blob: ea8c3a4cd2ae8acdf52a7a4e862e277f2390c265 [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
2 * Generic hugetlb support.
3 * (C) William Irwin, April 2004
4 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07005#include <linux/list.h>
6#include <linux/init.h>
7#include <linux/module.h>
8#include <linux/mm.h>
Alexey Dobriyane1759c22008-10-15 23:50:22 +04009#include <linux/seq_file.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070010#include <linux/sysctl.h>
11#include <linux/highmem.h>
Andrea Arcangelicddb8a52008-07-28 15:46:29 -070012#include <linux/mmu_notifier.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070013#include <linux/nodemask.h>
David Gibson63551ae2005-06-21 17:14:44 -070014#include <linux/pagemap.h>
Christoph Lameter5da7ca82006-01-06 00:10:46 -080015#include <linux/mempolicy.h>
Christoph Lameteraea47ff2006-01-08 01:00:57 -080016#include <linux/cpuset.h>
David Gibson3935baa2006-03-22 00:08:53 -080017#include <linux/mutex.h>
Andi Kleenaa888a72008-07-23 21:27:47 -070018#include <linux/bootmem.h>
Nishanth Aravamudana3437872008-07-23 21:27:44 -070019#include <linux/sysfs.h>
Tejun Heo5a0e3ad2010-03-24 17:04:11 +090020#include <linux/slab.h>
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +090021#include <linux/rmap.h>
Naoya Horiguchifd6a03e2010-05-28 09:29:21 +090022#include <linux/swap.h>
23#include <linux/swapops.h>
Linus Torvaldsd6606682008-08-06 12:04:54 -070024
David Gibson63551ae2005-06-21 17:14:44 -070025#include <asm/page.h>
26#include <asm/pgtable.h>
Chris Forbes32f84522011-07-25 17:12:14 -070027#include <linux/io.h>
David Gibson63551ae2005-06-21 17:14:44 -070028
29#include <linux/hugetlb.h>
Lee Schermerhorn9a3052302009-12-14 17:58:25 -080030#include <linux/node.h>
Nick Piggin7835e982006-03-22 00:08:40 -080031#include "internal.h"
Linus Torvalds1da177e2005-04-16 15:20:36 -070032
33const unsigned long hugetlb_zero = 0, hugetlb_infinity = ~0UL;
Mel Gorman396faf02007-07-17 04:03:13 -070034static gfp_t htlb_alloc_mask = GFP_HIGHUSER;
35unsigned long hugepages_treat_as_movable;
Andi Kleena5516432008-07-23 21:27:41 -070036
Andi Kleene5ff2152008-07-23 21:27:42 -070037static int max_hstate;
38unsigned int default_hstate_idx;
39struct hstate hstates[HUGE_MAX_HSTATE];
40
Jon Tollefson53ba51d2008-07-23 21:27:52 -070041__initdata LIST_HEAD(huge_boot_pages);
42
Andi Kleene5ff2152008-07-23 21:27:42 -070043/* for command line parsing */
44static struct hstate * __initdata parsed_hstate;
45static unsigned long __initdata default_hstate_max_huge_pages;
Nick Piggine11bfbf2008-07-23 21:27:52 -070046static unsigned long __initdata default_hstate_size;
Andi Kleene5ff2152008-07-23 21:27:42 -070047
48#define for_each_hstate(h) \
49 for ((h) = hstates; (h) < &hstates[max_hstate]; (h)++)
Mel Gorman396faf02007-07-17 04:03:13 -070050
David Gibson3935baa2006-03-22 00:08:53 -080051/*
52 * Protects updates to hugepage_freelists, nr_huge_pages, and free_huge_pages
53 */
54static DEFINE_SPINLOCK(hugetlb_lock);
Eric Paris0bd0f9f2005-11-21 21:32:28 -080055
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -070056/*
Andy Whitcroft96822902008-07-23 21:27:29 -070057 * Region tracking -- allows tracking of reservations and instantiated pages
58 * across the pages in a mapping.
Andy Whitcroft84afd992008-07-23 21:27:32 -070059 *
60 * The region data structures are protected by a combination of the mmap_sem
61 * and the hugetlb_instantion_mutex. To access or modify a region the caller
62 * must either hold the mmap_sem for write, or the mmap_sem for read and
63 * the hugetlb_instantiation mutex:
64 *
Chris Forbes32f84522011-07-25 17:12:14 -070065 * down_write(&mm->mmap_sem);
Andy Whitcroft84afd992008-07-23 21:27:32 -070066 * or
Chris Forbes32f84522011-07-25 17:12:14 -070067 * down_read(&mm->mmap_sem);
68 * mutex_lock(&hugetlb_instantiation_mutex);
Andy Whitcroft96822902008-07-23 21:27:29 -070069 */
70struct file_region {
71 struct list_head link;
72 long from;
73 long to;
74};
75
76static long region_add(struct list_head *head, long f, long t)
77{
78 struct file_region *rg, *nrg, *trg;
79
80 /* Locate the region we are either in or before. */
81 list_for_each_entry(rg, head, link)
82 if (f <= rg->to)
83 break;
84
85 /* Round our left edge to the current segment if it encloses us. */
86 if (f > rg->from)
87 f = rg->from;
88
89 /* Check for and consume any regions we now overlap with. */
90 nrg = rg;
91 list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
92 if (&rg->link == head)
93 break;
94 if (rg->from > t)
95 break;
96
97 /* If this area reaches higher then extend our area to
98 * include it completely. If this is not the first area
99 * which we intend to reuse, free it. */
100 if (rg->to > t)
101 t = rg->to;
102 if (rg != nrg) {
103 list_del(&rg->link);
104 kfree(rg);
105 }
106 }
107 nrg->from = f;
108 nrg->to = t;
109 return 0;
110}
111
112static long region_chg(struct list_head *head, long f, long t)
113{
114 struct file_region *rg, *nrg;
115 long chg = 0;
116
117 /* Locate the region we are before or in. */
118 list_for_each_entry(rg, head, link)
119 if (f <= rg->to)
120 break;
121
122 /* If we are below the current region then a new region is required.
123 * Subtle, allocate a new region at the position but make it zero
124 * size such that we can guarantee to record the reservation. */
125 if (&rg->link == head || t < rg->from) {
126 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
127 if (!nrg)
128 return -ENOMEM;
129 nrg->from = f;
130 nrg->to = f;
131 INIT_LIST_HEAD(&nrg->link);
132 list_add(&nrg->link, rg->link.prev);
133
134 return t - f;
135 }
136
137 /* Round our left edge to the current segment if it encloses us. */
138 if (f > rg->from)
139 f = rg->from;
140 chg = t - f;
141
142 /* Check for and consume any regions we now overlap with. */
143 list_for_each_entry(rg, rg->link.prev, link) {
144 if (&rg->link == head)
145 break;
146 if (rg->from > t)
147 return chg;
148
Lucas De Marchi25985ed2011-03-30 22:57:33 -0300149 /* We overlap with this area, if it extends further than
Andy Whitcroft96822902008-07-23 21:27:29 -0700150 * us then we must extend ourselves. Account for its
151 * existing reservation. */
152 if (rg->to > t) {
153 chg += rg->to - t;
154 t = rg->to;
155 }
156 chg -= rg->to - rg->from;
157 }
158 return chg;
159}
160
161static long region_truncate(struct list_head *head, long end)
162{
163 struct file_region *rg, *trg;
164 long chg = 0;
165
166 /* Locate the region we are either in or before. */
167 list_for_each_entry(rg, head, link)
168 if (end <= rg->to)
169 break;
170 if (&rg->link == head)
171 return 0;
172
173 /* If we are in the middle of a region then adjust it. */
174 if (end > rg->from) {
175 chg = rg->to - end;
176 rg->to = end;
177 rg = list_entry(rg->link.next, typeof(*rg), link);
178 }
179
180 /* Drop any remaining regions. */
181 list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
182 if (&rg->link == head)
183 break;
184 chg += rg->to - rg->from;
185 list_del(&rg->link);
186 kfree(rg);
187 }
188 return chg;
189}
190
Andy Whitcroft84afd992008-07-23 21:27:32 -0700191static long region_count(struct list_head *head, long f, long t)
192{
193 struct file_region *rg;
194 long chg = 0;
195
196 /* Locate each segment we overlap with, and count that overlap. */
197 list_for_each_entry(rg, head, link) {
198 int seg_from;
199 int seg_to;
200
201 if (rg->to <= f)
202 continue;
203 if (rg->from >= t)
204 break;
205
206 seg_from = max(rg->from, f);
207 seg_to = min(rg->to, t);
208
209 chg += seg_to - seg_from;
210 }
211
212 return chg;
213}
214
Andy Whitcroft96822902008-07-23 21:27:29 -0700215/*
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700216 * Convert the address within this vma to the page offset within
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700217 * the mapping, in pagecache page units; huge pages here.
218 */
Andi Kleena5516432008-07-23 21:27:41 -0700219static pgoff_t vma_hugecache_offset(struct hstate *h,
220 struct vm_area_struct *vma, unsigned long address)
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700221{
Andi Kleena5516432008-07-23 21:27:41 -0700222 return ((address - vma->vm_start) >> huge_page_shift(h)) +
223 (vma->vm_pgoff >> huge_page_order(h));
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700224}
225
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +0900226pgoff_t linear_hugepage_index(struct vm_area_struct *vma,
227 unsigned long address)
228{
229 return vma_hugecache_offset(hstate_vma(vma), vma, address);
230}
231
Andy Whitcroft84afd992008-07-23 21:27:32 -0700232/*
Mel Gorman08fba692009-01-06 14:38:53 -0800233 * Return the size of the pages allocated when backing a VMA. In the majority
234 * cases this will be same size as used by the page table entries.
235 */
236unsigned long vma_kernel_pagesize(struct vm_area_struct *vma)
237{
238 struct hstate *hstate;
239
240 if (!is_vm_hugetlb_page(vma))
241 return PAGE_SIZE;
242
243 hstate = hstate_vma(vma);
244
245 return 1UL << (hstate->order + PAGE_SHIFT);
246}
Joerg Roedelf340ca02009-06-19 15:16:22 +0200247EXPORT_SYMBOL_GPL(vma_kernel_pagesize);
Mel Gorman08fba692009-01-06 14:38:53 -0800248
249/*
Mel Gorman33402892009-01-06 14:38:54 -0800250 * Return the page size being used by the MMU to back a VMA. In the majority
251 * of cases, the page size used by the kernel matches the MMU size. On
252 * architectures where it differs, an architecture-specific version of this
253 * function is required.
254 */
255#ifndef vma_mmu_pagesize
256unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
257{
258 return vma_kernel_pagesize(vma);
259}
260#endif
261
262/*
Andy Whitcroft84afd992008-07-23 21:27:32 -0700263 * Flags for MAP_PRIVATE reservations. These are stored in the bottom
264 * bits of the reservation map pointer, which are always clear due to
265 * alignment.
266 */
267#define HPAGE_RESV_OWNER (1UL << 0)
268#define HPAGE_RESV_UNMAPPED (1UL << 1)
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700269#define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
Andy Whitcroft84afd992008-07-23 21:27:32 -0700270
Mel Gormana1e78772008-07-23 21:27:23 -0700271/*
272 * These helpers are used to track how many pages are reserved for
273 * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
274 * is guaranteed to have their future faults succeed.
275 *
276 * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
277 * the reserve counters are updated with the hugetlb_lock held. It is safe
278 * to reset the VMA at fork() time as it is not in use yet and there is no
279 * chance of the global counters getting corrupted as a result of the values.
Andy Whitcroft84afd992008-07-23 21:27:32 -0700280 *
281 * The private mapping reservation is represented in a subtly different
282 * manner to a shared mapping. A shared mapping has a region map associated
283 * with the underlying file, this region map represents the backing file
284 * pages which have ever had a reservation assigned which this persists even
285 * after the page is instantiated. A private mapping has a region map
286 * associated with the original mmap which is attached to all VMAs which
287 * reference it, this region map represents those offsets which have consumed
288 * reservation ie. where pages have been instantiated.
Mel Gormana1e78772008-07-23 21:27:23 -0700289 */
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700290static unsigned long get_vma_private_data(struct vm_area_struct *vma)
291{
292 return (unsigned long)vma->vm_private_data;
293}
294
295static void set_vma_private_data(struct vm_area_struct *vma,
296 unsigned long value)
297{
298 vma->vm_private_data = (void *)value;
299}
300
Andy Whitcroft84afd992008-07-23 21:27:32 -0700301struct resv_map {
302 struct kref refs;
303 struct list_head regions;
304};
305
Harvey Harrison2a4b3de2008-10-18 20:27:06 -0700306static struct resv_map *resv_map_alloc(void)
Andy Whitcroft84afd992008-07-23 21:27:32 -0700307{
308 struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
309 if (!resv_map)
310 return NULL;
311
312 kref_init(&resv_map->refs);
313 INIT_LIST_HEAD(&resv_map->regions);
314
315 return resv_map;
316}
317
Harvey Harrison2a4b3de2008-10-18 20:27:06 -0700318static void resv_map_release(struct kref *ref)
Andy Whitcroft84afd992008-07-23 21:27:32 -0700319{
320 struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
321
322 /* Clear out any active regions before we release the map. */
323 region_truncate(&resv_map->regions, 0);
324 kfree(resv_map);
325}
326
327static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
Mel Gormana1e78772008-07-23 21:27:23 -0700328{
329 VM_BUG_ON(!is_vm_hugetlb_page(vma));
Mel Gormanf83a2752009-05-28 14:34:40 -0700330 if (!(vma->vm_flags & VM_MAYSHARE))
Andy Whitcroft84afd992008-07-23 21:27:32 -0700331 return (struct resv_map *)(get_vma_private_data(vma) &
332 ~HPAGE_RESV_MASK);
Harvey Harrison2a4b3de2008-10-18 20:27:06 -0700333 return NULL;
Mel Gormana1e78772008-07-23 21:27:23 -0700334}
335
Andy Whitcroft84afd992008-07-23 21:27:32 -0700336static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
Mel Gormana1e78772008-07-23 21:27:23 -0700337{
338 VM_BUG_ON(!is_vm_hugetlb_page(vma));
Mel Gormanf83a2752009-05-28 14:34:40 -0700339 VM_BUG_ON(vma->vm_flags & VM_MAYSHARE);
Mel Gormana1e78772008-07-23 21:27:23 -0700340
Andy Whitcroft84afd992008-07-23 21:27:32 -0700341 set_vma_private_data(vma, (get_vma_private_data(vma) &
342 HPAGE_RESV_MASK) | (unsigned long)map);
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700343}
344
345static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
346{
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700347 VM_BUG_ON(!is_vm_hugetlb_page(vma));
Mel Gormanf83a2752009-05-28 14:34:40 -0700348 VM_BUG_ON(vma->vm_flags & VM_MAYSHARE);
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700349
350 set_vma_private_data(vma, get_vma_private_data(vma) | flags);
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700351}
352
353static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
354{
355 VM_BUG_ON(!is_vm_hugetlb_page(vma));
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700356
357 return (get_vma_private_data(vma) & flag) != 0;
Mel Gormana1e78772008-07-23 21:27:23 -0700358}
359
360/* Decrement the reserved pages in the hugepage pool by one */
Andi Kleena5516432008-07-23 21:27:41 -0700361static void decrement_hugepage_resv_vma(struct hstate *h,
362 struct vm_area_struct *vma)
Mel Gormana1e78772008-07-23 21:27:23 -0700363{
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -0700364 if (vma->vm_flags & VM_NORESERVE)
365 return;
366
Mel Gormanf83a2752009-05-28 14:34:40 -0700367 if (vma->vm_flags & VM_MAYSHARE) {
Mel Gormana1e78772008-07-23 21:27:23 -0700368 /* Shared mappings always use reserves */
Andi Kleena5516432008-07-23 21:27:41 -0700369 h->resv_huge_pages--;
Andy Whitcroft84afd992008-07-23 21:27:32 -0700370 } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
Mel Gormana1e78772008-07-23 21:27:23 -0700371 /*
372 * Only the process that called mmap() has reserves for
373 * private mappings.
