blob: c01cb9fedb18f8495c815d4caa380580734e1c8e [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
2 * Generic hugetlb support.
Nadia Yvette Chambers6d49e352012-12-06 10:39:54 +01003 * (C) Nadia Yvette Chambers, April 2004
Linus Torvalds1da177e2005-04-16 15:20:36 -07004 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07005#include <linux/list.h>
6#include <linux/init.h>
7#include <linux/module.h>
8#include <linux/mm.h>
Alexey Dobriyane1759c22008-10-15 23:50:22 +04009#include <linux/seq_file.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070010#include <linux/sysctl.h>
11#include <linux/highmem.h>
Andrea Arcangelicddb8a52008-07-28 15:46:29 -070012#include <linux/mmu_notifier.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070013#include <linux/nodemask.h>
David Gibson63551ae2005-06-21 17:14:44 -070014#include <linux/pagemap.h>
Christoph Lameter5da7ca82006-01-06 00:10:46 -080015#include <linux/mempolicy.h>
Christoph Lameteraea47ff2006-01-08 01:00:57 -080016#include <linux/cpuset.h>
David Gibson3935baa2006-03-22 00:08:53 -080017#include <linux/mutex.h>
Andi Kleenaa888a72008-07-23 21:27:47 -070018#include <linux/bootmem.h>
Nishanth Aravamudana3437872008-07-23 21:27:44 -070019#include <linux/sysfs.h>
Tejun Heo5a0e3ad2010-03-24 17:04:11 +090020#include <linux/slab.h>
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +090021#include <linux/rmap.h>
Naoya Horiguchifd6a03e2010-05-28 09:29:21 +090022#include <linux/swap.h>
23#include <linux/swapops.h>
Naoya Horiguchic8721bb2013-09-11 14:22:09 -070024#include <linux/page-isolation.h>
Linus Torvaldsd6606682008-08-06 12:04:54 -070025
David Gibson63551ae2005-06-21 17:14:44 -070026#include <asm/page.h>
27#include <asm/pgtable.h>
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -070028#include <asm/tlb.h>
David Gibson63551ae2005-06-21 17:14:44 -070029
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -070030#include <linux/io.h>
David Gibson63551ae2005-06-21 17:14:44 -070031#include <linux/hugetlb.h>
Aneesh Kumar K.V9dd540e2012-07-31 16:42:15 -070032#include <linux/hugetlb_cgroup.h>
Lee Schermerhorn9a3052302009-12-14 17:58:25 -080033#include <linux/node.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 -070037unsigned long hugepages_treat_as_movable;
Andi Kleena5516432008-07-23 21:27:41 -070038
Aneesh Kumar K.Vc3f38a32012-07-31 16:42:10 -070039int hugetlb_max_hstate __read_mostly;
Andi Kleene5ff2152008-07-23 21:27:42 -070040unsigned int default_hstate_idx;
41struct hstate hstates[HUGE_MAX_HSTATE];
42
Jon Tollefson53ba51d2008-07-23 21:27:52 -070043__initdata LIST_HEAD(huge_boot_pages);
44
Andi Kleene5ff2152008-07-23 21:27:42 -070045/* for command line parsing */
46static struct hstate * __initdata parsed_hstate;
47static unsigned long __initdata default_hstate_max_huge_pages;
Nick Piggine11bfbf2008-07-23 21:27:52 -070048static unsigned long __initdata default_hstate_size;
Andi Kleene5ff2152008-07-23 21:27:42 -070049
David Gibson3935baa2006-03-22 00:08:53 -080050/*
Naoya Horiguchi31caf662013-09-11 14:21:59 -070051 * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages,
52 * free_huge_pages, and surplus_huge_pages.
David Gibson3935baa2006-03-22 00:08:53 -080053 */
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{
Al Viro496ad9a2013-01-23 17:07:38 -0500131 return subpool_inode(file_inode(vma->vm_file));
David Gibson90481622012-03-21 16:34:12 -0700132}
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
Joonsoo Kimc748c262013-09-11 14:20:57 -0700139 * and the hugetlb_instantiation_mutex. To access or modify a region the caller
Andy Whitcroft84afd992008-07-23 21:27:32 -0700140 * must either hold the mmap_sem for write, or the mmap_sem for read and
Joonsoo Kimc748c262013-09-11 14:20:57 -0700141 * the hugetlb_instantiation_mutex:
Andy Whitcroft84afd992008-07-23 21:27:32 -0700142 *
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
Wanpeng Li2415cf12013-07-03 15:02:43 -0700323 return 1UL << huge_page_shift(hstate);
Mel Gorman08fba692009-01-06 14:38:53 -0800324}
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
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700438/* Reset counters to 0 and clear all HPAGE_RESV_* flags */
Mel Gormana1e78772008-07-23 21:27:23 -0700439void reset_vma_resv_huge_pages(struct vm_area_struct *vma)
440{
441 VM_BUG_ON(!is_vm_hugetlb_page(vma));
Mel Gormanf83a2752009-05-28 14:34:40 -0700442 if (!(vma->vm_flags & VM_MAYSHARE))
Mel Gormana1e78772008-07-23 21:27:23 -0700443 vma->vm_private_data = (void *)0;
444}
445
446/* Returns true if the VMA has associated reserve pages */
Joonsoo Kimaf0ed732013-09-11 14:21:18 -0700447static int vma_has_reserves(struct vm_area_struct *vma, long chg)
Mel Gormana1e78772008-07-23 21:27:23 -0700448{
Joonsoo Kimaf0ed732013-09-11 14:21:18 -0700449 if (vma->vm_flags & VM_NORESERVE) {
450 /*
451 * This address is already reserved by other process(chg == 0),
452 * so, we should decrement reserved count. Without decrementing,
453 * reserve count remains after releasing inode, because this
454 * allocated page will go into page cache and is regarded as
455 * coming from reserved pool in releasing step. Currently, we
456 * don't have any other solution to deal with this situation
457 * properly, so add work-around here.
458 */
459 if (vma->vm_flags & VM_MAYSHARE && chg == 0)
460 return 1;
461 else
462 return 0;
463 }
Joonsoo Kima63884e2013-09-11 14:21:07 -0700464
465 /* Shared mappings always use reserves */
Mel Gormanf83a2752009-05-28 14:34:40 -0700466 if (vma->vm_flags & VM_MAYSHARE)
Mel Gorman7f09ca52008-07-23 21:27:58 -0700467 return 1;
Joonsoo Kima63884e2013-09-11 14:21:07 -0700468
469 /*
470 * Only the process that called mmap() has reserves for
471 * private mappings.
472 */
Mel Gorman7f09ca52008-07-23 21:27:58 -0700473 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER))
474 return 1;
Joonsoo Kima63884e2013-09-11 14:21:07 -0700475
Mel Gorman7f09ca52008-07-23 21:27:58 -0700476 return 0;
Mel Gormana1e78772008-07-23 21:27:23 -0700477}
478
Andi Kleena5516432008-07-23 21:27:41 -0700479static void enqueue_huge_page(struct hstate *h, struct page *page)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700480{
481 int nid = page_to_nid(page);
Aneesh Kumar K.V0edaecf2012-07-31 16:42:07 -0700482 list_move(&page->lru, &h->hugepage_freelists[nid]);
Andi Kleena5516432008-07-23 21:27:41 -0700483 h->free_huge_pages++;
484 h->free_huge_pages_node[nid]++;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700485}
486
Naoya Horiguchibf50bab2010-09-08 10:19:33 +0900487static struct page *dequeue_huge_page_node(struct hstate *h, int nid)
488{
489 struct page *page;
490
Naoya Horiguchic8721bb2013-09-11 14:22:09 -0700491 list_for_each_entry(page, &h->hugepage_freelists[nid], lru)
492 if (!is_migrate_isolate_page(page))
493 break;
494 /*
495 * if 'non-isolated free hugepage' not found on the list,
496 * the allocation fails.
497 */
498 if (&h->hugepage_freelists[nid] == &page->lru)
Naoya Horiguchibf50bab2010-09-08 10:19:33 +0900499 return NULL;
Aneesh Kumar K.V0edaecf2012-07-31 16:42:07 -0700500 list_move(&page->lru, &h->hugepage_activelist);
Naoya Horiguchia9869b82010-09-08 10:19:37 +0900501 set_page_refcounted(page);
Naoya Horiguchibf50bab2010-09-08 10:19:33 +0900502 h->free_huge_pages--;
503 h->free_huge_pages_node[nid]--;
504 return page;
505}
506
Naoya Horiguchi86cdb462013-09-11 14:22:13 -0700507/* Movability of hugepages depends on migration support. */
508static inline gfp_t htlb_alloc_mask(struct hstate *h)
509{
510 if (hugepages_treat_as_movable || hugepage_migration_support(h))
511 return GFP_HIGHUSER_MOVABLE;
512 else
513 return GFP_HIGHUSER;
514}
515
Andi Kleena5516432008-07-23 21:27:41 -0700516static struct page *dequeue_huge_page_vma(struct hstate *h,
517 struct vm_area_struct *vma,
Joonsoo Kimaf0ed732013-09-11 14:21:18 -0700518 unsigned long address, int avoid_reserve,
519 long chg)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700520{
Konstantin Khlebnikovb1c12cb2012-04-25 16:01:46 -0700521 struct page *page = NULL;
Lee Schermerhorn480eccf2007-09-18 22:46:47 -0700522 struct mempolicy *mpol;
Mel Gorman19770b32008-04-28 02:12:18 -0700523 nodemask_t *nodemask;
Miao Xiec0ff7452010-05-24 14:32:08 -0700524 struct zonelist *zonelist;
Mel Gormandd1a2392008-04-28 02:12:17 -0700525 struct zone *zone;
526 struct zoneref *z;
Mel Gormancc9a6c82012-03-21 16:34:11 -0700527 unsigned int cpuset_mems_cookie;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700528
Mel Gormana1e78772008-07-23 21:27:23 -0700529 /*
530 * A child process with MAP_PRIVATE mappings created by their parent
531 * have no page reserves. This check ensures that reservations are
532 * not "stolen". The child may still get SIGKILLed
533 */
Joonsoo Kimaf0ed732013-09-11 14:21:18 -0700534 if (!vma_has_reserves(vma, chg) &&
Andi Kleena5516432008-07-23 21:27:41 -0700535 h->free_huge_pages - h->resv_huge_pages == 0)
Miao Xiec0ff7452010-05-24 14:32:08 -0700536 goto err;
Mel Gormana1e78772008-07-23 21:27:23 -0700537
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700538 /* If reserves cannot be used, ensure enough pages are in the pool */
Andi Kleena5516432008-07-23 21:27:41 -0700539 if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
Justin P. Mattock6eab04a2011-04-08 19:49:08 -0700540 goto err;
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700541
Joonsoo Kim9966c4b2013-09-11 14:20:50 -0700542retry_cpuset:
543 cpuset_mems_cookie = get_mems_allowed();
544 zonelist = huge_zonelist(vma, address,
Naoya Horiguchi86cdb462013-09-11 14:22:13 -0700545 htlb_alloc_mask(h), &mpol, &nodemask);
Joonsoo Kim9966c4b2013-09-11 14:20:50 -0700546
Mel Gorman19770b32008-04-28 02:12:18 -0700547 for_each_zone_zonelist_nodemask(zone, z, zonelist,
548 MAX_NR_ZONES - 1, nodemask) {
Naoya Horiguchi86cdb462013-09-11 14:22:13 -0700549 if (cpuset_zone_allowed_softwall(zone, htlb_alloc_mask(h))) {
Naoya Horiguchibf50bab2010-09-08 10:19:33 +0900550 page = dequeue_huge_page_node(h, zone_to_nid(zone));
551 if (page) {
Joonsoo Kimaf0ed732013-09-11 14:21:18 -0700552 if (avoid_reserve)
553 break;
554 if (!vma_has_reserves(vma, chg))
555 break;
556
Joonsoo Kim07443a82013-09-11 14:21:58 -0700557 SetPagePrivate(page);
Joonsoo Kimaf0ed732013-09-11 14:21:18 -0700558 h->resv_huge_pages--;
Naoya Horiguchibf50bab2010-09-08 10:19:33 +0900559 break;
560 }
Andrew Morton3abf7af2007-07-19 01:49:08 -0700561 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700562 }
Mel Gormancc9a6c82012-03-21 16:34:11 -0700563
564 mpol_cond_put(mpol);
565 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page))
566 goto retry_cpuset;
567 return page;
568
Miao Xiec0ff7452010-05-24 14:32:08 -0700569err:
Mel Gormancc9a6c82012-03-21 16:34:11 -0700570 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700571}
572
Andi Kleena5516432008-07-23 21:27:41 -0700573static void update_and_free_page(struct hstate *h, struct page *page)
Adam Litke6af2acb2007-10-16 01:26:16 -0700574{
575 int i;
Andi Kleena5516432008-07-23 21:27:41 -0700576
Andy Whitcroft18229df2008-11-06 12:53:27 -0800577 VM_BUG_ON(h->order >= MAX_ORDER);
578
Andi Kleena5516432008-07-23 21:27:41 -0700579 h->nr_huge_pages--;
580 h->nr_huge_pages_node[page_to_nid(page)]--;
581 for (i = 0; i < pages_per_huge_page(h); i++) {
Chris Forbes32f84522011-07-25 17:12:14 -0700582 page[i].flags &= ~(1 << PG_locked | 1 << PG_error |
583 1 << PG_referenced | 1 << PG_dirty |
584 1 << PG_active | 1 << PG_reserved |
585 1 << PG_private | 1 << PG_writeback);
Adam Litke6af2acb2007-10-16 01:26:16 -0700586 }
Sasha Levin309381fea2014-01-23 15:52:54 -0800587 VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page);
Adam Litke6af2acb2007-10-16 01:26:16 -0700588 set_compound_page_dtor(page, NULL);
589 set_page_refcounted(page);
Gerald Schaefer7f2e9522008-04-28 02:13:29 -0700590 arch_release_hugepage(page);
Andi Kleena5516432008-07-23 21:27:41 -0700591 __free_pages(page, huge_page_order(h));
Adam Litke6af2acb2007-10-16 01:26:16 -0700592}
593
Andi Kleene5ff2152008-07-23 21:27:42 -0700594struct hstate *size_to_hstate(unsigned long size)
595{
596 struct hstate *h;
597
598 for_each_hstate(h) {
599 if (huge_page_size(h) == size)
600 return h;
601 }
602 return NULL;
603}
604
David Gibson27a85ef2006-03-22 00:08:56 -0800605static void free_huge_page(struct page *page)
606{
Andi Kleena5516432008-07-23 21:27:41 -0700607 /*
608 * Can't pass hstate in here because it is called from the
609 * compound page destructor.
610 */
Andi Kleene5ff2152008-07-23 21:27:42 -0700611 struct hstate *h = page_hstate(page);
Adam Litke7893d1d2007-10-16 01:26:18 -0700612 int nid = page_to_nid(page);
David Gibson90481622012-03-21 16:34:12 -0700613 struct hugepage_subpool *spool =
614 (struct hugepage_subpool *)page_private(page);
Joonsoo Kim07443a82013-09-11 14:21:58 -0700615 bool restore_reserve;
David Gibson27a85ef2006-03-22 00:08:56 -0800616
Andy Whitcrofte5df70a2008-02-23 15:23:32 -0800617 set_page_private(page, 0);
Mel Gorman23be7462010-04-23 13:17:56 -0400618 page->mapping = NULL;
Adam Litke7893d1d2007-10-16 01:26:18 -0700619 BUG_ON(page_count(page));
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +0900620 BUG_ON(page_mapcount(page));
Joonsoo Kim07443a82013-09-11 14:21:58 -0700621 restore_reserve = PagePrivate(page);
Joonsoo Kim16c794b2013-10-16 13:46:48 -0700622 ClearPagePrivate(page);
David Gibson27a85ef2006-03-22 00:08:56 -0800623
624 spin_lock(&hugetlb_lock);
Aneesh Kumar K.V6d76dcf2012-07-31 16:42:18 -0700625 hugetlb_cgroup_uncharge_page(hstate_index(h),
626 pages_per_huge_page(h), page);
Joonsoo Kim07443a82013-09-11 14:21:58 -0700627 if (restore_reserve)
628 h->resv_huge_pages++;
629
Andi Kleenaa888a72008-07-23 21:27:47 -0700630 if (h->surplus_huge_pages_node[nid] && huge_page_order(h) < MAX_ORDER) {
Aneesh Kumar K.V0edaecf2012-07-31 16:42:07 -0700631 /* remove the page from active list */
632 list_del(&page->lru);
Andi Kleena5516432008-07-23 21:27:41 -0700633 update_and_free_page(h, page);
634 h->surplus_huge_pages--;
635 h->surplus_huge_pages_node[nid]--;
Adam Litke7893d1d2007-10-16 01:26:18 -0700636 } else {
Will Deacon5d3a5512012-10-08 16:29:32 -0700637 arch_clear_hugepage_flags(page);
Andi Kleena5516432008-07-23 21:27:41 -0700638 enqueue_huge_page(h, page);
Adam Litke7893d1d2007-10-16 01:26:18 -0700639 }
David Gibson27a85ef2006-03-22 00:08:56 -0800640 spin_unlock(&hugetlb_lock);
David Gibson90481622012-03-21 16:34:12 -0700641 hugepage_subpool_put_pages(spool, 1);
David Gibson27a85ef2006-03-22 00:08:56 -0800642}
643
Andi Kleena5516432008-07-23 21:27:41 -0700644static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
Andi Kleenb7ba30c2008-07-23 21:27:40 -0700645{
Aneesh Kumar K.V0edaecf2012-07-31 16:42:07 -0700646 INIT_LIST_HEAD(&page->lru);
Andi Kleenb7ba30c2008-07-23 21:27:40 -0700647 set_compound_page_dtor(page, free_huge_page);
648 spin_lock(&hugetlb_lock);
Aneesh Kumar K.V9dd540e2012-07-31 16:42:15 -0700649 set_hugetlb_cgroup(page, NULL);
Andi Kleena5516432008-07-23 21:27:41 -0700650 h->nr_huge_pages++;
651 h->nr_huge_pages_node[nid]++;
Andi Kleenb7ba30c2008-07-23 21:27:40 -0700652 spin_unlock(&hugetlb_lock);
653 put_page(page); /* free it into the hugepage allocator */
654}
655
Wu Fengguang20a03072009-06-16 15:32:22 -0700656static void prep_compound_gigantic_page(struct page *page, unsigned long order)
657{
658 int i;
659 int nr_pages = 1 << order;
660 struct page *p = page + 1;
661
662 /* we rely on prep_new_huge_page to set the destructor */
663 set_compound_order(page, order);
664 __SetPageHead(page);
Andrea Arcangelief5a22b2013-10-16 13:46:56 -0700665 __ClearPageReserved(page);
Wu Fengguang20a03072009-06-16 15:32:22 -0700666 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
667 __SetPageTail(p);
Andrea Arcangelief5a22b2013-10-16 13:46:56 -0700668 /*
669 * For gigantic hugepages allocated through bootmem at
670 * boot, it's safer to be consistent with the not-gigantic
671 * hugepages and clear the PG_reserved bit from all tail pages
672 * too. Otherwse drivers using get_user_pages() to access tail
673 * pages may get the reference counting wrong if they see
674 * PG_reserved set on a tail page (despite the head page not
675 * having PG_reserved set). Enforcing this consistency between
676 * head and tail pages allows drivers to optimize away a check
677 * on the head page when they need know if put_page() is needed
678 * after get_user_pages().
679 */
680 __ClearPageReserved(p);
Youquan Song58a84aa2011-12-08 14:34:18 -0800681 set_page_count(p, 0);
Wu Fengguang20a03072009-06-16 15:32:22 -0700682 p->first_page = page;
683 }
684}
685
Andrew Morton77959122012-10-08 16:34:11 -0700686/*
687 * PageHuge() only returns true for hugetlbfs pages, but not for normal or
688 * transparent huge pages. See the PageTransHuge() documentation for more
689 * details.
690 */
Wu Fengguang20a03072009-06-16 15:32:22 -0700691int PageHuge(struct page *page)
692{
Wu Fengguang20a03072009-06-16 15:32:22 -0700693 if (!PageCompound(page))
694 return 0;
695
696 page = compound_head(page);
Andrew Morton758f66a2014-01-21 15:48:57 -0800697 return get_compound_page_dtor(page) == free_huge_page;
Wu Fengguang20a03072009-06-16 15:32:22 -0700698}
Naoya Horiguchi43131e12010-05-28 09:29:22 +0900699EXPORT_SYMBOL_GPL(PageHuge);
700
Andrea Arcangeli27c73ae2013-11-21 14:32:02 -0800701/*
702 * PageHeadHuge() only returns true for hugetlbfs head page, but not for
703 * normal or transparent huge pages.
