blob: 4f3ea0b1e57ce33bbf71bd1d3f41e04e11ae05dc [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
2 * Generic hugetlb support.
Nadia Yvette Chambers6d49e352012-12-06 10:39:54 +01003 * (C) Nadia Yvette Chambers, April 2004
Linus Torvalds1da177e2005-04-16 15:20:36 -07004 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07005#include <linux/list.h>
6#include <linux/init.h>
7#include <linux/module.h>
8#include <linux/mm.h>
Alexey Dobriyane1759c22008-10-15 23:50:22 +04009#include <linux/seq_file.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070010#include <linux/sysctl.h>
11#include <linux/highmem.h>
Andrea Arcangelicddb8a52008-07-28 15:46:29 -070012#include <linux/mmu_notifier.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070013#include <linux/nodemask.h>
David Gibson63551ae2005-06-21 17:14:44 -070014#include <linux/pagemap.h>
Christoph Lameter5da7ca82006-01-06 00:10:46 -080015#include <linux/mempolicy.h>
Christoph Lameteraea47ff2006-01-08 01:00:57 -080016#include <linux/cpuset.h>
David Gibson3935baa2006-03-22 00:08:53 -080017#include <linux/mutex.h>
Andi Kleenaa888a72008-07-23 21:27:47 -070018#include <linux/bootmem.h>
Nishanth Aravamudana3437872008-07-23 21:27:44 -070019#include <linux/sysfs.h>
Tejun Heo5a0e3ad2010-03-24 17:04:11 +090020#include <linux/slab.h>
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +090021#include <linux/rmap.h>
Naoya Horiguchifd6a03ed2010-05-28 09:29:21 +090022#include <linux/swap.h>
23#include <linux/swapops.h>
Linus Torvaldsd6606682008-08-06 12:04:54 -070024
David Gibson63551ae2005-06-21 17:14:44 -070025#include <asm/page.h>
26#include <asm/pgtable.h>
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -070027#include <asm/tlb.h>
David Gibson63551ae2005-06-21 17:14:44 -070028
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -070029#include <linux/io.h>
David Gibson63551ae2005-06-21 17:14:44 -070030#include <linux/hugetlb.h>
Aneesh Kumar K.V9dd540e2012-07-31 16:42:15 -070031#include <linux/hugetlb_cgroup.h>
Lee Schermerhorn9a305232009-12-14 17:58:25 -080032#include <linux/node.h>
Nick Piggin7835e982006-03-22 00:08:40 -080033#include "internal.h"
Linus Torvalds1da177e2005-04-16 15:20:36 -070034
35const unsigned long hugetlb_zero = 0, hugetlb_infinity = ~0UL;
Mel Gorman396faf02007-07-17 04:03:13 -070036static gfp_t htlb_alloc_mask = GFP_HIGHUSER;
37unsigned long hugepages_treat_as_movable;
Andi Kleena5516432008-07-23 21:27:41 -070038
Aneesh Kumar K.Vc3f38a32012-07-31 16:42:10 -070039int hugetlb_max_hstate __read_mostly;
Andi Kleene5ff2152008-07-23 21:27:42 -070040unsigned int default_hstate_idx;
41struct hstate hstates[HUGE_MAX_HSTATE];
42
Jon Tollefson53ba51d2008-07-23 21:27:52 -070043__initdata LIST_HEAD(huge_boot_pages);
44
Andi Kleene5ff2152008-07-23 21:27:42 -070045/* for command line parsing */
46static struct hstate * __initdata parsed_hstate;
47static unsigned long __initdata default_hstate_max_huge_pages;
Nick Piggine11bfbf2008-07-23 21:27:52 -070048static unsigned long __initdata default_hstate_size;
Andi Kleene5ff2152008-07-23 21:27:42 -070049
David Gibson3935baa2006-03-22 00:08:53 -080050/*
51 * Protects updates to hugepage_freelists, nr_huge_pages, and free_huge_pages
52 */
Aneesh Kumar K.Vc3f38a32012-07-31 16:42:10 -070053DEFINE_SPINLOCK(hugetlb_lock);
Eric Paris0bd0f9f2005-11-21 21:32:28 -080054
David Gibson90481622012-03-21 16:34:12 -070055static inline void unlock_or_release_subpool(struct hugepage_subpool *spool)
56{
57 bool free = (spool->count == 0) && (spool->used_hpages == 0);
58
59 spin_unlock(&spool->lock);
60
61 /* If no pages are used, and no other handles to the subpool
62 * remain, free the subpool the subpool remain */
63 if (free)
64 kfree(spool);
65}
66
67struct hugepage_subpool *hugepage_new_subpool(long nr_blocks)
68{
69 struct hugepage_subpool *spool;
70
71 spool = kmalloc(sizeof(*spool), GFP_KERNEL);
72 if (!spool)
73 return NULL;
74
75 spin_lock_init(&spool->lock);
76 spool->count = 1;
77 spool->max_hpages = nr_blocks;
78 spool->used_hpages = 0;
79
80 return spool;
81}
82
83void hugepage_put_subpool(struct hugepage_subpool *spool)
84{
85 spin_lock(&spool->lock);
86 BUG_ON(!spool->count);
87 spool->count--;
88 unlock_or_release_subpool(spool);
89}
90
91static int hugepage_subpool_get_pages(struct hugepage_subpool *spool,
92 long delta)
93{
94 int ret = 0;
95
96 if (!spool)
97 return 0;
98
99 spin_lock(&spool->lock);
100 if ((spool->used_hpages + delta) <= spool->max_hpages) {
101 spool->used_hpages += delta;
102 } else {
103 ret = -ENOMEM;
104 }
105 spin_unlock(&spool->lock);
106
107 return ret;
108}
109
110static void hugepage_subpool_put_pages(struct hugepage_subpool *spool,
111 long delta)
112{
113 if (!spool)
114 return;
115
116 spin_lock(&spool->lock);
117 spool->used_hpages -= delta;
118 /* If hugetlbfs_put_super couldn't free spool due to
119 * an outstanding quota reference, free it now. */
120 unlock_or_release_subpool(spool);
121}
122
123static inline struct hugepage_subpool *subpool_inode(struct inode *inode)
124{
125 return HUGETLBFS_SB(inode->i_sb)->spool;
126}
127
128static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma)
129{
130 return subpool_inode(vma->vm_file->f_dentry->d_inode);
131}
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
758static int alloc_fresh_huge_page(struct hstate *h, nodemask_t *nodes_allowed)
Nishanth Aravamudan63b46132007-10-16 01:26:24 -0700759{
760 struct page *page;
761 int start_nid;
762 int next_nid;
763 int ret = 0;
764
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800765 start_nid = hstate_next_node_to_alloc(h, nodes_allowed);
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700766 next_nid = start_nid;
Nishanth Aravamudan63b46132007-10-16 01:26:24 -0700767
768 do {
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700769 page = alloc_fresh_huge_page_node(h, next_nid);
Lee Schermerhorn9a76db02009-12-14 17:58:15 -0800770 if (page) {
Nishanth Aravamudan63b46132007-10-16 01:26:24 -0700771 ret = 1;
Lee Schermerhorn9a76db02009-12-14 17:58:15 -0800772 break;
773 }
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800774 next_nid = hstate_next_node_to_alloc(h, nodes_allowed);
Lee Schermerhorn9a76db02009-12-14 17:58:15 -0800775 } while (next_nid != start_nid);
Nishanth Aravamudan63b46132007-10-16 01:26:24 -0700776
Adam Litke3b116302008-04-28 02:13:06 -0700777 if (ret)
778 count_vm_event(HTLB_BUDDY_PGALLOC);
779 else
780 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
781
Nishanth Aravamudan63b46132007-10-16 01:26:24 -0700782 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700783}
784
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700785/*
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800786 * helper for free_pool_huge_page() - return the previously saved
787 * node ["this node"] from which to free a huge page. Advance the
788 * next node id whether or not we find a free huge page to free so
789 * that the next attempt to free addresses the next node.
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700790 */
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800791static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700792{
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800793 int nid;
Lee Schermerhorn9a76db02009-12-14 17:58:15 -0800794
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800795 VM_BUG_ON(!nodes_allowed);
796
797 nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
798 h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
799
Lee Schermerhorn9a76db02009-12-14 17:58:15 -0800800 return nid;
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700801}
802
803/*
804 * Free huge page from pool from next node to free.
805 * Attempt to keep persistent huge pages more or less
806 * balanced over allowed nodes.
807 * Called with hugetlb_lock locked.
808 */
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800809static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
810 bool acct_surplus)
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700811{
812 int start_nid;
813 int next_nid;
814 int ret = 0;
815
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800816 start_nid = hstate_next_node_to_free(h, nodes_allowed);
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700817 next_nid = start_nid;
818
819 do {
Lee Schermerhorn685f3452009-09-21 17:01:23 -0700820 /*
821 * If we're returning unused surplus pages, only examine
822 * nodes with surplus pages.
823 */
824 if ((!acct_surplus || h->surplus_huge_pages_node[next_nid]) &&
825 !list_empty(&h->hugepage_freelists[next_nid])) {
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700826 struct page *page =
827 list_entry(h->hugepage_freelists[next_nid].next,
828 struct page, lru);
829 list_del(&page->lru);
830 h->free_huge_pages--;
831 h->free_huge_pages_node[next_nid]--;
Lee Schermerhorn685f3452009-09-21 17:01:23 -0700832 if (acct_surplus) {
833 h->surplus_huge_pages--;
834 h->surplus_huge_pages_node[next_nid]--;
835 }
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700836 update_and_free_page(h, page);
837 ret = 1;
Lee Schermerhorn9a76db02009-12-14 17:58:15 -0800838 break;
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700839 }
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800840 next_nid = hstate_next_node_to_free(h, nodes_allowed);
Lee Schermerhorn9a76db02009-12-14 17:58:15 -0800841 } while (next_nid != start_nid);
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700842
843 return ret;
844}
845
Naoya Horiguchibf50bab2010-09-08 10:19:33 +0900846static struct page *alloc_buddy_huge_page(struct hstate *h, int nid)
Adam Litke7893d1d2007-10-16 01:26:18 -0700847{
848 struct page *page;
Naoya Horiguchibf50bab2010-09-08 10:19:33 +0900849 unsigned int r_nid;
Adam Litke7893d1d2007-10-16 01:26:18 -0700850
Andi Kleenaa888a72008-07-23 21:27:47 -0700851 if (h->order >= MAX_ORDER)
852 return NULL;
853
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800854 /*
855 * Assume we will successfully allocate the surplus page to
856 * prevent racing processes from causing the surplus to exceed
857 * overcommit
858 *
859 * This however introduces a different race, where a process B
860 * tries to grow the static hugepage pool while alloc_pages() is
861 * called by process A. B will only examine the per-node
862 * counters in determining if surplus huge pages can be
863 * converted to normal huge pages in adjust_pool_surplus(). A
864 * won't be able to increment the per-node counter, until the
865 * lock is dropped by B, but B doesn't drop hugetlb_lock until
866 * no more huge pages can be converted from surplus to normal
867 * state (and doesn't try to convert again). Thus, we have a
868 * case where a surplus huge page exists, the pool is grown, and
869 * the surplus huge page still exists after, even though it
870 * should just have been converted to a normal huge page. This
871 * does not leak memory, though, as the hugepage will be freed
872 * once it is out of use. It also does not allow the counters to
873 * go out of whack in adjust_pool_surplus() as we don't modify
874 * the node values until we've gotten the hugepage and only the
875 * per-node value is checked there.
876 */
877 spin_lock(&hugetlb_lock);
Andi Kleena5516432008-07-23 21:27:41 -0700878 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800879 spin_unlock(&hugetlb_lock);
880 return NULL;
881 } else {
Andi Kleena5516432008-07-23 21:27:41 -0700882 h->nr_huge_pages++;
883 h->surplus_huge_pages++;
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800884 }
885 spin_unlock(&hugetlb_lock);
886
Naoya Horiguchibf50bab2010-09-08 10:19:33 +0900887 if (nid == NUMA_NO_NODE)
888 page = alloc_pages(htlb_alloc_mask|__GFP_COMP|
889 __GFP_REPEAT|__GFP_NOWARN,
890 huge_page_order(h));
891 else
892 page = alloc_pages_exact_node(nid,
893 htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE|
894 __GFP_REPEAT|__GFP_NOWARN, huge_page_order(h));
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800895
Gerald Schaefercaff3a22008-08-12 15:08:38 -0700896 if (page && arch_prepare_hugepage(page)) {
897 __free_pages(page, huge_page_order(h));
Hillf Dantonea5768c2012-01-10 15:08:30 -0800898 page = NULL;
Gerald Schaefercaff3a22008-08-12 15:08:38 -0700899 }
900
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800901 spin_lock(&hugetlb_lock);
Adam Litke7893d1d2007-10-16 01:26:18 -0700902 if (page) {
Aneesh Kumar K.V0edaecf2012-07-31 16:42:07 -0700903 INIT_LIST_HEAD(&page->lru);
Naoya Horiguchibf50bab2010-09-08 10:19:33 +0900904 r_nid = page_to_nid(page);
Adam Litke7893d1d2007-10-16 01:26:18 -0700905 set_compound_page_dtor(page, free_huge_page);
Aneesh Kumar K.V9dd540e2012-07-31 16:42:15 -0700906 set_hugetlb_cgroup(page, NULL);
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800907 /*
908 * We incremented the global counters already
909 */
Naoya Horiguchibf50bab2010-09-08 10:19:33 +0900910 h->nr_huge_pages_node[r_nid]++;
911 h->surplus_huge_pages_node[r_nid]++;
Adam Litke3b116302008-04-28 02:13:06 -0700912 __count_vm_event(HTLB_BUDDY_PGALLOC);
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800913 } else {
Andi Kleena5516432008-07-23 21:27:41 -0700914 h->nr_huge_pages--;
915 h->surplus_huge_pages--;
Adam Litke3b116302008-04-28 02:13:06 -0700916 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
Adam Litke7893d1d2007-10-16 01:26:18 -0700917 }
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800918 spin_unlock(&hugetlb_lock);
Adam Litke7893d1d2007-10-16 01:26:18 -0700919
920 return page;
921}
922
Adam Litkee4e574b2007-10-16 01:26:19 -0700923/*
Naoya Horiguchibf50bab2010-09-08 10:19:33 +0900924 * This allocation function is useful in the context where vma is irrelevant.
925 * E.g. soft-offlining uses this function because it only cares physical
926 * address of error page.
927 */
928struct page *alloc_huge_page_node(struct hstate *h, int nid)
929{
930 struct page *page;
931
932 spin_lock(&hugetlb_lock);
933 page = dequeue_huge_page_node(h, nid);
934 spin_unlock(&hugetlb_lock);
935
Aneesh Kumar K.V94ae8ba2012-07-31 16:42:35 -0700936 if (!page)
Naoya Horiguchibf50bab2010-09-08 10:19:33 +0900937 page = alloc_buddy_huge_page(h, nid);
938
939 return page;
940}
941
942/*
Lucas De Marchi25985ed2011-03-30 22:57:33 -0300943 * Increase the hugetlb pool such that it can accommodate a reservation
Adam Litkee4e574b2007-10-16 01:26:19 -0700944 * of size 'delta'.
945 */
Andi Kleena5516432008-07-23 21:27:41 -0700946static int gather_surplus_pages(struct hstate *h, int delta)
Adam Litkee4e574b2007-10-16 01:26:19 -0700947{
948 struct list_head surplus_list;
949 struct page *page, *tmp;
950 int ret, i;
951 int needed, allocated;
Hillf Danton28073b02012-03-21 16:34:00 -0700952 bool alloc_ok = true;
Adam Litkee4e574b2007-10-16 01:26:19 -0700953
Andi Kleena5516432008-07-23 21:27:41 -0700954 needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
Adam Litkeac09b3a2008-03-04 14:29:38 -0800955 if (needed <= 0) {
Andi Kleena5516432008-07-23 21:27:41 -0700956 h->resv_huge_pages += delta;
Adam Litkee4e574b2007-10-16 01:26:19 -0700957 return 0;
Adam Litkeac09b3a2008-03-04 14:29:38 -0800958 }
Adam Litkee4e574b2007-10-16 01:26:19 -0700959
960 allocated = 0;
961 INIT_LIST_HEAD(&surplus_list);
962
963 ret = -ENOMEM;
964retry:
965 spin_unlock(&hugetlb_lock);
966 for (i = 0; i < needed; i++) {
Naoya Horiguchibf50bab2010-09-08 10:19:33 +0900967 page = alloc_buddy_huge_page(h, NUMA_NO_NODE);
Hillf Danton28073b02012-03-21 16:34:00 -0700968 if (!page) {
969 alloc_ok = false;
970 break;
971 }
Adam Litkee4e574b2007-10-16 01:26:19 -0700972 list_add(&page->lru, &surplus_list);
973 }
Hillf Danton28073b02012-03-21 16:34:00 -0700974 allocated += i;
Adam Litkee4e574b2007-10-16 01:26:19 -0700975
976 /*
977 * After retaking hugetlb_lock, we need to recalculate 'needed'
978 * because either resv_huge_pages or free_huge_pages may have changed.