374 */
Andi Kleena5516432008-07-23 21:27:41 -0700375 h->resv_huge_pages--;
Mel Gormana1e78772008-07-23 21:27:23 -0700376 }
377}
378
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700379/* Reset counters to 0 and clear all HPAGE_RESV_* flags */
Mel Gormana1e78772008-07-23 21:27:23 -0700380void reset_vma_resv_huge_pages(struct vm_area_struct *vma)
381{
382 VM_BUG_ON(!is_vm_hugetlb_page(vma));
Mel Gormanf83a2752009-05-28 14:34:40 -0700383 if (!(vma->vm_flags & VM_MAYSHARE))
Mel Gormana1e78772008-07-23 21:27:23 -0700384 vma->vm_private_data = (void *)0;
385}
386
387/* Returns true if the VMA has associated reserve pages */
Mel Gorman7f09ca52008-07-23 21:27:58 -0700388static int vma_has_reserves(struct vm_area_struct *vma)
Mel Gormana1e78772008-07-23 21:27:23 -0700389{
Mel Gormanf83a2752009-05-28 14:34:40 -0700390 if (vma->vm_flags & VM_MAYSHARE)
Mel Gorman7f09ca52008-07-23 21:27:58 -0700391 return 1;
392 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER))
393 return 1;
394 return 0;
Mel Gormana1e78772008-07-23 21:27:23 -0700395}
396
Naoya Horiguchi0ebabb42010-09-08 10:19:34 +0900397static void copy_gigantic_page(struct page *dst, struct page *src)
398{
399 int i;
400 struct hstate *h = page_hstate(src);
401 struct page *dst_base = dst;
402 struct page *src_base = src;
403
404 for (i = 0; i < pages_per_huge_page(h); ) {
405 cond_resched();
406 copy_highpage(dst, src);
407
408 i++;
409 dst = mem_map_next(dst, dst_base, i);
410 src = mem_map_next(src, src_base, i);
411 }
412}
413
414void copy_huge_page(struct page *dst, struct page *src)
415{
416 int i;
417 struct hstate *h = page_hstate(src);
418
419 if (unlikely(pages_per_huge_page(h) > MAX_ORDER_NR_PAGES)) {
420 copy_gigantic_page(dst, src);
421 return;
422 }
423
424 might_sleep();
425 for (i = 0; i < pages_per_huge_page(h); i++) {
426 cond_resched();
427 copy_highpage(dst + i, src + i);
428 }
429}
430
Andi Kleena5516432008-07-23 21:27:41 -0700431static void enqueue_huge_page(struct hstate *h, struct page *page)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700432{
433 int nid = page_to_nid(page);
Andi Kleena5516432008-07-23 21:27:41 -0700434 list_add(&page->lru, &h->hugepage_freelists[nid]);
435 h->free_huge_pages++;
436 h->free_huge_pages_node[nid]++;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700437}
438
Naoya Horiguchibf50bab2010-09-08 10:19:33 +0900439static struct page *dequeue_huge_page_node(struct hstate *h, int nid)
440{
441 struct page *page;
442
443 if (list_empty(&h->hugepage_freelists[nid]))
444 return NULL;
445 page = list_entry(h->hugepage_freelists[nid].next, struct page, lru);
446 list_del(&page->lru);
Naoya Horiguchia9869b82010-09-08 10:19:37 +0900447 set_page_refcounted(page);
Naoya Horiguchibf50bab2010-09-08 10:19:33 +0900448 h->free_huge_pages--;
449 h->free_huge_pages_node[nid]--;
450 return page;
451}
452
Andi Kleena5516432008-07-23 21:27:41 -0700453static struct page *dequeue_huge_page_vma(struct hstate *h,
454 struct vm_area_struct *vma,
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700455 unsigned long address, int avoid_reserve)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700456{
Linus Torvalds1da177e2005-04-16 15:20:36 -0700457 struct page *page = NULL;
Lee Schermerhorn480eccf2007-09-18 22:46:47 -0700458 struct mempolicy *mpol;
Mel Gorman19770b32008-04-28 02:12:18 -0700459 nodemask_t *nodemask;
Miao Xiec0ff7452010-05-24 14:32:08 -0700460 struct zonelist *zonelist;
Mel Gormandd1a2392008-04-28 02:12:17 -0700461 struct zone *zone;
462 struct zoneref *z;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700463
Miao Xiec0ff7452010-05-24 14:32:08 -0700464 get_mems_allowed();
465 zonelist = huge_zonelist(vma, address,
466 htlb_alloc_mask, &mpol, &nodemask);
Mel Gormana1e78772008-07-23 21:27:23 -0700467 /*
468 * A child process with MAP_PRIVATE mappings created by their parent
469 * have no page reserves. This check ensures that reservations are
470 * not "stolen". The child may still get SIGKILLed
471 */
Mel Gorman7f09ca52008-07-23 21:27:58 -0700472 if (!vma_has_reserves(vma) &&
Andi Kleena5516432008-07-23 21:27:41 -0700473 h->free_huge_pages - h->resv_huge_pages == 0)
Miao Xiec0ff7452010-05-24 14:32:08 -0700474 goto err;
Mel Gormana1e78772008-07-23 21:27:23 -0700475
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700476 /* If reserves cannot be used, ensure enough pages are in the pool */
Andi Kleena5516432008-07-23 21:27:41 -0700477 if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
Justin P. Mattock6eab04a2011-04-08 19:49:08 -0700478 goto err;
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700479
Mel Gorman19770b32008-04-28 02:12:18 -0700480 for_each_zone_zonelist_nodemask(zone, z, zonelist,
481 MAX_NR_ZONES - 1, nodemask) {
Naoya Horiguchibf50bab2010-09-08 10:19:33 +0900482 if (cpuset_zone_allowed_softwall(zone, htlb_alloc_mask)) {
483 page = dequeue_huge_page_node(h, zone_to_nid(zone));
484 if (page) {
485 if (!avoid_reserve)
486 decrement_hugepage_resv_vma(h, vma);
487 break;
488 }
Andrew Morton3abf7af2007-07-19 01:49:08 -0700489 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700490 }
Miao Xiec0ff7452010-05-24 14:32:08 -0700491err:
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -0700492 mpol_cond_put(mpol);
Miao Xiec0ff7452010-05-24 14:32:08 -0700493 put_mems_allowed();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700494 return page;
495}
496
Andi Kleena5516432008-07-23 21:27:41 -0700497static void update_and_free_page(struct hstate *h, struct page *page)
Adam Litke6af2acb2007-10-16 01:26:16 -0700498{
499 int i;
Andi Kleena5516432008-07-23 21:27:41 -0700500
Andy Whitcroft18229df2008-11-06 12:53:27 -0800501 VM_BUG_ON(h->order >= MAX_ORDER);
502
Andi Kleena5516432008-07-23 21:27:41 -0700503 h->nr_huge_pages--;
504 h->nr_huge_pages_node[page_to_nid(page)]--;
505 for (i = 0; i < pages_per_huge_page(h); i++) {
Chris Forbes32f84522011-07-25 17:12:14 -0700506 page[i].flags &= ~(1 << PG_locked | 1 << PG_error |
507 1 << PG_referenced | 1 << PG_dirty |
508 1 << PG_active | 1 << PG_reserved |
509 1 << PG_private | 1 << PG_writeback);
Adam Litke6af2acb2007-10-16 01:26:16 -0700510 }
511 set_compound_page_dtor(page, NULL);
512 set_page_refcounted(page);
Gerald Schaefer7f2e9522008-04-28 02:13:29 -0700513 arch_release_hugepage(page);
Andi Kleena5516432008-07-23 21:27:41 -0700514 __free_pages(page, huge_page_order(h));
Adam Litke6af2acb2007-10-16 01:26:16 -0700515}
516
Andi Kleene5ff2152008-07-23 21:27:42 -0700517struct hstate *size_to_hstate(unsigned long size)
518{
519 struct hstate *h;
520
521 for_each_hstate(h) {
522 if (huge_page_size(h) == size)
523 return h;
524 }
525 return NULL;
526}
527
David Gibson27a85ef2006-03-22 00:08:56 -0800528static void free_huge_page(struct page *page)
529{
Andi Kleena5516432008-07-23 21:27:41 -0700530 /*
531 * Can't pass hstate in here because it is called from the
532 * compound page destructor.
533 */
Andi Kleene5ff2152008-07-23 21:27:42 -0700534 struct hstate *h = page_hstate(page);
Adam Litke7893d1d2007-10-16 01:26:18 -0700535 int nid = page_to_nid(page);
Adam Litkec79fb752007-11-14 16:59:38 -0800536 struct address_space *mapping;
David Gibson27a85ef2006-03-22 00:08:56 -0800537
Adam Litkec79fb752007-11-14 16:59:38 -0800538 mapping = (struct address_space *) page_private(page);
Andy Whitcrofte5df70a2008-02-23 15:23:32 -0800539 set_page_private(page, 0);
Mel Gorman23be7462010-04-23 13:17:56 -0400540 page->mapping = NULL;
Adam Litke7893d1d2007-10-16 01:26:18 -0700541 BUG_ON(page_count(page));
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +0900542 BUG_ON(page_mapcount(page));
David Gibson27a85ef2006-03-22 00:08:56 -0800543 INIT_LIST_HEAD(&page->lru);
544
545 spin_lock(&hugetlb_lock);
Andi Kleenaa888a72008-07-23 21:27:47 -0700546 if (h->surplus_huge_pages_node[nid] && huge_page_order(h) < MAX_ORDER) {
Andi Kleena5516432008-07-23 21:27:41 -0700547 update_and_free_page(h, page);
548 h->surplus_huge_pages--;
549 h->surplus_huge_pages_node[nid]--;
Adam Litke7893d1d2007-10-16 01:26:18 -0700550 } else {
Andi Kleena5516432008-07-23 21:27:41 -0700551 enqueue_huge_page(h, page);
Adam Litke7893d1d2007-10-16 01:26:18 -0700552 }
David Gibson27a85ef2006-03-22 00:08:56 -0800553 spin_unlock(&hugetlb_lock);
Adam Litkec79fb752007-11-14 16:59:38 -0800554 if (mapping)
Adam Litke9a119c02007-11-14 16:59:41 -0800555 hugetlb_put_quota(mapping, 1);
David Gibson27a85ef2006-03-22 00:08:56 -0800556}
557
Andi Kleena5516432008-07-23 21:27:41 -0700558static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
Andi Kleenb7ba30c2008-07-23 21:27:40 -0700559{
560 set_compound_page_dtor(page, free_huge_page);
561 spin_lock(&hugetlb_lock);
Andi Kleena5516432008-07-23 21:27:41 -0700562 h->nr_huge_pages++;
563 h->nr_huge_pages_node[nid]++;
Andi Kleenb7ba30c2008-07-23 21:27:40 -0700564 spin_unlock(&hugetlb_lock);
565 put_page(page); /* free it into the hugepage allocator */
566}
567
Wu Fengguang20a03072009-06-16 15:32:22 -0700568static void prep_compound_gigantic_page(struct page *page, unsigned long order)
569{
570 int i;
571 int nr_pages = 1 << order;
572 struct page *p = page + 1;
573
574 /* we rely on prep_new_huge_page to set the destructor */
575 set_compound_order(page, order);
576 __SetPageHead(page);
577 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
578 __SetPageTail(p);
Youquan Song58a84aa2011-12-08 14:34:18 -0800579 set_page_count(p, 0);
Wu Fengguang20a03072009-06-16 15:32:22 -0700580 p->first_page = page;
581 }
582}
583
584int PageHuge(struct page *page)
585{
586 compound_page_dtor *dtor;
587
588 if (!PageCompound(page))
589 return 0;
590
591 page = compound_head(page);
592 dtor = get_compound_page_dtor(page);
593
594 return dtor == free_huge_page;
595}
Naoya Horiguchi43131e12010-05-28 09:29:22 +0900596EXPORT_SYMBOL_GPL(PageHuge);
597
Andi Kleena5516432008-07-23 21:27:41 -0700598static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700599{
Linus Torvalds1da177e2005-04-16 15:20:36 -0700600 struct page *page;
Joe Jinf96efd52007-07-15 23:38:12 -0700601
Andi Kleenaa888a72008-07-23 21:27:47 -0700602 if (h->order >= MAX_ORDER)
603 return NULL;
604
Mel Gorman6484eb32009-06-16 15:31:54 -0700605 page = alloc_pages_exact_node(nid,
Nishanth Aravamudan551883a2008-04-29 00:58:26 -0700606 htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE|
607 __GFP_REPEAT|__GFP_NOWARN,
Andi Kleena5516432008-07-23 21:27:41 -0700608 huge_page_order(h));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700609 if (page) {
Gerald Schaefer7f2e9522008-04-28 02:13:29 -0700610 if (arch_prepare_hugepage(page)) {
Gerald Schaefercaff3a22008-08-12 15:08:38 -0700611 __free_pages(page, huge_page_order(h));
Harvey Harrison7b8ee842008-04-28 14:13:19 -0700612 return NULL;
Gerald Schaefer7f2e9522008-04-28 02:13:29 -0700613 }
Andi Kleena5516432008-07-23 21:27:41 -0700614 prep_new_huge_page(h, page, nid);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700615 }
Nishanth Aravamudan63b46132007-10-16 01:26:24 -0700616
617 return page;
618}
619
Andi Kleen5ced66c2008-07-23 21:27:45 -0700620/*
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800621 * common helper functions for hstate_next_node_to_{alloc|free}.
622 * We may have allocated or freed a huge page based on a different
623 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
624 * be outside of *nodes_allowed. Ensure that we use an allowed
625 * node for alloc or free.
Lee Schermerhorn9a76db02009-12-14 17:58:15 -0800626 */
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800627static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
Lee Schermerhorn9a76db02009-12-14 17:58:15 -0800628{
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800629 nid = next_node(nid, *nodes_allowed);
Lee Schermerhorn9a76db02009-12-14 17:58:15 -0800630 if (nid == MAX_NUMNODES)
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800631 nid = first_node(*nodes_allowed);
Lee Schermerhorn9a76db02009-12-14 17:58:15 -0800632 VM_BUG_ON(nid >= MAX_NUMNODES);
633
634 return nid;
635}
636
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800637static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
Andi Kleen5ced66c2008-07-23 21:27:45 -0700638{
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800639 if (!node_isset(nid, *nodes_allowed))
640 nid = next_node_allowed(nid, nodes_allowed);
Lee Schermerhorn9a76db02009-12-14 17:58:15 -0800641 return nid;
Andi Kleen5ced66c2008-07-23 21:27:45 -0700642}
643
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800644/*
645 * returns the previously saved node ["this node"] from which to
646 * allocate a persistent huge page for the pool and advance the
647 * next node from which to allocate, handling wrap at end of node
648 * mask.
649 */
650static int hstate_next_node_to_alloc(struct hstate *h,
651 nodemask_t *nodes_allowed)
652{
653 int nid;
654
655 VM_BUG_ON(!nodes_allowed);
656
657 nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
658 h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
659
660 return nid;
661}
662
663static int alloc_fresh_huge_page(struct hstate *h, nodemask_t *nodes_allowed)
Nishanth Aravamudan63b46132007-10-16 01:26:24 -0700664{
665 struct page *page;
666 int start_nid;
667 int next_nid;
668 int ret = 0;
669
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800670 start_nid = hstate_next_node_to_alloc(h, nodes_allowed);
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700671 next_nid = start_nid;
Nishanth Aravamudan63b46132007-10-16 01:26:24 -0700672
673 do {
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700674 page = alloc_fresh_huge_page_node(h, next_nid);
Lee Schermerhorn9a76db02009-12-14 17:58:15 -0800675 if (page) {
Nishanth Aravamudan63b46132007-10-16 01:26:24 -0700676 ret = 1;
Lee Schermerhorn9a76db02009-12-14 17:58:15 -0800677 break;
678 }
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800679 next_nid = hstate_next_node_to_alloc(h, nodes_allowed);
Lee Schermerhorn9a76db02009-12-14 17:58:15 -0800680 } while (next_nid != start_nid);
Nishanth Aravamudan63b46132007-10-16 01:26:24 -0700681
Adam Litke3b116302008-04-28 02:13:06 -0700682 if (ret)
683 count_vm_event(HTLB_BUDDY_PGALLOC);
684 else
685 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
686
Nishanth Aravamudan63b46132007-10-16 01:26:24 -0700687 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700688}
689
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700690/*
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800691 * helper for free_pool_huge_page() - return the previously saved
692 * node ["this node"] from which to free a huge page. Advance the
693 * next node id whether or not we find a free huge page to free so
694 * that the next attempt to free addresses the next node.
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700695 */
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800696static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700697{
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800698 int nid;
Lee Schermerhorn9a76db02009-12-14 17:58:15 -0800699
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800700 VM_BUG_ON(!nodes_allowed);
701
702 nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
703 h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
704
Lee Schermerhorn9a76db02009-12-14 17:58:15 -0800705 return nid;
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700706}
707
708/*
709 * Free huge page from pool from next node to free.
710 * Attempt to keep persistent huge pages more or less
711 * balanced over allowed nodes.
712 * Called with hugetlb_lock locked.
713 */
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800714static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
715 bool acct_surplus)
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700716{
717 int start_nid;
718 int next_nid;
719 int ret = 0;
720
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800721 start_nid = hstate_next_node_to_free(h, nodes_allowed);
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700722 next_nid = start_nid;
723
724 do {
Lee Schermerhorn685f3452009-09-21 17:01:23 -0700725 /*
726 * If we're returning unused surplus pages, only examine
727 * nodes with surplus pages.
728 */
729 if ((!acct_surplus || h->surplus_huge_pages_node[next_nid]) &&
730 !list_empty(&h->hugepage_freelists[next_nid])) {
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700731 struct page *page =
732 list_entry(h->hugepage_freelists[next_nid].next,
733 struct page, lru);
734 list_del(&page->lru);
735 h->free_huge_pages--;
736 h->free_huge_pages_node[next_nid]--;
Lee Schermerhorn685f3452009-09-21 17:01:23 -0700737 if (acct_surplus) {
738 h->surplus_huge_pages--;
739 h->surplus_huge_pages_node[next_nid]--;
740 }
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700741 update_and_free_page(h, page);
742 ret = 1;
Lee Schermerhorn9a76db02009-12-14 17:58:15 -0800743 break;
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700744 }
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800745 next_nid = hstate_next_node_to_free(h, nodes_allowed);
Lee Schermerhorn9a76db02009-12-14 17:58:15 -0800746 } while (next_nid != start_nid);
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700747
748 return ret;
749}
750
Naoya Horiguchibf50bab2010-09-08 10:19:33 +0900751static struct page *alloc_buddy_huge_page(struct hstate *h, int nid)
Adam Litke7893d1d2007-10-16 01:26:18 -0700752{
753 struct page *page;
Naoya Horiguchibf50bab2010-09-08 10:19:33 +0900754 unsigned int r_nid;
Adam Litke7893d1d2007-10-16 01:26:18 -0700755
Andi Kleenaa888a72008-07-23 21:27:47 -0700756 if (h->order >= MAX_ORDER)
757 return NULL;
758
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800759 /*
760 * Assume we will successfully allocate the surplus page to
761 * prevent racing processes from causing the surplus to exceed
762 * overcommit
763 *
764 * This however introduces a different race, where a process B
765 * tries to grow the static hugepage pool while alloc_pages() is
766 * called by process A. B will only examine the per-node
767 * counters in determining if surplus huge pages can be
768 * converted to normal huge pages in adjust_pool_surplus(). A
769 * won't be able to increment the per-node counter, until the
770 * lock is dropped by B, but B doesn't drop hugetlb_lock until
771 * no more huge pages can be converted from surplus to normal
772 * state (and doesn't try to convert again). Thus, we have a
773 * case where a surplus huge page exists, the pool is grown, and
774 * the surplus huge page still exists after, even though it
775 * should just have been converted to a normal huge page. This
776 * does not leak memory, though, as the hugepage will be freed
777 * once it is out of use. It also does not allow the counters to
778 * go out of whack in adjust_pool_surplus() as we don't modify
779 * the node values until we've gotten the hugepage and only the
780 * per-node value is checked there.
