700370 * Flags for MAP_PRIVATE reservations. These are stored in the bottom
371 * bits of the reservation map pointer, which are always clear due to
372 * alignment.
373 */
374#define HPAGE_RESV_OWNER (1UL << 0)
375#define HPAGE_RESV_UNMAPPED (1UL << 1)
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700376#define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
Andy Whitcroft84afd992008-07-23 21:27:32 -0700377
Mel Gormana1e78772008-07-23 21:27:23 -0700378/*
379 * These helpers are used to track how many pages are reserved for
380 * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
381 * is guaranteed to have their future faults succeed.
382 *
383 * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
384 * the reserve counters are updated with the hugetlb_lock held. It is safe
385 * to reset the VMA at fork() time as it is not in use yet and there is no
386 * chance of the global counters getting corrupted as a result of the values.
Andy Whitcroft84afd992008-07-23 21:27:32 -0700387 *
388 * The private mapping reservation is represented in a subtly different
389 * manner to a shared mapping. A shared mapping has a region map associated
390 * with the underlying file, this region map represents the backing file
391 * pages which have ever had a reservation assigned which this persists even
392 * after the page is instantiated. A private mapping has a region map
393 * associated with the original mmap which is attached to all VMAs which
394 * reference it, this region map represents those offsets which have consumed
395 * reservation ie. where pages have been instantiated.
Mel Gormana1e78772008-07-23 21:27:23 -0700396 */
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700397static unsigned long get_vma_private_data(struct vm_area_struct *vma)
398{
399 return (unsigned long)vma->vm_private_data;
400}
401
402static void set_vma_private_data(struct vm_area_struct *vma,
403 unsigned long value)
404{
405 vma->vm_private_data = (void *)value;
406}
407
Joonsoo Kim9119a412014-04-03 14:47:25 -0700408struct resv_map *resv_map_alloc(void)
Andy Whitcroft84afd992008-07-23 21:27:32 -0700409{
410 struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
411 if (!resv_map)
412 return NULL;
413
414 kref_init(&resv_map->refs);
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700415 spin_lock_init(&resv_map->lock);
Andy Whitcroft84afd992008-07-23 21:27:32 -0700416 INIT_LIST_HEAD(&resv_map->regions);
417
418 return resv_map;
419}
420
Joonsoo Kim9119a412014-04-03 14:47:25 -0700421void resv_map_release(struct kref *ref)
Andy Whitcroft84afd992008-07-23 21:27:32 -0700422{
423 struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
424
425 /* Clear out any active regions before we release the map. */
Joonsoo Kim1406ec92014-04-03 14:47:26 -0700426 region_truncate(resv_map, 0);
Andy Whitcroft84afd992008-07-23 21:27:32 -0700427 kfree(resv_map);
428}
429
Joonsoo Kim4e35f482014-04-03 14:47:30 -0700430static inline struct resv_map *inode_resv_map(struct inode *inode)
431{
432 return inode->i_mapping->private_data;
433}
434
Andy Whitcroft84afd992008-07-23 21:27:32 -0700435static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
Mel Gormana1e78772008-07-23 21:27:23 -0700436{
Sasha Levin81d1b092014-10-09 15:28:10 -0700437 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
Joonsoo Kim4e35f482014-04-03 14:47:30 -0700438 if (vma->vm_flags & VM_MAYSHARE) {
439 struct address_space *mapping = vma->vm_file->f_mapping;
440 struct inode *inode = mapping->host;
441
442 return inode_resv_map(inode);
443
444 } else {
Andy Whitcroft84afd992008-07-23 21:27:32 -0700445 return (struct resv_map *)(get_vma_private_data(vma) &
446 ~HPAGE_RESV_MASK);
Joonsoo Kim4e35f482014-04-03 14:47:30 -0700447 }
Mel Gormana1e78772008-07-23 21:27:23 -0700448}
449
Andy Whitcroft84afd992008-07-23 21:27:32 -0700450static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
Mel Gormana1e78772008-07-23 21:27:23 -0700451{
Sasha Levin81d1b092014-10-09 15:28:10 -0700452 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
453 VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
Mel Gormana1e78772008-07-23 21:27:23 -0700454
Andy Whitcroft84afd992008-07-23 21:27:32 -0700455 set_vma_private_data(vma, (get_vma_private_data(vma) &
456 HPAGE_RESV_MASK) | (unsigned long)map);
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700457}
458
459static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
460{
Sasha Levin81d1b092014-10-09 15:28:10 -0700461 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
462 VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700463
464 set_vma_private_data(vma, get_vma_private_data(vma) | flags);
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700465}
466
467static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
468{
Sasha Levin81d1b092014-10-09 15:28:10 -0700469 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700470
471 return (get_vma_private_data(vma) & flag) != 0;
Mel Gormana1e78772008-07-23 21:27:23 -0700472}
473
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700474/* Reset counters to 0 and clear all HPAGE_RESV_* flags */
Mel Gormana1e78772008-07-23 21:27:23 -0700475void reset_vma_resv_huge_pages(struct vm_area_struct *vma)
476{
Sasha Levin81d1b092014-10-09 15:28:10 -0700477 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
Mel Gormanf83a2752009-05-28 14:34:40 -0700478 if (!(vma->vm_flags & VM_MAYSHARE))
Mel Gormana1e78772008-07-23 21:27:23 -0700479 vma->vm_private_data = (void *)0;
480}
481
482/* Returns true if the VMA has associated reserve pages */
Joonsoo Kimaf0ed732013-09-11 14:21:18 -0700483static int vma_has_reserves(struct vm_area_struct *vma, long chg)
Mel Gormana1e78772008-07-23 21:27:23 -0700484{
Joonsoo Kimaf0ed732013-09-11 14:21:18 -0700485 if (vma->vm_flags & VM_NORESERVE) {
486 /*
487 * This address is already reserved by other process(chg == 0),
488 * so, we should decrement reserved count. Without decrementing,
489 * reserve count remains after releasing inode, because this
490 * allocated page will go into page cache and is regarded as
491 * coming from reserved pool in releasing step. Currently, we
492 * don't have any other solution to deal with this situation
493 * properly, so add work-around here.
494 */
495 if (vma->vm_flags & VM_MAYSHARE && chg == 0)
496 return 1;
497 else
498 return 0;
499 }
Joonsoo Kima63884e2013-09-11 14:21:07 -0700500
501 /* Shared mappings always use reserves */
Mel Gormanf83a2752009-05-28 14:34:40 -0700502 if (vma->vm_flags & VM_MAYSHARE)
Mel Gorman7f09ca52008-07-23 21:27:58 -0700503 return 1;
Joonsoo Kima63884e2013-09-11 14:21:07 -0700504
505 /*
506 * Only the process that called mmap() has reserves for
507 * private mappings.
508 */
Mel Gorman7f09ca52008-07-23 21:27:58 -0700509 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER))
510 return 1;
Joonsoo Kima63884e2013-09-11 14:21:07 -0700511
Mel Gorman7f09ca52008-07-23 21:27:58 -0700512 return 0;
Mel Gormana1e78772008-07-23 21:27:23 -0700513}
514
Andi Kleena5516432008-07-23 21:27:41 -0700515static void enqueue_huge_page(struct hstate *h, struct page *page)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700516{
517 int nid = page_to_nid(page);
Aneesh Kumar K.V0edaecf2012-07-31 16:42:07 -0700518 list_move(&page->lru, &h->hugepage_freelists[nid]);
Andi Kleena5516432008-07-23 21:27:41 -0700519 h->free_huge_pages++;
520 h->free_huge_pages_node[nid]++;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700521}
522
Naoya Horiguchibf50bab2010-09-08 10:19:33 +0900523static struct page *dequeue_huge_page_node(struct hstate *h, int nid)
524{
525 struct page *page;
526
Naoya Horiguchic8721bb2013-09-11 14:22:09 -0700527 list_for_each_entry(page, &h->hugepage_freelists[nid], lru)
528 if (!is_migrate_isolate_page(page))
529 break;
530 /*
531 * if 'non-isolated free hugepage' not found on the list,
532 * the allocation fails.
533 */
534 if (&h->hugepage_freelists[nid] == &page->lru)
Naoya Horiguchibf50bab2010-09-08 10:19:33 +0900535 return NULL;
Aneesh Kumar K.V0edaecf2012-07-31 16:42:07 -0700536 list_move(&page->lru, &h->hugepage_activelist);
Naoya Horiguchia9869b82010-09-08 10:19:37 +0900537 set_page_refcounted(page);
Naoya Horiguchibf50bab2010-09-08 10:19:33 +0900538 h->free_huge_pages--;
539 h->free_huge_pages_node[nid]--;
540 return page;
541}
542
Naoya Horiguchi86cdb462013-09-11 14:22:13 -0700543/* Movability of hugepages depends on migration support. */
544static inline gfp_t htlb_alloc_mask(struct hstate *h)
545{
Naoya Horiguchi100873d2014-06-04 16:10:56 -0700546 if (hugepages_treat_as_movable || hugepage_migration_supported(h))
Naoya Horiguchi86cdb462013-09-11 14:22:13 -0700547 return GFP_HIGHUSER_MOVABLE;
548 else
549 return GFP_HIGHUSER;
550}
551
Andi Kleena5516432008-07-23 21:27:41 -0700552static struct page *dequeue_huge_page_vma(struct hstate *h,
553 struct vm_area_struct *vma,
Joonsoo Kimaf0ed732013-09-11 14:21:18 -0700554 unsigned long address, int avoid_reserve,
555 long chg)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700556{
Konstantin Khlebnikovb1c12cb2012-04-25 16:01:46 -0700557 struct page *page = NULL;
Lee Schermerhorn480eccf2007-09-18 22:46:47 -0700558 struct mempolicy *mpol;
Mel Gorman19770b32008-04-28 02:12:18 -0700559 nodemask_t *nodemask;
Miao Xiec0ff7452010-05-24 14:32:08 -0700560 struct zonelist *zonelist;
Mel Gormandd1a2392008-04-28 02:12:17 -0700561 struct zone *zone;
562 struct zoneref *z;
Mel Gormancc9a6c82012-03-21 16:34:11 -0700563 unsigned int cpuset_mems_cookie;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700564
Mel Gormana1e78772008-07-23 21:27:23 -0700565 /*
566 * A child process with MAP_PRIVATE mappings created by their parent
567 * have no page reserves. This check ensures that reservations are
568 * not "stolen". The child may still get SIGKILLed
569 */
Joonsoo Kimaf0ed732013-09-11 14:21:18 -0700570 if (!vma_has_reserves(vma, chg) &&
Andi Kleena5516432008-07-23 21:27:41 -0700571 h->free_huge_pages - h->resv_huge_pages == 0)
Miao Xiec0ff7452010-05-24 14:32:08 -0700572 goto err;
Mel Gormana1e78772008-07-23 21:27:23 -0700573
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700574 /* If reserves cannot be used, ensure enough pages are in the pool */
Andi Kleena5516432008-07-23 21:27:41 -0700575 if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
Justin P. Mattock6eab04a2011-04-08 19:49:08 -0700576 goto err;
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700577
Joonsoo Kim9966c4b2013-09-11 14:20:50 -0700578retry_cpuset:
Mel Gormand26914d2014-04-03 14:47:24 -0700579 cpuset_mems_cookie = read_mems_allowed_begin();
Joonsoo Kim9966c4b2013-09-11 14:20:50 -0700580 zonelist = huge_zonelist(vma, address,
Naoya Horiguchi86cdb462013-09-11 14:22:13 -0700581 htlb_alloc_mask(h), &mpol, &nodemask);
Joonsoo Kim9966c4b2013-09-11 14:20:50 -0700582
Mel Gorman19770b32008-04-28 02:12:18 -0700583 for_each_zone_zonelist_nodemask(zone, z, zonelist,
584 MAX_NR_ZONES - 1, nodemask) {
Vladimir Davydov344736f2014-10-20 15:50:30 +0400585 if (cpuset_zone_allowed(zone, htlb_alloc_mask(h))) {
Naoya Horiguchibf50bab2010-09-08 10:19:33 +0900586 page = dequeue_huge_page_node(h, zone_to_nid(zone));
587 if (page) {
Joonsoo Kimaf0ed732013-09-11 14:21:18 -0700588 if (avoid_reserve)
589 break;
590 if (!vma_has_reserves(vma, chg))
591 break;
592
Joonsoo Kim07443a82013-09-11 14:21:58 -0700593 SetPagePrivate(page);
Joonsoo Kimaf0ed732013-09-11 14:21:18 -0700594 h->resv_huge_pages--;
Naoya Horiguchibf50bab2010-09-08 10:19:33 +0900595 break;
596 }
Andrew Morton3abf7af2007-07-19 01:49:08 -0700597 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700598 }
Mel Gormancc9a6c82012-03-21 16:34:11 -0700599
600 mpol_cond_put(mpol);
Mel Gormand26914d2014-04-03 14:47:24 -0700601 if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie)))
Mel Gormancc9a6c82012-03-21 16:34:11 -0700602 goto retry_cpuset;
603 return page;
604
Miao Xiec0ff7452010-05-24 14:32:08 -0700605err:
Mel Gormancc9a6c82012-03-21 16:34:11 -0700606 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700607}
608
Luiz Capitulino1cac6f22014-06-04 16:07:11 -0700609/*
610 * common helper functions for hstate_next_node_to_{alloc|free}.
611 * We may have allocated or freed a huge page based on a different
612 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
613 * be outside of *nodes_allowed. Ensure that we use an allowed
614 * node for alloc or free.
615 */
616static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
617{
618 nid = next_node(nid, *nodes_allowed);
619 if (nid == MAX_NUMNODES)
620 nid = first_node(*nodes_allowed);
621 VM_BUG_ON(nid >= MAX_NUMNODES);
622
623 return nid;
624}
625
626static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
627{
628 if (!node_isset(nid, *nodes_allowed))
629 nid = next_node_allowed(nid, nodes_allowed);
630 return nid;
631}
632
633/*
634 * returns the previously saved node ["this node"] from which to
635 * allocate a persistent huge page for the pool and advance the
636 * next node from which to allocate, handling wrap at end of node
637 * mask.
638 */
639static int hstate_next_node_to_alloc(struct hstate *h,
640 nodemask_t *nodes_allowed)
641{
642 int nid;
643
644 VM_BUG_ON(!nodes_allowed);
645
646 nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
647 h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
648
649 return nid;
650}
651
652/*
653 * helper for free_pool_huge_page() - return the previously saved
654 * node ["this node"] from which to free a huge page. Advance the
655 * next node id whether or not we find a free huge page to free so
656 * that the next attempt to free addresses the next node.
657 */
658static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
659{
660 int nid;
661
662 VM_BUG_ON(!nodes_allowed);
663
664 nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
665 h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
666
667 return nid;
668}
669
670#define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \
671 for (nr_nodes = nodes_weight(*mask); \
672 nr_nodes > 0 && \
673 ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \
674 nr_nodes--)
675
676#define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \
677 for (nr_nodes = nodes_weight(*mask); \
678 nr_nodes > 0 && \
679 ((node = hstate_next_node_to_free(hs, mask)) || 1); \
680 nr_nodes--)
681
Luiz Capitulino944d9fe2014-06-04 16:07:13 -0700682#if defined(CONFIG_CMA) && defined(CONFIG_X86_64)
683static void destroy_compound_gigantic_page(struct page *page,
684 unsigned long order)
685{
686 int i;
687 int nr_pages = 1 << order;
688 struct page *p = page + 1;
689
690 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
691 __ClearPageTail(p);
692 set_page_refcounted(p);
693 p->first_page = NULL;
694 }
695
696 set_compound_order(page, 0);
697 __ClearPageHead(page);
698}
699
700static void free_gigantic_page(struct page *page, unsigned order)
701{
702 free_contig_range(page_to_pfn(page), 1 << order);
703}
704
705static int __alloc_gigantic_page(unsigned long start_pfn,
706 unsigned long nr_pages)
707{
708 unsigned long end_pfn = start_pfn + nr_pages;
709 return alloc_contig_range(start_pfn, end_pfn, MIGRATE_MOVABLE);
710}
711
712static bool pfn_range_valid_gigantic(unsigned long start_pfn,
713 unsigned long nr_pages)
714{
715 unsigned long i, end_pfn = start_pfn + nr_pages;
716 struct page *page;
717
718 for (i = start_pfn; i < end_pfn; i++) {
719 if (!pfn_valid(i))
720 return false;
721
722 page = pfn_to_page(i);
723
724 if (PageReserved(page))
725 return false;
726
727 if (page_count(page) > 0)
728 return false;
729
730 if (PageHuge(page))
731 return false;
732 }
733
734 return true;
735}
736
737static bool zone_spans_last_pfn(const struct zone *zone,
738 unsigned long start_pfn, unsigned long nr_pages)
739{
740 unsigned long last_pfn = start_pfn + nr_pages - 1;
741 return zone_spans_pfn(zone, last_pfn);
742}
743
744static struct page *alloc_gigantic_page(int nid, unsigned order)
745{
746 unsigned long nr_pages = 1 << order;
747 unsigned long ret, pfn, flags;
748 struct zone *z;
749
750 z = NODE_DATA(nid)->node_zones;
751 for (; z - NODE_DATA(nid)->node_zones < MAX_NR_ZONES; z++) {
752 spin_lock_irqsave(&z->lock, flags);
753
754 pfn = ALIGN(z->zone_start_pfn, nr_pages);
755 while (zone_spans_last_pfn(z, pfn, nr_pages)) {
756 if (pfn_range_valid_gigantic(pfn, nr_pages)) {
757 /*
758 * We release the zone lock here because
759 * alloc_contig_range() will also lock the zone
760 * at some point. If there's an allocation
761 * spinning on this lock, it may win the race
762 * and cause alloc_contig_range() to fail...
763 */
764 spin_unlock_irqrestore(&z->lock, flags);
765 ret = __alloc_gigantic_page(pfn, nr_pages);
766 if (!ret)
767 return pfn_to_page(pfn);
768 spin_lock_irqsave(&z->lock, flags);
769 }
770 pfn += nr_pages;
771 }
772
773 spin_unlock_irqrestore(&z->lock, flags);
774 }
775
776 return NULL;
777}
778
779static void prep_new_huge_page(struct hstate *h, struct page *page, int nid);
780static void prep_compound_gigantic_page(struct page *page, unsigned long order);
781
782static struct page *alloc_fresh_gigantic_page_node(struct hstate *h, int nid)
783{
784 struct page *page;
785
786 page = alloc_gigantic_page(nid, huge_page_order(h));
787 if (page) {
788 prep_compound_gigantic_page(page, huge_page_order(h));
789 prep_new_huge_page(h, page, nid);
790 }
791
792 return page;
793}
794
795static int alloc_fresh_gigantic_page(struct hstate *h,
796 nodemask_t *nodes_allowed)
797{
798 struct page *page = NULL;
799 int nr_nodes, node;
800
801 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
802 page = alloc_fresh_gigantic_page_node(h, node);
803 if (page)
804 return 1;
805 }
806
807 return 0;
808}
809
810static inline bool gigantic_page_supported(void) { return true; }
811#else
812static inline bool gigantic_page_supported(void) { return false; }
813static inline void free_gigantic_page(struct page *page, unsigned order) { }
814static inline void destroy_compound_gigantic_page(struct page *page,
815 unsigned long order) { }
816static inline int alloc_fresh_gigantic_page(struct hstate *h,
817 nodemask_t *nodes_allowed) { return 0; }
818#endif
819
Andi Kleena5516432008-07-23 21:27:41 -0700820static void update_and_free_page(struct hstate *h, struct page *page)
Adam Litke6af2acb2007-10-16 01:26:16 -0700821{
822 int i;
Andi Kleena5516432008-07-23 21:27:41 -0700823
Luiz Capitulino944d9fe2014-06-04 16:07:13 -0700824 if (hstate_is_gigantic(h) && !gigantic_page_supported())
825 return;
Andy Whitcroft18229df2008-11-06 12:53:27 -0800826
Andi Kleena5516432008-07-23 21:27:41 -0700827 h->nr_huge_pages--;
828 h->nr_huge_pages_node[page_to_nid(page)]--;
829 for (i = 0; i < pages_per_huge_page(h); i++) {
Chris Forbes32f84522011-07-25 17:12:14 -0700830 page[i].flags &= ~(1 << PG_locked | 1 << PG_error |
831 1 << PG_referenced | 1 << PG_dirty |
Luiz Capitulinoa7407a22014-06-04 16:07:09 -0700832 1 << PG_active | 1 << PG_private |
833 1 << PG_writeback);
Adam Litke6af2acb2007-10-16 01:26:16 -0700834 }
Sasha Levin309381fea2014-01-23 15:52:54 -0800835 VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page);
Adam Litke6af2acb2007-10-16 01:26:16 -0700836 set_compound_page_dtor(page, NULL);
837 set_page_refcounted(page);
Luiz Capitulino944d9fe2014-06-04 16:07:13 -0700838 if (hstate_is_gigantic(h)) {
839 destroy_compound_gigantic_page(page, huge_page_order(h));
840 free_gigantic_page(page, huge_page_order(h));
841 } else {
842 arch_release_hugepage(page);
843 __free_pages(page, huge_page_order(h));
844 }
Adam Litke6af2acb2007-10-16 01:26:16 -0700845}
846
Andi Kleene5ff2152008-07-23 21:27:42 -0700847struct hstate *size_to_hstate(unsigned long size)
848{
849 struct hstate *h;
850
851 for_each_hstate(h) {
852 if (huge_page_size(h) == size)
853 return h;
854 }
855 return NULL;
856}
857
Atsushi Kumagai8f1d26d2014-07-30 16:08:39 -0700858void free_huge_page(struct page *page)
David Gibson27a85ef2006-03-22 00:08:56 -0800859{
Andi Kleena5516432008-07-23 21:27:41 -0700860 /*
861 * Can't pass hstate in here because it is called from the
862 * compound page destructor.
