blob: bc727122dd44de6c4ae9307c618e58ddf4da3c87 [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
2 * Generic hugetlb support.
3 * (C) William Irwin, April 2004
4 */
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>
Aneesh Kumar K.Vabb82062012-07-31 16:42:24 -070033#include <linux/hugetlb_cgroup.h>
Nick Piggin7835e982006-03-22 00:08:40 -080034#include "internal.h"
Linus Torvalds1da177e2005-04-16 15:20:36 -070035
36const unsigned long hugetlb_zero = 0, hugetlb_infinity = ~0UL;
Mel Gorman396faf02007-07-17 04:03:13 -070037static gfp_t htlb_alloc_mask = GFP_HIGHUSER;
38unsigned long hugepages_treat_as_movable;
Andi Kleena5516432008-07-23 21:27:41 -070039
Aneesh Kumar K.Vc3f38a32012-07-31 16:42:10 -070040int hugetlb_max_hstate __read_mostly;
Andi Kleene5ff2152008-07-23 21:27:42 -070041unsigned int default_hstate_idx;
42struct hstate hstates[HUGE_MAX_HSTATE];
43
Jon Tollefson53ba51d2008-07-23 21:27:52 -070044__initdata LIST_HEAD(huge_boot_pages);
45
Andi Kleene5ff2152008-07-23 21:27:42 -070046/* for command line parsing */
47static struct hstate * __initdata parsed_hstate;
48static unsigned long __initdata default_hstate_max_huge_pages;
Nick Piggine11bfbf2008-07-23 21:27:52 -070049static unsigned long __initdata default_hstate_size;
Andi Kleene5ff2152008-07-23 21:27:42 -070050
David Gibson3935baa2006-03-22 00:08:53 -080051/*
52 * Protects updates to hugepage_freelists, nr_huge_pages, and free_huge_pages
53 */
Aneesh Kumar K.Vc3f38a32012-07-31 16:42:10 -070054DEFINE_SPINLOCK(hugetlb_lock);
Eric Paris0bd0f9f2005-11-21 21:32:28 -080055
David Gibson90481622012-03-21 16:34:12 -070056static inline void unlock_or_release_subpool(struct hugepage_subpool *spool)
57{
58 bool free = (spool->count == 0) && (spool->used_hpages == 0);
59
60 spin_unlock(&spool->lock);
61
62 /* If no pages are used, and no other handles to the subpool
63 * remain, free the subpool the subpool remain */
64 if (free)
65 kfree(spool);
66}
67
68struct hugepage_subpool *hugepage_new_subpool(long nr_blocks)
69{
70 struct hugepage_subpool *spool;
71
72 spool = kmalloc(sizeof(*spool), GFP_KERNEL);
73 if (!spool)
74 return NULL;
75
76 spin_lock_init(&spool->lock);
77 spool->count = 1;
78 spool->max_hpages = nr_blocks;
79 spool->used_hpages = 0;
80
81 return spool;
82}
83
84void hugepage_put_subpool(struct hugepage_subpool *spool)
85{
86 spin_lock(&spool->lock);
87 BUG_ON(!spool->count);
88 spool->count--;
89 unlock_or_release_subpool(spool);
90}
91
92static int hugepage_subpool_get_pages(struct hugepage_subpool *spool,
93 long delta)
94{
95 int ret = 0;
96
97 if (!spool)
98 return 0;
99
100 spin_lock(&spool->lock);
101 if ((spool->used_hpages + delta) <= spool->max_hpages) {
102 spool->used_hpages += delta;
103 } else {
104 ret = -ENOMEM;
105 }
106 spin_unlock(&spool->lock);
107
108 return ret;
109}
110
111static void hugepage_subpool_put_pages(struct hugepage_subpool *spool,
112 long delta)
113{
114 if (!spool)
115 return;
116
117 spin_lock(&spool->lock);
118 spool->used_hpages -= delta;
119 /* If hugetlbfs_put_super couldn't free spool due to
120 * an outstanding quota reference, free it now. */
121 unlock_or_release_subpool(spool);
122}
123
124static inline struct hugepage_subpool *subpool_inode(struct inode *inode)
125{
126 return HUGETLBFS_SB(inode->i_sb)->spool;
127}
128
129static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma)
130{
131 return subpool_inode(vma->vm_file->f_dentry->d_inode);
132}
133
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700134/*
Andy Whitcroft96822902008-07-23 21:27:29 -0700135 * Region tracking -- allows tracking of reservations and instantiated pages
136 * across the pages in a mapping.
Andy Whitcroft84afd992008-07-23 21:27:32 -0700137 *
138 * The region data structures are protected by a combination of the mmap_sem
139 * and the hugetlb_instantion_mutex. To access or modify a region the caller
140 * must either hold the mmap_sem for write, or the mmap_sem for read and
141 * the hugetlb_instantiation mutex:
142 *
Chris Forbes32f84522011-07-25 17:12:14 -0700143 * down_write(&mm->mmap_sem);
Andy Whitcroft84afd992008-07-23 21:27:32 -0700144 * or
Chris Forbes32f84522011-07-25 17:12:14 -0700145 * down_read(&mm->mmap_sem);
146 * mutex_lock(&hugetlb_instantiation_mutex);
Andy Whitcroft96822902008-07-23 21:27:29 -0700147 */
148struct file_region {
149 struct list_head link;
150 long from;
151 long to;
152};
153
154static long region_add(struct list_head *head, long f, long t)
155{
156 struct file_region *rg, *nrg, *trg;
157
158 /* Locate the region we are either in or before. */
159 list_for_each_entry(rg, head, link)
160 if (f <= rg->to)
161 break;
162
163 /* Round our left edge to the current segment if it encloses us. */
164 if (f > rg->from)
165 f = rg->from;
166
167 /* Check for and consume any regions we now overlap with. */
168 nrg = rg;
169 list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
170 if (&rg->link == head)
171 break;
172 if (rg->from > t)
173 break;
174
175 /* If this area reaches higher then extend our area to
176 * include it completely. If this is not the first area
177 * which we intend to reuse, free it. */
178 if (rg->to > t)
179 t = rg->to;
180 if (rg != nrg) {
181 list_del(&rg->link);
182 kfree(rg);
183 }
184 }
185 nrg->from = f;
186 nrg->to = t;
187 return 0;
188}
189
190static long region_chg(struct list_head *head, long f, long t)
191{
192 struct file_region *rg, *nrg;
193 long chg = 0;
194
195 /* Locate the region we are before or in. */
196 list_for_each_entry(rg, head, link)
197 if (f <= rg->to)
198 break;
199
200 /* If we are below the current region then a new region is required.
201 * Subtle, allocate a new region at the position but make it zero
202 * size such that we can guarantee to record the reservation. */
203 if (&rg->link == head || t < rg->from) {
204 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
205 if (!nrg)
206 return -ENOMEM;
207 nrg->from = f;
208 nrg->to = f;
209 INIT_LIST_HEAD(&nrg->link);
210 list_add(&nrg->link, rg->link.prev);
211
212 return t - f;
213 }
214
215 /* Round our left edge to the current segment if it encloses us. */
216 if (f > rg->from)
217 f = rg->from;
218 chg = t - f;
219
220 /* Check for and consume any regions we now overlap with. */
221 list_for_each_entry(rg, rg->link.prev, link) {
222 if (&rg->link == head)
223 break;
224 if (rg->from > t)
225 return chg;
226
Lucas De Marchi25985ed2011-03-30 22:57:33 -0300227 /* We overlap with this area, if it extends further than
Andy Whitcroft96822902008-07-23 21:27:29 -0700228 * us then we must extend ourselves. Account for its
229 * existing reservation. */
230 if (rg->to > t) {
231 chg += rg->to - t;
232 t = rg->to;
233 }
234 chg -= rg->to - rg->from;
235 }
236 return chg;
237}
238
239static long region_truncate(struct list_head *head, long end)
240{
241 struct file_region *rg, *trg;
242 long chg = 0;
243
244 /* Locate the region we are either in or before. */
245 list_for_each_entry(rg, head, link)
246 if (end <= rg->to)
247 break;
248 if (&rg->link == head)
249 return 0;
250
251 /* If we are in the middle of a region then adjust it. */
252 if (end > rg->from) {
253 chg = rg->to - end;
254 rg->to = end;
255 rg = list_entry(rg->link.next, typeof(*rg), link);
256 }
257
258 /* Drop any remaining regions. */
259 list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
260 if (&rg->link == head)
261 break;
262 chg += rg->to - rg->from;
263 list_del(&rg->link);
264 kfree(rg);
265 }
266 return chg;
267}
268
Andy Whitcroft84afd992008-07-23 21:27:32 -0700269static long region_count(struct list_head *head, long f, long t)
270{
271 struct file_region *rg;
272 long chg = 0;
273
274 /* Locate each segment we overlap with, and count that overlap. */
275 list_for_each_entry(rg, head, link) {
Wang Sheng-Huif2135a42012-05-29 15:06:17 -0700276 long seg_from;
277 long seg_to;
Andy Whitcroft84afd992008-07-23 21:27:32 -0700278
279 if (rg->to <= f)
280 continue;
281 if (rg->from >= t)
282 break;
283
284 seg_from = max(rg->from, f);
285 seg_to = min(rg->to, t);
286
287 chg += seg_to - seg_from;
288 }
289
290 return chg;
291}
292
Andy Whitcroft96822902008-07-23 21:27:29 -0700293/*
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700294 * Convert the address within this vma to the page offset within
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700295 * the mapping, in pagecache page units; huge pages here.
296 */
Andi Kleena5516432008-07-23 21:27:41 -0700297static pgoff_t vma_hugecache_offset(struct hstate *h,
298 struct vm_area_struct *vma, unsigned long address)
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700299{
Andi Kleena5516432008-07-23 21:27:41 -0700300 return ((address - vma->vm_start) >> huge_page_shift(h)) +
301 (vma->vm_pgoff >> huge_page_order(h));
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700302}
303
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +0900304pgoff_t linear_hugepage_index(struct vm_area_struct *vma,
305 unsigned long address)
306{
307 return vma_hugecache_offset(hstate_vma(vma), vma, address);
308}
309
Andy Whitcroft84afd992008-07-23 21:27:32 -0700310/*
Mel Gorman08fba692009-01-06 14:38:53 -0800311 * Return the size of the pages allocated when backing a VMA. In the majority
312 * cases this will be same size as used by the page table entries.
313 */
314unsigned long vma_kernel_pagesize(struct vm_area_struct *vma)
315{
316 struct hstate *hstate;
317
318 if (!is_vm_hugetlb_page(vma))
319 return PAGE_SIZE;
320
321 hstate = hstate_vma(vma);
322
323 return 1UL << (hstate->order + PAGE_SHIFT);
324}
Joerg Roedelf340ca02009-06-19 15:16:22 +0200325EXPORT_SYMBOL_GPL(vma_kernel_pagesize);
Mel Gorman08fba692009-01-06 14:38:53 -0800326
327/*
Mel Gorman33402892009-01-06 14:38:54 -0800328 * Return the page size being used by the MMU to back a VMA. In the majority
329 * of cases, the page size used by the kernel matches the MMU size. On
330 * architectures where it differs, an architecture-specific version of this
331 * function is required.
332 */
333#ifndef vma_mmu_pagesize
334unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
335{
336 return vma_kernel_pagesize(vma);
337}
338#endif
339
340/*
Andy Whitcroft84afd992008-07-23 21:27:32 -0700341 * Flags for MAP_PRIVATE reservations. These are stored in the bottom
342 * bits of the reservation map pointer, which are always clear due to
343 * alignment.
344 */
345#define HPAGE_RESV_OWNER (1UL << 0)
346#define HPAGE_RESV_UNMAPPED (1UL << 1)
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700347#define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
Andy Whitcroft84afd992008-07-23 21:27:32 -0700348
Mel Gormana1e78772008-07-23 21:27:23 -0700349/*
350 * These helpers are used to track how many pages are reserved for
351 * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
352 * is guaranteed to have their future faults succeed.
353 *
354 * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
355 * the reserve counters are updated with the hugetlb_lock held. It is safe
356 * to reset the VMA at fork() time as it is not in use yet and there is no
357 * chance of the global counters getting corrupted as a result of the values.
Andy Whitcroft84afd992008-07-23 21:27:32 -0700358 *
359 * The private mapping reservation is represented in a subtly different
360 * manner to a shared mapping. A shared mapping has a region map associated
361 * with the underlying file, this region map represents the backing file
362 * pages which have ever had a reservation assigned which this persists even
363 * after the page is instantiated. A private mapping has a region map
364 * associated with the original mmap which is attached to all VMAs which
365 * reference it, this region map represents those offsets which have consumed
366 * reservation ie. where pages have been instantiated.
Mel Gormana1e78772008-07-23 21:27:23 -0700367 */
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700368static unsigned long get_vma_private_data(struct vm_area_struct *vma)
369{
370 return (unsigned long)vma->vm_private_data;
371}
372
373static void set_vma_private_data(struct vm_area_struct *vma,
374 unsigned long value)
375{
376 vma->vm_private_data = (void *)value;
377}
378
Andy Whitcroft84afd992008-07-23 21:27:32 -0700379struct resv_map {
380 struct kref refs;
381 struct list_head regions;
382};
383
Harvey Harrison2a4b3de2008-10-18 20:27:06 -0700384static struct resv_map *resv_map_alloc(void)
Andy Whitcroft84afd992008-07-23 21:27:32 -0700385{
386 struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
387 if (!resv_map)
388 return NULL;
389
390 kref_init(&resv_map->refs);
391 INIT_LIST_HEAD(&resv_map->regions);
392
393 return resv_map;
394}
395
Harvey Harrison2a4b3de2008-10-18 20:27:06 -0700396static void resv_map_release(struct kref *ref)
Andy Whitcroft84afd992008-07-23 21:27:32 -0700397{
398 struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
399
400 /* Clear out any active regions before we release the map. */
401 region_truncate(&resv_map->regions, 0);
402 kfree(resv_map);
403}
404
405static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
Mel Gormana1e78772008-07-23 21:27:23 -0700406{
407 VM_BUG_ON(!is_vm_hugetlb_page(vma));
Mel Gormanf83a2752009-05-28 14:34:40 -0700408 if (!(vma->vm_flags & VM_MAYSHARE))
Andy Whitcroft84afd992008-07-23 21:27:32 -0700409 return (struct resv_map *)(get_vma_private_data(vma) &
410 ~HPAGE_RESV_MASK);
Harvey Harrison2a4b3de2008-10-18 20:27:06 -0700411 return NULL;
Mel Gormana1e78772008-07-23 21:27:23 -0700412}
413
Andy Whitcroft84afd992008-07-23 21:27:32 -0700414static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
Mel Gormana1e78772008-07-23 21:27:23 -0700415{
416 VM_BUG_ON(!is_vm_hugetlb_page(vma));
Mel Gormanf83a2752009-05-28 14:34:40 -0700417 VM_BUG_ON(vma->vm_flags & VM_MAYSHARE);
Mel Gormana1e78772008-07-23 21:27:23 -0700418
Andy Whitcroft84afd992008-07-23 21:27:32 -0700419 set_vma_private_data(vma, (get_vma_private_data(vma) &
420 HPAGE_RESV_MASK) | (unsigned long)map);
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700421}
422
423static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
424{
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700425 VM_BUG_ON(!is_vm_hugetlb_page(vma));
Mel Gormanf83a2752009-05-28 14:34:40 -0700426 VM_BUG_ON(vma->vm_flags & VM_MAYSHARE);
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700427
428 set_vma_private_data(vma, get_vma_private_data(vma) | flags);
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700429}
430
431static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
432{
433 VM_BUG_ON(!is_vm_hugetlb_page(vma));
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700434
435 return (get_vma_private_data(vma) & flag) != 0;
Mel Gormana1e78772008-07-23 21:27:23 -0700436}
437
438/* Decrement the reserved pages in the hugepage pool by one */
Andi Kleena5516432008-07-23 21:27:41 -0700439static void decrement_hugepage_resv_vma(struct hstate *h,
440 struct vm_area_struct *vma)
Mel Gormana1e78772008-07-23 21:27:23 -0700441{
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -0700442 if (vma->vm_flags & VM_NORESERVE)
443 return;
444
Mel Gormanf83a2752009-05-28 14:34:40 -0700445 if (vma->vm_flags & VM_MAYSHARE) {
Mel Gormana1e78772008-07-23 21:27:23 -0700446 /* Shared mappings always use reserves */
Andi Kleena5516432008-07-23 21:27:41 -0700447 h->resv_huge_pages--;
Andy Whitcroft84afd992008-07-23 21:27:32 -0700448 } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
Mel Gormana1e78772008-07-23 21:27:23 -0700449 /*
450 * Only the process that called mmap() has reserves for
451 * private mappings.
