blob: 83aff0a4d0938e308619a703dfca17265419e628 [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>
Christoph Lameteraea47ff2006-01-08 01:00:57 -080016#include <linux/cpuset.h>
David Gibson3935baa2006-03-22 00:08:53 -080017#include <linux/mutex.h>
Andi Kleenaa888a72008-07-23 21:27:47 -070018#include <linux/bootmem.h>
Nishanth Aravamudana3437872008-07-23 21:27:44 -070019#include <linux/sysfs.h>
Tejun Heo5a0e3ad2010-03-24 17:04:11 +090020#include <linux/slab.h>
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +090021#include <linux/rmap.h>
Naoya Horiguchifd6a03ed2010-05-28 09:29:21 +090022#include <linux/swap.h>
23#include <linux/swapops.h>
Linus Torvaldsd6606682008-08-06 12:04:54 -070024
David Gibson63551ae2005-06-21 17:14:44 -070025#include <asm/page.h>
26#include <asm/pgtable.h>
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -070027#include <asm/tlb.h>
David Gibson63551ae2005-06-21 17:14:44 -070028
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -070029#include <linux/io.h>
David Gibson63551ae2005-06-21 17:14:44 -070030#include <linux/hugetlb.h>
Aneesh Kumar K.V9dd540e2012-07-31 16:42:15 -070031#include <linux/hugetlb_cgroup.h>
Lee Schermerhorn9a305232009-12-14 17:58:25 -080032#include <linux/node.h>
Nick Piggin7835e982006-03-22 00:08:40 -080033#include "internal.h"
Linus Torvalds1da177e2005-04-16 15:20:36 -070034
35const unsigned long hugetlb_zero = 0, hugetlb_infinity = ~0UL;
Mel Gorman396faf02007-07-17 04:03:13 -070036static gfp_t htlb_alloc_mask = GFP_HIGHUSER;
37unsigned long hugepages_treat_as_movable;
Andi Kleena5516432008-07-23 21:27:41 -070038
Aneesh Kumar K.Vc3f38a32012-07-31 16:42:10 -070039int hugetlb_max_hstate __read_mostly;
Andi Kleene5ff2152008-07-23 21:27:42 -070040unsigned int default_hstate_idx;
41struct hstate hstates[HUGE_MAX_HSTATE];
42
Jon Tollefson53ba51d2008-07-23 21:27:52 -070043__initdata LIST_HEAD(huge_boot_pages);
44
Andi Kleene5ff2152008-07-23 21:27:42 -070045/* for command line parsing */
46static struct hstate * __initdata parsed_hstate;
47static unsigned long __initdata default_hstate_max_huge_pages;
Nick Piggine11bfbf2008-07-23 21:27:52 -070048static unsigned long __initdata default_hstate_size;
Andi Kleene5ff2152008-07-23 21:27:42 -070049
David Gibson3935baa2006-03-22 00:08:53 -080050/*
51 * Protects updates to hugepage_freelists, nr_huge_pages, and free_huge_pages
52 */
Aneesh Kumar K.Vc3f38a32012-07-31 16:42:10 -070053DEFINE_SPINLOCK(hugetlb_lock);
Eric Paris0bd0f9f2005-11-21 21:32:28 -080054
David Gibson90481622012-03-21 16:34:12 -070055static inline void unlock_or_release_subpool(struct hugepage_subpool *spool)
56{
57 bool free = (spool->count == 0) && (spool->used_hpages == 0);
58
59 spin_unlock(&spool->lock);
60
61 /* If no pages are used, and no other handles to the subpool
62 * remain, free the subpool the subpool remain */
63 if (free)
64 kfree(spool);
65}
66
67struct hugepage_subpool *hugepage_new_subpool(long nr_blocks)
68{
69 struct hugepage_subpool *spool;
70
71 spool = kmalloc(sizeof(*spool), GFP_KERNEL);
72 if (!spool)
73 return NULL;
74
75 spin_lock_init(&spool->lock);
76 spool->count = 1;
77 spool->max_hpages = nr_blocks;
78 spool->used_hpages = 0;
79
80 return spool;
81}
82
83void hugepage_put_subpool(struct hugepage_subpool *spool)
84{
85 spin_lock(&spool->lock);
86 BUG_ON(!spool->count);
87 spool->count--;
88 unlock_or_release_subpool(spool);
89}
90
91static int hugepage_subpool_get_pages(struct hugepage_subpool *spool,
92 long delta)
93{
94 int ret = 0;
95
96 if (!spool)
97 return 0;
98
99 spin_lock(&spool->lock);
100 if ((spool->used_hpages + delta) <= spool->max_hpages) {
101 spool->used_hpages += delta;
102 } else {
103 ret = -ENOMEM;
104 }
105 spin_unlock(&spool->lock);
106
107 return ret;
108}
109
110static void hugepage_subpool_put_pages(struct hugepage_subpool *spool,
111 long delta)
112{
113 if (!spool)
114 return;
115
116 spin_lock(&spool->lock);
117 spool->used_hpages -= delta;
118 /* If hugetlbfs_put_super couldn't free spool due to
119 * an outstanding quota reference, free it now. */
120 unlock_or_release_subpool(spool);
121}
122
123static inline struct hugepage_subpool *subpool_inode(struct inode *inode)
124{
125 return HUGETLBFS_SB(inode->i_sb)->spool;
126}
127
128static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma)
129{
Al Viro496ad9a2013-01-23 17:07:38 -0500130 return subpool_inode(file_inode(vma->vm_file));
David Gibson90481622012-03-21 16:34:12 -0700131}
132
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700133/*
Andy Whitcroft96822902008-07-23 21:27:29 -0700134 * Region tracking -- allows tracking of reservations and instantiated pages
135 * across the pages in a mapping.
Andy Whitcroft84afd992008-07-23 21:27:32 -0700136 *
137 * The region data structures are protected by a combination of the mmap_sem
138 * and the hugetlb_instantion_mutex. To access or modify a region the caller
139 * must either hold the mmap_sem for write, or the mmap_sem for read and
140 * the hugetlb_instantiation mutex:
141 *
Chris Forbes32f84522011-07-25 17:12:14 -0700142 * down_write(&mm->mmap_sem);
Andy Whitcroft84afd992008-07-23 21:27:32 -0700143 * or
Chris Forbes32f84522011-07-25 17:12:14 -0700144 * down_read(&mm->mmap_sem);
145 * mutex_lock(&hugetlb_instantiation_mutex);
Andy Whitcroft96822902008-07-23 21:27:29 -0700146 */
147struct file_region {
148 struct list_head link;
149 long from;
150 long to;
151};
152
153static long region_add(struct list_head *head, long f, long t)
154{
155 struct file_region *rg, *nrg, *trg;
156
157 /* Locate the region we are either in or before. */
158 list_for_each_entry(rg, head, link)
159 if (f <= rg->to)
160 break;
161
162 /* Round our left edge to the current segment if it encloses us. */
163 if (f > rg->from)
164 f = rg->from;
165
166 /* Check for and consume any regions we now overlap with. */
167 nrg = rg;
168 list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
169 if (&rg->link == head)
170 break;
171 if (rg->from > t)
172 break;
173
174 /* If this area reaches higher then extend our area to
175 * include it completely. If this is not the first area
176 * which we intend to reuse, free it. */
177 if (rg->to > t)
178 t = rg->to;
179 if (rg != nrg) {
180 list_del(&rg->link);
181 kfree(rg);
182 }
183 }
184 nrg->from = f;
185 nrg->to = t;
186 return 0;
187}
188
189static long region_chg(struct list_head *head, long f, long t)
190{
191 struct file_region *rg, *nrg;
192 long chg = 0;
193
194 /* Locate the region we are before or in. */
195 list_for_each_entry(rg, head, link)
196 if (f <= rg->to)
197 break;
198
199 /* If we are below the current region then a new region is required.
200 * Subtle, allocate a new region at the position but make it zero
201 * size such that we can guarantee to record the reservation. */
202 if (&rg->link == head || t < rg->from) {
203 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
204 if (!nrg)
205 return -ENOMEM;
206 nrg->from = f;
207 nrg->to = f;
208 INIT_LIST_HEAD(&nrg->link);
209 list_add(&nrg->link, rg->link.prev);
210
211 return t - f;
212 }
213
214 /* Round our left edge to the current segment if it encloses us. */
215 if (f > rg->from)
216 f = rg->from;
217 chg = t - f;
218
219 /* Check for and consume any regions we now overlap with. */
220 list_for_each_entry(rg, rg->link.prev, link) {
221 if (&rg->link == head)
222 break;
223 if (rg->from > t)
224 return chg;
225
Lucas De Marchi25985ed2011-03-30 22:57:33 -0300226 /* We overlap with this area, if it extends further than
Andy Whitcroft96822902008-07-23 21:27:29 -0700227 * us then we must extend ourselves. Account for its
228 * existing reservation. */
229 if (rg->to > t) {
230 chg += rg->to - t;
231 t = rg->to;
232 }
233 chg -= rg->to - rg->from;
234 }
235 return chg;
236}
237
238static long region_truncate(struct list_head *head, long end)
239{
240 struct file_region *rg, *trg;
241 long chg = 0;
242
243 /* Locate the region we are either in or before. */
244 list_for_each_entry(rg, head, link)
245 if (end <= rg->to)
246 break;
247 if (&rg->link == head)
248 return 0;
249
250 /* If we are in the middle of a region then adjust it. */
251 if (end > rg->from) {
252 chg = rg->to - end;
253 rg->to = end;
254 rg = list_entry(rg->link.next, typeof(*rg), link);
255 }
256
257 /* Drop any remaining regions. */
258 list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
259 if (&rg->link == head)
260 break;
261 chg += rg->to - rg->from;
262 list_del(&rg->link);
263 kfree(rg);
264 }
265 return chg;
266}
267
Andy Whitcroft84afd992008-07-23 21:27:32 -0700268static long region_count(struct list_head *head, long f, long t)
269{
270 struct file_region *rg;
271 long chg = 0;
272
273 /* Locate each segment we overlap with, and count that overlap. */
274 list_for_each_entry(rg, head, link) {
Wang Sheng-Huif2135a42012-05-29 15:06:17 -0700275 long seg_from;
276 long seg_to;
Andy Whitcroft84afd992008-07-23 21:27:32 -0700277
278 if (rg->to <= f)
279 continue;
280 if (rg->from >= t)
281 break;
282
283 seg_from = max(rg->from, f);
284 seg_to = min(rg->to, t);
285
286 chg += seg_to - seg_from;
287 }
288
289 return chg;
290}
291
Andy Whitcroft96822902008-07-23 21:27:29 -0700292/*
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700293 * Convert the address within this vma to the page offset within
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700294 * the mapping, in pagecache page units; huge pages here.
295 */
Andi Kleena5516432008-07-23 21:27:41 -0700296static pgoff_t vma_hugecache_offset(struct hstate *h,
297 struct vm_area_struct *vma, unsigned long address)
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700298{
Andi Kleena5516432008-07-23 21:27:41 -0700299 return ((address - vma->vm_start) >> huge_page_shift(h)) +
300 (vma->vm_pgoff >> huge_page_order(h));
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700301}
302
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +0900303pgoff_t linear_hugepage_index(struct vm_area_struct *vma,
304 unsigned long address)
305{
306 return vma_hugecache_offset(hstate_vma(vma), vma, address);
307}
308
Andy Whitcroft84afd992008-07-23 21:27:32 -0700309/*
Mel Gorman08fba692009-01-06 14:38:53 -0800310 * Return the size of the pages allocated when backing a VMA. In the majority
311 * cases this will be same size as used by the page table entries.
312 */
313unsigned long vma_kernel_pagesize(struct vm_area_struct *vma)
314{
315 struct hstate *hstate;
316
317 if (!is_vm_hugetlb_page(vma))
318 return PAGE_SIZE;
319
320 hstate = hstate_vma(vma);
321
Wanpeng Li2415cf12013-07-03 15:02:43 -0700322 return 1UL << huge_page_shift(hstate);
Mel Gorman08fba692009-01-06 14:38:53 -0800323}
Joerg Roedelf340ca02009-06-19 15:16:22 +0200324EXPORT_SYMBOL_GPL(vma_kernel_pagesize);
Mel Gorman08fba692009-01-06 14:38:53 -0800325
326/*
Mel Gorman33402892009-01-06 14:38:54 -0800327 * Return the page size being used by the MMU to back a VMA. In the majority
328 * of cases, the page size used by the kernel matches the MMU size. On
329 * architectures where it differs, an architecture-specific version of this
330 * function is required.
331 */
332#ifndef vma_mmu_pagesize
333unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
334{
335 return vma_kernel_pagesize(vma);
336}
337#endif
338
339/*
Andy Whitcroft84afd992008-07-23 21:27:32 -0700340 * Flags for MAP_PRIVATE reservations. These are stored in the bottom
341 * bits of the reservation map pointer, which are always clear due to
342 * alignment.
343 */
344#define HPAGE_RESV_OWNER (1UL << 0)
345#define HPAGE_RESV_UNMAPPED (1UL << 1)
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700346#define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
Andy Whitcroft84afd992008-07-23 21:27:32 -0700347
Mel Gormana1e78772008-07-23 21:27:23 -0700348/*
349 * These helpers are used to track how many pages are reserved for
350 * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
351 * is guaranteed to have their future faults succeed.
352 *
353 * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
354 * the reserve counters are updated with the hugetlb_lock held. It is safe
355 * to reset the VMA at fork() time as it is not in use yet and there is no
356 * chance of the global counters getting corrupted as a result of the values.
Andy Whitcroft84afd992008-07-23 21:27:32 -0700357 *
358 * The private mapping reservation is represented in a subtly different
359 * manner to a shared mapping. A shared mapping has a region map associated
360 * with the underlying file, this region map represents the backing file
361 * pages which have ever had a reservation assigned which this persists even
362 * after the page is instantiated. A private mapping has a region map
363 * associated with the original mmap which is attached to all VMAs which
364 * reference it, this region map represents those offsets which have consumed
365 * reservation ie. where pages have been instantiated.
Mel Gormana1e78772008-07-23 21:27:23 -0700366 */
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700367static unsigned long get_vma_private_data(struct vm_area_struct *vma)
368{
369 return (unsigned long)vma->vm_private_data;
370}
371
372static void set_vma_private_data(struct vm_area_struct *vma,
373 unsigned long value)
374{
375 vma->vm_private_data = (void *)value;
376}
377
Andy Whitcroft84afd992008-07-23 21:27:32 -0700378struct resv_map {
379 struct kref refs;
380 struct list_head regions;
381};
382
Harvey Harrison2a4b3de2008-10-18 20:27:06 -0700383static struct resv_map *resv_map_alloc(void)
Andy Whitcroft84afd992008-07-23 21:27:32 -0700384{
385 struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
386 if (!resv_map)
387 return NULL;
388
389 kref_init(&resv_map->refs);
390 INIT_LIST_HEAD(&resv_map->regions);
391
392 return resv_map;
393}
394
Harvey Harrison2a4b3de2008-10-18 20:27:06 -0700395static void resv_map_release(struct kref *ref)
Andy Whitcroft84afd992008-07-23 21:27:32 -0700396{
397 struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
398
399 /* Clear out any active regions before we release the map. */
400 region_truncate(&resv_map->regions, 0);
401 kfree(resv_map);
402}
403
404static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
Mel Gormana1e78772008-07-23 21:27:23 -0700405{
406 VM_BUG_ON(!is_vm_hugetlb_page(vma));
Mel Gormanf83a2752009-05-28 14:34:40 -0700407 if (!(vma->vm_flags & VM_MAYSHARE))
Andy Whitcroft84afd992008-07-23 21:27:32 -0700408 return (struct resv_map *)(get_vma_private_data(vma) &
409 ~HPAGE_RESV_MASK);
Harvey Harrison2a4b3de2008-10-18 20:27:06 -0700410 return NULL;
Mel Gormana1e78772008-07-23 21:27:23 -0700411}
412
Andy Whitcroft84afd992008-07-23 21:27:32 -0700413static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
Mel Gormana1e78772008-07-23 21:27:23 -0700414{
415 VM_BUG_ON(!is_vm_hugetlb_page(vma));
Mel Gormanf83a2752009-05-28 14:34:40 -0700416 VM_BUG_ON(vma->vm_flags & VM_MAYSHARE);
Mel Gormana1e78772008-07-23 21:27:23 -0700417
Andy Whitcroft84afd992008-07-23 21:27:32 -0700418 set_vma_private_data(vma, (get_vma_private_data(vma) &
419 HPAGE_RESV_MASK) | (unsigned long)map);
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700420}
421
422static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
423{
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700424 VM_BUG_ON(!is_vm_hugetlb_page(vma));
Mel Gormanf83a2752009-05-28 14:34:40 -0700425 VM_BUG_ON(vma->vm_flags & VM_MAYSHARE);
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700426
427 set_vma_private_data(vma, get_vma_private_data(vma) | flags);
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700428}
429
430static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
431{
432 VM_BUG_ON(!is_vm_hugetlb_page(vma));
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700433
434 return (get_vma_private_data(vma) & flag) != 0;
Mel Gormana1e78772008-07-23 21:27:23 -0700435}
436
437/* Decrement the reserved pages in the hugepage pool by one */
Andi Kleena5516432008-07-23 21:27:41 -0700438static void decrement_hugepage_resv_vma(struct hstate *h,
439 struct vm_area_struct *vma)
Mel Gormana1e78772008-07-23 21:27:23 -0700440{
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -0700441 if (vma->vm_flags & VM_NORESERVE)
442 return;
443
Mel Gormanf83a2752009-05-28 14:34:40 -0700444 if (vma->vm_flags & VM_MAYSHARE) {
Mel Gormana1e78772008-07-23 21:27:23 -0700445 /* Shared mappings always use reserves */
Andi Kleena5516432008-07-23 21:27:41 -0700446 h->resv_huge_pages--;
Andy Whitcroft84afd992008-07-23 21:27:32 -0700447 } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
Mel Gormana1e78772008-07-23 21:27:23 -0700448 /*
449 * Only the process that called mmap() has reserves for
450 * private mappings.
