blob: 39b717e1267d1649908df2942cc044e9bff7cdf1 [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 Horiguchifd6a03e2010-05-28 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 Schermerhorn9a3052302009-12-14 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
322 return 1UL << (hstate->order + PAGE_SHIFT);
323}
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
Andi Kleena5516432008-07-23 21:27:41 -0700693static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700694{
Linus Torvalds1da177e2005-04-16 15:20:36 -0700695 struct page *page;
Joe Jinf96efd52007-07-15 23:38:12 -0700696
Andi Kleenaa888a72008-07-23 21:27:47 -0700697 if (h->order >= MAX_ORDER)
698 return NULL;
699
Mel Gorman6484eb32009-06-16 15:31:54 -0700700 page = alloc_pages_exact_node(nid,
Nishanth Aravamudan551883a2008-04-29 00:58:26 -0700701 htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE|
702 __GFP_REPEAT|__GFP_NOWARN,
Andi Kleena5516432008-07-23 21:27:41 -0700703 huge_page_order(h));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700704 if (page) {
Gerald Schaefer7f2e9522008-04-28 02:13:29 -0700705 if (arch_prepare_hugepage(page)) {
Gerald Schaefercaff3a22008-08-12 15:08:38 -0700706 __free_pages(page, huge_page_order(h));
Harvey Harrison7b8ee842008-04-28 14:13:19 -0700707 return NULL;
Gerald Schaefer7f2e9522008-04-28 02:13:29 -0700708 }
Andi Kleena5516432008-07-23 21:27:41 -0700709 prep_new_huge_page(h, page, nid);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700710 }
Nishanth Aravamudan63b46132007-10-16 01:26:24 -0700711
712 return page;
713}
714
Andi Kleen5ced66c2008-07-23 21:27:45 -0700715/*
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800716 * common helper functions for hstate_next_node_to_{alloc|free}.
717 * We may have allocated or freed a huge page based on a different
718 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
719 * be outside of *nodes_allowed. Ensure that we use an allowed
720 * node for alloc or free.
Lee Schermerhorn9a76db02009-12-14 17:58:15 -0800721 */
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800722static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
Lee Schermerhorn9a76db02009-12-14 17:58:15 -0800723{
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800724 nid = next_node(nid, *nodes_allowed);
Lee Schermerhorn9a76db02009-12-14 17:58:15 -0800725 if (nid == MAX_NUMNODES)
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800726 nid = first_node(*nodes_allowed);
Lee Schermerhorn9a76db02009-12-14 17:58:15 -0800727 VM_BUG_ON(nid >= MAX_NUMNODES);
728
729 return nid;
730}
731
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800732static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
Andi Kleen5ced66c2008-07-23 21:27:45 -0700733{
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800734 if (!node_isset(nid, *nodes_allowed))
735 nid = next_node_allowed(nid, nodes_allowed);
Lee Schermerhorn9a76db02009-12-14 17:58:15 -0800736 return nid;
Andi Kleen5ced66c2008-07-23 21:27:45 -0700737}
738
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800739/*
740 * returns the previously saved node ["this node"] from which to
741 * allocate a persistent huge page for the pool and advance the
742 * next node from which to allocate, handling wrap at end of node
743 * mask.
744 */
745static int hstate_next_node_to_alloc(struct hstate *h,
746 nodemask_t *nodes_allowed)
747{
748 int nid;
749
750 VM_BUG_ON(!nodes_allowed);
751
752 nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
753 h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
754
755 return nid;
756}
757
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700758/*
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800759 * helper for free_pool_huge_page() - return the previously saved
760 * node ["this node"] from which to free a huge page. Advance the
761 * next node id whether or not we find a free huge page to free so
762 * that the next attempt to free addresses the next node.
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700763 */
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800764static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700765{
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800766 int nid;
Lee Schermerhorn9a76db02009-12-14 17:58:15 -0800767
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800768 VM_BUG_ON(!nodes_allowed);
769
770 nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
771 h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
772
Lee Schermerhorn9a76db02009-12-14 17:58:15 -0800773 return nid;
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700774}
775
Joonsoo Kimcf864eb2013-09-11 14:21:00 -0700776#define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \
777 for (nr_nodes = nodes_weight(*mask); \
778 nr_nodes > 0 && \
779 ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \
780 nr_nodes--)
781
782#define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \
783 for (nr_nodes = nodes_weight(*mask); \
784 nr_nodes > 0 && \
785 ((node = hstate_next_node_to_free(hs, mask)) || 1); \
786 nr_nodes--)
787
788static int alloc_fresh_huge_page(struct hstate *h, nodemask_t *nodes_allowed)
789{
790 struct page *page;
791 int nr_nodes, node;
792 int ret = 0;
793
794 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
795 page = alloc_fresh_huge_page_node(h, node);
796 if (page) {
797 ret = 1;
798 break;
799 }
800 }
801
802 if (ret)
803 count_vm_event(HTLB_BUDDY_PGALLOC);
804 else
805 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
806
807 return ret;
808}
809
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700810/*
811 * Free huge page from pool from next node to free.
812 * Attempt to keep persistent huge pages more or less
813 * balanced over allowed nodes.
814 * Called with hugetlb_lock locked.
815 */
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800816static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
817 bool acct_surplus)
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700818{
Joonsoo Kimcf864eb2013-09-11 14:21:00 -0700819 int nr_nodes, node;
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700820 int ret = 0;
821
Joonsoo Kimcf864eb2013-09-11 14:21:00 -0700822 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
Lee Schermerhorn685f3452009-09-21 17:01:23 -0700823 /*
824 * If we're returning unused surplus pages, only examine
825 * nodes with surplus pages.
826 */
Joonsoo Kimcf864eb2013-09-11 14:21:00 -0700827 if ((!acct_surplus || h->surplus_huge_pages_node[node]) &&
828 !list_empty(&h->hugepage_freelists[node])) {
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700829 struct page *page =
Joonsoo Kimcf864eb2013-09-11 14:21:00 -0700830 list_entry(h->hugepage_freelists[node].next,
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700831 struct page, lru);
832 list_del(&page->lru);
833 h->free_huge_pages--;
Joonsoo Kimcf864eb2013-09-11 14:21:00 -0700834 h->free_huge_pages_node[node]--;
Lee Schermerhorn685f3452009-09-21 17:01:23 -0700835 if (acct_surplus) {
836 h->surplus_huge_pages--;
Joonsoo Kimcf864eb2013-09-11 14:21:00 -0700837 h->surplus_huge_pages_node[node]--;
Lee Schermerhorn685f3452009-09-21 17:01:23 -0700838 }
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700839 update_and_free_page(h, page);
840 ret = 1;
Lee Schermerhorn9a76db02009-12-14 17:58:15 -0800841 break;
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700842 }
Joonsoo Kimcf864eb2013-09-11 14:21:00 -0700843 }
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700844
845 return ret;
846}
847
Naoya Horiguchibf50bab2010-09-08 10:19:33 +0900848static struct page *alloc_buddy_huge_page(struct hstate *h, int nid)
Adam Litke7893d1d2007-10-16 01:26:18 -0700849{
850 struct page *page;
Naoya Horiguchibf50bab2010-09-08 10:19:33 +0900851 unsigned int r_nid;
Adam Litke7893d1d2007-10-16 01:26:18 -0700852
Andi Kleenaa888a72008-07-23 21:27:47 -0700853 if (h->order >= MAX_ORDER)
854 return NULL;
855
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800856 /*
857 * Assume we will successfully allocate the surplus page to
858 * prevent racing processes from causing the surplus to exceed
859 * overcommit
860 *
861 * This however introduces a different race, where a process B
862 * tries to grow the static hugepage pool while alloc_pages() is
863 * called by process A. B will only examine the per-node
864 * counters in determining if surplus huge pages can be
865 * converted to normal huge pages in adjust_pool_surplus(). A
866 * won't be able to increment the per-node counter, until the
867 * lock is dropped by B, but B doesn't drop hugetlb_lock until
868 * no more huge pages can be converted from surplus to normal
869 * state (and doesn't try to convert again). Thus, we have a
870 * case where a surplus huge page exists, the pool is grown, and
871 * the surplus huge page still exists after, even though it
872 * should just have been converted to a normal huge page. This
873 * does not leak memory, though, as the hugepage will be freed
874 * once it is out of use. It also does not allow the counters to
875 * go out of whack in adjust_pool_surplus() as we don't modify
876 * the node values until we've gotten the hugepage and only the
877 * per-node value is checked there.
878 */
879 spin_lock(&hugetlb_lock);
Andi Kleena5516432008-07-23 21:27:41 -0700880 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800881 spin_unlock(&hugetlb_lock);
882 return NULL;
883 } else {
Andi Kleena5516432008-07-23 21:27:41 -0700884 h->nr_huge_pages++;
885 h->surplus_huge_pages++;
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800886 }
887 spin_unlock(&hugetlb_lock);
888
Naoya Horiguchibf50bab2010-09-08 10:19:33 +0900889 if (nid == NUMA_NO_NODE)
890 page = alloc_pages(htlb_alloc_mask|__GFP_COMP|
891 __GFP_REPEAT|__GFP_NOWARN,
892 huge_page_order(h));
893 else
894 page = alloc_pages_exact_node(nid,
895 htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE|
896 __GFP_REPEAT|__GFP_NOWARN, huge_page_order(h));
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800897
Gerald Schaefercaff3a22008-08-12 15:08:38 -0700898 if (page && arch_prepare_hugepage(page)) {
899 __free_pages(page, huge_page_order(h));
Hillf Dantonea5768c2012-01-10 15:08:30 -0800900 page = NULL;
Gerald Schaefercaff3a22008-08-12 15:08:38 -0700901 }
902
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800903 spin_lock(&hugetlb_lock);
Adam Litke7893d1d2007-10-16 01:26:18 -0700904 if (page) {
Aneesh Kumar K.V0edaecf2012-07-31 16:42:07 -0700905 INIT_LIST_HEAD(&page->lru);
Naoya Horiguchibf50bab2010-09-08 10:19:33 +0900906 r_nid = page_to_nid(page);
Adam Litke7893d1d2007-10-16 01:26:18 -0700907 set_compound_page_dtor(page, free_huge_page);
Aneesh Kumar K.V9dd540e2012-07-31 16:42:15 -0700908 set_hugetlb_cgroup(page, NULL);
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800909 /*
910 * We incremented the global counters already
911 */
Naoya Horiguchibf50bab2010-09-08 10:19:33 +0900912 h->nr_huge_pages_node[r_nid]++;
913 h->surplus_huge_pages_node[r_nid]++;
Adam Litke3b116302008-04-28 02:13:06 -0700914 __count_vm_event(HTLB_BUDDY_PGALLOC);
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800915 } else {
Andi Kleena5516432008-07-23 21:27:41 -0700916 h->nr_huge_pages--;
917 h->surplus_huge_pages--;
Adam Litke3b116302008-04-28 02:13:06 -0700918 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
Adam Litke7893d1d2007-10-16 01:26:18 -0700919 }
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800920 spin_unlock(&hugetlb_lock);
Adam Litke7893d1d2007-10-16 01:26:18 -0700921
922 return page;
923}
924
Adam Litkee4e574b2007-10-16 01:26:19 -0700925/*
Naoya Horiguchibf50bab2010-09-08 10:19:33 +0900926 * This allocation function is useful in the context where vma is irrelevant.
927 * E.g. soft-offlining uses this function because it only cares physical
928 * address of error page.
929 */
930struct page *alloc_huge_page_node(struct hstate *h, int nid)
931{
932 struct page *page;
933
934 spin_lock(&hugetlb_lock);
935 page = dequeue_huge_page_node(h, nid);
936 spin_unlock(&hugetlb_lock);
937
Aneesh Kumar K.V94ae8ba2012-07-31 16:42:35 -0700938 if (!page)
Naoya Horiguchibf50bab2010-09-08 10:19:33 +0900939 page = alloc_buddy_huge_page(h, nid);
940
941 return page;
942}
943
944/*
Lucas De Marchi25985ed2011-03-30 22:57:33 -0300945 * Increase the hugetlb pool such that it can accommodate a reservation
Adam Litkee4e574b2007-10-16 01:26:19 -0700946 * of size 'delta'.
947 */
Andi Kleena5516432008-07-23 21:27:41 -0700948static int gather_surplus_pages(struct hstate *h, int delta)
Adam Litkee4e574b2007-10-16 01:26:19 -0700949{
950 struct list_head surplus_list;
951 struct page *page, *tmp;
952 int ret, i;
953 int needed, allocated;
Hillf Danton28073b02012-03-21 16:34:00 -0700954 bool alloc_ok = true;
Adam Litkee4e574b2007-10-16 01:26:19 -0700955
Andi Kleena5516432008-07-23 21:27:41 -0700956 needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
Adam Litkeac09b3a2008-03-04 14:29:38 -0800957 if (needed <= 0) {
Andi Kleena5516432008-07-23 21:27:41 -0700958 h->resv_huge_pages += delta;
Adam Litkee4e574b2007-10-16 01:26:19 -0700959 return 0;
Adam Litkeac09b3a2008-03-04 14:29:38 -0800960 }
Adam Litkee4e574b2007-10-16 01:26:19 -0700961
962 allocated = 0;
963 INIT_LIST_HEAD(&surplus_list);
964
965 ret = -ENOMEM;
966retry:
967 spin_unlock(&hugetlb_lock);
968 for (i = 0; i < needed; i++) {
Naoya Horiguchibf50bab2010-09-08 10:19:33 +0900969 page = alloc_buddy_huge_page(h, NUMA_NO_NODE);
Hillf Danton28073b02012-03-21 16:34:00 -0700970 if (!page) {
971 alloc_ok = false;
972 break;
973 }
Adam Litkee4e574b2007-10-16 01:26:19 -0700974 list_add(&page->lru, &surplus_list);
975 }
Hillf Danton28073b02012-03-21 16:34:00 -0700976 allocated += i;
Adam Litkee4e574b2007-10-16 01:26:19 -0700977
978 /*
979 * After retaking hugetlb_lock, we need to recalculate 'needed'
980 * because either resv_huge_pages or free_huge_pages may have changed.
981 */
982 spin_lock(&hugetlb_lock);
Andi Kleena5516432008-07-23 21:27:41 -0700983 needed = (h->resv_huge_pages + delta) -
984 (h->free_huge_pages + allocated);
Hillf Danton28073b02012-03-21 16:34:00 -0700985 if (needed > 0) {
986 if (alloc_ok)
987 goto retry;
988 /*
989 * We were not able to allocate enough pages to
990 * satisfy the entire reservation so we free what
991 * we've allocated so far.
992 */
993 goto free;
994 }
Adam Litkee4e574b2007-10-16 01:26:19 -0700995 /*
996 * The surplus_list now contains _at_least_ the number of extra pages
Lucas De Marchi25985ed2011-03-30 22:57:33 -0300997 * needed to accommodate the reservation. Add the appropriate number
Adam Litkee4e574b2007-10-16 01:26:19 -0700998 * of pages to the hugetlb pool and free the extras back to the buddy
Adam Litkeac09b3a2008-03-04 14:29:38 -0800999 * allocator. Commit the entire reservation here to prevent another
1000 * process from stealing the pages as they are added to the pool but
1001 * before they are reserved.
Adam Litkee4e574b2007-10-16 01:26:19 -07001002 */
1003 needed += allocated;
Andi Kleena5516432008-07-23 21:27:41 -07001004 h->resv_huge_pages += delta;
Adam Litkee4e574b2007-10-16 01:26:19 -07001005 ret = 0;
Naoya Horiguchia9869b82010-09-08 10:19:37 +09001006
Adam Litke19fc3f02008-04-28 02:12:20 -07001007 /* Free the needed pages to the hugetlb pool */
Adam Litkee4e574b2007-10-16 01:26:19 -07001008 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
Adam Litke19fc3f02008-04-28 02:12:20 -07001009 if ((--needed) < 0)
1010 break;
Naoya Horiguchia9869b82010-09-08 10:19:37 +09001011 /*
1012 * This page is now managed by the hugetlb allocator and has
1013 * no users -- drop the buddy allocator's reference.
