blob: dee6cf4e6d34135e1880c5c01c7627aa1a33c69a [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
2 * Generic hugetlb support.
Nadia Yvette Chambers6d49e352012-12-06 10:39:54 +01003 * (C) Nadia Yvette Chambers, April 2004
Linus Torvalds1da177e2005-04-16 15:20:36 -07004 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07005#include <linux/list.h>
6#include <linux/init.h>
7#include <linux/module.h>
8#include <linux/mm.h>
Alexey Dobriyane1759c22008-10-15 23:50:22 +04009#include <linux/seq_file.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070010#include <linux/sysctl.h>
11#include <linux/highmem.h>
Andrea Arcangelicddb8a52008-07-28 15:46:29 -070012#include <linux/mmu_notifier.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070013#include <linux/nodemask.h>
David Gibson63551ae2005-06-21 17:14:44 -070014#include <linux/pagemap.h>
Christoph Lameter5da7ca82006-01-06 00:10:46 -080015#include <linux/mempolicy.h>
Christoph Lameteraea47ff2006-01-08 01:00:57 -080016#include <linux/cpuset.h>
David Gibson3935baa2006-03-22 00:08:53 -080017#include <linux/mutex.h>
Andi Kleenaa888a72008-07-23 21:27:47 -070018#include <linux/bootmem.h>
Nishanth Aravamudana3437872008-07-23 21:27:44 -070019#include <linux/sysfs.h>
Tejun Heo5a0e3ad2010-03-24 17:04:11 +090020#include <linux/slab.h>
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +090021#include <linux/rmap.h>
Naoya Horiguchifd6a03ed2010-05-28 09:29:21 +090022#include <linux/swap.h>
23#include <linux/swapops.h>
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 Schermerhorn9a305232009-12-14 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 }
Aneesh Kumar K.V9dd540e2012-07-31 16:42:15 -0700587 VM_BUG_ON(hugetlb_cgroup_from_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{
693 compound_page_dtor *dtor;
694
695 if (!PageCompound(page))
696 return 0;
697
698 page = compound_head(page);
699 dtor = get_compound_page_dtor(page);
700
701 return dtor == free_huge_page;
702}
Naoya Horiguchi43131e12010-05-28 09:29:22 +0900703EXPORT_SYMBOL_GPL(PageHuge);
704
Andrea Arcangeli27c73ae2013-11-21 14:32:02 -0800705/*
706 * PageHeadHuge() only returns true for hugetlbfs head page, but not for
707 * normal or transparent huge pages.
708 */
709int PageHeadHuge(struct page *page_head)
710{
711 compound_page_dtor *dtor;
712
713 if (!PageHead(page_head))
714 return 0;
715
716 dtor = get_compound_page_dtor(page_head);
717
718 return dtor == free_huge_page;
719}
720EXPORT_SYMBOL_GPL(PageHeadHuge);
721
Zhang Yi13d60f42013-06-25 21:19:31 +0800722pgoff_t __basepage_index(struct page *page)
723{
724 struct page *page_head = compound_head(page);
725 pgoff_t index = page_index(page_head);
726 unsigned long compound_idx;
727
728 if (!PageHuge(page_head))
729 return page_index(page);
730
731 if (compound_order(page_head) >= MAX_ORDER)
732 compound_idx = page_to_pfn(page) - page_to_pfn(page_head);
733 else
734 compound_idx = page - page_head;
735
736 return (index << compound_order(page_head)) + compound_idx;
737}
738
Andi Kleena5516432008-07-23 21:27:41 -0700739static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700740{
Linus Torvalds1da177e2005-04-16 15:20:36 -0700741 struct page *page;
Joe Jinf96efd52007-07-15 23:38:12 -0700742
Andi Kleenaa888a72008-07-23 21:27:47 -0700743 if (h->order >= MAX_ORDER)
744 return NULL;
745
Mel Gorman6484eb32009-06-16 15:31:54 -0700746 page = alloc_pages_exact_node(nid,
Naoya Horiguchi86cdb462013-09-11 14:22:13 -0700747 htlb_alloc_mask(h)|__GFP_COMP|__GFP_THISNODE|
Nishanth Aravamudan551883a2008-04-29 00:58:26 -0700748 __GFP_REPEAT|__GFP_NOWARN,
Andi Kleena5516432008-07-23 21:27:41 -0700749 huge_page_order(h));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700750 if (page) {
Gerald Schaefer7f2e9522008-04-28 02:13:29 -0700751 if (arch_prepare_hugepage(page)) {
Gerald Schaefercaff3a22008-08-12 15:08:38 -0700752 __free_pages(page, huge_page_order(h));
Harvey Harrison7b8ee842008-04-28 14:13:19 -0700753 return NULL;
Gerald Schaefer7f2e9522008-04-28 02:13:29 -0700754 }
Andi Kleena5516432008-07-23 21:27:41 -0700755 prep_new_huge_page(h, page, nid);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700756 }
Nishanth Aravamudan63b46132007-10-16 01:26:24 -0700757
758 return page;
759}
760
Andi Kleen5ced66c2008-07-23 21:27:45 -0700761/*
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800762 * common helper functions for hstate_next_node_to_{alloc|free}.
763 * We may have allocated or freed a huge page based on a different
764 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
765 * be outside of *nodes_allowed. Ensure that we use an allowed
766 * node for alloc or free.
Lee Schermerhorn9a76db02009-12-14 17:58:15 -0800767 */
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800768static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
Lee Schermerhorn9a76db02009-12-14 17:58:15 -0800769{
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800770 nid = next_node(nid, *nodes_allowed);
Lee Schermerhorn9a76db02009-12-14 17:58:15 -0800771 if (nid == MAX_NUMNODES)
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800772 nid = first_node(*nodes_allowed);
Lee Schermerhorn9a76db02009-12-14 17:58:15 -0800773 VM_BUG_ON(nid >= MAX_NUMNODES);
774
775 return nid;
776}
777
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800778static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
Andi Kleen5ced66c2008-07-23 21:27:45 -0700779{
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800780 if (!node_isset(nid, *nodes_allowed))
781 nid = next_node_allowed(nid, nodes_allowed);
Lee Schermerhorn9a76db02009-12-14 17:58:15 -0800782 return nid;
Andi Kleen5ced66c2008-07-23 21:27:45 -0700783}
784
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800785/*
786 * returns the previously saved node ["this node"] from which to
787 * allocate a persistent huge page for the pool and advance the
788 * next node from which to allocate, handling wrap at end of node
789 * mask.
790 */
791static int hstate_next_node_to_alloc(struct hstate *h,
792 nodemask_t *nodes_allowed)
793{
794 int nid;
795
796 VM_BUG_ON(!nodes_allowed);
797
798 nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
799 h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
800
801 return nid;
802}
803
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700804/*
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800805 * helper for free_pool_huge_page() - return the previously saved
806 * node ["this node"] from which to free a huge page. Advance the
807 * next node id whether or not we find a free huge page to free so
808 * that the next attempt to free addresses the next node.
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700809 */
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800810static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700811{
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800812 int nid;
Lee Schermerhorn9a76db02009-12-14 17:58:15 -0800813
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800814 VM_BUG_ON(!nodes_allowed);
815
816 nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
817 h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
818
Lee Schermerhorn9a76db02009-12-14 17:58:15 -0800819 return nid;
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700820}
821
Joonsoo Kimb2261022013-09-11 14:21:00 -0700822#define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \
823 for (nr_nodes = nodes_weight(*mask); \
824 nr_nodes > 0 && \
825 ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \
826 nr_nodes--)
827
828#define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \
829 for (nr_nodes = nodes_weight(*mask); \
830 nr_nodes > 0 && \
831 ((node = hstate_next_node_to_free(hs, mask)) || 1); \
832 nr_nodes--)
833
834static int alloc_fresh_huge_page(struct hstate *h, nodemask_t *nodes_allowed)
835{
836 struct page *page;
837 int nr_nodes, node;
838 int ret = 0;
839
840 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
841 page = alloc_fresh_huge_page_node(h, node);
842 if (page) {
843 ret = 1;
844 break;
845 }
846 }
847
848 if (ret)
849 count_vm_event(HTLB_BUDDY_PGALLOC);
850 else
851 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
852
853 return ret;
854}
855
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700856/*
857 * Free huge page from pool from next node to free.
858 * Attempt to keep persistent huge pages more or less
859 * balanced over allowed nodes.
860 * Called with hugetlb_lock locked.
861 */
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800862static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
863 bool acct_surplus)
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700864{
Joonsoo Kimb2261022013-09-11 14:21:00 -0700865 int nr_nodes, node;
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700866 int ret = 0;
867
Joonsoo Kimb2261022013-09-11 14:21:00 -0700868 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
Lee Schermerhorn685f3452009-09-21 17:01:23 -0700869 /*
870 * If we're returning unused surplus pages, only examine
871 * nodes with surplus pages.
872 */
Joonsoo Kimb2261022013-09-11 14:21:00 -0700873 if ((!acct_surplus || h->surplus_huge_pages_node[node]) &&
874 !list_empty(&h->hugepage_freelists[node])) {
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700875 struct page *page =
Joonsoo Kimb2261022013-09-11 14:21:00 -0700876 list_entry(h->hugepage_freelists[node].next,
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700877 struct page, lru);
878 list_del(&page->lru);
879 h->free_huge_pages--;
Joonsoo Kimb2261022013-09-11 14:21:00 -0700880 h->free_huge_pages_node[node]--;
Lee Schermerhorn685f3452009-09-21 17:01:23 -0700881 if (acct_surplus) {
882 h->surplus_huge_pages--;
Joonsoo Kimb2261022013-09-11 14:21:00 -0700883 h->surplus_huge_pages_node[node]--;
Lee Schermerhorn685f3452009-09-21 17:01:23 -0700884 }
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700885 update_and_free_page(h, page);
886 ret = 1;
Lee Schermerhorn9a76db02009-12-14 17:58:15 -0800887 break;
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700888 }
Joonsoo Kimb2261022013-09-11 14:21:00 -0700889 }
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700890
891 return ret;
892}
893
Naoya Horiguchic8721bb2013-09-11 14:22:09 -0700894/*
895 * Dissolve a given free hugepage into free buddy pages. This function does
896 * nothing for in-use (including surplus) hugepages.
897 */
898static void dissolve_free_huge_page(struct page *page)
899{
900 spin_lock(&hugetlb_lock);
901 if (PageHuge(page) && !page_count(page)) {
902 struct hstate *h = page_hstate(page);
903 int nid = page_to_nid(page);
904 list_del(&page->lru);
905 h->free_huge_pages--;
906 h->free_huge_pages_node[nid]--;
907 update_and_free_page(h, page);
908 }
909 spin_unlock(&hugetlb_lock);
910}
911
912/*
913 * Dissolve free hugepages in a given pfn range. Used by memory hotplug to
914 * make specified memory blocks removable from the system.
915 * Note that start_pfn should aligned with (minimum) hugepage size.
916 */
917void dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn)
918{
919 unsigned int order = 8 * sizeof(void *);
920 unsigned long pfn;
921 struct hstate *h;
922
923 /* Set scan step to minimum hugepage size */
924 for_each_hstate(h)
925 if (order > huge_page_order(h))
926 order = huge_page_order(h);
927 VM_BUG_ON(!IS_ALIGNED(start_pfn, 1 << order));
928 for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << order)
929 dissolve_free_huge_page(pfn_to_page(pfn));
930}
931
Naoya Horiguchibf50bab2010-09-08 10:19:33 +0900932static struct page *alloc_buddy_huge_page(struct hstate *h, int nid)
Adam Litke7893d1d2007-10-16 01:26:18 -0700933{
934 struct page *page;
Naoya Horiguchibf50bab2010-09-08 10:19:33 +0900935 unsigned int r_nid;
Adam Litke7893d1d2007-10-16 01:26:18 -0700936
Andi Kleenaa888a72008-07-23 21:27:47 -0700937 if (h->order >= MAX_ORDER)
938 return NULL;
939
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800940 /*
941 * Assume we will successfully allocate the surplus page to
942 * prevent racing processes from causing the surplus to exceed
943 * overcommit
944 *
945 * This however introduces a different race, where a process B
946 * tries to grow the static hugepage pool while alloc_pages() is
947 * called by process A. B will only examine the per-node
948 * counters in determining if surplus huge pages can be
949 * converted to normal huge pages in adjust_pool_surplus(). A
950 * won't be able to increment the per-node counter, until the
951 * lock is dropped by B, but B doesn't drop hugetlb_lock until
952 * no more huge pages can be converted from surplus to normal
953 * state (and doesn't try to convert again). Thus, we have a
954 * case where a surplus huge page exists, the pool is grown, and
955 * the surplus huge page still exists after, even though it
956 * should just have been converted to a normal huge page. This
957 * does not leak memory, though, as the hugepage will be freed
958 * once it is out of use. It also does not allow the counters to
959 * go out of whack in adjust_pool_surplus() as we don't modify
960 * the node values until we've gotten the hugepage and only the
961 * per-node value is checked there.
962 */
963 spin_lock(&hugetlb_lock);
Andi Kleena5516432008-07-23 21:27:41 -0700964 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800965 spin_unlock(&hugetlb_lock);
966 return NULL;
967 } else {
Andi Kleena5516432008-07-23 21:27:41 -0700968 h->nr_huge_pages++;
969 h->surplus_huge_pages++;
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800970 }
971 spin_unlock(&hugetlb_lock);
972
Naoya Horiguchibf50bab2010-09-08 10:19:33 +0900973 if (nid == NUMA_NO_NODE)
Naoya Horiguchi86cdb462013-09-11 14:22:13 -0700974 page = alloc_pages(htlb_alloc_mask(h)|__GFP_COMP|
Naoya Horiguchibf50bab2010-09-08 10:19:33 +0900975 __GFP_REPEAT|__GFP_NOWARN,
976 huge_page_order(h));
977 else
978 page = alloc_pages_exact_node(nid,
Naoya Horiguchi86cdb462013-09-11 14:22:13 -0700979 htlb_alloc_mask(h)|__GFP_COMP|__GFP_THISNODE|
Naoya Horiguchibf50bab2010-09-08 10:19:33 +0900980 __GFP_REPEAT|__GFP_NOWARN, huge_page_order(h));
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800981
Gerald Schaefercaff3a22008-08-12 15:08:38 -0700982 if (page && arch_prepare_hugepage(page)) {
983 __free_pages(page, huge_page_order(h));
Hillf Dantonea5768c2012-01-10 15:08:30 -0800984 page = NULL;
Gerald Schaefercaff3a22008-08-12 15:08:38 -0700985 }
986
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800987 spin_lock(&hugetlb_lock);
Adam Litke7893d1d2007-10-16 01:26:18 -0700988 if (page) {
Aneesh Kumar K.V0edaecf2012-07-31 16:42:07 -0700989 INIT_LIST_HEAD(&page->lru);
Naoya Horiguchibf50bab2010-09-08 10:19:33 +0900990 r_nid = page_to_nid(page);
Adam Litke7893d1d2007-10-16 01:26:18 -0700991 set_compound_page_dtor(page, free_huge_page);
Aneesh Kumar K.V9dd540e2012-07-31 16:42:15 -0700992 set_hugetlb_cgroup(page, NULL);
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800993 /*
994 * We incremented the global counters already
995 */
Naoya Horiguchibf50bab2010-09-08 10:19:33 +0900996 h->nr_huge_pages_node[r_nid]++;
997 h->surplus_huge_pages_node[r_nid]++;
Adam Litke3b116302008-04-28 02:13:06 -0700998 __count_vm_event(HTLB_BUDDY_PGALLOC);
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800999 } else {
Andi Kleena5516432008-07-23 21:27:41 -07001000 h->nr_huge_pages--;
1001 h->surplus_huge_pages--;
Adam Litke3b116302008-04-28 02:13:06 -07001002 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
Adam Litke7893d1d2007-10-16 01:26:18 -07001003 }
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001004 spin_unlock(&hugetlb_lock);
Adam Litke7893d1d2007-10-16 01:26:18 -07001005
1006 return page;
1007}
1008
Adam Litkee4e574b2007-10-16 01:26:19 -07001009/*
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001010 * This allocation function is useful in the context where vma is irrelevant.
1011 * E.g. soft-offlining uses this function because it only cares physical
1012 * address of error page.
1013 */
1014struct page *alloc_huge_page_node(struct hstate *h, int nid)
1015{
Joonsoo Kim4ef91842013-09-11 14:21:51 -07001016 struct page *page = NULL;
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001017
1018 spin_lock(&hugetlb_lock);
Joonsoo Kim4ef91842013-09-11 14:21:51 -07001019 if (h->free_huge_pages - h->resv_huge_pages > 0)
1020 page = dequeue_huge_page_node(h, nid);
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001021 spin_unlock(&hugetlb_lock);
1022
Aneesh Kumar K.V94ae8ba2012-07-31 16:42:35 -07001023 if (!page)
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001024 page = alloc_buddy_huge_page(h, nid);
1025
1026 return page;
1027}
1028
1029/*
Lucas De Marchi25985ed2011-03-30 22:57:33 -03001030 * Increase the hugetlb pool such that it can accommodate a reservation
Adam Litkee4e574b2007-10-16 01:26:19 -07001031 * of size 'delta'.
1032 */
Andi Kleena5516432008-07-23 21:27:41 -07001033static int gather_surplus_pages(struct hstate *h, int delta)
Adam Litkee4e574b2007-10-16 01:26:19 -07001034{
1035 struct list_head surplus_list;
1036 struct page *page, *tmp;
1037 int ret, i;
1038 int needed, allocated;
Hillf Danton28073b02012-03-21 16:34:00 -07001039 bool alloc_ok = true;
Adam Litkee4e574b2007-10-16 01:26:19 -07001040
Andi Kleena5516432008-07-23 21:27:41 -07001041 needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
Adam Litkeac09b3a2008-03-04 14:29:38 -08001042 if (needed <= 0) {
Andi Kleena5516432008-07-23 21:27:41 -07001043 h->resv_huge_pages += delta;
Adam Litkee4e574b2007-10-16 01:26:19 -07001044 return 0;
Adam Litkeac09b3a2008-03-04 14:29:38 -08001045 }
Adam Litkee4e574b2007-10-16 01:26:19 -07001046
1047 allocated = 0;
1048 INIT_LIST_HEAD(&surplus_list);
1049
1050 ret = -ENOMEM;
1051retry:
1052 spin_unlock(&hugetlb_lock);
1053 for (i = 0; i < needed; i++) {
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001054 page = alloc_buddy_huge_page(h, NUMA_NO_NODE);
Hillf Danton28073b02012-03-21 16:34:00 -07001055 if (!page) {
1056 alloc_ok = false;
1057 break;
1058 }
Adam Litkee4e574b2007-10-16 01:26:19 -07001059 list_add(&page->lru, &surplus_list);
1060 }
Hillf Danton28073b02012-03-21 16:34:00 -07001061 allocated += i;
Adam Litkee4e574b2007-10-16 01:26:19 -07001062
1063 /*
1064 * After retaking hugetlb_lock, we need to recalculate 'needed'
1065 * because either resv_huge_pages or free_huge_pages may have changed.
