blob: 85855240933d7cf195ce1548faa75d8120c21478 [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>
Adrian Bunk78a34ae2008-07-28 15:46:30 -070027#include <asm/io.h>
David Gibson63551ae2005-06-21 17:14:44 -070028
29#include <linux/hugetlb.h>
Lee Schermerhorn9a305232009-12-14 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 *
65 * down_write(&mm->mmap_sem);
66 * or
67 * 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
149 /* We overlap with this area, if it extends futher than
150 * 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
Andy Whitcroft69d177c2008-11-06 12:53:26 -0800397static void clear_gigantic_page(struct page *page,
398 unsigned long addr, unsigned long sz)
399{
400 int i;
401 struct page *p = page;
402
403 might_sleep();
404 for (i = 0; i < sz/PAGE_SIZE; i++, p = mem_map_next(p, page, i)) {
405 cond_resched();
406 clear_user_highpage(p, addr + i * PAGE_SIZE);
407 }
408}
Andi Kleena5516432008-07-23 21:27:41 -0700409static void clear_huge_page(struct page *page,
410 unsigned long addr, unsigned long sz)
David Gibson79ac6ba2006-03-22 00:08:51 -0800411{
412 int i;
413
Andrea Arcangeli74dbdd22010-01-08 14:43:05 -0800414 if (unlikely(sz/PAGE_SIZE > MAX_ORDER_NR_PAGES)) {
Hannes Ederebdd4ae2009-01-06 14:39:58 -0800415 clear_gigantic_page(page, addr, sz);
416 return;
417 }
Andy Whitcroft69d177c2008-11-06 12:53:26 -0800418
David Gibson79ac6ba2006-03-22 00:08:51 -0800419 might_sleep();
Andi Kleena5516432008-07-23 21:27:41 -0700420 for (i = 0; i < sz/PAGE_SIZE; i++) {
David Gibson79ac6ba2006-03-22 00:08:51 -0800421 cond_resched();
Ralf Baechle281e0e32007-10-01 01:20:10 -0700422 clear_user_highpage(page + i, addr + i * PAGE_SIZE);
David Gibson79ac6ba2006-03-22 00:08:51 -0800423 }
424}
425
Naoya Horiguchi0ebabb42010-09-08 10:19:34 +0900426static void copy_user_gigantic_page(struct page *dst, struct page *src,
Andy Whitcroft69d177c2008-11-06 12:53:26 -0800427 unsigned long addr, struct vm_area_struct *vma)
428{
429 int i;
430 struct hstate *h = hstate_vma(vma);
431 struct page *dst_base = dst;
432 struct page *src_base = src;
Naoya Horiguchi0ebabb42010-09-08 10:19:34 +0900433
Andy Whitcroft69d177c2008-11-06 12:53:26 -0800434 for (i = 0; i < pages_per_huge_page(h); ) {
435 cond_resched();
436 copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
437
438 i++;
439 dst = mem_map_next(dst, dst_base, i);
440 src = mem_map_next(src, src_base, i);
441 }
442}
Naoya Horiguchi0ebabb42010-09-08 10:19:34 +0900443
444static void copy_user_huge_page(struct page *dst, struct page *src,
Atsushi Nemoto9de455b2006-12-12 17:14:55 +0000445 unsigned long addr, struct vm_area_struct *vma)
David Gibson79ac6ba2006-03-22 00:08:51 -0800446{
447 int i;
Andi Kleena5516432008-07-23 21:27:41 -0700448 struct hstate *h = hstate_vma(vma);
David Gibson79ac6ba2006-03-22 00:08:51 -0800449
Hannes Ederebdd4ae2009-01-06 14:39:58 -0800450 if (unlikely(pages_per_huge_page(h) > MAX_ORDER_NR_PAGES)) {
Naoya Horiguchi0ebabb42010-09-08 10:19:34 +0900451 copy_user_gigantic_page(dst, src, addr, vma);
Hannes Ederebdd4ae2009-01-06 14:39:58 -0800452 return;
453 }
Andy Whitcroft69d177c2008-11-06 12:53:26 -0800454
David Gibson79ac6ba2006-03-22 00:08:51 -0800455 might_sleep();
Andi Kleena5516432008-07-23 21:27:41 -0700456 for (i = 0; i < pages_per_huge_page(h); i++) {
David Gibson79ac6ba2006-03-22 00:08:51 -0800457 cond_resched();
Atsushi Nemoto9de455b2006-12-12 17:14:55 +0000458 copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
David Gibson79ac6ba2006-03-22 00:08:51 -0800459 }
460}
461
Naoya Horiguchi0ebabb42010-09-08 10:19:34 +0900462static void copy_gigantic_page(struct page *dst, struct page *src)
463{
464 int i;
465 struct hstate *h = page_hstate(src);
466 struct page *dst_base = dst;
467 struct page *src_base = src;
468
469 for (i = 0; i < pages_per_huge_page(h); ) {
470 cond_resched();
471 copy_highpage(dst, src);
472
473 i++;
474 dst = mem_map_next(dst, dst_base, i);
475 src = mem_map_next(src, src_base, i);
476 }
477}
478
479void copy_huge_page(struct page *dst, struct page *src)
480{
481 int i;
482 struct hstate *h = page_hstate(src);
483
484 if (unlikely(pages_per_huge_page(h) > MAX_ORDER_NR_PAGES)) {
485 copy_gigantic_page(dst, src);
486 return;
487 }
488
489 might_sleep();
490 for (i = 0; i < pages_per_huge_page(h); i++) {
491 cond_resched();
492 copy_highpage(dst + i, src + i);
493 }
494}
495
Andi Kleena5516432008-07-23 21:27:41 -0700496static void enqueue_huge_page(struct hstate *h, struct page *page)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700497{
498 int nid = page_to_nid(page);
Andi Kleena5516432008-07-23 21:27:41 -0700499 list_add(&page->lru, &h->hugepage_freelists[nid]);