1014 */
1015 put_page_testzero(page);
1016 VM_BUG_ON(page_count(page));
Andi Kleena5516432008-07-23 21:27:41 -07001017 enqueue_huge_page(h, page);
Adam Litke19fc3f02008-04-28 02:12:20 -07001018 }
Hillf Danton28073b02012-03-21 16:34:00 -07001019free:
Hillf Dantonb0365c82011-12-28 15:57:16 -08001020 spin_unlock(&hugetlb_lock);
Adam Litke19fc3f02008-04-28 02:12:20 -07001021
1022 /* Free unnecessary surplus pages to the buddy allocator */
Joonsoo Kim933a8a32013-09-11 14:21:02 -07001023 list_for_each_entry_safe(page, tmp, &surplus_list, lru)
1024 put_page(page);
Naoya Horiguchia9869b82010-09-08 10:19:37 +09001025 spin_lock(&hugetlb_lock);
Adam Litkee4e574b2007-10-16 01:26:19 -07001026
1027 return ret;
1028}
1029
1030/*
1031 * When releasing a hugetlb pool reservation, any surplus pages that were
1032 * allocated to satisfy the reservation must be explicitly freed if they were
1033 * never used.
Lee Schermerhorn685f3452009-09-21 17:01:23 -07001034 * Called with hugetlb_lock held.
Adam Litkee4e574b2007-10-16 01:26:19 -07001035 */
Andi Kleena5516432008-07-23 21:27:41 -07001036static void return_unused_surplus_pages(struct hstate *h,
1037 unsigned long unused_resv_pages)
Adam Litkee4e574b2007-10-16 01:26:19 -07001038{
Adam Litkee4e574b2007-10-16 01:26:19 -07001039 unsigned long nr_pages;
1040
Adam Litkeac09b3a2008-03-04 14:29:38 -08001041 /* Uncommit the reservation */
Andi Kleena5516432008-07-23 21:27:41 -07001042 h->resv_huge_pages -= unused_resv_pages;
Adam Litkeac09b3a2008-03-04 14:29:38 -08001043
Andi Kleenaa888a72008-07-23 21:27:47 -07001044 /* Cannot return gigantic pages currently */
1045 if (h->order >= MAX_ORDER)
1046 return;
1047
Andi Kleena5516432008-07-23 21:27:41 -07001048 nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
Adam Litkee4e574b2007-10-16 01:26:19 -07001049
Lee Schermerhorn685f3452009-09-21 17:01:23 -07001050 /*
1051 * We want to release as many surplus pages as possible, spread
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -08001052 * evenly across all nodes with memory. Iterate across these nodes
1053 * until we can no longer free unreserved surplus pages. This occurs
1054 * when the nodes with surplus pages have no free pages.
1055 * free_pool_huge_page() will balance the the freed pages across the
1056 * on-line nodes with memory and will handle the hstate accounting.
Lee Schermerhorn685f3452009-09-21 17:01:23 -07001057 */
1058 while (nr_pages--) {
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08001059 if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1))
Lee Schermerhorn685f3452009-09-21 17:01:23 -07001060 break;
Adam Litkee4e574b2007-10-16 01:26:19 -07001061 }
1062}
1063
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001064/*
1065 * Determine if the huge page at addr within the vma has an associated
1066 * reservation. Where it does not we will need to logically increase
David Gibson90481622012-03-21 16:34:12 -07001067 * reservation and actually increase subpool usage before an allocation
1068 * can occur. Where any new reservation would be required the
1069 * reservation change is prepared, but not committed. Once the page
1070 * has been allocated from the subpool and instantiated the change should
1071 * be committed via vma_commit_reservation. No action is required on
1072 * failure.
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001073 */
Roel Kluine2f17d92009-03-31 15:23:15 -07001074static long vma_needs_reservation(struct hstate *h,
Andi Kleena5516432008-07-23 21:27:41 -07001075 struct vm_area_struct *vma, unsigned long addr)
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001076{
1077 struct address_space *mapping = vma->vm_file->f_mapping;
1078 struct inode *inode = mapping->host;
1079
Mel Gormanf83a2752009-05-28 14:34:40 -07001080 if (vma->vm_flags & VM_MAYSHARE) {
Andi Kleena5516432008-07-23 21:27:41 -07001081 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001082 return region_chg(&inode->i_mapping->private_list,
1083 idx, idx + 1);
1084
Andy Whitcroft84afd992008-07-23 21:27:32 -07001085 } else if (!is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
1086 return 1;
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001087
Andy Whitcroft84afd992008-07-23 21:27:32 -07001088 } else {
Roel Kluine2f17d92009-03-31 15:23:15 -07001089 long err;
Andi Kleena5516432008-07-23 21:27:41 -07001090 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
Andy Whitcroft84afd992008-07-23 21:27:32 -07001091 struct resv_map *reservations = vma_resv_map(vma);
1092
1093 err = region_chg(&reservations->regions, idx, idx + 1);
1094 if (err < 0)
1095 return err;
1096 return 0;
1097 }
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001098}
Andi Kleena5516432008-07-23 21:27:41 -07001099static void vma_commit_reservation(struct hstate *h,
1100 struct vm_area_struct *vma, unsigned long addr)
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001101{
1102 struct address_space *mapping = vma->vm_file->f_mapping;
1103 struct inode *inode = mapping->host;
1104
Mel Gormanf83a2752009-05-28 14:34:40 -07001105 if (vma->vm_flags & VM_MAYSHARE) {
Andi Kleena5516432008-07-23 21:27:41 -07001106 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001107 region_add(&inode->i_mapping->private_list, idx, idx + 1);
Andy Whitcroft84afd992008-07-23 21:27:32 -07001108
1109 } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
Andi Kleena5516432008-07-23 21:27:41 -07001110 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
Andy Whitcroft84afd992008-07-23 21:27:32 -07001111 struct resv_map *reservations = vma_resv_map(vma);
1112
1113 /* Mark this page used in the map. */
1114 region_add(&reservations->regions, idx, idx + 1);
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001115 }
1116}
1117
David Gibson27a85ef2006-03-22 00:08:56 -08001118static struct page *alloc_huge_page(struct vm_area_struct *vma,
Mel Gorman04f2cbe2008-07-23 21:27:25 -07001119 unsigned long addr, int avoid_reserve)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001120{
David Gibson90481622012-03-21 16:34:12 -07001121 struct hugepage_subpool *spool = subpool_vma(vma);
Andi Kleena5516432008-07-23 21:27:41 -07001122 struct hstate *h = hstate_vma(vma);
Adam Litke348ea202007-11-14 16:59:37 -08001123 struct page *page;
Roel Kluine2f17d92009-03-31 15:23:15 -07001124 long chg;
Aneesh Kumar K.V6d76dcf2012-07-31 16:42:18 -07001125 int ret, idx;
1126 struct hugetlb_cgroup *h_cg;
Adam Litke2fc39ce2007-11-14 16:59:39 -08001127
Aneesh Kumar K.V6d76dcf2012-07-31 16:42:18 -07001128 idx = hstate_index(h);
Mel Gormana1e78772008-07-23 21:27:23 -07001129 /*
David Gibson90481622012-03-21 16:34:12 -07001130 * Processes that did not create the mapping will have no
1131 * reserves and will not have accounted against subpool
1132 * limit. Check that the subpool limit can be made before
1133 * satisfying the allocation MAP_NORESERVE mappings may also
1134 * need pages and subpool limit allocated allocated if no reserve
1135 * mapping overlaps.
Mel Gormana1e78772008-07-23 21:27:23 -07001136 */
Andi Kleena5516432008-07-23 21:27:41 -07001137 chg = vma_needs_reservation(h, vma, addr);
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001138 if (chg < 0)
Aneesh Kumar K.V76dcee72012-07-31 16:41:57 -07001139 return ERR_PTR(-ENOMEM);
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001140 if (chg)
David Gibson90481622012-03-21 16:34:12 -07001141 if (hugepage_subpool_get_pages(spool, chg))
Aneesh Kumar K.V76dcee72012-07-31 16:41:57 -07001142 return ERR_PTR(-ENOSPC);
Mel Gormana1e78772008-07-23 21:27:23 -07001143
Aneesh Kumar K.V6d76dcf2012-07-31 16:42:18 -07001144 ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg);
1145 if (ret) {
1146 hugepage_subpool_put_pages(spool, chg);
1147 return ERR_PTR(-ENOSPC);
1148 }
Mel Gormana1e78772008-07-23 21:27:23 -07001149 spin_lock(&hugetlb_lock);
Andi Kleena5516432008-07-23 21:27:41 -07001150 page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve);
Joonsoo Kim47606bd2013-09-11 14:20:58 -07001151 if (!page) {
Aneesh Kumar K.V94ae8ba2012-07-31 16:42:35 -07001152 spin_unlock(&hugetlb_lock);
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001153 page = alloc_buddy_huge_page(h, NUMA_NO_NODE);
Mel Gormana1e78772008-07-23 21:27:23 -07001154 if (!page) {
Aneesh Kumar K.V6d76dcf2012-07-31 16:42:18 -07001155 hugetlb_cgroup_uncharge_cgroup(idx,
1156 pages_per_huge_page(h),
1157 h_cg);
David Gibson90481622012-03-21 16:34:12 -07001158 hugepage_subpool_put_pages(spool, chg);
Aneesh Kumar K.V76dcee72012-07-31 16:41:57 -07001159 return ERR_PTR(-ENOSPC);
Mel Gormana1e78772008-07-23 21:27:23 -07001160 }
Aneesh Kumar K.V79dbb232012-07-31 16:42:32 -07001161 spin_lock(&hugetlb_lock);
1162 list_move(&page->lru, &h->hugepage_activelist);
Joonsoo Kim47606bd2013-09-11 14:20:58 -07001163 /* Fall through */
Mel Gormana1e78772008-07-23 21:27:23 -07001164 }
Joonsoo Kim47606bd2013-09-11 14:20:58 -07001165 hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page);
1166 spin_unlock(&hugetlb_lock);
Mel Gormana1e78772008-07-23 21:27:23 -07001167
David Gibson90481622012-03-21 16:34:12 -07001168 set_page_private(page, (unsigned long)spool);
Mel Gormana1e78772008-07-23 21:27:23 -07001169
Andi Kleena5516432008-07-23 21:27:41 -07001170 vma_commit_reservation(h, vma, addr);
Adam Litke90d8b7e2007-11-14 16:59:42 -08001171 return page;
David Gibsonb45b5bd2006-03-22 00:08:55 -08001172}
1173
Cyrill Gorcunov91f47662009-01-06 14:40:33 -08001174int __weak alloc_bootmem_huge_page(struct hstate *h)
Andi Kleenaa888a72008-07-23 21:27:47 -07001175{
1176 struct huge_bootmem_page *m;
Joonsoo Kimcf864eb2013-09-11 14:21:00 -07001177 int nr_nodes, node;
Andi Kleenaa888a72008-07-23 21:27:47 -07001178
Joonsoo Kimcf864eb2013-09-11 14:21:00 -07001179 for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) {
Andi Kleenaa888a72008-07-23 21:27:47 -07001180 void *addr;
1181
Joonsoo Kimcf864eb2013-09-11 14:21:00 -07001182 addr = __alloc_bootmem_node_nopanic(NODE_DATA(node),
Andi Kleenaa888a72008-07-23 21:27:47 -07001183 huge_page_size(h), huge_page_size(h), 0);
1184
1185 if (addr) {
1186 /*
1187 * Use the beginning of the huge page to store the
1188 * huge_bootmem_page struct (until gather_bootmem
1189 * puts them into the mem_map).
1190 */
1191 m = addr;
Cyrill Gorcunov91f47662009-01-06 14:40:33 -08001192 goto found;
Andi Kleenaa888a72008-07-23 21:27:47 -07001193 }
Andi Kleenaa888a72008-07-23 21:27:47 -07001194 }
1195 return 0;
1196
1197found:
1198 BUG_ON((unsigned long)virt_to_phys(m) & (huge_page_size(h) - 1));
1199 /* Put them into a private list first because mem_map is not up yet */
1200 list_add(&m->list, &huge_boot_pages);
1201 m->hstate = h;
1202 return 1;
1203}
1204
Andy Whitcroft18229df2008-11-06 12:53:27 -08001205static void prep_compound_huge_page(struct page *page, int order)
1206{
1207 if (unlikely(order > (MAX_ORDER - 1)))
1208 prep_compound_gigantic_page(page, order);
1209 else
1210 prep_compound_page(page, order);
1211}
1212
Andi Kleenaa888a72008-07-23 21:27:47 -07001213/* Put bootmem huge pages into the standard lists after mem_map is up */
1214static void __init gather_bootmem_prealloc(void)
1215{
1216 struct huge_bootmem_page *m;
1217
1218 list_for_each_entry(m, &huge_boot_pages, list) {
Andi Kleenaa888a72008-07-23 21:27:47 -07001219 struct hstate *h = m->hstate;
Becky Bruceee8f2482011-07-25 17:11:50 -07001220 struct page *page;
1221
1222#ifdef CONFIG_HIGHMEM
1223 page = pfn_to_page(m->phys >> PAGE_SHIFT);
1224 free_bootmem_late((unsigned long)m,
1225 sizeof(struct huge_bootmem_page));
1226#else
1227 page = virt_to_page(m);
1228#endif
Andi Kleenaa888a72008-07-23 21:27:47 -07001229 __ClearPageReserved(page);
1230 WARN_ON(page_count(page) != 1);
Andy Whitcroft18229df2008-11-06 12:53:27 -08001231 prep_compound_huge_page(page, h->order);
Andi Kleenaa888a72008-07-23 21:27:47 -07001232 prep_new_huge_page(h, page, page_to_nid(page));
Rafael Aquinib0320c72011-06-15 15:08:39 -07001233 /*
1234 * If we had gigantic hugepages allocated at boot time, we need
1235 * to restore the 'stolen' pages to totalram_pages in order to
1236 * fix confusing memory reports from free(1) and another
1237 * side-effects, like CommitLimit going negative.
1238 */
1239 if (h->order > (MAX_ORDER - 1))
1240 totalram_pages += 1 << h->order;
Andi Kleenaa888a72008-07-23 21:27:47 -07001241 }
1242}
1243
Andi Kleen8faa8b02008-07-23 21:27:48 -07001244static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001245{
1246 unsigned long i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001247
Andi Kleene5ff2152008-07-23 21:27:42 -07001248 for (i = 0; i < h->max_huge_pages; ++i) {
Andi Kleenaa888a72008-07-23 21:27:47 -07001249 if (h->order >= MAX_ORDER) {
1250 if (!alloc_bootmem_huge_page(h))
1251 break;
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -08001252 } else if (!alloc_fresh_huge_page(h,
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08001253 &node_states[N_MEMORY]))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001254 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001255 }
Andi Kleen8faa8b02008-07-23 21:27:48 -07001256 h->max_huge_pages = i;
Andi Kleene5ff2152008-07-23 21:27:42 -07001257}
1258
1259static void __init hugetlb_init_hstates(void)
1260{
1261 struct hstate *h;
1262
1263 for_each_hstate(h) {
Andi Kleen8faa8b02008-07-23 21:27:48 -07001264 /* oversize hugepages were init'ed in early boot */
1265 if (h->order < MAX_ORDER)
1266 hugetlb_hstate_alloc_pages(h);
Andi Kleene5ff2152008-07-23 21:27:42 -07001267 }
1268}
1269
Andi Kleen4abd32d2008-07-23 21:27:49 -07001270static char * __init memfmt(char *buf, unsigned long n)
1271{
1272 if (n >= (1UL << 30))
1273 sprintf(buf, "%lu GB", n >> 30);
1274 else if (n >= (1UL << 20))
1275 sprintf(buf, "%lu MB", n >> 20);
1276 else
1277 sprintf(buf, "%lu KB", n >> 10);
1278 return buf;
1279}
1280
Andi Kleene5ff2152008-07-23 21:27:42 -07001281static void __init report_hugepages(void)
1282{
1283 struct hstate *h;
1284
1285 for_each_hstate(h) {
Andi Kleen4abd32d2008-07-23 21:27:49 -07001286 char buf[32];
Andrew Mortonffb22af2013-02-22 16:32:08 -08001287 pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n",
Andi Kleen4abd32d2008-07-23 21:27:49 -07001288 memfmt(buf, huge_page_size(h)),
1289 h->free_huge_pages);
Andi Kleene5ff2152008-07-23 21:27:42 -07001290 }
1291}
1292
Linus Torvalds1da177e2005-04-16 15:20:36 -07001293#ifdef CONFIG_HIGHMEM
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001294static void try_to_free_low(struct hstate *h, unsigned long count,
1295 nodemask_t *nodes_allowed)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001296{
Christoph Lameter4415cc82006-09-25 23:31:55 -07001297 int i;
1298
Andi Kleenaa888a72008-07-23 21:27:47 -07001299 if (h->order >= MAX_ORDER)
1300 return;
1301
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001302 for_each_node_mask(i, *nodes_allowed) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001303 struct page *page, *next;
Andi Kleena5516432008-07-23 21:27:41 -07001304 struct list_head *freel = &h->hugepage_freelists[i];
1305 list_for_each_entry_safe(page, next, freel, lru) {
1306 if (count >= h->nr_huge_pages)
Adam Litke6b0c8802007-10-16 01:26:23 -07001307 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001308 if (PageHighMem(page))
1309 continue;
1310 list_del(&page->lru);
Andi Kleene5ff2152008-07-23 21:27:42 -07001311 update_and_free_page(h, page);
Andi Kleena5516432008-07-23 21:27:41 -07001312 h->free_huge_pages--;
1313 h->free_huge_pages_node[page_to_nid(page)]--;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001314 }
1315 }
1316}
1317#else
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001318static inline void try_to_free_low(struct hstate *h, unsigned long count,
1319 nodemask_t *nodes_allowed)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001320{
1321}
1322#endif
1323
Wu Fengguang20a03072009-06-16 15:32:22 -07001324/*
1325 * Increment or decrement surplus_huge_pages. Keep node-specific counters
1326 * balanced by operating on them in a round-robin fashion.