1066 */
1067 spin_lock(&hugetlb_lock);
Andi Kleena5516432008-07-23 21:27:41 -07001068 needed = (h->resv_huge_pages + delta) -
1069 (h->free_huge_pages + allocated);
Hillf Danton28073b02012-03-21 16:34:00 -07001070 if (needed > 0) {
1071 if (alloc_ok)
1072 goto retry;
1073 /*
1074 * We were not able to allocate enough pages to
1075 * satisfy the entire reservation so we free what
1076 * we've allocated so far.
1077 */
1078 goto free;
1079 }
Adam Litkee4e574b2007-10-16 01:26:19 -07001080 /*
1081 * The surplus_list now contains _at_least_ the number of extra pages
Lucas De Marchi25985ed2011-03-30 22:57:33 -03001082 * needed to accommodate the reservation. Add the appropriate number
Adam Litkee4e574b2007-10-16 01:26:19 -07001083 * of pages to the hugetlb pool and free the extras back to the buddy
Adam Litkeac09b3a2008-03-04 14:29:38 -08001084 * allocator. Commit the entire reservation here to prevent another
1085 * process from stealing the pages as they are added to the pool but
1086 * before they are reserved.
Adam Litkee4e574b2007-10-16 01:26:19 -07001087 */
1088 needed += allocated;
Andi Kleena5516432008-07-23 21:27:41 -07001089 h->resv_huge_pages += delta;
Adam Litkee4e574b2007-10-16 01:26:19 -07001090 ret = 0;
Naoya Horiguchia9869b82010-09-08 10:19:37 +09001091
Adam Litke19fc3f02008-04-28 02:12:20 -07001092 /* Free the needed pages to the hugetlb pool */
Adam Litkee4e574b2007-10-16 01:26:19 -07001093 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
Adam Litke19fc3f02008-04-28 02:12:20 -07001094 if ((--needed) < 0)
1095 break;
Naoya Horiguchia9869b82010-09-08 10:19:37 +09001096 /*
1097 * This page is now managed by the hugetlb allocator and has
1098 * no users -- drop the buddy allocator's reference.
1099 */
1100 put_page_testzero(page);
1101 VM_BUG_ON(page_count(page));
Andi Kleena5516432008-07-23 21:27:41 -07001102 enqueue_huge_page(h, page);
Adam Litke19fc3f02008-04-28 02:12:20 -07001103 }
Hillf Danton28073b02012-03-21 16:34:00 -07001104free:
Hillf Dantonb0365c82011-12-28 15:57:16 -08001105 spin_unlock(&hugetlb_lock);
Adam Litke19fc3f02008-04-28 02:12:20 -07001106
1107 /* Free unnecessary surplus pages to the buddy allocator */
Joonsoo Kimc0d934b2013-09-11 14:21:02 -07001108 list_for_each_entry_safe(page, tmp, &surplus_list, lru)
1109 put_page(page);
Naoya Horiguchia9869b82010-09-08 10:19:37 +09001110 spin_lock(&hugetlb_lock);
Adam Litkee4e574b2007-10-16 01:26:19 -07001111
1112 return ret;
1113}
1114
1115/*
1116 * When releasing a hugetlb pool reservation, any surplus pages that were
1117 * allocated to satisfy the reservation must be explicitly freed if they were
1118 * never used.
Lee Schermerhorn685f3452009-09-21 17:01:23 -07001119 * Called with hugetlb_lock held.
Adam Litkee4e574b2007-10-16 01:26:19 -07001120 */
Andi Kleena5516432008-07-23 21:27:41 -07001121static void return_unused_surplus_pages(struct hstate *h,
1122 unsigned long unused_resv_pages)
Adam Litkee4e574b2007-10-16 01:26:19 -07001123{
Adam Litkee4e574b2007-10-16 01:26:19 -07001124 unsigned long nr_pages;
1125
Adam Litkeac09b3a2008-03-04 14:29:38 -08001126 /* Uncommit the reservation */
Andi Kleena5516432008-07-23 21:27:41 -07001127 h->resv_huge_pages -= unused_resv_pages;
Adam Litkeac09b3a2008-03-04 14:29:38 -08001128
Andi Kleenaa888a72008-07-23 21:27:47 -07001129 /* Cannot return gigantic pages currently */
1130 if (h->order >= MAX_ORDER)
1131 return;
1132
Andi Kleena5516432008-07-23 21:27:41 -07001133 nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
Adam Litkee4e574b2007-10-16 01:26:19 -07001134
Lee Schermerhorn685f3452009-09-21 17:01:23 -07001135 /*
1136 * We want to release as many surplus pages as possible, spread
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -08001137 * evenly across all nodes with memory. Iterate across these nodes
1138 * until we can no longer free unreserved surplus pages. This occurs
1139 * when the nodes with surplus pages have no free pages.
1140 * free_pool_huge_page() will balance the the freed pages across the
1141 * on-line nodes with memory and will handle the hstate accounting.
Lee Schermerhorn685f3452009-09-21 17:01:23 -07001142 */
1143 while (nr_pages--) {
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08001144 if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1))
Lee Schermerhorn685f3452009-09-21 17:01:23 -07001145 break;
Adam Litkee4e574b2007-10-16 01:26:19 -07001146 }
1147}
1148
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001149/*
1150 * Determine if the huge page at addr within the vma has an associated
1151 * reservation. Where it does not we will need to logically increase
David Gibson90481622012-03-21 16:34:12 -07001152 * reservation and actually increase subpool usage before an allocation
1153 * can occur. Where any new reservation would be required the
1154 * reservation change is prepared, but not committed. Once the page
1155 * has been allocated from the subpool and instantiated the change should
1156 * be committed via vma_commit_reservation. No action is required on
1157 * failure.
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001158 */
Roel Kluine2f17d92009-03-31 15:23:15 -07001159static long vma_needs_reservation(struct hstate *h,
Andi Kleena5516432008-07-23 21:27:41 -07001160 struct vm_area_struct *vma, unsigned long addr)
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001161{
1162 struct address_space *mapping = vma->vm_file->f_mapping;
1163 struct inode *inode = mapping->host;
1164
Mel Gormanf83a2752009-05-28 14:34:40 -07001165 if (vma->vm_flags & VM_MAYSHARE) {
Andi Kleena5516432008-07-23 21:27:41 -07001166 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001167 return region_chg(&inode->i_mapping->private_list,
1168 idx, idx + 1);
1169
Andy Whitcroft84afd992008-07-23 21:27:32 -07001170 } else if (!is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
1171 return 1;
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001172
Andy Whitcroft84afd992008-07-23 21:27:32 -07001173 } else {
Roel Kluine2f17d92009-03-31 15:23:15 -07001174 long err;
Andi Kleena5516432008-07-23 21:27:41 -07001175 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
Joonsoo Kimf522c3a2013-09-11 14:21:53 -07001176 struct resv_map *resv = vma_resv_map(vma);
Andy Whitcroft84afd992008-07-23 21:27:32 -07001177
Joonsoo Kimf522c3a2013-09-11 14:21:53 -07001178 err = region_chg(&resv->regions, idx, idx + 1);
Andy Whitcroft84afd992008-07-23 21:27:32 -07001179 if (err < 0)
1180 return err;
1181 return 0;
1182 }
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001183}
Andi Kleena5516432008-07-23 21:27:41 -07001184static void vma_commit_reservation(struct hstate *h,
1185 struct vm_area_struct *vma, unsigned long addr)
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001186{
1187 struct address_space *mapping = vma->vm_file->f_mapping;
1188 struct inode *inode = mapping->host;
1189
Mel Gormanf83a2752009-05-28 14:34:40 -07001190 if (vma->vm_flags & VM_MAYSHARE) {
Andi Kleena5516432008-07-23 21:27:41 -07001191 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001192 region_add(&inode->i_mapping->private_list, idx, idx + 1);
Andy Whitcroft84afd992008-07-23 21:27:32 -07001193
1194 } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
Andi Kleena5516432008-07-23 21:27:41 -07001195 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
Joonsoo Kimf522c3a2013-09-11 14:21:53 -07001196 struct resv_map *resv = vma_resv_map(vma);
Andy Whitcroft84afd992008-07-23 21:27:32 -07001197
1198 /* Mark this page used in the map. */
Joonsoo Kimf522c3a2013-09-11 14:21:53 -07001199 region_add(&resv->regions, idx, idx + 1);
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001200 }
1201}
1202
David Gibson27a85ef2006-03-22 00:08:56 -08001203static struct page *alloc_huge_page(struct vm_area_struct *vma,
Mel Gorman04f2cbe2008-07-23 21:27:25 -07001204 unsigned long addr, int avoid_reserve)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001205{
David Gibson90481622012-03-21 16:34:12 -07001206 struct hugepage_subpool *spool = subpool_vma(vma);
Andi Kleena5516432008-07-23 21:27:41 -07001207 struct hstate *h = hstate_vma(vma);
Adam Litke348ea202007-11-14 16:59:37 -08001208 struct page *page;
Roel Kluine2f17d92009-03-31 15:23:15 -07001209 long chg;
Aneesh Kumar K.V6d76dcf2012-07-31 16:42:18 -07001210 int ret, idx;
1211 struct hugetlb_cgroup *h_cg;
Adam Litke2fc39ce2007-11-14 16:59:39 -08001212
Aneesh Kumar K.V6d76dcf2012-07-31 16:42:18 -07001213 idx = hstate_index(h);
Mel Gormana1e78772008-07-23 21:27:23 -07001214 /*
David Gibson90481622012-03-21 16:34:12 -07001215 * Processes that did not create the mapping will have no
1216 * reserves and will not have accounted against subpool
1217 * limit. Check that the subpool limit can be made before
1218 * satisfying the allocation MAP_NORESERVE mappings may also
1219 * need pages and subpool limit allocated allocated if no reserve
1220 * mapping overlaps.
Mel Gormana1e78772008-07-23 21:27:23 -07001221 */
Andi Kleena5516432008-07-23 21:27:41 -07001222 chg = vma_needs_reservation(h, vma, addr);
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001223 if (chg < 0)
Aneesh Kumar K.V76dcee72012-07-31 16:41:57 -07001224 return ERR_PTR(-ENOMEM);
Joonsoo Kim8bb3f122013-09-11 14:21:54 -07001225 if (chg || avoid_reserve)
1226 if (hugepage_subpool_get_pages(spool, 1))
Aneesh Kumar K.V76dcee72012-07-31 16:41:57 -07001227 return ERR_PTR(-ENOSPC);
Mel Gormana1e78772008-07-23 21:27:23 -07001228
Aneesh Kumar K.V6d76dcf2012-07-31 16:42:18 -07001229 ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg);
1230 if (ret) {
Joonsoo Kim8bb3f122013-09-11 14:21:54 -07001231 if (chg || avoid_reserve)
1232 hugepage_subpool_put_pages(spool, 1);
Aneesh Kumar K.V6d76dcf2012-07-31 16:42:18 -07001233 return ERR_PTR(-ENOSPC);
1234 }
Mel Gormana1e78772008-07-23 21:27:23 -07001235 spin_lock(&hugetlb_lock);
Joonsoo Kimaf0ed732013-09-11 14:21:18 -07001236 page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, chg);
Joonsoo Kim81a6fca2013-09-11 14:20:58 -07001237 if (!page) {
Aneesh Kumar K.V94ae8ba2012-07-31 16:42:35 -07001238 spin_unlock(&hugetlb_lock);
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001239 page = alloc_buddy_huge_page(h, NUMA_NO_NODE);
Mel Gormana1e78772008-07-23 21:27:23 -07001240 if (!page) {
Aneesh Kumar K.V6d76dcf2012-07-31 16:42:18 -07001241 hugetlb_cgroup_uncharge_cgroup(idx,
1242 pages_per_huge_page(h),
1243 h_cg);
Joonsoo Kim8bb3f122013-09-11 14:21:54 -07001244 if (chg || avoid_reserve)
1245 hugepage_subpool_put_pages(spool, 1);
Aneesh Kumar K.V76dcee72012-07-31 16:41:57 -07001246 return ERR_PTR(-ENOSPC);
Mel Gormana1e78772008-07-23 21:27:23 -07001247 }
Aneesh Kumar K.V79dbb232012-07-31 16:42:32 -07001248 spin_lock(&hugetlb_lock);
1249 list_move(&page->lru, &h->hugepage_activelist);
Joonsoo Kim81a6fca2013-09-11 14:20:58 -07001250 /* Fall through */
Mel Gormana1e78772008-07-23 21:27:23 -07001251 }
Joonsoo Kim81a6fca2013-09-11 14:20:58 -07001252 hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page);
1253 spin_unlock(&hugetlb_lock);
Mel Gormana1e78772008-07-23 21:27:23 -07001254
David Gibson90481622012-03-21 16:34:12 -07001255 set_page_private(page, (unsigned long)spool);
Mel Gormana1e78772008-07-23 21:27:23 -07001256
Andi Kleena5516432008-07-23 21:27:41 -07001257 vma_commit_reservation(h, vma, addr);
Adam Litke90d8b7e2007-11-14 16:59:42 -08001258 return page;
David Gibsonb45b5bd2006-03-22 00:08:55 -08001259}
1260
Naoya Horiguchi74060e42013-09-11 14:22:06 -07001261/*
1262 * alloc_huge_page()'s wrapper which simply returns the page if allocation
1263 * succeeds, otherwise NULL. This function is called from new_vma_page(),
1264 * where no ERR_VALUE is expected to be returned.
1265 */
1266struct page *alloc_huge_page_noerr(struct vm_area_struct *vma,
1267 unsigned long addr, int avoid_reserve)
1268{
1269 struct page *page = alloc_huge_page(vma, addr, avoid_reserve);
1270 if (IS_ERR(page))
1271 page = NULL;
1272 return page;
1273}
1274
Cyrill Gorcunov91f47662009-01-06 14:40:33 -08001275int __weak alloc_bootmem_huge_page(struct hstate *h)
Andi Kleenaa888a72008-07-23 21:27:47 -07001276{
1277 struct huge_bootmem_page *m;
Joonsoo Kimb2261022013-09-11 14:21:00 -07001278 int nr_nodes, node;
Andi Kleenaa888a72008-07-23 21:27:47 -07001279
Joonsoo Kimb2261022013-09-11 14:21:00 -07001280 for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) {
Andi Kleenaa888a72008-07-23 21:27:47 -07001281 void *addr;
1282
Joonsoo Kimb2261022013-09-11 14:21:00 -07001283 addr = __alloc_bootmem_node_nopanic(NODE_DATA(node),
Andi Kleenaa888a72008-07-23 21:27:47 -07001284 huge_page_size(h), huge_page_size(h), 0);
1285
1286 if (addr) {
1287 /*
1288 * Use the beginning of the huge page to store the
1289 * huge_bootmem_page struct (until gather_bootmem
1290 * puts them into the mem_map).
1291 */
1292 m = addr;
Cyrill Gorcunov91f47662009-01-06 14:40:33 -08001293 goto found;
Andi Kleenaa888a72008-07-23 21:27:47 -07001294 }
Andi Kleenaa888a72008-07-23 21:27:47 -07001295 }
1296 return 0;
1297
1298found:
1299 BUG_ON((unsigned long)virt_to_phys(m) & (huge_page_size(h) - 1));
1300 /* Put them into a private list first because mem_map is not up yet */
1301 list_add(&m->list, &huge_boot_pages);
1302 m->hstate = h;
1303 return 1;
1304}
1305
Andy Whitcroft18229df2008-11-06 12:53:27 -08001306static void prep_compound_huge_page(struct page *page, int order)
1307{
1308 if (unlikely(order > (MAX_ORDER - 1)))
1309 prep_compound_gigantic_page(page, order);
1310 else
1311 prep_compound_page(page, order);
1312}
1313
Andi Kleenaa888a72008-07-23 21:27:47 -07001314/* Put bootmem huge pages into the standard lists after mem_map is up */
1315static void __init gather_bootmem_prealloc(void)
1316{
1317 struct huge_bootmem_page *m;
1318
1319 list_for_each_entry(m, &huge_boot_pages, list) {
Andi Kleenaa888a72008-07-23 21:27:47 -07001320 struct hstate *h = m->hstate;
Becky Bruceee8f2482011-07-25 17:11:50 -07001321 struct page *page;
1322
1323#ifdef CONFIG_HIGHMEM
1324 page = pfn_to_page(m->phys >> PAGE_SHIFT);
1325 free_bootmem_late((unsigned long)m,
1326 sizeof(struct huge_bootmem_page));
1327#else
1328 page = virt_to_page(m);
1329#endif
Andi Kleenaa888a72008-07-23 21:27:47 -07001330 WARN_ON(page_count(page) != 1);
Andy Whitcroft18229df2008-11-06 12:53:27 -08001331 prep_compound_huge_page(page, h->order);
Andrea Arcangelief5a22b2013-10-16 13:46:56 -07001332 WARN_ON(PageReserved(page));
Andi Kleenaa888a72008-07-23 21:27:47 -07001333 prep_new_huge_page(h, page, page_to_nid(page));
Rafael Aquinib0320c72011-06-15 15:08:39 -07001334 /*
1335 * If we had gigantic hugepages allocated at boot time, we need
1336 * to restore the 'stolen' pages to totalram_pages in order to
1337 * fix confusing memory reports from free(1) and another
1338 * side-effects, like CommitLimit going negative.