781 */
782 spin_lock(&hugetlb_lock);
Andi Kleena5516432008-07-23 21:27:41 -0700783 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800784 spin_unlock(&hugetlb_lock);
785 return NULL;
786 } else {
Andi Kleena5516432008-07-23 21:27:41 -0700787 h->nr_huge_pages++;
788 h->surplus_huge_pages++;
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800789 }
790 spin_unlock(&hugetlb_lock);
791
Naoya Horiguchibf50bab2010-09-08 10:19:33 +0900792 if (nid == NUMA_NO_NODE)
793 page = alloc_pages(htlb_alloc_mask|__GFP_COMP|
794 __GFP_REPEAT|__GFP_NOWARN,
795 huge_page_order(h));
796 else
797 page = alloc_pages_exact_node(nid,
798 htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE|
799 __GFP_REPEAT|__GFP_NOWARN, huge_page_order(h));
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800800
Gerald Schaefercaff3a22008-08-12 15:08:38 -0700801 if (page && arch_prepare_hugepage(page)) {
802 __free_pages(page, huge_page_order(h));
Hillf Dantonea5768c2012-01-10 15:08:30 -0800803 page = NULL;
Gerald Schaefercaff3a22008-08-12 15:08:38 -0700804 }
805
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800806 spin_lock(&hugetlb_lock);
Adam Litke7893d1d2007-10-16 01:26:18 -0700807 if (page) {
Naoya Horiguchibf50bab2010-09-08 10:19:33 +0900808 r_nid = page_to_nid(page);
Adam Litke7893d1d2007-10-16 01:26:18 -0700809 set_compound_page_dtor(page, free_huge_page);
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800810 /*
811 * We incremented the global counters already
812 */
Naoya Horiguchibf50bab2010-09-08 10:19:33 +0900813 h->nr_huge_pages_node[r_nid]++;
814 h->surplus_huge_pages_node[r_nid]++;
Adam Litke3b116302008-04-28 02:13:06 -0700815 __count_vm_event(HTLB_BUDDY_PGALLOC);
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800816 } else {
Andi Kleena5516432008-07-23 21:27:41 -0700817 h->nr_huge_pages--;
818 h->surplus_huge_pages--;
Adam Litke3b116302008-04-28 02:13:06 -0700819 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
Adam Litke7893d1d2007-10-16 01:26:18 -0700820 }
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800821 spin_unlock(&hugetlb_lock);
Adam Litke7893d1d2007-10-16 01:26:18 -0700822
823 return page;
824}
825
Adam Litkee4e574b2007-10-16 01:26:19 -0700826/*
Naoya Horiguchibf50bab2010-09-08 10:19:33 +0900827 * This allocation function is useful in the context where vma is irrelevant.
828 * E.g. soft-offlining uses this function because it only cares physical
829 * address of error page.
830 */
831struct page *alloc_huge_page_node(struct hstate *h, int nid)
832{
833 struct page *page;
834
835 spin_lock(&hugetlb_lock);
836 page = dequeue_huge_page_node(h, nid);
837 spin_unlock(&hugetlb_lock);
838
839 if (!page)
840 page = alloc_buddy_huge_page(h, nid);
841
842 return page;
843}
844
845/*
Lucas De Marchi25985ed2011-03-30 22:57:33 -0300846 * Increase the hugetlb pool such that it can accommodate a reservation
Adam Litkee4e574b2007-10-16 01:26:19 -0700847 * of size 'delta'.
848 */
Andi Kleena5516432008-07-23 21:27:41 -0700849static int gather_surplus_pages(struct hstate *h, int delta)
Adam Litkee4e574b2007-10-16 01:26:19 -0700850{
851 struct list_head surplus_list;
852 struct page *page, *tmp;
853 int ret, i;
854 int needed, allocated;
855
Andi Kleena5516432008-07-23 21:27:41 -0700856 needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
Adam Litkeac09b3a2008-03-04 14:29:38 -0800857 if (needed <= 0) {
Andi Kleena5516432008-07-23 21:27:41 -0700858 h->resv_huge_pages += delta;
Adam Litkee4e574b2007-10-16 01:26:19 -0700859 return 0;
Adam Litkeac09b3a2008-03-04 14:29:38 -0800860 }
Adam Litkee4e574b2007-10-16 01:26:19 -0700861
862 allocated = 0;
863 INIT_LIST_HEAD(&surplus_list);
864
865 ret = -ENOMEM;
866retry:
867 spin_unlock(&hugetlb_lock);
868 for (i = 0; i < needed; i++) {
Naoya Horiguchibf50bab2010-09-08 10:19:33 +0900869 page = alloc_buddy_huge_page(h, NUMA_NO_NODE);
Naoya Horiguchia9869b82010-09-08 10:19:37 +0900870 if (!page)
Adam Litkee4e574b2007-10-16 01:26:19 -0700871 /*
872 * We were not able to allocate enough pages to
873 * satisfy the entire reservation so we free what
874 * we've allocated so far.
875 */
Adam Litkee4e574b2007-10-16 01:26:19 -0700876 goto free;
Adam Litkee4e574b2007-10-16 01:26:19 -0700877
878 list_add(&page->lru, &surplus_list);
879 }
880 allocated += needed;
881
882 /*
883 * After retaking hugetlb_lock, we need to recalculate 'needed'
884 * because either resv_huge_pages or free_huge_pages may have changed.
885 */
886 spin_lock(&hugetlb_lock);
Andi Kleena5516432008-07-23 21:27:41 -0700887 needed = (h->resv_huge_pages + delta) -
888 (h->free_huge_pages + allocated);
Adam Litkee4e574b2007-10-16 01:26:19 -0700889 if (needed > 0)
890 goto retry;
891
892 /*
893 * The surplus_list now contains _at_least_ the number of extra pages
Lucas De Marchi25985ed2011-03-30 22:57:33 -0300894 * needed to accommodate the reservation. Add the appropriate number
Adam Litkee4e574b2007-10-16 01:26:19 -0700895 * of pages to the hugetlb pool and free the extras back to the buddy
Adam Litkeac09b3a2008-03-04 14:29:38 -0800896 * allocator. Commit the entire reservation here to prevent another
897 * process from stealing the pages as they are added to the pool but
898 * before they are reserved.
Adam Litkee4e574b2007-10-16 01:26:19 -0700899 */
900 needed += allocated;
Andi Kleena5516432008-07-23 21:27:41 -0700901 h->resv_huge_pages += delta;
Adam Litkee4e574b2007-10-16 01:26:19 -0700902 ret = 0;
Naoya Horiguchia9869b82010-09-08 10:19:37 +0900903
Adam Litke19fc3f02008-04-28 02:12:20 -0700904 /* Free the needed pages to the hugetlb pool */
Adam Litkee4e574b2007-10-16 01:26:19 -0700905 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
Adam Litke19fc3f02008-04-28 02:12:20 -0700906 if ((--needed) < 0)
907 break;
Adam Litkee4e574b2007-10-16 01:26:19 -0700908 list_del(&page->lru);
Naoya Horiguchia9869b82010-09-08 10:19:37 +0900909 /*
910 * This page is now managed by the hugetlb allocator and has
911 * no users -- drop the buddy allocator's reference.
912 */
913 put_page_testzero(page);
914 VM_BUG_ON(page_count(page));
Andi Kleena5516432008-07-23 21:27:41 -0700915 enqueue_huge_page(h, page);
Adam Litke19fc3f02008-04-28 02:12:20 -0700916 }
Hillf Dantonb0365c82011-12-28 15:57:16 -0800917 spin_unlock(&hugetlb_lock);
Adam Litke19fc3f02008-04-28 02:12:20 -0700918
919 /* Free unnecessary surplus pages to the buddy allocator */
Naoya Horiguchia9869b82010-09-08 10:19:37 +0900920free:
Adam Litke19fc3f02008-04-28 02:12:20 -0700921 if (!list_empty(&surplus_list)) {
Adam Litke19fc3f02008-04-28 02:12:20 -0700922 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
923 list_del(&page->lru);
Naoya Horiguchia9869b82010-09-08 10:19:37 +0900924 put_page(page);
Adam Litkeaf767cb2007-10-16 01:26:25 -0700925 }
Adam Litkee4e574b2007-10-16 01:26:19 -0700926 }
Naoya Horiguchia9869b82010-09-08 10:19:37 +0900927 spin_lock(&hugetlb_lock);
Adam Litkee4e574b2007-10-16 01:26:19 -0700928
929 return ret;
930}
931
932/*
933 * When releasing a hugetlb pool reservation, any surplus pages that were
934 * allocated to satisfy the reservation must be explicitly freed if they were
935 * never used.
Lee Schermerhorn685f3452009-09-21 17:01:23 -0700936 * Called with hugetlb_lock held.
Adam Litkee4e574b2007-10-16 01:26:19 -0700937 */
Andi Kleena5516432008-07-23 21:27:41 -0700938static void return_unused_surplus_pages(struct hstate *h,
939 unsigned long unused_resv_pages)
Adam Litkee4e574b2007-10-16 01:26:19 -0700940{
Adam Litkee4e574b2007-10-16 01:26:19 -0700941 unsigned long nr_pages;
942
Adam Litkeac09b3a2008-03-04 14:29:38 -0800943 /* Uncommit the reservation */
Andi Kleena5516432008-07-23 21:27:41 -0700944 h->resv_huge_pages -= unused_resv_pages;
Adam Litkeac09b3a2008-03-04 14:29:38 -0800945
Andi Kleenaa888a72008-07-23 21:27:47 -0700946 /* Cannot return gigantic pages currently */
947 if (h->order >= MAX_ORDER)
948 return;
949
Andi Kleena5516432008-07-23 21:27:41 -0700950 nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
Adam Litkee4e574b2007-10-16 01:26:19 -0700951
Lee Schermerhorn685f3452009-09-21 17:01:23 -0700952 /*
953 * We want to release as many surplus pages as possible, spread
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -0800954 * evenly across all nodes with memory. Iterate across these nodes
955 * until we can no longer free unreserved surplus pages. This occurs
956 * when the nodes with surplus pages have no free pages.
957 * free_pool_huge_page() will balance the the freed pages across the
958 * on-line nodes with memory and will handle the hstate accounting.
Lee Schermerhorn685f3452009-09-21 17:01:23 -0700959 */
960 while (nr_pages--) {
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -0800961 if (!free_pool_huge_page(h, &node_states[N_HIGH_MEMORY], 1))
Lee Schermerhorn685f3452009-09-21 17:01:23 -0700962 break;
Adam Litkee4e574b2007-10-16 01:26:19 -0700963 }
964}
965
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -0700966/*
967 * Determine if the huge page at addr within the vma has an associated
968 * reservation. Where it does not we will need to logically increase
969 * reservation and actually increase quota before an allocation can occur.
970 * Where any new reservation would be required the reservation change is
971 * prepared, but not committed. Once the page has been quota'd allocated
972 * an instantiated the change should be committed via vma_commit_reservation.
973 * No action is required on failure.
974 */
Roel Kluine2f17d92009-03-31 15:23:15 -0700975static long vma_needs_reservation(struct hstate *h,
Andi Kleena5516432008-07-23 21:27:41 -0700976 struct vm_area_struct *vma, unsigned long addr)
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -0700977{
978 struct address_space *mapping = vma->vm_file->f_mapping;
979 struct inode *inode = mapping->host;
980
Mel Gormanf83a2752009-05-28 14:34:40 -0700981 if (vma->vm_flags & VM_MAYSHARE) {
Andi Kleena5516432008-07-23 21:27:41 -0700982 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -0700983 return region_chg(&inode->i_mapping->private_list,
984 idx, idx + 1);
985
Andy Whitcroft84afd992008-07-23 21:27:32 -0700986 } else if (!is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
987 return 1;
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -0700988
Andy Whitcroft84afd992008-07-23 21:27:32 -0700989 } else {
Roel Kluine2f17d92009-03-31 15:23:15 -0700990 long err;
Andi Kleena5516432008-07-23 21:27:41 -0700991 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
Andy Whitcroft84afd992008-07-23 21:27:32 -0700992 struct resv_map *reservations = vma_resv_map(vma);
993
994 err = region_chg(&reservations->regions, idx, idx + 1);
995 if (err < 0)
996 return err;
997 return 0;
998 }
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -0700999}
Andi Kleena5516432008-07-23 21:27:41 -07001000static void vma_commit_reservation(struct hstate *h,
1001 struct vm_area_struct *vma, unsigned long addr)
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001002{
1003 struct address_space *mapping = vma->vm_file->f_mapping;
1004 struct inode *inode = mapping->host;
1005
Mel Gormanf83a2752009-05-28 14:34:40 -07001006 if (vma->vm_flags & VM_MAYSHARE) {
Andi Kleena5516432008-07-23 21:27:41 -07001007 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001008 region_add(&inode->i_mapping->private_list, idx, idx + 1);
Andy Whitcroft84afd992008-07-23 21:27:32 -07001009
1010 } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
Andi Kleena5516432008-07-23 21:27:41 -07001011 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
Andy Whitcroft84afd992008-07-23 21:27:32 -07001012 struct resv_map *reservations = vma_resv_map(vma);
1013
1014 /* Mark this page used in the map. */
1015 region_add(&reservations->regions, idx, idx + 1);
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001016 }
1017}
1018
David Gibson27a85ef2006-03-22 00:08:56 -08001019static struct page *alloc_huge_page(struct vm_area_struct *vma,
Mel Gorman04f2cbe2008-07-23 21:27:25 -07001020 unsigned long addr, int avoid_reserve)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001021{
Andi Kleena5516432008-07-23 21:27:41 -07001022 struct hstate *h = hstate_vma(vma);
Adam Litke348ea202007-11-14 16:59:37 -08001023 struct page *page;
Adam Litke2fc39ce2007-11-14 16:59:39 -08001024 struct address_space *mapping = vma->vm_file->f_mapping;
Mel Gormana1e78772008-07-23 21:27:23 -07001025 struct inode *inode = mapping->host;
Roel Kluine2f17d92009-03-31 15:23:15 -07001026 long chg;
Adam Litke2fc39ce2007-11-14 16:59:39 -08001027
Mel Gormana1e78772008-07-23 21:27:23 -07001028 /*
1029 * Processes that did not create the mapping will have no reserves and
1030 * will not have accounted against quota. Check that the quota can be
1031 * made before satisfying the allocation
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001032 * MAP_NORESERVE mappings may also need pages and quota allocated
1033 * if no reserve mapping overlaps.
Mel Gormana1e78772008-07-23 21:27:23 -07001034 */
Andi Kleena5516432008-07-23 21:27:41 -07001035 chg = vma_needs_reservation(h, vma, addr);
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001036 if (chg < 0)
Hugh Dickinse0dcd8a2011-06-05 22:03:13 -07001037 return ERR_PTR(-VM_FAULT_OOM);
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001038 if (chg)
Mel Gormana1e78772008-07-23 21:27:23 -07001039 if (hugetlb_get_quota(inode->i_mapping, chg))
Hugh Dickinse0dcd8a2011-06-05 22:03:13 -07001040 return ERR_PTR(-VM_FAULT_SIGBUS);
Mel Gormana1e78772008-07-23 21:27:23 -07001041
1042 spin_lock(&hugetlb_lock);
Andi Kleena5516432008-07-23 21:27:41 -07001043 page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve);
Mel Gormana1e78772008-07-23 21:27:23 -07001044 spin_unlock(&hugetlb_lock);
1045
1046 if (!page) {
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001047 page = alloc_buddy_huge_page(h, NUMA_NO_NODE);
Mel Gormana1e78772008-07-23 21:27:23 -07001048 if (!page) {
1049 hugetlb_put_quota(inode->i_mapping, chg);
Mel Gorman4a6018f2010-05-11 14:06:53 -07001050 return ERR_PTR(-VM_FAULT_SIGBUS);
Mel Gormana1e78772008-07-23 21:27:23 -07001051 }
1052 }
1053
Mel Gormana1e78772008-07-23 21:27:23 -07001054 set_page_private(page, (unsigned long) mapping);
1055
Andi Kleena5516432008-07-23 21:27:41 -07001056 vma_commit_reservation(h, vma, addr);
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001057
Adam Litke90d8b7e2007-11-14 16:59:42 -08001058 return page;
David Gibsonb45b5bd2006-03-22 00:08:55 -08001059}
1060
Cyrill Gorcunov91f47662009-01-06 14:40:33 -08001061int __weak alloc_bootmem_huge_page(struct hstate *h)
Andi Kleenaa888a72008-07-23 21:27:47 -07001062{
1063 struct huge_bootmem_page *m;
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -08001064 int nr_nodes = nodes_weight(node_states[N_HIGH_MEMORY]);
Andi Kleenaa888a72008-07-23 21:27:47 -07001065
1066 while (nr_nodes) {
1067 void *addr;
1068
1069 addr = __alloc_bootmem_node_nopanic(
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001070 NODE_DATA(hstate_next_node_to_alloc(h,
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -08001071 &node_states[N_HIGH_MEMORY])),
Andi Kleenaa888a72008-07-23 21:27:47 -07001072 huge_page_size(h), huge_page_size(h), 0);
1073
1074 if (addr) {
1075 /*
1076 * Use the beginning of the huge page to store the
1077 * huge_bootmem_page struct (until gather_bootmem
1078 * puts them into the mem_map).
1079 */
1080 m = addr;
Cyrill Gorcunov91f47662009-01-06 14:40:33 -08001081 goto found;
Andi Kleenaa888a72008-07-23 21:27:47 -07001082 }
Andi Kleenaa888a72008-07-23 21:27:47 -07001083 nr_nodes--;
1084 }
1085 return 0;
1086
1087found:
1088 BUG_ON((unsigned long)virt_to_phys(m) & (huge_page_size(h) - 1));
1089 /* Put them into a private list first because mem_map is not up yet */
1090 list_add(&m->list, &huge_boot_pages);
1091 m->hstate = h;
1092 return 1;
1093}
1094
Andy Whitcroft18229df2008-11-06 12:53:27 -08001095static void prep_compound_huge_page(struct page *page, int order)
1096{
1097 if (unlikely(order > (MAX_ORDER - 1)))
1098 prep_compound_gigantic_page(page, order);
1099 else
1100 prep_compound_page(page, order);
1101}
1102
Andi Kleenaa888a72008-07-23 21:27:47 -07001103/* Put bootmem huge pages into the standard lists after mem_map is up */
1104static void __init gather_bootmem_prealloc(void)
1105{
1106 struct huge_bootmem_page *m;
1107
1108 list_for_each_entry(m, &huge_boot_pages, list) {
Andi Kleenaa888a72008-07-23 21:27:47 -07001109 struct hstate *h = m->hstate;
Becky Bruceee8f2482011-07-25 17:11:50 -07001110 struct page *page;
1111
1112#ifdef CONFIG_HIGHMEM
1113 page = pfn_to_page(m->phys >> PAGE_SHIFT);
1114 free_bootmem_late((unsigned long)m,
1115 sizeof(struct huge_bootmem_page));
1116#else
1117 page = virt_to_page(m);
1118#endif
Andi Kleenaa888a72008-07-23 21:27:47 -07001119 __ClearPageReserved(page);
1120 WARN_ON(page_count(page) != 1);
Andy Whitcroft18229df2008-11-06 12:53:27 -08001121 prep_compound_huge_page(page, h->order);
Andi Kleenaa888a72008-07-23 21:27:47 -07001122 prep_new_huge_page(h, page, page_to_nid(page));
Rafael Aquinib0320c72011-06-15 15:08:39 -07001123 /*
1124 * If we had gigantic hugepages allocated at boot time, we need
1125 * to restore the 'stolen' pages to totalram_pages in order to
1126 * fix confusing memory reports from free(1) and another
1127 * side-effects, like CommitLimit going negative.