1327 * Returns 1 if an adjustment was made.
1328 */
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001329static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
1330 int delta)
Wu Fengguang20a03072009-06-16 15:32:22 -07001331{
Joonsoo Kimcf864eb2013-09-11 14:21:00 -07001332 int nr_nodes, node;
Wu Fengguang20a03072009-06-16 15:32:22 -07001333
1334 VM_BUG_ON(delta != -1 && delta != 1);
Wu Fengguang20a03072009-06-16 15:32:22 -07001335
Joonsoo Kimcf864eb2013-09-11 14:21:00 -07001336 if (delta < 0) {
1337 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
1338 if (h->surplus_huge_pages_node[node])
1339 goto found;
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001340 }
Joonsoo Kimcf864eb2013-09-11 14:21:00 -07001341 } else {
1342 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
1343 if (h->surplus_huge_pages_node[node] <
1344 h->nr_huge_pages_node[node])
1345 goto found;
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001346 }
Joonsoo Kimcf864eb2013-09-11 14:21:00 -07001347 }
1348 return 0;
Wu Fengguang20a03072009-06-16 15:32:22 -07001349
Joonsoo Kimcf864eb2013-09-11 14:21:00 -07001350found:
1351 h->surplus_huge_pages += delta;
1352 h->surplus_huge_pages_node[node] += delta;
1353 return 1;
Wu Fengguang20a03072009-06-16 15:32:22 -07001354}
1355
Andi Kleena5516432008-07-23 21:27:41 -07001356#define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001357static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count,
1358 nodemask_t *nodes_allowed)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001359{
Adam Litke7893d1d2007-10-16 01:26:18 -07001360 unsigned long min_count, ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001361
Andi Kleenaa888a72008-07-23 21:27:47 -07001362 if (h->order >= MAX_ORDER)
1363 return h->max_huge_pages;
1364
Adam Litke7893d1d2007-10-16 01:26:18 -07001365 /*
1366 * Increase the pool size
1367 * First take pages out of surplus state. Then make up the
1368 * remaining difference by allocating fresh huge pages.
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001369 *
1370 * We might race with alloc_buddy_huge_page() here and be unable
1371 * to convert a surplus huge page to a normal huge page. That is
1372 * not critical, though, it just means the overall size of the
1373 * pool might be one hugepage larger than it needs to be, but
1374 * within all the constraints specified by the sysctls.
Adam Litke7893d1d2007-10-16 01:26:18 -07001375 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001376 spin_lock(&hugetlb_lock);
Andi Kleena5516432008-07-23 21:27:41 -07001377 while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001378 if (!adjust_pool_surplus(h, nodes_allowed, -1))
Adam Litke7893d1d2007-10-16 01:26:18 -07001379 break;
1380 }
1381
Andi Kleena5516432008-07-23 21:27:41 -07001382 while (count > persistent_huge_pages(h)) {
Adam Litke7893d1d2007-10-16 01:26:18 -07001383 /*
1384 * If this allocation races such that we no longer need the
1385 * page, free_huge_page will handle it by freeing the page
1386 * and reducing the surplus.
1387 */
1388 spin_unlock(&hugetlb_lock);
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001389 ret = alloc_fresh_huge_page(h, nodes_allowed);
Adam Litke7893d1d2007-10-16 01:26:18 -07001390 spin_lock(&hugetlb_lock);
1391 if (!ret)
1392 goto out;
1393
Mel Gorman536240f22009-12-14 17:59:56 -08001394 /* Bail for signals. Probably ctrl-c from user */
1395 if (signal_pending(current))
1396 goto out;
Adam Litke7893d1d2007-10-16 01:26:18 -07001397 }
Adam Litke7893d1d2007-10-16 01:26:18 -07001398
1399 /*
1400 * Decrease the pool size
1401 * First return free pages to the buddy allocator (being careful
1402 * to keep enough around to satisfy reservations). Then place
1403 * pages into surplus state as needed so the pool will shrink
1404 * to the desired size as pages become free.
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001405 *
1406 * By placing pages into the surplus state independent of the
1407 * overcommit value, we are allowing the surplus pool size to
1408 * exceed overcommit. There are few sane options here. Since
1409 * alloc_buddy_huge_page() is checking the global counter,
1410 * though, we'll note that we're not allowed to exceed surplus
1411 * and won't grow the pool anywhere else. Not until one of the
1412 * sysctls are changed, or the surplus pages go out of use.
Adam Litke7893d1d2007-10-16 01:26:18 -07001413 */
Andi Kleena5516432008-07-23 21:27:41 -07001414 min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
Adam Litke6b0c8802007-10-16 01:26:23 -07001415 min_count = max(count, min_count);
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001416 try_to_free_low(h, min_count, nodes_allowed);
Andi Kleena5516432008-07-23 21:27:41 -07001417 while (min_count < persistent_huge_pages(h)) {
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001418 if (!free_pool_huge_page(h, nodes_allowed, 0))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001419 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001420 }
Andi Kleena5516432008-07-23 21:27:41 -07001421 while (count < persistent_huge_pages(h)) {
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001422 if (!adjust_pool_surplus(h, nodes_allowed, 1))
Adam Litke7893d1d2007-10-16 01:26:18 -07001423 break;
1424 }
1425out:
Andi Kleena5516432008-07-23 21:27:41 -07001426 ret = persistent_huge_pages(h);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001427 spin_unlock(&hugetlb_lock);
Adam Litke7893d1d2007-10-16 01:26:18 -07001428 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001429}
1430
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001431#define HSTATE_ATTR_RO(_name) \
1432 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
1433
1434#define HSTATE_ATTR(_name) \
1435 static struct kobj_attribute _name##_attr = \
1436 __ATTR(_name, 0644, _name##_show, _name##_store)
1437
1438static struct kobject *hugepages_kobj;
1439static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
1440
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001441static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
1442
1443static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001444{
1445 int i;
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001446
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001447 for (i = 0; i < HUGE_MAX_HSTATE; i++)
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001448 if (hstate_kobjs[i] == kobj) {
1449 if (nidp)
1450 *nidp = NUMA_NO_NODE;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001451 return &hstates[i];
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001452 }
1453
1454 return kobj_to_node_hstate(kobj, nidp);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001455}
1456
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001457static ssize_t nr_hugepages_show_common(struct kobject *kobj,
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001458 struct kobj_attribute *attr, char *buf)
1459{
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001460 struct hstate *h;
1461 unsigned long nr_huge_pages;
1462 int nid;
1463
1464 h = kobj_to_hstate(kobj, &nid);
1465 if (nid == NUMA_NO_NODE)
1466 nr_huge_pages = h->nr_huge_pages;
1467 else
1468 nr_huge_pages = h->nr_huge_pages_node[nid];
1469
1470 return sprintf(buf, "%lu\n", nr_huge_pages);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001471}
Eric B Munsonadbe8722011-01-13 15:47:27 -08001472
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001473static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
1474 struct kobject *kobj, struct kobj_attribute *attr,
1475 const char *buf, size_t len)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001476{
1477 int err;
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001478 int nid;
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001479 unsigned long count;
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001480 struct hstate *h;
David Rientjesbad44b52009-12-14 17:58:38 -08001481 NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001482
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001483 err = strict_strtoul(buf, 10, &count);
Eric B Munson73ae31e2011-01-13 15:47:28 -08001484 if (err)
Eric B Munsonadbe8722011-01-13 15:47:27 -08001485 goto out;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001486
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001487 h = kobj_to_hstate(kobj, &nid);
Eric B Munsonadbe8722011-01-13 15:47:27 -08001488 if (h->order >= MAX_ORDER) {
1489 err = -EINVAL;
1490 goto out;
1491 }
1492
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001493 if (nid == NUMA_NO_NODE) {
1494 /*
1495 * global hstate attribute
1496 */
1497 if (!(obey_mempolicy &&
1498 init_nodemask_of_mempolicy(nodes_allowed))) {
1499 NODEMASK_FREE(nodes_allowed);
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08001500 nodes_allowed = &node_states[N_MEMORY];
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001501 }
1502 } else if (nodes_allowed) {
1503 /*
1504 * per node hstate attribute: adjust count to global,
1505 * but restrict alloc/free to the specified node.
1506 */
1507 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
1508 init_nodemask_of_node(nodes_allowed, nid);
1509 } else
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08001510 nodes_allowed = &node_states[N_MEMORY];
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001511
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001512 h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001513
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08001514 if (nodes_allowed != &node_states[N_MEMORY])
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001515 NODEMASK_FREE(nodes_allowed);
1516
1517 return len;
Eric B Munsonadbe8722011-01-13 15:47:27 -08001518out:
1519 NODEMASK_FREE(nodes_allowed);
1520 return err;
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001521}
1522
1523static ssize_t nr_hugepages_show(struct kobject *kobj,
1524 struct kobj_attribute *attr, char *buf)
1525{
1526 return nr_hugepages_show_common(kobj, attr, buf);
1527}
1528
1529static ssize_t nr_hugepages_store(struct kobject *kobj,
1530 struct kobj_attribute *attr, const char *buf, size_t len)
1531{
1532 return nr_hugepages_store_common(false, kobj, attr, buf, len);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001533}
1534HSTATE_ATTR(nr_hugepages);
1535
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001536#ifdef CONFIG_NUMA
1537
1538/*
1539 * hstate attribute for optionally mempolicy-based constraint on persistent
1540 * huge page alloc/free.
1541 */
1542static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
1543 struct kobj_attribute *attr, char *buf)
1544{
1545 return nr_hugepages_show_common(kobj, attr, buf);
1546}
1547
1548static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
1549 struct kobj_attribute *attr, const char *buf, size_t len)
1550{
1551 return nr_hugepages_store_common(true, kobj, attr, buf, len);
1552}
1553HSTATE_ATTR(nr_hugepages_mempolicy);
1554#endif
1555
1556
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001557static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
1558 struct kobj_attribute *attr, char *buf)
1559{
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001560 struct hstate *h = kobj_to_hstate(kobj, NULL);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001561 return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
1562}
Eric B Munsonadbe8722011-01-13 15:47:27 -08001563
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001564static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
1565 struct kobj_attribute *attr, const char *buf, size_t count)
1566{
1567 int err;
1568 unsigned long input;
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001569 struct hstate *h = kobj_to_hstate(kobj, NULL);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001570
Eric B Munsonadbe8722011-01-13 15:47:27 -08001571 if (h->order >= MAX_ORDER)
1572 return -EINVAL;
1573
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001574 err = strict_strtoul(buf, 10, &input);
1575 if (err)
Eric B Munson73ae31e2011-01-13 15:47:28 -08001576 return err;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001577
1578 spin_lock(&hugetlb_lock);
1579 h->nr_overcommit_huge_pages = input;
1580 spin_unlock(&hugetlb_lock);
1581
1582 return count;
1583}
1584HSTATE_ATTR(nr_overcommit_hugepages);
1585
1586static ssize_t free_hugepages_show(struct kobject *kobj,
1587 struct kobj_attribute *attr, char *buf)
1588{
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001589 struct hstate *h;
1590 unsigned long free_huge_pages;
1591 int nid;
1592
1593 h = kobj_to_hstate(kobj, &nid);
1594 if (nid == NUMA_NO_NODE)
1595 free_huge_pages = h->free_huge_pages;
1596 else
1597 free_huge_pages = h->free_huge_pages_node[nid];
1598
1599 return sprintf(buf, "%lu\n", free_huge_pages);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001600}
1601HSTATE_ATTR_RO(free_hugepages);
1602
1603static ssize_t resv_hugepages_show(struct kobject *kobj,
1604 struct kobj_attribute *attr, char *buf)
1605{
Lee Schermerhorn9a3052302009-12-14 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->resv_huge_pages);
1608}
1609HSTATE_ATTR_RO(resv_hugepages);
1610
1611static ssize_t surplus_hugepages_show(struct kobject *kobj,
1612 struct kobj_attribute *attr, char *buf)
1613{
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001614 struct hstate *h;
1615 unsigned long surplus_huge_pages;
1616 int nid;
1617
1618 h = kobj_to_hstate(kobj, &nid);
1619 if (nid == NUMA_NO_NODE)
1620 surplus_huge_pages = h->surplus_huge_pages;
1621 else
1622 surplus_huge_pages = h->surplus_huge_pages_node[nid];
1623
1624 return sprintf(buf, "%lu\n", surplus_huge_pages);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001625}
1626HSTATE_ATTR_RO(surplus_hugepages);
1627
1628static struct attribute *hstate_attrs[] = {
1629 &nr_hugepages_attr.attr,
1630 &nr_overcommit_hugepages_attr.attr,
1631 &free_hugepages_attr.attr,
1632 &resv_hugepages_attr.attr,
1633 &surplus_hugepages_attr.attr,
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001634#ifdef CONFIG_NUMA
1635 &nr_hugepages_mempolicy_attr.attr,
1636#endif
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001637 NULL,
1638};
1639
1640static struct attribute_group hstate_attr_group = {
1641 .attrs = hstate_attrs,
1642};
1643
Jeff Mahoney094e9532010-02-02 13:44:14 -08001644static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
1645 struct kobject **hstate_kobjs,
1646 struct attribute_group *hstate_attr_group)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001647{
1648 int retval;
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07001649 int hi = hstate_index(h);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001650
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001651 hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
1652 if (!hstate_kobjs[hi])
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001653 return -ENOMEM;
1654
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001655 retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001656 if (retval)
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001657 kobject_put(hstate_kobjs[hi]);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001658
1659 return retval;
1660}
1661
1662static void __init hugetlb_sysfs_init(void)
1663{
1664 struct hstate *h;
1665 int err;
1666
1667 hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
1668 if (!hugepages_kobj)
1669 return;
1670
1671 for_each_hstate(h) {
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001672 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
1673 hstate_kobjs, &hstate_attr_group);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001674 if (err)
Andrew Mortonffb22af2013-02-22 16:32:08 -08001675 pr_err("Hugetlb: Unable to add hstate %s", h->name);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001676 }
1677}
1678
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001679#ifdef CONFIG_NUMA
1680
1681/*
1682 * node_hstate/s - associate per node hstate attributes, via their kobjects,
Kay Sievers10fbcf42011-12-21 14:48:43 -08001683 * with node devices in node_devices[] using a parallel array. The array
1684 * index of a node device or _hstate == node id.
1685 * This is here to avoid any static dependency of the node device driver, in
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001686 * the base kernel, on the hugetlb module.