451 */
Andi Kleena5516432008-07-23 21:27:41 -0700452 h->resv_huge_pages--;
Mel Gormana1e78772008-07-23 21:27:23 -0700453 }
454}
455
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700456/* Reset counters to 0 and clear all HPAGE_RESV_* flags */
Mel Gormana1e78772008-07-23 21:27:23 -0700457void reset_vma_resv_huge_pages(struct vm_area_struct *vma)
458{
459 VM_BUG_ON(!is_vm_hugetlb_page(vma));
Mel Gormanf83a2752009-05-28 14:34:40 -0700460 if (!(vma->vm_flags & VM_MAYSHARE))
Mel Gormana1e78772008-07-23 21:27:23 -0700461 vma->vm_private_data = (void *)0;
462}
463
464/* Returns true if the VMA has associated reserve pages */
Mel Gorman7f09ca52008-07-23 21:27:58 -0700465static int vma_has_reserves(struct vm_area_struct *vma)
Mel Gormana1e78772008-07-23 21:27:23 -0700466{
Mel Gormanf83a2752009-05-28 14:34:40 -0700467 if (vma->vm_flags & VM_MAYSHARE)
Mel Gorman7f09ca52008-07-23 21:27:58 -0700468 return 1;
469 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER))
470 return 1;
471 return 0;
Mel Gormana1e78772008-07-23 21:27:23 -0700472}
473
Naoya Horiguchi0ebabb42010-09-08 10:19:34 +0900474static void copy_gigantic_page(struct page *dst, struct page *src)
475{
476 int i;
477 struct hstate *h = page_hstate(src);
478 struct page *dst_base = dst;
479 struct page *src_base = src;
480
481 for (i = 0; i < pages_per_huge_page(h); ) {
482 cond_resched();
483 copy_highpage(dst, src);
484
485 i++;
486 dst = mem_map_next(dst, dst_base, i);
487 src = mem_map_next(src, src_base, i);
488 }
489}
490
491void copy_huge_page(struct page *dst, struct page *src)
492{
493 int i;
494 struct hstate *h = page_hstate(src);
495
496 if (unlikely(pages_per_huge_page(h) > MAX_ORDER_NR_PAGES)) {
497 copy_gigantic_page(dst, src);
498 return;
499 }
500
501 might_sleep();
502 for (i = 0; i < pages_per_huge_page(h); i++) {
503 cond_resched();
504 copy_highpage(dst + i, src + i);
505 }
506}
507
Andi Kleena5516432008-07-23 21:27:41 -0700508static void enqueue_huge_page(struct hstate *h, struct page *page)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700509{
510 int nid = page_to_nid(page);
Aneesh Kumar K.V0edaecf2012-07-31 16:42:07 -0700511 list_move(&page->lru, &h->hugepage_freelists[nid]);
Andi Kleena5516432008-07-23 21:27:41 -0700512 h->free_huge_pages++;
513 h->free_huge_pages_node[nid]++;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700514}
515
Naoya Horiguchibf50bab2010-09-08 10:19:33 +0900516static struct page *dequeue_huge_page_node(struct hstate *h, int nid)
517{
518 struct page *page;
519
520 if (list_empty(&h->hugepage_freelists[nid]))
521 return NULL;
522 page = list_entry(h->hugepage_freelists[nid].next, struct page, lru);
Aneesh Kumar K.V0edaecf2012-07-31 16:42:07 -0700523 list_move(&page->lru, &h->hugepage_activelist);
Naoya Horiguchia9869b82010-09-08 10:19:37 +0900524 set_page_refcounted(page);
Naoya Horiguchibf50bab2010-09-08 10:19:33 +0900525 h->free_huge_pages--;
526 h->free_huge_pages_node[nid]--;
527 return page;
528}
529
Andi Kleena5516432008-07-23 21:27:41 -0700530static struct page *dequeue_huge_page_vma(struct hstate *h,
531 struct vm_area_struct *vma,
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700532 unsigned long address, int avoid_reserve)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700533{
Konstantin Khlebnikovb1c12cb2012-04-25 16:01:46 -0700534 struct page *page = NULL;
Lee Schermerhorn480eccf2007-09-18 22:46:47 -0700535 struct mempolicy *mpol;
Mel Gorman19770b32008-04-28 02:12:18 -0700536 nodemask_t *nodemask;
Miao Xiec0ff7452010-05-24 14:32:08 -0700537 struct zonelist *zonelist;
Mel Gormandd1a2392008-04-28 02:12:17 -0700538 struct zone *zone;
539 struct zoneref *z;
Mel Gormancc9a6c82012-03-21 16:34:11 -0700540 unsigned int cpuset_mems_cookie;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700541
Mel Gormancc9a6c82012-03-21 16:34:11 -0700542retry_cpuset:
543 cpuset_mems_cookie = get_mems_allowed();
Miao Xiec0ff7452010-05-24 14:32:08 -0700544 zonelist = huge_zonelist(vma, address,
545 htlb_alloc_mask, &mpol, &nodemask);
Mel Gormana1e78772008-07-23 21:27:23 -0700546 /*
547 * A child process with MAP_PRIVATE mappings created by their parent
548 * have no page reserves. This check ensures that reservations are
549 * not "stolen". The child may still get SIGKILLed
550 */
Mel Gorman7f09ca52008-07-23 21:27:58 -0700551 if (!vma_has_reserves(vma) &&
Andi Kleena5516432008-07-23 21:27:41 -0700552 h->free_huge_pages - h->resv_huge_pages == 0)
Miao Xiec0ff7452010-05-24 14:32:08 -0700553 goto err;
Mel Gormana1e78772008-07-23 21:27:23 -0700554
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700555 /* If reserves cannot be used, ensure enough pages are in the pool */
Andi Kleena5516432008-07-23 21:27:41 -0700556 if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
Justin P. Mattock6eab04a2011-04-08 19:49:08 -0700557 goto err;
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700558
Mel Gorman19770b32008-04-28 02:12:18 -0700559 for_each_zone_zonelist_nodemask(zone, z, zonelist,
560 MAX_NR_ZONES - 1, nodemask) {
Naoya Horiguchibf50bab2010-09-08 10:19:33 +0900561 if (cpuset_zone_allowed_softwall(zone, htlb_alloc_mask)) {
562 page = dequeue_huge_page_node(h, zone_to_nid(zone));
563 if (page) {
564 if (!avoid_reserve)
565 decrement_hugepage_resv_vma(h, vma);
566 break;
567 }
Andrew Morton3abf7af2007-07-19 01:49:08 -0700568 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700569 }
Mel Gormancc9a6c82012-03-21 16:34:11 -0700570
571 mpol_cond_put(mpol);
572 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page))
573 goto retry_cpuset;
574 return page;
575
Miao Xiec0ff7452010-05-24 14:32:08 -0700576err:
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -0700577 mpol_cond_put(mpol);
Mel Gormancc9a6c82012-03-21 16:34:11 -0700578 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700579}
580
Andi Kleena5516432008-07-23 21:27:41 -0700581static void update_and_free_page(struct hstate *h, struct page *page)
Adam Litke6af2acb2007-10-16 01:26:16 -0700582{
583 int i;
Andi Kleena5516432008-07-23 21:27:41 -0700584
Andy Whitcroft18229df2008-11-06 12:53:27 -0800585 VM_BUG_ON(h->order >= MAX_ORDER);
586
Andi Kleena5516432008-07-23 21:27:41 -0700587 h->nr_huge_pages--;
588 h->nr_huge_pages_node[page_to_nid(page)]--;
589 for (i = 0; i < pages_per_huge_page(h); i++) {
Chris Forbes32f84522011-07-25 17:12:14 -0700590 page[i].flags &= ~(1 << PG_locked | 1 << PG_error |
591 1 << PG_referenced | 1 << PG_dirty |
592 1 << PG_active | 1 << PG_reserved |
593 1 << PG_private | 1 << PG_writeback);
Adam Litke6af2acb2007-10-16 01:26:16 -0700594 }
Aneesh Kumar K.V9dd540e2012-07-31 16:42:15 -0700595 VM_BUG_ON(hugetlb_cgroup_from_page(page));
Adam Litke6af2acb2007-10-16 01:26:16 -0700596 set_compound_page_dtor(page, NULL);
597 set_page_refcounted(page);
Gerald Schaefer7f2e9522008-04-28 02:13:29 -0700598 arch_release_hugepage(page);
Andi Kleena5516432008-07-23 21:27:41 -0700599 __free_pages(page, huge_page_order(h));
Adam Litke6af2acb2007-10-16 01:26:16 -0700600}
601
Andi Kleene5ff2152008-07-23 21:27:42 -0700602struct hstate *size_to_hstate(unsigned long size)
603{
604 struct hstate *h;
605
606 for_each_hstate(h) {
607 if (huge_page_size(h) == size)
608 return h;
609 }
610 return NULL;
611}
612
David Gibson27a85ef2006-03-22 00:08:56 -0800613static void free_huge_page(struct page *page)
614{
Andi Kleena5516432008-07-23 21:27:41 -0700615 /*
616 * Can't pass hstate in here because it is called from the
617 * compound page destructor.
618 */
Andi Kleene5ff2152008-07-23 21:27:42 -0700619 struct hstate *h = page_hstate(page);
Adam Litke7893d1d2007-10-16 01:26:18 -0700620 int nid = page_to_nid(page);
David Gibson90481622012-03-21 16:34:12 -0700621 struct hugepage_subpool *spool =
622 (struct hugepage_subpool *)page_private(page);
David Gibson27a85ef2006-03-22 00:08:56 -0800623
Andy Whitcrofte5df70a2008-02-23 15:23:32 -0800624 set_page_private(page, 0);
Mel Gorman23be7462010-04-23 13:17:56 -0400625 page->mapping = NULL;
Adam Litke7893d1d2007-10-16 01:26:18 -0700626 BUG_ON(page_count(page));
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +0900627 BUG_ON(page_mapcount(page));
David Gibson27a85ef2006-03-22 00:08:56 -0800628
629 spin_lock(&hugetlb_lock);
Aneesh Kumar K.V6d76dcf2012-07-31 16:42:18 -0700630 hugetlb_cgroup_uncharge_page(hstate_index(h),
631 pages_per_huge_page(h), page);
Andi Kleenaa888a72008-07-23 21:27:47 -0700632 if (h->surplus_huge_pages_node[nid] && huge_page_order(h) < MAX_ORDER) {
Aneesh Kumar K.V0edaecf2012-07-31 16:42:07 -0700633 /* remove the page from active list */
634 list_del(&page->lru);
Andi Kleena5516432008-07-23 21:27:41 -0700635 update_and_free_page(h, page);
636 h->surplus_huge_pages--;
637 h->surplus_huge_pages_node[nid]--;
Adam Litke7893d1d2007-10-16 01:26:18 -0700638 } else {
Will Deacon5d3a5512012-10-08 16:29:32 -0700639 arch_clear_hugepage_flags(page);
Andi Kleena5516432008-07-23 21:27:41 -0700640 enqueue_huge_page(h, page);
Adam Litke7893d1d2007-10-16 01:26:18 -0700641 }
David Gibson27a85ef2006-03-22 00:08:56 -0800642 spin_unlock(&hugetlb_lock);
David Gibson90481622012-03-21 16:34:12 -0700643 hugepage_subpool_put_pages(spool, 1);
David Gibson27a85ef2006-03-22 00:08:56 -0800644}
645
Andi Kleena5516432008-07-23 21:27:41 -0700646static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
Andi Kleenb7ba30c2008-07-23 21:27:40 -0700647{
Aneesh Kumar K.V0edaecf2012-07-31 16:42:07 -0700648 INIT_LIST_HEAD(&page->lru);
Andi Kleenb7ba30c2008-07-23 21:27:40 -0700649 set_compound_page_dtor(page, free_huge_page);
650 spin_lock(&hugetlb_lock);
Aneesh Kumar K.V9dd540e2012-07-31 16:42:15 -0700651 set_hugetlb_cgroup(page, NULL);
Andi Kleena5516432008-07-23 21:27:41 -0700652 h->nr_huge_pages++;
653 h->nr_huge_pages_node[nid]++;
Andi Kleenb7ba30c2008-07-23 21:27:40 -0700654 spin_unlock(&hugetlb_lock);
655 put_page(page); /* free it into the hugepage allocator */
656}
657
Wu Fengguang20a03072009-06-16 15:32:22 -0700658static void prep_compound_gigantic_page(struct page *page, unsigned long order)
659{
660 int i;
661 int nr_pages = 1 << order;
662 struct page *p = page + 1;
663
664 /* we rely on prep_new_huge_page to set the destructor */
665 set_compound_order(page, order);
666 __SetPageHead(page);
667 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
668 __SetPageTail(p);
Youquan Song58a84aa2011-12-08 14:34:18 -0800669 set_page_count(p, 0);
Wu Fengguang20a03072009-06-16 15:32:22 -0700670 p->first_page = page;
671 }
672}
673
Andrew Morton77959122012-10-08 16:34:11 -0700674/*
675 * PageHuge() only returns true for hugetlbfs pages, but not for normal or
676 * transparent huge pages. See the PageTransHuge() documentation for more
677 * details.
678 */
Wu Fengguang20a03072009-06-16 15:32:22 -0700679int PageHuge(struct page *page)
680{
681 compound_page_dtor *dtor;
682
683 if (!PageCompound(page))
684 return 0;
685
686 page = compound_head(page);
687 dtor = get_compound_page_dtor(page);
688
689 return dtor == free_huge_page;
690}
Naoya Horiguchi43131e12010-05-28 09:29:22 +0900691EXPORT_SYMBOL_GPL(PageHuge);
692
Zhang Yi13d60f42013-06-25 21:19:31 +0800693pgoff_t __basepage_index(struct page *page)
694{
695 struct page *page_head = compound_head(page);
696 pgoff_t index = page_index(page_head);
697 unsigned long compound_idx;
698
699 if (!PageHuge(page_head))
700 return page_index(page);
701
702 if (compound_order(page_head) >= MAX_ORDER)
703 compound_idx = page_to_pfn(page) - page_to_pfn(page_head);
704 else
705 compound_idx = page - page_head;
706
707 return (index << compound_order(page_head)) + compound_idx;
708}
709
Andi Kleena5516432008-07-23 21:27:41 -0700710static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700711{
Linus Torvalds1da177e2005-04-16 15:20:36 -0700712 struct page *page;
Joe Jinf96efd52007-07-15 23:38:12 -0700713
Andi Kleenaa888a72008-07-23 21:27:47 -0700714 if (h->order >= MAX_ORDER)
715 return NULL;
716
Mel Gorman6484eb32009-06-16 15:31:54 -0700717 page = alloc_pages_exact_node(nid,
Nishanth Aravamudan551883a2008-04-29 00:58:26 -0700718 htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE|
719 __GFP_REPEAT|__GFP_NOWARN,
Andi Kleena5516432008-07-23 21:27:41 -0700720 huge_page_order(h));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700721 if (page) {
Gerald Schaefer7f2e9522008-04-28 02:13:29 -0700722 if (arch_prepare_hugepage(page)) {
Gerald Schaefercaff3a22008-08-12 15:08:38 -0700723 __free_pages(page, huge_page_order(h));
Harvey Harrison7b8ee842008-04-28 14:13:19 -0700724 return NULL;
Gerald Schaefer7f2e9522008-04-28 02:13:29 -0700725 }
Andi Kleena5516432008-07-23 21:27:41 -0700726 prep_new_huge_page(h, page, nid);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700727 }
Nishanth Aravamudan63b46132007-10-16 01:26:24 -0700728
729 return page;
730}
731
Andi Kleen5ced66c2008-07-23 21:27:45 -0700732/*
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800733 * common helper functions for hstate_next_node_to_{alloc|free}.
734 * We may have allocated or freed a huge page based on a different
735 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
736 * be outside of *nodes_allowed. Ensure that we use an allowed
737 * node for alloc or free.
Lee Schermerhorn9a76db02009-12-14 17:58:15 -0800738 */
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800739static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
Lee Schermerhorn9a76db02009-12-14 17:58:15 -0800740{
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800741 nid = next_node(nid, *nodes_allowed);
Lee Schermerhorn9a76db02009-12-14 17:58:15 -0800742 if (nid == MAX_NUMNODES)
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800743 nid = first_node(*nodes_allowed);
Lee Schermerhorn9a76db02009-12-14 17:58:15 -0800744 VM_BUG_ON(nid >= MAX_NUMNODES);
745
746 return nid;
747}
748
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800749static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
Andi Kleen5ced66c2008-07-23 21:27:45 -0700750{
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800751 if (!node_isset(nid, *nodes_allowed))
752 nid = next_node_allowed(nid, nodes_allowed);
Lee Schermerhorn9a76db02009-12-14 17:58:15 -0800753 return nid;
Andi Kleen5ced66c2008-07-23 21:27:45 -0700754}
755
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800756/*
757 * returns the previously saved node ["this node"] from which to
758 * allocate a persistent huge page for the pool and advance the
759 * next node from which to allocate, handling wrap at end of node
760 * mask.
761 */
762static int hstate_next_node_to_alloc(struct hstate *h,
763 nodemask_t *nodes_allowed)
764{
765 int nid;
766
767 VM_BUG_ON(!nodes_allowed);
768
769 nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
770 h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
771
772 return nid;
773}
774
775static int alloc_fresh_huge_page(struct hstate *h, nodemask_t *nodes_allowed)
Nishanth Aravamudan63b46132007-10-16 01:26:24 -0700776{
777 struct page *page;
778 int start_nid;
779 int next_nid;
780 int ret = 0;
781
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800782 start_nid = hstate_next_node_to_alloc(h, nodes_allowed);
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700783 next_nid = start_nid;
Nishanth Aravamudan63b46132007-10-16 01:26:24 -0700784
785 do {
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700786 page = alloc_fresh_huge_page_node(h, next_nid);
Lee Schermerhorn9a76db02009-12-14 17:58:15 -0800787 if (page) {
Nishanth Aravamudan63b46132007-10-16 01:26:24 -0700788 ret = 1;
Lee Schermerhorn9a76db02009-12-14 17:58:15 -0800789 break;
790 }
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800791 next_nid = hstate_next_node_to_alloc(h, nodes_allowed);
Lee Schermerhorn9a76db02009-12-14 17:58:15 -0800792 } while (next_nid != start_nid);
Nishanth Aravamudan63b46132007-10-16 01:26:24 -0700793
Adam Litke3b116302008-04-28 02:13:06 -0700794 if (ret)
795 count_vm_event(HTLB_BUDDY_PGALLOC);
796 else
797 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
798
Nishanth Aravamudan63b46132007-10-16 01:26:24 -0700799 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700800}
801
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700802/*
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800803 * helper for free_pool_huge_page() - return the previously saved
804 * node ["this node"] from which to free a huge page. Advance the
805 * next node id whether or not we find a free huge page to free so
806 * that the next attempt to free addresses the next node.
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700807 */
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800808static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700809{
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800810 int nid;
Lee Schermerhorn9a76db02009-12-14 17:58:15 -0800811
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800812 VM_BUG_ON(!nodes_allowed);
813
814 nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
815 h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
816
Lee Schermerhorn9a76db02009-12-14 17:58:15 -0800817 return nid;
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700818}
819
820/*
821 * Free huge page from pool from next node to free.
822 * Attempt to keep persistent huge pages more or less
823 * balanced over allowed nodes.
824 * Called with hugetlb_lock locked.
825 */
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800826static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
827 bool acct_surplus)
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700828{
829 int start_nid;
830 int next_nid;
831 int ret = 0;
832
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800833 start_nid = hstate_next_node_to_free(h, nodes_allowed);
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700834 next_nid = start_nid;
835
836 do {
Lee Schermerhorn685f3452009-09-21 17:01:23 -0700837 /*
838 * If we're returning unused surplus pages, only examine
839 * nodes with surplus pages.
840 */
841 if ((!acct_surplus || h->surplus_huge_pages_node[next_nid]) &&
842 !list_empty(&h->hugepage_freelists[next_nid])) {
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700843 struct page *page =
844 list_entry(h->hugepage_freelists[next_nid].next,
845 struct page, lru);
846 list_del(&page->lru);
847 h->free_huge_pages--;
848 h->free_huge_pages_node[next_nid]--;
Lee Schermerhorn685f3452009-09-21 17:01:23 -0700849 if (acct_surplus) {
850 h->surplus_huge_pages--;
851 h->surplus_huge_pages_node[next_nid]--;
852 }
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700853 update_and_free_page(h, page);
854 ret = 1;
Lee Schermerhorn9a76db02009-12-14 17:58:15 -0800855 break;
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700856 }
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800857 next_nid = hstate_next_node_to_free(h, nodes_allowed);
Lee Schermerhorn9a76db02009-12-14 17:58:15 -0800858 } while (next_nid != start_nid);
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700859
860 return ret;
861}
862
Naoya Horiguchibf50bab2010-09-08 10:19:33 +0900863static struct page *alloc_buddy_huge_page(struct hstate *h, int nid)
Adam Litke7893d1d2007-10-16 01:26:18 -0700864{
865 struct page *page;
Naoya Horiguchibf50bab2010-09-08 10:19:33 +0900866 unsigned int r_nid;
Adam Litke7893d1d2007-10-16 01:26:18 -0700867
Andi Kleenaa888a72008-07-23 21:27:47 -0700868 if (h->order >= MAX_ORDER)
869 return NULL;
870
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800871 /*
872 * Assume we will successfully allocate the surplus page to
873 * prevent racing processes from causing the surplus to exceed
874 * overcommit
875 *
876 * This however introduces a different race, where a process B
877 * tries to grow the static hugepage pool while alloc_pages() is
878 * called by process A. B will only examine the per-node
879 * counters in determining if surplus huge pages can be
880 * converted to normal huge pages in adjust_pool_surplus(). A
881 * won't be able to increment the per-node counter, until the
882 * lock is dropped by B, but B doesn't drop hugetlb_lock until
883 * no more huge pages can be converted from surplus to normal
884 * state (and doesn't try to convert again). Thus, we have a
885 * case where a surplus huge page exists, the pool is grown, and
886 * the surplus huge page still exists after, even though it
887 * should just have been converted to a normal huge page. This
888 * does not leak memory, though, as the hugepage will be freed
889 * once it is out of use. It also does not allow the counters to
890 * go out of whack in adjust_pool_surplus() as we don't modify
891 * the node values until we've gotten the hugepage and only the
892 * per-node value is checked there.
893 */
894 spin_lock(&hugetlb_lock);
Andi Kleena5516432008-07-23 21:27:41 -0700895 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800896 spin_unlock(&hugetlb_lock);
897 return NULL;
898 } else {
Andi Kleena5516432008-07-23 21:27:41 -0700899 h->nr_huge_pages++;
900 h->surplus_huge_pages++;
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800901 }
902 spin_unlock(&hugetlb_lock);
903
Naoya Horiguchibf50bab2010-09-08 10:19:33 +0900904 if (nid == NUMA_NO_NODE)
905 page = alloc_pages(htlb_alloc_mask|__GFP_COMP|
906 __GFP_REPEAT|__GFP_NOWARN,
907 huge_page_order(h));
908 else
909 page = alloc_pages_exact_node(nid,
910 htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE|
911 __GFP_REPEAT|__GFP_NOWARN, huge_page_order(h));
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800912
Gerald Schaefercaff3a22008-08-12 15:08:38 -0700913 if (page && arch_prepare_hugepage(page)) {
914 __free_pages(page, huge_page_order(h));
Hillf Dantonea5768c2012-01-10 15:08:30 -0800915 page = NULL;
Gerald Schaefercaff3a22008-08-12 15:08:38 -0700916 }
917
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800918 spin_lock(&hugetlb_lock);
Adam Litke7893d1d2007-10-16 01:26:18 -0700919 if (page) {
Aneesh Kumar K.V0edaecf2012-07-31 16:42:07 -0700920 INIT_LIST_HEAD(&page->lru);
Naoya Horiguchibf50bab2010-09-08 10:19:33 +0900921 r_nid = page_to_nid(page);
Adam Litke7893d1d2007-10-16 01:26:18 -0700922 set_compound_page_dtor(page, free_huge_page);
Aneesh Kumar K.V9dd540e2012-07-31 16:42:15 -0700923 set_hugetlb_cgroup(page, NULL);
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800924 /*
925 * We incremented the global counters already
926 */
Naoya Horiguchibf50bab2010-09-08 10:19:33 +0900927 h->nr_huge_pages_node[r_nid]++;
928 h->surplus_huge_pages_node[r_nid]++;
Adam Litke3b116302008-04-28 02:13:06 -0700929 __count_vm_event(HTLB_BUDDY_PGALLOC);
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800930 } else {
Andi Kleena5516432008-07-23 21:27:41 -0700931 h->nr_huge_pages--;
932 h->surplus_huge_pages--;
Adam Litke3b116302008-04-28 02:13:06 -0700933 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
Adam Litke7893d1d2007-10-16 01:26:18 -0700934 }
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800935 spin_unlock(&hugetlb_lock);
Adam Litke7893d1d2007-10-16 01:26:18 -0700936
937 return page;
938}
939
Adam Litkee4e574b2007-10-16 01:26:19 -0700940/*
Naoya Horiguchibf50bab2010-09-08 10:19:33 +0900941 * This allocation function is useful in the context where vma is irrelevant.