979 */
980 spin_lock(&hugetlb_lock);
Andi Kleena5516432008-07-23 21:27:41 -0700981 needed = (h->resv_huge_pages + delta) -
982 (h->free_huge_pages + allocated);
Hillf Danton28073b02012-03-21 16:34:00 -0700983 if (needed > 0) {
984 if (alloc_ok)
985 goto retry;
986 /*
987 * We were not able to allocate enough pages to
988 * satisfy the entire reservation so we free what
989 * we've allocated so far.
990 */
991 goto free;
992 }
Adam Litkee4e574b2007-10-16 01:26:19 -0700993 /*
994 * The surplus_list now contains _at_least_ the number of extra pages
Lucas De Marchi25985ed2011-03-30 22:57:33 -0300995 * needed to accommodate the reservation. Add the appropriate number
Adam Litkee4e574b2007-10-16 01:26:19 -0700996 * of pages to the hugetlb pool and free the extras back to the buddy
Adam Litkeac09b3a2008-03-04 14:29:38 -0800997 * allocator. Commit the entire reservation here to prevent another
998 * process from stealing the pages as they are added to the pool but
999 * before they are reserved.
Adam Litkee4e574b2007-10-16 01:26:19 -07001000 */
1001 needed += allocated;
Andi Kleena5516432008-07-23 21:27:41 -07001002 h->resv_huge_pages += delta;
Adam Litkee4e574b2007-10-16 01:26:19 -07001003 ret = 0;
Naoya Horiguchia9869b82010-09-08 10:19:37 +09001004
Adam Litke19fc3f02008-04-28 02:12:20 -07001005 /* Free the needed pages to the hugetlb pool */
Adam Litkee4e574b2007-10-16 01:26:19 -07001006 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
Adam Litke19fc3f02008-04-28 02:12:20 -07001007 if ((--needed) < 0)
1008 break;
Naoya Horiguchia9869b82010-09-08 10:19:37 +09001009 /*
1010 * This page is now managed by the hugetlb allocator and has
1011 * no users -- drop the buddy allocator's reference.
1012 */
1013 put_page_testzero(page);
1014 VM_BUG_ON(page_count(page));
Andi Kleena5516432008-07-23 21:27:41 -07001015 enqueue_huge_page(h, page);
Adam Litke19fc3f02008-04-28 02:12:20 -07001016 }
Hillf Danton28073b02012-03-21 16:34:00 -07001017free:
Hillf Dantonb0365c82011-12-28 15:57:16 -08001018 spin_unlock(&hugetlb_lock);
Adam Litke19fc3f02008-04-28 02:12:20 -07001019
1020 /* Free unnecessary surplus pages to the buddy allocator */
1021 if (!list_empty(&surplus_list)) {
Adam Litke19fc3f02008-04-28 02:12:20 -07001022 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
Naoya Horiguchia9869b82010-09-08 10:19:37 +09001023 put_page(page);
Adam Litkeaf767cb2007-10-16 01:26:25 -07001024 }
Adam Litkee4e574b2007-10-16 01:26:19 -07001025 }
Naoya Horiguchia9869b82010-09-08 10:19:37 +09001026 spin_lock(&hugetlb_lock);
Adam Litkee4e574b2007-10-16 01:26:19 -07001027
1028 return ret;
1029}
1030
1031/*
1032 * When releasing a hugetlb pool reservation, any surplus pages that were
1033 * allocated to satisfy the reservation must be explicitly freed if they were
1034 * never used.
Lee Schermerhorn685f3452009-09-21 17:01:23 -07001035 * Called with hugetlb_lock held.
Adam Litkee4e574b2007-10-16 01:26:19 -07001036 */
Andi Kleena5516432008-07-23 21:27:41 -07001037static void return_unused_surplus_pages(struct hstate *h,
1038 unsigned long unused_resv_pages)
Adam Litkee4e574b2007-10-16 01:26:19 -07001039{
Adam Litkee4e574b2007-10-16 01:26:19 -07001040 unsigned long nr_pages;
1041
Adam Litkeac09b3a2008-03-04 14:29:38 -08001042 /* Uncommit the reservation */
Andi Kleena5516432008-07-23 21:27:41 -07001043 h->resv_huge_pages -= unused_resv_pages;
Adam Litkeac09b3a2008-03-04 14:29:38 -08001044
Andi Kleenaa888a72008-07-23 21:27:47 -07001045 /* Cannot return gigantic pages currently */
1046 if (h->order >= MAX_ORDER)
1047 return;
1048
Andi Kleena5516432008-07-23 21:27:41 -07001049 nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
Adam Litkee4e574b2007-10-16 01:26:19 -07001050
Lee Schermerhorn685f3452009-09-21 17:01:23 -07001051 /*
1052 * We want to release as many surplus pages as possible, spread
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -08001053 * evenly across all nodes with memory. Iterate across these nodes
1054 * until we can no longer free unreserved surplus pages. This occurs
1055 * when the nodes with surplus pages have no free pages.
1056 * free_pool_huge_page() will balance the the freed pages across the
1057 * on-line nodes with memory and will handle the hstate accounting.
Lee Schermerhorn685f3452009-09-21 17:01:23 -07001058 */
1059 while (nr_pages--) {
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08001060 if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1))
Lee Schermerhorn685f3452009-09-21 17:01:23 -07001061 break;
Adam Litkee4e574b2007-10-16 01:26:19 -07001062 }
1063}
1064
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001065/*
1066 * Determine if the huge page at addr within the vma has an associated
1067 * reservation. Where it does not we will need to logically increase
David Gibson90481622012-03-21 16:34:12 -07001068 * reservation and actually increase subpool usage before an allocation
1069 * can occur. Where any new reservation would be required the
1070 * reservation change is prepared, but not committed. Once the page
1071 * has been allocated from the subpool and instantiated the change should
1072 * be committed via vma_commit_reservation. No action is required on
1073 * failure.
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001074 */
Roel Kluine2f17d92009-03-31 15:23:15 -07001075static long vma_needs_reservation(struct hstate *h,
Andi Kleena5516432008-07-23 21:27:41 -07001076 struct vm_area_struct *vma, unsigned long addr)
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001077{
1078 struct address_space *mapping = vma->vm_file->f_mapping;
1079 struct inode *inode = mapping->host;
1080
Mel Gormanf83a2752009-05-28 14:34:40 -07001081 if (vma->vm_flags & VM_MAYSHARE) {
Andi Kleena5516432008-07-23 21:27:41 -07001082 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001083 return region_chg(&inode->i_mapping->private_list,
1084 idx, idx + 1);
1085
Andy Whitcroft84afd992008-07-23 21:27:32 -07001086 } else if (!is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
1087 return 1;
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001088
Andy Whitcroft84afd992008-07-23 21:27:32 -07001089 } else {
Roel Kluine2f17d92009-03-31 15:23:15 -07001090 long err;
Andi Kleena5516432008-07-23 21:27:41 -07001091 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
Andy Whitcroft84afd992008-07-23 21:27:32 -07001092 struct resv_map *reservations = vma_resv_map(vma);
1093
1094 err = region_chg(&reservations->regions, idx, idx + 1);
1095 if (err < 0)
1096 return err;
1097 return 0;
1098 }
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001099}
Andi Kleena5516432008-07-23 21:27:41 -07001100static void vma_commit_reservation(struct hstate *h,
1101 struct vm_area_struct *vma, unsigned long addr)
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001102{
1103 struct address_space *mapping = vma->vm_file->f_mapping;
1104 struct inode *inode = mapping->host;
1105
Mel Gormanf83a2752009-05-28 14:34:40 -07001106 if (vma->vm_flags & VM_MAYSHARE) {
Andi Kleena5516432008-07-23 21:27:41 -07001107 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001108 region_add(&inode->i_mapping->private_list, idx, idx + 1);
Andy Whitcroft84afd992008-07-23 21:27:32 -07001109
1110 } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
Andi Kleena5516432008-07-23 21:27:41 -07001111 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
Andy Whitcroft84afd992008-07-23 21:27:32 -07001112 struct resv_map *reservations = vma_resv_map(vma);
1113
1114 /* Mark this page used in the map. */
1115 region_add(&reservations->regions, idx, idx + 1);
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001116 }
1117}
1118
David Gibson27a85ef2006-03-22 00:08:56 -08001119static struct page *alloc_huge_page(struct vm_area_struct *vma,
Mel Gorman04f2cbe2008-07-23 21:27:25 -07001120 unsigned long addr, int avoid_reserve)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001121{
David Gibson90481622012-03-21 16:34:12 -07001122 struct hugepage_subpool *spool = subpool_vma(vma);
Andi Kleena5516432008-07-23 21:27:41 -07001123 struct hstate *h = hstate_vma(vma);
Adam Litke348ea202007-11-14 16:59:37 -08001124 struct page *page;
Roel Kluine2f17d92009-03-31 15:23:15 -07001125 long chg;
Aneesh Kumar K.V6d76dcf2012-07-31 16:42:18 -07001126 int ret, idx;
1127 struct hugetlb_cgroup *h_cg;
Adam Litke2fc39ce2007-11-14 16:59:39 -08001128
Aneesh Kumar K.V6d76dcf2012-07-31 16:42:18 -07001129 idx = hstate_index(h);
Mel Gormana1e78772008-07-23 21:27:23 -07001130 /*
David Gibson90481622012-03-21 16:34:12 -07001131 * Processes that did not create the mapping will have no
1132 * reserves and will not have accounted against subpool
1133 * limit. Check that the subpool limit can be made before
1134 * satisfying the allocation MAP_NORESERVE mappings may also
1135 * need pages and subpool limit allocated allocated if no reserve
1136 * mapping overlaps.
Mel Gormana1e78772008-07-23 21:27:23 -07001137 */
Andi Kleena5516432008-07-23 21:27:41 -07001138 chg = vma_needs_reservation(h, vma, addr);
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001139 if (chg < 0)
Aneesh Kumar K.V76dcee72012-07-31 16:41:57 -07001140 return ERR_PTR(-ENOMEM);
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001141 if (chg)
David Gibson90481622012-03-21 16:34:12 -07001142 if (hugepage_subpool_get_pages(spool, chg))
Aneesh Kumar K.V76dcee72012-07-31 16:41:57 -07001143 return ERR_PTR(-ENOSPC);
Mel Gormana1e78772008-07-23 21:27:23 -07001144
Aneesh Kumar K.V6d76dcf2012-07-31 16:42:18 -07001145 ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg);
1146 if (ret) {
1147 hugepage_subpool_put_pages(spool, chg);
1148 return ERR_PTR(-ENOSPC);
1149 }
Mel Gormana1e78772008-07-23 21:27:23 -07001150 spin_lock(&hugetlb_lock);
Andi Kleena5516432008-07-23 21:27:41 -07001151 page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve);
Aneesh Kumar K.V94ae8ba2012-07-31 16:42:35 -07001152 if (page) {
1153 /* update page cgroup details */
1154 hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h),
1155 h_cg, page);
1156 spin_unlock(&hugetlb_lock);
1157 } else {
1158 spin_unlock(&hugetlb_lock);
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001159 page = alloc_buddy_huge_page(h, NUMA_NO_NODE);
Mel Gormana1e78772008-07-23 21:27:23 -07001160 if (!page) {
Aneesh Kumar K.V6d76dcf2012-07-31 16:42:18 -07001161 hugetlb_cgroup_uncharge_cgroup(idx,
1162 pages_per_huge_page(h),
1163 h_cg);
David Gibson90481622012-03-21 16:34:12 -07001164 hugepage_subpool_put_pages(spool, chg);
Aneesh Kumar K.V76dcee72012-07-31 16:41:57 -07001165 return ERR_PTR(-ENOSPC);
Mel Gormana1e78772008-07-23 21:27:23 -07001166 }
Aneesh Kumar K.V79dbb232012-07-31 16:42:32 -07001167 spin_lock(&hugetlb_lock);
Aneesh Kumar K.V94ae8ba2012-07-31 16:42:35 -07001168 hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h),
1169 h_cg, page);
Aneesh Kumar K.V79dbb232012-07-31 16:42:32 -07001170 list_move(&page->lru, &h->hugepage_activelist);
1171 spin_unlock(&hugetlb_lock);
Mel Gormana1e78772008-07-23 21:27:23 -07001172 }
1173
David Gibson90481622012-03-21 16:34:12 -07001174 set_page_private(page, (unsigned long)spool);
Mel Gormana1e78772008-07-23 21:27:23 -07001175
Andi Kleena5516432008-07-23 21:27:41 -07001176 vma_commit_reservation(h, vma, addr);
Adam Litke90d8b7e2007-11-14 16:59:42 -08001177 return page;
David Gibsonb45b5bd2006-03-22 00:08:55 -08001178}
1179
Cyrill Gorcunov91f47662009-01-06 14:40:33 -08001180int __weak alloc_bootmem_huge_page(struct hstate *h)
Andi Kleenaa888a72008-07-23 21:27:47 -07001181{
1182 struct huge_bootmem_page *m;
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08001183 int nr_nodes = nodes_weight(node_states[N_MEMORY]);
Andi Kleenaa888a72008-07-23 21:27:47 -07001184
1185 while (nr_nodes) {
1186 void *addr;
1187
1188 addr = __alloc_bootmem_node_nopanic(
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001189 NODE_DATA(hstate_next_node_to_alloc(h,
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08001190 &node_states[N_MEMORY])),
Andi Kleenaa888a72008-07-23 21:27:47 -07001191 huge_page_size(h), huge_page_size(h), 0);
1192
1193 if (addr) {
1194 /*
1195 * Use the beginning of the huge page to store the
1196 * huge_bootmem_page struct (until gather_bootmem
1197 * puts them into the mem_map).
1198 */
1199 m = addr;
Cyrill Gorcunov91f47662009-01-06 14:40:33 -08001200 goto found;
Andi Kleenaa888a72008-07-23 21:27:47 -07001201 }
Andi Kleenaa888a72008-07-23 21:27:47 -07001202 nr_nodes--;
1203 }
1204 return 0;
1205
1206found:
1207 BUG_ON((unsigned long)virt_to_phys(m) & (huge_page_size(h) - 1));
1208 /* Put them into a private list first because mem_map is not up yet */
1209 list_add(&m->list, &huge_boot_pages);
1210 m->hstate = h;
1211 return 1;
1212}
1213
Andy Whitcroft18229df2008-11-06 12:53:27 -08001214static void prep_compound_huge_page(struct page *page, int order)
1215{
1216 if (unlikely(order > (MAX_ORDER - 1)))
1217 prep_compound_gigantic_page(page, order);
1218 else
1219 prep_compound_page(page, order);
1220}
1221
Andi Kleenaa888a72008-07-23 21:27:47 -07001222/* Put bootmem huge pages into the standard lists after mem_map is up */
1223static void __init gather_bootmem_prealloc(void)
1224{
1225 struct huge_bootmem_page *m;
1226
1227 list_for_each_entry(m, &huge_boot_pages, list) {
Andi Kleenaa888a72008-07-23 21:27:47 -07001228 struct hstate *h = m->hstate;
Becky Bruceee8f2482011-07-25 17:11:50 -07001229 struct page *page;
1230
1231#ifdef CONFIG_HIGHMEM
1232 page = pfn_to_page(m->phys >> PAGE_SHIFT);
1233 free_bootmem_late((unsigned long)m,
1234 sizeof(struct huge_bootmem_page));
1235#else
1236 page = virt_to_page(m);
1237#endif
Andi Kleenaa888a72008-07-23 21:27:47 -07001238 __ClearPageReserved(page);
1239 WARN_ON(page_count(page) != 1);
Andy Whitcroft18229df2008-11-06 12:53:27 -08001240 prep_compound_huge_page(page, h->order);
Andi Kleenaa888a72008-07-23 21:27:47 -07001241 prep_new_huge_page(h, page, page_to_nid(page));
Rafael Aquinib0320c72011-06-15 15:08:39 -07001242 /*
1243 * If we had gigantic hugepages allocated at boot time, we need
1244 * to restore the 'stolen' pages to totalram_pages in order to
1245 * fix confusing memory reports from free(1) and another
1246 * side-effects, like CommitLimit going negative.
