blob: c41b2a0ee2736e4f7df74c440ceb90bd5fcceecb [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
2 * Generic hugetlb support.
Nadia Yvette Chambers6d49e352012-12-06 10:39:54 +01003 * (C) Nadia Yvette Chambers, April 2004
Linus Torvalds1da177e2005-04-16 15:20:36 -07004 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07005#include <linux/list.h>
6#include <linux/init.h>
7#include <linux/module.h>
8#include <linux/mm.h>
Alexey Dobriyane1759c22008-10-15 23:50:22 +04009#include <linux/seq_file.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070010#include <linux/sysctl.h>
11#include <linux/highmem.h>
Andrea Arcangelicddb8a52008-07-28 15:46:29 -070012#include <linux/mmu_notifier.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070013#include <linux/nodemask.h>
David Gibson63551ae2005-06-21 17:14:44 -070014#include <linux/pagemap.h>
Christoph Lameter5da7ca82006-01-06 00:10:46 -080015#include <linux/mempolicy.h>
Gideon Israel Dsouza3b321232014-04-07 15:37:26 -070016#include <linux/compiler.h>
Christoph Lameteraea47ff2006-01-08 01:00:57 -080017#include <linux/cpuset.h>
David Gibson3935baa2006-03-22 00:08:53 -080018#include <linux/mutex.h>
Andi Kleenaa888a72008-07-23 21:27:47 -070019#include <linux/bootmem.h>
Nishanth Aravamudana3437872008-07-23 21:27:44 -070020#include <linux/sysfs.h>
Tejun Heo5a0e3ad2010-03-24 17:04:11 +090021#include <linux/slab.h>
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +090022#include <linux/rmap.h>
Naoya Horiguchifd6a03e2010-05-28 09:29:21 +090023#include <linux/swap.h>
24#include <linux/swapops.h>
Naoya Horiguchic8721bb2013-09-11 14:22:09 -070025#include <linux/page-isolation.h>
Davidlohr Bueso8382d912014-04-03 14:47:31 -070026#include <linux/jhash.h>
Linus Torvaldsd6606682008-08-06 12:04:54 -070027
David Gibson63551ae2005-06-21 17:14:44 -070028#include <asm/page.h>
29#include <asm/pgtable.h>
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -070030#include <asm/tlb.h>
David Gibson63551ae2005-06-21 17:14:44 -070031
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -070032#include <linux/io.h>
David Gibson63551ae2005-06-21 17:14:44 -070033#include <linux/hugetlb.h>
Aneesh Kumar K.V9dd540e2012-07-31 16:42:15 -070034#include <linux/hugetlb_cgroup.h>
Lee Schermerhorn9a305232009-12-14 17:58:25 -080035#include <linux/node.h>
Nick Piggin7835e982006-03-22 00:08:40 -080036#include "internal.h"
Linus Torvalds1da177e2005-04-16 15:20:36 -070037
Andrey Ryabinin753162c2015-02-10 14:11:36 -080038int hugepages_treat_as_movable;
Andi Kleena5516432008-07-23 21:27:41 -070039
Aneesh Kumar K.Vc3f38a32012-07-31 16:42:10 -070040int hugetlb_max_hstate __read_mostly;
Andi Kleene5ff2152008-07-23 21:27:42 -070041unsigned int default_hstate_idx;
42struct hstate hstates[HUGE_MAX_HSTATE];
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)) {
Andrea Arcangelief5a22b2013-10-16 13:46:56 -0700920 /*
921 * For gigantic hugepages allocated through bootmem at
922 * boot, it's safer to be consistent with the not-gigantic
923 * hugepages and clear the PG_reserved bit from all tail pages
924 * too. Otherwse drivers using get_user_pages() to access tail
925 * pages may get the reference counting wrong if they see
926 * PG_reserved set on a tail page (despite the head page not
927 * having PG_reserved set). Enforcing this consistency between
928 * head and tail pages allows drivers to optimize away a check
929 * on the head page when they need know if put_page() is needed
930 * after get_user_pages().
931 */
932 __ClearPageReserved(p);
Youquan Song58a84aa2011-12-08 14:34:18 -0800933 set_page_count(p, 0);
Wu Fengguang20a03072009-06-16 15:32:22 -0700934 p->first_page = page;
David Rientjes44fc8052015-03-12 16:25:54 -0700935 /* Make sure p->first_page is always valid for PageTail() */
936 smp_wmb();
937 __SetPageTail(p);
Wu Fengguang20a03072009-06-16 15:32:22 -0700938 }
939}
940
Andrew Morton77959122012-10-08 16:34:11 -0700941/*
942 * PageHuge() only returns true for hugetlbfs pages, but not for normal or
943 * transparent huge pages. See the PageTransHuge() documentation for more
944 * details.
945 */
Wu Fengguang20a03072009-06-16 15:32:22 -0700946int PageHuge(struct page *page)
947{
Wu Fengguang20a03072009-06-16 15:32:22 -0700948 if (!PageCompound(page))
949 return 0;
950
951 page = compound_head(page);
Andrew Morton758f66a2014-01-21 15:48:57 -0800952 return get_compound_page_dtor(page) == free_huge_page;
Wu Fengguang20a03072009-06-16 15:32:22 -0700953}
Naoya Horiguchi43131e12010-05-28 09:29:22 +0900954EXPORT_SYMBOL_GPL(PageHuge);
955
Andrea Arcangeli27c73ae2013-11-21 14:32:02 -0800956/*
957 * PageHeadHuge() only returns true for hugetlbfs head page, but not for
958 * normal or transparent huge pages.
959 */
960int PageHeadHuge(struct page *page_head)
961{
Andrea Arcangeli27c73ae2013-11-21 14:32:02 -0800962 if (!PageHead(page_head))
963 return 0;
964
Andrew Morton758f66a2014-01-21 15:48:57 -0800965 return get_compound_page_dtor(page_head) == free_huge_page;
Andrea Arcangeli27c73ae2013-11-21 14:32:02 -0800966}
Andrea Arcangeli27c73ae2013-11-21 14:32:02 -0800967
Zhang Yi13d60f42013-06-25 21:19:31 +0800968pgoff_t __basepage_index(struct page *page)
969{
970 struct page *page_head = compound_head(page);
971 pgoff_t index = page_index(page_head);
972 unsigned long compound_idx;
973
974 if (!PageHuge(page_head))
975 return page_index(page);
976
977 if (compound_order(page_head) >= MAX_ORDER)
978 compound_idx = page_to_pfn(page) - page_to_pfn(page_head);
979 else
980 compound_idx = page - page_head;
981
982 return (index << compound_order(page_head)) + compound_idx;
983}
984
Andi Kleena5516432008-07-23 21:27:41 -0700985static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700986{
Linus Torvalds1da177e2005-04-16 15:20:36 -0700987 struct page *page;
Joe Jinf96efd52007-07-15 23:38:12 -0700988
Mel Gorman6484eb32009-06-16 15:31:54 -0700989 page = alloc_pages_exact_node(nid,
Naoya Horiguchi86cdb462013-09-11 14:22:13 -0700990 htlb_alloc_mask(h)|__GFP_COMP|__GFP_THISNODE|
Nishanth Aravamudan551883a2008-04-29 00:58:26 -0700991 __GFP_REPEAT|__GFP_NOWARN,
Andi Kleena5516432008-07-23 21:27:41 -0700992 huge_page_order(h));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700993 if (page) {
Gerald Schaefer7f2e9522008-04-28 02:13:29 -0700994 if (arch_prepare_hugepage(page)) {
Gerald Schaefercaff3a22008-08-12 15:08:38 -0700995 __free_pages(page, huge_page_order(h));
Harvey Harrison7b8ee842008-04-28 14:13:19 -0700996 return NULL;
Gerald Schaefer7f2e9522008-04-28 02:13:29 -0700997 }
Andi Kleena5516432008-07-23 21:27:41 -0700998 prep_new_huge_page(h, page, nid);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700999 }
Nishanth Aravamudan63b46132007-10-16 01:26:24 -07001000
1001 return page;
1002}
1003
Joonsoo Kimb2261022013-09-11 14:21:00 -07001004static int alloc_fresh_huge_page(struct hstate *h, nodemask_t *nodes_allowed)
1005{
1006 struct page *page;
1007 int nr_nodes, node;
1008 int ret = 0;
1009
1010 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
1011 page = alloc_fresh_huge_page_node(h, node);
1012 if (page) {
1013 ret = 1;
1014 break;
1015 }
1016 }
1017
1018 if (ret)
1019 count_vm_event(HTLB_BUDDY_PGALLOC);
1020 else
1021 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
1022
1023 return ret;
1024}
1025
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001026/*
1027 * Free huge page from pool from next node to free.
1028 * Attempt to keep persistent huge pages more or less
1029 * balanced over allowed nodes.
1030 * Called with hugetlb_lock locked.
1031 */
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001032static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
1033 bool acct_surplus)
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001034{
Joonsoo Kimb2261022013-09-11 14:21:00 -07001035 int nr_nodes, node;
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001036 int ret = 0;
1037
Joonsoo Kimb2261022013-09-11 14:21:00 -07001038 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
Lee Schermerhorn685f3452009-09-21 17:01:23 -07001039 /*
1040 * If we're returning unused surplus pages, only examine
1041 * nodes with surplus pages.
1042 */
Joonsoo Kimb2261022013-09-11 14:21:00 -07001043 if ((!acct_surplus || h->surplus_huge_pages_node[node]) &&
1044 !list_empty(&h->hugepage_freelists[node])) {
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001045 struct page *page =
Joonsoo Kimb2261022013-09-11 14:21:00 -07001046 list_entry(h->hugepage_freelists[node].next,
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001047 struct page, lru);
1048 list_del(&page->lru);
1049 h->free_huge_pages--;
Joonsoo Kimb2261022013-09-11 14:21:00 -07001050 h->free_huge_pages_node[node]--;
Lee Schermerhorn685f3452009-09-21 17:01:23 -07001051 if (acct_surplus) {
1052 h->surplus_huge_pages--;
Joonsoo Kimb2261022013-09-11 14:21:00 -07001053 h->surplus_huge_pages_node[node]--;
Lee Schermerhorn685f3452009-09-21 17:01:23 -07001054 }
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001055 update_and_free_page(h, page);
1056 ret = 1;
Lee Schermerhorn9a76db02009-12-14 17:58:15 -08001057 break;
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001058 }
Joonsoo Kimb2261022013-09-11 14:21:00 -07001059 }
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001060
1061 return ret;
1062}
1063
Naoya Horiguchic8721bb2013-09-11 14:22:09 -07001064/*
1065 * Dissolve a given free hugepage into free buddy pages. This function does
1066 * nothing for in-use (including surplus) hugepages.
1067 */
1068static void dissolve_free_huge_page(struct page *page)
1069{
1070 spin_lock(&hugetlb_lock);
1071 if (PageHuge(page) && !page_count(page)) {
1072 struct hstate *h = page_hstate(page);
1073 int nid = page_to_nid(page);
1074 list_del(&page->lru);
1075 h->free_huge_pages--;
1076 h->free_huge_pages_node[nid]--;
1077 update_and_free_page(h, page);
1078 }
1079 spin_unlock(&hugetlb_lock);
1080}
1081
1082/*
1083 * Dissolve free hugepages in a given pfn range. Used by memory hotplug to
1084 * make specified memory blocks removable from the system.
1085 * Note that start_pfn should aligned with (minimum) hugepage size.
1086 */
1087void dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn)
1088{
1089 unsigned int order = 8 * sizeof(void *);
1090 unsigned long pfn;
1091 struct hstate *h;
1092
Li Zhongd0177632014-08-06 16:07:56 -07001093 if (!hugepages_supported())
1094 return;
1095
Naoya Horiguchic8721bb2013-09-11 14:22:09 -07001096 /* Set scan step to minimum hugepage size */
1097 for_each_hstate(h)
1098 if (order > huge_page_order(h))
1099 order = huge_page_order(h);
1100 VM_BUG_ON(!IS_ALIGNED(start_pfn, 1 << order));
1101 for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << order)
1102 dissolve_free_huge_page(pfn_to_page(pfn));
1103}
1104
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001105static struct page *alloc_buddy_huge_page(struct hstate *h, int nid)
Adam Litke7893d1d2007-10-16 01:26:18 -07001106{
1107 struct page *page;
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001108 unsigned int r_nid;
Adam Litke7893d1d2007-10-16 01:26:18 -07001109
Luiz Capitulinobae7f4a2014-06-04 16:07:08 -07001110 if (hstate_is_gigantic(h))
Andi Kleenaa888a72008-07-23 21:27:47 -07001111 return NULL;
1112
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001113 /*
1114 * Assume we will successfully allocate the surplus page to
1115 * prevent racing processes from causing the surplus to exceed
1116 * overcommit
1117 *
1118 * This however introduces a different race, where a process B
1119 * tries to grow the static hugepage pool while alloc_pages() is
1120 * called by process A. B will only examine the per-node
1121 * counters in determining if surplus huge pages can be
1122 * converted to normal huge pages in adjust_pool_surplus(). A
1123 * won't be able to increment the per-node counter, until the
1124 * lock is dropped by B, but B doesn't drop hugetlb_lock until
1125 * no more huge pages can be converted from surplus to normal
1126 * state (and doesn't try to convert again). Thus, we have a
1127 * case where a surplus huge page exists, the pool is grown, and
1128 * the surplus huge page still exists after, even though it
1129 * should just have been converted to a normal huge page. This
1130 * does not leak memory, though, as the hugepage will be freed
1131 * once it is out of use. It also does not allow the counters to
1132 * go out of whack in adjust_pool_surplus() as we don't modify
1133 * the node values until we've gotten the hugepage and only the
1134 * per-node value is checked there.
1135 */
1136 spin_lock(&hugetlb_lock);
Andi Kleena5516432008-07-23 21:27:41 -07001137 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001138 spin_unlock(&hugetlb_lock);
1139 return NULL;
1140 } else {
Andi Kleena5516432008-07-23 21:27:41 -07001141 h->nr_huge_pages++;
1142 h->surplus_huge_pages++;
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001143 }
1144 spin_unlock(&hugetlb_lock);
1145
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001146 if (nid == NUMA_NO_NODE)
Naoya Horiguchi86cdb462013-09-11 14:22:13 -07001147 page = alloc_pages(htlb_alloc_mask(h)|__GFP_COMP|
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001148 __GFP_REPEAT|__GFP_NOWARN,
1149 huge_page_order(h));
1150 else
1151 page = alloc_pages_exact_node(nid,
Naoya Horiguchi86cdb462013-09-11 14:22:13 -07001152 htlb_alloc_mask(h)|__GFP_COMP|__GFP_THISNODE|
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001153 __GFP_REPEAT|__GFP_NOWARN, huge_page_order(h));
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001154
Gerald Schaefercaff3a22008-08-12 15:08:38 -07001155 if (page && arch_prepare_hugepage(page)) {
1156 __free_pages(page, huge_page_order(h));
Hillf Dantonea5768c2012-01-10 15:08:30 -08001157 page = NULL;
Gerald Schaefercaff3a22008-08-12 15:08:38 -07001158 }
1159
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001160 spin_lock(&hugetlb_lock);
Adam Litke7893d1d2007-10-16 01:26:18 -07001161 if (page) {
Aneesh Kumar K.V0edaecf2012-07-31 16:42:07 -07001162 INIT_LIST_HEAD(&page->lru);
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001163 r_nid = page_to_nid(page);
Adam Litke7893d1d2007-10-16 01:26:18 -07001164 set_compound_page_dtor(page, free_huge_page);
Aneesh Kumar K.V9dd540e2012-07-31 16:42:15 -07001165 set_hugetlb_cgroup(page, NULL);
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001166 /*
1167 * We incremented the global counters already
1168 */
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001169 h->nr_huge_pages_node[r_nid]++;
1170 h->surplus_huge_pages_node[r_nid]++;
Adam Litke3b116302008-04-28 02:13:06 -07001171 __count_vm_event(HTLB_BUDDY_PGALLOC);
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001172 } else {
Andi Kleena5516432008-07-23 21:27:41 -07001173 h->nr_huge_pages--;
1174 h->surplus_huge_pages--;
Adam Litke3b116302008-04-28 02:13:06 -07001175 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
Adam Litke7893d1d2007-10-16 01:26:18 -07001176 }
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001177 spin_unlock(&hugetlb_lock);
Adam Litke7893d1d2007-10-16 01:26:18 -07001178
1179 return page;
1180}
1181
Adam Litkee4e574b2007-10-16 01:26:19 -07001182/*
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001183 * This allocation function is useful in the context where vma is irrelevant.
1184 * E.g. soft-offlining uses this function because it only cares physical
1185 * address of error page.
1186 */
1187struct page *alloc_huge_page_node(struct hstate *h, int nid)
1188{
Joonsoo Kim4ef91842013-09-11 14:21:51 -07001189 struct page *page = NULL;
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001190
1191 spin_lock(&hugetlb_lock);
Joonsoo Kim4ef91842013-09-11 14:21:51 -07001192 if (h->free_huge_pages - h->resv_huge_pages > 0)
1193 page = dequeue_huge_page_node(h, nid);
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001194 spin_unlock(&hugetlb_lock);
1195
Aneesh Kumar K.V94ae8ba2012-07-31 16:42:35 -07001196 if (!page)
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001197 page = alloc_buddy_huge_page(h, nid);
1198
1199 return page;
1200}
1201
1202/*
Lucas De Marchi25985ed2011-03-30 22:57:33 -03001203 * Increase the hugetlb pool such that it can accommodate a reservation
Adam Litkee4e574b2007-10-16 01:26:19 -07001204 * of size 'delta'.
1205 */
Andi Kleena5516432008-07-23 21:27:41 -07001206static int gather_surplus_pages(struct hstate *h, int delta)
Adam Litkee4e574b2007-10-16 01:26:19 -07001207{
1208 struct list_head surplus_list;
1209 struct page *page, *tmp;
1210 int ret, i;
1211 int needed, allocated;
Hillf Danton28073b02012-03-21 16:34:00 -07001212 bool alloc_ok = true;
Adam Litkee4e574b2007-10-16 01:26:19 -07001213
Andi Kleena5516432008-07-23 21:27:41 -07001214 needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
Adam Litkeac09b3a2008-03-04 14:29:38 -08001215 if (needed <= 0) {
Andi Kleena5516432008-07-23 21:27:41 -07001216 h->resv_huge_pages += delta;
Adam Litkee4e574b2007-10-16 01:26:19 -07001217 return 0;
Adam Litkeac09b3a2008-03-04 14:29:38 -08001218 }
Adam Litkee4e574b2007-10-16 01:26:19 -07001219
1220 allocated = 0;
1221 INIT_LIST_HEAD(&surplus_list);
1222
1223 ret = -ENOMEM;
1224retry:
1225 spin_unlock(&hugetlb_lock);
1226 for (i = 0; i < needed; i++) {
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001227 page = alloc_buddy_huge_page(h, NUMA_NO_NODE);
Hillf Danton28073b02012-03-21 16:34:00 -07001228 if (!page) {
1229 alloc_ok = false;
1230 break;
1231 }
Adam Litkee4e574b2007-10-16 01:26:19 -07001232 list_add(&page->lru, &surplus_list);
1233 }
Hillf Danton28073b02012-03-21 16:34:00 -07001234 allocated += i;
Adam Litkee4e574b2007-10-16 01:26:19 -07001235
1236 /*
1237 * After retaking hugetlb_lock, we need to recalculate 'needed'
1238 * because either resv_huge_pages or free_huge_pages may have changed.