863 */
Andi Kleene5ff2152008-07-23 21:27:42 -0700864 struct hstate *h = page_hstate(page);
Adam Litke7893d1d2007-10-16 01:26:18 -0700865 int nid = page_to_nid(page);
David Gibson90481622012-03-21 16:34:12 -0700866 struct hugepage_subpool *spool =
867 (struct hugepage_subpool *)page_private(page);
Joonsoo Kim07443a82013-09-11 14:21:58 -0700868 bool restore_reserve;
David Gibson27a85ef2006-03-22 00:08:56 -0800869
Andy Whitcrofte5df70a2008-02-23 15:23:32 -0800870 set_page_private(page, 0);
Mel Gorman23be7462010-04-23 13:17:56 -0400871 page->mapping = NULL;
Adam Litke7893d1d2007-10-16 01:26:18 -0700872 BUG_ON(page_count(page));
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +0900873 BUG_ON(page_mapcount(page));
Joonsoo Kim07443a82013-09-11 14:21:58 -0700874 restore_reserve = PagePrivate(page);
Joonsoo Kim16c794b2013-10-16 13:46:48 -0700875 ClearPagePrivate(page);
David Gibson27a85ef2006-03-22 00:08:56 -0800876
877 spin_lock(&hugetlb_lock);
Aneesh Kumar K.V6d76dcf2012-07-31 16:42:18 -0700878 hugetlb_cgroup_uncharge_page(hstate_index(h),
879 pages_per_huge_page(h), page);
Joonsoo Kim07443a82013-09-11 14:21:58 -0700880 if (restore_reserve)
881 h->resv_huge_pages++;
882
Luiz Capitulino944d9fe2014-06-04 16:07:13 -0700883 if (h->surplus_huge_pages_node[nid]) {
Aneesh Kumar K.V0edaecf2012-07-31 16:42:07 -0700884 /* remove the page from active list */
885 list_del(&page->lru);
Andi Kleena5516432008-07-23 21:27:41 -0700886 update_and_free_page(h, page);
887 h->surplus_huge_pages--;
888 h->surplus_huge_pages_node[nid]--;
Adam Litke7893d1d2007-10-16 01:26:18 -0700889 } else {
Will Deacon5d3a5512012-10-08 16:29:32 -0700890 arch_clear_hugepage_flags(page);
Andi Kleena5516432008-07-23 21:27:41 -0700891 enqueue_huge_page(h, page);
Adam Litke7893d1d2007-10-16 01:26:18 -0700892 }
David Gibson27a85ef2006-03-22 00:08:56 -0800893 spin_unlock(&hugetlb_lock);
David Gibson90481622012-03-21 16:34:12 -0700894 hugepage_subpool_put_pages(spool, 1);
David Gibson27a85ef2006-03-22 00:08:56 -0800895}
896
Andi Kleena5516432008-07-23 21:27:41 -0700897static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
Andi Kleenb7ba30c2008-07-23 21:27:40 -0700898{
Aneesh Kumar K.V0edaecf2012-07-31 16:42:07 -0700899 INIT_LIST_HEAD(&page->lru);
Andi Kleenb7ba30c2008-07-23 21:27:40 -0700900 set_compound_page_dtor(page, free_huge_page);
901 spin_lock(&hugetlb_lock);
Aneesh Kumar K.V9dd540e2012-07-31 16:42:15 -0700902 set_hugetlb_cgroup(page, NULL);
Andi Kleena5516432008-07-23 21:27:41 -0700903 h->nr_huge_pages++;
904 h->nr_huge_pages_node[nid]++;
Andi Kleenb7ba30c2008-07-23 21:27:40 -0700905 spin_unlock(&hugetlb_lock);
906 put_page(page); /* free it into the hugepage allocator */
907}
908
Luiz Capitulino2906dd52014-06-04 16:07:06 -0700909static void prep_compound_gigantic_page(struct page *page, unsigned long order)
Wu Fengguang20a03072009-06-16 15:32:22 -0700910{
911 int i;
912 int nr_pages = 1 << order;
913 struct page *p = page + 1;
914
915 /* we rely on prep_new_huge_page to set the destructor */
916 set_compound_order(page, order);
917 __SetPageHead(page);
Andrea Arcangelief5a22b2013-10-16 13:46:56 -0700918 __ClearPageReserved(page);
Wu Fengguang20a03072009-06-16 15:32:22 -0700919 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
920 __SetPageTail(p);
Andrea Arcangelief5a22b2013-10-16 13:46:56 -0700921 /*
922 * For gigantic hugepages allocated through bootmem at
923 * boot, it's safer to be consistent with the not-gigantic
924 * hugepages and clear the PG_reserved bit from all tail pages
925 * too. Otherwse drivers using get_user_pages() to access tail
926 * pages may get the reference counting wrong if they see
927 * PG_reserved set on a tail page (despite the head page not
928 * having PG_reserved set). Enforcing this consistency between
929 * head and tail pages allows drivers to optimize away a check
930 * on the head page when they need know if put_page() is needed
931 * after get_user_pages().
932 */
933 __ClearPageReserved(p);
Youquan Song58a84aa2011-12-08 14:34:18 -0800934 set_page_count(p, 0);
Wu Fengguang20a03072009-06-16 15:32:22 -0700935 p->first_page = page;
936 }
937}
938
Andrew Morton77959122012-10-08 16:34:11 -0700939/*
940 * PageHuge() only returns true for hugetlbfs pages, but not for normal or
941 * transparent huge pages. See the PageTransHuge() documentation for more
942 * details.
943 */
Wu Fengguang20a03072009-06-16 15:32:22 -0700944int PageHuge(struct page *page)
945{
Wu Fengguang20a03072009-06-16 15:32:22 -0700946 if (!PageCompound(page))
947 return 0;
948
949 page = compound_head(page);
Andrew Morton758f66a2014-01-21 15:48:57 -0800950 return get_compound_page_dtor(page) == free_huge_page;
Wu Fengguang20a03072009-06-16 15:32:22 -0700951}
Naoya Horiguchi43131e12010-05-28 09:29:22 +0900952EXPORT_SYMBOL_GPL(PageHuge);
953
Andrea Arcangeli27c73ae2013-11-21 14:32:02 -0800954/*
955 * PageHeadHuge() only returns true for hugetlbfs head page, but not for
956 * normal or transparent huge pages.
957 */
958int PageHeadHuge(struct page *page_head)
959{
Andrea Arcangeli27c73ae2013-11-21 14:32:02 -0800960 if (!PageHead(page_head))
961 return 0;
962
Andrew Morton758f66a2014-01-21 15:48:57 -0800963 return get_compound_page_dtor(page_head) == free_huge_page;
Andrea Arcangeli27c73ae2013-11-21 14:32:02 -0800964}
Andrea Arcangeli27c73ae2013-11-21 14:32:02 -0800965
Zhang Yi13d60f42013-06-25 21:19:31 +0800966pgoff_t __basepage_index(struct page *page)
967{
968 struct page *page_head = compound_head(page);
969 pgoff_t index = page_index(page_head);
970 unsigned long compound_idx;
971
972 if (!PageHuge(page_head))
973 return page_index(page);
974
975 if (compound_order(page_head) >= MAX_ORDER)
976 compound_idx = page_to_pfn(page) - page_to_pfn(page_head);
977 else
978 compound_idx = page - page_head;
979
980 return (index << compound_order(page_head)) + compound_idx;
981}
982
Andi Kleena5516432008-07-23 21:27:41 -0700983static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700984{
Linus Torvalds1da177e2005-04-16 15:20:36 -0700985 struct page *page;
Joe Jinf96efd52007-07-15 23:38:12 -0700986
Mel Gorman6484eb32009-06-16 15:31:54 -0700987 page = alloc_pages_exact_node(nid,
Naoya Horiguchi86cdb462013-09-11 14:22:13 -0700988 htlb_alloc_mask(h)|__GFP_COMP|__GFP_THISNODE|
Nishanth Aravamudan551883a2008-04-29 00:58:26 -0700989 __GFP_REPEAT|__GFP_NOWARN,
Andi Kleena5516432008-07-23 21:27:41 -0700990 huge_page_order(h));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700991 if (page) {
Gerald Schaefer7f2e9522008-04-28 02:13:29 -0700992 if (arch_prepare_hugepage(page)) {
Gerald Schaefercaff3a22008-08-12 15:08:38 -0700993 __free_pages(page, huge_page_order(h));
Harvey Harrison7b8ee842008-04-28 14:13:19 -0700994 return NULL;
Gerald Schaefer7f2e9522008-04-28 02:13:29 -0700995 }
Andi Kleena5516432008-07-23 21:27:41 -0700996 prep_new_huge_page(h, page, nid);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700997 }
Nishanth Aravamudan63b46132007-10-16 01:26:24 -0700998
999 return page;
1000}
1001
Joonsoo Kimb2261022013-09-11 14:21:00 -07001002static int alloc_fresh_huge_page(struct hstate *h, nodemask_t *nodes_allowed)
1003{
1004 struct page *page;
1005 int nr_nodes, node;
1006 int ret = 0;
1007
1008 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
1009 page = alloc_fresh_huge_page_node(h, node);
1010 if (page) {
1011 ret = 1;
1012 break;
1013 }
1014 }
1015
1016 if (ret)
1017 count_vm_event(HTLB_BUDDY_PGALLOC);
1018 else
1019 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
1020
1021 return ret;
1022}
1023
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001024/*
1025 * Free huge page from pool from next node to free.
1026 * Attempt to keep persistent huge pages more or less
1027 * balanced over allowed nodes.
1028 * Called with hugetlb_lock locked.
1029 */
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001030static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
1031 bool acct_surplus)
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001032{
Joonsoo Kimb2261022013-09-11 14:21:00 -07001033 int nr_nodes, node;
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001034 int ret = 0;
1035
Joonsoo Kimb2261022013-09-11 14:21:00 -07001036 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
Lee Schermerhorn685f3452009-09-21 17:01:23 -07001037 /*
1038 * If we're returning unused surplus pages, only examine
1039 * nodes with surplus pages.
1040 */
Joonsoo Kimb2261022013-09-11 14:21:00 -07001041 if ((!acct_surplus || h->surplus_huge_pages_node[node]) &&
1042 !list_empty(&h->hugepage_freelists[node])) {
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001043 struct page *page =
Joonsoo Kimb2261022013-09-11 14:21:00 -07001044 list_entry(h->hugepage_freelists[node].next,
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001045 struct page, lru);
1046 list_del(&page->lru);
1047 h->free_huge_pages--;
Joonsoo Kimb2261022013-09-11 14:21:00 -07001048 h->free_huge_pages_node[node]--;
Lee Schermerhorn685f3452009-09-21 17:01:23 -07001049 if (acct_surplus) {
1050 h->surplus_huge_pages--;
Joonsoo Kimb2261022013-09-11 14:21:00 -07001051 h->surplus_huge_pages_node[node]--;
Lee Schermerhorn685f3452009-09-21 17:01:23 -07001052 }
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001053 update_and_free_page(h, page);
1054 ret = 1;
Lee Schermerhorn9a76db02009-12-14 17:58:15 -08001055 break;
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001056 }
Joonsoo Kimb2261022013-09-11 14:21:00 -07001057 }
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001058
1059 return ret;
1060}
1061
Naoya Horiguchic8721bb2013-09-11 14:22:09 -07001062/*
1063 * Dissolve a given free hugepage into free buddy pages. This function does
1064 * nothing for in-use (including surplus) hugepages.
1065 */
1066static void dissolve_free_huge_page(struct page *page)
1067{
1068 spin_lock(&hugetlb_lock);
1069 if (PageHuge(page) && !page_count(page)) {
1070 struct hstate *h = page_hstate(page);
1071 int nid = page_to_nid(page);
1072 list_del(&page->lru);
1073 h->free_huge_pages--;
1074 h->free_huge_pages_node[nid]--;
1075 update_and_free_page(h, page);
1076 }
1077 spin_unlock(&hugetlb_lock);
1078}
1079
1080/*
1081 * Dissolve free hugepages in a given pfn range. Used by memory hotplug to
1082 * make specified memory blocks removable from the system.
1083 * Note that start_pfn should aligned with (minimum) hugepage size.
1084 */
1085void dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn)
1086{
1087 unsigned int order = 8 * sizeof(void *);
1088 unsigned long pfn;
1089 struct hstate *h;
1090
Li Zhongd0177632014-08-06 16:07:56 -07001091 if (!hugepages_supported())
1092 return;
1093
Naoya Horiguchic8721bb2013-09-11 14:22:09 -07001094 /* Set scan step to minimum hugepage size */
1095 for_each_hstate(h)
1096 if (order > huge_page_order(h))
1097 order = huge_page_order(h);
1098 VM_BUG_ON(!IS_ALIGNED(start_pfn, 1 << order));
1099 for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << order)
1100 dissolve_free_huge_page(pfn_to_page(pfn));
1101}
1102
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001103static struct page *alloc_buddy_huge_page(struct hstate *h, int nid)
Adam Litke7893d1d2007-10-16 01:26:18 -07001104{
1105 struct page *page;
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001106 unsigned int r_nid;
Adam Litke7893d1d2007-10-16 01:26:18 -07001107
Luiz Capitulinobae7f4a2014-06-04 16:07:08 -07001108 if (hstate_is_gigantic(h))
Andi Kleenaa888a72008-07-23 21:27:47 -07001109 return NULL;
1110
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001111 /*
1112 * Assume we will successfully allocate the surplus page to
1113 * prevent racing processes from causing the surplus to exceed
1114 * overcommit
1115 *
1116 * This however introduces a different race, where a process B
1117 * tries to grow the static hugepage pool while alloc_pages() is
1118 * called by process A. B will only examine the per-node
1119 * counters in determining if surplus huge pages can be
1120 * converted to normal huge pages in adjust_pool_surplus(). A
1121 * won't be able to increment the per-node counter, until the
1122 * lock is dropped by B, but B doesn't drop hugetlb_lock until
1123 * no more huge pages can be converted from surplus to normal
1124 * state (and doesn't try to convert again). Thus, we have a
1125 * case where a surplus huge page exists, the pool is grown, and
1126 * the surplus huge page still exists after, even though it
1127 * should just have been converted to a normal huge page. This
1128 * does not leak memory, though, as the hugepage will be freed
1129 * once it is out of use. It also does not allow the counters to
1130 * go out of whack in adjust_pool_surplus() as we don't modify
1131 * the node values until we've gotten the hugepage and only the
1132 * per-node value is checked there.
1133 */
1134 spin_lock(&hugetlb_lock);
Andi Kleena5516432008-07-23 21:27:41 -07001135 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001136 spin_unlock(&hugetlb_lock);
1137 return NULL;
1138 } else {
Andi Kleena5516432008-07-23 21:27:41 -07001139 h->nr_huge_pages++;
1140 h->surplus_huge_pages++;
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001141 }
1142 spin_unlock(&hugetlb_lock);
1143
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001144 if (nid == NUMA_NO_NODE)
Naoya Horiguchi86cdb462013-09-11 14:22:13 -07001145 page = alloc_pages(htlb_alloc_mask(h)|__GFP_COMP|
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001146 __GFP_REPEAT|__GFP_NOWARN,
1147 huge_page_order(h));
1148 else
1149 page = alloc_pages_exact_node(nid,
Naoya Horiguchi86cdb462013-09-11 14:22:13 -07001150 htlb_alloc_mask(h)|__GFP_COMP|__GFP_THISNODE|
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001151 __GFP_REPEAT|__GFP_NOWARN, huge_page_order(h));
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001152
Gerald Schaefercaff3a22008-08-12 15:08:38 -07001153 if (page && arch_prepare_hugepage(page)) {
1154 __free_pages(page, huge_page_order(h));
Hillf Dantonea5768c2012-01-10 15:08:30 -08001155 page = NULL;
Gerald Schaefercaff3a22008-08-12 15:08:38 -07001156 }
1157
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001158 spin_lock(&hugetlb_lock);
Adam Litke7893d1d2007-10-16 01:26:18 -07001159 if (page) {
Aneesh Kumar K.V0edaecf2012-07-31 16:42:07 -07001160 INIT_LIST_HEAD(&page->lru);
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001161 r_nid = page_to_nid(page);
Adam Litke7893d1d2007-10-16 01:26:18 -07001162 set_compound_page_dtor(page, free_huge_page);
Aneesh Kumar K.V9dd540e2012-07-31 16:42:15 -07001163 set_hugetlb_cgroup(page, NULL);
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001164 /*
1165 * We incremented the global counters already
1166 */
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001167 h->nr_huge_pages_node[r_nid]++;
1168 h->surplus_huge_pages_node[r_nid]++;
Adam Litke3b116302008-04-28 02:13:06 -07001169 __count_vm_event(HTLB_BUDDY_PGALLOC);
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001170 } else {
Andi Kleena5516432008-07-23 21:27:41 -07001171 h->nr_huge_pages--;
1172 h->surplus_huge_pages--;
Adam Litke3b116302008-04-28 02:13:06 -07001173 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
Adam Litke7893d1d2007-10-16 01:26:18 -07001174 }
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001175 spin_unlock(&hugetlb_lock);
Adam Litke7893d1d2007-10-16 01:26:18 -07001176
1177 return page;
1178}
1179
Adam Litkee4e574b2007-10-16 01:26:19 -07001180/*
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001181 * This allocation function is useful in the context where vma is irrelevant.
1182 * E.g. soft-offlining uses this function because it only cares physical
1183 * address of error page.
1184 */
1185struct page *alloc_huge_page_node(struct hstate *h, int nid)
1186{
Joonsoo Kim4ef91842013-09-11 14:21:51 -07001187 struct page *page = NULL;
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001188
1189 spin_lock(&hugetlb_lock);
Joonsoo Kim4ef91842013-09-11 14:21:51 -07001190 if (h->free_huge_pages - h->resv_huge_pages > 0)
1191 page = dequeue_huge_page_node(h, nid);
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001192 spin_unlock(&hugetlb_lock);
1193
Aneesh Kumar K.V94ae8ba2012-07-31 16:42:35 -07001194 if (!page)
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001195 page = alloc_buddy_huge_page(h, nid);
1196
1197 return page;
1198}
1199
1200/*
Lucas De Marchi25985ed2011-03-30 22:57:33 -03001201 * Increase the hugetlb pool such that it can accommodate a reservation
Adam Litkee4e574b2007-10-16 01:26:19 -07001202 * of size 'delta'.
1203 */
Andi Kleena5516432008-07-23 21:27:41 -07001204static int gather_surplus_pages(struct hstate *h, int delta)
Adam Litkee4e574b2007-10-16 01:26:19 -07001205{
1206 struct list_head surplus_list;
1207 struct page *page, *tmp;
1208 int ret, i;
1209 int needed, allocated;
Hillf Danton28073b02012-03-21 16:34:00 -07001210 bool alloc_ok = true;
Adam Litkee4e574b2007-10-16 01:26:19 -07001211
Andi Kleena5516432008-07-23 21:27:41 -07001212 needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
Adam Litkeac09b3a2008-03-04 14:29:38 -08001213 if (needed <= 0) {
Andi Kleena5516432008-07-23 21:27:41 -07001214 h->resv_huge_pages += delta;
Adam Litkee4e574b2007-10-16 01:26:19 -07001215 return 0;
Adam Litkeac09b3a2008-03-04 14:29:38 -08001216 }
Adam Litkee4e574b2007-10-16 01:26:19 -07001217
1218 allocated = 0;
1219 INIT_LIST_HEAD(&surplus_list);
1220
1221 ret = -ENOMEM;
1222retry:
1223 spin_unlock(&hugetlb_lock);
1224 for (i = 0; i < needed; i++) {
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001225 page = alloc_buddy_huge_page(h, NUMA_NO_NODE);
Hillf Danton28073b02012-03-21 16:34:00 -07001226 if (!page) {
1227 alloc_ok = false;
1228 break;
1229 }
Adam Litkee4e574b2007-10-16 01:26:19 -07001230 list_add(&page->lru, &surplus_list);
1231 }
Hillf Danton28073b02012-03-21 16:34:00 -07001232 allocated += i;
Adam Litkee4e574b2007-10-16 01:26:19 -07001233
1234 /*
1235 * After retaking hugetlb_lock, we need to recalculate 'needed'
1236 * because either resv_huge_pages or free_huge_pages may have changed.
1237 */
1238 spin_lock(&hugetlb_lock);
Andi Kleena5516432008-07-23 21:27:41 -07001239 needed = (h->resv_huge_pages + delta) -
1240 (h->free_huge_pages + allocated);
Hillf Danton28073b02012-03-21 16:34:00 -07001241 if (needed > 0) {
1242 if (alloc_ok)
1243 goto retry;
1244 /*
1245 * We were not able to allocate enough pages to
1246 * satisfy the entire reservation so we free what
1247 * we've allocated so far.
1248 */
1249 goto free;
1250 }
Adam Litkee4e574b2007-10-16 01:26:19 -07001251 /*
1252 * The surplus_list now contains _at_least_ the number of extra pages
Lucas De Marchi25985ed2011-03-30 22:57:33 -03001253 * needed to accommodate the reservation. Add the appropriate number
Adam Litkee4e574b2007-10-16 01:26:19 -07001254 * of pages to the hugetlb pool and free the extras back to the buddy
Adam Litkeac09b3a2008-03-04 14:29:38 -08001255 * allocator. Commit the entire reservation here to prevent another
1256 * process from stealing the pages as they are added to the pool but
1257 * before they are reserved.