942 * E.g. soft-offlining uses this function because it only cares physical
943 * address of error page.
944 */
945struct page *alloc_huge_page_node(struct hstate *h, int nid)
946{
947 struct page *page;
948
949 spin_lock(&hugetlb_lock);
950 page = dequeue_huge_page_node(h, nid);
951 spin_unlock(&hugetlb_lock);
952
Aneesh Kumar K.V94ae8ba2012-07-31 16:42:35 -0700953 if (!page)
Naoya Horiguchibf50bab2010-09-08 10:19:33 +0900954 page = alloc_buddy_huge_page(h, nid);
955
956 return page;
957}
958
959/*
Lucas De Marchi25985ed2011-03-30 22:57:33 -0300960 * Increase the hugetlb pool such that it can accommodate a reservation
Adam Litkee4e574b2007-10-16 01:26:19 -0700961 * of size 'delta'.
962 */
Andi Kleena5516432008-07-23 21:27:41 -0700963static int gather_surplus_pages(struct hstate *h, int delta)
Adam Litkee4e574b2007-10-16 01:26:19 -0700964{
965 struct list_head surplus_list;
966 struct page *page, *tmp;
967 int ret, i;
968 int needed, allocated;
Hillf Danton28073b02012-03-21 16:34:00 -0700969 bool alloc_ok = true;
Adam Litkee4e574b2007-10-16 01:26:19 -0700970
Andi Kleena5516432008-07-23 21:27:41 -0700971 needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
Adam Litkeac09b3a2008-03-04 14:29:38 -0800972 if (needed <= 0) {
Andi Kleena5516432008-07-23 21:27:41 -0700973 h->resv_huge_pages += delta;
Adam Litkee4e574b2007-10-16 01:26:19 -0700974 return 0;
Adam Litkeac09b3a2008-03-04 14:29:38 -0800975 }
Adam Litkee4e574b2007-10-16 01:26:19 -0700976
977 allocated = 0;
978 INIT_LIST_HEAD(&surplus_list);
979
980 ret = -ENOMEM;
981retry:
982 spin_unlock(&hugetlb_lock);
983 for (i = 0; i < needed; i++) {
Naoya Horiguchibf50bab2010-09-08 10:19:33 +0900984 page = alloc_buddy_huge_page(h, NUMA_NO_NODE);
Hillf Danton28073b02012-03-21 16:34:00 -0700985 if (!page) {
986 alloc_ok = false;
987 break;
988 }
Adam Litkee4e574b2007-10-16 01:26:19 -0700989 list_add(&page->lru, &surplus_list);
990 }
Hillf Danton28073b02012-03-21 16:34:00 -0700991 allocated += i;
Adam Litkee4e574b2007-10-16 01:26:19 -0700992
993 /*
994 * After retaking hugetlb_lock, we need to recalculate 'needed'
995 * because either resv_huge_pages or free_huge_pages may have changed.
996 */
997 spin_lock(&hugetlb_lock);
Andi Kleena5516432008-07-23 21:27:41 -0700998 needed = (h->resv_huge_pages + delta) -
999 (h->free_huge_pages + allocated);
Hillf Danton28073b02012-03-21 16:34:00 -07001000 if (needed > 0) {
1001 if (alloc_ok)
1002 goto retry;
1003 /*
1004 * We were not able to allocate enough pages to
1005 * satisfy the entire reservation so we free what
1006 * we've allocated so far.
1007 */
1008 goto free;
1009 }
Adam Litkee4e574b2007-10-16 01:26:19 -07001010 /*
1011 * The surplus_list now contains _at_least_ the number of extra pages
Lucas De Marchi25985ed2011-03-30 22:57:33 -03001012 * needed to accommodate the reservation. Add the appropriate number
Adam Litkee4e574b2007-10-16 01:26:19 -07001013 * of pages to the hugetlb pool and free the extras back to the buddy
Adam Litkeac09b3a2008-03-04 14:29:38 -08001014 * allocator. Commit the entire reservation here to prevent another
1015 * process from stealing the pages as they are added to the pool but
1016 * before they are reserved.
Adam Litkee4e574b2007-10-16 01:26:19 -07001017 */
1018 needed += allocated;
Andi Kleena5516432008-07-23 21:27:41 -07001019 h->resv_huge_pages += delta;
Adam Litkee4e574b2007-10-16 01:26:19 -07001020 ret = 0;
Naoya Horiguchia9869b82010-09-08 10:19:37 +09001021
Adam Litke19fc3f02008-04-28 02:12:20 -07001022 /* Free the needed pages to the hugetlb pool */
Adam Litkee4e574b2007-10-16 01:26:19 -07001023 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
Adam Litke19fc3f02008-04-28 02:12:20 -07001024 if ((--needed) < 0)
1025 break;
Naoya Horiguchia9869b82010-09-08 10:19:37 +09001026 /*
1027 * This page is now managed by the hugetlb allocator and has
1028 * no users -- drop the buddy allocator's reference.
1029 */
1030 put_page_testzero(page);
1031 VM_BUG_ON(page_count(page));
Andi Kleena5516432008-07-23 21:27:41 -07001032 enqueue_huge_page(h, page);
Adam Litke19fc3f02008-04-28 02:12:20 -07001033 }
Hillf Danton28073b02012-03-21 16:34:00 -07001034free:
Hillf Dantonb0365c82011-12-28 15:57:16 -08001035 spin_unlock(&hugetlb_lock);
Adam Litke19fc3f02008-04-28 02:12:20 -07001036
1037 /* Free unnecessary surplus pages to the buddy allocator */
1038 if (!list_empty(&surplus_list)) {
Adam Litke19fc3f02008-04-28 02:12:20 -07001039 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
Naoya Horiguchia9869b82010-09-08 10:19:37 +09001040 put_page(page);
Adam Litkeaf767cb2007-10-16 01:26:25 -07001041 }
Adam Litkee4e574b2007-10-16 01:26:19 -07001042 }
Naoya Horiguchia9869b82010-09-08 10:19:37 +09001043 spin_lock(&hugetlb_lock);
Adam Litkee4e574b2007-10-16 01:26:19 -07001044
1045 return ret;
1046}
1047
1048/*
1049 * When releasing a hugetlb pool reservation, any surplus pages that were
1050 * allocated to satisfy the reservation must be explicitly freed if they were
1051 * never used.
Lee Schermerhorn685f3452009-09-21 17:01:23 -07001052 * Called with hugetlb_lock held.
Adam Litkee4e574b2007-10-16 01:26:19 -07001053 */
Andi Kleena5516432008-07-23 21:27:41 -07001054static void return_unused_surplus_pages(struct hstate *h,
1055 unsigned long unused_resv_pages)
Adam Litkee4e574b2007-10-16 01:26:19 -07001056{
Adam Litkee4e574b2007-10-16 01:26:19 -07001057 unsigned long nr_pages;
1058
Adam Litkeac09b3a2008-03-04 14:29:38 -08001059 /* Uncommit the reservation */
Andi Kleena5516432008-07-23 21:27:41 -07001060 h->resv_huge_pages -= unused_resv_pages;
Adam Litkeac09b3a2008-03-04 14:29:38 -08001061
Andi Kleenaa888a72008-07-23 21:27:47 -07001062 /* Cannot return gigantic pages currently */
1063 if (h->order >= MAX_ORDER)
1064 return;
1065
Andi Kleena5516432008-07-23 21:27:41 -07001066 nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
Adam Litkee4e574b2007-10-16 01:26:19 -07001067
Lee Schermerhorn685f3452009-09-21 17:01:23 -07001068 /*
1069 * We want to release as many surplus pages as possible, spread
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -08001070 * evenly across all nodes with memory. Iterate across these nodes
1071 * until we can no longer free unreserved surplus pages. This occurs
1072 * when the nodes with surplus pages have no free pages.
1073 * free_pool_huge_page() will balance the the freed pages across the
1074 * on-line nodes with memory and will handle the hstate accounting.
Lee Schermerhorn685f3452009-09-21 17:01:23 -07001075 */
1076 while (nr_pages--) {
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08001077 if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1))
Lee Schermerhorn685f3452009-09-21 17:01:23 -07001078 break;
Adam Litkee4e574b2007-10-16 01:26:19 -07001079 }
1080}
1081
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001082/*
1083 * Determine if the huge page at addr within the vma has an associated
1084 * reservation. Where it does not we will need to logically increase
David Gibson90481622012-03-21 16:34:12 -07001085 * reservation and actually increase subpool usage before an allocation
1086 * can occur. Where any new reservation would be required the
1087 * reservation change is prepared, but not committed. Once the page
1088 * has been allocated from the subpool and instantiated the change should
1089 * be committed via vma_commit_reservation. No action is required on
1090 * failure.
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001091 */
Roel Kluine2f17d92009-03-31 15:23:15 -07001092static long vma_needs_reservation(struct hstate *h,
Andi Kleena5516432008-07-23 21:27:41 -07001093 struct vm_area_struct *vma, unsigned long addr)
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001094{
1095 struct address_space *mapping = vma->vm_file->f_mapping;
1096 struct inode *inode = mapping->host;
1097
Mel Gormanf83a2752009-05-28 14:34:40 -07001098 if (vma->vm_flags & VM_MAYSHARE) {
Andi Kleena5516432008-07-23 21:27:41 -07001099 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001100 return region_chg(&inode->i_mapping->private_list,
1101 idx, idx + 1);
1102
Andy Whitcroft84afd992008-07-23 21:27:32 -07001103 } else if (!is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
1104 return 1;
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001105
Andy Whitcroft84afd992008-07-23 21:27:32 -07001106 } else {
Roel Kluine2f17d92009-03-31 15:23:15 -07001107 long err;
Andi Kleena5516432008-07-23 21:27:41 -07001108 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
Andy Whitcroft84afd992008-07-23 21:27:32 -07001109 struct resv_map *reservations = vma_resv_map(vma);
1110
1111 err = region_chg(&reservations->regions, idx, idx + 1);
1112 if (err < 0)
1113 return err;
1114 return 0;
1115 }
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001116}
Andi Kleena5516432008-07-23 21:27:41 -07001117static void vma_commit_reservation(struct hstate *h,
1118 struct vm_area_struct *vma, unsigned long addr)
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001119{
1120 struct address_space *mapping = vma->vm_file->f_mapping;
1121 struct inode *inode = mapping->host;
1122
Mel Gormanf83a2752009-05-28 14:34:40 -07001123 if (vma->vm_flags & VM_MAYSHARE) {
Andi Kleena5516432008-07-23 21:27:41 -07001124 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001125 region_add(&inode->i_mapping->private_list, idx, idx + 1);
Andy Whitcroft84afd992008-07-23 21:27:32 -07001126
1127 } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
Andi Kleena5516432008-07-23 21:27:41 -07001128 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
Andy Whitcroft84afd992008-07-23 21:27:32 -07001129 struct resv_map *reservations = vma_resv_map(vma);
1130
1131 /* Mark this page used in the map. */
1132 region_add(&reservations->regions, idx, idx + 1);
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001133 }
1134}
1135
David Gibson27a85ef2006-03-22 00:08:56 -08001136static struct page *alloc_huge_page(struct vm_area_struct *vma,
Mel Gorman04f2cbe2008-07-23 21:27:25 -07001137 unsigned long addr, int avoid_reserve)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001138{
David Gibson90481622012-03-21 16:34:12 -07001139 struct hugepage_subpool *spool = subpool_vma(vma);
Andi Kleena5516432008-07-23 21:27:41 -07001140 struct hstate *h = hstate_vma(vma);
Adam Litke348ea202007-11-14 16:59:37 -08001141 struct page *page;
Roel Kluine2f17d92009-03-31 15:23:15 -07001142 long chg;
Aneesh Kumar K.V6d76dcf2012-07-31 16:42:18 -07001143 int ret, idx;
1144 struct hugetlb_cgroup *h_cg;
Adam Litke2fc39ce2007-11-14 16:59:39 -08001145
Aneesh Kumar K.V6d76dcf2012-07-31 16:42:18 -07001146 idx = hstate_index(h);
Mel Gormana1e78772008-07-23 21:27:23 -07001147 /*
David Gibson90481622012-03-21 16:34:12 -07001148 * Processes that did not create the mapping will have no
1149 * reserves and will not have accounted against subpool
1150 * limit. Check that the subpool limit can be made before
1151 * satisfying the allocation MAP_NORESERVE mappings may also
1152 * need pages and subpool limit allocated allocated if no reserve
1153 * mapping overlaps.
Mel Gormana1e78772008-07-23 21:27:23 -07001154 */
Andi Kleena5516432008-07-23 21:27:41 -07001155 chg = vma_needs_reservation(h, vma, addr);
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001156 if (chg < 0)
Aneesh Kumar K.V76dcee72012-07-31 16:41:57 -07001157 return ERR_PTR(-ENOMEM);
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001158 if (chg)
David Gibson90481622012-03-21 16:34:12 -07001159 if (hugepage_subpool_get_pages(spool, chg))
Aneesh Kumar K.V76dcee72012-07-31 16:41:57 -07001160 return ERR_PTR(-ENOSPC);
Mel Gormana1e78772008-07-23 21:27:23 -07001161
Aneesh Kumar K.V6d76dcf2012-07-31 16:42:18 -07001162 ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg);
1163 if (ret) {
1164 hugepage_subpool_put_pages(spool, chg);
1165 return ERR_PTR(-ENOSPC);
1166 }
Mel Gormana1e78772008-07-23 21:27:23 -07001167 spin_lock(&hugetlb_lock);
Andi Kleena5516432008-07-23 21:27:41 -07001168 page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve);
Aneesh Kumar K.V94ae8ba2012-07-31 16:42:35 -07001169 if (page) {
1170 /* update page cgroup details */
1171 hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h),
1172 h_cg, page);
1173 spin_unlock(&hugetlb_lock);
1174 } else {
1175 spin_unlock(&hugetlb_lock);
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001176 page = alloc_buddy_huge_page(h, NUMA_NO_NODE);
Mel Gormana1e78772008-07-23 21:27:23 -07001177 if (!page) {
Aneesh Kumar K.V6d76dcf2012-07-31 16:42:18 -07001178 hugetlb_cgroup_uncharge_cgroup(idx,
1179 pages_per_huge_page(h),
1180 h_cg);
David Gibson90481622012-03-21 16:34:12 -07001181 hugepage_subpool_put_pages(spool, chg);
Aneesh Kumar K.V76dcee72012-07-31 16:41:57 -07001182 return ERR_PTR(-ENOSPC);
Mel Gormana1e78772008-07-23 21:27:23 -07001183 }
Aneesh Kumar K.V79dbb232012-07-31 16:42:32 -07001184 spin_lock(&hugetlb_lock);
Aneesh Kumar K.V94ae8ba2012-07-31 16:42:35 -07001185 hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h),
1186 h_cg, page);
Aneesh Kumar K.V79dbb232012-07-31 16:42:32 -07001187 list_move(&page->lru, &h->hugepage_activelist);
1188 spin_unlock(&hugetlb_lock);
Mel Gormana1e78772008-07-23 21:27:23 -07001189 }
1190
David Gibson90481622012-03-21 16:34:12 -07001191 set_page_private(page, (unsigned long)spool);
Mel Gormana1e78772008-07-23 21:27:23 -07001192
Andi Kleena5516432008-07-23 21:27:41 -07001193 vma_commit_reservation(h, vma, addr);
Adam Litke90d8b7e2007-11-14 16:59:42 -08001194 return page;
David Gibsonb45b5bd2006-03-22 00:08:55 -08001195}
1196
Cyrill Gorcunov91f47662009-01-06 14:40:33 -08001197int __weak alloc_bootmem_huge_page(struct hstate *h)
Andi Kleenaa888a72008-07-23 21:27:47 -07001198{
1199 struct huge_bootmem_page *m;
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08001200 int nr_nodes = nodes_weight(node_states[N_MEMORY]);
Andi Kleenaa888a72008-07-23 21:27:47 -07001201
1202 while (nr_nodes) {
1203 void *addr;
1204
1205 addr = __alloc_bootmem_node_nopanic(
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001206 NODE_DATA(hstate_next_node_to_alloc(h,
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08001207 &node_states[N_MEMORY])),
Andi Kleenaa888a72008-07-23 21:27:47 -07001208 huge_page_size(h), huge_page_size(h), 0);
1209
1210 if (addr) {
1211 /*
1212 * Use the beginning of the huge page to store the
1213 * huge_bootmem_page struct (until gather_bootmem
1214 * puts them into the mem_map).
1215 */
1216 m = addr;
Cyrill Gorcunov91f47662009-01-06 14:40:33 -08001217 goto found;
Andi Kleenaa888a72008-07-23 21:27:47 -07001218 }
Andi Kleenaa888a72008-07-23 21:27:47 -07001219 nr_nodes--;
1220 }
1221 return 0;
1222
1223found:
1224 BUG_ON((unsigned long)virt_to_phys(m) & (huge_page_size(h) - 1));
1225 /* Put them into a private list first because mem_map is not up yet */
1226 list_add(&m->list, &huge_boot_pages);
1227 m->hstate = h;
1228 return 1;
1229}
1230
Andy Whitcroft18229df2008-11-06 12:53:27 -08001231static void prep_compound_huge_page(struct page *page, int order)
1232{
1233 if (unlikely(order > (MAX_ORDER - 1)))
1234 prep_compound_gigantic_page(page, order);
1235 else
1236 prep_compound_page(page, order);
1237}
1238
Andi Kleenaa888a72008-07-23 21:27:47 -07001239/* Put bootmem huge pages into the standard lists after mem_map is up */
1240static void __init gather_bootmem_prealloc(void)
1241{
1242 struct huge_bootmem_page *m;
1243
1244 list_for_each_entry(m, &huge_boot_pages, list) {
Andi Kleenaa888a72008-07-23 21:27:47 -07001245 struct hstate *h = m->hstate;
Becky Bruceee8f2482011-07-25 17:11:50 -07001246 struct page *page;
1247
1248#ifdef CONFIG_HIGHMEM
1249 page = pfn_to_page(m->phys >> PAGE_SHIFT);
1250 free_bootmem_late((unsigned long)m,
1251 sizeof(struct huge_bootmem_page));
1252#else
1253 page = virt_to_page(m);
1254#endif
Andi Kleenaa888a72008-07-23 21:27:47 -07001255 __ClearPageReserved(page);
1256 WARN_ON(page_count(page) != 1);
Andy Whitcroft18229df2008-11-06 12:53:27 -08001257 prep_compound_huge_page(page, h->order);
Andi Kleenaa888a72008-07-23 21:27:47 -07001258 prep_new_huge_page(h, page, page_to_nid(page));
Rafael Aquinib0320c72011-06-15 15:08:39 -07001259 /*
1260 * If we had gigantic hugepages allocated at boot time, we need
1261 * to restore the 'stolen' pages to totalram_pages in order to
1262 * fix confusing memory reports from free(1) and another
1263 * side-effects, like CommitLimit going negative.