1128 */
1129 if (h->order > (MAX_ORDER - 1))
1130 totalram_pages += 1 << h->order;
Andi Kleenaa888a72008-07-23 21:27:47 -07001131 }
1132}
1133
Andi Kleen8faa8b02008-07-23 21:27:48 -07001134static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001135{
1136 unsigned long i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001137
Andi Kleene5ff2152008-07-23 21:27:42 -07001138 for (i = 0; i < h->max_huge_pages; ++i) {
Andi Kleenaa888a72008-07-23 21:27:47 -07001139 if (h->order >= MAX_ORDER) {
1140 if (!alloc_bootmem_huge_page(h))
1141 break;
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -08001142 } else if (!alloc_fresh_huge_page(h,
1143 &node_states[N_HIGH_MEMORY]))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001144 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001145 }
Andi Kleen8faa8b02008-07-23 21:27:48 -07001146 h->max_huge_pages = i;
Andi Kleene5ff2152008-07-23 21:27:42 -07001147}
1148
1149static void __init hugetlb_init_hstates(void)
1150{
1151 struct hstate *h;
1152
1153 for_each_hstate(h) {
Andi Kleen8faa8b02008-07-23 21:27:48 -07001154 /* oversize hugepages were init'ed in early boot */
1155 if (h->order < MAX_ORDER)
1156 hugetlb_hstate_alloc_pages(h);
Andi Kleene5ff2152008-07-23 21:27:42 -07001157 }
1158}
1159
Andi Kleen4abd32d2008-07-23 21:27:49 -07001160static char * __init memfmt(char *buf, unsigned long n)
1161{
1162 if (n >= (1UL << 30))
1163 sprintf(buf, "%lu GB", n >> 30);
1164 else if (n >= (1UL << 20))
1165 sprintf(buf, "%lu MB", n >> 20);
1166 else
1167 sprintf(buf, "%lu KB", n >> 10);
1168 return buf;
1169}
1170
Andi Kleene5ff2152008-07-23 21:27:42 -07001171static void __init report_hugepages(void)
1172{
1173 struct hstate *h;
1174
1175 for_each_hstate(h) {
Andi Kleen4abd32d2008-07-23 21:27:49 -07001176 char buf[32];
1177 printk(KERN_INFO "HugeTLB registered %s page size, "
1178 "pre-allocated %ld pages\n",
1179 memfmt(buf, huge_page_size(h)),
1180 h->free_huge_pages);
Andi Kleene5ff2152008-07-23 21:27:42 -07001181 }
1182}
1183
Linus Torvalds1da177e2005-04-16 15:20:36 -07001184#ifdef CONFIG_HIGHMEM
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001185static void try_to_free_low(struct hstate *h, unsigned long count,
1186 nodemask_t *nodes_allowed)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001187{
Christoph Lameter4415cc82006-09-25 23:31:55 -07001188 int i;
1189
Andi Kleenaa888a72008-07-23 21:27:47 -07001190 if (h->order >= MAX_ORDER)
1191 return;
1192
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001193 for_each_node_mask(i, *nodes_allowed) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001194 struct page *page, *next;
Andi Kleena5516432008-07-23 21:27:41 -07001195 struct list_head *freel = &h->hugepage_freelists[i];
1196 list_for_each_entry_safe(page, next, freel, lru) {
1197 if (count >= h->nr_huge_pages)
Adam Litke6b0c8802007-10-16 01:26:23 -07001198 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001199 if (PageHighMem(page))
1200 continue;
1201 list_del(&page->lru);
Andi Kleene5ff2152008-07-23 21:27:42 -07001202 update_and_free_page(h, page);
Andi Kleena5516432008-07-23 21:27:41 -07001203 h->free_huge_pages--;
1204 h->free_huge_pages_node[page_to_nid(page)]--;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001205 }
1206 }
1207}
1208#else
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001209static inline void try_to_free_low(struct hstate *h, unsigned long count,
1210 nodemask_t *nodes_allowed)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001211{
1212}
1213#endif
1214
Wu Fengguang20a03072009-06-16 15:32:22 -07001215/*
1216 * Increment or decrement surplus_huge_pages. Keep node-specific counters
1217 * balanced by operating on them in a round-robin fashion.
1218 * Returns 1 if an adjustment was made.
1219 */
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001220static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
1221 int delta)
Wu Fengguang20a03072009-06-16 15:32:22 -07001222{
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001223 int start_nid, next_nid;
Wu Fengguang20a03072009-06-16 15:32:22 -07001224 int ret = 0;
1225
1226 VM_BUG_ON(delta != -1 && delta != 1);
Wu Fengguang20a03072009-06-16 15:32:22 -07001227
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001228 if (delta < 0)
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001229 start_nid = hstate_next_node_to_alloc(h, nodes_allowed);
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001230 else
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001231 start_nid = hstate_next_node_to_free(h, nodes_allowed);
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001232 next_nid = start_nid;
1233
1234 do {
1235 int nid = next_nid;
1236 if (delta < 0) {
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001237 /*
1238 * To shrink on this node, there must be a surplus page
1239 */
Lee Schermerhorn9a76db02009-12-14 17:58:15 -08001240 if (!h->surplus_huge_pages_node[nid]) {
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001241 next_nid = hstate_next_node_to_alloc(h,
1242 nodes_allowed);
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001243 continue;
Lee Schermerhorn9a76db02009-12-14 17:58:15 -08001244 }
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001245 }
1246 if (delta > 0) {
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001247 /*
1248 * Surplus cannot exceed the total number of pages
1249 */
1250 if (h->surplus_huge_pages_node[nid] >=
Lee Schermerhorn9a76db02009-12-14 17:58:15 -08001251 h->nr_huge_pages_node[nid]) {
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001252 next_nid = hstate_next_node_to_free(h,
1253 nodes_allowed);
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001254 continue;
Lee Schermerhorn9a76db02009-12-14 17:58:15 -08001255 }
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001256 }
Wu Fengguang20a03072009-06-16 15:32:22 -07001257
1258 h->surplus_huge_pages += delta;
1259 h->surplus_huge_pages_node[nid] += delta;
1260 ret = 1;
1261 break;
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001262 } while (next_nid != start_nid);
Wu Fengguang20a03072009-06-16 15:32:22 -07001263
Wu Fengguang20a03072009-06-16 15:32:22 -07001264 return ret;
1265}
1266
Andi Kleena5516432008-07-23 21:27:41 -07001267#define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001268static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count,
1269 nodemask_t *nodes_allowed)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001270{
Adam Litke7893d1d2007-10-16 01:26:18 -07001271 unsigned long min_count, ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001272
Andi Kleenaa888a72008-07-23 21:27:47 -07001273 if (h->order >= MAX_ORDER)
1274 return h->max_huge_pages;
1275
Adam Litke7893d1d2007-10-16 01:26:18 -07001276 /*
1277 * Increase the pool size
1278 * First take pages out of surplus state. Then make up the
1279 * remaining difference by allocating fresh huge pages.
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001280 *
1281 * We might race with alloc_buddy_huge_page() here and be unable
1282 * to convert a surplus huge page to a normal huge page. That is
1283 * not critical, though, it just means the overall size of the
1284 * pool might be one hugepage larger than it needs to be, but
1285 * within all the constraints specified by the sysctls.
Adam Litke7893d1d2007-10-16 01:26:18 -07001286 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001287 spin_lock(&hugetlb_lock);
Andi Kleena5516432008-07-23 21:27:41 -07001288 while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001289 if (!adjust_pool_surplus(h, nodes_allowed, -1))
Adam Litke7893d1d2007-10-16 01:26:18 -07001290 break;
1291 }
1292
Andi Kleena5516432008-07-23 21:27:41 -07001293 while (count > persistent_huge_pages(h)) {
Adam Litke7893d1d2007-10-16 01:26:18 -07001294 /*
1295 * If this allocation races such that we no longer need the
1296 * page, free_huge_page will handle it by freeing the page
1297 * and reducing the surplus.
1298 */
1299 spin_unlock(&hugetlb_lock);
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001300 ret = alloc_fresh_huge_page(h, nodes_allowed);
Adam Litke7893d1d2007-10-16 01:26:18 -07001301 spin_lock(&hugetlb_lock);
1302 if (!ret)
1303 goto out;
1304
Mel Gorman536240f22009-12-14 17:59:56 -08001305 /* Bail for signals. Probably ctrl-c from user */
1306 if (signal_pending(current))
1307 goto out;
Adam Litke7893d1d2007-10-16 01:26:18 -07001308 }
Adam Litke7893d1d2007-10-16 01:26:18 -07001309
1310 /*
1311 * Decrease the pool size
1312 * First return free pages to the buddy allocator (being careful
1313 * to keep enough around to satisfy reservations). Then place
1314 * pages into surplus state as needed so the pool will shrink
1315 * to the desired size as pages become free.
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001316 *
1317 * By placing pages into the surplus state independent of the
1318 * overcommit value, we are allowing the surplus pool size to
1319 * exceed overcommit. There are few sane options here. Since
1320 * alloc_buddy_huge_page() is checking the global counter,
1321 * though, we'll note that we're not allowed to exceed surplus
1322 * and won't grow the pool anywhere else. Not until one of the
1323 * sysctls are changed, or the surplus pages go out of use.
Adam Litke7893d1d2007-10-16 01:26:18 -07001324 */
Andi Kleena5516432008-07-23 21:27:41 -07001325 min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
Adam Litke6b0c8802007-10-16 01:26:23 -07001326 min_count = max(count, min_count);
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001327 try_to_free_low(h, min_count, nodes_allowed);
Andi Kleena5516432008-07-23 21:27:41 -07001328 while (min_count < persistent_huge_pages(h)) {
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001329 if (!free_pool_huge_page(h, nodes_allowed, 0))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001330 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001331 }
Andi Kleena5516432008-07-23 21:27:41 -07001332 while (count < persistent_huge_pages(h)) {
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001333 if (!adjust_pool_surplus(h, nodes_allowed, 1))
Adam Litke7893d1d2007-10-16 01:26:18 -07001334 break;
1335 }
1336out:
Andi Kleena5516432008-07-23 21:27:41 -07001337 ret = persistent_huge_pages(h);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001338 spin_unlock(&hugetlb_lock);
Adam Litke7893d1d2007-10-16 01:26:18 -07001339 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001340}
1341
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001342#define HSTATE_ATTR_RO(_name) \
1343 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
1344
1345#define HSTATE_ATTR(_name) \
1346 static struct kobj_attribute _name##_attr = \
1347 __ATTR(_name, 0644, _name##_show, _name##_store)
1348
1349static struct kobject *hugepages_kobj;
1350static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
1351
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001352static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
1353
1354static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001355{
1356 int i;
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001357
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001358 for (i = 0; i < HUGE_MAX_HSTATE; i++)
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001359 if (hstate_kobjs[i] == kobj) {
1360 if (nidp)
1361 *nidp = NUMA_NO_NODE;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001362 return &hstates[i];
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001363 }
1364
1365 return kobj_to_node_hstate(kobj, nidp);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001366}
1367
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001368static ssize_t nr_hugepages_show_common(struct kobject *kobj,
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001369 struct kobj_attribute *attr, char *buf)
1370{
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001371 struct hstate *h;
1372 unsigned long nr_huge_pages;
1373 int nid;
1374
1375 h = kobj_to_hstate(kobj, &nid);
1376 if (nid == NUMA_NO_NODE)
1377 nr_huge_pages = h->nr_huge_pages;
1378 else
1379 nr_huge_pages = h->nr_huge_pages_node[nid];
1380
1381 return sprintf(buf, "%lu\n", nr_huge_pages);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001382}
Eric B Munsonadbe8722011-01-13 15:47:27 -08001383
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001384static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
1385 struct kobject *kobj, struct kobj_attribute *attr,
1386 const char *buf, size_t len)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001387{
1388 int err;
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001389 int nid;
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001390 unsigned long count;
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001391 struct hstate *h;
David Rientjesbad44b52009-12-14 17:58:38 -08001392 NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001393
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001394 err = strict_strtoul(buf, 10, &count);
Eric B Munson73ae31e2011-01-13 15:47:28 -08001395 if (err)
Eric B Munsonadbe8722011-01-13 15:47:27 -08001396 goto out;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001397
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001398 h = kobj_to_hstate(kobj, &nid);
Eric B Munsonadbe8722011-01-13 15:47:27 -08001399 if (h->order >= MAX_ORDER) {
1400 err = -EINVAL;
1401 goto out;
1402 }
1403
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001404 if (nid == NUMA_NO_NODE) {
1405 /*
1406 * global hstate attribute
1407 */
1408 if (!(obey_mempolicy &&
1409 init_nodemask_of_mempolicy(nodes_allowed))) {
1410 NODEMASK_FREE(nodes_allowed);
1411 nodes_allowed = &node_states[N_HIGH_MEMORY];
1412 }
1413 } else if (nodes_allowed) {
1414 /*
1415 * per node hstate attribute: adjust count to global,
1416 * but restrict alloc/free to the specified node.
1417 */
1418 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
1419 init_nodemask_of_node(nodes_allowed, nid);
1420 } else
1421 nodes_allowed = &node_states[N_HIGH_MEMORY];
1422
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001423 h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001424
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -08001425 if (nodes_allowed != &node_states[N_HIGH_MEMORY])
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001426 NODEMASK_FREE(nodes_allowed);
1427
1428 return len;
Eric B Munsonadbe8722011-01-13 15:47:27 -08001429out:
1430 NODEMASK_FREE(nodes_allowed);
1431 return err;
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001432}
1433
1434static ssize_t nr_hugepages_show(struct kobject *kobj,
1435 struct kobj_attribute *attr, char *buf)
1436{
1437 return nr_hugepages_show_common(kobj, attr, buf);
1438}
1439
1440static ssize_t nr_hugepages_store(struct kobject *kobj,
1441 struct kobj_attribute *attr, const char *buf, size_t len)
1442{
1443 return nr_hugepages_store_common(false, kobj, attr, buf, len);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001444}
1445HSTATE_ATTR(nr_hugepages);
1446
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001447#ifdef CONFIG_NUMA
1448
1449/*
1450 * hstate attribute for optionally mempolicy-based constraint on persistent
1451 * huge page alloc/free.
1452 */
1453static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
1454 struct kobj_attribute *attr, char *buf)
1455{
1456 return nr_hugepages_show_common(kobj, attr, buf);
1457}
1458
1459static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
1460 struct kobj_attribute *attr, const char *buf, size_t len)
1461{
1462 return nr_hugepages_store_common(true, kobj, attr, buf, len);
1463}
1464HSTATE_ATTR(nr_hugepages_mempolicy);
1465#endif
1466
1467
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001468static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
1469 struct kobj_attribute *attr, char *buf)
1470{
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001471 struct hstate *h = kobj_to_hstate(kobj, NULL);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001472 return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
1473}
Eric B Munsonadbe8722011-01-13 15:47:27 -08001474
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001475static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
1476 struct kobj_attribute *attr, const char *buf, size_t count)
1477{
1478 int err;
1479 unsigned long input;
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001480 struct hstate *h = kobj_to_hstate(kobj, NULL);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001481
Eric B Munsonadbe8722011-01-13 15:47:27 -08001482 if (h->order >= MAX_ORDER)
1483 return -EINVAL;
1484
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001485 err = strict_strtoul(buf, 10, &input);
1486 if (err)
Eric B Munson73ae31e2011-01-13 15:47:28 -08001487 return err;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001488
1489 spin_lock(&hugetlb_lock);
1490 h->nr_overcommit_huge_pages = input;
1491 spin_unlock(&hugetlb_lock);
1492
1493 return count;
1494}
1495HSTATE_ATTR(nr_overcommit_hugepages);
1496
1497static ssize_t free_hugepages_show(struct kobject *kobj,
1498 struct kobj_attribute *attr, char *buf)
1499{
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001500 struct hstate *h;
1501 unsigned long free_huge_pages;
1502 int nid;
1503
1504 h = kobj_to_hstate(kobj, &nid);
1505 if (nid == NUMA_NO_NODE)
1506 free_huge_pages = h->free_huge_pages;
1507 else
1508 free_huge_pages = h->free_huge_pages_node[nid];
1509
1510 return sprintf(buf, "%lu\n", free_huge_pages);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001511}
1512HSTATE_ATTR_RO(free_hugepages);
1513
1514static ssize_t resv_hugepages_show(struct kobject *kobj,
1515 struct kobj_attribute *attr, char *buf)
1516{
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001517 struct hstate *h = kobj_to_hstate(kobj, NULL);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001518 return sprintf(buf, "%lu\n", h->resv_huge_pages);
1519}
1520HSTATE_ATTR_RO(resv_hugepages);
1521
1522static ssize_t surplus_hugepages_show(struct kobject *kobj,
1523 struct kobj_attribute *attr, char *buf)
1524{
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001525 struct hstate *h;
1526 unsigned long surplus_huge_pages;
1527 int nid;
1528
1529 h = kobj_to_hstate(kobj, &nid);
1530 if (nid == NUMA_NO_NODE)
1531 surplus_huge_pages = h->surplus_huge_pages;
1532 else
1533 surplus_huge_pages = h->surplus_huge_pages_node[nid];
1534
1535 return sprintf(buf, "%lu\n", surplus_huge_pages);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001536}
1537HSTATE_ATTR_RO(surplus_hugepages);
1538
1539static struct attribute *hstate_attrs[] = {
1540 &nr_hugepages_attr.attr,
1541 &nr_overcommit_hugepages_attr.attr,
1542 &free_hugepages_attr.attr,
1543 &resv_hugepages_attr.attr,
1544 &surplus_hugepages_attr.attr,
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001545#ifdef CONFIG_NUMA
1546 &nr_hugepages_mempolicy_attr.attr,
1547#endif
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001548 NULL,
1549};
1550
1551static struct attribute_group hstate_attr_group = {
1552 .attrs = hstate_attrs,
1553};
1554
Jeff Mahoney094e9532010-02-02 13:44:14 -08001555static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
1556 struct kobject **hstate_kobjs,
1557 struct attribute_group *hstate_attr_group)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001558{
1559 int retval;
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001560 int hi = h - hstates;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001561
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001562 hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
1563 if (!hstate_kobjs[hi])
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001564 return -ENOMEM;
1565
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001566 retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001567 if (retval)
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001568 kobject_put(hstate_kobjs[hi]);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001569
1570 return retval;
1571}
1572
1573static void __init hugetlb_sysfs_init(void)
1574{
1575 struct hstate *h;
1576 int err;
1577
1578 hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
1579 if (!hugepages_kobj)
1580 return;
1581
1582 for_each_hstate(h) {
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001583 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
1584 hstate_kobjs, &hstate_attr_group);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001585 if (err)
1586 printk(KERN_ERR "Hugetlb: Unable to add hstate %s",
1587 h->name);
1588 }
1589}
1590
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001591#ifdef CONFIG_NUMA
1592
1593/*
1594 * node_hstate/s - associate per node hstate attributes, via their kobjects,
Kay Sievers10fbcf42011-12-21 14:48:43 -08001595 * with node devices in node_devices[] using a parallel array. The array
1596 * index of a node device or _hstate == node id.