Adam Litkee4e574b2007-10-16 01:26:19 -07001258 */
1259 needed += allocated;
Andi Kleena5516432008-07-23 21:27:41 -07001260 h->resv_huge_pages += delta;
Adam Litkee4e574b2007-10-16 01:26:19 -07001261 ret = 0;
Naoya Horiguchia9869b82010-09-08 10:19:37 +09001262
Adam Litke19fc3f02008-04-28 02:12:20 -07001263 /* Free the needed pages to the hugetlb pool */
Adam Litkee4e574b2007-10-16 01:26:19 -07001264 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
Adam Litke19fc3f02008-04-28 02:12:20 -07001265 if ((--needed) < 0)
1266 break;
Naoya Horiguchia9869b82010-09-08 10:19:37 +09001267 /*
1268 * This page is now managed by the hugetlb allocator and has
1269 * no users -- drop the buddy allocator's reference.
1270 */
1271 put_page_testzero(page);
Sasha Levin309381fea2014-01-23 15:52:54 -08001272 VM_BUG_ON_PAGE(page_count(page), page);
Andi Kleena5516432008-07-23 21:27:41 -07001273 enqueue_huge_page(h, page);
Adam Litke19fc3f02008-04-28 02:12:20 -07001274 }
Hillf Danton28073b02012-03-21 16:34:00 -07001275free:
Hillf Dantonb0365c82011-12-28 15:57:16 -08001276 spin_unlock(&hugetlb_lock);
Adam Litke19fc3f02008-04-28 02:12:20 -07001277
1278 /* Free unnecessary surplus pages to the buddy allocator */
Joonsoo Kimc0d934b2013-09-11 14:21:02 -07001279 list_for_each_entry_safe(page, tmp, &surplus_list, lru)
1280 put_page(page);
Naoya Horiguchia9869b82010-09-08 10:19:37 +09001281 spin_lock(&hugetlb_lock);
Adam Litkee4e574b2007-10-16 01:26:19 -07001282
1283 return ret;
1284}
1285
1286/*
1287 * When releasing a hugetlb pool reservation, any surplus pages that were
1288 * allocated to satisfy the reservation must be explicitly freed if they were
1289 * never used.
Lee Schermerhorn685f3452009-09-21 17:01:23 -07001290 * Called with hugetlb_lock held.
Adam Litkee4e574b2007-10-16 01:26:19 -07001291 */
Andi Kleena5516432008-07-23 21:27:41 -07001292static void return_unused_surplus_pages(struct hstate *h,
1293 unsigned long unused_resv_pages)
Adam Litkee4e574b2007-10-16 01:26:19 -07001294{
Adam Litkee4e574b2007-10-16 01:26:19 -07001295 unsigned long nr_pages;
1296
Adam Litkeac09b3a2008-03-04 14:29:38 -08001297 /* Uncommit the reservation */
Andi Kleena5516432008-07-23 21:27:41 -07001298 h->resv_huge_pages -= unused_resv_pages;
Adam Litkeac09b3a2008-03-04 14:29:38 -08001299
Andi Kleenaa888a72008-07-23 21:27:47 -07001300 /* Cannot return gigantic pages currently */
Luiz Capitulinobae7f4a2014-06-04 16:07:08 -07001301 if (hstate_is_gigantic(h))
Andi Kleenaa888a72008-07-23 21:27:47 -07001302 return;
1303
Andi Kleena5516432008-07-23 21:27:41 -07001304 nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
Adam Litkee4e574b2007-10-16 01:26:19 -07001305
Lee Schermerhorn685f3452009-09-21 17:01:23 -07001306 /*
1307 * We want to release as many surplus pages as possible, spread
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -08001308 * evenly across all nodes with memory. Iterate across these nodes
1309 * until we can no longer free unreserved surplus pages. This occurs
1310 * when the nodes with surplus pages have no free pages.
1311 * free_pool_huge_page() will balance the the freed pages across the
1312 * on-line nodes with memory and will handle the hstate accounting.
Lee Schermerhorn685f3452009-09-21 17:01:23 -07001313 */
1314 while (nr_pages--) {
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08001315 if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1))
Lee Schermerhorn685f3452009-09-21 17:01:23 -07001316 break;
Mizuma, Masayoshi7848a4b2014-04-18 15:07:18 -07001317 cond_resched_lock(&hugetlb_lock);
Adam Litkee4e574b2007-10-16 01:26:19 -07001318 }
1319}
1320
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001321/*
1322 * Determine if the huge page at addr within the vma has an associated
1323 * reservation. Where it does not we will need to logically increase
David Gibson90481622012-03-21 16:34:12 -07001324 * reservation and actually increase subpool usage before an allocation
1325 * can occur. Where any new reservation would be required the
1326 * reservation change is prepared, but not committed. Once the page
1327 * has been allocated from the subpool and instantiated the change should
1328 * be committed via vma_commit_reservation. No action is required on
1329 * failure.
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001330 */
Roel Kluine2f17d92009-03-31 15:23:15 -07001331static long vma_needs_reservation(struct hstate *h,
Andi Kleena5516432008-07-23 21:27:41 -07001332 struct vm_area_struct *vma, unsigned long addr)
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001333{
Joonsoo Kim4e35f482014-04-03 14:47:30 -07001334 struct resv_map *resv;
1335 pgoff_t idx;
1336 long chg;
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001337
Joonsoo Kim4e35f482014-04-03 14:47:30 -07001338 resv = vma_resv_map(vma);
1339 if (!resv)
Andy Whitcroft84afd992008-07-23 21:27:32 -07001340 return 1;
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001341
Joonsoo Kim4e35f482014-04-03 14:47:30 -07001342 idx = vma_hugecache_offset(h, vma, addr);
1343 chg = region_chg(resv, idx, idx + 1);
Andy Whitcroft84afd992008-07-23 21:27:32 -07001344
Joonsoo Kim4e35f482014-04-03 14:47:30 -07001345 if (vma->vm_flags & VM_MAYSHARE)
1346 return chg;
1347 else
1348 return chg < 0 ? chg : 0;
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001349}
Andi Kleena5516432008-07-23 21:27:41 -07001350static void vma_commit_reservation(struct hstate *h,
1351 struct vm_area_struct *vma, unsigned long addr)
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001352{
Joonsoo Kim4e35f482014-04-03 14:47:30 -07001353 struct resv_map *resv;
1354 pgoff_t idx;
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001355
Joonsoo Kim4e35f482014-04-03 14:47:30 -07001356 resv = vma_resv_map(vma);
1357 if (!resv)
1358 return;
Joonsoo Kim9119a412014-04-03 14:47:25 -07001359
Joonsoo Kim4e35f482014-04-03 14:47:30 -07001360 idx = vma_hugecache_offset(h, vma, addr);
1361 region_add(resv, idx, idx + 1);
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001362}
1363
David Gibson27a85ef2006-03-22 00:08:56 -08001364static struct page *alloc_huge_page(struct vm_area_struct *vma,
Mel Gorman04f2cbe2008-07-23 21:27:25 -07001365 unsigned long addr, int avoid_reserve)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001366{
David Gibson90481622012-03-21 16:34:12 -07001367 struct hugepage_subpool *spool = subpool_vma(vma);
Andi Kleena5516432008-07-23 21:27:41 -07001368 struct hstate *h = hstate_vma(vma);
Adam Litke348ea202007-11-14 16:59:37 -08001369 struct page *page;
Roel Kluine2f17d92009-03-31 15:23:15 -07001370 long chg;
Aneesh Kumar K.V6d76dcf2012-07-31 16:42:18 -07001371 int ret, idx;
1372 struct hugetlb_cgroup *h_cg;
Adam Litke2fc39ce2007-11-14 16:59:39 -08001373
Aneesh Kumar K.V6d76dcf2012-07-31 16:42:18 -07001374 idx = hstate_index(h);
Mel Gormana1e78772008-07-23 21:27:23 -07001375 /*
David Gibson90481622012-03-21 16:34:12 -07001376 * Processes that did not create the mapping will have no
1377 * reserves and will not have accounted against subpool
1378 * limit. Check that the subpool limit can be made before
1379 * satisfying the allocation MAP_NORESERVE mappings may also
1380 * need pages and subpool limit allocated allocated if no reserve
1381 * mapping overlaps.
Mel Gormana1e78772008-07-23 21:27:23 -07001382 */
Andi Kleena5516432008-07-23 21:27:41 -07001383 chg = vma_needs_reservation(h, vma, addr);
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001384 if (chg < 0)
Aneesh Kumar K.V76dcee72012-07-31 16:41:57 -07001385 return ERR_PTR(-ENOMEM);
Joonsoo Kim8bb3f122013-09-11 14:21:54 -07001386 if (chg || avoid_reserve)
1387 if (hugepage_subpool_get_pages(spool, 1))
Aneesh Kumar K.V76dcee72012-07-31 16:41:57 -07001388 return ERR_PTR(-ENOSPC);
Mel Gormana1e78772008-07-23 21:27:23 -07001389
Aneesh Kumar K.V6d76dcf2012-07-31 16:42:18 -07001390 ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg);
Jianyu Zhan8f34af62014-06-04 16:10:36 -07001391 if (ret)
1392 goto out_subpool_put;
1393
Mel Gormana1e78772008-07-23 21:27:23 -07001394 spin_lock(&hugetlb_lock);
Joonsoo Kimaf0ed732013-09-11 14:21:18 -07001395 page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, chg);
Joonsoo Kim81a6fca2013-09-11 14:20:58 -07001396 if (!page) {
Aneesh Kumar K.V94ae8ba2012-07-31 16:42:35 -07001397 spin_unlock(&hugetlb_lock);
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001398 page = alloc_buddy_huge_page(h, NUMA_NO_NODE);
Jianyu Zhan8f34af62014-06-04 16:10:36 -07001399 if (!page)
1400 goto out_uncharge_cgroup;
1401
Aneesh Kumar K.V79dbb232012-07-31 16:42:32 -07001402 spin_lock(&hugetlb_lock);
1403 list_move(&page->lru, &h->hugepage_activelist);
Joonsoo Kim81a6fca2013-09-11 14:20:58 -07001404 /* Fall through */
Mel Gormana1e78772008-07-23 21:27:23 -07001405 }
Joonsoo Kim81a6fca2013-09-11 14:20:58 -07001406 hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page);
1407 spin_unlock(&hugetlb_lock);
Mel Gormana1e78772008-07-23 21:27:23 -07001408
David Gibson90481622012-03-21 16:34:12 -07001409 set_page_private(page, (unsigned long)spool);
Mel Gormana1e78772008-07-23 21:27:23 -07001410
Andi Kleena5516432008-07-23 21:27:41 -07001411 vma_commit_reservation(h, vma, addr);
Adam Litke90d8b7e2007-11-14 16:59:42 -08001412 return page;
Jianyu Zhan8f34af62014-06-04 16:10:36 -07001413
1414out_uncharge_cgroup:
1415 hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg);
1416out_subpool_put:
1417 if (chg || avoid_reserve)
1418 hugepage_subpool_put_pages(spool, 1);
1419 return ERR_PTR(-ENOSPC);
David Gibsonb45b5bd2006-03-22 00:08:55 -08001420}
1421
Naoya Horiguchi74060e42013-09-11 14:22:06 -07001422/*
1423 * alloc_huge_page()'s wrapper which simply returns the page if allocation
1424 * succeeds, otherwise NULL. This function is called from new_vma_page(),
1425 * where no ERR_VALUE is expected to be returned.
1426 */
1427struct page *alloc_huge_page_noerr(struct vm_area_struct *vma,
1428 unsigned long addr, int avoid_reserve)
1429{
1430 struct page *page = alloc_huge_page(vma, addr, avoid_reserve);
1431 if (IS_ERR(page))
1432 page = NULL;
1433 return page;
1434}
1435
Cyrill Gorcunov91f47662009-01-06 14:40:33 -08001436int __weak alloc_bootmem_huge_page(struct hstate *h)
Andi Kleenaa888a72008-07-23 21:27:47 -07001437{
1438 struct huge_bootmem_page *m;
Joonsoo Kimb2261022013-09-11 14:21:00 -07001439 int nr_nodes, node;
Andi Kleenaa888a72008-07-23 21:27:47 -07001440
Joonsoo Kimb2261022013-09-11 14:21:00 -07001441 for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) {
Andi Kleenaa888a72008-07-23 21:27:47 -07001442 void *addr;
1443
Grygorii Strashko8b89a112014-01-21 15:50:36 -08001444 addr = memblock_virt_alloc_try_nid_nopanic(
1445 huge_page_size(h), huge_page_size(h),
1446 0, BOOTMEM_ALLOC_ACCESSIBLE, node);
Andi Kleenaa888a72008-07-23 21:27:47 -07001447 if (addr) {
1448 /*
1449 * Use the beginning of the huge page to store the
1450 * huge_bootmem_page struct (until gather_bootmem
1451 * puts them into the mem_map).
1452 */
1453 m = addr;
Cyrill Gorcunov91f47662009-01-06 14:40:33 -08001454 goto found;
Andi Kleenaa888a72008-07-23 21:27:47 -07001455 }
Andi Kleenaa888a72008-07-23 21:27:47 -07001456 }
1457 return 0;
1458
1459found:
Luiz Capitulinodf994ea2014-12-12 16:55:21 -08001460 BUG_ON(!IS_ALIGNED(virt_to_phys(m), huge_page_size(h)));
Andi Kleenaa888a72008-07-23 21:27:47 -07001461 /* Put them into a private list first because mem_map is not up yet */
1462 list_add(&m->list, &huge_boot_pages);
1463 m->hstate = h;
1464 return 1;
1465}
1466
David Rientjesf412c972014-04-03 14:47:59 -07001467static void __init prep_compound_huge_page(struct page *page, int order)
Andy Whitcroft18229df2008-11-06 12:53:27 -08001468{
1469 if (unlikely(order > (MAX_ORDER - 1)))
1470 prep_compound_gigantic_page(page, order);
1471 else
1472 prep_compound_page(page, order);
1473}
1474
Andi Kleenaa888a72008-07-23 21:27:47 -07001475/* Put bootmem huge pages into the standard lists after mem_map is up */
1476static void __init gather_bootmem_prealloc(void)
1477{
1478 struct huge_bootmem_page *m;
1479
1480 list_for_each_entry(m, &huge_boot_pages, list) {
Andi Kleenaa888a72008-07-23 21:27:47 -07001481 struct hstate *h = m->hstate;
Becky Bruceee8f2482011-07-25 17:11:50 -07001482 struct page *page;
1483
1484#ifdef CONFIG_HIGHMEM
1485 page = pfn_to_page(m->phys >> PAGE_SHIFT);
Grygorii Strashko8b89a112014-01-21 15:50:36 -08001486 memblock_free_late(__pa(m),
1487 sizeof(struct huge_bootmem_page));
Becky Bruceee8f2482011-07-25 17:11:50 -07001488#else
1489 page = virt_to_page(m);
1490#endif
Andi Kleenaa888a72008-07-23 21:27:47 -07001491 WARN_ON(page_count(page) != 1);
Andy Whitcroft18229df2008-11-06 12:53:27 -08001492 prep_compound_huge_page(page, h->order);
Andrea Arcangelief5a22b2013-10-16 13:46:56 -07001493 WARN_ON(PageReserved(page));
Andi Kleenaa888a72008-07-23 21:27:47 -07001494 prep_new_huge_page(h, page, page_to_nid(page));
Rafael Aquinib0320c72011-06-15 15:08:39 -07001495 /*
1496 * If we had gigantic hugepages allocated at boot time, we need
1497 * to restore the 'stolen' pages to totalram_pages in order to
1498 * fix confusing memory reports from free(1) and another
1499 * side-effects, like CommitLimit going negative.
1500 */
Luiz Capitulinobae7f4a2014-06-04 16:07:08 -07001501 if (hstate_is_gigantic(h))
Jiang Liu3dcc0572013-07-03 15:03:21 -07001502 adjust_managed_page_count(page, 1 << h->order);
Andi Kleenaa888a72008-07-23 21:27:47 -07001503 }
1504}
1505
Andi Kleen8faa8b02008-07-23 21:27:48 -07001506static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001507{
1508 unsigned long i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001509
Andi Kleene5ff2152008-07-23 21:27:42 -07001510 for (i = 0; i < h->max_huge_pages; ++i) {
Luiz Capitulinobae7f4a2014-06-04 16:07:08 -07001511 if (hstate_is_gigantic(h)) {
Andi Kleenaa888a72008-07-23 21:27:47 -07001512 if (!alloc_bootmem_huge_page(h))
1513 break;
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -08001514 } else if (!alloc_fresh_huge_page(h,
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08001515 &node_states[N_MEMORY]))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001516 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001517 }
Andi Kleen8faa8b02008-07-23 21:27:48 -07001518 h->max_huge_pages = i;
Andi Kleene5ff2152008-07-23 21:27:42 -07001519}
1520
1521static void __init hugetlb_init_hstates(void)
1522{
1523 struct hstate *h;
1524
1525 for_each_hstate(h) {
Andi Kleen8faa8b02008-07-23 21:27:48 -07001526 /* oversize hugepages were init'ed in early boot */
Luiz Capitulinobae7f4a2014-06-04 16:07:08 -07001527 if (!hstate_is_gigantic(h))
Andi Kleen8faa8b02008-07-23 21:27:48 -07001528 hugetlb_hstate_alloc_pages(h);
Andi Kleene5ff2152008-07-23 21:27:42 -07001529 }
1530}
1531
Andi Kleen4abd32d2008-07-23 21:27:49 -07001532static char * __init memfmt(char *buf, unsigned long n)
1533{
1534 if (n >= (1UL << 30))
1535 sprintf(buf, "%lu GB", n >> 30);
1536 else if (n >= (1UL << 20))
1537 sprintf(buf, "%lu MB", n >> 20);
1538 else
1539 sprintf(buf, "%lu KB", n >> 10);
1540 return buf;
1541}
1542
Andi Kleene5ff2152008-07-23 21:27:42 -07001543static void __init report_hugepages(void)
1544{
1545 struct hstate *h;
1546
1547 for_each_hstate(h) {
Andi Kleen4abd32d2008-07-23 21:27:49 -07001548 char buf[32];
Andrew Mortonffb22af2013-02-22 16:32:08 -08001549 pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n",
Andi Kleen4abd32d2008-07-23 21:27:49 -07001550 memfmt(buf, huge_page_size(h)),
1551 h->free_huge_pages);
Andi Kleene5ff2152008-07-23 21:27:42 -07001552 }
1553}
1554
Linus Torvalds1da177e2005-04-16 15:20:36 -07001555#ifdef CONFIG_HIGHMEM
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001556static void try_to_free_low(struct hstate *h, unsigned long count,
1557 nodemask_t *nodes_allowed)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001558{
Christoph Lameter4415cc82006-09-25 23:31:55 -07001559 int i;
1560
Luiz Capitulinobae7f4a2014-06-04 16:07:08 -07001561 if (hstate_is_gigantic(h))
Andi Kleenaa888a72008-07-23 21:27:47 -07001562 return;
1563
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001564 for_each_node_mask(i, *nodes_allowed) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001565 struct page *page, *next;
Andi Kleena5516432008-07-23 21:27:41 -07001566 struct list_head *freel = &h->hugepage_freelists[i];
1567 list_for_each_entry_safe(page, next, freel, lru) {
1568 if (count >= h->nr_huge_pages)
Adam Litke6b0c8802007-10-16 01:26:23 -07001569 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001570 if (PageHighMem(page))
1571 continue;
1572 list_del(&page->lru);
Andi Kleene5ff2152008-07-23 21:27:42 -07001573 update_and_free_page(h, page);
Andi Kleena5516432008-07-23 21:27:41 -07001574 h->free_huge_pages--;
1575 h->free_huge_pages_node[page_to_nid(page)]--;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001576 }
1577 }
1578}
1579#else
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001580static inline void try_to_free_low(struct hstate *h, unsigned long count,
1581 nodemask_t *nodes_allowed)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001582{
1583}
1584#endif
1585
Wu Fengguang20a03072009-06-16 15:32:22 -07001586/*
1587 * Increment or decrement surplus_huge_pages. Keep node-specific counters
1588 * balanced by operating on them in a round-robin fashion.
1589 * Returns 1 if an adjustment was made.
1590 */
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001591static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
1592 int delta)
Wu Fengguang20a03072009-06-16 15:32:22 -07001593{
Joonsoo Kimb2261022013-09-11 14:21:00 -07001594 int nr_nodes, node;
Wu Fengguang20a03072009-06-16 15:32:22 -07001595
1596 VM_BUG_ON(delta != -1 && delta != 1);
Wu Fengguang20a03072009-06-16 15:32:22 -07001597
Joonsoo Kimb2261022013-09-11 14:21:00 -07001598 if (delta < 0) {
1599 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
1600 if (h->surplus_huge_pages_node[node])
1601 goto found;
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001602 }
Joonsoo Kimb2261022013-09-11 14:21:00 -07001603 } else {
1604 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
1605 if (h->surplus_huge_pages_node[node] <
1606 h->nr_huge_pages_node[node])
1607 goto found;
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001608 }
Joonsoo Kimb2261022013-09-11 14:21:00 -07001609 }
1610 return 0;
Wu Fengguang20a03072009-06-16 15:32:22 -07001611
Joonsoo Kimb2261022013-09-11 14:21:00 -07001612found:
1613 h->surplus_huge_pages += delta;
1614 h->surplus_huge_pages_node[node] += delta;
1615 return 1;
Wu Fengguang20a03072009-06-16 15:32:22 -07001616}
1617
Andi Kleena5516432008-07-23 21:27:41 -07001618#define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001619static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count,
1620 nodemask_t *nodes_allowed)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001621{
Adam Litke7893d1d2007-10-16 01:26:18 -07001622 unsigned long min_count, ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001623
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001624 if (hstate_is_gigantic(h) && !gigantic_page_supported())
Andi Kleenaa888a72008-07-23 21:27:47 -07001625 return h->max_huge_pages;
1626
Adam Litke7893d1d2007-10-16 01:26:18 -07001627 /*
1628 * Increase the pool size
1629 * First take pages out of surplus state. Then make up the
1630 * remaining difference by allocating fresh huge pages.