1264 */
1265 if (h->order > (MAX_ORDER - 1))
Jiang Liu3dcc0572013-07-03 15:03:21 -07001266 adjust_managed_page_count(page, 1 << h->order);
Andi Kleenaa888a72008-07-23 21:27:47 -07001267 }
1268}
1269
Andi Kleen8faa8b02008-07-23 21:27:48 -07001270static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001271{
1272 unsigned long i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001273
Andi Kleene5ff2152008-07-23 21:27:42 -07001274 for (i = 0; i < h->max_huge_pages; ++i) {
Andi Kleenaa888a72008-07-23 21:27:47 -07001275 if (h->order >= MAX_ORDER) {
1276 if (!alloc_bootmem_huge_page(h))
1277 break;
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -08001278 } else if (!alloc_fresh_huge_page(h,
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08001279 &node_states[N_MEMORY]))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001280 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001281 }
Andi Kleen8faa8b02008-07-23 21:27:48 -07001282 h->max_huge_pages = i;
Andi Kleene5ff2152008-07-23 21:27:42 -07001283}
1284
1285static void __init hugetlb_init_hstates(void)
1286{
1287 struct hstate *h;
1288
1289 for_each_hstate(h) {
Andi Kleen8faa8b02008-07-23 21:27:48 -07001290 /* oversize hugepages were init'ed in early boot */
1291 if (h->order < MAX_ORDER)
1292 hugetlb_hstate_alloc_pages(h);
Andi Kleene5ff2152008-07-23 21:27:42 -07001293 }
1294}
1295
Andi Kleen4abd32d2008-07-23 21:27:49 -07001296static char * __init memfmt(char *buf, unsigned long n)
1297{
1298 if (n >= (1UL << 30))
1299 sprintf(buf, "%lu GB", n >> 30);
1300 else if (n >= (1UL << 20))
1301 sprintf(buf, "%lu MB", n >> 20);
1302 else
1303 sprintf(buf, "%lu KB", n >> 10);
1304 return buf;
1305}
1306
Andi Kleene5ff2152008-07-23 21:27:42 -07001307static void __init report_hugepages(void)
1308{
1309 struct hstate *h;
1310
1311 for_each_hstate(h) {
Andi Kleen4abd32d2008-07-23 21:27:49 -07001312 char buf[32];
Andrew Mortonffb22af2013-02-22 16:32:08 -08001313 pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n",
Andi Kleen4abd32d2008-07-23 21:27:49 -07001314 memfmt(buf, huge_page_size(h)),
1315 h->free_huge_pages);
Andi Kleene5ff2152008-07-23 21:27:42 -07001316 }
1317}
1318
Linus Torvalds1da177e2005-04-16 15:20:36 -07001319#ifdef CONFIG_HIGHMEM
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001320static void try_to_free_low(struct hstate *h, unsigned long count,
1321 nodemask_t *nodes_allowed)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001322{
Christoph Lameter4415cc82006-09-25 23:31:55 -07001323 int i;
1324
Andi Kleenaa888a72008-07-23 21:27:47 -07001325 if (h->order >= MAX_ORDER)
1326 return;
1327
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001328 for_each_node_mask(i, *nodes_allowed) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001329 struct page *page, *next;
Andi Kleena5516432008-07-23 21:27:41 -07001330 struct list_head *freel = &h->hugepage_freelists[i];
1331 list_for_each_entry_safe(page, next, freel, lru) {
1332 if (count >= h->nr_huge_pages)
Adam Litke6b0c8802007-10-16 01:26:23 -07001333 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001334 if (PageHighMem(page))
1335 continue;
1336 list_del(&page->lru);
Andi Kleene5ff2152008-07-23 21:27:42 -07001337 update_and_free_page(h, page);
Andi Kleena5516432008-07-23 21:27:41 -07001338 h->free_huge_pages--;
1339 h->free_huge_pages_node[page_to_nid(page)]--;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001340 }
1341 }
1342}
1343#else
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001344static inline void try_to_free_low(struct hstate *h, unsigned long count,
1345 nodemask_t *nodes_allowed)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001346{
1347}
1348#endif
1349
Wu Fengguang20a03072009-06-16 15:32:22 -07001350/*
1351 * Increment or decrement surplus_huge_pages. Keep node-specific counters
1352 * balanced by operating on them in a round-robin fashion.
1353 * Returns 1 if an adjustment was made.
1354 */
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001355static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
1356 int delta)
Wu Fengguang20a03072009-06-16 15:32:22 -07001357{
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001358 int start_nid, next_nid;
Wu Fengguang20a03072009-06-16 15:32:22 -07001359 int ret = 0;
1360
1361 VM_BUG_ON(delta != -1 && delta != 1);
Wu Fengguang20a03072009-06-16 15:32:22 -07001362
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001363 if (delta < 0)
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001364 start_nid = hstate_next_node_to_alloc(h, nodes_allowed);
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001365 else
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001366 start_nid = hstate_next_node_to_free(h, nodes_allowed);
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001367 next_nid = start_nid;
1368
1369 do {
1370 int nid = next_nid;
1371 if (delta < 0) {
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001372 /*
1373 * To shrink on this node, there must be a surplus page
1374 */
Lee Schermerhorn9a76db02009-12-14 17:58:15 -08001375 if (!h->surplus_huge_pages_node[nid]) {
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001376 next_nid = hstate_next_node_to_alloc(h,
1377 nodes_allowed);
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001378 continue;
Lee Schermerhorn9a76db02009-12-14 17:58:15 -08001379 }
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001380 }
1381 if (delta > 0) {
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001382 /*
1383 * Surplus cannot exceed the total number of pages
1384 */
1385 if (h->surplus_huge_pages_node[nid] >=
Lee Schermerhorn9a76db02009-12-14 17:58:15 -08001386 h->nr_huge_pages_node[nid]) {
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001387 next_nid = hstate_next_node_to_free(h,
1388 nodes_allowed);
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001389 continue;
Lee Schermerhorn9a76db02009-12-14 17:58:15 -08001390 }
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001391 }
Wu Fengguang20a03072009-06-16 15:32:22 -07001392
1393 h->surplus_huge_pages += delta;
1394 h->surplus_huge_pages_node[nid] += delta;
1395 ret = 1;
1396 break;
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001397 } while (next_nid != start_nid);
Wu Fengguang20a03072009-06-16 15:32:22 -07001398
Wu Fengguang20a03072009-06-16 15:32:22 -07001399 return ret;
1400}
1401
Andi Kleena5516432008-07-23 21:27:41 -07001402#define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001403static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count,
1404 nodemask_t *nodes_allowed)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001405{
Adam Litke7893d1d2007-10-16 01:26:18 -07001406 unsigned long min_count, ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001407
Andi Kleenaa888a72008-07-23 21:27:47 -07001408 if (h->order >= MAX_ORDER)
1409 return h->max_huge_pages;
1410
Adam Litke7893d1d2007-10-16 01:26:18 -07001411 /*
1412 * Increase the pool size
1413 * First take pages out of surplus state. Then make up the
1414 * remaining difference by allocating fresh huge pages.
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001415 *
1416 * We might race with alloc_buddy_huge_page() here and be unable
1417 * to convert a surplus huge page to a normal huge page. That is
1418 * not critical, though, it just means the overall size of the
1419 * pool might be one hugepage larger than it needs to be, but
1420 * within all the constraints specified by the sysctls.
Adam Litke7893d1d2007-10-16 01:26:18 -07001421 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001422 spin_lock(&hugetlb_lock);
Andi Kleena5516432008-07-23 21:27:41 -07001423 while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001424 if (!adjust_pool_surplus(h, nodes_allowed, -1))
Adam Litke7893d1d2007-10-16 01:26:18 -07001425 break;
1426 }
1427
Andi Kleena5516432008-07-23 21:27:41 -07001428 while (count > persistent_huge_pages(h)) {
Adam Litke7893d1d2007-10-16 01:26:18 -07001429 /*
1430 * If this allocation races such that we no longer need the
1431 * page, free_huge_page will handle it by freeing the page
1432 * and reducing the surplus.
1433 */
1434 spin_unlock(&hugetlb_lock);
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001435 ret = alloc_fresh_huge_page(h, nodes_allowed);
Adam Litke7893d1d2007-10-16 01:26:18 -07001436 spin_lock(&hugetlb_lock);
1437 if (!ret)
1438 goto out;
1439
Mel Gorman536240f22009-12-14 17:59:56 -08001440 /* Bail for signals. Probably ctrl-c from user */
1441 if (signal_pending(current))
1442 goto out;
Adam Litke7893d1d2007-10-16 01:26:18 -07001443 }
Adam Litke7893d1d2007-10-16 01:26:18 -07001444
1445 /*
1446 * Decrease the pool size
1447 * First return free pages to the buddy allocator (being careful
1448 * to keep enough around to satisfy reservations). Then place
1449 * pages into surplus state as needed so the pool will shrink
1450 * to the desired size as pages become free.
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001451 *
1452 * By placing pages into the surplus state independent of the
1453 * overcommit value, we are allowing the surplus pool size to
1454 * exceed overcommit. There are few sane options here. Since
1455 * alloc_buddy_huge_page() is checking the global counter,
1456 * though, we'll note that we're not allowed to exceed surplus
1457 * and won't grow the pool anywhere else. Not until one of the
1458 * sysctls are changed, or the surplus pages go out of use.
Adam Litke7893d1d2007-10-16 01:26:18 -07001459 */
Andi Kleena5516432008-07-23 21:27:41 -07001460 min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
Adam Litke6b0c8802007-10-16 01:26:23 -07001461 min_count = max(count, min_count);
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001462 try_to_free_low(h, min_count, nodes_allowed);
Andi Kleena5516432008-07-23 21:27:41 -07001463 while (min_count < persistent_huge_pages(h)) {
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001464 if (!free_pool_huge_page(h, nodes_allowed, 0))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001465 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001466 }
Andi Kleena5516432008-07-23 21:27:41 -07001467 while (count < persistent_huge_pages(h)) {
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001468 if (!adjust_pool_surplus(h, nodes_allowed, 1))
Adam Litke7893d1d2007-10-16 01:26:18 -07001469 break;
1470 }
1471out:
Andi Kleena5516432008-07-23 21:27:41 -07001472 ret = persistent_huge_pages(h);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001473 spin_unlock(&hugetlb_lock);
Adam Litke7893d1d2007-10-16 01:26:18 -07001474 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001475}
1476
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001477#define HSTATE_ATTR_RO(_name) \
1478 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
1479
1480#define HSTATE_ATTR(_name) \
1481 static struct kobj_attribute _name##_attr = \
1482 __ATTR(_name, 0644, _name##_show, _name##_store)
1483
1484static struct kobject *hugepages_kobj;
1485static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
1486
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001487static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
1488
1489static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001490{
1491 int i;
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001492
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001493 for (i = 0; i < HUGE_MAX_HSTATE; i++)
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001494 if (hstate_kobjs[i] == kobj) {
1495 if (nidp)
1496 *nidp = NUMA_NO_NODE;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001497 return &hstates[i];
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001498 }
1499
1500 return kobj_to_node_hstate(kobj, nidp);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001501}
1502
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001503static ssize_t nr_hugepages_show_common(struct kobject *kobj,
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001504 struct kobj_attribute *attr, char *buf)
1505{
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001506 struct hstate *h;
1507 unsigned long nr_huge_pages;
1508 int nid;
1509
1510 h = kobj_to_hstate(kobj, &nid);
1511 if (nid == NUMA_NO_NODE)
1512 nr_huge_pages = h->nr_huge_pages;
1513 else
1514 nr_huge_pages = h->nr_huge_pages_node[nid];
1515
1516 return sprintf(buf, "%lu\n", nr_huge_pages);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001517}
Eric B Munsonadbe8722011-01-13 15:47:27 -08001518
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001519static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
1520 struct kobject *kobj, struct kobj_attribute *attr,
1521 const char *buf, size_t len)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001522{
1523 int err;
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001524 int nid;
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001525 unsigned long count;
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001526 struct hstate *h;
David Rientjesbad44b52009-12-14 17:58:38 -08001527 NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001528
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001529 err = strict_strtoul(buf, 10, &count);
Eric B Munson73ae31e2011-01-13 15:47:28 -08001530 if (err)
Eric B Munsonadbe8722011-01-13 15:47:27 -08001531 goto out;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001532
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001533 h = kobj_to_hstate(kobj, &nid);
Eric B Munsonadbe8722011-01-13 15:47:27 -08001534 if (h->order >= MAX_ORDER) {
1535 err = -EINVAL;
1536 goto out;
1537 }
1538
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001539 if (nid == NUMA_NO_NODE) {
1540 /*
1541 * global hstate attribute
1542 */
1543 if (!(obey_mempolicy &&
1544 init_nodemask_of_mempolicy(nodes_allowed))) {
1545 NODEMASK_FREE(nodes_allowed);
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08001546 nodes_allowed = &node_states[N_MEMORY];
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001547 }
1548 } else if (nodes_allowed) {
1549 /*
1550 * per node hstate attribute: adjust count to global,
1551 * but restrict alloc/free to the specified node.
1552 */
1553 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
1554 init_nodemask_of_node(nodes_allowed, nid);
1555 } else
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08001556 nodes_allowed = &node_states[N_MEMORY];
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001557
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001558 h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001559
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08001560 if (nodes_allowed != &node_states[N_MEMORY])
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001561 NODEMASK_FREE(nodes_allowed);
1562
1563 return len;
Eric B Munsonadbe8722011-01-13 15:47:27 -08001564out:
1565 NODEMASK_FREE(nodes_allowed);
1566 return err;
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001567}
1568
1569static ssize_t nr_hugepages_show(struct kobject *kobj,
1570 struct kobj_attribute *attr, char *buf)
1571{
1572 return nr_hugepages_show_common(kobj, attr, buf);
1573}
1574
1575static ssize_t nr_hugepages_store(struct kobject *kobj,
1576 struct kobj_attribute *attr, const char *buf, size_t len)
1577{
1578 return nr_hugepages_store_common(false, kobj, attr, buf, len);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001579}
1580HSTATE_ATTR(nr_hugepages);
1581
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001582#ifdef CONFIG_NUMA
1583
1584/*
1585 * hstate attribute for optionally mempolicy-based constraint on persistent
1586 * huge page alloc/free.
1587 */
1588static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
1589 struct kobj_attribute *attr, char *buf)
1590{
1591 return nr_hugepages_show_common(kobj, attr, buf);
1592}
1593
1594static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
1595 struct kobj_attribute *attr, const char *buf, size_t len)
1596{
1597 return nr_hugepages_store_common(true, kobj, attr, buf, len);
1598}
1599HSTATE_ATTR(nr_hugepages_mempolicy);
1600#endif
1601
1602
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001603static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
1604 struct kobj_attribute *attr, char *buf)
1605{
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001606 struct hstate *h = kobj_to_hstate(kobj, NULL);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001607 return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
1608}
Eric B Munsonadbe8722011-01-13 15:47:27 -08001609
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001610static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
1611 struct kobj_attribute *attr, const char *buf, size_t count)
1612{
1613 int err;
1614 unsigned long input;
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001615 struct hstate *h = kobj_to_hstate(kobj, NULL);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001616
Eric B Munsonadbe8722011-01-13 15:47:27 -08001617 if (h->order >= MAX_ORDER)
1618 return -EINVAL;
1619
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001620 err = strict_strtoul(buf, 10, &input);
1621 if (err)
Eric B Munson73ae31e2011-01-13 15:47:28 -08001622 return err;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001623
1624 spin_lock(&hugetlb_lock);
1625 h->nr_overcommit_huge_pages = input;
1626 spin_unlock(&hugetlb_lock);
1627
1628 return count;
1629}
1630HSTATE_ATTR(nr_overcommit_hugepages);
1631
1632static ssize_t free_hugepages_show(struct kobject *kobj,
1633 struct kobj_attribute *attr, char *buf)
1634{
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001635 struct hstate *h;
1636 unsigned long free_huge_pages;
1637 int nid;
1638
1639 h = kobj_to_hstate(kobj, &nid);
1640 if (nid == NUMA_NO_NODE)
1641 free_huge_pages = h->free_huge_pages;
1642 else
1643 free_huge_pages = h->free_huge_pages_node[nid];
1644
1645 return sprintf(buf, "%lu\n", free_huge_pages);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001646}
1647HSTATE_ATTR_RO(free_hugepages);
1648
1649static ssize_t resv_hugepages_show(struct kobject *kobj,
1650 struct kobj_attribute *attr, char *buf)
1651{
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001652 struct hstate *h = kobj_to_hstate(kobj, NULL);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001653 return sprintf(buf, "%lu\n", h->resv_huge_pages);
1654}
1655HSTATE_ATTR_RO(resv_hugepages);
1656
1657static ssize_t surplus_hugepages_show(struct kobject *kobj,
1658 struct kobj_attribute *attr, char *buf)
1659{
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001660 struct hstate *h;
1661 unsigned long surplus_huge_pages;
1662 int nid;
1663
1664 h = kobj_to_hstate(kobj, &nid);
1665 if (nid == NUMA_NO_NODE)
1666 surplus_huge_pages = h->surplus_huge_pages;
1667 else
1668 surplus_huge_pages = h->surplus_huge_pages_node[nid];
1669
1670 return sprintf(buf, "%lu\n", surplus_huge_pages);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001671}
1672HSTATE_ATTR_RO(surplus_hugepages);
1673
1674static struct attribute *hstate_attrs[] = {
1675 &nr_hugepages_attr.attr,
1676 &nr_overcommit_hugepages_attr.attr,
1677 &free_hugepages_attr.attr,
1678 &resv_hugepages_attr.attr,
1679 &surplus_hugepages_attr.attr,
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001680#ifdef CONFIG_NUMA
1681 &nr_hugepages_mempolicy_attr.attr,
1682#endif
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001683 NULL,
1684};
1685
1686static struct attribute_group hstate_attr_group = {
1687 .attrs = hstate_attrs,
1688};
1689
Jeff Mahoney094e9532010-02-02 13:44:14 -08001690static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
1691 struct kobject **hstate_kobjs,
1692 struct attribute_group *hstate_attr_group)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001693{
1694 int retval;
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07001695 int hi = hstate_index(h);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001696
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001697 hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
1698 if (!hstate_kobjs[hi])
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001699 return -ENOMEM;
1700
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001701 retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001702 if (retval)
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001703 kobject_put(hstate_kobjs[hi]);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001704
1705 return retval;
1706}
1707
1708static void __init hugetlb_sysfs_init(void)
1709{
1710 struct hstate *h;
1711 int err;
1712
1713 hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
1714 if (!hugepages_kobj)
1715 return;
1716
1717 for_each_hstate(h) {
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001718 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
1719 hstate_kobjs, &hstate_attr_group);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001720 if (err)
Andrew Mortonffb22af2013-02-22 16:32:08 -08001721 pr_err("Hugetlb: Unable to add hstate %s", h->name);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001722 }
1723}
1724
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001725#ifdef CONFIG_NUMA
1726
1727/*
1728 * node_hstate/s - associate per node hstate attributes, via their kobjects,
Kay Sievers10fbcf42011-12-21 14:48:43 -08001729 * with node devices in node_devices[] using a parallel array. The array
1730 * index of a node device or _hstate == node id.
1731 * This is here to avoid any static dependency of the node device driver, in
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001732 * the base kernel, on the hugetlb module.
1733 */
1734struct node_hstate {
1735 struct kobject *hugepages_kobj;
1736 struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
1737};
1738struct node_hstate node_hstates[MAX_NUMNODES];
1739
1740/*
Kay Sievers10fbcf42011-12-21 14:48:43 -08001741 * A subset of global hstate attributes for node devices
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001742 */
1743static struct attribute *per_node_hstate_attrs[] = {
1744 &nr_hugepages_attr.attr,
1745 &free_hugepages_attr.attr,
1746 &surplus_hugepages_attr.attr,
1747 NULL,
1748};
1749
1750static struct attribute_group per_node_hstate_attr_group = {
1751 .attrs = per_node_hstate_attrs,
1752};
1753
1754/*
Kay Sievers10fbcf42011-12-21 14:48:43 -08001755 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001756 * Returns node id via non-NULL nidp.
1757 */
1758static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
1759{
1760 int nid;
1761
1762 for (nid = 0; nid < nr_node_ids; nid++) {
1763 struct node_hstate *nhs = &node_hstates[nid];
1764 int i;
1765 for (i = 0; i < HUGE_MAX_HSTATE; i++)
1766 if (nhs->hstate_kobjs[i] == kobj) {
1767 if (nidp)
1768 *nidp = nid;
1769 return &hstates[i];
1770 }
1771 }
1772
1773 BUG();
1774 return NULL;
1775}
1776
1777/*
Kay Sievers10fbcf42011-12-21 14:48:43 -08001778 * Unregister hstate attributes from a single node device.
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001779 * No-op if no hstate attributes attached.
1780 */
Claudiu Ghioc3cd8b442013-03-04 12:46:15 +02001781static void hugetlb_unregister_node(struct node *node)
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001782{
1783 struct hstate *h;
Kay Sievers10fbcf42011-12-21 14:48:43 -08001784 struct node_hstate *nhs = &node_hstates[node->dev.id];
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001785
1786 if (!nhs->hugepages_kobj)
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -08001787 return; /* no hstate attributes */
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001788
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07001789 for_each_hstate(h) {
1790 int idx = hstate_index(h);
1791 if (nhs->hstate_kobjs[idx]) {
1792 kobject_put(nhs->hstate_kobjs[idx]);
1793 nhs->hstate_kobjs[idx] = NULL;
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001794 }
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07001795 }
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001796
1797 kobject_put(nhs->hugepages_kobj);
1798 nhs->hugepages_kobj = NULL;
1799}
1800
1801/*
Kay Sievers10fbcf42011-12-21 14:48:43 -08001802 * hugetlb module exit: unregister hstate attributes from node devices
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001803 * that have them.
1804 */
1805static void hugetlb_unregister_all_nodes(void)
1806{
1807 int nid;
1808
1809 /*
Kay Sievers10fbcf42011-12-21 14:48:43 -08001810 * disable node device registrations.
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001811 */
1812 register_hugetlbfs_with_node(NULL, NULL);
1813
1814 /*
1815 * remove hstate attributes from any nodes that have them.
1816 */
1817 for (nid = 0; nid < nr_node_ids; nid++)
Wen Congyang87327942012-12-11 16:00:56 -08001818 hugetlb_unregister_node(node_devices[nid]);
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001819}
1820
1821/*
Kay Sievers10fbcf42011-12-21 14:48:43 -08001822 * Register hstate attributes for a single node device.