1339 */
1340 if (h->order > (MAX_ORDER - 1))
Jiang Liu3dcc0572013-07-03 15:03:21 -07001341 adjust_managed_page_count(page, 1 << h->order);
Andi Kleenaa888a72008-07-23 21:27:47 -07001342 }
1343}
1344
Andi Kleen8faa8b02008-07-23 21:27:48 -07001345static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001346{
1347 unsigned long i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001348
Andi Kleene5ff2152008-07-23 21:27:42 -07001349 for (i = 0; i < h->max_huge_pages; ++i) {
Andi Kleenaa888a72008-07-23 21:27:47 -07001350 if (h->order >= MAX_ORDER) {
1351 if (!alloc_bootmem_huge_page(h))
1352 break;
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -08001353 } else if (!alloc_fresh_huge_page(h,
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08001354 &node_states[N_MEMORY]))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001355 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001356 }
Andi Kleen8faa8b02008-07-23 21:27:48 -07001357 h->max_huge_pages = i;
Andi Kleene5ff2152008-07-23 21:27:42 -07001358}
1359
1360static void __init hugetlb_init_hstates(void)
1361{
1362 struct hstate *h;
1363
1364 for_each_hstate(h) {
Andi Kleen8faa8b02008-07-23 21:27:48 -07001365 /* oversize hugepages were init'ed in early boot */
1366 if (h->order < MAX_ORDER)
1367 hugetlb_hstate_alloc_pages(h);
Andi Kleene5ff2152008-07-23 21:27:42 -07001368 }
1369}
1370
Andi Kleen4abd32d2008-07-23 21:27:49 -07001371static char * __init memfmt(char *buf, unsigned long n)
1372{
1373 if (n >= (1UL << 30))
1374 sprintf(buf, "%lu GB", n >> 30);
1375 else if (n >= (1UL << 20))
1376 sprintf(buf, "%lu MB", n >> 20);
1377 else
1378 sprintf(buf, "%lu KB", n >> 10);
1379 return buf;
1380}
1381
Andi Kleene5ff2152008-07-23 21:27:42 -07001382static void __init report_hugepages(void)
1383{
1384 struct hstate *h;
1385
1386 for_each_hstate(h) {
Andi Kleen4abd32d2008-07-23 21:27:49 -07001387 char buf[32];
Andrew Mortonffb22af2013-02-22 16:32:08 -08001388 pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n",
Andi Kleen4abd32d2008-07-23 21:27:49 -07001389 memfmt(buf, huge_page_size(h)),
1390 h->free_huge_pages);
Andi Kleene5ff2152008-07-23 21:27:42 -07001391 }
1392}
1393
Linus Torvalds1da177e2005-04-16 15:20:36 -07001394#ifdef CONFIG_HIGHMEM
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001395static void try_to_free_low(struct hstate *h, unsigned long count,
1396 nodemask_t *nodes_allowed)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001397{
Christoph Lameter4415cc82006-09-25 23:31:55 -07001398 int i;
1399
Andi Kleenaa888a72008-07-23 21:27:47 -07001400 if (h->order >= MAX_ORDER)
1401 return;
1402
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001403 for_each_node_mask(i, *nodes_allowed) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001404 struct page *page, *next;
Andi Kleena5516432008-07-23 21:27:41 -07001405 struct list_head *freel = &h->hugepage_freelists[i];
1406 list_for_each_entry_safe(page, next, freel, lru) {
1407 if (count >= h->nr_huge_pages)
Adam Litke6b0c8802007-10-16 01:26:23 -07001408 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001409 if (PageHighMem(page))
1410 continue;
1411 list_del(&page->lru);
Andi Kleene5ff2152008-07-23 21:27:42 -07001412 update_and_free_page(h, page);
Andi Kleena5516432008-07-23 21:27:41 -07001413 h->free_huge_pages--;
1414 h->free_huge_pages_node[page_to_nid(page)]--;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001415 }
1416 }
1417}
1418#else
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001419static inline void try_to_free_low(struct hstate *h, unsigned long count,
1420 nodemask_t *nodes_allowed)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001421{
1422}
1423#endif
1424
Wu Fengguang20a03072009-06-16 15:32:22 -07001425/*
1426 * Increment or decrement surplus_huge_pages. Keep node-specific counters
1427 * balanced by operating on them in a round-robin fashion.
1428 * Returns 1 if an adjustment was made.
1429 */
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001430static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
1431 int delta)
Wu Fengguang20a03072009-06-16 15:32:22 -07001432{
Joonsoo Kimb2261022013-09-11 14:21:00 -07001433 int nr_nodes, node;
Wu Fengguang20a03072009-06-16 15:32:22 -07001434
1435 VM_BUG_ON(delta != -1 && delta != 1);
Wu Fengguang20a03072009-06-16 15:32:22 -07001436
Joonsoo Kimb2261022013-09-11 14:21:00 -07001437 if (delta < 0) {
1438 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
1439 if (h->surplus_huge_pages_node[node])
1440 goto found;
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001441 }
Joonsoo Kimb2261022013-09-11 14:21:00 -07001442 } else {
1443 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
1444 if (h->surplus_huge_pages_node[node] <
1445 h->nr_huge_pages_node[node])
1446 goto found;
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001447 }
Joonsoo Kimb2261022013-09-11 14:21:00 -07001448 }
1449 return 0;
Wu Fengguang20a03072009-06-16 15:32:22 -07001450
Joonsoo Kimb2261022013-09-11 14:21:00 -07001451found:
1452 h->surplus_huge_pages += delta;
1453 h->surplus_huge_pages_node[node] += delta;
1454 return 1;
Wu Fengguang20a03072009-06-16 15:32:22 -07001455}
1456
Andi Kleena5516432008-07-23 21:27:41 -07001457#define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001458static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count,
1459 nodemask_t *nodes_allowed)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001460{
Adam Litke7893d1d2007-10-16 01:26:18 -07001461 unsigned long min_count, ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001462
Andi Kleenaa888a72008-07-23 21:27:47 -07001463 if (h->order >= MAX_ORDER)
1464 return h->max_huge_pages;
1465
Adam Litke7893d1d2007-10-16 01:26:18 -07001466 /*
1467 * Increase the pool size
1468 * First take pages out of surplus state. Then make up the
1469 * remaining difference by allocating fresh huge pages.
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001470 *
1471 * We might race with alloc_buddy_huge_page() here and be unable
1472 * to convert a surplus huge page to a normal huge page. That is
1473 * not critical, though, it just means the overall size of the
1474 * pool might be one hugepage larger than it needs to be, but
1475 * within all the constraints specified by the sysctls.
Adam Litke7893d1d2007-10-16 01:26:18 -07001476 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001477 spin_lock(&hugetlb_lock);
Andi Kleena5516432008-07-23 21:27:41 -07001478 while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001479 if (!adjust_pool_surplus(h, nodes_allowed, -1))
Adam Litke7893d1d2007-10-16 01:26:18 -07001480 break;
1481 }
1482
Andi Kleena5516432008-07-23 21:27:41 -07001483 while (count > persistent_huge_pages(h)) {
Adam Litke7893d1d2007-10-16 01:26:18 -07001484 /*
1485 * If this allocation races such that we no longer need the
1486 * page, free_huge_page will handle it by freeing the page
1487 * and reducing the surplus.
1488 */
1489 spin_unlock(&hugetlb_lock);
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001490 ret = alloc_fresh_huge_page(h, nodes_allowed);
Adam Litke7893d1d2007-10-16 01:26:18 -07001491 spin_lock(&hugetlb_lock);
1492 if (!ret)
1493 goto out;
1494
Mel Gorman536240f22009-12-14 17:59:56 -08001495 /* Bail for signals. Probably ctrl-c from user */
1496 if (signal_pending(current))
1497 goto out;
Adam Litke7893d1d2007-10-16 01:26:18 -07001498 }
Adam Litke7893d1d2007-10-16 01:26:18 -07001499
1500 /*
1501 * Decrease the pool size
1502 * First return free pages to the buddy allocator (being careful
1503 * to keep enough around to satisfy reservations). Then place
1504 * pages into surplus state as needed so the pool will shrink
1505 * to the desired size as pages become free.
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001506 *
1507 * By placing pages into the surplus state independent of the
1508 * overcommit value, we are allowing the surplus pool size to
1509 * exceed overcommit. There are few sane options here. Since
1510 * alloc_buddy_huge_page() is checking the global counter,
1511 * though, we'll note that we're not allowed to exceed surplus
1512 * and won't grow the pool anywhere else. Not until one of the
1513 * sysctls are changed, or the surplus pages go out of use.
Adam Litke7893d1d2007-10-16 01:26:18 -07001514 */
Andi Kleena5516432008-07-23 21:27:41 -07001515 min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
Adam Litke6b0c8802007-10-16 01:26:23 -07001516 min_count = max(count, min_count);
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001517 try_to_free_low(h, min_count, nodes_allowed);
Andi Kleena5516432008-07-23 21:27:41 -07001518 while (min_count < persistent_huge_pages(h)) {
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001519 if (!free_pool_huge_page(h, nodes_allowed, 0))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001520 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001521 }
Andi Kleena5516432008-07-23 21:27:41 -07001522 while (count < persistent_huge_pages(h)) {
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001523 if (!adjust_pool_surplus(h, nodes_allowed, 1))
Adam Litke7893d1d2007-10-16 01:26:18 -07001524 break;
1525 }
1526out:
Andi Kleena5516432008-07-23 21:27:41 -07001527 ret = persistent_huge_pages(h);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001528 spin_unlock(&hugetlb_lock);
Adam Litke7893d1d2007-10-16 01:26:18 -07001529 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001530}
1531
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001532#define HSTATE_ATTR_RO(_name) \
1533 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
1534
1535#define HSTATE_ATTR(_name) \
1536 static struct kobj_attribute _name##_attr = \
1537 __ATTR(_name, 0644, _name##_show, _name##_store)
1538
1539static struct kobject *hugepages_kobj;
1540static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
1541
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001542static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
1543
1544static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001545{
1546 int i;
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001547
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001548 for (i = 0; i < HUGE_MAX_HSTATE; i++)
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001549 if (hstate_kobjs[i] == kobj) {
1550 if (nidp)
1551 *nidp = NUMA_NO_NODE;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001552 return &hstates[i];
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001553 }
1554
1555 return kobj_to_node_hstate(kobj, nidp);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001556}
1557
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001558static ssize_t nr_hugepages_show_common(struct kobject *kobj,
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001559 struct kobj_attribute *attr, char *buf)
1560{
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001561 struct hstate *h;
1562 unsigned long nr_huge_pages;
1563 int nid;
1564
1565 h = kobj_to_hstate(kobj, &nid);
1566 if (nid == NUMA_NO_NODE)
1567 nr_huge_pages = h->nr_huge_pages;
1568 else
1569 nr_huge_pages = h->nr_huge_pages_node[nid];
1570
1571 return sprintf(buf, "%lu\n", nr_huge_pages);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001572}
Eric B Munsonadbe8722011-01-13 15:47:27 -08001573
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001574static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
1575 struct kobject *kobj, struct kobj_attribute *attr,
1576 const char *buf, size_t len)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001577{
1578 int err;
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001579 int nid;
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001580 unsigned long count;
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001581 struct hstate *h;
David Rientjesbad44b52009-12-14 17:58:38 -08001582 NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001583
Jingoo Han3dbb95f2013-09-11 14:20:25 -07001584 err = kstrtoul(buf, 10, &count);
Eric B Munson73ae31e2011-01-13 15:47:28 -08001585 if (err)
Eric B Munsonadbe8722011-01-13 15:47:27 -08001586 goto out;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001587
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001588 h = kobj_to_hstate(kobj, &nid);
Eric B Munsonadbe8722011-01-13 15:47:27 -08001589 if (h->order >= MAX_ORDER) {
1590 err = -EINVAL;
1591 goto out;
1592 }
1593
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001594 if (nid == NUMA_NO_NODE) {
1595 /*
1596 * global hstate attribute
1597 */
1598 if (!(obey_mempolicy &&
1599 init_nodemask_of_mempolicy(nodes_allowed))) {
1600 NODEMASK_FREE(nodes_allowed);
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08001601 nodes_allowed = &node_states[N_MEMORY];
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001602 }
1603 } else if (nodes_allowed) {
1604 /*
1605 * per node hstate attribute: adjust count to global,
1606 * but restrict alloc/free to the specified node.
1607 */
1608 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
1609 init_nodemask_of_node(nodes_allowed, nid);
1610 } else
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08001611 nodes_allowed = &node_states[N_MEMORY];
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001612
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001613 h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001614
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08001615 if (nodes_allowed != &node_states[N_MEMORY])
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001616 NODEMASK_FREE(nodes_allowed);
1617
1618 return len;
Eric B Munsonadbe8722011-01-13 15:47:27 -08001619out:
1620 NODEMASK_FREE(nodes_allowed);
1621 return err;
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001622}
1623
1624static ssize_t nr_hugepages_show(struct kobject *kobj,
1625 struct kobj_attribute *attr, char *buf)
1626{
1627 return nr_hugepages_show_common(kobj, attr, buf);
1628}
1629
1630static ssize_t nr_hugepages_store(struct kobject *kobj,
1631 struct kobj_attribute *attr, const char *buf, size_t len)
1632{
1633 return nr_hugepages_store_common(false, kobj, attr, buf, len);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001634}
1635HSTATE_ATTR(nr_hugepages);
1636
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001637#ifdef CONFIG_NUMA
1638
1639/*
1640 * hstate attribute for optionally mempolicy-based constraint on persistent
1641 * huge page alloc/free.
1642 */
1643static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
1644 struct kobj_attribute *attr, char *buf)
1645{
1646 return nr_hugepages_show_common(kobj, attr, buf);
1647}
1648
1649static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
1650 struct kobj_attribute *attr, const char *buf, size_t len)
1651{
1652 return nr_hugepages_store_common(true, kobj, attr, buf, len);
1653}
1654HSTATE_ATTR(nr_hugepages_mempolicy);
1655#endif
1656
1657
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001658static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
1659 struct kobj_attribute *attr, char *buf)
1660{
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001661 struct hstate *h = kobj_to_hstate(kobj, NULL);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001662 return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
1663}
Eric B Munsonadbe8722011-01-13 15:47:27 -08001664
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001665static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
1666 struct kobj_attribute *attr, const char *buf, size_t count)
1667{
1668 int err;
1669 unsigned long input;
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001670 struct hstate *h = kobj_to_hstate(kobj, NULL);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001671
Eric B Munsonadbe8722011-01-13 15:47:27 -08001672 if (h->order >= MAX_ORDER)
1673 return -EINVAL;
1674
Jingoo Han3dbb95f2013-09-11 14:20:25 -07001675 err = kstrtoul(buf, 10, &input);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001676 if (err)
Eric B Munson73ae31e2011-01-13 15:47:28 -08001677 return err;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001678
1679 spin_lock(&hugetlb_lock);
1680 h->nr_overcommit_huge_pages = input;
1681 spin_unlock(&hugetlb_lock);
1682
1683 return count;
1684}
1685HSTATE_ATTR(nr_overcommit_hugepages);
1686
1687static ssize_t free_hugepages_show(struct kobject *kobj,
1688 struct kobj_attribute *attr, char *buf)
1689{
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001690 struct hstate *h;
1691 unsigned long free_huge_pages;
1692 int nid;
1693
1694 h = kobj_to_hstate(kobj, &nid);
1695 if (nid == NUMA_NO_NODE)
1696 free_huge_pages = h->free_huge_pages;
1697 else
1698 free_huge_pages = h->free_huge_pages_node[nid];
1699
1700 return sprintf(buf, "%lu\n", free_huge_pages);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001701}
1702HSTATE_ATTR_RO(free_hugepages);
1703
1704static ssize_t resv_hugepages_show(struct kobject *kobj,
1705 struct kobj_attribute *attr, char *buf)
1706{
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001707 struct hstate *h = kobj_to_hstate(kobj, NULL);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001708 return sprintf(buf, "%lu\n", h->resv_huge_pages);
1709}
1710HSTATE_ATTR_RO(resv_hugepages);
1711
1712static ssize_t surplus_hugepages_show(struct kobject *kobj,
1713 struct kobj_attribute *attr, char *buf)
1714{
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001715 struct hstate *h;
1716 unsigned long surplus_huge_pages;
1717 int nid;
1718
1719 h = kobj_to_hstate(kobj, &nid);
1720 if (nid == NUMA_NO_NODE)
1721 surplus_huge_pages = h->surplus_huge_pages;
1722 else
1723 surplus_huge_pages = h->surplus_huge_pages_node[nid];
1724
1725 return sprintf(buf, "%lu\n", surplus_huge_pages);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001726}
1727HSTATE_ATTR_RO(surplus_hugepages);
1728
1729static struct attribute *hstate_attrs[] = {
1730 &nr_hugepages_attr.attr,
1731 &nr_overcommit_hugepages_attr.attr,
1732 &free_hugepages_attr.attr,
1733 &resv_hugepages_attr.attr,
1734 &surplus_hugepages_attr.attr,
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001735#ifdef CONFIG_NUMA
1736 &nr_hugepages_mempolicy_attr.attr,
1737#endif
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001738 NULL,
1739};
1740
1741static struct attribute_group hstate_attr_group = {
1742 .attrs = hstate_attrs,
1743};
1744
Jeff Mahoney094e9532010-02-02 13:44:14 -08001745static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
1746 struct kobject **hstate_kobjs,
1747 struct attribute_group *hstate_attr_group)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001748{
1749 int retval;
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07001750 int hi = hstate_index(h);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001751
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001752 hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
1753 if (!hstate_kobjs[hi])
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001754 return -ENOMEM;
1755
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001756 retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001757 if (retval)
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001758 kobject_put(hstate_kobjs[hi]);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001759
1760 return retval;
1761}
1762
1763static void __init hugetlb_sysfs_init(void)
1764{
1765 struct hstate *h;
1766 int err;
1767
1768 hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
1769 if (!hugepages_kobj)
1770 return;
1771
1772 for_each_hstate(h) {
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001773 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
1774 hstate_kobjs, &hstate_attr_group);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001775 if (err)
Andrew Mortonffb22af2013-02-22 16:32:08 -08001776 pr_err("Hugetlb: Unable to add hstate %s", h->name);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001777 }
1778}
1779
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001780#ifdef CONFIG_NUMA
1781
1782/*
1783 * node_hstate/s - associate per node hstate attributes, via their kobjects,
Kay Sievers10fbcf42011-12-21 14:48:43 -08001784 * with node devices in node_devices[] using a parallel array. The array
1785 * index of a node device or _hstate == node id.