1247 */
1248 if (h->order > (MAX_ORDER - 1))
1249 totalram_pages += 1 << h->order;
Andi Kleenaa888a72008-07-23 21:27:47 -07001250 }
1251}
1252
Andi Kleen8faa8b02008-07-23 21:27:48 -07001253static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001254{
1255 unsigned long i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001256
Andi Kleene5ff2152008-07-23 21:27:42 -07001257 for (i = 0; i < h->max_huge_pages; ++i) {
Andi Kleenaa888a72008-07-23 21:27:47 -07001258 if (h->order >= MAX_ORDER) {
1259 if (!alloc_bootmem_huge_page(h))
1260 break;
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -08001261 } else if (!alloc_fresh_huge_page(h,
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08001262 &node_states[N_MEMORY]))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001263 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001264 }
Andi Kleen8faa8b02008-07-23 21:27:48 -07001265 h->max_huge_pages = i;
Andi Kleene5ff2152008-07-23 21:27:42 -07001266}
1267
1268static void __init hugetlb_init_hstates(void)
1269{
1270 struct hstate *h;
1271
1272 for_each_hstate(h) {
Andi Kleen8faa8b02008-07-23 21:27:48 -07001273 /* oversize hugepages were init'ed in early boot */
1274 if (h->order < MAX_ORDER)
1275 hugetlb_hstate_alloc_pages(h);
Andi Kleene5ff2152008-07-23 21:27:42 -07001276 }
1277}
1278
Andi Kleen4abd32d2008-07-23 21:27:49 -07001279static char * __init memfmt(char *buf, unsigned long n)
1280{
1281 if (n >= (1UL << 30))
1282 sprintf(buf, "%lu GB", n >> 30);
1283 else if (n >= (1UL << 20))
1284 sprintf(buf, "%lu MB", n >> 20);
1285 else
1286 sprintf(buf, "%lu KB", n >> 10);
1287 return buf;
1288}
1289
Andi Kleene5ff2152008-07-23 21:27:42 -07001290static void __init report_hugepages(void)
1291{
1292 struct hstate *h;
1293
1294 for_each_hstate(h) {
Andi Kleen4abd32d2008-07-23 21:27:49 -07001295 char buf[32];
1296 printk(KERN_INFO "HugeTLB registered %s page size, "
1297 "pre-allocated %ld pages\n",
1298 memfmt(buf, huge_page_size(h)),
1299 h->free_huge_pages);
Andi Kleene5ff2152008-07-23 21:27:42 -07001300 }
1301}
1302
Linus Torvalds1da177e2005-04-16 15:20:36 -07001303#ifdef CONFIG_HIGHMEM
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001304static void try_to_free_low(struct hstate *h, unsigned long count,
1305 nodemask_t *nodes_allowed)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001306{
Christoph Lameter4415cc82006-09-25 23:31:55 -07001307 int i;
1308
Andi Kleenaa888a72008-07-23 21:27:47 -07001309 if (h->order >= MAX_ORDER)
1310 return;
1311
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001312 for_each_node_mask(i, *nodes_allowed) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001313 struct page *page, *next;
Andi Kleena5516432008-07-23 21:27:41 -07001314 struct list_head *freel = &h->hugepage_freelists[i];
1315 list_for_each_entry_safe(page, next, freel, lru) {
1316 if (count >= h->nr_huge_pages)
Adam Litke6b0c8802007-10-16 01:26:23 -07001317 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001318 if (PageHighMem(page))
1319 continue;
1320 list_del(&page->lru);
Andi Kleene5ff2152008-07-23 21:27:42 -07001321 update_and_free_page(h, page);
Andi Kleena5516432008-07-23 21:27:41 -07001322 h->free_huge_pages--;
1323 h->free_huge_pages_node[page_to_nid(page)]--;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001324 }
1325 }
1326}
1327#else
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001328static inline void try_to_free_low(struct hstate *h, unsigned long count,
1329 nodemask_t *nodes_allowed)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001330{
1331}
1332#endif
1333
Wu Fengguang20a03072009-06-16 15:32:22 -07001334/*
1335 * Increment or decrement surplus_huge_pages. Keep node-specific counters
1336 * balanced by operating on them in a round-robin fashion.
1337 * Returns 1 if an adjustment was made.
1338 */
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001339static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
1340 int delta)
Wu Fengguang20a03072009-06-16 15:32:22 -07001341{
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001342 int start_nid, next_nid;
Wu Fengguang20a03072009-06-16 15:32:22 -07001343 int ret = 0;
1344
1345 VM_BUG_ON(delta != -1 && delta != 1);
Wu Fengguang20a03072009-06-16 15:32:22 -07001346
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001347 if (delta < 0)
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001348 start_nid = hstate_next_node_to_alloc(h, nodes_allowed);
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001349 else
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001350 start_nid = hstate_next_node_to_free(h, nodes_allowed);
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001351 next_nid = start_nid;
1352
1353 do {
1354 int nid = next_nid;
1355 if (delta < 0) {
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001356 /*
1357 * To shrink on this node, there must be a surplus page
1358 */
Lee Schermerhorn9a76db02009-12-14 17:58:15 -08001359 if (!h->surplus_huge_pages_node[nid]) {
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001360 next_nid = hstate_next_node_to_alloc(h,
1361 nodes_allowed);
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001362 continue;
Lee Schermerhorn9a76db02009-12-14 17:58:15 -08001363 }
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001364 }
1365 if (delta > 0) {
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001366 /*
1367 * Surplus cannot exceed the total number of pages
1368 */
1369 if (h->surplus_huge_pages_node[nid] >=
Lee Schermerhorn9a76db02009-12-14 17:58:15 -08001370 h->nr_huge_pages_node[nid]) {
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001371 next_nid = hstate_next_node_to_free(h,
1372 nodes_allowed);
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001373 continue;
Lee Schermerhorn9a76db02009-12-14 17:58:15 -08001374 }
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001375 }
Wu Fengguang20a03072009-06-16 15:32:22 -07001376
1377 h->surplus_huge_pages += delta;
1378 h->surplus_huge_pages_node[nid] += delta;
1379 ret = 1;
1380 break;
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001381 } while (next_nid != start_nid);
Wu Fengguang20a03072009-06-16 15:32:22 -07001382
Wu Fengguang20a03072009-06-16 15:32:22 -07001383 return ret;
1384}
1385
Andi Kleena5516432008-07-23 21:27:41 -07001386#define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001387static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count,
1388 nodemask_t *nodes_allowed)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001389{
Adam Litke7893d1d2007-10-16 01:26:18 -07001390 unsigned long min_count, ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001391
Andi Kleenaa888a72008-07-23 21:27:47 -07001392 if (h->order >= MAX_ORDER)
1393 return h->max_huge_pages;
1394
Adam Litke7893d1d2007-10-16 01:26:18 -07001395 /*
1396 * Increase the pool size
1397 * First take pages out of surplus state. Then make up the
1398 * remaining difference by allocating fresh huge pages.
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001399 *
1400 * We might race with alloc_buddy_huge_page() here and be unable
1401 * to convert a surplus huge page to a normal huge page. That is
1402 * not critical, though, it just means the overall size of the
1403 * pool might be one hugepage larger than it needs to be, but
1404 * within all the constraints specified by the sysctls.
Adam Litke7893d1d2007-10-16 01:26:18 -07001405 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001406 spin_lock(&hugetlb_lock);
Andi Kleena5516432008-07-23 21:27:41 -07001407 while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001408 if (!adjust_pool_surplus(h, nodes_allowed, -1))
Adam Litke7893d1d2007-10-16 01:26:18 -07001409 break;
1410 }
1411
Andi Kleena5516432008-07-23 21:27:41 -07001412 while (count > persistent_huge_pages(h)) {
Adam Litke7893d1d2007-10-16 01:26:18 -07001413 /*
1414 * If this allocation races such that we no longer need the
1415 * page, free_huge_page will handle it by freeing the page
1416 * and reducing the surplus.
1417 */
1418 spin_unlock(&hugetlb_lock);
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001419 ret = alloc_fresh_huge_page(h, nodes_allowed);
Adam Litke7893d1d2007-10-16 01:26:18 -07001420 spin_lock(&hugetlb_lock);
1421 if (!ret)
1422 goto out;
1423
Mel Gorman536240f22009-12-14 17:59:56 -08001424 /* Bail for signals. Probably ctrl-c from user */
1425 if (signal_pending(current))
1426 goto out;
Adam Litke7893d1d2007-10-16 01:26:18 -07001427 }
Adam Litke7893d1d2007-10-16 01:26:18 -07001428
1429 /*
1430 * Decrease the pool size
1431 * First return free pages to the buddy allocator (being careful
1432 * to keep enough around to satisfy reservations). Then place
1433 * pages into surplus state as needed so the pool will shrink
1434 * to the desired size as pages become free.
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001435 *
1436 * By placing pages into the surplus state independent of the
1437 * overcommit value, we are allowing the surplus pool size to
1438 * exceed overcommit. There are few sane options here. Since
1439 * alloc_buddy_huge_page() is checking the global counter,
1440 * though, we'll note that we're not allowed to exceed surplus
1441 * and won't grow the pool anywhere else. Not until one of the
1442 * sysctls are changed, or the surplus pages go out of use.
Adam Litke7893d1d2007-10-16 01:26:18 -07001443 */
Andi Kleena5516432008-07-23 21:27:41 -07001444 min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
Adam Litke6b0c8802007-10-16 01:26:23 -07001445 min_count = max(count, min_count);
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001446 try_to_free_low(h, min_count, nodes_allowed);
Andi Kleena5516432008-07-23 21:27:41 -07001447 while (min_count < persistent_huge_pages(h)) {
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001448 if (!free_pool_huge_page(h, nodes_allowed, 0))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001449 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001450 }
Andi Kleena5516432008-07-23 21:27:41 -07001451 while (count < persistent_huge_pages(h)) {
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001452 if (!adjust_pool_surplus(h, nodes_allowed, 1))
Adam Litke7893d1d2007-10-16 01:26:18 -07001453 break;
1454 }
1455out:
Andi Kleena5516432008-07-23 21:27:41 -07001456 ret = persistent_huge_pages(h);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001457 spin_unlock(&hugetlb_lock);
Adam Litke7893d1d2007-10-16 01:26:18 -07001458 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001459}
1460
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001461#define HSTATE_ATTR_RO(_name) \
1462 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
1463
1464#define HSTATE_ATTR(_name) \
1465 static struct kobj_attribute _name##_attr = \
1466 __ATTR(_name, 0644, _name##_show, _name##_store)
1467
1468static struct kobject *hugepages_kobj;
1469static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
1470
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001471static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
1472
1473static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001474{
1475 int i;
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001476
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001477 for (i = 0; i < HUGE_MAX_HSTATE; i++)
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001478 if (hstate_kobjs[i] == kobj) {
1479 if (nidp)
1480 *nidp = NUMA_NO_NODE;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001481 return &hstates[i];
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001482 }
1483
1484 return kobj_to_node_hstate(kobj, nidp);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001485}
1486
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001487static ssize_t nr_hugepages_show_common(struct kobject *kobj,
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001488 struct kobj_attribute *attr, char *buf)
1489{
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001490 struct hstate *h;
1491 unsigned long nr_huge_pages;
1492 int nid;
1493
1494 h = kobj_to_hstate(kobj, &nid);
1495 if (nid == NUMA_NO_NODE)
1496 nr_huge_pages = h->nr_huge_pages;
1497 else
1498 nr_huge_pages = h->nr_huge_pages_node[nid];
1499
1500 return sprintf(buf, "%lu\n", nr_huge_pages);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001501}
Eric B Munsonadbe8722011-01-13 15:47:27 -08001502
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001503static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
1504 struct kobject *kobj, struct kobj_attribute *attr,
1505 const char *buf, size_t len)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001506{
1507 int err;
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001508 int nid;
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001509 unsigned long count;
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001510 struct hstate *h;
David Rientjesbad44b52009-12-14 17:58:38 -08001511 NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001512
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001513 err = strict_strtoul(buf, 10, &count);
Eric B Munson73ae31e2011-01-13 15:47:28 -08001514 if (err)
Eric B Munsonadbe8722011-01-13 15:47:27 -08001515 goto out;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001516
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001517 h = kobj_to_hstate(kobj, &nid);
Eric B Munsonadbe8722011-01-13 15:47:27 -08001518 if (h->order >= MAX_ORDER) {
1519 err = -EINVAL;
1520 goto out;
1521 }
1522
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001523 if (nid == NUMA_NO_NODE) {
1524 /*
1525 * global hstate attribute
1526 */
1527 if (!(obey_mempolicy &&
1528 init_nodemask_of_mempolicy(nodes_allowed))) {
1529 NODEMASK_FREE(nodes_allowed);
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08001530 nodes_allowed = &node_states[N_MEMORY];
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001531 }
1532 } else if (nodes_allowed) {
1533 /*
1534 * per node hstate attribute: adjust count to global,
1535 * but restrict alloc/free to the specified node.
1536 */
1537 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
1538 init_nodemask_of_node(nodes_allowed, nid);
1539 } else
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08001540 nodes_allowed = &node_states[N_MEMORY];
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001541
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001542 h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001543
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08001544 if (nodes_allowed != &node_states[N_MEMORY])
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001545 NODEMASK_FREE(nodes_allowed);
1546
1547 return len;
Eric B Munsonadbe8722011-01-13 15:47:27 -08001548out:
1549 NODEMASK_FREE(nodes_allowed);
1550 return err;
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001551}
1552
1553static ssize_t nr_hugepages_show(struct kobject *kobj,
1554 struct kobj_attribute *attr, char *buf)
1555{
1556 return nr_hugepages_show_common(kobj, attr, buf);
1557}
1558
1559static ssize_t nr_hugepages_store(struct kobject *kobj,
1560 struct kobj_attribute *attr, const char *buf, size_t len)
1561{
1562 return nr_hugepages_store_common(false, kobj, attr, buf, len);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001563}
1564HSTATE_ATTR(nr_hugepages);
1565
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001566#ifdef CONFIG_NUMA
1567
1568/*
1569 * hstate attribute for optionally mempolicy-based constraint on persistent
1570 * huge page alloc/free.