452 */
Andi Kleena5516432008-07-23 21:27:41 -0700453 h->resv_huge_pages--;
Mel Gormana1e78772008-07-23 21:27:23 -0700454 }
455}
456
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700457/* Reset counters to 0 and clear all HPAGE_RESV_* flags */
Mel Gormana1e78772008-07-23 21:27:23 -0700458void reset_vma_resv_huge_pages(struct vm_area_struct *vma)
459{
460 VM_BUG_ON(!is_vm_hugetlb_page(vma));
Mel Gormanf83a2752009-05-28 14:34:40 -0700461 if (!(vma->vm_flags & VM_MAYSHARE))
Mel Gormana1e78772008-07-23 21:27:23 -0700462 vma->vm_private_data = (void *)0;
463}
464
465/* Returns true if the VMA has associated reserve pages */
Mel Gorman7f09ca52008-07-23 21:27:58 -0700466static int vma_has_reserves(struct vm_area_struct *vma)
Mel Gormana1e78772008-07-23 21:27:23 -0700467{
Mel Gormanf83a2752009-05-28 14:34:40 -0700468 if (vma->vm_flags & VM_MAYSHARE)
Mel Gorman7f09ca52008-07-23 21:27:58 -0700469 return 1;
470 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER))
471 return 1;
472 return 0;
Mel Gormana1e78772008-07-23 21:27:23 -0700473}
474
Naoya Horiguchi0ebabb42010-09-08 10:19:34 +0900475static void copy_gigantic_page(struct page *dst, struct page *src)
476{
477 int i;
478 struct hstate *h = page_hstate(src);
479 struct page *dst_base = dst;
480 struct page *src_base = src;
481
482 for (i = 0; i < pages_per_huge_page(h); ) {
483 cond_resched();
484 copy_highpage(dst, src);
485
486 i++;
487 dst = mem_map_next(dst, dst_base, i);
488 src = mem_map_next(src, src_base, i);
489 }
490}
491
492void copy_huge_page(struct page *dst, struct page *src)
493{
494 int i;
495 struct hstate *h = page_hstate(src);
496
497 if (unlikely(pages_per_huge_page(h) > MAX_ORDER_NR_PAGES)) {
498 copy_gigantic_page(dst, src);
499 return;
500 }
501
502 might_sleep();
503 for (i = 0; i < pages_per_huge_page(h); i++) {
504 cond_resched();
505 copy_highpage(dst + i, src + i);
506 }
507}
508
Andi Kleena5516432008-07-23 21:27:41 -0700509static void enqueue_huge_page(struct hstate *h, struct page *page)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700510{
511 int nid = page_to_nid(page);
Aneesh Kumar K.V0edaecf2012-07-31 16:42:07 -0700512 list_move(&page->lru, &h->hugepage_freelists[nid]);
Andi Kleena5516432008-07-23 21:27:41 -0700513 h->free_huge_pages++;
514 h->free_huge_pages_node[nid]++;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700515}
516
Naoya Horiguchibf50bab2010-09-08 10:19:33 +0900517static struct page *dequeue_huge_page_node(struct hstate *h, int nid)
518{
519 struct page *page;
520
521 if (list_empty(&h->hugepage_freelists[nid]))
522 return NULL;
523 page = list_entry(h->hugepage_freelists[nid].next, struct page, lru);
Aneesh Kumar K.V0edaecf2012-07-31 16:42:07 -0700524 list_move(&page->lru, &h->hugepage_activelist);
Naoya Horiguchia9869b82010-09-08 10:19:37 +0900525 set_page_refcounted(page);
Naoya Horiguchibf50bab2010-09-08 10:19:33 +0900526 h->free_huge_pages--;
527 h->free_huge_pages_node[nid]--;
528 return page;
529}
530
Andi Kleena5516432008-07-23 21:27:41 -0700531static struct page *dequeue_huge_page_vma(struct hstate *h,
532 struct vm_area_struct *vma,
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700533 unsigned long address, int avoid_reserve)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700534{
Konstantin Khlebnikovb1c12cb2012-04-25 16:01:46 -0700535 struct page *page = NULL;
Lee Schermerhorn480eccf2007-09-18 22:46:47 -0700536 struct mempolicy *mpol;
Mel Gorman19770b32008-04-28 02:12:18 -0700537 nodemask_t *nodemask;
Miao Xiec0ff7452010-05-24 14:32:08 -0700538 struct zonelist *zonelist;
Mel Gormandd1a2392008-04-28 02:12:17 -0700539 struct zone *zone;
540 struct zoneref *z;
Mel Gormancc9a6c82012-03-21 16:34:11 -0700541 unsigned int cpuset_mems_cookie;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700542
Mel Gormancc9a6c82012-03-21 16:34:11 -0700543retry_cpuset:
544 cpuset_mems_cookie = get_mems_allowed();
Miao Xiec0ff7452010-05-24 14:32:08 -0700545 zonelist = huge_zonelist(vma, address,
546 htlb_alloc_mask, &mpol, &nodemask);
Mel Gormana1e78772008-07-23 21:27:23 -0700547 /*
548 * A child process with MAP_PRIVATE mappings created by their parent
549 * have no page reserves. This check ensures that reservations are
550 * not "stolen". The child may still get SIGKILLed
551 */
Mel Gorman7f09ca52008-07-23 21:27:58 -0700552 if (!vma_has_reserves(vma) &&
Andi Kleena5516432008-07-23 21:27:41 -0700553 h->free_huge_pages - h->resv_huge_pages == 0)
Miao Xiec0ff7452010-05-24 14:32:08 -0700554 goto err;
Mel Gormana1e78772008-07-23 21:27:23 -0700555
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700556 /* If reserves cannot be used, ensure enough pages are in the pool */
Andi Kleena5516432008-07-23 21:27:41 -0700557 if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
Justin P. Mattock6eab04a2011-04-08 19:49:08 -0700558 goto err;
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700559
Mel Gorman19770b32008-04-28 02:12:18 -0700560 for_each_zone_zonelist_nodemask(zone, z, zonelist,
561 MAX_NR_ZONES - 1, nodemask) {
Naoya Horiguchibf50bab2010-09-08 10:19:33 +0900562 if (cpuset_zone_allowed_softwall(zone, htlb_alloc_mask)) {
563 page = dequeue_huge_page_node(h, zone_to_nid(zone));
564 if (page) {
565 if (!avoid_reserve)
566 decrement_hugepage_resv_vma(h, vma);
567 break;
568 }
Andrew Morton3abf7af2007-07-19 01:49:08 -0700569 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700570 }
Mel Gormancc9a6c82012-03-21 16:34:11 -0700571
572 mpol_cond_put(mpol);
573 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page))
574 goto retry_cpuset;
575 return page;
576
Miao Xiec0ff7452010-05-24 14:32:08 -0700577err:
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -0700578 mpol_cond_put(mpol);
Mel Gormancc9a6c82012-03-21 16:34:11 -0700579 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700580}
581
Andi Kleena5516432008-07-23 21:27:41 -0700582static void update_and_free_page(struct hstate *h, struct page *page)
Adam Litke6af2acb2007-10-16 01:26:16 -0700583{
584 int i;
Andi Kleena5516432008-07-23 21:27:41 -0700585
Andy Whitcroft18229df2008-11-06 12:53:27 -0800586 VM_BUG_ON(h->order >= MAX_ORDER);
587
Andi Kleena5516432008-07-23 21:27:41 -0700588 h->nr_huge_pages--;
589 h->nr_huge_pages_node[page_to_nid(page)]--;
590 for (i = 0; i < pages_per_huge_page(h); i++) {
Chris Forbes32f84522011-07-25 17:12:14 -0700591 page[i].flags &= ~(1 << PG_locked | 1 << PG_error |
592 1 << PG_referenced | 1 << PG_dirty |
593 1 << PG_active | 1 << PG_reserved |
594 1 << PG_private | 1 << PG_writeback);
Adam Litke6af2acb2007-10-16 01:26:16 -0700595 }
Aneesh Kumar K.V9dd540e2012-07-31 16:42:15 -0700596 VM_BUG_ON(hugetlb_cgroup_from_page(page));
Adam Litke6af2acb2007-10-16 01:26:16 -0700597 set_compound_page_dtor(page, NULL);
598 set_page_refcounted(page);
Gerald Schaefer7f2e9522008-04-28 02:13:29 -0700599 arch_release_hugepage(page);
Andi Kleena5516432008-07-23 21:27:41 -0700600 __free_pages(page, huge_page_order(h));
Adam Litke6af2acb2007-10-16 01:26:16 -0700601}
602
Andi Kleene5ff2152008-07-23 21:27:42 -0700603struct hstate *size_to_hstate(unsigned long size)
604{
605 struct hstate *h;
606
607 for_each_hstate(h) {
608 if (huge_page_size(h) == size)
609 return h;
610 }
611 return NULL;
612}
613
David Gibson27a85ef2006-03-22 00:08:56 -0800614static void free_huge_page(struct page *page)
615{
Andi Kleena5516432008-07-23 21:27:41 -0700616 /*
617 * Can't pass hstate in here because it is called from the
618 * compound page destructor.
619 */
Andi Kleene5ff2152008-07-23 21:27:42 -0700620 struct hstate *h = page_hstate(page);
Adam Litke7893d1d2007-10-16 01:26:18 -0700621 int nid = page_to_nid(page);
David Gibson90481622012-03-21 16:34:12 -0700622 struct hugepage_subpool *spool =
623 (struct hugepage_subpool *)page_private(page);
David Gibson27a85ef2006-03-22 00:08:56 -0800624
Andy Whitcrofte5df70a2008-02-23 15:23:32 -0800625 set_page_private(page, 0);
Mel Gorman23be7462010-04-23 13:17:56 -0400626 page->mapping = NULL;
Adam Litke7893d1d2007-10-16 01:26:18 -0700627 BUG_ON(page_count(page));
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +0900628 BUG_ON(page_mapcount(page));
David Gibson27a85ef2006-03-22 00:08:56 -0800629
630 spin_lock(&hugetlb_lock);
Aneesh Kumar K.V6d76dcf2012-07-31 16:42:18 -0700631 hugetlb_cgroup_uncharge_page(hstate_index(h),
632 pages_per_huge_page(h), page);
Andi Kleenaa888a72008-07-23 21:27:47 -0700633 if (h->surplus_huge_pages_node[nid] && huge_page_order(h) < MAX_ORDER) {
Aneesh Kumar K.V0edaecf2012-07-31 16:42:07 -0700634 /* remove the page from active list */
635 list_del(&page->lru);
Andi Kleena5516432008-07-23 21:27:41 -0700636 update_and_free_page(h, page);
637 h->surplus_huge_pages--;
638 h->surplus_huge_pages_node[nid]--;
Adam Litke7893d1d2007-10-16 01:26:18 -0700639 } else {
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
674int PageHuge(struct page *page)
675{
676 compound_page_dtor *dtor;
677
678 if (!PageCompound(page))
679 return 0;
680
681 page = compound_head(page);
682 dtor = get_compound_page_dtor(page);
683
684 return dtor == free_huge_page;
685}
Naoya Horiguchi43131e12010-05-28 09:29:22 +0900686EXPORT_SYMBOL_GPL(PageHuge);
687
Andi Kleena5516432008-07-23 21:27:41 -0700688static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700689{
Linus Torvalds1da177e2005-04-16 15:20:36 -0700690 struct page *page;
Joe Jinf96efd52007-07-15 23:38:12 -0700691
Andi Kleenaa888a72008-07-23 21:27:47 -0700692 if (h->order >= MAX_ORDER)
693 return NULL;
694
Mel Gorman6484eb32009-06-16 15:31:54 -0700695 page = alloc_pages_exact_node(nid,
Nishanth Aravamudan551883a2008-04-29 00:58:26 -0700696 htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE|
697 __GFP_REPEAT|__GFP_NOWARN,
Andi Kleena5516432008-07-23 21:27:41 -0700698 huge_page_order(h));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700699 if (page) {
Gerald Schaefer7f2e9522008-04-28 02:13:29 -0700700 if (arch_prepare_hugepage(page)) {
Gerald Schaefercaff3a22008-08-12 15:08:38 -0700701 __free_pages(page, huge_page_order(h));
Harvey Harrison7b8ee842008-04-28 14:13:19 -0700702 return NULL;
Gerald Schaefer7f2e9522008-04-28 02:13:29 -0700703 }
Andi Kleena5516432008-07-23 21:27:41 -0700704 prep_new_huge_page(h, page, nid);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700705 }
Nishanth Aravamudan63b46132007-10-16 01:26:24 -0700706
707 return page;
708}
709
Andi Kleen5ced66c2008-07-23 21:27:45 -0700710/*
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800711 * common helper functions for hstate_next_node_to_{alloc|free}.
712 * We may have allocated or freed a huge page based on a different
713 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
714 * be outside of *nodes_allowed. Ensure that we use an allowed
715 * node for alloc or free.
Lee Schermerhorn9a76db02009-12-14 17:58:15 -0800716 */
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800717static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
Lee Schermerhorn9a76db02009-12-14 17:58:15 -0800718{
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800719 nid = next_node(nid, *nodes_allowed);
Lee Schermerhorn9a76db02009-12-14 17:58:15 -0800720 if (nid == MAX_NUMNODES)
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800721 nid = first_node(*nodes_allowed);
Lee Schermerhorn9a76db02009-12-14 17:58:15 -0800722 VM_BUG_ON(nid >= MAX_NUMNODES);
723
724 return nid;
725}
726
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800727static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
Andi Kleen5ced66c2008-07-23 21:27:45 -0700728{
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800729 if (!node_isset(nid, *nodes_allowed))
730 nid = next_node_allowed(nid, nodes_allowed);
Lee Schermerhorn9a76db02009-12-14 17:58:15 -0800731 return nid;
Andi Kleen5ced66c2008-07-23 21:27:45 -0700732}
733
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800734/*
735 * returns the previously saved node ["this node"] from which to
736 * allocate a persistent huge page for the pool and advance the
737 * next node from which to allocate, handling wrap at end of node
738 * mask.
739 */
740static int hstate_next_node_to_alloc(struct hstate *h,
741 nodemask_t *nodes_allowed)
742{
743 int nid;
744
745 VM_BUG_ON(!nodes_allowed);
746
747 nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
748 h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
749
750 return nid;
751}
752
753static int alloc_fresh_huge_page(struct hstate *h, nodemask_t *nodes_allowed)
Nishanth Aravamudan63b46132007-10-16 01:26:24 -0700754{
755 struct page *page;
756 int start_nid;
757 int next_nid;
758 int ret = 0;
759
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800760 start_nid = hstate_next_node_to_alloc(h, nodes_allowed);
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700761 next_nid = start_nid;
Nishanth Aravamudan63b46132007-10-16 01:26:24 -0700762
763 do {
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700764 page = alloc_fresh_huge_page_node(h, next_nid);
Lee Schermerhorn9a76db02009-12-14 17:58:15 -0800765 if (page) {
Nishanth Aravamudan63b46132007-10-16 01:26:24 -0700766 ret = 1;
Lee Schermerhorn9a76db02009-12-14 17:58:15 -0800767 break;
768 }
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800769 next_nid = hstate_next_node_to_alloc(h, nodes_allowed);
Lee Schermerhorn9a76db02009-12-14 17:58:15 -0800770 } while (next_nid != start_nid);
Nishanth Aravamudan63b46132007-10-16 01:26:24 -0700771
Adam Litke3b116302008-04-28 02:13:06 -0700772 if (ret)
773 count_vm_event(HTLB_BUDDY_PGALLOC);
774 else
775 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
776
Nishanth Aravamudan63b46132007-10-16 01:26:24 -0700777 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700778}
779
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700780/*
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800781 * helper for free_pool_huge_page() - return the previously saved
782 * node ["this node"] from which to free a huge page. Advance the
783 * next node id whether or not we find a free huge page to free so
784 * that the next attempt to free addresses the next node.
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700785 */
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800786static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700787{
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800788 int nid;
Lee Schermerhorn9a76db02009-12-14 17:58:15 -0800789
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800790 VM_BUG_ON(!nodes_allowed);
791
792 nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
793 h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
794
Lee Schermerhorn9a76db02009-12-14 17:58:15 -0800795 return nid;
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700796}
797
798/*
799 * Free huge page from pool from next node to free.
800 * Attempt to keep persistent huge pages more or less
801 * balanced over allowed nodes.
802 * Called with hugetlb_lock locked.
803 */
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800804static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
805 bool acct_surplus)
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700806{
807 int start_nid;
808 int next_nid;
809 int ret = 0;
810
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800811 start_nid = hstate_next_node_to_free(h, nodes_allowed);
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700812 next_nid = start_nid;
813
814 do {
Lee Schermerhorn685f3452009-09-21 17:01:23 -0700815 /*
816 * If we're returning unused surplus pages, only examine
817 * nodes with surplus pages.
818 */
819 if ((!acct_surplus || h->surplus_huge_pages_node[next_nid]) &&
820 !list_empty(&h->hugepage_freelists[next_nid])) {
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700821 struct page *page =
822 list_entry(h->hugepage_freelists[next_nid].next,
823 struct page, lru);
824 list_del(&page->lru);
825 h->free_huge_pages--;
826 h->free_huge_pages_node[next_nid]--;
Lee Schermerhorn685f3452009-09-21 17:01:23 -0700827 if (acct_surplus) {
828 h->surplus_huge_pages--;
829 h->surplus_huge_pages_node[next_nid]--;
830 }
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700831 update_and_free_page(h, page);
832 ret = 1;
Lee Schermerhorn9a76db02009-12-14 17:58:15 -0800833 break;
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700834 }
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800835 next_nid = hstate_next_node_to_free(h, nodes_allowed);
Lee Schermerhorn9a76db02009-12-14 17:58:15 -0800836 } while (next_nid != start_nid);
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700837
838 return ret;
839}
840
Naoya Horiguchibf50bab2010-09-08 10:19:33 +0900841static struct page *alloc_buddy_huge_page(struct hstate *h, int nid)
Adam Litke7893d1d2007-10-16 01:26:18 -0700842{
843 struct page *page;
Naoya Horiguchibf50bab2010-09-08 10:19:33 +0900844 unsigned int r_nid;
Adam Litke7893d1d2007-10-16 01:26:18 -0700845
Andi Kleenaa888a72008-07-23 21:27:47 -0700846 if (h->order >= MAX_ORDER)
847 return NULL;
848
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800849 /*
850 * Assume we will successfully allocate the surplus page to
851 * prevent racing processes from causing the surplus to exceed
852 * overcommit
853 *
854 * This however introduces a different race, where a process B
855 * tries to grow the static hugepage pool while alloc_pages() is
856 * called by process A. B will only examine the per-node
857 * counters in determining if surplus huge pages can be
858 * converted to normal huge pages in adjust_pool_surplus(). A
859 * won't be able to increment the per-node counter, until the
860 * lock is dropped by B, but B doesn't drop hugetlb_lock until
861 * no more huge pages can be converted from surplus to normal
862 * state (and doesn't try to convert again). Thus, we have a
863 * case where a surplus huge page exists, the pool is grown, and
864 * the surplus huge page still exists after, even though it
865 * should just have been converted to a normal huge page. This
866 * does not leak memory, though, as the hugepage will be freed
867 * once it is out of use. It also does not allow the counters to
868 * go out of whack in adjust_pool_surplus() as we don't modify
869 * the node values until we've gotten the hugepage and only the
870 * per-node value is checked there.