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001631 *
1632 * We might race with alloc_buddy_huge_page() here and be unable
1633 * to convert a surplus huge page to a normal huge page. That is
1634 * not critical, though, it just means the overall size of the
1635 * pool might be one hugepage larger than it needs to be, but
1636 * within all the constraints specified by the sysctls.
Adam Litke7893d1d2007-10-16 01:26:18 -07001637 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001638 spin_lock(&hugetlb_lock);
Andi Kleena5516432008-07-23 21:27:41 -07001639 while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001640 if (!adjust_pool_surplus(h, nodes_allowed, -1))
Adam Litke7893d1d2007-10-16 01:26:18 -07001641 break;
1642 }
1643
Andi Kleena5516432008-07-23 21:27:41 -07001644 while (count > persistent_huge_pages(h)) {
Adam Litke7893d1d2007-10-16 01:26:18 -07001645 /*
1646 * If this allocation races such that we no longer need the
1647 * page, free_huge_page will handle it by freeing the page
1648 * and reducing the surplus.
1649 */
1650 spin_unlock(&hugetlb_lock);
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001651 if (hstate_is_gigantic(h))
1652 ret = alloc_fresh_gigantic_page(h, nodes_allowed);
1653 else
1654 ret = alloc_fresh_huge_page(h, nodes_allowed);
Adam Litke7893d1d2007-10-16 01:26:18 -07001655 spin_lock(&hugetlb_lock);
1656 if (!ret)
1657 goto out;
1658
Mel Gorman536240f22009-12-14 17:59:56 -08001659 /* Bail for signals. Probably ctrl-c from user */
1660 if (signal_pending(current))
1661 goto out;
Adam Litke7893d1d2007-10-16 01:26:18 -07001662 }
Adam Litke7893d1d2007-10-16 01:26:18 -07001663
1664 /*
1665 * Decrease the pool size
1666 * First return free pages to the buddy allocator (being careful
1667 * to keep enough around to satisfy reservations). Then place
1668 * pages into surplus state as needed so the pool will shrink
1669 * to the desired size as pages become free.
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001670 *
1671 * By placing pages into the surplus state independent of the
1672 * overcommit value, we are allowing the surplus pool size to
1673 * exceed overcommit. There are few sane options here. Since
1674 * alloc_buddy_huge_page() is checking the global counter,
1675 * though, we'll note that we're not allowed to exceed surplus
1676 * and won't grow the pool anywhere else. Not until one of the
1677 * sysctls are changed, or the surplus pages go out of use.
Adam Litke7893d1d2007-10-16 01:26:18 -07001678 */
Andi Kleena5516432008-07-23 21:27:41 -07001679 min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
Adam Litke6b0c8802007-10-16 01:26:23 -07001680 min_count = max(count, min_count);
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001681 try_to_free_low(h, min_count, nodes_allowed);
Andi Kleena5516432008-07-23 21:27:41 -07001682 while (min_count < persistent_huge_pages(h)) {
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001683 if (!free_pool_huge_page(h, nodes_allowed, 0))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001684 break;
Mizuma, Masayoshi55f67142014-04-07 15:37:54 -07001685 cond_resched_lock(&hugetlb_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001686 }
Andi Kleena5516432008-07-23 21:27:41 -07001687 while (count < persistent_huge_pages(h)) {
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001688 if (!adjust_pool_surplus(h, nodes_allowed, 1))
Adam Litke7893d1d2007-10-16 01:26:18 -07001689 break;
1690 }
1691out:
Andi Kleena5516432008-07-23 21:27:41 -07001692 ret = persistent_huge_pages(h);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001693 spin_unlock(&hugetlb_lock);
Adam Litke7893d1d2007-10-16 01:26:18 -07001694 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001695}
1696
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001697#define HSTATE_ATTR_RO(_name) \
1698 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
1699
1700#define HSTATE_ATTR(_name) \
1701 static struct kobj_attribute _name##_attr = \
1702 __ATTR(_name, 0644, _name##_show, _name##_store)
1703
1704static struct kobject *hugepages_kobj;
1705static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
1706
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001707static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
1708
1709static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001710{
1711 int i;
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001712
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001713 for (i = 0; i < HUGE_MAX_HSTATE; i++)
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001714 if (hstate_kobjs[i] == kobj) {
1715 if (nidp)
1716 *nidp = NUMA_NO_NODE;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001717 return &hstates[i];
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001718 }
1719
1720 return kobj_to_node_hstate(kobj, nidp);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001721}
1722
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001723static ssize_t nr_hugepages_show_common(struct kobject *kobj,
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001724 struct kobj_attribute *attr, char *buf)
1725{
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001726 struct hstate *h;
1727 unsigned long nr_huge_pages;
1728 int nid;
1729
1730 h = kobj_to_hstate(kobj, &nid);
1731 if (nid == NUMA_NO_NODE)
1732 nr_huge_pages = h->nr_huge_pages;
1733 else
1734 nr_huge_pages = h->nr_huge_pages_node[nid];
1735
1736 return sprintf(buf, "%lu\n", nr_huge_pages);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001737}
Eric B Munsonadbe8722011-01-13 15:47:27 -08001738
David Rientjes238d3c12014-08-06 16:06:51 -07001739static ssize_t __nr_hugepages_store_common(bool obey_mempolicy,
1740 struct hstate *h, int nid,
1741 unsigned long count, size_t len)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001742{
1743 int err;
David Rientjesbad44b52009-12-14 17:58:38 -08001744 NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001745
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001746 if (hstate_is_gigantic(h) && !gigantic_page_supported()) {
Eric B Munsonadbe8722011-01-13 15:47:27 -08001747 err = -EINVAL;
1748 goto out;
1749 }
1750
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001751 if (nid == NUMA_NO_NODE) {
1752 /*
1753 * global hstate attribute
1754 */
1755 if (!(obey_mempolicy &&
1756 init_nodemask_of_mempolicy(nodes_allowed))) {
1757 NODEMASK_FREE(nodes_allowed);
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08001758 nodes_allowed = &node_states[N_MEMORY];
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001759 }
1760 } else if (nodes_allowed) {
1761 /*
1762 * per node hstate attribute: adjust count to global,
1763 * but restrict alloc/free to the specified node.
1764 */
1765 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
1766 init_nodemask_of_node(nodes_allowed, nid);
1767 } else
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08001768 nodes_allowed = &node_states[N_MEMORY];
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001769
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001770 h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001771
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08001772 if (nodes_allowed != &node_states[N_MEMORY])
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001773 NODEMASK_FREE(nodes_allowed);
1774
1775 return len;
Eric B Munsonadbe8722011-01-13 15:47:27 -08001776out:
1777 NODEMASK_FREE(nodes_allowed);
1778 return err;
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001779}
1780
David Rientjes238d3c12014-08-06 16:06:51 -07001781static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
1782 struct kobject *kobj, const char *buf,
1783 size_t len)
1784{
1785 struct hstate *h;
1786 unsigned long count;
1787 int nid;
1788 int err;
1789
1790 err = kstrtoul(buf, 10, &count);
1791 if (err)
1792 return err;
1793
1794 h = kobj_to_hstate(kobj, &nid);
1795 return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len);
1796}
1797
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001798static ssize_t nr_hugepages_show(struct kobject *kobj,
1799 struct kobj_attribute *attr, char *buf)
1800{
1801 return nr_hugepages_show_common(kobj, attr, buf);
1802}
1803
1804static ssize_t nr_hugepages_store(struct kobject *kobj,
1805 struct kobj_attribute *attr, const char *buf, size_t len)
1806{
David Rientjes238d3c12014-08-06 16:06:51 -07001807 return nr_hugepages_store_common(false, kobj, buf, len);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001808}
1809HSTATE_ATTR(nr_hugepages);
1810
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001811#ifdef CONFIG_NUMA
1812
1813/*
1814 * hstate attribute for optionally mempolicy-based constraint on persistent
1815 * huge page alloc/free.
1816 */
1817static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
1818 struct kobj_attribute *attr, char *buf)
1819{
1820 return nr_hugepages_show_common(kobj, attr, buf);
1821}
1822
1823static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
1824 struct kobj_attribute *attr, const char *buf, size_t len)
1825{
David Rientjes238d3c12014-08-06 16:06:51 -07001826 return nr_hugepages_store_common(true, kobj, buf, len);
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001827}
1828HSTATE_ATTR(nr_hugepages_mempolicy);
1829#endif
1830
1831
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001832static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
1833 struct kobj_attribute *attr, char *buf)
1834{
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001835 struct hstate *h = kobj_to_hstate(kobj, NULL);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001836 return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
1837}
Eric B Munsonadbe8722011-01-13 15:47:27 -08001838
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001839static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
1840 struct kobj_attribute *attr, const char *buf, size_t count)
1841{
1842 int err;
1843 unsigned long input;
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001844 struct hstate *h = kobj_to_hstate(kobj, NULL);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001845
Luiz Capitulinobae7f4a2014-06-04 16:07:08 -07001846 if (hstate_is_gigantic(h))
Eric B Munsonadbe8722011-01-13 15:47:27 -08001847 return -EINVAL;
1848
Jingoo Han3dbb95f2013-09-11 14:20:25 -07001849 err = kstrtoul(buf, 10, &input);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001850 if (err)
Eric B Munson73ae31e2011-01-13 15:47:28 -08001851 return err;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001852
1853 spin_lock(&hugetlb_lock);
1854 h->nr_overcommit_huge_pages = input;
1855 spin_unlock(&hugetlb_lock);
1856
1857 return count;
1858}
1859HSTATE_ATTR(nr_overcommit_hugepages);
1860
1861static ssize_t free_hugepages_show(struct kobject *kobj,
1862 struct kobj_attribute *attr, char *buf)
1863{
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001864 struct hstate *h;
1865 unsigned long free_huge_pages;
1866 int nid;
1867
1868 h = kobj_to_hstate(kobj, &nid);
1869 if (nid == NUMA_NO_NODE)
1870 free_huge_pages = h->free_huge_pages;
1871 else
1872 free_huge_pages = h->free_huge_pages_node[nid];
1873
1874 return sprintf(buf, "%lu\n", free_huge_pages);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001875}
1876HSTATE_ATTR_RO(free_hugepages);
1877
1878static ssize_t resv_hugepages_show(struct kobject *kobj,
1879 struct kobj_attribute *attr, char *buf)
1880{
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001881 struct hstate *h = kobj_to_hstate(kobj, NULL);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001882 return sprintf(buf, "%lu\n", h->resv_huge_pages);
1883}
1884HSTATE_ATTR_RO(resv_hugepages);
1885
1886static ssize_t surplus_hugepages_show(struct kobject *kobj,
1887 struct kobj_attribute *attr, char *buf)
1888{
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001889 struct hstate *h;
1890 unsigned long surplus_huge_pages;
1891 int nid;
1892
1893 h = kobj_to_hstate(kobj, &nid);
1894 if (nid == NUMA_NO_NODE)
1895 surplus_huge_pages = h->surplus_huge_pages;
1896 else
1897 surplus_huge_pages = h->surplus_huge_pages_node[nid];
1898
1899 return sprintf(buf, "%lu\n", surplus_huge_pages);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001900}
1901HSTATE_ATTR_RO(surplus_hugepages);
1902
1903static struct attribute *hstate_attrs[] = {
1904 &nr_hugepages_attr.attr,
1905 &nr_overcommit_hugepages_attr.attr,
1906 &free_hugepages_attr.attr,
1907 &resv_hugepages_attr.attr,
1908 &surplus_hugepages_attr.attr,
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001909#ifdef CONFIG_NUMA
1910 &nr_hugepages_mempolicy_attr.attr,
1911#endif
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001912 NULL,
1913};
1914
1915static struct attribute_group hstate_attr_group = {
1916 .attrs = hstate_attrs,
1917};
1918
Jeff Mahoney094e9532010-02-02 13:44:14 -08001919static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
1920 struct kobject **hstate_kobjs,
1921 struct attribute_group *hstate_attr_group)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001922{
1923 int retval;
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07001924 int hi = hstate_index(h);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001925
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001926 hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
1927 if (!hstate_kobjs[hi])
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001928 return -ENOMEM;
1929
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001930 retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001931 if (retval)
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001932 kobject_put(hstate_kobjs[hi]);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001933
1934 return retval;
1935}
1936
1937static void __init hugetlb_sysfs_init(void)
1938{
1939 struct hstate *h;
1940 int err;
1941
1942 hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
1943 if (!hugepages_kobj)
1944 return;
1945
1946 for_each_hstate(h) {
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001947 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
1948 hstate_kobjs, &hstate_attr_group);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001949 if (err)
Andrew Mortonffb22af2013-02-22 16:32:08 -08001950 pr_err("Hugetlb: Unable to add hstate %s", h->name);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001951 }
1952}
1953
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001954#ifdef CONFIG_NUMA
1955
1956/*
1957 * node_hstate/s - associate per node hstate attributes, via their kobjects,
Kay Sievers10fbcf42011-12-21 14:48:43 -08001958 * with node devices in node_devices[] using a parallel array. The array
1959 * index of a node device or _hstate == node id.
1960 * This is here to avoid any static dependency of the node device driver, in
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001961 * the base kernel, on the hugetlb module.
1962 */
1963struct node_hstate {
1964 struct kobject *hugepages_kobj;
1965 struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
1966};
1967struct node_hstate node_hstates[MAX_NUMNODES];
1968
1969/*
Kay Sievers10fbcf42011-12-21 14:48:43 -08001970 * A subset of global hstate attributes for node devices
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001971 */
1972static struct attribute *per_node_hstate_attrs[] = {
1973 &nr_hugepages_attr.attr,
1974 &free_hugepages_attr.attr,
1975 &surplus_hugepages_attr.attr,
1976 NULL,
1977};
1978
1979static struct attribute_group per_node_hstate_attr_group = {
1980 .attrs = per_node_hstate_attrs,
1981};
1982
1983/*
Kay Sievers10fbcf42011-12-21 14:48:43 -08001984 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001985 * Returns node id via non-NULL nidp.
1986 */
1987static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
1988{
1989 int nid;
1990
1991 for (nid = 0; nid < nr_node_ids; nid++) {
1992 struct node_hstate *nhs = &node_hstates[nid];
1993 int i;
1994 for (i = 0; i < HUGE_MAX_HSTATE; i++)
1995 if (nhs->hstate_kobjs[i] == kobj) {
1996 if (nidp)
1997 *nidp = nid;
1998 return &hstates[i];
1999 }
2000 }
2001
2002 BUG();
2003 return NULL;
2004}
2005
2006/*
Kay Sievers10fbcf42011-12-21 14:48:43 -08002007 * Unregister hstate attributes from a single node device.
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002008 * No-op if no hstate attributes attached.
2009 */
Claudiu Ghioc3cd8b442013-03-04 12:46:15 +02002010static void hugetlb_unregister_node(struct node *node)
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002011{
2012 struct hstate *h;
Kay Sievers10fbcf42011-12-21 14:48:43 -08002013 struct node_hstate *nhs = &node_hstates[node->dev.id];
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002014
2015 if (!nhs->hugepages_kobj)
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -08002016 return; /* no hstate attributes */
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002017
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07002018 for_each_hstate(h) {
2019 int idx = hstate_index(h);
2020 if (nhs->hstate_kobjs[idx]) {
2021 kobject_put(nhs->hstate_kobjs[idx]);
2022 nhs->hstate_kobjs[idx] = NULL;
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002023 }
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07002024 }
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002025
2026 kobject_put(nhs->hugepages_kobj);
2027 nhs->hugepages_kobj = NULL;
2028}
2029
2030/*
Kay Sievers10fbcf42011-12-21 14:48:43 -08002031 * hugetlb module exit: unregister hstate attributes from node devices
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002032 * that have them.
2033 */
2034static void hugetlb_unregister_all_nodes(void)
2035{
2036 int nid;
2037
2038 /*
Kay Sievers10fbcf42011-12-21 14:48:43 -08002039 * disable node device registrations.
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002040 */
2041 register_hugetlbfs_with_node(NULL, NULL);
2042
2043 /*
2044 * remove hstate attributes from any nodes that have them.
2045 */
2046 for (nid = 0; nid < nr_node_ids; nid++)
Wen Congyang87327942012-12-11 16:00:56 -08002047 hugetlb_unregister_node(node_devices[nid]);
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002048}
2049
2050/*
Kay Sievers10fbcf42011-12-21 14:48:43 -08002051 * Register hstate attributes for a single node device.
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002052 * No-op if attributes already registered.
2053 */
Claudiu Ghioc3cd8b442013-03-04 12:46:15 +02002054static void hugetlb_register_node(struct node *node)
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002055{
2056 struct hstate *h;
Kay Sievers10fbcf42011-12-21 14:48:43 -08002057 struct node_hstate *nhs = &node_hstates[node->dev.id];
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002058 int err;
2059
2060 if (nhs->hugepages_kobj)
2061 return; /* already allocated */
2062
2063 nhs->hugepages_kobj = kobject_create_and_add("hugepages",
Kay Sievers10fbcf42011-12-21 14:48:43 -08002064 &node->dev.kobj);
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002065 if (!nhs->hugepages_kobj)
2066 return;
2067
2068 for_each_hstate(h) {
2069 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
2070 nhs->hstate_kobjs,
2071 &per_node_hstate_attr_group);
2072 if (err) {
Andrew Mortonffb22af2013-02-22 16:32:08 -08002073 pr_err("Hugetlb: Unable to add hstate %s for node %d\n",
2074 h->name, node->dev.id);
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002075 hugetlb_unregister_node(node);
2076 break;
2077 }
2078 }
2079}
2080
2081/*
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -08002082 * hugetlb init time: register hstate attributes for all registered node
Kay Sievers10fbcf42011-12-21 14:48:43 -08002083 * devices of nodes that have memory. All on-line nodes should have
2084 * registered their associated device by this time.
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002085 */
Luiz Capitulino7d9ca002014-12-12 16:55:24 -08002086static void __init hugetlb_register_all_nodes(void)
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002087{
2088 int nid;
2089
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08002090 for_each_node_state(nid, N_MEMORY) {
Wen Congyang87327942012-12-11 16:00:56 -08002091 struct node *node = node_devices[nid];
Kay Sievers10fbcf42011-12-21 14:48:43 -08002092 if (node->dev.id == nid)
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002093 hugetlb_register_node(node);
2094 }
2095
2096 /*
Kay Sievers10fbcf42011-12-21 14:48:43 -08002097 * Let the node device driver know we're here so it can
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002098 * [un]register hstate attributes on node hotplug.
2099 */
2100 register_hugetlbfs_with_node(hugetlb_register_node,
2101 hugetlb_unregister_node);
2102}
2103#else /* !CONFIG_NUMA */
2104
2105static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
2106{
2107 BUG();
2108 if (nidp)
2109 *nidp = -1;
2110 return NULL;
2111}
2112
2113static void hugetlb_unregister_all_nodes(void) { }
2114
2115static void hugetlb_register_all_nodes(void) { }
2116
2117#endif
2118
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002119static void __exit hugetlb_exit(void)
2120{
2121 struct hstate *h;
2122
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002123 hugetlb_unregister_all_nodes();
2124
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002125 for_each_hstate(h) {
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07002126 kobject_put(hstate_kobjs[hstate_index(h)]);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002127 }
2128
2129 kobject_put(hugepages_kobj);
Davidlohr Bueso8382d912014-04-03 14:47:31 -07002130 kfree(htlb_fault_mutex_table);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002131}
2132module_exit(hugetlb_exit);
2133
2134static int __init hugetlb_init(void)
2135{
Davidlohr Bueso8382d912014-04-03 14:47:31 -07002136 int i;
2137
Nishanth Aravamudan457c1b22014-05-06 12:50:00 -07002138 if (!hugepages_supported())
Benjamin Herrenschmidt0ef89d22008-07-31 00:07:30 -07002139 return 0;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002140
Nick Piggine11bfbf2008-07-23 21:27:52 -07002141 if (!size_to_hstate(default_hstate_size)) {
2142 default_hstate_size = HPAGE_SIZE;
2143 if (!size_to_hstate(default_hstate_size))
2144 hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002145 }
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07002146 default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size));
Nick Piggine11bfbf2008-07-23 21:27:52 -07002147 if (default_hstate_max_huge_pages)
2148 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002149
2150 hugetlb_init_hstates();
Andi Kleenaa888a72008-07-23 21:27:47 -07002151 gather_bootmem_prealloc();
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002152 report_hugepages();
2153
2154 hugetlb_sysfs_init();
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002155 hugetlb_register_all_nodes();
Jianguo Wu7179e7b2012-12-18 14:23:19 -08002156 hugetlb_cgroup_file_init();
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002157
Davidlohr Bueso8382d912014-04-03 14:47:31 -07002158#ifdef CONFIG_SMP
2159 num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus());
2160#else
2161 num_fault_mutexes = 1;
2162#endif
2163 htlb_fault_mutex_table =
2164 kmalloc(sizeof(struct mutex) * num_fault_mutexes, GFP_KERNEL);
2165 BUG_ON(!htlb_fault_mutex_table);
2166
2167 for (i = 0; i < num_fault_mutexes; i++)
2168 mutex_init(&htlb_fault_mutex_table[i]);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002169 return 0;
2170}
2171module_init(hugetlb_init);
2172
2173/* Should be called on processing a hugepagesz=... option */
2174void __init hugetlb_add_hstate(unsigned order)
2175{
2176 struct hstate *h;
Andi Kleen8faa8b02008-07-23 21:27:48 -07002177 unsigned long i;
2178
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002179 if (size_to_hstate(PAGE_SIZE << order)) {
Andrew Mortonffb22af2013-02-22 16:32:08 -08002180 pr_warning("hugepagesz= specified twice, ignoring\n");
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002181 return;
2182 }
Aneesh Kumar K.V47d38342012-07-31 16:41:54 -07002183 BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002184 BUG_ON(order == 0);
Aneesh Kumar K.V47d38342012-07-31 16:41:54 -07002185 h = &hstates[hugetlb_max_hstate++];
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002186 h->order = order;
2187 h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
Andi Kleen8faa8b02008-07-23 21:27:48 -07002188 h->nr_huge_pages = 0;
2189 h->free_huge_pages = 0;
2190 for (i = 0; i < MAX_NUMNODES; ++i)
2191 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
Aneesh Kumar K.V0edaecf2012-07-31 16:42:07 -07002192 INIT_LIST_HEAD(&h->hugepage_activelist);
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08002193 h->next_nid_to_alloc = first_node(node_states[N_MEMORY]);
2194 h->next_nid_to_free = first_node(node_states[N_MEMORY]);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002195 snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
2196 huge_page_size(h)/1024);
Andi Kleen8faa8b02008-07-23 21:27:48 -07002197
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002198 parsed_hstate = h;
2199}
2200
Nick Piggine11bfbf2008-07-23 21:27:52 -07002201static int __init hugetlb_nrpages_setup(char *s)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002202{
2203 unsigned long *mhp;
Andi Kleen8faa8b02008-07-23 21:27:48 -07002204 static unsigned long *last_mhp;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002205
2206 /*
Aneesh Kumar K.V47d38342012-07-31 16:41:54 -07002207 * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet,
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002208 * so this hugepages= parameter goes to the "default hstate".