1571 */
1572static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
1573 struct kobj_attribute *attr, char *buf)
1574{
1575 return nr_hugepages_show_common(kobj, attr, buf);
1576}
1577
1578static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
1579 struct kobj_attribute *attr, const char *buf, size_t len)
1580{
1581 return nr_hugepages_store_common(true, kobj, attr, buf, len);
1582}
1583HSTATE_ATTR(nr_hugepages_mempolicy);
1584#endif
1585
1586
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001587static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
1588 struct kobj_attribute *attr, char *buf)
1589{
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001590 struct hstate *h = kobj_to_hstate(kobj, NULL);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001591 return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
1592}
Eric B Munsonadbe8722011-01-13 15:47:27 -08001593
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001594static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
1595 struct kobj_attribute *attr, const char *buf, size_t count)
1596{
1597 int err;
1598 unsigned long input;
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001599 struct hstate *h = kobj_to_hstate(kobj, NULL);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001600
Eric B Munsonadbe8722011-01-13 15:47:27 -08001601 if (h->order >= MAX_ORDER)
1602 return -EINVAL;
1603
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001604 err = strict_strtoul(buf, 10, &input);
1605 if (err)
Eric B Munson73ae31e2011-01-13 15:47:28 -08001606 return err;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001607
1608 spin_lock(&hugetlb_lock);
1609 h->nr_overcommit_huge_pages = input;
1610 spin_unlock(&hugetlb_lock);
1611
1612 return count;
1613}
1614HSTATE_ATTR(nr_overcommit_hugepages);
1615
1616static ssize_t free_hugepages_show(struct kobject *kobj,
1617 struct kobj_attribute *attr, char *buf)
1618{
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001619 struct hstate *h;
1620 unsigned long free_huge_pages;
1621 int nid;
1622
1623 h = kobj_to_hstate(kobj, &nid);
1624 if (nid == NUMA_NO_NODE)
1625 free_huge_pages = h->free_huge_pages;
1626 else
1627 free_huge_pages = h->free_huge_pages_node[nid];
1628
1629 return sprintf(buf, "%lu\n", free_huge_pages);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001630}
1631HSTATE_ATTR_RO(free_hugepages);
1632
1633static ssize_t resv_hugepages_show(struct kobject *kobj,
1634 struct kobj_attribute *attr, char *buf)
1635{
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001636 struct hstate *h = kobj_to_hstate(kobj, NULL);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001637 return sprintf(buf, "%lu\n", h->resv_huge_pages);
1638}
1639HSTATE_ATTR_RO(resv_hugepages);
1640
1641static ssize_t surplus_hugepages_show(struct kobject *kobj,
1642 struct kobj_attribute *attr, char *buf)
1643{
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001644 struct hstate *h;
1645 unsigned long surplus_huge_pages;
1646 int nid;
1647
1648 h = kobj_to_hstate(kobj, &nid);
1649 if (nid == NUMA_NO_NODE)
1650 surplus_huge_pages = h->surplus_huge_pages;
1651 else
1652 surplus_huge_pages = h->surplus_huge_pages_node[nid];
1653
1654 return sprintf(buf, "%lu\n", surplus_huge_pages);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001655}
1656HSTATE_ATTR_RO(surplus_hugepages);
1657
1658static struct attribute *hstate_attrs[] = {
1659 &nr_hugepages_attr.attr,
1660 &nr_overcommit_hugepages_attr.attr,
1661 &free_hugepages_attr.attr,
1662 &resv_hugepages_attr.attr,
1663 &surplus_hugepages_attr.attr,
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001664#ifdef CONFIG_NUMA
1665 &nr_hugepages_mempolicy_attr.attr,
1666#endif
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001667 NULL,
1668};
1669
1670static struct attribute_group hstate_attr_group = {
1671 .attrs = hstate_attrs,
1672};
1673
Jeff Mahoney094e9532010-02-02 13:44:14 -08001674static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
1675 struct kobject **hstate_kobjs,
1676 struct attribute_group *hstate_attr_group)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001677{
1678 int retval;
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07001679 int hi = hstate_index(h);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001680
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001681 hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
1682 if (!hstate_kobjs[hi])
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001683 return -ENOMEM;
1684
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001685 retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001686 if (retval)
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001687 kobject_put(hstate_kobjs[hi]);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001688
1689 return retval;
1690}
1691
1692static void __init hugetlb_sysfs_init(void)
1693{
1694 struct hstate *h;
1695 int err;
1696
1697 hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
1698 if (!hugepages_kobj)
1699 return;
1700
1701 for_each_hstate(h) {
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001702 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
1703 hstate_kobjs, &hstate_attr_group);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001704 if (err)
1705 printk(KERN_ERR "Hugetlb: Unable to add hstate %s",
1706 h->name);
1707 }
1708}
1709
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001710#ifdef CONFIG_NUMA
1711
1712/*
1713 * node_hstate/s - associate per node hstate attributes, via their kobjects,
Kay Sievers10fbcf42011-12-21 14:48:43 -08001714 * with node devices in node_devices[] using a parallel array. The array
1715 * index of a node device or _hstate == node id.
1716 * This is here to avoid any static dependency of the node device driver, in
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001717 * the base kernel, on the hugetlb module.
1718 */
1719struct node_hstate {
1720 struct kobject *hugepages_kobj;
1721 struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
1722};
1723struct node_hstate node_hstates[MAX_NUMNODES];
1724
1725/*
Kay Sievers10fbcf42011-12-21 14:48:43 -08001726 * A subset of global hstate attributes for node devices
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001727 */
1728static struct attribute *per_node_hstate_attrs[] = {
1729 &nr_hugepages_attr.attr,
1730 &free_hugepages_attr.attr,
1731 &surplus_hugepages_attr.attr,
1732 NULL,
1733};
1734
1735static struct attribute_group per_node_hstate_attr_group = {
1736 .attrs = per_node_hstate_attrs,
1737};
1738
1739/*
Kay Sievers10fbcf42011-12-21 14:48:43 -08001740 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001741 * Returns node id via non-NULL nidp.
1742 */
1743static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
1744{
1745 int nid;
1746
1747 for (nid = 0; nid < nr_node_ids; nid++) {
1748 struct node_hstate *nhs = &node_hstates[nid];
1749 int i;
1750 for (i = 0; i < HUGE_MAX_HSTATE; i++)
1751 if (nhs->hstate_kobjs[i] == kobj) {
1752 if (nidp)
1753 *nidp = nid;
1754 return &hstates[i];
1755 }
1756 }
1757
1758 BUG();
1759 return NULL;
1760}
1761
1762/*
Kay Sievers10fbcf42011-12-21 14:48:43 -08001763 * Unregister hstate attributes from a single node device.
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001764 * No-op if no hstate attributes attached.
1765 */
1766void hugetlb_unregister_node(struct node *node)
1767{
1768 struct hstate *h;
Kay Sievers10fbcf42011-12-21 14:48:43 -08001769 struct node_hstate *nhs = &node_hstates[node->dev.id];
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001770
1771 if (!nhs->hugepages_kobj)
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -08001772 return; /* no hstate attributes */
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001773
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07001774 for_each_hstate(h) {
1775 int idx = hstate_index(h);
1776 if (nhs->hstate_kobjs[idx]) {
1777 kobject_put(nhs->hstate_kobjs[idx]);
1778 nhs->hstate_kobjs[idx] = NULL;
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001779 }
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07001780 }
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001781
1782 kobject_put(nhs->hugepages_kobj);
1783 nhs->hugepages_kobj = NULL;
1784}
1785
1786/*
Kay Sievers10fbcf42011-12-21 14:48:43 -08001787 * hugetlb module exit: unregister hstate attributes from node devices
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001788 * that have them.
1789 */
1790static void hugetlb_unregister_all_nodes(void)
1791{
1792 int nid;
1793
1794 /*
Kay Sievers10fbcf42011-12-21 14:48:43 -08001795 * disable node device registrations.
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001796 */
1797 register_hugetlbfs_with_node(NULL, NULL);
1798
1799 /*
1800 * remove hstate attributes from any nodes that have them.
1801 */
1802 for (nid = 0; nid < nr_node_ids; nid++)
Wen Congyang87327942012-12-11 16:00:56 -08001803 hugetlb_unregister_node(node_devices[nid]);
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001804}
1805
1806/*
Kay Sievers10fbcf42011-12-21 14:48:43 -08001807 * Register hstate attributes for a single node device.
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001808 * No-op if attributes already registered.
1809 */
1810void hugetlb_register_node(struct node *node)
1811{
1812 struct hstate *h;
Kay Sievers10fbcf42011-12-21 14:48:43 -08001813 struct node_hstate *nhs = &node_hstates[node->dev.id];
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001814 int err;
1815
1816 if (nhs->hugepages_kobj)
1817 return; /* already allocated */
1818
1819 nhs->hugepages_kobj = kobject_create_and_add("hugepages",
Kay Sievers10fbcf42011-12-21 14:48:43 -08001820 &node->dev.kobj);
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001821 if (!nhs->hugepages_kobj)
1822 return;
1823
1824 for_each_hstate(h) {
1825 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
1826 nhs->hstate_kobjs,
1827 &per_node_hstate_attr_group);
1828 if (err) {
1829 printk(KERN_ERR "Hugetlb: Unable to add hstate %s"
1830 " for node %d\n",
Kay Sievers10fbcf42011-12-21 14:48:43 -08001831 h->name, node->dev.id);
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001832 hugetlb_unregister_node(node);
1833 break;
1834 }
1835 }
1836}
1837
1838/*
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -08001839 * hugetlb init time: register hstate attributes for all registered node
Kay Sievers10fbcf42011-12-21 14:48:43 -08001840 * devices of nodes that have memory. All on-line nodes should have
1841 * registered their associated device by this time.
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001842 */
1843static void hugetlb_register_all_nodes(void)
1844{
1845 int nid;
1846
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08001847 for_each_node_state(nid, N_MEMORY) {
Wen Congyang87327942012-12-11 16:00:56 -08001848 struct node *node = node_devices[nid];
Kay Sievers10fbcf42011-12-21 14:48:43 -08001849 if (node->dev.id == nid)
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001850 hugetlb_register_node(node);
1851 }
1852
1853 /*
Kay Sievers10fbcf42011-12-21 14:48:43 -08001854 * Let the node device driver know we're here so it can
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001855 * [un]register hstate attributes on node hotplug.
1856 */
1857 register_hugetlbfs_with_node(hugetlb_register_node,
1858 hugetlb_unregister_node);
1859}
1860#else /* !CONFIG_NUMA */
1861
1862static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
1863{
1864 BUG();
1865 if (nidp)
1866 *nidp = -1;
1867 return NULL;
1868}
1869
1870static void hugetlb_unregister_all_nodes(void) { }
1871
1872static void hugetlb_register_all_nodes(void) { }
1873
1874#endif
1875
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001876static void __exit hugetlb_exit(void)
1877{
1878 struct hstate *h;
1879
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001880 hugetlb_unregister_all_nodes();
1881
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001882 for_each_hstate(h) {
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07001883 kobject_put(hstate_kobjs[hstate_index(h)]);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001884 }
1885
1886 kobject_put(hugepages_kobj);
1887}
1888module_exit(hugetlb_exit);
1889
1890static int __init hugetlb_init(void)
1891{
Benjamin Herrenschmidt0ef89d22008-07-31 00:07:30 -07001892 /* Some platform decide whether they support huge pages at boot
1893 * time. On these, such as powerpc, HPAGE_SHIFT is set to 0 when
1894 * there is no such support
1895 */
1896 if (HPAGE_SHIFT == 0)
1897 return 0;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001898
Nick Piggine11bfbf2008-07-23 21:27:52 -07001899 if (!size_to_hstate(default_hstate_size)) {
1900 default_hstate_size = HPAGE_SIZE;
1901 if (!size_to_hstate(default_hstate_size))
1902 hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001903 }
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07001904 default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size));
Nick Piggine11bfbf2008-07-23 21:27:52 -07001905 if (default_hstate_max_huge_pages)
1906 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001907
1908 hugetlb_init_hstates();
Andi Kleenaa888a72008-07-23 21:27:47 -07001909 gather_bootmem_prealloc();
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001910 report_hugepages();
1911
1912 hugetlb_sysfs_init();
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001913 hugetlb_register_all_nodes();
Jianguo Wu7179e7b2012-12-18 14:23:19 -08001914 hugetlb_cgroup_file_init();
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001915
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001916 return 0;
1917}
1918module_init(hugetlb_init);
1919
1920/* Should be called on processing a hugepagesz=... option */
1921void __init hugetlb_add_hstate(unsigned order)
1922{
1923 struct hstate *h;
Andi Kleen8faa8b02008-07-23 21:27:48 -07001924 unsigned long i;
1925
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001926 if (size_to_hstate(PAGE_SIZE << order)) {
1927 printk(KERN_WARNING "hugepagesz= specified twice, ignoring\n");
1928 return;
1929 }
Aneesh Kumar K.V47d38342012-07-31 16:41:54 -07001930 BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001931 BUG_ON(order == 0);
Aneesh Kumar K.V47d38342012-07-31 16:41:54 -07001932 h = &hstates[hugetlb_max_hstate++];
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001933 h->order = order;
1934 h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
Andi Kleen8faa8b02008-07-23 21:27:48 -07001935 h->nr_huge_pages = 0;
1936 h->free_huge_pages = 0;
1937 for (i = 0; i < MAX_NUMNODES; ++i)
1938 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
Aneesh Kumar K.V0edaecf2012-07-31 16:42:07 -07001939 INIT_LIST_HEAD(&h->hugepage_activelist);
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08001940 h->next_nid_to_alloc = first_node(node_states[N_MEMORY]);
1941 h->next_nid_to_free = first_node(node_states[N_MEMORY]);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001942 snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
1943 huge_page_size(h)/1024);
Andi Kleen8faa8b02008-07-23 21:27:48 -07001944
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001945 parsed_hstate = h;
1946}
1947
Nick Piggine11bfbf2008-07-23 21:27:52 -07001948static int __init hugetlb_nrpages_setup(char *s)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001949{
1950 unsigned long *mhp;
Andi Kleen8faa8b02008-07-23 21:27:48 -07001951 static unsigned long *last_mhp;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001952
1953 /*
Aneesh Kumar K.V47d38342012-07-31 16:41:54 -07001954 * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet,
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001955 * so this hugepages= parameter goes to the "default hstate".
1956 */
Aneesh Kumar K.V47d38342012-07-31 16:41:54 -07001957 if (!hugetlb_max_hstate)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001958 mhp = &default_hstate_max_huge_pages;
1959 else
1960 mhp = &parsed_hstate->max_huge_pages;
1961
Andi Kleen8faa8b02008-07-23 21:27:48 -07001962 if (mhp == last_mhp) {
1963 printk(KERN_WARNING "hugepages= specified twice without "
1964 "interleaving hugepagesz=, ignoring\n");
1965 return 1;
1966 }
1967
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001968 if (sscanf(s, "%lu", mhp) <= 0)
1969 *mhp = 0;
1970
Andi Kleen8faa8b02008-07-23 21:27:48 -07001971 /*
1972 * Global state is always initialized later in hugetlb_init.
1973 * But we need to allocate >= MAX_ORDER hstates here early to still
1974 * use the bootmem allocator.