1687 */
1688struct node_hstate {
1689 struct kobject *hugepages_kobj;
1690 struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
1691};
1692struct node_hstate node_hstates[MAX_NUMNODES];
1693
1694/*
Kay Sievers10fbcf42011-12-21 14:48:43 -08001695 * A subset of global hstate attributes for node devices
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001696 */
1697static struct attribute *per_node_hstate_attrs[] = {
1698 &nr_hugepages_attr.attr,
1699 &free_hugepages_attr.attr,
1700 &surplus_hugepages_attr.attr,
1701 NULL,
1702};
1703
1704static struct attribute_group per_node_hstate_attr_group = {
1705 .attrs = per_node_hstate_attrs,
1706};
1707
1708/*
Kay Sievers10fbcf42011-12-21 14:48:43 -08001709 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001710 * Returns node id via non-NULL nidp.
1711 */
1712static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
1713{
1714 int nid;
1715
1716 for (nid = 0; nid < nr_node_ids; nid++) {
1717 struct node_hstate *nhs = &node_hstates[nid];
1718 int i;
1719 for (i = 0; i < HUGE_MAX_HSTATE; i++)
1720 if (nhs->hstate_kobjs[i] == kobj) {
1721 if (nidp)
1722 *nidp = nid;
1723 return &hstates[i];
1724 }
1725 }
1726
1727 BUG();
1728 return NULL;
1729}
1730
1731/*
Kay Sievers10fbcf42011-12-21 14:48:43 -08001732 * Unregister hstate attributes from a single node device.
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001733 * No-op if no hstate attributes attached.
1734 */
Claudiu Ghioc3cd8b442013-03-04 12:46:15 +02001735static void hugetlb_unregister_node(struct node *node)
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001736{
1737 struct hstate *h;
Kay Sievers10fbcf42011-12-21 14:48:43 -08001738 struct node_hstate *nhs = &node_hstates[node->dev.id];
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001739
1740 if (!nhs->hugepages_kobj)
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -08001741 return; /* no hstate attributes */
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001742
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07001743 for_each_hstate(h) {
1744 int idx = hstate_index(h);
1745 if (nhs->hstate_kobjs[idx]) {
1746 kobject_put(nhs->hstate_kobjs[idx]);
1747 nhs->hstate_kobjs[idx] = NULL;
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001748 }
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07001749 }
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001750
1751 kobject_put(nhs->hugepages_kobj);
1752 nhs->hugepages_kobj = NULL;
1753}
1754
1755/*
Kay Sievers10fbcf42011-12-21 14:48:43 -08001756 * hugetlb module exit: unregister hstate attributes from node devices
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001757 * that have them.
1758 */
1759static void hugetlb_unregister_all_nodes(void)
1760{
1761 int nid;
1762
1763 /*
Kay Sievers10fbcf42011-12-21 14:48:43 -08001764 * disable node device registrations.
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001765 */
1766 register_hugetlbfs_with_node(NULL, NULL);
1767
1768 /*
1769 * remove hstate attributes from any nodes that have them.
1770 */
1771 for (nid = 0; nid < nr_node_ids; nid++)
Wen Congyang87327942012-12-11 16:00:56 -08001772 hugetlb_unregister_node(node_devices[nid]);
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001773}
1774
1775/*
Kay Sievers10fbcf42011-12-21 14:48:43 -08001776 * Register hstate attributes for a single node device.
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001777 * No-op if attributes already registered.
1778 */
Claudiu Ghioc3cd8b442013-03-04 12:46:15 +02001779static void hugetlb_register_node(struct node *node)
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001780{
1781 struct hstate *h;
Kay Sievers10fbcf42011-12-21 14:48:43 -08001782 struct node_hstate *nhs = &node_hstates[node->dev.id];
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001783 int err;
1784
1785 if (nhs->hugepages_kobj)
1786 return; /* already allocated */
1787
1788 nhs->hugepages_kobj = kobject_create_and_add("hugepages",
Kay Sievers10fbcf42011-12-21 14:48:43 -08001789 &node->dev.kobj);
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001790 if (!nhs->hugepages_kobj)
1791 return;
1792
1793 for_each_hstate(h) {
1794 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
1795 nhs->hstate_kobjs,
1796 &per_node_hstate_attr_group);
1797 if (err) {
Andrew Mortonffb22af2013-02-22 16:32:08 -08001798 pr_err("Hugetlb: Unable to add hstate %s for node %d\n",
1799 h->name, node->dev.id);
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001800 hugetlb_unregister_node(node);
1801 break;
1802 }
1803 }
1804}
1805
1806/*
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -08001807 * hugetlb init time: register hstate attributes for all registered node
Kay Sievers10fbcf42011-12-21 14:48:43 -08001808 * devices of nodes that have memory. All on-line nodes should have
1809 * registered their associated device by this time.
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001810 */
1811static void hugetlb_register_all_nodes(void)
1812{
1813 int nid;
1814
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08001815 for_each_node_state(nid, N_MEMORY) {
Wen Congyang87327942012-12-11 16:00:56 -08001816 struct node *node = node_devices[nid];
Kay Sievers10fbcf42011-12-21 14:48:43 -08001817 if (node->dev.id == nid)
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001818 hugetlb_register_node(node);
1819 }
1820
1821 /*
Kay Sievers10fbcf42011-12-21 14:48:43 -08001822 * Let the node device driver know we're here so it can
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001823 * [un]register hstate attributes on node hotplug.
1824 */
1825 register_hugetlbfs_with_node(hugetlb_register_node,
1826 hugetlb_unregister_node);
1827}
1828#else /* !CONFIG_NUMA */
1829
1830static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
1831{
1832 BUG();
1833 if (nidp)
1834 *nidp = -1;
1835 return NULL;
1836}
1837
1838static void hugetlb_unregister_all_nodes(void) { }
1839
1840static void hugetlb_register_all_nodes(void) { }
1841
1842#endif
1843
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001844static void __exit hugetlb_exit(void)
1845{
1846 struct hstate *h;
1847
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001848 hugetlb_unregister_all_nodes();
1849
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001850 for_each_hstate(h) {
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07001851 kobject_put(hstate_kobjs[hstate_index(h)]);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001852 }
1853
1854 kobject_put(hugepages_kobj);
1855}
1856module_exit(hugetlb_exit);
1857
1858static int __init hugetlb_init(void)
1859{
Benjamin Herrenschmidt0ef89d22008-07-31 00:07:30 -07001860 /* Some platform decide whether they support huge pages at boot
1861 * time. On these, such as powerpc, HPAGE_SHIFT is set to 0 when
1862 * there is no such support
1863 */
1864 if (HPAGE_SHIFT == 0)
1865 return 0;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001866
Nick Piggine11bfbf2008-07-23 21:27:52 -07001867 if (!size_to_hstate(default_hstate_size)) {
1868 default_hstate_size = HPAGE_SIZE;
1869 if (!size_to_hstate(default_hstate_size))
1870 hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001871 }
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07001872 default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size));
Nick Piggine11bfbf2008-07-23 21:27:52 -07001873 if (default_hstate_max_huge_pages)
1874 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001875
1876 hugetlb_init_hstates();
Andi Kleenaa888a72008-07-23 21:27:47 -07001877 gather_bootmem_prealloc();
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001878 report_hugepages();
1879
1880 hugetlb_sysfs_init();
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001881 hugetlb_register_all_nodes();
Jianguo Wu7179e7b2012-12-18 14:23:19 -08001882 hugetlb_cgroup_file_init();
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001883
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001884 return 0;
1885}
1886module_init(hugetlb_init);
1887
1888/* Should be called on processing a hugepagesz=... option */
1889void __init hugetlb_add_hstate(unsigned order)
1890{
1891 struct hstate *h;
Andi Kleen8faa8b02008-07-23 21:27:48 -07001892 unsigned long i;
1893
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001894 if (size_to_hstate(PAGE_SIZE << order)) {
Andrew Mortonffb22af2013-02-22 16:32:08 -08001895 pr_warning("hugepagesz= specified twice, ignoring\n");
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001896 return;
1897 }
Aneesh Kumar K.V47d38342012-07-31 16:41:54 -07001898 BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001899 BUG_ON(order == 0);
Aneesh Kumar K.V47d38342012-07-31 16:41:54 -07001900 h = &hstates[hugetlb_max_hstate++];
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001901 h->order = order;
1902 h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
Andi Kleen8faa8b02008-07-23 21:27:48 -07001903 h->nr_huge_pages = 0;
1904 h->free_huge_pages = 0;
1905 for (i = 0; i < MAX_NUMNODES; ++i)
1906 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
Aneesh Kumar K.V0edaecf2012-07-31 16:42:07 -07001907 INIT_LIST_HEAD(&h->hugepage_activelist);
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08001908 h->next_nid_to_alloc = first_node(node_states[N_MEMORY]);
1909 h->next_nid_to_free = first_node(node_states[N_MEMORY]);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001910 snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
1911 huge_page_size(h)/1024);
Andi Kleen8faa8b02008-07-23 21:27:48 -07001912
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001913 parsed_hstate = h;
1914}
1915
Nick Piggine11bfbf2008-07-23 21:27:52 -07001916static int __init hugetlb_nrpages_setup(char *s)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001917{
1918 unsigned long *mhp;
Andi Kleen8faa8b02008-07-23 21:27:48 -07001919 static unsigned long *last_mhp;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001920
1921 /*
Aneesh Kumar K.V47d38342012-07-31 16:41:54 -07001922 * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet,
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001923 * so this hugepages= parameter goes to the "default hstate".
1924 */
Aneesh Kumar K.V47d38342012-07-31 16:41:54 -07001925 if (!hugetlb_max_hstate)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001926 mhp = &default_hstate_max_huge_pages;
1927 else
1928 mhp = &parsed_hstate->max_huge_pages;
1929
Andi Kleen8faa8b02008-07-23 21:27:48 -07001930 if (mhp == last_mhp) {
Andrew Mortonffb22af2013-02-22 16:32:08 -08001931 pr_warning("hugepages= specified twice without "
1932 "interleaving hugepagesz=, ignoring\n");
Andi Kleen8faa8b02008-07-23 21:27:48 -07001933 return 1;
1934 }
1935
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001936 if (sscanf(s, "%lu", mhp) <= 0)
1937 *mhp = 0;
1938
Andi Kleen8faa8b02008-07-23 21:27:48 -07001939 /*
1940 * Global state is always initialized later in hugetlb_init.
1941 * But we need to allocate >= MAX_ORDER hstates here early to still
1942 * use the bootmem allocator.
1943 */
Aneesh Kumar K.V47d38342012-07-31 16:41:54 -07001944 if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER)
Andi Kleen8faa8b02008-07-23 21:27:48 -07001945 hugetlb_hstate_alloc_pages(parsed_hstate);
1946
1947 last_mhp = mhp;
1948
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001949 return 1;
1950}
Nick Piggine11bfbf2008-07-23 21:27:52 -07001951__setup("hugepages=", hugetlb_nrpages_setup);
1952
1953static int __init hugetlb_default_setup(char *s)
1954{
1955 default_hstate_size = memparse(s, &s);
1956 return 1;
1957}
1958__setup("default_hugepagesz=", hugetlb_default_setup);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001959
Nishanth Aravamudan8a213462008-07-25 19:44:37 -07001960static unsigned int cpuset_mems_nr(unsigned int *array)
1961{
1962 int node;
1963 unsigned int nr = 0;
1964
1965 for_each_node_mask(node, cpuset_current_mems_allowed)
1966 nr += array[node];
1967
1968 return nr;
1969}
1970
1971#ifdef CONFIG_SYSCTL
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001972static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
1973 struct ctl_table *table, int write,
1974 void __user *buffer, size_t *length, loff_t *ppos)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001975{
Andi Kleene5ff2152008-07-23 21:27:42 -07001976 struct hstate *h = &default_hstate;
1977 unsigned long tmp;
Michal Hocko08d4a242011-01-13 15:47:26 -08001978 int ret;
Andi Kleene5ff2152008-07-23 21:27:42 -07001979
Petr Holasekc033a932011-03-22 16:33:05 -07001980 tmp = h->max_huge_pages;
Andi Kleene5ff2152008-07-23 21:27:42 -07001981
Eric B Munsonadbe8722011-01-13 15:47:27 -08001982 if (write && h->order >= MAX_ORDER)
1983 return -EINVAL;
1984
Andi Kleene5ff2152008-07-23 21:27:42 -07001985 table->data = &tmp;
1986 table->maxlen = sizeof(unsigned long);
Michal Hocko08d4a242011-01-13 15:47:26 -08001987 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
1988 if (ret)
1989 goto out;
Andi Kleene5ff2152008-07-23 21:27:42 -07001990
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001991 if (write) {
David Rientjesbad44b52009-12-14 17:58:38 -08001992 NODEMASK_ALLOC(nodemask_t, nodes_allowed,
1993 GFP_KERNEL | __GFP_NORETRY);
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001994 if (!(obey_mempolicy &&
1995 init_nodemask_of_mempolicy(nodes_allowed))) {
1996 NODEMASK_FREE(nodes_allowed);
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08001997 nodes_allowed = &node_states[N_MEMORY];
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001998 }
1999 h->max_huge_pages = set_max_huge_pages(h, tmp, nodes_allowed);
2000
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08002001 if (nodes_allowed != &node_states[N_MEMORY])
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002002 NODEMASK_FREE(nodes_allowed);
2003 }
Michal Hocko08d4a242011-01-13 15:47:26 -08002004out:
2005 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002006}
Mel Gorman396faf02007-07-17 04:03:13 -07002007
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002008int hugetlb_sysctl_handler(struct ctl_table *table, int write,
2009 void __user *buffer, size_t *length, loff_t *ppos)
2010{
2011
2012 return hugetlb_sysctl_handler_common(false, table, write,
2013 buffer, length, ppos);
2014}
2015
2016#ifdef CONFIG_NUMA
2017int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
2018 void __user *buffer, size_t *length, loff_t *ppos)
2019{
2020 return hugetlb_sysctl_handler_common(true, table, write,
2021 buffer, length, ppos);
2022}
2023#endif /* CONFIG_NUMA */
2024
Mel Gorman396faf02007-07-17 04:03:13 -07002025int hugetlb_treat_movable_handler(struct ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07002026 void __user *buffer,
Mel Gorman396faf02007-07-17 04:03:13 -07002027 size_t *length, loff_t *ppos)
2028{
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07002029 proc_dointvec(table, write, buffer, length, ppos);
Mel Gorman396faf02007-07-17 04:03:13 -07002030 if (hugepages_treat_as_movable)
2031 htlb_alloc_mask = GFP_HIGHUSER_MOVABLE;
2032 else
2033 htlb_alloc_mask = GFP_HIGHUSER;
2034 return 0;
2035}
2036
Nishanth Aravamudana3d0c6a2008-02-08 04:18:18 -08002037int hugetlb_overcommit_handler(struct ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07002038 void __user *buffer,
Nishanth Aravamudana3d0c6a2008-02-08 04:18:18 -08002039 size_t *length, loff_t *ppos)
2040{
Andi Kleena5516432008-07-23 21:27:41 -07002041 struct hstate *h = &default_hstate;
Andi Kleene5ff2152008-07-23 21:27:42 -07002042 unsigned long tmp;
Michal Hocko08d4a242011-01-13 15:47:26 -08002043 int ret;
Andi Kleene5ff2152008-07-23 21:27:42 -07002044
Petr Holasekc033a932011-03-22 16:33:05 -07002045 tmp = h->nr_overcommit_huge_pages;
Andi Kleene5ff2152008-07-23 21:27:42 -07002046
Eric B Munsonadbe8722011-01-13 15:47:27 -08002047 if (write && h->order >= MAX_ORDER)
2048 return -EINVAL;
2049
Andi Kleene5ff2152008-07-23 21:27:42 -07002050 table->data = &tmp;
2051 table->maxlen = sizeof(unsigned long);
Michal Hocko08d4a242011-01-13 15:47:26 -08002052 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2053 if (ret)
2054 goto out;
Andi Kleene5ff2152008-07-23 21:27:42 -07002055
2056 if (write) {
2057 spin_lock(&hugetlb_lock);
2058 h->nr_overcommit_huge_pages = tmp;
2059 spin_unlock(&hugetlb_lock);
2060 }
Michal Hocko08d4a242011-01-13 15:47:26 -08002061out:
2062 return ret;
Nishanth Aravamudana3d0c6a2008-02-08 04:18:18 -08002063}
2064
Linus Torvalds1da177e2005-04-16 15:20:36 -07002065#endif /* CONFIG_SYSCTL */
2066
Alexey Dobriyane1759c22008-10-15 23:50:22 +04002067void hugetlb_report_meminfo(struct seq_file *m)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002068{
Andi Kleena5516432008-07-23 21:27:41 -07002069 struct hstate *h = &default_hstate;
Alexey Dobriyane1759c22008-10-15 23:50:22 +04002070 seq_printf(m,
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002071 "HugePages_Total: %5lu\n"
2072 "HugePages_Free: %5lu\n"
2073 "HugePages_Rsvd: %5lu\n"
2074 "HugePages_Surp: %5lu\n"
2075 "Hugepagesize: %8lu kB\n",
Andi Kleena5516432008-07-23 21:27:41 -07002076 h->nr_huge_pages,
2077 h->free_huge_pages,
2078 h->resv_huge_pages,
2079 h->surplus_huge_pages,
2080 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002081}
2082
2083int hugetlb_report_node_meminfo(int nid, char *buf)
2084{
Andi Kleena5516432008-07-23 21:27:41 -07002085 struct hstate *h = &default_hstate;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002086 return sprintf(buf,
2087 "Node %d HugePages_Total: %5u\n"
Nishanth Aravamudana1de0912008-03-26 14:37:53 -07002088 "Node %d HugePages_Free: %5u\n"
2089 "Node %d HugePages_Surp: %5u\n",
Andi Kleena5516432008-07-23 21:27:41 -07002090 nid, h->nr_huge_pages_node[nid],
2091 nid, h->free_huge_pages_node[nid],
2092 nid, h->surplus_huge_pages_node[nid]);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002093}
2094
David Rientjes949f7ec2013-04-29 15:07:48 -07002095void hugetlb_show_meminfo(void)
2096{
2097 struct hstate *h;
2098 int nid;
2099
2100 for_each_node_state(nid, N_MEMORY)
2101 for_each_hstate(h)
2102 pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
2103 nid,
2104 h->nr_huge_pages_node[nid],
2105 h->free_huge_pages_node[nid],
2106 h->surplus_huge_pages_node[nid],
2107 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
2108}
2109
Linus Torvalds1da177e2005-04-16 15:20:36 -07002110/* Return the number pages of memory we physically have, in PAGE_SIZE units. */
2111unsigned long hugetlb_total_pages(void)
2112{
Wanpeng Lid0028582013-03-22 15:04:40 -07002113 struct hstate *h;
2114 unsigned long nr_total_pages = 0;
2115
2116 for_each_hstate(h)
2117 nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h);
2118 return nr_total_pages;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002119}
Linus Torvalds1da177e2005-04-16 15:20:36 -07002120
Andi Kleena5516432008-07-23 21:27:41 -07002121static int hugetlb_acct_memory(struct hstate *h, long delta)
Mel Gormanfc1b8a72008-07-23 21:27:22 -07002122{
2123 int ret = -ENOMEM;
2124
2125 spin_lock(&hugetlb_lock);
2126 /*
2127 * When cpuset is configured, it breaks the strict hugetlb page
2128 * reservation as the accounting is done on a global variable. Such
2129 * reservation is completely rubbish in the presence of cpuset because
2130 * the reservation is not checked against page availability for the
2131 * current cpuset. Application can still potentially OOM'ed by kernel
2132 * with lack of free htlb page in cpuset that the task is in.