1786 * This is here to avoid any static dependency of the node device driver, in
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001787 * the base kernel, on the hugetlb module.
1788 */
1789struct node_hstate {
1790 struct kobject *hugepages_kobj;
1791 struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
1792};
1793struct node_hstate node_hstates[MAX_NUMNODES];
1794
1795/*
Kay Sievers10fbcf42011-12-21 14:48:43 -08001796 * A subset of global hstate attributes for node devices
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001797 */
1798static struct attribute *per_node_hstate_attrs[] = {
1799 &nr_hugepages_attr.attr,
1800 &free_hugepages_attr.attr,
1801 &surplus_hugepages_attr.attr,
1802 NULL,
1803};
1804
1805static struct attribute_group per_node_hstate_attr_group = {
1806 .attrs = per_node_hstate_attrs,
1807};
1808
1809/*
Kay Sievers10fbcf42011-12-21 14:48:43 -08001810 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001811 * Returns node id via non-NULL nidp.
1812 */
1813static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
1814{
1815 int nid;
1816
1817 for (nid = 0; nid < nr_node_ids; nid++) {
1818 struct node_hstate *nhs = &node_hstates[nid];
1819 int i;
1820 for (i = 0; i < HUGE_MAX_HSTATE; i++)
1821 if (nhs->hstate_kobjs[i] == kobj) {
1822 if (nidp)
1823 *nidp = nid;
1824 return &hstates[i];
1825 }
1826 }
1827
1828 BUG();
1829 return NULL;
1830}
1831
1832/*
Kay Sievers10fbcf42011-12-21 14:48:43 -08001833 * Unregister hstate attributes from a single node device.
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001834 * No-op if no hstate attributes attached.
1835 */
Claudiu Ghioc3cd8b442013-03-04 12:46:15 +02001836static void hugetlb_unregister_node(struct node *node)
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001837{
1838 struct hstate *h;
Kay Sievers10fbcf42011-12-21 14:48:43 -08001839 struct node_hstate *nhs = &node_hstates[node->dev.id];
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001840
1841 if (!nhs->hugepages_kobj)
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -08001842 return; /* no hstate attributes */
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001843
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07001844 for_each_hstate(h) {
1845 int idx = hstate_index(h);
1846 if (nhs->hstate_kobjs[idx]) {
1847 kobject_put(nhs->hstate_kobjs[idx]);
1848 nhs->hstate_kobjs[idx] = NULL;
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001849 }
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07001850 }
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001851
1852 kobject_put(nhs->hugepages_kobj);
1853 nhs->hugepages_kobj = NULL;
1854}
1855
1856/*
Kay Sievers10fbcf42011-12-21 14:48:43 -08001857 * hugetlb module exit: unregister hstate attributes from node devices
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001858 * that have them.
1859 */
1860static void hugetlb_unregister_all_nodes(void)
1861{
1862 int nid;
1863
1864 /*
Kay Sievers10fbcf42011-12-21 14:48:43 -08001865 * disable node device registrations.
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001866 */
1867 register_hugetlbfs_with_node(NULL, NULL);
1868
1869 /*
1870 * remove hstate attributes from any nodes that have them.
1871 */
1872 for (nid = 0; nid < nr_node_ids; nid++)
Wen Congyang87327942012-12-11 16:00:56 -08001873 hugetlb_unregister_node(node_devices[nid]);
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001874}
1875
1876/*
Kay Sievers10fbcf42011-12-21 14:48:43 -08001877 * Register hstate attributes for a single node device.
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001878 * No-op if attributes already registered.
1879 */
Claudiu Ghioc3cd8b442013-03-04 12:46:15 +02001880static void hugetlb_register_node(struct node *node)
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001881{
1882 struct hstate *h;
Kay Sievers10fbcf42011-12-21 14:48:43 -08001883 struct node_hstate *nhs = &node_hstates[node->dev.id];
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001884 int err;
1885
1886 if (nhs->hugepages_kobj)
1887 return; /* already allocated */
1888
1889 nhs->hugepages_kobj = kobject_create_and_add("hugepages",
Kay Sievers10fbcf42011-12-21 14:48:43 -08001890 &node->dev.kobj);
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001891 if (!nhs->hugepages_kobj)
1892 return;
1893
1894 for_each_hstate(h) {
1895 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
1896 nhs->hstate_kobjs,
1897 &per_node_hstate_attr_group);
1898 if (err) {
Andrew Mortonffb22af2013-02-22 16:32:08 -08001899 pr_err("Hugetlb: Unable to add hstate %s for node %d\n",
1900 h->name, node->dev.id);
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001901 hugetlb_unregister_node(node);
1902 break;
1903 }
1904 }
1905}
1906
1907/*
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -08001908 * hugetlb init time: register hstate attributes for all registered node
Kay Sievers10fbcf42011-12-21 14:48:43 -08001909 * devices of nodes that have memory. All on-line nodes should have
1910 * registered their associated device by this time.
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001911 */
1912static void hugetlb_register_all_nodes(void)
1913{
1914 int nid;
1915
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08001916 for_each_node_state(nid, N_MEMORY) {
Wen Congyang87327942012-12-11 16:00:56 -08001917 struct node *node = node_devices[nid];
Kay Sievers10fbcf42011-12-21 14:48:43 -08001918 if (node->dev.id == nid)
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001919 hugetlb_register_node(node);
1920 }
1921
1922 /*
Kay Sievers10fbcf42011-12-21 14:48:43 -08001923 * Let the node device driver know we're here so it can
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001924 * [un]register hstate attributes on node hotplug.
1925 */
1926 register_hugetlbfs_with_node(hugetlb_register_node,
1927 hugetlb_unregister_node);
1928}
1929#else /* !CONFIG_NUMA */
1930
1931static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
1932{
1933 BUG();
1934 if (nidp)
1935 *nidp = -1;
1936 return NULL;
1937}
1938
1939static void hugetlb_unregister_all_nodes(void) { }
1940
1941static void hugetlb_register_all_nodes(void) { }
1942
1943#endif
1944
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001945static void __exit hugetlb_exit(void)
1946{
1947 struct hstate *h;
1948
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001949 hugetlb_unregister_all_nodes();
1950
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001951 for_each_hstate(h) {
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07001952 kobject_put(hstate_kobjs[hstate_index(h)]);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001953 }
1954
1955 kobject_put(hugepages_kobj);
1956}
1957module_exit(hugetlb_exit);
1958
1959static int __init hugetlb_init(void)
1960{
Benjamin Herrenschmidt0ef89d22008-07-31 00:07:30 -07001961 /* Some platform decide whether they support huge pages at boot
1962 * time. On these, such as powerpc, HPAGE_SHIFT is set to 0 when
1963 * there is no such support
1964 */
1965 if (HPAGE_SHIFT == 0)
1966 return 0;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001967
Nick Piggine11bfbf2008-07-23 21:27:52 -07001968 if (!size_to_hstate(default_hstate_size)) {
1969 default_hstate_size = HPAGE_SIZE;
1970 if (!size_to_hstate(default_hstate_size))
1971 hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001972 }
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07001973 default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size));
Nick Piggine11bfbf2008-07-23 21:27:52 -07001974 if (default_hstate_max_huge_pages)
1975 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001976
1977 hugetlb_init_hstates();
Andi Kleenaa888a72008-07-23 21:27:47 -07001978 gather_bootmem_prealloc();
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001979 report_hugepages();
1980
1981 hugetlb_sysfs_init();
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001982 hugetlb_register_all_nodes();
Jianguo Wu7179e7b2012-12-18 14:23:19 -08001983 hugetlb_cgroup_file_init();
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001984
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001985 return 0;
1986}
1987module_init(hugetlb_init);
1988
1989/* Should be called on processing a hugepagesz=... option */
1990void __init hugetlb_add_hstate(unsigned order)
1991{
1992 struct hstate *h;
Andi Kleen8faa8b02008-07-23 21:27:48 -07001993 unsigned long i;
1994
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001995 if (size_to_hstate(PAGE_SIZE << order)) {
Andrew Mortonffb22af2013-02-22 16:32:08 -08001996 pr_warning("hugepagesz= specified twice, ignoring\n");
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001997 return;
1998 }
Aneesh Kumar K.V47d38342012-07-31 16:41:54 -07001999 BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002000 BUG_ON(order == 0);
Aneesh Kumar K.V47d38342012-07-31 16:41:54 -07002001 h = &hstates[hugetlb_max_hstate++];
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002002 h->order = order;
2003 h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
Andi Kleen8faa8b02008-07-23 21:27:48 -07002004 h->nr_huge_pages = 0;
2005 h->free_huge_pages = 0;
2006 for (i = 0; i < MAX_NUMNODES; ++i)
2007 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
Aneesh Kumar K.V0edaecf2012-07-31 16:42:07 -07002008 INIT_LIST_HEAD(&h->hugepage_activelist);
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08002009 h->next_nid_to_alloc = first_node(node_states[N_MEMORY]);
2010 h->next_nid_to_free = first_node(node_states[N_MEMORY]);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002011 snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
2012 huge_page_size(h)/1024);
Andi Kleen8faa8b02008-07-23 21:27:48 -07002013
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002014 parsed_hstate = h;
2015}
2016
Nick Piggine11bfbf2008-07-23 21:27:52 -07002017static int __init hugetlb_nrpages_setup(char *s)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002018{
2019 unsigned long *mhp;
Andi Kleen8faa8b02008-07-23 21:27:48 -07002020 static unsigned long *last_mhp;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002021
2022 /*
Aneesh Kumar K.V47d38342012-07-31 16:41:54 -07002023 * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet,
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002024 * so this hugepages= parameter goes to the "default hstate".
2025 */
Aneesh Kumar K.V47d38342012-07-31 16:41:54 -07002026 if (!hugetlb_max_hstate)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002027 mhp = &default_hstate_max_huge_pages;
2028 else
2029 mhp = &parsed_hstate->max_huge_pages;
2030
Andi Kleen8faa8b02008-07-23 21:27:48 -07002031 if (mhp == last_mhp) {
Andrew Mortonffb22af2013-02-22 16:32:08 -08002032 pr_warning("hugepages= specified twice without "
2033 "interleaving hugepagesz=, ignoring\n");
Andi Kleen8faa8b02008-07-23 21:27:48 -07002034 return 1;
2035 }
2036
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002037 if (sscanf(s, "%lu", mhp) <= 0)
2038 *mhp = 0;
2039
Andi Kleen8faa8b02008-07-23 21:27:48 -07002040 /*
2041 * Global state is always initialized later in hugetlb_init.
2042 * But we need to allocate >= MAX_ORDER hstates here early to still
2043 * use the bootmem allocator.
2044 */
Aneesh Kumar K.V47d38342012-07-31 16:41:54 -07002045 if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER)
Andi Kleen8faa8b02008-07-23 21:27:48 -07002046 hugetlb_hstate_alloc_pages(parsed_hstate);
2047
2048 last_mhp = mhp;
2049
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002050 return 1;
2051}
Nick Piggine11bfbf2008-07-23 21:27:52 -07002052__setup("hugepages=", hugetlb_nrpages_setup);
2053
2054static int __init hugetlb_default_setup(char *s)
2055{
2056 default_hstate_size = memparse(s, &s);
2057 return 1;
2058}
2059__setup("default_hugepagesz=", hugetlb_default_setup);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002060
Nishanth Aravamudan8a213462008-07-25 19:44:37 -07002061static unsigned int cpuset_mems_nr(unsigned int *array)
2062{
2063 int node;
2064 unsigned int nr = 0;
2065
2066 for_each_node_mask(node, cpuset_current_mems_allowed)
2067 nr += array[node];
2068
2069 return nr;
2070}
2071
2072#ifdef CONFIG_SYSCTL
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002073static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
2074 struct ctl_table *table, int write,
2075 void __user *buffer, size_t *length, loff_t *ppos)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002076{
Andi Kleene5ff2152008-07-23 21:27:42 -07002077 struct hstate *h = &default_hstate;
2078 unsigned long tmp;
Michal Hocko08d4a242011-01-13 15:47:26 -08002079 int ret;
Andi Kleene5ff2152008-07-23 21:27:42 -07002080
Petr Holasekc033a932011-03-22 16:33:05 -07002081 tmp = h->max_huge_pages;
Andi Kleene5ff2152008-07-23 21:27:42 -07002082
Eric B Munsonadbe8722011-01-13 15:47:27 -08002083 if (write && h->order >= MAX_ORDER)
2084 return -EINVAL;
2085
Andi Kleene5ff2152008-07-23 21:27:42 -07002086 table->data = &tmp;
2087 table->maxlen = sizeof(unsigned long);
Michal Hocko08d4a242011-01-13 15:47:26 -08002088 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2089 if (ret)
2090 goto out;
Andi Kleene5ff2152008-07-23 21:27:42 -07002091
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002092 if (write) {
David Rientjesbad44b52009-12-14 17:58:38 -08002093 NODEMASK_ALLOC(nodemask_t, nodes_allowed,
2094 GFP_KERNEL | __GFP_NORETRY);
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002095 if (!(obey_mempolicy &&
2096 init_nodemask_of_mempolicy(nodes_allowed))) {
2097 NODEMASK_FREE(nodes_allowed);
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08002098 nodes_allowed = &node_states[N_MEMORY];
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002099 }
2100 h->max_huge_pages = set_max_huge_pages(h, tmp, nodes_allowed);
2101
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08002102 if (nodes_allowed != &node_states[N_MEMORY])
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002103 NODEMASK_FREE(nodes_allowed);
2104 }
Michal Hocko08d4a242011-01-13 15:47:26 -08002105out:
2106 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002107}
Mel Gorman396faf02007-07-17 04:03:13 -07002108
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002109int hugetlb_sysctl_handler(struct ctl_table *table, int write,
2110 void __user *buffer, size_t *length, loff_t *ppos)
2111{
2112
2113 return hugetlb_sysctl_handler_common(false, table, write,
2114 buffer, length, ppos);
2115}
2116
2117#ifdef CONFIG_NUMA
2118int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
2119 void __user *buffer, size_t *length, loff_t *ppos)
2120{
2121 return hugetlb_sysctl_handler_common(true, table, write,
2122 buffer, length, ppos);
2123}
2124#endif /* CONFIG_NUMA */
2125
Nishanth Aravamudana3d0c6a2008-02-08 04:18:18 -08002126int hugetlb_overcommit_handler(struct ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07002127 void __user *buffer,
Nishanth Aravamudana3d0c6a2008-02-08 04:18:18 -08002128 size_t *length, loff_t *ppos)
2129{
Andi Kleena5516432008-07-23 21:27:41 -07002130 struct hstate *h = &default_hstate;
Andi Kleene5ff2152008-07-23 21:27:42 -07002131 unsigned long tmp;
Michal Hocko08d4a242011-01-13 15:47:26 -08002132 int ret;
Andi Kleene5ff2152008-07-23 21:27:42 -07002133
Petr Holasekc033a932011-03-22 16:33:05 -07002134 tmp = h->nr_overcommit_huge_pages;
Andi Kleene5ff2152008-07-23 21:27:42 -07002135
Eric B Munsonadbe8722011-01-13 15:47:27 -08002136 if (write && h->order >= MAX_ORDER)
2137 return -EINVAL;
2138
Andi Kleene5ff2152008-07-23 21:27:42 -07002139 table->data = &tmp;
2140 table->maxlen = sizeof(unsigned long);
Michal Hocko08d4a242011-01-13 15:47:26 -08002141 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2142 if (ret)
2143 goto out;
Andi Kleene5ff2152008-07-23 21:27:42 -07002144
2145 if (write) {
2146 spin_lock(&hugetlb_lock);
2147 h->nr_overcommit_huge_pages = tmp;
2148 spin_unlock(&hugetlb_lock);
2149 }
Michal Hocko08d4a242011-01-13 15:47:26 -08002150out:
2151 return ret;
Nishanth Aravamudana3d0c6a2008-02-08 04:18:18 -08002152}
2153
Linus Torvalds1da177e2005-04-16 15:20:36 -07002154#endif /* CONFIG_SYSCTL */
2155
Alexey Dobriyane1759c22008-10-15 23:50:22 +04002156void hugetlb_report_meminfo(struct seq_file *m)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002157{
Andi Kleena5516432008-07-23 21:27:41 -07002158 struct hstate *h = &default_hstate;
Alexey Dobriyane1759c22008-10-15 23:50:22 +04002159 seq_printf(m,
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002160 "HugePages_Total: %5lu\n"
2161 "HugePages_Free: %5lu\n"
2162 "HugePages_Rsvd: %5lu\n"
2163 "HugePages_Surp: %5lu\n"
2164 "Hugepagesize: %8lu kB\n",
Andi Kleena5516432008-07-23 21:27:41 -07002165 h->nr_huge_pages,
2166 h->free_huge_pages,
2167 h->resv_huge_pages,
2168 h->surplus_huge_pages,
2169 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002170}
2171
2172int hugetlb_report_node_meminfo(int nid, char *buf)
2173{
Andi Kleena5516432008-07-23 21:27:41 -07002174 struct hstate *h = &default_hstate;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002175 return sprintf(buf,
2176 "Node %d HugePages_Total: %5u\n"
Nishanth Aravamudana1de0912008-03-26 14:37:53 -07002177 "Node %d HugePages_Free: %5u\n"
2178 "Node %d HugePages_Surp: %5u\n",
Andi Kleena5516432008-07-23 21:27:41 -07002179 nid, h->nr_huge_pages_node[nid],
2180 nid, h->free_huge_pages_node[nid],
2181 nid, h->surplus_huge_pages_node[nid]);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002182}
2183
David Rientjes949f7ec2013-04-29 15:07:48 -07002184void hugetlb_show_meminfo(void)
2185{
2186 struct hstate *h;
2187 int nid;
2188
2189 for_each_node_state(nid, N_MEMORY)
2190 for_each_hstate(h)
2191 pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
2192 nid,
2193 h->nr_huge_pages_node[nid],
2194 h->free_huge_pages_node[nid],
2195 h->surplus_huge_pages_node[nid],
2196 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
2197}
2198
Linus Torvalds1da177e2005-04-16 15:20:36 -07002199/* Return the number pages of memory we physically have, in PAGE_SIZE units. */
2200unsigned long hugetlb_total_pages(void)
2201{
Wanpeng Lid0028582013-03-22 15:04:40 -07002202 struct hstate *h;
2203 unsigned long nr_total_pages = 0;
2204
2205 for_each_hstate(h)
2206 nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h);
2207 return nr_total_pages;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002208}
Linus Torvalds1da177e2005-04-16 15:20:36 -07002209
Andi Kleena5516432008-07-23 21:27:41 -07002210static int hugetlb_acct_memory(struct hstate *h, long delta)
Mel Gormanfc1b8a72008-07-23 21:27:22 -07002211{
2212 int ret = -ENOMEM;
2213
2214 spin_lock(&hugetlb_lock);
2215 /*
2216 * When cpuset is configured, it breaks the strict hugetlb page
2217 * reservation as the accounting is done on a global variable. Such
2218 * reservation is completely rubbish in the presence of cpuset because
2219 * the reservation is not checked against page availability for the
2220 * current cpuset. Application can still potentially OOM'ed by kernel
2221 * with lack of free htlb page in cpuset that the task is in.