1239 */
1240 spin_lock(&hugetlb_lock);
Andi Kleena5516432008-07-23 21:27:41 -07001241 needed = (h->resv_huge_pages + delta) -
1242 (h->free_huge_pages + allocated);
Hillf Danton28073b02012-03-21 16:34:00 -07001243 if (needed > 0) {
1244 if (alloc_ok)
1245 goto retry;
1246 /*
1247 * We were not able to allocate enough pages to
1248 * satisfy the entire reservation so we free what
1249 * we've allocated so far.
1250 */
1251 goto free;
1252 }
Adam Litkee4e574b2007-10-16 01:26:19 -07001253 /*
1254 * The surplus_list now contains _at_least_ the number of extra pages
Lucas De Marchi25985ed2011-03-30 22:57:33 -03001255 * needed to accommodate the reservation. Add the appropriate number
Adam Litkee4e574b2007-10-16 01:26:19 -07001256 * of pages to the hugetlb pool and free the extras back to the buddy
Adam Litkeac09b3a2008-03-04 14:29:38 -08001257 * allocator. Commit the entire reservation here to prevent another
1258 * process from stealing the pages as they are added to the pool but
1259 * before they are reserved.
Adam Litkee4e574b2007-10-16 01:26:19 -07001260 */
1261 needed += allocated;
Andi Kleena5516432008-07-23 21:27:41 -07001262 h->resv_huge_pages += delta;
Adam Litkee4e574b2007-10-16 01:26:19 -07001263 ret = 0;
Naoya Horiguchia9869b82010-09-08 10:19:37 +09001264
Adam Litke19fc3f02008-04-28 02:12:20 -07001265 /* Free the needed pages to the hugetlb pool */
Adam Litkee4e574b2007-10-16 01:26:19 -07001266 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
Adam Litke19fc3f02008-04-28 02:12:20 -07001267 if ((--needed) < 0)
1268 break;
Naoya Horiguchia9869b82010-09-08 10:19:37 +09001269 /*
1270 * This page is now managed by the hugetlb allocator and has
1271 * no users -- drop the buddy allocator's reference.
1272 */
1273 put_page_testzero(page);
Sasha Levin309381fea2014-01-23 15:52:54 -08001274 VM_BUG_ON_PAGE(page_count(page), page);
Andi Kleena5516432008-07-23 21:27:41 -07001275 enqueue_huge_page(h, page);
Adam Litke19fc3f02008-04-28 02:12:20 -07001276 }
Hillf Danton28073b02012-03-21 16:34:00 -07001277free:
Hillf Dantonb0365c82011-12-28 15:57:16 -08001278 spin_unlock(&hugetlb_lock);
Adam Litke19fc3f02008-04-28 02:12:20 -07001279
1280 /* Free unnecessary surplus pages to the buddy allocator */
Joonsoo Kimc0d934b2013-09-11 14:21:02 -07001281 list_for_each_entry_safe(page, tmp, &surplus_list, lru)
1282 put_page(page);
Naoya Horiguchia9869b82010-09-08 10:19:37 +09001283 spin_lock(&hugetlb_lock);
Adam Litkee4e574b2007-10-16 01:26:19 -07001284
1285 return ret;
1286}
1287
1288/*
1289 * When releasing a hugetlb pool reservation, any surplus pages that were
1290 * allocated to satisfy the reservation must be explicitly freed if they were
1291 * never used.
Lee Schermerhorn685f3452009-09-21 17:01:23 -07001292 * Called with hugetlb_lock held.
Adam Litkee4e574b2007-10-16 01:26:19 -07001293 */
Andi Kleena5516432008-07-23 21:27:41 -07001294static void return_unused_surplus_pages(struct hstate *h,
1295 unsigned long unused_resv_pages)
Adam Litkee4e574b2007-10-16 01:26:19 -07001296{
Adam Litkee4e574b2007-10-16 01:26:19 -07001297 unsigned long nr_pages;
1298
Adam Litkeac09b3a2008-03-04 14:29:38 -08001299 /* Uncommit the reservation */
Andi Kleena5516432008-07-23 21:27:41 -07001300 h->resv_huge_pages -= unused_resv_pages;
Adam Litkeac09b3a2008-03-04 14:29:38 -08001301
Andi Kleenaa888a72008-07-23 21:27:47 -07001302 /* Cannot return gigantic pages currently */
Luiz Capitulinobae7f4a2014-06-04 16:07:08 -07001303 if (hstate_is_gigantic(h))
Andi Kleenaa888a72008-07-23 21:27:47 -07001304 return;
1305
Andi Kleena5516432008-07-23 21:27:41 -07001306 nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
Adam Litkee4e574b2007-10-16 01:26:19 -07001307
Lee Schermerhorn685f3452009-09-21 17:01:23 -07001308 /*
1309 * We want to release as many surplus pages as possible, spread
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -08001310 * evenly across all nodes with memory. Iterate across these nodes
1311 * until we can no longer free unreserved surplus pages. This occurs
1312 * when the nodes with surplus pages have no free pages.
1313 * free_pool_huge_page() will balance the the freed pages across the
1314 * on-line nodes with memory and will handle the hstate accounting.
Lee Schermerhorn685f3452009-09-21 17:01:23 -07001315 */
1316 while (nr_pages--) {
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08001317 if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1))
Lee Schermerhorn685f3452009-09-21 17:01:23 -07001318 break;
Mizuma, Masayoshi7848a4b2014-04-18 15:07:18 -07001319 cond_resched_lock(&hugetlb_lock);
Adam Litkee4e574b2007-10-16 01:26:19 -07001320 }
1321}
1322
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001323/*
1324 * Determine if the huge page at addr within the vma has an associated
1325 * reservation. Where it does not we will need to logically increase
David Gibson90481622012-03-21 16:34:12 -07001326 * reservation and actually increase subpool usage before an allocation
1327 * can occur. Where any new reservation would be required the
1328 * reservation change is prepared, but not committed. Once the page
1329 * has been allocated from the subpool and instantiated the change should
1330 * be committed via vma_commit_reservation. No action is required on
1331 * failure.
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001332 */
Roel Kluine2f17d92009-03-31 15:23:15 -07001333static long vma_needs_reservation(struct hstate *h,
Andi Kleena5516432008-07-23 21:27:41 -07001334 struct vm_area_struct *vma, unsigned long addr)
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001335{
Joonsoo Kim4e35f482014-04-03 14:47:30 -07001336 struct resv_map *resv;
1337 pgoff_t idx;
1338 long chg;
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001339
Joonsoo Kim4e35f482014-04-03 14:47:30 -07001340 resv = vma_resv_map(vma);
1341 if (!resv)
Andy Whitcroft84afd992008-07-23 21:27:32 -07001342 return 1;
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001343
Joonsoo Kim4e35f482014-04-03 14:47:30 -07001344 idx = vma_hugecache_offset(h, vma, addr);
1345 chg = region_chg(resv, idx, idx + 1);
Andy Whitcroft84afd992008-07-23 21:27:32 -07001346
Joonsoo Kim4e35f482014-04-03 14:47:30 -07001347 if (vma->vm_flags & VM_MAYSHARE)
1348 return chg;
1349 else
1350 return chg < 0 ? chg : 0;
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001351}
Andi Kleena5516432008-07-23 21:27:41 -07001352static void vma_commit_reservation(struct hstate *h,
1353 struct vm_area_struct *vma, unsigned long addr)
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001354{
Joonsoo Kim4e35f482014-04-03 14:47:30 -07001355 struct resv_map *resv;
1356 pgoff_t idx;
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001357
Joonsoo Kim4e35f482014-04-03 14:47:30 -07001358 resv = vma_resv_map(vma);
1359 if (!resv)
1360 return;
Joonsoo Kim9119a412014-04-03 14:47:25 -07001361
Joonsoo Kim4e35f482014-04-03 14:47:30 -07001362 idx = vma_hugecache_offset(h, vma, addr);
1363 region_add(resv, idx, idx + 1);
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001364}
1365
David Gibson27a85ef2006-03-22 00:08:56 -08001366static struct page *alloc_huge_page(struct vm_area_struct *vma,
Mel Gorman04f2cbe2008-07-23 21:27:25 -07001367 unsigned long addr, int avoid_reserve)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001368{
David Gibson90481622012-03-21 16:34:12 -07001369 struct hugepage_subpool *spool = subpool_vma(vma);
Andi Kleena5516432008-07-23 21:27:41 -07001370 struct hstate *h = hstate_vma(vma);
Adam Litke348ea202007-11-14 16:59:37 -08001371 struct page *page;
Roel Kluine2f17d92009-03-31 15:23:15 -07001372 long chg;
Aneesh Kumar K.V6d76dcf2012-07-31 16:42:18 -07001373 int ret, idx;
1374 struct hugetlb_cgroup *h_cg;
Adam Litke2fc39ce2007-11-14 16:59:39 -08001375
Aneesh Kumar K.V6d76dcf2012-07-31 16:42:18 -07001376 idx = hstate_index(h);
Mel Gormana1e78772008-07-23 21:27:23 -07001377 /*
David Gibson90481622012-03-21 16:34:12 -07001378 * Processes that did not create the mapping will have no
1379 * reserves and will not have accounted against subpool
1380 * limit. Check that the subpool limit can be made before
1381 * satisfying the allocation MAP_NORESERVE mappings may also
1382 * need pages and subpool limit allocated allocated if no reserve
1383 * mapping overlaps.
Mel Gormana1e78772008-07-23 21:27:23 -07001384 */
Andi Kleena5516432008-07-23 21:27:41 -07001385 chg = vma_needs_reservation(h, vma, addr);
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001386 if (chg < 0)
Aneesh Kumar K.V76dcee72012-07-31 16:41:57 -07001387 return ERR_PTR(-ENOMEM);
Joonsoo Kim8bb3f122013-09-11 14:21:54 -07001388 if (chg || avoid_reserve)
1389 if (hugepage_subpool_get_pages(spool, 1))
Aneesh Kumar K.V76dcee72012-07-31 16:41:57 -07001390 return ERR_PTR(-ENOSPC);
Mel Gormana1e78772008-07-23 21:27:23 -07001391
Aneesh Kumar K.V6d76dcf2012-07-31 16:42:18 -07001392 ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg);
Jianyu Zhan8f34af62014-06-04 16:10:36 -07001393 if (ret)
1394 goto out_subpool_put;
1395
Mel Gormana1e78772008-07-23 21:27:23 -07001396 spin_lock(&hugetlb_lock);
Joonsoo Kimaf0ed732013-09-11 14:21:18 -07001397 page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, chg);
Joonsoo Kim81a6fca2013-09-11 14:20:58 -07001398 if (!page) {
Aneesh Kumar K.V94ae8ba2012-07-31 16:42:35 -07001399 spin_unlock(&hugetlb_lock);
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001400 page = alloc_buddy_huge_page(h, NUMA_NO_NODE);
Jianyu Zhan8f34af62014-06-04 16:10:36 -07001401 if (!page)
1402 goto out_uncharge_cgroup;
1403
Aneesh Kumar K.V79dbb232012-07-31 16:42:32 -07001404 spin_lock(&hugetlb_lock);
1405 list_move(&page->lru, &h->hugepage_activelist);
Joonsoo Kim81a6fca2013-09-11 14:20:58 -07001406 /* Fall through */
Mel Gormana1e78772008-07-23 21:27:23 -07001407 }
Joonsoo Kim81a6fca2013-09-11 14:20:58 -07001408 hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page);
1409 spin_unlock(&hugetlb_lock);
Mel Gormana1e78772008-07-23 21:27:23 -07001410
David Gibson90481622012-03-21 16:34:12 -07001411 set_page_private(page, (unsigned long)spool);
Mel Gormana1e78772008-07-23 21:27:23 -07001412
Andi Kleena5516432008-07-23 21:27:41 -07001413 vma_commit_reservation(h, vma, addr);
Adam Litke90d8b7e2007-11-14 16:59:42 -08001414 return page;
Jianyu Zhan8f34af62014-06-04 16:10:36 -07001415
1416out_uncharge_cgroup:
1417 hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg);
1418out_subpool_put:
1419 if (chg || avoid_reserve)
1420 hugepage_subpool_put_pages(spool, 1);
1421 return ERR_PTR(-ENOSPC);
David Gibsonb45b5bd2006-03-22 00:08:55 -08001422}
1423
Naoya Horiguchi74060e42013-09-11 14:22:06 -07001424/*
1425 * alloc_huge_page()'s wrapper which simply returns the page if allocation
1426 * succeeds, otherwise NULL. This function is called from new_vma_page(),
1427 * where no ERR_VALUE is expected to be returned.
1428 */
1429struct page *alloc_huge_page_noerr(struct vm_area_struct *vma,
1430 unsigned long addr, int avoid_reserve)
1431{
1432 struct page *page = alloc_huge_page(vma, addr, avoid_reserve);
1433 if (IS_ERR(page))
1434 page = NULL;
1435 return page;
1436}
1437
Cyrill Gorcunov91f47662009-01-06 14:40:33 -08001438int __weak alloc_bootmem_huge_page(struct hstate *h)
Andi Kleenaa888a72008-07-23 21:27:47 -07001439{
1440 struct huge_bootmem_page *m;
Joonsoo Kimb2261022013-09-11 14:21:00 -07001441 int nr_nodes, node;
Andi Kleenaa888a72008-07-23 21:27:47 -07001442
Joonsoo Kimb2261022013-09-11 14:21:00 -07001443 for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) {
Andi Kleenaa888a72008-07-23 21:27:47 -07001444 void *addr;
1445
Grygorii Strashko8b89a112014-01-21 15:50:36 -08001446 addr = memblock_virt_alloc_try_nid_nopanic(
1447 huge_page_size(h), huge_page_size(h),
1448 0, BOOTMEM_ALLOC_ACCESSIBLE, node);
Andi Kleenaa888a72008-07-23 21:27:47 -07001449 if (addr) {
1450 /*
1451 * Use the beginning of the huge page to store the
1452 * huge_bootmem_page struct (until gather_bootmem
1453 * puts them into the mem_map).
1454 */
1455 m = addr;
Cyrill Gorcunov91f47662009-01-06 14:40:33 -08001456 goto found;
Andi Kleenaa888a72008-07-23 21:27:47 -07001457 }
Andi Kleenaa888a72008-07-23 21:27:47 -07001458 }
1459 return 0;
1460
1461found:
Luiz Capitulinodf994ea2014-12-12 16:55:21 -08001462 BUG_ON(!IS_ALIGNED(virt_to_phys(m), huge_page_size(h)));
Andi Kleenaa888a72008-07-23 21:27:47 -07001463 /* Put them into a private list first because mem_map is not up yet */
1464 list_add(&m->list, &huge_boot_pages);
1465 m->hstate = h;
1466 return 1;
1467}
1468
David Rientjesf412c972014-04-03 14:47:59 -07001469static void __init prep_compound_huge_page(struct page *page, int order)
Andy Whitcroft18229df2008-11-06 12:53:27 -08001470{
1471 if (unlikely(order > (MAX_ORDER - 1)))
1472 prep_compound_gigantic_page(page, order);
1473 else
1474 prep_compound_page(page, order);
1475}
1476
Andi Kleenaa888a72008-07-23 21:27:47 -07001477/* Put bootmem huge pages into the standard lists after mem_map is up */
1478static void __init gather_bootmem_prealloc(void)
1479{
1480 struct huge_bootmem_page *m;
1481
1482 list_for_each_entry(m, &huge_boot_pages, list) {
Andi Kleenaa888a72008-07-23 21:27:47 -07001483 struct hstate *h = m->hstate;
Becky Bruceee8f2482011-07-25 17:11:50 -07001484 struct page *page;
1485
1486#ifdef CONFIG_HIGHMEM
1487 page = pfn_to_page(m->phys >> PAGE_SHIFT);
Grygorii Strashko8b89a112014-01-21 15:50:36 -08001488 memblock_free_late(__pa(m),
1489 sizeof(struct huge_bootmem_page));
Becky Bruceee8f2482011-07-25 17:11:50 -07001490#else
1491 page = virt_to_page(m);
1492#endif
Andi Kleenaa888a72008-07-23 21:27:47 -07001493 WARN_ON(page_count(page) != 1);
Andy Whitcroft18229df2008-11-06 12:53:27 -08001494 prep_compound_huge_page(page, h->order);
Andrea Arcangelief5a22b2013-10-16 13:46:56 -07001495 WARN_ON(PageReserved(page));
Andi Kleenaa888a72008-07-23 21:27:47 -07001496 prep_new_huge_page(h, page, page_to_nid(page));
Rafael Aquinib0320c72011-06-15 15:08:39 -07001497 /*
1498 * If we had gigantic hugepages allocated at boot time, we need
1499 * to restore the 'stolen' pages to totalram_pages in order to
1500 * fix confusing memory reports from free(1) and another
1501 * side-effects, like CommitLimit going negative.
1502 */
Luiz Capitulinobae7f4a2014-06-04 16:07:08 -07001503 if (hstate_is_gigantic(h))
Jiang Liu3dcc0572013-07-03 15:03:21 -07001504 adjust_managed_page_count(page, 1 << h->order);
Andi Kleenaa888a72008-07-23 21:27:47 -07001505 }
1506}
1507
Andi Kleen8faa8b02008-07-23 21:27:48 -07001508static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001509{
1510 unsigned long i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001511
Andi Kleene5ff2152008-07-23 21:27:42 -07001512 for (i = 0; i < h->max_huge_pages; ++i) {
Luiz Capitulinobae7f4a2014-06-04 16:07:08 -07001513 if (hstate_is_gigantic(h)) {
Andi Kleenaa888a72008-07-23 21:27:47 -07001514 if (!alloc_bootmem_huge_page(h))
1515 break;
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -08001516 } else if (!alloc_fresh_huge_page(h,
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08001517 &node_states[N_MEMORY]))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001518 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001519 }
Andi Kleen8faa8b02008-07-23 21:27:48 -07001520 h->max_huge_pages = i;
Andi Kleene5ff2152008-07-23 21:27:42 -07001521}
1522
1523static void __init hugetlb_init_hstates(void)
1524{
1525 struct hstate *h;
1526
1527 for_each_hstate(h) {
Andi Kleen8faa8b02008-07-23 21:27:48 -07001528 /* oversize hugepages were init'ed in early boot */
Luiz Capitulinobae7f4a2014-06-04 16:07:08 -07001529 if (!hstate_is_gigantic(h))
Andi Kleen8faa8b02008-07-23 21:27:48 -07001530 hugetlb_hstate_alloc_pages(h);
Andi Kleene5ff2152008-07-23 21:27:42 -07001531 }
1532}
1533
Andi Kleen4abd32d2008-07-23 21:27:49 -07001534static char * __init memfmt(char *buf, unsigned long n)
1535{
1536 if (n >= (1UL << 30))
1537 sprintf(buf, "%lu GB", n >> 30);
1538 else if (n >= (1UL << 20))
1539 sprintf(buf, "%lu MB", n >> 20);
1540 else
1541 sprintf(buf, "%lu KB", n >> 10);
1542 return buf;
1543}
1544
Andi Kleene5ff2152008-07-23 21:27:42 -07001545static void __init report_hugepages(void)
1546{
1547 struct hstate *h;
1548
1549 for_each_hstate(h) {
Andi Kleen4abd32d2008-07-23 21:27:49 -07001550 char buf[32];
Andrew Mortonffb22af2013-02-22 16:32:08 -08001551 pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n",
Andi Kleen4abd32d2008-07-23 21:27:49 -07001552 memfmt(buf, huge_page_size(h)),
1553 h->free_huge_pages);
Andi Kleene5ff2152008-07-23 21:27:42 -07001554 }
1555}
1556
Linus Torvalds1da177e2005-04-16 15:20:36 -07001557#ifdef CONFIG_HIGHMEM
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001558static void try_to_free_low(struct hstate *h, unsigned long count,
1559 nodemask_t *nodes_allowed)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001560{
Christoph Lameter4415cc82006-09-25 23:31:55 -07001561 int i;
1562
Luiz Capitulinobae7f4a2014-06-04 16:07:08 -07001563 if (hstate_is_gigantic(h))
Andi Kleenaa888a72008-07-23 21:27:47 -07001564 return;
1565
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001566 for_each_node_mask(i, *nodes_allowed) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001567 struct page *page, *next;
Andi Kleena5516432008-07-23 21:27:41 -07001568 struct list_head *freel = &h->hugepage_freelists[i];
1569 list_for_each_entry_safe(page, next, freel, lru) {
1570 if (count >= h->nr_huge_pages)
Adam Litke6b0c8802007-10-16 01:26:23 -07001571 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001572 if (PageHighMem(page))
1573 continue;
1574 list_del(&page->lru);
Andi Kleene5ff2152008-07-23 21:27:42 -07001575 update_and_free_page(h, page);
Andi Kleena5516432008-07-23 21:27:41 -07001576 h->free_huge_pages--;
1577 h->free_huge_pages_node[page_to_nid(page)]--;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001578 }
1579 }
1580}
1581#else
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001582static inline void try_to_free_low(struct hstate *h, unsigned long count,
1583 nodemask_t *nodes_allowed)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001584{
1585}
1586#endif
1587
Wu Fengguang20a03072009-06-16 15:32:22 -07001588/*
1589 * Increment or decrement surplus_huge_pages. Keep node-specific counters
1590 * balanced by operating on them in a round-robin fashion.