500 h->free_huge_pages++;
501 h->free_huge_pages_node[nid]++;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700502}
503
Naoya Horiguchibf50bab2010-09-08 10:19:33 +0900504static struct page *dequeue_huge_page_node(struct hstate *h, int nid)
505{
506 struct page *page;
507
508 if (list_empty(&h->hugepage_freelists[nid]))
509 return NULL;
510 page = list_entry(h->hugepage_freelists[nid].next, struct page, lru);
511 list_del(&page->lru);
Naoya Horiguchia9869b82010-09-08 10:19:37 +0900512 set_page_refcounted(page);
Naoya Horiguchibf50bab2010-09-08 10:19:33 +0900513 h->free_huge_pages--;
514 h->free_huge_pages_node[nid]--;
515 return page;
516}
517
Andi Kleena5516432008-07-23 21:27:41 -0700518static struct page *dequeue_huge_page_vma(struct hstate *h,
519 struct vm_area_struct *vma,
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700520 unsigned long address, int avoid_reserve)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700521{
Linus Torvalds1da177e2005-04-16 15:20:36 -0700522 struct page *page = NULL;
Lee Schermerhorn480eccf2007-09-18 22:46:47 -0700523 struct mempolicy *mpol;
Mel Gorman19770b32008-04-28 02:12:18 -0700524 nodemask_t *nodemask;
Miao Xiec0ff7452010-05-24 14:32:08 -0700525 struct zonelist *zonelist;
Mel Gormandd1a2392008-04-28 02:12:17 -0700526 struct zone *zone;
527 struct zoneref *z;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700528
Miao Xiec0ff7452010-05-24 14:32:08 -0700529 get_mems_allowed();
530 zonelist = huge_zonelist(vma, address,
531 htlb_alloc_mask, &mpol, &nodemask);
Mel Gormana1e78772008-07-23 21:27:23 -0700532 /*
533 * A child process with MAP_PRIVATE mappings created by their parent
534 * have no page reserves. This check ensures that reservations are
535 * not "stolen". The child may still get SIGKILLed
536 */
Mel Gorman7f09ca52008-07-23 21:27:58 -0700537 if (!vma_has_reserves(vma) &&
Andi Kleena5516432008-07-23 21:27:41 -0700538 h->free_huge_pages - h->resv_huge_pages == 0)
Miao Xiec0ff7452010-05-24 14:32:08 -0700539 goto err;
Mel Gormana1e78772008-07-23 21:27:23 -0700540
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700541 /* If reserves cannot be used, ensure enough pages are in the pool */
Andi Kleena5516432008-07-23 21:27:41 -0700542 if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
Miao Xiec0ff7452010-05-24 14:32:08 -0700543 goto err;;
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700544
Mel Gorman19770b32008-04-28 02:12:18 -0700545 for_each_zone_zonelist_nodemask(zone, z, zonelist,
546 MAX_NR_ZONES - 1, nodemask) {
Naoya Horiguchibf50bab2010-09-08 10:19:33 +0900547 if (cpuset_zone_allowed_softwall(zone, htlb_alloc_mask)) {
548 page = dequeue_huge_page_node(h, zone_to_nid(zone));
549 if (page) {
550 if (!avoid_reserve)
551 decrement_hugepage_resv_vma(h, vma);
552 break;
553 }
Andrew Morton3abf7af2007-07-19 01:49:08 -0700554 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700555 }
Miao Xiec0ff7452010-05-24 14:32:08 -0700556err:
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -0700557 mpol_cond_put(mpol);
Miao Xiec0ff7452010-05-24 14:32:08 -0700558 put_mems_allowed();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700559 return page;
560}
561
Andi Kleena5516432008-07-23 21:27:41 -0700562static void update_and_free_page(struct hstate *h, struct page *page)
Adam Litke6af2acb2007-10-16 01:26:16 -0700563{
564 int i;
Andi Kleena5516432008-07-23 21:27:41 -0700565
Andy Whitcroft18229df2008-11-06 12:53:27 -0800566 VM_BUG_ON(h->order >= MAX_ORDER);
567
Andi Kleena5516432008-07-23 21:27:41 -0700568 h->nr_huge_pages--;
569 h->nr_huge_pages_node[page_to_nid(page)]--;
570 for (i = 0; i < pages_per_huge_page(h); i++) {
Adam Litke6af2acb2007-10-16 01:26:16 -0700571 page[i].flags &= ~(1 << PG_locked | 1 << PG_error | 1 << PG_referenced |
572 1 << PG_dirty | 1 << PG_active | 1 << PG_reserved |
573 1 << PG_private | 1<< PG_writeback);
574 }
575 set_compound_page_dtor(page, NULL);
576 set_page_refcounted(page);
Gerald Schaefer7f2e9522008-04-28 02:13:29 -0700577 arch_release_hugepage(page);
Andi Kleena5516432008-07-23 21:27:41 -0700578 __free_pages(page, huge_page_order(h));
Adam Litke6af2acb2007-10-16 01:26:16 -0700579}
580
Andi Kleene5ff2152008-07-23 21:27:42 -0700581struct hstate *size_to_hstate(unsigned long size)
582{
583 struct hstate *h;
584
585 for_each_hstate(h) {
586 if (huge_page_size(h) == size)
587 return h;
588 }
589 return NULL;
590}
591
David Gibson27a85ef2006-03-22 00:08:56 -0800592static void free_huge_page(struct page *page)
593{
Andi Kleena5516432008-07-23 21:27:41 -0700594 /*
595 * Can't pass hstate in here because it is called from the
596 * compound page destructor.