871 */
872 spin_lock(&hugetlb_lock);
Andi Kleena5516432008-07-23 21:27:41 -0700873 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800874 spin_unlock(&hugetlb_lock);
875 return NULL;
876 } else {
Andi Kleena5516432008-07-23 21:27:41 -0700877 h->nr_huge_pages++;
878 h->surplus_huge_pages++;
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800879 }
880 spin_unlock(&hugetlb_lock);
881
Naoya Horiguchibf50bab2010-09-08 10:19:33 +0900882 if (nid == NUMA_NO_NODE)
883 page = alloc_pages(htlb_alloc_mask|__GFP_COMP|
884 __GFP_REPEAT|__GFP_NOWARN,
885 huge_page_order(h));
886 else
887 page = alloc_pages_exact_node(nid,
888 htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE|
889 __GFP_REPEAT|__GFP_NOWARN, huge_page_order(h));
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800890
Gerald Schaefercaff3a22008-08-12 15:08:38 -0700891 if (page && arch_prepare_hugepage(page)) {
892 __free_pages(page, huge_page_order(h));
Hillf Dantonea5768c2012-01-10 15:08:30 -0800893 page = NULL;
Gerald Schaefercaff3a22008-08-12 15:08:38 -0700894 }
895
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800896 spin_lock(&hugetlb_lock);
Adam Litke7893d1d2007-10-16 01:26:18 -0700897 if (page) {
Aneesh Kumar K.V0edaecf2012-07-31 16:42:07 -0700898 INIT_LIST_HEAD(&page->lru);
Naoya Horiguchibf50bab2010-09-08 10:19:33 +0900899 r_nid = page_to_nid(page);
Adam Litke7893d1d2007-10-16 01:26:18 -0700900 set_compound_page_dtor(page, free_huge_page);
Aneesh Kumar K.V9dd540e2012-07-31 16:42:15 -0700901 set_hugetlb_cgroup(page, NULL);
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800902 /*
903 * We incremented the global counters already
904 */
Naoya Horiguchibf50bab2010-09-08 10:19:33 +0900905 h->nr_huge_pages_node[r_nid]++;
906 h->surplus_huge_pages_node[r_nid]++;
Adam Litke3b116302008-04-28 02:13:06 -0700907 __count_vm_event(HTLB_BUDDY_PGALLOC);
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800908 } else {
Andi Kleena5516432008-07-23 21:27:41 -0700909 h->nr_huge_pages--;
910 h->surplus_huge_pages--;
Adam Litke3b116302008-04-28 02:13:06 -0700911 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
Adam Litke7893d1d2007-10-16 01:26:18 -0700912 }
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800913 spin_unlock(&hugetlb_lock);
Adam Litke7893d1d2007-10-16 01:26:18 -0700914
915 return page;
916}
917
Adam Litkee4e574b2007-10-16 01:26:19 -0700918/*
Naoya Horiguchibf50bab2010-09-08 10:19:33 +0900919 * This allocation function is useful in the context where vma is irrelevant.
920 * E.g. soft-offlining uses this function because it only cares physical
921 * address of error page.
922 */
923struct page *alloc_huge_page_node(struct hstate *h, int nid)
924{
925 struct page *page;
926
927 spin_lock(&hugetlb_lock);
928 page = dequeue_huge_page_node(h, nid);
929 spin_unlock(&hugetlb_lock);
930
Aneesh Kumar K.V94ae8ba2012-07-31 16:42:35 -0700931 if (!page)
Naoya Horiguchibf50bab2010-09-08 10:19:33 +0900932 page = alloc_buddy_huge_page(h, nid);
933
934 return page;
935}
936
937/*
Lucas De Marchi25985ed2011-03-30 22:57:33 -0300938 * Increase the hugetlb pool such that it can accommodate a reservation
Adam Litkee4e574b2007-10-16 01:26:19 -0700939 * of size 'delta'.
940 */
Andi Kleena5516432008-07-23 21:27:41 -0700941static int gather_surplus_pages(struct hstate *h, int delta)
Adam Litkee4e574b2007-10-16 01:26:19 -0700942{
943 struct list_head surplus_list;
944 struct page *page, *tmp;
945 int ret, i;
946 int needed, allocated;
Hillf Danton28073b02012-03-21 16:34:00 -0700947 bool alloc_ok = true;
Adam Litkee4e574b2007-10-16 01:26:19 -0700948
Andi Kleena5516432008-07-23 21:27:41 -0700949 needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
Adam Litkeac09b3a2008-03-04 14:29:38 -0800950 if (needed <= 0) {
Andi Kleena5516432008-07-23 21:27:41 -0700951 h->resv_huge_pages += delta;
Adam Litkee4e574b2007-10-16 01:26:19 -0700952 return 0;
Adam Litkeac09b3a2008-03-04 14:29:38 -0800953 }
Adam Litkee4e574b2007-10-16 01:26:19 -0700954
955 allocated = 0;
956 INIT_LIST_HEAD(&surplus_list);
957
958 ret = -ENOMEM;
959retry:
960 spin_unlock(&hugetlb_lock);
961 for (i = 0; i < needed; i++) {
Naoya Horiguchibf50bab2010-09-08 10:19:33 +0900962 page = alloc_buddy_huge_page(h, NUMA_NO_NODE);
Hillf Danton28073b02012-03-21 16:34:00 -0700963 if (!page) {
964 alloc_ok = false;
965 break;
966 }
Adam Litkee4e574b2007-10-16 01:26:19 -0700967 list_add(&page->lru, &surplus_list);
968 }
Hillf Danton28073b02012-03-21 16:34:00 -0700969 allocated += i;
Adam Litkee4e574b2007-10-16 01:26:19 -0700970
971 /*
972 * After retaking hugetlb_lock, we need to recalculate 'needed'
973 * because either resv_huge_pages or free_huge_pages may have changed.
974 */
975 spin_lock(&hugetlb_lock);
Andi Kleena5516432008-07-23 21:27:41 -0700976 needed = (h->resv_huge_pages + delta) -
977 (h->free_huge_pages + allocated);
Hillf Danton28073b02012-03-21 16:34:00 -0700978 if (needed > 0) {
979 if (alloc_ok)
980 goto retry;
981 /*
982 * We were not able to allocate enough pages to
983 * satisfy the entire reservation so we free what
984 * we've allocated so far.
985 */
986 goto free;
987 }
Adam Litkee4e574b2007-10-16 01:26:19 -0700988 /*
989 * The surplus_list now contains _at_least_ the number of extra pages
Lucas De Marchi25985ed2011-03-30 22:57:33 -0300990 * needed to accommodate the reservation. Add the appropriate number
Adam Litkee4e574b2007-10-16 01:26:19 -0700991 * of pages to the hugetlb pool and free the extras back to the buddy
Adam Litkeac09b3a2008-03-04 14:29:38 -0800992 * allocator. Commit the entire reservation here to prevent another
993 * process from stealing the pages as they are added to the pool but
994 * before they are reserved.
Adam Litkee4e574b2007-10-16 01:26:19 -0700995 */
996 needed += allocated;
Andi Kleena5516432008-07-23 21:27:41 -0700997 h->resv_huge_pages += delta;
Adam Litkee4e574b2007-10-16 01:26:19 -0700998 ret = 0;
Naoya Horiguchia9869b82010-09-08 10:19:37 +0900999
Adam Litke19fc3f02008-04-28 02:12:20 -07001000 /* Free the needed pages to the hugetlb pool */
Adam Litkee4e574b2007-10-16 01:26:19 -07001001 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
Adam Litke19fc3f02008-04-28 02:12:20 -07001002 if ((--needed) < 0)
1003 break;
Naoya Horiguchia9869b82010-09-08 10:19:37 +09001004 /*
1005 * This page is now managed by the hugetlb allocator and has
1006 * no users -- drop the buddy allocator's reference.
1007 */
1008 put_page_testzero(page);
1009 VM_BUG_ON(page_count(page));
Andi Kleena5516432008-07-23 21:27:41 -07001010 enqueue_huge_page(h, page);
Adam Litke19fc3f02008-04-28 02:12:20 -07001011 }
Hillf Danton28073b02012-03-21 16:34:00 -07001012free:
Hillf Dantonb0365c82011-12-28 15:57:16 -08001013 spin_unlock(&hugetlb_lock);
Adam Litke19fc3f02008-04-28 02:12:20 -07001014
1015 /* Free unnecessary surplus pages to the buddy allocator */
1016 if (!list_empty(&surplus_list)) {
Adam Litke19fc3f02008-04-28 02:12:20 -07001017 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
Naoya Horiguchia9869b82010-09-08 10:19:37 +09001018 put_page(page);
Adam Litkeaf767cb2007-10-16 01:26:25 -07001019 }
Adam Litkee4e574b2007-10-16 01:26:19 -07001020 }
Naoya Horiguchia9869b82010-09-08 10:19:37 +09001021 spin_lock(&hugetlb_lock);
Adam Litkee4e574b2007-10-16 01:26:19 -07001022
1023 return ret;
1024}
1025
1026/*
1027 * When releasing a hugetlb pool reservation, any surplus pages that were
1028 * allocated to satisfy the reservation must be explicitly freed if they were
1029 * never used.
Lee Schermerhorn685f3452009-09-21 17:01:23 -07001030 * Called with hugetlb_lock held.
Adam Litkee4e574b2007-10-16 01:26:19 -07001031 */
Andi Kleena5516432008-07-23 21:27:41 -07001032static void return_unused_surplus_pages(struct hstate *h,
1033 unsigned long unused_resv_pages)
Adam Litkee4e574b2007-10-16 01:26:19 -07001034{
Adam Litkee4e574b2007-10-16 01:26:19 -07001035 unsigned long nr_pages;
1036
Adam Litkeac09b3a2008-03-04 14:29:38 -08001037 /* Uncommit the reservation */
Andi Kleena5516432008-07-23 21:27:41 -07001038 h->resv_huge_pages -= unused_resv_pages;
Adam Litkeac09b3a2008-03-04 14:29:38 -08001039
Andi Kleenaa888a72008-07-23 21:27:47 -07001040 /* Cannot return gigantic pages currently */
1041 if (h->order >= MAX_ORDER)
1042 return;
1043
Andi Kleena5516432008-07-23 21:27:41 -07001044 nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
Adam Litkee4e574b2007-10-16 01:26:19 -07001045
Lee Schermerhorn685f3452009-09-21 17:01:23 -07001046 /*
1047 * We want to release as many surplus pages as possible, spread
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -08001048 * evenly across all nodes with memory. Iterate across these nodes
1049 * until we can no longer free unreserved surplus pages. This occurs
1050 * when the nodes with surplus pages have no free pages.
1051 * free_pool_huge_page() will balance the the freed pages across the
1052 * on-line nodes with memory and will handle the hstate accounting.
Lee Schermerhorn685f3452009-09-21 17:01:23 -07001053 */
1054 while (nr_pages--) {
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -08001055 if (!free_pool_huge_page(h, &node_states[N_HIGH_MEMORY], 1))
Lee Schermerhorn685f3452009-09-21 17:01:23 -07001056 break;
Adam Litkee4e574b2007-10-16 01:26:19 -07001057 }
1058}
1059
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001060/*
1061 * Determine if the huge page at addr within the vma has an associated
1062 * reservation. Where it does not we will need to logically increase
David Gibson90481622012-03-21 16:34:12 -07001063 * reservation and actually increase subpool usage before an allocation
1064 * can occur. Where any new reservation would be required the
1065 * reservation change is prepared, but not committed. Once the page
1066 * has been allocated from the subpool and instantiated the change should
1067 * be committed via vma_commit_reservation. No action is required on
1068 * failure.
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001069 */
Roel Kluine2f17d92009-03-31 15:23:15 -07001070static long vma_needs_reservation(struct hstate *h,
Andi Kleena5516432008-07-23 21:27:41 -07001071 struct vm_area_struct *vma, unsigned long addr)
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001072{
1073 struct address_space *mapping = vma->vm_file->f_mapping;
1074 struct inode *inode = mapping->host;
1075
Mel Gormanf83a2752009-05-28 14:34:40 -07001076 if (vma->vm_flags & VM_MAYSHARE) {
Andi Kleena5516432008-07-23 21:27:41 -07001077 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001078 return region_chg(&inode->i_mapping->private_list,
1079 idx, idx + 1);
1080
Andy Whitcroft84afd992008-07-23 21:27:32 -07001081 } else if (!is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
1082 return 1;
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001083
Andy Whitcroft84afd992008-07-23 21:27:32 -07001084 } else {
Roel Kluine2f17d92009-03-31 15:23:15 -07001085 long err;
Andi Kleena5516432008-07-23 21:27:41 -07001086 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
Andy Whitcroft84afd992008-07-23 21:27:32 -07001087 struct resv_map *reservations = vma_resv_map(vma);
1088
1089 err = region_chg(&reservations->regions, idx, idx + 1);
1090 if (err < 0)
1091 return err;
1092 return 0;
1093 }
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001094}
Andi Kleena5516432008-07-23 21:27:41 -07001095static void vma_commit_reservation(struct hstate *h,
1096 struct vm_area_struct *vma, unsigned long addr)
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001097{
1098 struct address_space *mapping = vma->vm_file->f_mapping;
1099 struct inode *inode = mapping->host;
1100
Mel Gormanf83a2752009-05-28 14:34:40 -07001101 if (vma->vm_flags & VM_MAYSHARE) {
Andi Kleena5516432008-07-23 21:27:41 -07001102 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001103 region_add(&inode->i_mapping->private_list, idx, idx + 1);
Andy Whitcroft84afd992008-07-23 21:27:32 -07001104
1105 } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
Andi Kleena5516432008-07-23 21:27:41 -07001106 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
Andy Whitcroft84afd992008-07-23 21:27:32 -07001107 struct resv_map *reservations = vma_resv_map(vma);
1108
1109 /* Mark this page used in the map. */
1110 region_add(&reservations->regions, idx, idx + 1);
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001111 }
1112}
1113
David Gibson27a85ef2006-03-22 00:08:56 -08001114static struct page *alloc_huge_page(struct vm_area_struct *vma,
Mel Gorman04f2cbe2008-07-23 21:27:25 -07001115 unsigned long addr, int avoid_reserve)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001116{
David Gibson90481622012-03-21 16:34:12 -07001117 struct hugepage_subpool *spool = subpool_vma(vma);
Andi Kleena5516432008-07-23 21:27:41 -07001118 struct hstate *h = hstate_vma(vma);
Adam Litke348ea202007-11-14 16:59:37 -08001119 struct page *page;
Roel Kluine2f17d92009-03-31 15:23:15 -07001120 long chg;
Aneesh Kumar K.V6d76dcf2012-07-31 16:42:18 -07001121 int ret, idx;
1122 struct hugetlb_cgroup *h_cg;
Adam Litke2fc39ce2007-11-14 16:59:39 -08001123
Aneesh Kumar K.V6d76dcf2012-07-31 16:42:18 -07001124 idx = hstate_index(h);
Mel Gormana1e78772008-07-23 21:27:23 -07001125 /*
David Gibson90481622012-03-21 16:34:12 -07001126 * Processes that did not create the mapping will have no
1127 * reserves and will not have accounted against subpool
1128 * limit. Check that the subpool limit can be made before
1129 * satisfying the allocation MAP_NORESERVE mappings may also
1130 * need pages and subpool limit allocated allocated if no reserve
1131 * mapping overlaps.
Mel Gormana1e78772008-07-23 21:27:23 -07001132 */
Andi Kleena5516432008-07-23 21:27:41 -07001133 chg = vma_needs_reservation(h, vma, addr);
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001134 if (chg < 0)
Aneesh Kumar K.V76dcee72012-07-31 16:41:57 -07001135 return ERR_PTR(-ENOMEM);
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001136 if (chg)
David Gibson90481622012-03-21 16:34:12 -07001137 if (hugepage_subpool_get_pages(spool, chg))
Aneesh Kumar K.V76dcee72012-07-31 16:41:57 -07001138 return ERR_PTR(-ENOSPC);
Mel Gormana1e78772008-07-23 21:27:23 -07001139
Aneesh Kumar K.V6d76dcf2012-07-31 16:42:18 -07001140 ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg);
1141 if (ret) {
1142 hugepage_subpool_put_pages(spool, chg);
1143 return ERR_PTR(-ENOSPC);
1144 }
Mel Gormana1e78772008-07-23 21:27:23 -07001145 spin_lock(&hugetlb_lock);
Andi Kleena5516432008-07-23 21:27:41 -07001146 page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve);
Aneesh Kumar K.V94ae8ba2012-07-31 16:42:35 -07001147 if (page) {
1148 /* update page cgroup details */
1149 hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h),
1150 h_cg, page);
1151 spin_unlock(&hugetlb_lock);
1152 } else {
1153 spin_unlock(&hugetlb_lock);
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001154 page = alloc_buddy_huge_page(h, NUMA_NO_NODE);
Mel Gormana1e78772008-07-23 21:27:23 -07001155 if (!page) {
Aneesh Kumar K.V6d76dcf2012-07-31 16:42:18 -07001156 hugetlb_cgroup_uncharge_cgroup(idx,
1157 pages_per_huge_page(h),
1158 h_cg);
David Gibson90481622012-03-21 16:34:12 -07001159 hugepage_subpool_put_pages(spool, chg);
Aneesh Kumar K.V76dcee72012-07-31 16:41:57 -07001160 return ERR_PTR(-ENOSPC);
Mel Gormana1e78772008-07-23 21:27:23 -07001161 }
Aneesh Kumar K.V79dbb232012-07-31 16:42:32 -07001162 spin_lock(&hugetlb_lock);
Aneesh Kumar K.V94ae8ba2012-07-31 16:42:35 -07001163 hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h),
1164 h_cg, page);
Aneesh Kumar K.V79dbb232012-07-31 16:42:32 -07001165 list_move(&page->lru, &h->hugepage_activelist);
1166 spin_unlock(&hugetlb_lock);
Mel Gormana1e78772008-07-23 21:27:23 -07001167 }
1168
David Gibson90481622012-03-21 16:34:12 -07001169 set_page_private(page, (unsigned long)spool);
Mel Gormana1e78772008-07-23 21:27:23 -07001170
Andi Kleena5516432008-07-23 21:27:41 -07001171 vma_commit_reservation(h, vma, addr);
Adam Litke90d8b7e2007-11-14 16:59:42 -08001172 return page;
David Gibsonb45b5bd2006-03-22 00:08:55 -08001173}
1174
Cyrill Gorcunov91f47662009-01-06 14:40:33 -08001175int __weak alloc_bootmem_huge_page(struct hstate *h)
Andi Kleenaa888a72008-07-23 21:27:47 -07001176{
1177 struct huge_bootmem_page *m;
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -08001178 int nr_nodes = nodes_weight(node_states[N_HIGH_MEMORY]);
Andi Kleenaa888a72008-07-23 21:27:47 -07001179
1180 while (nr_nodes) {
1181 void *addr;
1182
1183 addr = __alloc_bootmem_node_nopanic(
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001184 NODE_DATA(hstate_next_node_to_alloc(h,
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -08001185 &node_states[N_HIGH_MEMORY])),
Andi Kleenaa888a72008-07-23 21:27:47 -07001186 huge_page_size(h), huge_page_size(h), 0);
1187
1188 if (addr) {
1189 /*
1190 * Use the beginning of the huge page to store the
1191 * huge_bootmem_page struct (until gather_bootmem
1192 * puts them into the mem_map).