704 */
705int PageHeadHuge(struct page *page_head)
706{
Andrea Arcangeli27c73ae2013-11-21 14:32:02 -0800707 if (!PageHead(page_head))
708 return 0;
709
Andrew Morton758f66a2014-01-21 15:48:57 -0800710 return get_compound_page_dtor(page_head) == free_huge_page;
Andrea Arcangeli27c73ae2013-11-21 14:32:02 -0800711}
Andrea Arcangeli27c73ae2013-11-21 14:32:02 -0800712
Zhang Yi13d60f42013-06-25 21:19:31 +0800713pgoff_t __basepage_index(struct page *page)
714{
715 struct page *page_head = compound_head(page);
716 pgoff_t index = page_index(page_head);
717 unsigned long compound_idx;
718
719 if (!PageHuge(page_head))
720 return page_index(page);
721
722 if (compound_order(page_head) >= MAX_ORDER)
723 compound_idx = page_to_pfn(page) - page_to_pfn(page_head);
724 else
725 compound_idx = page - page_head;
726
727 return (index << compound_order(page_head)) + compound_idx;
728}
729
Andi Kleena5516432008-07-23 21:27:41 -0700730static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700731{
Linus Torvalds1da177e2005-04-16 15:20:36 -0700732 struct page *page;
Joe Jinf96efd52007-07-15 23:38:12 -0700733
Andi Kleenaa888a72008-07-23 21:27:47 -0700734 if (h->order >= MAX_ORDER)
735 return NULL;
736
Mel Gorman6484eb32009-06-16 15:31:54 -0700737 page = alloc_pages_exact_node(nid,
Naoya Horiguchi86cdb462013-09-11 14:22:13 -0700738 htlb_alloc_mask(h)|__GFP_COMP|__GFP_THISNODE|
Nishanth Aravamudan551883a2008-04-29 00:58:26 -0700739 __GFP_REPEAT|__GFP_NOWARN,
Andi Kleena5516432008-07-23 21:27:41 -0700740 huge_page_order(h));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700741 if (page) {
Gerald Schaefer7f2e9522008-04-28 02:13:29 -0700742 if (arch_prepare_hugepage(page)) {
Gerald Schaefercaff3a22008-08-12 15:08:38 -0700743 __free_pages(page, huge_page_order(h));
Harvey Harrison7b8ee842008-04-28 14:13:19 -0700744 return NULL;
Gerald Schaefer7f2e9522008-04-28 02:13:29 -0700745 }
Andi Kleena5516432008-07-23 21:27:41 -0700746 prep_new_huge_page(h, page, nid);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700747 }
Nishanth Aravamudan63b46132007-10-16 01:26:24 -0700748
749 return page;
750}
751
Andi Kleen5ced66c2008-07-23 21:27:45 -0700752/*
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800753 * common helper functions for hstate_next_node_to_{alloc|free}.
754 * We may have allocated or freed a huge page based on a different
755 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
756 * be outside of *nodes_allowed. Ensure that we use an allowed
757 * node for alloc or free.
Lee Schermerhorn9a76db02009-12-14 17:58:15 -0800758 */
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800759static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
Lee Schermerhorn9a76db02009-12-14 17:58:15 -0800760{
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800761 nid = next_node(nid, *nodes_allowed);
Lee Schermerhorn9a76db02009-12-14 17:58:15 -0800762 if (nid == MAX_NUMNODES)
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800763 nid = first_node(*nodes_allowed);
Lee Schermerhorn9a76db02009-12-14 17:58:15 -0800764 VM_BUG_ON(nid >= MAX_NUMNODES);
765
766 return nid;
767}
768
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800769static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
Andi Kleen5ced66c2008-07-23 21:27:45 -0700770{
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800771 if (!node_isset(nid, *nodes_allowed))
772 nid = next_node_allowed(nid, nodes_allowed);
Lee Schermerhorn9a76db02009-12-14 17:58:15 -0800773 return nid;
Andi Kleen5ced66c2008-07-23 21:27:45 -0700774}
775
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800776/*
777 * returns the previously saved node ["this node"] from which to
778 * allocate a persistent huge page for the pool and advance the
779 * next node from which to allocate, handling wrap at end of node
780 * mask.
781 */
782static int hstate_next_node_to_alloc(struct hstate *h,
783 nodemask_t *nodes_allowed)
784{
785 int nid;
786
787 VM_BUG_ON(!nodes_allowed);
788
789 nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
790 h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
791
792 return nid;
793}
794
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700795/*
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800796 * helper for free_pool_huge_page() - return the previously saved
797 * node ["this node"] from which to free a huge page. Advance the
798 * next node id whether or not we find a free huge page to free so
799 * that the next attempt to free addresses the next node.
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700800 */
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800801static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700802{
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800803 int nid;
Lee Schermerhorn9a76db02009-12-14 17:58:15 -0800804
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800805 VM_BUG_ON(!nodes_allowed);
806
807 nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
808 h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
809
Lee Schermerhorn9a76db02009-12-14 17:58:15 -0800810 return nid;
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700811}
812
Joonsoo Kimb2261022013-09-11 14:21:00 -0700813#define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \
814 for (nr_nodes = nodes_weight(*mask); \
815 nr_nodes > 0 && \
816 ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \
817 nr_nodes--)
818
819#define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \
820 for (nr_nodes = nodes_weight(*mask); \
821 nr_nodes > 0 && \
822 ((node = hstate_next_node_to_free(hs, mask)) || 1); \
823 nr_nodes--)
824
825static int alloc_fresh_huge_page(struct hstate *h, nodemask_t *nodes_allowed)
826{
827 struct page *page;
828 int nr_nodes, node;
829 int ret = 0;
830
831 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
832 page = alloc_fresh_huge_page_node(h, node);
833 if (page) {
834 ret = 1;
835 break;
836 }
837 }
838
839 if (ret)
840 count_vm_event(HTLB_BUDDY_PGALLOC);
841 else
842 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
843
844 return ret;
845}
846
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700847/*
848 * Free huge page from pool from next node to free.
849 * Attempt to keep persistent huge pages more or less
850 * balanced over allowed nodes.
851 * Called with hugetlb_lock locked.
852 */
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800853static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
854 bool acct_surplus)
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700855{
Joonsoo Kimb2261022013-09-11 14:21:00 -0700856 int nr_nodes, node;
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700857 int ret = 0;
858
Joonsoo Kimb2261022013-09-11 14:21:00 -0700859 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
Lee Schermerhorn685f3452009-09-21 17:01:23 -0700860 /*
861 * If we're returning unused surplus pages, only examine
862 * nodes with surplus pages.
863 */
Joonsoo Kimb2261022013-09-11 14:21:00 -0700864 if ((!acct_surplus || h->surplus_huge_pages_node[node]) &&
865 !list_empty(&h->hugepage_freelists[node])) {
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700866 struct page *page =
Joonsoo Kimb2261022013-09-11 14:21:00 -0700867 list_entry(h->hugepage_freelists[node].next,
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700868 struct page, lru);
869 list_del(&page->lru);
870 h->free_huge_pages--;
Joonsoo Kimb2261022013-09-11 14:21:00 -0700871 h->free_huge_pages_node[node]--;
Lee Schermerhorn685f3452009-09-21 17:01:23 -0700872 if (acct_surplus) {
873 h->surplus_huge_pages--;
Joonsoo Kimb2261022013-09-11 14:21:00 -0700874 h->surplus_huge_pages_node[node]--;
Lee Schermerhorn685f3452009-09-21 17:01:23 -0700875 }
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700876 update_and_free_page(h, page);
877 ret = 1;
Lee Schermerhorn9a76db02009-12-14 17:58:15 -0800878 break;
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700879 }
Joonsoo Kimb2261022013-09-11 14:21:00 -0700880 }
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700881
882 return ret;
883}
884
Naoya Horiguchic8721bb2013-09-11 14:22:09 -0700885/*
886 * Dissolve a given free hugepage into free buddy pages. This function does
887 * nothing for in-use (including surplus) hugepages.
888 */
889static void dissolve_free_huge_page(struct page *page)
890{
891 spin_lock(&hugetlb_lock);
892 if (PageHuge(page) && !page_count(page)) {
893 struct hstate *h = page_hstate(page);
894 int nid = page_to_nid(page);
895 list_del(&page->lru);
896 h->free_huge_pages--;
897 h->free_huge_pages_node[nid]--;
898 update_and_free_page(h, page);
899 }
900 spin_unlock(&hugetlb_lock);
901}
902
903/*
904 * Dissolve free hugepages in a given pfn range. Used by memory hotplug to
905 * make specified memory blocks removable from the system.
906 * Note that start_pfn should aligned with (minimum) hugepage size.
907 */
908void dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn)
909{
910 unsigned int order = 8 * sizeof(void *);
911 unsigned long pfn;
912 struct hstate *h;
913
914 /* Set scan step to minimum hugepage size */
915 for_each_hstate(h)
916 if (order > huge_page_order(h))
917 order = huge_page_order(h);
918 VM_BUG_ON(!IS_ALIGNED(start_pfn, 1 << order));
919 for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << order)
920 dissolve_free_huge_page(pfn_to_page(pfn));
921}
922
Naoya Horiguchibf50bab2010-09-08 10:19:33 +0900923static struct page *alloc_buddy_huge_page(struct hstate *h, int nid)
Adam Litke7893d1d2007-10-16 01:26:18 -0700924{
925 struct page *page;
Naoya Horiguchibf50bab2010-09-08 10:19:33 +0900926 unsigned int r_nid;
Adam Litke7893d1d2007-10-16 01:26:18 -0700927
Andi Kleenaa888a72008-07-23 21:27:47 -0700928 if (h->order >= MAX_ORDER)
929 return NULL;
930
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800931 /*
932 * Assume we will successfully allocate the surplus page to
933 * prevent racing processes from causing the surplus to exceed
934 * overcommit
935 *
936 * This however introduces a different race, where a process B
937 * tries to grow the static hugepage pool while alloc_pages() is
938 * called by process A. B will only examine the per-node
939 * counters in determining if surplus huge pages can be
940 * converted to normal huge pages in adjust_pool_surplus(). A
941 * won't be able to increment the per-node counter, until the
942 * lock is dropped by B, but B doesn't drop hugetlb_lock until
943 * no more huge pages can be converted from surplus to normal
944 * state (and doesn't try to convert again). Thus, we have a
945 * case where a surplus huge page exists, the pool is grown, and
946 * the surplus huge page still exists after, even though it
947 * should just have been converted to a normal huge page. This
948 * does not leak memory, though, as the hugepage will be freed
949 * once it is out of use. It also does not allow the counters to
950 * go out of whack in adjust_pool_surplus() as we don't modify
951 * the node values until we've gotten the hugepage and only the
952 * per-node value is checked there.
953 */
954 spin_lock(&hugetlb_lock);
Andi Kleena5516432008-07-23 21:27:41 -0700955 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800956 spin_unlock(&hugetlb_lock);
957 return NULL;
958 } else {
Andi Kleena5516432008-07-23 21:27:41 -0700959 h->nr_huge_pages++;
960 h->surplus_huge_pages++;
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800961 }
962 spin_unlock(&hugetlb_lock);
963
Naoya Horiguchibf50bab2010-09-08 10:19:33 +0900964 if (nid == NUMA_NO_NODE)
Naoya Horiguchi86cdb462013-09-11 14:22:13 -0700965 page = alloc_pages(htlb_alloc_mask(h)|__GFP_COMP|
Naoya Horiguchibf50bab2010-09-08 10:19:33 +0900966 __GFP_REPEAT|__GFP_NOWARN,
967 huge_page_order(h));
968 else
969 page = alloc_pages_exact_node(nid,
Naoya Horiguchi86cdb462013-09-11 14:22:13 -0700970 htlb_alloc_mask(h)|__GFP_COMP|__GFP_THISNODE|
Naoya Horiguchibf50bab2010-09-08 10:19:33 +0900971 __GFP_REPEAT|__GFP_NOWARN, huge_page_order(h));
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800972
Gerald Schaefercaff3a22008-08-12 15:08:38 -0700973 if (page && arch_prepare_hugepage(page)) {
974 __free_pages(page, huge_page_order(h));
Hillf Dantonea5768c2012-01-10 15:08:30 -0800975 page = NULL;
Gerald Schaefercaff3a22008-08-12 15:08:38 -0700976 }
977
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800978 spin_lock(&hugetlb_lock);
Adam Litke7893d1d2007-10-16 01:26:18 -0700979 if (page) {
Aneesh Kumar K.V0edaecf2012-07-31 16:42:07 -0700980 INIT_LIST_HEAD(&page->lru);
Naoya Horiguchibf50bab2010-09-08 10:19:33 +0900981 r_nid = page_to_nid(page);
Adam Litke7893d1d2007-10-16 01:26:18 -0700982 set_compound_page_dtor(page, free_huge_page);
Aneesh Kumar K.V9dd540e2012-07-31 16:42:15 -0700983 set_hugetlb_cgroup(page, NULL);
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800984 /*
985 * We incremented the global counters already
986 */
Naoya Horiguchibf50bab2010-09-08 10:19:33 +0900987 h->nr_huge_pages_node[r_nid]++;
988 h->surplus_huge_pages_node[r_nid]++;
Adam Litke3b116302008-04-28 02:13:06 -0700989 __count_vm_event(HTLB_BUDDY_PGALLOC);
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800990 } else {
Andi Kleena5516432008-07-23 21:27:41 -0700991 h->nr_huge_pages--;
992 h->surplus_huge_pages--;
Adam Litke3b116302008-04-28 02:13:06 -0700993 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
Adam Litke7893d1d2007-10-16 01:26:18 -0700994 }
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800995 spin_unlock(&hugetlb_lock);
Adam Litke7893d1d2007-10-16 01:26:18 -0700996
997 return page;
998}
999
Adam Litkee4e574b2007-10-16 01:26:19 -07001000/*
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001001 * This allocation function is useful in the context where vma is irrelevant.
1002 * E.g. soft-offlining uses this function because it only cares physical
1003 * address of error page.
1004 */
1005struct page *alloc_huge_page_node(struct hstate *h, int nid)
1006{
Joonsoo Kim4ef91842013-09-11 14:21:51 -07001007 struct page *page = NULL;
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001008
1009 spin_lock(&hugetlb_lock);
Joonsoo Kim4ef91842013-09-11 14:21:51 -07001010 if (h->free_huge_pages - h->resv_huge_pages > 0)
1011 page = dequeue_huge_page_node(h, nid);
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001012 spin_unlock(&hugetlb_lock);
1013
Aneesh Kumar K.V94ae8ba2012-07-31 16:42:35 -07001014 if (!page)
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001015 page = alloc_buddy_huge_page(h, nid);
1016
1017 return page;
1018}
1019
1020/*
Lucas De Marchi25985ed2011-03-30 22:57:33 -03001021 * Increase the hugetlb pool such that it can accommodate a reservation
Adam Litkee4e574b2007-10-16 01:26:19 -07001022 * of size 'delta'.
1023 */
Andi Kleena5516432008-07-23 21:27:41 -07001024static int gather_surplus_pages(struct hstate *h, int delta)
Adam Litkee4e574b2007-10-16 01:26:19 -07001025{
1026 struct list_head surplus_list;
1027 struct page *page, *tmp;
1028 int ret, i;
1029 int needed, allocated;
Hillf Danton28073b02012-03-21 16:34:00 -07001030 bool alloc_ok = true;
Adam Litkee4e574b2007-10-16 01:26:19 -07001031
Andi Kleena5516432008-07-23 21:27:41 -07001032 needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
Adam Litkeac09b3a2008-03-04 14:29:38 -08001033 if (needed <= 0) {
Andi Kleena5516432008-07-23 21:27:41 -07001034 h->resv_huge_pages += delta;
Adam Litkee4e574b2007-10-16 01:26:19 -07001035 return 0;
Adam Litkeac09b3a2008-03-04 14:29:38 -08001036 }
Adam Litkee4e574b2007-10-16 01:26:19 -07001037
1038 allocated = 0;
1039 INIT_LIST_HEAD(&surplus_list);
1040
1041 ret = -ENOMEM;
1042retry:
1043 spin_unlock(&hugetlb_lock);
1044 for (i = 0; i < needed; i++) {
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001045 page = alloc_buddy_huge_page(h, NUMA_NO_NODE);
Hillf Danton28073b02012-03-21 16:34:00 -07001046 if (!page) {
1047 alloc_ok = false;
1048 break;
1049 }
Adam Litkee4e574b2007-10-16 01:26:19 -07001050 list_add(&page->lru, &surplus_list);
1051 }
Hillf Danton28073b02012-03-21 16:34:00 -07001052 allocated += i;
Adam Litkee4e574b2007-10-16 01:26:19 -07001053
1054 /*
1055 * After retaking hugetlb_lock, we need to recalculate 'needed'
1056 * because either resv_huge_pages or free_huge_pages may have changed.
1057 */
1058 spin_lock(&hugetlb_lock);
Andi Kleena5516432008-07-23 21:27:41 -07001059 needed = (h->resv_huge_pages + delta) -
1060 (h->free_huge_pages + allocated);
Hillf Danton28073b02012-03-21 16:34:00 -07001061 if (needed > 0) {
1062 if (alloc_ok)
1063 goto retry;
1064 /*
1065 * We were not able to allocate enough pages to
1066 * satisfy the entire reservation so we free what
1067 * we've allocated so far.
1068 */
1069 goto free;
1070 }
Adam Litkee4e574b2007-10-16 01:26:19 -07001071 /*
1072 * The surplus_list now contains _at_least_ the number of extra pages
Lucas De Marchi25985ed2011-03-30 22:57:33 -03001073 * needed to accommodate the reservation. Add the appropriate number
Adam Litkee4e574b2007-10-16 01:26:19 -07001074 * of pages to the hugetlb pool and free the extras back to the buddy
Adam Litkeac09b3a2008-03-04 14:29:38 -08001075 * allocator. Commit the entire reservation here to prevent another
1076 * process from stealing the pages as they are added to the pool but
1077 * before they are reserved.
Adam Litkee4e574b2007-10-16 01:26:19 -07001078 */
1079 needed += allocated;
Andi Kleena5516432008-07-23 21:27:41 -07001080 h->resv_huge_pages += delta;
Adam Litkee4e574b2007-10-16 01:26:19 -07001081 ret = 0;
Naoya Horiguchia9869b82010-09-08 10:19:37 +09001082
Adam Litke19fc3f02008-04-28 02:12:20 -07001083 /* Free the needed pages to the hugetlb pool */
Adam Litkee4e574b2007-10-16 01:26:19 -07001084 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
Adam Litke19fc3f02008-04-28 02:12:20 -07001085 if ((--needed) < 0)
1086 break;
Naoya Horiguchia9869b82010-09-08 10:19:37 +09001087 /*
1088 * This page is now managed by the hugetlb allocator and has
1089 * no users -- drop the buddy allocator's reference.
1090 */
1091 put_page_testzero(page);
Sasha Levin309381fea2014-01-23 15:52:54 -08001092 VM_BUG_ON_PAGE(page_count(page), page);
Andi Kleena5516432008-07-23 21:27:41 -07001093 enqueue_huge_page(h, page);
Adam Litke19fc3f02008-04-28 02:12:20 -07001094 }
Hillf Danton28073b02012-03-21 16:34:00 -07001095free:
Hillf Dantonb0365c82011-12-28 15:57:16 -08001096 spin_unlock(&hugetlb_lock);
Adam Litke19fc3f02008-04-28 02:12:20 -07001097
1098 /* Free unnecessary surplus pages to the buddy allocator */
Joonsoo Kimc0d934b2013-09-11 14:21:02 -07001099 list_for_each_entry_safe(page, tmp, &surplus_list, lru)
1100 put_page(page);
Naoya Horiguchia9869b82010-09-08 10:19:37 +09001101 spin_lock(&hugetlb_lock);
Adam Litkee4e574b2007-10-16 01:26:19 -07001102
1103 return ret;
1104}
1105
1106/*
1107 * When releasing a hugetlb pool reservation, any surplus pages that were
1108 * allocated to satisfy the reservation must be explicitly freed if they were
1109 * never used.
Lee Schermerhorn685f3452009-09-21 17:01:23 -07001110 * Called with hugetlb_lock held.