1597 * This is here to avoid any static dependency of the node device driver, in
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001598 * the base kernel, on the hugetlb module.
1599 */
1600struct node_hstate {
1601 struct kobject *hugepages_kobj;
1602 struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
1603};
1604struct node_hstate node_hstates[MAX_NUMNODES];
1605
1606/*
Kay Sievers10fbcf42011-12-21 14:48:43 -08001607 * A subset of global hstate attributes for node devices
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001608 */
1609static struct attribute *per_node_hstate_attrs[] = {
1610 &nr_hugepages_attr.attr,
1611 &free_hugepages_attr.attr,
1612 &surplus_hugepages_attr.attr,
1613 NULL,
1614};
1615
1616static struct attribute_group per_node_hstate_attr_group = {
1617 .attrs = per_node_hstate_attrs,
1618};
1619
1620/*
Kay Sievers10fbcf42011-12-21 14:48:43 -08001621 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001622 * Returns node id via non-NULL nidp.
1623 */
1624static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
1625{
1626 int nid;
1627
1628 for (nid = 0; nid < nr_node_ids; nid++) {
1629 struct node_hstate *nhs = &node_hstates[nid];
1630 int i;
1631 for (i = 0; i < HUGE_MAX_HSTATE; i++)
1632 if (nhs->hstate_kobjs[i] == kobj) {
1633 if (nidp)
1634 *nidp = nid;
1635 return &hstates[i];
1636 }
1637 }
1638
1639 BUG();
1640 return NULL;
1641}
1642
1643/*
Kay Sievers10fbcf42011-12-21 14:48:43 -08001644 * Unregister hstate attributes from a single node device.
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001645 * No-op if no hstate attributes attached.
1646 */
1647void hugetlb_unregister_node(struct node *node)
1648{
1649 struct hstate *h;
Kay Sievers10fbcf42011-12-21 14:48:43 -08001650 struct node_hstate *nhs = &node_hstates[node->dev.id];
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001651
1652 if (!nhs->hugepages_kobj)
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -08001653 return; /* no hstate attributes */
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001654
1655 for_each_hstate(h)
1656 if (nhs->hstate_kobjs[h - hstates]) {
1657 kobject_put(nhs->hstate_kobjs[h - hstates]);
1658 nhs->hstate_kobjs[h - hstates] = NULL;
1659 }
1660
1661 kobject_put(nhs->hugepages_kobj);
1662 nhs->hugepages_kobj = NULL;
1663}
1664
1665/*
Kay Sievers10fbcf42011-12-21 14:48:43 -08001666 * hugetlb module exit: unregister hstate attributes from node devices
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001667 * that have them.
1668 */
1669static void hugetlb_unregister_all_nodes(void)
1670{
1671 int nid;
1672
1673 /*
Kay Sievers10fbcf42011-12-21 14:48:43 -08001674 * disable node device registrations.
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001675 */
1676 register_hugetlbfs_with_node(NULL, NULL);
1677
1678 /*
1679 * remove hstate attributes from any nodes that have them.
1680 */
1681 for (nid = 0; nid < nr_node_ids; nid++)
1682 hugetlb_unregister_node(&node_devices[nid]);
1683}
1684
1685/*
Kay Sievers10fbcf42011-12-21 14:48:43 -08001686 * Register hstate attributes for a single node device.
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001687 * No-op if attributes already registered.
1688 */
1689void hugetlb_register_node(struct node *node)
1690{
1691 struct hstate *h;
Kay Sievers10fbcf42011-12-21 14:48:43 -08001692 struct node_hstate *nhs = &node_hstates[node->dev.id];
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001693 int err;
1694
1695 if (nhs->hugepages_kobj)
1696 return; /* already allocated */
1697
1698 nhs->hugepages_kobj = kobject_create_and_add("hugepages",
Kay Sievers10fbcf42011-12-21 14:48:43 -08001699 &node->dev.kobj);
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001700 if (!nhs->hugepages_kobj)
1701 return;
1702
1703 for_each_hstate(h) {
1704 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
1705 nhs->hstate_kobjs,
1706 &per_node_hstate_attr_group);
1707 if (err) {
1708 printk(KERN_ERR "Hugetlb: Unable to add hstate %s"
1709 " for node %d\n",
Kay Sievers10fbcf42011-12-21 14:48:43 -08001710 h->name, node->dev.id);
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001711 hugetlb_unregister_node(node);
1712 break;
1713 }
1714 }
1715}
1716
1717/*
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -08001718 * hugetlb init time: register hstate attributes for all registered node
Kay Sievers10fbcf42011-12-21 14:48:43 -08001719 * devices of nodes that have memory. All on-line nodes should have
1720 * registered their associated device by this time.
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001721 */
1722static void hugetlb_register_all_nodes(void)
1723{
1724 int nid;
1725
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -08001726 for_each_node_state(nid, N_HIGH_MEMORY) {
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001727 struct node *node = &node_devices[nid];
Kay Sievers10fbcf42011-12-21 14:48:43 -08001728 if (node->dev.id == nid)
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001729 hugetlb_register_node(node);
1730 }
1731
1732 /*
Kay Sievers10fbcf42011-12-21 14:48:43 -08001733 * Let the node device driver know we're here so it can
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001734 * [un]register hstate attributes on node hotplug.
1735 */
1736 register_hugetlbfs_with_node(hugetlb_register_node,
1737 hugetlb_unregister_node);
1738}
1739#else /* !CONFIG_NUMA */
1740
1741static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
1742{
1743 BUG();
1744 if (nidp)
1745 *nidp = -1;
1746 return NULL;
1747}
1748
1749static void hugetlb_unregister_all_nodes(void) { }
1750
1751static void hugetlb_register_all_nodes(void) { }
1752
1753#endif
1754
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001755static void __exit hugetlb_exit(void)
1756{
1757 struct hstate *h;
1758
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001759 hugetlb_unregister_all_nodes();
1760
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001761 for_each_hstate(h) {
1762 kobject_put(hstate_kobjs[h - hstates]);
1763 }
1764
1765 kobject_put(hugepages_kobj);
1766}
1767module_exit(hugetlb_exit);
1768
1769static int __init hugetlb_init(void)
1770{
Benjamin Herrenschmidt0ef89d22008-07-31 00:07:30 -07001771 /* Some platform decide whether they support huge pages at boot
1772 * time. On these, such as powerpc, HPAGE_SHIFT is set to 0 when
1773 * there is no such support
1774 */
1775 if (HPAGE_SHIFT == 0)
1776 return 0;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001777
Nick Piggine11bfbf2008-07-23 21:27:52 -07001778 if (!size_to_hstate(default_hstate_size)) {
1779 default_hstate_size = HPAGE_SIZE;
1780 if (!size_to_hstate(default_hstate_size))
1781 hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001782 }
Nick Piggine11bfbf2008-07-23 21:27:52 -07001783 default_hstate_idx = size_to_hstate(default_hstate_size) - hstates;
1784 if (default_hstate_max_huge_pages)
1785 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001786
1787 hugetlb_init_hstates();
1788
Andi Kleenaa888a72008-07-23 21:27:47 -07001789 gather_bootmem_prealloc();
1790
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001791 report_hugepages();
1792
1793 hugetlb_sysfs_init();
1794
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001795 hugetlb_register_all_nodes();
1796
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001797 return 0;
1798}
1799module_init(hugetlb_init);
1800
1801/* Should be called on processing a hugepagesz=... option */
1802void __init hugetlb_add_hstate(unsigned order)
1803{
1804 struct hstate *h;
Andi Kleen8faa8b02008-07-23 21:27:48 -07001805 unsigned long i;
1806
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001807 if (size_to_hstate(PAGE_SIZE << order)) {
1808 printk(KERN_WARNING "hugepagesz= specified twice, ignoring\n");
1809 return;
1810 }
1811 BUG_ON(max_hstate >= HUGE_MAX_HSTATE);
1812 BUG_ON(order == 0);
1813 h = &hstates[max_hstate++];
1814 h->order = order;
1815 h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
Andi Kleen8faa8b02008-07-23 21:27:48 -07001816 h->nr_huge_pages = 0;
1817 h->free_huge_pages = 0;
1818 for (i = 0; i < MAX_NUMNODES; ++i)
1819 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -08001820 h->next_nid_to_alloc = first_node(node_states[N_HIGH_MEMORY]);
1821 h->next_nid_to_free = first_node(node_states[N_HIGH_MEMORY]);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001822 snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
1823 huge_page_size(h)/1024);
Andi Kleen8faa8b02008-07-23 21:27:48 -07001824
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001825 parsed_hstate = h;
1826}
1827
Nick Piggine11bfbf2008-07-23 21:27:52 -07001828static int __init hugetlb_nrpages_setup(char *s)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001829{
1830 unsigned long *mhp;
Andi Kleen8faa8b02008-07-23 21:27:48 -07001831 static unsigned long *last_mhp;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001832
1833 /*
1834 * !max_hstate means we haven't parsed a hugepagesz= parameter yet,
1835 * so this hugepages= parameter goes to the "default hstate".
1836 */
1837 if (!max_hstate)
1838 mhp = &default_hstate_max_huge_pages;
1839 else
1840 mhp = &parsed_hstate->max_huge_pages;
1841
Andi Kleen8faa8b02008-07-23 21:27:48 -07001842 if (mhp == last_mhp) {
1843 printk(KERN_WARNING "hugepages= specified twice without "
1844 "interleaving hugepagesz=, ignoring\n");
1845 return 1;
1846 }
1847
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001848 if (sscanf(s, "%lu", mhp) <= 0)
1849 *mhp = 0;
1850
Andi Kleen8faa8b02008-07-23 21:27:48 -07001851 /*
1852 * Global state is always initialized later in hugetlb_init.
1853 * But we need to allocate >= MAX_ORDER hstates here early to still
1854 * use the bootmem allocator.
1855 */
1856 if (max_hstate && parsed_hstate->order >= MAX_ORDER)
1857 hugetlb_hstate_alloc_pages(parsed_hstate);
1858
1859 last_mhp = mhp;
1860
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001861 return 1;
1862}
Nick Piggine11bfbf2008-07-23 21:27:52 -07001863__setup("hugepages=", hugetlb_nrpages_setup);
1864
1865static int __init hugetlb_default_setup(char *s)
1866{
1867 default_hstate_size = memparse(s, &s);
1868 return 1;
1869}
1870__setup("default_hugepagesz=", hugetlb_default_setup);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001871
Nishanth Aravamudan8a213462008-07-25 19:44:37 -07001872static unsigned int cpuset_mems_nr(unsigned int *array)
1873{
1874 int node;
1875 unsigned int nr = 0;
1876
1877 for_each_node_mask(node, cpuset_current_mems_allowed)
1878 nr += array[node];
1879
1880 return nr;
1881}
1882
1883#ifdef CONFIG_SYSCTL
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001884static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
1885 struct ctl_table *table, int write,
1886 void __user *buffer, size_t *length, loff_t *ppos)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001887{
Andi Kleene5ff2152008-07-23 21:27:42 -07001888 struct hstate *h = &default_hstate;
1889 unsigned long tmp;
Michal Hocko08d4a242011-01-13 15:47:26 -08001890 int ret;
Andi Kleene5ff2152008-07-23 21:27:42 -07001891
Petr Holasekc033a932011-03-22 16:33:05 -07001892 tmp = h->max_huge_pages;
Andi Kleene5ff2152008-07-23 21:27:42 -07001893
Eric B Munsonadbe8722011-01-13 15:47:27 -08001894 if (write && h->order >= MAX_ORDER)
1895 return -EINVAL;
1896
Andi Kleene5ff2152008-07-23 21:27:42 -07001897 table->data = &tmp;
1898 table->maxlen = sizeof(unsigned long);
Michal Hocko08d4a242011-01-13 15:47:26 -08001899 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
1900 if (ret)
1901 goto out;
Andi Kleene5ff2152008-07-23 21:27:42 -07001902
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001903 if (write) {
David Rientjesbad44b52009-12-14 17:58:38 -08001904 NODEMASK_ALLOC(nodemask_t, nodes_allowed,
1905 GFP_KERNEL | __GFP_NORETRY);
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001906 if (!(obey_mempolicy &&
1907 init_nodemask_of_mempolicy(nodes_allowed))) {
1908 NODEMASK_FREE(nodes_allowed);
1909 nodes_allowed = &node_states[N_HIGH_MEMORY];
1910 }
1911 h->max_huge_pages = set_max_huge_pages(h, tmp, nodes_allowed);
1912
1913 if (nodes_allowed != &node_states[N_HIGH_MEMORY])
1914 NODEMASK_FREE(nodes_allowed);
1915 }
Michal Hocko08d4a242011-01-13 15:47:26 -08001916out:
1917 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001918}
Mel Gorman396faf02007-07-17 04:03:13 -07001919
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001920int hugetlb_sysctl_handler(struct ctl_table *table, int write,
1921 void __user *buffer, size_t *length, loff_t *ppos)
1922{
1923
1924 return hugetlb_sysctl_handler_common(false, table, write,
1925 buffer, length, ppos);
1926}
1927
1928#ifdef CONFIG_NUMA
1929int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
1930 void __user *buffer, size_t *length, loff_t *ppos)
1931{
1932 return hugetlb_sysctl_handler_common(true, table, write,
1933 buffer, length, ppos);
1934}
1935#endif /* CONFIG_NUMA */
1936
Mel Gorman396faf02007-07-17 04:03:13 -07001937int hugetlb_treat_movable_handler(struct ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07001938 void __user *buffer,
Mel Gorman396faf02007-07-17 04:03:13 -07001939 size_t *length, loff_t *ppos)
1940{
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07001941 proc_dointvec(table, write, buffer, length, ppos);
Mel Gorman396faf02007-07-17 04:03:13 -07001942 if (hugepages_treat_as_movable)
1943 htlb_alloc_mask = GFP_HIGHUSER_MOVABLE;
1944 else
1945 htlb_alloc_mask = GFP_HIGHUSER;
1946 return 0;
1947}
1948
Nishanth Aravamudana3d0c6a2008-02-08 04:18:18 -08001949int hugetlb_overcommit_handler(struct ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07001950 void __user *buffer,
Nishanth Aravamudana3d0c6a2008-02-08 04:18:18 -08001951 size_t *length, loff_t *ppos)
1952{
Andi Kleena5516432008-07-23 21:27:41 -07001953 struct hstate *h = &default_hstate;
Andi Kleene5ff2152008-07-23 21:27:42 -07001954 unsigned long tmp;
Michal Hocko08d4a242011-01-13 15:47:26 -08001955 int ret;
Andi Kleene5ff2152008-07-23 21:27:42 -07001956
Petr Holasekc033a932011-03-22 16:33:05 -07001957 tmp = h->nr_overcommit_huge_pages;
Andi Kleene5ff2152008-07-23 21:27:42 -07001958
Eric B Munsonadbe8722011-01-13 15:47:27 -08001959 if (write && h->order >= MAX_ORDER)
1960 return -EINVAL;
1961
Andi Kleene5ff2152008-07-23 21:27:42 -07001962 table->data = &tmp;
1963 table->maxlen = sizeof(unsigned long);
Michal Hocko08d4a242011-01-13 15:47:26 -08001964 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
1965 if (ret)
1966 goto out;
Andi Kleene5ff2152008-07-23 21:27:42 -07001967
1968 if (write) {
1969 spin_lock(&hugetlb_lock);
1970 h->nr_overcommit_huge_pages = tmp;
1971 spin_unlock(&hugetlb_lock);
1972 }
Michal Hocko08d4a242011-01-13 15:47:26 -08001973out:
1974 return ret;
Nishanth Aravamudana3d0c6a2008-02-08 04:18:18 -08001975}
1976
Linus Torvalds1da177e2005-04-16 15:20:36 -07001977#endif /* CONFIG_SYSCTL */
1978
Alexey Dobriyane1759c22008-10-15 23:50:22 +04001979void hugetlb_report_meminfo(struct seq_file *m)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001980{
Andi Kleena5516432008-07-23 21:27:41 -07001981 struct hstate *h = &default_hstate;
Alexey Dobriyane1759c22008-10-15 23:50:22 +04001982 seq_printf(m,
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001983 "HugePages_Total: %5lu\n"
1984 "HugePages_Free: %5lu\n"
1985 "HugePages_Rsvd: %5lu\n"
1986 "HugePages_Surp: %5lu\n"
1987 "Hugepagesize: %8lu kB\n",
Andi Kleena5516432008-07-23 21:27:41 -07001988 h->nr_huge_pages,
1989 h->free_huge_pages,
1990 h->resv_huge_pages,
1991 h->surplus_huge_pages,
1992 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001993}
1994
1995int hugetlb_report_node_meminfo(int nid, char *buf)
1996{
Andi Kleena5516432008-07-23 21:27:41 -07001997 struct hstate *h = &default_hstate;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001998 return sprintf(buf,
1999 "Node %d HugePages_Total: %5u\n"
Nishanth Aravamudana1de0912008-03-26 14:37:53 -07002000 "Node %d HugePages_Free: %5u\n"
2001 "Node %d HugePages_Surp: %5u\n",
Andi Kleena5516432008-07-23 21:27:41 -07002002 nid, h->nr_huge_pages_node[nid],
2003 nid, h->free_huge_pages_node[nid],
2004 nid, h->surplus_huge_pages_node[nid]);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002005}
2006
Linus Torvalds1da177e2005-04-16 15:20:36 -07002007/* Return the number pages of memory we physically have, in PAGE_SIZE units. */
2008unsigned long hugetlb_total_pages(void)
2009{
Andi Kleena5516432008-07-23 21:27:41 -07002010 struct hstate *h = &default_hstate;
2011 return h->nr_huge_pages * pages_per_huge_page(h);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002012}
Linus Torvalds1da177e2005-04-16 15:20:36 -07002013
Andi Kleena5516432008-07-23 21:27:41 -07002014static int hugetlb_acct_memory(struct hstate *h, long delta)
Mel Gormanfc1b8a72008-07-23 21:27:22 -07002015{
2016 int ret = -ENOMEM;
2017
2018 spin_lock(&hugetlb_lock);
2019 /*
2020 * When cpuset is configured, it breaks the strict hugetlb page
2021 * reservation as the accounting is done on a global variable. Such
2022 * reservation is completely rubbish in the presence of cpuset because
2023 * the reservation is not checked against page availability for the
2024 * current cpuset. Application can still potentially OOM'ed by kernel
2025 * with lack of free htlb page in cpuset that the task is in.