1193 */
1194 m = addr;
Cyrill Gorcunov91f47662009-01-06 14:40:33 -08001195 goto found;
Andi Kleenaa888a72008-07-23 21:27:47 -07001196 }
Andi Kleenaa888a72008-07-23 21:27:47 -07001197 nr_nodes--;
1198 }
1199 return 0;
1200
1201found:
1202 BUG_ON((unsigned long)virt_to_phys(m) & (huge_page_size(h) - 1));
1203 /* Put them into a private list first because mem_map is not up yet */
1204 list_add(&m->list, &huge_boot_pages);
1205 m->hstate = h;
1206 return 1;
1207}
1208
Andy Whitcroft18229df2008-11-06 12:53:27 -08001209static void prep_compound_huge_page(struct page *page, int order)
1210{
1211 if (unlikely(order > (MAX_ORDER - 1)))
1212 prep_compound_gigantic_page(page, order);
1213 else
1214 prep_compound_page(page, order);
1215}
1216
Andi Kleenaa888a72008-07-23 21:27:47 -07001217/* Put bootmem huge pages into the standard lists after mem_map is up */
1218static void __init gather_bootmem_prealloc(void)
1219{
1220 struct huge_bootmem_page *m;
1221
1222 list_for_each_entry(m, &huge_boot_pages, list) {
Andi Kleenaa888a72008-07-23 21:27:47 -07001223 struct hstate *h = m->hstate;
Becky Bruceee8f2482011-07-25 17:11:50 -07001224 struct page *page;
1225
1226#ifdef CONFIG_HIGHMEM
1227 page = pfn_to_page(m->phys >> PAGE_SHIFT);
1228 free_bootmem_late((unsigned long)m,
1229 sizeof(struct huge_bootmem_page));
1230#else
1231 page = virt_to_page(m);
1232#endif
Andi Kleenaa888a72008-07-23 21:27:47 -07001233 __ClearPageReserved(page);
1234 WARN_ON(page_count(page) != 1);
Andy Whitcroft18229df2008-11-06 12:53:27 -08001235 prep_compound_huge_page(page, h->order);
Andi Kleenaa888a72008-07-23 21:27:47 -07001236 prep_new_huge_page(h, page, page_to_nid(page));
Rafael Aquinib0320c72011-06-15 15:08:39 -07001237 /*
1238 * If we had gigantic hugepages allocated at boot time, we need
1239 * to restore the 'stolen' pages to totalram_pages in order to
1240 * fix confusing memory reports from free(1) and another
1241 * side-effects, like CommitLimit going negative.
1242 */
1243 if (h->order > (MAX_ORDER - 1))
1244 totalram_pages += 1 << h->order;
Andi Kleenaa888a72008-07-23 21:27:47 -07001245 }
1246}
1247
Andi Kleen8faa8b02008-07-23 21:27:48 -07001248static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001249{
1250 unsigned long i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001251
Andi Kleene5ff2152008-07-23 21:27:42 -07001252 for (i = 0; i < h->max_huge_pages; ++i) {
Andi Kleenaa888a72008-07-23 21:27:47 -07001253 if (h->order >= MAX_ORDER) {
1254 if (!alloc_bootmem_huge_page(h))
1255 break;
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -08001256 } else if (!alloc_fresh_huge_page(h,
1257 &node_states[N_HIGH_MEMORY]))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001258 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001259 }
Andi Kleen8faa8b02008-07-23 21:27:48 -07001260 h->max_huge_pages = i;
Andi Kleene5ff2152008-07-23 21:27:42 -07001261}
1262
1263static void __init hugetlb_init_hstates(void)
1264{
1265 struct hstate *h;
1266
1267 for_each_hstate(h) {
Andi Kleen8faa8b02008-07-23 21:27:48 -07001268 /* oversize hugepages were init'ed in early boot */
1269 if (h->order < MAX_ORDER)
1270 hugetlb_hstate_alloc_pages(h);
Andi Kleene5ff2152008-07-23 21:27:42 -07001271 }
1272}
1273
Andi Kleen4abd32d2008-07-23 21:27:49 -07001274static char * __init memfmt(char *buf, unsigned long n)
1275{
1276 if (n >= (1UL << 30))
1277 sprintf(buf, "%lu GB", n >> 30);
1278 else if (n >= (1UL << 20))
1279 sprintf(buf, "%lu MB", n >> 20);
1280 else
1281 sprintf(buf, "%lu KB", n >> 10);
1282 return buf;
1283}
1284
Andi Kleene5ff2152008-07-23 21:27:42 -07001285static void __init report_hugepages(void)
1286{
1287 struct hstate *h;
1288
1289 for_each_hstate(h) {
Andi Kleen4abd32d2008-07-23 21:27:49 -07001290 char buf[32];
1291 printk(KERN_INFO "HugeTLB registered %s page size, "
1292 "pre-allocated %ld pages\n",
1293 memfmt(buf, huge_page_size(h)),
1294 h->free_huge_pages);
Andi Kleene5ff2152008-07-23 21:27:42 -07001295 }
1296}
1297
Linus Torvalds1da177e2005-04-16 15:20:36 -07001298#ifdef CONFIG_HIGHMEM
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001299static void try_to_free_low(struct hstate *h, unsigned long count,
1300 nodemask_t *nodes_allowed)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001301{
Christoph Lameter4415cc82006-09-25 23:31:55 -07001302 int i;
1303
Andi Kleenaa888a72008-07-23 21:27:47 -07001304 if (h->order >= MAX_ORDER)
1305 return;
1306
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001307 for_each_node_mask(i, *nodes_allowed) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001308 struct page *page, *next;
Andi Kleena5516432008-07-23 21:27:41 -07001309 struct list_head *freel = &h->hugepage_freelists[i];
1310 list_for_each_entry_safe(page, next, freel, lru) {
1311 if (count >= h->nr_huge_pages)
Adam Litke6b0c8802007-10-16 01:26:23 -07001312 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001313 if (PageHighMem(page))
1314 continue;
1315 list_del(&page->lru);
Andi Kleene5ff2152008-07-23 21:27:42 -07001316 update_and_free_page(h, page);
Andi Kleena5516432008-07-23 21:27:41 -07001317 h->free_huge_pages--;
1318 h->free_huge_pages_node[page_to_nid(page)]--;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001319 }
1320 }
1321}
1322#else
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001323static inline void try_to_free_low(struct hstate *h, unsigned long count,
1324 nodemask_t *nodes_allowed)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001325{
1326}
1327#endif
1328
Wu Fengguang20a03072009-06-16 15:32:22 -07001329/*
1330 * Increment or decrement surplus_huge_pages. Keep node-specific counters
1331 * balanced by operating on them in a round-robin fashion.
1332 * Returns 1 if an adjustment was made.
1333 */
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001334static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
1335 int delta)
Wu Fengguang20a03072009-06-16 15:32:22 -07001336{
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001337 int start_nid, next_nid;
Wu Fengguang20a03072009-06-16 15:32:22 -07001338 int ret = 0;
1339
1340 VM_BUG_ON(delta != -1 && delta != 1);
Wu Fengguang20a03072009-06-16 15:32:22 -07001341
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001342 if (delta < 0)
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001343 start_nid = hstate_next_node_to_alloc(h, nodes_allowed);
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001344 else
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001345 start_nid = hstate_next_node_to_free(h, nodes_allowed);
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001346 next_nid = start_nid;
1347
1348 do {
1349 int nid = next_nid;
1350 if (delta < 0) {
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001351 /*
1352 * To shrink on this node, there must be a surplus page
1353 */
Lee Schermerhorn9a76db02009-12-14 17:58:15 -08001354 if (!h->surplus_huge_pages_node[nid]) {
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001355 next_nid = hstate_next_node_to_alloc(h,
1356 nodes_allowed);
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001357 continue;
Lee Schermerhorn9a76db02009-12-14 17:58:15 -08001358 }
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001359 }
1360 if (delta > 0) {
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001361 /*
1362 * Surplus cannot exceed the total number of pages
1363 */
1364 if (h->surplus_huge_pages_node[nid] >=
Lee Schermerhorn9a76db02009-12-14 17:58:15 -08001365 h->nr_huge_pages_node[nid]) {
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001366 next_nid = hstate_next_node_to_free(h,
1367 nodes_allowed);
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001368 continue;
Lee Schermerhorn9a76db02009-12-14 17:58:15 -08001369 }
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001370 }
Wu Fengguang20a03072009-06-16 15:32:22 -07001371
1372 h->surplus_huge_pages += delta;
1373 h->surplus_huge_pages_node[nid] += delta;
1374 ret = 1;
1375 break;
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001376 } while (next_nid != start_nid);
Wu Fengguang20a03072009-06-16 15:32:22 -07001377
Wu Fengguang20a03072009-06-16 15:32:22 -07001378 return ret;
1379}
1380
Andi Kleena5516432008-07-23 21:27:41 -07001381#define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001382static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count,
1383 nodemask_t *nodes_allowed)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001384{
Adam Litke7893d1d2007-10-16 01:26:18 -07001385 unsigned long min_count, ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001386
Andi Kleenaa888a72008-07-23 21:27:47 -07001387 if (h->order >= MAX_ORDER)
1388 return h->max_huge_pages;
1389
Adam Litke7893d1d2007-10-16 01:26:18 -07001390 /*
1391 * Increase the pool size
1392 * First take pages out of surplus state. Then make up the
1393 * remaining difference by allocating fresh huge pages.
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001394 *
1395 * We might race with alloc_buddy_huge_page() here and be unable
1396 * to convert a surplus huge page to a normal huge page. That is
1397 * not critical, though, it just means the overall size of the
1398 * pool might be one hugepage larger than it needs to be, but
1399 * within all the constraints specified by the sysctls.
Adam Litke7893d1d2007-10-16 01:26:18 -07001400 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001401 spin_lock(&hugetlb_lock);
Andi Kleena5516432008-07-23 21:27:41 -07001402 while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001403 if (!adjust_pool_surplus(h, nodes_allowed, -1))
Adam Litke7893d1d2007-10-16 01:26:18 -07001404 break;
1405 }
1406
Andi Kleena5516432008-07-23 21:27:41 -07001407 while (count > persistent_huge_pages(h)) {
Adam Litke7893d1d2007-10-16 01:26:18 -07001408 /*
1409 * If this allocation races such that we no longer need the
1410 * page, free_huge_page will handle it by freeing the page
1411 * and reducing the surplus.
1412 */
1413 spin_unlock(&hugetlb_lock);
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001414 ret = alloc_fresh_huge_page(h, nodes_allowed);
Adam Litke7893d1d2007-10-16 01:26:18 -07001415 spin_lock(&hugetlb_lock);
1416 if (!ret)
1417 goto out;
1418
Mel Gorman536240f22009-12-14 17:59:56 -08001419 /* Bail for signals. Probably ctrl-c from user */
1420 if (signal_pending(current))
1421 goto out;
Adam Litke7893d1d2007-10-16 01:26:18 -07001422 }
Adam Litke7893d1d2007-10-16 01:26:18 -07001423
1424 /*
1425 * Decrease the pool size
1426 * First return free pages to the buddy allocator (being careful
1427 * to keep enough around to satisfy reservations). Then place
1428 * pages into surplus state as needed so the pool will shrink
1429 * to the desired size as pages become free.
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001430 *
1431 * By placing pages into the surplus state independent of the
1432 * overcommit value, we are allowing the surplus pool size to
1433 * exceed overcommit. There are few sane options here. Since
1434 * alloc_buddy_huge_page() is checking the global counter,
1435 * though, we'll note that we're not allowed to exceed surplus
1436 * and won't grow the pool anywhere else. Not until one of the
1437 * sysctls are changed, or the surplus pages go out of use.
Adam Litke7893d1d2007-10-16 01:26:18 -07001438 */
Andi Kleena5516432008-07-23 21:27:41 -07001439 min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
Adam Litke6b0c8802007-10-16 01:26:23 -07001440 min_count = max(count, min_count);
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001441 try_to_free_low(h, min_count, nodes_allowed);
Andi Kleena5516432008-07-23 21:27:41 -07001442 while (min_count < persistent_huge_pages(h)) {
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001443 if (!free_pool_huge_page(h, nodes_allowed, 0))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001444 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001445 }
Andi Kleena5516432008-07-23 21:27:41 -07001446 while (count < persistent_huge_pages(h)) {
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001447 if (!adjust_pool_surplus(h, nodes_allowed, 1))
Adam Litke7893d1d2007-10-16 01:26:18 -07001448 break;
1449 }
1450out:
Andi Kleena5516432008-07-23 21:27:41 -07001451 ret = persistent_huge_pages(h);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001452 spin_unlock(&hugetlb_lock);
Adam Litke7893d1d2007-10-16 01:26:18 -07001453 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001454}
1455
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001456#define HSTATE_ATTR_RO(_name) \
1457 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
1458
1459#define HSTATE_ATTR(_name) \
1460 static struct kobj_attribute _name##_attr = \
1461 __ATTR(_name, 0644, _name##_show, _name##_store)
1462
1463static struct kobject *hugepages_kobj;
1464static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
1465
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001466static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
1467
1468static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001469{
1470 int i;
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001471
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001472 for (i = 0; i < HUGE_MAX_HSTATE; i++)
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001473 if (hstate_kobjs[i] == kobj) {
1474 if (nidp)
1475 *nidp = NUMA_NO_NODE;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001476 return &hstates[i];
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001477 }
1478
1479 return kobj_to_node_hstate(kobj, nidp);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001480}
1481
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001482static ssize_t nr_hugepages_show_common(struct kobject *kobj,
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001483 struct kobj_attribute *attr, char *buf)
1484{
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001485 struct hstate *h;
1486 unsigned long nr_huge_pages;
1487 int nid;
1488
1489 h = kobj_to_hstate(kobj, &nid);
1490 if (nid == NUMA_NO_NODE)
1491 nr_huge_pages = h->nr_huge_pages;
1492 else
1493 nr_huge_pages = h->nr_huge_pages_node[nid];
1494
1495 return sprintf(buf, "%lu\n", nr_huge_pages);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001496}
Eric B Munsonadbe8722011-01-13 15:47:27 -08001497
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001498static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
1499 struct kobject *kobj, struct kobj_attribute *attr,
1500 const char *buf, size_t len)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001501{
1502 int err;
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001503 int nid;
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001504 unsigned long count;
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001505 struct hstate *h;
David Rientjesbad44b52009-12-14 17:58:38 -08001506 NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001507
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001508 err = strict_strtoul(buf, 10, &count);
Eric B Munson73ae31e2011-01-13 15:47:28 -08001509 if (err)
Eric B Munsonadbe8722011-01-13 15:47:27 -08001510 goto out;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001511
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001512 h = kobj_to_hstate(kobj, &nid);
Eric B Munsonadbe8722011-01-13 15:47:27 -08001513 if (h->order >= MAX_ORDER) {
1514 err = -EINVAL;
1515 goto out;
1516 }
1517
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001518 if (nid == NUMA_NO_NODE) {
1519 /*
1520 * global hstate attribute
1521 */
1522 if (!(obey_mempolicy &&
1523 init_nodemask_of_mempolicy(nodes_allowed))) {
1524 NODEMASK_FREE(nodes_allowed);
1525 nodes_allowed = &node_states[N_HIGH_MEMORY];
1526 }
1527 } else if (nodes_allowed) {
1528 /*
1529 * per node hstate attribute: adjust count to global,
1530 * but restrict alloc/free to the specified node.
1531 */
1532 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
1533 init_nodemask_of_node(nodes_allowed, nid);
1534 } else
1535 nodes_allowed = &node_states[N_HIGH_MEMORY];
1536
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001537 h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001538
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -08001539 if (nodes_allowed != &node_states[N_HIGH_MEMORY])
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001540 NODEMASK_FREE(nodes_allowed);
1541
1542 return len;
Eric B Munsonadbe8722011-01-13 15:47:27 -08001543out:
1544 NODEMASK_FREE(nodes_allowed);
1545 return err;
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001546}
1547
1548static ssize_t nr_hugepages_show(struct kobject *kobj,
1549 struct kobj_attribute *attr, char *buf)
1550{
1551 return nr_hugepages_show_common(kobj, attr, buf);
1552}
1553
1554static ssize_t nr_hugepages_store(struct kobject *kobj,
1555 struct kobj_attribute *attr, const char *buf, size_t len)
1556{
1557 return nr_hugepages_store_common(false, kobj, attr, buf, len);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001558}
1559HSTATE_ATTR(nr_hugepages);
1560
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001561#ifdef CONFIG_NUMA
1562
1563/*
1564 * hstate attribute for optionally mempolicy-based constraint on persistent
1565 * huge page alloc/free.