Adam Litkee4e574b2007-10-16 01:26:19 -07001111 */
Andi Kleena5516432008-07-23 21:27:41 -07001112static void return_unused_surplus_pages(struct hstate *h,
1113 unsigned long unused_resv_pages)
Adam Litkee4e574b2007-10-16 01:26:19 -07001114{
Adam Litkee4e574b2007-10-16 01:26:19 -07001115 unsigned long nr_pages;
1116
Adam Litkeac09b3a2008-03-04 14:29:38 -08001117 /* Uncommit the reservation */
Andi Kleena5516432008-07-23 21:27:41 -07001118 h->resv_huge_pages -= unused_resv_pages;
Adam Litkeac09b3a2008-03-04 14:29:38 -08001119
Andi Kleenaa888a72008-07-23 21:27:47 -07001120 /* Cannot return gigantic pages currently */
1121 if (h->order >= MAX_ORDER)
1122 return;
1123
Andi Kleena5516432008-07-23 21:27:41 -07001124 nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
Adam Litkee4e574b2007-10-16 01:26:19 -07001125
Lee Schermerhorn685f3452009-09-21 17:01:23 -07001126 /*
1127 * We want to release as many surplus pages as possible, spread
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -08001128 * evenly across all nodes with memory. Iterate across these nodes
1129 * until we can no longer free unreserved surplus pages. This occurs
1130 * when the nodes with surplus pages have no free pages.
1131 * free_pool_huge_page() will balance the the freed pages across the
1132 * on-line nodes with memory and will handle the hstate accounting.
Lee Schermerhorn685f3452009-09-21 17:01:23 -07001133 */
1134 while (nr_pages--) {
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08001135 if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1))
Lee Schermerhorn685f3452009-09-21 17:01:23 -07001136 break;
Adam Litkee4e574b2007-10-16 01:26:19 -07001137 }
1138}
1139
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001140/*
1141 * Determine if the huge page at addr within the vma has an associated
1142 * reservation. Where it does not we will need to logically increase
David Gibson90481622012-03-21 16:34:12 -07001143 * reservation and actually increase subpool usage before an allocation
1144 * can occur. Where any new reservation would be required the
1145 * reservation change is prepared, but not committed. Once the page
1146 * has been allocated from the subpool and instantiated the change should
1147 * be committed via vma_commit_reservation. No action is required on
1148 * failure.
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001149 */
Roel Kluine2f17d92009-03-31 15:23:15 -07001150static long vma_needs_reservation(struct hstate *h,
Andi Kleena5516432008-07-23 21:27:41 -07001151 struct vm_area_struct *vma, unsigned long addr)
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001152{
1153 struct address_space *mapping = vma->vm_file->f_mapping;
1154 struct inode *inode = mapping->host;
1155
Mel Gormanf83a2752009-05-28 14:34:40 -07001156 if (vma->vm_flags & VM_MAYSHARE) {
Andi Kleena5516432008-07-23 21:27:41 -07001157 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001158 return region_chg(&inode->i_mapping->private_list,
1159 idx, idx + 1);
1160
Andy Whitcroft84afd992008-07-23 21:27:32 -07001161 } else if (!is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
1162 return 1;
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001163
Andy Whitcroft84afd992008-07-23 21:27:32 -07001164 } else {
Roel Kluine2f17d92009-03-31 15:23:15 -07001165 long err;
Andi Kleena5516432008-07-23 21:27:41 -07001166 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
Joonsoo Kimf522c3a2013-09-11 14:21:53 -07001167 struct resv_map *resv = vma_resv_map(vma);
Andy Whitcroft84afd992008-07-23 21:27:32 -07001168
Joonsoo Kimf522c3a2013-09-11 14:21:53 -07001169 err = region_chg(&resv->regions, idx, idx + 1);
Andy Whitcroft84afd992008-07-23 21:27:32 -07001170 if (err < 0)
1171 return err;
1172 return 0;
1173 }
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001174}
Andi Kleena5516432008-07-23 21:27:41 -07001175static void vma_commit_reservation(struct hstate *h,
1176 struct vm_area_struct *vma, unsigned long addr)
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001177{
1178 struct address_space *mapping = vma->vm_file->f_mapping;
1179 struct inode *inode = mapping->host;
1180
Mel Gormanf83a2752009-05-28 14:34:40 -07001181 if (vma->vm_flags & VM_MAYSHARE) {
Andi Kleena5516432008-07-23 21:27:41 -07001182 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001183 region_add(&inode->i_mapping->private_list, idx, idx + 1);
Andy Whitcroft84afd992008-07-23 21:27:32 -07001184
1185 } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
Andi Kleena5516432008-07-23 21:27:41 -07001186 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
Joonsoo Kimf522c3a2013-09-11 14:21:53 -07001187 struct resv_map *resv = vma_resv_map(vma);
Andy Whitcroft84afd992008-07-23 21:27:32 -07001188
1189 /* Mark this page used in the map. */
Joonsoo Kimf522c3a2013-09-11 14:21:53 -07001190 region_add(&resv->regions, idx, idx + 1);
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001191 }
1192}
1193
David Gibson27a85ef2006-03-22 00:08:56 -08001194static struct page *alloc_huge_page(struct vm_area_struct *vma,
Mel Gorman04f2cbe2008-07-23 21:27:25 -07001195 unsigned long addr, int avoid_reserve)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001196{
David Gibson90481622012-03-21 16:34:12 -07001197 struct hugepage_subpool *spool = subpool_vma(vma);
Andi Kleena5516432008-07-23 21:27:41 -07001198 struct hstate *h = hstate_vma(vma);
Adam Litke348ea202007-11-14 16:59:37 -08001199 struct page *page;
Roel Kluine2f17d92009-03-31 15:23:15 -07001200 long chg;
Aneesh Kumar K.V6d76dcf2012-07-31 16:42:18 -07001201 int ret, idx;
1202 struct hugetlb_cgroup *h_cg;
Adam Litke2fc39ce2007-11-14 16:59:39 -08001203
Aneesh Kumar K.V6d76dcf2012-07-31 16:42:18 -07001204 idx = hstate_index(h);
Mel Gormana1e78772008-07-23 21:27:23 -07001205 /*
David Gibson90481622012-03-21 16:34:12 -07001206 * Processes that did not create the mapping will have no
1207 * reserves and will not have accounted against subpool
1208 * limit. Check that the subpool limit can be made before
1209 * satisfying the allocation MAP_NORESERVE mappings may also
1210 * need pages and subpool limit allocated allocated if no reserve
1211 * mapping overlaps.
Mel Gormana1e78772008-07-23 21:27:23 -07001212 */
Andi Kleena5516432008-07-23 21:27:41 -07001213 chg = vma_needs_reservation(h, vma, addr);
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001214 if (chg < 0)
Aneesh Kumar K.V76dcee72012-07-31 16:41:57 -07001215 return ERR_PTR(-ENOMEM);
Joonsoo Kim8bb3f122013-09-11 14:21:54 -07001216 if (chg || avoid_reserve)
1217 if (hugepage_subpool_get_pages(spool, 1))
Aneesh Kumar K.V76dcee72012-07-31 16:41:57 -07001218 return ERR_PTR(-ENOSPC);
Mel Gormana1e78772008-07-23 21:27:23 -07001219
Aneesh Kumar K.V6d76dcf2012-07-31 16:42:18 -07001220 ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg);
1221 if (ret) {
Joonsoo Kim8bb3f122013-09-11 14:21:54 -07001222 if (chg || avoid_reserve)
1223 hugepage_subpool_put_pages(spool, 1);
Aneesh Kumar K.V6d76dcf2012-07-31 16:42:18 -07001224 return ERR_PTR(-ENOSPC);
1225 }
Mel Gormana1e78772008-07-23 21:27:23 -07001226 spin_lock(&hugetlb_lock);
Joonsoo Kimaf0ed732013-09-11 14:21:18 -07001227 page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, chg);
Joonsoo Kim81a6fca2013-09-11 14:20:58 -07001228 if (!page) {
Aneesh Kumar K.V94ae8ba2012-07-31 16:42:35 -07001229 spin_unlock(&hugetlb_lock);
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001230 page = alloc_buddy_huge_page(h, NUMA_NO_NODE);
Mel Gormana1e78772008-07-23 21:27:23 -07001231 if (!page) {
Aneesh Kumar K.V6d76dcf2012-07-31 16:42:18 -07001232 hugetlb_cgroup_uncharge_cgroup(idx,
1233 pages_per_huge_page(h),
1234 h_cg);
Joonsoo Kim8bb3f122013-09-11 14:21:54 -07001235 if (chg || avoid_reserve)
1236 hugepage_subpool_put_pages(spool, 1);
Aneesh Kumar K.V76dcee72012-07-31 16:41:57 -07001237 return ERR_PTR(-ENOSPC);
Mel Gormana1e78772008-07-23 21:27:23 -07001238 }
Aneesh Kumar K.V79dbb232012-07-31 16:42:32 -07001239 spin_lock(&hugetlb_lock);
1240 list_move(&page->lru, &h->hugepage_activelist);
Joonsoo Kim81a6fca2013-09-11 14:20:58 -07001241 /* Fall through */
Mel Gormana1e78772008-07-23 21:27:23 -07001242 }
Joonsoo Kim81a6fca2013-09-11 14:20:58 -07001243 hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page);
1244 spin_unlock(&hugetlb_lock);
Mel Gormana1e78772008-07-23 21:27:23 -07001245
David Gibson90481622012-03-21 16:34:12 -07001246 set_page_private(page, (unsigned long)spool);
Mel Gormana1e78772008-07-23 21:27:23 -07001247
Andi Kleena5516432008-07-23 21:27:41 -07001248 vma_commit_reservation(h, vma, addr);
Adam Litke90d8b7e2007-11-14 16:59:42 -08001249 return page;
David Gibsonb45b5bd2006-03-22 00:08:55 -08001250}
1251
Naoya Horiguchi74060e42013-09-11 14:22:06 -07001252/*
1253 * alloc_huge_page()'s wrapper which simply returns the page if allocation
1254 * succeeds, otherwise NULL. This function is called from new_vma_page(),
1255 * where no ERR_VALUE is expected to be returned.
1256 */
1257struct page *alloc_huge_page_noerr(struct vm_area_struct *vma,
1258 unsigned long addr, int avoid_reserve)
1259{
1260 struct page *page = alloc_huge_page(vma, addr, avoid_reserve);
1261 if (IS_ERR(page))
1262 page = NULL;
1263 return page;
1264}
1265
Cyrill Gorcunov91f47662009-01-06 14:40:33 -08001266int __weak alloc_bootmem_huge_page(struct hstate *h)
Andi Kleenaa888a72008-07-23 21:27:47 -07001267{
1268 struct huge_bootmem_page *m;
Joonsoo Kimb2261022013-09-11 14:21:00 -07001269 int nr_nodes, node;
Andi Kleenaa888a72008-07-23 21:27:47 -07001270
Joonsoo Kimb2261022013-09-11 14:21:00 -07001271 for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) {
Andi Kleenaa888a72008-07-23 21:27:47 -07001272 void *addr;
1273
Grygorii Strashko8b89a112014-01-21 15:50:36 -08001274 addr = memblock_virt_alloc_try_nid_nopanic(
1275 huge_page_size(h), huge_page_size(h),
1276 0, BOOTMEM_ALLOC_ACCESSIBLE, node);
Andi Kleenaa888a72008-07-23 21:27:47 -07001277 if (addr) {
1278 /*
1279 * Use the beginning of the huge page to store the
1280 * huge_bootmem_page struct (until gather_bootmem
1281 * puts them into the mem_map).
1282 */
1283 m = addr;
Cyrill Gorcunov91f47662009-01-06 14:40:33 -08001284 goto found;
Andi Kleenaa888a72008-07-23 21:27:47 -07001285 }
Andi Kleenaa888a72008-07-23 21:27:47 -07001286 }
1287 return 0;
1288
1289found:
1290 BUG_ON((unsigned long)virt_to_phys(m) & (huge_page_size(h) - 1));
1291 /* Put them into a private list first because mem_map is not up yet */
1292 list_add(&m->list, &huge_boot_pages);
1293 m->hstate = h;
1294 return 1;
1295}
1296
Andy Whitcroft18229df2008-11-06 12:53:27 -08001297static void prep_compound_huge_page(struct page *page, int order)
1298{
1299 if (unlikely(order > (MAX_ORDER - 1)))
1300 prep_compound_gigantic_page(page, order);
1301 else
1302 prep_compound_page(page, order);
1303}
1304
Andi Kleenaa888a72008-07-23 21:27:47 -07001305/* Put bootmem huge pages into the standard lists after mem_map is up */
1306static void __init gather_bootmem_prealloc(void)
1307{
1308 struct huge_bootmem_page *m;
1309
1310 list_for_each_entry(m, &huge_boot_pages, list) {
Andi Kleenaa888a72008-07-23 21:27:47 -07001311 struct hstate *h = m->hstate;
Becky Bruceee8f2482011-07-25 17:11:50 -07001312 struct page *page;
1313
1314#ifdef CONFIG_HIGHMEM
1315 page = pfn_to_page(m->phys >> PAGE_SHIFT);
Grygorii Strashko8b89a112014-01-21 15:50:36 -08001316 memblock_free_late(__pa(m),
1317 sizeof(struct huge_bootmem_page));
Becky Bruceee8f2482011-07-25 17:11:50 -07001318#else
1319 page = virt_to_page(m);
1320#endif
Andi Kleenaa888a72008-07-23 21:27:47 -07001321 WARN_ON(page_count(page) != 1);
Andy Whitcroft18229df2008-11-06 12:53:27 -08001322 prep_compound_huge_page(page, h->order);
Andrea Arcangelief5a22b2013-10-16 13:46:56 -07001323 WARN_ON(PageReserved(page));
Andi Kleenaa888a72008-07-23 21:27:47 -07001324 prep_new_huge_page(h, page, page_to_nid(page));
Rafael Aquinib0320c72011-06-15 15:08:39 -07001325 /*
1326 * If we had gigantic hugepages allocated at boot time, we need
1327 * to restore the 'stolen' pages to totalram_pages in order to
1328 * fix confusing memory reports from free(1) and another
1329 * side-effects, like CommitLimit going negative.
1330 */
1331 if (h->order > (MAX_ORDER - 1))
Jiang Liu3dcc0572013-07-03 15:03:21 -07001332 adjust_managed_page_count(page, 1 << h->order);
Andi Kleenaa888a72008-07-23 21:27:47 -07001333 }
1334}
1335
Andi Kleen8faa8b02008-07-23 21:27:48 -07001336static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001337{
1338 unsigned long i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001339
Andi Kleene5ff2152008-07-23 21:27:42 -07001340 for (i = 0; i < h->max_huge_pages; ++i) {
Andi Kleenaa888a72008-07-23 21:27:47 -07001341 if (h->order >= MAX_ORDER) {
1342 if (!alloc_bootmem_huge_page(h))
1343 break;
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -08001344 } else if (!alloc_fresh_huge_page(h,
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08001345 &node_states[N_MEMORY]))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001346 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001347 }
Andi Kleen8faa8b02008-07-23 21:27:48 -07001348 h->max_huge_pages = i;
Andi Kleene5ff2152008-07-23 21:27:42 -07001349}
1350
1351static void __init hugetlb_init_hstates(void)
1352{
1353 struct hstate *h;
1354
1355 for_each_hstate(h) {
Andi Kleen8faa8b02008-07-23 21:27:48 -07001356 /* oversize hugepages were init'ed in early boot */
1357 if (h->order < MAX_ORDER)
1358 hugetlb_hstate_alloc_pages(h);
Andi Kleene5ff2152008-07-23 21:27:42 -07001359 }
1360}
1361
Andi Kleen4abd32d2008-07-23 21:27:49 -07001362static char * __init memfmt(char *buf, unsigned long n)
1363{
1364 if (n >= (1UL << 30))
1365 sprintf(buf, "%lu GB", n >> 30);
1366 else if (n >= (1UL << 20))
1367 sprintf(buf, "%lu MB", n >> 20);
1368 else
1369 sprintf(buf, "%lu KB", n >> 10);
1370 return buf;
1371}
1372
Andi Kleene5ff2152008-07-23 21:27:42 -07001373static void __init report_hugepages(void)
1374{
1375 struct hstate *h;
1376
1377 for_each_hstate(h) {
Andi Kleen4abd32d2008-07-23 21:27:49 -07001378 char buf[32];
Andrew Mortonffb22af2013-02-22 16:32:08 -08001379 pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n",
Andi Kleen4abd32d2008-07-23 21:27:49 -07001380 memfmt(buf, huge_page_size(h)),
1381 h->free_huge_pages);
Andi Kleene5ff2152008-07-23 21:27:42 -07001382 }
1383}
1384
Linus Torvalds1da177e2005-04-16 15:20:36 -07001385#ifdef CONFIG_HIGHMEM
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001386static void try_to_free_low(struct hstate *h, unsigned long count,
1387 nodemask_t *nodes_allowed)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001388{
Christoph Lameter4415cc82006-09-25 23:31:55 -07001389 int i;
1390
Andi Kleenaa888a72008-07-23 21:27:47 -07001391 if (h->order >= MAX_ORDER)
1392 return;
1393
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001394 for_each_node_mask(i, *nodes_allowed) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001395 struct page *page, *next;
Andi Kleena5516432008-07-23 21:27:41 -07001396 struct list_head *freel = &h->hugepage_freelists[i];
1397 list_for_each_entry_safe(page, next, freel, lru) {
1398 if (count >= h->nr_huge_pages)
Adam Litke6b0c8802007-10-16 01:26:23 -07001399 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001400 if (PageHighMem(page))
1401 continue;
1402 list_del(&page->lru);
Andi Kleene5ff2152008-07-23 21:27:42 -07001403 update_and_free_page(h, page);
Andi Kleena5516432008-07-23 21:27:41 -07001404 h->free_huge_pages--;
1405 h->free_huge_pages_node[page_to_nid(page)]--;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001406 }
1407 }
1408}
1409#else
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001410static inline void try_to_free_low(struct hstate *h, unsigned long count,
1411 nodemask_t *nodes_allowed)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001412{
1413}
1414#endif
1415
Wu Fengguang20a03072009-06-16 15:32:22 -07001416/*
1417 * Increment or decrement surplus_huge_pages. Keep node-specific counters
1418 * balanced by operating on them in a round-robin fashion.
1419 * Returns 1 if an adjustment was made.
1420 */
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001421static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
1422 int delta)
Wu Fengguang20a03072009-06-16 15:32:22 -07001423{
Joonsoo Kimb2261022013-09-11 14:21:00 -07001424 int nr_nodes, node;
Wu Fengguang20a03072009-06-16 15:32:22 -07001425
1426 VM_BUG_ON(delta != -1 && delta != 1);
Wu Fengguang20a03072009-06-16 15:32:22 -07001427
Joonsoo Kimb2261022013-09-11 14:21:00 -07001428 if (delta < 0) {
1429 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
1430 if (h->surplus_huge_pages_node[node])
1431 goto found;
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001432 }
Joonsoo Kimb2261022013-09-11 14:21:00 -07001433 } else {
1434 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
1435 if (h->surplus_huge_pages_node[node] <
1436 h->nr_huge_pages_node[node])
1437 goto found;
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001438 }
Joonsoo Kimb2261022013-09-11 14:21:00 -07001439 }
1440 return 0;
Wu Fengguang20a03072009-06-16 15:32:22 -07001441
Joonsoo Kimb2261022013-09-11 14:21:00 -07001442found:
1443 h->surplus_huge_pages += delta;
1444 h->surplus_huge_pages_node[node] += delta;
1445 return 1;
Wu Fengguang20a03072009-06-16 15:32:22 -07001446}
1447
Andi Kleena5516432008-07-23 21:27:41 -07001448#define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001449static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count,
1450 nodemask_t *nodes_allowed)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001451{
Adam Litke7893d1d2007-10-16 01:26:18 -07001452 unsigned long min_count, ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001453
Andi Kleenaa888a72008-07-23 21:27:47 -07001454 if (h->order >= MAX_ORDER)
1455 return h->max_huge_pages;
1456
Adam Litke7893d1d2007-10-16 01:26:18 -07001457 /*
1458 * Increase the pool size
1459 * First take pages out of surplus state. Then make up the
1460 * remaining difference by allocating fresh huge pages.
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001461 *
1462 * We might race with alloc_buddy_huge_page() here and be unable
1463 * to convert a surplus huge page to a normal huge page. That is
1464 * not critical, though, it just means the overall size of the
1465 * pool might be one hugepage larger than it needs to be, but
1466 * within all the constraints specified by the sysctls.
Adam Litke7893d1d2007-10-16 01:26:18 -07001467 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001468 spin_lock(&hugetlb_lock);
Andi Kleena5516432008-07-23 21:27:41 -07001469 while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001470 if (!adjust_pool_surplus(h, nodes_allowed, -1))
Adam Litke7893d1d2007-10-16 01:26:18 -07001471 break;
1472 }
1473
Andi Kleena5516432008-07-23 21:27:41 -07001474 while (count > persistent_huge_pages(h)) {
Adam Litke7893d1d2007-10-16 01:26:18 -07001475 /*
1476 * If this allocation races such that we no longer need the
1477 * page, free_huge_page will handle it by freeing the page
1478 * and reducing the surplus.