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001823 * No-op if attributes already registered.
1824 */
Claudiu Ghioc3cd8b442013-03-04 12:46:15 +02001825static void hugetlb_register_node(struct node *node)
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001826{
1827 struct hstate *h;
Kay Sievers10fbcf42011-12-21 14:48:43 -08001828 struct node_hstate *nhs = &node_hstates[node->dev.id];
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001829 int err;
1830
1831 if (nhs->hugepages_kobj)
1832 return; /* already allocated */
1833
1834 nhs->hugepages_kobj = kobject_create_and_add("hugepages",
Kay Sievers10fbcf42011-12-21 14:48:43 -08001835 &node->dev.kobj);
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001836 if (!nhs->hugepages_kobj)
1837 return;
1838
1839 for_each_hstate(h) {
1840 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
1841 nhs->hstate_kobjs,
1842 &per_node_hstate_attr_group);
1843 if (err) {
Andrew Mortonffb22af2013-02-22 16:32:08 -08001844 pr_err("Hugetlb: Unable to add hstate %s for node %d\n",
1845 h->name, node->dev.id);
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001846 hugetlb_unregister_node(node);
1847 break;
1848 }
1849 }
1850}
1851
1852/*
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -08001853 * hugetlb init time: register hstate attributes for all registered node
Kay Sievers10fbcf42011-12-21 14:48:43 -08001854 * devices of nodes that have memory. All on-line nodes should have
1855 * registered their associated device by this time.
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001856 */
1857static void hugetlb_register_all_nodes(void)
1858{
1859 int nid;
1860
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08001861 for_each_node_state(nid, N_MEMORY) {
Wen Congyang87327942012-12-11 16:00:56 -08001862 struct node *node = node_devices[nid];
Kay Sievers10fbcf42011-12-21 14:48:43 -08001863 if (node->dev.id == nid)
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001864 hugetlb_register_node(node);
1865 }
1866
1867 /*
Kay Sievers10fbcf42011-12-21 14:48:43 -08001868 * Let the node device driver know we're here so it can
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001869 * [un]register hstate attributes on node hotplug.
1870 */
1871 register_hugetlbfs_with_node(hugetlb_register_node,
1872 hugetlb_unregister_node);
1873}
1874#else /* !CONFIG_NUMA */
1875
1876static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
1877{
1878 BUG();
1879 if (nidp)
1880 *nidp = -1;
1881 return NULL;
1882}
1883
1884static void hugetlb_unregister_all_nodes(void) { }
1885
1886static void hugetlb_register_all_nodes(void) { }
1887
1888#endif
1889
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001890static void __exit hugetlb_exit(void)
1891{
1892 struct hstate *h;
1893
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001894 hugetlb_unregister_all_nodes();
1895
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001896 for_each_hstate(h) {
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07001897 kobject_put(hstate_kobjs[hstate_index(h)]);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001898 }
1899
1900 kobject_put(hugepages_kobj);
1901}
1902module_exit(hugetlb_exit);
1903
1904static int __init hugetlb_init(void)
1905{
Benjamin Herrenschmidt0ef89d22008-07-31 00:07:30 -07001906 /* Some platform decide whether they support huge pages at boot
1907 * time. On these, such as powerpc, HPAGE_SHIFT is set to 0 when
1908 * there is no such support
1909 */
1910 if (HPAGE_SHIFT == 0)
1911 return 0;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001912
Nick Piggine11bfbf2008-07-23 21:27:52 -07001913 if (!size_to_hstate(default_hstate_size)) {
1914 default_hstate_size = HPAGE_SIZE;
1915 if (!size_to_hstate(default_hstate_size))
1916 hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001917 }
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07001918 default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size));
Nick Piggine11bfbf2008-07-23 21:27:52 -07001919 if (default_hstate_max_huge_pages)
1920 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001921
1922 hugetlb_init_hstates();
Andi Kleenaa888a72008-07-23 21:27:47 -07001923 gather_bootmem_prealloc();
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001924 report_hugepages();
1925
1926 hugetlb_sysfs_init();
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001927 hugetlb_register_all_nodes();
Jianguo Wu7179e7b2012-12-18 14:23:19 -08001928 hugetlb_cgroup_file_init();
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001929
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001930 return 0;
1931}
1932module_init(hugetlb_init);
1933
1934/* Should be called on processing a hugepagesz=... option */
1935void __init hugetlb_add_hstate(unsigned order)
1936{
1937 struct hstate *h;
Andi Kleen8faa8b02008-07-23 21:27:48 -07001938 unsigned long i;
1939
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001940 if (size_to_hstate(PAGE_SIZE << order)) {
Andrew Mortonffb22af2013-02-22 16:32:08 -08001941 pr_warning("hugepagesz= specified twice, ignoring\n");
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001942 return;
1943 }
Aneesh Kumar K.V47d38342012-07-31 16:41:54 -07001944 BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001945 BUG_ON(order == 0);
Aneesh Kumar K.V47d38342012-07-31 16:41:54 -07001946 h = &hstates[hugetlb_max_hstate++];
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001947 h->order = order;
1948 h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
Andi Kleen8faa8b02008-07-23 21:27:48 -07001949 h->nr_huge_pages = 0;
1950 h->free_huge_pages = 0;
1951 for (i = 0; i < MAX_NUMNODES; ++i)
1952 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
Aneesh Kumar K.V0edaecf2012-07-31 16:42:07 -07001953 INIT_LIST_HEAD(&h->hugepage_activelist);
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08001954 h->next_nid_to_alloc = first_node(node_states[N_MEMORY]);
1955 h->next_nid_to_free = first_node(node_states[N_MEMORY]);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001956 snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
1957 huge_page_size(h)/1024);
Andi Kleen8faa8b02008-07-23 21:27:48 -07001958
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001959 parsed_hstate = h;
1960}
1961
Nick Piggine11bfbf2008-07-23 21:27:52 -07001962static int __init hugetlb_nrpages_setup(char *s)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001963{
1964 unsigned long *mhp;
Andi Kleen8faa8b02008-07-23 21:27:48 -07001965 static unsigned long *last_mhp;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001966
1967 /*
Aneesh Kumar K.V47d38342012-07-31 16:41:54 -07001968 * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet,
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001969 * so this hugepages= parameter goes to the "default hstate".
1970 */
Aneesh Kumar K.V47d38342012-07-31 16:41:54 -07001971 if (!hugetlb_max_hstate)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001972 mhp = &default_hstate_max_huge_pages;
1973 else
1974 mhp = &parsed_hstate->max_huge_pages;
1975
Andi Kleen8faa8b02008-07-23 21:27:48 -07001976 if (mhp == last_mhp) {
Andrew Mortonffb22af2013-02-22 16:32:08 -08001977 pr_warning("hugepages= specified twice without "
1978 "interleaving hugepagesz=, ignoring\n");
Andi Kleen8faa8b02008-07-23 21:27:48 -07001979 return 1;
1980 }
1981
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001982 if (sscanf(s, "%lu", mhp) <= 0)
1983 *mhp = 0;
1984
Andi Kleen8faa8b02008-07-23 21:27:48 -07001985 /*
1986 * Global state is always initialized later in hugetlb_init.
1987 * But we need to allocate >= MAX_ORDER hstates here early to still
1988 * use the bootmem allocator.
1989 */
Aneesh Kumar K.V47d38342012-07-31 16:41:54 -07001990 if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER)
Andi Kleen8faa8b02008-07-23 21:27:48 -07001991 hugetlb_hstate_alloc_pages(parsed_hstate);
1992
1993 last_mhp = mhp;
1994
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001995 return 1;
1996}
Nick Piggine11bfbf2008-07-23 21:27:52 -07001997__setup("hugepages=", hugetlb_nrpages_setup);
1998
1999static int __init hugetlb_default_setup(char *s)
2000{
2001 default_hstate_size = memparse(s, &s);
2002 return 1;
2003}
2004__setup("default_hugepagesz=", hugetlb_default_setup);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002005
Nishanth Aravamudan8a213462008-07-25 19:44:37 -07002006static unsigned int cpuset_mems_nr(unsigned int *array)
2007{
2008 int node;
2009 unsigned int nr = 0;
2010
2011 for_each_node_mask(node, cpuset_current_mems_allowed)
2012 nr += array[node];
2013
2014 return nr;
2015}
2016
2017#ifdef CONFIG_SYSCTL
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002018static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
2019 struct ctl_table *table, int write,
2020 void __user *buffer, size_t *length, loff_t *ppos)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002021{
Andi Kleene5ff2152008-07-23 21:27:42 -07002022 struct hstate *h = &default_hstate;
2023 unsigned long tmp;
Michal Hocko08d4a242011-01-13 15:47:26 -08002024 int ret;
Andi Kleene5ff2152008-07-23 21:27:42 -07002025
Petr Holasekc033a932011-03-22 16:33:05 -07002026 tmp = h->max_huge_pages;
Andi Kleene5ff2152008-07-23 21:27:42 -07002027
Eric B Munsonadbe8722011-01-13 15:47:27 -08002028 if (write && h->order >= MAX_ORDER)
2029 return -EINVAL;
2030
Andi Kleene5ff2152008-07-23 21:27:42 -07002031 table->data = &tmp;
2032 table->maxlen = sizeof(unsigned long);
Michal Hocko08d4a242011-01-13 15:47:26 -08002033 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2034 if (ret)
2035 goto out;
Andi Kleene5ff2152008-07-23 21:27:42 -07002036
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002037 if (write) {
David Rientjesbad44b52009-12-14 17:58:38 -08002038 NODEMASK_ALLOC(nodemask_t, nodes_allowed,
2039 GFP_KERNEL | __GFP_NORETRY);
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002040 if (!(obey_mempolicy &&
2041 init_nodemask_of_mempolicy(nodes_allowed))) {
2042 NODEMASK_FREE(nodes_allowed);
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08002043 nodes_allowed = &node_states[N_MEMORY];
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002044 }
2045 h->max_huge_pages = set_max_huge_pages(h, tmp, nodes_allowed);
2046
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08002047 if (nodes_allowed != &node_states[N_MEMORY])
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002048 NODEMASK_FREE(nodes_allowed);
2049 }
Michal Hocko08d4a242011-01-13 15:47:26 -08002050out:
2051 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002052}
Mel Gorman396faf02007-07-17 04:03:13 -07002053
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002054int hugetlb_sysctl_handler(struct ctl_table *table, int write,
2055 void __user *buffer, size_t *length, loff_t *ppos)
2056{
2057
2058 return hugetlb_sysctl_handler_common(false, table, write,
2059 buffer, length, ppos);
2060}
2061
2062#ifdef CONFIG_NUMA
2063int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
2064 void __user *buffer, size_t *length, loff_t *ppos)
2065{
2066 return hugetlb_sysctl_handler_common(true, table, write,
2067 buffer, length, ppos);
2068}
2069#endif /* CONFIG_NUMA */
2070
Mel Gorman396faf02007-07-17 04:03:13 -07002071int hugetlb_treat_movable_handler(struct ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07002072 void __user *buffer,
Mel Gorman396faf02007-07-17 04:03:13 -07002073 size_t *length, loff_t *ppos)
2074{
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07002075 proc_dointvec(table, write, buffer, length, ppos);
Mel Gorman396faf02007-07-17 04:03:13 -07002076 if (hugepages_treat_as_movable)
2077 htlb_alloc_mask = GFP_HIGHUSER_MOVABLE;
2078 else
2079 htlb_alloc_mask = GFP_HIGHUSER;
2080 return 0;
2081}
2082
Nishanth Aravamudana3d0c6a2008-02-08 04:18:18 -08002083int hugetlb_overcommit_handler(struct ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07002084 void __user *buffer,
Nishanth Aravamudana3d0c6a2008-02-08 04:18:18 -08002085 size_t *length, loff_t *ppos)
2086{
Andi Kleena5516432008-07-23 21:27:41 -07002087 struct hstate *h = &default_hstate;
Andi Kleene5ff2152008-07-23 21:27:42 -07002088 unsigned long tmp;
Michal Hocko08d4a242011-01-13 15:47:26 -08002089 int ret;
Andi Kleene5ff2152008-07-23 21:27:42 -07002090
Petr Holasekc033a932011-03-22 16:33:05 -07002091 tmp = h->nr_overcommit_huge_pages;
Andi Kleene5ff2152008-07-23 21:27:42 -07002092
Eric B Munsonadbe8722011-01-13 15:47:27 -08002093 if (write && h->order >= MAX_ORDER)
2094 return -EINVAL;
2095
Andi Kleene5ff2152008-07-23 21:27:42 -07002096 table->data = &tmp;
2097 table->maxlen = sizeof(unsigned long);
Michal Hocko08d4a242011-01-13 15:47:26 -08002098 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2099 if (ret)
2100 goto out;
Andi Kleene5ff2152008-07-23 21:27:42 -07002101
2102 if (write) {
2103 spin_lock(&hugetlb_lock);
2104 h->nr_overcommit_huge_pages = tmp;
2105 spin_unlock(&hugetlb_lock);
2106 }
Michal Hocko08d4a242011-01-13 15:47:26 -08002107out:
2108 return ret;
Nishanth Aravamudana3d0c6a2008-02-08 04:18:18 -08002109}
2110
Linus Torvalds1da177e2005-04-16 15:20:36 -07002111#endif /* CONFIG_SYSCTL */
2112
Alexey Dobriyane1759c22008-10-15 23:50:22 +04002113void hugetlb_report_meminfo(struct seq_file *m)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002114{
Andi Kleena5516432008-07-23 21:27:41 -07002115 struct hstate *h = &default_hstate;
Alexey Dobriyane1759c22008-10-15 23:50:22 +04002116 seq_printf(m,
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002117 "HugePages_Total: %5lu\n"
2118 "HugePages_Free: %5lu\n"
2119 "HugePages_Rsvd: %5lu\n"
2120 "HugePages_Surp: %5lu\n"
2121 "Hugepagesize: %8lu kB\n",
Andi Kleena5516432008-07-23 21:27:41 -07002122 h->nr_huge_pages,
2123 h->free_huge_pages,
2124 h->resv_huge_pages,
2125 h->surplus_huge_pages,
2126 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002127}
2128
2129int hugetlb_report_node_meminfo(int nid, char *buf)
2130{
Andi Kleena5516432008-07-23 21:27:41 -07002131 struct hstate *h = &default_hstate;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002132 return sprintf(buf,
2133 "Node %d HugePages_Total: %5u\n"
Nishanth Aravamudana1de0912008-03-26 14:37:53 -07002134 "Node %d HugePages_Free: %5u\n"
2135 "Node %d HugePages_Surp: %5u\n",
Andi Kleena5516432008-07-23 21:27:41 -07002136 nid, h->nr_huge_pages_node[nid],
2137 nid, h->free_huge_pages_node[nid],
2138 nid, h->surplus_huge_pages_node[nid]);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002139}
2140
David Rientjes949f7ec2013-04-29 15:07:48 -07002141void hugetlb_show_meminfo(void)
2142{
2143 struct hstate *h;
2144 int nid;
2145
2146 for_each_node_state(nid, N_MEMORY)
2147 for_each_hstate(h)
2148 pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
2149 nid,
2150 h->nr_huge_pages_node[nid],
2151 h->free_huge_pages_node[nid],
2152 h->surplus_huge_pages_node[nid],
2153 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
2154}
2155
Linus Torvalds1da177e2005-04-16 15:20:36 -07002156/* Return the number pages of memory we physically have, in PAGE_SIZE units. */
2157unsigned long hugetlb_total_pages(void)
2158{
Wanpeng Lid0028582013-03-22 15:04:40 -07002159 struct hstate *h;
2160 unsigned long nr_total_pages = 0;
2161
2162 for_each_hstate(h)
2163 nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h);
2164 return nr_total_pages;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002165}
Linus Torvalds1da177e2005-04-16 15:20:36 -07002166
Andi Kleena5516432008-07-23 21:27:41 -07002167static int hugetlb_acct_memory(struct hstate *h, long delta)
Mel Gormanfc1b8a72008-07-23 21:27:22 -07002168{
2169 int ret = -ENOMEM;
2170
2171 spin_lock(&hugetlb_lock);
2172 /*
2173 * When cpuset is configured, it breaks the strict hugetlb page
2174 * reservation as the accounting is done on a global variable. Such
2175 * reservation is completely rubbish in the presence of cpuset because
2176 * the reservation is not checked against page availability for the
2177 * current cpuset. Application can still potentially OOM'ed by kernel
2178 * with lack of free htlb page in cpuset that the task is in.
2179 * Attempt to enforce strict accounting with cpuset is almost
2180 * impossible (or too ugly) because cpuset is too fluid that
2181 * task or memory node can be dynamically moved between cpusets.
2182 *
2183 * The change of semantics for shared hugetlb mapping with cpuset is
2184 * undesirable. However, in order to preserve some of the semantics,
2185 * we fall back to check against current free page availability as
2186 * a best attempt and hopefully to minimize the impact of changing
2187 * semantics that cpuset has.
2188 */
2189 if (delta > 0) {
Andi Kleena5516432008-07-23 21:27:41 -07002190 if (gather_surplus_pages(h, delta) < 0)
Mel Gormanfc1b8a72008-07-23 21:27:22 -07002191 goto out;
2192
Andi Kleena5516432008-07-23 21:27:41 -07002193 if (delta > cpuset_mems_nr(h->free_huge_pages_node)) {
2194 return_unused_surplus_pages(h, delta);
Mel Gormanfc1b8a72008-07-23 21:27:22 -07002195 goto out;
2196 }
2197 }
2198
2199 ret = 0;
2200 if (delta < 0)
Andi Kleena5516432008-07-23 21:27:41 -07002201 return_unused_surplus_pages(h, (unsigned long) -delta);
Mel Gormanfc1b8a72008-07-23 21:27:22 -07002202
2203out:
2204 spin_unlock(&hugetlb_lock);
2205 return ret;
2206}
2207
Andy Whitcroft84afd992008-07-23 21:27:32 -07002208static void hugetlb_vm_op_open(struct vm_area_struct *vma)
2209{
2210 struct resv_map *reservations = vma_resv_map(vma);
2211
2212 /*
2213 * This new VMA should share its siblings reservation map if present.
2214 * The VMA will only ever have a valid reservation map pointer where
2215 * it is being copied for another still existing VMA. As that VMA
Lucas De Marchi25985ed2011-03-30 22:57:33 -03002216 * has a reference to the reservation map it cannot disappear until
Andy Whitcroft84afd992008-07-23 21:27:32 -07002217 * after this open call completes. It is therefore safe to take a
2218 * new reference here without additional locking.
2219 */
2220 if (reservations)
2221 kref_get(&reservations->refs);
2222}
2223
Dave Hansenc50ac052012-05-29 15:06:46 -07002224static void resv_map_put(struct vm_area_struct *vma)
2225{
2226 struct resv_map *reservations = vma_resv_map(vma);
2227
2228 if (!reservations)
2229 return;
2230 kref_put(&reservations->refs, resv_map_release);
2231}
2232
Mel Gormana1e78772008-07-23 21:27:23 -07002233static void hugetlb_vm_op_close(struct vm_area_struct *vma)
2234{
Andi Kleena5516432008-07-23 21:27:41 -07002235 struct hstate *h = hstate_vma(vma);
Andy Whitcroft84afd992008-07-23 21:27:32 -07002236 struct resv_map *reservations = vma_resv_map(vma);
David Gibson90481622012-03-21 16:34:12 -07002237 struct hugepage_subpool *spool = subpool_vma(vma);
Andy Whitcroft84afd992008-07-23 21:27:32 -07002238 unsigned long reserve;
2239 unsigned long start;
2240 unsigned long end;
2241
2242 if (reservations) {
Andi Kleena5516432008-07-23 21:27:41 -07002243 start = vma_hugecache_offset(h, vma, vma->vm_start);
2244 end = vma_hugecache_offset(h, vma, vma->vm_end);
Andy Whitcroft84afd992008-07-23 21:27:32 -07002245
2246 reserve = (end - start) -
2247 region_count(&reservations->regions, start, end);
2248
Dave Hansenc50ac052012-05-29 15:06:46 -07002249 resv_map_put(vma);
Andy Whitcroft84afd992008-07-23 21:27:32 -07002250
Adam Litke7251ff72008-07-23 21:27:59 -07002251 if (reserve) {
Andi Kleena5516432008-07-23 21:27:41 -07002252 hugetlb_acct_memory(h, -reserve);
David Gibson90481622012-03-21 16:34:12 -07002253 hugepage_subpool_put_pages(spool, reserve);
Adam Litke7251ff72008-07-23 21:27:59 -07002254 }
Andy Whitcroft84afd992008-07-23 21:27:32 -07002255 }
Mel Gormana1e78772008-07-23 21:27:23 -07002256}
2257
Linus Torvalds1da177e2005-04-16 15:20:36 -07002258/*
2259 * We cannot handle pagefaults against hugetlb pages at all. They cause
2260 * handle_mm_fault() to try to instantiate regular-sized pages in the
2261 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
2262 * this far.