1566 */
1567static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
1568 struct kobj_attribute *attr, char *buf)
1569{
1570 return nr_hugepages_show_common(kobj, attr, buf);
1571}
1572
1573static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
1574 struct kobj_attribute *attr, const char *buf, size_t len)
1575{
1576 return nr_hugepages_store_common(true, kobj, attr, buf, len);
1577}
1578HSTATE_ATTR(nr_hugepages_mempolicy);
1579#endif
1580
1581
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001582static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
1583 struct kobj_attribute *attr, char *buf)
1584{
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001585 struct hstate *h = kobj_to_hstate(kobj, NULL);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001586 return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
1587}
Eric B Munsonadbe8722011-01-13 15:47:27 -08001588
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001589static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
1590 struct kobj_attribute *attr, const char *buf, size_t count)
1591{
1592 int err;
1593 unsigned long input;
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001594 struct hstate *h = kobj_to_hstate(kobj, NULL);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001595
Eric B Munsonadbe8722011-01-13 15:47:27 -08001596 if (h->order >= MAX_ORDER)
1597 return -EINVAL;
1598
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001599 err = strict_strtoul(buf, 10, &input);
1600 if (err)
Eric B Munson73ae31e2011-01-13 15:47:28 -08001601 return err;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001602
1603 spin_lock(&hugetlb_lock);
1604 h->nr_overcommit_huge_pages = input;
1605 spin_unlock(&hugetlb_lock);
1606
1607 return count;
1608}
1609HSTATE_ATTR(nr_overcommit_hugepages);
1610
1611static ssize_t free_hugepages_show(struct kobject *kobj,
1612 struct kobj_attribute *attr, char *buf)
1613{
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001614 struct hstate *h;
1615 unsigned long free_huge_pages;
1616 int nid;
1617
1618 h = kobj_to_hstate(kobj, &nid);
1619 if (nid == NUMA_NO_NODE)
1620 free_huge_pages = h->free_huge_pages;
1621 else
1622 free_huge_pages = h->free_huge_pages_node[nid];
1623
1624 return sprintf(buf, "%lu\n", free_huge_pages);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001625}
1626HSTATE_ATTR_RO(free_hugepages);
1627
1628static ssize_t resv_hugepages_show(struct kobject *kobj,
1629 struct kobj_attribute *attr, char *buf)
1630{
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001631 struct hstate *h = kobj_to_hstate(kobj, NULL);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001632 return sprintf(buf, "%lu\n", h->resv_huge_pages);
1633}
1634HSTATE_ATTR_RO(resv_hugepages);
1635
1636static ssize_t surplus_hugepages_show(struct kobject *kobj,
1637 struct kobj_attribute *attr, char *buf)
1638{
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001639 struct hstate *h;
1640 unsigned long surplus_huge_pages;
1641 int nid;
1642
1643 h = kobj_to_hstate(kobj, &nid);
1644 if (nid == NUMA_NO_NODE)
1645 surplus_huge_pages = h->surplus_huge_pages;
1646 else
1647 surplus_huge_pages = h->surplus_huge_pages_node[nid];
1648
1649 return sprintf(buf, "%lu\n", surplus_huge_pages);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001650}
1651HSTATE_ATTR_RO(surplus_hugepages);
1652
1653static struct attribute *hstate_attrs[] = {
1654 &nr_hugepages_attr.attr,
1655 &nr_overcommit_hugepages_attr.attr,
1656 &free_hugepages_attr.attr,
1657 &resv_hugepages_attr.attr,
1658 &surplus_hugepages_attr.attr,
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001659#ifdef CONFIG_NUMA
1660 &nr_hugepages_mempolicy_attr.attr,
1661#endif
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001662 NULL,
1663};
1664
1665static struct attribute_group hstate_attr_group = {
1666 .attrs = hstate_attrs,
1667};
1668
Jeff Mahoney094e9532010-02-02 13:44:14 -08001669static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
1670 struct kobject **hstate_kobjs,
1671 struct attribute_group *hstate_attr_group)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001672{
1673 int retval;
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07001674 int hi = hstate_index(h);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001675
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001676 hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
1677 if (!hstate_kobjs[hi])
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001678 return -ENOMEM;
1679
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001680 retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001681 if (retval)
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001682 kobject_put(hstate_kobjs[hi]);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001683
1684 return retval;
1685}
1686
1687static void __init hugetlb_sysfs_init(void)
1688{
1689 struct hstate *h;
1690 int err;
1691
1692 hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
1693 if (!hugepages_kobj)
1694 return;
1695
1696 for_each_hstate(h) {
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001697 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
1698 hstate_kobjs, &hstate_attr_group);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001699 if (err)
1700 printk(KERN_ERR "Hugetlb: Unable to add hstate %s",
1701 h->name);
1702 }
1703}
1704
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001705#ifdef CONFIG_NUMA
1706
1707/*
1708 * node_hstate/s - associate per node hstate attributes, via their kobjects,
Kay Sievers10fbcf42011-12-21 14:48:43 -08001709 * with node devices in node_devices[] using a parallel array. The array
1710 * index of a node device or _hstate == node id.
1711 * This is here to avoid any static dependency of the node device driver, in
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001712 * the base kernel, on the hugetlb module.
1713 */
1714struct node_hstate {
1715 struct kobject *hugepages_kobj;
1716 struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
1717};
1718struct node_hstate node_hstates[MAX_NUMNODES];
1719
1720/*
Kay Sievers10fbcf42011-12-21 14:48:43 -08001721 * A subset of global hstate attributes for node devices
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001722 */
1723static struct attribute *per_node_hstate_attrs[] = {
1724 &nr_hugepages_attr.attr,
1725 &free_hugepages_attr.attr,
1726 &surplus_hugepages_attr.attr,
1727 NULL,
1728};
1729
1730static struct attribute_group per_node_hstate_attr_group = {
1731 .attrs = per_node_hstate_attrs,
1732};
1733
1734/*
Kay Sievers10fbcf42011-12-21 14:48:43 -08001735 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001736 * Returns node id via non-NULL nidp.
1737 */
1738static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
1739{
1740 int nid;
1741
1742 for (nid = 0; nid < nr_node_ids; nid++) {
1743 struct node_hstate *nhs = &node_hstates[nid];
1744 int i;
1745 for (i = 0; i < HUGE_MAX_HSTATE; i++)
1746 if (nhs->hstate_kobjs[i] == kobj) {
1747 if (nidp)
1748 *nidp = nid;
1749 return &hstates[i];
1750 }
1751 }
1752
1753 BUG();
1754 return NULL;
1755}
1756
1757/*
Kay Sievers10fbcf42011-12-21 14:48:43 -08001758 * Unregister hstate attributes from a single node device.
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001759 * No-op if no hstate attributes attached.
1760 */
1761void hugetlb_unregister_node(struct node *node)
1762{
1763 struct hstate *h;
Kay Sievers10fbcf42011-12-21 14:48:43 -08001764 struct node_hstate *nhs = &node_hstates[node->dev.id];
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001765
1766 if (!nhs->hugepages_kobj)
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -08001767 return; /* no hstate attributes */
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001768
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07001769 for_each_hstate(h) {
1770 int idx = hstate_index(h);
1771 if (nhs->hstate_kobjs[idx]) {
1772 kobject_put(nhs->hstate_kobjs[idx]);
1773 nhs->hstate_kobjs[idx] = NULL;
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001774 }
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07001775 }
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001776
1777 kobject_put(nhs->hugepages_kobj);
1778 nhs->hugepages_kobj = NULL;
1779}
1780
1781/*
Kay Sievers10fbcf42011-12-21 14:48:43 -08001782 * hugetlb module exit: unregister hstate attributes from node devices
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001783 * that have them.
1784 */
1785static void hugetlb_unregister_all_nodes(void)
1786{
1787 int nid;
1788
1789 /*
Kay Sievers10fbcf42011-12-21 14:48:43 -08001790 * disable node device registrations.
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001791 */
1792 register_hugetlbfs_with_node(NULL, NULL);
1793
1794 /*
1795 * remove hstate attributes from any nodes that have them.
1796 */
1797 for (nid = 0; nid < nr_node_ids; nid++)
1798 hugetlb_unregister_node(&node_devices[nid]);
1799}
1800
1801/*
Kay Sievers10fbcf42011-12-21 14:48:43 -08001802 * Register hstate attributes for a single node device.
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001803 * No-op if attributes already registered.
1804 */
1805void hugetlb_register_node(struct node *node)
1806{
1807 struct hstate *h;
Kay Sievers10fbcf42011-12-21 14:48:43 -08001808 struct node_hstate *nhs = &node_hstates[node->dev.id];
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001809 int err;
1810
1811 if (nhs->hugepages_kobj)
1812 return; /* already allocated */
1813
1814 nhs->hugepages_kobj = kobject_create_and_add("hugepages",
Kay Sievers10fbcf42011-12-21 14:48:43 -08001815 &node->dev.kobj);
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001816 if (!nhs->hugepages_kobj)
1817 return;
1818
1819 for_each_hstate(h) {
1820 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
1821 nhs->hstate_kobjs,
1822 &per_node_hstate_attr_group);
1823 if (err) {
1824 printk(KERN_ERR "Hugetlb: Unable to add hstate %s"
1825 " for node %d\n",
Kay Sievers10fbcf42011-12-21 14:48:43 -08001826 h->name, node->dev.id);
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001827 hugetlb_unregister_node(node);
1828 break;
1829 }
1830 }
1831}
1832
1833/*
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -08001834 * hugetlb init time: register hstate attributes for all registered node
Kay Sievers10fbcf42011-12-21 14:48:43 -08001835 * devices of nodes that have memory. All on-line nodes should have
1836 * registered their associated device by this time.
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001837 */
1838static void hugetlb_register_all_nodes(void)
1839{
1840 int nid;
1841
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -08001842 for_each_node_state(nid, N_HIGH_MEMORY) {
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001843 struct node *node = &node_devices[nid];
Kay Sievers10fbcf42011-12-21 14:48:43 -08001844 if (node->dev.id == nid)
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001845 hugetlb_register_node(node);
1846 }
1847
1848 /*
Kay Sievers10fbcf42011-12-21 14:48:43 -08001849 * Let the node device driver know we're here so it can
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001850 * [un]register hstate attributes on node hotplug.
1851 */
1852 register_hugetlbfs_with_node(hugetlb_register_node,
1853 hugetlb_unregister_node);
1854}
1855#else /* !CONFIG_NUMA */
1856
1857static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
1858{
1859 BUG();
1860 if (nidp)
1861 *nidp = -1;
1862 return NULL;
1863}
1864
1865static void hugetlb_unregister_all_nodes(void) { }
1866
1867static void hugetlb_register_all_nodes(void) { }
1868
1869#endif
1870
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001871static void __exit hugetlb_exit(void)
1872{
1873 struct hstate *h;
1874
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001875 hugetlb_unregister_all_nodes();
1876
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001877 for_each_hstate(h) {
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07001878 kobject_put(hstate_kobjs[hstate_index(h)]);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001879 }
1880
1881 kobject_put(hugepages_kobj);
1882}
1883module_exit(hugetlb_exit);
1884
1885static int __init hugetlb_init(void)
1886{
Benjamin Herrenschmidt0ef89d22008-07-31 00:07:30 -07001887 /* Some platform decide whether they support huge pages at boot
1888 * time. On these, such as powerpc, HPAGE_SHIFT is set to 0 when
1889 * there is no such support
1890 */
1891 if (HPAGE_SHIFT == 0)
1892 return 0;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001893
Nick Piggine11bfbf2008-07-23 21:27:52 -07001894 if (!size_to_hstate(default_hstate_size)) {
1895 default_hstate_size = HPAGE_SIZE;
1896 if (!size_to_hstate(default_hstate_size))
1897 hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001898 }
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07001899 default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size));
Nick Piggine11bfbf2008-07-23 21:27:52 -07001900 if (default_hstate_max_huge_pages)
1901 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001902
1903 hugetlb_init_hstates();
1904
Andi Kleenaa888a72008-07-23 21:27:47 -07001905 gather_bootmem_prealloc();
1906
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001907 report_hugepages();
1908
1909 hugetlb_sysfs_init();
1910
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001911 hugetlb_register_all_nodes();
1912
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001913 return 0;
1914}
1915module_init(hugetlb_init);
1916
1917/* Should be called on processing a hugepagesz=... option */
1918void __init hugetlb_add_hstate(unsigned order)
1919{
1920 struct hstate *h;
Andi Kleen8faa8b02008-07-23 21:27:48 -07001921 unsigned long i;
1922
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001923 if (size_to_hstate(PAGE_SIZE << order)) {
1924 printk(KERN_WARNING "hugepagesz= specified twice, ignoring\n");
1925 return;
1926 }
Aneesh Kumar K.V47d38342012-07-31 16:41:54 -07001927 BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001928 BUG_ON(order == 0);
Aneesh Kumar K.V47d38342012-07-31 16:41:54 -07001929 h = &hstates[hugetlb_max_hstate++];
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001930 h->order = order;
1931 h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
Andi Kleen8faa8b02008-07-23 21:27:48 -07001932 h->nr_huge_pages = 0;
1933 h->free_huge_pages = 0;
1934 for (i = 0; i < MAX_NUMNODES; ++i)
1935 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
Aneesh Kumar K.V0edaecf2012-07-31 16:42:07 -07001936 INIT_LIST_HEAD(&h->hugepage_activelist);
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -08001937 h->next_nid_to_alloc = first_node(node_states[N_HIGH_MEMORY]);
1938 h->next_nid_to_free = first_node(node_states[N_HIGH_MEMORY]);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001939 snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
1940 huge_page_size(h)/1024);
Aneesh Kumar K.Vabb82062012-07-31 16:42:24 -07001941 /*
1942 * Add cgroup control files only if the huge page consists
1943 * of more than two normal pages. This is because we use
1944 * page[2].lru.next for storing cgoup details.
1945 */
1946 if (order >= HUGETLB_CGROUP_MIN_ORDER)
1947 hugetlb_cgroup_file_init(hugetlb_max_hstate - 1);
Andi Kleen8faa8b02008-07-23 21:27:48 -07001948
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001949 parsed_hstate = h;
1950}
1951
Nick Piggine11bfbf2008-07-23 21:27:52 -07001952static int __init hugetlb_nrpages_setup(char *s)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001953{
1954 unsigned long *mhp;
Andi Kleen8faa8b02008-07-23 21:27:48 -07001955 static unsigned long *last_mhp;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001956
1957 /*
Aneesh Kumar K.V47d38342012-07-31 16:41:54 -07001958 * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet,
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001959 * so this hugepages= parameter goes to the "default hstate".
1960 */
Aneesh Kumar K.V47d38342012-07-31 16:41:54 -07001961 if (!hugetlb_max_hstate)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001962 mhp = &default_hstate_max_huge_pages;
1963 else
1964 mhp = &parsed_hstate->max_huge_pages;
1965
Andi Kleen8faa8b02008-07-23 21:27:48 -07001966 if (mhp == last_mhp) {
1967 printk(KERN_WARNING "hugepages= specified twice without "
1968 "interleaving hugepagesz=, ignoring\n");
1969 return 1;
1970 }
1971
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001972 if (sscanf(s, "%lu", mhp) <= 0)
1973 *mhp = 0;
1974
Andi Kleen8faa8b02008-07-23 21:27:48 -07001975 /*
1976 * Global state is always initialized later in hugetlb_init.
1977 * But we need to allocate >= MAX_ORDER hstates here early to still
1978 * use the bootmem allocator.
1979 */
Aneesh Kumar K.V47d38342012-07-31 16:41:54 -07001980 if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER)
Andi Kleen8faa8b02008-07-23 21:27:48 -07001981 hugetlb_hstate_alloc_pages(parsed_hstate);
1982
1983 last_mhp = mhp;
1984
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001985 return 1;
1986}
Nick Piggine11bfbf2008-07-23 21:27:52 -07001987__setup("hugepages=", hugetlb_nrpages_setup);
1988
1989static int __init hugetlb_default_setup(char *s)
1990{
1991 default_hstate_size = memparse(s, &s);
1992 return 1;
1993}
1994__setup("default_hugepagesz=", hugetlb_default_setup);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001995
Nishanth Aravamudan8a213462008-07-25 19:44:37 -07001996static unsigned int cpuset_mems_nr(unsigned int *array)
1997{
1998 int node;
1999 unsigned int nr = 0;
2000
2001 for_each_node_mask(node, cpuset_current_mems_allowed)
2002 nr += array[node];
2003
2004 return nr;
2005}
2006
2007#ifdef CONFIG_SYSCTL
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002008static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
2009 struct ctl_table *table, int write,
2010 void __user *buffer, size_t *length, loff_t *ppos)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002011{
Andi Kleene5ff2152008-07-23 21:27:42 -07002012 struct hstate *h = &default_hstate;
2013 unsigned long tmp;
Michal Hocko08d4a242011-01-13 15:47:26 -08002014 int ret;
Andi Kleene5ff2152008-07-23 21:27:42 -07002015
Petr Holasekc033a932011-03-22 16:33:05 -07002016 tmp = h->max_huge_pages;
Andi Kleene5ff2152008-07-23 21:27:42 -07002017
Eric B Munsonadbe8722011-01-13 15:47:27 -08002018 if (write && h->order >= MAX_ORDER)
2019 return -EINVAL;
2020
Andi Kleene5ff2152008-07-23 21:27:42 -07002021 table->data = &tmp;
2022 table->maxlen = sizeof(unsigned long);
Michal Hocko08d4a242011-01-13 15:47:26 -08002023 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2024 if (ret)
2025 goto out;
Andi Kleene5ff2152008-07-23 21:27:42 -07002026
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002027 if (write) {
David Rientjesbad44b52009-12-14 17:58:38 -08002028 NODEMASK_ALLOC(nodemask_t, nodes_allowed,
2029 GFP_KERNEL | __GFP_NORETRY);
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002030 if (!(obey_mempolicy &&
2031 init_nodemask_of_mempolicy(nodes_allowed))) {
2032 NODEMASK_FREE(nodes_allowed);
2033 nodes_allowed = &node_states[N_HIGH_MEMORY];
2034 }
2035 h->max_huge_pages = set_max_huge_pages(h, tmp, nodes_allowed);
2036
2037 if (nodes_allowed != &node_states[N_HIGH_MEMORY])
2038 NODEMASK_FREE(nodes_allowed);
2039 }
Michal Hocko08d4a242011-01-13 15:47:26 -08002040out:
2041 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002042}
Mel Gorman396faf02007-07-17 04:03:13 -07002043
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002044int hugetlb_sysctl_handler(struct ctl_table *table, int write,
2045 void __user *buffer, size_t *length, loff_t *ppos)
2046{
2047
2048 return hugetlb_sysctl_handler_common(false, table, write,
2049 buffer, length, ppos);
2050}
2051
2052#ifdef CONFIG_NUMA
2053int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
2054 void __user *buffer, size_t *length, loff_t *ppos)
2055{
2056 return hugetlb_sysctl_handler_common(true, table, write,
2057 buffer, length, ppos);
2058}
2059#endif /* CONFIG_NUMA */
2060
Mel Gorman396faf02007-07-17 04:03:13 -07002061int hugetlb_treat_movable_handler(struct ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07002062 void __user *buffer,
Mel Gorman396faf02007-07-17 04:03:13 -07002063 size_t *length, loff_t *ppos)
2064{
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07002065 proc_dointvec(table, write, buffer, length, ppos);
Mel Gorman396faf02007-07-17 04:03:13 -07002066 if (hugepages_treat_as_movable)
2067 htlb_alloc_mask = GFP_HIGHUSER_MOVABLE;
2068 else
2069 htlb_alloc_mask = GFP_HIGHUSER;
2070 return 0;
2071}
2072
Nishanth Aravamudana3d0c6a2008-02-08 04:18:18 -08002073int hugetlb_overcommit_handler(struct ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07002074 void __user *buffer,
Nishanth Aravamudana3d0c6a2008-02-08 04:18:18 -08002075 size_t *length, loff_t *ppos)
2076{
Andi Kleena5516432008-07-23 21:27:41 -07002077 struct hstate *h = &default_hstate;
Andi Kleene5ff2152008-07-23 21:27:42 -07002078 unsigned long tmp;
Michal Hocko08d4a242011-01-13 15:47:26 -08002079 int ret;
Andi Kleene5ff2152008-07-23 21:27:42 -07002080
Petr Holasekc033a932011-03-22 16:33:05 -07002081 tmp = h->nr_overcommit_huge_pages;
Andi Kleene5ff2152008-07-23 21:27:42 -07002082
Eric B Munsonadbe8722011-01-13 15:47:27 -08002083 if (write && h->order >= MAX_ORDER)
2084 return -EINVAL;
2085
Andi Kleene5ff2152008-07-23 21:27:42 -07002086 table->data = &tmp;
2087 table->maxlen = sizeof(unsigned long);
Michal Hocko08d4a242011-01-13 15:47:26 -08002088 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2089 if (ret)
2090 goto out;
Andi Kleene5ff2152008-07-23 21:27:42 -07002091
2092 if (write) {
2093 spin_lock(&hugetlb_lock);
2094 h->nr_overcommit_huge_pages = tmp;
2095 spin_unlock(&hugetlb_lock);
2096 }
Michal Hocko08d4a242011-01-13 15:47:26 -08002097out:
2098 return ret;
Nishanth Aravamudana3d0c6a2008-02-08 04:18:18 -08002099}
2100
Linus Torvalds1da177e2005-04-16 15:20:36 -07002101#endif /* CONFIG_SYSCTL */
2102
Alexey Dobriyane1759c22008-10-15 23:50:22 +04002103void hugetlb_report_meminfo(struct seq_file *m)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002104{
Andi Kleena5516432008-07-23 21:27:41 -07002105 struct hstate *h = &default_hstate;
Alexey Dobriyane1759c22008-10-15 23:50:22 +04002106 seq_printf(m,
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002107 "HugePages_Total: %5lu\n"
2108 "HugePages_Free: %5lu\n"
2109 "HugePages_Rsvd: %5lu\n"
2110 "HugePages_Surp: %5lu\n"
2111 "Hugepagesize: %8lu kB\n",
Andi Kleena5516432008-07-23 21:27:41 -07002112 h->nr_huge_pages,
2113 h->free_huge_pages,
2114 h->resv_huge_pages,
2115 h->surplus_huge_pages,
2116 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002117}
2118
2119int hugetlb_report_node_meminfo(int nid, char *buf)
2120{
Andi Kleena5516432008-07-23 21:27:41 -07002121 struct hstate *h = &default_hstate;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002122 return sprintf(buf,
2123 "Node %d HugePages_Total: %5u\n"
Nishanth Aravamudana1de0912008-03-26 14:37:53 -07002124 "Node %d HugePages_Free: %5u\n"
2125 "Node %d HugePages_Surp: %5u\n",
Andi Kleena5516432008-07-23 21:27:41 -07002126 nid, h->nr_huge_pages_node[nid],
2127 nid, h->free_huge_pages_node[nid],
2128 nid, h->surplus_huge_pages_node[nid]);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002129}
2130
Linus Torvalds1da177e2005-04-16 15:20:36 -07002131/* Return the number pages of memory we physically have, in PAGE_SIZE units. */
2132unsigned long hugetlb_total_pages(void)
2133{
Andi Kleena5516432008-07-23 21:27:41 -07002134 struct hstate *h = &default_hstate;
2135 return h->nr_huge_pages * pages_per_huge_page(h);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002136}
Linus Torvalds1da177e2005-04-16 15:20:36 -07002137
Andi Kleena5516432008-07-23 21:27:41 -07002138static int hugetlb_acct_memory(struct hstate *h, long delta)
Mel Gormanfc1b8a72008-07-23 21:27:22 -07002139{
2140 int ret = -ENOMEM;
2141
2142 spin_lock(&hugetlb_lock);
2143 /*
2144 * When cpuset is configured, it breaks the strict hugetlb page
2145 * reservation as the accounting is done on a global variable. Such
2146 * reservation is completely rubbish in the presence of cpuset because
2147 * the reservation is not checked against page availability for the
2148 * current cpuset. Application can still potentially OOM'ed by kernel
2149 * with lack of free htlb page in cpuset that the task is in.