2222 * Attempt to enforce strict accounting with cpuset is almost
2223 * impossible (or too ugly) because cpuset is too fluid that
2224 * task or memory node can be dynamically moved between cpusets.
2225 *
2226 * The change of semantics for shared hugetlb mapping with cpuset is
2227 * undesirable. However, in order to preserve some of the semantics,
2228 * we fall back to check against current free page availability as
2229 * a best attempt and hopefully to minimize the impact of changing
2230 * semantics that cpuset has.
2231 */
2232 if (delta > 0) {
Andi Kleena5516432008-07-23 21:27:41 -07002233 if (gather_surplus_pages(h, delta) < 0)
Mel Gormanfc1b8a72008-07-23 21:27:22 -07002234 goto out;
2235
Andi Kleena5516432008-07-23 21:27:41 -07002236 if (delta > cpuset_mems_nr(h->free_huge_pages_node)) {
2237 return_unused_surplus_pages(h, delta);
Mel Gormanfc1b8a72008-07-23 21:27:22 -07002238 goto out;
2239 }
2240 }
2241
2242 ret = 0;
2243 if (delta < 0)
Andi Kleena5516432008-07-23 21:27:41 -07002244 return_unused_surplus_pages(h, (unsigned long) -delta);
Mel Gormanfc1b8a72008-07-23 21:27:22 -07002245
2246out:
2247 spin_unlock(&hugetlb_lock);
2248 return ret;
2249}
2250
Andy Whitcroft84afd992008-07-23 21:27:32 -07002251static void hugetlb_vm_op_open(struct vm_area_struct *vma)
2252{
Joonsoo Kimf522c3a2013-09-11 14:21:53 -07002253 struct resv_map *resv = vma_resv_map(vma);
Andy Whitcroft84afd992008-07-23 21:27:32 -07002254
2255 /*
2256 * This new VMA should share its siblings reservation map if present.
2257 * The VMA will only ever have a valid reservation map pointer where
2258 * it is being copied for another still existing VMA. As that VMA
Lucas De Marchi25985ed2011-03-30 22:57:33 -03002259 * has a reference to the reservation map it cannot disappear until
Andy Whitcroft84afd992008-07-23 21:27:32 -07002260 * after this open call completes. It is therefore safe to take a
2261 * new reference here without additional locking.
2262 */
Joonsoo Kimf522c3a2013-09-11 14:21:53 -07002263 if (resv)
2264 kref_get(&resv->refs);
Andy Whitcroft84afd992008-07-23 21:27:32 -07002265}
2266
Dave Hansenc50ac052012-05-29 15:06:46 -07002267static void resv_map_put(struct vm_area_struct *vma)
2268{
Joonsoo Kimf522c3a2013-09-11 14:21:53 -07002269 struct resv_map *resv = vma_resv_map(vma);
Dave Hansenc50ac052012-05-29 15:06:46 -07002270
Joonsoo Kimf522c3a2013-09-11 14:21:53 -07002271 if (!resv)
Dave Hansenc50ac052012-05-29 15:06:46 -07002272 return;
Joonsoo Kimf522c3a2013-09-11 14:21:53 -07002273 kref_put(&resv->refs, resv_map_release);
Dave Hansenc50ac052012-05-29 15:06:46 -07002274}
2275
Mel Gormana1e78772008-07-23 21:27:23 -07002276static void hugetlb_vm_op_close(struct vm_area_struct *vma)
2277{
Andi Kleena5516432008-07-23 21:27:41 -07002278 struct hstate *h = hstate_vma(vma);
Joonsoo Kimf522c3a2013-09-11 14:21:53 -07002279 struct resv_map *resv = vma_resv_map(vma);
David Gibson90481622012-03-21 16:34:12 -07002280 struct hugepage_subpool *spool = subpool_vma(vma);
Andy Whitcroft84afd992008-07-23 21:27:32 -07002281 unsigned long reserve;
2282 unsigned long start;
2283 unsigned long end;
2284
Joonsoo Kimf522c3a2013-09-11 14:21:53 -07002285 if (resv) {
Andi Kleena5516432008-07-23 21:27:41 -07002286 start = vma_hugecache_offset(h, vma, vma->vm_start);
2287 end = vma_hugecache_offset(h, vma, vma->vm_end);
Andy Whitcroft84afd992008-07-23 21:27:32 -07002288
2289 reserve = (end - start) -
Joonsoo Kimf522c3a2013-09-11 14:21:53 -07002290 region_count(&resv->regions, start, end);
Andy Whitcroft84afd992008-07-23 21:27:32 -07002291
Dave Hansenc50ac052012-05-29 15:06:46 -07002292 resv_map_put(vma);
Andy Whitcroft84afd992008-07-23 21:27:32 -07002293
Adam Litke7251ff72008-07-23 21:27:59 -07002294 if (reserve) {
Andi Kleena5516432008-07-23 21:27:41 -07002295 hugetlb_acct_memory(h, -reserve);
David Gibson90481622012-03-21 16:34:12 -07002296 hugepage_subpool_put_pages(spool, reserve);
Adam Litke7251ff72008-07-23 21:27:59 -07002297 }
Andy Whitcroft84afd992008-07-23 21:27:32 -07002298 }
Mel Gormana1e78772008-07-23 21:27:23 -07002299}
2300
Linus Torvalds1da177e2005-04-16 15:20:36 -07002301/*
2302 * We cannot handle pagefaults against hugetlb pages at all. They cause
2303 * handle_mm_fault() to try to instantiate regular-sized pages in the
2304 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
2305 * this far.
2306 */
Nick Piggind0217ac2007-07-19 01:47:03 -07002307static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002308{
2309 BUG();
Nick Piggind0217ac2007-07-19 01:47:03 -07002310 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002311}
2312
Alexey Dobriyanf0f37e22009-09-27 22:29:37 +04002313const struct vm_operations_struct hugetlb_vm_ops = {
Nick Piggind0217ac2007-07-19 01:47:03 -07002314 .fault = hugetlb_vm_op_fault,
Andy Whitcroft84afd992008-07-23 21:27:32 -07002315 .open = hugetlb_vm_op_open,
Mel Gormana1e78772008-07-23 21:27:23 -07002316 .close = hugetlb_vm_op_close,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002317};
2318
David Gibson1e8f8892006-01-06 00:10:44 -08002319static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
2320 int writable)
David Gibson63551ae2005-06-21 17:14:44 -07002321{
2322 pte_t entry;
2323
David Gibson1e8f8892006-01-06 00:10:44 -08002324 if (writable) {
Gerald Schaefer106c9922013-04-29 15:07:23 -07002325 entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page,
2326 vma->vm_page_prot)));
David Gibson63551ae2005-06-21 17:14:44 -07002327 } else {
Gerald Schaefer106c9922013-04-29 15:07:23 -07002328 entry = huge_pte_wrprotect(mk_huge_pte(page,
2329 vma->vm_page_prot));
David Gibson63551ae2005-06-21 17:14:44 -07002330 }
2331 entry = pte_mkyoung(entry);
2332 entry = pte_mkhuge(entry);
Chris Metcalfd9ed9fa2012-04-01 14:01:34 -04002333 entry = arch_make_huge_pte(entry, vma, page, writable);
David Gibson63551ae2005-06-21 17:14:44 -07002334
2335 return entry;
2336}
2337
David Gibson1e8f8892006-01-06 00:10:44 -08002338static void set_huge_ptep_writable(struct vm_area_struct *vma,
2339 unsigned long address, pte_t *ptep)
2340{
2341 pte_t entry;
2342
Gerald Schaefer106c9922013-04-29 15:07:23 -07002343 entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep)));
Chris Forbes32f84522011-07-25 17:12:14 -07002344 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
Russell King4b3073e2009-12-18 16:40:18 +00002345 update_mmu_cache(vma, address, ptep);
David Gibson1e8f8892006-01-06 00:10:44 -08002346}
2347
2348
David Gibson63551ae2005-06-21 17:14:44 -07002349int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
2350 struct vm_area_struct *vma)
2351{
2352 pte_t *src_pte, *dst_pte, entry;
2353 struct page *ptepage;
Hugh Dickins1c598272005-10-19 21:23:43 -07002354 unsigned long addr;
David Gibson1e8f8892006-01-06 00:10:44 -08002355 int cow;
Andi Kleena5516432008-07-23 21:27:41 -07002356 struct hstate *h = hstate_vma(vma);
2357 unsigned long sz = huge_page_size(h);
David Gibson1e8f8892006-01-06 00:10:44 -08002358
2359 cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
David Gibson63551ae2005-06-21 17:14:44 -07002360
Andi Kleena5516432008-07-23 21:27:41 -07002361 for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002362 spinlock_t *src_ptl, *dst_ptl;
Hugh Dickinsc74df322005-10-29 18:16:23 -07002363 src_pte = huge_pte_offset(src, addr);
2364 if (!src_pte)
2365 continue;
Andi Kleena5516432008-07-23 21:27:41 -07002366 dst_pte = huge_pte_alloc(dst, addr, sz);
David Gibson63551ae2005-06-21 17:14:44 -07002367 if (!dst_pte)
2368 goto nomem;
Larry Woodmanc5c99422008-01-24 05:49:25 -08002369
2370 /* If the pagetables are shared don't copy or take references */
2371 if (dst_pte == src_pte)
2372 continue;
2373
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002374 dst_ptl = huge_pte_lock(h, dst, dst_pte);
2375 src_ptl = huge_pte_lockptr(h, src, src_pte);
2376 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07002377 if (!huge_pte_none(huge_ptep_get(src_pte))) {
David Gibson1e8f8892006-01-06 00:10:44 -08002378 if (cow)
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07002379 huge_ptep_set_wrprotect(src, addr, src_pte);
2380 entry = huge_ptep_get(src_pte);
Hugh Dickins1c598272005-10-19 21:23:43 -07002381 ptepage = pte_page(entry);
2382 get_page(ptepage);
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002383 page_dup_rmap(ptepage);
Hugh Dickins1c598272005-10-19 21:23:43 -07002384 set_huge_pte_at(dst, addr, dst_pte, entry);
2385 }
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002386 spin_unlock(src_ptl);
2387 spin_unlock(dst_ptl);
David Gibson63551ae2005-06-21 17:14:44 -07002388 }
2389 return 0;
2390
2391nomem:
2392 return -ENOMEM;
2393}
2394
Naoya Horiguchi290408d2010-09-08 10:19:35 +09002395static int is_hugetlb_entry_migration(pte_t pte)
2396{
2397 swp_entry_t swp;
2398
2399 if (huge_pte_none(pte) || pte_present(pte))
2400 return 0;
2401 swp = pte_to_swp_entry(pte);
Chris Forbes32f84522011-07-25 17:12:14 -07002402 if (non_swap_entry(swp) && is_migration_entry(swp))
Naoya Horiguchi290408d2010-09-08 10:19:35 +09002403 return 1;
Chris Forbes32f84522011-07-25 17:12:14 -07002404 else
Naoya Horiguchi290408d2010-09-08 10:19:35 +09002405 return 0;
2406}
2407
Naoya Horiguchifd6a03ed2010-05-28 09:29:21 +09002408static int is_hugetlb_entry_hwpoisoned(pte_t pte)
2409{
2410 swp_entry_t swp;
2411
2412 if (huge_pte_none(pte) || pte_present(pte))
2413 return 0;
2414 swp = pte_to_swp_entry(pte);
Chris Forbes32f84522011-07-25 17:12:14 -07002415 if (non_swap_entry(swp) && is_hwpoison_entry(swp))
Naoya Horiguchifd6a03ed2010-05-28 09:29:21 +09002416 return 1;
Chris Forbes32f84522011-07-25 17:12:14 -07002417 else
Naoya Horiguchifd6a03ed2010-05-28 09:29:21 +09002418 return 0;
2419}
2420
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07002421void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
2422 unsigned long start, unsigned long end,
2423 struct page *ref_page)
David Gibson63551ae2005-06-21 17:14:44 -07002424{
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07002425 int force_flush = 0;
David Gibson63551ae2005-06-21 17:14:44 -07002426 struct mm_struct *mm = vma->vm_mm;
2427 unsigned long address;
David Gibsonc7546f82005-08-05 11:59:35 -07002428 pte_t *ptep;
David Gibson63551ae2005-06-21 17:14:44 -07002429 pte_t pte;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002430 spinlock_t *ptl;
David Gibson63551ae2005-06-21 17:14:44 -07002431 struct page *page;
Andi Kleena5516432008-07-23 21:27:41 -07002432 struct hstate *h = hstate_vma(vma);
2433 unsigned long sz = huge_page_size(h);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07002434 const unsigned long mmun_start = start; /* For mmu_notifiers */
2435 const unsigned long mmun_end = end; /* For mmu_notifiers */
Andi Kleena5516432008-07-23 21:27:41 -07002436
David Gibson63551ae2005-06-21 17:14:44 -07002437 WARN_ON(!is_vm_hugetlb_page(vma));
Andi Kleena5516432008-07-23 21:27:41 -07002438 BUG_ON(start & ~huge_page_mask(h));
2439 BUG_ON(end & ~huge_page_mask(h));
David Gibson63551ae2005-06-21 17:14:44 -07002440
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07002441 tlb_start_vma(tlb, vma);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07002442 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07002443again:
Andi Kleena5516432008-07-23 21:27:41 -07002444 for (address = start; address < end; address += sz) {
David Gibsonc7546f82005-08-05 11:59:35 -07002445 ptep = huge_pte_offset(mm, address);
Adam Litke4c887262005-10-29 18:16:46 -07002446 if (!ptep)
David Gibsonc7546f82005-08-05 11:59:35 -07002447 continue;
2448
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002449 ptl = huge_pte_lock(h, mm, ptep);
Chen, Kenneth W39dde652006-12-06 20:32:03 -08002450 if (huge_pmd_unshare(mm, &address, ptep))
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002451 goto unlock;
Chen, Kenneth W39dde652006-12-06 20:32:03 -08002452
Hillf Danton66293262012-03-23 15:01:48 -07002453 pte = huge_ptep_get(ptep);
2454 if (huge_pte_none(pte))
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002455 goto unlock;
Hillf Danton66293262012-03-23 15:01:48 -07002456
2457 /*
2458 * HWPoisoned hugepage is already unmapped and dropped reference
2459 */
Naoya Horiguchi8c4894c2012-12-12 13:52:28 -08002460 if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) {
Gerald Schaefer106c9922013-04-29 15:07:23 -07002461 huge_pte_clear(mm, address, ptep);
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002462 goto unlock;
Naoya Horiguchi8c4894c2012-12-12 13:52:28 -08002463 }
Hillf Danton66293262012-03-23 15:01:48 -07002464
2465 page = pte_page(pte);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002466 /*
2467 * If a reference page is supplied, it is because a specific
2468 * page is being unmapped, not a range. Ensure the page we
2469 * are about to unmap is the actual page of interest.
2470 */
2471 if (ref_page) {
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002472 if (page != ref_page)
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002473 goto unlock;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002474
2475 /*
2476 * Mark the VMA as having unmapped its page so that
2477 * future faults in this VMA will fail rather than
2478 * looking like data was lost
2479 */
2480 set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
2481 }
2482
David Gibsonc7546f82005-08-05 11:59:35 -07002483 pte = huge_ptep_get_and_clear(mm, address, ptep);
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07002484 tlb_remove_tlb_entry(tlb, ptep, address);
Gerald Schaefer106c9922013-04-29 15:07:23 -07002485 if (huge_pte_dirty(pte))
Ken Chen6649a382007-02-08 14:20:27 -08002486 set_page_dirty(page);
Hillf Danton9e811302012-03-21 16:34:03 -07002487
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07002488 page_remove_rmap(page);
2489 force_flush = !__tlb_remove_page(tlb, page);
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002490 if (force_flush) {
2491 spin_unlock(ptl);
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07002492 break;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002493 }
Hillf Danton9e811302012-03-21 16:34:03 -07002494 /* Bail out after unmapping reference page if supplied */
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002495 if (ref_page) {
2496 spin_unlock(ptl);
Hillf Danton9e811302012-03-21 16:34:03 -07002497 break;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002498 }
2499unlock:
2500 spin_unlock(ptl);
David Gibson63551ae2005-06-21 17:14:44 -07002501 }
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07002502 /*
2503 * mmu_gather ran out of room to batch pages, we break out of
2504 * the PTE lock to avoid doing the potential expensive TLB invalidate
2505 * and page-free while holding it.