2133 * Attempt to enforce strict accounting with cpuset is almost
2134 * impossible (or too ugly) because cpuset is too fluid that
2135 * task or memory node can be dynamically moved between cpusets.
2136 *
2137 * The change of semantics for shared hugetlb mapping with cpuset is
2138 * undesirable. However, in order to preserve some of the semantics,
2139 * we fall back to check against current free page availability as
2140 * a best attempt and hopefully to minimize the impact of changing
2141 * semantics that cpuset has.
2142 */
2143 if (delta > 0) {
Andi Kleena5516432008-07-23 21:27:41 -07002144 if (gather_surplus_pages(h, delta) < 0)
Mel Gormanfc1b8a72008-07-23 21:27:22 -07002145 goto out;
2146
Andi Kleena5516432008-07-23 21:27:41 -07002147 if (delta > cpuset_mems_nr(h->free_huge_pages_node)) {
2148 return_unused_surplus_pages(h, delta);
Mel Gormanfc1b8a72008-07-23 21:27:22 -07002149 goto out;
2150 }
2151 }
2152
2153 ret = 0;
2154 if (delta < 0)
Andi Kleena5516432008-07-23 21:27:41 -07002155 return_unused_surplus_pages(h, (unsigned long) -delta);
Mel Gormanfc1b8a72008-07-23 21:27:22 -07002156
2157out:
2158 spin_unlock(&hugetlb_lock);
2159 return ret;
2160}
2161
Andy Whitcroft84afd992008-07-23 21:27:32 -07002162static void hugetlb_vm_op_open(struct vm_area_struct *vma)
2163{
2164 struct resv_map *reservations = vma_resv_map(vma);
2165
2166 /*
2167 * This new VMA should share its siblings reservation map if present.
2168 * The VMA will only ever have a valid reservation map pointer where
2169 * it is being copied for another still existing VMA. As that VMA
Lucas De Marchi25985ed2011-03-30 22:57:33 -03002170 * has a reference to the reservation map it cannot disappear until
Andy Whitcroft84afd992008-07-23 21:27:32 -07002171 * after this open call completes. It is therefore safe to take a
2172 * new reference here without additional locking.
2173 */
2174 if (reservations)
2175 kref_get(&reservations->refs);
2176}
2177
Dave Hansenc50ac052012-05-29 15:06:46 -07002178static void resv_map_put(struct vm_area_struct *vma)
2179{
2180 struct resv_map *reservations = vma_resv_map(vma);
2181
2182 if (!reservations)
2183 return;
2184 kref_put(&reservations->refs, resv_map_release);
2185}
2186
Mel Gormana1e78772008-07-23 21:27:23 -07002187static void hugetlb_vm_op_close(struct vm_area_struct *vma)
2188{
Andi Kleena5516432008-07-23 21:27:41 -07002189 struct hstate *h = hstate_vma(vma);
Andy Whitcroft84afd992008-07-23 21:27:32 -07002190 struct resv_map *reservations = vma_resv_map(vma);
David Gibson90481622012-03-21 16:34:12 -07002191 struct hugepage_subpool *spool = subpool_vma(vma);
Andy Whitcroft84afd992008-07-23 21:27:32 -07002192 unsigned long reserve;
2193 unsigned long start;
2194 unsigned long end;
2195
2196 if (reservations) {
Andi Kleena5516432008-07-23 21:27:41 -07002197 start = vma_hugecache_offset(h, vma, vma->vm_start);
2198 end = vma_hugecache_offset(h, vma, vma->vm_end);
Andy Whitcroft84afd992008-07-23 21:27:32 -07002199
2200 reserve = (end - start) -
2201 region_count(&reservations->regions, start, end);
2202
Dave Hansenc50ac052012-05-29 15:06:46 -07002203 resv_map_put(vma);
Andy Whitcroft84afd992008-07-23 21:27:32 -07002204
Adam Litke7251ff72008-07-23 21:27:59 -07002205 if (reserve) {
Andi Kleena5516432008-07-23 21:27:41 -07002206 hugetlb_acct_memory(h, -reserve);
David Gibson90481622012-03-21 16:34:12 -07002207 hugepage_subpool_put_pages(spool, reserve);
Adam Litke7251ff72008-07-23 21:27:59 -07002208 }
Andy Whitcroft84afd992008-07-23 21:27:32 -07002209 }
Mel Gormana1e78772008-07-23 21:27:23 -07002210}
2211
Linus Torvalds1da177e2005-04-16 15:20:36 -07002212/*
2213 * We cannot handle pagefaults against hugetlb pages at all. They cause
2214 * handle_mm_fault() to try to instantiate regular-sized pages in the
2215 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
2216 * this far.
2217 */
Nick Piggind0217ac2007-07-19 01:47:03 -07002218static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002219{
2220 BUG();
Nick Piggind0217ac2007-07-19 01:47:03 -07002221 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002222}
2223
Alexey Dobriyanf0f37e22009-09-27 22:29:37 +04002224const struct vm_operations_struct hugetlb_vm_ops = {
Nick Piggind0217ac2007-07-19 01:47:03 -07002225 .fault = hugetlb_vm_op_fault,
Andy Whitcroft84afd992008-07-23 21:27:32 -07002226 .open = hugetlb_vm_op_open,
Mel Gormana1e78772008-07-23 21:27:23 -07002227 .close = hugetlb_vm_op_close,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002228};
2229
David Gibson1e8f8892006-01-06 00:10:44 -08002230static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
2231 int writable)
David Gibson63551ae2005-06-21 17:14:44 -07002232{
2233 pte_t entry;
2234
David Gibson1e8f8892006-01-06 00:10:44 -08002235 if (writable) {
Gerald Schaefer106c9922013-04-29 15:07:23 -07002236 entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page,
2237 vma->vm_page_prot)));
David Gibson63551ae2005-06-21 17:14:44 -07002238 } else {
Gerald Schaefer106c9922013-04-29 15:07:23 -07002239 entry = huge_pte_wrprotect(mk_huge_pte(page,
2240 vma->vm_page_prot));
David Gibson63551ae2005-06-21 17:14:44 -07002241 }
2242 entry = pte_mkyoung(entry);
2243 entry = pte_mkhuge(entry);
Chris Metcalfd9ed9fa2012-04-01 14:01:34 -04002244 entry = arch_make_huge_pte(entry, vma, page, writable);
David Gibson63551ae2005-06-21 17:14:44 -07002245
2246 return entry;
2247}
2248
David Gibson1e8f8892006-01-06 00:10:44 -08002249static void set_huge_ptep_writable(struct vm_area_struct *vma,
2250 unsigned long address, pte_t *ptep)
2251{
2252 pte_t entry;
2253
Gerald Schaefer106c9922013-04-29 15:07:23 -07002254 entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep)));
Chris Forbes32f84522011-07-25 17:12:14 -07002255 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
Russell King4b3073e2009-12-18 16:40:18 +00002256 update_mmu_cache(vma, address, ptep);
David Gibson1e8f8892006-01-06 00:10:44 -08002257}
2258
2259
David Gibson63551ae2005-06-21 17:14:44 -07002260int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
2261 struct vm_area_struct *vma)
2262{
2263 pte_t *src_pte, *dst_pte, entry;
2264 struct page *ptepage;
Hugh Dickins1c598272005-10-19 21:23:43 -07002265 unsigned long addr;
David Gibson1e8f8892006-01-06 00:10:44 -08002266 int cow;
Andi Kleena5516432008-07-23 21:27:41 -07002267 struct hstate *h = hstate_vma(vma);
2268 unsigned long sz = huge_page_size(h);
David Gibson1e8f8892006-01-06 00:10:44 -08002269
2270 cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
David Gibson63551ae2005-06-21 17:14:44 -07002271
Andi Kleena5516432008-07-23 21:27:41 -07002272 for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
Hugh Dickinsc74df322005-10-29 18:16:23 -07002273 src_pte = huge_pte_offset(src, addr);
2274 if (!src_pte)
2275 continue;
Andi Kleena5516432008-07-23 21:27:41 -07002276 dst_pte = huge_pte_alloc(dst, addr, sz);
David Gibson63551ae2005-06-21 17:14:44 -07002277 if (!dst_pte)
2278 goto nomem;
Larry Woodmanc5c99422008-01-24 05:49:25 -08002279
2280 /* If the pagetables are shared don't copy or take references */
2281 if (dst_pte == src_pte)
2282 continue;
2283
Hugh Dickinsc74df322005-10-29 18:16:23 -07002284 spin_lock(&dst->page_table_lock);
Nick Piggin46478752008-06-05 22:45:57 -07002285 spin_lock_nested(&src->page_table_lock, SINGLE_DEPTH_NESTING);
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07002286 if (!huge_pte_none(huge_ptep_get(src_pte))) {
David Gibson1e8f8892006-01-06 00:10:44 -08002287 if (cow)
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07002288 huge_ptep_set_wrprotect(src, addr, src_pte);
2289 entry = huge_ptep_get(src_pte);
Hugh Dickins1c598272005-10-19 21:23:43 -07002290 ptepage = pte_page(entry);
2291 get_page(ptepage);
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002292 page_dup_rmap(ptepage);
Hugh Dickins1c598272005-10-19 21:23:43 -07002293 set_huge_pte_at(dst, addr, dst_pte, entry);
2294 }
2295 spin_unlock(&src->page_table_lock);
Hugh Dickinsc74df322005-10-29 18:16:23 -07002296 spin_unlock(&dst->page_table_lock);
David Gibson63551ae2005-06-21 17:14:44 -07002297 }
2298 return 0;
2299
2300nomem:
2301 return -ENOMEM;
2302}
2303
Naoya Horiguchi290408d2010-09-08 10:19:35 +09002304static int is_hugetlb_entry_migration(pte_t pte)
2305{
2306 swp_entry_t swp;
2307
2308 if (huge_pte_none(pte) || pte_present(pte))
2309 return 0;
2310 swp = pte_to_swp_entry(pte);
Chris Forbes32f84522011-07-25 17:12:14 -07002311 if (non_swap_entry(swp) && is_migration_entry(swp))
Naoya Horiguchi290408d2010-09-08 10:19:35 +09002312 return 1;
Chris Forbes32f84522011-07-25 17:12:14 -07002313 else
Naoya Horiguchi290408d2010-09-08 10:19:35 +09002314 return 0;
2315}
2316
Naoya Horiguchifd6a03e2010-05-28 09:29:21 +09002317static int is_hugetlb_entry_hwpoisoned(pte_t pte)
2318{
2319 swp_entry_t swp;
2320
2321 if (huge_pte_none(pte) || pte_present(pte))
2322 return 0;
2323 swp = pte_to_swp_entry(pte);
Chris Forbes32f84522011-07-25 17:12:14 -07002324 if (non_swap_entry(swp) && is_hwpoison_entry(swp))
Naoya Horiguchifd6a03e2010-05-28 09:29:21 +09002325 return 1;
Chris Forbes32f84522011-07-25 17:12:14 -07002326 else
Naoya Horiguchifd6a03e2010-05-28 09:29:21 +09002327 return 0;
2328}
2329
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07002330void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
2331 unsigned long start, unsigned long end,
2332 struct page *ref_page)
David Gibson63551ae2005-06-21 17:14:44 -07002333{
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07002334 int force_flush = 0;
David Gibson63551ae2005-06-21 17:14:44 -07002335 struct mm_struct *mm = vma->vm_mm;
2336 unsigned long address;
David Gibsonc7546f82005-08-05 11:59:35 -07002337 pte_t *ptep;
David Gibson63551ae2005-06-21 17:14:44 -07002338 pte_t pte;
2339 struct page *page;
Andi Kleena5516432008-07-23 21:27:41 -07002340 struct hstate *h = hstate_vma(vma);
2341 unsigned long sz = huge_page_size(h);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07002342 const unsigned long mmun_start = start; /* For mmu_notifiers */
2343 const unsigned long mmun_end = end; /* For mmu_notifiers */
Andi Kleena5516432008-07-23 21:27:41 -07002344
David Gibson63551ae2005-06-21 17:14:44 -07002345 WARN_ON(!is_vm_hugetlb_page(vma));
Andi Kleena5516432008-07-23 21:27:41 -07002346 BUG_ON(start & ~huge_page_mask(h));
2347 BUG_ON(end & ~huge_page_mask(h));
David Gibson63551ae2005-06-21 17:14:44 -07002348
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07002349 tlb_start_vma(tlb, vma);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07002350 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07002351again:
Hugh Dickins508034a2005-10-29 18:16:30 -07002352 spin_lock(&mm->page_table_lock);
Andi Kleena5516432008-07-23 21:27:41 -07002353 for (address = start; address < end; address += sz) {
David Gibsonc7546f82005-08-05 11:59:35 -07002354 ptep = huge_pte_offset(mm, address);
Adam Litke4c887262005-10-29 18:16:46 -07002355 if (!ptep)
David Gibsonc7546f82005-08-05 11:59:35 -07002356 continue;
2357
Chen, Kenneth W39dde652006-12-06 20:32:03 -08002358 if (huge_pmd_unshare(mm, &address, ptep))
2359 continue;
2360
Hillf Danton66293262012-03-23 15:01:48 -07002361 pte = huge_ptep_get(ptep);
2362 if (huge_pte_none(pte))
2363 continue;
2364
2365 /*
2366 * HWPoisoned hugepage is already unmapped and dropped reference
2367 */
Naoya Horiguchi8c4894c2012-12-12 13:52:28 -08002368 if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) {
Gerald Schaefer106c9922013-04-29 15:07:23 -07002369 huge_pte_clear(mm, address, ptep);
Hillf Danton66293262012-03-23 15:01:48 -07002370 continue;
Naoya Horiguchi8c4894c2012-12-12 13:52:28 -08002371 }
Hillf Danton66293262012-03-23 15:01:48 -07002372
2373 page = pte_page(pte);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002374 /*
2375 * If a reference page is supplied, it is because a specific
2376 * page is being unmapped, not a range. Ensure the page we
2377 * are about to unmap is the actual page of interest.