2209 */
Aneesh Kumar K.V47d38342012-07-31 16:41:54 -07002210 if (!hugetlb_max_hstate)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002211 mhp = &default_hstate_max_huge_pages;
2212 else
2213 mhp = &parsed_hstate->max_huge_pages;
2214
Andi Kleen8faa8b02008-07-23 21:27:48 -07002215 if (mhp == last_mhp) {
Andrew Mortonffb22af2013-02-22 16:32:08 -08002216 pr_warning("hugepages= specified twice without "
2217 "interleaving hugepagesz=, ignoring\n");
Andi Kleen8faa8b02008-07-23 21:27:48 -07002218 return 1;
2219 }
2220
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002221 if (sscanf(s, "%lu", mhp) <= 0)
2222 *mhp = 0;
2223
Andi Kleen8faa8b02008-07-23 21:27:48 -07002224 /*
2225 * Global state is always initialized later in hugetlb_init.
2226 * But we need to allocate >= MAX_ORDER hstates here early to still
2227 * use the bootmem allocator.
2228 */
Aneesh Kumar K.V47d38342012-07-31 16:41:54 -07002229 if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER)
Andi Kleen8faa8b02008-07-23 21:27:48 -07002230 hugetlb_hstate_alloc_pages(parsed_hstate);
2231
2232 last_mhp = mhp;
2233
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002234 return 1;
2235}
Nick Piggine11bfbf2008-07-23 21:27:52 -07002236__setup("hugepages=", hugetlb_nrpages_setup);
2237
2238static int __init hugetlb_default_setup(char *s)
2239{
2240 default_hstate_size = memparse(s, &s);
2241 return 1;
2242}
2243__setup("default_hugepagesz=", hugetlb_default_setup);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002244
Nishanth Aravamudan8a213462008-07-25 19:44:37 -07002245static unsigned int cpuset_mems_nr(unsigned int *array)
2246{
2247 int node;
2248 unsigned int nr = 0;
2249
2250 for_each_node_mask(node, cpuset_current_mems_allowed)
2251 nr += array[node];
2252
2253 return nr;
2254}
2255
2256#ifdef CONFIG_SYSCTL
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002257static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
2258 struct ctl_table *table, int write,
2259 void __user *buffer, size_t *length, loff_t *ppos)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002260{
Andi Kleene5ff2152008-07-23 21:27:42 -07002261 struct hstate *h = &default_hstate;
David Rientjes238d3c12014-08-06 16:06:51 -07002262 unsigned long tmp = h->max_huge_pages;
Michal Hocko08d4a242011-01-13 15:47:26 -08002263 int ret;
Andi Kleene5ff2152008-07-23 21:27:42 -07002264
Nishanth Aravamudan457c1b22014-05-06 12:50:00 -07002265 if (!hugepages_supported())
2266 return -ENOTSUPP;
2267
Andi Kleene5ff2152008-07-23 21:27:42 -07002268 table->data = &tmp;
2269 table->maxlen = sizeof(unsigned long);
Michal Hocko08d4a242011-01-13 15:47:26 -08002270 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2271 if (ret)
2272 goto out;
Andi Kleene5ff2152008-07-23 21:27:42 -07002273
David Rientjes238d3c12014-08-06 16:06:51 -07002274 if (write)
2275 ret = __nr_hugepages_store_common(obey_mempolicy, h,
2276 NUMA_NO_NODE, tmp, *length);
Michal Hocko08d4a242011-01-13 15:47:26 -08002277out:
2278 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002279}
Mel Gorman396faf02007-07-17 04:03:13 -07002280
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002281int hugetlb_sysctl_handler(struct ctl_table *table, int write,
2282 void __user *buffer, size_t *length, loff_t *ppos)
2283{
2284
2285 return hugetlb_sysctl_handler_common(false, table, write,
2286 buffer, length, ppos);
2287}
2288
2289#ifdef CONFIG_NUMA
2290int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
2291 void __user *buffer, size_t *length, loff_t *ppos)
2292{
2293 return hugetlb_sysctl_handler_common(true, table, write,
2294 buffer, length, ppos);
2295}
2296#endif /* CONFIG_NUMA */
2297
Nishanth Aravamudana3d0c6a2008-02-08 04:18:18 -08002298int hugetlb_overcommit_handler(struct ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07002299 void __user *buffer,
Nishanth Aravamudana3d0c6a2008-02-08 04:18:18 -08002300 size_t *length, loff_t *ppos)
2301{
Andi Kleena5516432008-07-23 21:27:41 -07002302 struct hstate *h = &default_hstate;
Andi Kleene5ff2152008-07-23 21:27:42 -07002303 unsigned long tmp;
Michal Hocko08d4a242011-01-13 15:47:26 -08002304 int ret;
Andi Kleene5ff2152008-07-23 21:27:42 -07002305
Nishanth Aravamudan457c1b22014-05-06 12:50:00 -07002306 if (!hugepages_supported())
2307 return -ENOTSUPP;
2308
Petr Holasekc033a932011-03-22 16:33:05 -07002309 tmp = h->nr_overcommit_huge_pages;
Andi Kleene5ff2152008-07-23 21:27:42 -07002310
Luiz Capitulinobae7f4a2014-06-04 16:07:08 -07002311 if (write && hstate_is_gigantic(h))
Eric B Munsonadbe8722011-01-13 15:47:27 -08002312 return -EINVAL;
2313
Andi Kleene5ff2152008-07-23 21:27:42 -07002314 table->data = &tmp;
2315 table->maxlen = sizeof(unsigned long);
Michal Hocko08d4a242011-01-13 15:47:26 -08002316 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2317 if (ret)
2318 goto out;
Andi Kleene5ff2152008-07-23 21:27:42 -07002319
2320 if (write) {
2321 spin_lock(&hugetlb_lock);
2322 h->nr_overcommit_huge_pages = tmp;
2323 spin_unlock(&hugetlb_lock);
2324 }
Michal Hocko08d4a242011-01-13 15:47:26 -08002325out:
2326 return ret;
Nishanth Aravamudana3d0c6a2008-02-08 04:18:18 -08002327}
2328
Linus Torvalds1da177e2005-04-16 15:20:36 -07002329#endif /* CONFIG_SYSCTL */
2330
Alexey Dobriyane1759c22008-10-15 23:50:22 +04002331void hugetlb_report_meminfo(struct seq_file *m)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002332{
Andi Kleena5516432008-07-23 21:27:41 -07002333 struct hstate *h = &default_hstate;
Nishanth Aravamudan457c1b22014-05-06 12:50:00 -07002334 if (!hugepages_supported())
2335 return;
Alexey Dobriyane1759c22008-10-15 23:50:22 +04002336 seq_printf(m,
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002337 "HugePages_Total: %5lu\n"
2338 "HugePages_Free: %5lu\n"
2339 "HugePages_Rsvd: %5lu\n"
2340 "HugePages_Surp: %5lu\n"
2341 "Hugepagesize: %8lu kB\n",
Andi Kleena5516432008-07-23 21:27:41 -07002342 h->nr_huge_pages,
2343 h->free_huge_pages,
2344 h->resv_huge_pages,
2345 h->surplus_huge_pages,
2346 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002347}
2348
2349int hugetlb_report_node_meminfo(int nid, char *buf)
2350{
Andi Kleena5516432008-07-23 21:27:41 -07002351 struct hstate *h = &default_hstate;
Nishanth Aravamudan457c1b22014-05-06 12:50:00 -07002352 if (!hugepages_supported())
2353 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002354 return sprintf(buf,
2355 "Node %d HugePages_Total: %5u\n"
Nishanth Aravamudana1de0912008-03-26 14:37:53 -07002356 "Node %d HugePages_Free: %5u\n"
2357 "Node %d HugePages_Surp: %5u\n",
Andi Kleena5516432008-07-23 21:27:41 -07002358 nid, h->nr_huge_pages_node[nid],
2359 nid, h->free_huge_pages_node[nid],
2360 nid, h->surplus_huge_pages_node[nid]);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002361}
2362
David Rientjes949f7ec2013-04-29 15:07:48 -07002363void hugetlb_show_meminfo(void)
2364{
2365 struct hstate *h;
2366 int nid;
2367
Nishanth Aravamudan457c1b22014-05-06 12:50:00 -07002368 if (!hugepages_supported())
2369 return;
2370
David Rientjes949f7ec2013-04-29 15:07:48 -07002371 for_each_node_state(nid, N_MEMORY)
2372 for_each_hstate(h)
2373 pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
2374 nid,
2375 h->nr_huge_pages_node[nid],
2376 h->free_huge_pages_node[nid],
2377 h->surplus_huge_pages_node[nid],
2378 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
2379}
2380
Linus Torvalds1da177e2005-04-16 15:20:36 -07002381/* Return the number pages of memory we physically have, in PAGE_SIZE units. */
2382unsigned long hugetlb_total_pages(void)
2383{
Wanpeng Lid0028582013-03-22 15:04:40 -07002384 struct hstate *h;
2385 unsigned long nr_total_pages = 0;
2386
2387 for_each_hstate(h)
2388 nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h);
2389 return nr_total_pages;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002390}
Linus Torvalds1da177e2005-04-16 15:20:36 -07002391
Andi Kleena5516432008-07-23 21:27:41 -07002392static int hugetlb_acct_memory(struct hstate *h, long delta)
Mel Gormanfc1b8a72008-07-23 21:27:22 -07002393{
2394 int ret = -ENOMEM;
2395
2396 spin_lock(&hugetlb_lock);
2397 /*
2398 * When cpuset is configured, it breaks the strict hugetlb page
2399 * reservation as the accounting is done on a global variable. Such
2400 * reservation is completely rubbish in the presence of cpuset because
2401 * the reservation is not checked against page availability for the
2402 * current cpuset. Application can still potentially OOM'ed by kernel
2403 * with lack of free htlb page in cpuset that the task is in.
2404 * Attempt to enforce strict accounting with cpuset is almost
2405 * impossible (or too ugly) because cpuset is too fluid that
2406 * task or memory node can be dynamically moved between cpusets.
2407 *
2408 * The change of semantics for shared hugetlb mapping with cpuset is
2409 * undesirable. However, in order to preserve some of the semantics,
2410 * we fall back to check against current free page availability as
2411 * a best attempt and hopefully to minimize the impact of changing
2412 * semantics that cpuset has.
2413 */
2414 if (delta > 0) {
Andi Kleena5516432008-07-23 21:27:41 -07002415 if (gather_surplus_pages(h, delta) < 0)
Mel Gormanfc1b8a72008-07-23 21:27:22 -07002416 goto out;
2417
Andi Kleena5516432008-07-23 21:27:41 -07002418 if (delta > cpuset_mems_nr(h->free_huge_pages_node)) {
2419 return_unused_surplus_pages(h, delta);
Mel Gormanfc1b8a72008-07-23 21:27:22 -07002420 goto out;
2421 }
2422 }
2423
2424 ret = 0;
2425 if (delta < 0)
Andi Kleena5516432008-07-23 21:27:41 -07002426 return_unused_surplus_pages(h, (unsigned long) -delta);
Mel Gormanfc1b8a72008-07-23 21:27:22 -07002427
2428out:
2429 spin_unlock(&hugetlb_lock);
2430 return ret;
2431}
2432
Andy Whitcroft84afd992008-07-23 21:27:32 -07002433static void hugetlb_vm_op_open(struct vm_area_struct *vma)
2434{
Joonsoo Kimf522c3a2013-09-11 14:21:53 -07002435 struct resv_map *resv = vma_resv_map(vma);
Andy Whitcroft84afd992008-07-23 21:27:32 -07002436
2437 /*
2438 * This new VMA should share its siblings reservation map if present.
2439 * The VMA will only ever have a valid reservation map pointer where
2440 * it is being copied for another still existing VMA. As that VMA
Lucas De Marchi25985ed2011-03-30 22:57:33 -03002441 * has a reference to the reservation map it cannot disappear until
Andy Whitcroft84afd992008-07-23 21:27:32 -07002442 * after this open call completes. It is therefore safe to take a
2443 * new reference here without additional locking.
2444 */
Joonsoo Kim4e35f482014-04-03 14:47:30 -07002445 if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
Joonsoo Kimf522c3a2013-09-11 14:21:53 -07002446 kref_get(&resv->refs);
Andy Whitcroft84afd992008-07-23 21:27:32 -07002447}
2448
Mel Gormana1e78772008-07-23 21:27:23 -07002449static void hugetlb_vm_op_close(struct vm_area_struct *vma)
2450{
Andi Kleena5516432008-07-23 21:27:41 -07002451 struct hstate *h = hstate_vma(vma);
Joonsoo Kimf522c3a2013-09-11 14:21:53 -07002452 struct resv_map *resv = vma_resv_map(vma);
David Gibson90481622012-03-21 16:34:12 -07002453 struct hugepage_subpool *spool = subpool_vma(vma);
Joonsoo Kim4e35f482014-04-03 14:47:30 -07002454 unsigned long reserve, start, end;
Andy Whitcroft84afd992008-07-23 21:27:32 -07002455
Joonsoo Kim4e35f482014-04-03 14:47:30 -07002456 if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER))
2457 return;
Andy Whitcroft84afd992008-07-23 21:27:32 -07002458
Joonsoo Kim4e35f482014-04-03 14:47:30 -07002459 start = vma_hugecache_offset(h, vma, vma->vm_start);
2460 end = vma_hugecache_offset(h, vma, vma->vm_end);
Andy Whitcroft84afd992008-07-23 21:27:32 -07002461
Joonsoo Kim4e35f482014-04-03 14:47:30 -07002462 reserve = (end - start) - region_count(resv, start, end);
Andy Whitcroft84afd992008-07-23 21:27:32 -07002463
Joonsoo Kim4e35f482014-04-03 14:47:30 -07002464 kref_put(&resv->refs, resv_map_release);
2465
2466 if (reserve) {
2467 hugetlb_acct_memory(h, -reserve);
2468 hugepage_subpool_put_pages(spool, reserve);
Andy Whitcroft84afd992008-07-23 21:27:32 -07002469 }
Mel Gormana1e78772008-07-23 21:27:23 -07002470}
2471
Linus Torvalds1da177e2005-04-16 15:20:36 -07002472/*
2473 * We cannot handle pagefaults against hugetlb pages at all. They cause
2474 * handle_mm_fault() to try to instantiate regular-sized pages in the
2475 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
2476 * this far.
2477 */
Nick Piggind0217ac2007-07-19 01:47:03 -07002478static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002479{
2480 BUG();
Nick Piggind0217ac2007-07-19 01:47:03 -07002481 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002482}
2483
Alexey Dobriyanf0f37e22009-09-27 22:29:37 +04002484const struct vm_operations_struct hugetlb_vm_ops = {
Nick Piggind0217ac2007-07-19 01:47:03 -07002485 .fault = hugetlb_vm_op_fault,
Andy Whitcroft84afd992008-07-23 21:27:32 -07002486 .open = hugetlb_vm_op_open,
Mel Gormana1e78772008-07-23 21:27:23 -07002487 .close = hugetlb_vm_op_close,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002488};
2489
David Gibson1e8f8892006-01-06 00:10:44 -08002490static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
2491 int writable)
David Gibson63551ae2005-06-21 17:14:44 -07002492{
2493 pte_t entry;
2494
David Gibson1e8f8892006-01-06 00:10:44 -08002495 if (writable) {
Gerald Schaefer106c9922013-04-29 15:07:23 -07002496 entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page,
2497 vma->vm_page_prot)));
David Gibson63551ae2005-06-21 17:14:44 -07002498 } else {
Gerald Schaefer106c9922013-04-29 15:07:23 -07002499 entry = huge_pte_wrprotect(mk_huge_pte(page,
2500 vma->vm_page_prot));
David Gibson63551ae2005-06-21 17:14:44 -07002501 }
2502 entry = pte_mkyoung(entry);
2503 entry = pte_mkhuge(entry);
Chris Metcalfd9ed9fa2012-04-01 14:01:34 -04002504 entry = arch_make_huge_pte(entry, vma, page, writable);
David Gibson63551ae2005-06-21 17:14:44 -07002505
2506 return entry;
2507}
2508
David Gibson1e8f8892006-01-06 00:10:44 -08002509static void set_huge_ptep_writable(struct vm_area_struct *vma,
2510 unsigned long address, pte_t *ptep)
2511{
2512 pte_t entry;
2513
Gerald Schaefer106c9922013-04-29 15:07:23 -07002514 entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep)));
Chris Forbes32f84522011-07-25 17:12:14 -07002515 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
Russell King4b3073e2009-12-18 16:40:18 +00002516 update_mmu_cache(vma, address, ptep);
David Gibson1e8f8892006-01-06 00:10:44 -08002517}
2518
Naoya Horiguchi4a705fe2014-06-23 13:22:03 -07002519static int is_hugetlb_entry_migration(pte_t pte)
2520{
2521 swp_entry_t swp;
2522
2523 if (huge_pte_none(pte) || pte_present(pte))
2524 return 0;
2525 swp = pte_to_swp_entry(pte);
2526 if (non_swap_entry(swp) && is_migration_entry(swp))
2527 return 1;
2528 else
2529 return 0;
2530}
2531
2532static int is_hugetlb_entry_hwpoisoned(pte_t pte)
2533{
2534 swp_entry_t swp;
2535
2536 if (huge_pte_none(pte) || pte_present(pte))
2537 return 0;
2538 swp = pte_to_swp_entry(pte);
2539 if (non_swap_entry(swp) && is_hwpoison_entry(swp))
2540 return 1;
2541 else
2542 return 0;
2543}
David Gibson1e8f8892006-01-06 00:10:44 -08002544
David Gibson63551ae2005-06-21 17:14:44 -07002545int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
2546 struct vm_area_struct *vma)
2547{
2548 pte_t *src_pte, *dst_pte, entry;
2549 struct page *ptepage;
Hugh Dickins1c598272005-10-19 21:23:43 -07002550 unsigned long addr;
David Gibson1e8f8892006-01-06 00:10:44 -08002551 int cow;
Andi Kleena5516432008-07-23 21:27:41 -07002552 struct hstate *h = hstate_vma(vma);
2553 unsigned long sz = huge_page_size(h);
Andreas Sandberge8569dd2014-01-21 15:49:09 -08002554 unsigned long mmun_start; /* For mmu_notifiers */
2555 unsigned long mmun_end; /* For mmu_notifiers */
2556 int ret = 0;
David Gibson1e8f8892006-01-06 00:10:44 -08002557
2558 cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
David Gibson63551ae2005-06-21 17:14:44 -07002559
Andreas Sandberge8569dd2014-01-21 15:49:09 -08002560 mmun_start = vma->vm_start;
2561 mmun_end = vma->vm_end;
2562 if (cow)
2563 mmu_notifier_invalidate_range_start(src, mmun_start, mmun_end);
2564
Andi Kleena5516432008-07-23 21:27:41 -07002565 for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002566 spinlock_t *src_ptl, *dst_ptl;
Hugh Dickinsc74df322005-10-29 18:16:23 -07002567 src_pte = huge_pte_offset(src, addr);
2568 if (!src_pte)
2569 continue;
Andi Kleena5516432008-07-23 21:27:41 -07002570 dst_pte = huge_pte_alloc(dst, addr, sz);
Andreas Sandberge8569dd2014-01-21 15:49:09 -08002571 if (!dst_pte) {
2572 ret = -ENOMEM;
2573 break;
2574 }
Larry Woodmanc5c99422008-01-24 05:49:25 -08002575
2576 /* If the pagetables are shared don't copy or take references */
2577 if (dst_pte == src_pte)
2578 continue;
2579
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002580 dst_ptl = huge_pte_lock(h, dst, dst_pte);
2581 src_ptl = huge_pte_lockptr(h, src, src_pte);
2582 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
Naoya Horiguchi4a705fe2014-06-23 13:22:03 -07002583 entry = huge_ptep_get(src_pte);
2584 if (huge_pte_none(entry)) { /* skip none entry */
2585 ;
2586 } else if (unlikely(is_hugetlb_entry_migration(entry) ||
2587 is_hugetlb_entry_hwpoisoned(entry))) {
2588 swp_entry_t swp_entry = pte_to_swp_entry(entry);
2589
2590 if (is_write_migration_entry(swp_entry) && cow) {
2591 /*
2592 * COW mappings require pages in both
2593 * parent and child to be set to read.