597 */
Andi Kleene5ff2152008-07-23 21:27:42 -0700598 struct hstate *h = page_hstate(page);
Adam Litke7893d1d2007-10-16 01:26:18 -0700599 int nid = page_to_nid(page);
Adam Litkec79fb752007-11-14 16:59:38 -0800600 struct address_space *mapping;
David Gibson27a85ef2006-03-22 00:08:56 -0800601
Adam Litkec79fb752007-11-14 16:59:38 -0800602 mapping = (struct address_space *) page_private(page);
Andy Whitcrofte5df70a2008-02-23 15:23:32 -0800603 set_page_private(page, 0);
Mel Gorman23be7462010-04-23 13:17:56 -0400604 page->mapping = NULL;
Adam Litke7893d1d2007-10-16 01:26:18 -0700605 BUG_ON(page_count(page));
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +0900606 BUG_ON(page_mapcount(page));
David Gibson27a85ef2006-03-22 00:08:56 -0800607 INIT_LIST_HEAD(&page->lru);
608
609 spin_lock(&hugetlb_lock);
Andi Kleenaa888a72008-07-23 21:27:47 -0700610 if (h->surplus_huge_pages_node[nid] && huge_page_order(h) < MAX_ORDER) {
Andi Kleena5516432008-07-23 21:27:41 -0700611 update_and_free_page(h, page);
612 h->surplus_huge_pages--;
613 h->surplus_huge_pages_node[nid]--;
Adam Litke7893d1d2007-10-16 01:26:18 -0700614 } else {
Andi Kleena5516432008-07-23 21:27:41 -0700615 enqueue_huge_page(h, page);
Adam Litke7893d1d2007-10-16 01:26:18 -0700616 }
David Gibson27a85ef2006-03-22 00:08:56 -0800617 spin_unlock(&hugetlb_lock);
Adam Litkec79fb752007-11-14 16:59:38 -0800618 if (mapping)
Adam Litke9a119c02007-11-14 16:59:41 -0800619 hugetlb_put_quota(mapping, 1);
David Gibson27a85ef2006-03-22 00:08:56 -0800620}
621
Andi Kleena5516432008-07-23 21:27:41 -0700622static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
Andi Kleenb7ba30c2008-07-23 21:27:40 -0700623{
624 set_compound_page_dtor(page, free_huge_page);
625 spin_lock(&hugetlb_lock);
Andi Kleena5516432008-07-23 21:27:41 -0700626 h->nr_huge_pages++;
627 h->nr_huge_pages_node[nid]++;
Andi Kleenb7ba30c2008-07-23 21:27:40 -0700628 spin_unlock(&hugetlb_lock);
629 put_page(page); /* free it into the hugepage allocator */
630}
631
Wu Fengguang20a03072009-06-16 15:32:22 -0700632static void prep_compound_gigantic_page(struct page *page, unsigned long order)
633{
634 int i;
635 int nr_pages = 1 << order;
636 struct page *p = page + 1;
637
638 /* we rely on prep_new_huge_page to set the destructor */
639 set_compound_order(page, order);
640 __SetPageHead(page);
641 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
642 __SetPageTail(p);
643 p->first_page = page;
644 }
645}
646
647int PageHuge(struct page *page)
648{
649 compound_page_dtor *dtor;
650
651 if (!PageCompound(page))
652 return 0;
653
654 page = compound_head(page);
655 dtor = get_compound_page_dtor(page);
656
657 return dtor == free_huge_page;
658}
659
Naoya Horiguchi43131e12010-05-28 09:29:22 +0900660EXPORT_SYMBOL_GPL(PageHuge);
661
Andi Kleena5516432008-07-23 21:27:41 -0700662static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700663{
Linus Torvalds1da177e2005-04-16 15:20:36 -0700664 struct page *page;
Joe Jinf96efd52007-07-15 23:38:12 -0700665
Andi Kleenaa888a72008-07-23 21:27:47 -0700666 if (h->order >= MAX_ORDER)
667 return NULL;
668
Mel Gorman6484eb32009-06-16 15:31:54 -0700669 page = alloc_pages_exact_node(nid,
Nishanth Aravamudan551883a2008-04-29 00:58:26 -0700670 htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE|
671 __GFP_REPEAT|__GFP_NOWARN,
Andi Kleena5516432008-07-23 21:27:41 -0700672 huge_page_order(h));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700673 if (page) {
Gerald Schaefer7f2e9522008-04-28 02:13:29 -0700674 if (arch_prepare_hugepage(page)) {
Gerald Schaefercaff3a22008-08-12 15:08:38 -0700675 __free_pages(page, huge_page_order(h));
Harvey Harrison7b8ee842008-04-28 14:13:19 -0700676 return NULL;
Gerald Schaefer7f2e9522008-04-28 02:13:29 -0700677 }
Andi Kleena5516432008-07-23 21:27:41 -0700678 prep_new_huge_page(h, page, nid);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700679 }
Nishanth Aravamudan63b46132007-10-16 01:26:24 -0700680
681 return page;
682}
683
Andi Kleen5ced66c2008-07-23 21:27:45 -0700684/*
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800685 * common helper functions for hstate_next_node_to_{alloc|free}.
686 * We may have allocated or freed a huge page based on a different
687 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
688 * be outside of *nodes_allowed. Ensure that we use an allowed
689 * node for alloc or free.
Lee Schermerhorn9a76db02009-12-14 17:58:15 -0800690 */
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800691static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
Lee Schermerhorn9a76db02009-12-14 17:58:15 -0800692{
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800693 nid = next_node(nid, *nodes_allowed);
Lee Schermerhorn9a76db02009-12-14 17:58:15 -0800694 if (nid == MAX_NUMNODES)
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800695 nid = first_node(*nodes_allowed);
Lee Schermerhorn9a76db02009-12-14 17:58:15 -0800696 VM_BUG_ON(nid >= MAX_NUMNODES);
697
698 return nid;
699}
700
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800701static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
Andi Kleen5ced66c2008-07-23 21:27:45 -0700702{
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800703 if (!node_isset(nid, *nodes_allowed))
704 nid = next_node_allowed(nid, nodes_allowed);
Lee Schermerhorn9a76db02009-12-14 17:58:15 -0800705 return nid;
Andi Kleen5ced66c2008-07-23 21:27:45 -0700706}
707
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800708/*
709 * returns the previously saved node ["this node"] from which to
710 * allocate a persistent huge page for the pool and advance the
711 * next node from which to allocate, handling wrap at end of node
712 * mask.
713 */
714static int hstate_next_node_to_alloc(struct hstate *h,
715 nodemask_t *nodes_allowed)
716{
717 int nid;
718
719 VM_BUG_ON(!nodes_allowed);
720
721 nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
722 h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
723
724 return nid;
725}
726
727static int alloc_fresh_huge_page(struct hstate *h, nodemask_t *nodes_allowed)
Nishanth Aravamudan63b46132007-10-16 01:26:24 -0700728{
729 struct page *page;
730 int start_nid;
731 int next_nid;
732 int ret = 0;
733
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800734 start_nid = hstate_next_node_to_alloc(h, nodes_allowed);
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700735 next_nid = start_nid;
Nishanth Aravamudan63b46132007-10-16 01:26:24 -0700736
737 do {
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700738 page = alloc_fresh_huge_page_node(h, next_nid);
Lee Schermerhorn9a76db02009-12-14 17:58:15 -0800739 if (page) {
Nishanth Aravamudan63b46132007-10-16 01:26:24 -0700740 ret = 1;
Lee Schermerhorn9a76db02009-12-14 17:58:15 -0800741 break;
742 }
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800743 next_nid = hstate_next_node_to_alloc(h, nodes_allowed);
Lee Schermerhorn9a76db02009-12-14 17:58:15 -0800744 } while (next_nid != start_nid);
Nishanth Aravamudan63b46132007-10-16 01:26:24 -0700745
Adam Litke3b116302008-04-28 02:13:06 -0700746 if (ret)
747 count_vm_event(HTLB_BUDDY_PGALLOC);
748 else
749 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
750
Nishanth Aravamudan63b46132007-10-16 01:26:24 -0700751 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700752}
753
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700754/*
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800755 * helper for free_pool_huge_page() - return the previously saved
756 * node ["this node"] from which to free a huge page. Advance the
757 * next node id whether or not we find a free huge page to free so
758 * that the next attempt to free addresses the next node.