2378 */
2379 if (ref_page) {
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002380 if (page != ref_page)
2381 continue;
2382
2383 /*
2384 * Mark the VMA as having unmapped its page so that
2385 * future faults in this VMA will fail rather than
2386 * looking like data was lost
2387 */
2388 set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
2389 }
2390
David Gibsonc7546f82005-08-05 11:59:35 -07002391 pte = huge_ptep_get_and_clear(mm, address, ptep);
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07002392 tlb_remove_tlb_entry(tlb, ptep, address);
Gerald Schaefer106c9922013-04-29 15:07:23 -07002393 if (huge_pte_dirty(pte))
Ken Chen6649a382007-02-08 14:20:27 -08002394 set_page_dirty(page);
Hillf Danton9e811302012-03-21 16:34:03 -07002395
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07002396 page_remove_rmap(page);
2397 force_flush = !__tlb_remove_page(tlb, page);
2398 if (force_flush)
2399 break;
Hillf Danton9e811302012-03-21 16:34:03 -07002400 /* Bail out after unmapping reference page if supplied */
2401 if (ref_page)
2402 break;
David Gibson63551ae2005-06-21 17:14:44 -07002403 }
Al Virocd2934a2012-03-05 06:40:29 +00002404 spin_unlock(&mm->page_table_lock);
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07002405 /*
2406 * mmu_gather ran out of room to batch pages, we break out of
2407 * the PTE lock to avoid doing the potential expensive TLB invalidate
2408 * and page-free while holding it.
2409 */
2410 if (force_flush) {
2411 force_flush = 0;
2412 tlb_flush_mmu(tlb);
2413 if (address < end && !ref_page)
2414 goto again;
Chen, Kenneth Wfe1668a2006-10-04 02:15:24 -07002415 }
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07002416 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07002417 tlb_end_vma(tlb, vma);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002418}
David Gibson63551ae2005-06-21 17:14:44 -07002419
Mel Gormand8333522012-07-31 16:46:20 -07002420void __unmap_hugepage_range_final(struct mmu_gather *tlb,
2421 struct vm_area_struct *vma, unsigned long start,
2422 unsigned long end, struct page *ref_page)
2423{
2424 __unmap_hugepage_range(tlb, vma, start, end, ref_page);
2425
2426 /*
2427 * Clear this flag so that x86's huge_pmd_share page_table_shareable
2428 * test will fail on a vma being torn down, and not grab a page table
2429 * on its way out. We're lucky that the flag has such an appropriate
2430 * name, and can in fact be safely cleared here. We could clear it
2431 * before the __unmap_hugepage_range above, but all that's necessary
2432 * is to clear it before releasing the i_mmap_mutex. This works
2433 * because in the context this is called, the VMA is about to be
2434 * destroyed and the i_mmap_mutex is held.
2435 */
2436 vma->vm_flags &= ~VM_MAYSHARE;
2437}
2438
Chen, Kenneth W502717f2006-10-11 01:20:46 -07002439void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002440 unsigned long end, struct page *ref_page)
Chen, Kenneth W502717f2006-10-11 01:20:46 -07002441{
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07002442 struct mm_struct *mm;
2443 struct mmu_gather tlb;
2444
2445 mm = vma->vm_mm;
2446
2447 tlb_gather_mmu(&tlb, mm, 0);
2448 __unmap_hugepage_range(&tlb, vma, start, end, ref_page);
2449 tlb_finish_mmu(&tlb, start, end);
Chen, Kenneth W502717f2006-10-11 01:20:46 -07002450}
2451
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002452/*
2453 * This is called when the original mapper is failing to COW a MAP_PRIVATE
2454 * mappping it owns the reserve page for. The intention is to unmap the page
2455 * from other VMAs and let the children be SIGKILLed if they are faulting the
2456 * same region.
2457 */
Harvey Harrison2a4b3de2008-10-18 20:27:06 -07002458static int unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
2459 struct page *page, unsigned long address)
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002460{
Adam Litke75266742008-11-12 13:24:56 -08002461 struct hstate *h = hstate_vma(vma);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002462 struct vm_area_struct *iter_vma;
2463 struct address_space *mapping;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002464 pgoff_t pgoff;
2465
2466 /*
2467 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
2468 * from page cache lookup which is in HPAGE_SIZE units.
2469 */
Adam Litke75266742008-11-12 13:24:56 -08002470 address = address & huge_page_mask(h);
Michal Hocko36e4f202012-10-08 16:33:31 -07002471 pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) +
2472 vma->vm_pgoff;
Al Viro496ad9a2013-01-23 17:07:38 -05002473 mapping = file_inode(vma->vm_file)->i_mapping;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002474
Mel Gorman4eb2b1d2009-12-14 17:59:53 -08002475 /*
2476 * Take the mapping lock for the duration of the table walk. As
2477 * this mapping should be shared between all the VMAs,
2478 * __unmap_hugepage_range() is called as the lock is already held
2479 */
Peter Zijlstra3d48ae42011-05-24 17:12:06 -07002480 mutex_lock(&mapping->i_mmap_mutex);
Michel Lespinasse6b2dbba2012-10-08 16:31:25 -07002481 vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) {
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002482 /* Do not unmap the current VMA */
2483 if (iter_vma == vma)
2484 continue;
2485
2486 /*
2487 * Unmap the page from other VMAs without their own reserves.
2488 * They get marked to be SIGKILLed if they fault in these
2489 * areas. This is because a future no-page fault on this VMA
2490 * could insert a zeroed page instead of the data existing
2491 * from the time of fork. This would look like data corruption
2492 */
2493 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07002494 unmap_hugepage_range(iter_vma, address,
2495 address + huge_page_size(h), page);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002496 }
Peter Zijlstra3d48ae42011-05-24 17:12:06 -07002497 mutex_unlock(&mapping->i_mmap_mutex);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002498
2499 return 1;
2500}
2501
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002502/*
2503 * Hugetlb_cow() should be called with page lock of the original hugepage held.
Michal Hockoef009b22012-01-10 15:07:21 -08002504 * Called with hugetlb_instantiation_mutex held and pte_page locked so we
2505 * cannot race with other handlers or page migration.
2506 * Keep the pte_same checks anyway to make transition from the mutex easier.
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002507 */
David Gibson1e8f8892006-01-06 00:10:44 -08002508static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002509 unsigned long address, pte_t *ptep, pte_t pte,
2510 struct page *pagecache_page)
David Gibson1e8f8892006-01-06 00:10:44 -08002511{
Andi Kleena5516432008-07-23 21:27:41 -07002512 struct hstate *h = hstate_vma(vma);
David Gibson1e8f8892006-01-06 00:10:44 -08002513 struct page *old_page, *new_page;
David Gibson79ac6ba2006-03-22 00:08:51 -08002514 int avoidcopy;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002515 int outside_reserve = 0;
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07002516 unsigned long mmun_start; /* For mmu_notifiers */
2517 unsigned long mmun_end; /* For mmu_notifiers */
David Gibson1e8f8892006-01-06 00:10:44 -08002518
2519 old_page = pte_page(pte);
2520
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002521retry_avoidcopy:
David Gibson1e8f8892006-01-06 00:10:44 -08002522 /* If no-one else is actually using this page, avoid the copy
2523 * and just make the page writable */
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002524 avoidcopy = (page_mapcount(old_page) == 1);
David Gibson1e8f8892006-01-06 00:10:44 -08002525 if (avoidcopy) {
Naoya Horiguchi56c9cfb2010-09-10 13:23:04 +09002526 if (PageAnon(old_page))
2527 page_move_anon_rmap(old_page, vma, address);
David Gibson1e8f8892006-01-06 00:10:44 -08002528 set_huge_ptep_writable(vma, address, ptep);
Nick Piggin83c54072007-07-19 01:47:05 -07002529 return 0;
David Gibson1e8f8892006-01-06 00:10:44 -08002530 }
2531
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002532 /*
2533 * If the process that created a MAP_PRIVATE mapping is about to
2534 * perform a COW due to a shared page count, attempt to satisfy
2535 * the allocation without using the existing reserves. The pagecache
2536 * page is used to determine if the reserve at this address was
2537 * consumed or not. If reserves were used, a partial faulted mapping
2538 * at the time of fork() could consume its reserves on COW instead
2539 * of the full address range.
2540 */
Mel Gormanf83a2752009-05-28 14:34:40 -07002541 if (!(vma->vm_flags & VM_MAYSHARE) &&
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002542 is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
2543 old_page != pagecache_page)
2544 outside_reserve = 1;
2545
David Gibson1e8f8892006-01-06 00:10:44 -08002546 page_cache_get(old_page);
Larry Woodmanb76c8cf2009-12-14 17:59:37 -08002547
2548 /* Drop page_table_lock as buddy allocator may be called */
2549 spin_unlock(&mm->page_table_lock);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002550 new_page = alloc_huge_page(vma, address, outside_reserve);
David Gibson1e8f8892006-01-06 00:10:44 -08002551
Adam Litke2fc39ce2007-11-14 16:59:39 -08002552 if (IS_ERR(new_page)) {
Aneesh Kumar K.V76dcee72012-07-31 16:41:57 -07002553 long err = PTR_ERR(new_page);
David Gibson1e8f8892006-01-06 00:10:44 -08002554 page_cache_release(old_page);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002555
2556 /*
2557 * If a process owning a MAP_PRIVATE mapping fails to COW,
2558 * it is due to references held by a child and an insufficient
2559 * huge page pool. To guarantee the original mappers
2560 * reliability, unmap the page from child processes. The child
2561 * may get SIGKILLed if it later faults.
2562 */
2563 if (outside_reserve) {
2564 BUG_ON(huge_pte_none(pte));
2565 if (unmap_ref_private(mm, vma, old_page, address)) {
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002566 BUG_ON(huge_pte_none(pte));
Larry Woodmanb76c8cf2009-12-14 17:59:37 -08002567 spin_lock(&mm->page_table_lock);
Hillf Dantona734bcc2012-01-10 15:07:20 -08002568 ptep = huge_pte_offset(mm, address & huge_page_mask(h));
2569 if (likely(pte_same(huge_ptep_get(ptep), pte)))
2570 goto retry_avoidcopy;
2571 /*
2572 * race occurs while re-acquiring page_table_lock, and
2573 * our job is done.
2574 */
2575 return 0;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002576 }
2577 WARN_ON_ONCE(1);
2578 }
2579
Larry Woodmanb76c8cf2009-12-14 17:59:37 -08002580 /* Caller expects lock to be held */
2581 spin_lock(&mm->page_table_lock);
Aneesh Kumar K.V76dcee72012-07-31 16:41:57 -07002582 if (err == -ENOMEM)
2583 return VM_FAULT_OOM;
2584 else
2585 return VM_FAULT_SIGBUS;
David Gibson1e8f8892006-01-06 00:10:44 -08002586 }
2587
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002588 /*
2589 * When the original hugepage is shared one, it does not have
2590 * anon_vma prepared.
2591 */
Dean Nelson44e2aa92010-10-26 14:22:08 -07002592 if (unlikely(anon_vma_prepare(vma))) {
Hillf Dantonea4039a2011-11-15 14:36:12 -08002593 page_cache_release(new_page);
2594 page_cache_release(old_page);
Dean Nelson44e2aa92010-10-26 14:22:08 -07002595 /* Caller expects lock to be held */
2596 spin_lock(&mm->page_table_lock);
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002597 return VM_FAULT_OOM;
Dean Nelson44e2aa92010-10-26 14:22:08 -07002598 }
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002599
Andrea Arcangeli47ad8472011-01-13 15:46:47 -08002600 copy_user_huge_page(new_page, old_page, address, vma,
2601 pages_per_huge_page(h));
Nick Piggin0ed361d2008-02-04 22:29:34 -08002602 __SetPageUptodate(new_page);
David Gibson1e8f8892006-01-06 00:10:44 -08002603
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07002604 mmun_start = address & huge_page_mask(h);
2605 mmun_end = mmun_start + huge_page_size(h);
2606 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
Larry Woodmanb76c8cf2009-12-14 17:59:37 -08002607 /*
2608 * Retake the page_table_lock to check for racing updates
2609 * before the page tables are altered
2610 */
2611 spin_lock(&mm->page_table_lock);
Andi Kleena5516432008-07-23 21:27:41 -07002612 ptep = huge_pte_offset(mm, address & huge_page_mask(h));
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07002613 if (likely(pte_same(huge_ptep_get(ptep), pte))) {
David Gibson1e8f8892006-01-06 00:10:44 -08002614 /* Break COW */
Gerald Schaefer8fe627e2008-04-28 02:13:28 -07002615 huge_ptep_clear_flush(vma, address, ptep);
David Gibson1e8f8892006-01-06 00:10:44 -08002616 set_huge_pte_at(mm, address, ptep,
2617 make_huge_pte(vma, new_page, 1));
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002618 page_remove_rmap(old_page);
Naoya Horiguchicd67f0d2010-09-10 13:23:04 +09002619 hugepage_add_new_anon_rmap(new_page, vma, address);
David Gibson1e8f8892006-01-06 00:10:44 -08002620 /* Make the old page be freed below */
2621 new_page = old_page;
2622 }
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07002623 spin_unlock(&mm->page_table_lock);
2624 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2625 /* Caller expects lock to be held */
2626 spin_lock(&mm->page_table_lock);
David Gibson1e8f8892006-01-06 00:10:44 -08002627 page_cache_release(new_page);
2628 page_cache_release(old_page);
Nick Piggin83c54072007-07-19 01:47:05 -07002629 return 0;
David Gibson1e8f8892006-01-06 00:10:44 -08002630}
2631
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002632/* Return the pagecache page at a given address within a VMA */
Andi Kleena5516432008-07-23 21:27:41 -07002633static struct page *hugetlbfs_pagecache_page(struct hstate *h,
2634 struct vm_area_struct *vma, unsigned long address)
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002635{
2636 struct address_space *mapping;
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -07002637 pgoff_t idx;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002638
2639 mapping = vma->vm_file->f_mapping;
Andi Kleena5516432008-07-23 21:27:41 -07002640 idx = vma_hugecache_offset(h, vma, address);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002641
2642 return find_lock_page(mapping, idx);
2643}
2644
Hugh Dickins3ae77f42009-09-21 17:03:33 -07002645/*
2646 * Return whether there is a pagecache page to back given address within VMA.