2594 */
2595 make_migration_entry_read(&swp_entry);
2596 entry = swp_entry_to_pte(swp_entry);
2597 set_huge_pte_at(src, addr, src_pte, entry);
2598 }
2599 set_huge_pte_at(dst, addr, dst_pte, entry);
2600 } else {
Joerg Roedel34ee6452014-11-13 13:46:09 +11002601 if (cow) {
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07002602 huge_ptep_set_wrprotect(src, addr, src_pte);
Joerg Roedel34ee6452014-11-13 13:46:09 +11002603 mmu_notifier_invalidate_range(src, mmun_start,
2604 mmun_end);
2605 }
Naoya Horiguchi0253d632014-07-23 14:00:19 -07002606 entry = huge_ptep_get(src_pte);
Hugh Dickins1c598272005-10-19 21:23:43 -07002607 ptepage = pte_page(entry);
2608 get_page(ptepage);
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002609 page_dup_rmap(ptepage);
Hugh Dickins1c598272005-10-19 21:23:43 -07002610 set_huge_pte_at(dst, addr, dst_pte, entry);
2611 }
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002612 spin_unlock(src_ptl);
2613 spin_unlock(dst_ptl);
David Gibson63551ae2005-06-21 17:14:44 -07002614 }
David Gibson63551ae2005-06-21 17:14:44 -07002615
Andreas Sandberge8569dd2014-01-21 15:49:09 -08002616 if (cow)
2617 mmu_notifier_invalidate_range_end(src, mmun_start, mmun_end);
2618
2619 return ret;
David Gibson63551ae2005-06-21 17:14:44 -07002620}
2621
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07002622void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
2623 unsigned long start, unsigned long end,
2624 struct page *ref_page)
David Gibson63551ae2005-06-21 17:14:44 -07002625{
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07002626 int force_flush = 0;
David Gibson63551ae2005-06-21 17:14:44 -07002627 struct mm_struct *mm = vma->vm_mm;
2628 unsigned long address;
David Gibsonc7546f82005-08-05 11:59:35 -07002629 pte_t *ptep;
David Gibson63551ae2005-06-21 17:14:44 -07002630 pte_t pte;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002631 spinlock_t *ptl;
David Gibson63551ae2005-06-21 17:14:44 -07002632 struct page *page;
Andi Kleena5516432008-07-23 21:27:41 -07002633 struct hstate *h = hstate_vma(vma);
2634 unsigned long sz = huge_page_size(h);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07002635 const unsigned long mmun_start = start; /* For mmu_notifiers */
2636 const unsigned long mmun_end = end; /* For mmu_notifiers */
Andi Kleena5516432008-07-23 21:27:41 -07002637
David Gibson63551ae2005-06-21 17:14:44 -07002638 WARN_ON(!is_vm_hugetlb_page(vma));
Andi Kleena5516432008-07-23 21:27:41 -07002639 BUG_ON(start & ~huge_page_mask(h));
2640 BUG_ON(end & ~huge_page_mask(h));
David Gibson63551ae2005-06-21 17:14:44 -07002641
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07002642 tlb_start_vma(tlb, vma);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07002643 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
Hillf Danton569f48b2014-12-10 15:44:41 -08002644 address = start;
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07002645again:
Hillf Danton569f48b2014-12-10 15:44:41 -08002646 for (; address < end; address += sz) {
David Gibsonc7546f82005-08-05 11:59:35 -07002647 ptep = huge_pte_offset(mm, address);
Adam Litke4c887262005-10-29 18:16:46 -07002648 if (!ptep)
David Gibsonc7546f82005-08-05 11:59:35 -07002649 continue;
2650
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002651 ptl = huge_pte_lock(h, mm, ptep);
Chen, Kenneth W39dde652006-12-06 20:32:03 -08002652 if (huge_pmd_unshare(mm, &address, ptep))
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002653 goto unlock;
Chen, Kenneth W39dde652006-12-06 20:32:03 -08002654
Hillf Danton66293262012-03-23 15:01:48 -07002655 pte = huge_ptep_get(ptep);
2656 if (huge_pte_none(pte))
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002657 goto unlock;
Hillf Danton66293262012-03-23 15:01:48 -07002658
2659 /*
2660 * HWPoisoned hugepage is already unmapped and dropped reference
2661 */
Naoya Horiguchi8c4894c2012-12-12 13:52:28 -08002662 if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) {
Gerald Schaefer106c9922013-04-29 15:07:23 -07002663 huge_pte_clear(mm, address, ptep);
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002664 goto unlock;
Naoya Horiguchi8c4894c2012-12-12 13:52:28 -08002665 }
Hillf Danton66293262012-03-23 15:01:48 -07002666
2667 page = pte_page(pte);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002668 /*
2669 * If a reference page is supplied, it is because a specific
2670 * page is being unmapped, not a range. Ensure the page we
2671 * are about to unmap is the actual page of interest.
2672 */
2673 if (ref_page) {
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002674 if (page != ref_page)
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002675 goto unlock;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002676
2677 /*
2678 * Mark the VMA as having unmapped its page so that
2679 * future faults in this VMA will fail rather than
2680 * looking like data was lost
2681 */
2682 set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
2683 }
2684
David Gibsonc7546f82005-08-05 11:59:35 -07002685 pte = huge_ptep_get_and_clear(mm, address, ptep);
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07002686 tlb_remove_tlb_entry(tlb, ptep, address);
Gerald Schaefer106c9922013-04-29 15:07:23 -07002687 if (huge_pte_dirty(pte))
Ken Chen6649a382007-02-08 14:20:27 -08002688 set_page_dirty(page);
Hillf Danton9e811302012-03-21 16:34:03 -07002689
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07002690 page_remove_rmap(page);
2691 force_flush = !__tlb_remove_page(tlb, page);
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002692 if (force_flush) {
Hillf Danton569f48b2014-12-10 15:44:41 -08002693 address += sz;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002694 spin_unlock(ptl);
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07002695 break;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002696 }
Hillf Danton9e811302012-03-21 16:34:03 -07002697 /* Bail out after unmapping reference page if supplied */
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002698 if (ref_page) {
2699 spin_unlock(ptl);
Hillf Danton9e811302012-03-21 16:34:03 -07002700 break;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002701 }
2702unlock:
2703 spin_unlock(ptl);
David Gibson63551ae2005-06-21 17:14:44 -07002704 }
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07002705 /*
2706 * mmu_gather ran out of room to batch pages, we break out of
2707 * the PTE lock to avoid doing the potential expensive TLB invalidate
2708 * and page-free while holding it.
2709 */
2710 if (force_flush) {
2711 force_flush = 0;
2712 tlb_flush_mmu(tlb);
2713 if (address < end && !ref_page)
2714 goto again;
Chen, Kenneth Wfe1668a2006-10-04 02:15:24 -07002715 }
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07002716 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07002717 tlb_end_vma(tlb, vma);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002718}
David Gibson63551ae2005-06-21 17:14:44 -07002719
Mel Gormand8333522012-07-31 16:46:20 -07002720void __unmap_hugepage_range_final(struct mmu_gather *tlb,
2721 struct vm_area_struct *vma, unsigned long start,
2722 unsigned long end, struct page *ref_page)
2723{
2724 __unmap_hugepage_range(tlb, vma, start, end, ref_page);
2725
2726 /*
2727 * Clear this flag so that x86's huge_pmd_share page_table_shareable
2728 * test will fail on a vma being torn down, and not grab a page table
2729 * on its way out. We're lucky that the flag has such an appropriate
2730 * name, and can in fact be safely cleared here. We could clear it
2731 * before the __unmap_hugepage_range above, but all that's necessary
Davidlohr Buesoc8c06ef2014-12-12 16:54:24 -08002732 * is to clear it before releasing the i_mmap_rwsem. This works
Mel Gormand8333522012-07-31 16:46:20 -07002733 * because in the context this is called, the VMA is about to be
Davidlohr Buesoc8c06ef2014-12-12 16:54:24 -08002734 * destroyed and the i_mmap_rwsem is held.
Mel Gormand8333522012-07-31 16:46:20 -07002735 */
2736 vma->vm_flags &= ~VM_MAYSHARE;
2737}
2738
Chen, Kenneth W502717f2006-10-11 01:20:46 -07002739void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002740 unsigned long end, struct page *ref_page)
Chen, Kenneth W502717f2006-10-11 01:20:46 -07002741{
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07002742 struct mm_struct *mm;
2743 struct mmu_gather tlb;
2744
2745 mm = vma->vm_mm;
2746
Linus Torvalds2b047252013-08-15 11:42:25 -07002747 tlb_gather_mmu(&tlb, mm, start, end);
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07002748 __unmap_hugepage_range(&tlb, vma, start, end, ref_page);
2749 tlb_finish_mmu(&tlb, start, end);
Chen, Kenneth W502717f2006-10-11 01:20:46 -07002750}
2751
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002752/*
2753 * This is called when the original mapper is failing to COW a MAP_PRIVATE
2754 * mappping it owns the reserve page for. The intention is to unmap the page
2755 * from other VMAs and let the children be SIGKILLed if they are faulting the
2756 * same region.
2757 */
Davidlohr Bueso2f4612a2014-08-06 16:06:45 -07002758static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
2759 struct page *page, unsigned long address)
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002760{
Adam Litke75266742008-11-12 13:24:56 -08002761 struct hstate *h = hstate_vma(vma);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002762 struct vm_area_struct *iter_vma;
2763 struct address_space *mapping;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002764 pgoff_t pgoff;
2765
2766 /*
2767 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
2768 * from page cache lookup which is in HPAGE_SIZE units.
2769 */
Adam Litke75266742008-11-12 13:24:56 -08002770 address = address & huge_page_mask(h);
Michal Hocko36e4f202012-10-08 16:33:31 -07002771 pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) +
2772 vma->vm_pgoff;
Al Viro496ad9a2013-01-23 17:07:38 -05002773 mapping = file_inode(vma->vm_file)->i_mapping;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002774
Mel Gorman4eb2b1d2009-12-14 17:59:53 -08002775 /*
2776 * Take the mapping lock for the duration of the table walk. As
2777 * this mapping should be shared between all the VMAs,
2778 * __unmap_hugepage_range() is called as the lock is already held
2779 */
Davidlohr Bueso83cde9e2014-12-12 16:54:21 -08002780 i_mmap_lock_write(mapping);
Michel Lespinasse6b2dbba2012-10-08 16:31:25 -07002781 vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) {
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002782 /* Do not unmap the current VMA */
2783 if (iter_vma == vma)
2784 continue;
2785
2786 /*
2787 * Unmap the page from other VMAs without their own reserves.
2788 * They get marked to be SIGKILLed if they fault in these
2789 * areas. This is because a future no-page fault on this VMA
2790 * could insert a zeroed page instead of the data existing
2791 * from the time of fork. This would look like data corruption
2792 */
2793 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07002794 unmap_hugepage_range(iter_vma, address,
2795 address + huge_page_size(h), page);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002796 }
Davidlohr Bueso83cde9e2014-12-12 16:54:21 -08002797 i_mmap_unlock_write(mapping);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002798}
2799
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002800/*
2801 * Hugetlb_cow() should be called with page lock of the original hugepage held.
Michal Hockoef009b22012-01-10 15:07:21 -08002802 * Called with hugetlb_instantiation_mutex held and pte_page locked so we
2803 * cannot race with other handlers or page migration.
2804 * Keep the pte_same checks anyway to make transition from the mutex easier.
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002805 */
David Gibson1e8f8892006-01-06 00:10:44 -08002806static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002807 unsigned long address, pte_t *ptep, pte_t pte,
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002808 struct page *pagecache_page, spinlock_t *ptl)
David Gibson1e8f8892006-01-06 00:10:44 -08002809{
Andi Kleena5516432008-07-23 21:27:41 -07002810 struct hstate *h = hstate_vma(vma);
David Gibson1e8f8892006-01-06 00:10:44 -08002811 struct page *old_page, *new_page;
Davidlohr Buesoad4404a2014-08-06 16:06:47 -07002812 int ret = 0, outside_reserve = 0;
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07002813 unsigned long mmun_start; /* For mmu_notifiers */
2814 unsigned long mmun_end; /* For mmu_notifiers */
David Gibson1e8f8892006-01-06 00:10:44 -08002815
2816 old_page = pte_page(pte);
2817
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002818retry_avoidcopy:
David Gibson1e8f8892006-01-06 00:10:44 -08002819 /* If no-one else is actually using this page, avoid the copy
2820 * and just make the page writable */
Joonsoo Kim37a21402013-09-11 14:21:04 -07002821 if (page_mapcount(old_page) == 1 && PageAnon(old_page)) {
2822 page_move_anon_rmap(old_page, vma, address);
David Gibson1e8f8892006-01-06 00:10:44 -08002823 set_huge_ptep_writable(vma, address, ptep);
Nick Piggin83c54072007-07-19 01:47:05 -07002824 return 0;
David Gibson1e8f8892006-01-06 00:10:44 -08002825 }
2826
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002827 /*
2828 * If the process that created a MAP_PRIVATE mapping is about to
2829 * perform a COW due to a shared page count, attempt to satisfy
2830 * the allocation without using the existing reserves. The pagecache
2831 * page is used to determine if the reserve at this address was
2832 * consumed or not. If reserves were used, a partial faulted mapping
2833 * at the time of fork() could consume its reserves on COW instead
2834 * of the full address range.
2835 */
Joonsoo Kim5944d012013-09-11 14:21:55 -07002836 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002837 old_page != pagecache_page)
2838 outside_reserve = 1;
2839
David Gibson1e8f8892006-01-06 00:10:44 -08002840 page_cache_get(old_page);
Larry Woodmanb76c8cf2009-12-14 17:59:37 -08002841
Davidlohr Buesoad4404a2014-08-06 16:06:47 -07002842 /*
2843 * Drop page table lock as buddy allocator may be called. It will
2844 * be acquired again before returning to the caller, as expected.
2845 */
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002846 spin_unlock(ptl);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002847 new_page = alloc_huge_page(vma, address, outside_reserve);
David Gibson1e8f8892006-01-06 00:10:44 -08002848
Adam Litke2fc39ce2007-11-14 16:59:39 -08002849 if (IS_ERR(new_page)) {
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002850 /*
2851 * If a process owning a MAP_PRIVATE mapping fails to COW,
2852 * it is due to references held by a child and an insufficient
2853 * huge page pool. To guarantee the original mappers
2854 * reliability, unmap the page from child processes. The child
2855 * may get SIGKILLed if it later faults.
2856 */
2857 if (outside_reserve) {
Davidlohr Buesoad4404a2014-08-06 16:06:47 -07002858 page_cache_release(old_page);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002859 BUG_ON(huge_pte_none(pte));
Davidlohr Bueso2f4612a2014-08-06 16:06:45 -07002860 unmap_ref_private(mm, vma, old_page, address);
2861 BUG_ON(huge_pte_none(pte));
2862 spin_lock(ptl);
2863 ptep = huge_pte_offset(mm, address & huge_page_mask(h));
2864 if (likely(ptep &&
2865 pte_same(huge_ptep_get(ptep), pte)))
2866 goto retry_avoidcopy;
2867 /*
2868 * race occurs while re-acquiring page table
2869 * lock, and our job is done.
2870 */
2871 return 0;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002872 }
2873
Davidlohr Buesoad4404a2014-08-06 16:06:47 -07002874 ret = (PTR_ERR(new_page) == -ENOMEM) ?
2875 VM_FAULT_OOM : VM_FAULT_SIGBUS;
2876 goto out_release_old;
David Gibson1e8f8892006-01-06 00:10:44 -08002877 }
2878
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002879 /*
2880 * When the original hugepage is shared one, it does not have
2881 * anon_vma prepared.
2882 */
Dean Nelson44e2aa92010-10-26 14:22:08 -07002883 if (unlikely(anon_vma_prepare(vma))) {
Davidlohr Buesoad4404a2014-08-06 16:06:47 -07002884 ret = VM_FAULT_OOM;
2885 goto out_release_all;
Dean Nelson44e2aa92010-10-26 14:22:08 -07002886 }
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002887
Andrea Arcangeli47ad8472011-01-13 15:46:47 -08002888 copy_user_huge_page(new_page, old_page, address, vma,
2889 pages_per_huge_page(h));
Nick Piggin0ed361d2008-02-04 22:29:34 -08002890 __SetPageUptodate(new_page);
David Gibson1e8f8892006-01-06 00:10:44 -08002891
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07002892 mmun_start = address & huge_page_mask(h);
2893 mmun_end = mmun_start + huge_page_size(h);
2894 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
Davidlohr Buesoad4404a2014-08-06 16:06:47 -07002895
Larry Woodmanb76c8cf2009-12-14 17:59:37 -08002896 /*
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002897 * Retake the page table lock to check for racing updates
Larry Woodmanb76c8cf2009-12-14 17:59:37 -08002898 * before the page tables are altered
2899 */
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002900 spin_lock(ptl);
Andi Kleena5516432008-07-23 21:27:41 -07002901 ptep = huge_pte_offset(mm, address & huge_page_mask(h));
Naoya Horiguchia9af0c52014-04-07 15:36:54 -07002902 if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) {
Joonsoo Kim07443a82013-09-11 14:21:58 -07002903 ClearPagePrivate(new_page);
2904
David Gibson1e8f8892006-01-06 00:10:44 -08002905 /* Break COW */
Gerald Schaefer8fe627e2008-04-28 02:13:28 -07002906 huge_ptep_clear_flush(vma, address, ptep);
Joerg Roedel34ee6452014-11-13 13:46:09 +11002907 mmu_notifier_invalidate_range(mm, mmun_start, mmun_end);
David Gibson1e8f8892006-01-06 00:10:44 -08002908 set_huge_pte_at(mm, address, ptep,
2909 make_huge_pte(vma, new_page, 1));
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002910 page_remove_rmap(old_page);
Naoya Horiguchicd67f0d2010-09-10 13:23:04 +09002911 hugepage_add_new_anon_rmap(new_page, vma, address);
David Gibson1e8f8892006-01-06 00:10:44 -08002912 /* Make the old page be freed below */
2913 new_page = old_page;
2914 }
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002915 spin_unlock(ptl);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07002916 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
Davidlohr Buesoad4404a2014-08-06 16:06:47 -07002917out_release_all:
David Gibson1e8f8892006-01-06 00:10:44 -08002918 page_cache_release(new_page);
Davidlohr Buesoad4404a2014-08-06 16:06:47 -07002919out_release_old:
David Gibson1e8f8892006-01-06 00:10:44 -08002920 page_cache_release(old_page);
Joonsoo Kim83120342013-09-11 14:21:57 -07002921
Davidlohr Buesoad4404a2014-08-06 16:06:47 -07002922 spin_lock(ptl); /* Caller expects lock to be held */
2923 return ret;
David Gibson1e8f8892006-01-06 00:10:44 -08002924}
2925
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002926/* Return the pagecache page at a given address within a VMA */
Andi Kleena5516432008-07-23 21:27:41 -07002927static struct page *hugetlbfs_pagecache_page(struct hstate *h,
2928 struct vm_area_struct *vma, unsigned long address)
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002929{
2930 struct address_space *mapping;
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -07002931 pgoff_t idx;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002932
2933 mapping = vma->vm_file->f_mapping;
Andi Kleena5516432008-07-23 21:27:41 -07002934 idx = vma_hugecache_offset(h, vma, address);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002935
2936 return find_lock_page(mapping, idx);
2937}
2938
Hugh Dickins3ae77f42009-09-21 17:03:33 -07002939/*
2940 * Return whether there is a pagecache page to back given address within VMA.
2941 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
2942 */
2943static bool hugetlbfs_pagecache_present(struct hstate *h,
Hugh Dickins2a15efc2009-09-21 17:03:27 -07002944 struct vm_area_struct *vma, unsigned long address)
2945{
2946 struct address_space *mapping;
2947 pgoff_t idx;
2948 struct page *page;
2949
2950 mapping = vma->vm_file->f_mapping;
2951 idx = vma_hugecache_offset(h, vma, address);
2952
2953 page = find_get_page(mapping, idx);
2954 if (page)
2955 put_page(page);
2956 return page != NULL;
2957}
2958
Robert P. J. Daya1ed3dd2007-07-17 04:03:33 -07002959static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
Davidlohr Bueso8382d912014-04-03 14:47:31 -07002960 struct address_space *mapping, pgoff_t idx,
2961 unsigned long address, pte_t *ptep, unsigned int flags)
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01002962{
Andi Kleena5516432008-07-23 21:27:41 -07002963 struct hstate *h = hstate_vma(vma);
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01002964 int ret = VM_FAULT_SIGBUS;
Hillf Danton409eb8c2012-01-20 14:34:13 -08002965 int anon_rmap = 0;
Adam Litke4c887262005-10-29 18:16:46 -07002966 unsigned long size;
Adam Litke4c887262005-10-29 18:16:46 -07002967 struct page *page;
David Gibson1e8f8892006-01-06 00:10:44 -08002968 pte_t new_pte;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002969 spinlock_t *ptl;
Adam Litke4c887262005-10-29 18:16:46 -07002970
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002971 /*
2972 * Currently, we are forced to kill the process in the event the
2973 * original mapper has unmapped pages from the child due to a failed
Lucas De Marchi25985ed2011-03-30 22:57:33 -03002974 * COW. Warn that such a situation has occurred as it may not be obvious
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002975 */
2976 if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
Andrew Mortonffb22af2013-02-22 16:32:08 -08002977 pr_warning("PID %d killed due to inadequate hugepage pool\n",
2978 current->pid);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002979 return ret;
2980 }
2981
Adam Litke4c887262005-10-29 18:16:46 -07002982 /*
2983 * Use page lock to guard against racing truncation
2984 * before we get page_table_lock.