2506 */
2507 if (force_flush) {
2508 force_flush = 0;
2509 tlb_flush_mmu(tlb);
2510 if (address < end && !ref_page)
2511 goto again;
Chen, Kenneth Wfe1668a2006-10-04 02:15:24 -07002512 }
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07002513 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07002514 tlb_end_vma(tlb, vma);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002515}
David Gibson63551ae2005-06-21 17:14:44 -07002516
Mel Gormand8333522012-07-31 16:46:20 -07002517void __unmap_hugepage_range_final(struct mmu_gather *tlb,
2518 struct vm_area_struct *vma, unsigned long start,
2519 unsigned long end, struct page *ref_page)
2520{
2521 __unmap_hugepage_range(tlb, vma, start, end, ref_page);
2522
2523 /*
2524 * Clear this flag so that x86's huge_pmd_share page_table_shareable
2525 * test will fail on a vma being torn down, and not grab a page table
2526 * on its way out. We're lucky that the flag has such an appropriate
2527 * name, and can in fact be safely cleared here. We could clear it
2528 * before the __unmap_hugepage_range above, but all that's necessary
2529 * is to clear it before releasing the i_mmap_mutex. This works
2530 * because in the context this is called, the VMA is about to be
2531 * destroyed and the i_mmap_mutex is held.
2532 */
2533 vma->vm_flags &= ~VM_MAYSHARE;
2534}
2535
Chen, Kenneth W502717f2006-10-11 01:20:46 -07002536void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002537 unsigned long end, struct page *ref_page)
Chen, Kenneth W502717f2006-10-11 01:20:46 -07002538{
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07002539 struct mm_struct *mm;
2540 struct mmu_gather tlb;
2541
2542 mm = vma->vm_mm;
2543
Linus Torvalds2b047252013-08-15 11:42:25 -07002544 tlb_gather_mmu(&tlb, mm, start, end);
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07002545 __unmap_hugepage_range(&tlb, vma, start, end, ref_page);
2546 tlb_finish_mmu(&tlb, start, end);
Chen, Kenneth W502717f2006-10-11 01:20:46 -07002547}
2548
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002549/*
2550 * This is called when the original mapper is failing to COW a MAP_PRIVATE
2551 * mappping it owns the reserve page for. The intention is to unmap the page
2552 * from other VMAs and let the children be SIGKILLed if they are faulting the
2553 * same region.
2554 */
Harvey Harrison2a4b3de2008-10-18 20:27:06 -07002555static int unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
2556 struct page *page, unsigned long address)
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002557{
Adam Litke75266742008-11-12 13:24:56 -08002558 struct hstate *h = hstate_vma(vma);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002559 struct vm_area_struct *iter_vma;
2560 struct address_space *mapping;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002561 pgoff_t pgoff;
2562
2563 /*
2564 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
2565 * from page cache lookup which is in HPAGE_SIZE units.
2566 */
Adam Litke75266742008-11-12 13:24:56 -08002567 address = address & huge_page_mask(h);
Michal Hocko36e4f202012-10-08 16:33:31 -07002568 pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) +
2569 vma->vm_pgoff;
Al Viro496ad9a2013-01-23 17:07:38 -05002570 mapping = file_inode(vma->vm_file)->i_mapping;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002571
Mel Gorman4eb2b1d2009-12-14 17:59:53 -08002572 /*
2573 * Take the mapping lock for the duration of the table walk. As
2574 * this mapping should be shared between all the VMAs,
2575 * __unmap_hugepage_range() is called as the lock is already held
2576 */
Peter Zijlstra3d48ae42011-05-24 17:12:06 -07002577 mutex_lock(&mapping->i_mmap_mutex);
Michel Lespinasse6b2dbba2012-10-08 16:31:25 -07002578 vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) {
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002579 /* Do not unmap the current VMA */
2580 if (iter_vma == vma)
2581 continue;
2582
2583 /*
2584 * Unmap the page from other VMAs without their own reserves.
2585 * They get marked to be SIGKILLed if they fault in these
2586 * areas. This is because a future no-page fault on this VMA
2587 * could insert a zeroed page instead of the data existing
2588 * from the time of fork. This would look like data corruption
2589 */
2590 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07002591 unmap_hugepage_range(iter_vma, address,
2592 address + huge_page_size(h), page);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002593 }
Peter Zijlstra3d48ae42011-05-24 17:12:06 -07002594 mutex_unlock(&mapping->i_mmap_mutex);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002595
2596 return 1;
2597}
2598
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002599/*
2600 * Hugetlb_cow() should be called with page lock of the original hugepage held.
Michal Hockoef009b22012-01-10 15:07:21 -08002601 * Called with hugetlb_instantiation_mutex held and pte_page locked so we
2602 * cannot race with other handlers or page migration.
2603 * Keep the pte_same checks anyway to make transition from the mutex easier.
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002604 */
David Gibson1e8f8892006-01-06 00:10:44 -08002605static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002606 unsigned long address, pte_t *ptep, pte_t pte,
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002607 struct page *pagecache_page, spinlock_t *ptl)
David Gibson1e8f8892006-01-06 00:10:44 -08002608{
Andi Kleena5516432008-07-23 21:27:41 -07002609 struct hstate *h = hstate_vma(vma);
David Gibson1e8f8892006-01-06 00:10:44 -08002610 struct page *old_page, *new_page;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002611 int outside_reserve = 0;
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07002612 unsigned long mmun_start; /* For mmu_notifiers */
2613 unsigned long mmun_end; /* For mmu_notifiers */
David Gibson1e8f8892006-01-06 00:10:44 -08002614
2615 old_page = pte_page(pte);
2616
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002617retry_avoidcopy:
David Gibson1e8f8892006-01-06 00:10:44 -08002618 /* If no-one else is actually using this page, avoid the copy
2619 * and just make the page writable */
Joonsoo Kim37a21402013-09-11 14:21:04 -07002620 if (page_mapcount(old_page) == 1 && PageAnon(old_page)) {
2621 page_move_anon_rmap(old_page, vma, address);
David Gibson1e8f8892006-01-06 00:10:44 -08002622 set_huge_ptep_writable(vma, address, ptep);
Nick Piggin83c54072007-07-19 01:47:05 -07002623 return 0;
David Gibson1e8f8892006-01-06 00:10:44 -08002624 }
2625
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002626 /*
2627 * If the process that created a MAP_PRIVATE mapping is about to
2628 * perform a COW due to a shared page count, attempt to satisfy
2629 * the allocation without using the existing reserves. The pagecache
2630 * page is used to determine if the reserve at this address was
2631 * consumed or not. If reserves were used, a partial faulted mapping
2632 * at the time of fork() could consume its reserves on COW instead
2633 * of the full address range.
2634 */
Joonsoo Kim5944d012013-09-11 14:21:55 -07002635 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002636 old_page != pagecache_page)
2637 outside_reserve = 1;
2638
David Gibson1e8f8892006-01-06 00:10:44 -08002639 page_cache_get(old_page);
Larry Woodmanb76c8cf2009-12-14 17:59:37 -08002640
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002641 /* Drop page table lock as buddy allocator may be called */
2642 spin_unlock(ptl);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002643 new_page = alloc_huge_page(vma, address, outside_reserve);
David Gibson1e8f8892006-01-06 00:10:44 -08002644
Adam Litke2fc39ce2007-11-14 16:59:39 -08002645 if (IS_ERR(new_page)) {
Aneesh Kumar K.V76dcee72012-07-31 16:41:57 -07002646 long err = PTR_ERR(new_page);
David Gibson1e8f8892006-01-06 00:10:44 -08002647 page_cache_release(old_page);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002648
2649 /*
2650 * If a process owning a MAP_PRIVATE mapping fails to COW,
2651 * it is due to references held by a child and an insufficient
2652 * huge page pool. To guarantee the original mappers
2653 * reliability, unmap the page from child processes. The child
2654 * may get SIGKILLed if it later faults.
2655 */
2656 if (outside_reserve) {
2657 BUG_ON(huge_pte_none(pte));
2658 if (unmap_ref_private(mm, vma, old_page, address)) {
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002659 BUG_ON(huge_pte_none(pte));
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002660 spin_lock(ptl);
Hillf Dantona734bcc2012-01-10 15:07:20 -08002661 ptep = huge_pte_offset(mm, address & huge_page_mask(h));
2662 if (likely(pte_same(huge_ptep_get(ptep), pte)))
2663 goto retry_avoidcopy;
2664 /*
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002665 * race occurs while re-acquiring page table
2666 * lock, and our job is done.
Hillf Dantona734bcc2012-01-10 15:07:20 -08002667 */
2668 return 0;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002669 }
2670 WARN_ON_ONCE(1);
2671 }
2672
Larry Woodmanb76c8cf2009-12-14 17:59:37 -08002673 /* Caller expects lock to be held */
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002674 spin_lock(ptl);
Aneesh Kumar K.V76dcee72012-07-31 16:41:57 -07002675 if (err == -ENOMEM)
2676 return VM_FAULT_OOM;
2677 else
2678 return VM_FAULT_SIGBUS;
David Gibson1e8f8892006-01-06 00:10:44 -08002679 }
2680
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002681 /*
2682 * When the original hugepage is shared one, it does not have
2683 * anon_vma prepared.
2684 */
Dean Nelson44e2aa92010-10-26 14:22:08 -07002685 if (unlikely(anon_vma_prepare(vma))) {
Hillf Dantonea4039a2011-11-15 14:36:12 -08002686 page_cache_release(new_page);
2687 page_cache_release(old_page);
Dean Nelson44e2aa92010-10-26 14:22:08 -07002688 /* Caller expects lock to be held */
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002689 spin_lock(ptl);
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002690 return VM_FAULT_OOM;
Dean Nelson44e2aa92010-10-26 14:22:08 -07002691 }
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002692
Andrea Arcangeli47ad8472011-01-13 15:46:47 -08002693 copy_user_huge_page(new_page, old_page, address, vma,
2694 pages_per_huge_page(h));
Nick Piggin0ed361d2008-02-04 22:29:34 -08002695 __SetPageUptodate(new_page);
David Gibson1e8f8892006-01-06 00:10:44 -08002696
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07002697 mmun_start = address & huge_page_mask(h);
2698 mmun_end = mmun_start + huge_page_size(h);
2699 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
Larry Woodmanb76c8cf2009-12-14 17:59:37 -08002700 /*
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002701 * Retake the page table lock to check for racing updates
Larry Woodmanb76c8cf2009-12-14 17:59:37 -08002702 * before the page tables are altered
2703 */
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002704 spin_lock(ptl);
Andi Kleena5516432008-07-23 21:27:41 -07002705 ptep = huge_pte_offset(mm, address & huge_page_mask(h));
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07002706 if (likely(pte_same(huge_ptep_get(ptep), pte))) {
Joonsoo Kim07443a82013-09-11 14:21:58 -07002707 ClearPagePrivate(new_page);
2708
David Gibson1e8f8892006-01-06 00:10:44 -08002709 /* Break COW */
Gerald Schaefer8fe627e2008-04-28 02:13:28 -07002710 huge_ptep_clear_flush(vma, address, ptep);
David Gibson1e8f8892006-01-06 00:10:44 -08002711 set_huge_pte_at(mm, address, ptep,
2712 make_huge_pte(vma, new_page, 1));
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002713 page_remove_rmap(old_page);
Naoya Horiguchicd67f0d2010-09-10 13:23:04 +09002714 hugepage_add_new_anon_rmap(new_page, vma, address);
David Gibson1e8f8892006-01-06 00:10:44 -08002715 /* Make the old page be freed below */
2716 new_page = old_page;
2717 }
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002718 spin_unlock(ptl);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07002719 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
David Gibson1e8f8892006-01-06 00:10:44 -08002720 page_cache_release(new_page);
2721 page_cache_release(old_page);
Joonsoo Kim83120342013-09-11 14:21:57 -07002722
2723 /* Caller expects lock to be held */
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002724 spin_lock(ptl);
Nick Piggin83c54072007-07-19 01:47:05 -07002725 return 0;
David Gibson1e8f8892006-01-06 00:10:44 -08002726}
2727
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002728/* Return the pagecache page at a given address within a VMA */
Andi Kleena5516432008-07-23 21:27:41 -07002729static struct page *hugetlbfs_pagecache_page(struct hstate *h,
2730 struct vm_area_struct *vma, unsigned long address)
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002731{
2732 struct address_space *mapping;
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -07002733 pgoff_t idx;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002734
2735 mapping = vma->vm_file->f_mapping;
Andi Kleena5516432008-07-23 21:27:41 -07002736 idx = vma_hugecache_offset(h, vma, address);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002737
2738 return find_lock_page(mapping, idx);
2739}
2740
Hugh Dickins3ae77f42009-09-21 17:03:33 -07002741/*
2742 * Return whether there is a pagecache page to back given address within VMA.
2743 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
2744 */
2745static bool hugetlbfs_pagecache_present(struct hstate *h,
Hugh Dickins2a15efc2009-09-21 17:03:27 -07002746 struct vm_area_struct *vma, unsigned long address)
2747{
2748 struct address_space *mapping;
2749 pgoff_t idx;
2750 struct page *page;
2751
2752 mapping = vma->vm_file->f_mapping;
2753 idx = vma_hugecache_offset(h, vma, address);
2754
2755 page = find_get_page(mapping, idx);
2756 if (page)
2757 put_page(page);
2758 return page != NULL;
2759}
2760
Robert P. J. Daya1ed3dd2007-07-17 04:03:33 -07002761static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
Hugh Dickins788c7df2009-06-23 13:49:05 +01002762 unsigned long address, pte_t *ptep, unsigned int flags)
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01002763{
Andi Kleena5516432008-07-23 21:27:41 -07002764 struct hstate *h = hstate_vma(vma);
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01002765 int ret = VM_FAULT_SIGBUS;
Hillf Danton409eb8c2012-01-20 14:34:13 -08002766 int anon_rmap = 0;
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -07002767 pgoff_t idx;
Adam Litke4c887262005-10-29 18:16:46 -07002768 unsigned long size;
Adam Litke4c887262005-10-29 18:16:46 -07002769 struct page *page;
2770 struct address_space *mapping;
David Gibson1e8f8892006-01-06 00:10:44 -08002771 pte_t new_pte;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002772 spinlock_t *ptl;
Adam Litke4c887262005-10-29 18:16:46 -07002773
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002774 /*
2775 * Currently, we are forced to kill the process in the event the
2776 * original mapper has unmapped pages from the child due to a failed
Lucas De Marchi25985ed2011-03-30 22:57:33 -03002777 * COW. Warn that such a situation has occurred as it may not be obvious
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002778 */
2779 if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
Andrew Mortonffb22af2013-02-22 16:32:08 -08002780 pr_warning("PID %d killed due to inadequate hugepage pool\n",
2781 current->pid);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002782 return ret;
2783 }
2784
Adam Litke4c887262005-10-29 18:16:46 -07002785 mapping = vma->vm_file->f_mapping;
Andi Kleena5516432008-07-23 21:27:41 -07002786 idx = vma_hugecache_offset(h, vma, address);
Adam Litke4c887262005-10-29 18:16:46 -07002787
2788 /*
2789 * Use page lock to guard against racing truncation
2790 * before we get page_table_lock.
2791 */
Christoph Lameter6bda6662006-01-06 00:10:49 -08002792retry:
2793 page = find_lock_page(mapping, idx);
2794 if (!page) {
Andi Kleena5516432008-07-23 21:27:41 -07002795 size = i_size_read(mapping->host) >> huge_page_shift(h);
Hugh Dickinsebed4bf2006-10-28 10:38:43 -07002796 if (idx >= size)
2797 goto out;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002798 page = alloc_huge_page(vma, address, 0);
Adam Litke2fc39ce2007-11-14 16:59:39 -08002799 if (IS_ERR(page)) {
Aneesh Kumar K.V76dcee72012-07-31 16:41:57 -07002800 ret = PTR_ERR(page);
2801 if (ret == -ENOMEM)
2802 ret = VM_FAULT_OOM;
2803 else
2804 ret = VM_FAULT_SIGBUS;
Christoph Lameter6bda6662006-01-06 00:10:49 -08002805 goto out;
2806 }
Andrea Arcangeli47ad8472011-01-13 15:46:47 -08002807 clear_huge_page(page, address, pages_per_huge_page(h));
Nick Piggin0ed361d2008-02-04 22:29:34 -08002808 __SetPageUptodate(page);
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01002809
Mel Gormanf83a2752009-05-28 14:34:40 -07002810 if (vma->vm_flags & VM_MAYSHARE) {
Christoph Lameter6bda6662006-01-06 00:10:49 -08002811 int err;
Ken Chen45c682a2007-11-14 16:59:44 -08002812 struct inode *inode = mapping->host;
Christoph Lameter6bda6662006-01-06 00:10:49 -08002813
2814 err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
2815 if (err) {
2816 put_page(page);
Christoph Lameter6bda6662006-01-06 00:10:49 -08002817 if (err == -EEXIST)
2818 goto retry;
2819 goto out;
2820 }
Joonsoo Kim07443a82013-09-11 14:21:58 -07002821 ClearPagePrivate(page);
Ken Chen45c682a2007-11-14 16:59:44 -08002822
2823 spin_lock(&inode->i_lock);
Andi Kleena5516432008-07-23 21:27:41 -07002824 inode->i_blocks += blocks_per_huge_page(h);
Ken Chen45c682a2007-11-14 16:59:44 -08002825 spin_unlock(&inode->i_lock);
Mel Gorman23be7462010-04-23 13:17:56 -04002826 } else {
Christoph Lameter6bda6662006-01-06 00:10:49 -08002827 lock_page(page);
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002828 if (unlikely(anon_vma_prepare(vma))) {
2829 ret = VM_FAULT_OOM;
2830 goto backout_unlocked;
2831 }
Hillf Danton409eb8c2012-01-20 14:34:13 -08002832 anon_rmap = 1;
Mel Gorman23be7462010-04-23 13:17:56 -04002833 }
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002834 } else {
Naoya Horiguchi998b4382010-09-08 10:19:32 +09002835 /*
2836 * If memory error occurs between mmap() and fault, some process
2837 * don't have hwpoisoned swap entry for errored virtual address.
2838 * So we need to block hugepage fault by PG_hwpoison bit check.
2839 */
2840 if (unlikely(PageHWPoison(page))) {
Chris Forbes32f84522011-07-25 17:12:14 -07002841 ret = VM_FAULT_HWPOISON |
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07002842 VM_FAULT_SET_HINDEX(hstate_index(h));
Naoya Horiguchi998b4382010-09-08 10:19:32 +09002843 goto backout_unlocked;
2844 }
Christoph Lameter6bda6662006-01-06 00:10:49 -08002845 }
David Gibson1e8f8892006-01-06 00:10:44 -08002846
Andy Whitcroft57303d82008-08-12 15:08:47 -07002847 /*
2848 * If we are going to COW a private mapping later, we examine the
2849 * pending reservations for this page now. This will ensure that
2850 * any allocations necessary to record that reservation occur outside
2851 * the spinlock.