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700759 */
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800760static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700761{
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800762 int nid;
Lee Schermerhorn9a76db02009-12-14 17:58:15 -0800763
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800764 VM_BUG_ON(!nodes_allowed);
765
766 nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
767 h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
768
Lee Schermerhorn9a76db02009-12-14 17:58:15 -0800769 return nid;
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700770}
771
772/*
773 * Free huge page from pool from next node to free.
774 * Attempt to keep persistent huge pages more or less
775 * balanced over allowed nodes.
776 * Called with hugetlb_lock locked.
777 */
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800778static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
779 bool acct_surplus)
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700780{
781 int start_nid;
782 int next_nid;
783 int ret = 0;
784
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800785 start_nid = hstate_next_node_to_free(h, nodes_allowed);
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700786 next_nid = start_nid;
787
788 do {
Lee Schermerhorn685f3452009-09-21 17:01:23 -0700789 /*
790 * If we're returning unused surplus pages, only examine
791 * nodes with surplus pages.
792 */
793 if ((!acct_surplus || h->surplus_huge_pages_node[next_nid]) &&
794 !list_empty(&h->hugepage_freelists[next_nid])) {
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700795 struct page *page =
796 list_entry(h->hugepage_freelists[next_nid].next,
797 struct page, lru);
798 list_del(&page->lru);
799 h->free_huge_pages--;
800 h->free_huge_pages_node[next_nid]--;
Lee Schermerhorn685f3452009-09-21 17:01:23 -0700801 if (acct_surplus) {
802 h->surplus_huge_pages--;
803 h->surplus_huge_pages_node[next_nid]--;
804 }
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700805 update_and_free_page(h, page);
806 ret = 1;
Lee Schermerhorn9a76db02009-12-14 17:58:15 -0800807 break;
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700808 }
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800809 next_nid = hstate_next_node_to_free(h, nodes_allowed);
Lee Schermerhorn9a76db02009-12-14 17:58:15 -0800810 } while (next_nid != start_nid);
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700811
812 return ret;
813}
814
Naoya Horiguchibf50bab2010-09-08 10:19:33 +0900815static struct page *alloc_buddy_huge_page(struct hstate *h, int nid)
Adam Litke7893d1d2007-10-16 01:26:18 -0700816{
817 struct page *page;
Naoya Horiguchibf50bab2010-09-08 10:19:33 +0900818 unsigned int r_nid;
Adam Litke7893d1d2007-10-16 01:26:18 -0700819
Andi Kleenaa888a72008-07-23 21:27:47 -0700820 if (h->order >= MAX_ORDER)
821 return NULL;
822
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800823 /*
824 * Assume we will successfully allocate the surplus page to
825 * prevent racing processes from causing the surplus to exceed
826 * overcommit
827 *
828 * This however introduces a different race, where a process B
829 * tries to grow the static hugepage pool while alloc_pages() is
830 * called by process A. B will only examine the per-node
831 * counters in determining if surplus huge pages can be
832 * converted to normal huge pages in adjust_pool_surplus(). A
833 * won't be able to increment the per-node counter, until the
834 * lock is dropped by B, but B doesn't drop hugetlb_lock until
835 * no more huge pages can be converted from surplus to normal
836 * state (and doesn't try to convert again). Thus, we have a
837 * case where a surplus huge page exists, the pool is grown, and
838 * the surplus huge page still exists after, even though it
839 * should just have been converted to a normal huge page. This
840 * does not leak memory, though, as the hugepage will be freed
841 * once it is out of use. It also does not allow the counters to
842 * go out of whack in adjust_pool_surplus() as we don't modify
843 * the node values until we've gotten the hugepage and only the
844 * per-node value is checked there.
845 */
846 spin_lock(&hugetlb_lock);
Andi Kleena5516432008-07-23 21:27:41 -0700847 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800848 spin_unlock(&hugetlb_lock);
849 return NULL;
850 } else {
Andi Kleena5516432008-07-23 21:27:41 -0700851 h->nr_huge_pages++;
852 h->surplus_huge_pages++;
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800853 }
854 spin_unlock(&hugetlb_lock);
855
Naoya Horiguchibf50bab2010-09-08 10:19:33 +0900856 if (nid == NUMA_NO_NODE)
857 page = alloc_pages(htlb_alloc_mask|__GFP_COMP|
858 __GFP_REPEAT|__GFP_NOWARN,
859 huge_page_order(h));
860 else
861 page = alloc_pages_exact_node(nid,
862 htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE|
863 __GFP_REPEAT|__GFP_NOWARN, huge_page_order(h));
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800864
Gerald Schaefercaff3a22008-08-12 15:08:38 -0700865 if (page && arch_prepare_hugepage(page)) {
866 __free_pages(page, huge_page_order(h));
867 return NULL;
868 }
869
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800870 spin_lock(&hugetlb_lock);
Adam Litke7893d1d2007-10-16 01:26:18 -0700871 if (page) {
Naoya Horiguchibf50bab2010-09-08 10:19:33 +0900872 r_nid = page_to_nid(page);
Adam Litke7893d1d2007-10-16 01:26:18 -0700873 set_compound_page_dtor(page, free_huge_page);
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800874 /*
875 * We incremented the global counters already
876 */
Naoya Horiguchibf50bab2010-09-08 10:19:33 +0900877 h->nr_huge_pages_node[r_nid]++;
878 h->surplus_huge_pages_node[r_nid]++;
Adam Litke3b116302008-04-28 02:13:06 -0700879 __count_vm_event(HTLB_BUDDY_PGALLOC);
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800880 } else {
Andi Kleena5516432008-07-23 21:27:41 -0700881 h->nr_huge_pages--;
882 h->surplus_huge_pages--;
Adam Litke3b116302008-04-28 02:13:06 -0700883 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
Adam Litke7893d1d2007-10-16 01:26:18 -0700884 }
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800885 spin_unlock(&hugetlb_lock);
Adam Litke7893d1d2007-10-16 01:26:18 -0700886
887 return page;
888}
889
Adam Litkee4e574b2007-10-16 01:26:19 -0700890/*
Naoya Horiguchibf50bab2010-09-08 10:19:33 +0900891 * This allocation function is useful in the context where vma is irrelevant.