1591 * Returns 1 if an adjustment was made.
1592 */
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001593static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
1594 int delta)
Wu Fengguang20a03072009-06-16 15:32:22 -07001595{
Joonsoo Kimb2261022013-09-11 14:21:00 -07001596 int nr_nodes, node;
Wu Fengguang20a03072009-06-16 15:32:22 -07001597
1598 VM_BUG_ON(delta != -1 && delta != 1);
Wu Fengguang20a03072009-06-16 15:32:22 -07001599
Joonsoo Kimb2261022013-09-11 14:21:00 -07001600 if (delta < 0) {
1601 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
1602 if (h->surplus_huge_pages_node[node])
1603 goto found;
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001604 }
Joonsoo Kimb2261022013-09-11 14:21:00 -07001605 } else {
1606 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
1607 if (h->surplus_huge_pages_node[node] <
1608 h->nr_huge_pages_node[node])
1609 goto found;
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001610 }
Joonsoo Kimb2261022013-09-11 14:21:00 -07001611 }
1612 return 0;
Wu Fengguang20a03072009-06-16 15:32:22 -07001613
Joonsoo Kimb2261022013-09-11 14:21:00 -07001614found:
1615 h->surplus_huge_pages += delta;
1616 h->surplus_huge_pages_node[node] += delta;
1617 return 1;
Wu Fengguang20a03072009-06-16 15:32:22 -07001618}
1619
Andi Kleena5516432008-07-23 21:27:41 -07001620#define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001621static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count,
1622 nodemask_t *nodes_allowed)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001623{
Adam Litke7893d1d2007-10-16 01:26:18 -07001624 unsigned long min_count, ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001625
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001626 if (hstate_is_gigantic(h) && !gigantic_page_supported())
Andi Kleenaa888a72008-07-23 21:27:47 -07001627 return h->max_huge_pages;
1628
Adam Litke7893d1d2007-10-16 01:26:18 -07001629 /*
1630 * Increase the pool size
1631 * First take pages out of surplus state. Then make up the
1632 * remaining difference by allocating fresh huge pages.
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001633 *
1634 * We might race with alloc_buddy_huge_page() here and be unable
1635 * to convert a surplus huge page to a normal huge page. That is
1636 * not critical, though, it just means the overall size of the
1637 * pool might be one hugepage larger than it needs to be, but
1638 * within all the constraints specified by the sysctls.
Adam Litke7893d1d2007-10-16 01:26:18 -07001639 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001640 spin_lock(&hugetlb_lock);
Andi Kleena5516432008-07-23 21:27:41 -07001641 while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001642 if (!adjust_pool_surplus(h, nodes_allowed, -1))
Adam Litke7893d1d2007-10-16 01:26:18 -07001643 break;
1644 }
1645
Andi Kleena5516432008-07-23 21:27:41 -07001646 while (count > persistent_huge_pages(h)) {
Adam Litke7893d1d2007-10-16 01:26:18 -07001647 /*
1648 * If this allocation races such that we no longer need the
1649 * page, free_huge_page will handle it by freeing the page
1650 * and reducing the surplus.
1651 */
1652 spin_unlock(&hugetlb_lock);
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001653 if (hstate_is_gigantic(h))
1654 ret = alloc_fresh_gigantic_page(h, nodes_allowed);
1655 else
1656 ret = alloc_fresh_huge_page(h, nodes_allowed);
Adam Litke7893d1d2007-10-16 01:26:18 -07001657 spin_lock(&hugetlb_lock);
1658 if (!ret)
1659 goto out;
1660
Mel Gorman536240f22009-12-14 17:59:56 -08001661 /* Bail for signals. Probably ctrl-c from user */
1662 if (signal_pending(current))
1663 goto out;
Adam Litke7893d1d2007-10-16 01:26:18 -07001664 }
Adam Litke7893d1d2007-10-16 01:26:18 -07001665
1666 /*
1667 * Decrease the pool size
1668 * First return free pages to the buddy allocator (being careful
1669 * to keep enough around to satisfy reservations). Then place
1670 * pages into surplus state as needed so the pool will shrink
1671 * to the desired size as pages become free.
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001672 *
1673 * By placing pages into the surplus state independent of the
1674 * overcommit value, we are allowing the surplus pool size to
1675 * exceed overcommit. There are few sane options here. Since
1676 * alloc_buddy_huge_page() is checking the global counter,
1677 * though, we'll note that we're not allowed to exceed surplus
1678 * and won't grow the pool anywhere else. Not until one of the
1679 * sysctls are changed, or the surplus pages go out of use.
Adam Litke7893d1d2007-10-16 01:26:18 -07001680 */
Andi Kleena5516432008-07-23 21:27:41 -07001681 min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
Adam Litke6b0c8802007-10-16 01:26:23 -07001682 min_count = max(count, min_count);
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001683 try_to_free_low(h, min_count, nodes_allowed);
Andi Kleena5516432008-07-23 21:27:41 -07001684 while (min_count < persistent_huge_pages(h)) {
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001685 if (!free_pool_huge_page(h, nodes_allowed, 0))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001686 break;
Mizuma, Masayoshi55f67142014-04-07 15:37:54 -07001687 cond_resched_lock(&hugetlb_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001688 }
Andi Kleena5516432008-07-23 21:27:41 -07001689 while (count < persistent_huge_pages(h)) {
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001690 if (!adjust_pool_surplus(h, nodes_allowed, 1))
Adam Litke7893d1d2007-10-16 01:26:18 -07001691 break;
1692 }
1693out:
Andi Kleena5516432008-07-23 21:27:41 -07001694 ret = persistent_huge_pages(h);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001695 spin_unlock(&hugetlb_lock);
Adam Litke7893d1d2007-10-16 01:26:18 -07001696 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001697}
1698
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001699#define HSTATE_ATTR_RO(_name) \
1700 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
1701
1702#define HSTATE_ATTR(_name) \
1703 static struct kobj_attribute _name##_attr = \
1704 __ATTR(_name, 0644, _name##_show, _name##_store)
1705
1706static struct kobject *hugepages_kobj;
1707static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
1708
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001709static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
1710
1711static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001712{
1713 int i;
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001714
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001715 for (i = 0; i < HUGE_MAX_HSTATE; i++)
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001716 if (hstate_kobjs[i] == kobj) {
1717 if (nidp)
1718 *nidp = NUMA_NO_NODE;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001719 return &hstates[i];
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001720 }
1721
1722 return kobj_to_node_hstate(kobj, nidp);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001723}
1724
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001725static ssize_t nr_hugepages_show_common(struct kobject *kobj,
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001726 struct kobj_attribute *attr, char *buf)
1727{
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001728 struct hstate *h;
1729 unsigned long nr_huge_pages;
1730 int nid;
1731
1732 h = kobj_to_hstate(kobj, &nid);
1733 if (nid == NUMA_NO_NODE)
1734 nr_huge_pages = h->nr_huge_pages;
1735 else
1736 nr_huge_pages = h->nr_huge_pages_node[nid];
1737
1738 return sprintf(buf, "%lu\n", nr_huge_pages);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001739}
Eric B Munsonadbe8722011-01-13 15:47:27 -08001740
David Rientjes238d3c12014-08-06 16:06:51 -07001741static ssize_t __nr_hugepages_store_common(bool obey_mempolicy,
1742 struct hstate *h, int nid,
1743 unsigned long count, size_t len)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001744{
1745 int err;
David Rientjesbad44b52009-12-14 17:58:38 -08001746 NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001747
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001748 if (hstate_is_gigantic(h) && !gigantic_page_supported()) {
Eric B Munsonadbe8722011-01-13 15:47:27 -08001749 err = -EINVAL;
1750 goto out;
1751 }
1752
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001753 if (nid == NUMA_NO_NODE) {
1754 /*
1755 * global hstate attribute
1756 */
1757 if (!(obey_mempolicy &&
1758 init_nodemask_of_mempolicy(nodes_allowed))) {
1759 NODEMASK_FREE(nodes_allowed);
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08001760 nodes_allowed = &node_states[N_MEMORY];
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001761 }
1762 } else if (nodes_allowed) {
1763 /*
1764 * per node hstate attribute: adjust count to global,
1765 * but restrict alloc/free to the specified node.
1766 */
1767 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
1768 init_nodemask_of_node(nodes_allowed, nid);
1769 } else
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08001770 nodes_allowed = &node_states[N_MEMORY];
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001771
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001772 h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001773
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08001774 if (nodes_allowed != &node_states[N_MEMORY])
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001775 NODEMASK_FREE(nodes_allowed);
1776
1777 return len;
Eric B Munsonadbe8722011-01-13 15:47:27 -08001778out:
1779 NODEMASK_FREE(nodes_allowed);
1780 return err;
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001781}
1782
David Rientjes238d3c12014-08-06 16:06:51 -07001783static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
1784 struct kobject *kobj, const char *buf,
1785 size_t len)
1786{
1787 struct hstate *h;
1788 unsigned long count;
1789 int nid;
1790 int err;
1791
1792 err = kstrtoul(buf, 10, &count);
1793 if (err)
1794 return err;
1795
1796 h = kobj_to_hstate(kobj, &nid);
1797 return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len);
1798}
1799
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001800static ssize_t nr_hugepages_show(struct kobject *kobj,
1801 struct kobj_attribute *attr, char *buf)
1802{
1803 return nr_hugepages_show_common(kobj, attr, buf);
1804}
1805
1806static ssize_t nr_hugepages_store(struct kobject *kobj,
1807 struct kobj_attribute *attr, const char *buf, size_t len)
1808{
David Rientjes238d3c12014-08-06 16:06:51 -07001809 return nr_hugepages_store_common(false, kobj, buf, len);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001810}
1811HSTATE_ATTR(nr_hugepages);
1812
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001813#ifdef CONFIG_NUMA
1814
1815/*
1816 * hstate attribute for optionally mempolicy-based constraint on persistent
1817 * huge page alloc/free.
1818 */
1819static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
1820 struct kobj_attribute *attr, char *buf)
1821{
1822 return nr_hugepages_show_common(kobj, attr, buf);
1823}
1824
1825static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
1826 struct kobj_attribute *attr, const char *buf, size_t len)
1827{
David Rientjes238d3c12014-08-06 16:06:51 -07001828 return nr_hugepages_store_common(true, kobj, buf, len);
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001829}
1830HSTATE_ATTR(nr_hugepages_mempolicy);
1831#endif
1832
1833
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001834static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
1835 struct kobj_attribute *attr, char *buf)
1836{
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001837 struct hstate *h = kobj_to_hstate(kobj, NULL);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001838 return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
1839}
Eric B Munsonadbe8722011-01-13 15:47:27 -08001840
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001841static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
1842 struct kobj_attribute *attr, const char *buf, size_t count)
1843{
1844 int err;
1845 unsigned long input;
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001846 struct hstate *h = kobj_to_hstate(kobj, NULL);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001847
Luiz Capitulinobae7f4a2014-06-04 16:07:08 -07001848 if (hstate_is_gigantic(h))
Eric B Munsonadbe8722011-01-13 15:47:27 -08001849 return -EINVAL;
1850
Jingoo Han3dbb95f2013-09-11 14:20:25 -07001851 err = kstrtoul(buf, 10, &input);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001852 if (err)
Eric B Munson73ae31e2011-01-13 15:47:28 -08001853 return err;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001854
1855 spin_lock(&hugetlb_lock);
1856 h->nr_overcommit_huge_pages = input;
1857 spin_unlock(&hugetlb_lock);
1858
1859 return count;
1860}
1861HSTATE_ATTR(nr_overcommit_hugepages);
1862
1863static ssize_t free_hugepages_show(struct kobject *kobj,
1864 struct kobj_attribute *attr, char *buf)
1865{
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001866 struct hstate *h;
1867 unsigned long free_huge_pages;
1868 int nid;
1869
1870 h = kobj_to_hstate(kobj, &nid);
1871 if (nid == NUMA_NO_NODE)
1872 free_huge_pages = h->free_huge_pages;
1873 else
1874 free_huge_pages = h->free_huge_pages_node[nid];
1875
1876 return sprintf(buf, "%lu\n", free_huge_pages);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001877}
1878HSTATE_ATTR_RO(free_hugepages);
1879
1880static ssize_t resv_hugepages_show(struct kobject *kobj,
1881 struct kobj_attribute *attr, char *buf)
1882{
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001883 struct hstate *h = kobj_to_hstate(kobj, NULL);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001884 return sprintf(buf, "%lu\n", h->resv_huge_pages);
1885}
1886HSTATE_ATTR_RO(resv_hugepages);
1887
1888static ssize_t surplus_hugepages_show(struct kobject *kobj,
1889 struct kobj_attribute *attr, char *buf)
1890{
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001891 struct hstate *h;
1892 unsigned long surplus_huge_pages;
1893 int nid;
1894
1895 h = kobj_to_hstate(kobj, &nid);
1896 if (nid == NUMA_NO_NODE)
1897 surplus_huge_pages = h->surplus_huge_pages;
1898 else
1899 surplus_huge_pages = h->surplus_huge_pages_node[nid];
1900
1901 return sprintf(buf, "%lu\n", surplus_huge_pages);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001902}
1903HSTATE_ATTR_RO(surplus_hugepages);
1904
1905static struct attribute *hstate_attrs[] = {
1906 &nr_hugepages_attr.attr,
1907 &nr_overcommit_hugepages_attr.attr,
1908 &free_hugepages_attr.attr,
1909 &resv_hugepages_attr.attr,
1910 &surplus_hugepages_attr.attr,
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001911#ifdef CONFIG_NUMA
1912 &nr_hugepages_mempolicy_attr.attr,
1913#endif
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001914 NULL,
1915};
1916
1917static struct attribute_group hstate_attr_group = {
1918 .attrs = hstate_attrs,
1919};
1920
Jeff Mahoney094e9532010-02-02 13:44:14 -08001921static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
1922 struct kobject **hstate_kobjs,
1923 struct attribute_group *hstate_attr_group)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001924{
1925 int retval;
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07001926 int hi = hstate_index(h);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001927
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001928 hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
1929 if (!hstate_kobjs[hi])
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001930 return -ENOMEM;
1931
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001932 retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001933 if (retval)
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001934 kobject_put(hstate_kobjs[hi]);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001935
1936 return retval;
1937}
1938
1939static void __init hugetlb_sysfs_init(void)
1940{
1941 struct hstate *h;
1942 int err;
1943
1944 hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
1945 if (!hugepages_kobj)
1946 return;
1947
1948 for_each_hstate(h) {
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001949 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
1950 hstate_kobjs, &hstate_attr_group);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001951 if (err)
Andrew Mortonffb22af2013-02-22 16:32:08 -08001952 pr_err("Hugetlb: Unable to add hstate %s", h->name);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001953 }
1954}
1955
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001956#ifdef CONFIG_NUMA
1957
1958/*
1959 * node_hstate/s - associate per node hstate attributes, via their kobjects,
Kay Sievers10fbcf42011-12-21 14:48:43 -08001960 * with node devices in node_devices[] using a parallel array. The array
1961 * index of a node device or _hstate == node id.
1962 * This is here to avoid any static dependency of the node device driver, in
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001963 * the base kernel, on the hugetlb module.
1964 */
1965struct node_hstate {
1966 struct kobject *hugepages_kobj;
1967 struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
1968};
1969struct node_hstate node_hstates[MAX_NUMNODES];
1970
1971/*
Kay Sievers10fbcf42011-12-21 14:48:43 -08001972 * A subset of global hstate attributes for node devices
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001973 */
1974static struct attribute *per_node_hstate_attrs[] = {
1975 &nr_hugepages_attr.attr,
1976 &free_hugepages_attr.attr,
1977 &surplus_hugepages_attr.attr,
1978 NULL,
1979};
1980
1981static struct attribute_group per_node_hstate_attr_group = {
1982 .attrs = per_node_hstate_attrs,
1983};
1984
1985/*
Kay Sievers10fbcf42011-12-21 14:48:43 -08001986 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001987 * Returns node id via non-NULL nidp.
1988 */
1989static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
1990{
1991 int nid;
1992
1993 for (nid = 0; nid < nr_node_ids; nid++) {
1994 struct node_hstate *nhs = &node_hstates[nid];
1995 int i;
1996 for (i = 0; i < HUGE_MAX_HSTATE; i++)
1997 if (nhs->hstate_kobjs[i] == kobj) {
1998 if (nidp)
1999 *nidp = nid;
2000 return &hstates[i];
2001 }
2002 }
2003
2004 BUG();
2005 return NULL;
2006}
2007
2008/*
Kay Sievers10fbcf42011-12-21 14:48:43 -08002009 * Unregister hstate attributes from a single node device.
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002010 * No-op if no hstate attributes attached.
2011 */
Claudiu Ghioc3cd8b442013-03-04 12:46:15 +02002012static void hugetlb_unregister_node(struct node *node)
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002013{
2014 struct hstate *h;
Kay Sievers10fbcf42011-12-21 14:48:43 -08002015 struct node_hstate *nhs = &node_hstates[node->dev.id];
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002016
2017 if (!nhs->hugepages_kobj)
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -08002018 return; /* no hstate attributes */
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002019
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07002020 for_each_hstate(h) {
2021 int idx = hstate_index(h);
2022 if (nhs->hstate_kobjs[idx]) {
2023 kobject_put(nhs->hstate_kobjs[idx]);
2024 nhs->hstate_kobjs[idx] = NULL;
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002025 }
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07002026 }
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002027
2028 kobject_put(nhs->hugepages_kobj);
2029 nhs->hugepages_kobj = NULL;
2030}
2031
2032/*
Kay Sievers10fbcf42011-12-21 14:48:43 -08002033 * hugetlb module exit: unregister hstate attributes from node devices
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002034 * that have them.