2852 */
Hugh Dickins788c7df2009-06-23 13:49:05 +01002853 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED))
Andy Whitcroft2b267362008-08-12 15:08:49 -07002854 if (vma_needs_reservation(h, vma, address) < 0) {
2855 ret = VM_FAULT_OOM;
2856 goto backout_unlocked;
2857 }
Andy Whitcroft57303d82008-08-12 15:08:47 -07002858
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002859 ptl = huge_pte_lockptr(h, mm, ptep);
2860 spin_lock(ptl);
Andi Kleena5516432008-07-23 21:27:41 -07002861 size = i_size_read(mapping->host) >> huge_page_shift(h);
Adam Litke4c887262005-10-29 18:16:46 -07002862 if (idx >= size)
2863 goto backout;
2864
Nick Piggin83c54072007-07-19 01:47:05 -07002865 ret = 0;
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07002866 if (!huge_pte_none(huge_ptep_get(ptep)))
Adam Litke4c887262005-10-29 18:16:46 -07002867 goto backout;
2868
Joonsoo Kim07443a82013-09-11 14:21:58 -07002869 if (anon_rmap) {
2870 ClearPagePrivate(page);
Hillf Danton409eb8c2012-01-20 14:34:13 -08002871 hugepage_add_new_anon_rmap(page, vma, address);
Joonsoo Kim07443a82013-09-11 14:21:58 -07002872 }
Hillf Danton409eb8c2012-01-20 14:34:13 -08002873 else
2874 page_dup_rmap(page);
David Gibson1e8f8892006-01-06 00:10:44 -08002875 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
2876 && (vma->vm_flags & VM_SHARED)));
2877 set_huge_pte_at(mm, address, ptep, new_pte);
2878
Hugh Dickins788c7df2009-06-23 13:49:05 +01002879 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
David Gibson1e8f8892006-01-06 00:10:44 -08002880 /* Optimization, do the COW without a second fault */
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002881 ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page, ptl);
David Gibson1e8f8892006-01-06 00:10:44 -08002882 }
2883
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002884 spin_unlock(ptl);
Adam Litke4c887262005-10-29 18:16:46 -07002885 unlock_page(page);
2886out:
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01002887 return ret;
Adam Litke4c887262005-10-29 18:16:46 -07002888
2889backout:
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002890 spin_unlock(ptl);
Andy Whitcroft2b267362008-08-12 15:08:49 -07002891backout_unlocked:
Adam Litke4c887262005-10-29 18:16:46 -07002892 unlock_page(page);
2893 put_page(page);
2894 goto out;
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01002895}
2896
Adam Litke86e52162006-01-06 00:10:43 -08002897int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
Hugh Dickins788c7df2009-06-23 13:49:05 +01002898 unsigned long address, unsigned int flags)
Adam Litke86e52162006-01-06 00:10:43 -08002899{
2900 pte_t *ptep;
2901 pte_t entry;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002902 spinlock_t *ptl;
David Gibson1e8f8892006-01-06 00:10:44 -08002903 int ret;
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002904 struct page *page = NULL;
Andy Whitcroft57303d82008-08-12 15:08:47 -07002905 struct page *pagecache_page = NULL;
David Gibson3935baa2006-03-22 00:08:53 -08002906 static DEFINE_MUTEX(hugetlb_instantiation_mutex);
Andi Kleena5516432008-07-23 21:27:41 -07002907 struct hstate *h = hstate_vma(vma);
Adam Litke86e52162006-01-06 00:10:43 -08002908
KAMEZAWA Hiroyuki1e16a532012-01-10 15:07:22 -08002909 address &= huge_page_mask(h);
2910
Naoya Horiguchifd6a03ed2010-05-28 09:29:21 +09002911 ptep = huge_pte_offset(mm, address);
2912 if (ptep) {
2913 entry = huge_ptep_get(ptep);
Naoya Horiguchi290408d2010-09-08 10:19:35 +09002914 if (unlikely(is_hugetlb_entry_migration(entry))) {
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002915 migration_entry_wait_huge(vma, mm, ptep);
Naoya Horiguchi290408d2010-09-08 10:19:35 +09002916 return 0;
2917 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
Chris Forbes32f84522011-07-25 17:12:14 -07002918 return VM_FAULT_HWPOISON_LARGE |
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07002919 VM_FAULT_SET_HINDEX(hstate_index(h));
Naoya Horiguchifd6a03ed2010-05-28 09:29:21 +09002920 }
2921
Andi Kleena5516432008-07-23 21:27:41 -07002922 ptep = huge_pte_alloc(mm, address, huge_page_size(h));
Adam Litke86e52162006-01-06 00:10:43 -08002923 if (!ptep)
2924 return VM_FAULT_OOM;
2925
David Gibson3935baa2006-03-22 00:08:53 -08002926 /*
2927 * Serialize hugepage allocation and instantiation, so that we don't
2928 * get spurious allocation failures if two CPUs race to instantiate
2929 * the same page in the page cache.
2930 */
2931 mutex_lock(&hugetlb_instantiation_mutex);
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07002932 entry = huge_ptep_get(ptep);
2933 if (huge_pte_none(entry)) {
Hugh Dickins788c7df2009-06-23 13:49:05 +01002934 ret = hugetlb_no_page(mm, vma, address, ptep, flags);
David Gibsonb4d1d992008-10-15 22:01:11 -07002935 goto out_mutex;
David Gibson3935baa2006-03-22 00:08:53 -08002936 }
Adam Litke86e52162006-01-06 00:10:43 -08002937
Nick Piggin83c54072007-07-19 01:47:05 -07002938 ret = 0;
David Gibson1e8f8892006-01-06 00:10:44 -08002939
Andy Whitcroft57303d82008-08-12 15:08:47 -07002940 /*
2941 * If we are going to COW the mapping later, we examine the pending
2942 * reservations for this page now. This will ensure that any
2943 * allocations necessary to record that reservation occur outside the
2944 * spinlock. For private mappings, we also lookup the pagecache
2945 * page now as it is used to determine if a reservation has been
2946 * consumed.
2947 */
Gerald Schaefer106c9922013-04-29 15:07:23 -07002948 if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) {
Andy Whitcroft2b267362008-08-12 15:08:49 -07002949 if (vma_needs_reservation(h, vma, address) < 0) {
2950 ret = VM_FAULT_OOM;
David Gibsonb4d1d992008-10-15 22:01:11 -07002951 goto out_mutex;
Andy Whitcroft2b267362008-08-12 15:08:49 -07002952 }
Andy Whitcroft57303d82008-08-12 15:08:47 -07002953
Mel Gormanf83a2752009-05-28 14:34:40 -07002954 if (!(vma->vm_flags & VM_MAYSHARE))
Andy Whitcroft57303d82008-08-12 15:08:47 -07002955 pagecache_page = hugetlbfs_pagecache_page(h,
2956 vma, address);
2957 }
2958
Naoya Horiguchi56c9cfb2010-09-10 13:23:04 +09002959 /*
2960 * hugetlb_cow() requires page locks of pte_page(entry) and
2961 * pagecache_page, so here we need take the former one
2962 * when page != pagecache_page or !pagecache_page.
2963 * Note that locking order is always pagecache_page -> page,
2964 * so no worry about deadlock.
2965 */
2966 page = pte_page(entry);
Chris Metcalf66aebce2012-04-12 12:49:15 -07002967 get_page(page);
Naoya Horiguchi56c9cfb2010-09-10 13:23:04 +09002968 if (page != pagecache_page)
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002969 lock_page(page);
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002970
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002971 ptl = huge_pte_lockptr(h, mm, ptep);
2972 spin_lock(ptl);
David Gibson1e8f8892006-01-06 00:10:44 -08002973 /* Check for a racing update before calling hugetlb_cow */
David Gibsonb4d1d992008-10-15 22:01:11 -07002974 if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002975 goto out_ptl;
David Gibsonb4d1d992008-10-15 22:01:11 -07002976
2977
Hugh Dickins788c7df2009-06-23 13:49:05 +01002978 if (flags & FAULT_FLAG_WRITE) {
Gerald Schaefer106c9922013-04-29 15:07:23 -07002979 if (!huge_pte_write(entry)) {
Andy Whitcroft57303d82008-08-12 15:08:47 -07002980 ret = hugetlb_cow(mm, vma, address, ptep, entry,
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002981 pagecache_page, ptl);
2982 goto out_ptl;
David Gibsonb4d1d992008-10-15 22:01:11 -07002983 }
Gerald Schaefer106c9922013-04-29 15:07:23 -07002984 entry = huge_pte_mkdirty(entry);
David Gibsonb4d1d992008-10-15 22:01:11 -07002985 }
2986 entry = pte_mkyoung(entry);
Hugh Dickins788c7df2009-06-23 13:49:05 +01002987 if (huge_ptep_set_access_flags(vma, address, ptep, entry,
2988 flags & FAULT_FLAG_WRITE))
Russell King4b3073e2009-12-18 16:40:18 +00002989 update_mmu_cache(vma, address, ptep);
David Gibsonb4d1d992008-10-15 22:01:11 -07002990
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002991out_ptl:
2992 spin_unlock(ptl);
Andy Whitcroft57303d82008-08-12 15:08:47 -07002993
2994 if (pagecache_page) {
2995 unlock_page(pagecache_page);
2996 put_page(pagecache_page);
2997 }
Dean Nelson1f64d692010-12-02 14:31:12 -08002998 if (page != pagecache_page)
2999 unlock_page(page);
Chris Metcalf66aebce2012-04-12 12:49:15 -07003000 put_page(page);
Andy Whitcroft57303d82008-08-12 15:08:47 -07003001
David Gibsonb4d1d992008-10-15 22:01:11 -07003002out_mutex:
David Gibson3935baa2006-03-22 00:08:53 -08003003 mutex_unlock(&hugetlb_instantiation_mutex);
David Gibson1e8f8892006-01-06 00:10:44 -08003004
3005 return ret;
Adam Litke86e52162006-01-06 00:10:43 -08003006}
3007
Michel Lespinasse28a35712013-02-22 16:35:55 -08003008long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
3009 struct page **pages, struct vm_area_struct **vmas,
3010 unsigned long *position, unsigned long *nr_pages,
3011 long i, unsigned int flags)
David Gibson63551ae2005-06-21 17:14:44 -07003012{
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08003013 unsigned long pfn_offset;
3014 unsigned long vaddr = *position;
Michel Lespinasse28a35712013-02-22 16:35:55 -08003015 unsigned long remainder = *nr_pages;
Andi Kleena5516432008-07-23 21:27:41 -07003016 struct hstate *h = hstate_vma(vma);
David Gibson63551ae2005-06-21 17:14:44 -07003017
David Gibson63551ae2005-06-21 17:14:44 -07003018 while (vaddr < vma->vm_end && remainder) {
Adam Litke4c887262005-10-29 18:16:46 -07003019 pte_t *pte;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003020 spinlock_t *ptl = NULL;
Hugh Dickins2a15efc2009-09-21 17:03:27 -07003021 int absent;
Adam Litke4c887262005-10-29 18:16:46 -07003022 struct page *page;
3023
3024 /*
3025 * Some archs (sparc64, sh*) have multiple pte_ts to
Hugh Dickins2a15efc2009-09-21 17:03:27 -07003026 * each hugepage. We have to make sure we get the
Adam Litke4c887262005-10-29 18:16:46 -07003027 * first, for the page indexing below to work.
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003028 *
3029 * Note that page table lock is not held when pte is null.
Adam Litke4c887262005-10-29 18:16:46 -07003030 */
Andi Kleena5516432008-07-23 21:27:41 -07003031 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h));
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003032 if (pte)
3033 ptl = huge_pte_lock(h, mm, pte);
Hugh Dickins2a15efc2009-09-21 17:03:27 -07003034 absent = !pte || huge_pte_none(huge_ptep_get(pte));
Adam Litke4c887262005-10-29 18:16:46 -07003035
Hugh Dickins2a15efc2009-09-21 17:03:27 -07003036 /*
3037 * When coredumping, it suits get_dump_page if we just return
Hugh Dickins3ae77f42009-09-21 17:03:33 -07003038 * an error where there's an empty slot with no huge pagecache
3039 * to back it. This way, we avoid allocating a hugepage, and
3040 * the sparse dumpfile avoids allocating disk blocks, but its
3041 * huge holes still show up with zeroes where they need to be.
Hugh Dickins2a15efc2009-09-21 17:03:27 -07003042 */
Hugh Dickins3ae77f42009-09-21 17:03:33 -07003043 if (absent && (flags & FOLL_DUMP) &&
3044 !hugetlbfs_pagecache_present(h, vma, vaddr)) {
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003045 if (pte)
3046 spin_unlock(ptl);
Hugh Dickins2a15efc2009-09-21 17:03:27 -07003047 remainder = 0;
3048 break;
3049 }
3050
Naoya Horiguchi9cc3a5b2013-04-17 15:58:30 -07003051 /*
3052 * We need call hugetlb_fault for both hugepages under migration
3053 * (in which case hugetlb_fault waits for the migration,) and
3054 * hwpoisoned hugepages (in which case we need to prevent the
3055 * caller from accessing to them.) In order to do this, we use
3056 * here is_swap_pte instead of is_hugetlb_entry_migration and
3057 * is_hugetlb_entry_hwpoisoned. This is because it simply covers
3058 * both cases, and because we can't follow correct pages
3059 * directly from any kind of swap entries.
3060 */
3061 if (absent || is_swap_pte(huge_ptep_get(pte)) ||
Gerald Schaefer106c9922013-04-29 15:07:23 -07003062 ((flags & FOLL_WRITE) &&
3063 !huge_pte_write(huge_ptep_get(pte)))) {
Adam Litke4c887262005-10-29 18:16:46 -07003064 int ret;
3065
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003066 if (pte)
3067 spin_unlock(ptl);
Hugh Dickins2a15efc2009-09-21 17:03:27 -07003068 ret = hugetlb_fault(mm, vma, vaddr,
3069 (flags & FOLL_WRITE) ? FAULT_FLAG_WRITE : 0);
Adam Litkea89182c2007-08-22 14:01:51 -07003070 if (!(ret & VM_FAULT_ERROR))
Adam Litke4c887262005-10-29 18:16:46 -07003071 continue;
3072
3073 remainder = 0;
Adam Litke4c887262005-10-29 18:16:46 -07003074 break;
3075 }
David Gibson63551ae2005-06-21 17:14:44 -07003076
Andi Kleena5516432008-07-23 21:27:41 -07003077 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07003078 page = pte_page(huge_ptep_get(pte));
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08003079same_page:
Chen, Kenneth Wd6692182006-03-31 02:29:57 -08003080 if (pages) {
Hugh Dickins2a15efc2009-09-21 17:03:27 -07003081 pages[i] = mem_map_offset(page, pfn_offset);
KOSAKI Motohiro4b2e38a2008-10-18 20:27:10 -07003082 get_page(pages[i]);
Chen, Kenneth Wd6692182006-03-31 02:29:57 -08003083 }
David Gibson63551ae2005-06-21 17:14:44 -07003084
3085 if (vmas)
3086 vmas[i] = vma;
3087
3088 vaddr += PAGE_SIZE;
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08003089 ++pfn_offset;
David Gibson63551ae2005-06-21 17:14:44 -07003090 --remainder;
3091 ++i;
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08003092 if (vaddr < vma->vm_end && remainder &&
Andi Kleena5516432008-07-23 21:27:41 -07003093 pfn_offset < pages_per_huge_page(h)) {
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08003094 /*
3095 * We use pfn_offset to avoid touching the pageframes
3096 * of this compound page.
3097 */
3098 goto same_page;
3099 }
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003100 spin_unlock(ptl);
David Gibson63551ae2005-06-21 17:14:44 -07003101 }
Michel Lespinasse28a35712013-02-22 16:35:55 -08003102 *nr_pages = remainder;
David Gibson63551ae2005-06-21 17:14:44 -07003103 *position = vaddr;
3104
Hugh Dickins2a15efc2009-09-21 17:03:27 -07003105 return i ? i : -EFAULT;
David Gibson63551ae2005-06-21 17:14:44 -07003106}
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003107
Peter Zijlstra7da4d642012-11-19 03:14:23 +01003108unsigned long hugetlb_change_protection(struct vm_area_struct *vma,
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003109 unsigned long address, unsigned long end, pgprot_t newprot)
3110{
3111 struct mm_struct *mm = vma->vm_mm;
3112 unsigned long start = address;
3113 pte_t *ptep;
3114 pte_t pte;
Andi Kleena5516432008-07-23 21:27:41 -07003115 struct hstate *h = hstate_vma(vma);
Peter Zijlstra7da4d642012-11-19 03:14:23 +01003116 unsigned long pages = 0;
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003117
3118 BUG_ON(address >= end);
3119 flush_cache_range(vma, address, end);
3120
Peter Zijlstra3d48ae42011-05-24 17:12:06 -07003121 mutex_lock(&vma->vm_file->f_mapping->i_mmap_mutex);
Andi Kleena5516432008-07-23 21:27:41 -07003122 for (; address < end; address += huge_page_size(h)) {
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003123 spinlock_t *ptl;
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003124 ptep = huge_pte_offset(mm, address);
3125 if (!ptep)
3126 continue;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003127 ptl = huge_pte_lock(h, mm, ptep);
Peter Zijlstra7da4d642012-11-19 03:14:23 +01003128 if (huge_pmd_unshare(mm, &address, ptep)) {
3129 pages++;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003130 spin_unlock(ptl);
Chen, Kenneth W39dde652006-12-06 20:32:03 -08003131 continue;
Peter Zijlstra7da4d642012-11-19 03:14:23 +01003132 }
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07003133 if (!huge_pte_none(huge_ptep_get(ptep))) {
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003134 pte = huge_ptep_get_and_clear(mm, address, ptep);
Gerald Schaefer106c9922013-04-29 15:07:23 -07003135 pte = pte_mkhuge(huge_pte_modify(pte, newprot));
Tony Lube7517d2013-02-04 14:28:46 -08003136 pte = arch_make_huge_pte(pte, vma, NULL, 0);
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003137 set_huge_pte_at(mm, address, ptep, pte);
Peter Zijlstra7da4d642012-11-19 03:14:23 +01003138 pages++;
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003139 }
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003140 spin_unlock(ptl);
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003141 }
Mel Gormand8333522012-07-31 16:46:20 -07003142 /*
3143 * Must flush TLB before releasing i_mmap_mutex: x86's huge_pmd_unshare
3144 * may have cleared our pud entry and done put_page on the page table:
3145 * once we release i_mmap_mutex, another task can do the final put_page
3146 * and that page table be reused and filled with junk.