1479 */
1480 spin_unlock(&hugetlb_lock);
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001481 ret = alloc_fresh_huge_page(h, nodes_allowed);
Adam Litke7893d1d2007-10-16 01:26:18 -07001482 spin_lock(&hugetlb_lock);
1483 if (!ret)
1484 goto out;
1485
Mel Gorman536240f22009-12-14 17:59:56 -08001486 /* Bail for signals. Probably ctrl-c from user */
1487 if (signal_pending(current))
1488 goto out;
Adam Litke7893d1d2007-10-16 01:26:18 -07001489 }
Adam Litke7893d1d2007-10-16 01:26:18 -07001490
1491 /*
1492 * Decrease the pool size
1493 * First return free pages to the buddy allocator (being careful
1494 * to keep enough around to satisfy reservations). Then place
1495 * pages into surplus state as needed so the pool will shrink
1496 * to the desired size as pages become free.
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001497 *
1498 * By placing pages into the surplus state independent of the
1499 * overcommit value, we are allowing the surplus pool size to
1500 * exceed overcommit. There are few sane options here. Since
1501 * alloc_buddy_huge_page() is checking the global counter,
1502 * though, we'll note that we're not allowed to exceed surplus
1503 * and won't grow the pool anywhere else. Not until one of the
1504 * sysctls are changed, or the surplus pages go out of use.
Adam Litke7893d1d2007-10-16 01:26:18 -07001505 */
Andi Kleena5516432008-07-23 21:27:41 -07001506 min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
Adam Litke6b0c8802007-10-16 01:26:23 -07001507 min_count = max(count, min_count);
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001508 try_to_free_low(h, min_count, nodes_allowed);
Andi Kleena5516432008-07-23 21:27:41 -07001509 while (min_count < persistent_huge_pages(h)) {
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001510 if (!free_pool_huge_page(h, nodes_allowed, 0))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001511 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001512 }
Andi Kleena5516432008-07-23 21:27:41 -07001513 while (count < persistent_huge_pages(h)) {
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001514 if (!adjust_pool_surplus(h, nodes_allowed, 1))
Adam Litke7893d1d2007-10-16 01:26:18 -07001515 break;
1516 }
1517out:
Andi Kleena5516432008-07-23 21:27:41 -07001518 ret = persistent_huge_pages(h);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001519 spin_unlock(&hugetlb_lock);
Adam Litke7893d1d2007-10-16 01:26:18 -07001520 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001521}
1522
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001523#define HSTATE_ATTR_RO(_name) \
1524 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
1525
1526#define HSTATE_ATTR(_name) \
1527 static struct kobj_attribute _name##_attr = \
1528 __ATTR(_name, 0644, _name##_show, _name##_store)
1529
1530static struct kobject *hugepages_kobj;
1531static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
1532
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001533static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
1534
1535static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001536{
1537 int i;
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001538
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001539 for (i = 0; i < HUGE_MAX_HSTATE; i++)
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001540 if (hstate_kobjs[i] == kobj) {
1541 if (nidp)
1542 *nidp = NUMA_NO_NODE;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001543 return &hstates[i];
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001544 }
1545
1546 return kobj_to_node_hstate(kobj, nidp);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001547}
1548
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001549static ssize_t nr_hugepages_show_common(struct kobject *kobj,
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001550 struct kobj_attribute *attr, char *buf)
1551{
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001552 struct hstate *h;
1553 unsigned long nr_huge_pages;
1554 int nid;
1555
1556 h = kobj_to_hstate(kobj, &nid);
1557 if (nid == NUMA_NO_NODE)
1558 nr_huge_pages = h->nr_huge_pages;
1559 else
1560 nr_huge_pages = h->nr_huge_pages_node[nid];
1561
1562 return sprintf(buf, "%lu\n", nr_huge_pages);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001563}
Eric B Munsonadbe8722011-01-13 15:47:27 -08001564
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001565static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
1566 struct kobject *kobj, struct kobj_attribute *attr,
1567 const char *buf, size_t len)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001568{
1569 int err;
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001570 int nid;
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001571 unsigned long count;
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001572 struct hstate *h;
David Rientjesbad44b52009-12-14 17:58:38 -08001573 NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001574
Jingoo Han3dbb95f2013-09-11 14:20:25 -07001575 err = kstrtoul(buf, 10, &count);
Eric B Munson73ae31e2011-01-13 15:47:28 -08001576 if (err)
Eric B Munsonadbe8722011-01-13 15:47:27 -08001577 goto out;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001578
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001579 h = kobj_to_hstate(kobj, &nid);
Eric B Munsonadbe8722011-01-13 15:47:27 -08001580 if (h->order >= MAX_ORDER) {
1581 err = -EINVAL;
1582 goto out;
1583 }
1584
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001585 if (nid == NUMA_NO_NODE) {
1586 /*
1587 * global hstate attribute
1588 */
1589 if (!(obey_mempolicy &&
1590 init_nodemask_of_mempolicy(nodes_allowed))) {
1591 NODEMASK_FREE(nodes_allowed);
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08001592 nodes_allowed = &node_states[N_MEMORY];
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001593 }
1594 } else if (nodes_allowed) {
1595 /*
1596 * per node hstate attribute: adjust count to global,
1597 * but restrict alloc/free to the specified node.
1598 */
1599 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
1600 init_nodemask_of_node(nodes_allowed, nid);
1601 } else
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08001602 nodes_allowed = &node_states[N_MEMORY];
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001603
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001604 h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001605
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08001606 if (nodes_allowed != &node_states[N_MEMORY])
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001607 NODEMASK_FREE(nodes_allowed);
1608
1609 return len;
Eric B Munsonadbe8722011-01-13 15:47:27 -08001610out:
1611 NODEMASK_FREE(nodes_allowed);
1612 return err;
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001613}
1614
1615static ssize_t nr_hugepages_show(struct kobject *kobj,
1616 struct kobj_attribute *attr, char *buf)
1617{
1618 return nr_hugepages_show_common(kobj, attr, buf);
1619}
1620
1621static ssize_t nr_hugepages_store(struct kobject *kobj,
1622 struct kobj_attribute *attr, const char *buf, size_t len)
1623{
1624 return nr_hugepages_store_common(false, kobj, attr, buf, len);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001625}
1626HSTATE_ATTR(nr_hugepages);
1627
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001628#ifdef CONFIG_NUMA
1629
1630/*
1631 * hstate attribute for optionally mempolicy-based constraint on persistent
1632 * huge page alloc/free.
1633 */
1634static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
1635 struct kobj_attribute *attr, char *buf)
1636{
1637 return nr_hugepages_show_common(kobj, attr, buf);
1638}
1639
1640static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
1641 struct kobj_attribute *attr, const char *buf, size_t len)
1642{
1643 return nr_hugepages_store_common(true, kobj, attr, buf, len);
1644}
1645HSTATE_ATTR(nr_hugepages_mempolicy);
1646#endif
1647
1648
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001649static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
1650 struct kobj_attribute *attr, char *buf)
1651{
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001652 struct hstate *h = kobj_to_hstate(kobj, NULL);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001653 return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
1654}
Eric B Munsonadbe8722011-01-13 15:47:27 -08001655
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001656static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
1657 struct kobj_attribute *attr, const char *buf, size_t count)
1658{
1659 int err;
1660 unsigned long input;
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001661 struct hstate *h = kobj_to_hstate(kobj, NULL);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001662
Eric B Munsonadbe8722011-01-13 15:47:27 -08001663 if (h->order >= MAX_ORDER)
1664 return -EINVAL;
1665
Jingoo Han3dbb95f2013-09-11 14:20:25 -07001666 err = kstrtoul(buf, 10, &input);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001667 if (err)
Eric B Munson73ae31e2011-01-13 15:47:28 -08001668 return err;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001669
1670 spin_lock(&hugetlb_lock);
1671 h->nr_overcommit_huge_pages = input;
1672 spin_unlock(&hugetlb_lock);
1673
1674 return count;
1675}
1676HSTATE_ATTR(nr_overcommit_hugepages);
1677
1678static ssize_t free_hugepages_show(struct kobject *kobj,
1679 struct kobj_attribute *attr, char *buf)
1680{
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001681 struct hstate *h;
1682 unsigned long free_huge_pages;
1683 int nid;
1684
1685 h = kobj_to_hstate(kobj, &nid);
1686 if (nid == NUMA_NO_NODE)
1687 free_huge_pages = h->free_huge_pages;
1688 else
1689 free_huge_pages = h->free_huge_pages_node[nid];
1690
1691 return sprintf(buf, "%lu\n", free_huge_pages);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001692}
1693HSTATE_ATTR_RO(free_hugepages);
1694
1695static ssize_t resv_hugepages_show(struct kobject *kobj,
1696 struct kobj_attribute *attr, char *buf)
1697{
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001698 struct hstate *h = kobj_to_hstate(kobj, NULL);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001699 return sprintf(buf, "%lu\n", h->resv_huge_pages);
1700}
1701HSTATE_ATTR_RO(resv_hugepages);
1702
1703static ssize_t surplus_hugepages_show(struct kobject *kobj,
1704 struct kobj_attribute *attr, char *buf)
1705{
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001706 struct hstate *h;
1707 unsigned long surplus_huge_pages;
1708 int nid;
1709
1710 h = kobj_to_hstate(kobj, &nid);
1711 if (nid == NUMA_NO_NODE)
1712 surplus_huge_pages = h->surplus_huge_pages;
1713 else
1714 surplus_huge_pages = h->surplus_huge_pages_node[nid];
1715
1716 return sprintf(buf, "%lu\n", surplus_huge_pages);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001717}
1718HSTATE_ATTR_RO(surplus_hugepages);
1719
1720static struct attribute *hstate_attrs[] = {
1721 &nr_hugepages_attr.attr,
1722 &nr_overcommit_hugepages_attr.attr,
1723 &free_hugepages_attr.attr,
1724 &resv_hugepages_attr.attr,
1725 &surplus_hugepages_attr.attr,
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001726#ifdef CONFIG_NUMA
1727 &nr_hugepages_mempolicy_attr.attr,
1728#endif
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001729 NULL,
1730};
1731
1732static struct attribute_group hstate_attr_group = {
1733 .attrs = hstate_attrs,
1734};
1735
Jeff Mahoney094e9532010-02-02 13:44:14 -08001736static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
1737 struct kobject **hstate_kobjs,
1738 struct attribute_group *hstate_attr_group)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001739{
1740 int retval;
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07001741 int hi = hstate_index(h);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001742
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001743 hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
1744 if (!hstate_kobjs[hi])
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001745 return -ENOMEM;
1746
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001747 retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001748 if (retval)
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001749 kobject_put(hstate_kobjs[hi]);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001750
1751 return retval;
1752}
1753
1754static void __init hugetlb_sysfs_init(void)
1755{
1756 struct hstate *h;
1757 int err;
1758
1759 hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
1760 if (!hugepages_kobj)
1761 return;
1762
1763 for_each_hstate(h) {
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001764 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
1765 hstate_kobjs, &hstate_attr_group);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001766 if (err)
Andrew Mortonffb22af2013-02-22 16:32:08 -08001767 pr_err("Hugetlb: Unable to add hstate %s", h->name);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001768 }
1769}
1770
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001771#ifdef CONFIG_NUMA
1772
1773/*
1774 * node_hstate/s - associate per node hstate attributes, via their kobjects,
Kay Sievers10fbcf42011-12-21 14:48:43 -08001775 * with node devices in node_devices[] using a parallel array. The array
1776 * index of a node device or _hstate == node id.
1777 * This is here to avoid any static dependency of the node device driver, in
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001778 * the base kernel, on the hugetlb module.
1779 */
1780struct node_hstate {
1781 struct kobject *hugepages_kobj;
1782 struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
1783};
1784struct node_hstate node_hstates[MAX_NUMNODES];
1785
1786/*
Kay Sievers10fbcf42011-12-21 14:48:43 -08001787 * A subset of global hstate attributes for node devices
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001788 */
1789static struct attribute *per_node_hstate_attrs[] = {
1790 &nr_hugepages_attr.attr,
1791 &free_hugepages_attr.attr,
1792 &surplus_hugepages_attr.attr,
1793 NULL,
1794};
1795
1796static struct attribute_group per_node_hstate_attr_group = {
1797 .attrs = per_node_hstate_attrs,
1798};
1799
1800/*
Kay Sievers10fbcf42011-12-21 14:48:43 -08001801 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001802 * Returns node id via non-NULL nidp.
1803 */
1804static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
1805{
1806 int nid;
1807
1808 for (nid = 0; nid < nr_node_ids; nid++) {
1809 struct node_hstate *nhs = &node_hstates[nid];
1810 int i;
1811 for (i = 0; i < HUGE_MAX_HSTATE; i++)
1812 if (nhs->hstate_kobjs[i] == kobj) {
1813 if (nidp)
1814 *nidp = nid;
1815 return &hstates[i];
1816 }
1817 }
1818
1819 BUG();
1820 return NULL;
1821}
1822
1823/*
Kay Sievers10fbcf42011-12-21 14:48:43 -08001824 * Unregister hstate attributes from a single node device.
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001825 * No-op if no hstate attributes attached.
1826 */
Claudiu Ghioc3cd8b442013-03-04 12:46:15 +02001827static void hugetlb_unregister_node(struct node *node)
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001828{
1829 struct hstate *h;
Kay Sievers10fbcf42011-12-21 14:48:43 -08001830 struct node_hstate *nhs = &node_hstates[node->dev.id];
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001831
1832 if (!nhs->hugepages_kobj)
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -08001833 return; /* no hstate attributes */
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001834
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07001835 for_each_hstate(h) {
1836 int idx = hstate_index(h);
1837 if (nhs->hstate_kobjs[idx]) {
1838 kobject_put(nhs->hstate_kobjs[idx]);
1839 nhs->hstate_kobjs[idx] = NULL;
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001840 }
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07001841 }
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001842
1843 kobject_put(nhs->hugepages_kobj);
1844 nhs->hugepages_kobj = NULL;
1845}
1846
1847/*
Kay Sievers10fbcf42011-12-21 14:48:43 -08001848 * hugetlb module exit: unregister hstate attributes from node devices
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001849 * that have them.
1850 */
1851static void hugetlb_unregister_all_nodes(void)
1852{
1853 int nid;
1854
1855 /*
Kay Sievers10fbcf42011-12-21 14:48:43 -08001856 * disable node device registrations.
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001857 */
1858 register_hugetlbfs_with_node(NULL, NULL);
1859
1860 /*
1861 * remove hstate attributes from any nodes that have them.
1862 */
1863 for (nid = 0; nid < nr_node_ids; nid++)
Wen Congyang87327942012-12-11 16:00:56 -08001864 hugetlb_unregister_node(node_devices[nid]);
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001865}
1866
1867/*
Kay Sievers10fbcf42011-12-21 14:48:43 -08001868 * Register hstate attributes for a single node device.
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001869 * No-op if attributes already registered.
1870 */
Claudiu Ghioc3cd8b442013-03-04 12:46:15 +02001871static void hugetlb_register_node(struct node *node)
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001872{
1873 struct hstate *h;
Kay Sievers10fbcf42011-12-21 14:48:43 -08001874 struct node_hstate *nhs = &node_hstates[node->dev.id];
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001875 int err;
1876
1877 if (nhs->hugepages_kobj)
1878 return; /* already allocated */
1879
1880 nhs->hugepages_kobj = kobject_create_and_add("hugepages",
Kay Sievers10fbcf42011-12-21 14:48:43 -08001881 &node->dev.kobj);
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001882 if (!nhs->hugepages_kobj)
1883 return;
1884
1885 for_each_hstate(h) {
1886 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
1887 nhs->hstate_kobjs,
1888 &per_node_hstate_attr_group);
1889 if (err) {
Andrew Mortonffb22af2013-02-22 16:32:08 -08001890 pr_err("Hugetlb: Unable to add hstate %s for node %d\n",
1891 h->name, node->dev.id);
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001892 hugetlb_unregister_node(node);
1893 break;
1894 }
1895 }
1896}
1897
1898/*
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -08001899 * hugetlb init time: register hstate attributes for all registered node
Kay Sievers10fbcf42011-12-21 14:48:43 -08001900 * devices of nodes that have memory. All on-line nodes should have
1901 * registered their associated device by this time.
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001902 */
1903static void hugetlb_register_all_nodes(void)
1904{
1905 int nid;
1906
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08001907 for_each_node_state(nid, N_MEMORY) {
Wen Congyang87327942012-12-11 16:00:56 -08001908 struct node *node = node_devices[nid];
Kay Sievers10fbcf42011-12-21 14:48:43 -08001909 if (node->dev.id == nid)
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001910 hugetlb_register_node(node);
1911 }
1912
1913 /*
Kay Sievers10fbcf42011-12-21 14:48:43 -08001914 * Let the node device driver know we're here so it can
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001915 * [un]register hstate attributes on node hotplug.
1916 */
1917 register_hugetlbfs_with_node(hugetlb_register_node,
1918 hugetlb_unregister_node);
1919}
1920#else /* !CONFIG_NUMA */
1921
1922static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
1923{
1924 BUG();
1925 if (nidp)
1926 *nidp = -1;
1927 return NULL;
1928}
1929
1930static void hugetlb_unregister_all_nodes(void) { }
1931
1932static void hugetlb_register_all_nodes(void) { }
1933
1934#endif
1935
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001936static void __exit hugetlb_exit(void)
1937{
1938 struct hstate *h;
1939
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001940 hugetlb_unregister_all_nodes();
1941
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001942 for_each_hstate(h) {
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07001943 kobject_put(hstate_kobjs[hstate_index(h)]);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001944 }
1945
1946 kobject_put(hugepages_kobj);
1947}
1948module_exit(hugetlb_exit);
1949
1950static int __init hugetlb_init(void)
1951{
Benjamin Herrenschmidt0ef89d22008-07-31 00:07:30 -07001952 /* Some platform decide whether they support huge pages at boot
1953 * time. On these, such as powerpc, HPAGE_SHIFT is set to 0 when
1954 * there is no such support
1955 */
1956 if (HPAGE_SHIFT == 0)
1957 return 0;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001958
Nick Piggine11bfbf2008-07-23 21:27:52 -07001959 if (!size_to_hstate(default_hstate_size)) {
1960 default_hstate_size = HPAGE_SIZE;
1961 if (!size_to_hstate(default_hstate_size))
1962 hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001963 }
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07001964 default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size));
Nick Piggine11bfbf2008-07-23 21:27:52 -07001965 if (default_hstate_max_huge_pages)
1966 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001967
1968 hugetlb_init_hstates();
Andi Kleenaa888a72008-07-23 21:27:47 -07001969 gather_bootmem_prealloc();
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001970 report_hugepages();
1971
1972 hugetlb_sysfs_init();
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001973 hugetlb_register_all_nodes();
Jianguo Wu7179e7b2012-12-18 14:23:19 -08001974 hugetlb_cgroup_file_init();
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001975
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001976 return 0;
1977}
1978module_init(hugetlb_init);
1979
1980/* Should be called on processing a hugepagesz=... option */
1981void __init hugetlb_add_hstate(unsigned order)
1982{
1983 struct hstate *h;
Andi Kleen8faa8b02008-07-23 21:27:48 -07001984 unsigned long i;
1985
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001986 if (size_to_hstate(PAGE_SIZE << order)) {
Andrew Mortonffb22af2013-02-22 16:32:08 -08001987 pr_warning("hugepagesz= specified twice, ignoring\n");
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001988 return;
1989 }
Aneesh Kumar K.V47d38342012-07-31 16:41:54 -07001990 BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001991 BUG_ON(order == 0);
Aneesh Kumar K.V47d38342012-07-31 16:41:54 -07001992 h = &hstates[hugetlb_max_hstate++];
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001993 h->order = order;
1994 h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
Andi Kleen8faa8b02008-07-23 21:27:48 -07001995 h->nr_huge_pages = 0;
1996 h->free_huge_pages = 0;
1997 for (i = 0; i < MAX_NUMNODES; ++i)
1998 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
Aneesh Kumar K.V0edaecf2012-07-31 16:42:07 -07001999 INIT_LIST_HEAD(&h->hugepage_activelist);
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08002000 h->next_nid_to_alloc = first_node(node_states[N_MEMORY]);
2001 h->next_nid_to_free = first_node(node_states[N_MEMORY]);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002002 snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
2003 huge_page_size(h)/1024);
Andi Kleen8faa8b02008-07-23 21:27:48 -07002004
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002005 parsed_hstate = h;
2006}
2007
Nick Piggine11bfbf2008-07-23 21:27:52 -07002008static int __init hugetlb_nrpages_setup(char *s)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002009{
2010 unsigned long *mhp;
Andi Kleen8faa8b02008-07-23 21:27:48 -07002011 static unsigned long *last_mhp;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002012
2013 /*
Aneesh Kumar K.V47d38342012-07-31 16:41:54 -07002014 * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet,
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002015 * so this hugepages= parameter goes to the "default hstate".