2985 */
Christoph Lameter6bda6662006-01-06 00:10:49 -08002986retry:
2987 page = find_lock_page(mapping, idx);
2988 if (!page) {
Andi Kleena5516432008-07-23 21:27:41 -07002989 size = i_size_read(mapping->host) >> huge_page_shift(h);
Hugh Dickinsebed4bf2006-10-28 10:38:43 -07002990 if (idx >= size)
2991 goto out;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002992 page = alloc_huge_page(vma, address, 0);
Adam Litke2fc39ce2007-11-14 16:59:39 -08002993 if (IS_ERR(page)) {
Aneesh Kumar K.V76dcee72012-07-31 16:41:57 -07002994 ret = PTR_ERR(page);
2995 if (ret == -ENOMEM)
2996 ret = VM_FAULT_OOM;
2997 else
2998 ret = VM_FAULT_SIGBUS;
Christoph Lameter6bda6662006-01-06 00:10:49 -08002999 goto out;
3000 }
Andrea Arcangeli47ad8472011-01-13 15:46:47 -08003001 clear_huge_page(page, address, pages_per_huge_page(h));
Nick Piggin0ed361d2008-02-04 22:29:34 -08003002 __SetPageUptodate(page);
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01003003
Mel Gormanf83a2752009-05-28 14:34:40 -07003004 if (vma->vm_flags & VM_MAYSHARE) {
Christoph Lameter6bda6662006-01-06 00:10:49 -08003005 int err;
Ken Chen45c682a2007-11-14 16:59:44 -08003006 struct inode *inode = mapping->host;
Christoph Lameter6bda6662006-01-06 00:10:49 -08003007
3008 err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
3009 if (err) {
3010 put_page(page);
Christoph Lameter6bda6662006-01-06 00:10:49 -08003011 if (err == -EEXIST)
3012 goto retry;
3013 goto out;
3014 }
Joonsoo Kim07443a82013-09-11 14:21:58 -07003015 ClearPagePrivate(page);
Ken Chen45c682a2007-11-14 16:59:44 -08003016
3017 spin_lock(&inode->i_lock);
Andi Kleena5516432008-07-23 21:27:41 -07003018 inode->i_blocks += blocks_per_huge_page(h);
Ken Chen45c682a2007-11-14 16:59:44 -08003019 spin_unlock(&inode->i_lock);
Mel Gorman23be7462010-04-23 13:17:56 -04003020 } else {
Christoph Lameter6bda6662006-01-06 00:10:49 -08003021 lock_page(page);
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09003022 if (unlikely(anon_vma_prepare(vma))) {
3023 ret = VM_FAULT_OOM;
3024 goto backout_unlocked;
3025 }
Hillf Danton409eb8c2012-01-20 14:34:13 -08003026 anon_rmap = 1;
Mel Gorman23be7462010-04-23 13:17:56 -04003027 }
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09003028 } else {
Naoya Horiguchi998b4382010-09-08 10:19:32 +09003029 /*
3030 * If memory error occurs between mmap() and fault, some process
3031 * don't have hwpoisoned swap entry for errored virtual address.
3032 * So we need to block hugepage fault by PG_hwpoison bit check.
3033 */
3034 if (unlikely(PageHWPoison(page))) {
Chris Forbes32f84522011-07-25 17:12:14 -07003035 ret = VM_FAULT_HWPOISON |
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07003036 VM_FAULT_SET_HINDEX(hstate_index(h));
Naoya Horiguchi998b4382010-09-08 10:19:32 +09003037 goto backout_unlocked;
3038 }
Christoph Lameter6bda6662006-01-06 00:10:49 -08003039 }
David Gibson1e8f8892006-01-06 00:10:44 -08003040
Andy Whitcroft57303d82008-08-12 15:08:47 -07003041 /*
3042 * If we are going to COW a private mapping later, we examine the
3043 * pending reservations for this page now. This will ensure that
3044 * any allocations necessary to record that reservation occur outside
3045 * the spinlock.
3046 */
Hugh Dickins788c7df2009-06-23 13:49:05 +01003047 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED))
Andy Whitcroft2b267362008-08-12 15:08:49 -07003048 if (vma_needs_reservation(h, vma, address) < 0) {
3049 ret = VM_FAULT_OOM;
3050 goto backout_unlocked;
3051 }
Andy Whitcroft57303d82008-08-12 15:08:47 -07003052
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003053 ptl = huge_pte_lockptr(h, mm, ptep);
3054 spin_lock(ptl);
Andi Kleena5516432008-07-23 21:27:41 -07003055 size = i_size_read(mapping->host) >> huge_page_shift(h);
Adam Litke4c887262005-10-29 18:16:46 -07003056 if (idx >= size)
3057 goto backout;
3058
Nick Piggin83c54072007-07-19 01:47:05 -07003059 ret = 0;
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07003060 if (!huge_pte_none(huge_ptep_get(ptep)))
Adam Litke4c887262005-10-29 18:16:46 -07003061 goto backout;
3062
Joonsoo Kim07443a82013-09-11 14:21:58 -07003063 if (anon_rmap) {
3064 ClearPagePrivate(page);
Hillf Danton409eb8c2012-01-20 14:34:13 -08003065 hugepage_add_new_anon_rmap(page, vma, address);
Choi Gi-yongac714902014-04-07 15:37:36 -07003066 } else
Hillf Danton409eb8c2012-01-20 14:34:13 -08003067 page_dup_rmap(page);
David Gibson1e8f8892006-01-06 00:10:44 -08003068 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
3069 && (vma->vm_flags & VM_SHARED)));
3070 set_huge_pte_at(mm, address, ptep, new_pte);
3071
Hugh Dickins788c7df2009-06-23 13:49:05 +01003072 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
David Gibson1e8f8892006-01-06 00:10:44 -08003073 /* Optimization, do the COW without a second fault */
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003074 ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page, ptl);
David Gibson1e8f8892006-01-06 00:10:44 -08003075 }
3076
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003077 spin_unlock(ptl);
Adam Litke4c887262005-10-29 18:16:46 -07003078 unlock_page(page);
3079out:
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01003080 return ret;
Adam Litke4c887262005-10-29 18:16:46 -07003081
3082backout:
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003083 spin_unlock(ptl);
Andy Whitcroft2b267362008-08-12 15:08:49 -07003084backout_unlocked:
Adam Litke4c887262005-10-29 18:16:46 -07003085 unlock_page(page);
3086 put_page(page);
3087 goto out;
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01003088}
3089
Davidlohr Bueso8382d912014-04-03 14:47:31 -07003090#ifdef CONFIG_SMP
3091static u32 fault_mutex_hash(struct hstate *h, struct mm_struct *mm,
3092 struct vm_area_struct *vma,
3093 struct address_space *mapping,
3094 pgoff_t idx, unsigned long address)
3095{
3096 unsigned long key[2];
3097 u32 hash;
3098
3099 if (vma->vm_flags & VM_SHARED) {
3100 key[0] = (unsigned long) mapping;
3101 key[1] = idx;
3102 } else {
3103 key[0] = (unsigned long) mm;
3104 key[1] = address >> huge_page_shift(h);
3105 }
3106
3107 hash = jhash2((u32 *)&key, sizeof(key)/sizeof(u32), 0);
3108
3109 return hash & (num_fault_mutexes - 1);
3110}
3111#else
3112/*
3113 * For uniprocesor systems we always use a single mutex, so just
3114 * return 0 and avoid the hashing overhead.
3115 */
3116static u32 fault_mutex_hash(struct hstate *h, struct mm_struct *mm,
3117 struct vm_area_struct *vma,
3118 struct address_space *mapping,
3119 pgoff_t idx, unsigned long address)
3120{
3121 return 0;
3122}
3123#endif
3124
Adam Litke86e52162006-01-06 00:10:43 -08003125int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
Hugh Dickins788c7df2009-06-23 13:49:05 +01003126 unsigned long address, unsigned int flags)
Adam Litke86e52162006-01-06 00:10:43 -08003127{
Davidlohr Bueso8382d912014-04-03 14:47:31 -07003128 pte_t *ptep, entry;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003129 spinlock_t *ptl;
David Gibson1e8f8892006-01-06 00:10:44 -08003130 int ret;
Davidlohr Bueso8382d912014-04-03 14:47:31 -07003131 u32 hash;
3132 pgoff_t idx;
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09003133 struct page *page = NULL;
Andy Whitcroft57303d82008-08-12 15:08:47 -07003134 struct page *pagecache_page = NULL;
Andi Kleena5516432008-07-23 21:27:41 -07003135 struct hstate *h = hstate_vma(vma);
Davidlohr Bueso8382d912014-04-03 14:47:31 -07003136 struct address_space *mapping;
Adam Litke86e52162006-01-06 00:10:43 -08003137
KAMEZAWA Hiroyuki1e16a532012-01-10 15:07:22 -08003138 address &= huge_page_mask(h);
3139
Naoya Horiguchifd6a03ed2010-05-28 09:29:21 +09003140 ptep = huge_pte_offset(mm, address);
3141 if (ptep) {
3142 entry = huge_ptep_get(ptep);
Naoya Horiguchi290408d2010-09-08 10:19:35 +09003143 if (unlikely(is_hugetlb_entry_migration(entry))) {
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003144 migration_entry_wait_huge(vma, mm, ptep);
Naoya Horiguchi290408d2010-09-08 10:19:35 +09003145 return 0;
3146 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
Chris Forbes32f84522011-07-25 17:12:14 -07003147 return VM_FAULT_HWPOISON_LARGE |
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07003148 VM_FAULT_SET_HINDEX(hstate_index(h));
Naoya Horiguchifd6a03ed2010-05-28 09:29:21 +09003149 }
3150
Andi Kleena5516432008-07-23 21:27:41 -07003151 ptep = huge_pte_alloc(mm, address, huge_page_size(h));
Adam Litke86e52162006-01-06 00:10:43 -08003152 if (!ptep)
3153 return VM_FAULT_OOM;
3154
Davidlohr Bueso8382d912014-04-03 14:47:31 -07003155 mapping = vma->vm_file->f_mapping;
3156 idx = vma_hugecache_offset(h, vma, address);
3157
David Gibson3935baa2006-03-22 00:08:53 -08003158 /*
3159 * Serialize hugepage allocation and instantiation, so that we don't
3160 * get spurious allocation failures if two CPUs race to instantiate
3161 * the same page in the page cache.
3162 */
Davidlohr Bueso8382d912014-04-03 14:47:31 -07003163 hash = fault_mutex_hash(h, mm, vma, mapping, idx, address);
3164 mutex_lock(&htlb_fault_mutex_table[hash]);
3165
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07003166 entry = huge_ptep_get(ptep);
3167 if (huge_pte_none(entry)) {
Davidlohr Bueso8382d912014-04-03 14:47:31 -07003168 ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags);
David Gibsonb4d1d992008-10-15 22:01:11 -07003169 goto out_mutex;
David Gibson3935baa2006-03-22 00:08:53 -08003170 }
Adam Litke86e52162006-01-06 00:10:43 -08003171
Nick Piggin83c54072007-07-19 01:47:05 -07003172 ret = 0;
David Gibson1e8f8892006-01-06 00:10:44 -08003173
Andy Whitcroft57303d82008-08-12 15:08:47 -07003174 /*
3175 * If we are going to COW the mapping later, we examine the pending
3176 * reservations for this page now. This will ensure that any
3177 * allocations necessary to record that reservation occur outside the
3178 * spinlock. For private mappings, we also lookup the pagecache
3179 * page now as it is used to determine if a reservation has been
3180 * consumed.
3181 */
Gerald Schaefer106c9922013-04-29 15:07:23 -07003182 if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) {
Andy Whitcroft2b267362008-08-12 15:08:49 -07003183 if (vma_needs_reservation(h, vma, address) < 0) {
3184 ret = VM_FAULT_OOM;
David Gibsonb4d1d992008-10-15 22:01:11 -07003185 goto out_mutex;
Andy Whitcroft2b267362008-08-12 15:08:49 -07003186 }
Andy Whitcroft57303d82008-08-12 15:08:47 -07003187
Mel Gormanf83a2752009-05-28 14:34:40 -07003188 if (!(vma->vm_flags & VM_MAYSHARE))
Andy Whitcroft57303d82008-08-12 15:08:47 -07003189 pagecache_page = hugetlbfs_pagecache_page(h,
3190 vma, address);
3191 }
3192
Naoya Horiguchi56c9cfb2010-09-10 13:23:04 +09003193 /*
3194 * hugetlb_cow() requires page locks of pte_page(entry) and
3195 * pagecache_page, so here we need take the former one
3196 * when page != pagecache_page or !pagecache_page.
3197 * Note that locking order is always pagecache_page -> page,
3198 * so no worry about deadlock.
3199 */
3200 page = pte_page(entry);
Chris Metcalf66aebce2012-04-12 12:49:15 -07003201 get_page(page);
Naoya Horiguchi56c9cfb2010-09-10 13:23:04 +09003202 if (page != pagecache_page)
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09003203 lock_page(page);
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09003204
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003205 ptl = huge_pte_lockptr(h, mm, ptep);
3206 spin_lock(ptl);
David Gibson1e8f8892006-01-06 00:10:44 -08003207 /* Check for a racing update before calling hugetlb_cow */
David Gibsonb4d1d992008-10-15 22:01:11 -07003208 if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003209 goto out_ptl;
David Gibsonb4d1d992008-10-15 22:01:11 -07003210
3211
Hugh Dickins788c7df2009-06-23 13:49:05 +01003212 if (flags & FAULT_FLAG_WRITE) {
Gerald Schaefer106c9922013-04-29 15:07:23 -07003213 if (!huge_pte_write(entry)) {
Andy Whitcroft57303d82008-08-12 15:08:47 -07003214 ret = hugetlb_cow(mm, vma, address, ptep, entry,
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003215 pagecache_page, ptl);
3216 goto out_ptl;
David Gibsonb4d1d992008-10-15 22:01:11 -07003217 }
Gerald Schaefer106c9922013-04-29 15:07:23 -07003218 entry = huge_pte_mkdirty(entry);
David Gibsonb4d1d992008-10-15 22:01:11 -07003219 }
3220 entry = pte_mkyoung(entry);
Hugh Dickins788c7df2009-06-23 13:49:05 +01003221 if (huge_ptep_set_access_flags(vma, address, ptep, entry,
3222 flags & FAULT_FLAG_WRITE))
Russell King4b3073e2009-12-18 16:40:18 +00003223 update_mmu_cache(vma, address, ptep);
David Gibsonb4d1d992008-10-15 22:01:11 -07003224
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003225out_ptl:
3226 spin_unlock(ptl);
Andy Whitcroft57303d82008-08-12 15:08:47 -07003227
3228 if (pagecache_page) {
3229 unlock_page(pagecache_page);
3230 put_page(pagecache_page);
3231 }
Dean Nelson1f64d692010-12-02 14:31:12 -08003232 if (page != pagecache_page)
3233 unlock_page(page);
Chris Metcalf66aebce2012-04-12 12:49:15 -07003234 put_page(page);
Andy Whitcroft57303d82008-08-12 15:08:47 -07003235
David Gibsonb4d1d992008-10-15 22:01:11 -07003236out_mutex:
Davidlohr Bueso8382d912014-04-03 14:47:31 -07003237 mutex_unlock(&htlb_fault_mutex_table[hash]);
David Gibson1e8f8892006-01-06 00:10:44 -08003238 return ret;
Adam Litke86e52162006-01-06 00:10:43 -08003239}
3240
Michel Lespinasse28a35712013-02-22 16:35:55 -08003241long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
3242 struct page **pages, struct vm_area_struct **vmas,
3243 unsigned long *position, unsigned long *nr_pages,
3244 long i, unsigned int flags)
David Gibson63551ae2005-06-21 17:14:44 -07003245{
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08003246 unsigned long pfn_offset;
3247 unsigned long vaddr = *position;
Michel Lespinasse28a35712013-02-22 16:35:55 -08003248 unsigned long remainder = *nr_pages;
Andi Kleena5516432008-07-23 21:27:41 -07003249 struct hstate *h = hstate_vma(vma);
David Gibson63551ae2005-06-21 17:14:44 -07003250
David Gibson63551ae2005-06-21 17:14:44 -07003251 while (vaddr < vma->vm_end && remainder) {
Adam Litke4c887262005-10-29 18:16:46 -07003252 pte_t *pte;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003253 spinlock_t *ptl = NULL;
Hugh Dickins2a15efc2009-09-21 17:03:27 -07003254 int absent;
Adam Litke4c887262005-10-29 18:16:46 -07003255 struct page *page;
3256
3257 /*
3258 * Some archs (sparc64, sh*) have multiple pte_ts to
Hugh Dickins2a15efc2009-09-21 17:03:27 -07003259 * each hugepage. We have to make sure we get the
Adam Litke4c887262005-10-29 18:16:46 -07003260 * first, for the page indexing below to work.
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003261 *
3262 * Note that page table lock is not held when pte is null.
Adam Litke4c887262005-10-29 18:16:46 -07003263 */
Andi Kleena5516432008-07-23 21:27:41 -07003264 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h));
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003265 if (pte)
3266 ptl = huge_pte_lock(h, mm, pte);
Hugh Dickins2a15efc2009-09-21 17:03:27 -07003267 absent = !pte || huge_pte_none(huge_ptep_get(pte));
Adam Litke4c887262005-10-29 18:16:46 -07003268
Hugh Dickins2a15efc2009-09-21 17:03:27 -07003269 /*
3270 * When coredumping, it suits get_dump_page if we just return
Hugh Dickins3ae77f42009-09-21 17:03:33 -07003271 * an error where there's an empty slot with no huge pagecache
3272 * to back it. This way, we avoid allocating a hugepage, and
3273 * the sparse dumpfile avoids allocating disk blocks, but its
3274 * huge holes still show up with zeroes where they need to be.
Hugh Dickins2a15efc2009-09-21 17:03:27 -07003275 */
Hugh Dickins3ae77f42009-09-21 17:03:33 -07003276 if (absent && (flags & FOLL_DUMP) &&
3277 !hugetlbfs_pagecache_present(h, vma, vaddr)) {
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003278 if (pte)
3279 spin_unlock(ptl);
Hugh Dickins2a15efc2009-09-21 17:03:27 -07003280 remainder = 0;
3281 break;
3282 }
3283
Naoya Horiguchi9cc3a5b2013-04-17 15:58:30 -07003284 /*
3285 * We need call hugetlb_fault for both hugepages under migration
3286 * (in which case hugetlb_fault waits for the migration,) and
3287 * hwpoisoned hugepages (in which case we need to prevent the
3288 * caller from accessing to them.) In order to do this, we use
3289 * here is_swap_pte instead of is_hugetlb_entry_migration and
3290 * is_hugetlb_entry_hwpoisoned. This is because it simply covers
3291 * both cases, and because we can't follow correct pages
3292 * directly from any kind of swap entries.
3293 */
3294 if (absent || is_swap_pte(huge_ptep_get(pte)) ||
Gerald Schaefer106c9922013-04-29 15:07:23 -07003295 ((flags & FOLL_WRITE) &&
3296 !huge_pte_write(huge_ptep_get(pte)))) {
Adam Litke4c887262005-10-29 18:16:46 -07003297 int ret;
3298
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003299 if (pte)
3300 spin_unlock(ptl);
Hugh Dickins2a15efc2009-09-21 17:03:27 -07003301 ret = hugetlb_fault(mm, vma, vaddr,
3302 (flags & FOLL_WRITE) ? FAULT_FLAG_WRITE : 0);
Adam Litkea89182c2007-08-22 14:01:51 -07003303 if (!(ret & VM_FAULT_ERROR))
Adam Litke4c887262005-10-29 18:16:46 -07003304 continue;
3305
3306 remainder = 0;
Adam Litke4c887262005-10-29 18:16:46 -07003307 break;
3308 }
David Gibson63551ae2005-06-21 17:14:44 -07003309
Andi Kleena5516432008-07-23 21:27:41 -07003310 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07003311 page = pte_page(huge_ptep_get(pte));
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08003312same_page:
Chen, Kenneth Wd6692182006-03-31 02:29:57 -08003313 if (pages) {
Hugh Dickins2a15efc2009-09-21 17:03:27 -07003314 pages[i] = mem_map_offset(page, pfn_offset);
Andrea Arcangelia0368d42014-01-21 15:48:49 -08003315 get_page_foll(pages[i]);
Chen, Kenneth Wd6692182006-03-31 02:29:57 -08003316 }
David Gibson63551ae2005-06-21 17:14:44 -07003317
3318 if (vmas)
3319 vmas[i] = vma;
3320
3321 vaddr += PAGE_SIZE;
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08003322 ++pfn_offset;
David Gibson63551ae2005-06-21 17:14:44 -07003323 --remainder;
3324 ++i;
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08003325 if (vaddr < vma->vm_end && remainder &&
Andi Kleena5516432008-07-23 21:27:41 -07003326 pfn_offset < pages_per_huge_page(h)) {
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08003327 /*
3328 * We use pfn_offset to avoid touching the pageframes
3329 * of this compound page.