2150 * Attempt to enforce strict accounting with cpuset is almost
2151 * impossible (or too ugly) because cpuset is too fluid that
2152 * task or memory node can be dynamically moved between cpusets.
2153 *
2154 * The change of semantics for shared hugetlb mapping with cpuset is
2155 * undesirable. However, in order to preserve some of the semantics,
2156 * we fall back to check against current free page availability as
2157 * a best attempt and hopefully to minimize the impact of changing
2158 * semantics that cpuset has.
2159 */
2160 if (delta > 0) {
Andi Kleena5516432008-07-23 21:27:41 -07002161 if (gather_surplus_pages(h, delta) < 0)
Mel Gormanfc1b8a72008-07-23 21:27:22 -07002162 goto out;
2163
Andi Kleena5516432008-07-23 21:27:41 -07002164 if (delta > cpuset_mems_nr(h->free_huge_pages_node)) {
2165 return_unused_surplus_pages(h, delta);
Mel Gormanfc1b8a72008-07-23 21:27:22 -07002166 goto out;
2167 }
2168 }
2169
2170 ret = 0;
2171 if (delta < 0)
Andi Kleena5516432008-07-23 21:27:41 -07002172 return_unused_surplus_pages(h, (unsigned long) -delta);
Mel Gormanfc1b8a72008-07-23 21:27:22 -07002173
2174out:
2175 spin_unlock(&hugetlb_lock);
2176 return ret;
2177}
2178
Andy Whitcroft84afd992008-07-23 21:27:32 -07002179static void hugetlb_vm_op_open(struct vm_area_struct *vma)
2180{
2181 struct resv_map *reservations = vma_resv_map(vma);
2182
2183 /*
2184 * This new VMA should share its siblings reservation map if present.
2185 * The VMA will only ever have a valid reservation map pointer where
2186 * it is being copied for another still existing VMA. As that VMA
Lucas De Marchi25985ed2011-03-30 22:57:33 -03002187 * has a reference to the reservation map it cannot disappear until
Andy Whitcroft84afd992008-07-23 21:27:32 -07002188 * after this open call completes. It is therefore safe to take a
2189 * new reference here without additional locking.
2190 */
2191 if (reservations)
2192 kref_get(&reservations->refs);
2193}
2194
Dave Hansenc50ac052012-05-29 15:06:46 -07002195static void resv_map_put(struct vm_area_struct *vma)
2196{
2197 struct resv_map *reservations = vma_resv_map(vma);
2198
2199 if (!reservations)
2200 return;
2201 kref_put(&reservations->refs, resv_map_release);
2202}
2203
Mel Gormana1e78772008-07-23 21:27:23 -07002204static void hugetlb_vm_op_close(struct vm_area_struct *vma)
2205{
Andi Kleena5516432008-07-23 21:27:41 -07002206 struct hstate *h = hstate_vma(vma);
Andy Whitcroft84afd992008-07-23 21:27:32 -07002207 struct resv_map *reservations = vma_resv_map(vma);
David Gibson90481622012-03-21 16:34:12 -07002208 struct hugepage_subpool *spool = subpool_vma(vma);
Andy Whitcroft84afd992008-07-23 21:27:32 -07002209 unsigned long reserve;
2210 unsigned long start;
2211 unsigned long end;
2212
2213 if (reservations) {
Andi Kleena5516432008-07-23 21:27:41 -07002214 start = vma_hugecache_offset(h, vma, vma->vm_start);
2215 end = vma_hugecache_offset(h, vma, vma->vm_end);
Andy Whitcroft84afd992008-07-23 21:27:32 -07002216
2217 reserve = (end - start) -
2218 region_count(&reservations->regions, start, end);
2219
Dave Hansenc50ac052012-05-29 15:06:46 -07002220 resv_map_put(vma);
Andy Whitcroft84afd992008-07-23 21:27:32 -07002221
Adam Litke7251ff72008-07-23 21:27:59 -07002222 if (reserve) {
Andi Kleena5516432008-07-23 21:27:41 -07002223 hugetlb_acct_memory(h, -reserve);
David Gibson90481622012-03-21 16:34:12 -07002224 hugepage_subpool_put_pages(spool, reserve);
Adam Litke7251ff72008-07-23 21:27:59 -07002225 }
Andy Whitcroft84afd992008-07-23 21:27:32 -07002226 }
Mel Gormana1e78772008-07-23 21:27:23 -07002227}
2228
Linus Torvalds1da177e2005-04-16 15:20:36 -07002229/*
2230 * We cannot handle pagefaults against hugetlb pages at all. They cause
2231 * handle_mm_fault() to try to instantiate regular-sized pages in the
2232 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
2233 * this far.
2234 */
Nick Piggind0217ac2007-07-19 01:47:03 -07002235static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002236{
2237 BUG();
Nick Piggind0217ac2007-07-19 01:47:03 -07002238 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002239}
2240
Alexey Dobriyanf0f37e22009-09-27 22:29:37 +04002241const struct vm_operations_struct hugetlb_vm_ops = {
Nick Piggind0217ac2007-07-19 01:47:03 -07002242 .fault = hugetlb_vm_op_fault,
Andy Whitcroft84afd992008-07-23 21:27:32 -07002243 .open = hugetlb_vm_op_open,
Mel Gormana1e78772008-07-23 21:27:23 -07002244 .close = hugetlb_vm_op_close,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002245};
2246
David Gibson1e8f8892006-01-06 00:10:44 -08002247static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
2248 int writable)
David Gibson63551ae2005-06-21 17:14:44 -07002249{
2250 pte_t entry;
2251
David Gibson1e8f8892006-01-06 00:10:44 -08002252 if (writable) {
David Gibson63551ae2005-06-21 17:14:44 -07002253 entry =
2254 pte_mkwrite(pte_mkdirty(mk_pte(page, vma->vm_page_prot)));
2255 } else {
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07002256 entry = huge_pte_wrprotect(mk_pte(page, vma->vm_page_prot));
David Gibson63551ae2005-06-21 17:14:44 -07002257 }
2258 entry = pte_mkyoung(entry);
2259 entry = pte_mkhuge(entry);
Chris Metcalfd9ed9fa2012-04-01 14:01:34 -04002260 entry = arch_make_huge_pte(entry, vma, page, writable);
David Gibson63551ae2005-06-21 17:14:44 -07002261
2262 return entry;
2263}
2264
David Gibson1e8f8892006-01-06 00:10:44 -08002265static void set_huge_ptep_writable(struct vm_area_struct *vma,
2266 unsigned long address, pte_t *ptep)
2267{
2268 pte_t entry;
2269
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07002270 entry = pte_mkwrite(pte_mkdirty(huge_ptep_get(ptep)));
Chris Forbes32f84522011-07-25 17:12:14 -07002271 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
Russell King4b3073e2009-12-18 16:40:18 +00002272 update_mmu_cache(vma, address, ptep);
David Gibson1e8f8892006-01-06 00:10:44 -08002273}
2274
2275
David Gibson63551ae2005-06-21 17:14:44 -07002276int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
2277 struct vm_area_struct *vma)
2278{
2279 pte_t *src_pte, *dst_pte, entry;
2280 struct page *ptepage;
Hugh Dickins1c598272005-10-19 21:23:43 -07002281 unsigned long addr;
David Gibson1e8f8892006-01-06 00:10:44 -08002282 int cow;
Andi Kleena5516432008-07-23 21:27:41 -07002283 struct hstate *h = hstate_vma(vma);
2284 unsigned long sz = huge_page_size(h);
David Gibson1e8f8892006-01-06 00:10:44 -08002285
2286 cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
David Gibson63551ae2005-06-21 17:14:44 -07002287
Andi Kleena5516432008-07-23 21:27:41 -07002288 for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
Hugh Dickinsc74df322005-10-29 18:16:23 -07002289 src_pte = huge_pte_offset(src, addr);
2290 if (!src_pte)
2291 continue;
Andi Kleena5516432008-07-23 21:27:41 -07002292 dst_pte = huge_pte_alloc(dst, addr, sz);
David Gibson63551ae2005-06-21 17:14:44 -07002293 if (!dst_pte)
2294 goto nomem;
Larry Woodmanc5c99422008-01-24 05:49:25 -08002295
2296 /* If the pagetables are shared don't copy or take references */
2297 if (dst_pte == src_pte)
2298 continue;
2299
Hugh Dickinsc74df322005-10-29 18:16:23 -07002300 spin_lock(&dst->page_table_lock);
Nick Piggin46478752008-06-05 22:45:57 -07002301 spin_lock_nested(&src->page_table_lock, SINGLE_DEPTH_NESTING);
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07002302 if (!huge_pte_none(huge_ptep_get(src_pte))) {
David Gibson1e8f8892006-01-06 00:10:44 -08002303 if (cow)
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07002304 huge_ptep_set_wrprotect(src, addr, src_pte);
2305 entry = huge_ptep_get(src_pte);
Hugh Dickins1c598272005-10-19 21:23:43 -07002306 ptepage = pte_page(entry);
2307 get_page(ptepage);
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002308 page_dup_rmap(ptepage);
Hugh Dickins1c598272005-10-19 21:23:43 -07002309 set_huge_pte_at(dst, addr, dst_pte, entry);
2310 }
2311 spin_unlock(&src->page_table_lock);
Hugh Dickinsc74df322005-10-29 18:16:23 -07002312 spin_unlock(&dst->page_table_lock);
David Gibson63551ae2005-06-21 17:14:44 -07002313 }
2314 return 0;
2315
2316nomem:
2317 return -ENOMEM;
2318}
2319
Naoya Horiguchi290408d2010-09-08 10:19:35 +09002320static int is_hugetlb_entry_migration(pte_t pte)
2321{
2322 swp_entry_t swp;
2323
2324 if (huge_pte_none(pte) || pte_present(pte))
2325 return 0;
2326 swp = pte_to_swp_entry(pte);
Chris Forbes32f84522011-07-25 17:12:14 -07002327 if (non_swap_entry(swp) && is_migration_entry(swp))
Naoya Horiguchi290408d2010-09-08 10:19:35 +09002328 return 1;
Chris Forbes32f84522011-07-25 17:12:14 -07002329 else
Naoya Horiguchi290408d2010-09-08 10:19:35 +09002330 return 0;
2331}
2332
Naoya Horiguchifd6a03ed2010-05-28 09:29:21 +09002333static int is_hugetlb_entry_hwpoisoned(pte_t pte)
2334{
2335 swp_entry_t swp;
2336
2337 if (huge_pte_none(pte) || pte_present(pte))
2338 return 0;
2339 swp = pte_to_swp_entry(pte);
Chris Forbes32f84522011-07-25 17:12:14 -07002340 if (non_swap_entry(swp) && is_hwpoison_entry(swp))
Naoya Horiguchifd6a03ed2010-05-28 09:29:21 +09002341 return 1;
Chris Forbes32f84522011-07-25 17:12:14 -07002342 else
Naoya Horiguchifd6a03ed2010-05-28 09:29:21 +09002343 return 0;
2344}
2345
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07002346void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
2347 unsigned long start, unsigned long end,
2348 struct page *ref_page)
David Gibson63551ae2005-06-21 17:14:44 -07002349{
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07002350 int force_flush = 0;
David Gibson63551ae2005-06-21 17:14:44 -07002351 struct mm_struct *mm = vma->vm_mm;
2352 unsigned long address;
David Gibsonc7546f82005-08-05 11:59:35 -07002353 pte_t *ptep;
David Gibson63551ae2005-06-21 17:14:44 -07002354 pte_t pte;
2355 struct page *page;
Andi Kleena5516432008-07-23 21:27:41 -07002356 struct hstate *h = hstate_vma(vma);
2357 unsigned long sz = huge_page_size(h);
2358
David Gibson63551ae2005-06-21 17:14:44 -07002359 WARN_ON(!is_vm_hugetlb_page(vma));
Andi Kleena5516432008-07-23 21:27:41 -07002360 BUG_ON(start & ~huge_page_mask(h));
2361 BUG_ON(end & ~huge_page_mask(h));
David Gibson63551ae2005-06-21 17:14:44 -07002362
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07002363 tlb_start_vma(tlb, vma);
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07002364 mmu_notifier_invalidate_range_start(mm, start, end);
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07002365again:
Hugh Dickins508034a2005-10-29 18:16:30 -07002366 spin_lock(&mm->page_table_lock);
Andi Kleena5516432008-07-23 21:27:41 -07002367 for (address = start; address < end; address += sz) {
David Gibsonc7546f82005-08-05 11:59:35 -07002368 ptep = huge_pte_offset(mm, address);
Adam Litke4c887262005-10-29 18:16:46 -07002369 if (!ptep)
David Gibsonc7546f82005-08-05 11:59:35 -07002370 continue;
2371
Chen, Kenneth W39dde652006-12-06 20:32:03 -08002372 if (huge_pmd_unshare(mm, &address, ptep))
2373 continue;
2374
Hillf Danton66293262012-03-23 15:01:48 -07002375 pte = huge_ptep_get(ptep);
2376 if (huge_pte_none(pte))
2377 continue;
2378
2379 /*
2380 * HWPoisoned hugepage is already unmapped and dropped reference
2381 */
2382 if (unlikely(is_hugetlb_entry_hwpoisoned(pte)))
2383 continue;
2384
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 }
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07002428 mmu_notifier_invalidate_range_end(mm, start, end);
2429 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;
2476 struct prio_tree_iter iter;
2477 pgoff_t pgoff;
2478
2479 /*
2480 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
2481 * from page cache lookup which is in HPAGE_SIZE units.
2482 */
Adam Litke75266742008-11-12 13:24:56 -08002483 address = address & huge_page_mask(h);
Hillf Danton0c176d52012-01-10 15:08:19 -08002484 pgoff = vma_hugecache_offset(h, vma, address);
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);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002493 vma_prio_tree_foreach(iter_vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
2494 /* 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;
David Gibson1e8f8892006-01-06 00:10:44 -08002528
2529 old_page = pte_page(pte);
2530
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002531retry_avoidcopy:
David Gibson1e8f8892006-01-06 00:10:44 -08002532 /* If no-one else is actually using this page, avoid the copy
2533 * and just make the page writable */
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002534 avoidcopy = (page_mapcount(old_page) == 1);
David Gibson1e8f8892006-01-06 00:10:44 -08002535 if (avoidcopy) {
Naoya Horiguchi56c9cfb2010-09-10 13:23:04 +09002536 if (PageAnon(old_page))
2537 page_move_anon_rmap(old_page, vma, address);
David Gibson1e8f8892006-01-06 00:10:44 -08002538 set_huge_ptep_writable(vma, address, ptep);
Nick Piggin83c54072007-07-19 01:47:05 -07002539 return 0;
David Gibson1e8f8892006-01-06 00:10:44 -08002540 }
2541
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002542 /*
2543 * If the process that created a MAP_PRIVATE mapping is about to
2544 * perform a COW due to a shared page count, attempt to satisfy
2545 * the allocation without using the existing reserves. The pagecache
2546 * page is used to determine if the reserve at this address was
2547 * consumed or not. If reserves were used, a partial faulted mapping
2548 * at the time of fork() could consume its reserves on COW instead
2549 * of the full address range.