1975 */
Aneesh Kumar K.V47d38342012-07-31 16:41:54 -07001976 if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER)
Andi Kleen8faa8b02008-07-23 21:27:48 -07001977 hugetlb_hstate_alloc_pages(parsed_hstate);
1978
1979 last_mhp = mhp;
1980
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001981 return 1;
1982}
Nick Piggine11bfbf2008-07-23 21:27:52 -07001983__setup("hugepages=", hugetlb_nrpages_setup);
1984
1985static int __init hugetlb_default_setup(char *s)
1986{
1987 default_hstate_size = memparse(s, &s);
1988 return 1;
1989}
1990__setup("default_hugepagesz=", hugetlb_default_setup);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001991
Nishanth Aravamudan8a213462008-07-25 19:44:37 -07001992static unsigned int cpuset_mems_nr(unsigned int *array)
1993{
1994 int node;
1995 unsigned int nr = 0;
1996
1997 for_each_node_mask(node, cpuset_current_mems_allowed)
1998 nr += array[node];
1999
2000 return nr;
2001}
2002
2003#ifdef CONFIG_SYSCTL
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002004static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
2005 struct ctl_table *table, int write,
2006 void __user *buffer, size_t *length, loff_t *ppos)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002007{
Andi Kleene5ff2152008-07-23 21:27:42 -07002008 struct hstate *h = &default_hstate;
2009 unsigned long tmp;
Michal Hocko08d4a242011-01-13 15:47:26 -08002010 int ret;
Andi Kleene5ff2152008-07-23 21:27:42 -07002011
Petr Holasekc033a932011-03-22 16:33:05 -07002012 tmp = h->max_huge_pages;
Andi Kleene5ff2152008-07-23 21:27:42 -07002013
Eric B Munsonadbe8722011-01-13 15:47:27 -08002014 if (write && h->order >= MAX_ORDER)
2015 return -EINVAL;
2016
Andi Kleene5ff2152008-07-23 21:27:42 -07002017 table->data = &tmp;
2018 table->maxlen = sizeof(unsigned long);
Michal Hocko08d4a242011-01-13 15:47:26 -08002019 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2020 if (ret)
2021 goto out;
Andi Kleene5ff2152008-07-23 21:27:42 -07002022
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002023 if (write) {
David Rientjesbad44b52009-12-14 17:58:38 -08002024 NODEMASK_ALLOC(nodemask_t, nodes_allowed,
2025 GFP_KERNEL | __GFP_NORETRY);
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002026 if (!(obey_mempolicy &&
2027 init_nodemask_of_mempolicy(nodes_allowed))) {
2028 NODEMASK_FREE(nodes_allowed);
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08002029 nodes_allowed = &node_states[N_MEMORY];
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002030 }
2031 h->max_huge_pages = set_max_huge_pages(h, tmp, nodes_allowed);
2032
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08002033 if (nodes_allowed != &node_states[N_MEMORY])
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002034 NODEMASK_FREE(nodes_allowed);
2035 }
Michal Hocko08d4a242011-01-13 15:47:26 -08002036out:
2037 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002038}
Mel Gorman396faf02007-07-17 04:03:13 -07002039
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002040int hugetlb_sysctl_handler(struct ctl_table *table, int write,
2041 void __user *buffer, size_t *length, loff_t *ppos)
2042{
2043
2044 return hugetlb_sysctl_handler_common(false, table, write,
2045 buffer, length, ppos);
2046}
2047
2048#ifdef CONFIG_NUMA
2049int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
2050 void __user *buffer, size_t *length, loff_t *ppos)
2051{
2052 return hugetlb_sysctl_handler_common(true, table, write,
2053 buffer, length, ppos);
2054}
2055#endif /* CONFIG_NUMA */
2056
Mel Gorman396faf02007-07-17 04:03:13 -07002057int hugetlb_treat_movable_handler(struct ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07002058 void __user *buffer,
Mel Gorman396faf02007-07-17 04:03:13 -07002059 size_t *length, loff_t *ppos)
2060{
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07002061 proc_dointvec(table, write, buffer, length, ppos);
Mel Gorman396faf02007-07-17 04:03:13 -07002062 if (hugepages_treat_as_movable)
2063 htlb_alloc_mask = GFP_HIGHUSER_MOVABLE;
2064 else
2065 htlb_alloc_mask = GFP_HIGHUSER;
2066 return 0;
2067}
2068
Nishanth Aravamudana3d0c6a2008-02-08 04:18:18 -08002069int hugetlb_overcommit_handler(struct ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07002070 void __user *buffer,
Nishanth Aravamudana3d0c6a2008-02-08 04:18:18 -08002071 size_t *length, loff_t *ppos)
2072{
Andi Kleena5516432008-07-23 21:27:41 -07002073 struct hstate *h = &default_hstate;
Andi Kleene5ff2152008-07-23 21:27:42 -07002074 unsigned long tmp;
Michal Hocko08d4a242011-01-13 15:47:26 -08002075 int ret;
Andi Kleene5ff2152008-07-23 21:27:42 -07002076
Petr Holasekc033a932011-03-22 16:33:05 -07002077 tmp = h->nr_overcommit_huge_pages;
Andi Kleene5ff2152008-07-23 21:27:42 -07002078
Eric B Munsonadbe8722011-01-13 15:47:27 -08002079 if (write && h->order >= MAX_ORDER)
2080 return -EINVAL;
2081
Andi Kleene5ff2152008-07-23 21:27:42 -07002082 table->data = &tmp;
2083 table->maxlen = sizeof(unsigned long);
Michal Hocko08d4a242011-01-13 15:47:26 -08002084 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2085 if (ret)
2086 goto out;
Andi Kleene5ff2152008-07-23 21:27:42 -07002087
2088 if (write) {
2089 spin_lock(&hugetlb_lock);
2090 h->nr_overcommit_huge_pages = tmp;
2091 spin_unlock(&hugetlb_lock);
2092 }
Michal Hocko08d4a242011-01-13 15:47:26 -08002093out:
2094 return ret;
Nishanth Aravamudana3d0c6a2008-02-08 04:18:18 -08002095}
2096
Linus Torvalds1da177e2005-04-16 15:20:36 -07002097#endif /* CONFIG_SYSCTL */
2098
Alexey Dobriyane1759c22008-10-15 23:50:22 +04002099void hugetlb_report_meminfo(struct seq_file *m)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002100{
Andi Kleena5516432008-07-23 21:27:41 -07002101 struct hstate *h = &default_hstate;
Alexey Dobriyane1759c22008-10-15 23:50:22 +04002102 seq_printf(m,
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002103 "HugePages_Total: %5lu\n"
2104 "HugePages_Free: %5lu\n"
2105 "HugePages_Rsvd: %5lu\n"
2106 "HugePages_Surp: %5lu\n"
2107 "Hugepagesize: %8lu kB\n",
Andi Kleena5516432008-07-23 21:27:41 -07002108 h->nr_huge_pages,
2109 h->free_huge_pages,
2110 h->resv_huge_pages,
2111 h->surplus_huge_pages,
2112 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002113}
2114
2115int hugetlb_report_node_meminfo(int nid, char *buf)
2116{
Andi Kleena5516432008-07-23 21:27:41 -07002117 struct hstate *h = &default_hstate;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002118 return sprintf(buf,
2119 "Node %d HugePages_Total: %5u\n"
Nishanth Aravamudana1de0912008-03-26 14:37:53 -07002120 "Node %d HugePages_Free: %5u\n"
2121 "Node %d HugePages_Surp: %5u\n",
Andi Kleena5516432008-07-23 21:27:41 -07002122 nid, h->nr_huge_pages_node[nid],
2123 nid, h->free_huge_pages_node[nid],
2124 nid, h->surplus_huge_pages_node[nid]);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002125}
2126
Linus Torvalds1da177e2005-04-16 15:20:36 -07002127/* Return the number pages of memory we physically have, in PAGE_SIZE units. */
2128unsigned long hugetlb_total_pages(void)
2129{
Andi Kleena5516432008-07-23 21:27:41 -07002130 struct hstate *h = &default_hstate;
2131 return h->nr_huge_pages * pages_per_huge_page(h);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002132}
Linus Torvalds1da177e2005-04-16 15:20:36 -07002133
Andi Kleena5516432008-07-23 21:27:41 -07002134static int hugetlb_acct_memory(struct hstate *h, long delta)
Mel Gormanfc1b8a72008-07-23 21:27:22 -07002135{
2136 int ret = -ENOMEM;
2137
2138 spin_lock(&hugetlb_lock);
2139 /*
2140 * When cpuset is configured, it breaks the strict hugetlb page
2141 * reservation as the accounting is done on a global variable. Such
2142 * reservation is completely rubbish in the presence of cpuset because
2143 * the reservation is not checked against page availability for the
2144 * current cpuset. Application can still potentially OOM'ed by kernel
2145 * with lack of free htlb page in cpuset that the task is in.
2146 * Attempt to enforce strict accounting with cpuset is almost
2147 * impossible (or too ugly) because cpuset is too fluid that
2148 * task or memory node can be dynamically moved between cpusets.
2149 *
2150 * The change of semantics for shared hugetlb mapping with cpuset is
2151 * undesirable. However, in order to preserve some of the semantics,
2152 * we fall back to check against current free page availability as
2153 * a best attempt and hopefully to minimize the impact of changing
2154 * semantics that cpuset has.
2155 */
2156 if (delta > 0) {
Andi Kleena5516432008-07-23 21:27:41 -07002157 if (gather_surplus_pages(h, delta) < 0)
Mel Gormanfc1b8a72008-07-23 21:27:22 -07002158 goto out;
2159
Andi Kleena5516432008-07-23 21:27:41 -07002160 if (delta > cpuset_mems_nr(h->free_huge_pages_node)) {
2161 return_unused_surplus_pages(h, delta);
Mel Gormanfc1b8a72008-07-23 21:27:22 -07002162 goto out;
2163 }
2164 }
2165
2166 ret = 0;
2167 if (delta < 0)
Andi Kleena5516432008-07-23 21:27:41 -07002168 return_unused_surplus_pages(h, (unsigned long) -delta);
Mel Gormanfc1b8a72008-07-23 21:27:22 -07002169
2170out:
2171 spin_unlock(&hugetlb_lock);
2172 return ret;
2173}
2174
Andy Whitcroft84afd992008-07-23 21:27:32 -07002175static void hugetlb_vm_op_open(struct vm_area_struct *vma)
2176{
2177 struct resv_map *reservations = vma_resv_map(vma);
2178
2179 /*
2180 * This new VMA should share its siblings reservation map if present.
2181 * The VMA will only ever have a valid reservation map pointer where
2182 * it is being copied for another still existing VMA. As that VMA
Lucas De Marchi25985ed2011-03-30 22:57:33 -03002183 * has a reference to the reservation map it cannot disappear until
Andy Whitcroft84afd992008-07-23 21:27:32 -07002184 * after this open call completes. It is therefore safe to take a
2185 * new reference here without additional locking.
2186 */
2187 if (reservations)
2188 kref_get(&reservations->refs);
2189}
2190
Dave Hansenc50ac052012-05-29 15:06:46 -07002191static void resv_map_put(struct vm_area_struct *vma)
2192{
2193 struct resv_map *reservations = vma_resv_map(vma);
2194
2195 if (!reservations)
2196 return;
2197 kref_put(&reservations->refs, resv_map_release);
2198}
2199
Mel Gormana1e78772008-07-23 21:27:23 -07002200static void hugetlb_vm_op_close(struct vm_area_struct *vma)
2201{
Andi Kleena5516432008-07-23 21:27:41 -07002202 struct hstate *h = hstate_vma(vma);
Andy Whitcroft84afd992008-07-23 21:27:32 -07002203 struct resv_map *reservations = vma_resv_map(vma);
David Gibson90481622012-03-21 16:34:12 -07002204 struct hugepage_subpool *spool = subpool_vma(vma);
Andy Whitcroft84afd992008-07-23 21:27:32 -07002205 unsigned long reserve;
2206 unsigned long start;
2207 unsigned long end;
2208
2209 if (reservations) {
Andi Kleena5516432008-07-23 21:27:41 -07002210 start = vma_hugecache_offset(h, vma, vma->vm_start);
2211 end = vma_hugecache_offset(h, vma, vma->vm_end);
Andy Whitcroft84afd992008-07-23 21:27:32 -07002212
2213 reserve = (end - start) -
2214 region_count(&reservations->regions, start, end);
2215
Dave Hansenc50ac052012-05-29 15:06:46 -07002216 resv_map_put(vma);
Andy Whitcroft84afd992008-07-23 21:27:32 -07002217
Adam Litke7251ff72008-07-23 21:27:59 -07002218 if (reserve) {
Andi Kleena5516432008-07-23 21:27:41 -07002219 hugetlb_acct_memory(h, -reserve);
David Gibson90481622012-03-21 16:34:12 -07002220 hugepage_subpool_put_pages(spool, reserve);
Adam Litke7251ff72008-07-23 21:27:59 -07002221 }
Andy Whitcroft84afd992008-07-23 21:27:32 -07002222 }
Mel Gormana1e78772008-07-23 21:27:23 -07002223}
2224
Linus Torvalds1da177e2005-04-16 15:20:36 -07002225/*
2226 * We cannot handle pagefaults against hugetlb pages at all. They cause
2227 * handle_mm_fault() to try to instantiate regular-sized pages in the
2228 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
2229 * this far.
2230 */
Nick Piggind0217ac2007-07-19 01:47:03 -07002231static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002232{
2233 BUG();
Nick Piggind0217ac2007-07-19 01:47:03 -07002234 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002235}
2236
Alexey Dobriyanf0f37e22009-09-27 22:29:37 +04002237const struct vm_operations_struct hugetlb_vm_ops = {
Nick Piggind0217ac2007-07-19 01:47:03 -07002238 .fault = hugetlb_vm_op_fault,
Andy Whitcroft84afd992008-07-23 21:27:32 -07002239 .open = hugetlb_vm_op_open,
Mel Gormana1e78772008-07-23 21:27:23 -07002240 .close = hugetlb_vm_op_close,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002241};
2242
David Gibson1e8f8892006-01-06 00:10:44 -08002243static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
2244 int writable)
David Gibson63551ae2005-06-21 17:14:44 -07002245{
2246 pte_t entry;
2247
David Gibson1e8f8892006-01-06 00:10:44 -08002248 if (writable) {
David Gibson63551ae2005-06-21 17:14:44 -07002249 entry =
2250 pte_mkwrite(pte_mkdirty(mk_pte(page, vma->vm_page_prot)));
2251 } else {
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07002252 entry = huge_pte_wrprotect(mk_pte(page, vma->vm_page_prot));
David Gibson63551ae2005-06-21 17:14:44 -07002253 }
2254 entry = pte_mkyoung(entry);
2255 entry = pte_mkhuge(entry);
Chris Metcalfd9ed9fa2012-04-01 14:01:34 -04002256 entry = arch_make_huge_pte(entry, vma, page, writable);
David Gibson63551ae2005-06-21 17:14:44 -07002257
2258 return entry;
2259}
2260
David Gibson1e8f8892006-01-06 00:10:44 -08002261static void set_huge_ptep_writable(struct vm_area_struct *vma,
2262 unsigned long address, pte_t *ptep)
2263{
2264 pte_t entry;
2265
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07002266 entry = pte_mkwrite(pte_mkdirty(huge_ptep_get(ptep)));
Chris Forbes32f84522011-07-25 17:12:14 -07002267 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
Russell King4b3073e2009-12-18 16:40:18 +00002268 update_mmu_cache(vma, address, ptep);
David Gibson1e8f8892006-01-06 00:10:44 -08002269}
2270
2271
David Gibson63551ae2005-06-21 17:14:44 -07002272int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
2273 struct vm_area_struct *vma)
2274{
2275 pte_t *src_pte, *dst_pte, entry;
2276 struct page *ptepage;
Hugh Dickins1c598272005-10-19 21:23:43 -07002277 unsigned long addr;
David Gibson1e8f8892006-01-06 00:10:44 -08002278 int cow;
Andi Kleena5516432008-07-23 21:27:41 -07002279 struct hstate *h = hstate_vma(vma);
2280 unsigned long sz = huge_page_size(h);
David Gibson1e8f8892006-01-06 00:10:44 -08002281
2282 cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
David Gibson63551ae2005-06-21 17:14:44 -07002283
Andi Kleena5516432008-07-23 21:27:41 -07002284 for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
Hugh Dickinsc74df322005-10-29 18:16:23 -07002285 src_pte = huge_pte_offset(src, addr);
2286 if (!src_pte)
2287 continue;
Andi Kleena5516432008-07-23 21:27:41 -07002288 dst_pte = huge_pte_alloc(dst, addr, sz);
David Gibson63551ae2005-06-21 17:14:44 -07002289 if (!dst_pte)
2290 goto nomem;
Larry Woodmanc5c99422008-01-24 05:49:25 -08002291
2292 /* If the pagetables are shared don't copy or take references */
2293 if (dst_pte == src_pte)
2294 continue;
2295
Hugh Dickinsc74df322005-10-29 18:16:23 -07002296 spin_lock(&dst->page_table_lock);
Nick Piggin46478752008-06-05 22:45:57 -07002297 spin_lock_nested(&src->page_table_lock, SINGLE_DEPTH_NESTING);
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07002298 if (!huge_pte_none(huge_ptep_get(src_pte))) {
David Gibson1e8f8892006-01-06 00:10:44 -08002299 if (cow)
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07002300 huge_ptep_set_wrprotect(src, addr, src_pte);
2301 entry = huge_ptep_get(src_pte);
Hugh Dickins1c598272005-10-19 21:23:43 -07002302 ptepage = pte_page(entry);
2303 get_page(ptepage);
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002304 page_dup_rmap(ptepage);
Hugh Dickins1c598272005-10-19 21:23:43 -07002305 set_huge_pte_at(dst, addr, dst_pte, entry);
2306 }
2307 spin_unlock(&src->page_table_lock);
Hugh Dickinsc74df322005-10-29 18:16:23 -07002308 spin_unlock(&dst->page_table_lock);
David Gibson63551ae2005-06-21 17:14:44 -07002309 }
2310 return 0;
2311
2312nomem:
2313 return -ENOMEM;
2314}
2315
Naoya Horiguchi290408d2010-09-08 10:19:35 +09002316static int is_hugetlb_entry_migration(pte_t pte)
2317{
2318 swp_entry_t swp;
2319
2320 if (huge_pte_none(pte) || pte_present(pte))
2321 return 0;
2322 swp = pte_to_swp_entry(pte);
Chris Forbes32f84522011-07-25 17:12:14 -07002323 if (non_swap_entry(swp) && is_migration_entry(swp))
Naoya Horiguchi290408d2010-09-08 10:19:35 +09002324 return 1;
Chris Forbes32f84522011-07-25 17:12:14 -07002325 else
Naoya Horiguchi290408d2010-09-08 10:19:35 +09002326 return 0;
2327}
2328
Naoya Horiguchifd6a03ed2010-05-28 09:29:21 +09002329static int is_hugetlb_entry_hwpoisoned(pte_t pte)
2330{
2331 swp_entry_t swp;
2332
2333 if (huge_pte_none(pte) || pte_present(pte))
2334 return 0;
2335 swp = pte_to_swp_entry(pte);
Chris Forbes32f84522011-07-25 17:12:14 -07002336 if (non_swap_entry(swp) && is_hwpoison_entry(swp))
Naoya Horiguchifd6a03ed2010-05-28 09:29:21 +09002337 return 1;
Chris Forbes32f84522011-07-25 17:12:14 -07002338 else
Naoya Horiguchifd6a03ed2010-05-28 09:29:21 +09002339 return 0;
2340}
2341
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07002342void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
2343 unsigned long start, unsigned long end,
2344 struct page *ref_page)
David Gibson63551ae2005-06-21 17:14:44 -07002345{
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07002346 int force_flush = 0;
David Gibson63551ae2005-06-21 17:14:44 -07002347 struct mm_struct *mm = vma->vm_mm;
2348 unsigned long address;
David Gibsonc7546f82005-08-05 11:59:35 -07002349 pte_t *ptep;
David Gibson63551ae2005-06-21 17:14:44 -07002350 pte_t pte;
2351 struct page *page;
Andi Kleena5516432008-07-23 21:27:41 -07002352 struct hstate *h = hstate_vma(vma);
2353 unsigned long sz = huge_page_size(h);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07002354 const unsigned long mmun_start = start; /* For mmu_notifiers */
2355 const unsigned long mmun_end = end; /* For mmu_notifiers */
Andi Kleena5516432008-07-23 21:27:41 -07002356
David Gibson63551ae2005-06-21 17:14:44 -07002357 WARN_ON(!is_vm_hugetlb_page(vma));
Andi Kleena5516432008-07-23 21:27:41 -07002358 BUG_ON(start & ~huge_page_mask(h));
2359 BUG_ON(end & ~huge_page_mask(h));
David Gibson63551ae2005-06-21 17:14:44 -07002360
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07002361 tlb_start_vma(tlb, vma);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07002362 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07002363again:
Hugh Dickins508034a2005-10-29 18:16:30 -07002364 spin_lock(&mm->page_table_lock);
Andi Kleena5516432008-07-23 21:27:41 -07002365 for (address = start; address < end; address += sz) {
David Gibsonc7546f82005-08-05 11:59:35 -07002366 ptep = huge_pte_offset(mm, address);
Adam Litke4c887262005-10-29 18:16:46 -07002367 if (!ptep)
David Gibsonc7546f82005-08-05 11:59:35 -07002368 continue;
2369
Chen, Kenneth W39dde652006-12-06 20:32:03 -08002370 if (huge_pmd_unshare(mm, &address, ptep))
2371 continue;
2372
Hillf Danton66293262012-03-23 15:01:48 -07002373 pte = huge_ptep_get(ptep);
2374 if (huge_pte_none(pte))
2375 continue;
2376
2377 /*
2378 * HWPoisoned hugepage is already unmapped and dropped reference
2379 */
Naoya Horiguchi8c4894c2012-12-12 13:52:28 -08002380 if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) {
2381 pte_clear(mm, address, ptep);
Hillf Danton66293262012-03-23 15:01:48 -07002382 continue;
Naoya Horiguchi8c4894c2012-12-12 13:52:28 -08002383 }
Hillf Danton66293262012-03-23 15:01:48 -07002384
2385 page = pte_page(pte);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002386 /*
2387 * If a reference page is supplied, it is because a specific
2388 * page is being unmapped, not a range. Ensure the page we
2389 * are about to unmap is the actual page of interest.