892 * E.g. soft-offlining uses this function because it only cares physical
893 * address of error page.
894 */
895struct page *alloc_huge_page_node(struct hstate *h, int nid)
896{
897 struct page *page;
898
899 spin_lock(&hugetlb_lock);
900 page = dequeue_huge_page_node(h, nid);
901 spin_unlock(&hugetlb_lock);
902
903 if (!page)
904 page = alloc_buddy_huge_page(h, nid);
905
906 return page;
907}
908
909/*
Adam Litkee4e574b2007-10-16 01:26:19 -0700910 * Increase the hugetlb pool such that it can accomodate a reservation
911 * of size 'delta'.
912 */
Andi Kleena5516432008-07-23 21:27:41 -0700913static int gather_surplus_pages(struct hstate *h, int delta)
Adam Litkee4e574b2007-10-16 01:26:19 -0700914{
915 struct list_head surplus_list;
916 struct page *page, *tmp;
917 int ret, i;
918 int needed, allocated;
919
Andi Kleena5516432008-07-23 21:27:41 -0700920 needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
Adam Litkeac09b3a2008-03-04 14:29:38 -0800921 if (needed <= 0) {
Andi Kleena5516432008-07-23 21:27:41 -0700922 h->resv_huge_pages += delta;
Adam Litkee4e574b2007-10-16 01:26:19 -0700923 return 0;
Adam Litkeac09b3a2008-03-04 14:29:38 -0800924 }
Adam Litkee4e574b2007-10-16 01:26:19 -0700925
926 allocated = 0;
927 INIT_LIST_HEAD(&surplus_list);
928
929 ret = -ENOMEM;
930retry:
931 spin_unlock(&hugetlb_lock);
932 for (i = 0; i < needed; i++) {
Naoya Horiguchibf50bab2010-09-08 10:19:33 +0900933 page = alloc_buddy_huge_page(h, NUMA_NO_NODE);
Naoya Horiguchia9869b82010-09-08 10:19:37 +0900934 if (!page)
Adam Litkee4e574b2007-10-16 01:26:19 -0700935 /*
936 * We were not able to allocate enough pages to
937 * satisfy the entire reservation so we free what
938 * we've allocated so far.
939 */
Adam Litkee4e574b2007-10-16 01:26:19 -0700940 goto free;
Adam Litkee4e574b2007-10-16 01:26:19 -0700941
942 list_add(&page->lru, &surplus_list);
943 }
944 allocated += needed;
945
946 /*
947 * After retaking hugetlb_lock, we need to recalculate 'needed'
948 * because either resv_huge_pages or free_huge_pages may have changed.
949 */
950 spin_lock(&hugetlb_lock);
Andi Kleena5516432008-07-23 21:27:41 -0700951 needed = (h->resv_huge_pages + delta) -
952 (h->free_huge_pages + allocated);
Adam Litkee4e574b2007-10-16 01:26:19 -0700953 if (needed > 0)
954 goto retry;
955
956 /*
957 * The surplus_list now contains _at_least_ the number of extra pages
958 * needed to accomodate the reservation. Add the appropriate number
959 * of pages to the hugetlb pool and free the extras back to the buddy
Adam Litkeac09b3a2008-03-04 14:29:38 -0800960 * allocator. Commit the entire reservation here to prevent another
961 * process from stealing the pages as they are added to the pool but
962 * before they are reserved.
Adam Litkee4e574b2007-10-16 01:26:19 -0700963 */
964 needed += allocated;
Andi Kleena5516432008-07-23 21:27:41 -0700965 h->resv_huge_pages += delta;
Adam Litkee4e574b2007-10-16 01:26:19 -0700966 ret = 0;
Naoya Horiguchia9869b82010-09-08 10:19:37 +0900967
968 spin_unlock(&hugetlb_lock);
Adam Litke19fc3f02008-04-28 02:12:20 -0700969 /* Free the needed pages to the hugetlb pool */
Adam Litkee4e574b2007-10-16 01:26:19 -0700970 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
Adam Litke19fc3f02008-04-28 02:12:20 -0700971 if ((--needed) < 0)
972 break;
Adam Litkee4e574b2007-10-16 01:26:19 -0700973 list_del(&page->lru);
Naoya Horiguchia9869b82010-09-08 10:19:37 +0900974 /*
975 * This page is now managed by the hugetlb allocator and has
976 * no users -- drop the buddy allocator's reference.
977 */
978 put_page_testzero(page);
979 VM_BUG_ON(page_count(page));
Andi Kleena5516432008-07-23 21:27:41 -0700980 enqueue_huge_page(h, page);
Adam Litke19fc3f02008-04-28 02:12:20 -0700981 }
982
983 /* Free unnecessary surplus pages to the buddy allocator */
Naoya Horiguchia9869b82010-09-08 10:19:37 +0900984free:
Adam Litke19fc3f02008-04-28 02:12:20 -0700985 if (!list_empty(&surplus_list)) {
Adam Litke19fc3f02008-04-28 02:12:20 -0700986 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
987 list_del(&page->lru);
Naoya Horiguchia9869b82010-09-08 10:19:37 +0900988 put_page(page);
Adam Litkeaf767cb2007-10-16 01:26:25 -0700989 }
Adam Litkee4e574b2007-10-16 01:26:19 -0700990 }
Naoya Horiguchia9869b82010-09-08 10:19:37 +0900991 spin_lock(&hugetlb_lock);
Adam Litkee4e574b2007-10-16 01:26:19 -0700992
993 return ret;
994}
995
996/*
997 * When releasing a hugetlb pool reservation, any surplus pages that were
998 * allocated to satisfy the reservation must be explicitly freed if they were
999 * never used.
Lee Schermerhorn685f3452009-09-21 17:01:23 -07001000 * Called with hugetlb_lock held.