2035 */
2036static void hugetlb_unregister_all_nodes(void)
2037{
2038 int nid;
2039
2040 /*
Kay Sievers10fbcf42011-12-21 14:48:43 -08002041 * disable node device registrations.
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002042 */
2043 register_hugetlbfs_with_node(NULL, NULL);
2044
2045 /*
2046 * remove hstate attributes from any nodes that have them.
2047 */
2048 for (nid = 0; nid < nr_node_ids; nid++)
Wen Congyang87327942012-12-11 16:00:56 -08002049 hugetlb_unregister_node(node_devices[nid]);
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002050}
2051
2052/*
Kay Sievers10fbcf42011-12-21 14:48:43 -08002053 * Register hstate attributes for a single node device.
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002054 * No-op if attributes already registered.
2055 */
Claudiu Ghioc3cd8b442013-03-04 12:46:15 +02002056static void hugetlb_register_node(struct node *node)
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002057{
2058 struct hstate *h;
Kay Sievers10fbcf42011-12-21 14:48:43 -08002059 struct node_hstate *nhs = &node_hstates[node->dev.id];
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002060 int err;
2061
2062 if (nhs->hugepages_kobj)
2063 return; /* already allocated */
2064
2065 nhs->hugepages_kobj = kobject_create_and_add("hugepages",
Kay Sievers10fbcf42011-12-21 14:48:43 -08002066 &node->dev.kobj);
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002067 if (!nhs->hugepages_kobj)
2068 return;
2069
2070 for_each_hstate(h) {
2071 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
2072 nhs->hstate_kobjs,
2073 &per_node_hstate_attr_group);
2074 if (err) {
Andrew Mortonffb22af2013-02-22 16:32:08 -08002075 pr_err("Hugetlb: Unable to add hstate %s for node %d\n",
2076 h->name, node->dev.id);
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002077 hugetlb_unregister_node(node);
2078 break;
2079 }
2080 }
2081}
2082
2083/*
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -08002084 * hugetlb init time: register hstate attributes for all registered node
Kay Sievers10fbcf42011-12-21 14:48:43 -08002085 * devices of nodes that have memory. All on-line nodes should have
2086 * registered their associated device by this time.
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002087 */
Luiz Capitulino7d9ca002014-12-12 16:55:24 -08002088static void __init hugetlb_register_all_nodes(void)
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002089{
2090 int nid;
2091
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08002092 for_each_node_state(nid, N_MEMORY) {
Wen Congyang87327942012-12-11 16:00:56 -08002093 struct node *node = node_devices[nid];
Kay Sievers10fbcf42011-12-21 14:48:43 -08002094 if (node->dev.id == nid)
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002095 hugetlb_register_node(node);
2096 }
2097
2098 /*
Kay Sievers10fbcf42011-12-21 14:48:43 -08002099 * Let the node device driver know we're here so it can
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002100 * [un]register hstate attributes on node hotplug.
2101 */
2102 register_hugetlbfs_with_node(hugetlb_register_node,
2103 hugetlb_unregister_node);
2104}
2105#else /* !CONFIG_NUMA */
2106
2107static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
2108{
2109 BUG();
2110 if (nidp)
2111 *nidp = -1;
2112 return NULL;
2113}
2114
2115static void hugetlb_unregister_all_nodes(void) { }
2116
2117static void hugetlb_register_all_nodes(void) { }
2118
2119#endif
2120
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002121static void __exit hugetlb_exit(void)
2122{
2123 struct hstate *h;
2124
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002125 hugetlb_unregister_all_nodes();
2126
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002127 for_each_hstate(h) {
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07002128 kobject_put(hstate_kobjs[hstate_index(h)]);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002129 }
2130
2131 kobject_put(hugepages_kobj);
Davidlohr Bueso8382d912014-04-03 14:47:31 -07002132 kfree(htlb_fault_mutex_table);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002133}
2134module_exit(hugetlb_exit);
2135
2136static int __init hugetlb_init(void)
2137{
Davidlohr Bueso8382d912014-04-03 14:47:31 -07002138 int i;
2139
Nishanth Aravamudan457c1b22014-05-06 12:50:00 -07002140 if (!hugepages_supported())
Benjamin Herrenschmidt0ef89d22008-07-31 00:07:30 -07002141 return 0;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002142
Nick Piggine11bfbf2008-07-23 21:27:52 -07002143 if (!size_to_hstate(default_hstate_size)) {
2144 default_hstate_size = HPAGE_SIZE;
2145 if (!size_to_hstate(default_hstate_size))
2146 hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002147 }
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07002148 default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size));
Nick Piggine11bfbf2008-07-23 21:27:52 -07002149 if (default_hstate_max_huge_pages)
2150 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002151
2152 hugetlb_init_hstates();
Andi Kleenaa888a72008-07-23 21:27:47 -07002153 gather_bootmem_prealloc();
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002154 report_hugepages();
2155
2156 hugetlb_sysfs_init();
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002157 hugetlb_register_all_nodes();
Jianguo Wu7179e7b2012-12-18 14:23:19 -08002158 hugetlb_cgroup_file_init();
Lee Schermerhorn9a305232009-12-14 17:58:25 -08002159
Davidlohr Bueso8382d912014-04-03 14:47:31 -07002160#ifdef CONFIG_SMP
2161 num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus());
2162#else
2163 num_fault_mutexes = 1;
2164#endif
2165 htlb_fault_mutex_table =
2166 kmalloc(sizeof(struct mutex) * num_fault_mutexes, GFP_KERNEL);
2167 BUG_ON(!htlb_fault_mutex_table);
2168
2169 for (i = 0; i < num_fault_mutexes; i++)
2170 mutex_init(&htlb_fault_mutex_table[i]);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002171 return 0;
2172}
2173module_init(hugetlb_init);
2174
2175/* Should be called on processing a hugepagesz=... option */
2176void __init hugetlb_add_hstate(unsigned order)
2177{
2178 struct hstate *h;
Andi Kleen8faa8b02008-07-23 21:27:48 -07002179 unsigned long i;
2180
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002181 if (size_to_hstate(PAGE_SIZE << order)) {
Andrew Mortonffb22af2013-02-22 16:32:08 -08002182 pr_warning("hugepagesz= specified twice, ignoring\n");
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002183 return;
2184 }
Aneesh Kumar K.V47d38342012-07-31 16:41:54 -07002185 BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002186 BUG_ON(order == 0);
Aneesh Kumar K.V47d38342012-07-31 16:41:54 -07002187 h = &hstates[hugetlb_max_hstate++];
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002188 h->order = order;
2189 h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
Andi Kleen8faa8b02008-07-23 21:27:48 -07002190 h->nr_huge_pages = 0;
2191 h->free_huge_pages = 0;
2192 for (i = 0; i < MAX_NUMNODES; ++i)
2193 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
Aneesh Kumar K.V0edaecf2012-07-31 16:42:07 -07002194 INIT_LIST_HEAD(&h->hugepage_activelist);
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08002195 h->next_nid_to_alloc = first_node(node_states[N_MEMORY]);
2196 h->next_nid_to_free = first_node(node_states[N_MEMORY]);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002197 snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
2198 huge_page_size(h)/1024);
Andi Kleen8faa8b02008-07-23 21:27:48 -07002199
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002200 parsed_hstate = h;
2201}
2202
Nick Piggine11bfbf2008-07-23 21:27:52 -07002203static int __init hugetlb_nrpages_setup(char *s)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002204{
2205 unsigned long *mhp;
Andi Kleen8faa8b02008-07-23 21:27:48 -07002206 static unsigned long *last_mhp;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002207
2208 /*
Aneesh Kumar K.V47d38342012-07-31 16:41:54 -07002209 * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet,
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002210 * so this hugepages= parameter goes to the "default hstate".
2211 */
Aneesh Kumar K.V47d38342012-07-31 16:41:54 -07002212 if (!hugetlb_max_hstate)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002213 mhp = &default_hstate_max_huge_pages;
2214 else
2215 mhp = &parsed_hstate->max_huge_pages;
2216
Andi Kleen8faa8b02008-07-23 21:27:48 -07002217 if (mhp == last_mhp) {
Andrew Mortonffb22af2013-02-22 16:32:08 -08002218 pr_warning("hugepages= specified twice without "
2219 "interleaving hugepagesz=, ignoring\n");
Andi Kleen8faa8b02008-07-23 21:27:48 -07002220 return 1;
2221 }
2222
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002223 if (sscanf(s, "%lu", mhp) <= 0)
2224 *mhp = 0;
2225
Andi Kleen8faa8b02008-07-23 21:27:48 -07002226 /*
2227 * Global state is always initialized later in hugetlb_init.
2228 * But we need to allocate >= MAX_ORDER hstates here early to still
2229 * use the bootmem allocator.
2230 */
Aneesh Kumar K.V47d38342012-07-31 16:41:54 -07002231 if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER)
Andi Kleen8faa8b02008-07-23 21:27:48 -07002232 hugetlb_hstate_alloc_pages(parsed_hstate);
2233
2234 last_mhp = mhp;
2235
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002236 return 1;
2237}
Nick Piggine11bfbf2008-07-23 21:27:52 -07002238__setup("hugepages=", hugetlb_nrpages_setup);
2239
2240static int __init hugetlb_default_setup(char *s)
2241{
2242 default_hstate_size = memparse(s, &s);
2243 return 1;
2244}
2245__setup("default_hugepagesz=", hugetlb_default_setup);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002246
Nishanth Aravamudan8a213462008-07-25 19:44:37 -07002247static unsigned int cpuset_mems_nr(unsigned int *array)
2248{
2249 int node;
2250 unsigned int nr = 0;
2251
2252 for_each_node_mask(node, cpuset_current_mems_allowed)
2253 nr += array[node];
2254
2255 return nr;
2256}
2257
2258#ifdef CONFIG_SYSCTL
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002259static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
2260 struct ctl_table *table, int write,
2261 void __user *buffer, size_t *length, loff_t *ppos)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002262{
Andi Kleene5ff2152008-07-23 21:27:42 -07002263 struct hstate *h = &default_hstate;
David Rientjes238d3c12014-08-06 16:06:51 -07002264 unsigned long tmp = h->max_huge_pages;
Michal Hocko08d4a242011-01-13 15:47:26 -08002265 int ret;
Andi Kleene5ff2152008-07-23 21:27:42 -07002266
Nishanth Aravamudan457c1b22014-05-06 12:50:00 -07002267 if (!hugepages_supported())
2268 return -ENOTSUPP;
2269
Andi Kleene5ff2152008-07-23 21:27:42 -07002270 table->data = &tmp;
2271 table->maxlen = sizeof(unsigned long);
Michal Hocko08d4a242011-01-13 15:47:26 -08002272 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2273 if (ret)
2274 goto out;
Andi Kleene5ff2152008-07-23 21:27:42 -07002275
David Rientjes238d3c12014-08-06 16:06:51 -07002276 if (write)
2277 ret = __nr_hugepages_store_common(obey_mempolicy, h,
2278 NUMA_NO_NODE, tmp, *length);
Michal Hocko08d4a242011-01-13 15:47:26 -08002279out:
2280 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002281}
Mel Gorman396faf02007-07-17 04:03:13 -07002282
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002283int hugetlb_sysctl_handler(struct ctl_table *table, int write,
2284 void __user *buffer, size_t *length, loff_t *ppos)
2285{
2286
2287 return hugetlb_sysctl_handler_common(false, table, write,
2288 buffer, length, ppos);
2289}
2290
2291#ifdef CONFIG_NUMA
2292int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
2293 void __user *buffer, size_t *length, loff_t *ppos)
2294{
2295 return hugetlb_sysctl_handler_common(true, table, write,
2296 buffer, length, ppos);
2297}
2298#endif /* CONFIG_NUMA */
2299
Nishanth Aravamudana3d0c6a2008-02-08 04:18:18 -08002300int hugetlb_overcommit_handler(struct ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07002301 void __user *buffer,
Nishanth Aravamudana3d0c6a2008-02-08 04:18:18 -08002302 size_t *length, loff_t *ppos)
2303{
Andi Kleena5516432008-07-23 21:27:41 -07002304 struct hstate *h = &default_hstate;
Andi Kleene5ff2152008-07-23 21:27:42 -07002305 unsigned long tmp;
Michal Hocko08d4a242011-01-13 15:47:26 -08002306 int ret;
Andi Kleene5ff2152008-07-23 21:27:42 -07002307
Nishanth Aravamudan457c1b22014-05-06 12:50:00 -07002308 if (!hugepages_supported())
2309 return -ENOTSUPP;
2310
Petr Holasekc033a932011-03-22 16:33:05 -07002311 tmp = h->nr_overcommit_huge_pages;
Andi Kleene5ff2152008-07-23 21:27:42 -07002312
Luiz Capitulinobae7f4a2014-06-04 16:07:08 -07002313 if (write && hstate_is_gigantic(h))
Eric B Munsonadbe8722011-01-13 15:47:27 -08002314 return -EINVAL;
2315
Andi Kleene5ff2152008-07-23 21:27:42 -07002316 table->data = &tmp;
2317 table->maxlen = sizeof(unsigned long);
Michal Hocko08d4a242011-01-13 15:47:26 -08002318 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2319 if (ret)
2320 goto out;
Andi Kleene5ff2152008-07-23 21:27:42 -07002321
2322 if (write) {
2323 spin_lock(&hugetlb_lock);
2324 h->nr_overcommit_huge_pages = tmp;
2325 spin_unlock(&hugetlb_lock);
2326 }
Michal Hocko08d4a242011-01-13 15:47:26 -08002327out:
2328 return ret;
Nishanth Aravamudana3d0c6a2008-02-08 04:18:18 -08002329}
2330
Linus Torvalds1da177e2005-04-16 15:20:36 -07002331#endif /* CONFIG_SYSCTL */
2332
Alexey Dobriyane1759c22008-10-15 23:50:22 +04002333void hugetlb_report_meminfo(struct seq_file *m)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002334{
Andi Kleena5516432008-07-23 21:27:41 -07002335 struct hstate *h = &default_hstate;
Nishanth Aravamudan457c1b22014-05-06 12:50:00 -07002336 if (!hugepages_supported())
2337 return;
Alexey Dobriyane1759c22008-10-15 23:50:22 +04002338 seq_printf(m,
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002339 "HugePages_Total: %5lu\n"
2340 "HugePages_Free: %5lu\n"
2341 "HugePages_Rsvd: %5lu\n"
2342 "HugePages_Surp: %5lu\n"
2343 "Hugepagesize: %8lu kB\n",
Andi Kleena5516432008-07-23 21:27:41 -07002344 h->nr_huge_pages,
2345 h->free_huge_pages,
2346 h->resv_huge_pages,
2347 h->surplus_huge_pages,
2348 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002349}
2350
2351int hugetlb_report_node_meminfo(int nid, char *buf)
2352{
Andi Kleena5516432008-07-23 21:27:41 -07002353 struct hstate *h = &default_hstate;
Nishanth Aravamudan457c1b22014-05-06 12:50:00 -07002354 if (!hugepages_supported())
2355 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002356 return sprintf(buf,
2357 "Node %d HugePages_Total: %5u\n"
Nishanth Aravamudana1de0912008-03-26 14:37:53 -07002358 "Node %d HugePages_Free: %5u\n"
2359 "Node %d HugePages_Surp: %5u\n",
Andi Kleena5516432008-07-23 21:27:41 -07002360 nid, h->nr_huge_pages_node[nid],
2361 nid, h->free_huge_pages_node[nid],
2362 nid, h->surplus_huge_pages_node[nid]);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002363}
2364
David Rientjes949f7ec2013-04-29 15:07:48 -07002365void hugetlb_show_meminfo(void)
2366{
2367 struct hstate *h;
2368 int nid;
2369
Nishanth Aravamudan457c1b22014-05-06 12:50:00 -07002370 if (!hugepages_supported())
2371 return;
2372
David Rientjes949f7ec2013-04-29 15:07:48 -07002373 for_each_node_state(nid, N_MEMORY)
2374 for_each_hstate(h)
2375 pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
2376 nid,
2377 h->nr_huge_pages_node[nid],
2378 h->free_huge_pages_node[nid],
2379 h->surplus_huge_pages_node[nid],
2380 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
2381}
2382
Linus Torvalds1da177e2005-04-16 15:20:36 -07002383/* Return the number pages of memory we physically have, in PAGE_SIZE units. */
2384unsigned long hugetlb_total_pages(void)
2385{
Wanpeng Lid0028582013-03-22 15:04:40 -07002386 struct hstate *h;
2387 unsigned long nr_total_pages = 0;
2388
2389 for_each_hstate(h)
2390 nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h);
2391 return nr_total_pages;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002392}
Linus Torvalds1da177e2005-04-16 15:20:36 -07002393
Andi Kleena5516432008-07-23 21:27:41 -07002394static int hugetlb_acct_memory(struct hstate *h, long delta)
Mel Gormanfc1b8a72008-07-23 21:27:22 -07002395{
2396 int ret = -ENOMEM;
2397
2398 spin_lock(&hugetlb_lock);
2399 /*
2400 * When cpuset is configured, it breaks the strict hugetlb page
2401 * reservation as the accounting is done on a global variable. Such
2402 * reservation is completely rubbish in the presence of cpuset because
2403 * the reservation is not checked against page availability for the
2404 * current cpuset. Application can still potentially OOM'ed by kernel
2405 * with lack of free htlb page in cpuset that the task is in.
2406 * Attempt to enforce strict accounting with cpuset is almost
2407 * impossible (or too ugly) because cpuset is too fluid that
2408 * task or memory node can be dynamically moved between cpusets.
2409 *
2410 * The change of semantics for shared hugetlb mapping with cpuset is
2411 * undesirable. However, in order to preserve some of the semantics,
2412 * we fall back to check against current free page availability as
2413 * a best attempt and hopefully to minimize the impact of changing
2414 * semantics that cpuset has.
2415 */
2416 if (delta > 0) {
Andi Kleena5516432008-07-23 21:27:41 -07002417 if (gather_surplus_pages(h, delta) < 0)
Mel Gormanfc1b8a72008-07-23 21:27:22 -07002418 goto out;
2419
Andi Kleena5516432008-07-23 21:27:41 -07002420 if (delta > cpuset_mems_nr(h->free_huge_pages_node)) {
2421 return_unused_surplus_pages(h, delta);
Mel Gormanfc1b8a72008-07-23 21:27:22 -07002422 goto out;
2423 }
2424 }
2425
2426 ret = 0;
2427 if (delta < 0)
Andi Kleena5516432008-07-23 21:27:41 -07002428 return_unused_surplus_pages(h, (unsigned long) -delta);
Mel Gormanfc1b8a72008-07-23 21:27:22 -07002429
2430out:
2431 spin_unlock(&hugetlb_lock);
2432 return ret;
2433}
2434
Andy Whitcroft84afd992008-07-23 21:27:32 -07002435static void hugetlb_vm_op_open(struct vm_area_struct *vma)
2436{
Joonsoo Kimf522c3a2013-09-11 14:21:53 -07002437 struct resv_map *resv = vma_resv_map(vma);
Andy Whitcroft84afd992008-07-23 21:27:32 -07002438
2439 /*
2440 * This new VMA should share its siblings reservation map if present.
2441 * The VMA will only ever have a valid reservation map pointer where
2442 * it is being copied for another still existing VMA. As that VMA
Lucas De Marchi25985ed2011-03-30 22:57:33 -03002443 * has a reference to the reservation map it cannot disappear until
Andy Whitcroft84afd992008-07-23 21:27:32 -07002444 * after this open call completes. It is therefore safe to take a
2445 * new reference here without additional locking.