2647 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
2648 */
2649static bool hugetlbfs_pagecache_present(struct hstate *h,
Hugh Dickins2a15efc2009-09-21 17:03:27 -07002650 struct vm_area_struct *vma, unsigned long address)
2651{
2652 struct address_space *mapping;
2653 pgoff_t idx;
2654 struct page *page;
2655
2656 mapping = vma->vm_file->f_mapping;
2657 idx = vma_hugecache_offset(h, vma, address);
2658
2659 page = find_get_page(mapping, idx);
2660 if (page)
2661 put_page(page);
2662 return page != NULL;
2663}
2664
Robert P. J. Daya1ed3dd2007-07-17 04:03:33 -07002665static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
Hugh Dickins788c7df2009-06-23 13:49:05 +01002666 unsigned long address, pte_t *ptep, unsigned int flags)
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01002667{
Andi Kleena5516432008-07-23 21:27:41 -07002668 struct hstate *h = hstate_vma(vma);
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01002669 int ret = VM_FAULT_SIGBUS;
Hillf Danton409eb8c2012-01-20 14:34:13 -08002670 int anon_rmap = 0;
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -07002671 pgoff_t idx;
Adam Litke4c887262005-10-29 18:16:46 -07002672 unsigned long size;
Adam Litke4c887262005-10-29 18:16:46 -07002673 struct page *page;
2674 struct address_space *mapping;
David Gibson1e8f8892006-01-06 00:10:44 -08002675 pte_t new_pte;
Adam Litke4c887262005-10-29 18:16:46 -07002676
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002677 /*
2678 * Currently, we are forced to kill the process in the event the
2679 * original mapper has unmapped pages from the child due to a failed
Lucas De Marchi25985ed2011-03-30 22:57:33 -03002680 * COW. Warn that such a situation has occurred as it may not be obvious
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002681 */
2682 if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
Andrew Mortonffb22af2013-02-22 16:32:08 -08002683 pr_warning("PID %d killed due to inadequate hugepage pool\n",
2684 current->pid);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002685 return ret;
2686 }
2687
Adam Litke4c887262005-10-29 18:16:46 -07002688 mapping = vma->vm_file->f_mapping;
Andi Kleena5516432008-07-23 21:27:41 -07002689 idx = vma_hugecache_offset(h, vma, address);
Adam Litke4c887262005-10-29 18:16:46 -07002690
2691 /*
2692 * Use page lock to guard against racing truncation
2693 * before we get page_table_lock.
2694 */
Christoph Lameter6bda6662006-01-06 00:10:49 -08002695retry:
2696 page = find_lock_page(mapping, idx);
2697 if (!page) {
Andi Kleena5516432008-07-23 21:27:41 -07002698 size = i_size_read(mapping->host) >> huge_page_shift(h);
Hugh Dickinsebed4bf2006-10-28 10:38:43 -07002699 if (idx >= size)
2700 goto out;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002701 page = alloc_huge_page(vma, address, 0);
Adam Litke2fc39ce2007-11-14 16:59:39 -08002702 if (IS_ERR(page)) {
Aneesh Kumar K.V76dcee72012-07-31 16:41:57 -07002703 ret = PTR_ERR(page);
2704 if (ret == -ENOMEM)
2705 ret = VM_FAULT_OOM;
2706 else
2707 ret = VM_FAULT_SIGBUS;
Christoph Lameter6bda6662006-01-06 00:10:49 -08002708 goto out;
2709 }
Andrea Arcangeli47ad8472011-01-13 15:46:47 -08002710 clear_huge_page(page, address, pages_per_huge_page(h));
Nick Piggin0ed361d2008-02-04 22:29:34 -08002711 __SetPageUptodate(page);
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01002712
Mel Gormanf83a2752009-05-28 14:34:40 -07002713 if (vma->vm_flags & VM_MAYSHARE) {
Christoph Lameter6bda6662006-01-06 00:10:49 -08002714 int err;
Ken Chen45c682a2007-11-14 16:59:44 -08002715 struct inode *inode = mapping->host;
Christoph Lameter6bda6662006-01-06 00:10:49 -08002716
2717 err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
2718 if (err) {
2719 put_page(page);
Christoph Lameter6bda6662006-01-06 00:10:49 -08002720 if (err == -EEXIST)
2721 goto retry;
2722 goto out;
2723 }
Ken Chen45c682a2007-11-14 16:59:44 -08002724
2725 spin_lock(&inode->i_lock);
Andi Kleena5516432008-07-23 21:27:41 -07002726 inode->i_blocks += blocks_per_huge_page(h);
Ken Chen45c682a2007-11-14 16:59:44 -08002727 spin_unlock(&inode->i_lock);
Mel Gorman23be7462010-04-23 13:17:56 -04002728 } else {
Christoph Lameter6bda6662006-01-06 00:10:49 -08002729 lock_page(page);
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002730 if (unlikely(anon_vma_prepare(vma))) {
2731 ret = VM_FAULT_OOM;
2732 goto backout_unlocked;
2733 }
Hillf Danton409eb8c2012-01-20 14:34:13 -08002734 anon_rmap = 1;
Mel Gorman23be7462010-04-23 13:17:56 -04002735 }
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002736 } else {
Naoya Horiguchi998b4382010-09-08 10:19:32 +09002737 /*
2738 * If memory error occurs between mmap() and fault, some process
2739 * don't have hwpoisoned swap entry for errored virtual address.
2740 * So we need to block hugepage fault by PG_hwpoison bit check.
2741 */
2742 if (unlikely(PageHWPoison(page))) {
Chris Forbes32f84522011-07-25 17:12:14 -07002743 ret = VM_FAULT_HWPOISON |
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07002744 VM_FAULT_SET_HINDEX(hstate_index(h));
Naoya Horiguchi998b4382010-09-08 10:19:32 +09002745 goto backout_unlocked;
2746 }
Christoph Lameter6bda6662006-01-06 00:10:49 -08002747 }
David Gibson1e8f8892006-01-06 00:10:44 -08002748
Andy Whitcroft57303d82008-08-12 15:08:47 -07002749 /*
2750 * If we are going to COW a private mapping later, we examine the
2751 * pending reservations for this page now. This will ensure that
2752 * any allocations necessary to record that reservation occur outside
2753 * the spinlock.
2754 */
Hugh Dickins788c7df2009-06-23 13:49:05 +01002755 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED))
Andy Whitcroft2b267362008-08-12 15:08:49 -07002756 if (vma_needs_reservation(h, vma, address) < 0) {
2757 ret = VM_FAULT_OOM;
2758 goto backout_unlocked;
2759 }
Andy Whitcroft57303d82008-08-12 15:08:47 -07002760
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01002761 spin_lock(&mm->page_table_lock);
Andi Kleena5516432008-07-23 21:27:41 -07002762 size = i_size_read(mapping->host) >> huge_page_shift(h);
Adam Litke4c887262005-10-29 18:16:46 -07002763 if (idx >= size)
2764 goto backout;
2765
Nick Piggin83c54072007-07-19 01:47:05 -07002766 ret = 0;
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07002767 if (!huge_pte_none(huge_ptep_get(ptep)))
Adam Litke4c887262005-10-29 18:16:46 -07002768 goto backout;
2769
Hillf Danton409eb8c2012-01-20 14:34:13 -08002770 if (anon_rmap)
2771 hugepage_add_new_anon_rmap(page, vma, address);
2772 else
2773 page_dup_rmap(page);
David Gibson1e8f8892006-01-06 00:10:44 -08002774 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
2775 && (vma->vm_flags & VM_SHARED)));
2776 set_huge_pte_at(mm, address, ptep, new_pte);
2777
Hugh Dickins788c7df2009-06-23 13:49:05 +01002778 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
David Gibson1e8f8892006-01-06 00:10:44 -08002779 /* Optimization, do the COW without a second fault */
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002780 ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page);
David Gibson1e8f8892006-01-06 00:10:44 -08002781 }
2782
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01002783 spin_unlock(&mm->page_table_lock);
Adam Litke4c887262005-10-29 18:16:46 -07002784 unlock_page(page);
2785out:
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01002786 return ret;
Adam Litke4c887262005-10-29 18:16:46 -07002787
2788backout:
2789 spin_unlock(&mm->page_table_lock);
Andy Whitcroft2b267362008-08-12 15:08:49 -07002790backout_unlocked:
Adam Litke4c887262005-10-29 18:16:46 -07002791 unlock_page(page);
2792 put_page(page);
2793 goto out;
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01002794}
2795
Adam Litke86e52162006-01-06 00:10:43 -08002796int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
Hugh Dickins788c7df2009-06-23 13:49:05 +01002797 unsigned long address, unsigned int flags)
Adam Litke86e52162006-01-06 00:10:43 -08002798{
2799 pte_t *ptep;
2800 pte_t entry;
David Gibson1e8f8892006-01-06 00:10:44 -08002801 int ret;
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002802 struct page *page = NULL;
Andy Whitcroft57303d82008-08-12 15:08:47 -07002803 struct page *pagecache_page = NULL;
David Gibson3935baa2006-03-22 00:08:53 -08002804 static DEFINE_MUTEX(hugetlb_instantiation_mutex);
Andi Kleena5516432008-07-23 21:27:41 -07002805 struct hstate *h = hstate_vma(vma);
Adam Litke86e52162006-01-06 00:10:43 -08002806
KAMEZAWA Hiroyuki1e16a5392012-01-10 15:07:22 -08002807 address &= huge_page_mask(h);
2808
Naoya Horiguchifd6a03e2010-05-28 09:29:21 +09002809 ptep = huge_pte_offset(mm, address);
2810 if (ptep) {
2811 entry = huge_ptep_get(ptep);
Naoya Horiguchi290408d2010-09-08 10:19:35 +09002812 if (unlikely(is_hugetlb_entry_migration(entry))) {
Naoya Horiguchi30dad302013-06-12 14:05:04 -07002813 migration_entry_wait_huge(mm, ptep);
Naoya Horiguchi290408d2010-09-08 10:19:35 +09002814 return 0;
2815 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
Chris Forbes32f84522011-07-25 17:12:14 -07002816 return VM_FAULT_HWPOISON_LARGE |
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07002817 VM_FAULT_SET_HINDEX(hstate_index(h));
Naoya Horiguchifd6a03e2010-05-28 09:29:21 +09002818 }
2819
Andi Kleena5516432008-07-23 21:27:41 -07002820 ptep = huge_pte_alloc(mm, address, huge_page_size(h));
Adam Litke86e52162006-01-06 00:10:43 -08002821 if (!ptep)
2822 return VM_FAULT_OOM;
2823
David Gibson3935baa2006-03-22 00:08:53 -08002824 /*
2825 * Serialize hugepage allocation and instantiation, so that we don't
2826 * get spurious allocation failures if two CPUs race to instantiate
2827 * the same page in the page cache.
2828 */
2829 mutex_lock(&hugetlb_instantiation_mutex);
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07002830 entry = huge_ptep_get(ptep);
2831 if (huge_pte_none(entry)) {
Hugh Dickins788c7df2009-06-23 13:49:05 +01002832 ret = hugetlb_no_page(mm, vma, address, ptep, flags);
David Gibsonb4d1d992008-10-15 22:01:11 -07002833 goto out_mutex;
David Gibson3935baa2006-03-22 00:08:53 -08002834 }
Adam Litke86e52162006-01-06 00:10:43 -08002835
Nick Piggin83c54072007-07-19 01:47:05 -07002836 ret = 0;
David Gibson1e8f8892006-01-06 00:10:44 -08002837
Andy Whitcroft57303d82008-08-12 15:08:47 -07002838 /*
2839 * If we are going to COW the mapping later, we examine the pending
2840 * reservations for this page now. This will ensure that any
2841 * allocations necessary to record that reservation occur outside the
2842 * spinlock. For private mappings, we also lookup the pagecache
2843 * page now as it is used to determine if a reservation has been
2844 * consumed.
2845 */
Gerald Schaefer106c9922013-04-29 15:07:23 -07002846 if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) {
Andy Whitcroft2b267362008-08-12 15:08:49 -07002847 if (vma_needs_reservation(h, vma, address) < 0) {
2848 ret = VM_FAULT_OOM;
David Gibsonb4d1d992008-10-15 22:01:11 -07002849 goto out_mutex;
Andy Whitcroft2b267362008-08-12 15:08:49 -07002850 }
Andy Whitcroft57303d82008-08-12 15:08:47 -07002851
Mel Gormanf83a2752009-05-28 14:34:40 -07002852 if (!(vma->vm_flags & VM_MAYSHARE))
Andy Whitcroft57303d82008-08-12 15:08:47 -07002853 pagecache_page = hugetlbfs_pagecache_page(h,
2854 vma, address);
2855 }
2856
Naoya Horiguchi56c9cfb2010-09-10 13:23:04 +09002857 /*
2858 * hugetlb_cow() requires page locks of pte_page(entry) and
2859 * pagecache_page, so here we need take the former one
2860 * when page != pagecache_page or !pagecache_page.
2861 * Note that locking order is always pagecache_page -> page,
2862 * so no worry about deadlock.
2863 */
2864 page = pte_page(entry);
Chris Metcalf66aebce2012-04-12 12:49:15 -07002865 get_page(page);
Naoya Horiguchi56c9cfb2010-09-10 13:23:04 +09002866 if (page != pagecache_page)
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002867 lock_page(page);
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002868
David Gibson1e8f8892006-01-06 00:10:44 -08002869 spin_lock(&mm->page_table_lock);
2870 /* Check for a racing update before calling hugetlb_cow */
David Gibsonb4d1d992008-10-15 22:01:11 -07002871 if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
2872 goto out_page_table_lock;
2873
2874
Hugh Dickins788c7df2009-06-23 13:49:05 +01002875 if (flags & FAULT_FLAG_WRITE) {
Gerald Schaefer106c9922013-04-29 15:07:23 -07002876 if (!huge_pte_write(entry)) {
Andy Whitcroft57303d82008-08-12 15:08:47 -07002877 ret = hugetlb_cow(mm, vma, address, ptep, entry,
2878 pagecache_page);
David Gibsonb4d1d992008-10-15 22:01:11 -07002879 goto out_page_table_lock;
2880 }
Gerald Schaefer106c9922013-04-29 15:07:23 -07002881 entry = huge_pte_mkdirty(entry);
David Gibsonb4d1d992008-10-15 22:01:11 -07002882 }
2883 entry = pte_mkyoung(entry);
Hugh Dickins788c7df2009-06-23 13:49:05 +01002884 if (huge_ptep_set_access_flags(vma, address, ptep, entry,
2885 flags & FAULT_FLAG_WRITE))
Russell King4b3073e2009-12-18 16:40:18 +00002886 update_mmu_cache(vma, address, ptep);
David Gibsonb4d1d992008-10-15 22:01:11 -07002887
2888out_page_table_lock:
David Gibson1e8f8892006-01-06 00:10:44 -08002889 spin_unlock(&mm->page_table_lock);
Andy Whitcroft57303d82008-08-12 15:08:47 -07002890
2891 if (pagecache_page) {
2892 unlock_page(pagecache_page);
2893 put_page(pagecache_page);
2894 }
Dean Nelson1f64d692010-12-02 14:31:12 -08002895 if (page != pagecache_page)
2896 unlock_page(page);
Chris Metcalf66aebce2012-04-12 12:49:15 -07002897 put_page(page);
Andy Whitcroft57303d82008-08-12 15:08:47 -07002898
David Gibsonb4d1d992008-10-15 22:01:11 -07002899out_mutex:
David Gibson3935baa2006-03-22 00:08:53 -08002900 mutex_unlock(&hugetlb_instantiation_mutex);
David Gibson1e8f8892006-01-06 00:10:44 -08002901
2902 return ret;
Adam Litke86e52162006-01-06 00:10:43 -08002903}
2904
Andi Kleenceb86872008-07-23 21:27:50 -07002905/* Can be overriden by architectures */
2906__attribute__((weak)) struct page *
2907follow_huge_pud(struct mm_struct *mm, unsigned long address,
2908 pud_t *pud, int write)
2909{
2910 BUG();
2911 return NULL;
2912}
2913
Michel Lespinasse28a35712013-02-22 16:35:55 -08002914long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
2915 struct page **pages, struct vm_area_struct **vmas,
2916 unsigned long *position, unsigned long *nr_pages,
2917 long i, unsigned int flags)
David Gibson63551ae2005-06-21 17:14:44 -07002918{
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08002919 unsigned long pfn_offset;
2920 unsigned long vaddr = *position;
Michel Lespinasse28a35712013-02-22 16:35:55 -08002921 unsigned long remainder = *nr_pages;
Andi Kleena5516432008-07-23 21:27:41 -07002922 struct hstate *h = hstate_vma(vma);
David Gibson63551ae2005-06-21 17:14:44 -07002923
Hugh Dickins1c598272005-10-19 21:23:43 -07002924 spin_lock(&mm->page_table_lock);
David Gibson63551ae2005-06-21 17:14:44 -07002925 while (vaddr < vma->vm_end && remainder) {
Adam Litke4c887262005-10-29 18:16:46 -07002926 pte_t *pte;
Hugh Dickins2a15efc2009-09-21 17:03:27 -07002927 int absent;
Adam Litke4c887262005-10-29 18:16:46 -07002928 struct page *page;
2929
2930 /*
2931 * Some archs (sparc64, sh*) have multiple pte_ts to
Hugh Dickins2a15efc2009-09-21 17:03:27 -07002932 * each hugepage. We have to make sure we get the
Adam Litke4c887262005-10-29 18:16:46 -07002933 * first, for the page indexing below to work.