3330 */
3331 goto same_page;
3332 }
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003333 spin_unlock(ptl);
David Gibson63551ae2005-06-21 17:14:44 -07003334 }
Michel Lespinasse28a35712013-02-22 16:35:55 -08003335 *nr_pages = remainder;
David Gibson63551ae2005-06-21 17:14:44 -07003336 *position = vaddr;
3337
Hugh Dickins2a15efc2009-09-21 17:03:27 -07003338 return i ? i : -EFAULT;
David Gibson63551ae2005-06-21 17:14:44 -07003339}
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003340
Peter Zijlstra7da4d642012-11-19 03:14:23 +01003341unsigned long hugetlb_change_protection(struct vm_area_struct *vma,
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003342 unsigned long address, unsigned long end, pgprot_t newprot)
3343{
3344 struct mm_struct *mm = vma->vm_mm;
3345 unsigned long start = address;
3346 pte_t *ptep;
3347 pte_t pte;
Andi Kleena5516432008-07-23 21:27:41 -07003348 struct hstate *h = hstate_vma(vma);
Peter Zijlstra7da4d642012-11-19 03:14:23 +01003349 unsigned long pages = 0;
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003350
3351 BUG_ON(address >= end);
3352 flush_cache_range(vma, address, end);
3353
Rik van Riela5338092014-04-07 15:36:57 -07003354 mmu_notifier_invalidate_range_start(mm, start, end);
Davidlohr Bueso83cde9e2014-12-12 16:54:21 -08003355 i_mmap_lock_write(vma->vm_file->f_mapping);
Andi Kleena5516432008-07-23 21:27:41 -07003356 for (; address < end; address += huge_page_size(h)) {
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003357 spinlock_t *ptl;
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003358 ptep = huge_pte_offset(mm, address);
3359 if (!ptep)
3360 continue;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003361 ptl = huge_pte_lock(h, mm, ptep);
Peter Zijlstra7da4d642012-11-19 03:14:23 +01003362 if (huge_pmd_unshare(mm, &address, ptep)) {
3363 pages++;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003364 spin_unlock(ptl);
Chen, Kenneth W39dde652006-12-06 20:32:03 -08003365 continue;
Peter Zijlstra7da4d642012-11-19 03:14:23 +01003366 }
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07003367 if (!huge_pte_none(huge_ptep_get(ptep))) {
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003368 pte = huge_ptep_get_and_clear(mm, address, ptep);
Gerald Schaefer106c9922013-04-29 15:07:23 -07003369 pte = pte_mkhuge(huge_pte_modify(pte, newprot));
Tony Lube7517d2013-02-04 14:28:46 -08003370 pte = arch_make_huge_pte(pte, vma, NULL, 0);
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003371 set_huge_pte_at(mm, address, ptep, pte);
Peter Zijlstra7da4d642012-11-19 03:14:23 +01003372 pages++;
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003373 }
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003374 spin_unlock(ptl);
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003375 }
Mel Gormand8333522012-07-31 16:46:20 -07003376 /*
Davidlohr Buesoc8c06ef2014-12-12 16:54:24 -08003377 * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare
Mel Gormand8333522012-07-31 16:46:20 -07003378 * may have cleared our pud entry and done put_page on the page table:
Davidlohr Buesoc8c06ef2014-12-12 16:54:24 -08003379 * once we release i_mmap_rwsem, another task can do the final put_page
Mel Gormand8333522012-07-31 16:46:20 -07003380 * and that page table be reused and filled with junk.
3381 */
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003382 flush_tlb_range(vma, start, end);
Joerg Roedel34ee6452014-11-13 13:46:09 +11003383 mmu_notifier_invalidate_range(mm, start, end);
Davidlohr Bueso83cde9e2014-12-12 16:54:21 -08003384 i_mmap_unlock_write(vma->vm_file->f_mapping);
Rik van Riela5338092014-04-07 15:36:57 -07003385 mmu_notifier_invalidate_range_end(mm, start, end);
Peter Zijlstra7da4d642012-11-19 03:14:23 +01003386
3387 return pages << h->order;
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003388}
3389
Mel Gormana1e78772008-07-23 21:27:23 -07003390int hugetlb_reserve_pages(struct inode *inode,
3391 long from, long to,
Mel Gorman5a6fe122009-02-10 14:02:27 +00003392 struct vm_area_struct *vma,
KOSAKI Motohiroca16d142011-05-26 19:16:19 +09003393 vm_flags_t vm_flags)
Adam Litkee4e574b2007-10-16 01:26:19 -07003394{
Mel Gorman17c9d122009-02-11 16:34:16 +00003395 long ret, chg;
Andi Kleena5516432008-07-23 21:27:41 -07003396 struct hstate *h = hstate_inode(inode);
David Gibson90481622012-03-21 16:34:12 -07003397 struct hugepage_subpool *spool = subpool_inode(inode);
Joonsoo Kim9119a412014-04-03 14:47:25 -07003398 struct resv_map *resv_map;
Adam Litkee4e574b2007-10-16 01:26:19 -07003399
Mel Gormana1e78772008-07-23 21:27:23 -07003400 /*
Mel Gorman17c9d122009-02-11 16:34:16 +00003401 * Only apply hugepage reservation if asked. At fault time, an
3402 * attempt will be made for VM_NORESERVE to allocate a page
David Gibson90481622012-03-21 16:34:12 -07003403 * without using reserves
Mel Gorman17c9d122009-02-11 16:34:16 +00003404 */
KOSAKI Motohiroca16d142011-05-26 19:16:19 +09003405 if (vm_flags & VM_NORESERVE)
Mel Gorman17c9d122009-02-11 16:34:16 +00003406 return 0;
3407
3408 /*
Mel Gormana1e78772008-07-23 21:27:23 -07003409 * Shared mappings base their reservation on the number of pages that
3410 * are already allocated on behalf of the file. Private mappings need
3411 * to reserve the full area even if read-only as mprotect() may be
3412 * called to make the mapping read-write. Assume !vma is a shm mapping
3413 */
Joonsoo Kim9119a412014-04-03 14:47:25 -07003414 if (!vma || vma->vm_flags & VM_MAYSHARE) {
Joonsoo Kim4e35f482014-04-03 14:47:30 -07003415 resv_map = inode_resv_map(inode);
Joonsoo Kim9119a412014-04-03 14:47:25 -07003416
Joonsoo Kim1406ec92014-04-03 14:47:26 -07003417 chg = region_chg(resv_map, from, to);
Joonsoo Kim9119a412014-04-03 14:47:25 -07003418
3419 } else {
3420 resv_map = resv_map_alloc();
Mel Gorman5a6fe122009-02-10 14:02:27 +00003421 if (!resv_map)
3422 return -ENOMEM;
3423
Mel Gorman17c9d122009-02-11 16:34:16 +00003424 chg = to - from;
3425
Mel Gorman5a6fe122009-02-10 14:02:27 +00003426 set_vma_resv_map(vma, resv_map);
3427 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
3428 }
3429
Dave Hansenc50ac052012-05-29 15:06:46 -07003430 if (chg < 0) {
3431 ret = chg;
3432 goto out_err;
3433 }
Mel Gorman17c9d122009-02-11 16:34:16 +00003434
David Gibson90481622012-03-21 16:34:12 -07003435 /* There must be enough pages in the subpool for the mapping */
Dave Hansenc50ac052012-05-29 15:06:46 -07003436 if (hugepage_subpool_get_pages(spool, chg)) {
3437 ret = -ENOSPC;
3438 goto out_err;
3439 }
Mel Gorman17c9d122009-02-11 16:34:16 +00003440
3441 /*
3442 * Check enough hugepages are available for the reservation.
David Gibson90481622012-03-21 16:34:12 -07003443 * Hand the pages back to the subpool if there are not
Mel Gorman17c9d122009-02-11 16:34:16 +00003444 */
3445 ret = hugetlb_acct_memory(h, chg);
3446 if (ret < 0) {
David Gibson90481622012-03-21 16:34:12 -07003447 hugepage_subpool_put_pages(spool, chg);
Dave Hansenc50ac052012-05-29 15:06:46 -07003448 goto out_err;
Mel Gorman17c9d122009-02-11 16:34:16 +00003449 }
3450
3451 /*
3452 * Account for the reservations made. Shared mappings record regions
3453 * that have reservations as they are shared by multiple VMAs.
3454 * When the last VMA disappears, the region map says how much
3455 * the reservation was and the page cache tells how much of
3456 * the reservation was consumed. Private mappings are per-VMA and
3457 * only the consumed reservations are tracked. When the VMA
3458 * disappears, the original reservation is the VMA size and the
3459 * consumed reservations are stored in the map. Hence, nothing
3460 * else has to be done for private mappings here
3461 */
Mel Gormanf83a2752009-05-28 14:34:40 -07003462 if (!vma || vma->vm_flags & VM_MAYSHARE)
Joonsoo Kim1406ec92014-04-03 14:47:26 -07003463 region_add(resv_map, from, to);
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -07003464 return 0;
Dave Hansenc50ac052012-05-29 15:06:46 -07003465out_err:
Joonsoo Kimf031dd22014-04-03 14:47:28 -07003466 if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
3467 kref_put(&resv_map->refs, resv_map_release);
Dave Hansenc50ac052012-05-29 15:06:46 -07003468 return ret;
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -07003469}
3470
3471void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed)
3472{
Andi Kleena5516432008-07-23 21:27:41 -07003473 struct hstate *h = hstate_inode(inode);
Joonsoo Kim4e35f482014-04-03 14:47:30 -07003474 struct resv_map *resv_map = inode_resv_map(inode);
Joonsoo Kim9119a412014-04-03 14:47:25 -07003475 long chg = 0;
David Gibson90481622012-03-21 16:34:12 -07003476 struct hugepage_subpool *spool = subpool_inode(inode);
Ken Chen45c682a2007-11-14 16:59:44 -08003477
Joonsoo Kim9119a412014-04-03 14:47:25 -07003478 if (resv_map)
Joonsoo Kim1406ec92014-04-03 14:47:26 -07003479 chg = region_truncate(resv_map, offset);
Ken Chen45c682a2007-11-14 16:59:44 -08003480 spin_lock(&inode->i_lock);
Eric Sandeene4c6f8b2009-07-29 15:02:16 -07003481 inode->i_blocks -= (blocks_per_huge_page(h) * freed);
Ken Chen45c682a2007-11-14 16:59:44 -08003482 spin_unlock(&inode->i_lock);
3483
David Gibson90481622012-03-21 16:34:12 -07003484 hugepage_subpool_put_pages(spool, (chg - freed));
Andi Kleena5516432008-07-23 21:27:41 -07003485 hugetlb_acct_memory(h, -(chg - freed));
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -07003486}
Naoya Horiguchi93f70f92010-05-28 09:29:20 +09003487
Steve Capper3212b532013-04-23 12:35:02 +01003488#ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
3489static unsigned long page_table_shareable(struct vm_area_struct *svma,
3490 struct vm_area_struct *vma,
3491 unsigned long addr, pgoff_t idx)
3492{
3493 unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) +
3494 svma->vm_start;
3495 unsigned long sbase = saddr & PUD_MASK;
3496 unsigned long s_end = sbase + PUD_SIZE;
3497
3498 /* Allow segments to share if only one is marked locked */
3499 unsigned long vm_flags = vma->vm_flags & ~VM_LOCKED;
3500 unsigned long svm_flags = svma->vm_flags & ~VM_LOCKED;
3501
3502 /*
3503 * match the virtual addresses, permission and the alignment of the
3504 * page table page.
3505 */
3506 if (pmd_index(addr) != pmd_index(saddr) ||
3507 vm_flags != svm_flags ||
3508 sbase < svma->vm_start || svma->vm_end < s_end)
3509 return 0;
3510
3511 return saddr;
3512}
3513
3514static int vma_shareable(struct vm_area_struct *vma, unsigned long addr)
3515{
3516 unsigned long base = addr & PUD_MASK;
3517 unsigned long end = base + PUD_SIZE;
3518
3519 /*
3520 * check on proper vm_flags and page table alignment
3521 */
3522 if (vma->vm_flags & VM_MAYSHARE &&
3523 vma->vm_start <= base && end <= vma->vm_end)
3524 return 1;
3525 return 0;
3526}
3527
3528/*
3529 * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
3530 * and returns the corresponding pte. While this is not necessary for the
3531 * !shared pmd case because we can allocate the pmd later as well, it makes the
3532 * code much cleaner. pmd allocation is essential for the shared case because
Davidlohr Buesoc8c06ef2014-12-12 16:54:24 -08003533 * pud has to be populated inside the same i_mmap_rwsem section - otherwise
Steve Capper3212b532013-04-23 12:35:02 +01003534 * racing tasks could either miss the sharing (see huge_pte_offset) or select a
3535 * bad pmd for sharing.
3536 */
3537pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
3538{
3539 struct vm_area_struct *vma = find_vma(mm, addr);
3540 struct address_space *mapping = vma->vm_file->f_mapping;
3541 pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) +
3542 vma->vm_pgoff;
3543 struct vm_area_struct *svma;
3544 unsigned long saddr;
3545 pte_t *spte = NULL;
3546 pte_t *pte;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003547 spinlock_t *ptl;
Steve Capper3212b532013-04-23 12:35:02 +01003548
3549 if (!vma_shareable(vma, addr))
3550 return (pte_t *)pmd_alloc(mm, pud, addr);
3551
Davidlohr Bueso83cde9e2014-12-12 16:54:21 -08003552 i_mmap_lock_write(mapping);
Steve Capper3212b532013-04-23 12:35:02 +01003553 vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) {
3554 if (svma == vma)
3555 continue;
3556
3557 saddr = page_table_shareable(svma, vma, addr, idx);
3558 if (saddr) {
3559 spte = huge_pte_offset(svma->vm_mm, saddr);
3560 if (spte) {
3561 get_page(virt_to_page(spte));
3562 break;
3563 }
3564 }
3565 }
3566
3567 if (!spte)
3568 goto out;
3569
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003570 ptl = huge_pte_lockptr(hstate_vma(vma), mm, spte);
3571 spin_lock(ptl);
Steve Capper3212b532013-04-23 12:35:02 +01003572 if (pud_none(*pud))
3573 pud_populate(mm, pud,
3574 (pmd_t *)((unsigned long)spte & PAGE_MASK));
3575 else
3576 put_page(virt_to_page(spte));
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003577 spin_unlock(ptl);
Steve Capper3212b532013-04-23 12:35:02 +01003578out:
3579 pte = (pte_t *)pmd_alloc(mm, pud, addr);
Davidlohr Bueso83cde9e2014-12-12 16:54:21 -08003580 i_mmap_unlock_write(mapping);
Steve Capper3212b532013-04-23 12:35:02 +01003581 return pte;
3582}
3583
3584/*
3585 * unmap huge page backed by shared pte.
3586 *
3587 * Hugetlb pte page is ref counted at the time of mapping. If pte is shared
3588 * indicated by page_count > 1, unmap is achieved by clearing pud and
3589 * decrementing the ref count. If count == 1, the pte page is not shared.
3590 *
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003591 * called with page table lock held.
Steve Capper3212b532013-04-23 12:35:02 +01003592 *
3593 * returns: 1 successfully unmapped a shared pte page
3594 * 0 the underlying pte page is not shared, or it is the last user
3595 */
3596int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
3597{
3598 pgd_t *pgd = pgd_offset(mm, *addr);
3599 pud_t *pud = pud_offset(pgd, *addr);
3600
3601 BUG_ON(page_count(virt_to_page(ptep)) == 0);
3602 if (page_count(virt_to_page(ptep)) == 1)
3603 return 0;
3604
3605 pud_clear(pud);
3606 put_page(virt_to_page(ptep));
3607 *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE;
3608 return 1;
3609}
Steve Capper9e5fc742013-04-30 08:02:03 +01003610#define want_pmd_share() (1)
3611#else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
3612pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
3613{
3614 return NULL;
3615}
3616#define want_pmd_share() (0)
Steve Capper3212b532013-04-23 12:35:02 +01003617#endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
3618
Steve Capper9e5fc742013-04-30 08:02:03 +01003619#ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
3620pte_t *huge_pte_alloc(struct mm_struct *mm,
3621 unsigned long addr, unsigned long sz)
3622{
3623 pgd_t *pgd;
3624 pud_t *pud;
3625 pte_t *pte = NULL;
3626
3627 pgd = pgd_offset(mm, addr);
3628 pud = pud_alloc(mm, pgd, addr);
3629 if (pud) {
3630 if (sz == PUD_SIZE) {
3631 pte = (pte_t *)pud;
3632 } else {
3633 BUG_ON(sz != PMD_SIZE);
3634 if (want_pmd_share() && pud_none(*pud))
3635 pte = huge_pmd_share(mm, addr, pud);
3636 else
3637 pte = (pte_t *)pmd_alloc(mm, pud, addr);
3638 }
3639 }
3640 BUG_ON(pte && !pte_none(*pte) && !pte_huge(*pte));
3641
3642 return pte;
3643}
3644
3645pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr)
3646{
3647 pgd_t *pgd;
3648 pud_t *pud;
3649 pmd_t *pmd = NULL;
3650
3651 pgd = pgd_offset(mm, addr);
3652 if (pgd_present(*pgd)) {
3653 pud = pud_offset(pgd, addr);
3654 if (pud_present(*pud)) {
3655 if (pud_huge(*pud))
3656 return (pte_t *)pud;
3657 pmd = pmd_offset(pud, addr);
3658 }
3659 }
3660 return (pte_t *) pmd;
3661}
3662
3663struct page *
3664follow_huge_pmd(struct mm_struct *mm, unsigned long address,
3665 pmd_t *pmd, int write)
3666{
3667 struct page *page;
3668
3669 page = pte_page(*(pte_t *)pmd);
3670 if (page)
3671 page += ((address & ~PMD_MASK) >> PAGE_SHIFT);
3672 return page;
3673}
3674
3675struct page *
3676follow_huge_pud(struct mm_struct *mm, unsigned long address,
3677 pud_t *pud, int write)
3678{
3679 struct page *page;
3680
3681 page = pte_page(*(pte_t *)pud);
3682 if (page)
3683 page += ((address & ~PUD_MASK) >> PAGE_SHIFT);
3684 return page;
3685}
3686
3687#else /* !CONFIG_ARCH_WANT_GENERAL_HUGETLB */
3688
3689/* Can be overriden by architectures */
Gideon Israel Dsouza3b321232014-04-07 15:37:26 -07003690struct page * __weak
Steve Capper9e5fc742013-04-30 08:02:03 +01003691follow_huge_pud(struct mm_struct *mm, unsigned long address,
3692 pud_t *pud, int write)
3693{
3694 BUG();
3695 return NULL;
3696}
3697
3698#endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
3699
Andi Kleend5bd9102010-09-27 09:00:12 +02003700#ifdef CONFIG_MEMORY_FAILURE
3701
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +09003702/* Should be called in hugetlb_lock */
3703static int is_hugepage_on_freelist(struct page *hpage)
3704{
3705 struct page *page;
3706 struct page *tmp;
3707 struct hstate *h = page_hstate(hpage);
3708 int nid = page_to_nid(hpage);
3709
3710 list_for_each_entry_safe(page, tmp, &h->hugepage_freelists[nid], lru)
3711 if (page == hpage)
3712 return 1;
3713 return 0;
3714}
3715
Naoya Horiguchi93f70f92010-05-28 09:29:20 +09003716/*
3717 * This function is called from memory failure code.
3718 * Assume the caller holds page lock of the head page.
3719 */
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +09003720int dequeue_hwpoisoned_huge_page(struct page *hpage)
Naoya Horiguchi93f70f92010-05-28 09:29:20 +09003721{
3722 struct hstate *h = page_hstate(hpage);
3723 int nid = page_to_nid(hpage);
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +09003724 int ret = -EBUSY;
Naoya Horiguchi93f70f92010-05-28 09:29:20 +09003725
3726 spin_lock(&hugetlb_lock);
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +09003727 if (is_hugepage_on_freelist(hpage)) {
Naoya Horiguchi56f2fb12012-12-12 13:52:33 -08003728 /*
3729 * Hwpoisoned hugepage isn't linked to activelist or freelist,
3730 * but dangling hpage->lru can trigger list-debug warnings
3731 * (this happens when we call unpoison_memory() on it),
3732 * so let it point to itself with list_del_init().
3733 */
3734 list_del_init(&hpage->lru);
Naoya Horiguchi8c6c2ec2010-09-08 10:19:38 +09003735 set_page_refcounted(hpage);
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +09003736 h->free_huge_pages--;
3737 h->free_huge_pages_node[nid]--;
3738 ret = 0;
3739 }
Naoya Horiguchi93f70f92010-05-28 09:29:20 +09003740 spin_unlock(&hugetlb_lock);
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +09003741 return ret;
Naoya Horiguchi93f70f92010-05-28 09:29:20 +09003742}
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +09003743#endif
Naoya Horiguchi31caf662013-09-11 14:21:59 -07003744
3745bool isolate_huge_page(struct page *page, struct list_head *list)
3746{
Sasha Levin309381fea2014-01-23 15:52:54 -08003747 VM_BUG_ON_PAGE(!PageHead(page), page);
Naoya Horiguchi31caf662013-09-11 14:21:59 -07003748 if (!get_page_unless_zero(page))
3749 return false;
3750 spin_lock(&hugetlb_lock);
3751 list_move_tail(&page->lru, list);
3752 spin_unlock(&hugetlb_lock);
3753 return true;
3754}
3755
3756void putback_active_hugepage(struct page *page)
3757{
Sasha Levin309381fea2014-01-23 15:52:54 -08003758 VM_BUG_ON_PAGE(!PageHead(page), page);
Naoya Horiguchi31caf662013-09-11 14:21:59 -07003759 spin_lock(&hugetlb_lock);
3760 list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist);
3761 spin_unlock(&hugetlb_lock);
3762 put_page(page);
3763}
Naoya Horiguchic8721bb2013-09-11 14:22:09 -07003764
3765bool is_hugepage_active(struct page *page)
3766{
Sasha Levin309381fea2014-01-23 15:52:54 -08003767 VM_BUG_ON_PAGE(!PageHuge(page), page);
Naoya Horiguchic8721bb2013-09-11 14:22:09 -07003768 /*
3769 * This function can be called for a tail page because the caller,
3770 * scan_movable_pages, scans through a given pfn-range which typically
3771 * covers one memory block. In systems using gigantic hugepage (1GB
3772 * for x86_64,) a hugepage is larger than a memory block, and we don't
3773 * support migrating such large hugepages for now, so return false
3774 * when called for tail pages.
3775 */
3776 if (PageTail(page))
3777 return false;
3778 /*
3779 * Refcount of a hwpoisoned hugepages is 1, but they are not active,
3780 * so we should return false for them.
3781 */
3782 if (unlikely(PageHWPoison(page)))
3783 return false;
3784 return page_count(page) > 0;
3785}