2026 * Attempt to enforce strict accounting with cpuset is almost
2027 * impossible (or too ugly) because cpuset is too fluid that
2028 * task or memory node can be dynamically moved between cpusets.
2029 *
2030 * The change of semantics for shared hugetlb mapping with cpuset is
2031 * undesirable. However, in order to preserve some of the semantics,
2032 * we fall back to check against current free page availability as
2033 * a best attempt and hopefully to minimize the impact of changing
2034 * semantics that cpuset has.
2035 */
2036 if (delta > 0) {
Andi Kleena5516432008-07-23 21:27:41 -07002037 if (gather_surplus_pages(h, delta) < 0)
Mel Gormanfc1b8a72008-07-23 21:27:22 -07002038 goto out;
2039
Andi Kleena5516432008-07-23 21:27:41 -07002040 if (delta > cpuset_mems_nr(h->free_huge_pages_node)) {
2041 return_unused_surplus_pages(h, delta);
Mel Gormanfc1b8a72008-07-23 21:27:22 -07002042 goto out;
2043 }
2044 }
2045
2046 ret = 0;
2047 if (delta < 0)
Andi Kleena5516432008-07-23 21:27:41 -07002048 return_unused_surplus_pages(h, (unsigned long) -delta);
Mel Gormanfc1b8a72008-07-23 21:27:22 -07002049
2050out:
2051 spin_unlock(&hugetlb_lock);
2052 return ret;
2053}
2054
Andy Whitcroft84afd992008-07-23 21:27:32 -07002055static void hugetlb_vm_op_open(struct vm_area_struct *vma)
2056{
2057 struct resv_map *reservations = vma_resv_map(vma);
2058
2059 /*
2060 * This new VMA should share its siblings reservation map if present.
2061 * The VMA will only ever have a valid reservation map pointer where
2062 * it is being copied for another still existing VMA. As that VMA
Lucas De Marchi25985ed2011-03-30 22:57:33 -03002063 * has a reference to the reservation map it cannot disappear until
Andy Whitcroft84afd992008-07-23 21:27:32 -07002064 * after this open call completes. It is therefore safe to take a
2065 * new reference here without additional locking.
2066 */
2067 if (reservations)
2068 kref_get(&reservations->refs);
2069}
2070
Mel Gormana1e78772008-07-23 21:27:23 -07002071static void hugetlb_vm_op_close(struct vm_area_struct *vma)
2072{
Andi Kleena5516432008-07-23 21:27:41 -07002073 struct hstate *h = hstate_vma(vma);
Andy Whitcroft84afd992008-07-23 21:27:32 -07002074 struct resv_map *reservations = vma_resv_map(vma);
2075 unsigned long reserve;
2076 unsigned long start;
2077 unsigned long end;
2078
2079 if (reservations) {
Andi Kleena5516432008-07-23 21:27:41 -07002080 start = vma_hugecache_offset(h, vma, vma->vm_start);
2081 end = vma_hugecache_offset(h, vma, vma->vm_end);
Andy Whitcroft84afd992008-07-23 21:27:32 -07002082
2083 reserve = (end - start) -
2084 region_count(&reservations->regions, start, end);
2085
2086 kref_put(&reservations->refs, resv_map_release);
2087
Adam Litke7251ff72008-07-23 21:27:59 -07002088 if (reserve) {
Andi Kleena5516432008-07-23 21:27:41 -07002089 hugetlb_acct_memory(h, -reserve);
Adam Litke7251ff72008-07-23 21:27:59 -07002090 hugetlb_put_quota(vma->vm_file->f_mapping, reserve);
2091 }
Andy Whitcroft84afd992008-07-23 21:27:32 -07002092 }
Mel Gormana1e78772008-07-23 21:27:23 -07002093}
2094
Linus Torvalds1da177e2005-04-16 15:20:36 -07002095/*
2096 * We cannot handle pagefaults against hugetlb pages at all. They cause
2097 * handle_mm_fault() to try to instantiate regular-sized pages in the
2098 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
2099 * this far.
2100 */
Nick Piggind0217ac2007-07-19 01:47:03 -07002101static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002102{
2103 BUG();
Nick Piggind0217ac2007-07-19 01:47:03 -07002104 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002105}
2106
Alexey Dobriyanf0f37e22009-09-27 22:29:37 +04002107const struct vm_operations_struct hugetlb_vm_ops = {
Nick Piggind0217ac2007-07-19 01:47:03 -07002108 .fault = hugetlb_vm_op_fault,
Andy Whitcroft84afd992008-07-23 21:27:32 -07002109 .open = hugetlb_vm_op_open,
Mel Gormana1e78772008-07-23 21:27:23 -07002110 .close = hugetlb_vm_op_close,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002111};
2112
David Gibson1e8f8892006-01-06 00:10:44 -08002113static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
2114 int writable)
David Gibson63551ae2005-06-21 17:14:44 -07002115{
2116 pte_t entry;
2117
David Gibson1e8f8892006-01-06 00:10:44 -08002118 if (writable) {
David Gibson63551ae2005-06-21 17:14:44 -07002119 entry =
2120 pte_mkwrite(pte_mkdirty(mk_pte(page, vma->vm_page_prot)));
2121 } else {
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07002122 entry = huge_pte_wrprotect(mk_pte(page, vma->vm_page_prot));
David Gibson63551ae2005-06-21 17:14:44 -07002123 }
2124 entry = pte_mkyoung(entry);
2125 entry = pte_mkhuge(entry);
2126
2127 return entry;
2128}
2129
David Gibson1e8f8892006-01-06 00:10:44 -08002130static void set_huge_ptep_writable(struct vm_area_struct *vma,
2131 unsigned long address, pte_t *ptep)
2132{
2133 pte_t entry;
2134
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07002135 entry = pte_mkwrite(pte_mkdirty(huge_ptep_get(ptep)));
Chris Forbes32f84522011-07-25 17:12:14 -07002136 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
Russell King4b3073e2009-12-18 16:40:18 +00002137 update_mmu_cache(vma, address, ptep);
David Gibson1e8f8892006-01-06 00:10:44 -08002138}
2139
2140
David Gibson63551ae2005-06-21 17:14:44 -07002141int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
2142 struct vm_area_struct *vma)
2143{
2144 pte_t *src_pte, *dst_pte, entry;
2145 struct page *ptepage;
Hugh Dickins1c598272005-10-19 21:23:43 -07002146 unsigned long addr;
David Gibson1e8f8892006-01-06 00:10:44 -08002147 int cow;
Andi Kleena5516432008-07-23 21:27:41 -07002148 struct hstate *h = hstate_vma(vma);
2149 unsigned long sz = huge_page_size(h);
David Gibson1e8f8892006-01-06 00:10:44 -08002150
2151 cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
David Gibson63551ae2005-06-21 17:14:44 -07002152
Andi Kleena5516432008-07-23 21:27:41 -07002153 for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
Hugh Dickinsc74df322005-10-29 18:16:23 -07002154 src_pte = huge_pte_offset(src, addr);
2155 if (!src_pte)
2156 continue;
Andi Kleena5516432008-07-23 21:27:41 -07002157 dst_pte = huge_pte_alloc(dst, addr, sz);
David Gibson63551ae2005-06-21 17:14:44 -07002158 if (!dst_pte)
2159 goto nomem;
Larry Woodmanc5c99422008-01-24 05:49:25 -08002160
2161 /* If the pagetables are shared don't copy or take references */
2162 if (dst_pte == src_pte)
2163 continue;
2164
Hugh Dickinsc74df322005-10-29 18:16:23 -07002165 spin_lock(&dst->page_table_lock);
Nick Piggin46478752008-06-05 22:45:57 -07002166 spin_lock_nested(&src->page_table_lock, SINGLE_DEPTH_NESTING);
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07002167 if (!huge_pte_none(huge_ptep_get(src_pte))) {
David Gibson1e8f8892006-01-06 00:10:44 -08002168 if (cow)
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07002169 huge_ptep_set_wrprotect(src, addr, src_pte);
2170 entry = huge_ptep_get(src_pte);
Hugh Dickins1c598272005-10-19 21:23:43 -07002171 ptepage = pte_page(entry);
2172 get_page(ptepage);
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002173 page_dup_rmap(ptepage);
Hugh Dickins1c598272005-10-19 21:23:43 -07002174 set_huge_pte_at(dst, addr, dst_pte, entry);
2175 }
2176 spin_unlock(&src->page_table_lock);
Hugh Dickinsc74df322005-10-29 18:16:23 -07002177 spin_unlock(&dst->page_table_lock);
David Gibson63551ae2005-06-21 17:14:44 -07002178 }
2179 return 0;
2180
2181nomem:
2182 return -ENOMEM;
2183}
2184
Naoya Horiguchi290408d2010-09-08 10:19:35 +09002185static int is_hugetlb_entry_migration(pte_t pte)
2186{
2187 swp_entry_t swp;
2188
2189 if (huge_pte_none(pte) || pte_present(pte))
2190 return 0;
2191 swp = pte_to_swp_entry(pte);
Chris Forbes32f84522011-07-25 17:12:14 -07002192 if (non_swap_entry(swp) && is_migration_entry(swp))
Naoya Horiguchi290408d2010-09-08 10:19:35 +09002193 return 1;
Chris Forbes32f84522011-07-25 17:12:14 -07002194 else
Naoya Horiguchi290408d2010-09-08 10:19:35 +09002195 return 0;
2196}
2197
Naoya Horiguchifd6a03e2010-05-28 09:29:21 +09002198static int is_hugetlb_entry_hwpoisoned(pte_t pte)
2199{
2200 swp_entry_t swp;
2201
2202 if (huge_pte_none(pte) || pte_present(pte))
2203 return 0;
2204 swp = pte_to_swp_entry(pte);
Chris Forbes32f84522011-07-25 17:12:14 -07002205 if (non_swap_entry(swp) && is_hwpoison_entry(swp))
Naoya Horiguchifd6a03e2010-05-28 09:29:21 +09002206 return 1;
Chris Forbes32f84522011-07-25 17:12:14 -07002207 else
Naoya Horiguchifd6a03e2010-05-28 09:29:21 +09002208 return 0;
2209}
2210
Chen, Kenneth W502717f2006-10-11 01:20:46 -07002211void __unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002212 unsigned long end, struct page *ref_page)
David Gibson63551ae2005-06-21 17:14:44 -07002213{
2214 struct mm_struct *mm = vma->vm_mm;
2215 unsigned long address;
David Gibsonc7546f82005-08-05 11:59:35 -07002216 pte_t *ptep;
David Gibson63551ae2005-06-21 17:14:44 -07002217 pte_t pte;
2218 struct page *page;
Chen, Kenneth Wfe1668a2006-10-04 02:15:24 -07002219 struct page *tmp;
Andi Kleena5516432008-07-23 21:27:41 -07002220 struct hstate *h = hstate_vma(vma);
2221 unsigned long sz = huge_page_size(h);
2222
Chen, Kenneth Wc0a499c2006-12-06 20:31:39 -08002223 /*
Peter Zijlstra3d48ae42011-05-24 17:12:06 -07002224 * A page gathering list, protected by per file i_mmap_mutex. The
Chen, Kenneth Wc0a499c2006-12-06 20:31:39 -08002225 * lock is used to avoid list corruption from multiple unmapping
2226 * of the same page since we are using page->lru.
2227 */
Chen, Kenneth Wfe1668a2006-10-04 02:15:24 -07002228 LIST_HEAD(page_list);
David Gibson63551ae2005-06-21 17:14:44 -07002229
2230 WARN_ON(!is_vm_hugetlb_page(vma));
Andi Kleena5516432008-07-23 21:27:41 -07002231 BUG_ON(start & ~huge_page_mask(h));
2232 BUG_ON(end & ~huge_page_mask(h));
David Gibson63551ae2005-06-21 17:14:44 -07002233
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07002234 mmu_notifier_invalidate_range_start(mm, start, end);
Hugh Dickins508034a2005-10-29 18:16:30 -07002235 spin_lock(&mm->page_table_lock);
Andi Kleena5516432008-07-23 21:27:41 -07002236 for (address = start; address < end; address += sz) {
David Gibsonc7546f82005-08-05 11:59:35 -07002237 ptep = huge_pte_offset(mm, address);
Adam Litke4c887262005-10-29 18:16:46 -07002238 if (!ptep)
David Gibsonc7546f82005-08-05 11:59:35 -07002239 continue;
2240
Chen, Kenneth W39dde652006-12-06 20:32:03 -08002241 if (huge_pmd_unshare(mm, &address, ptep))
2242 continue;
2243
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002244 /*
2245 * If a reference page is supplied, it is because a specific
2246 * page is being unmapped, not a range. Ensure the page we
2247 * are about to unmap is the actual page of interest.
2248 */
2249 if (ref_page) {
2250 pte = huge_ptep_get(ptep);
2251 if (huge_pte_none(pte))
2252 continue;
2253 page = pte_page(pte);
2254 if (page != ref_page)
2255 continue;
2256
2257 /*
2258 * Mark the VMA as having unmapped its page so that
2259 * future faults in this VMA will fail rather than
2260 * looking like data was lost
2261 */
2262 set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
2263 }
2264
David Gibsonc7546f82005-08-05 11:59:35 -07002265 pte = huge_ptep_get_and_clear(mm, address, ptep);
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07002266 if (huge_pte_none(pte))
David Gibson63551ae2005-06-21 17:14:44 -07002267 continue;
David Gibsonc7546f82005-08-05 11:59:35 -07002268
Naoya Horiguchifd6a03e2010-05-28 09:29:21 +09002269 /*
2270 * HWPoisoned hugepage is already unmapped and dropped reference
2271 */
2272 if (unlikely(is_hugetlb_entry_hwpoisoned(pte)))
2273 continue;
2274
David Gibson63551ae2005-06-21 17:14:44 -07002275 page = pte_page(pte);
Ken Chen6649a382007-02-08 14:20:27 -08002276 if (pte_dirty(pte))
2277 set_page_dirty(page);
Chen, Kenneth Wfe1668a2006-10-04 02:15:24 -07002278 list_add(&page->lru, &page_list);
David Gibson63551ae2005-06-21 17:14:44 -07002279 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002280 spin_unlock(&mm->page_table_lock);
Hugh Dickins508034a2005-10-29 18:16:30 -07002281 flush_tlb_range(vma, start, end);
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07002282 mmu_notifier_invalidate_range_end(mm, start, end);
Chen, Kenneth Wfe1668a2006-10-04 02:15:24 -07002283 list_for_each_entry_safe(page, tmp, &page_list, lru) {
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002284 page_remove_rmap(page);
Chen, Kenneth Wfe1668a2006-10-04 02:15:24 -07002285 list_del(&page->lru);
2286 put_page(page);
2287 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002288}
David Gibson63551ae2005-06-21 17:14:44 -07002289
Chen, Kenneth W502717f2006-10-11 01:20:46 -07002290void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002291 unsigned long end, struct page *ref_page)
Chen, Kenneth W502717f2006-10-11 01:20:46 -07002292{
Peter Zijlstra3d48ae42011-05-24 17:12:06 -07002293 mutex_lock(&vma->vm_file->f_mapping->i_mmap_mutex);
Andi Kleena137e1c2008-07-23 21:27:43 -07002294 __unmap_hugepage_range(vma, start, end, ref_page);
Peter Zijlstra3d48ae42011-05-24 17:12:06 -07002295 mutex_unlock(&vma->vm_file->f_mapping->i_mmap_mutex);
Chen, Kenneth W502717f2006-10-11 01:20:46 -07002296}
2297
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002298/*
2299 * This is called when the original mapper is failing to COW a MAP_PRIVATE
2300 * mappping it owns the reserve page for. The intention is to unmap the page
2301 * from other VMAs and let the children be SIGKILLed if they are faulting the
2302 * same region.
2303 */
Harvey Harrison2a4b3de2008-10-18 20:27:06 -07002304static int unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
2305 struct page *page, unsigned long address)
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002306{
Adam Litke75266742008-11-12 13:24:56 -08002307 struct hstate *h = hstate_vma(vma);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002308 struct vm_area_struct *iter_vma;
2309 struct address_space *mapping;
2310 struct prio_tree_iter iter;
2311 pgoff_t pgoff;
2312
2313 /*
2314 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
2315 * from page cache lookup which is in HPAGE_SIZE units.
2316 */
Adam Litke75266742008-11-12 13:24:56 -08002317 address = address & huge_page_mask(h);
Hillf Danton0c176d52012-01-10 15:08:19 -08002318 pgoff = vma_hugecache_offset(h, vma, address);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002319 mapping = (struct address_space *)page_private(page);
2320
Mel Gorman4eb2b1d2009-12-14 17:59:53 -08002321 /*
2322 * Take the mapping lock for the duration of the table walk. As
2323 * this mapping should be shared between all the VMAs,
2324 * __unmap_hugepage_range() is called as the lock is already held
2325 */
Peter Zijlstra3d48ae42011-05-24 17:12:06 -07002326 mutex_lock(&mapping->i_mmap_mutex);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002327 vma_prio_tree_foreach(iter_vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
2328 /* Do not unmap the current VMA */
2329 if (iter_vma == vma)
2330 continue;
2331
2332 /*
2333 * Unmap the page from other VMAs without their own reserves.
2334 * They get marked to be SIGKILLed if they fault in these
2335 * areas. This is because a future no-page fault on this VMA
2336 * could insert a zeroed page instead of the data existing
2337 * from the time of fork. This would look like data corruption
2338 */
2339 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
Mel Gorman4eb2b1d2009-12-14 17:59:53 -08002340 __unmap_hugepage_range(iter_vma,
Adam Litke75266742008-11-12 13:24:56 -08002341 address, address + huge_page_size(h),
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002342 page);
2343 }
Peter Zijlstra3d48ae42011-05-24 17:12:06 -07002344 mutex_unlock(&mapping->i_mmap_mutex);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002345
2346 return 1;
2347}
2348
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002349/*
2350 * Hugetlb_cow() should be called with page lock of the original hugepage held.
Michal Hockoef009b22012-01-10 15:07:21 -08002351 * Called with hugetlb_instantiation_mutex held and pte_page locked so we
2352 * cannot race with other handlers or page migration.
2353 * Keep the pte_same checks anyway to make transition from the mutex easier.