2263 */
Nick Piggind0217ac2007-07-19 01:47:03 -07002264static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002265{
2266 BUG();
Nick Piggind0217ac2007-07-19 01:47:03 -07002267 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002268}
2269
Alexey Dobriyanf0f37e22009-09-27 22:29:37 +04002270const struct vm_operations_struct hugetlb_vm_ops = {
Nick Piggind0217ac2007-07-19 01:47:03 -07002271 .fault = hugetlb_vm_op_fault,
Andy Whitcroft84afd992008-07-23 21:27:32 -07002272 .open = hugetlb_vm_op_open,
Mel Gormana1e78772008-07-23 21:27:23 -07002273 .close = hugetlb_vm_op_close,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002274};
2275
David Gibson1e8f8892006-01-06 00:10:44 -08002276static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
2277 int writable)
David Gibson63551ae2005-06-21 17:14:44 -07002278{
2279 pte_t entry;
2280
David Gibson1e8f8892006-01-06 00:10:44 -08002281 if (writable) {
Gerald Schaefer106c9922013-04-29 15:07:23 -07002282 entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page,
2283 vma->vm_page_prot)));
David Gibson63551ae2005-06-21 17:14:44 -07002284 } else {
Gerald Schaefer106c9922013-04-29 15:07:23 -07002285 entry = huge_pte_wrprotect(mk_huge_pte(page,
2286 vma->vm_page_prot));
David Gibson63551ae2005-06-21 17:14:44 -07002287 }
2288 entry = pte_mkyoung(entry);
2289 entry = pte_mkhuge(entry);
Chris Metcalfd9ed9fa2012-04-01 14:01:34 -04002290 entry = arch_make_huge_pte(entry, vma, page, writable);
David Gibson63551ae2005-06-21 17:14:44 -07002291
2292 return entry;
2293}
2294
David Gibson1e8f8892006-01-06 00:10:44 -08002295static void set_huge_ptep_writable(struct vm_area_struct *vma,
2296 unsigned long address, pte_t *ptep)
2297{
2298 pte_t entry;
2299
Gerald Schaefer106c9922013-04-29 15:07:23 -07002300 entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep)));
Chris Forbes32f84522011-07-25 17:12:14 -07002301 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
Russell King4b3073e2009-12-18 16:40:18 +00002302 update_mmu_cache(vma, address, ptep);
David Gibson1e8f8892006-01-06 00:10:44 -08002303}
2304
2305
David Gibson63551ae2005-06-21 17:14:44 -07002306int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
2307 struct vm_area_struct *vma)
2308{
2309 pte_t *src_pte, *dst_pte, entry;
2310 struct page *ptepage;
Hugh Dickins1c598272005-10-19 21:23:43 -07002311 unsigned long addr;
David Gibson1e8f8892006-01-06 00:10:44 -08002312 int cow;
Andi Kleena5516432008-07-23 21:27:41 -07002313 struct hstate *h = hstate_vma(vma);
2314 unsigned long sz = huge_page_size(h);
David Gibson1e8f8892006-01-06 00:10:44 -08002315
2316 cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
David Gibson63551ae2005-06-21 17:14:44 -07002317
Andi Kleena5516432008-07-23 21:27:41 -07002318 for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
Hugh Dickinsc74df322005-10-29 18:16:23 -07002319 src_pte = huge_pte_offset(src, addr);
2320 if (!src_pte)
2321 continue;
Andi Kleena5516432008-07-23 21:27:41 -07002322 dst_pte = huge_pte_alloc(dst, addr, sz);
David Gibson63551ae2005-06-21 17:14:44 -07002323 if (!dst_pte)
2324 goto nomem;
Larry Woodmanc5c99422008-01-24 05:49:25 -08002325
2326 /* If the pagetables are shared don't copy or take references */
2327 if (dst_pte == src_pte)
2328 continue;
2329
Hugh Dickinsc74df322005-10-29 18:16:23 -07002330 spin_lock(&dst->page_table_lock);
Nick Piggin46478752008-06-05 22:45:57 -07002331 spin_lock_nested(&src->page_table_lock, SINGLE_DEPTH_NESTING);
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07002332 if (!huge_pte_none(huge_ptep_get(src_pte))) {
David Gibson1e8f8892006-01-06 00:10:44 -08002333 if (cow)
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07002334 huge_ptep_set_wrprotect(src, addr, src_pte);
2335 entry = huge_ptep_get(src_pte);
Hugh Dickins1c598272005-10-19 21:23:43 -07002336 ptepage = pte_page(entry);
2337 get_page(ptepage);
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002338 page_dup_rmap(ptepage);
Hugh Dickins1c598272005-10-19 21:23:43 -07002339 set_huge_pte_at(dst, addr, dst_pte, entry);
2340 }
2341 spin_unlock(&src->page_table_lock);
Hugh Dickinsc74df322005-10-29 18:16:23 -07002342 spin_unlock(&dst->page_table_lock);
David Gibson63551ae2005-06-21 17:14:44 -07002343 }
2344 return 0;
2345
2346nomem:
2347 return -ENOMEM;
2348}
2349
Naoya Horiguchi290408d2010-09-08 10:19:35 +09002350static int is_hugetlb_entry_migration(pte_t pte)
2351{
2352 swp_entry_t swp;
2353
2354 if (huge_pte_none(pte) || pte_present(pte))
2355 return 0;
2356 swp = pte_to_swp_entry(pte);
Chris Forbes32f84522011-07-25 17:12:14 -07002357 if (non_swap_entry(swp) && is_migration_entry(swp))
Naoya Horiguchi290408d2010-09-08 10:19:35 +09002358 return 1;
Chris Forbes32f84522011-07-25 17:12:14 -07002359 else
Naoya Horiguchi290408d2010-09-08 10:19:35 +09002360 return 0;
2361}
2362
Naoya Horiguchifd6a03ed2010-05-28 09:29:21 +09002363static int is_hugetlb_entry_hwpoisoned(pte_t pte)
2364{
2365 swp_entry_t swp;
2366
2367 if (huge_pte_none(pte) || pte_present(pte))
2368 return 0;
2369 swp = pte_to_swp_entry(pte);
Chris Forbes32f84522011-07-25 17:12:14 -07002370 if (non_swap_entry(swp) && is_hwpoison_entry(swp))
Naoya Horiguchifd6a03ed2010-05-28 09:29:21 +09002371 return 1;
Chris Forbes32f84522011-07-25 17:12:14 -07002372 else
Naoya Horiguchifd6a03ed2010-05-28 09:29:21 +09002373 return 0;
2374}
2375
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07002376void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
2377 unsigned long start, unsigned long end,
2378 struct page *ref_page)
David Gibson63551ae2005-06-21 17:14:44 -07002379{
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07002380 int force_flush = 0;
David Gibson63551ae2005-06-21 17:14:44 -07002381 struct mm_struct *mm = vma->vm_mm;
2382 unsigned long address;
David Gibsonc7546f82005-08-05 11:59:35 -07002383 pte_t *ptep;
David Gibson63551ae2005-06-21 17:14:44 -07002384 pte_t pte;
2385 struct page *page;
Andi Kleena5516432008-07-23 21:27:41 -07002386 struct hstate *h = hstate_vma(vma);
2387 unsigned long sz = huge_page_size(h);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07002388 const unsigned long mmun_start = start; /* For mmu_notifiers */
2389 const unsigned long mmun_end = end; /* For mmu_notifiers */
Andi Kleena5516432008-07-23 21:27:41 -07002390
David Gibson63551ae2005-06-21 17:14:44 -07002391 WARN_ON(!is_vm_hugetlb_page(vma));
Andi Kleena5516432008-07-23 21:27:41 -07002392 BUG_ON(start & ~huge_page_mask(h));
2393 BUG_ON(end & ~huge_page_mask(h));
David Gibson63551ae2005-06-21 17:14:44 -07002394
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07002395 tlb_start_vma(tlb, vma);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07002396 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07002397again:
Hugh Dickins508034a2005-10-29 18:16:30 -07002398 spin_lock(&mm->page_table_lock);
Andi Kleena5516432008-07-23 21:27:41 -07002399 for (address = start; address < end; address += sz) {
David Gibsonc7546f82005-08-05 11:59:35 -07002400 ptep = huge_pte_offset(mm, address);
Adam Litke4c887262005-10-29 18:16:46 -07002401 if (!ptep)
David Gibsonc7546f82005-08-05 11:59:35 -07002402 continue;
2403
Chen, Kenneth W39dde652006-12-06 20:32:03 -08002404 if (huge_pmd_unshare(mm, &address, ptep))
2405 continue;
2406
Hillf Danton66293262012-03-23 15:01:48 -07002407 pte = huge_ptep_get(ptep);
2408 if (huge_pte_none(pte))
2409 continue;
2410
2411 /*
2412 * HWPoisoned hugepage is already unmapped and dropped reference
2413 */
Naoya Horiguchi8c4894c2012-12-12 13:52:28 -08002414 if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) {
Gerald Schaefer106c9922013-04-29 15:07:23 -07002415 huge_pte_clear(mm, address, ptep);
Hillf Danton66293262012-03-23 15:01:48 -07002416 continue;
Naoya Horiguchi8c4894c2012-12-12 13:52:28 -08002417 }
Hillf Danton66293262012-03-23 15:01:48 -07002418
2419 page = pte_page(pte);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002420 /*
2421 * If a reference page is supplied, it is because a specific
2422 * page is being unmapped, not a range. Ensure the page we
2423 * are about to unmap is the actual page of interest.
2424 */
2425 if (ref_page) {
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002426 if (page != ref_page)
2427 continue;
2428
2429 /*
2430 * Mark the VMA as having unmapped its page so that
2431 * future faults in this VMA will fail rather than
2432 * looking like data was lost
2433 */
2434 set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
2435 }
2436
David Gibsonc7546f82005-08-05 11:59:35 -07002437 pte = huge_ptep_get_and_clear(mm, address, ptep);
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07002438 tlb_remove_tlb_entry(tlb, ptep, address);
Gerald Schaefer106c9922013-04-29 15:07:23 -07002439 if (huge_pte_dirty(pte))
Ken Chen6649a382007-02-08 14:20:27 -08002440 set_page_dirty(page);
Hillf Danton9e811302012-03-21 16:34:03 -07002441
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07002442 page_remove_rmap(page);
2443 force_flush = !__tlb_remove_page(tlb, page);
2444 if (force_flush)
2445 break;
Hillf Danton9e811302012-03-21 16:34:03 -07002446 /* Bail out after unmapping reference page if supplied */
2447 if (ref_page)
2448 break;
David Gibson63551ae2005-06-21 17:14:44 -07002449 }
Al Virocd2934a2012-03-05 06:40:29 +00002450 spin_unlock(&mm->page_table_lock);
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07002451 /*
2452 * mmu_gather ran out of room to batch pages, we break out of
2453 * the PTE lock to avoid doing the potential expensive TLB invalidate
2454 * and page-free while holding it.
2455 */
2456 if (force_flush) {
2457 force_flush = 0;
2458 tlb_flush_mmu(tlb);
2459 if (address < end && !ref_page)
2460 goto again;
Chen, Kenneth Wfe1668a2006-10-04 02:15:24 -07002461 }
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07002462 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07002463 tlb_end_vma(tlb, vma);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002464}
David Gibson63551ae2005-06-21 17:14:44 -07002465
Mel Gormand8333522012-07-31 16:46:20 -07002466void __unmap_hugepage_range_final(struct mmu_gather *tlb,
2467 struct vm_area_struct *vma, unsigned long start,
2468 unsigned long end, struct page *ref_page)
2469{
2470 __unmap_hugepage_range(tlb, vma, start, end, ref_page);
2471
2472 /*
2473 * Clear this flag so that x86's huge_pmd_share page_table_shareable
2474 * test will fail on a vma being torn down, and not grab a page table
2475 * on its way out. We're lucky that the flag has such an appropriate
2476 * name, and can in fact be safely cleared here. We could clear it
2477 * before the __unmap_hugepage_range above, but all that's necessary
2478 * is to clear it before releasing the i_mmap_mutex. This works
2479 * because in the context this is called, the VMA is about to be
2480 * destroyed and the i_mmap_mutex is held.
2481 */
2482 vma->vm_flags &= ~VM_MAYSHARE;
2483}
2484
Chen, Kenneth W502717f2006-10-11 01:20:46 -07002485void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002486 unsigned long end, struct page *ref_page)
Chen, Kenneth W502717f2006-10-11 01:20:46 -07002487{
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07002488 struct mm_struct *mm;
2489 struct mmu_gather tlb;
2490
2491 mm = vma->vm_mm;
2492
2493 tlb_gather_mmu(&tlb, mm, 0);
2494 __unmap_hugepage_range(&tlb, vma, start, end, ref_page);
2495 tlb_finish_mmu(&tlb, start, end);
Chen, Kenneth W502717f2006-10-11 01:20:46 -07002496}
2497
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002498/*
2499 * This is called when the original mapper is failing to COW a MAP_PRIVATE
2500 * mappping it owns the reserve page for. The intention is to unmap the page
2501 * from other VMAs and let the children be SIGKILLed if they are faulting the
2502 * same region.
2503 */
Harvey Harrison2a4b3de2008-10-18 20:27:06 -07002504static int unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
2505 struct page *page, unsigned long address)
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002506{
Adam Litke75266742008-11-12 13:24:56 -08002507 struct hstate *h = hstate_vma(vma);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002508 struct vm_area_struct *iter_vma;
2509 struct address_space *mapping;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002510 pgoff_t pgoff;
2511
2512 /*
2513 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
2514 * from page cache lookup which is in HPAGE_SIZE units.
2515 */
Adam Litke75266742008-11-12 13:24:56 -08002516 address = address & huge_page_mask(h);
Michal Hocko36e4f202012-10-08 16:33:31 -07002517 pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) +
2518 vma->vm_pgoff;
Al Viro496ad9a2013-01-23 17:07:38 -05002519 mapping = file_inode(vma->vm_file)->i_mapping;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002520
Mel Gorman4eb2b1d2009-12-14 17:59:53 -08002521 /*
2522 * Take the mapping lock for the duration of the table walk. As
2523 * this mapping should be shared between all the VMAs,
2524 * __unmap_hugepage_range() is called as the lock is already held
2525 */
Peter Zijlstra3d48ae42011-05-24 17:12:06 -07002526 mutex_lock(&mapping->i_mmap_mutex);
Michel Lespinasse6b2dbba2012-10-08 16:31:25 -07002527 vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) {
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002528 /* Do not unmap the current VMA */
2529 if (iter_vma == vma)
2530 continue;
2531
2532 /*
2533 * Unmap the page from other VMAs without their own reserves.
2534 * They get marked to be SIGKILLed if they fault in these
2535 * areas. This is because a future no-page fault on this VMA
2536 * could insert a zeroed page instead of the data existing
2537 * from the time of fork. This would look like data corruption
2538 */
2539 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07002540 unmap_hugepage_range(iter_vma, address,
2541 address + huge_page_size(h), page);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002542 }
Peter Zijlstra3d48ae42011-05-24 17:12:06 -07002543 mutex_unlock(&mapping->i_mmap_mutex);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002544
2545 return 1;
2546}
2547
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002548/*
2549 * Hugetlb_cow() should be called with page lock of the original hugepage held.
Michal Hockoef009b22012-01-10 15:07:21 -08002550 * Called with hugetlb_instantiation_mutex held and pte_page locked so we
2551 * cannot race with other handlers or page migration.
2552 * Keep the pte_same checks anyway to make transition from the mutex easier.
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002553 */
David Gibson1e8f8892006-01-06 00:10:44 -08002554static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002555 unsigned long address, pte_t *ptep, pte_t pte,
2556 struct page *pagecache_page)
David Gibson1e8f8892006-01-06 00:10:44 -08002557{
Andi Kleena5516432008-07-23 21:27:41 -07002558 struct hstate *h = hstate_vma(vma);
David Gibson1e8f8892006-01-06 00:10:44 -08002559 struct page *old_page, *new_page;
David Gibson79ac6ba2006-03-22 00:08:51 -08002560 int avoidcopy;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002561 int outside_reserve = 0;
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07002562 unsigned long mmun_start; /* For mmu_notifiers */
2563 unsigned long mmun_end; /* For mmu_notifiers */
David Gibson1e8f8892006-01-06 00:10:44 -08002564
2565 old_page = pte_page(pte);
2566
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002567retry_avoidcopy:
David Gibson1e8f8892006-01-06 00:10:44 -08002568 /* If no-one else is actually using this page, avoid the copy
2569 * and just make the page writable */
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002570 avoidcopy = (page_mapcount(old_page) == 1);
David Gibson1e8f8892006-01-06 00:10:44 -08002571 if (avoidcopy) {
Naoya Horiguchi56c9cfb2010-09-10 13:23:04 +09002572 if (PageAnon(old_page))
2573 page_move_anon_rmap(old_page, vma, address);
David Gibson1e8f8892006-01-06 00:10:44 -08002574 set_huge_ptep_writable(vma, address, ptep);
Nick Piggin83c54072007-07-19 01:47:05 -07002575 return 0;
David Gibson1e8f8892006-01-06 00:10:44 -08002576 }
2577
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002578 /*
2579 * If the process that created a MAP_PRIVATE mapping is about to
2580 * perform a COW due to a shared page count, attempt to satisfy
2581 * the allocation without using the existing reserves. The pagecache
2582 * page is used to determine if the reserve at this address was
2583 * consumed or not. If reserves were used, a partial faulted mapping
2584 * at the time of fork() could consume its reserves on COW instead
2585 * of the full address range.
2586 */
Mel Gormanf83a2752009-05-28 14:34:40 -07002587 if (!(vma->vm_flags & VM_MAYSHARE) &&
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002588 is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
2589 old_page != pagecache_page)
2590 outside_reserve = 1;
2591
David Gibson1e8f8892006-01-06 00:10:44 -08002592 page_cache_get(old_page);
Larry Woodmanb76c8cf2009-12-14 17:59:37 -08002593
2594 /* Drop page_table_lock as buddy allocator may be called */
2595 spin_unlock(&mm->page_table_lock);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002596 new_page = alloc_huge_page(vma, address, outside_reserve);
David Gibson1e8f8892006-01-06 00:10:44 -08002597
Adam Litke2fc39ce2007-11-14 16:59:39 -08002598 if (IS_ERR(new_page)) {
Aneesh Kumar K.V76dcee72012-07-31 16:41:57 -07002599 long err = PTR_ERR(new_page);
David Gibson1e8f8892006-01-06 00:10:44 -08002600 page_cache_release(old_page);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002601
2602 /*
2603 * If a process owning a MAP_PRIVATE mapping fails to COW,
2604 * it is due to references held by a child and an insufficient
2605 * huge page pool. To guarantee the original mappers
2606 * reliability, unmap the page from child processes. The child
2607 * may get SIGKILLed if it later faults.
2608 */
2609 if (outside_reserve) {
2610 BUG_ON(huge_pte_none(pte));
2611 if (unmap_ref_private(mm, vma, old_page, address)) {
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002612 BUG_ON(huge_pte_none(pte));
Larry Woodmanb76c8cf2009-12-14 17:59:37 -08002613 spin_lock(&mm->page_table_lock);
Hillf Dantona734bcc2012-01-10 15:07:20 -08002614 ptep = huge_pte_offset(mm, address & huge_page_mask(h));
2615 if (likely(pte_same(huge_ptep_get(ptep), pte)))
2616 goto retry_avoidcopy;
2617 /*
2618 * race occurs while re-acquiring page_table_lock, and
2619 * our job is done.
2620 */
2621 return 0;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002622 }
2623 WARN_ON_ONCE(1);
2624 }
2625
Larry Woodmanb76c8cf2009-12-14 17:59:37 -08002626 /* Caller expects lock to be held */
2627 spin_lock(&mm->page_table_lock);
Aneesh Kumar K.V76dcee72012-07-31 16:41:57 -07002628 if (err == -ENOMEM)
2629 return VM_FAULT_OOM;
2630 else
2631 return VM_FAULT_SIGBUS;
David Gibson1e8f8892006-01-06 00:10:44 -08002632 }
2633
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002634 /*
2635 * When the original hugepage is shared one, it does not have
2636 * anon_vma prepared.