2934 */
Andi Kleena5516432008-07-23 21:27:41 -07002935 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h));
Hugh Dickins2a15efc2009-09-21 17:03:27 -07002936 absent = !pte || huge_pte_none(huge_ptep_get(pte));
Adam Litke4c887262005-10-29 18:16:46 -07002937
Hugh Dickins2a15efc2009-09-21 17:03:27 -07002938 /*
2939 * When coredumping, it suits get_dump_page if we just return
Hugh Dickins3ae77f42009-09-21 17:03:33 -07002940 * an error where there's an empty slot with no huge pagecache
2941 * to back it. This way, we avoid allocating a hugepage, and
2942 * the sparse dumpfile avoids allocating disk blocks, but its
2943 * huge holes still show up with zeroes where they need to be.
Hugh Dickins2a15efc2009-09-21 17:03:27 -07002944 */
Hugh Dickins3ae77f42009-09-21 17:03:33 -07002945 if (absent && (flags & FOLL_DUMP) &&
2946 !hugetlbfs_pagecache_present(h, vma, vaddr)) {
Hugh Dickins2a15efc2009-09-21 17:03:27 -07002947 remainder = 0;
2948 break;
2949 }
2950
Naoya Horiguchi9cc3a5b2013-04-17 15:58:30 -07002951 /*
2952 * We need call hugetlb_fault for both hugepages under migration
2953 * (in which case hugetlb_fault waits for the migration,) and
2954 * hwpoisoned hugepages (in which case we need to prevent the
2955 * caller from accessing to them.) In order to do this, we use
2956 * here is_swap_pte instead of is_hugetlb_entry_migration and
2957 * is_hugetlb_entry_hwpoisoned. This is because it simply covers
2958 * both cases, and because we can't follow correct pages
2959 * directly from any kind of swap entries.
2960 */
2961 if (absent || is_swap_pte(huge_ptep_get(pte)) ||
Gerald Schaefer106c9922013-04-29 15:07:23 -07002962 ((flags & FOLL_WRITE) &&
2963 !huge_pte_write(huge_ptep_get(pte)))) {
Adam Litke4c887262005-10-29 18:16:46 -07002964 int ret;
2965
2966 spin_unlock(&mm->page_table_lock);
Hugh Dickins2a15efc2009-09-21 17:03:27 -07002967 ret = hugetlb_fault(mm, vma, vaddr,
2968 (flags & FOLL_WRITE) ? FAULT_FLAG_WRITE : 0);
Adam Litke4c887262005-10-29 18:16:46 -07002969 spin_lock(&mm->page_table_lock);
Adam Litkea89182c2007-08-22 14:01:51 -07002970 if (!(ret & VM_FAULT_ERROR))
Adam Litke4c887262005-10-29 18:16:46 -07002971 continue;
2972
2973 remainder = 0;
Adam Litke4c887262005-10-29 18:16:46 -07002974 break;
2975 }
David Gibson63551ae2005-06-21 17:14:44 -07002976
Andi Kleena5516432008-07-23 21:27:41 -07002977 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07002978 page = pte_page(huge_ptep_get(pte));
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08002979same_page:
Chen, Kenneth Wd6692182006-03-31 02:29:57 -08002980 if (pages) {
Hugh Dickins2a15efc2009-09-21 17:03:27 -07002981 pages[i] = mem_map_offset(page, pfn_offset);
KOSAKI Motohiro4b2e38a2008-10-18 20:27:10 -07002982 get_page(pages[i]);
Chen, Kenneth Wd6692182006-03-31 02:29:57 -08002983 }
David Gibson63551ae2005-06-21 17:14:44 -07002984
2985 if (vmas)
2986 vmas[i] = vma;
2987
2988 vaddr += PAGE_SIZE;
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08002989 ++pfn_offset;
David Gibson63551ae2005-06-21 17:14:44 -07002990 --remainder;
2991 ++i;
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08002992 if (vaddr < vma->vm_end && remainder &&
Andi Kleena5516432008-07-23 21:27:41 -07002993 pfn_offset < pages_per_huge_page(h)) {
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08002994 /*
2995 * We use pfn_offset to avoid touching the pageframes
2996 * of this compound page.
2997 */
2998 goto same_page;
2999 }
David Gibson63551ae2005-06-21 17:14:44 -07003000 }
Hugh Dickins1c598272005-10-19 21:23:43 -07003001 spin_unlock(&mm->page_table_lock);
Michel Lespinasse28a35712013-02-22 16:35:55 -08003002 *nr_pages = remainder;
David Gibson63551ae2005-06-21 17:14:44 -07003003 *position = vaddr;
3004
Hugh Dickins2a15efc2009-09-21 17:03:27 -07003005 return i ? i : -EFAULT;
David Gibson63551ae2005-06-21 17:14:44 -07003006}
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003007
Peter Zijlstra7da4d642012-11-19 03:14:23 +01003008unsigned long hugetlb_change_protection(struct vm_area_struct *vma,
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003009 unsigned long address, unsigned long end, pgprot_t newprot)
3010{
3011 struct mm_struct *mm = vma->vm_mm;
3012 unsigned long start = address;
3013 pte_t *ptep;
3014 pte_t pte;
Andi Kleena5516432008-07-23 21:27:41 -07003015 struct hstate *h = hstate_vma(vma);
Peter Zijlstra7da4d642012-11-19 03:14:23 +01003016 unsigned long pages = 0;
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003017
3018 BUG_ON(address >= end);
3019 flush_cache_range(vma, address, end);
3020
Peter Zijlstra3d48ae42011-05-24 17:12:06 -07003021 mutex_lock(&vma->vm_file->f_mapping->i_mmap_mutex);
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003022 spin_lock(&mm->page_table_lock);
Andi Kleena5516432008-07-23 21:27:41 -07003023 for (; address < end; address += huge_page_size(h)) {
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003024 ptep = huge_pte_offset(mm, address);
3025 if (!ptep)
3026 continue;
Peter Zijlstra7da4d642012-11-19 03:14:23 +01003027 if (huge_pmd_unshare(mm, &address, ptep)) {
3028 pages++;
Chen, Kenneth W39dde652006-12-06 20:32:03 -08003029 continue;
Peter Zijlstra7da4d642012-11-19 03:14:23 +01003030 }
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07003031 if (!huge_pte_none(huge_ptep_get(ptep))) {
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003032 pte = huge_ptep_get_and_clear(mm, address, ptep);
Gerald Schaefer106c9922013-04-29 15:07:23 -07003033 pte = pte_mkhuge(huge_pte_modify(pte, newprot));
Tony Lube7517d2013-02-04 14:28:46 -08003034 pte = arch_make_huge_pte(pte, vma, NULL, 0);
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003035 set_huge_pte_at(mm, address, ptep, pte);
Peter Zijlstra7da4d642012-11-19 03:14:23 +01003036 pages++;
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003037 }
3038 }
3039 spin_unlock(&mm->page_table_lock);
Mel Gormand8333522012-07-31 16:46:20 -07003040 /*
3041 * Must flush TLB before releasing i_mmap_mutex: x86's huge_pmd_unshare
3042 * may have cleared our pud entry and done put_page on the page table:
3043 * once we release i_mmap_mutex, another task can do the final put_page
3044 * and that page table be reused and filled with junk.
3045 */
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003046 flush_tlb_range(vma, start, end);
Mel Gormand8333522012-07-31 16:46:20 -07003047 mutex_unlock(&vma->vm_file->f_mapping->i_mmap_mutex);
Peter Zijlstra7da4d642012-11-19 03:14:23 +01003048
3049 return pages << h->order;
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003050}
3051
Mel Gormana1e78772008-07-23 21:27:23 -07003052int hugetlb_reserve_pages(struct inode *inode,
3053 long from, long to,
Mel Gorman5a6fe122009-02-10 14:02:27 +00003054 struct vm_area_struct *vma,
KOSAKI Motohiroca16d142011-05-26 19:16:19 +09003055 vm_flags_t vm_flags)
Adam Litkee4e574b2007-10-16 01:26:19 -07003056{
Mel Gorman17c9d122009-02-11 16:34:16 +00003057 long ret, chg;
Andi Kleena5516432008-07-23 21:27:41 -07003058 struct hstate *h = hstate_inode(inode);
David Gibson90481622012-03-21 16:34:12 -07003059 struct hugepage_subpool *spool = subpool_inode(inode);
Adam Litkee4e574b2007-10-16 01:26:19 -07003060
Mel Gormana1e78772008-07-23 21:27:23 -07003061 /*
Mel Gorman17c9d122009-02-11 16:34:16 +00003062 * Only apply hugepage reservation if asked. At fault time, an
3063 * attempt will be made for VM_NORESERVE to allocate a page
David Gibson90481622012-03-21 16:34:12 -07003064 * without using reserves
Mel Gorman17c9d122009-02-11 16:34:16 +00003065 */
KOSAKI Motohiroca16d142011-05-26 19:16:19 +09003066 if (vm_flags & VM_NORESERVE)
Mel Gorman17c9d122009-02-11 16:34:16 +00003067 return 0;
3068
3069 /*
Mel Gormana1e78772008-07-23 21:27:23 -07003070 * Shared mappings base their reservation on the number of pages that
3071 * are already allocated on behalf of the file. Private mappings need
3072 * to reserve the full area even if read-only as mprotect() may be
3073 * called to make the mapping read-write. Assume !vma is a shm mapping
3074 */
Mel Gormanf83a2752009-05-28 14:34:40 -07003075 if (!vma || vma->vm_flags & VM_MAYSHARE)
Mel Gormana1e78772008-07-23 21:27:23 -07003076 chg = region_chg(&inode->i_mapping->private_list, from, to);
Mel Gorman5a6fe122009-02-10 14:02:27 +00003077 else {
3078 struct resv_map *resv_map = resv_map_alloc();
Mel Gorman5a6fe122009-02-10 14:02:27 +00003079 if (!resv_map)
3080 return -ENOMEM;
3081
Mel Gorman17c9d122009-02-11 16:34:16 +00003082 chg = to - from;
3083
Mel Gorman5a6fe122009-02-10 14:02:27 +00003084 set_vma_resv_map(vma, resv_map);
3085 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
3086 }
3087
Dave Hansenc50ac052012-05-29 15:06:46 -07003088 if (chg < 0) {
3089 ret = chg;
3090 goto out_err;
3091 }
Mel Gorman17c9d122009-02-11 16:34:16 +00003092
David Gibson90481622012-03-21 16:34:12 -07003093 /* There must be enough pages in the subpool for the mapping */
Dave Hansenc50ac052012-05-29 15:06:46 -07003094 if (hugepage_subpool_get_pages(spool, chg)) {
3095 ret = -ENOSPC;
3096 goto out_err;
3097 }
Mel Gorman17c9d122009-02-11 16:34:16 +00003098
3099 /*
3100 * Check enough hugepages are available for the reservation.
David Gibson90481622012-03-21 16:34:12 -07003101 * Hand the pages back to the subpool if there are not
Mel Gorman17c9d122009-02-11 16:34:16 +00003102 */
3103 ret = hugetlb_acct_memory(h, chg);
3104 if (ret < 0) {
David Gibson90481622012-03-21 16:34:12 -07003105 hugepage_subpool_put_pages(spool, chg);
Dave Hansenc50ac052012-05-29 15:06:46 -07003106 goto out_err;
Mel Gorman17c9d122009-02-11 16:34:16 +00003107 }
3108
3109 /*
3110 * Account for the reservations made. Shared mappings record regions
3111 * that have reservations as they are shared by multiple VMAs.
3112 * When the last VMA disappears, the region map says how much
3113 * the reservation was and the page cache tells how much of
3114 * the reservation was consumed. Private mappings are per-VMA and
3115 * only the consumed reservations are tracked. When the VMA
3116 * disappears, the original reservation is the VMA size and the
3117 * consumed reservations are stored in the map. Hence, nothing
3118 * else has to be done for private mappings here
3119 */
Mel Gormanf83a2752009-05-28 14:34:40 -07003120 if (!vma || vma->vm_flags & VM_MAYSHARE)
Mel Gorman17c9d122009-02-11 16:34:16 +00003121 region_add(&inode->i_mapping->private_list, from, to);
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -07003122 return 0;
Dave Hansenc50ac052012-05-29 15:06:46 -07003123out_err:
Dave Hansen4523e142012-05-30 07:51:07 -07003124 if (vma)
3125 resv_map_put(vma);
Dave Hansenc50ac052012-05-29 15:06:46 -07003126 return ret;
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -07003127}
3128
3129void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed)
3130{
Andi Kleena5516432008-07-23 21:27:41 -07003131 struct hstate *h = hstate_inode(inode);
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -07003132 long chg = region_truncate(&inode->i_mapping->private_list, offset);
David Gibson90481622012-03-21 16:34:12 -07003133 struct hugepage_subpool *spool = subpool_inode(inode);
Ken Chen45c682a2007-11-14 16:59:44 -08003134
3135 spin_lock(&inode->i_lock);
Eric Sandeene4c6f8b2009-07-29 15:02:16 -07003136 inode->i_blocks -= (blocks_per_huge_page(h) * freed);
Ken Chen45c682a2007-11-14 16:59:44 -08003137 spin_unlock(&inode->i_lock);
3138
David Gibson90481622012-03-21 16:34:12 -07003139 hugepage_subpool_put_pages(spool, (chg - freed));
Andi Kleena5516432008-07-23 21:27:41 -07003140 hugetlb_acct_memory(h, -(chg - freed));
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -07003141}
Naoya Horiguchi93f70f92010-05-28 09:29:20 +09003142
Andi Kleend5bd9102010-09-27 09:00:12 +02003143#ifdef CONFIG_MEMORY_FAILURE
3144
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +09003145/* Should be called in hugetlb_lock */
3146static int is_hugepage_on_freelist(struct page *hpage)
3147{
3148 struct page *page;
3149 struct page *tmp;
3150 struct hstate *h = page_hstate(hpage);
3151 int nid = page_to_nid(hpage);
3152
3153 list_for_each_entry_safe(page, tmp, &h->hugepage_freelists[nid], lru)
3154 if (page == hpage)
3155 return 1;
3156 return 0;
3157}
3158
Naoya Horiguchi93f70f92010-05-28 09:29:20 +09003159/*
3160 * This function is called from memory failure code.
3161 * Assume the caller holds page lock of the head page.
3162 */
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +09003163int dequeue_hwpoisoned_huge_page(struct page *hpage)
Naoya Horiguchi93f70f92010-05-28 09:29:20 +09003164{
3165 struct hstate *h = page_hstate(hpage);
3166 int nid = page_to_nid(hpage);
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +09003167 int ret = -EBUSY;
Naoya Horiguchi93f70f92010-05-28 09:29:20 +09003168
3169 spin_lock(&hugetlb_lock);
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +09003170 if (is_hugepage_on_freelist(hpage)) {
Naoya Horiguchi56f2fb12012-12-12 13:52:33 -08003171 /*
3172 * Hwpoisoned hugepage isn't linked to activelist or freelist,
3173 * but dangling hpage->lru can trigger list-debug warnings
3174 * (this happens when we call unpoison_memory() on it),
3175 * so let it point to itself with list_del_init().
3176 */
3177 list_del_init(&hpage->lru);
Naoya Horiguchi8c6c2ec2010-09-08 10:19:38 +09003178 set_page_refcounted(hpage);
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +09003179 h->free_huge_pages--;
3180 h->free_huge_pages_node[nid]--;
3181 ret = 0;
3182 }
Naoya Horiguchi93f70f92010-05-28 09:29:20 +09003183 spin_unlock(&hugetlb_lock);
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +09003184 return ret;
Naoya Horiguchi93f70f92010-05-28 09:29:20 +09003185}
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +09003186#endif