Adam Litkee4e574b2007-10-16 01:26:19 -07001001 */
Andi Kleena5516432008-07-23 21:27:41 -07001002static void return_unused_surplus_pages(struct hstate *h,
1003 unsigned long unused_resv_pages)
Adam Litkee4e574b2007-10-16 01:26:19 -07001004{
Adam Litkee4e574b2007-10-16 01:26:19 -07001005 unsigned long nr_pages;
1006
Adam Litkeac09b3a2008-03-04 14:29:38 -08001007 /* Uncommit the reservation */
Andi Kleena5516432008-07-23 21:27:41 -07001008 h->resv_huge_pages -= unused_resv_pages;
Adam Litkeac09b3a2008-03-04 14:29:38 -08001009
Andi Kleenaa888a72008-07-23 21:27:47 -07001010 /* Cannot return gigantic pages currently */
1011 if (h->order >= MAX_ORDER)
1012 return;
1013
Andi Kleena5516432008-07-23 21:27:41 -07001014 nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
Adam Litkee4e574b2007-10-16 01:26:19 -07001015
Lee Schermerhorn685f3452009-09-21 17:01:23 -07001016 /*
1017 * We want to release as many surplus pages as possible, spread
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -08001018 * evenly across all nodes with memory. Iterate across these nodes
1019 * until we can no longer free unreserved surplus pages. This occurs
1020 * when the nodes with surplus pages have no free pages.
1021 * free_pool_huge_page() will balance the the freed pages across the
1022 * on-line nodes with memory and will handle the hstate accounting.
Lee Schermerhorn685f3452009-09-21 17:01:23 -07001023 */
1024 while (nr_pages--) {
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -08001025 if (!free_pool_huge_page(h, &node_states[N_HIGH_MEMORY], 1))
Lee Schermerhorn685f3452009-09-21 17:01:23 -07001026 break;
Adam Litkee4e574b2007-10-16 01:26:19 -07001027 }
1028}
1029
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001030/*
1031 * Determine if the huge page at addr within the vma has an associated
1032 * reservation. Where it does not we will need to logically increase
1033 * reservation and actually increase quota before an allocation can occur.
1034 * Where any new reservation would be required the reservation change is
1035 * prepared, but not committed. Once the page has been quota'd allocated
1036 * an instantiated the change should be committed via vma_commit_reservation.
1037 * No action is required on failure.
1038 */
Roel Kluine2f17d92009-03-31 15:23:15 -07001039static long vma_needs_reservation(struct hstate *h,
Andi Kleena5516432008-07-23 21:27:41 -07001040 struct vm_area_struct *vma, unsigned long addr)
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001041{
1042 struct address_space *mapping = vma->vm_file->f_mapping;
1043 struct inode *inode = mapping->host;
1044
Mel Gormanf83a2752009-05-28 14:34:40 -07001045 if (vma->vm_flags & VM_MAYSHARE) {
Andi Kleena5516432008-07-23 21:27:41 -07001046 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001047 return region_chg(&inode->i_mapping->private_list,
1048 idx, idx + 1);
1049
Andy Whitcroft84afd992008-07-23 21:27:32 -07001050 } else if (!is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
1051 return 1;
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001052
Andy Whitcroft84afd992008-07-23 21:27:32 -07001053 } else {
Roel Kluine2f17d92009-03-31 15:23:15 -07001054 long err;
Andi Kleena5516432008-07-23 21:27:41 -07001055 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
Andy Whitcroft84afd992008-07-23 21:27:32 -07001056 struct resv_map *reservations = vma_resv_map(vma);
1057
1058 err = region_chg(&reservations->regions, idx, idx + 1);
1059 if (err < 0)
1060 return err;
1061 return 0;
1062 }
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001063}
Andi Kleena5516432008-07-23 21:27:41 -07001064static void vma_commit_reservation(struct hstate *h,
1065 struct vm_area_struct *vma, unsigned long addr)
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001066{
1067 struct address_space *mapping = vma->vm_file->f_mapping;
1068 struct inode *inode = mapping->host;
1069
Mel Gormanf83a2752009-05-28 14:34:40 -07001070 if (vma->vm_flags & VM_MAYSHARE) {
Andi Kleena5516432008-07-23 21:27:41 -07001071 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001072 region_add(&inode->i_mapping->private_list, idx, idx + 1);
Andy Whitcroft84afd992008-07-23 21:27:32 -07001073
1074 } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
Andi Kleena5516432008-07-23 21:27:41 -07001075 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
Andy Whitcroft84afd992008-07-23 21:27:32 -07001076 struct resv_map *reservations = vma_resv_map(vma);
1077
1078 /* Mark this page used in the map. */
1079 region_add(&reservations->regions, idx, idx + 1);
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001080 }
1081}
1082
David Gibson27a85ef2006-03-22 00:08:56 -08001083static struct page *alloc_huge_page(struct vm_area_struct *vma,
Mel Gorman04f2cbe2008-07-23 21:27:25 -07001084 unsigned long addr, int avoid_reserve)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001085{
Andi Kleena5516432008-07-23 21:27:41 -07001086 struct hstate *h = hstate_vma(vma);
Adam Litke348ea202007-11-14 16:59:37 -08001087 struct page *page;
Adam Litke2fc39ce2007-11-14 16:59:39 -08001088 struct address_space *mapping = vma->vm_file->f_mapping;
Mel Gormana1e78772008-07-23 21:27:23 -07001089 struct inode *inode = mapping->host;
Roel Kluine2f17d92009-03-31 15:23:15 -07001090 long chg;
Adam Litke2fc39ce2007-11-14 16:59:39 -08001091
Mel Gormana1e78772008-07-23 21:27:23 -07001092 /*
1093 * Processes that did not create the mapping will have no reserves and
1094 * will not have accounted against quota. Check that the quota can be
1095 * made before satisfying the allocation
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001096 * MAP_NORESERVE mappings may also need pages and quota allocated
1097 * if no reserve mapping overlaps.