2016 */
Aneesh Kumar K.V47d38342012-07-31 16:41:54 -07002017 if (!hugetlb_max_hstate)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002018 mhp = &default_hstate_max_huge_pages;
2019 else
2020 mhp = &parsed_hstate->max_huge_pages;
2021
Andi Kleen8faa8b02008-07-23 21:27:48 -07002022 if (mhp == last_mhp) {
Andrew Mortonffb22af2013-02-22 16:32:08 -08002023 pr_warning("hugepages= specified twice without "
2024 "interleaving hugepagesz=, ignoring\n");
Andi Kleen8faa8b02008-07-23 21:27:48 -07002025 return 1;
2026 }
2027
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002028 if (sscanf(s, "%lu", mhp) <= 0)
2029 *mhp = 0;
2030
Andi Kleen8faa8b02008-07-23 21:27:48 -07002031 /*
2032 * Global state is always initialized later in hugetlb_init.
2033 * But we need to allocate >= MAX_ORDER hstates here early to still
2034 * use the bootmem allocator.
2035 */
Aneesh Kumar K.V47d38342012-07-31 16:41:54 -07002036 if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER)
Andi Kleen8faa8b02008-07-23 21:27:48 -07002037 hugetlb_hstate_alloc_pages(parsed_hstate);
2038
2039 last_mhp = mhp;
2040
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002041 return 1;
2042}
Nick Piggine11bfbf2008-07-23 21:27:52 -07002043__setup("hugepages=", hugetlb_nrpages_setup);
2044
2045static int __init hugetlb_default_setup(char *s)
2046{
2047 default_hstate_size = memparse(s, &s);
2048 return 1;
2049}
2050__setup("default_hugepagesz=", hugetlb_default_setup);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002051
Nishanth Aravamudan8a213462008-07-25 19:44:37 -07002052static unsigned int cpuset_mems_nr(unsigned int *array)
2053{
2054 int node;
2055 unsigned int nr = 0;
2056
2057 for_each_node_mask(node, cpuset_current_mems_allowed)
2058 nr += array[node];
2059
2060 return nr;
2061}
2062
2063#ifdef CONFIG_SYSCTL
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002064static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
2065 struct ctl_table *table, int write,
2066 void __user *buffer, size_t *length, loff_t *ppos)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002067{
Andi Kleene5ff2152008-07-23 21:27:42 -07002068 struct hstate *h = &default_hstate;
2069 unsigned long tmp;
Michal Hocko08d4a242011-01-13 15:47:26 -08002070 int ret;
Andi Kleene5ff2152008-07-23 21:27:42 -07002071
Petr Holasekc033a932011-03-22 16:33:05 -07002072 tmp = h->max_huge_pages;
Andi Kleene5ff2152008-07-23 21:27:42 -07002073
Eric B Munsonadbe8722011-01-13 15:47:27 -08002074 if (write && h->order >= MAX_ORDER)
2075 return -EINVAL;
2076
Andi Kleene5ff2152008-07-23 21:27:42 -07002077 table->data = &tmp;
2078 table->maxlen = sizeof(unsigned long);
Michal Hocko08d4a242011-01-13 15:47:26 -08002079 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2080 if (ret)
2081 goto out;
Andi Kleene5ff2152008-07-23 21:27:42 -07002082
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002083 if (write) {
David Rientjesbad44b52009-12-14 17:58:38 -08002084 NODEMASK_ALLOC(nodemask_t, nodes_allowed,
2085 GFP_KERNEL | __GFP_NORETRY);
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002086 if (!(obey_mempolicy &&
2087 init_nodemask_of_mempolicy(nodes_allowed))) {
2088 NODEMASK_FREE(nodes_allowed);
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08002089 nodes_allowed = &node_states[N_MEMORY];
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002090 }
2091 h->max_huge_pages = set_max_huge_pages(h, tmp, nodes_allowed);
2092
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08002093 if (nodes_allowed != &node_states[N_MEMORY])
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002094 NODEMASK_FREE(nodes_allowed);
2095 }
Michal Hocko08d4a242011-01-13 15:47:26 -08002096out:
2097 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002098}
Mel Gorman396faf02007-07-17 04:03:13 -07002099
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002100int hugetlb_sysctl_handler(struct ctl_table *table, int write,
2101 void __user *buffer, size_t *length, loff_t *ppos)
2102{
2103
2104 return hugetlb_sysctl_handler_common(false, table, write,
2105 buffer, length, ppos);
2106}
2107
2108#ifdef CONFIG_NUMA
2109int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
2110 void __user *buffer, size_t *length, loff_t *ppos)
2111{
2112 return hugetlb_sysctl_handler_common(true, table, write,
2113 buffer, length, ppos);
2114}
2115#endif /* CONFIG_NUMA */
2116
Nishanth Aravamudana3d0c6a2008-02-08 04:18:18 -08002117int hugetlb_overcommit_handler(struct ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07002118 void __user *buffer,
Nishanth Aravamudana3d0c6a2008-02-08 04:18:18 -08002119 size_t *length, loff_t *ppos)
2120{
Andi Kleena5516432008-07-23 21:27:41 -07002121 struct hstate *h = &default_hstate;
Andi Kleene5ff2152008-07-23 21:27:42 -07002122 unsigned long tmp;
Michal Hocko08d4a242011-01-13 15:47:26 -08002123 int ret;
Andi Kleene5ff2152008-07-23 21:27:42 -07002124
Petr Holasekc033a932011-03-22 16:33:05 -07002125 tmp = h->nr_overcommit_huge_pages;
Andi Kleene5ff2152008-07-23 21:27:42 -07002126
Eric B Munsonadbe8722011-01-13 15:47:27 -08002127 if (write && h->order >= MAX_ORDER)
2128 return -EINVAL;
2129
Andi Kleene5ff2152008-07-23 21:27:42 -07002130 table->data = &tmp;
2131 table->maxlen = sizeof(unsigned long);
Michal Hocko08d4a242011-01-13 15:47:26 -08002132 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2133 if (ret)
2134 goto out;
Andi Kleene5ff2152008-07-23 21:27:42 -07002135
2136 if (write) {
2137 spin_lock(&hugetlb_lock);
2138 h->nr_overcommit_huge_pages = tmp;
2139 spin_unlock(&hugetlb_lock);
2140 }
Michal Hocko08d4a242011-01-13 15:47:26 -08002141out:
2142 return ret;
Nishanth Aravamudana3d0c6a2008-02-08 04:18:18 -08002143}
2144
Linus Torvalds1da177e2005-04-16 15:20:36 -07002145#endif /* CONFIG_SYSCTL */
2146
Alexey Dobriyane1759c22008-10-15 23:50:22 +04002147void hugetlb_report_meminfo(struct seq_file *m)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002148{
Andi Kleena5516432008-07-23 21:27:41 -07002149 struct hstate *h = &default_hstate;
Alexey Dobriyane1759c22008-10-15 23:50:22 +04002150 seq_printf(m,
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002151 "HugePages_Total: %5lu\n"
2152 "HugePages_Free: %5lu\n"
2153 "HugePages_Rsvd: %5lu\n"
2154 "HugePages_Surp: %5lu\n"
2155 "Hugepagesize: %8lu kB\n",
Andi Kleena5516432008-07-23 21:27:41 -07002156 h->nr_huge_pages,
2157 h->free_huge_pages,
2158 h->resv_huge_pages,
2159 h->surplus_huge_pages,
2160 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002161}
2162
2163int hugetlb_report_node_meminfo(int nid, char *buf)
2164{
Andi Kleena5516432008-07-23 21:27:41 -07002165 struct hstate *h = &default_hstate;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002166 return sprintf(buf,
2167 "Node %d HugePages_Total: %5u\n"
Nishanth Aravamudana1de0912008-03-26 14:37:53 -07002168 "Node %d HugePages_Free: %5u\n"
2169 "Node %d HugePages_Surp: %5u\n",
Andi Kleena5516432008-07-23 21:27:41 -07002170 nid, h->nr_huge_pages_node[nid],
2171 nid, h->free_huge_pages_node[nid],
2172 nid, h->surplus_huge_pages_node[nid]);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002173}
2174
David Rientjes949f7ec2013-04-29 15:07:48 -07002175void hugetlb_show_meminfo(void)
2176{
2177 struct hstate *h;
2178 int nid;
2179
2180 for_each_node_state(nid, N_MEMORY)
2181 for_each_hstate(h)
2182 pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
2183 nid,
2184 h->nr_huge_pages_node[nid],
2185 h->free_huge_pages_node[nid],
2186 h->surplus_huge_pages_node[nid],
2187 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
2188}
2189
Linus Torvalds1da177e2005-04-16 15:20:36 -07002190/* Return the number pages of memory we physically have, in PAGE_SIZE units. */
2191unsigned long hugetlb_total_pages(void)
2192{
Wanpeng Lid0028582013-03-22 15:04:40 -07002193 struct hstate *h;
2194 unsigned long nr_total_pages = 0;
2195
2196 for_each_hstate(h)
2197 nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h);
2198 return nr_total_pages;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002199}
Linus Torvalds1da177e2005-04-16 15:20:36 -07002200
Andi Kleena5516432008-07-23 21:27:41 -07002201static int hugetlb_acct_memory(struct hstate *h, long delta)
Mel Gormanfc1b8a72008-07-23 21:27:22 -07002202{
2203 int ret = -ENOMEM;
2204
2205 spin_lock(&hugetlb_lock);
2206 /*
2207 * When cpuset is configured, it breaks the strict hugetlb page
2208 * reservation as the accounting is done on a global variable. Such
2209 * reservation is completely rubbish in the presence of cpuset because
2210 * the reservation is not checked against page availability for the
2211 * current cpuset. Application can still potentially OOM'ed by kernel
2212 * with lack of free htlb page in cpuset that the task is in.
2213 * Attempt to enforce strict accounting with cpuset is almost
2214 * impossible (or too ugly) because cpuset is too fluid that
2215 * task or memory node can be dynamically moved between cpusets.
2216 *
2217 * The change of semantics for shared hugetlb mapping with cpuset is
2218 * undesirable. However, in order to preserve some of the semantics,
2219 * we fall back to check against current free page availability as
2220 * a best attempt and hopefully to minimize the impact of changing
2221 * semantics that cpuset has.
2222 */
2223 if (delta > 0) {
Andi Kleena5516432008-07-23 21:27:41 -07002224 if (gather_surplus_pages(h, delta) < 0)
Mel Gormanfc1b8a72008-07-23 21:27:22 -07002225 goto out;
2226
Andi Kleena5516432008-07-23 21:27:41 -07002227 if (delta > cpuset_mems_nr(h->free_huge_pages_node)) {
2228 return_unused_surplus_pages(h, delta);
Mel Gormanfc1b8a72008-07-23 21:27:22 -07002229 goto out;
2230 }
2231 }
2232
2233 ret = 0;
2234 if (delta < 0)
Andi Kleena5516432008-07-23 21:27:41 -07002235 return_unused_surplus_pages(h, (unsigned long) -delta);
Mel Gormanfc1b8a72008-07-23 21:27:22 -07002236
2237out:
2238 spin_unlock(&hugetlb_lock);
2239 return ret;
2240}
2241
Andy Whitcroft84afd992008-07-23 21:27:32 -07002242static void hugetlb_vm_op_open(struct vm_area_struct *vma)
2243{
Joonsoo Kimf522c3a2013-09-11 14:21:53 -07002244 struct resv_map *resv = vma_resv_map(vma);
Andy Whitcroft84afd992008-07-23 21:27:32 -07002245
2246 /*
2247 * This new VMA should share its siblings reservation map if present.
2248 * The VMA will only ever have a valid reservation map pointer where
2249 * it is being copied for another still existing VMA. As that VMA
Lucas De Marchi25985ed2011-03-30 22:57:33 -03002250 * has a reference to the reservation map it cannot disappear until
Andy Whitcroft84afd992008-07-23 21:27:32 -07002251 * after this open call completes. It is therefore safe to take a
2252 * new reference here without additional locking.
2253 */
Joonsoo Kimf522c3a2013-09-11 14:21:53 -07002254 if (resv)
2255 kref_get(&resv->refs);
Andy Whitcroft84afd992008-07-23 21:27:32 -07002256}
2257
Dave Hansenc50ac052012-05-29 15:06:46 -07002258static void resv_map_put(struct vm_area_struct *vma)
2259{
Joonsoo Kimf522c3a2013-09-11 14:21:53 -07002260 struct resv_map *resv = vma_resv_map(vma);
Dave Hansenc50ac052012-05-29 15:06:46 -07002261
Joonsoo Kimf522c3a2013-09-11 14:21:53 -07002262 if (!resv)
Dave Hansenc50ac052012-05-29 15:06:46 -07002263 return;
Joonsoo Kimf522c3a2013-09-11 14:21:53 -07002264 kref_put(&resv->refs, resv_map_release);
Dave Hansenc50ac052012-05-29 15:06:46 -07002265}
2266
Mel Gormana1e78772008-07-23 21:27:23 -07002267static void hugetlb_vm_op_close(struct vm_area_struct *vma)
2268{
Andi Kleena5516432008-07-23 21:27:41 -07002269 struct hstate *h = hstate_vma(vma);
Joonsoo Kimf522c3a2013-09-11 14:21:53 -07002270 struct resv_map *resv = vma_resv_map(vma);
David Gibson90481622012-03-21 16:34:12 -07002271 struct hugepage_subpool *spool = subpool_vma(vma);
Andy Whitcroft84afd992008-07-23 21:27:32 -07002272 unsigned long reserve;
2273 unsigned long start;
2274 unsigned long end;
2275
Joonsoo Kimf522c3a2013-09-11 14:21:53 -07002276 if (resv) {
Andi Kleena5516432008-07-23 21:27:41 -07002277 start = vma_hugecache_offset(h, vma, vma->vm_start);
2278 end = vma_hugecache_offset(h, vma, vma->vm_end);
Andy Whitcroft84afd992008-07-23 21:27:32 -07002279
2280 reserve = (end - start) -
Joonsoo Kimf522c3a2013-09-11 14:21:53 -07002281 region_count(&resv->regions, start, end);
Andy Whitcroft84afd992008-07-23 21:27:32 -07002282
Dave Hansenc50ac052012-05-29 15:06:46 -07002283 resv_map_put(vma);
Andy Whitcroft84afd992008-07-23 21:27:32 -07002284
Adam Litke7251ff72008-07-23 21:27:59 -07002285 if (reserve) {
Andi Kleena5516432008-07-23 21:27:41 -07002286 hugetlb_acct_memory(h, -reserve);
David Gibson90481622012-03-21 16:34:12 -07002287 hugepage_subpool_put_pages(spool, reserve);
Adam Litke7251ff72008-07-23 21:27:59 -07002288 }
Andy Whitcroft84afd992008-07-23 21:27:32 -07002289 }
Mel Gormana1e78772008-07-23 21:27:23 -07002290}
2291
Linus Torvalds1da177e2005-04-16 15:20:36 -07002292/*
2293 * We cannot handle pagefaults against hugetlb pages at all. They cause
2294 * handle_mm_fault() to try to instantiate regular-sized pages in the
2295 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
2296 * this far.
2297 */
Nick Piggind0217ac2007-07-19 01:47:03 -07002298static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002299{
2300 BUG();
Nick Piggind0217ac2007-07-19 01:47:03 -07002301 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002302}
2303
Alexey Dobriyanf0f37e22009-09-27 22:29:37 +04002304const struct vm_operations_struct hugetlb_vm_ops = {
Nick Piggind0217ac2007-07-19 01:47:03 -07002305 .fault = hugetlb_vm_op_fault,
Andy Whitcroft84afd992008-07-23 21:27:32 -07002306 .open = hugetlb_vm_op_open,
Mel Gormana1e78772008-07-23 21:27:23 -07002307 .close = hugetlb_vm_op_close,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002308};
2309
David Gibson1e8f8892006-01-06 00:10:44 -08002310static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
2311 int writable)
David Gibson63551ae2005-06-21 17:14:44 -07002312{
2313 pte_t entry;
2314
David Gibson1e8f8892006-01-06 00:10:44 -08002315 if (writable) {
Gerald Schaefer106c9922013-04-29 15:07:23 -07002316 entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page,
2317 vma->vm_page_prot)));
David Gibson63551ae2005-06-21 17:14:44 -07002318 } else {
Gerald Schaefer106c9922013-04-29 15:07:23 -07002319 entry = huge_pte_wrprotect(mk_huge_pte(page,
2320 vma->vm_page_prot));
David Gibson63551ae2005-06-21 17:14:44 -07002321 }
2322 entry = pte_mkyoung(entry);
2323 entry = pte_mkhuge(entry);
Chris Metcalfd9ed9fa2012-04-01 14:01:34 -04002324 entry = arch_make_huge_pte(entry, vma, page, writable);
David Gibson63551ae2005-06-21 17:14:44 -07002325
2326 return entry;
2327}
2328
David Gibson1e8f8892006-01-06 00:10:44 -08002329static void set_huge_ptep_writable(struct vm_area_struct *vma,
2330 unsigned long address, pte_t *ptep)
2331{
2332 pte_t entry;
2333
Gerald Schaefer106c9922013-04-29 15:07:23 -07002334 entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep)));
Chris Forbes32f84522011-07-25 17:12:14 -07002335 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
Russell King4b3073e2009-12-18 16:40:18 +00002336 update_mmu_cache(vma, address, ptep);
David Gibson1e8f8892006-01-06 00:10:44 -08002337}
2338
2339
David Gibson63551ae2005-06-21 17:14:44 -07002340int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
2341 struct vm_area_struct *vma)
2342{
2343 pte_t *src_pte, *dst_pte, entry;
2344 struct page *ptepage;
Hugh Dickins1c598272005-10-19 21:23:43 -07002345 unsigned long addr;
David Gibson1e8f8892006-01-06 00:10:44 -08002346 int cow;
Andi Kleena5516432008-07-23 21:27:41 -07002347 struct hstate *h = hstate_vma(vma);
2348 unsigned long sz = huge_page_size(h);
Andreas Sandberge8569dd2014-01-21 15:49:09 -08002349 unsigned long mmun_start; /* For mmu_notifiers */
2350 unsigned long mmun_end; /* For mmu_notifiers */
2351 int ret = 0;
David Gibson1e8f8892006-01-06 00:10:44 -08002352
2353 cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
David Gibson63551ae2005-06-21 17:14:44 -07002354
Andreas Sandberge8569dd2014-01-21 15:49:09 -08002355 mmun_start = vma->vm_start;
2356 mmun_end = vma->vm_end;
2357 if (cow)
2358 mmu_notifier_invalidate_range_start(src, mmun_start, mmun_end);
2359
Andi Kleena5516432008-07-23 21:27:41 -07002360 for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002361 spinlock_t *src_ptl, *dst_ptl;
Hugh Dickinsc74df322005-10-29 18:16:23 -07002362 src_pte = huge_pte_offset(src, addr);
2363 if (!src_pte)
2364 continue;
Andi Kleena5516432008-07-23 21:27:41 -07002365 dst_pte = huge_pte_alloc(dst, addr, sz);
Andreas Sandberge8569dd2014-01-21 15:49:09 -08002366 if (!dst_pte) {
2367 ret = -ENOMEM;
2368 break;
2369 }
Larry Woodmanc5c99422008-01-24 05:49:25 -08002370
2371 /* If the pagetables are shared don't copy or take references */
2372 if (dst_pte == src_pte)
2373 continue;
2374
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002375 dst_ptl = huge_pte_lock(h, dst, dst_pte);
2376 src_ptl = huge_pte_lockptr(h, src, src_pte);
2377 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07002378 if (!huge_pte_none(huge_ptep_get(src_pte))) {
David Gibson1e8f8892006-01-06 00:10:44 -08002379 if (cow)
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07002380 huge_ptep_set_wrprotect(src, addr, src_pte);
2381 entry = huge_ptep_get(src_pte);
Hugh Dickins1c598272005-10-19 21:23:43 -07002382 ptepage = pte_page(entry);
2383 get_page(ptepage);
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002384 page_dup_rmap(ptepage);
Hugh Dickins1c598272005-10-19 21:23:43 -07002385 set_huge_pte_at(dst, addr, dst_pte, entry);
2386 }
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002387 spin_unlock(src_ptl);
2388 spin_unlock(dst_ptl);
David Gibson63551ae2005-06-21 17:14:44 -07002389 }
David Gibson63551ae2005-06-21 17:14:44 -07002390
Andreas Sandberge8569dd2014-01-21 15:49:09 -08002391 if (cow)
2392 mmu_notifier_invalidate_range_end(src, mmun_start, mmun_end);
2393
2394 return ret;
David Gibson63551ae2005-06-21 17:14:44 -07002395}
2396
Naoya Horiguchi290408d2010-09-08 10:19:35 +09002397static int is_hugetlb_entry_migration(pte_t pte)
2398{
2399 swp_entry_t swp;
2400
2401 if (huge_pte_none(pte) || pte_present(pte))
2402 return 0;
2403 swp = pte_to_swp_entry(pte);
Chris Forbes32f84522011-07-25 17:12:14 -07002404 if (non_swap_entry(swp) && is_migration_entry(swp))
Naoya Horiguchi290408d2010-09-08 10:19:35 +09002405 return 1;
Chris Forbes32f84522011-07-25 17:12:14 -07002406 else
Naoya Horiguchi290408d2010-09-08 10:19:35 +09002407 return 0;
2408}
2409
Naoya Horiguchifd6a03e2010-05-28 09:29:21 +09002410static int is_hugetlb_entry_hwpoisoned(pte_t pte)
2411{
2412 swp_entry_t swp;
2413
2414 if (huge_pte_none(pte) || pte_present(pte))
2415 return 0;
2416 swp = pte_to_swp_entry(pte);
Chris Forbes32f84522011-07-25 17:12:14 -07002417 if (non_swap_entry(swp) && is_hwpoison_entry(swp))
Naoya Horiguchifd6a03e2010-05-28 09:29:21 +09002418 return 1;
Chris Forbes32f84522011-07-25 17:12:14 -07002419 else
Naoya Horiguchifd6a03e2010-05-28 09:29:21 +09002420 return 0;
2421}
2422
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07002423void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
2424 unsigned long start, unsigned long end,
2425 struct page *ref_page)
David Gibson63551ae2005-06-21 17:14:44 -07002426{
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07002427 int force_flush = 0;
David Gibson63551ae2005-06-21 17:14:44 -07002428 struct mm_struct *mm = vma->vm_mm;
2429 unsigned long address;
David Gibsonc7546f82005-08-05 11:59:35 -07002430 pte_t *ptep;
David Gibson63551ae2005-06-21 17:14:44 -07002431 pte_t pte;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002432 spinlock_t *ptl;
David Gibson63551ae2005-06-21 17:14:44 -07002433 struct page *page;
Andi Kleena5516432008-07-23 21:27:41 -07002434 struct hstate *h = hstate_vma(vma);
2435 unsigned long sz = huge_page_size(h);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07002436 const unsigned long mmun_start = start; /* For mmu_notifiers */
2437 const unsigned long mmun_end = end; /* For mmu_notifiers */
Andi Kleena5516432008-07-23 21:27:41 -07002438
David Gibson63551ae2005-06-21 17:14:44 -07002439 WARN_ON(!is_vm_hugetlb_page(vma));
Andi Kleena5516432008-07-23 21:27:41 -07002440 BUG_ON(start & ~huge_page_mask(h));
2441 BUG_ON(end & ~huge_page_mask(h));
David Gibson63551ae2005-06-21 17:14:44 -07002442
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07002443 tlb_start_vma(tlb, vma);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07002444 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07002445again:
Andi Kleena5516432008-07-23 21:27:41 -07002446 for (address = start; address < end; address += sz) {
David Gibsonc7546f82005-08-05 11:59:35 -07002447 ptep = huge_pte_offset(mm, address);
Adam Litke4c887262005-10-29 18:16:46 -07002448 if (!ptep)
David Gibsonc7546f82005-08-05 11:59:35 -07002449 continue;
2450
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002451 ptl = huge_pte_lock(h, mm, ptep);
Chen, Kenneth W39dde652006-12-06 20:32:03 -08002452 if (huge_pmd_unshare(mm, &address, ptep))
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002453 goto unlock;
Chen, Kenneth W39dde652006-12-06 20:32:03 -08002454
Hillf Danton66293262012-03-23 15:01:48 -07002455 pte = huge_ptep_get(ptep);
2456 if (huge_pte_none(pte))
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002457 goto unlock;
Hillf Danton66293262012-03-23 15:01:48 -07002458
2459 /*
2460 * HWPoisoned hugepage is already unmapped and dropped reference
2461 */
Naoya Horiguchi8c4894c2012-12-12 13:52:28 -08002462 if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) {
Gerald Schaefer106c9922013-04-29 15:07:23 -07002463 huge_pte_clear(mm, address, ptep);
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002464 goto unlock;
Naoya Horiguchi8c4894c2012-12-12 13:52:28 -08002465 }
Hillf Danton66293262012-03-23 15:01:48 -07002466
2467 page = pte_page(pte);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002468 /*
2469 * If a reference page is supplied, it is because a specific
2470 * page is being unmapped, not a range. Ensure the page we
2471 * are about to unmap is the actual page of interest.