2550 */
Mel Gormanf83a2752009-05-28 14:34:40 -07002551 if (!(vma->vm_flags & VM_MAYSHARE) &&
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002552 is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
2553 old_page != pagecache_page)
2554 outside_reserve = 1;
2555
David Gibson1e8f8892006-01-06 00:10:44 -08002556 page_cache_get(old_page);
Larry Woodmanb76c8cf2009-12-14 17:59:37 -08002557
2558 /* Drop page_table_lock as buddy allocator may be called */
2559 spin_unlock(&mm->page_table_lock);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002560 new_page = alloc_huge_page(vma, address, outside_reserve);
David Gibson1e8f8892006-01-06 00:10:44 -08002561
Adam Litke2fc39ce2007-11-14 16:59:39 -08002562 if (IS_ERR(new_page)) {
Aneesh Kumar K.V76dcee72012-07-31 16:41:57 -07002563 long err = PTR_ERR(new_page);
David Gibson1e8f8892006-01-06 00:10:44 -08002564 page_cache_release(old_page);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002565
2566 /*
2567 * If a process owning a MAP_PRIVATE mapping fails to COW,
2568 * it is due to references held by a child and an insufficient
2569 * huge page pool. To guarantee the original mappers
2570 * reliability, unmap the page from child processes. The child
2571 * may get SIGKILLed if it later faults.
2572 */
2573 if (outside_reserve) {
2574 BUG_ON(huge_pte_none(pte));
2575 if (unmap_ref_private(mm, vma, old_page, address)) {
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002576 BUG_ON(huge_pte_none(pte));
Larry Woodmanb76c8cf2009-12-14 17:59:37 -08002577 spin_lock(&mm->page_table_lock);
Hillf Dantona734bcc2012-01-10 15:07:20 -08002578 ptep = huge_pte_offset(mm, address & huge_page_mask(h));
2579 if (likely(pte_same(huge_ptep_get(ptep), pte)))
2580 goto retry_avoidcopy;
2581 /*
2582 * race occurs while re-acquiring page_table_lock, and
2583 * our job is done.
2584 */
2585 return 0;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002586 }
2587 WARN_ON_ONCE(1);
2588 }
2589
Larry Woodmanb76c8cf2009-12-14 17:59:37 -08002590 /* Caller expects lock to be held */
2591 spin_lock(&mm->page_table_lock);
Aneesh Kumar K.V76dcee72012-07-31 16:41:57 -07002592 if (err == -ENOMEM)
2593 return VM_FAULT_OOM;
2594 else
2595 return VM_FAULT_SIGBUS;
David Gibson1e8f8892006-01-06 00:10:44 -08002596 }
2597
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002598 /*
2599 * When the original hugepage is shared one, it does not have
2600 * anon_vma prepared.
2601 */
Dean Nelson44e2aa92010-10-26 14:22:08 -07002602 if (unlikely(anon_vma_prepare(vma))) {
Hillf Dantonea4039a2011-11-15 14:36:12 -08002603 page_cache_release(new_page);
2604 page_cache_release(old_page);
Dean Nelson44e2aa92010-10-26 14:22:08 -07002605 /* Caller expects lock to be held */
2606 spin_lock(&mm->page_table_lock);
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002607 return VM_FAULT_OOM;
Dean Nelson44e2aa92010-10-26 14:22:08 -07002608 }
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002609
Andrea Arcangeli47ad8472011-01-13 15:46:47 -08002610 copy_user_huge_page(new_page, old_page, address, vma,
2611 pages_per_huge_page(h));
Nick Piggin0ed361d2008-02-04 22:29:34 -08002612 __SetPageUptodate(new_page);
David Gibson1e8f8892006-01-06 00:10:44 -08002613
Larry Woodmanb76c8cf2009-12-14 17:59:37 -08002614 /*
2615 * Retake the page_table_lock to check for racing updates
2616 * before the page tables are altered
2617 */
2618 spin_lock(&mm->page_table_lock);
Andi Kleena5516432008-07-23 21:27:41 -07002619 ptep = huge_pte_offset(mm, address & huge_page_mask(h));
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07002620 if (likely(pte_same(huge_ptep_get(ptep), pte))) {
David Gibson1e8f8892006-01-06 00:10:44 -08002621 /* Break COW */
Doug Doan3edd4fc2010-08-09 17:18:30 -07002622 mmu_notifier_invalidate_range_start(mm,
2623 address & huge_page_mask(h),
2624 (address & huge_page_mask(h)) + huge_page_size(h));
Gerald Schaefer8fe627e2008-04-28 02:13:28 -07002625 huge_ptep_clear_flush(vma, address, ptep);
David Gibson1e8f8892006-01-06 00:10:44 -08002626 set_huge_pte_at(mm, address, ptep,
2627 make_huge_pte(vma, new_page, 1));
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002628 page_remove_rmap(old_page);
Naoya Horiguchicd67f0d2010-09-10 13:23:04 +09002629 hugepage_add_new_anon_rmap(new_page, vma, address);
David Gibson1e8f8892006-01-06 00:10:44 -08002630 /* Make the old page be freed below */
2631 new_page = old_page;
Doug Doan3edd4fc2010-08-09 17:18:30 -07002632 mmu_notifier_invalidate_range_end(mm,
2633 address & huge_page_mask(h),
2634 (address & huge_page_mask(h)) + huge_page_size(h));
David Gibson1e8f8892006-01-06 00:10:44 -08002635 }
2636 page_cache_release(new_page);
2637 page_cache_release(old_page);
Nick Piggin83c54072007-07-19 01:47:05 -07002638 return 0;
David Gibson1e8f8892006-01-06 00:10:44 -08002639}
2640
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002641/* Return the pagecache page at a given address within a VMA */
Andi Kleena5516432008-07-23 21:27:41 -07002642static struct page *hugetlbfs_pagecache_page(struct hstate *h,
2643 struct vm_area_struct *vma, unsigned long address)
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002644{
2645 struct address_space *mapping;
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -07002646 pgoff_t idx;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002647
2648 mapping = vma->vm_file->f_mapping;
Andi Kleena5516432008-07-23 21:27:41 -07002649 idx = vma_hugecache_offset(h, vma, address);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002650
2651 return find_lock_page(mapping, idx);
2652}
2653
Hugh Dickins3ae77f42009-09-21 17:03:33 -07002654/*
2655 * Return whether there is a pagecache page to back given address within VMA.
2656 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
2657 */
2658static bool hugetlbfs_pagecache_present(struct hstate *h,
Hugh Dickins2a15efc2009-09-21 17:03:27 -07002659 struct vm_area_struct *vma, unsigned long address)
2660{
2661 struct address_space *mapping;
2662 pgoff_t idx;
2663 struct page *page;
2664
2665 mapping = vma->vm_file->f_mapping;
2666 idx = vma_hugecache_offset(h, vma, address);
2667
2668 page = find_get_page(mapping, idx);
2669 if (page)
2670 put_page(page);
2671 return page != NULL;
2672}
2673
Robert P. J. Daya1ed3dd2007-07-17 04:03:33 -07002674static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
Hugh Dickins788c7df2009-06-23 13:49:05 +01002675 unsigned long address, pte_t *ptep, unsigned int flags)
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01002676{
Andi Kleena5516432008-07-23 21:27:41 -07002677 struct hstate *h = hstate_vma(vma);
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01002678 int ret = VM_FAULT_SIGBUS;
Hillf Danton409eb8c2012-01-20 14:34:13 -08002679 int anon_rmap = 0;
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -07002680 pgoff_t idx;
Adam Litke4c887262005-10-29 18:16:46 -07002681 unsigned long size;
Adam Litke4c887262005-10-29 18:16:46 -07002682 struct page *page;
2683 struct address_space *mapping;
David Gibson1e8f8892006-01-06 00:10:44 -08002684 pte_t new_pte;
Adam Litke4c887262005-10-29 18:16:46 -07002685
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002686 /*
2687 * Currently, we are forced to kill the process in the event the
2688 * original mapper has unmapped pages from the child due to a failed
Lucas De Marchi25985ed2011-03-30 22:57:33 -03002689 * COW. Warn that such a situation has occurred as it may not be obvious
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002690 */
2691 if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
2692 printk(KERN_WARNING
2693 "PID %d killed due to inadequate hugepage pool\n",
2694 current->pid);
2695 return ret;
2696 }
2697
Adam Litke4c887262005-10-29 18:16:46 -07002698 mapping = vma->vm_file->f_mapping;
Andi Kleena5516432008-07-23 21:27:41 -07002699 idx = vma_hugecache_offset(h, vma, address);
Adam Litke4c887262005-10-29 18:16:46 -07002700
2701 /*
2702 * Use page lock to guard against racing truncation
2703 * before we get page_table_lock.
2704 */
Christoph Lameter6bda6662006-01-06 00:10:49 -08002705retry:
2706 page = find_lock_page(mapping, idx);
2707 if (!page) {
Andi Kleena5516432008-07-23 21:27:41 -07002708 size = i_size_read(mapping->host) >> huge_page_shift(h);
Hugh Dickinsebed4bf2006-10-28 10:38:43 -07002709 if (idx >= size)
2710 goto out;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002711 page = alloc_huge_page(vma, address, 0);
Adam Litke2fc39ce2007-11-14 16:59:39 -08002712 if (IS_ERR(page)) {
Aneesh Kumar K.V76dcee72012-07-31 16:41:57 -07002713 ret = PTR_ERR(page);
2714 if (ret == -ENOMEM)
2715 ret = VM_FAULT_OOM;
2716 else
2717 ret = VM_FAULT_SIGBUS;
Christoph Lameter6bda6662006-01-06 00:10:49 -08002718 goto out;
2719 }
Andrea Arcangeli47ad8472011-01-13 15:46:47 -08002720 clear_huge_page(page, address, pages_per_huge_page(h));
Nick Piggin0ed361d2008-02-04 22:29:34 -08002721 __SetPageUptodate(page);
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01002722
Mel Gormanf83a2752009-05-28 14:34:40 -07002723 if (vma->vm_flags & VM_MAYSHARE) {
Christoph Lameter6bda6662006-01-06 00:10:49 -08002724 int err;
Ken Chen45c682a2007-11-14 16:59:44 -08002725 struct inode *inode = mapping->host;
Christoph Lameter6bda6662006-01-06 00:10:49 -08002726
2727 err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
2728 if (err) {
2729 put_page(page);
Christoph Lameter6bda6662006-01-06 00:10:49 -08002730 if (err == -EEXIST)
2731 goto retry;
2732 goto out;
2733 }
Ken Chen45c682a2007-11-14 16:59:44 -08002734
2735 spin_lock(&inode->i_lock);
Andi Kleena5516432008-07-23 21:27:41 -07002736 inode->i_blocks += blocks_per_huge_page(h);
Ken Chen45c682a2007-11-14 16:59:44 -08002737 spin_unlock(&inode->i_lock);
Mel Gorman23be7462010-04-23 13:17:56 -04002738 } else {
Christoph Lameter6bda6662006-01-06 00:10:49 -08002739 lock_page(page);
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002740 if (unlikely(anon_vma_prepare(vma))) {
2741 ret = VM_FAULT_OOM;
2742 goto backout_unlocked;
2743 }
Hillf Danton409eb8c2012-01-20 14:34:13 -08002744 anon_rmap = 1;
Mel Gorman23be7462010-04-23 13:17:56 -04002745 }
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002746 } else {
Naoya Horiguchi998b4382010-09-08 10:19:32 +09002747 /*
2748 * If memory error occurs between mmap() and fault, some process
2749 * don't have hwpoisoned swap entry for errored virtual address.
2750 * So we need to block hugepage fault by PG_hwpoison bit check.
2751 */
2752 if (unlikely(PageHWPoison(page))) {
Chris Forbes32f84522011-07-25 17:12:14 -07002753 ret = VM_FAULT_HWPOISON |
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07002754 VM_FAULT_SET_HINDEX(hstate_index(h));
Naoya Horiguchi998b4382010-09-08 10:19:32 +09002755 goto backout_unlocked;
2756 }
Christoph Lameter6bda6662006-01-06 00:10:49 -08002757 }
David Gibson1e8f8892006-01-06 00:10:44 -08002758
Andy Whitcroft57303d82008-08-12 15:08:47 -07002759 /*
2760 * If we are going to COW a private mapping later, we examine the
2761 * pending reservations for this page now. This will ensure that
2762 * any allocations necessary to record that reservation occur outside
2763 * the spinlock.
2764 */
Hugh Dickins788c7df2009-06-23 13:49:05 +01002765 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED))
Andy Whitcroft2b267362008-08-12 15:08:49 -07002766 if (vma_needs_reservation(h, vma, address) < 0) {
2767 ret = VM_FAULT_OOM;
2768 goto backout_unlocked;
2769 }
Andy Whitcroft57303d82008-08-12 15:08:47 -07002770
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01002771 spin_lock(&mm->page_table_lock);
Andi Kleena5516432008-07-23 21:27:41 -07002772 size = i_size_read(mapping->host) >> huge_page_shift(h);
Adam Litke4c887262005-10-29 18:16:46 -07002773 if (idx >= size)
2774 goto backout;
2775
Nick Piggin83c54072007-07-19 01:47:05 -07002776 ret = 0;
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07002777 if (!huge_pte_none(huge_ptep_get(ptep)))
Adam Litke4c887262005-10-29 18:16:46 -07002778 goto backout;
2779
Hillf Danton409eb8c2012-01-20 14:34:13 -08002780 if (anon_rmap)
2781 hugepage_add_new_anon_rmap(page, vma, address);
2782 else
2783 page_dup_rmap(page);
David Gibson1e8f8892006-01-06 00:10:44 -08002784 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
2785 && (vma->vm_flags & VM_SHARED)));
2786 set_huge_pte_at(mm, address, ptep, new_pte);
2787
Hugh Dickins788c7df2009-06-23 13:49:05 +01002788 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
David Gibson1e8f8892006-01-06 00:10:44 -08002789 /* Optimization, do the COW without a second fault */
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002790 ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page);
David Gibson1e8f8892006-01-06 00:10:44 -08002791 }
2792
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01002793 spin_unlock(&mm->page_table_lock);
Adam Litke4c887262005-10-29 18:16:46 -07002794 unlock_page(page);
2795out:
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01002796 return ret;
Adam Litke4c887262005-10-29 18:16:46 -07002797
2798backout:
2799 spin_unlock(&mm->page_table_lock);
Andy Whitcroft2b267362008-08-12 15:08:49 -07002800backout_unlocked:
Adam Litke4c887262005-10-29 18:16:46 -07002801 unlock_page(page);
2802 put_page(page);
2803 goto out;
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01002804}
2805
Adam Litke86e52162006-01-06 00:10:43 -08002806int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
Hugh Dickins788c7df2009-06-23 13:49:05 +01002807 unsigned long address, unsigned int flags)
Adam Litke86e52162006-01-06 00:10:43 -08002808{
2809 pte_t *ptep;
2810 pte_t entry;
David Gibson1e8f8892006-01-06 00:10:44 -08002811 int ret;
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002812 struct page *page = NULL;
Andy Whitcroft57303d82008-08-12 15:08:47 -07002813 struct page *pagecache_page = NULL;
David Gibson3935baa2006-03-22 00:08:53 -08002814 static DEFINE_MUTEX(hugetlb_instantiation_mutex);
Andi Kleena5516432008-07-23 21:27:41 -07002815 struct hstate *h = hstate_vma(vma);
Adam Litke86e52162006-01-06 00:10:43 -08002816
KAMEZAWA Hiroyuki1e16a532012-01-10 15:07:22 -08002817 address &= huge_page_mask(h);
2818
Naoya Horiguchifd6a03ed2010-05-28 09:29:21 +09002819 ptep = huge_pte_offset(mm, address);
2820 if (ptep) {
2821 entry = huge_ptep_get(ptep);
Naoya Horiguchi290408d2010-09-08 10:19:35 +09002822 if (unlikely(is_hugetlb_entry_migration(entry))) {
2823 migration_entry_wait(mm, (pmd_t *)ptep, address);
2824 return 0;
2825 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
Chris Forbes32f84522011-07-25 17:12:14 -07002826 return VM_FAULT_HWPOISON_LARGE |
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07002827 VM_FAULT_SET_HINDEX(hstate_index(h));
Naoya Horiguchifd6a03ed2010-05-28 09:29:21 +09002828 }
2829
Andi Kleena5516432008-07-23 21:27:41 -07002830 ptep = huge_pte_alloc(mm, address, huge_page_size(h));
Adam Litke86e52162006-01-06 00:10:43 -08002831 if (!ptep)
2832 return VM_FAULT_OOM;
2833
David Gibson3935baa2006-03-22 00:08:53 -08002834 /*
2835 * Serialize hugepage allocation and instantiation, so that we don't
2836 * get spurious allocation failures if two CPUs race to instantiate
2837 * the same page in the page cache.
2838 */
2839 mutex_lock(&hugetlb_instantiation_mutex);
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07002840 entry = huge_ptep_get(ptep);
2841 if (huge_pte_none(entry)) {
Hugh Dickins788c7df2009-06-23 13:49:05 +01002842 ret = hugetlb_no_page(mm, vma, address, ptep, flags);
David Gibsonb4d1d992008-10-15 22:01:11 -07002843 goto out_mutex;
David Gibson3935baa2006-03-22 00:08:53 -08002844 }
Adam Litke86e52162006-01-06 00:10:43 -08002845
Nick Piggin83c54072007-07-19 01:47:05 -07002846 ret = 0;
David Gibson1e8f8892006-01-06 00:10:44 -08002847
Andy Whitcroft57303d82008-08-12 15:08:47 -07002848 /*
2849 * If we are going to COW the mapping later, we examine the pending
2850 * reservations for this page now. This will ensure that any
2851 * allocations necessary to record that reservation occur outside the
2852 * spinlock. For private mappings, we also lookup the pagecache
2853 * page now as it is used to determine if a reservation has been
2854 * consumed.
2855 */
Hugh Dickins788c7df2009-06-23 13:49:05 +01002856 if ((flags & FAULT_FLAG_WRITE) && !pte_write(entry)) {
Andy Whitcroft2b267362008-08-12 15:08:49 -07002857 if (vma_needs_reservation(h, vma, address) < 0) {
2858 ret = VM_FAULT_OOM;
David Gibsonb4d1d992008-10-15 22:01:11 -07002859 goto out_mutex;
Andy Whitcroft2b267362008-08-12 15:08:49 -07002860 }
Andy Whitcroft57303d82008-08-12 15:08:47 -07002861
Mel Gormanf83a2752009-05-28 14:34:40 -07002862 if (!(vma->vm_flags & VM_MAYSHARE))
Andy Whitcroft57303d82008-08-12 15:08:47 -07002863 pagecache_page = hugetlbfs_pagecache_page(h,
2864 vma, address);
2865 }
2866
Naoya Horiguchi56c9cfb2010-09-10 13:23:04 +09002867 /*
2868 * hugetlb_cow() requires page locks of pte_page(entry) and
2869 * pagecache_page, so here we need take the former one
2870 * when page != pagecache_page or !pagecache_page.