2637 */
Dean Nelson44e2aa92010-10-26 14:22:08 -07002638 if (unlikely(anon_vma_prepare(vma))) {
Hillf Dantonea4039a2011-11-15 14:36:12 -08002639 page_cache_release(new_page);
2640 page_cache_release(old_page);
Dean Nelson44e2aa92010-10-26 14:22:08 -07002641 /* Caller expects lock to be held */
2642 spin_lock(&mm->page_table_lock);
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002643 return VM_FAULT_OOM;
Dean Nelson44e2aa92010-10-26 14:22:08 -07002644 }
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002645
Andrea Arcangeli47ad8472011-01-13 15:46:47 -08002646 copy_user_huge_page(new_page, old_page, address, vma,
2647 pages_per_huge_page(h));
Nick Piggin0ed361d2008-02-04 22:29:34 -08002648 __SetPageUptodate(new_page);
David Gibson1e8f8892006-01-06 00:10:44 -08002649
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07002650 mmun_start = address & huge_page_mask(h);
2651 mmun_end = mmun_start + huge_page_size(h);
2652 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
Larry Woodmanb76c8cf2009-12-14 17:59:37 -08002653 /*
2654 * Retake the page_table_lock to check for racing updates
2655 * before the page tables are altered
2656 */
2657 spin_lock(&mm->page_table_lock);
Andi Kleena5516432008-07-23 21:27:41 -07002658 ptep = huge_pte_offset(mm, address & huge_page_mask(h));
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07002659 if (likely(pte_same(huge_ptep_get(ptep), pte))) {
David Gibson1e8f8892006-01-06 00:10:44 -08002660 /* Break COW */
Gerald Schaefer8fe627e2008-04-28 02:13:28 -07002661 huge_ptep_clear_flush(vma, address, ptep);
David Gibson1e8f8892006-01-06 00:10:44 -08002662 set_huge_pte_at(mm, address, ptep,
2663 make_huge_pte(vma, new_page, 1));
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002664 page_remove_rmap(old_page);
Naoya Horiguchicd67f0d2010-09-10 13:23:04 +09002665 hugepage_add_new_anon_rmap(new_page, vma, address);
David Gibson1e8f8892006-01-06 00:10:44 -08002666 /* Make the old page be freed below */
2667 new_page = old_page;
2668 }
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07002669 spin_unlock(&mm->page_table_lock);
2670 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2671 /* Caller expects lock to be held */
2672 spin_lock(&mm->page_table_lock);
David Gibson1e8f8892006-01-06 00:10:44 -08002673 page_cache_release(new_page);
2674 page_cache_release(old_page);
Nick Piggin83c54072007-07-19 01:47:05 -07002675 return 0;
David Gibson1e8f8892006-01-06 00:10:44 -08002676}
2677
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002678/* Return the pagecache page at a given address within a VMA */
Andi Kleena5516432008-07-23 21:27:41 -07002679static struct page *hugetlbfs_pagecache_page(struct hstate *h,
2680 struct vm_area_struct *vma, unsigned long address)
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002681{
2682 struct address_space *mapping;
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -07002683 pgoff_t idx;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002684
2685 mapping = vma->vm_file->f_mapping;
Andi Kleena5516432008-07-23 21:27:41 -07002686 idx = vma_hugecache_offset(h, vma, address);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002687
2688 return find_lock_page(mapping, idx);
2689}
2690
Hugh Dickins3ae77f42009-09-21 17:03:33 -07002691/*
2692 * Return whether there is a pagecache page to back given address within VMA.
2693 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
2694 */
2695static bool hugetlbfs_pagecache_present(struct hstate *h,
Hugh Dickins2a15efc2009-09-21 17:03:27 -07002696 struct vm_area_struct *vma, unsigned long address)
2697{
2698 struct address_space *mapping;
2699 pgoff_t idx;
2700 struct page *page;
2701
2702 mapping = vma->vm_file->f_mapping;
2703 idx = vma_hugecache_offset(h, vma, address);
2704
2705 page = find_get_page(mapping, idx);
2706 if (page)
2707 put_page(page);
2708 return page != NULL;
2709}
2710
Robert P. J. Daya1ed3dd2007-07-17 04:03:33 -07002711static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
Hugh Dickins788c7df2009-06-23 13:49:05 +01002712 unsigned long address, pte_t *ptep, unsigned int flags)
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01002713{
Andi Kleena5516432008-07-23 21:27:41 -07002714 struct hstate *h = hstate_vma(vma);
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01002715 int ret = VM_FAULT_SIGBUS;
Hillf Danton409eb8c2012-01-20 14:34:13 -08002716 int anon_rmap = 0;
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -07002717 pgoff_t idx;
Adam Litke4c887262005-10-29 18:16:46 -07002718 unsigned long size;
Adam Litke4c887262005-10-29 18:16:46 -07002719 struct page *page;
2720 struct address_space *mapping;
David Gibson1e8f8892006-01-06 00:10:44 -08002721 pte_t new_pte;
Adam Litke4c887262005-10-29 18:16:46 -07002722
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002723 /*
2724 * Currently, we are forced to kill the process in the event the
2725 * original mapper has unmapped pages from the child due to a failed
Lucas De Marchi25985ed2011-03-30 22:57:33 -03002726 * COW. Warn that such a situation has occurred as it may not be obvious
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002727 */
2728 if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
Andrew Mortonffb22af2013-02-22 16:32:08 -08002729 pr_warning("PID %d killed due to inadequate hugepage pool\n",
2730 current->pid);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002731 return ret;
2732 }
2733
Adam Litke4c887262005-10-29 18:16:46 -07002734 mapping = vma->vm_file->f_mapping;
Andi Kleena5516432008-07-23 21:27:41 -07002735 idx = vma_hugecache_offset(h, vma, address);
Adam Litke4c887262005-10-29 18:16:46 -07002736
2737 /*
2738 * Use page lock to guard against racing truncation
2739 * before we get page_table_lock.
2740 */
Christoph Lameter6bda6662006-01-06 00:10:49 -08002741retry:
2742 page = find_lock_page(mapping, idx);
2743 if (!page) {
Andi Kleena5516432008-07-23 21:27:41 -07002744 size = i_size_read(mapping->host) >> huge_page_shift(h);
Hugh Dickinsebed4bf2006-10-28 10:38:43 -07002745 if (idx >= size)
2746 goto out;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002747 page = alloc_huge_page(vma, address, 0);
Adam Litke2fc39ce2007-11-14 16:59:39 -08002748 if (IS_ERR(page)) {
Aneesh Kumar K.V76dcee72012-07-31 16:41:57 -07002749 ret = PTR_ERR(page);
2750 if (ret == -ENOMEM)
2751 ret = VM_FAULT_OOM;
2752 else
2753 ret = VM_FAULT_SIGBUS;
Christoph Lameter6bda6662006-01-06 00:10:49 -08002754 goto out;
2755 }
Andrea Arcangeli47ad8472011-01-13 15:46:47 -08002756 clear_huge_page(page, address, pages_per_huge_page(h));
Nick Piggin0ed361d2008-02-04 22:29:34 -08002757 __SetPageUptodate(page);
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01002758
Mel Gormanf83a2752009-05-28 14:34:40 -07002759 if (vma->vm_flags & VM_MAYSHARE) {
Christoph Lameter6bda6662006-01-06 00:10:49 -08002760 int err;
Ken Chen45c682a2007-11-14 16:59:44 -08002761 struct inode *inode = mapping->host;
Christoph Lameter6bda6662006-01-06 00:10:49 -08002762
2763 err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
2764 if (err) {
2765 put_page(page);
Christoph Lameter6bda6662006-01-06 00:10:49 -08002766 if (err == -EEXIST)
2767 goto retry;
2768 goto out;
2769 }
Ken Chen45c682a2007-11-14 16:59:44 -08002770
2771 spin_lock(&inode->i_lock);
Andi Kleena5516432008-07-23 21:27:41 -07002772 inode->i_blocks += blocks_per_huge_page(h);
Ken Chen45c682a2007-11-14 16:59:44 -08002773 spin_unlock(&inode->i_lock);
Mel Gorman23be7462010-04-23 13:17:56 -04002774 } else {
Christoph Lameter6bda6662006-01-06 00:10:49 -08002775 lock_page(page);
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002776 if (unlikely(anon_vma_prepare(vma))) {
2777 ret = VM_FAULT_OOM;
2778 goto backout_unlocked;
2779 }
Hillf Danton409eb8c2012-01-20 14:34:13 -08002780 anon_rmap = 1;
Mel Gorman23be7462010-04-23 13:17:56 -04002781 }
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002782 } else {
Naoya Horiguchi998b4382010-09-08 10:19:32 +09002783 /*
2784 * If memory error occurs between mmap() and fault, some process
2785 * don't have hwpoisoned swap entry for errored virtual address.
2786 * So we need to block hugepage fault by PG_hwpoison bit check.
2787 */
2788 if (unlikely(PageHWPoison(page))) {
Chris Forbes32f84522011-07-25 17:12:14 -07002789 ret = VM_FAULT_HWPOISON |
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07002790 VM_FAULT_SET_HINDEX(hstate_index(h));
Naoya Horiguchi998b4382010-09-08 10:19:32 +09002791 goto backout_unlocked;
2792 }
Christoph Lameter6bda6662006-01-06 00:10:49 -08002793 }
David Gibson1e8f8892006-01-06 00:10:44 -08002794
Andy Whitcroft57303d82008-08-12 15:08:47 -07002795 /*
2796 * If we are going to COW a private mapping later, we examine the
2797 * pending reservations for this page now. This will ensure that
2798 * any allocations necessary to record that reservation occur outside
2799 * the spinlock.
2800 */
Hugh Dickins788c7df2009-06-23 13:49:05 +01002801 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED))
Andy Whitcroft2b267362008-08-12 15:08:49 -07002802 if (vma_needs_reservation(h, vma, address) < 0) {
2803 ret = VM_FAULT_OOM;
2804 goto backout_unlocked;
2805 }
Andy Whitcroft57303d82008-08-12 15:08:47 -07002806
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01002807 spin_lock(&mm->page_table_lock);
Andi Kleena5516432008-07-23 21:27:41 -07002808 size = i_size_read(mapping->host) >> huge_page_shift(h);
Adam Litke4c887262005-10-29 18:16:46 -07002809 if (idx >= size)
2810 goto backout;
2811
Nick Piggin83c54072007-07-19 01:47:05 -07002812 ret = 0;
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07002813 if (!huge_pte_none(huge_ptep_get(ptep)))
Adam Litke4c887262005-10-29 18:16:46 -07002814 goto backout;
2815
Hillf Danton409eb8c2012-01-20 14:34:13 -08002816 if (anon_rmap)
2817 hugepage_add_new_anon_rmap(page, vma, address);
2818 else
2819 page_dup_rmap(page);
David Gibson1e8f8892006-01-06 00:10:44 -08002820 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
2821 && (vma->vm_flags & VM_SHARED)));
2822 set_huge_pte_at(mm, address, ptep, new_pte);
2823
Hugh Dickins788c7df2009-06-23 13:49:05 +01002824 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
David Gibson1e8f8892006-01-06 00:10:44 -08002825 /* Optimization, do the COW without a second fault */
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002826 ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page);
David Gibson1e8f8892006-01-06 00:10:44 -08002827 }
2828
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01002829 spin_unlock(&mm->page_table_lock);
Adam Litke4c887262005-10-29 18:16:46 -07002830 unlock_page(page);
2831out:
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01002832 return ret;
Adam Litke4c887262005-10-29 18:16:46 -07002833
2834backout:
2835 spin_unlock(&mm->page_table_lock);
Andy Whitcroft2b267362008-08-12 15:08:49 -07002836backout_unlocked:
Adam Litke4c887262005-10-29 18:16:46 -07002837 unlock_page(page);
2838 put_page(page);
2839 goto out;
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01002840}
2841
Adam Litke86e52162006-01-06 00:10:43 -08002842int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
Hugh Dickins788c7df2009-06-23 13:49:05 +01002843 unsigned long address, unsigned int flags)
Adam Litke86e52162006-01-06 00:10:43 -08002844{
2845 pte_t *ptep;
2846 pte_t entry;
David Gibson1e8f8892006-01-06 00:10:44 -08002847 int ret;
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002848 struct page *page = NULL;
Andy Whitcroft57303d82008-08-12 15:08:47 -07002849 struct page *pagecache_page = NULL;
David Gibson3935baa2006-03-22 00:08:53 -08002850 static DEFINE_MUTEX(hugetlb_instantiation_mutex);
Andi Kleena5516432008-07-23 21:27:41 -07002851 struct hstate *h = hstate_vma(vma);
Adam Litke86e52162006-01-06 00:10:43 -08002852
KAMEZAWA Hiroyuki1e16a532012-01-10 15:07:22 -08002853 address &= huge_page_mask(h);
2854
Naoya Horiguchifd6a03ed2010-05-28 09:29:21 +09002855 ptep = huge_pte_offset(mm, address);
2856 if (ptep) {
2857 entry = huge_ptep_get(ptep);
Naoya Horiguchi290408d2010-09-08 10:19:35 +09002858 if (unlikely(is_hugetlb_entry_migration(entry))) {
Naoya Horiguchi30dad302013-06-12 14:05:04 -07002859 migration_entry_wait_huge(mm, ptep);
Naoya Horiguchi290408d2010-09-08 10:19:35 +09002860 return 0;
2861 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
Chris Forbes32f84522011-07-25 17:12:14 -07002862 return VM_FAULT_HWPOISON_LARGE |
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07002863 VM_FAULT_SET_HINDEX(hstate_index(h));
Naoya Horiguchifd6a03ed2010-05-28 09:29:21 +09002864 }
2865
Andi Kleena5516432008-07-23 21:27:41 -07002866 ptep = huge_pte_alloc(mm, address, huge_page_size(h));
Adam Litke86e52162006-01-06 00:10:43 -08002867 if (!ptep)
2868 return VM_FAULT_OOM;
2869
David Gibson3935baa2006-03-22 00:08:53 -08002870 /*
2871 * Serialize hugepage allocation and instantiation, so that we don't
2872 * get spurious allocation failures if two CPUs race to instantiate
2873 * the same page in the page cache.
2874 */
2875 mutex_lock(&hugetlb_instantiation_mutex);
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07002876 entry = huge_ptep_get(ptep);
2877 if (huge_pte_none(entry)) {
Hugh Dickins788c7df2009-06-23 13:49:05 +01002878 ret = hugetlb_no_page(mm, vma, address, ptep, flags);
David Gibsonb4d1d992008-10-15 22:01:11 -07002879 goto out_mutex;
David Gibson3935baa2006-03-22 00:08:53 -08002880 }
Adam Litke86e52162006-01-06 00:10:43 -08002881
Nick Piggin83c54072007-07-19 01:47:05 -07002882 ret = 0;
David Gibson1e8f8892006-01-06 00:10:44 -08002883
Andy Whitcroft57303d82008-08-12 15:08:47 -07002884 /*
2885 * If we are going to COW the mapping later, we examine the pending
2886 * reservations for this page now. This will ensure that any
2887 * allocations necessary to record that reservation occur outside the
2888 * spinlock. For private mappings, we also lookup the pagecache
2889 * page now as it is used to determine if a reservation has been
2890 * consumed.
2891 */
Gerald Schaefer106c9922013-04-29 15:07:23 -07002892 if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) {
Andy Whitcroft2b267362008-08-12 15:08:49 -07002893 if (vma_needs_reservation(h, vma, address) < 0) {
2894 ret = VM_FAULT_OOM;
David Gibsonb4d1d992008-10-15 22:01:11 -07002895 goto out_mutex;
Andy Whitcroft2b267362008-08-12 15:08:49 -07002896 }
Andy Whitcroft57303d82008-08-12 15:08:47 -07002897
Mel Gormanf83a2752009-05-28 14:34:40 -07002898 if (!(vma->vm_flags & VM_MAYSHARE))
Andy Whitcroft57303d82008-08-12 15:08:47 -07002899 pagecache_page = hugetlbfs_pagecache_page(h,
2900 vma, address);
2901 }
2902
Naoya Horiguchi56c9cfb2010-09-10 13:23:04 +09002903 /*
2904 * hugetlb_cow() requires page locks of pte_page(entry) and
2905 * pagecache_page, so here we need take the former one
2906 * when page != pagecache_page or !pagecache_page.
2907 * Note that locking order is always pagecache_page -> page,
2908 * so no worry about deadlock.
2909 */
2910 page = pte_page(entry);
Chris Metcalf66aebce2012-04-12 12:49:15 -07002911 get_page(page);
Naoya Horiguchi56c9cfb2010-09-10 13:23:04 +09002912 if (page != pagecache_page)
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002913 lock_page(page);
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002914
David Gibson1e8f8892006-01-06 00:10:44 -08002915 spin_lock(&mm->page_table_lock);
2916 /* Check for a racing update before calling hugetlb_cow */
David Gibsonb4d1d992008-10-15 22:01:11 -07002917 if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
2918 goto out_page_table_lock;
2919
2920
Hugh Dickins788c7df2009-06-23 13:49:05 +01002921 if (flags & FAULT_FLAG_WRITE) {
Gerald Schaefer106c9922013-04-29 15:07:23 -07002922 if (!huge_pte_write(entry)) {
Andy Whitcroft57303d82008-08-12 15:08:47 -07002923 ret = hugetlb_cow(mm, vma, address, ptep, entry,
2924 pagecache_page);
David Gibsonb4d1d992008-10-15 22:01:11 -07002925 goto out_page_table_lock;
2926 }
Gerald Schaefer106c9922013-04-29 15:07:23 -07002927 entry = huge_pte_mkdirty(entry);
David Gibsonb4d1d992008-10-15 22:01:11 -07002928 }
2929 entry = pte_mkyoung(entry);
Hugh Dickins788c7df2009-06-23 13:49:05 +01002930 if (huge_ptep_set_access_flags(vma, address, ptep, entry,
2931 flags & FAULT_FLAG_WRITE))
Russell King4b3073e2009-12-18 16:40:18 +00002932 update_mmu_cache(vma, address, ptep);
David Gibsonb4d1d992008-10-15 22:01:11 -07002933
2934out_page_table_lock:
David Gibson1e8f8892006-01-06 00:10:44 -08002935 spin_unlock(&mm->page_table_lock);
Andy Whitcroft57303d82008-08-12 15:08:47 -07002936
2937 if (pagecache_page) {
2938 unlock_page(pagecache_page);
2939 put_page(pagecache_page);
2940 }
Dean Nelson1f64d692010-12-02 14:31:12 -08002941 if (page != pagecache_page)
2942 unlock_page(page);
Chris Metcalf66aebce2012-04-12 12:49:15 -07002943 put_page(page);
Andy Whitcroft57303d82008-08-12 15:08:47 -07002944
David Gibsonb4d1d992008-10-15 22:01:11 -07002945out_mutex:
David Gibson3935baa2006-03-22 00:08:53 -08002946 mutex_unlock(&hugetlb_instantiation_mutex);
David Gibson1e8f8892006-01-06 00:10:44 -08002947
2948 return ret;
Adam Litke86e52162006-01-06 00:10:43 -08002949}
2950
Michel Lespinasse28a35712013-02-22 16:35:55 -08002951long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
2952 struct page **pages, struct vm_area_struct **vmas,
2953 unsigned long *position, unsigned long *nr_pages,
2954 long i, unsigned int flags)
David Gibson63551ae2005-06-21 17:14:44 -07002955{
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08002956 unsigned long pfn_offset;
2957 unsigned long vaddr = *position;
Michel Lespinasse28a35712013-02-22 16:35:55 -08002958 unsigned long remainder = *nr_pages;
Andi Kleena5516432008-07-23 21:27:41 -07002959 struct hstate *h = hstate_vma(vma);
David Gibson63551ae2005-06-21 17:14:44 -07002960
Hugh Dickins1c598272005-10-19 21:23:43 -07002961 spin_lock(&mm->page_table_lock);
David Gibson63551ae2005-06-21 17:14:44 -07002962 while (vaddr < vma->vm_end && remainder) {
Adam Litke4c887262005-10-29 18:16:46 -07002963 pte_t *pte;
Hugh Dickins2a15efc2009-09-21 17:03:27 -07002964 int absent;
Adam Litke4c887262005-10-29 18:16:46 -07002965 struct page *page;
2966
2967 /*
2968 * Some archs (sparc64, sh*) have multiple pte_ts to
Hugh Dickins2a15efc2009-09-21 17:03:27 -07002969 * each hugepage. We have to make sure we get the
Adam Litke4c887262005-10-29 18:16:46 -07002970 * first, for the page indexing below to work.
2971 */
Andi Kleena5516432008-07-23 21:27:41 -07002972 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h));
Hugh Dickins2a15efc2009-09-21 17:03:27 -07002973 absent = !pte || huge_pte_none(huge_ptep_get(pte));
Adam Litke4c887262005-10-29 18:16:46 -07002974
Hugh Dickins2a15efc2009-09-21 17:03:27 -07002975 /*
2976 * When coredumping, it suits get_dump_page if we just return
Hugh Dickins3ae77f42009-09-21 17:03:33 -07002977 * an error where there's an empty slot with no huge pagecache
2978 * to back it. This way, we avoid allocating a hugepage, and
2979 * the sparse dumpfile avoids allocating disk blocks, but its
2980 * huge holes still show up with zeroes where they need to be.
Hugh Dickins2a15efc2009-09-21 17:03:27 -07002981 */
Hugh Dickins3ae77f42009-09-21 17:03:33 -07002982 if (absent && (flags & FOLL_DUMP) &&
2983 !hugetlbfs_pagecache_present(h, vma, vaddr)) {
Hugh Dickins2a15efc2009-09-21 17:03:27 -07002984 remainder = 0;
2985 break;
2986 }
2987
Naoya Horiguchi9cc3a5b2013-04-17 15:58:30 -07002988 /*
2989 * We need call hugetlb_fault for both hugepages under migration
2990 * (in which case hugetlb_fault waits for the migration,) and
2991 * hwpoisoned hugepages (in which case we need to prevent the
2992 * caller from accessing to them.) In order to do this, we use
2993 * here is_swap_pte instead of is_hugetlb_entry_migration and
2994 * is_hugetlb_entry_hwpoisoned. This is because it simply covers
2995 * both cases, and because we can't follow correct pages
2996 * directly from any kind of swap entries.