2472 */
2473 if (ref_page) {
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002474 if (page != ref_page)
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002475 goto unlock;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002476
2477 /*
2478 * Mark the VMA as having unmapped its page so that
2479 * future faults in this VMA will fail rather than
2480 * looking like data was lost
2481 */
2482 set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
2483 }
2484
David Gibsonc7546f82005-08-05 11:59:35 -07002485 pte = huge_ptep_get_and_clear(mm, address, ptep);
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07002486 tlb_remove_tlb_entry(tlb, ptep, address);
Gerald Schaefer106c9922013-04-29 15:07:23 -07002487 if (huge_pte_dirty(pte))
Ken Chen6649a382007-02-08 14:20:27 -08002488 set_page_dirty(page);
Hillf Danton9e811302012-03-21 16:34:03 -07002489
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07002490 page_remove_rmap(page);
2491 force_flush = !__tlb_remove_page(tlb, page);
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002492 if (force_flush) {
2493 spin_unlock(ptl);
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07002494 break;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002495 }
Hillf Danton9e811302012-03-21 16:34:03 -07002496 /* Bail out after unmapping reference page if supplied */
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002497 if (ref_page) {
2498 spin_unlock(ptl);
Hillf Danton9e811302012-03-21 16:34:03 -07002499 break;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002500 }
2501unlock:
2502 spin_unlock(ptl);
David Gibson63551ae2005-06-21 17:14:44 -07002503 }
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07002504 /*
2505 * mmu_gather ran out of room to batch pages, we break out of
2506 * the PTE lock to avoid doing the potential expensive TLB invalidate
2507 * and page-free while holding it.
2508 */
2509 if (force_flush) {
2510 force_flush = 0;
2511 tlb_flush_mmu(tlb);
2512 if (address < end && !ref_page)
2513 goto again;
Chen, Kenneth Wfe1668a2006-10-04 02:15:24 -07002514 }
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07002515 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07002516 tlb_end_vma(tlb, vma);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002517}
David Gibson63551ae2005-06-21 17:14:44 -07002518
Mel Gormand8333522012-07-31 16:46:20 -07002519void __unmap_hugepage_range_final(struct mmu_gather *tlb,
2520 struct vm_area_struct *vma, unsigned long start,
2521 unsigned long end, struct page *ref_page)
2522{
2523 __unmap_hugepage_range(tlb, vma, start, end, ref_page);
2524
2525 /*
2526 * Clear this flag so that x86's huge_pmd_share page_table_shareable
2527 * test will fail on a vma being torn down, and not grab a page table
2528 * on its way out. We're lucky that the flag has such an appropriate
2529 * name, and can in fact be safely cleared here. We could clear it
2530 * before the __unmap_hugepage_range above, but all that's necessary
2531 * is to clear it before releasing the i_mmap_mutex. This works
2532 * because in the context this is called, the VMA is about to be
2533 * destroyed and the i_mmap_mutex is held.
2534 */
2535 vma->vm_flags &= ~VM_MAYSHARE;
2536}
2537
Chen, Kenneth W502717f2006-10-11 01:20:46 -07002538void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002539 unsigned long end, struct page *ref_page)
Chen, Kenneth W502717f2006-10-11 01:20:46 -07002540{
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07002541 struct mm_struct *mm;
2542 struct mmu_gather tlb;
2543
2544 mm = vma->vm_mm;
2545
Linus Torvalds2b047252013-08-15 11:42:25 -07002546 tlb_gather_mmu(&tlb, mm, start, end);
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07002547 __unmap_hugepage_range(&tlb, vma, start, end, ref_page);
2548 tlb_finish_mmu(&tlb, start, end);
Chen, Kenneth W502717f2006-10-11 01:20:46 -07002549}
2550
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002551/*
2552 * This is called when the original mapper is failing to COW a MAP_PRIVATE
2553 * mappping it owns the reserve page for. The intention is to unmap the page
2554 * from other VMAs and let the children be SIGKILLed if they are faulting the
2555 * same region.
2556 */
Harvey Harrison2a4b3de2008-10-18 20:27:06 -07002557static int unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
2558 struct page *page, unsigned long address)
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002559{
Adam Litke75266742008-11-12 13:24:56 -08002560 struct hstate *h = hstate_vma(vma);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002561 struct vm_area_struct *iter_vma;
2562 struct address_space *mapping;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002563 pgoff_t pgoff;
2564
2565 /*
2566 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
2567 * from page cache lookup which is in HPAGE_SIZE units.
2568 */
Adam Litke75266742008-11-12 13:24:56 -08002569 address = address & huge_page_mask(h);
Michal Hocko36e4f202012-10-08 16:33:31 -07002570 pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) +
2571 vma->vm_pgoff;
Al Viro496ad9a2013-01-23 17:07:38 -05002572 mapping = file_inode(vma->vm_file)->i_mapping;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002573
Mel Gorman4eb2b1d2009-12-14 17:59:53 -08002574 /*
2575 * Take the mapping lock for the duration of the table walk. As
2576 * this mapping should be shared between all the VMAs,
2577 * __unmap_hugepage_range() is called as the lock is already held
2578 */
Peter Zijlstra3d48ae42011-05-24 17:12:06 -07002579 mutex_lock(&mapping->i_mmap_mutex);
Michel Lespinasse6b2dbba2012-10-08 16:31:25 -07002580 vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) {
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002581 /* Do not unmap the current VMA */
2582 if (iter_vma == vma)
2583 continue;
2584
2585 /*
2586 * Unmap the page from other VMAs without their own reserves.
2587 * They get marked to be SIGKILLed if they fault in these
2588 * areas. This is because a future no-page fault on this VMA
2589 * could insert a zeroed page instead of the data existing
2590 * from the time of fork. This would look like data corruption
2591 */
2592 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07002593 unmap_hugepage_range(iter_vma, address,
2594 address + huge_page_size(h), page);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002595 }
Peter Zijlstra3d48ae42011-05-24 17:12:06 -07002596 mutex_unlock(&mapping->i_mmap_mutex);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002597
2598 return 1;
2599}
2600
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002601/*
2602 * Hugetlb_cow() should be called with page lock of the original hugepage held.
Michal Hockoef009b22012-01-10 15:07:21 -08002603 * Called with hugetlb_instantiation_mutex held and pte_page locked so we
2604 * cannot race with other handlers or page migration.
2605 * Keep the pte_same checks anyway to make transition from the mutex easier.
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002606 */
David Gibson1e8f8892006-01-06 00:10:44 -08002607static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002608 unsigned long address, pte_t *ptep, pte_t pte,
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002609 struct page *pagecache_page, spinlock_t *ptl)
David Gibson1e8f8892006-01-06 00:10:44 -08002610{
Andi Kleena5516432008-07-23 21:27:41 -07002611 struct hstate *h = hstate_vma(vma);
David Gibson1e8f8892006-01-06 00:10:44 -08002612 struct page *old_page, *new_page;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002613 int outside_reserve = 0;
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07002614 unsigned long mmun_start; /* For mmu_notifiers */
2615 unsigned long mmun_end; /* For mmu_notifiers */
David Gibson1e8f8892006-01-06 00:10:44 -08002616
2617 old_page = pte_page(pte);
2618
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002619retry_avoidcopy:
David Gibson1e8f8892006-01-06 00:10:44 -08002620 /* If no-one else is actually using this page, avoid the copy
2621 * and just make the page writable */
Joonsoo Kim37a21402013-09-11 14:21:04 -07002622 if (page_mapcount(old_page) == 1 && PageAnon(old_page)) {
2623 page_move_anon_rmap(old_page, vma, address);
David Gibson1e8f8892006-01-06 00:10:44 -08002624 set_huge_ptep_writable(vma, address, ptep);
Nick Piggin83c54072007-07-19 01:47:05 -07002625 return 0;
David Gibson1e8f8892006-01-06 00:10:44 -08002626 }
2627
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002628 /*
2629 * If the process that created a MAP_PRIVATE mapping is about to
2630 * perform a COW due to a shared page count, attempt to satisfy
2631 * the allocation without using the existing reserves. The pagecache
2632 * page is used to determine if the reserve at this address was
2633 * consumed or not. If reserves were used, a partial faulted mapping
2634 * at the time of fork() could consume its reserves on COW instead
2635 * of the full address range.
2636 */
Joonsoo Kim5944d012013-09-11 14:21:55 -07002637 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002638 old_page != pagecache_page)
2639 outside_reserve = 1;
2640
David Gibson1e8f8892006-01-06 00:10:44 -08002641 page_cache_get(old_page);
Larry Woodmanb76c8cf2009-12-14 17:59:37 -08002642
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002643 /* Drop page table lock as buddy allocator may be called */
2644 spin_unlock(ptl);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002645 new_page = alloc_huge_page(vma, address, outside_reserve);
David Gibson1e8f8892006-01-06 00:10:44 -08002646
Adam Litke2fc39ce2007-11-14 16:59:39 -08002647 if (IS_ERR(new_page)) {
Aneesh Kumar K.V76dcee72012-07-31 16:41:57 -07002648 long err = PTR_ERR(new_page);
David Gibson1e8f8892006-01-06 00:10:44 -08002649 page_cache_release(old_page);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002650
2651 /*
2652 * If a process owning a MAP_PRIVATE mapping fails to COW,
2653 * it is due to references held by a child and an insufficient
2654 * huge page pool. To guarantee the original mappers
2655 * reliability, unmap the page from child processes. The child
2656 * may get SIGKILLed if it later faults.
2657 */
2658 if (outside_reserve) {
2659 BUG_ON(huge_pte_none(pte));
2660 if (unmap_ref_private(mm, vma, old_page, address)) {
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002661 BUG_ON(huge_pte_none(pte));
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002662 spin_lock(ptl);
Hillf Dantona734bcc2012-01-10 15:07:20 -08002663 ptep = huge_pte_offset(mm, address & huge_page_mask(h));
2664 if (likely(pte_same(huge_ptep_get(ptep), pte)))
2665 goto retry_avoidcopy;
2666 /*
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002667 * race occurs while re-acquiring page table
2668 * lock, and our job is done.
Hillf Dantona734bcc2012-01-10 15:07:20 -08002669 */
2670 return 0;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002671 }
2672 WARN_ON_ONCE(1);
2673 }
2674
Larry Woodmanb76c8cf2009-12-14 17:59:37 -08002675 /* Caller expects lock to be held */
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002676 spin_lock(ptl);
Aneesh Kumar K.V76dcee72012-07-31 16:41:57 -07002677 if (err == -ENOMEM)
2678 return VM_FAULT_OOM;
2679 else
2680 return VM_FAULT_SIGBUS;
David Gibson1e8f8892006-01-06 00:10:44 -08002681 }
2682
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002683 /*
2684 * When the original hugepage is shared one, it does not have
2685 * anon_vma prepared.
2686 */
Dean Nelson44e2aa92010-10-26 14:22:08 -07002687 if (unlikely(anon_vma_prepare(vma))) {
Hillf Dantonea4039a2011-11-15 14:36:12 -08002688 page_cache_release(new_page);
2689 page_cache_release(old_page);
Dean Nelson44e2aa92010-10-26 14:22:08 -07002690 /* Caller expects lock to be held */
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002691 spin_lock(ptl);
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002692 return VM_FAULT_OOM;
Dean Nelson44e2aa92010-10-26 14:22:08 -07002693 }
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002694
Andrea Arcangeli47ad8472011-01-13 15:46:47 -08002695 copy_user_huge_page(new_page, old_page, address, vma,
2696 pages_per_huge_page(h));
Nick Piggin0ed361d2008-02-04 22:29:34 -08002697 __SetPageUptodate(new_page);
David Gibson1e8f8892006-01-06 00:10:44 -08002698
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07002699 mmun_start = address & huge_page_mask(h);
2700 mmun_end = mmun_start + huge_page_size(h);
2701 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
Larry Woodmanb76c8cf2009-12-14 17:59:37 -08002702 /*
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002703 * Retake the page table lock to check for racing updates
Larry Woodmanb76c8cf2009-12-14 17:59:37 -08002704 * before the page tables are altered
2705 */
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002706 spin_lock(ptl);
Andi Kleena5516432008-07-23 21:27:41 -07002707 ptep = huge_pte_offset(mm, address & huge_page_mask(h));
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07002708 if (likely(pte_same(huge_ptep_get(ptep), pte))) {
Joonsoo Kim07443a82013-09-11 14:21:58 -07002709 ClearPagePrivate(new_page);
2710
David Gibson1e8f8892006-01-06 00:10:44 -08002711 /* Break COW */
Gerald Schaefer8fe627e2008-04-28 02:13:28 -07002712 huge_ptep_clear_flush(vma, address, ptep);
David Gibson1e8f8892006-01-06 00:10:44 -08002713 set_huge_pte_at(mm, address, ptep,
2714 make_huge_pte(vma, new_page, 1));
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002715 page_remove_rmap(old_page);
Naoya Horiguchicd67f0d2010-09-10 13:23:04 +09002716 hugepage_add_new_anon_rmap(new_page, vma, address);
David Gibson1e8f8892006-01-06 00:10:44 -08002717 /* Make the old page be freed below */
2718 new_page = old_page;
2719 }
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002720 spin_unlock(ptl);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07002721 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
David Gibson1e8f8892006-01-06 00:10:44 -08002722 page_cache_release(new_page);
2723 page_cache_release(old_page);
Joonsoo Kim83120342013-09-11 14:21:57 -07002724
2725 /* Caller expects lock to be held */
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002726 spin_lock(ptl);
Nick Piggin83c54072007-07-19 01:47:05 -07002727 return 0;
David Gibson1e8f8892006-01-06 00:10:44 -08002728}
2729
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002730/* Return the pagecache page at a given address within a VMA */
Andi Kleena5516432008-07-23 21:27:41 -07002731static struct page *hugetlbfs_pagecache_page(struct hstate *h,
2732 struct vm_area_struct *vma, unsigned long address)
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002733{
2734 struct address_space *mapping;
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -07002735 pgoff_t idx;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002736
2737 mapping = vma->vm_file->f_mapping;
Andi Kleena5516432008-07-23 21:27:41 -07002738 idx = vma_hugecache_offset(h, vma, address);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002739
2740 return find_lock_page(mapping, idx);
2741}
2742
Hugh Dickins3ae77f42009-09-21 17:03:33 -07002743/*
2744 * Return whether there is a pagecache page to back given address within VMA.
2745 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
2746 */
2747static bool hugetlbfs_pagecache_present(struct hstate *h,
Hugh Dickins2a15efc2009-09-21 17:03:27 -07002748 struct vm_area_struct *vma, unsigned long address)
2749{
2750 struct address_space *mapping;
2751 pgoff_t idx;
2752 struct page *page;
2753
2754 mapping = vma->vm_file->f_mapping;
2755 idx = vma_hugecache_offset(h, vma, address);
2756
2757 page = find_get_page(mapping, idx);
2758 if (page)
2759 put_page(page);
2760 return page != NULL;
2761}
2762
Robert P. J. Daya1ed3dd2007-07-17 04:03:33 -07002763static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
Hugh Dickins788c7df2009-06-23 13:49:05 +01002764 unsigned long address, pte_t *ptep, unsigned int flags)
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01002765{
Andi Kleena5516432008-07-23 21:27:41 -07002766 struct hstate *h = hstate_vma(vma);
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01002767 int ret = VM_FAULT_SIGBUS;
Hillf Danton409eb8c2012-01-20 14:34:13 -08002768 int anon_rmap = 0;
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -07002769 pgoff_t idx;
Adam Litke4c887262005-10-29 18:16:46 -07002770 unsigned long size;
Adam Litke4c887262005-10-29 18:16:46 -07002771 struct page *page;
2772 struct address_space *mapping;
David Gibson1e8f8892006-01-06 00:10:44 -08002773 pte_t new_pte;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002774 spinlock_t *ptl;
Adam Litke4c887262005-10-29 18:16:46 -07002775
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002776 /*
2777 * Currently, we are forced to kill the process in the event the
2778 * original mapper has unmapped pages from the child due to a failed
Lucas De Marchi25985ed2011-03-30 22:57:33 -03002779 * COW. Warn that such a situation has occurred as it may not be obvious
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002780 */
2781 if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
Andrew Mortonffb22af2013-02-22 16:32:08 -08002782 pr_warning("PID %d killed due to inadequate hugepage pool\n",
2783 current->pid);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002784 return ret;
2785 }
2786
Adam Litke4c887262005-10-29 18:16:46 -07002787 mapping = vma->vm_file->f_mapping;
Andi Kleena5516432008-07-23 21:27:41 -07002788 idx = vma_hugecache_offset(h, vma, address);
Adam Litke4c887262005-10-29 18:16:46 -07002789
2790 /*
2791 * Use page lock to guard against racing truncation
2792 * before we get page_table_lock.