Mel Gormana1e78772008-07-23 21:27:23 -07001098 */
Andi Kleena5516432008-07-23 21:27:41 -07001099 chg = vma_needs_reservation(h, vma, addr);
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001100 if (chg < 0)
1101 return ERR_PTR(chg);
1102 if (chg)
Mel Gormana1e78772008-07-23 21:27:23 -07001103 if (hugetlb_get_quota(inode->i_mapping, chg))
1104 return ERR_PTR(-ENOSPC);
Mel Gormana1e78772008-07-23 21:27:23 -07001105
1106 spin_lock(&hugetlb_lock);
Andi Kleena5516432008-07-23 21:27:41 -07001107 page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve);
Mel Gormana1e78772008-07-23 21:27:23 -07001108 spin_unlock(&hugetlb_lock);
1109
1110 if (!page) {
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001111 page = alloc_buddy_huge_page(h, NUMA_NO_NODE);
Mel Gormana1e78772008-07-23 21:27:23 -07001112 if (!page) {
1113 hugetlb_put_quota(inode->i_mapping, chg);
Mel Gorman4a6018f2010-05-11 14:06:53 -07001114 return ERR_PTR(-VM_FAULT_SIGBUS);
Mel Gormana1e78772008-07-23 21:27:23 -07001115 }
1116 }
1117
Mel Gormana1e78772008-07-23 21:27:23 -07001118 set_page_private(page, (unsigned long) mapping);
1119
Andi Kleena5516432008-07-23 21:27:41 -07001120 vma_commit_reservation(h, vma, addr);
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001121
Adam Litke90d8b7e2007-11-14 16:59:42 -08001122 return page;
David Gibsonb45b5bd2006-03-22 00:08:55 -08001123}
1124
Cyrill Gorcunov91f47662009-01-06 14:40:33 -08001125int __weak alloc_bootmem_huge_page(struct hstate *h)
Andi Kleenaa888a72008-07-23 21:27:47 -07001126{
1127 struct huge_bootmem_page *m;
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -08001128 int nr_nodes = nodes_weight(node_states[N_HIGH_MEMORY]);
Andi Kleenaa888a72008-07-23 21:27:47 -07001129
1130 while (nr_nodes) {
1131 void *addr;
1132
1133 addr = __alloc_bootmem_node_nopanic(
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001134 NODE_DATA(hstate_next_node_to_alloc(h,
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -08001135 &node_states[N_HIGH_MEMORY])),
Andi Kleenaa888a72008-07-23 21:27:47 -07001136 huge_page_size(h), huge_page_size(h), 0);
1137
1138 if (addr) {
1139 /*
1140 * Use the beginning of the huge page to store the
1141 * huge_bootmem_page struct (until gather_bootmem
1142 * puts them into the mem_map).
1143 */
1144 m = addr;
Cyrill Gorcunov91f47662009-01-06 14:40:33 -08001145 goto found;
Andi Kleenaa888a72008-07-23 21:27:47 -07001146 }
Andi Kleenaa888a72008-07-23 21:27:47 -07001147 nr_nodes--;
1148 }
1149 return 0;
1150
1151found:
1152 BUG_ON((unsigned long)virt_to_phys(m) & (huge_page_size(h) - 1));
1153 /* Put them into a private list first because mem_map is not up yet */
1154 list_add(&m->list, &huge_boot_pages);
1155 m->hstate = h;
1156 return 1;
1157}
1158
Andy Whitcroft18229df2008-11-06 12:53:27 -08001159static void prep_compound_huge_page(struct page *page, int order)
1160{
1161 if (unlikely(order > (MAX_ORDER - 1)))
1162 prep_compound_gigantic_page(page, order);
1163 else
1164 prep_compound_page(page, order);
1165}
1166
Andi Kleenaa888a72008-07-23 21:27:47 -07001167/* Put bootmem huge pages into the standard lists after mem_map is up */
1168static void __init gather_bootmem_prealloc(void)
1169{
1170 struct huge_bootmem_page *m;
1171
1172 list_for_each_entry(m, &huge_boot_pages, list) {
1173 struct page *page = virt_to_page(m);
1174 struct hstate *h = m->hstate;
1175 __ClearPageReserved(page);
1176 WARN_ON(page_count(page) != 1);
Andy Whitcroft18229df2008-11-06 12:53:27 -08001177 prep_compound_huge_page(page, h->order);
Andi Kleenaa888a72008-07-23 21:27:47 -07001178 prep_new_huge_page(h, page, page_to_nid(page));
1179 }
1180}
1181
Andi Kleen8faa8b02008-07-23 21:27:48 -07001182static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001183{
1184 unsigned long i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001185
Andi Kleene5ff2152008-07-23 21:27:42 -07001186 for (i = 0; i < h->max_huge_pages; ++i) {
Andi Kleenaa888a72008-07-23 21:27:47 -07001187 if (h->order >= MAX_ORDER) {
1188 if (!alloc_bootmem_huge_page(h))
1189 break;
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -08001190 } else if (!alloc_fresh_huge_page(h,
1191 &node_states[N_HIGH_MEMORY]))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001192 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001193 }
Andi Kleen8faa8b02008-07-23 21:27:48 -07001194 h->max_huge_pages = i;
Andi Kleene5ff2152008-07-23 21:27:42 -07001195}
1196
1197static void __init hugetlb_init_hstates(void)
1198{
1199 struct hstate *h;
1200
1201 for_each_hstate(h) {
Andi Kleen8faa8b02008-07-23 21:27:48 -07001202 /* oversize hugepages were init'ed in early boot */
1203 if (h->order < MAX_ORDER)
1204 hugetlb_hstate_alloc_pages(h);
Andi Kleene5ff2152008-07-23 21:27:42 -07001205 }
1206}
1207
Andi Kleen4abd32d2008-07-23 21:27:49 -07001208static char * __init memfmt(char *buf, unsigned long n)
1209{
1210 if (n >= (1UL << 30))
1211 sprintf(buf, "%lu GB", n >> 30);
1212 else if (n >= (1UL << 20))
1213 sprintf(buf, "%lu MB", n >> 20);
1214 else
1215 sprintf(buf, "%lu KB", n >> 10);
1216 return buf;
1217}
1218
Andi Kleene5ff2152008-07-23 21:27:42 -07001219static void __init report_hugepages(void)
1220{
1221 struct hstate *h;
1222
1223 for_each_hstate(h) {
Andi Kleen4abd32d2008-07-23 21:27:49 -07001224 char buf[32];
1225 printk(KERN_INFO "HugeTLB registered %s page size, "
1226 "pre-allocated %ld pages\n",
1227 memfmt(buf, huge_page_size(h)),
1228 h->free_huge_pages);
Andi Kleene5ff2152008-07-23 21:27:42 -07001229 }
1230}
1231
Linus Torvalds1da177e2005-04-16 15:20:36 -07001232#ifdef CONFIG_HIGHMEM
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001233static void try_to_free_low(struct hstate *h, unsigned long count,
1234 nodemask_t *nodes_allowed)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001235{
Christoph Lameter4415cc82006-09-25 23:31:55 -07001236 int i;
1237
Andi Kleenaa888a72008-07-23 21:27:47 -07001238 if (h->order >= MAX_ORDER)
1239 return;
1240
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001241 for_each_node_mask(i, *nodes_allowed) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001242 struct page *page, *next;
Andi Kleena5516432008-07-23 21:27:41 -07001243 struct list_head *freel = &h->hugepage_freelists[i];
1244 list_for_each_entry_safe(page, next, freel, lru) {
1245 if (count >= h->nr_huge_pages)
Adam Litke6b0c8802007-10-16 01:26:23 -07001246 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001247 if (PageHighMem(page))
1248 continue;
1249 list_del(&page->lru);
Andi Kleene5ff2152008-07-23 21:27:42 -07001250 update_and_free_page(h, page);
Andi Kleena5516432008-07-23 21:27:41 -07001251 h->free_huge_pages--;
1252 h->free_huge_pages_node[page_to_nid(page)]--;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001253 }
1254 }
1255}
1256#else
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001257static inline void try_to_free_low(struct hstate *h, unsigned long count,
1258 nodemask_t *nodes_allowed)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001259{
1260}
1261#endif
1262
Wu Fengguang20a03072009-06-16 15:32:22 -07001263/*
1264 * Increment or decrement surplus_huge_pages. Keep node-specific counters
1265 * balanced by operating on them in a round-robin fashion.