2997 */
2998 if (absent || is_swap_pte(huge_ptep_get(pte)) ||
Gerald Schaefer106c9922013-04-29 15:07:23 -07002999 ((flags & FOLL_WRITE) &&
3000 !huge_pte_write(huge_ptep_get(pte)))) {
Adam Litke4c887262005-10-29 18:16:46 -07003001 int ret;
3002
3003 spin_unlock(&mm->page_table_lock);
Hugh Dickins2a15efc2009-09-21 17:03:27 -07003004 ret = hugetlb_fault(mm, vma, vaddr,
3005 (flags & FOLL_WRITE) ? FAULT_FLAG_WRITE : 0);
Adam Litke4c887262005-10-29 18:16:46 -07003006 spin_lock(&mm->page_table_lock);
Adam Litkea89182c2007-08-22 14:01:51 -07003007 if (!(ret & VM_FAULT_ERROR))
Adam Litke4c887262005-10-29 18:16:46 -07003008 continue;
3009
3010 remainder = 0;
Adam Litke4c887262005-10-29 18:16:46 -07003011 break;
3012 }
David Gibson63551ae2005-06-21 17:14:44 -07003013
Andi Kleena5516432008-07-23 21:27:41 -07003014 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07003015 page = pte_page(huge_ptep_get(pte));
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08003016same_page:
Chen, Kenneth Wd6692182006-03-31 02:29:57 -08003017 if (pages) {
Hugh Dickins2a15efc2009-09-21 17:03:27 -07003018 pages[i] = mem_map_offset(page, pfn_offset);
KOSAKI Motohiro4b2e38a2008-10-18 20:27:10 -07003019 get_page(pages[i]);
Chen, Kenneth Wd6692182006-03-31 02:29:57 -08003020 }
David Gibson63551ae2005-06-21 17:14:44 -07003021
3022 if (vmas)
3023 vmas[i] = vma;
3024
3025 vaddr += PAGE_SIZE;
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08003026 ++pfn_offset;
David Gibson63551ae2005-06-21 17:14:44 -07003027 --remainder;
3028 ++i;
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08003029 if (vaddr < vma->vm_end && remainder &&
Andi Kleena5516432008-07-23 21:27:41 -07003030 pfn_offset < pages_per_huge_page(h)) {
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08003031 /*
3032 * We use pfn_offset to avoid touching the pageframes
3033 * of this compound page.
3034 */
3035 goto same_page;
3036 }
David Gibson63551ae2005-06-21 17:14:44 -07003037 }
Hugh Dickins1c598272005-10-19 21:23:43 -07003038 spin_unlock(&mm->page_table_lock);
Michel Lespinasse28a35712013-02-22 16:35:55 -08003039 *nr_pages = remainder;
David Gibson63551ae2005-06-21 17:14:44 -07003040 *position = vaddr;
3041
Hugh Dickins2a15efc2009-09-21 17:03:27 -07003042 return i ? i : -EFAULT;
David Gibson63551ae2005-06-21 17:14:44 -07003043}
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003044
Peter Zijlstra7da4d642012-11-19 03:14:23 +01003045unsigned long hugetlb_change_protection(struct vm_area_struct *vma,
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003046 unsigned long address, unsigned long end, pgprot_t newprot)
3047{
3048 struct mm_struct *mm = vma->vm_mm;
3049 unsigned long start = address;
3050 pte_t *ptep;
3051 pte_t pte;
Andi Kleena5516432008-07-23 21:27:41 -07003052 struct hstate *h = hstate_vma(vma);
Peter Zijlstra7da4d642012-11-19 03:14:23 +01003053 unsigned long pages = 0;
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003054
3055 BUG_ON(address >= end);
3056 flush_cache_range(vma, address, end);
3057
Peter Zijlstra3d48ae42011-05-24 17:12:06 -07003058 mutex_lock(&vma->vm_file->f_mapping->i_mmap_mutex);
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003059 spin_lock(&mm->page_table_lock);
Andi Kleena5516432008-07-23 21:27:41 -07003060 for (; address < end; address += huge_page_size(h)) {
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003061 ptep = huge_pte_offset(mm, address);
3062 if (!ptep)
3063 continue;
Peter Zijlstra7da4d642012-11-19 03:14:23 +01003064 if (huge_pmd_unshare(mm, &address, ptep)) {
3065 pages++;
Chen, Kenneth W39dde652006-12-06 20:32:03 -08003066 continue;
Peter Zijlstra7da4d642012-11-19 03:14:23 +01003067 }
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07003068 if (!huge_pte_none(huge_ptep_get(ptep))) {
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003069 pte = huge_ptep_get_and_clear(mm, address, ptep);
Gerald Schaefer106c9922013-04-29 15:07:23 -07003070 pte = pte_mkhuge(huge_pte_modify(pte, newprot));
Tony Lube7517d2013-02-04 14:28:46 -08003071 pte = arch_make_huge_pte(pte, vma, NULL, 0);
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003072 set_huge_pte_at(mm, address, ptep, pte);
Peter Zijlstra7da4d642012-11-19 03:14:23 +01003073 pages++;
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003074 }
3075 }
3076 spin_unlock(&mm->page_table_lock);
Mel Gormand8333522012-07-31 16:46:20 -07003077 /*
3078 * Must flush TLB before releasing i_mmap_mutex: x86's huge_pmd_unshare
3079 * may have cleared our pud entry and done put_page on the page table:
3080 * once we release i_mmap_mutex, another task can do the final put_page
3081 * and that page table be reused and filled with junk.
3082 */
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003083 flush_tlb_range(vma, start, end);
Mel Gormand8333522012-07-31 16:46:20 -07003084 mutex_unlock(&vma->vm_file->f_mapping->i_mmap_mutex);
Peter Zijlstra7da4d642012-11-19 03:14:23 +01003085
3086 return pages << h->order;
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003087}
3088
Mel Gormana1e78772008-07-23 21:27:23 -07003089int hugetlb_reserve_pages(struct inode *inode,
3090 long from, long to,
Mel Gorman5a6fe122009-02-10 14:02:27 +00003091 struct vm_area_struct *vma,
KOSAKI Motohiroca16d142011-05-26 19:16:19 +09003092 vm_flags_t vm_flags)
Adam Litkee4e574b2007-10-16 01:26:19 -07003093{
Mel Gorman17c9d122009-02-11 16:34:16 +00003094 long ret, chg;
Andi Kleena5516432008-07-23 21:27:41 -07003095 struct hstate *h = hstate_inode(inode);
David Gibson90481622012-03-21 16:34:12 -07003096 struct hugepage_subpool *spool = subpool_inode(inode);
Adam Litkee4e574b2007-10-16 01:26:19 -07003097
Mel Gormana1e78772008-07-23 21:27:23 -07003098 /*
Mel Gorman17c9d122009-02-11 16:34:16 +00003099 * Only apply hugepage reservation if asked. At fault time, an
3100 * attempt will be made for VM_NORESERVE to allocate a page
David Gibson90481622012-03-21 16:34:12 -07003101 * without using reserves
Mel Gorman17c9d122009-02-11 16:34:16 +00003102 */
KOSAKI Motohiroca16d142011-05-26 19:16:19 +09003103 if (vm_flags & VM_NORESERVE)
Mel Gorman17c9d122009-02-11 16:34:16 +00003104 return 0;
3105
3106 /*
Mel Gormana1e78772008-07-23 21:27:23 -07003107 * Shared mappings base their reservation on the number of pages that
3108 * are already allocated on behalf of the file. Private mappings need
3109 * to reserve the full area even if read-only as mprotect() may be
3110 * called to make the mapping read-write. Assume !vma is a shm mapping
3111 */
Mel Gormanf83a2752009-05-28 14:34:40 -07003112 if (!vma || vma->vm_flags & VM_MAYSHARE)
Mel Gormana1e78772008-07-23 21:27:23 -07003113 chg = region_chg(&inode->i_mapping->private_list, from, to);
Mel Gorman5a6fe122009-02-10 14:02:27 +00003114 else {
3115 struct resv_map *resv_map = resv_map_alloc();
Mel Gorman5a6fe122009-02-10 14:02:27 +00003116 if (!resv_map)
3117 return -ENOMEM;
3118
Mel Gorman17c9d122009-02-11 16:34:16 +00003119 chg = to - from;
3120
Mel Gorman5a6fe122009-02-10 14:02:27 +00003121 set_vma_resv_map(vma, resv_map);
3122 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
3123 }
3124
Dave Hansenc50ac052012-05-29 15:06:46 -07003125 if (chg < 0) {
3126 ret = chg;
3127 goto out_err;
3128 }
Mel Gorman17c9d122009-02-11 16:34:16 +00003129
David Gibson90481622012-03-21 16:34:12 -07003130 /* There must be enough pages in the subpool for the mapping */
Dave Hansenc50ac052012-05-29 15:06:46 -07003131 if (hugepage_subpool_get_pages(spool, chg)) {
3132 ret = -ENOSPC;
3133 goto out_err;
3134 }
Mel Gorman17c9d122009-02-11 16:34:16 +00003135
3136 /*
3137 * Check enough hugepages are available for the reservation.
David Gibson90481622012-03-21 16:34:12 -07003138 * Hand the pages back to the subpool if there are not
Mel Gorman17c9d122009-02-11 16:34:16 +00003139 */
3140 ret = hugetlb_acct_memory(h, chg);
3141 if (ret < 0) {
David Gibson90481622012-03-21 16:34:12 -07003142 hugepage_subpool_put_pages(spool, chg);
Dave Hansenc50ac052012-05-29 15:06:46 -07003143 goto out_err;
Mel Gorman17c9d122009-02-11 16:34:16 +00003144 }
3145
3146 /*
3147 * Account for the reservations made. Shared mappings record regions
3148 * that have reservations as they are shared by multiple VMAs.
3149 * When the last VMA disappears, the region map says how much
3150 * the reservation was and the page cache tells how much of
3151 * the reservation was consumed. Private mappings are per-VMA and
3152 * only the consumed reservations are tracked. When the VMA
3153 * disappears, the original reservation is the VMA size and the
3154 * consumed reservations are stored in the map. Hence, nothing
3155 * else has to be done for private mappings here
3156 */
Mel Gormanf83a2752009-05-28 14:34:40 -07003157 if (!vma || vma->vm_flags & VM_MAYSHARE)
Mel Gorman17c9d122009-02-11 16:34:16 +00003158 region_add(&inode->i_mapping->private_list, from, to);
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -07003159 return 0;
Dave Hansenc50ac052012-05-29 15:06:46 -07003160out_err:
Dave Hansen4523e142012-05-30 07:51:07 -07003161 if (vma)
3162 resv_map_put(vma);
Dave Hansenc50ac052012-05-29 15:06:46 -07003163 return ret;
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -07003164}
3165
3166void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed)
3167{
Andi Kleena5516432008-07-23 21:27:41 -07003168 struct hstate *h = hstate_inode(inode);
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -07003169 long chg = region_truncate(&inode->i_mapping->private_list, offset);
David Gibson90481622012-03-21 16:34:12 -07003170 struct hugepage_subpool *spool = subpool_inode(inode);
Ken Chen45c682a2007-11-14 16:59:44 -08003171
3172 spin_lock(&inode->i_lock);
Eric Sandeene4c6f8b2009-07-29 15:02:16 -07003173 inode->i_blocks -= (blocks_per_huge_page(h) * freed);
Ken Chen45c682a2007-11-14 16:59:44 -08003174 spin_unlock(&inode->i_lock);
3175
David Gibson90481622012-03-21 16:34:12 -07003176 hugepage_subpool_put_pages(spool, (chg - freed));
Andi Kleena5516432008-07-23 21:27:41 -07003177 hugetlb_acct_memory(h, -(chg - freed));
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -07003178}
Naoya Horiguchi93f70f92010-05-28 09:29:20 +09003179
Steve Capper3212b532013-04-23 12:35:02 +01003180#ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
3181static unsigned long page_table_shareable(struct vm_area_struct *svma,
3182 struct vm_area_struct *vma,
3183 unsigned long addr, pgoff_t idx)
3184{
3185 unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) +
3186 svma->vm_start;
3187 unsigned long sbase = saddr & PUD_MASK;
3188 unsigned long s_end = sbase + PUD_SIZE;
3189
3190 /* Allow segments to share if only one is marked locked */
3191 unsigned long vm_flags = vma->vm_flags & ~VM_LOCKED;
3192 unsigned long svm_flags = svma->vm_flags & ~VM_LOCKED;
3193
3194 /*
3195 * match the virtual addresses, permission and the alignment of the
3196 * page table page.
3197 */
3198 if (pmd_index(addr) != pmd_index(saddr) ||
3199 vm_flags != svm_flags ||
3200 sbase < svma->vm_start || svma->vm_end < s_end)
3201 return 0;
3202
3203 return saddr;
3204}
3205
3206static int vma_shareable(struct vm_area_struct *vma, unsigned long addr)
3207{
3208 unsigned long base = addr & PUD_MASK;
3209 unsigned long end = base + PUD_SIZE;
3210
3211 /*
3212 * check on proper vm_flags and page table alignment
3213 */
3214 if (vma->vm_flags & VM_MAYSHARE &&
3215 vma->vm_start <= base && end <= vma->vm_end)
3216 return 1;
3217 return 0;
3218}
3219
3220/*
3221 * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
3222 * and returns the corresponding pte. While this is not necessary for the
3223 * !shared pmd case because we can allocate the pmd later as well, it makes the
3224 * code much cleaner. pmd allocation is essential for the shared case because
3225 * pud has to be populated inside the same i_mmap_mutex section - otherwise
3226 * racing tasks could either miss the sharing (see huge_pte_offset) or select a
3227 * bad pmd for sharing.
3228 */
3229pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
3230{
3231 struct vm_area_struct *vma = find_vma(mm, addr);
3232 struct address_space *mapping = vma->vm_file->f_mapping;
3233 pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) +
3234 vma->vm_pgoff;
3235 struct vm_area_struct *svma;
3236 unsigned long saddr;
3237 pte_t *spte = NULL;
3238 pte_t *pte;
3239
3240 if (!vma_shareable(vma, addr))
3241 return (pte_t *)pmd_alloc(mm, pud, addr);
3242
3243 mutex_lock(&mapping->i_mmap_mutex);
3244 vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) {
3245 if (svma == vma)
3246 continue;
3247
3248 saddr = page_table_shareable(svma, vma, addr, idx);
3249 if (saddr) {
3250 spte = huge_pte_offset(svma->vm_mm, saddr);
3251 if (spte) {
3252 get_page(virt_to_page(spte));
3253 break;
3254 }
3255 }
3256 }
3257
3258 if (!spte)
3259 goto out;
3260
3261 spin_lock(&mm->page_table_lock);
3262 if (pud_none(*pud))
3263 pud_populate(mm, pud,
3264 (pmd_t *)((unsigned long)spte & PAGE_MASK));
3265 else
3266 put_page(virt_to_page(spte));
3267 spin_unlock(&mm->page_table_lock);
3268out:
3269 pte = (pte_t *)pmd_alloc(mm, pud, addr);
3270 mutex_unlock(&mapping->i_mmap_mutex);
3271 return pte;
3272}
3273
3274/*
3275 * unmap huge page backed by shared pte.
3276 *
3277 * Hugetlb pte page is ref counted at the time of mapping. If pte is shared
3278 * indicated by page_count > 1, unmap is achieved by clearing pud and
3279 * decrementing the ref count. If count == 1, the pte page is not shared.
3280 *
3281 * called with vma->vm_mm->page_table_lock held.
3282 *
3283 * returns: 1 successfully unmapped a shared pte page
3284 * 0 the underlying pte page is not shared, or it is the last user
3285 */
3286int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
3287{
3288 pgd_t *pgd = pgd_offset(mm, *addr);
3289 pud_t *pud = pud_offset(pgd, *addr);
3290
3291 BUG_ON(page_count(virt_to_page(ptep)) == 0);
3292 if (page_count(virt_to_page(ptep)) == 1)
3293 return 0;
3294
3295 pud_clear(pud);
3296 put_page(virt_to_page(ptep));
3297 *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE;
3298 return 1;
3299}
Steve Capper9e5fc742013-04-30 08:02:03 +01003300#define want_pmd_share() (1)
3301#else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
3302pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
3303{
3304 return NULL;
3305}
3306#define want_pmd_share() (0)
Steve Capper3212b532013-04-23 12:35:02 +01003307#endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
3308
Steve Capper9e5fc742013-04-30 08:02:03 +01003309#ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
3310pte_t *huge_pte_alloc(struct mm_struct *mm,
3311 unsigned long addr, unsigned long sz)
3312{
3313 pgd_t *pgd;
3314 pud_t *pud;
3315 pte_t *pte = NULL;
3316
3317 pgd = pgd_offset(mm, addr);
3318 pud = pud_alloc(mm, pgd, addr);
3319 if (pud) {
3320 if (sz == PUD_SIZE) {
3321 pte = (pte_t *)pud;
3322 } else {
3323 BUG_ON(sz != PMD_SIZE);
3324 if (want_pmd_share() && pud_none(*pud))
3325 pte = huge_pmd_share(mm, addr, pud);
3326 else
3327 pte = (pte_t *)pmd_alloc(mm, pud, addr);
3328 }
3329 }
3330 BUG_ON(pte && !pte_none(*pte) && !pte_huge(*pte));
3331
3332 return pte;
3333}
3334
3335pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr)
3336{
3337 pgd_t *pgd;
3338 pud_t *pud;
3339 pmd_t *pmd = NULL;
3340
3341 pgd = pgd_offset(mm, addr);
3342 if (pgd_present(*pgd)) {
3343 pud = pud_offset(pgd, addr);
3344 if (pud_present(*pud)) {
3345 if (pud_huge(*pud))
3346 return (pte_t *)pud;
3347 pmd = pmd_offset(pud, addr);
3348 }
3349 }
3350 return (pte_t *) pmd;
3351}
3352
3353struct page *
3354follow_huge_pmd(struct mm_struct *mm, unsigned long address,
3355 pmd_t *pmd, int write)
3356{
3357 struct page *page;
3358
3359 page = pte_page(*(pte_t *)pmd);
3360 if (page)
3361 page += ((address & ~PMD_MASK) >> PAGE_SHIFT);
3362 return page;
3363}
3364
3365struct page *
3366follow_huge_pud(struct mm_struct *mm, unsigned long address,
3367 pud_t *pud, int write)
3368{
3369 struct page *page;
3370
3371 page = pte_page(*(pte_t *)pud);
3372 if (page)
3373 page += ((address & ~PUD_MASK) >> PAGE_SHIFT);
3374 return page;
3375}
3376
3377#else /* !CONFIG_ARCH_WANT_GENERAL_HUGETLB */
3378
3379/* Can be overriden by architectures */
3380__attribute__((weak)) struct page *
3381follow_huge_pud(struct mm_struct *mm, unsigned long address,
3382 pud_t *pud, int write)
3383{
3384 BUG();
3385 return NULL;
3386}
3387
3388#endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
3389
Andi Kleend5bd9102010-09-27 09:00:12 +02003390#ifdef CONFIG_MEMORY_FAILURE
3391
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +09003392/* Should be called in hugetlb_lock */
3393static int is_hugepage_on_freelist(struct page *hpage)
3394{
3395 struct page *page;
3396 struct page *tmp;
3397 struct hstate *h = page_hstate(hpage);
3398 int nid = page_to_nid(hpage);
3399
3400 list_for_each_entry_safe(page, tmp, &h->hugepage_freelists[nid], lru)
3401 if (page == hpage)
3402 return 1;
3403 return 0;
3404}
3405
Naoya Horiguchi93f70f92010-05-28 09:29:20 +09003406/*
3407 * This function is called from memory failure code.
3408 * Assume the caller holds page lock of the head page.
3409 */
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +09003410int dequeue_hwpoisoned_huge_page(struct page *hpage)
Naoya Horiguchi93f70f92010-05-28 09:29:20 +09003411{
3412 struct hstate *h = page_hstate(hpage);
3413 int nid = page_to_nid(hpage);
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +09003414 int ret = -EBUSY;
Naoya Horiguchi93f70f92010-05-28 09:29:20 +09003415
3416 spin_lock(&hugetlb_lock);
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +09003417 if (is_hugepage_on_freelist(hpage)) {
Naoya Horiguchi56f2fb12012-12-12 13:52:33 -08003418 /*
3419 * Hwpoisoned hugepage isn't linked to activelist or freelist,
3420 * but dangling hpage->lru can trigger list-debug warnings
3421 * (this happens when we call unpoison_memory() on it),
3422 * so let it point to itself with list_del_init().
3423 */
3424 list_del_init(&hpage->lru);
Naoya Horiguchi8c6c2ec2010-09-08 10:19:38 +09003425 set_page_refcounted(hpage);
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +09003426 h->free_huge_pages--;
3427 h->free_huge_pages_node[nid]--;
3428 ret = 0;
3429 }
Naoya Horiguchi93f70f92010-05-28 09:29:20 +09003430 spin_unlock(&hugetlb_lock);
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +09003431 return ret;
Naoya Horiguchi93f70f92010-05-28 09:29:20 +09003432}
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +09003433#endif