2793 */
Christoph Lameter6bda6662006-01-06 00:10:49 -08002794retry:
2795 page = find_lock_page(mapping, idx);
2796 if (!page) {
Andi Kleena5516432008-07-23 21:27:41 -07002797 size = i_size_read(mapping->host) >> huge_page_shift(h);
Hugh Dickinsebed4bf2006-10-28 10:38:43 -07002798 if (idx >= size)
2799 goto out;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002800 page = alloc_huge_page(vma, address, 0);
Adam Litke2fc39ce2007-11-14 16:59:39 -08002801 if (IS_ERR(page)) {
Aneesh Kumar K.V76dcee72012-07-31 16:41:57 -07002802 ret = PTR_ERR(page);
2803 if (ret == -ENOMEM)
2804 ret = VM_FAULT_OOM;
2805 else
2806 ret = VM_FAULT_SIGBUS;
Christoph Lameter6bda6662006-01-06 00:10:49 -08002807 goto out;
2808 }
Andrea Arcangeli47ad8472011-01-13 15:46:47 -08002809 clear_huge_page(page, address, pages_per_huge_page(h));
Nick Piggin0ed361d2008-02-04 22:29:34 -08002810 __SetPageUptodate(page);
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01002811
Mel Gormanf83a2752009-05-28 14:34:40 -07002812 if (vma->vm_flags & VM_MAYSHARE) {
Christoph Lameter6bda6662006-01-06 00:10:49 -08002813 int err;
Ken Chen45c682a2007-11-14 16:59:44 -08002814 struct inode *inode = mapping->host;
Christoph Lameter6bda6662006-01-06 00:10:49 -08002815
2816 err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
2817 if (err) {
2818 put_page(page);
Christoph Lameter6bda6662006-01-06 00:10:49 -08002819 if (err == -EEXIST)
2820 goto retry;
2821 goto out;
2822 }
Joonsoo Kim07443a82013-09-11 14:21:58 -07002823 ClearPagePrivate(page);
Ken Chen45c682a2007-11-14 16:59:44 -08002824
2825 spin_lock(&inode->i_lock);
Andi Kleena5516432008-07-23 21:27:41 -07002826 inode->i_blocks += blocks_per_huge_page(h);
Ken Chen45c682a2007-11-14 16:59:44 -08002827 spin_unlock(&inode->i_lock);
Mel Gorman23be7462010-04-23 13:17:56 -04002828 } else {
Christoph Lameter6bda6662006-01-06 00:10:49 -08002829 lock_page(page);
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002830 if (unlikely(anon_vma_prepare(vma))) {
2831 ret = VM_FAULT_OOM;
2832 goto backout_unlocked;
2833 }
Hillf Danton409eb8c2012-01-20 14:34:13 -08002834 anon_rmap = 1;
Mel Gorman23be7462010-04-23 13:17:56 -04002835 }
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002836 } else {
Naoya Horiguchi998b4382010-09-08 10:19:32 +09002837 /*
2838 * If memory error occurs between mmap() and fault, some process
2839 * don't have hwpoisoned swap entry for errored virtual address.
2840 * So we need to block hugepage fault by PG_hwpoison bit check.
2841 */
2842 if (unlikely(PageHWPoison(page))) {
Chris Forbes32f84522011-07-25 17:12:14 -07002843 ret = VM_FAULT_HWPOISON |
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07002844 VM_FAULT_SET_HINDEX(hstate_index(h));
Naoya Horiguchi998b4382010-09-08 10:19:32 +09002845 goto backout_unlocked;
2846 }
Christoph Lameter6bda6662006-01-06 00:10:49 -08002847 }
David Gibson1e8f8892006-01-06 00:10:44 -08002848
Andy Whitcroft57303d82008-08-12 15:08:47 -07002849 /*
2850 * If we are going to COW a private mapping later, we examine the
2851 * pending reservations for this page now. This will ensure that
2852 * any allocations necessary to record that reservation occur outside
2853 * the spinlock.
2854 */
Hugh Dickins788c7df2009-06-23 13:49:05 +01002855 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED))
Andy Whitcroft2b267362008-08-12 15:08:49 -07002856 if (vma_needs_reservation(h, vma, address) < 0) {
2857 ret = VM_FAULT_OOM;
2858 goto backout_unlocked;
2859 }
Andy Whitcroft57303d82008-08-12 15:08:47 -07002860
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002861 ptl = huge_pte_lockptr(h, mm, ptep);
2862 spin_lock(ptl);
Andi Kleena5516432008-07-23 21:27:41 -07002863 size = i_size_read(mapping->host) >> huge_page_shift(h);
Adam Litke4c887262005-10-29 18:16:46 -07002864 if (idx >= size)
2865 goto backout;
2866
Nick Piggin83c54072007-07-19 01:47:05 -07002867 ret = 0;
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07002868 if (!huge_pte_none(huge_ptep_get(ptep)))
Adam Litke4c887262005-10-29 18:16:46 -07002869 goto backout;
2870
Joonsoo Kim07443a82013-09-11 14:21:58 -07002871 if (anon_rmap) {
2872 ClearPagePrivate(page);
Hillf Danton409eb8c2012-01-20 14:34:13 -08002873 hugepage_add_new_anon_rmap(page, vma, address);
Joonsoo Kim07443a82013-09-11 14:21:58 -07002874 }
Hillf Danton409eb8c2012-01-20 14:34:13 -08002875 else
2876 page_dup_rmap(page);
David Gibson1e8f8892006-01-06 00:10:44 -08002877 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
2878 && (vma->vm_flags & VM_SHARED)));
2879 set_huge_pte_at(mm, address, ptep, new_pte);
2880
Hugh Dickins788c7df2009-06-23 13:49:05 +01002881 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
David Gibson1e8f8892006-01-06 00:10:44 -08002882 /* Optimization, do the COW without a second fault */
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002883 ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page, ptl);
David Gibson1e8f8892006-01-06 00:10:44 -08002884 }
2885
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002886 spin_unlock(ptl);
Adam Litke4c887262005-10-29 18:16:46 -07002887 unlock_page(page);
2888out:
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01002889 return ret;
Adam Litke4c887262005-10-29 18:16:46 -07002890
2891backout:
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002892 spin_unlock(ptl);
Andy Whitcroft2b267362008-08-12 15:08:49 -07002893backout_unlocked:
Adam Litke4c887262005-10-29 18:16:46 -07002894 unlock_page(page);
2895 put_page(page);
2896 goto out;
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01002897}
2898
Adam Litke86e52162006-01-06 00:10:43 -08002899int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
Hugh Dickins788c7df2009-06-23 13:49:05 +01002900 unsigned long address, unsigned int flags)
Adam Litke86e52162006-01-06 00:10:43 -08002901{
2902 pte_t *ptep;
2903 pte_t entry;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002904 spinlock_t *ptl;
David Gibson1e8f8892006-01-06 00:10:44 -08002905 int ret;
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002906 struct page *page = NULL;
Andy Whitcroft57303d82008-08-12 15:08:47 -07002907 struct page *pagecache_page = NULL;
David Gibson3935baa2006-03-22 00:08:53 -08002908 static DEFINE_MUTEX(hugetlb_instantiation_mutex);
Andi Kleena5516432008-07-23 21:27:41 -07002909 struct hstate *h = hstate_vma(vma);
Adam Litke86e52162006-01-06 00:10:43 -08002910
KAMEZAWA Hiroyuki1e16a532012-01-10 15:07:22 -08002911 address &= huge_page_mask(h);
2912
Naoya Horiguchifd6a03e2010-05-28 09:29:21 +09002913 ptep = huge_pte_offset(mm, address);
2914 if (ptep) {
2915 entry = huge_ptep_get(ptep);
Naoya Horiguchi290408d2010-09-08 10:19:35 +09002916 if (unlikely(is_hugetlb_entry_migration(entry))) {
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002917 migration_entry_wait_huge(vma, mm, ptep);
Naoya Horiguchi290408d2010-09-08 10:19:35 +09002918 return 0;
2919 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
Chris Forbes32f84522011-07-25 17:12:14 -07002920 return VM_FAULT_HWPOISON_LARGE |
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07002921 VM_FAULT_SET_HINDEX(hstate_index(h));
Naoya Horiguchifd6a03e2010-05-28 09:29:21 +09002922 }
2923
Andi Kleena5516432008-07-23 21:27:41 -07002924 ptep = huge_pte_alloc(mm, address, huge_page_size(h));
Adam Litke86e52162006-01-06 00:10:43 -08002925 if (!ptep)
2926 return VM_FAULT_OOM;
2927
David Gibson3935baa2006-03-22 00:08:53 -08002928 /*
2929 * Serialize hugepage allocation and instantiation, so that we don't
2930 * get spurious allocation failures if two CPUs race to instantiate
2931 * the same page in the page cache.
2932 */
2933 mutex_lock(&hugetlb_instantiation_mutex);
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07002934 entry = huge_ptep_get(ptep);
2935 if (huge_pte_none(entry)) {
Hugh Dickins788c7df2009-06-23 13:49:05 +01002936 ret = hugetlb_no_page(mm, vma, address, ptep, flags);
David Gibsonb4d1d992008-10-15 22:01:11 -07002937 goto out_mutex;
David Gibson3935baa2006-03-22 00:08:53 -08002938 }
Adam Litke86e52162006-01-06 00:10:43 -08002939
Nick Piggin83c54072007-07-19 01:47:05 -07002940 ret = 0;
David Gibson1e8f8892006-01-06 00:10:44 -08002941
Andy Whitcroft57303d82008-08-12 15:08:47 -07002942 /*
2943 * If we are going to COW the mapping later, we examine the pending
2944 * reservations for this page now. This will ensure that any
2945 * allocations necessary to record that reservation occur outside the
2946 * spinlock. For private mappings, we also lookup the pagecache
2947 * page now as it is used to determine if a reservation has been
2948 * consumed.
2949 */
Gerald Schaefer106c9922013-04-29 15:07:23 -07002950 if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) {
Andy Whitcroft2b267362008-08-12 15:08:49 -07002951 if (vma_needs_reservation(h, vma, address) < 0) {
2952 ret = VM_FAULT_OOM;
David Gibsonb4d1d992008-10-15 22:01:11 -07002953 goto out_mutex;
Andy Whitcroft2b267362008-08-12 15:08:49 -07002954 }
Andy Whitcroft57303d82008-08-12 15:08:47 -07002955
Mel Gormanf83a2752009-05-28 14:34:40 -07002956 if (!(vma->vm_flags & VM_MAYSHARE))
Andy Whitcroft57303d82008-08-12 15:08:47 -07002957 pagecache_page = hugetlbfs_pagecache_page(h,
2958 vma, address);
2959 }
2960
Naoya Horiguchi56c9cfb2010-09-10 13:23:04 +09002961 /*
2962 * hugetlb_cow() requires page locks of pte_page(entry) and
2963 * pagecache_page, so here we need take the former one
2964 * when page != pagecache_page or !pagecache_page.
2965 * Note that locking order is always pagecache_page -> page,
2966 * so no worry about deadlock.
2967 */
2968 page = pte_page(entry);
Chris Metcalf66aebce2012-04-12 12:49:15 -07002969 get_page(page);
Naoya Horiguchi56c9cfb2010-09-10 13:23:04 +09002970 if (page != pagecache_page)
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002971 lock_page(page);
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002972
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002973 ptl = huge_pte_lockptr(h, mm, ptep);
2974 spin_lock(ptl);
David Gibson1e8f8892006-01-06 00:10:44 -08002975 /* Check for a racing update before calling hugetlb_cow */
David Gibsonb4d1d992008-10-15 22:01:11 -07002976 if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002977 goto out_ptl;
David Gibsonb4d1d992008-10-15 22:01:11 -07002978
2979
Hugh Dickins788c7df2009-06-23 13:49:05 +01002980 if (flags & FAULT_FLAG_WRITE) {
Gerald Schaefer106c9922013-04-29 15:07:23 -07002981 if (!huge_pte_write(entry)) {
Andy Whitcroft57303d82008-08-12 15:08:47 -07002982 ret = hugetlb_cow(mm, vma, address, ptep, entry,
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002983 pagecache_page, ptl);
2984 goto out_ptl;
David Gibsonb4d1d992008-10-15 22:01:11 -07002985 }
Gerald Schaefer106c9922013-04-29 15:07:23 -07002986 entry = huge_pte_mkdirty(entry);
David Gibsonb4d1d992008-10-15 22:01:11 -07002987 }
2988 entry = pte_mkyoung(entry);
Hugh Dickins788c7df2009-06-23 13:49:05 +01002989 if (huge_ptep_set_access_flags(vma, address, ptep, entry,
2990 flags & FAULT_FLAG_WRITE))
Russell King4b3073e2009-12-18 16:40:18 +00002991 update_mmu_cache(vma, address, ptep);
David Gibsonb4d1d992008-10-15 22:01:11 -07002992
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002993out_ptl:
2994 spin_unlock(ptl);
Andy Whitcroft57303d82008-08-12 15:08:47 -07002995
2996 if (pagecache_page) {
2997 unlock_page(pagecache_page);
2998 put_page(pagecache_page);
2999 }
Dean Nelson1f64d692010-12-02 14:31:12 -08003000 if (page != pagecache_page)
3001 unlock_page(page);
Chris Metcalf66aebce2012-04-12 12:49:15 -07003002 put_page(page);
Andy Whitcroft57303d82008-08-12 15:08:47 -07003003
David Gibsonb4d1d992008-10-15 22:01:11 -07003004out_mutex:
David Gibson3935baa2006-03-22 00:08:53 -08003005 mutex_unlock(&hugetlb_instantiation_mutex);
David Gibson1e8f8892006-01-06 00:10:44 -08003006
3007 return ret;
Adam Litke86e52162006-01-06 00:10:43 -08003008}
3009
Michel Lespinasse28a35712013-02-22 16:35:55 -08003010long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
3011 struct page **pages, struct vm_area_struct **vmas,
3012 unsigned long *position, unsigned long *nr_pages,
3013 long i, unsigned int flags)
David Gibson63551ae2005-06-21 17:14:44 -07003014{
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08003015 unsigned long pfn_offset;
3016 unsigned long vaddr = *position;
Michel Lespinasse28a35712013-02-22 16:35:55 -08003017 unsigned long remainder = *nr_pages;
Andi Kleena5516432008-07-23 21:27:41 -07003018 struct hstate *h = hstate_vma(vma);
David Gibson63551ae2005-06-21 17:14:44 -07003019
David Gibson63551ae2005-06-21 17:14:44 -07003020 while (vaddr < vma->vm_end && remainder) {
Adam Litke4c887262005-10-29 18:16:46 -07003021 pte_t *pte;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003022 spinlock_t *ptl = NULL;
Hugh Dickins2a15efc2009-09-21 17:03:27 -07003023 int absent;
Adam Litke4c887262005-10-29 18:16:46 -07003024 struct page *page;
3025
3026 /*
3027 * Some archs (sparc64, sh*) have multiple pte_ts to
Hugh Dickins2a15efc2009-09-21 17:03:27 -07003028 * each hugepage. We have to make sure we get the
Adam Litke4c887262005-10-29 18:16:46 -07003029 * first, for the page indexing below to work.
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003030 *
3031 * Note that page table lock is not held when pte is null.
Adam Litke4c887262005-10-29 18:16:46 -07003032 */
Andi Kleena5516432008-07-23 21:27:41 -07003033 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h));
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003034 if (pte)
3035 ptl = huge_pte_lock(h, mm, pte);
Hugh Dickins2a15efc2009-09-21 17:03:27 -07003036 absent = !pte || huge_pte_none(huge_ptep_get(pte));
Adam Litke4c887262005-10-29 18:16:46 -07003037
Hugh Dickins2a15efc2009-09-21 17:03:27 -07003038 /*
3039 * When coredumping, it suits get_dump_page if we just return
Hugh Dickins3ae77f42009-09-21 17:03:33 -07003040 * an error where there's an empty slot with no huge pagecache
3041 * to back it. This way, we avoid allocating a hugepage, and
3042 * the sparse dumpfile avoids allocating disk blocks, but its
3043 * huge holes still show up with zeroes where they need to be.
Hugh Dickins2a15efc2009-09-21 17:03:27 -07003044 */
Hugh Dickins3ae77f42009-09-21 17:03:33 -07003045 if (absent && (flags & FOLL_DUMP) &&
3046 !hugetlbfs_pagecache_present(h, vma, vaddr)) {
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003047 if (pte)
3048 spin_unlock(ptl);
Hugh Dickins2a15efc2009-09-21 17:03:27 -07003049 remainder = 0;
3050 break;
3051 }
3052
Naoya Horiguchi9cc3a5b2013-04-17 15:58:30 -07003053 /*
3054 * We need call hugetlb_fault for both hugepages under migration
3055 * (in which case hugetlb_fault waits for the migration,) and
3056 * hwpoisoned hugepages (in which case we need to prevent the
3057 * caller from accessing to them.) In order to do this, we use
3058 * here is_swap_pte instead of is_hugetlb_entry_migration and
3059 * is_hugetlb_entry_hwpoisoned. This is because it simply covers
3060 * both cases, and because we can't follow correct pages
3061 * directly from any kind of swap entries.
3062 */
3063 if (absent || is_swap_pte(huge_ptep_get(pte)) ||
Gerald Schaefer106c9922013-04-29 15:07:23 -07003064 ((flags & FOLL_WRITE) &&
3065 !huge_pte_write(huge_ptep_get(pte)))) {
Adam Litke4c887262005-10-29 18:16:46 -07003066 int ret;
3067
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003068 if (pte)
3069 spin_unlock(ptl);
Hugh Dickins2a15efc2009-09-21 17:03:27 -07003070 ret = hugetlb_fault(mm, vma, vaddr,
3071 (flags & FOLL_WRITE) ? FAULT_FLAG_WRITE : 0);
Adam Litkea89182c2007-08-22 14:01:51 -07003072 if (!(ret & VM_FAULT_ERROR))
Adam Litke4c887262005-10-29 18:16:46 -07003073 continue;
3074
3075 remainder = 0;
Adam Litke4c887262005-10-29 18:16:46 -07003076 break;
3077 }
David Gibson63551ae2005-06-21 17:14:44 -07003078
Andi Kleena5516432008-07-23 21:27:41 -07003079 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07003080 page = pte_page(huge_ptep_get(pte));
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08003081same_page:
Chen, Kenneth Wd6692182006-03-31 02:29:57 -08003082 if (pages) {
Hugh Dickins2a15efc2009-09-21 17:03:27 -07003083 pages[i] = mem_map_offset(page, pfn_offset);
Andrea Arcangelia0368d42014-01-21 15:48:49 -08003084 get_page_foll(pages[i]);
Chen, Kenneth Wd6692182006-03-31 02:29:57 -08003085 }
David Gibson63551ae2005-06-21 17:14:44 -07003086
3087 if (vmas)
3088 vmas[i] = vma;
3089
3090 vaddr += PAGE_SIZE;
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08003091 ++pfn_offset;
David Gibson63551ae2005-06-21 17:14:44 -07003092 --remainder;
3093 ++i;
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08003094 if (vaddr < vma->vm_end && remainder &&
Andi Kleena5516432008-07-23 21:27:41 -07003095 pfn_offset < pages_per_huge_page(h)) {
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08003096 /*
3097 * We use pfn_offset to avoid touching the pageframes
3098 * of this compound page.