2446 */
Joonsoo Kim4e35f482014-04-03 14:47:30 -07002447 if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
Joonsoo Kimf522c3a2013-09-11 14:21:53 -07002448 kref_get(&resv->refs);
Andy Whitcroft84afd992008-07-23 21:27:32 -07002449}
2450
Mel Gormana1e78772008-07-23 21:27:23 -07002451static void hugetlb_vm_op_close(struct vm_area_struct *vma)
2452{
Andi Kleena5516432008-07-23 21:27:41 -07002453 struct hstate *h = hstate_vma(vma);
Joonsoo Kimf522c3a2013-09-11 14:21:53 -07002454 struct resv_map *resv = vma_resv_map(vma);
David Gibson90481622012-03-21 16:34:12 -07002455 struct hugepage_subpool *spool = subpool_vma(vma);
Joonsoo Kim4e35f482014-04-03 14:47:30 -07002456 unsigned long reserve, start, end;
Andy Whitcroft84afd992008-07-23 21:27:32 -07002457
Joonsoo Kim4e35f482014-04-03 14:47:30 -07002458 if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER))
2459 return;
Andy Whitcroft84afd992008-07-23 21:27:32 -07002460
Joonsoo Kim4e35f482014-04-03 14:47:30 -07002461 start = vma_hugecache_offset(h, vma, vma->vm_start);
2462 end = vma_hugecache_offset(h, vma, vma->vm_end);
Andy Whitcroft84afd992008-07-23 21:27:32 -07002463
Joonsoo Kim4e35f482014-04-03 14:47:30 -07002464 reserve = (end - start) - region_count(resv, start, end);
Andy Whitcroft84afd992008-07-23 21:27:32 -07002465
Joonsoo Kim4e35f482014-04-03 14:47:30 -07002466 kref_put(&resv->refs, resv_map_release);
2467
2468 if (reserve) {
2469 hugetlb_acct_memory(h, -reserve);
2470 hugepage_subpool_put_pages(spool, reserve);
Andy Whitcroft84afd992008-07-23 21:27:32 -07002471 }
Mel Gormana1e78772008-07-23 21:27:23 -07002472}
2473
Linus Torvalds1da177e2005-04-16 15:20:36 -07002474/*
2475 * We cannot handle pagefaults against hugetlb pages at all. They cause
2476 * handle_mm_fault() to try to instantiate regular-sized pages in the
2477 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
2478 * this far.
2479 */
Nick Piggind0217ac2007-07-19 01:47:03 -07002480static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002481{
2482 BUG();
Nick Piggind0217ac2007-07-19 01:47:03 -07002483 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002484}
2485
Alexey Dobriyanf0f37e22009-09-27 22:29:37 +04002486const struct vm_operations_struct hugetlb_vm_ops = {
Nick Piggind0217ac2007-07-19 01:47:03 -07002487 .fault = hugetlb_vm_op_fault,
Andy Whitcroft84afd992008-07-23 21:27:32 -07002488 .open = hugetlb_vm_op_open,
Mel Gormana1e78772008-07-23 21:27:23 -07002489 .close = hugetlb_vm_op_close,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002490};
2491
David Gibson1e8f8892006-01-06 00:10:44 -08002492static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
2493 int writable)
David Gibson63551ae2005-06-21 17:14:44 -07002494{
2495 pte_t entry;
2496
David Gibson1e8f8892006-01-06 00:10:44 -08002497 if (writable) {
Gerald Schaefer106c9922013-04-29 15:07:23 -07002498 entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page,
2499 vma->vm_page_prot)));
David Gibson63551ae2005-06-21 17:14:44 -07002500 } else {
Gerald Schaefer106c9922013-04-29 15:07:23 -07002501 entry = huge_pte_wrprotect(mk_huge_pte(page,
2502 vma->vm_page_prot));
David Gibson63551ae2005-06-21 17:14:44 -07002503 }
2504 entry = pte_mkyoung(entry);
2505 entry = pte_mkhuge(entry);
Chris Metcalfd9ed9fa2012-04-01 14:01:34 -04002506 entry = arch_make_huge_pte(entry, vma, page, writable);
David Gibson63551ae2005-06-21 17:14:44 -07002507
2508 return entry;
2509}
2510
David Gibson1e8f8892006-01-06 00:10:44 -08002511static void set_huge_ptep_writable(struct vm_area_struct *vma,
2512 unsigned long address, pte_t *ptep)
2513{
2514 pte_t entry;
2515
Gerald Schaefer106c9922013-04-29 15:07:23 -07002516 entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep)));
Chris Forbes32f84522011-07-25 17:12:14 -07002517 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
Russell King4b3073e2009-12-18 16:40:18 +00002518 update_mmu_cache(vma, address, ptep);
David Gibson1e8f8892006-01-06 00:10:44 -08002519}
2520
Naoya Horiguchi4a705fe2014-06-23 13:22:03 -07002521static int is_hugetlb_entry_migration(pte_t pte)
2522{
2523 swp_entry_t swp;
2524
2525 if (huge_pte_none(pte) || pte_present(pte))
2526 return 0;
2527 swp = pte_to_swp_entry(pte);
2528 if (non_swap_entry(swp) && is_migration_entry(swp))
2529 return 1;
2530 else
2531 return 0;
2532}
2533
2534static int is_hugetlb_entry_hwpoisoned(pte_t pte)
2535{
2536 swp_entry_t swp;
2537
2538 if (huge_pte_none(pte) || pte_present(pte))
2539 return 0;
2540 swp = pte_to_swp_entry(pte);
2541 if (non_swap_entry(swp) && is_hwpoison_entry(swp))
2542 return 1;
2543 else
2544 return 0;
2545}
David Gibson1e8f8892006-01-06 00:10:44 -08002546
David Gibson63551ae2005-06-21 17:14:44 -07002547int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
2548 struct vm_area_struct *vma)
2549{
2550 pte_t *src_pte, *dst_pte, entry;
2551 struct page *ptepage;
Hugh Dickins1c598272005-10-19 21:23:43 -07002552 unsigned long addr;
David Gibson1e8f8892006-01-06 00:10:44 -08002553 int cow;
Andi Kleena5516432008-07-23 21:27:41 -07002554 struct hstate *h = hstate_vma(vma);
2555 unsigned long sz = huge_page_size(h);
Andreas Sandberge8569dd2014-01-21 15:49:09 -08002556 unsigned long mmun_start; /* For mmu_notifiers */
2557 unsigned long mmun_end; /* For mmu_notifiers */
2558 int ret = 0;
David Gibson1e8f8892006-01-06 00:10:44 -08002559
2560 cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
David Gibson63551ae2005-06-21 17:14:44 -07002561
Andreas Sandberge8569dd2014-01-21 15:49:09 -08002562 mmun_start = vma->vm_start;
2563 mmun_end = vma->vm_end;
2564 if (cow)
2565 mmu_notifier_invalidate_range_start(src, mmun_start, mmun_end);
2566
Andi Kleena5516432008-07-23 21:27:41 -07002567 for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002568 spinlock_t *src_ptl, *dst_ptl;
Hugh Dickinsc74df322005-10-29 18:16:23 -07002569 src_pte = huge_pte_offset(src, addr);
2570 if (!src_pte)
2571 continue;
Andi Kleena5516432008-07-23 21:27:41 -07002572 dst_pte = huge_pte_alloc(dst, addr, sz);
Andreas Sandberge8569dd2014-01-21 15:49:09 -08002573 if (!dst_pte) {
2574 ret = -ENOMEM;
2575 break;
2576 }
Larry Woodmanc5c99422008-01-24 05:49:25 -08002577
2578 /* If the pagetables are shared don't copy or take references */
2579 if (dst_pte == src_pte)
2580 continue;
2581
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002582 dst_ptl = huge_pte_lock(h, dst, dst_pte);
2583 src_ptl = huge_pte_lockptr(h, src, src_pte);
2584 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
Naoya Horiguchi4a705fe2014-06-23 13:22:03 -07002585 entry = huge_ptep_get(src_pte);
2586 if (huge_pte_none(entry)) { /* skip none entry */
2587 ;
2588 } else if (unlikely(is_hugetlb_entry_migration(entry) ||
2589 is_hugetlb_entry_hwpoisoned(entry))) {
2590 swp_entry_t swp_entry = pte_to_swp_entry(entry);
2591
2592 if (is_write_migration_entry(swp_entry) && cow) {
2593 /*
2594 * COW mappings require pages in both
2595 * parent and child to be set to read.
2596 */
2597 make_migration_entry_read(&swp_entry);
2598 entry = swp_entry_to_pte(swp_entry);
2599 set_huge_pte_at(src, addr, src_pte, entry);
2600 }
2601 set_huge_pte_at(dst, addr, dst_pte, entry);
2602 } else {
Joerg Roedel34ee6452014-11-13 13:46:09 +11002603 if (cow) {
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07002604 huge_ptep_set_wrprotect(src, addr, src_pte);
Joerg Roedel34ee6452014-11-13 13:46:09 +11002605 mmu_notifier_invalidate_range(src, mmun_start,
2606 mmun_end);
2607 }
Naoya Horiguchi0253d632014-07-23 14:00:19 -07002608 entry = huge_ptep_get(src_pte);
Hugh Dickins1c598272005-10-19 21:23:43 -07002609 ptepage = pte_page(entry);
2610 get_page(ptepage);
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002611 page_dup_rmap(ptepage);
Hugh Dickins1c598272005-10-19 21:23:43 -07002612 set_huge_pte_at(dst, addr, dst_pte, entry);
2613 }
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002614 spin_unlock(src_ptl);
2615 spin_unlock(dst_ptl);
David Gibson63551ae2005-06-21 17:14:44 -07002616 }
David Gibson63551ae2005-06-21 17:14:44 -07002617
Andreas Sandberge8569dd2014-01-21 15:49:09 -08002618 if (cow)
2619 mmu_notifier_invalidate_range_end(src, mmun_start, mmun_end);
2620
2621 return ret;
David Gibson63551ae2005-06-21 17:14:44 -07002622}
2623
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07002624void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
2625 unsigned long start, unsigned long end,
2626 struct page *ref_page)
David Gibson63551ae2005-06-21 17:14:44 -07002627{
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07002628 int force_flush = 0;
David Gibson63551ae2005-06-21 17:14:44 -07002629 struct mm_struct *mm = vma->vm_mm;
2630 unsigned long address;
David Gibsonc7546f82005-08-05 11:59:35 -07002631 pte_t *ptep;
David Gibson63551ae2005-06-21 17:14:44 -07002632 pte_t pte;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002633 spinlock_t *ptl;
David Gibson63551ae2005-06-21 17:14:44 -07002634 struct page *page;
Andi Kleena5516432008-07-23 21:27:41 -07002635 struct hstate *h = hstate_vma(vma);
2636 unsigned long sz = huge_page_size(h);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07002637 const unsigned long mmun_start = start; /* For mmu_notifiers */
2638 const unsigned long mmun_end = end; /* For mmu_notifiers */
Andi Kleena5516432008-07-23 21:27:41 -07002639
David Gibson63551ae2005-06-21 17:14:44 -07002640 WARN_ON(!is_vm_hugetlb_page(vma));
Andi Kleena5516432008-07-23 21:27:41 -07002641 BUG_ON(start & ~huge_page_mask(h));
2642 BUG_ON(end & ~huge_page_mask(h));
David Gibson63551ae2005-06-21 17:14:44 -07002643
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07002644 tlb_start_vma(tlb, vma);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07002645 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
Hillf Danton569f48b2014-12-10 15:44:41 -08002646 address = start;
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07002647again:
Hillf Danton569f48b2014-12-10 15:44:41 -08002648 for (; address < end; address += sz) {
David Gibsonc7546f82005-08-05 11:59:35 -07002649 ptep = huge_pte_offset(mm, address);
Adam Litke4c887262005-10-29 18:16:46 -07002650 if (!ptep)
David Gibsonc7546f82005-08-05 11:59:35 -07002651 continue;
2652
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002653 ptl = huge_pte_lock(h, mm, ptep);
Chen, Kenneth W39dde652006-12-06 20:32:03 -08002654 if (huge_pmd_unshare(mm, &address, ptep))
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002655 goto unlock;
Chen, Kenneth W39dde652006-12-06 20:32:03 -08002656
Hillf Danton66293262012-03-23 15:01:48 -07002657 pte = huge_ptep_get(ptep);
2658 if (huge_pte_none(pte))
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002659 goto unlock;
Hillf Danton66293262012-03-23 15:01:48 -07002660
2661 /*
Naoya Horiguchi9fbc1f62015-02-11 15:25:32 -08002662 * Migrating hugepage or HWPoisoned hugepage is already
2663 * unmapped and its refcount is dropped, so just clear pte here.
Hillf Danton66293262012-03-23 15:01:48 -07002664 */
Naoya Horiguchi9fbc1f62015-02-11 15:25:32 -08002665 if (unlikely(!pte_present(pte))) {
Gerald Schaefer106c9922013-04-29 15:07:23 -07002666 huge_pte_clear(mm, address, ptep);
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002667 goto unlock;
Naoya Horiguchi8c4894c2012-12-12 13:52:28 -08002668 }
Hillf Danton66293262012-03-23 15:01:48 -07002669
2670 page = pte_page(pte);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002671 /*
2672 * If a reference page is supplied, it is because a specific
2673 * page is being unmapped, not a range. Ensure the page we
2674 * are about to unmap is the actual page of interest.
2675 */
2676 if (ref_page) {
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002677 if (page != ref_page)
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002678 goto unlock;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002679
2680 /*
2681 * Mark the VMA as having unmapped its page so that
2682 * future faults in this VMA will fail rather than
2683 * looking like data was lost
2684 */
2685 set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
2686 }
2687
David Gibsonc7546f82005-08-05 11:59:35 -07002688 pte = huge_ptep_get_and_clear(mm, address, ptep);
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07002689 tlb_remove_tlb_entry(tlb, ptep, address);
Gerald Schaefer106c9922013-04-29 15:07:23 -07002690 if (huge_pte_dirty(pte))
Ken Chen6649a382007-02-08 14:20:27 -08002691 set_page_dirty(page);
Hillf Danton9e811302012-03-21 16:34:03 -07002692
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07002693 page_remove_rmap(page);
2694 force_flush = !__tlb_remove_page(tlb, page);
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002695 if (force_flush) {
Hillf Danton569f48b2014-12-10 15:44:41 -08002696 address += sz;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002697 spin_unlock(ptl);
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07002698 break;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002699 }
Hillf Danton9e811302012-03-21 16:34:03 -07002700 /* Bail out after unmapping reference page if supplied */
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002701 if (ref_page) {
2702 spin_unlock(ptl);
Hillf Danton9e811302012-03-21 16:34:03 -07002703 break;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002704 }
2705unlock:
2706 spin_unlock(ptl);
David Gibson63551ae2005-06-21 17:14:44 -07002707 }
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07002708 /*
2709 * mmu_gather ran out of room to batch pages, we break out of
2710 * the PTE lock to avoid doing the potential expensive TLB invalidate
2711 * and page-free while holding it.
2712 */
2713 if (force_flush) {
2714 force_flush = 0;
2715 tlb_flush_mmu(tlb);
2716 if (address < end && !ref_page)
2717 goto again;
Chen, Kenneth Wfe1668a2006-10-04 02:15:24 -07002718 }
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07002719 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07002720 tlb_end_vma(tlb, vma);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002721}
David Gibson63551ae2005-06-21 17:14:44 -07002722
Mel Gormand8333522012-07-31 16:46:20 -07002723void __unmap_hugepage_range_final(struct mmu_gather *tlb,
2724 struct vm_area_struct *vma, unsigned long start,
2725 unsigned long end, struct page *ref_page)
2726{
2727 __unmap_hugepage_range(tlb, vma, start, end, ref_page);
2728
2729 /*
2730 * Clear this flag so that x86's huge_pmd_share page_table_shareable
2731 * test will fail on a vma being torn down, and not grab a page table
2732 * on its way out. We're lucky that the flag has such an appropriate
2733 * name, and can in fact be safely cleared here. We could clear it
2734 * before the __unmap_hugepage_range above, but all that's necessary
Davidlohr Buesoc8c06ef2014-12-12 16:54:24 -08002735 * is to clear it before releasing the i_mmap_rwsem. This works
Mel Gormand8333522012-07-31 16:46:20 -07002736 * because in the context this is called, the VMA is about to be
Davidlohr Buesoc8c06ef2014-12-12 16:54:24 -08002737 * destroyed and the i_mmap_rwsem is held.
Mel Gormand8333522012-07-31 16:46:20 -07002738 */
2739 vma->vm_flags &= ~VM_MAYSHARE;
2740}
2741
Chen, Kenneth W502717f2006-10-11 01:20:46 -07002742void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002743 unsigned long end, struct page *ref_page)
Chen, Kenneth W502717f2006-10-11 01:20:46 -07002744{
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07002745 struct mm_struct *mm;
2746 struct mmu_gather tlb;
2747
2748 mm = vma->vm_mm;
2749
Linus Torvalds2b047252013-08-15 11:42:25 -07002750 tlb_gather_mmu(&tlb, mm, start, end);
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07002751 __unmap_hugepage_range(&tlb, vma, start, end, ref_page);
2752 tlb_finish_mmu(&tlb, start, end);
Chen, Kenneth W502717f2006-10-11 01:20:46 -07002753}
2754
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002755/*
2756 * This is called when the original mapper is failing to COW a MAP_PRIVATE
2757 * mappping it owns the reserve page for. The intention is to unmap the page
2758 * from other VMAs and let the children be SIGKILLed if they are faulting the
2759 * same region.
2760 */
Davidlohr Bueso2f4612a2014-08-06 16:06:45 -07002761static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
2762 struct page *page, unsigned long address)
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002763{
Adam Litke75266742008-11-12 13:24:56 -08002764 struct hstate *h = hstate_vma(vma);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002765 struct vm_area_struct *iter_vma;
2766 struct address_space *mapping;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002767 pgoff_t pgoff;
2768
2769 /*
2770 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
2771 * from page cache lookup which is in HPAGE_SIZE units.
2772 */
Adam Litke75266742008-11-12 13:24:56 -08002773 address = address & huge_page_mask(h);
Michal Hocko36e4f202012-10-08 16:33:31 -07002774 pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) +
2775 vma->vm_pgoff;
Al Viro496ad9a2013-01-23 17:07:38 -05002776 mapping = file_inode(vma->vm_file)->i_mapping;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002777
Mel Gorman4eb2b1d2009-12-14 17:59:53 -08002778 /*
2779 * Take the mapping lock for the duration of the table walk. As
2780 * this mapping should be shared between all the VMAs,
2781 * __unmap_hugepage_range() is called as the lock is already held
2782 */
Davidlohr Bueso83cde9e2014-12-12 16:54:21 -08002783 i_mmap_lock_write(mapping);
Michel Lespinasse6b2dbba2012-10-08 16:31:25 -07002784 vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) {
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002785 /* Do not unmap the current VMA */
2786 if (iter_vma == vma)
2787 continue;
2788
2789 /*
2790 * Unmap the page from other VMAs without their own reserves.
2791 * They get marked to be SIGKILLed if they fault in these
2792 * areas. This is because a future no-page fault on this VMA
2793 * could insert a zeroed page instead of the data existing
2794 * from the time of fork. This would look like data corruption
2795 */
2796 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07002797 unmap_hugepage_range(iter_vma, address,
2798 address + huge_page_size(h), page);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002799 }
Davidlohr Bueso83cde9e2014-12-12 16:54:21 -08002800 i_mmap_unlock_write(mapping);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002801}
2802
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002803/*
2804 * Hugetlb_cow() should be called with page lock of the original hugepage held.