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002354 */
David Gibson1e8f8892006-01-06 00:10:44 -08002355static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002356 unsigned long address, pte_t *ptep, pte_t pte,
2357 struct page *pagecache_page)
David Gibson1e8f8892006-01-06 00:10:44 -08002358{
Andi Kleena5516432008-07-23 21:27:41 -07002359 struct hstate *h = hstate_vma(vma);
David Gibson1e8f8892006-01-06 00:10:44 -08002360 struct page *old_page, *new_page;
David Gibson79ac6ba2006-03-22 00:08:51 -08002361 int avoidcopy;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002362 int outside_reserve = 0;
David Gibson1e8f8892006-01-06 00:10:44 -08002363
2364 old_page = pte_page(pte);
2365
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002366retry_avoidcopy:
David Gibson1e8f8892006-01-06 00:10:44 -08002367 /* If no-one else is actually using this page, avoid the copy
2368 * and just make the page writable */
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002369 avoidcopy = (page_mapcount(old_page) == 1);
David Gibson1e8f8892006-01-06 00:10:44 -08002370 if (avoidcopy) {
Naoya Horiguchi56c9cfb2010-09-10 13:23:04 +09002371 if (PageAnon(old_page))
2372 page_move_anon_rmap(old_page, vma, address);
David Gibson1e8f8892006-01-06 00:10:44 -08002373 set_huge_ptep_writable(vma, address, ptep);
Nick Piggin83c54072007-07-19 01:47:05 -07002374 return 0;
David Gibson1e8f8892006-01-06 00:10:44 -08002375 }
2376
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002377 /*
2378 * If the process that created a MAP_PRIVATE mapping is about to
2379 * perform a COW due to a shared page count, attempt to satisfy
2380 * the allocation without using the existing reserves. The pagecache
2381 * page is used to determine if the reserve at this address was
2382 * consumed or not. If reserves were used, a partial faulted mapping
2383 * at the time of fork() could consume its reserves on COW instead
2384 * of the full address range.
2385 */
Mel Gormanf83a2752009-05-28 14:34:40 -07002386 if (!(vma->vm_flags & VM_MAYSHARE) &&
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002387 is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
2388 old_page != pagecache_page)
2389 outside_reserve = 1;
2390
David Gibson1e8f8892006-01-06 00:10:44 -08002391 page_cache_get(old_page);
Larry Woodmanb76c8cf2009-12-14 17:59:37 -08002392
2393 /* Drop page_table_lock as buddy allocator may be called */
2394 spin_unlock(&mm->page_table_lock);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002395 new_page = alloc_huge_page(vma, address, outside_reserve);
David Gibson1e8f8892006-01-06 00:10:44 -08002396
Adam Litke2fc39ce2007-11-14 16:59:39 -08002397 if (IS_ERR(new_page)) {
David Gibson1e8f8892006-01-06 00:10:44 -08002398 page_cache_release(old_page);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002399
2400 /*
2401 * If a process owning a MAP_PRIVATE mapping fails to COW,
2402 * it is due to references held by a child and an insufficient
2403 * huge page pool. To guarantee the original mappers
2404 * reliability, unmap the page from child processes. The child
2405 * may get SIGKILLed if it later faults.
2406 */
2407 if (outside_reserve) {
2408 BUG_ON(huge_pte_none(pte));
2409 if (unmap_ref_private(mm, vma, old_page, address)) {
2410 BUG_ON(page_count(old_page) != 1);
2411 BUG_ON(huge_pte_none(pte));
Larry Woodmanb76c8cf2009-12-14 17:59:37 -08002412 spin_lock(&mm->page_table_lock);
Hillf Dantona734bcc2012-01-10 15:07:20 -08002413 ptep = huge_pte_offset(mm, address & huge_page_mask(h));
2414 if (likely(pte_same(huge_ptep_get(ptep), pte)))
2415 goto retry_avoidcopy;
2416 /*
2417 * race occurs while re-acquiring page_table_lock, and
2418 * our job is done.
2419 */
2420 return 0;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002421 }
2422 WARN_ON_ONCE(1);
2423 }
2424
Larry Woodmanb76c8cf2009-12-14 17:59:37 -08002425 /* Caller expects lock to be held */
2426 spin_lock(&mm->page_table_lock);
Adam Litke2fc39ce2007-11-14 16:59:39 -08002427 return -PTR_ERR(new_page);
David Gibson1e8f8892006-01-06 00:10:44 -08002428 }
2429
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002430 /*
2431 * When the original hugepage is shared one, it does not have
2432 * anon_vma prepared.
2433 */
Dean Nelson44e2aa92010-10-26 14:22:08 -07002434 if (unlikely(anon_vma_prepare(vma))) {
Hillf Dantonea4039a2011-11-15 14:36:12 -08002435 page_cache_release(new_page);
2436 page_cache_release(old_page);
Dean Nelson44e2aa92010-10-26 14:22:08 -07002437 /* Caller expects lock to be held */
2438 spin_lock(&mm->page_table_lock);
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002439 return VM_FAULT_OOM;
Dean Nelson44e2aa92010-10-26 14:22:08 -07002440 }
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002441
Andrea Arcangeli47ad8472011-01-13 15:46:47 -08002442 copy_user_huge_page(new_page, old_page, address, vma,
2443 pages_per_huge_page(h));
Nick Piggin0ed361d2008-02-04 22:29:34 -08002444 __SetPageUptodate(new_page);
David Gibson1e8f8892006-01-06 00:10:44 -08002445
Larry Woodmanb76c8cf2009-12-14 17:59:37 -08002446 /*
2447 * Retake the page_table_lock to check for racing updates
2448 * before the page tables are altered
2449 */
2450 spin_lock(&mm->page_table_lock);
Andi Kleena5516432008-07-23 21:27:41 -07002451 ptep = huge_pte_offset(mm, address & huge_page_mask(h));
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07002452 if (likely(pte_same(huge_ptep_get(ptep), pte))) {
David Gibson1e8f8892006-01-06 00:10:44 -08002453 /* Break COW */
Doug Doan3edd4fc2010-08-09 17:18:30 -07002454 mmu_notifier_invalidate_range_start(mm,
2455 address & huge_page_mask(h),
2456 (address & huge_page_mask(h)) + huge_page_size(h));
Gerald Schaefer8fe627e2008-04-28 02:13:28 -07002457 huge_ptep_clear_flush(vma, address, ptep);
David Gibson1e8f8892006-01-06 00:10:44 -08002458 set_huge_pte_at(mm, address, ptep,
2459 make_huge_pte(vma, new_page, 1));
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002460 page_remove_rmap(old_page);
Naoya Horiguchicd67f0d2010-09-10 13:23:04 +09002461 hugepage_add_new_anon_rmap(new_page, vma, address);
David Gibson1e8f8892006-01-06 00:10:44 -08002462 /* Make the old page be freed below */
2463 new_page = old_page;
Doug Doan3edd4fc2010-08-09 17:18:30 -07002464 mmu_notifier_invalidate_range_end(mm,
2465 address & huge_page_mask(h),
2466 (address & huge_page_mask(h)) + huge_page_size(h));
David Gibson1e8f8892006-01-06 00:10:44 -08002467 }
2468 page_cache_release(new_page);
2469 page_cache_release(old_page);
Nick Piggin83c54072007-07-19 01:47:05 -07002470 return 0;
David Gibson1e8f8892006-01-06 00:10:44 -08002471}
2472
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002473/* Return the pagecache page at a given address within a VMA */
Andi Kleena5516432008-07-23 21:27:41 -07002474static struct page *hugetlbfs_pagecache_page(struct hstate *h,
2475 struct vm_area_struct *vma, unsigned long address)
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002476{
2477 struct address_space *mapping;
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -07002478 pgoff_t idx;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002479
2480 mapping = vma->vm_file->f_mapping;
Andi Kleena5516432008-07-23 21:27:41 -07002481 idx = vma_hugecache_offset(h, vma, address);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002482
2483 return find_lock_page(mapping, idx);
2484}
2485
Hugh Dickins3ae77f42009-09-21 17:03:33 -07002486/*
2487 * Return whether there is a pagecache page to back given address within VMA.
2488 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
2489 */
2490static bool hugetlbfs_pagecache_present(struct hstate *h,
Hugh Dickins2a15efc2009-09-21 17:03:27 -07002491 struct vm_area_struct *vma, unsigned long address)
2492{
2493 struct address_space *mapping;
2494 pgoff_t idx;
2495 struct page *page;
2496
2497 mapping = vma->vm_file->f_mapping;
2498 idx = vma_hugecache_offset(h, vma, address);
2499
2500 page = find_get_page(mapping, idx);
2501 if (page)
2502 put_page(page);
2503 return page != NULL;
2504}
2505
Robert P. J. Daya1ed3dd2007-07-17 04:03:33 -07002506static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
Hugh Dickins788c7df2009-06-23 13:49:05 +01002507 unsigned long address, pte_t *ptep, unsigned int flags)
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01002508{
Andi Kleena5516432008-07-23 21:27:41 -07002509 struct hstate *h = hstate_vma(vma);
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01002510 int ret = VM_FAULT_SIGBUS;
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -07002511 pgoff_t idx;
Adam Litke4c887262005-10-29 18:16:46 -07002512 unsigned long size;
Adam Litke4c887262005-10-29 18:16:46 -07002513 struct page *page;
2514 struct address_space *mapping;
David Gibson1e8f8892006-01-06 00:10:44 -08002515 pte_t new_pte;
Adam Litke4c887262005-10-29 18:16:46 -07002516
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002517 /*
2518 * Currently, we are forced to kill the process in the event the
2519 * original mapper has unmapped pages from the child due to a failed
Lucas De Marchi25985ed2011-03-30 22:57:33 -03002520 * COW. Warn that such a situation has occurred as it may not be obvious
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002521 */
2522 if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
2523 printk(KERN_WARNING
2524 "PID %d killed due to inadequate hugepage pool\n",
2525 current->pid);
2526 return ret;
2527 }
2528
Adam Litke4c887262005-10-29 18:16:46 -07002529 mapping = vma->vm_file->f_mapping;
Andi Kleena5516432008-07-23 21:27:41 -07002530 idx = vma_hugecache_offset(h, vma, address);
Adam Litke4c887262005-10-29 18:16:46 -07002531
2532 /*
2533 * Use page lock to guard against racing truncation
2534 * before we get page_table_lock.
2535 */
Christoph Lameter6bda6662006-01-06 00:10:49 -08002536retry:
2537 page = find_lock_page(mapping, idx);
2538 if (!page) {
Andi Kleena5516432008-07-23 21:27:41 -07002539 size = i_size_read(mapping->host) >> huge_page_shift(h);
Hugh Dickinsebed4bf2006-10-28 10:38:43 -07002540 if (idx >= size)
2541 goto out;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002542 page = alloc_huge_page(vma, address, 0);
Adam Litke2fc39ce2007-11-14 16:59:39 -08002543 if (IS_ERR(page)) {
2544 ret = -PTR_ERR(page);
Christoph Lameter6bda6662006-01-06 00:10:49 -08002545 goto out;
2546 }
Andrea Arcangeli47ad8472011-01-13 15:46:47 -08002547 clear_huge_page(page, address, pages_per_huge_page(h));
Nick Piggin0ed361d2008-02-04 22:29:34 -08002548 __SetPageUptodate(page);
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01002549
Mel Gormanf83a2752009-05-28 14:34:40 -07002550 if (vma->vm_flags & VM_MAYSHARE) {
Christoph Lameter6bda6662006-01-06 00:10:49 -08002551 int err;
Ken Chen45c682a2007-11-14 16:59:44 -08002552 struct inode *inode = mapping->host;
Christoph Lameter6bda6662006-01-06 00:10:49 -08002553
2554 err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
2555 if (err) {
2556 put_page(page);
Christoph Lameter6bda6662006-01-06 00:10:49 -08002557 if (err == -EEXIST)
2558 goto retry;
2559 goto out;
2560 }
Ken Chen45c682a2007-11-14 16:59:44 -08002561
2562 spin_lock(&inode->i_lock);
Andi Kleena5516432008-07-23 21:27:41 -07002563 inode->i_blocks += blocks_per_huge_page(h);
Ken Chen45c682a2007-11-14 16:59:44 -08002564 spin_unlock(&inode->i_lock);
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002565 page_dup_rmap(page);
Mel Gorman23be7462010-04-23 13:17:56 -04002566 } else {
Christoph Lameter6bda6662006-01-06 00:10:49 -08002567 lock_page(page);
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002568 if (unlikely(anon_vma_prepare(vma))) {
2569 ret = VM_FAULT_OOM;
2570 goto backout_unlocked;
2571 }
2572 hugepage_add_new_anon_rmap(page, vma, address);
Mel Gorman23be7462010-04-23 13:17:56 -04002573 }
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002574 } else {
Naoya Horiguchi998b4382010-09-08 10:19:32 +09002575 /*
2576 * If memory error occurs between mmap() and fault, some process
2577 * don't have hwpoisoned swap entry for errored virtual address.
2578 * So we need to block hugepage fault by PG_hwpoison bit check.
2579 */
2580 if (unlikely(PageHWPoison(page))) {
Chris Forbes32f84522011-07-25 17:12:14 -07002581 ret = VM_FAULT_HWPOISON |
Andi Kleenaa50d3a2010-10-06 21:45:00 +02002582 VM_FAULT_SET_HINDEX(h - hstates);
Naoya Horiguchi998b4382010-09-08 10:19:32 +09002583 goto backout_unlocked;
2584 }
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002585 page_dup_rmap(page);
Christoph Lameter6bda6662006-01-06 00:10:49 -08002586 }
David Gibson1e8f8892006-01-06 00:10:44 -08002587
Andy Whitcroft57303d82008-08-12 15:08:47 -07002588 /*
2589 * If we are going to COW a private mapping later, we examine the
2590 * pending reservations for this page now. This will ensure that
2591 * any allocations necessary to record that reservation occur outside
2592 * the spinlock.
2593 */
Hugh Dickins788c7df2009-06-23 13:49:05 +01002594 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED))
Andy Whitcroft2b267362008-08-12 15:08:49 -07002595 if (vma_needs_reservation(h, vma, address) < 0) {
2596 ret = VM_FAULT_OOM;
2597 goto backout_unlocked;
2598 }
Andy Whitcroft57303d82008-08-12 15:08:47 -07002599
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01002600 spin_lock(&mm->page_table_lock);
Andi Kleena5516432008-07-23 21:27:41 -07002601 size = i_size_read(mapping->host) >> huge_page_shift(h);
Adam Litke4c887262005-10-29 18:16:46 -07002602 if (idx >= size)
2603 goto backout;
2604
Nick Piggin83c54072007-07-19 01:47:05 -07002605 ret = 0;
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07002606 if (!huge_pte_none(huge_ptep_get(ptep)))
Adam Litke4c887262005-10-29 18:16:46 -07002607 goto backout;
2608
David Gibson1e8f8892006-01-06 00:10:44 -08002609 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
2610 && (vma->vm_flags & VM_SHARED)));
2611 set_huge_pte_at(mm, address, ptep, new_pte);
2612
Hugh Dickins788c7df2009-06-23 13:49:05 +01002613 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
David Gibson1e8f8892006-01-06 00:10:44 -08002614 /* Optimization, do the COW without a second fault */
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002615 ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page);
David Gibson1e8f8892006-01-06 00:10:44 -08002616 }
2617
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01002618 spin_unlock(&mm->page_table_lock);
Adam Litke4c887262005-10-29 18:16:46 -07002619 unlock_page(page);
2620out:
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01002621 return ret;
Adam Litke4c887262005-10-29 18:16:46 -07002622
2623backout:
2624 spin_unlock(&mm->page_table_lock);
Andy Whitcroft2b267362008-08-12 15:08:49 -07002625backout_unlocked:
Adam Litke4c887262005-10-29 18:16:46 -07002626 unlock_page(page);
2627 put_page(page);
2628 goto out;
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01002629}
2630
Adam Litke86e52162006-01-06 00:10:43 -08002631int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
Hugh Dickins788c7df2009-06-23 13:49:05 +01002632 unsigned long address, unsigned int flags)
Adam Litke86e52162006-01-06 00:10:43 -08002633{
2634 pte_t *ptep;
2635 pte_t entry;
David Gibson1e8f8892006-01-06 00:10:44 -08002636 int ret;
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002637 struct page *page = NULL;
Andy Whitcroft57303d82008-08-12 15:08:47 -07002638 struct page *pagecache_page = NULL;
David Gibson3935baa2006-03-22 00:08:53 -08002639 static DEFINE_MUTEX(hugetlb_instantiation_mutex);
Andi Kleena5516432008-07-23 21:27:41 -07002640 struct hstate *h = hstate_vma(vma);
Adam Litke86e52162006-01-06 00:10:43 -08002641
KAMEZAWA Hiroyuki1e16a532012-01-10 15:07:22 -08002642 address &= huge_page_mask(h);
2643
Naoya Horiguchifd6a03e2010-05-28 09:29:21 +09002644 ptep = huge_pte_offset(mm, address);
2645 if (ptep) {
2646 entry = huge_ptep_get(ptep);
Naoya Horiguchi290408d2010-09-08 10:19:35 +09002647 if (unlikely(is_hugetlb_entry_migration(entry))) {
2648 migration_entry_wait(mm, (pmd_t *)ptep, address);
2649 return 0;
2650 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
Chris Forbes32f84522011-07-25 17:12:14 -07002651 return VM_FAULT_HWPOISON_LARGE |
Andi Kleenaa50d3a2010-10-06 21:45:00 +02002652 VM_FAULT_SET_HINDEX(h - hstates);
Naoya Horiguchifd6a03e2010-05-28 09:29:21 +09002653 }
2654
Andi Kleena5516432008-07-23 21:27:41 -07002655 ptep = huge_pte_alloc(mm, address, huge_page_size(h));
Adam Litke86e52162006-01-06 00:10:43 -08002656 if (!ptep)
2657 return VM_FAULT_OOM;
2658
David Gibson3935baa2006-03-22 00:08:53 -08002659 /*
2660 * Serialize hugepage allocation and instantiation, so that we don't
2661 * get spurious allocation failures if two CPUs race to instantiate
2662 * the same page in the page cache.
2663 */
2664 mutex_lock(&hugetlb_instantiation_mutex);
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07002665 entry = huge_ptep_get(ptep);
2666 if (huge_pte_none(entry)) {
Hugh Dickins788c7df2009-06-23 13:49:05 +01002667 ret = hugetlb_no_page(mm, vma, address, ptep, flags);
David Gibsonb4d1d992008-10-15 22:01:11 -07002668 goto out_mutex;
David Gibson3935baa2006-03-22 00:08:53 -08002669 }
Adam Litke86e52162006-01-06 00:10:43 -08002670
Nick Piggin83c54072007-07-19 01:47:05 -07002671 ret = 0;
David Gibson1e8f8892006-01-06 00:10:44 -08002672
Andy Whitcroft57303d82008-08-12 15:08:47 -07002673 /*
2674 * If we are going to COW the mapping later, we examine the pending
2675 * reservations for this page now. This will ensure that any
2676 * allocations necessary to record that reservation occur outside the
2677 * spinlock. For private mappings, we also lookup the pagecache
2678 * page now as it is used to determine if a reservation has been
2679 * consumed.