2390 */
2391 if (ref_page) {
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002392 if (page != ref_page)
2393 continue;
2394
2395 /*
2396 * Mark the VMA as having unmapped its page so that
2397 * future faults in this VMA will fail rather than
2398 * looking like data was lost
2399 */
2400 set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
2401 }
2402
David Gibsonc7546f82005-08-05 11:59:35 -07002403 pte = huge_ptep_get_and_clear(mm, address, ptep);
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07002404 tlb_remove_tlb_entry(tlb, ptep, address);
Ken Chen6649a382007-02-08 14:20:27 -08002405 if (pte_dirty(pte))
2406 set_page_dirty(page);
Hillf Danton9e811302012-03-21 16:34:03 -07002407
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07002408 page_remove_rmap(page);
2409 force_flush = !__tlb_remove_page(tlb, page);
2410 if (force_flush)
2411 break;
Hillf Danton9e811302012-03-21 16:34:03 -07002412 /* Bail out after unmapping reference page if supplied */
2413 if (ref_page)
2414 break;
David Gibson63551ae2005-06-21 17:14:44 -07002415 }
Al Virocd2934a2012-03-05 06:40:29 +00002416 spin_unlock(&mm->page_table_lock);
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07002417 /*
2418 * mmu_gather ran out of room to batch pages, we break out of
2419 * the PTE lock to avoid doing the potential expensive TLB invalidate
2420 * and page-free while holding it.
2421 */
2422 if (force_flush) {
2423 force_flush = 0;
2424 tlb_flush_mmu(tlb);
2425 if (address < end && !ref_page)
2426 goto again;
Chen, Kenneth Wfe1668a2006-10-04 02:15:24 -07002427 }
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07002428 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07002429 tlb_end_vma(tlb, vma);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002430}
David Gibson63551ae2005-06-21 17:14:44 -07002431
Mel Gormand8333522012-07-31 16:46:20 -07002432void __unmap_hugepage_range_final(struct mmu_gather *tlb,
2433 struct vm_area_struct *vma, unsigned long start,
2434 unsigned long end, struct page *ref_page)
2435{
2436 __unmap_hugepage_range(tlb, vma, start, end, ref_page);
2437
2438 /*
2439 * Clear this flag so that x86's huge_pmd_share page_table_shareable
2440 * test will fail on a vma being torn down, and not grab a page table
2441 * on its way out. We're lucky that the flag has such an appropriate
2442 * name, and can in fact be safely cleared here. We could clear it
2443 * before the __unmap_hugepage_range above, but all that's necessary
2444 * is to clear it before releasing the i_mmap_mutex. This works
2445 * because in the context this is called, the VMA is about to be
2446 * destroyed and the i_mmap_mutex is held.
2447 */
2448 vma->vm_flags &= ~VM_MAYSHARE;
2449}
2450
Chen, Kenneth W502717f2006-10-11 01:20:46 -07002451void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002452 unsigned long end, struct page *ref_page)
Chen, Kenneth W502717f2006-10-11 01:20:46 -07002453{
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07002454 struct mm_struct *mm;
2455 struct mmu_gather tlb;
2456
2457 mm = vma->vm_mm;
2458
2459 tlb_gather_mmu(&tlb, mm, 0);
2460 __unmap_hugepage_range(&tlb, vma, start, end, ref_page);
2461 tlb_finish_mmu(&tlb, start, end);
Chen, Kenneth W502717f2006-10-11 01:20:46 -07002462}
2463
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002464/*
2465 * This is called when the original mapper is failing to COW a MAP_PRIVATE
2466 * mappping it owns the reserve page for. The intention is to unmap the page
2467 * from other VMAs and let the children be SIGKILLed if they are faulting the
2468 * same region.
2469 */
Harvey Harrison2a4b3de2008-10-18 20:27:06 -07002470static int unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
2471 struct page *page, unsigned long address)
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002472{
Adam Litke75266742008-11-12 13:24:56 -08002473 struct hstate *h = hstate_vma(vma);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002474 struct vm_area_struct *iter_vma;
2475 struct address_space *mapping;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002476 pgoff_t pgoff;
2477
2478 /*
2479 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
2480 * from page cache lookup which is in HPAGE_SIZE units.
2481 */
Adam Litke75266742008-11-12 13:24:56 -08002482 address = address & huge_page_mask(h);
Michal Hocko36e4f202012-10-08 16:33:31 -07002483 pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) +
2484 vma->vm_pgoff;
David Gibson90481622012-03-21 16:34:12 -07002485 mapping = vma->vm_file->f_dentry->d_inode->i_mapping;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002486
Mel Gorman4eb2b1d2009-12-14 17:59:53 -08002487 /*
2488 * Take the mapping lock for the duration of the table walk. As
2489 * this mapping should be shared between all the VMAs,
2490 * __unmap_hugepage_range() is called as the lock is already held
2491 */
Peter Zijlstra3d48ae42011-05-24 17:12:06 -07002492 mutex_lock(&mapping->i_mmap_mutex);
Michel Lespinasse6b2dbba2012-10-08 16:31:25 -07002493 vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) {
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002494 /* Do not unmap the current VMA */
2495 if (iter_vma == vma)
2496 continue;
2497
2498 /*
2499 * Unmap the page from other VMAs without their own reserves.
2500 * They get marked to be SIGKILLed if they fault in these
2501 * areas. This is because a future no-page fault on this VMA
2502 * could insert a zeroed page instead of the data existing
2503 * from the time of fork. This would look like data corruption
2504 */
2505 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07002506 unmap_hugepage_range(iter_vma, address,
2507 address + huge_page_size(h), page);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002508 }
Peter Zijlstra3d48ae42011-05-24 17:12:06 -07002509 mutex_unlock(&mapping->i_mmap_mutex);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002510
2511 return 1;
2512}
2513
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002514/*
2515 * Hugetlb_cow() should be called with page lock of the original hugepage held.
Michal Hockoef009b22012-01-10 15:07:21 -08002516 * Called with hugetlb_instantiation_mutex held and pte_page locked so we
2517 * cannot race with other handlers or page migration.
2518 * Keep the pte_same checks anyway to make transition from the mutex easier.
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002519 */
David Gibson1e8f8892006-01-06 00:10:44 -08002520static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002521 unsigned long address, pte_t *ptep, pte_t pte,
2522 struct page *pagecache_page)
David Gibson1e8f8892006-01-06 00:10:44 -08002523{
Andi Kleena5516432008-07-23 21:27:41 -07002524 struct hstate *h = hstate_vma(vma);
David Gibson1e8f8892006-01-06 00:10:44 -08002525 struct page *old_page, *new_page;
David Gibson79ac6ba2006-03-22 00:08:51 -08002526 int avoidcopy;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002527 int outside_reserve = 0;
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07002528 unsigned long mmun_start; /* For mmu_notifiers */
2529 unsigned long mmun_end; /* For mmu_notifiers */
David Gibson1e8f8892006-01-06 00:10:44 -08002530
2531 old_page = pte_page(pte);
2532
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002533retry_avoidcopy:
David Gibson1e8f8892006-01-06 00:10:44 -08002534 /* If no-one else is actually using this page, avoid the copy
2535 * and just make the page writable */
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002536 avoidcopy = (page_mapcount(old_page) == 1);
David Gibson1e8f8892006-01-06 00:10:44 -08002537 if (avoidcopy) {
Naoya Horiguchi56c9cfb2010-09-10 13:23:04 +09002538 if (PageAnon(old_page))
2539 page_move_anon_rmap(old_page, vma, address);
David Gibson1e8f8892006-01-06 00:10:44 -08002540 set_huge_ptep_writable(vma, address, ptep);
Nick Piggin83c54072007-07-19 01:47:05 -07002541 return 0;
David Gibson1e8f8892006-01-06 00:10:44 -08002542 }
2543
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002544 /*
2545 * If the process that created a MAP_PRIVATE mapping is about to
2546 * perform a COW due to a shared page count, attempt to satisfy
2547 * the allocation without using the existing reserves. The pagecache
2548 * page is used to determine if the reserve at this address was
2549 * consumed or not. If reserves were used, a partial faulted mapping
2550 * at the time of fork() could consume its reserves on COW instead
2551 * of the full address range.
2552 */
Mel Gormanf83a2752009-05-28 14:34:40 -07002553 if (!(vma->vm_flags & VM_MAYSHARE) &&
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002554 is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
2555 old_page != pagecache_page)
2556 outside_reserve = 1;
2557
David Gibson1e8f8892006-01-06 00:10:44 -08002558 page_cache_get(old_page);
Larry Woodmanb76c8cf2009-12-14 17:59:37 -08002559
2560 /* Drop page_table_lock as buddy allocator may be called */
2561 spin_unlock(&mm->page_table_lock);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002562 new_page = alloc_huge_page(vma, address, outside_reserve);
David Gibson1e8f8892006-01-06 00:10:44 -08002563
Adam Litke2fc39ce2007-11-14 16:59:39 -08002564 if (IS_ERR(new_page)) {
Aneesh Kumar K.V76dcee72012-07-31 16:41:57 -07002565 long err = PTR_ERR(new_page);
David Gibson1e8f8892006-01-06 00:10:44 -08002566 page_cache_release(old_page);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002567
2568 /*
2569 * If a process owning a MAP_PRIVATE mapping fails to COW,
2570 * it is due to references held by a child and an insufficient
2571 * huge page pool. To guarantee the original mappers
2572 * reliability, unmap the page from child processes. The child
2573 * may get SIGKILLed if it later faults.
2574 */
2575 if (outside_reserve) {
2576 BUG_ON(huge_pte_none(pte));
2577 if (unmap_ref_private(mm, vma, old_page, address)) {
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002578 BUG_ON(huge_pte_none(pte));
Larry Woodmanb76c8cf2009-12-14 17:59:37 -08002579 spin_lock(&mm->page_table_lock);
Hillf Dantona734bcc2012-01-10 15:07:20 -08002580 ptep = huge_pte_offset(mm, address & huge_page_mask(h));
2581 if (likely(pte_same(huge_ptep_get(ptep), pte)))
2582 goto retry_avoidcopy;
2583 /*
2584 * race occurs while re-acquiring page_table_lock, and
2585 * our job is done.
2586 */
2587 return 0;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002588 }
2589 WARN_ON_ONCE(1);
2590 }
2591
Larry Woodmanb76c8cf2009-12-14 17:59:37 -08002592 /* Caller expects lock to be held */
2593 spin_lock(&mm->page_table_lock);
Aneesh Kumar K.V76dcee72012-07-31 16:41:57 -07002594 if (err == -ENOMEM)
2595 return VM_FAULT_OOM;
2596 else
2597 return VM_FAULT_SIGBUS;
David Gibson1e8f8892006-01-06 00:10:44 -08002598 }
2599
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002600 /*
2601 * When the original hugepage is shared one, it does not have
2602 * anon_vma prepared.
2603 */
Dean Nelson44e2aa92010-10-26 14:22:08 -07002604 if (unlikely(anon_vma_prepare(vma))) {
Hillf Dantonea4039a2011-11-15 14:36:12 -08002605 page_cache_release(new_page);
2606 page_cache_release(old_page);
Dean Nelson44e2aa92010-10-26 14:22:08 -07002607 /* Caller expects lock to be held */
2608 spin_lock(&mm->page_table_lock);
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002609 return VM_FAULT_OOM;
Dean Nelson44e2aa92010-10-26 14:22:08 -07002610 }
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002611
Andrea Arcangeli47ad8472011-01-13 15:46:47 -08002612 copy_user_huge_page(new_page, old_page, address, vma,
2613 pages_per_huge_page(h));
Nick Piggin0ed361d2008-02-04 22:29:34 -08002614 __SetPageUptodate(new_page);
David Gibson1e8f8892006-01-06 00:10:44 -08002615
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07002616 mmun_start = address & huge_page_mask(h);
2617 mmun_end = mmun_start + huge_page_size(h);
2618 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
Larry Woodmanb76c8cf2009-12-14 17:59:37 -08002619 /*
2620 * Retake the page_table_lock to check for racing updates
2621 * before the page tables are altered
2622 */
2623 spin_lock(&mm->page_table_lock);
Andi Kleena5516432008-07-23 21:27:41 -07002624 ptep = huge_pte_offset(mm, address & huge_page_mask(h));
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07002625 if (likely(pte_same(huge_ptep_get(ptep), pte))) {
David Gibson1e8f8892006-01-06 00:10:44 -08002626 /* Break COW */
Gerald Schaefer8fe627e2008-04-28 02:13:28 -07002627 huge_ptep_clear_flush(vma, address, ptep);
David Gibson1e8f8892006-01-06 00:10:44 -08002628 set_huge_pte_at(mm, address, ptep,
2629 make_huge_pte(vma, new_page, 1));
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002630 page_remove_rmap(old_page);
Naoya Horiguchicd67f0d2010-09-10 13:23:04 +09002631 hugepage_add_new_anon_rmap(new_page, vma, address);
David Gibson1e8f8892006-01-06 00:10:44 -08002632 /* Make the old page be freed below */
2633 new_page = old_page;
2634 }
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07002635 spin_unlock(&mm->page_table_lock);
2636 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2637 /* Caller expects lock to be held */
2638 spin_lock(&mm->page_table_lock);
David Gibson1e8f8892006-01-06 00:10:44 -08002639 page_cache_release(new_page);
2640 page_cache_release(old_page);
Nick Piggin83c54072007-07-19 01:47:05 -07002641 return 0;
David Gibson1e8f8892006-01-06 00:10:44 -08002642}
2643
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002644/* Return the pagecache page at a given address within a VMA */
Andi Kleena5516432008-07-23 21:27:41 -07002645static struct page *hugetlbfs_pagecache_page(struct hstate *h,
2646 struct vm_area_struct *vma, unsigned long address)
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002647{
2648 struct address_space *mapping;
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -07002649 pgoff_t idx;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002650
2651 mapping = vma->vm_file->f_mapping;
Andi Kleena5516432008-07-23 21:27:41 -07002652 idx = vma_hugecache_offset(h, vma, address);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002653
2654 return find_lock_page(mapping, idx);
2655}
2656
Hugh Dickins3ae77f42009-09-21 17:03:33 -07002657/*
2658 * Return whether there is a pagecache page to back given address within VMA.