Michal Hockoef009b22012-01-10 15:07:21 -08002805 * Called with hugetlb_instantiation_mutex held and pte_page locked so we
2806 * cannot race with other handlers or page migration.
2807 * Keep the pte_same checks anyway to make transition from the mutex easier.
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002808 */
David Gibson1e8f8892006-01-06 00:10:44 -08002809static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002810 unsigned long address, pte_t *ptep, pte_t pte,
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002811 struct page *pagecache_page, spinlock_t *ptl)
David Gibson1e8f8892006-01-06 00:10:44 -08002812{
Andi Kleena5516432008-07-23 21:27:41 -07002813 struct hstate *h = hstate_vma(vma);
David Gibson1e8f8892006-01-06 00:10:44 -08002814 struct page *old_page, *new_page;
Davidlohr Buesoad4404a2014-08-06 16:06:47 -07002815 int ret = 0, outside_reserve = 0;
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07002816 unsigned long mmun_start; /* For mmu_notifiers */
2817 unsigned long mmun_end; /* For mmu_notifiers */
David Gibson1e8f8892006-01-06 00:10:44 -08002818
2819 old_page = pte_page(pte);
2820
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002821retry_avoidcopy:
David Gibson1e8f8892006-01-06 00:10:44 -08002822 /* If no-one else is actually using this page, avoid the copy
2823 * and just make the page writable */
Joonsoo Kim37a21402013-09-11 14:21:04 -07002824 if (page_mapcount(old_page) == 1 && PageAnon(old_page)) {
2825 page_move_anon_rmap(old_page, vma, address);
David Gibson1e8f8892006-01-06 00:10:44 -08002826 set_huge_ptep_writable(vma, address, ptep);
Nick Piggin83c54072007-07-19 01:47:05 -07002827 return 0;
David Gibson1e8f8892006-01-06 00:10:44 -08002828 }
2829
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002830 /*
2831 * If the process that created a MAP_PRIVATE mapping is about to
2832 * perform a COW due to a shared page count, attempt to satisfy
2833 * the allocation without using the existing reserves. The pagecache
2834 * page is used to determine if the reserve at this address was
2835 * consumed or not. If reserves were used, a partial faulted mapping
2836 * at the time of fork() could consume its reserves on COW instead
2837 * of the full address range.
2838 */
Joonsoo Kim5944d012013-09-11 14:21:55 -07002839 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002840 old_page != pagecache_page)
2841 outside_reserve = 1;
2842
David Gibson1e8f8892006-01-06 00:10:44 -08002843 page_cache_get(old_page);
Larry Woodmanb76c8cf2009-12-14 17:59:37 -08002844
Davidlohr Buesoad4404a2014-08-06 16:06:47 -07002845 /*
2846 * Drop page table lock as buddy allocator may be called. It will
2847 * be acquired again before returning to the caller, as expected.
2848 */
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002849 spin_unlock(ptl);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002850 new_page = alloc_huge_page(vma, address, outside_reserve);
David Gibson1e8f8892006-01-06 00:10:44 -08002851
Adam Litke2fc39ce2007-11-14 16:59:39 -08002852 if (IS_ERR(new_page)) {
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002853 /*
2854 * If a process owning a MAP_PRIVATE mapping fails to COW,
2855 * it is due to references held by a child and an insufficient
2856 * huge page pool. To guarantee the original mappers
2857 * reliability, unmap the page from child processes. The child
2858 * may get SIGKILLed if it later faults.
2859 */
2860 if (outside_reserve) {
Davidlohr Buesoad4404a2014-08-06 16:06:47 -07002861 page_cache_release(old_page);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002862 BUG_ON(huge_pte_none(pte));
Davidlohr Bueso2f4612a2014-08-06 16:06:45 -07002863 unmap_ref_private(mm, vma, old_page, address);
2864 BUG_ON(huge_pte_none(pte));
2865 spin_lock(ptl);
2866 ptep = huge_pte_offset(mm, address & huge_page_mask(h));
2867 if (likely(ptep &&
2868 pte_same(huge_ptep_get(ptep), pte)))
2869 goto retry_avoidcopy;
2870 /*
2871 * race occurs while re-acquiring page table
2872 * lock, and our job is done.
2873 */
2874 return 0;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002875 }
2876
Davidlohr Buesoad4404a2014-08-06 16:06:47 -07002877 ret = (PTR_ERR(new_page) == -ENOMEM) ?
2878 VM_FAULT_OOM : VM_FAULT_SIGBUS;
2879 goto out_release_old;
David Gibson1e8f8892006-01-06 00:10:44 -08002880 }
2881
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002882 /*
2883 * When the original hugepage is shared one, it does not have
2884 * anon_vma prepared.
2885 */
Dean Nelson44e2aa92010-10-26 14:22:08 -07002886 if (unlikely(anon_vma_prepare(vma))) {
Davidlohr Buesoad4404a2014-08-06 16:06:47 -07002887 ret = VM_FAULT_OOM;
2888 goto out_release_all;
Dean Nelson44e2aa92010-10-26 14:22:08 -07002889 }
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002890
Andrea Arcangeli47ad8472011-01-13 15:46:47 -08002891 copy_user_huge_page(new_page, old_page, address, vma,
2892 pages_per_huge_page(h));
Nick Piggin0ed361d2008-02-04 22:29:34 -08002893 __SetPageUptodate(new_page);
David Gibson1e8f8892006-01-06 00:10:44 -08002894
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07002895 mmun_start = address & huge_page_mask(h);
2896 mmun_end = mmun_start + huge_page_size(h);
2897 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
Davidlohr Buesoad4404a2014-08-06 16:06:47 -07002898
Larry Woodmanb76c8cf2009-12-14 17:59:37 -08002899 /*
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002900 * Retake the page table lock to check for racing updates
Larry Woodmanb76c8cf2009-12-14 17:59:37 -08002901 * before the page tables are altered
2902 */
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002903 spin_lock(ptl);
Andi Kleena5516432008-07-23 21:27:41 -07002904 ptep = huge_pte_offset(mm, address & huge_page_mask(h));
Naoya Horiguchia9af0c52014-04-07 15:36:54 -07002905 if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) {
Joonsoo Kim07443a82013-09-11 14:21:58 -07002906 ClearPagePrivate(new_page);
2907
David Gibson1e8f8892006-01-06 00:10:44 -08002908 /* Break COW */
Gerald Schaefer8fe627e2008-04-28 02:13:28 -07002909 huge_ptep_clear_flush(vma, address, ptep);
Joerg Roedel34ee6452014-11-13 13:46:09 +11002910 mmu_notifier_invalidate_range(mm, mmun_start, mmun_end);
David Gibson1e8f8892006-01-06 00:10:44 -08002911 set_huge_pte_at(mm, address, ptep,
2912 make_huge_pte(vma, new_page, 1));
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002913 page_remove_rmap(old_page);
Naoya Horiguchicd67f0d2010-09-10 13:23:04 +09002914 hugepage_add_new_anon_rmap(new_page, vma, address);
David Gibson1e8f8892006-01-06 00:10:44 -08002915 /* Make the old page be freed below */
2916 new_page = old_page;
2917 }
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002918 spin_unlock(ptl);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07002919 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
Davidlohr Buesoad4404a2014-08-06 16:06:47 -07002920out_release_all:
David Gibson1e8f8892006-01-06 00:10:44 -08002921 page_cache_release(new_page);
Davidlohr Buesoad4404a2014-08-06 16:06:47 -07002922out_release_old:
David Gibson1e8f8892006-01-06 00:10:44 -08002923 page_cache_release(old_page);
Joonsoo Kim83120342013-09-11 14:21:57 -07002924
Davidlohr Buesoad4404a2014-08-06 16:06:47 -07002925 spin_lock(ptl); /* Caller expects lock to be held */
2926 return ret;
David Gibson1e8f8892006-01-06 00:10:44 -08002927}
2928
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002929/* Return the pagecache page at a given address within a VMA */
Andi Kleena5516432008-07-23 21:27:41 -07002930static struct page *hugetlbfs_pagecache_page(struct hstate *h,
2931 struct vm_area_struct *vma, unsigned long address)
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002932{
2933 struct address_space *mapping;
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -07002934 pgoff_t idx;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002935
2936 mapping = vma->vm_file->f_mapping;
Andi Kleena5516432008-07-23 21:27:41 -07002937 idx = vma_hugecache_offset(h, vma, address);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002938
2939 return find_lock_page(mapping, idx);
2940}
2941
Hugh Dickins3ae77f42009-09-21 17:03:33 -07002942/*
2943 * Return whether there is a pagecache page to back given address within VMA.
2944 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
2945 */
2946static bool hugetlbfs_pagecache_present(struct hstate *h,
Hugh Dickins2a15efc2009-09-21 17:03:27 -07002947 struct vm_area_struct *vma, unsigned long address)
2948{
2949 struct address_space *mapping;
2950 pgoff_t idx;
2951 struct page *page;
2952
2953 mapping = vma->vm_file->f_mapping;
2954 idx = vma_hugecache_offset(h, vma, address);
2955
2956 page = find_get_page(mapping, idx);
2957 if (page)
2958 put_page(page);
2959 return page != NULL;
2960}
2961
Robert P. J. Daya1ed3dd2007-07-17 04:03:33 -07002962static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
Davidlohr Bueso8382d912014-04-03 14:47:31 -07002963 struct address_space *mapping, pgoff_t idx,
2964 unsigned long address, pte_t *ptep, unsigned int flags)
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01002965{
Andi Kleena5516432008-07-23 21:27:41 -07002966 struct hstate *h = hstate_vma(vma);
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01002967 int ret = VM_FAULT_SIGBUS;
Hillf Danton409eb8c2012-01-20 14:34:13 -08002968 int anon_rmap = 0;
Adam Litke4c887262005-10-29 18:16:46 -07002969 unsigned long size;
Adam Litke4c887262005-10-29 18:16:46 -07002970 struct page *page;
David Gibson1e8f8892006-01-06 00:10:44 -08002971 pte_t new_pte;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002972 spinlock_t *ptl;
Adam Litke4c887262005-10-29 18:16:46 -07002973
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002974 /*
2975 * Currently, we are forced to kill the process in the event the
2976 * original mapper has unmapped pages from the child due to a failed
Lucas De Marchi25985ed2011-03-30 22:57:33 -03002977 * COW. Warn that such a situation has occurred as it may not be obvious
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002978 */
2979 if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
Andrew Mortonffb22af2013-02-22 16:32:08 -08002980 pr_warning("PID %d killed due to inadequate hugepage pool\n",
2981 current->pid);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002982 return ret;
2983 }
2984
Adam Litke4c887262005-10-29 18:16:46 -07002985 /*
2986 * Use page lock to guard against racing truncation
2987 * before we get page_table_lock.
2988 */
Christoph Lameter6bda6662006-01-06 00:10:49 -08002989retry:
2990 page = find_lock_page(mapping, idx);
2991 if (!page) {
Andi Kleena5516432008-07-23 21:27:41 -07002992 size = i_size_read(mapping->host) >> huge_page_shift(h);
Hugh Dickinsebed4bf2006-10-28 10:38:43 -07002993 if (idx >= size)
2994 goto out;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002995 page = alloc_huge_page(vma, address, 0);
Adam Litke2fc39ce2007-11-14 16:59:39 -08002996 if (IS_ERR(page)) {
Aneesh Kumar K.V76dcee72012-07-31 16:41:57 -07002997 ret = PTR_ERR(page);
2998 if (ret == -ENOMEM)
2999 ret = VM_FAULT_OOM;
3000 else
3001 ret = VM_FAULT_SIGBUS;
Christoph Lameter6bda6662006-01-06 00:10:49 -08003002 goto out;
3003 }
Andrea Arcangeli47ad8472011-01-13 15:46:47 -08003004 clear_huge_page(page, address, pages_per_huge_page(h));
Nick Piggin0ed361d2008-02-04 22:29:34 -08003005 __SetPageUptodate(page);
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01003006
Mel Gormanf83a2752009-05-28 14:34:40 -07003007 if (vma->vm_flags & VM_MAYSHARE) {
Christoph Lameter6bda6662006-01-06 00:10:49 -08003008 int err;
Ken Chen45c682a2007-11-14 16:59:44 -08003009 struct inode *inode = mapping->host;
Christoph Lameter6bda6662006-01-06 00:10:49 -08003010
3011 err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
3012 if (err) {
3013 put_page(page);
Christoph Lameter6bda6662006-01-06 00:10:49 -08003014 if (err == -EEXIST)
3015 goto retry;
3016 goto out;
3017 }
Joonsoo Kim07443a82013-09-11 14:21:58 -07003018 ClearPagePrivate(page);
Ken Chen45c682a2007-11-14 16:59:44 -08003019
3020 spin_lock(&inode->i_lock);
Andi Kleena5516432008-07-23 21:27:41 -07003021 inode->i_blocks += blocks_per_huge_page(h);
Ken Chen45c682a2007-11-14 16:59:44 -08003022 spin_unlock(&inode->i_lock);
Mel Gorman23be7462010-04-23 13:17:56 -04003023 } else {
Christoph Lameter6bda6662006-01-06 00:10:49 -08003024 lock_page(page);
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09003025 if (unlikely(anon_vma_prepare(vma))) {
3026 ret = VM_FAULT_OOM;
3027 goto backout_unlocked;
3028 }
Hillf Danton409eb8c2012-01-20 14:34:13 -08003029 anon_rmap = 1;
Mel Gorman23be7462010-04-23 13:17:56 -04003030 }
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09003031 } else {
Naoya Horiguchi998b4382010-09-08 10:19:32 +09003032 /*
3033 * If memory error occurs between mmap() and fault, some process
3034 * don't have hwpoisoned swap entry for errored virtual address.
3035 * So we need to block hugepage fault by PG_hwpoison bit check.
3036 */
3037 if (unlikely(PageHWPoison(page))) {
Chris Forbes32f84522011-07-25 17:12:14 -07003038 ret = VM_FAULT_HWPOISON |
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07003039 VM_FAULT_SET_HINDEX(hstate_index(h));
Naoya Horiguchi998b4382010-09-08 10:19:32 +09003040 goto backout_unlocked;
3041 }
Christoph Lameter6bda6662006-01-06 00:10:49 -08003042 }
David Gibson1e8f8892006-01-06 00:10:44 -08003043
Andy Whitcroft57303d82008-08-12 15:08:47 -07003044 /*
3045 * If we are going to COW a private mapping later, we examine the
3046 * pending reservations for this page now. This will ensure that
3047 * any allocations necessary to record that reservation occur outside
3048 * the spinlock.
3049 */
Hugh Dickins788c7df2009-06-23 13:49:05 +01003050 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED))
Andy Whitcroft2b267362008-08-12 15:08:49 -07003051 if (vma_needs_reservation(h, vma, address) < 0) {
3052 ret = VM_FAULT_OOM;
3053 goto backout_unlocked;
3054 }
Andy Whitcroft57303d82008-08-12 15:08:47 -07003055
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003056 ptl = huge_pte_lockptr(h, mm, ptep);
3057 spin_lock(ptl);
Andi Kleena5516432008-07-23 21:27:41 -07003058 size = i_size_read(mapping->host) >> huge_page_shift(h);
Adam Litke4c887262005-10-29 18:16:46 -07003059 if (idx >= size)
3060 goto backout;
3061
Nick Piggin83c54072007-07-19 01:47:05 -07003062 ret = 0;
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07003063 if (!huge_pte_none(huge_ptep_get(ptep)))
Adam Litke4c887262005-10-29 18:16:46 -07003064 goto backout;
3065
Joonsoo Kim07443a82013-09-11 14:21:58 -07003066 if (anon_rmap) {
3067 ClearPagePrivate(page);
Hillf Danton409eb8c2012-01-20 14:34:13 -08003068 hugepage_add_new_anon_rmap(page, vma, address);
Choi Gi-yongac714902014-04-07 15:37:36 -07003069 } else
Hillf Danton409eb8c2012-01-20 14:34:13 -08003070 page_dup_rmap(page);
David Gibson1e8f8892006-01-06 00:10:44 -08003071 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
3072 && (vma->vm_flags & VM_SHARED)));
3073 set_huge_pte_at(mm, address, ptep, new_pte);
3074
Hugh Dickins788c7df2009-06-23 13:49:05 +01003075 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
David Gibson1e8f8892006-01-06 00:10:44 -08003076 /* Optimization, do the COW without a second fault */
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003077 ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page, ptl);
David Gibson1e8f8892006-01-06 00:10:44 -08003078 }
3079
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003080 spin_unlock(ptl);
Adam Litke4c887262005-10-29 18:16:46 -07003081 unlock_page(page);
3082out:
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01003083 return ret;
Adam Litke4c887262005-10-29 18:16:46 -07003084
3085backout:
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003086 spin_unlock(ptl);
Andy Whitcroft2b267362008-08-12 15:08:49 -07003087backout_unlocked:
Adam Litke4c887262005-10-29 18:16:46 -07003088 unlock_page(page);
3089 put_page(page);
3090 goto out;
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01003091}
3092
Davidlohr Bueso8382d912014-04-03 14:47:31 -07003093#ifdef CONFIG_SMP
3094static u32 fault_mutex_hash(struct hstate *h, struct mm_struct *mm,
3095 struct vm_area_struct *vma,
3096 struct address_space *mapping,
3097 pgoff_t idx, unsigned long address)
3098{
3099 unsigned long key[2];
3100 u32 hash;
3101
3102 if (vma->vm_flags & VM_SHARED) {
3103 key[0] = (unsigned long) mapping;
3104 key[1] = idx;
3105 } else {
3106 key[0] = (unsigned long) mm;
3107 key[1] = address >> huge_page_shift(h);
3108 }
3109
3110 hash = jhash2((u32 *)&key, sizeof(key)/sizeof(u32), 0);
3111
3112 return hash & (num_fault_mutexes - 1);
3113}
3114#else
3115/*
3116 * For uniprocesor systems we always use a single mutex, so just
3117 * return 0 and avoid the hashing overhead.