2871 * Note that locking order is always pagecache_page -> page,
2872 * so no worry about deadlock.
2873 */
2874 page = pte_page(entry);
Chris Metcalf66aebce2012-04-12 12:49:15 -07002875 get_page(page);
Naoya Horiguchi56c9cfb2010-09-10 13:23:04 +09002876 if (page != pagecache_page)
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002877 lock_page(page);
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002878
David Gibson1e8f8892006-01-06 00:10:44 -08002879 spin_lock(&mm->page_table_lock);
2880 /* Check for a racing update before calling hugetlb_cow */
David Gibsonb4d1d992008-10-15 22:01:11 -07002881 if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
2882 goto out_page_table_lock;
2883
2884
Hugh Dickins788c7df2009-06-23 13:49:05 +01002885 if (flags & FAULT_FLAG_WRITE) {
David Gibsonb4d1d992008-10-15 22:01:11 -07002886 if (!pte_write(entry)) {
Andy Whitcroft57303d82008-08-12 15:08:47 -07002887 ret = hugetlb_cow(mm, vma, address, ptep, entry,
2888 pagecache_page);
David Gibsonb4d1d992008-10-15 22:01:11 -07002889 goto out_page_table_lock;
2890 }
2891 entry = pte_mkdirty(entry);
2892 }
2893 entry = pte_mkyoung(entry);
Hugh Dickins788c7df2009-06-23 13:49:05 +01002894 if (huge_ptep_set_access_flags(vma, address, ptep, entry,
2895 flags & FAULT_FLAG_WRITE))
Russell King4b3073e2009-12-18 16:40:18 +00002896 update_mmu_cache(vma, address, ptep);
David Gibsonb4d1d992008-10-15 22:01:11 -07002897
2898out_page_table_lock:
David Gibson1e8f8892006-01-06 00:10:44 -08002899 spin_unlock(&mm->page_table_lock);
Andy Whitcroft57303d82008-08-12 15:08:47 -07002900
2901 if (pagecache_page) {
2902 unlock_page(pagecache_page);
2903 put_page(pagecache_page);
2904 }
Dean Nelson1f64d692010-12-02 14:31:12 -08002905 if (page != pagecache_page)
2906 unlock_page(page);
Chris Metcalf66aebce2012-04-12 12:49:15 -07002907 put_page(page);
Andy Whitcroft57303d82008-08-12 15:08:47 -07002908
David Gibsonb4d1d992008-10-15 22:01:11 -07002909out_mutex:
David Gibson3935baa2006-03-22 00:08:53 -08002910 mutex_unlock(&hugetlb_instantiation_mutex);
David Gibson1e8f8892006-01-06 00:10:44 -08002911
2912 return ret;
Adam Litke86e52162006-01-06 00:10:43 -08002913}
2914
Andi Kleenceb86872008-07-23 21:27:50 -07002915/* Can be overriden by architectures */
2916__attribute__((weak)) struct page *
2917follow_huge_pud(struct mm_struct *mm, unsigned long address,
2918 pud_t *pud, int write)
2919{
2920 BUG();
2921 return NULL;
2922}
2923
David Gibson63551ae2005-06-21 17:14:44 -07002924int follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
2925 struct page **pages, struct vm_area_struct **vmas,
Adam Litke5b23dbe2007-11-14 16:59:33 -08002926 unsigned long *position, int *length, int i,
Hugh Dickins2a15efc2009-09-21 17:03:27 -07002927 unsigned int flags)
David Gibson63551ae2005-06-21 17:14:44 -07002928{
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08002929 unsigned long pfn_offset;
2930 unsigned long vaddr = *position;
David Gibson63551ae2005-06-21 17:14:44 -07002931 int remainder = *length;
Andi Kleena5516432008-07-23 21:27:41 -07002932 struct hstate *h = hstate_vma(vma);
David Gibson63551ae2005-06-21 17:14:44 -07002933
Hugh Dickins1c598272005-10-19 21:23:43 -07002934 spin_lock(&mm->page_table_lock);
David Gibson63551ae2005-06-21 17:14:44 -07002935 while (vaddr < vma->vm_end && remainder) {
Adam Litke4c887262005-10-29 18:16:46 -07002936 pte_t *pte;
Hugh Dickins2a15efc2009-09-21 17:03:27 -07002937 int absent;
Adam Litke4c887262005-10-29 18:16:46 -07002938 struct page *page;
2939
2940 /*
2941 * Some archs (sparc64, sh*) have multiple pte_ts to
Hugh Dickins2a15efc2009-09-21 17:03:27 -07002942 * each hugepage. We have to make sure we get the
Adam Litke4c887262005-10-29 18:16:46 -07002943 * first, for the page indexing below to work.
2944 */
Andi Kleena5516432008-07-23 21:27:41 -07002945 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h));
Hugh Dickins2a15efc2009-09-21 17:03:27 -07002946 absent = !pte || huge_pte_none(huge_ptep_get(pte));
Adam Litke4c887262005-10-29 18:16:46 -07002947
Hugh Dickins2a15efc2009-09-21 17:03:27 -07002948 /*
2949 * When coredumping, it suits get_dump_page if we just return
Hugh Dickins3ae77f42009-09-21 17:03:33 -07002950 * an error where there's an empty slot with no huge pagecache
2951 * to back it. This way, we avoid allocating a hugepage, and
2952 * the sparse dumpfile avoids allocating disk blocks, but its
2953 * huge holes still show up with zeroes where they need to be.
Hugh Dickins2a15efc2009-09-21 17:03:27 -07002954 */
Hugh Dickins3ae77f42009-09-21 17:03:33 -07002955 if (absent && (flags & FOLL_DUMP) &&
2956 !hugetlbfs_pagecache_present(h, vma, vaddr)) {
Hugh Dickins2a15efc2009-09-21 17:03:27 -07002957 remainder = 0;
2958 break;
2959 }
2960
2961 if (absent ||
2962 ((flags & FOLL_WRITE) && !pte_write(huge_ptep_get(pte)))) {
Adam Litke4c887262005-10-29 18:16:46 -07002963 int ret;
2964
2965 spin_unlock(&mm->page_table_lock);
Hugh Dickins2a15efc2009-09-21 17:03:27 -07002966 ret = hugetlb_fault(mm, vma, vaddr,
2967 (flags & FOLL_WRITE) ? FAULT_FLAG_WRITE : 0);
Adam Litke4c887262005-10-29 18:16:46 -07002968 spin_lock(&mm->page_table_lock);
Adam Litkea89182c2007-08-22 14:01:51 -07002969 if (!(ret & VM_FAULT_ERROR))
Adam Litke4c887262005-10-29 18:16:46 -07002970 continue;
2971
2972 remainder = 0;
Adam Litke4c887262005-10-29 18:16:46 -07002973 break;
2974 }
David Gibson63551ae2005-06-21 17:14:44 -07002975
Andi Kleena5516432008-07-23 21:27:41 -07002976 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07002977 page = pte_page(huge_ptep_get(pte));
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08002978same_page:
Chen, Kenneth Wd6692182006-03-31 02:29:57 -08002979 if (pages) {
Hugh Dickins2a15efc2009-09-21 17:03:27 -07002980 pages[i] = mem_map_offset(page, pfn_offset);
KOSAKI Motohiro4b2e38a2008-10-18 20:27:10 -07002981 get_page(pages[i]);
Chen, Kenneth Wd6692182006-03-31 02:29:57 -08002982 }
David Gibson63551ae2005-06-21 17:14:44 -07002983
2984 if (vmas)
2985 vmas[i] = vma;
2986
2987 vaddr += PAGE_SIZE;
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08002988 ++pfn_offset;
David Gibson63551ae2005-06-21 17:14:44 -07002989 --remainder;
2990 ++i;
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08002991 if (vaddr < vma->vm_end && remainder &&
Andi Kleena5516432008-07-23 21:27:41 -07002992 pfn_offset < pages_per_huge_page(h)) {
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08002993 /*
2994 * We use pfn_offset to avoid touching the pageframes
2995 * of this compound page.
2996 */
2997 goto same_page;
2998 }
David Gibson63551ae2005-06-21 17:14:44 -07002999 }
Hugh Dickins1c598272005-10-19 21:23:43 -07003000 spin_unlock(&mm->page_table_lock);
David Gibson63551ae2005-06-21 17:14:44 -07003001 *length = remainder;
3002 *position = vaddr;
3003
Hugh Dickins2a15efc2009-09-21 17:03:27 -07003004 return i ? i : -EFAULT;
David Gibson63551ae2005-06-21 17:14:44 -07003005}
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003006
3007void hugetlb_change_protection(struct vm_area_struct *vma,
3008 unsigned long address, unsigned long end, pgprot_t newprot)
3009{
3010 struct mm_struct *mm = vma->vm_mm;
3011 unsigned long start = address;
3012 pte_t *ptep;
3013 pte_t pte;
Andi Kleena5516432008-07-23 21:27:41 -07003014 struct hstate *h = hstate_vma(vma);
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003015
3016 BUG_ON(address >= end);
3017 flush_cache_range(vma, address, end);
3018
Peter Zijlstra3d48ae42011-05-24 17:12:06 -07003019 mutex_lock(&vma->vm_file->f_mapping->i_mmap_mutex);
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003020 spin_lock(&mm->page_table_lock);
Andi Kleena5516432008-07-23 21:27:41 -07003021 for (; address < end; address += huge_page_size(h)) {
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003022 ptep = huge_pte_offset(mm, address);
3023 if (!ptep)
3024 continue;
Chen, Kenneth W39dde652006-12-06 20:32:03 -08003025 if (huge_pmd_unshare(mm, &address, ptep))
3026 continue;
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07003027 if (!huge_pte_none(huge_ptep_get(ptep))) {
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003028 pte = huge_ptep_get_and_clear(mm, address, ptep);
3029 pte = pte_mkhuge(pte_modify(pte, newprot));
3030 set_huge_pte_at(mm, address, ptep, pte);
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003031 }
3032 }
3033 spin_unlock(&mm->page_table_lock);
Mel Gormand8333522012-07-31 16:46:20 -07003034 /*
3035 * Must flush TLB before releasing i_mmap_mutex: x86's huge_pmd_unshare
3036 * may have cleared our pud entry and done put_page on the page table:
3037 * once we release i_mmap_mutex, another task can do the final put_page
3038 * and that page table be reused and filled with junk.
3039 */
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003040 flush_tlb_range(vma, start, end);
Mel Gormand8333522012-07-31 16:46:20 -07003041 mutex_unlock(&vma->vm_file->f_mapping->i_mmap_mutex);
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003042}
3043
Mel Gormana1e78772008-07-23 21:27:23 -07003044int hugetlb_reserve_pages(struct inode *inode,
3045 long from, long to,
Mel Gorman5a6fe122009-02-10 14:02:27 +00003046 struct vm_area_struct *vma,
KOSAKI Motohiroca16d142011-05-26 19:16:19 +09003047 vm_flags_t vm_flags)
Adam Litkee4e574b2007-10-16 01:26:19 -07003048{
Mel Gorman17c9d122009-02-11 16:34:16 +00003049 long ret, chg;
Andi Kleena5516432008-07-23 21:27:41 -07003050 struct hstate *h = hstate_inode(inode);
David Gibson90481622012-03-21 16:34:12 -07003051 struct hugepage_subpool *spool = subpool_inode(inode);
Adam Litkee4e574b2007-10-16 01:26:19 -07003052
Mel Gormana1e78772008-07-23 21:27:23 -07003053 /*
Mel Gorman17c9d122009-02-11 16:34:16 +00003054 * Only apply hugepage reservation if asked. At fault time, an
3055 * attempt will be made for VM_NORESERVE to allocate a page
David Gibson90481622012-03-21 16:34:12 -07003056 * without using reserves
Mel Gorman17c9d122009-02-11 16:34:16 +00003057 */
KOSAKI Motohiroca16d142011-05-26 19:16:19 +09003058 if (vm_flags & VM_NORESERVE)
Mel Gorman17c9d122009-02-11 16:34:16 +00003059 return 0;
3060
3061 /*
Mel Gormana1e78772008-07-23 21:27:23 -07003062 * Shared mappings base their reservation on the number of pages that
3063 * are already allocated on behalf of the file. Private mappings need
3064 * to reserve the full area even if read-only as mprotect() may be
3065 * called to make the mapping read-write. Assume !vma is a shm mapping
3066 */
Mel Gormanf83a2752009-05-28 14:34:40 -07003067 if (!vma || vma->vm_flags & VM_MAYSHARE)
Mel Gormana1e78772008-07-23 21:27:23 -07003068 chg = region_chg(&inode->i_mapping->private_list, from, to);
Mel Gorman5a6fe122009-02-10 14:02:27 +00003069 else {
3070 struct resv_map *resv_map = resv_map_alloc();
Mel Gorman5a6fe122009-02-10 14:02:27 +00003071 if (!resv_map)
3072 return -ENOMEM;
3073
Mel Gorman17c9d122009-02-11 16:34:16 +00003074 chg = to - from;
3075
Mel Gorman5a6fe122009-02-10 14:02:27 +00003076 set_vma_resv_map(vma, resv_map);
3077 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
3078 }
3079
Dave Hansenc50ac052012-05-29 15:06:46 -07003080 if (chg < 0) {
3081 ret = chg;
3082 goto out_err;
3083 }
Mel Gorman17c9d122009-02-11 16:34:16 +00003084
David Gibson90481622012-03-21 16:34:12 -07003085 /* There must be enough pages in the subpool for the mapping */
Dave Hansenc50ac052012-05-29 15:06:46 -07003086 if (hugepage_subpool_get_pages(spool, chg)) {
3087 ret = -ENOSPC;
3088 goto out_err;
3089 }
Mel Gorman17c9d122009-02-11 16:34:16 +00003090
3091 /*
3092 * Check enough hugepages are available for the reservation.
David Gibson90481622012-03-21 16:34:12 -07003093 * Hand the pages back to the subpool if there are not
Mel Gorman17c9d122009-02-11 16:34:16 +00003094 */
3095 ret = hugetlb_acct_memory(h, chg);
3096 if (ret < 0) {
David Gibson90481622012-03-21 16:34:12 -07003097 hugepage_subpool_put_pages(spool, chg);
Dave Hansenc50ac052012-05-29 15:06:46 -07003098 goto out_err;
Mel Gorman17c9d122009-02-11 16:34:16 +00003099 }
3100
3101 /*
3102 * Account for the reservations made. Shared mappings record regions
3103 * that have reservations as they are shared by multiple VMAs.
3104 * When the last VMA disappears, the region map says how much
3105 * the reservation was and the page cache tells how much of
3106 * the reservation was consumed. Private mappings are per-VMA and
3107 * only the consumed reservations are tracked. When the VMA
3108 * disappears, the original reservation is the VMA size and the
3109 * consumed reservations are stored in the map. Hence, nothing
3110 * else has to be done for private mappings here
3111 */
Mel Gormanf83a2752009-05-28 14:34:40 -07003112 if (!vma || vma->vm_flags & VM_MAYSHARE)
Mel Gorman17c9d122009-02-11 16:34:16 +00003113 region_add(&inode->i_mapping->private_list, from, to);
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -07003114 return 0;
Dave Hansenc50ac052012-05-29 15:06:46 -07003115out_err:
Dave Hansen4523e142012-05-30 07:51:07 -07003116 if (vma)
3117 resv_map_put(vma);
Dave Hansenc50ac052012-05-29 15:06:46 -07003118 return ret;
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -07003119}
3120
3121void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed)
3122{
Andi Kleena5516432008-07-23 21:27:41 -07003123 struct hstate *h = hstate_inode(inode);
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -07003124 long chg = region_truncate(&inode->i_mapping->private_list, offset);
David Gibson90481622012-03-21 16:34:12 -07003125 struct hugepage_subpool *spool = subpool_inode(inode);
Ken Chen45c682a2007-11-14 16:59:44 -08003126
3127 spin_lock(&inode->i_lock);
Eric Sandeene4c6f8b2009-07-29 15:02:16 -07003128 inode->i_blocks -= (blocks_per_huge_page(h) * freed);
Ken Chen45c682a2007-11-14 16:59:44 -08003129 spin_unlock(&inode->i_lock);
3130
David Gibson90481622012-03-21 16:34:12 -07003131 hugepage_subpool_put_pages(spool, (chg - freed));
Andi Kleena5516432008-07-23 21:27:41 -07003132 hugetlb_acct_memory(h, -(chg - freed));
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -07003133}
Naoya Horiguchi93f70f92010-05-28 09:29:20 +09003134
Andi Kleend5bd9102010-09-27 09:00:12 +02003135#ifdef CONFIG_MEMORY_FAILURE
3136
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +09003137/* Should be called in hugetlb_lock */
3138static int is_hugepage_on_freelist(struct page *hpage)
3139{
3140 struct page *page;
3141 struct page *tmp;
3142 struct hstate *h = page_hstate(hpage);
3143 int nid = page_to_nid(hpage);
3144
3145 list_for_each_entry_safe(page, tmp, &h->hugepage_freelists[nid], lru)
3146 if (page == hpage)
3147 return 1;
3148 return 0;
3149}
3150
Naoya Horiguchi93f70f92010-05-28 09:29:20 +09003151/*
3152 * This function is called from memory failure code.
3153 * Assume the caller holds page lock of the head page.
3154 */
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +09003155int dequeue_hwpoisoned_huge_page(struct page *hpage)
Naoya Horiguchi93f70f92010-05-28 09:29:20 +09003156{
3157 struct hstate *h = page_hstate(hpage);
3158 int nid = page_to_nid(hpage);
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +09003159 int ret = -EBUSY;
Naoya Horiguchi93f70f92010-05-28 09:29:20 +09003160
3161 spin_lock(&hugetlb_lock);
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +09003162 if (is_hugepage_on_freelist(hpage)) {
3163 list_del(&hpage->lru);
Naoya Horiguchi8c6c2ec2010-09-08 10:19:38 +09003164 set_page_refcounted(hpage);
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +09003165 h->free_huge_pages--;
3166 h->free_huge_pages_node[nid]--;
3167 ret = 0;
3168 }
Naoya Horiguchi93f70f92010-05-28 09:29:20 +09003169 spin_unlock(&hugetlb_lock);
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +09003170 return ret;
Naoya Horiguchi93f70f92010-05-28 09:29:20 +09003171}
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +09003172#endif