2659 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
2660 */
2661static bool hugetlbfs_pagecache_present(struct hstate *h,
Hugh Dickins2a15efc2009-09-21 17:03:27 -07002662 struct vm_area_struct *vma, unsigned long address)
2663{
2664 struct address_space *mapping;
2665 pgoff_t idx;
2666 struct page *page;
2667
2668 mapping = vma->vm_file->f_mapping;
2669 idx = vma_hugecache_offset(h, vma, address);
2670
2671 page = find_get_page(mapping, idx);
2672 if (page)
2673 put_page(page);
2674 return page != NULL;
2675}
2676
Robert P. J. Daya1ed3dd2007-07-17 04:03:33 -07002677static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
Hugh Dickins788c7df2009-06-23 13:49:05 +01002678 unsigned long address, pte_t *ptep, unsigned int flags)
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01002679{
Andi Kleena5516432008-07-23 21:27:41 -07002680 struct hstate *h = hstate_vma(vma);
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01002681 int ret = VM_FAULT_SIGBUS;
Hillf Danton409eb8c2012-01-20 14:34:13 -08002682 int anon_rmap = 0;
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -07002683 pgoff_t idx;
Adam Litke4c887262005-10-29 18:16:46 -07002684 unsigned long size;
Adam Litke4c887262005-10-29 18:16:46 -07002685 struct page *page;
2686 struct address_space *mapping;
David Gibson1e8f8892006-01-06 00:10:44 -08002687 pte_t new_pte;
Adam Litke4c887262005-10-29 18:16:46 -07002688
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002689 /*
2690 * Currently, we are forced to kill the process in the event the
2691 * original mapper has unmapped pages from the child due to a failed
Lucas De Marchi25985ed2011-03-30 22:57:33 -03002692 * COW. Warn that such a situation has occurred as it may not be obvious
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002693 */
2694 if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
2695 printk(KERN_WARNING
2696 "PID %d killed due to inadequate hugepage pool\n",
2697 current->pid);
2698 return ret;
2699 }
2700
Adam Litke4c887262005-10-29 18:16:46 -07002701 mapping = vma->vm_file->f_mapping;
Andi Kleena5516432008-07-23 21:27:41 -07002702 idx = vma_hugecache_offset(h, vma, address);
Adam Litke4c887262005-10-29 18:16:46 -07002703
2704 /*
2705 * Use page lock to guard against racing truncation
2706 * before we get page_table_lock.
2707 */
Christoph Lameter6bda6662006-01-06 00:10:49 -08002708retry:
2709 page = find_lock_page(mapping, idx);
2710 if (!page) {
Andi Kleena5516432008-07-23 21:27:41 -07002711 size = i_size_read(mapping->host) >> huge_page_shift(h);
Hugh Dickinsebed4bf2006-10-28 10:38:43 -07002712 if (idx >= size)
2713 goto out;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002714 page = alloc_huge_page(vma, address, 0);
Adam Litke2fc39ce2007-11-14 16:59:39 -08002715 if (IS_ERR(page)) {
Aneesh Kumar K.V76dcee72012-07-31 16:41:57 -07002716 ret = PTR_ERR(page);
2717 if (ret == -ENOMEM)
2718 ret = VM_FAULT_OOM;
2719 else
2720 ret = VM_FAULT_SIGBUS;
Christoph Lameter6bda6662006-01-06 00:10:49 -08002721 goto out;
2722 }
Andrea Arcangeli47ad8472011-01-13 15:46:47 -08002723 clear_huge_page(page, address, pages_per_huge_page(h));
Nick Piggin0ed361d2008-02-04 22:29:34 -08002724 __SetPageUptodate(page);
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01002725
Mel Gormanf83a2752009-05-28 14:34:40 -07002726 if (vma->vm_flags & VM_MAYSHARE) {
Christoph Lameter6bda6662006-01-06 00:10:49 -08002727 int err;
Ken Chen45c682a2007-11-14 16:59:44 -08002728 struct inode *inode = mapping->host;
Christoph Lameter6bda6662006-01-06 00:10:49 -08002729
2730 err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
2731 if (err) {
2732 put_page(page);
Christoph Lameter6bda6662006-01-06 00:10:49 -08002733 if (err == -EEXIST)
2734 goto retry;
2735 goto out;
2736 }
Ken Chen45c682a2007-11-14 16:59:44 -08002737
2738 spin_lock(&inode->i_lock);
Andi Kleena5516432008-07-23 21:27:41 -07002739 inode->i_blocks += blocks_per_huge_page(h);
Ken Chen45c682a2007-11-14 16:59:44 -08002740 spin_unlock(&inode->i_lock);
Mel Gorman23be7462010-04-23 13:17:56 -04002741 } else {
Christoph Lameter6bda6662006-01-06 00:10:49 -08002742 lock_page(page);
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002743 if (unlikely(anon_vma_prepare(vma))) {
2744 ret = VM_FAULT_OOM;
2745 goto backout_unlocked;
2746 }
Hillf Danton409eb8c2012-01-20 14:34:13 -08002747 anon_rmap = 1;
Mel Gorman23be7462010-04-23 13:17:56 -04002748 }
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002749 } else {
Naoya Horiguchi998b4382010-09-08 10:19:32 +09002750 /*
2751 * If memory error occurs between mmap() and fault, some process
2752 * don't have hwpoisoned swap entry for errored virtual address.
2753 * So we need to block hugepage fault by PG_hwpoison bit check.
2754 */
2755 if (unlikely(PageHWPoison(page))) {
Chris Forbes32f84522011-07-25 17:12:14 -07002756 ret = VM_FAULT_HWPOISON |
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07002757 VM_FAULT_SET_HINDEX(hstate_index(h));
Naoya Horiguchi998b4382010-09-08 10:19:32 +09002758 goto backout_unlocked;
2759 }
Christoph Lameter6bda6662006-01-06 00:10:49 -08002760 }
David Gibson1e8f8892006-01-06 00:10:44 -08002761
Andy Whitcroft57303d82008-08-12 15:08:47 -07002762 /*
2763 * If we are going to COW a private mapping later, we examine the
2764 * pending reservations for this page now. This will ensure that
2765 * any allocations necessary to record that reservation occur outside
2766 * the spinlock.
2767 */
Hugh Dickins788c7df2009-06-23 13:49:05 +01002768 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED))
Andy Whitcroft2b267362008-08-12 15:08:49 -07002769 if (vma_needs_reservation(h, vma, address) < 0) {
2770 ret = VM_FAULT_OOM;
2771 goto backout_unlocked;
2772 }
Andy Whitcroft57303d82008-08-12 15:08:47 -07002773
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01002774 spin_lock(&mm->page_table_lock);
Andi Kleena5516432008-07-23 21:27:41 -07002775 size = i_size_read(mapping->host) >> huge_page_shift(h);
Adam Litke4c887262005-10-29 18:16:46 -07002776 if (idx >= size)
2777 goto backout;
2778
Nick Piggin83c54072007-07-19 01:47:05 -07002779 ret = 0;
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07002780 if (!huge_pte_none(huge_ptep_get(ptep)))
Adam Litke4c887262005-10-29 18:16:46 -07002781 goto backout;
2782
Hillf Danton409eb8c2012-01-20 14:34:13 -08002783 if (anon_rmap)
2784 hugepage_add_new_anon_rmap(page, vma, address);
2785 else
2786 page_dup_rmap(page);
David Gibson1e8f8892006-01-06 00:10:44 -08002787 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
2788 && (vma->vm_flags & VM_SHARED)));
2789 set_huge_pte_at(mm, address, ptep, new_pte);
2790
Hugh Dickins788c7df2009-06-23 13:49:05 +01002791 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
David Gibson1e8f8892006-01-06 00:10:44 -08002792 /* Optimization, do the COW without a second fault */
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002793 ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page);
David Gibson1e8f8892006-01-06 00:10:44 -08002794 }
2795
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01002796 spin_unlock(&mm->page_table_lock);
Adam Litke4c887262005-10-29 18:16:46 -07002797 unlock_page(page);
2798out:
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01002799 return ret;
Adam Litke4c887262005-10-29 18:16:46 -07002800
2801backout:
2802 spin_unlock(&mm->page_table_lock);
Andy Whitcroft2b267362008-08-12 15:08:49 -07002803backout_unlocked:
Adam Litke4c887262005-10-29 18:16:46 -07002804 unlock_page(page);
2805 put_page(page);
2806 goto out;
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01002807}
2808
Adam Litke86e52162006-01-06 00:10:43 -08002809int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
Hugh Dickins788c7df2009-06-23 13:49:05 +01002810 unsigned long address, unsigned int flags)
Adam Litke86e52162006-01-06 00:10:43 -08002811{
2812 pte_t *ptep;
2813 pte_t entry;
David Gibson1e8f8892006-01-06 00:10:44 -08002814 int ret;
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002815 struct page *page = NULL;
Andy Whitcroft57303d82008-08-12 15:08:47 -07002816 struct page *pagecache_page = NULL;
David Gibson3935baa2006-03-22 00:08:53 -08002817 static DEFINE_MUTEX(hugetlb_instantiation_mutex);
Andi Kleena5516432008-07-23 21:27:41 -07002818 struct hstate *h = hstate_vma(vma);
Adam Litke86e52162006-01-06 00:10:43 -08002819
KAMEZAWA Hiroyuki1e16a532012-01-10 15:07:22 -08002820 address &= huge_page_mask(h);
2821
Naoya Horiguchifd6a03ed2010-05-28 09:29:21 +09002822 ptep = huge_pte_offset(mm, address);
2823 if (ptep) {
2824 entry = huge_ptep_get(ptep);
Naoya Horiguchi290408d2010-09-08 10:19:35 +09002825 if (unlikely(is_hugetlb_entry_migration(entry))) {
2826 migration_entry_wait(mm, (pmd_t *)ptep, address);
2827 return 0;
2828 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
Chris Forbes32f84522011-07-25 17:12:14 -07002829 return VM_FAULT_HWPOISON_LARGE |
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07002830 VM_FAULT_SET_HINDEX(hstate_index(h));
Naoya Horiguchifd6a03ed2010-05-28 09:29:21 +09002831 }
2832
Andi Kleena5516432008-07-23 21:27:41 -07002833 ptep = huge_pte_alloc(mm, address, huge_page_size(h));
Adam Litke86e52162006-01-06 00:10:43 -08002834 if (!ptep)
2835 return VM_FAULT_OOM;
2836
David Gibson3935baa2006-03-22 00:08:53 -08002837 /*
2838 * Serialize hugepage allocation and instantiation, so that we don't
2839 * get spurious allocation failures if two CPUs race to instantiate
2840 * the same page in the page cache.
2841 */
2842 mutex_lock(&hugetlb_instantiation_mutex);
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07002843 entry = huge_ptep_get(ptep);
2844 if (huge_pte_none(entry)) {
Hugh Dickins788c7df2009-06-23 13:49:05 +01002845 ret = hugetlb_no_page(mm, vma, address, ptep, flags);
David Gibsonb4d1d992008-10-15 22:01:11 -07002846 goto out_mutex;
David Gibson3935baa2006-03-22 00:08:53 -08002847 }
Adam Litke86e52162006-01-06 00:10:43 -08002848
Nick Piggin83c54072007-07-19 01:47:05 -07002849 ret = 0;
David Gibson1e8f8892006-01-06 00:10:44 -08002850
Andy Whitcroft57303d82008-08-12 15:08:47 -07002851 /*
2852 * If we are going to COW the mapping later, we examine the pending
2853 * reservations for this page now. This will ensure that any
2854 * allocations necessary to record that reservation occur outside the
2855 * spinlock. For private mappings, we also lookup the pagecache
2856 * page now as it is used to determine if a reservation has been
2857 * consumed.
2858 */
Hugh Dickins788c7df2009-06-23 13:49:05 +01002859 if ((flags & FAULT_FLAG_WRITE) && !pte_write(entry)) {
Andy Whitcroft2b267362008-08-12 15:08:49 -07002860 if (vma_needs_reservation(h, vma, address) < 0) {
2861 ret = VM_FAULT_OOM;
David Gibsonb4d1d992008-10-15 22:01:11 -07002862 goto out_mutex;
Andy Whitcroft2b267362008-08-12 15:08:49 -07002863 }
Andy Whitcroft57303d82008-08-12 15:08:47 -07002864
Mel Gormanf83a2752009-05-28 14:34:40 -07002865 if (!(vma->vm_flags & VM_MAYSHARE))
Andy Whitcroft57303d82008-08-12 15:08:47 -07002866 pagecache_page = hugetlbfs_pagecache_page(h,
2867 vma, address);
2868 }
2869
Naoya Horiguchi56c9cfb2010-09-10 13:23:04 +09002870 /*
2871 * hugetlb_cow() requires page locks of pte_page(entry) and
2872 * pagecache_page, so here we need take the former one
2873 * when page != pagecache_page or !pagecache_page.
2874 * Note that locking order is always pagecache_page -> page,
2875 * so no worry about deadlock.
2876 */
2877 page = pte_page(entry);
Chris Metcalf66aebce2012-04-12 12:49:15 -07002878 get_page(page);
Naoya Horiguchi56c9cfb2010-09-10 13:23:04 +09002879 if (page != pagecache_page)
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002880 lock_page(page);
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002881
David Gibson1e8f8892006-01-06 00:10:44 -08002882 spin_lock(&mm->page_table_lock);
2883 /* Check for a racing update before calling hugetlb_cow */
David Gibsonb4d1d992008-10-15 22:01:11 -07002884 if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
2885 goto out_page_table_lock;
2886
2887
Hugh Dickins788c7df2009-06-23 13:49:05 +01002888 if (flags & FAULT_FLAG_WRITE) {
David Gibsonb4d1d992008-10-15 22:01:11 -07002889 if (!pte_write(entry)) {
Andy Whitcroft57303d82008-08-12 15:08:47 -07002890 ret = hugetlb_cow(mm, vma, address, ptep, entry,
2891 pagecache_page);
David Gibsonb4d1d992008-10-15 22:01:11 -07002892 goto out_page_table_lock;
2893 }
2894 entry = pte_mkdirty(entry);
2895 }
2896 entry = pte_mkyoung(entry);
Hugh Dickins788c7df2009-06-23 13:49:05 +01002897 if (huge_ptep_set_access_flags(vma, address, ptep, entry,
2898 flags & FAULT_FLAG_WRITE))
Russell King4b3073e2009-12-18 16:40:18 +00002899 update_mmu_cache(vma, address, ptep);
David Gibsonb4d1d992008-10-15 22:01:11 -07002900
2901out_page_table_lock:
David Gibson1e8f8892006-01-06 00:10:44 -08002902 spin_unlock(&mm->page_table_lock);
Andy Whitcroft57303d82008-08-12 15:08:47 -07002903
2904 if (pagecache_page) {
2905 unlock_page(pagecache_page);
2906 put_page(pagecache_page);
2907 }
Dean Nelson1f64d692010-12-02 14:31:12 -08002908 if (page != pagecache_page)
2909 unlock_page(page);
Chris Metcalf66aebce2012-04-12 12:49:15 -07002910 put_page(page);
Andy Whitcroft57303d82008-08-12 15:08:47 -07002911
David Gibsonb4d1d992008-10-15 22:01:11 -07002912out_mutex:
David Gibson3935baa2006-03-22 00:08:53 -08002913 mutex_unlock(&hugetlb_instantiation_mutex);
David Gibson1e8f8892006-01-06 00:10:44 -08002914
2915 return ret;
Adam Litke86e52162006-01-06 00:10:43 -08002916}
2917
Andi Kleenceb86872008-07-23 21:27:50 -07002918/* Can be overriden by architectures */
2919__attribute__((weak)) struct page *
2920follow_huge_pud(struct mm_struct *mm, unsigned long address,
2921 pud_t *pud, int write)
2922{
2923 BUG();
2924 return NULL;
2925}
2926
David Gibson63551ae2005-06-21 17:14:44 -07002927int follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
2928 struct page **pages, struct vm_area_struct **vmas,
Adam Litke5b23dbe2007-11-14 16:59:33 -08002929 unsigned long *position, int *length, int i,
Hugh Dickins2a15efc2009-09-21 17:03:27 -07002930 unsigned int flags)
David Gibson63551ae2005-06-21 17:14:44 -07002931{
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08002932 unsigned long pfn_offset;
2933 unsigned long vaddr = *position;
David Gibson63551ae2005-06-21 17:14:44 -07002934 int remainder = *length;
Andi Kleena5516432008-07-23 21:27:41 -07002935 struct hstate *h = hstate_vma(vma);
David Gibson63551ae2005-06-21 17:14:44 -07002936
Hugh Dickins1c598272005-10-19 21:23:43 -07002937 spin_lock(&mm->page_table_lock);
David Gibson63551ae2005-06-21 17:14:44 -07002938 while (vaddr < vma->vm_end && remainder) {
Adam Litke4c887262005-10-29 18:16:46 -07002939 pte_t *pte;
Hugh Dickins2a15efc2009-09-21 17:03:27 -07002940 int absent;
Adam Litke4c887262005-10-29 18:16:46 -07002941 struct page *page;
2942
2943 /*
2944 * Some archs (sparc64, sh*) have multiple pte_ts to
Hugh Dickins2a15efc2009-09-21 17:03:27 -07002945 * each hugepage. We have to make sure we get the
Adam Litke4c887262005-10-29 18:16:46 -07002946 * first, for the page indexing below to work.