3118 */
3119static u32 fault_mutex_hash(struct hstate *h, struct mm_struct *mm,
3120 struct vm_area_struct *vma,
3121 struct address_space *mapping,
3122 pgoff_t idx, unsigned long address)
3123{
3124 return 0;
3125}
3126#endif
3127
Adam Litke86e52162006-01-06 00:10:43 -08003128int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
Hugh Dickins788c7df2009-06-23 13:49:05 +01003129 unsigned long address, unsigned int flags)
Adam Litke86e52162006-01-06 00:10:43 -08003130{
Davidlohr Bueso8382d912014-04-03 14:47:31 -07003131 pte_t *ptep, entry;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003132 spinlock_t *ptl;
David Gibson1e8f8892006-01-06 00:10:44 -08003133 int ret;
Davidlohr Bueso8382d912014-04-03 14:47:31 -07003134 u32 hash;
3135 pgoff_t idx;
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09003136 struct page *page = NULL;
Andy Whitcroft57303d82008-08-12 15:08:47 -07003137 struct page *pagecache_page = NULL;
Andi Kleena5516432008-07-23 21:27:41 -07003138 struct hstate *h = hstate_vma(vma);
Davidlohr Bueso8382d912014-04-03 14:47:31 -07003139 struct address_space *mapping;
Naoya Horiguchi0f792cf2015-02-11 15:25:25 -08003140 int need_wait_lock = 0;
Adam Litke86e52162006-01-06 00:10:43 -08003141
KAMEZAWA Hiroyuki1e16a532012-01-10 15:07:22 -08003142 address &= huge_page_mask(h);
3143
Naoya Horiguchifd6a03e2010-05-28 09:29:21 +09003144 ptep = huge_pte_offset(mm, address);
3145 if (ptep) {
3146 entry = huge_ptep_get(ptep);
Naoya Horiguchi290408d2010-09-08 10:19:35 +09003147 if (unlikely(is_hugetlb_entry_migration(entry))) {
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003148 migration_entry_wait_huge(vma, mm, ptep);
Naoya Horiguchi290408d2010-09-08 10:19:35 +09003149 return 0;
3150 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
Chris Forbes32f84522011-07-25 17:12:14 -07003151 return VM_FAULT_HWPOISON_LARGE |
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07003152 VM_FAULT_SET_HINDEX(hstate_index(h));
Naoya Horiguchifd6a03e2010-05-28 09:29:21 +09003153 }
3154
Andi Kleena5516432008-07-23 21:27:41 -07003155 ptep = huge_pte_alloc(mm, address, huge_page_size(h));
Adam Litke86e52162006-01-06 00:10:43 -08003156 if (!ptep)
3157 return VM_FAULT_OOM;
3158
Davidlohr Bueso8382d912014-04-03 14:47:31 -07003159 mapping = vma->vm_file->f_mapping;
3160 idx = vma_hugecache_offset(h, vma, address);
3161
David Gibson3935baa2006-03-22 00:08:53 -08003162 /*
3163 * Serialize hugepage allocation and instantiation, so that we don't
3164 * get spurious allocation failures if two CPUs race to instantiate
3165 * the same page in the page cache.
3166 */
Davidlohr Bueso8382d912014-04-03 14:47:31 -07003167 hash = fault_mutex_hash(h, mm, vma, mapping, idx, address);
3168 mutex_lock(&htlb_fault_mutex_table[hash]);
3169
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07003170 entry = huge_ptep_get(ptep);
3171 if (huge_pte_none(entry)) {
Davidlohr Bueso8382d912014-04-03 14:47:31 -07003172 ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags);
David Gibsonb4d1d992008-10-15 22:01:11 -07003173 goto out_mutex;
David Gibson3935baa2006-03-22 00:08:53 -08003174 }
Adam Litke86e52162006-01-06 00:10:43 -08003175
Nick Piggin83c54072007-07-19 01:47:05 -07003176 ret = 0;
David Gibson1e8f8892006-01-06 00:10:44 -08003177
Andy Whitcroft57303d82008-08-12 15:08:47 -07003178 /*
Naoya Horiguchi0f792cf2015-02-11 15:25:25 -08003179 * entry could be a migration/hwpoison entry at this point, so this
3180 * check prevents the kernel from going below assuming that we have
3181 * a active hugepage in pagecache. This goto expects the 2nd page fault,
3182 * and is_hugetlb_entry_(migration|hwpoisoned) check will properly
3183 * handle it.
3184 */
3185 if (!pte_present(entry))
3186 goto out_mutex;
3187
3188 /*
Andy Whitcroft57303d82008-08-12 15:08:47 -07003189 * If we are going to COW the mapping later, we examine the pending
3190 * reservations for this page now. This will ensure that any
3191 * allocations necessary to record that reservation occur outside the
3192 * spinlock. For private mappings, we also lookup the pagecache
3193 * page now as it is used to determine if a reservation has been
3194 * consumed.
3195 */
Gerald Schaefer106c9922013-04-29 15:07:23 -07003196 if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) {
Andy Whitcroft2b267362008-08-12 15:08:49 -07003197 if (vma_needs_reservation(h, vma, address) < 0) {
3198 ret = VM_FAULT_OOM;
David Gibsonb4d1d992008-10-15 22:01:11 -07003199 goto out_mutex;
Andy Whitcroft2b267362008-08-12 15:08:49 -07003200 }
Andy Whitcroft57303d82008-08-12 15:08:47 -07003201
Mel Gormanf83a2752009-05-28 14:34:40 -07003202 if (!(vma->vm_flags & VM_MAYSHARE))
Andy Whitcroft57303d82008-08-12 15:08:47 -07003203 pagecache_page = hugetlbfs_pagecache_page(h,
3204 vma, address);
3205 }
3206
Naoya Horiguchi0f792cf2015-02-11 15:25:25 -08003207 ptl = huge_pte_lock(h, mm, ptep);
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09003208
David Gibson1e8f8892006-01-06 00:10:44 -08003209 /* Check for a racing update before calling hugetlb_cow */
David Gibsonb4d1d992008-10-15 22:01:11 -07003210 if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003211 goto out_ptl;
David Gibsonb4d1d992008-10-15 22:01:11 -07003212
Naoya Horiguchi0f792cf2015-02-11 15:25:25 -08003213 /*
3214 * hugetlb_cow() requires page locks of pte_page(entry) and
3215 * pagecache_page, so here we need take the former one
3216 * when page != pagecache_page or !pagecache_page.
3217 */
3218 page = pte_page(entry);
3219 if (page != pagecache_page)
3220 if (!trylock_page(page)) {
3221 need_wait_lock = 1;
3222 goto out_ptl;
3223 }
3224
3225 get_page(page);
David Gibsonb4d1d992008-10-15 22:01:11 -07003226
Hugh Dickins788c7df2009-06-23 13:49:05 +01003227 if (flags & FAULT_FLAG_WRITE) {
Gerald Schaefer106c9922013-04-29 15:07:23 -07003228 if (!huge_pte_write(entry)) {
Andy Whitcroft57303d82008-08-12 15:08:47 -07003229 ret = hugetlb_cow(mm, vma, address, ptep, entry,
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003230 pagecache_page, ptl);
Naoya Horiguchi0f792cf2015-02-11 15:25:25 -08003231 goto out_put_page;
David Gibsonb4d1d992008-10-15 22:01:11 -07003232 }
Gerald Schaefer106c9922013-04-29 15:07:23 -07003233 entry = huge_pte_mkdirty(entry);
David Gibsonb4d1d992008-10-15 22:01:11 -07003234 }
3235 entry = pte_mkyoung(entry);
Hugh Dickins788c7df2009-06-23 13:49:05 +01003236 if (huge_ptep_set_access_flags(vma, address, ptep, entry,
3237 flags & FAULT_FLAG_WRITE))
Russell King4b3073e2009-12-18 16:40:18 +00003238 update_mmu_cache(vma, address, ptep);
Naoya Horiguchi0f792cf2015-02-11 15:25:25 -08003239out_put_page:
3240 if (page != pagecache_page)
3241 unlock_page(page);
3242 put_page(page);
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003243out_ptl:
3244 spin_unlock(ptl);
Andy Whitcroft57303d82008-08-12 15:08:47 -07003245
3246 if (pagecache_page) {
3247 unlock_page(pagecache_page);
3248 put_page(pagecache_page);
3249 }
David Gibsonb4d1d992008-10-15 22:01:11 -07003250out_mutex:
Davidlohr Bueso8382d912014-04-03 14:47:31 -07003251 mutex_unlock(&htlb_fault_mutex_table[hash]);
Naoya Horiguchi0f792cf2015-02-11 15:25:25 -08003252 /*
3253 * Generally it's safe to hold refcount during waiting page lock. But
3254 * here we just wait to defer the next page fault to avoid busy loop and
3255 * the page is not used after unlocked before returning from the current
3256 * page fault. So we are safe from accessing freed page, even if we wait
3257 * here without taking refcount.
3258 */
3259 if (need_wait_lock)
3260 wait_on_page_locked(page);
David Gibson1e8f8892006-01-06 00:10:44 -08003261 return ret;
Adam Litke86e52162006-01-06 00:10:43 -08003262}
3263
Michel Lespinasse28a35712013-02-22 16:35:55 -08003264long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
3265 struct page **pages, struct vm_area_struct **vmas,
3266 unsigned long *position, unsigned long *nr_pages,
3267 long i, unsigned int flags)
David Gibson63551ae2005-06-21 17:14:44 -07003268{
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08003269 unsigned long pfn_offset;
3270 unsigned long vaddr = *position;
Michel Lespinasse28a35712013-02-22 16:35:55 -08003271 unsigned long remainder = *nr_pages;
Andi Kleena5516432008-07-23 21:27:41 -07003272 struct hstate *h = hstate_vma(vma);
David Gibson63551ae2005-06-21 17:14:44 -07003273
David Gibson63551ae2005-06-21 17:14:44 -07003274 while (vaddr < vma->vm_end && remainder) {
Adam Litke4c887262005-10-29 18:16:46 -07003275 pte_t *pte;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003276 spinlock_t *ptl = NULL;
Hugh Dickins2a15efc2009-09-21 17:03:27 -07003277 int absent;
Adam Litke4c887262005-10-29 18:16:46 -07003278 struct page *page;
3279
3280 /*
3281 * Some archs (sparc64, sh*) have multiple pte_ts to
Hugh Dickins2a15efc2009-09-21 17:03:27 -07003282 * each hugepage. We have to make sure we get the
Adam Litke4c887262005-10-29 18:16:46 -07003283 * first, for the page indexing below to work.
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003284 *
3285 * Note that page table lock is not held when pte is null.
Adam Litke4c887262005-10-29 18:16:46 -07003286 */
Andi Kleena5516432008-07-23 21:27:41 -07003287 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h));
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003288 if (pte)
3289 ptl = huge_pte_lock(h, mm, pte);
Hugh Dickins2a15efc2009-09-21 17:03:27 -07003290 absent = !pte || huge_pte_none(huge_ptep_get(pte));
Adam Litke4c887262005-10-29 18:16:46 -07003291
Hugh Dickins2a15efc2009-09-21 17:03:27 -07003292 /*
3293 * When coredumping, it suits get_dump_page if we just return
Hugh Dickins3ae77f42009-09-21 17:03:33 -07003294 * an error where there's an empty slot with no huge pagecache
3295 * to back it. This way, we avoid allocating a hugepage, and
3296 * the sparse dumpfile avoids allocating disk blocks, but its
3297 * huge holes still show up with zeroes where they need to be.
Hugh Dickins2a15efc2009-09-21 17:03:27 -07003298 */
Hugh Dickins3ae77f42009-09-21 17:03:33 -07003299 if (absent && (flags & FOLL_DUMP) &&
3300 !hugetlbfs_pagecache_present(h, vma, vaddr)) {
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003301 if (pte)
3302 spin_unlock(ptl);
Hugh Dickins2a15efc2009-09-21 17:03:27 -07003303 remainder = 0;
3304 break;
3305 }
3306
Naoya Horiguchi9cc3a5b2013-04-17 15:58:30 -07003307 /*
3308 * We need call hugetlb_fault for both hugepages under migration
3309 * (in which case hugetlb_fault waits for the migration,) and
3310 * hwpoisoned hugepages (in which case we need to prevent the
3311 * caller from accessing to them.) In order to do this, we use
3312 * here is_swap_pte instead of is_hugetlb_entry_migration and
3313 * is_hugetlb_entry_hwpoisoned. This is because it simply covers
3314 * both cases, and because we can't follow correct pages
3315 * directly from any kind of swap entries.
3316 */
3317 if (absent || is_swap_pte(huge_ptep_get(pte)) ||
Gerald Schaefer106c9922013-04-29 15:07:23 -07003318 ((flags & FOLL_WRITE) &&
3319 !huge_pte_write(huge_ptep_get(pte)))) {
Adam Litke4c887262005-10-29 18:16:46 -07003320 int ret;
3321
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003322 if (pte)
3323 spin_unlock(ptl);
Hugh Dickins2a15efc2009-09-21 17:03:27 -07003324 ret = hugetlb_fault(mm, vma, vaddr,
3325 (flags & FOLL_WRITE) ? FAULT_FLAG_WRITE : 0);
Adam Litkea89182c2007-08-22 14:01:51 -07003326 if (!(ret & VM_FAULT_ERROR))
Adam Litke4c887262005-10-29 18:16:46 -07003327 continue;
3328
3329 remainder = 0;
Adam Litke4c887262005-10-29 18:16:46 -07003330 break;
3331 }
David Gibson63551ae2005-06-21 17:14:44 -07003332
Andi Kleena5516432008-07-23 21:27:41 -07003333 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07003334 page = pte_page(huge_ptep_get(pte));
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08003335same_page:
Chen, Kenneth Wd6692182006-03-31 02:29:57 -08003336 if (pages) {
Hugh Dickins2a15efc2009-09-21 17:03:27 -07003337 pages[i] = mem_map_offset(page, pfn_offset);
Andrea Arcangelia0368d42014-01-21 15:48:49 -08003338 get_page_foll(pages[i]);
Chen, Kenneth Wd6692182006-03-31 02:29:57 -08003339 }
David Gibson63551ae2005-06-21 17:14:44 -07003340
3341 if (vmas)
3342 vmas[i] = vma;
3343
3344 vaddr += PAGE_SIZE;
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08003345 ++pfn_offset;
David Gibson63551ae2005-06-21 17:14:44 -07003346 --remainder;
3347 ++i;
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08003348 if (vaddr < vma->vm_end && remainder &&
Andi Kleena5516432008-07-23 21:27:41 -07003349 pfn_offset < pages_per_huge_page(h)) {
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08003350 /*
3351 * We use pfn_offset to avoid touching the pageframes
3352 * of this compound page.
3353 */
3354 goto same_page;
3355 }
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003356 spin_unlock(ptl);
David Gibson63551ae2005-06-21 17:14:44 -07003357 }
Michel Lespinasse28a35712013-02-22 16:35:55 -08003358 *nr_pages = remainder;
David Gibson63551ae2005-06-21 17:14:44 -07003359 *position = vaddr;
3360
Hugh Dickins2a15efc2009-09-21 17:03:27 -07003361 return i ? i : -EFAULT;
David Gibson63551ae2005-06-21 17:14:44 -07003362}
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003363
Peter Zijlstra7da4d642012-11-19 03:14:23 +01003364unsigned long hugetlb_change_protection(struct vm_area_struct *vma,
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003365 unsigned long address, unsigned long end, pgprot_t newprot)
3366{
3367 struct mm_struct *mm = vma->vm_mm;
3368 unsigned long start = address;
3369 pte_t *ptep;
3370 pte_t pte;
Andi Kleena5516432008-07-23 21:27:41 -07003371 struct hstate *h = hstate_vma(vma);
Peter Zijlstra7da4d642012-11-19 03:14:23 +01003372 unsigned long pages = 0;
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003373
3374 BUG_ON(address >= end);
3375 flush_cache_range(vma, address, end);
3376
Rik van Riela5338092014-04-07 15:36:57 -07003377 mmu_notifier_invalidate_range_start(mm, start, end);
Davidlohr Bueso83cde9e2014-12-12 16:54:21 -08003378 i_mmap_lock_write(vma->vm_file->f_mapping);
Andi Kleena5516432008-07-23 21:27:41 -07003379 for (; address < end; address += huge_page_size(h)) {
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003380 spinlock_t *ptl;
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003381 ptep = huge_pte_offset(mm, address);
3382 if (!ptep)
3383 continue;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003384 ptl = huge_pte_lock(h, mm, ptep);
Peter Zijlstra7da4d642012-11-19 03:14:23 +01003385 if (huge_pmd_unshare(mm, &address, ptep)) {
3386 pages++;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003387 spin_unlock(ptl);
Chen, Kenneth W39dde652006-12-06 20:32:03 -08003388 continue;
Peter Zijlstra7da4d642012-11-19 03:14:23 +01003389 }
Naoya Horiguchia8bda282015-02-11 15:25:28 -08003390 pte = huge_ptep_get(ptep);
3391 if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) {
3392 spin_unlock(ptl);
3393 continue;
3394 }
3395 if (unlikely(is_hugetlb_entry_migration(pte))) {
3396 swp_entry_t entry = pte_to_swp_entry(pte);
3397
3398 if (is_write_migration_entry(entry)) {
3399 pte_t newpte;
3400
3401 make_migration_entry_read(&entry);
3402 newpte = swp_entry_to_pte(entry);
3403 set_huge_pte_at(mm, address, ptep, newpte);
3404 pages++;
3405 }
3406 spin_unlock(ptl);
3407 continue;
3408 }
3409 if (!huge_pte_none(pte)) {
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003410 pte = huge_ptep_get_and_clear(mm, address, ptep);
Gerald Schaefer106c9922013-04-29 15:07:23 -07003411 pte = pte_mkhuge(huge_pte_modify(pte, newprot));
Tony Lube7517d2013-02-04 14:28:46 -08003412 pte = arch_make_huge_pte(pte, vma, NULL, 0);
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003413 set_huge_pte_at(mm, address, ptep, pte);
Peter Zijlstra7da4d642012-11-19 03:14:23 +01003414 pages++;
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003415 }
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003416 spin_unlock(ptl);
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003417 }
Mel Gormand8333522012-07-31 16:46:20 -07003418 /*
Davidlohr Buesoc8c06ef2014-12-12 16:54:24 -08003419 * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare
Mel Gormand8333522012-07-31 16:46:20 -07003420 * may have cleared our pud entry and done put_page on the page table:
Davidlohr Buesoc8c06ef2014-12-12 16:54:24 -08003421 * once we release i_mmap_rwsem, another task can do the final put_page
Mel Gormand8333522012-07-31 16:46:20 -07003422 * and that page table be reused and filled with junk.
3423 */
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003424 flush_tlb_range(vma, start, end);
Joerg Roedel34ee6452014-11-13 13:46:09 +11003425 mmu_notifier_invalidate_range(mm, start, end);
Davidlohr Bueso83cde9e2014-12-12 16:54:21 -08003426 i_mmap_unlock_write(vma->vm_file->f_mapping);
Rik van Riela5338092014-04-07 15:36:57 -07003427 mmu_notifier_invalidate_range_end(mm, start, end);
Peter Zijlstra7da4d642012-11-19 03:14:23 +01003428
3429 return pages << h->order;
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003430}
3431
Mel Gormana1e78772008-07-23 21:27:23 -07003432int hugetlb_reserve_pages(struct inode *inode,
3433 long from, long to,
Mel Gorman5a6fe122009-02-10 14:02:27 +00003434 struct vm_area_struct *vma,
KOSAKI Motohiroca16d142011-05-26 19:16:19 +09003435 vm_flags_t vm_flags)
Adam Litkee4e574b2007-10-16 01:26:19 -07003436{
Mel Gorman17c9d122009-02-11 16:34:16 +00003437 long ret, chg;
Andi Kleena5516432008-07-23 21:27:41 -07003438 struct hstate *h = hstate_inode(inode);
David Gibson90481622012-03-21 16:34:12 -07003439 struct hugepage_subpool *spool = subpool_inode(inode);
Joonsoo Kim9119a412014-04-03 14:47:25 -07003440 struct resv_map *resv_map;
Adam Litkee4e574b2007-10-16 01:26:19 -07003441
Mel Gormana1e78772008-07-23 21:27:23 -07003442 /*
Mel Gorman17c9d122009-02-11 16:34:16 +00003443 * Only apply hugepage reservation if asked. At fault time, an
3444 * attempt will be made for VM_NORESERVE to allocate a page
David Gibson90481622012-03-21 16:34:12 -07003445 * without using reserves
Mel Gorman17c9d122009-02-11 16:34:16 +00003446 */
KOSAKI Motohiroca16d142011-05-26 19:16:19 +09003447 if (vm_flags & VM_NORESERVE)
Mel Gorman17c9d122009-02-11 16:34:16 +00003448 return 0;
3449
3450 /*
Mel Gormana1e78772008-07-23 21:27:23 -07003451 * Shared mappings base their reservation on the number of pages that
3452 * are already allocated on behalf of the file. Private mappings need
3453 * to reserve the full area even if read-only as mprotect() may be
3454 * called to make the mapping read-write. Assume !vma is a shm mapping
3455 */
Joonsoo Kim9119a412014-04-03 14:47:25 -07003456 if (!vma || vma->vm_flags & VM_MAYSHARE) {
Joonsoo Kim4e35f482014-04-03 14:47:30 -07003457 resv_map = inode_resv_map(inode);
Joonsoo Kim9119a412014-04-03 14:47:25 -07003458
Joonsoo Kim1406ec92014-04-03 14:47:26 -07003459 chg = region_chg(resv_map, from, to);
Joonsoo Kim9119a412014-04-03 14:47:25 -07003460
3461 } else {
3462 resv_map = resv_map_alloc();
Mel Gorman5a6fe122009-02-10 14:02:27 +00003463 if (!resv_map)
3464 return -ENOMEM;
3465
Mel Gorman17c9d122009-02-11 16:34:16 +00003466 chg = to - from;
3467
Mel Gorman5a6fe122009-02-10 14:02:27 +00003468 set_vma_resv_map(vma, resv_map);
3469 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
3470 }
3471
Dave Hansenc50ac052012-05-29 15:06:46 -07003472 if (chg < 0) {
3473 ret = chg;
3474 goto out_err;
3475 }
Mel Gorman17c9d122009-02-11 16:34:16 +00003476
David Gibson90481622012-03-21 16:34:12 -07003477 /* There must be enough pages in the subpool for the mapping */
Dave Hansenc50ac052012-05-29 15:06:46 -07003478 if (hugepage_subpool_get_pages(spool, chg)) {
3479 ret = -ENOSPC;
3480 goto out_err;
3481 }
Mel Gorman17c9d122009-02-11 16:34:16 +00003482
3483 /*
3484 * Check enough hugepages are available for the reservation.