2680 */
Hugh Dickins788c7df2009-06-23 13:49:05 +01002681 if ((flags & FAULT_FLAG_WRITE) && !pte_write(entry)) {
Andy Whitcroft2b267362008-08-12 15:08:49 -07002682 if (vma_needs_reservation(h, vma, address) < 0) {
2683 ret = VM_FAULT_OOM;
David Gibsonb4d1d992008-10-15 22:01:11 -07002684 goto out_mutex;
Andy Whitcroft2b267362008-08-12 15:08:49 -07002685 }
Andy Whitcroft57303d82008-08-12 15:08:47 -07002686
Mel Gormanf83a2752009-05-28 14:34:40 -07002687 if (!(vma->vm_flags & VM_MAYSHARE))
Andy Whitcroft57303d82008-08-12 15:08:47 -07002688 pagecache_page = hugetlbfs_pagecache_page(h,
2689 vma, address);
2690 }
2691
Naoya Horiguchi56c9cfb2010-09-10 13:23:04 +09002692 /*
2693 * hugetlb_cow() requires page locks of pte_page(entry) and
2694 * pagecache_page, so here we need take the former one
2695 * when page != pagecache_page or !pagecache_page.
2696 * Note that locking order is always pagecache_page -> page,
2697 * so no worry about deadlock.
2698 */
2699 page = pte_page(entry);
2700 if (page != pagecache_page)
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002701 lock_page(page);
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002702
David Gibson1e8f8892006-01-06 00:10:44 -08002703 spin_lock(&mm->page_table_lock);
2704 /* Check for a racing update before calling hugetlb_cow */
David Gibsonb4d1d992008-10-15 22:01:11 -07002705 if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
2706 goto out_page_table_lock;
2707
2708
Hugh Dickins788c7df2009-06-23 13:49:05 +01002709 if (flags & FAULT_FLAG_WRITE) {
David Gibsonb4d1d992008-10-15 22:01:11 -07002710 if (!pte_write(entry)) {
Andy Whitcroft57303d82008-08-12 15:08:47 -07002711 ret = hugetlb_cow(mm, vma, address, ptep, entry,
2712 pagecache_page);
David Gibsonb4d1d992008-10-15 22:01:11 -07002713 goto out_page_table_lock;
2714 }
2715 entry = pte_mkdirty(entry);
2716 }
2717 entry = pte_mkyoung(entry);
Hugh Dickins788c7df2009-06-23 13:49:05 +01002718 if (huge_ptep_set_access_flags(vma, address, ptep, entry,
2719 flags & FAULT_FLAG_WRITE))
Russell King4b3073e2009-12-18 16:40:18 +00002720 update_mmu_cache(vma, address, ptep);
David Gibsonb4d1d992008-10-15 22:01:11 -07002721
2722out_page_table_lock:
David Gibson1e8f8892006-01-06 00:10:44 -08002723 spin_unlock(&mm->page_table_lock);
Andy Whitcroft57303d82008-08-12 15:08:47 -07002724
2725 if (pagecache_page) {
2726 unlock_page(pagecache_page);
2727 put_page(pagecache_page);
2728 }
Dean Nelson1f64d692010-12-02 14:31:12 -08002729 if (page != pagecache_page)
2730 unlock_page(page);
Andy Whitcroft57303d82008-08-12 15:08:47 -07002731
David Gibsonb4d1d992008-10-15 22:01:11 -07002732out_mutex:
David Gibson3935baa2006-03-22 00:08:53 -08002733 mutex_unlock(&hugetlb_instantiation_mutex);
David Gibson1e8f8892006-01-06 00:10:44 -08002734
2735 return ret;
Adam Litke86e52162006-01-06 00:10:43 -08002736}
2737
Andi Kleenceb86872008-07-23 21:27:50 -07002738/* Can be overriden by architectures */
2739__attribute__((weak)) struct page *
2740follow_huge_pud(struct mm_struct *mm, unsigned long address,
2741 pud_t *pud, int write)
2742{
2743 BUG();
2744 return NULL;
2745}
2746
David Gibson63551ae2005-06-21 17:14:44 -07002747int follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
2748 struct page **pages, struct vm_area_struct **vmas,
Adam Litke5b23dbe2007-11-14 16:59:33 -08002749 unsigned long *position, int *length, int i,
Hugh Dickins2a15efc2009-09-21 17:03:27 -07002750 unsigned int flags)
David Gibson63551ae2005-06-21 17:14:44 -07002751{
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08002752 unsigned long pfn_offset;
2753 unsigned long vaddr = *position;
David Gibson63551ae2005-06-21 17:14:44 -07002754 int remainder = *length;
Andi Kleena5516432008-07-23 21:27:41 -07002755 struct hstate *h = hstate_vma(vma);
David Gibson63551ae2005-06-21 17:14:44 -07002756
Hugh Dickins1c598272005-10-19 21:23:43 -07002757 spin_lock(&mm->page_table_lock);
David Gibson63551ae2005-06-21 17:14:44 -07002758 while (vaddr < vma->vm_end && remainder) {
Adam Litke4c887262005-10-29 18:16:46 -07002759 pte_t *pte;
Hugh Dickins2a15efc2009-09-21 17:03:27 -07002760 int absent;
Adam Litke4c887262005-10-29 18:16:46 -07002761 struct page *page;
2762
2763 /*
2764 * Some archs (sparc64, sh*) have multiple pte_ts to
Hugh Dickins2a15efc2009-09-21 17:03:27 -07002765 * each hugepage. We have to make sure we get the
Adam Litke4c887262005-10-29 18:16:46 -07002766 * first, for the page indexing below to work.
2767 */
Andi Kleena5516432008-07-23 21:27:41 -07002768 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h));
Hugh Dickins2a15efc2009-09-21 17:03:27 -07002769 absent = !pte || huge_pte_none(huge_ptep_get(pte));
Adam Litke4c887262005-10-29 18:16:46 -07002770
Hugh Dickins2a15efc2009-09-21 17:03:27 -07002771 /*
2772 * When coredumping, it suits get_dump_page if we just return
Hugh Dickins3ae77f42009-09-21 17:03:33 -07002773 * an error where there's an empty slot with no huge pagecache
2774 * to back it. This way, we avoid allocating a hugepage, and
2775 * the sparse dumpfile avoids allocating disk blocks, but its
2776 * huge holes still show up with zeroes where they need to be.
Hugh Dickins2a15efc2009-09-21 17:03:27 -07002777 */
Hugh Dickins3ae77f42009-09-21 17:03:33 -07002778 if (absent && (flags & FOLL_DUMP) &&
2779 !hugetlbfs_pagecache_present(h, vma, vaddr)) {
Hugh Dickins2a15efc2009-09-21 17:03:27 -07002780 remainder = 0;
2781 break;
2782 }
2783
2784 if (absent ||
2785 ((flags & FOLL_WRITE) && !pte_write(huge_ptep_get(pte)))) {
Adam Litke4c887262005-10-29 18:16:46 -07002786 int ret;
2787
2788 spin_unlock(&mm->page_table_lock);
Hugh Dickins2a15efc2009-09-21 17:03:27 -07002789 ret = hugetlb_fault(mm, vma, vaddr,
2790 (flags & FOLL_WRITE) ? FAULT_FLAG_WRITE : 0);
Adam Litke4c887262005-10-29 18:16:46 -07002791 spin_lock(&mm->page_table_lock);
Adam Litkea89182c2007-08-22 14:01:51 -07002792 if (!(ret & VM_FAULT_ERROR))
Adam Litke4c887262005-10-29 18:16:46 -07002793 continue;
2794
2795 remainder = 0;
Adam Litke4c887262005-10-29 18:16:46 -07002796 break;
2797 }
David Gibson63551ae2005-06-21 17:14:44 -07002798
Andi Kleena5516432008-07-23 21:27:41 -07002799 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07002800 page = pte_page(huge_ptep_get(pte));
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08002801same_page:
Chen, Kenneth Wd6692182006-03-31 02:29:57 -08002802 if (pages) {
Hugh Dickins2a15efc2009-09-21 17:03:27 -07002803 pages[i] = mem_map_offset(page, pfn_offset);
KOSAKI Motohiro4b2e38a2008-10-18 20:27:10 -07002804 get_page(pages[i]);
Chen, Kenneth Wd6692182006-03-31 02:29:57 -08002805 }
David Gibson63551ae2005-06-21 17:14:44 -07002806
2807 if (vmas)
2808 vmas[i] = vma;
2809
2810 vaddr += PAGE_SIZE;
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08002811 ++pfn_offset;
David Gibson63551ae2005-06-21 17:14:44 -07002812 --remainder;
2813 ++i;
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08002814 if (vaddr < vma->vm_end && remainder &&
Andi Kleena5516432008-07-23 21:27:41 -07002815 pfn_offset < pages_per_huge_page(h)) {
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08002816 /*
2817 * We use pfn_offset to avoid touching the pageframes
2818 * of this compound page.
2819 */
2820 goto same_page;
2821 }
David Gibson63551ae2005-06-21 17:14:44 -07002822 }
Hugh Dickins1c598272005-10-19 21:23:43 -07002823 spin_unlock(&mm->page_table_lock);
David Gibson63551ae2005-06-21 17:14:44 -07002824 *length = remainder;
2825 *position = vaddr;
2826
Hugh Dickins2a15efc2009-09-21 17:03:27 -07002827 return i ? i : -EFAULT;
David Gibson63551ae2005-06-21 17:14:44 -07002828}
Zhang, Yanmin8f860592006-03-22 00:08:50 -08002829
2830void hugetlb_change_protection(struct vm_area_struct *vma,
2831 unsigned long address, unsigned long end, pgprot_t newprot)
2832{
2833 struct mm_struct *mm = vma->vm_mm;
2834 unsigned long start = address;
2835 pte_t *ptep;
2836 pte_t pte;
Andi Kleena5516432008-07-23 21:27:41 -07002837 struct hstate *h = hstate_vma(vma);
Zhang, Yanmin8f860592006-03-22 00:08:50 -08002838
2839 BUG_ON(address >= end);
2840 flush_cache_range(vma, address, end);
2841
Peter Zijlstra3d48ae42011-05-24 17:12:06 -07002842 mutex_lock(&vma->vm_file->f_mapping->i_mmap_mutex);
Zhang, Yanmin8f860592006-03-22 00:08:50 -08002843 spin_lock(&mm->page_table_lock);
Andi Kleena5516432008-07-23 21:27:41 -07002844 for (; address < end; address += huge_page_size(h)) {
Zhang, Yanmin8f860592006-03-22 00:08:50 -08002845 ptep = huge_pte_offset(mm, address);
2846 if (!ptep)
2847 continue;
Chen, Kenneth W39dde652006-12-06 20:32:03 -08002848 if (huge_pmd_unshare(mm, &address, ptep))
2849 continue;
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07002850 if (!huge_pte_none(huge_ptep_get(ptep))) {
Zhang, Yanmin8f860592006-03-22 00:08:50 -08002851 pte = huge_ptep_get_and_clear(mm, address, ptep);
2852 pte = pte_mkhuge(pte_modify(pte, newprot));
2853 set_huge_pte_at(mm, address, ptep, pte);
Zhang, Yanmin8f860592006-03-22 00:08:50 -08002854 }
2855 }
2856 spin_unlock(&mm->page_table_lock);
Peter Zijlstra3d48ae42011-05-24 17:12:06 -07002857 mutex_unlock(&vma->vm_file->f_mapping->i_mmap_mutex);
Zhang, Yanmin8f860592006-03-22 00:08:50 -08002858
2859 flush_tlb_range(vma, start, end);
2860}
2861
Mel Gormana1e78772008-07-23 21:27:23 -07002862int hugetlb_reserve_pages(struct inode *inode,
2863 long from, long to,
Mel Gorman5a6fe122009-02-10 14:02:27 +00002864 struct vm_area_struct *vma,
KOSAKI Motohiroca16d142011-05-26 19:16:19 +09002865 vm_flags_t vm_flags)
Adam Litkee4e574b2007-10-16 01:26:19 -07002866{
Mel Gorman17c9d122009-02-11 16:34:16 +00002867 long ret, chg;
Andi Kleena5516432008-07-23 21:27:41 -07002868 struct hstate *h = hstate_inode(inode);
Adam Litkee4e574b2007-10-16 01:26:19 -07002869
Mel Gormana1e78772008-07-23 21:27:23 -07002870 /*
Mel Gorman17c9d122009-02-11 16:34:16 +00002871 * Only apply hugepage reservation if asked. At fault time, an
2872 * attempt will be made for VM_NORESERVE to allocate a page
2873 * and filesystem quota without using reserves
2874 */
KOSAKI Motohiroca16d142011-05-26 19:16:19 +09002875 if (vm_flags & VM_NORESERVE)
Mel Gorman17c9d122009-02-11 16:34:16 +00002876 return 0;
2877
2878 /*
Mel Gormana1e78772008-07-23 21:27:23 -07002879 * Shared mappings base their reservation on the number of pages that
2880 * are already allocated on behalf of the file. Private mappings need
2881 * to reserve the full area even if read-only as mprotect() may be
2882 * called to make the mapping read-write. Assume !vma is a shm mapping
2883 */
Mel Gormanf83a2752009-05-28 14:34:40 -07002884 if (!vma || vma->vm_flags & VM_MAYSHARE)
Mel Gormana1e78772008-07-23 21:27:23 -07002885 chg = region_chg(&inode->i_mapping->private_list, from, to);
Mel Gorman5a6fe122009-02-10 14:02:27 +00002886 else {
2887 struct resv_map *resv_map = resv_map_alloc();
Mel Gorman5a6fe122009-02-10 14:02:27 +00002888 if (!resv_map)
2889 return -ENOMEM;
2890
Mel Gorman17c9d122009-02-11 16:34:16 +00002891 chg = to - from;
2892
Mel Gorman5a6fe122009-02-10 14:02:27 +00002893 set_vma_resv_map(vma, resv_map);
2894 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
2895 }
2896
Mel Gorman17c9d122009-02-11 16:34:16 +00002897 if (chg < 0)
2898 return chg;
2899
2900 /* There must be enough filesystem quota for the mapping */
2901 if (hugetlb_get_quota(inode->i_mapping, chg))
2902 return -ENOSPC;
2903
2904 /*
2905 * Check enough hugepages are available for the reservation.
2906 * Hand back the quota if there are not
2907 */
2908 ret = hugetlb_acct_memory(h, chg);
2909 if (ret < 0) {
2910 hugetlb_put_quota(inode->i_mapping, chg);
2911 return ret;
2912 }
2913
2914 /*
2915 * Account for the reservations made. Shared mappings record regions
2916 * that have reservations as they are shared by multiple VMAs.
2917 * When the last VMA disappears, the region map says how much
2918 * the reservation was and the page cache tells how much of
2919 * the reservation was consumed. Private mappings are per-VMA and
2920 * only the consumed reservations are tracked. When the VMA
2921 * disappears, the original reservation is the VMA size and the
2922 * consumed reservations are stored in the map. Hence, nothing
2923 * else has to be done for private mappings here
2924 */
Mel Gormanf83a2752009-05-28 14:34:40 -07002925 if (!vma || vma->vm_flags & VM_MAYSHARE)
Mel Gorman17c9d122009-02-11 16:34:16 +00002926 region_add(&inode->i_mapping->private_list, from, to);
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -07002927 return 0;
2928}
2929
2930void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed)
2931{
Andi Kleena5516432008-07-23 21:27:41 -07002932 struct hstate *h = hstate_inode(inode);
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -07002933 long chg = region_truncate(&inode->i_mapping->private_list, offset);
Ken Chen45c682a2007-11-14 16:59:44 -08002934
2935 spin_lock(&inode->i_lock);
Eric Sandeene4c6f8b2009-07-29 15:02:16 -07002936 inode->i_blocks -= (blocks_per_huge_page(h) * freed);
Ken Chen45c682a2007-11-14 16:59:44 -08002937 spin_unlock(&inode->i_lock);
2938
Adam Litke90d8b7e2007-11-14 16:59:42 -08002939 hugetlb_put_quota(inode->i_mapping, (chg - freed));
Andi Kleena5516432008-07-23 21:27:41 -07002940 hugetlb_acct_memory(h, -(chg - freed));
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -07002941}
Naoya Horiguchi93f70f92010-05-28 09:29:20 +09002942
Andi Kleend5bd9102010-09-27 09:00:12 +02002943#ifdef CONFIG_MEMORY_FAILURE
2944
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +09002945/* Should be called in hugetlb_lock */
2946static int is_hugepage_on_freelist(struct page *hpage)
2947{
2948 struct page *page;
2949 struct page *tmp;
2950 struct hstate *h = page_hstate(hpage);
2951 int nid = page_to_nid(hpage);
2952
2953 list_for_each_entry_safe(page, tmp, &h->hugepage_freelists[nid], lru)
2954 if (page == hpage)
2955 return 1;
2956 return 0;
2957}
2958
Naoya Horiguchi93f70f92010-05-28 09:29:20 +09002959/*
2960 * This function is called from memory failure code.
2961 * Assume the caller holds page lock of the head page.
2962 */
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +09002963int dequeue_hwpoisoned_huge_page(struct page *hpage)
Naoya Horiguchi93f70f92010-05-28 09:29:20 +09002964{
2965 struct hstate *h = page_hstate(hpage);
2966 int nid = page_to_nid(hpage);
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +09002967 int ret = -EBUSY;
Naoya Horiguchi93f70f92010-05-28 09:29:20 +09002968
2969 spin_lock(&hugetlb_lock);
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +09002970 if (is_hugepage_on_freelist(hpage)) {
2971 list_del(&hpage->lru);
Naoya Horiguchi8c6c2ec2010-09-08 10:19:38 +09002972 set_page_refcounted(hpage);
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +09002973 h->free_huge_pages--;
2974 h->free_huge_pages_node[nid]--;
2975 ret = 0;
2976 }
Naoya Horiguchi93f70f92010-05-28 09:29:20 +09002977 spin_unlock(&hugetlb_lock);
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +09002978 return ret;
Naoya Horiguchi93f70f92010-05-28 09:29:20 +09002979}
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +09002980#endif