2947 */
Andi Kleena5516432008-07-23 21:27:41 -07002948 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h));
Hugh Dickins2a15efc2009-09-21 17:03:27 -07002949 absent = !pte || huge_pte_none(huge_ptep_get(pte));
Adam Litke4c887262005-10-29 18:16:46 -07002950
Hugh Dickins2a15efc2009-09-21 17:03:27 -07002951 /*
2952 * When coredumping, it suits get_dump_page if we just return
Hugh Dickins3ae77f42009-09-21 17:03:33 -07002953 * an error where there's an empty slot with no huge pagecache
2954 * to back it. This way, we avoid allocating a hugepage, and
2955 * the sparse dumpfile avoids allocating disk blocks, but its
2956 * huge holes still show up with zeroes where they need to be.
Hugh Dickins2a15efc2009-09-21 17:03:27 -07002957 */
Hugh Dickins3ae77f42009-09-21 17:03:33 -07002958 if (absent && (flags & FOLL_DUMP) &&
2959 !hugetlbfs_pagecache_present(h, vma, vaddr)) {
Hugh Dickins2a15efc2009-09-21 17:03:27 -07002960 remainder = 0;
2961 break;
2962 }
2963
2964 if (absent ||
2965 ((flags & FOLL_WRITE) && !pte_write(huge_ptep_get(pte)))) {
Adam Litke4c887262005-10-29 18:16:46 -07002966 int ret;
2967
2968 spin_unlock(&mm->page_table_lock);
Hugh Dickins2a15efc2009-09-21 17:03:27 -07002969 ret = hugetlb_fault(mm, vma, vaddr,
2970 (flags & FOLL_WRITE) ? FAULT_FLAG_WRITE : 0);
Adam Litke4c887262005-10-29 18:16:46 -07002971 spin_lock(&mm->page_table_lock);
Adam Litkea89182c2007-08-22 14:01:51 -07002972 if (!(ret & VM_FAULT_ERROR))
Adam Litke4c887262005-10-29 18:16:46 -07002973 continue;
2974
2975 remainder = 0;
Adam Litke4c887262005-10-29 18:16:46 -07002976 break;
2977 }
David Gibson63551ae2005-06-21 17:14:44 -07002978
Andi Kleena5516432008-07-23 21:27:41 -07002979 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07002980 page = pte_page(huge_ptep_get(pte));
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08002981same_page:
Chen, Kenneth Wd6692182006-03-31 02:29:57 -08002982 if (pages) {
Hugh Dickins2a15efc2009-09-21 17:03:27 -07002983 pages[i] = mem_map_offset(page, pfn_offset);
KOSAKI Motohiro4b2e38a2008-10-18 20:27:10 -07002984 get_page(pages[i]);
Chen, Kenneth Wd6692182006-03-31 02:29:57 -08002985 }
David Gibson63551ae2005-06-21 17:14:44 -07002986
2987 if (vmas)
2988 vmas[i] = vma;
2989
2990 vaddr += PAGE_SIZE;
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08002991 ++pfn_offset;
David Gibson63551ae2005-06-21 17:14:44 -07002992 --remainder;
2993 ++i;
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08002994 if (vaddr < vma->vm_end && remainder &&
Andi Kleena5516432008-07-23 21:27:41 -07002995 pfn_offset < pages_per_huge_page(h)) {
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08002996 /*
2997 * We use pfn_offset to avoid touching the pageframes
2998 * of this compound page.
2999 */
3000 goto same_page;
3001 }
David Gibson63551ae2005-06-21 17:14:44 -07003002 }
Hugh Dickins1c598272005-10-19 21:23:43 -07003003 spin_unlock(&mm->page_table_lock);
David Gibson63551ae2005-06-21 17:14:44 -07003004 *length = remainder;
3005 *position = vaddr;
3006
Hugh Dickins2a15efc2009-09-21 17:03:27 -07003007 return i ? i : -EFAULT;
David Gibson63551ae2005-06-21 17:14:44 -07003008}
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003009
Peter Zijlstra7da4d642012-11-19 03:14:23 +01003010unsigned long hugetlb_change_protection(struct vm_area_struct *vma,
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003011 unsigned long address, unsigned long end, pgprot_t newprot)
3012{
3013 struct mm_struct *mm = vma->vm_mm;
3014 unsigned long start = address;
3015 pte_t *ptep;
3016 pte_t pte;
Andi Kleena5516432008-07-23 21:27:41 -07003017 struct hstate *h = hstate_vma(vma);
Peter Zijlstra7da4d642012-11-19 03:14:23 +01003018 unsigned long pages = 0;
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003019
3020 BUG_ON(address >= end);
3021 flush_cache_range(vma, address, end);
3022
Peter Zijlstra3d48ae42011-05-24 17:12:06 -07003023 mutex_lock(&vma->vm_file->f_mapping->i_mmap_mutex);
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003024 spin_lock(&mm->page_table_lock);
Andi Kleena5516432008-07-23 21:27:41 -07003025 for (; address < end; address += huge_page_size(h)) {
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003026 ptep = huge_pte_offset(mm, address);
3027 if (!ptep)
3028 continue;
Peter Zijlstra7da4d642012-11-19 03:14:23 +01003029 if (huge_pmd_unshare(mm, &address, ptep)) {
3030 pages++;
Chen, Kenneth W39dde652006-12-06 20:32:03 -08003031 continue;
Peter Zijlstra7da4d642012-11-19 03:14:23 +01003032 }
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07003033 if (!huge_pte_none(huge_ptep_get(ptep))) {
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003034 pte = huge_ptep_get_and_clear(mm, address, ptep);
3035 pte = pte_mkhuge(pte_modify(pte, newprot));
3036 set_huge_pte_at(mm, address, ptep, pte);
Peter Zijlstra7da4d642012-11-19 03:14:23 +01003037 pages++;
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003038 }
3039 }
3040 spin_unlock(&mm->page_table_lock);
Mel Gormand8333522012-07-31 16:46:20 -07003041 /*
3042 * Must flush TLB before releasing i_mmap_mutex: x86's huge_pmd_unshare
3043 * may have cleared our pud entry and done put_page on the page table:
3044 * once we release i_mmap_mutex, another task can do the final put_page
3045 * and that page table be reused and filled with junk.
3046 */
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003047 flush_tlb_range(vma, start, end);
Mel Gormand8333522012-07-31 16:46:20 -07003048 mutex_unlock(&vma->vm_file->f_mapping->i_mmap_mutex);
Peter Zijlstra7da4d642012-11-19 03:14:23 +01003049
3050 return pages << h->order;
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003051}
3052
Mel Gormana1e78772008-07-23 21:27:23 -07003053int hugetlb_reserve_pages(struct inode *inode,
3054 long from, long to,
Mel Gorman5a6fe122009-02-10 14:02:27 +00003055 struct vm_area_struct *vma,
KOSAKI Motohiroca16d142011-05-26 19:16:19 +09003056 vm_flags_t vm_flags)
Adam Litkee4e574b2007-10-16 01:26:19 -07003057{
Mel Gorman17c9d122009-02-11 16:34:16 +00003058 long ret, chg;
Andi Kleena5516432008-07-23 21:27:41 -07003059 struct hstate *h = hstate_inode(inode);
David Gibson90481622012-03-21 16:34:12 -07003060 struct hugepage_subpool *spool = subpool_inode(inode);
Adam Litkee4e574b2007-10-16 01:26:19 -07003061
Mel Gormana1e78772008-07-23 21:27:23 -07003062 /*
Mel Gorman17c9d122009-02-11 16:34:16 +00003063 * Only apply hugepage reservation if asked. At fault time, an
3064 * attempt will be made for VM_NORESERVE to allocate a page
David Gibson90481622012-03-21 16:34:12 -07003065 * without using reserves
Mel Gorman17c9d122009-02-11 16:34:16 +00003066 */
KOSAKI Motohiroca16d142011-05-26 19:16:19 +09003067 if (vm_flags & VM_NORESERVE)
Mel Gorman17c9d122009-02-11 16:34:16 +00003068 return 0;
3069
3070 /*
Mel Gormana1e78772008-07-23 21:27:23 -07003071 * Shared mappings base their reservation on the number of pages that
3072 * are already allocated on behalf of the file. Private mappings need
3073 * to reserve the full area even if read-only as mprotect() may be
3074 * called to make the mapping read-write. Assume !vma is a shm mapping
3075 */
Mel Gormanf83a2752009-05-28 14:34:40 -07003076 if (!vma || vma->vm_flags & VM_MAYSHARE)
Mel Gormana1e78772008-07-23 21:27:23 -07003077 chg = region_chg(&inode->i_mapping->private_list, from, to);
Mel Gorman5a6fe122009-02-10 14:02:27 +00003078 else {
3079 struct resv_map *resv_map = resv_map_alloc();
Mel Gorman5a6fe122009-02-10 14:02:27 +00003080 if (!resv_map)
3081 return -ENOMEM;
3082
Mel Gorman17c9d122009-02-11 16:34:16 +00003083 chg = to - from;
3084
Mel Gorman5a6fe122009-02-10 14:02:27 +00003085 set_vma_resv_map(vma, resv_map);
3086 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
3087 }
3088
Dave Hansenc50ac052012-05-29 15:06:46 -07003089 if (chg < 0) {
3090 ret = chg;
3091 goto out_err;
3092 }
Mel Gorman17c9d122009-02-11 16:34:16 +00003093
David Gibson90481622012-03-21 16:34:12 -07003094 /* There must be enough pages in the subpool for the mapping */
Dave Hansenc50ac052012-05-29 15:06:46 -07003095 if (hugepage_subpool_get_pages(spool, chg)) {
3096 ret = -ENOSPC;
3097 goto out_err;
3098 }
Mel Gorman17c9d122009-02-11 16:34:16 +00003099
3100 /*
3101 * Check enough hugepages are available for the reservation.
David Gibson90481622012-03-21 16:34:12 -07003102 * Hand the pages back to the subpool if there are not
Mel Gorman17c9d122009-02-11 16:34:16 +00003103 */
3104 ret = hugetlb_acct_memory(h, chg);
3105 if (ret < 0) {
David Gibson90481622012-03-21 16:34:12 -07003106 hugepage_subpool_put_pages(spool, chg);
Dave Hansenc50ac052012-05-29 15:06:46 -07003107 goto out_err;
Mel Gorman17c9d122009-02-11 16:34:16 +00003108 }
3109
3110 /*
3111 * Account for the reservations made. Shared mappings record regions
3112 * that have reservations as they are shared by multiple VMAs.
3113 * When the last VMA disappears, the region map says how much
3114 * the reservation was and the page cache tells how much of
3115 * the reservation was consumed. Private mappings are per-VMA and
3116 * only the consumed reservations are tracked. When the VMA
3117 * disappears, the original reservation is the VMA size and the
3118 * consumed reservations are stored in the map. Hence, nothing
3119 * else has to be done for private mappings here
3120 */
Mel Gormanf83a2752009-05-28 14:34:40 -07003121 if (!vma || vma->vm_flags & VM_MAYSHARE)
Mel Gorman17c9d122009-02-11 16:34:16 +00003122 region_add(&inode->i_mapping->private_list, from, to);
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -07003123 return 0;
Dave Hansenc50ac052012-05-29 15:06:46 -07003124out_err:
Dave Hansen4523e142012-05-30 07:51:07 -07003125 if (vma)
3126 resv_map_put(vma);
Dave Hansenc50ac052012-05-29 15:06:46 -07003127 return ret;
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -07003128}
3129
3130void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed)
3131{
Andi Kleena5516432008-07-23 21:27:41 -07003132 struct hstate *h = hstate_inode(inode);
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -07003133 long chg = region_truncate(&inode->i_mapping->private_list, offset);
David Gibson90481622012-03-21 16:34:12 -07003134 struct hugepage_subpool *spool = subpool_inode(inode);
Ken Chen45c682a2007-11-14 16:59:44 -08003135
3136 spin_lock(&inode->i_lock);
Eric Sandeene4c6f8b2009-07-29 15:02:16 -07003137 inode->i_blocks -= (blocks_per_huge_page(h) * freed);
Ken Chen45c682a2007-11-14 16:59:44 -08003138 spin_unlock(&inode->i_lock);
3139
David Gibson90481622012-03-21 16:34:12 -07003140 hugepage_subpool_put_pages(spool, (chg - freed));
Andi Kleena5516432008-07-23 21:27:41 -07003141 hugetlb_acct_memory(h, -(chg - freed));
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -07003142}
Naoya Horiguchi93f70f92010-05-28 09:29:20 +09003143
Andi Kleend5bd9102010-09-27 09:00:12 +02003144#ifdef CONFIG_MEMORY_FAILURE
3145
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +09003146/* Should be called in hugetlb_lock */
3147static int is_hugepage_on_freelist(struct page *hpage)
3148{
3149 struct page *page;
3150 struct page *tmp;
3151 struct hstate *h = page_hstate(hpage);
3152 int nid = page_to_nid(hpage);
3153
3154 list_for_each_entry_safe(page, tmp, &h->hugepage_freelists[nid], lru)
3155 if (page == hpage)
3156 return 1;
3157 return 0;
3158}
3159
Naoya Horiguchi93f70f92010-05-28 09:29:20 +09003160/*
3161 * This function is called from memory failure code.
3162 * Assume the caller holds page lock of the head page.
3163 */
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +09003164int dequeue_hwpoisoned_huge_page(struct page *hpage)
Naoya Horiguchi93f70f92010-05-28 09:29:20 +09003165{
3166 struct hstate *h = page_hstate(hpage);
3167 int nid = page_to_nid(hpage);
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +09003168 int ret = -EBUSY;
Naoya Horiguchi93f70f92010-05-28 09:29:20 +09003169
3170 spin_lock(&hugetlb_lock);
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +09003171 if (is_hugepage_on_freelist(hpage)) {
Naoya Horiguchi56f2fb12012-12-12 13:52:33 -08003172 /*
3173 * Hwpoisoned hugepage isn't linked to activelist or freelist,
3174 * but dangling hpage->lru can trigger list-debug warnings
3175 * (this happens when we call unpoison_memory() on it),
3176 * so let it point to itself with list_del_init().
3177 */
3178 list_del_init(&hpage->lru);
Naoya Horiguchi8c6c2ec2010-09-08 10:19:38 +09003179 set_page_refcounted(hpage);
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +09003180 h->free_huge_pages--;
3181 h->free_huge_pages_node[nid]--;
3182 ret = 0;
3183 }
Naoya Horiguchi93f70f92010-05-28 09:29:20 +09003184 spin_unlock(&hugetlb_lock);
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +09003185 return ret;
Naoya Horiguchi93f70f92010-05-28 09:29:20 +09003186}
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +09003187#endif