David Gibson90481622012-03-21 16:34:12 -07003485 * Hand the pages back to the subpool if there are not
Mel Gorman17c9d122009-02-11 16:34:16 +00003486 */
3487 ret = hugetlb_acct_memory(h, chg);
3488 if (ret < 0) {
David Gibson90481622012-03-21 16:34:12 -07003489 hugepage_subpool_put_pages(spool, chg);
Dave Hansenc50ac052012-05-29 15:06:46 -07003490 goto out_err;
Mel Gorman17c9d122009-02-11 16:34:16 +00003491 }
3492
3493 /*
3494 * Account for the reservations made. Shared mappings record regions
3495 * that have reservations as they are shared by multiple VMAs.
3496 * When the last VMA disappears, the region map says how much
3497 * the reservation was and the page cache tells how much of
3498 * the reservation was consumed. Private mappings are per-VMA and
3499 * only the consumed reservations are tracked. When the VMA
3500 * disappears, the original reservation is the VMA size and the
3501 * consumed reservations are stored in the map. Hence, nothing
3502 * else has to be done for private mappings here
3503 */
Mel Gormanf83a2752009-05-28 14:34:40 -07003504 if (!vma || vma->vm_flags & VM_MAYSHARE)
Joonsoo Kim1406ec92014-04-03 14:47:26 -07003505 region_add(resv_map, from, to);
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -07003506 return 0;
Dave Hansenc50ac052012-05-29 15:06:46 -07003507out_err:
Joonsoo Kimf031dd22014-04-03 14:47:28 -07003508 if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
3509 kref_put(&resv_map->refs, resv_map_release);
Dave Hansenc50ac052012-05-29 15:06:46 -07003510 return ret;
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -07003511}
3512
3513void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed)
3514{
Andi Kleena5516432008-07-23 21:27:41 -07003515 struct hstate *h = hstate_inode(inode);
Joonsoo Kim4e35f482014-04-03 14:47:30 -07003516 struct resv_map *resv_map = inode_resv_map(inode);
Joonsoo Kim9119a412014-04-03 14:47:25 -07003517 long chg = 0;
David Gibson90481622012-03-21 16:34:12 -07003518 struct hugepage_subpool *spool = subpool_inode(inode);
Ken Chen45c682a2007-11-14 16:59:44 -08003519
Joonsoo Kim9119a412014-04-03 14:47:25 -07003520 if (resv_map)
Joonsoo Kim1406ec92014-04-03 14:47:26 -07003521 chg = region_truncate(resv_map, offset);
Ken Chen45c682a2007-11-14 16:59:44 -08003522 spin_lock(&inode->i_lock);
Eric Sandeene4c6f8b2009-07-29 15:02:16 -07003523 inode->i_blocks -= (blocks_per_huge_page(h) * freed);
Ken Chen45c682a2007-11-14 16:59:44 -08003524 spin_unlock(&inode->i_lock);
3525
David Gibson90481622012-03-21 16:34:12 -07003526 hugepage_subpool_put_pages(spool, (chg - freed));
Andi Kleena5516432008-07-23 21:27:41 -07003527 hugetlb_acct_memory(h, -(chg - freed));
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -07003528}
Naoya Horiguchi93f70f92010-05-28 09:29:20 +09003529
Steve Capper3212b532013-04-23 12:35:02 +01003530#ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
3531static unsigned long page_table_shareable(struct vm_area_struct *svma,
3532 struct vm_area_struct *vma,
3533 unsigned long addr, pgoff_t idx)
3534{
3535 unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) +
3536 svma->vm_start;
3537 unsigned long sbase = saddr & PUD_MASK;
3538 unsigned long s_end = sbase + PUD_SIZE;
3539
3540 /* Allow segments to share if only one is marked locked */
3541 unsigned long vm_flags = vma->vm_flags & ~VM_LOCKED;
3542 unsigned long svm_flags = svma->vm_flags & ~VM_LOCKED;
3543
3544 /*
3545 * match the virtual addresses, permission and the alignment of the
3546 * page table page.
3547 */
3548 if (pmd_index(addr) != pmd_index(saddr) ||
3549 vm_flags != svm_flags ||
3550 sbase < svma->vm_start || svma->vm_end < s_end)
3551 return 0;
3552
3553 return saddr;
3554}
3555
3556static int vma_shareable(struct vm_area_struct *vma, unsigned long addr)
3557{
3558 unsigned long base = addr & PUD_MASK;
3559 unsigned long end = base + PUD_SIZE;
3560
3561 /*
3562 * check on proper vm_flags and page table alignment
3563 */
3564 if (vma->vm_flags & VM_MAYSHARE &&
3565 vma->vm_start <= base && end <= vma->vm_end)
3566 return 1;
3567 return 0;
3568}
3569
3570/*
3571 * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
3572 * and returns the corresponding pte. While this is not necessary for the
3573 * !shared pmd case because we can allocate the pmd later as well, it makes the
3574 * code much cleaner. pmd allocation is essential for the shared case because
Davidlohr Buesoc8c06ef2014-12-12 16:54:24 -08003575 * pud has to be populated inside the same i_mmap_rwsem section - otherwise
Steve Capper3212b532013-04-23 12:35:02 +01003576 * racing tasks could either miss the sharing (see huge_pte_offset) or select a
3577 * bad pmd for sharing.
3578 */
3579pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
3580{
3581 struct vm_area_struct *vma = find_vma(mm, addr);
3582 struct address_space *mapping = vma->vm_file->f_mapping;
3583 pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) +
3584 vma->vm_pgoff;
3585 struct vm_area_struct *svma;
3586 unsigned long saddr;
3587 pte_t *spte = NULL;
3588 pte_t *pte;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003589 spinlock_t *ptl;
Steve Capper3212b532013-04-23 12:35:02 +01003590
3591 if (!vma_shareable(vma, addr))
3592 return (pte_t *)pmd_alloc(mm, pud, addr);
3593
Davidlohr Bueso83cde9e2014-12-12 16:54:21 -08003594 i_mmap_lock_write(mapping);
Steve Capper3212b532013-04-23 12:35:02 +01003595 vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) {
3596 if (svma == vma)
3597 continue;
3598
3599 saddr = page_table_shareable(svma, vma, addr, idx);
3600 if (saddr) {
3601 spte = huge_pte_offset(svma->vm_mm, saddr);
3602 if (spte) {
Kirill A. Shutemovdc6c9a32015-02-11 15:26:50 -08003603 mm_inc_nr_pmds(mm);
Steve Capper3212b532013-04-23 12:35:02 +01003604 get_page(virt_to_page(spte));
3605 break;
3606 }
3607 }
3608 }
3609
3610 if (!spte)
3611 goto out;
3612
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003613 ptl = huge_pte_lockptr(hstate_vma(vma), mm, spte);
3614 spin_lock(ptl);
Kirill A. Shutemovdc6c9a32015-02-11 15:26:50 -08003615 if (pud_none(*pud)) {
Steve Capper3212b532013-04-23 12:35:02 +01003616 pud_populate(mm, pud,
3617 (pmd_t *)((unsigned long)spte & PAGE_MASK));
Kirill A. Shutemovdc6c9a32015-02-11 15:26:50 -08003618 } else {
Steve Capper3212b532013-04-23 12:35:02 +01003619 put_page(virt_to_page(spte));
Kirill A. Shutemovdc6c9a32015-02-11 15:26:50 -08003620 mm_inc_nr_pmds(mm);
3621 }
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003622 spin_unlock(ptl);
Steve Capper3212b532013-04-23 12:35:02 +01003623out:
3624 pte = (pte_t *)pmd_alloc(mm, pud, addr);
Davidlohr Bueso83cde9e2014-12-12 16:54:21 -08003625 i_mmap_unlock_write(mapping);
Steve Capper3212b532013-04-23 12:35:02 +01003626 return pte;
3627}
3628
3629/*
3630 * unmap huge page backed by shared pte.
3631 *
3632 * Hugetlb pte page is ref counted at the time of mapping. If pte is shared
3633 * indicated by page_count > 1, unmap is achieved by clearing pud and
3634 * decrementing the ref count. If count == 1, the pte page is not shared.
3635 *
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003636 * called with page table lock held.
Steve Capper3212b532013-04-23 12:35:02 +01003637 *
3638 * returns: 1 successfully unmapped a shared pte page
3639 * 0 the underlying pte page is not shared, or it is the last user
3640 */
3641int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
3642{
3643 pgd_t *pgd = pgd_offset(mm, *addr);
3644 pud_t *pud = pud_offset(pgd, *addr);
3645
3646 BUG_ON(page_count(virt_to_page(ptep)) == 0);
3647 if (page_count(virt_to_page(ptep)) == 1)
3648 return 0;
3649
3650 pud_clear(pud);
3651 put_page(virt_to_page(ptep));
Kirill A. Shutemovdc6c9a32015-02-11 15:26:50 -08003652 mm_dec_nr_pmds(mm);
Steve Capper3212b532013-04-23 12:35:02 +01003653 *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE;
3654 return 1;
3655}
Steve Capper9e5fc742013-04-30 08:02:03 +01003656#define want_pmd_share() (1)
3657#else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
3658pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
3659{
3660 return NULL;
3661}
3662#define want_pmd_share() (0)
Steve Capper3212b532013-04-23 12:35:02 +01003663#endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
3664
Steve Capper9e5fc742013-04-30 08:02:03 +01003665#ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
3666pte_t *huge_pte_alloc(struct mm_struct *mm,
3667 unsigned long addr, unsigned long sz)
3668{
3669 pgd_t *pgd;
3670 pud_t *pud;
3671 pte_t *pte = NULL;
3672
3673 pgd = pgd_offset(mm, addr);
3674 pud = pud_alloc(mm, pgd, addr);
3675 if (pud) {
3676 if (sz == PUD_SIZE) {
3677 pte = (pte_t *)pud;
3678 } else {
3679 BUG_ON(sz != PMD_SIZE);
3680 if (want_pmd_share() && pud_none(*pud))
3681 pte = huge_pmd_share(mm, addr, pud);
3682 else
3683 pte = (pte_t *)pmd_alloc(mm, pud, addr);
3684 }
3685 }
3686 BUG_ON(pte && !pte_none(*pte) && !pte_huge(*pte));
3687
3688 return pte;
3689}
3690
3691pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr)
3692{
3693 pgd_t *pgd;
3694 pud_t *pud;
3695 pmd_t *pmd = NULL;
3696
3697 pgd = pgd_offset(mm, addr);
3698 if (pgd_present(*pgd)) {
3699 pud = pud_offset(pgd, addr);
3700 if (pud_present(*pud)) {
3701 if (pud_huge(*pud))
3702 return (pte_t *)pud;
3703 pmd = pmd_offset(pud, addr);
3704 }
3705 }
3706 return (pte_t *) pmd;
3707}
3708
Naoya Horiguchi61f77ed2015-02-11 15:25:15 -08003709#endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
3710
3711/*
3712 * These functions are overwritable if your architecture needs its own
3713 * behavior.
3714 */
3715struct page * __weak
3716follow_huge_addr(struct mm_struct *mm, unsigned long address,
3717 int write)
3718{
3719 return ERR_PTR(-EINVAL);
3720}
3721
3722struct page * __weak
Steve Capper9e5fc742013-04-30 08:02:03 +01003723follow_huge_pmd(struct mm_struct *mm, unsigned long address,
Naoya Horiguchie66f17f2015-02-11 15:25:22 -08003724 pmd_t *pmd, int flags)
Steve Capper9e5fc742013-04-30 08:02:03 +01003725{
Naoya Horiguchie66f17f2015-02-11 15:25:22 -08003726 struct page *page = NULL;
3727 spinlock_t *ptl;
3728retry:
3729 ptl = pmd_lockptr(mm, pmd);
3730 spin_lock(ptl);
3731 /*
3732 * make sure that the address range covered by this pmd is not
3733 * unmapped from other threads.
3734 */
3735 if (!pmd_huge(*pmd))
3736 goto out;
3737 if (pmd_present(*pmd)) {
3738 page = pte_page(*(pte_t *)pmd) +
3739 ((address & ~PMD_MASK) >> PAGE_SHIFT);
3740 if (flags & FOLL_GET)
3741 get_page(page);
3742 } else {
3743 if (is_hugetlb_entry_migration(huge_ptep_get((pte_t *)pmd))) {
3744 spin_unlock(ptl);
3745 __migration_entry_wait(mm, (pte_t *)pmd, ptl);
3746 goto retry;
3747 }
3748 /*
3749 * hwpoisoned entry is treated as no_page_table in
3750 * follow_page_mask().
3751 */
3752 }
3753out:
3754 spin_unlock(ptl);
Steve Capper9e5fc742013-04-30 08:02:03 +01003755 return page;
3756}
3757
Naoya Horiguchi61f77ed2015-02-11 15:25:15 -08003758struct page * __weak
Steve Capper9e5fc742013-04-30 08:02:03 +01003759follow_huge_pud(struct mm_struct *mm, unsigned long address,
Naoya Horiguchie66f17f2015-02-11 15:25:22 -08003760 pud_t *pud, int flags)
Steve Capper9e5fc742013-04-30 08:02:03 +01003761{
Naoya Horiguchie66f17f2015-02-11 15:25:22 -08003762 if (flags & FOLL_GET)
3763 return NULL;
Steve Capper9e5fc742013-04-30 08:02:03 +01003764
Naoya Horiguchie66f17f2015-02-11 15:25:22 -08003765 return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT);
Steve Capper9e5fc742013-04-30 08:02:03 +01003766}
3767
Andi Kleend5bd9102010-09-27 09:00:12 +02003768#ifdef CONFIG_MEMORY_FAILURE
3769
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +09003770/* Should be called in hugetlb_lock */
3771static int is_hugepage_on_freelist(struct page *hpage)
3772{
3773 struct page *page;
3774 struct page *tmp;
3775 struct hstate *h = page_hstate(hpage);
3776 int nid = page_to_nid(hpage);
3777
3778 list_for_each_entry_safe(page, tmp, &h->hugepage_freelists[nid], lru)
3779 if (page == hpage)
3780 return 1;
3781 return 0;
3782}
3783
Naoya Horiguchi93f70f92010-05-28 09:29:20 +09003784/*
3785 * This function is called from memory failure code.
3786 * Assume the caller holds page lock of the head page.
3787 */
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +09003788int dequeue_hwpoisoned_huge_page(struct page *hpage)
Naoya Horiguchi93f70f92010-05-28 09:29:20 +09003789{
3790 struct hstate *h = page_hstate(hpage);
3791 int nid = page_to_nid(hpage);
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +09003792 int ret = -EBUSY;
Naoya Horiguchi93f70f92010-05-28 09:29:20 +09003793
3794 spin_lock(&hugetlb_lock);
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +09003795 if (is_hugepage_on_freelist(hpage)) {
Naoya Horiguchi56f2fb12012-12-12 13:52:33 -08003796 /*
3797 * Hwpoisoned hugepage isn't linked to activelist or freelist,
3798 * but dangling hpage->lru can trigger list-debug warnings
3799 * (this happens when we call unpoison_memory() on it),
3800 * so let it point to itself with list_del_init().
3801 */
3802 list_del_init(&hpage->lru);
Naoya Horiguchi8c6c2ec2010-09-08 10:19:38 +09003803 set_page_refcounted(hpage);
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +09003804 h->free_huge_pages--;
3805 h->free_huge_pages_node[nid]--;
3806 ret = 0;
3807 }
Naoya Horiguchi93f70f92010-05-28 09:29:20 +09003808 spin_unlock(&hugetlb_lock);
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +09003809 return ret;
Naoya Horiguchi93f70f92010-05-28 09:29:20 +09003810}
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +09003811#endif
Naoya Horiguchi31caf662013-09-11 14:21:59 -07003812
3813bool isolate_huge_page(struct page *page, struct list_head *list)
3814{
Sasha Levin309381fea2014-01-23 15:52:54 -08003815 VM_BUG_ON_PAGE(!PageHead(page), page);
Naoya Horiguchi31caf662013-09-11 14:21:59 -07003816 if (!get_page_unless_zero(page))
3817 return false;
3818 spin_lock(&hugetlb_lock);
3819 list_move_tail(&page->lru, list);
3820 spin_unlock(&hugetlb_lock);
3821 return true;
3822}
3823
3824void putback_active_hugepage(struct page *page)
3825{
Sasha Levin309381fea2014-01-23 15:52:54 -08003826 VM_BUG_ON_PAGE(!PageHead(page), page);
Naoya Horiguchi31caf662013-09-11 14:21:59 -07003827 spin_lock(&hugetlb_lock);
3828 list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist);
3829 spin_unlock(&hugetlb_lock);
3830 put_page(page);
3831}
Naoya Horiguchic8721bb2013-09-11 14:22:09 -07003832
3833bool is_hugepage_active(struct page *page)
3834{
Sasha Levin309381fea2014-01-23 15:52:54 -08003835 VM_BUG_ON_PAGE(!PageHuge(page), page);
Naoya Horiguchic8721bb2013-09-11 14:22:09 -07003836 /*
3837 * This function can be called for a tail page because the caller,
3838 * scan_movable_pages, scans through a given pfn-range which typically
3839 * covers one memory block. In systems using gigantic hugepage (1GB
3840 * for x86_64,) a hugepage is larger than a memory block, and we don't
3841 * support migrating such large hugepages for now, so return false
3842 * when called for tail pages.
3843 */
3844 if (PageTail(page))
3845 return false;
3846 /*
3847 * Refcount of a hwpoisoned hugepages is 1, but they are not active,
3848 * so we should return false for them.
3849 */
3850 if (unlikely(PageHWPoison(page)))
3851 return false;
3852 return page_count(page) > 0;
3853}