3099 */
3100 goto same_page;
3101 }
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003102 spin_unlock(ptl);
David Gibson63551ae2005-06-21 17:14:44 -07003103 }
Michel Lespinasse28a35712013-02-22 16:35:55 -08003104 *nr_pages = remainder;
David Gibson63551ae2005-06-21 17:14:44 -07003105 *position = vaddr;
3106
Hugh Dickins2a15efc2009-09-21 17:03:27 -07003107 return i ? i : -EFAULT;
David Gibson63551ae2005-06-21 17:14:44 -07003108}
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003109
Peter Zijlstra7da4d642012-11-19 03:14:23 +01003110unsigned long hugetlb_change_protection(struct vm_area_struct *vma,
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003111 unsigned long address, unsigned long end, pgprot_t newprot)
3112{
3113 struct mm_struct *mm = vma->vm_mm;
3114 unsigned long start = address;
3115 pte_t *ptep;
3116 pte_t pte;
Andi Kleena5516432008-07-23 21:27:41 -07003117 struct hstate *h = hstate_vma(vma);
Peter Zijlstra7da4d642012-11-19 03:14:23 +01003118 unsigned long pages = 0;
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003119
3120 BUG_ON(address >= end);
3121 flush_cache_range(vma, address, end);
3122
Peter Zijlstra3d48ae42011-05-24 17:12:06 -07003123 mutex_lock(&vma->vm_file->f_mapping->i_mmap_mutex);
Andi Kleena5516432008-07-23 21:27:41 -07003124 for (; address < end; address += huge_page_size(h)) {
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003125 spinlock_t *ptl;
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003126 ptep = huge_pte_offset(mm, address);
3127 if (!ptep)
3128 continue;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003129 ptl = huge_pte_lock(h, mm, ptep);
Peter Zijlstra7da4d642012-11-19 03:14:23 +01003130 if (huge_pmd_unshare(mm, &address, ptep)) {
3131 pages++;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003132 spin_unlock(ptl);
Chen, Kenneth W39dde652006-12-06 20:32:03 -08003133 continue;
Peter Zijlstra7da4d642012-11-19 03:14:23 +01003134 }
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07003135 if (!huge_pte_none(huge_ptep_get(ptep))) {
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003136 pte = huge_ptep_get_and_clear(mm, address, ptep);
Gerald Schaefer106c9922013-04-29 15:07:23 -07003137 pte = pte_mkhuge(huge_pte_modify(pte, newprot));
Tony Lube7517d2013-02-04 14:28:46 -08003138 pte = arch_make_huge_pte(pte, vma, NULL, 0);
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003139 set_huge_pte_at(mm, address, ptep, pte);
Peter Zijlstra7da4d642012-11-19 03:14:23 +01003140 pages++;
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003141 }
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003142 spin_unlock(ptl);
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003143 }
Mel Gormand8333522012-07-31 16:46:20 -07003144 /*
3145 * Must flush TLB before releasing i_mmap_mutex: x86's huge_pmd_unshare
3146 * may have cleared our pud entry and done put_page on the page table:
3147 * once we release i_mmap_mutex, another task can do the final put_page
3148 * and that page table be reused and filled with junk.
3149 */
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003150 flush_tlb_range(vma, start, end);
Mel Gormand8333522012-07-31 16:46:20 -07003151 mutex_unlock(&vma->vm_file->f_mapping->i_mmap_mutex);
Peter Zijlstra7da4d642012-11-19 03:14:23 +01003152
3153 return pages << h->order;
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003154}
3155
Mel Gormana1e78772008-07-23 21:27:23 -07003156int hugetlb_reserve_pages(struct inode *inode,
3157 long from, long to,
Mel Gorman5a6fe122009-02-10 14:02:27 +00003158 struct vm_area_struct *vma,
KOSAKI Motohiroca16d142011-05-26 19:16:19 +09003159 vm_flags_t vm_flags)
Adam Litkee4e574b2007-10-16 01:26:19 -07003160{
Mel Gorman17c9d122009-02-11 16:34:16 +00003161 long ret, chg;
Andi Kleena5516432008-07-23 21:27:41 -07003162 struct hstate *h = hstate_inode(inode);
David Gibson90481622012-03-21 16:34:12 -07003163 struct hugepage_subpool *spool = subpool_inode(inode);
Adam Litkee4e574b2007-10-16 01:26:19 -07003164
Mel Gormana1e78772008-07-23 21:27:23 -07003165 /*
Mel Gorman17c9d122009-02-11 16:34:16 +00003166 * Only apply hugepage reservation if asked. At fault time, an
3167 * attempt will be made for VM_NORESERVE to allocate a page
David Gibson90481622012-03-21 16:34:12 -07003168 * without using reserves
Mel Gorman17c9d122009-02-11 16:34:16 +00003169 */
KOSAKI Motohiroca16d142011-05-26 19:16:19 +09003170 if (vm_flags & VM_NORESERVE)
Mel Gorman17c9d122009-02-11 16:34:16 +00003171 return 0;
3172
3173 /*
Mel Gormana1e78772008-07-23 21:27:23 -07003174 * Shared mappings base their reservation on the number of pages that
3175 * are already allocated on behalf of the file. Private mappings need
3176 * to reserve the full area even if read-only as mprotect() may be
3177 * called to make the mapping read-write. Assume !vma is a shm mapping
3178 */
Mel Gormanf83a2752009-05-28 14:34:40 -07003179 if (!vma || vma->vm_flags & VM_MAYSHARE)
Mel Gormana1e78772008-07-23 21:27:23 -07003180 chg = region_chg(&inode->i_mapping->private_list, from, to);
Mel Gorman5a6fe122009-02-10 14:02:27 +00003181 else {
3182 struct resv_map *resv_map = resv_map_alloc();
Mel Gorman5a6fe122009-02-10 14:02:27 +00003183 if (!resv_map)
3184 return -ENOMEM;
3185
Mel Gorman17c9d122009-02-11 16:34:16 +00003186 chg = to - from;
3187
Mel Gorman5a6fe122009-02-10 14:02:27 +00003188 set_vma_resv_map(vma, resv_map);
3189 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
3190 }
3191
Dave Hansenc50ac052012-05-29 15:06:46 -07003192 if (chg < 0) {
3193 ret = chg;
3194 goto out_err;
3195 }
Mel Gorman17c9d122009-02-11 16:34:16 +00003196
David Gibson90481622012-03-21 16:34:12 -07003197 /* There must be enough pages in the subpool for the mapping */
Dave Hansenc50ac052012-05-29 15:06:46 -07003198 if (hugepage_subpool_get_pages(spool, chg)) {
3199 ret = -ENOSPC;
3200 goto out_err;
3201 }
Mel Gorman17c9d122009-02-11 16:34:16 +00003202
3203 /*
3204 * Check enough hugepages are available for the reservation.
David Gibson90481622012-03-21 16:34:12 -07003205 * Hand the pages back to the subpool if there are not
Mel Gorman17c9d122009-02-11 16:34:16 +00003206 */
3207 ret = hugetlb_acct_memory(h, chg);
3208 if (ret < 0) {
David Gibson90481622012-03-21 16:34:12 -07003209 hugepage_subpool_put_pages(spool, chg);
Dave Hansenc50ac052012-05-29 15:06:46 -07003210 goto out_err;
Mel Gorman17c9d122009-02-11 16:34:16 +00003211 }
3212
3213 /*
3214 * Account for the reservations made. Shared mappings record regions
3215 * that have reservations as they are shared by multiple VMAs.
3216 * When the last VMA disappears, the region map says how much
3217 * the reservation was and the page cache tells how much of
3218 * the reservation was consumed. Private mappings are per-VMA and
3219 * only the consumed reservations are tracked. When the VMA
3220 * disappears, the original reservation is the VMA size and the
3221 * consumed reservations are stored in the map. Hence, nothing
3222 * else has to be done for private mappings here
3223 */
Mel Gormanf83a2752009-05-28 14:34:40 -07003224 if (!vma || vma->vm_flags & VM_MAYSHARE)
Mel Gorman17c9d122009-02-11 16:34:16 +00003225 region_add(&inode->i_mapping->private_list, from, to);
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -07003226 return 0;
Dave Hansenc50ac052012-05-29 15:06:46 -07003227out_err:
Dave Hansen4523e142012-05-30 07:51:07 -07003228 if (vma)
3229 resv_map_put(vma);
Dave Hansenc50ac052012-05-29 15:06:46 -07003230 return ret;
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -07003231}
3232
3233void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed)
3234{
Andi Kleena5516432008-07-23 21:27:41 -07003235 struct hstate *h = hstate_inode(inode);
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -07003236 long chg = region_truncate(&inode->i_mapping->private_list, offset);
David Gibson90481622012-03-21 16:34:12 -07003237 struct hugepage_subpool *spool = subpool_inode(inode);
Ken Chen45c682a2007-11-14 16:59:44 -08003238
3239 spin_lock(&inode->i_lock);
Eric Sandeene4c6f8b2009-07-29 15:02:16 -07003240 inode->i_blocks -= (blocks_per_huge_page(h) * freed);
Ken Chen45c682a2007-11-14 16:59:44 -08003241 spin_unlock(&inode->i_lock);
3242
David Gibson90481622012-03-21 16:34:12 -07003243 hugepage_subpool_put_pages(spool, (chg - freed));
Andi Kleena5516432008-07-23 21:27:41 -07003244 hugetlb_acct_memory(h, -(chg - freed));
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -07003245}
Naoya Horiguchi93f70f92010-05-28 09:29:20 +09003246
Steve Capper3212b532013-04-23 12:35:02 +01003247#ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
3248static unsigned long page_table_shareable(struct vm_area_struct *svma,
3249 struct vm_area_struct *vma,
3250 unsigned long addr, pgoff_t idx)
3251{
3252 unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) +
3253 svma->vm_start;
3254 unsigned long sbase = saddr & PUD_MASK;
3255 unsigned long s_end = sbase + PUD_SIZE;
3256
3257 /* Allow segments to share if only one is marked locked */
3258 unsigned long vm_flags = vma->vm_flags & ~VM_LOCKED;
3259 unsigned long svm_flags = svma->vm_flags & ~VM_LOCKED;
3260
3261 /*
3262 * match the virtual addresses, permission and the alignment of the
3263 * page table page.
3264 */
3265 if (pmd_index(addr) != pmd_index(saddr) ||
3266 vm_flags != svm_flags ||
3267 sbase < svma->vm_start || svma->vm_end < s_end)
3268 return 0;
3269
3270 return saddr;
3271}
3272
3273static int vma_shareable(struct vm_area_struct *vma, unsigned long addr)
3274{
3275 unsigned long base = addr & PUD_MASK;
3276 unsigned long end = base + PUD_SIZE;
3277
3278 /*
3279 * check on proper vm_flags and page table alignment
3280 */
3281 if (vma->vm_flags & VM_MAYSHARE &&
3282 vma->vm_start <= base && end <= vma->vm_end)
3283 return 1;
3284 return 0;
3285}
3286
3287/*
3288 * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
3289 * and returns the corresponding pte. While this is not necessary for the
3290 * !shared pmd case because we can allocate the pmd later as well, it makes the
3291 * code much cleaner. pmd allocation is essential for the shared case because
3292 * pud has to be populated inside the same i_mmap_mutex section - otherwise
3293 * racing tasks could either miss the sharing (see huge_pte_offset) or select a
3294 * bad pmd for sharing.
3295 */
3296pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
3297{
3298 struct vm_area_struct *vma = find_vma(mm, addr);
3299 struct address_space *mapping = vma->vm_file->f_mapping;
3300 pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) +
3301 vma->vm_pgoff;
3302 struct vm_area_struct *svma;
3303 unsigned long saddr;
3304 pte_t *spte = NULL;
3305 pte_t *pte;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003306 spinlock_t *ptl;
Steve Capper3212b532013-04-23 12:35:02 +01003307
3308 if (!vma_shareable(vma, addr))
3309 return (pte_t *)pmd_alloc(mm, pud, addr);
3310
3311 mutex_lock(&mapping->i_mmap_mutex);
3312 vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) {
3313 if (svma == vma)
3314 continue;
3315
3316 saddr = page_table_shareable(svma, vma, addr, idx);
3317 if (saddr) {
3318 spte = huge_pte_offset(svma->vm_mm, saddr);
3319 if (spte) {
3320 get_page(virt_to_page(spte));
3321 break;
3322 }
3323 }
3324 }
3325
3326 if (!spte)
3327 goto out;
3328
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003329 ptl = huge_pte_lockptr(hstate_vma(vma), mm, spte);
3330 spin_lock(ptl);
Steve Capper3212b532013-04-23 12:35:02 +01003331 if (pud_none(*pud))
3332 pud_populate(mm, pud,
3333 (pmd_t *)((unsigned long)spte & PAGE_MASK));
3334 else
3335 put_page(virt_to_page(spte));
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003336 spin_unlock(ptl);
Steve Capper3212b532013-04-23 12:35:02 +01003337out:
3338 pte = (pte_t *)pmd_alloc(mm, pud, addr);
3339 mutex_unlock(&mapping->i_mmap_mutex);
3340 return pte;
3341}
3342
3343/*
3344 * unmap huge page backed by shared pte.
3345 *
3346 * Hugetlb pte page is ref counted at the time of mapping. If pte is shared
3347 * indicated by page_count > 1, unmap is achieved by clearing pud and
3348 * decrementing the ref count. If count == 1, the pte page is not shared.
3349 *
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003350 * called with page table lock held.
Steve Capper3212b532013-04-23 12:35:02 +01003351 *
3352 * returns: 1 successfully unmapped a shared pte page
3353 * 0 the underlying pte page is not shared, or it is the last user
3354 */
3355int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
3356{
3357 pgd_t *pgd = pgd_offset(mm, *addr);
3358 pud_t *pud = pud_offset(pgd, *addr);
3359
3360 BUG_ON(page_count(virt_to_page(ptep)) == 0);
3361 if (page_count(virt_to_page(ptep)) == 1)
3362 return 0;
3363
3364 pud_clear(pud);
3365 put_page(virt_to_page(ptep));
3366 *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE;
3367 return 1;
3368}
Steve Capper9e5fc742013-04-30 08:02:03 +01003369#define want_pmd_share() (1)
3370#else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
3371pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
3372{
3373 return NULL;
3374}
3375#define want_pmd_share() (0)
Steve Capper3212b532013-04-23 12:35:02 +01003376#endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
3377
Steve Capper9e5fc742013-04-30 08:02:03 +01003378#ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
3379pte_t *huge_pte_alloc(struct mm_struct *mm,
3380 unsigned long addr, unsigned long sz)
3381{
3382 pgd_t *pgd;
3383 pud_t *pud;
3384 pte_t *pte = NULL;
3385
3386 pgd = pgd_offset(mm, addr);
3387 pud = pud_alloc(mm, pgd, addr);
3388 if (pud) {
3389 if (sz == PUD_SIZE) {
3390 pte = (pte_t *)pud;
3391 } else {
3392 BUG_ON(sz != PMD_SIZE);
3393 if (want_pmd_share() && pud_none(*pud))
3394 pte = huge_pmd_share(mm, addr, pud);
3395 else
3396 pte = (pte_t *)pmd_alloc(mm, pud, addr);
3397 }
3398 }
3399 BUG_ON(pte && !pte_none(*pte) && !pte_huge(*pte));
3400
3401 return pte;
3402}
3403
3404pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr)
3405{
3406 pgd_t *pgd;
3407 pud_t *pud;
3408 pmd_t *pmd = NULL;
3409
3410 pgd = pgd_offset(mm, addr);
3411 if (pgd_present(*pgd)) {
3412 pud = pud_offset(pgd, addr);
3413 if (pud_present(*pud)) {
3414 if (pud_huge(*pud))
3415 return (pte_t *)pud;
3416 pmd = pmd_offset(pud, addr);
3417 }
3418 }
3419 return (pte_t *) pmd;
3420}
3421
3422struct page *
3423follow_huge_pmd(struct mm_struct *mm, unsigned long address,
3424 pmd_t *pmd, int write)
3425{
3426 struct page *page;
3427
3428 page = pte_page(*(pte_t *)pmd);
3429 if (page)
3430 page += ((address & ~PMD_MASK) >> PAGE_SHIFT);
3431 return page;
3432}
3433
3434struct page *
3435follow_huge_pud(struct mm_struct *mm, unsigned long address,
3436 pud_t *pud, int write)
3437{
3438 struct page *page;
3439
3440 page = pte_page(*(pte_t *)pud);
3441 if (page)
3442 page += ((address & ~PUD_MASK) >> PAGE_SHIFT);
3443 return page;
3444}
3445
3446#else /* !CONFIG_ARCH_WANT_GENERAL_HUGETLB */
3447
3448/* Can be overriden by architectures */
3449__attribute__((weak)) struct page *
3450follow_huge_pud(struct mm_struct *mm, unsigned long address,
3451 pud_t *pud, int write)
3452{
3453 BUG();
3454 return NULL;
3455}
3456
3457#endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
3458
Andi Kleend5bd9102010-09-27 09:00:12 +02003459#ifdef CONFIG_MEMORY_FAILURE
3460
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +09003461/* Should be called in hugetlb_lock */
3462static int is_hugepage_on_freelist(struct page *hpage)
3463{
3464 struct page *page;
3465 struct page *tmp;
3466 struct hstate *h = page_hstate(hpage);
3467 int nid = page_to_nid(hpage);
3468
3469 list_for_each_entry_safe(page, tmp, &h->hugepage_freelists[nid], lru)
3470 if (page == hpage)
3471 return 1;
3472 return 0;
3473}
3474
Naoya Horiguchi93f70f92010-05-28 09:29:20 +09003475/*
3476 * This function is called from memory failure code.
3477 * Assume the caller holds page lock of the head page.
3478 */
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +09003479int dequeue_hwpoisoned_huge_page(struct page *hpage)
Naoya Horiguchi93f70f92010-05-28 09:29:20 +09003480{
3481 struct hstate *h = page_hstate(hpage);
3482 int nid = page_to_nid(hpage);
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +09003483 int ret = -EBUSY;
Naoya Horiguchi93f70f92010-05-28 09:29:20 +09003484
3485 spin_lock(&hugetlb_lock);
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +09003486 if (is_hugepage_on_freelist(hpage)) {
Naoya Horiguchi56f2fb12012-12-12 13:52:33 -08003487 /*
3488 * Hwpoisoned hugepage isn't linked to activelist or freelist,
3489 * but dangling hpage->lru can trigger list-debug warnings
3490 * (this happens when we call unpoison_memory() on it),
3491 * so let it point to itself with list_del_init().
3492 */
3493 list_del_init(&hpage->lru);
Naoya Horiguchi8c6c2ec2010-09-08 10:19:38 +09003494 set_page_refcounted(hpage);
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +09003495 h->free_huge_pages--;
3496 h->free_huge_pages_node[nid]--;
3497 ret = 0;
3498 }
Naoya Horiguchi93f70f92010-05-28 09:29:20 +09003499 spin_unlock(&hugetlb_lock);
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +09003500 return ret;
Naoya Horiguchi93f70f92010-05-28 09:29:20 +09003501}
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +09003502#endif
Naoya Horiguchi31caf662013-09-11 14:21:59 -07003503
3504bool isolate_huge_page(struct page *page, struct list_head *list)
3505{
Sasha Levin309381fea2014-01-23 15:52:54 -08003506 VM_BUG_ON_PAGE(!PageHead(page), page);
Naoya Horiguchi31caf662013-09-11 14:21:59 -07003507 if (!get_page_unless_zero(page))
3508 return false;
3509 spin_lock(&hugetlb_lock);
3510 list_move_tail(&page->lru, list);
3511 spin_unlock(&hugetlb_lock);
3512 return true;
3513}
3514
3515void putback_active_hugepage(struct page *page)
3516{
Sasha Levin309381fea2014-01-23 15:52:54 -08003517 VM_BUG_ON_PAGE(!PageHead(page), page);
Naoya Horiguchi31caf662013-09-11 14:21:59 -07003518 spin_lock(&hugetlb_lock);
3519 list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist);
3520 spin_unlock(&hugetlb_lock);
3521 put_page(page);
3522}
Naoya Horiguchic8721bb2013-09-11 14:22:09 -07003523
3524bool is_hugepage_active(struct page *page)
3525{
Sasha Levin309381fea2014-01-23 15:52:54 -08003526 VM_BUG_ON_PAGE(!PageHuge(page), page);
Naoya Horiguchic8721bb2013-09-11 14:22:09 -07003527 /*
3528 * This function can be called for a tail page because the caller,
3529 * scan_movable_pages, scans through a given pfn-range which typically
3530 * covers one memory block. In systems using gigantic hugepage (1GB
3531 * for x86_64,) a hugepage is larger than a memory block, and we don't
3532 * support migrating such large hugepages for now, so return false
3533 * when called for tail pages.
3534 */
3535 if (PageTail(page))
3536 return false;
3537 /*
3538 * Refcount of a hwpoisoned hugepages is 1, but they are not active,
3539 * so we should return false for them.
3540 */
3541 if (unlikely(PageHWPoison(page)))
3542 return false;
3543 return page_count(page) > 0;
3544}