3147 */
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003148 flush_tlb_range(vma, start, end);
Mel Gormand8333522012-07-31 16:46:20 -07003149 mutex_unlock(&vma->vm_file->f_mapping->i_mmap_mutex);
Peter Zijlstra7da4d642012-11-19 03:14:23 +01003150
3151 return pages << h->order;
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003152}
3153
Mel Gormana1e78772008-07-23 21:27:23 -07003154int hugetlb_reserve_pages(struct inode *inode,
3155 long from, long to,
Mel Gorman5a6fe122009-02-10 14:02:27 +00003156 struct vm_area_struct *vma,
KOSAKI Motohiroca16d142011-05-26 19:16:19 +09003157 vm_flags_t vm_flags)
Adam Litkee4e574b2007-10-16 01:26:19 -07003158{
Mel Gorman17c9d122009-02-11 16:34:16 +00003159 long ret, chg;
Andi Kleena5516432008-07-23 21:27:41 -07003160 struct hstate *h = hstate_inode(inode);
David Gibson90481622012-03-21 16:34:12 -07003161 struct hugepage_subpool *spool = subpool_inode(inode);
Adam Litkee4e574b2007-10-16 01:26:19 -07003162
Mel Gormana1e78772008-07-23 21:27:23 -07003163 /*
Mel Gorman17c9d122009-02-11 16:34:16 +00003164 * Only apply hugepage reservation if asked. At fault time, an
3165 * attempt will be made for VM_NORESERVE to allocate a page
David Gibson90481622012-03-21 16:34:12 -07003166 * without using reserves
Mel Gorman17c9d122009-02-11 16:34:16 +00003167 */
KOSAKI Motohiroca16d142011-05-26 19:16:19 +09003168 if (vm_flags & VM_NORESERVE)
Mel Gorman17c9d122009-02-11 16:34:16 +00003169 return 0;
3170
3171 /*
Mel Gormana1e78772008-07-23 21:27:23 -07003172 * Shared mappings base their reservation on the number of pages that
3173 * are already allocated on behalf of the file. Private mappings need
3174 * to reserve the full area even if read-only as mprotect() may be
3175 * called to make the mapping read-write. Assume !vma is a shm mapping
3176 */
Mel Gormanf83a2752009-05-28 14:34:40 -07003177 if (!vma || vma->vm_flags & VM_MAYSHARE)
Mel Gormana1e78772008-07-23 21:27:23 -07003178 chg = region_chg(&inode->i_mapping->private_list, from, to);
Mel Gorman5a6fe122009-02-10 14:02:27 +00003179 else {
3180 struct resv_map *resv_map = resv_map_alloc();
Mel Gorman5a6fe122009-02-10 14:02:27 +00003181 if (!resv_map)
3182 return -ENOMEM;
3183
Mel Gorman17c9d122009-02-11 16:34:16 +00003184 chg = to - from;
3185
Mel Gorman5a6fe122009-02-10 14:02:27 +00003186 set_vma_resv_map(vma, resv_map);
3187 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
3188 }
3189
Dave Hansenc50ac052012-05-29 15:06:46 -07003190 if (chg < 0) {
3191 ret = chg;
3192 goto out_err;
3193 }
Mel Gorman17c9d122009-02-11 16:34:16 +00003194
David Gibson90481622012-03-21 16:34:12 -07003195 /* There must be enough pages in the subpool for the mapping */
Dave Hansenc50ac052012-05-29 15:06:46 -07003196 if (hugepage_subpool_get_pages(spool, chg)) {
3197 ret = -ENOSPC;
3198 goto out_err;
3199 }
Mel Gorman17c9d122009-02-11 16:34:16 +00003200
3201 /*
3202 * Check enough hugepages are available for the reservation.
David Gibson90481622012-03-21 16:34:12 -07003203 * Hand the pages back to the subpool if there are not
Mel Gorman17c9d122009-02-11 16:34:16 +00003204 */
3205 ret = hugetlb_acct_memory(h, chg);
3206 if (ret < 0) {
David Gibson90481622012-03-21 16:34:12 -07003207 hugepage_subpool_put_pages(spool, chg);
Dave Hansenc50ac052012-05-29 15:06:46 -07003208 goto out_err;
Mel Gorman17c9d122009-02-11 16:34:16 +00003209 }
3210
3211 /*
3212 * Account for the reservations made. Shared mappings record regions
3213 * that have reservations as they are shared by multiple VMAs.
3214 * When the last VMA disappears, the region map says how much
3215 * the reservation was and the page cache tells how much of
3216 * the reservation was consumed. Private mappings are per-VMA and
3217 * only the consumed reservations are tracked. When the VMA
3218 * disappears, the original reservation is the VMA size and the
3219 * consumed reservations are stored in the map. Hence, nothing
3220 * else has to be done for private mappings here
3221 */
Mel Gormanf83a2752009-05-28 14:34:40 -07003222 if (!vma || vma->vm_flags & VM_MAYSHARE)
Mel Gorman17c9d122009-02-11 16:34:16 +00003223 region_add(&inode->i_mapping->private_list, from, to);
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -07003224 return 0;
Dave Hansenc50ac052012-05-29 15:06:46 -07003225out_err:
Dave Hansen4523e142012-05-30 07:51:07 -07003226 if (vma)
3227 resv_map_put(vma);
Dave Hansenc50ac052012-05-29 15:06:46 -07003228 return ret;
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -07003229}
3230
3231void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed)
3232{
Andi Kleena5516432008-07-23 21:27:41 -07003233 struct hstate *h = hstate_inode(inode);
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -07003234 long chg = region_truncate(&inode->i_mapping->private_list, offset);
David Gibson90481622012-03-21 16:34:12 -07003235 struct hugepage_subpool *spool = subpool_inode(inode);
Ken Chen45c682a2007-11-14 16:59:44 -08003236
3237 spin_lock(&inode->i_lock);
Eric Sandeene4c6f8b2009-07-29 15:02:16 -07003238 inode->i_blocks -= (blocks_per_huge_page(h) * freed);
Ken Chen45c682a2007-11-14 16:59:44 -08003239 spin_unlock(&inode->i_lock);
3240
David Gibson90481622012-03-21 16:34:12 -07003241 hugepage_subpool_put_pages(spool, (chg - freed));
Andi Kleena5516432008-07-23 21:27:41 -07003242 hugetlb_acct_memory(h, -(chg - freed));
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -07003243}
Naoya Horiguchi93f70f92010-05-28 09:29:20 +09003244
Steve Capper3212b532013-04-23 12:35:02 +01003245#ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
3246static unsigned long page_table_shareable(struct vm_area_struct *svma,
3247 struct vm_area_struct *vma,
3248 unsigned long addr, pgoff_t idx)
3249{
3250 unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) +
3251 svma->vm_start;
3252 unsigned long sbase = saddr & PUD_MASK;
3253 unsigned long s_end = sbase + PUD_SIZE;
3254
3255 /* Allow segments to share if only one is marked locked */
3256 unsigned long vm_flags = vma->vm_flags & ~VM_LOCKED;
3257 unsigned long svm_flags = svma->vm_flags & ~VM_LOCKED;
3258
3259 /*
3260 * match the virtual addresses, permission and the alignment of the
3261 * page table page.
3262 */
3263 if (pmd_index(addr) != pmd_index(saddr) ||
3264 vm_flags != svm_flags ||
3265 sbase < svma->vm_start || svma->vm_end < s_end)
3266 return 0;
3267
3268 return saddr;
3269}
3270
3271static int vma_shareable(struct vm_area_struct *vma, unsigned long addr)
3272{
3273 unsigned long base = addr & PUD_MASK;
3274 unsigned long end = base + PUD_SIZE;
3275
3276 /*
3277 * check on proper vm_flags and page table alignment
3278 */
3279 if (vma->vm_flags & VM_MAYSHARE &&
3280 vma->vm_start <= base && end <= vma->vm_end)
3281 return 1;
3282 return 0;
3283}
3284
3285/*
3286 * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
3287 * and returns the corresponding pte. While this is not necessary for the
3288 * !shared pmd case because we can allocate the pmd later as well, it makes the
3289 * code much cleaner. pmd allocation is essential for the shared case because
3290 * pud has to be populated inside the same i_mmap_mutex section - otherwise
3291 * racing tasks could either miss the sharing (see huge_pte_offset) or select a
3292 * bad pmd for sharing.
3293 */
3294pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
3295{
3296 struct vm_area_struct *vma = find_vma(mm, addr);
3297 struct address_space *mapping = vma->vm_file->f_mapping;
3298 pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) +
3299 vma->vm_pgoff;
3300 struct vm_area_struct *svma;
3301 unsigned long saddr;
3302 pte_t *spte = NULL;
3303 pte_t *pte;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003304 spinlock_t *ptl;
Steve Capper3212b532013-04-23 12:35:02 +01003305
3306 if (!vma_shareable(vma, addr))
3307 return (pte_t *)pmd_alloc(mm, pud, addr);
3308
3309 mutex_lock(&mapping->i_mmap_mutex);
3310 vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) {
3311 if (svma == vma)
3312 continue;
3313
3314 saddr = page_table_shareable(svma, vma, addr, idx);
3315 if (saddr) {
3316 spte = huge_pte_offset(svma->vm_mm, saddr);
3317 if (spte) {
3318 get_page(virt_to_page(spte));
3319 break;
3320 }
3321 }
3322 }
3323
3324 if (!spte)
3325 goto out;
3326
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003327 ptl = huge_pte_lockptr(hstate_vma(vma), mm, spte);
3328 spin_lock(ptl);
Steve Capper3212b532013-04-23 12:35:02 +01003329 if (pud_none(*pud))
3330 pud_populate(mm, pud,
3331 (pmd_t *)((unsigned long)spte & PAGE_MASK));
3332 else
3333 put_page(virt_to_page(spte));
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003334 spin_unlock(ptl);
Steve Capper3212b532013-04-23 12:35:02 +01003335out:
3336 pte = (pte_t *)pmd_alloc(mm, pud, addr);
3337 mutex_unlock(&mapping->i_mmap_mutex);
3338 return pte;
3339}
3340
3341/*
3342 * unmap huge page backed by shared pte.
3343 *
3344 * Hugetlb pte page is ref counted at the time of mapping. If pte is shared
3345 * indicated by page_count > 1, unmap is achieved by clearing pud and
3346 * decrementing the ref count. If count == 1, the pte page is not shared.
3347 *
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003348 * called with page table lock held.
Steve Capper3212b532013-04-23 12:35:02 +01003349 *
3350 * returns: 1 successfully unmapped a shared pte page
3351 * 0 the underlying pte page is not shared, or it is the last user
3352 */
3353int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
3354{
3355 pgd_t *pgd = pgd_offset(mm, *addr);
3356 pud_t *pud = pud_offset(pgd, *addr);
3357
3358 BUG_ON(page_count(virt_to_page(ptep)) == 0);
3359 if (page_count(virt_to_page(ptep)) == 1)
3360 return 0;
3361
3362 pud_clear(pud);
3363 put_page(virt_to_page(ptep));
3364 *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE;
3365 return 1;
3366}
Steve Capper9e5fc742013-04-30 08:02:03 +01003367#define want_pmd_share() (1)
3368#else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
3369pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
3370{
3371 return NULL;
3372}
3373#define want_pmd_share() (0)
Steve Capper3212b532013-04-23 12:35:02 +01003374#endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
3375
Steve Capper9e5fc742013-04-30 08:02:03 +01003376#ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
3377pte_t *huge_pte_alloc(struct mm_struct *mm,
3378 unsigned long addr, unsigned long sz)
3379{
3380 pgd_t *pgd;
3381 pud_t *pud;
3382 pte_t *pte = NULL;
3383
3384 pgd = pgd_offset(mm, addr);
3385 pud = pud_alloc(mm, pgd, addr);
3386 if (pud) {
3387 if (sz == PUD_SIZE) {
3388 pte = (pte_t *)pud;
3389 } else {
3390 BUG_ON(sz != PMD_SIZE);
3391 if (want_pmd_share() && pud_none(*pud))
3392 pte = huge_pmd_share(mm, addr, pud);
3393 else
3394 pte = (pte_t *)pmd_alloc(mm, pud, addr);
3395 }
3396 }
3397 BUG_ON(pte && !pte_none(*pte) && !pte_huge(*pte));
3398
3399 return pte;
3400}
3401
3402pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr)
3403{
3404 pgd_t *pgd;
3405 pud_t *pud;
3406 pmd_t *pmd = NULL;
3407
3408 pgd = pgd_offset(mm, addr);
3409 if (pgd_present(*pgd)) {
3410 pud = pud_offset(pgd, addr);
3411 if (pud_present(*pud)) {
3412 if (pud_huge(*pud))
3413 return (pte_t *)pud;
3414 pmd = pmd_offset(pud, addr);
3415 }
3416 }
3417 return (pte_t *) pmd;
3418}
3419
3420struct page *
3421follow_huge_pmd(struct mm_struct *mm, unsigned long address,
3422 pmd_t *pmd, int write)
3423{
3424 struct page *page;
3425
3426 page = pte_page(*(pte_t *)pmd);
3427 if (page)
3428 page += ((address & ~PMD_MASK) >> PAGE_SHIFT);
3429 return page;
3430}
3431
3432struct page *
3433follow_huge_pud(struct mm_struct *mm, unsigned long address,
3434 pud_t *pud, int write)
3435{
3436 struct page *page;
3437
3438 page = pte_page(*(pte_t *)pud);
3439 if (page)
3440 page += ((address & ~PUD_MASK) >> PAGE_SHIFT);
3441 return page;
3442}
3443
3444#else /* !CONFIG_ARCH_WANT_GENERAL_HUGETLB */
3445
3446/* Can be overriden by architectures */
3447__attribute__((weak)) struct page *
3448follow_huge_pud(struct mm_struct *mm, unsigned long address,
3449 pud_t *pud, int write)
3450{
3451 BUG();
3452 return NULL;
3453}
3454
3455#endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
3456
Andi Kleend5bd9102010-09-27 09:00:12 +02003457#ifdef CONFIG_MEMORY_FAILURE
3458
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +09003459/* Should be called in hugetlb_lock */
3460static int is_hugepage_on_freelist(struct page *hpage)
3461{
3462 struct page *page;
3463 struct page *tmp;
3464 struct hstate *h = page_hstate(hpage);
3465 int nid = page_to_nid(hpage);
3466
3467 list_for_each_entry_safe(page, tmp, &h->hugepage_freelists[nid], lru)
3468 if (page == hpage)
3469 return 1;
3470 return 0;
3471}
3472
Naoya Horiguchi93f70f92010-05-28 09:29:20 +09003473/*
3474 * This function is called from memory failure code.
3475 * Assume the caller holds page lock of the head page.
3476 */
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +09003477int dequeue_hwpoisoned_huge_page(struct page *hpage)
Naoya Horiguchi93f70f92010-05-28 09:29:20 +09003478{
3479 struct hstate *h = page_hstate(hpage);
3480 int nid = page_to_nid(hpage);
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +09003481 int ret = -EBUSY;
Naoya Horiguchi93f70f92010-05-28 09:29:20 +09003482
3483 spin_lock(&hugetlb_lock);
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +09003484 if (is_hugepage_on_freelist(hpage)) {
Naoya Horiguchi56f2fb12012-12-12 13:52:33 -08003485 /*
3486 * Hwpoisoned hugepage isn't linked to activelist or freelist,
3487 * but dangling hpage->lru can trigger list-debug warnings
3488 * (this happens when we call unpoison_memory() on it),
3489 * so let it point to itself with list_del_init().
3490 */
3491 list_del_init(&hpage->lru);
Naoya Horiguchi8c6c2ec2010-09-08 10:19:38 +09003492 set_page_refcounted(hpage);
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +09003493 h->free_huge_pages--;
3494 h->free_huge_pages_node[nid]--;
3495 ret = 0;
3496 }
Naoya Horiguchi93f70f92010-05-28 09:29:20 +09003497 spin_unlock(&hugetlb_lock);
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +09003498 return ret;
Naoya Horiguchi93f70f92010-05-28 09:29:20 +09003499}
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +09003500#endif
Naoya Horiguchi31caf662013-09-11 14:21:59 -07003501
3502bool isolate_huge_page(struct page *page, struct list_head *list)
3503{
3504 VM_BUG_ON(!PageHead(page));
3505 if (!get_page_unless_zero(page))
3506 return false;
3507 spin_lock(&hugetlb_lock);
3508 list_move_tail(&page->lru, list);
3509 spin_unlock(&hugetlb_lock);
3510 return true;
3511}
3512
3513void putback_active_hugepage(struct page *page)
3514{
3515 VM_BUG_ON(!PageHead(page));
3516 spin_lock(&hugetlb_lock);
3517 list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist);
3518 spin_unlock(&hugetlb_lock);
3519 put_page(page);
3520}
Naoya Horiguchic8721bb2013-09-11 14:22:09 -07003521
3522bool is_hugepage_active(struct page *page)
3523{
3524 VM_BUG_ON(!PageHuge(page));
3525 /*
3526 * This function can be called for a tail page because the caller,
3527 * scan_movable_pages, scans through a given pfn-range which typically
3528 * covers one memory block. In systems using gigantic hugepage (1GB
3529 * for x86_64,) a hugepage is larger than a memory block, and we don't
3530 * support migrating such large hugepages for now, so return false
3531 * when called for tail pages.
3532 */
3533 if (PageTail(page))
3534 return false;
3535 /*
3536 * Refcount of a hwpoisoned hugepages is 1, but they are not active,
3537 * so we should return false for them.
3538 */
3539 if (unlikely(PageHWPoison(page)))
3540 return false;
3541 return page_count(page) > 0;
3542}