1266 * Returns 1 if an adjustment was made.
1267 */
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001268static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
1269 int delta)
Wu Fengguang20a03072009-06-16 15:32:22 -07001270{
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001271 int start_nid, next_nid;
Wu Fengguang20a03072009-06-16 15:32:22 -07001272 int ret = 0;
1273
1274 VM_BUG_ON(delta != -1 && delta != 1);
Wu Fengguang20a03072009-06-16 15:32:22 -07001275
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001276 if (delta < 0)
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001277 start_nid = hstate_next_node_to_alloc(h, nodes_allowed);
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001278 else
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001279 start_nid = hstate_next_node_to_free(h, nodes_allowed);
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001280 next_nid = start_nid;
1281
1282 do {
1283 int nid = next_nid;
1284 if (delta < 0) {
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001285 /*
1286 * To shrink on this node, there must be a surplus page
1287 */
Lee Schermerhorn9a76db02009-12-14 17:58:15 -08001288 if (!h->surplus_huge_pages_node[nid]) {
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001289 next_nid = hstate_next_node_to_alloc(h,
1290 nodes_allowed);
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001291 continue;
Lee Schermerhorn9a76db02009-12-14 17:58:15 -08001292 }
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001293 }
1294 if (delta > 0) {
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001295 /*
1296 * Surplus cannot exceed the total number of pages
1297 */
1298 if (h->surplus_huge_pages_node[nid] >=
Lee Schermerhorn9a76db02009-12-14 17:58:15 -08001299 h->nr_huge_pages_node[nid]) {
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001300 next_nid = hstate_next_node_to_free(h,
1301 nodes_allowed);
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001302 continue;
Lee Schermerhorn9a76db02009-12-14 17:58:15 -08001303 }
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001304 }
Wu Fengguang20a03072009-06-16 15:32:22 -07001305
1306 h->surplus_huge_pages += delta;
1307 h->surplus_huge_pages_node[nid] += delta;
1308 ret = 1;
1309 break;
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001310 } while (next_nid != start_nid);
Wu Fengguang20a03072009-06-16 15:32:22 -07001311
Wu Fengguang20a03072009-06-16 15:32:22 -07001312 return ret;
1313}
1314
Andi Kleena5516432008-07-23 21:27:41 -07001315#define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001316static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count,
1317 nodemask_t *nodes_allowed)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001318{
Adam Litke7893d1d2007-10-16 01:26:18 -07001319 unsigned long min_count, ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001320
Andi Kleenaa888a72008-07-23 21:27:47 -07001321 if (h->order >= MAX_ORDER)
1322 return h->max_huge_pages;
1323
Adam Litke7893d1d2007-10-16 01:26:18 -07001324 /*
1325 * Increase the pool size
1326 * First take pages out of surplus state. Then make up the
1327 * remaining difference by allocating fresh huge pages.
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001328 *
1329 * We might race with alloc_buddy_huge_page() here and be unable
1330 * to convert a surplus huge page to a normal huge page. That is
1331 * not critical, though, it just means the overall size of the
1332 * pool might be one hugepage larger than it needs to be, but
1333 * within all the constraints specified by the sysctls.
Adam Litke7893d1d2007-10-16 01:26:18 -07001334 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001335 spin_lock(&hugetlb_lock);
Andi Kleena5516432008-07-23 21:27:41 -07001336 while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001337 if (!adjust_pool_surplus(h, nodes_allowed, -1))
Adam Litke7893d1d2007-10-16 01:26:18 -07001338 break;
1339 }
1340
Andi Kleena5516432008-07-23 21:27:41 -07001341 while (count > persistent_huge_pages(h)) {
Adam Litke7893d1d2007-10-16 01:26:18 -07001342 /*
1343 * If this allocation races such that we no longer need the
1344 * page, free_huge_page will handle it by freeing the page
1345 * and reducing the surplus.
1346 */
1347 spin_unlock(&hugetlb_lock);
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001348 ret = alloc_fresh_huge_page(h, nodes_allowed);
Adam Litke7893d1d2007-10-16 01:26:18 -07001349 spin_lock(&hugetlb_lock);
1350 if (!ret)
1351 goto out;
1352
Mel Gorman536240f2009-12-14 17:59:56 -08001353 /* Bail for signals. Probably ctrl-c from user */
1354 if (signal_pending(current))
1355 goto out;
Adam Litke7893d1d2007-10-16 01:26:18 -07001356 }
Adam Litke7893d1d2007-10-16 01:26:18 -07001357
1358 /*
1359 * Decrease the pool size
1360 * First return free pages to the buddy allocator (being careful
1361 * to keep enough around to satisfy reservations). Then place
1362 * pages into surplus state as needed so the pool will shrink
1363 * to the desired size as pages become free.
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001364 *
1365 * By placing pages into the surplus state independent of the
1366 * overcommit value, we are allowing the surplus pool size to
1367 * exceed overcommit. There are few sane options here. Since
1368 * alloc_buddy_huge_page() is checking the global counter,
1369 * though, we'll note that we're not allowed to exceed surplus
1370 * and won't grow the pool anywhere else. Not until one of the
1371 * sysctls are changed, or the surplus pages go out of use.
Adam Litke7893d1d2007-10-16 01:26:18 -07001372 */
Andi Kleena5516432008-07-23 21:27:41 -07001373 min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
Adam Litke6b0c8802007-10-16 01:26:23 -07001374 min_count = max(count, min_count);
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001375 try_to_free_low(h, min_count, nodes_allowed);
Andi Kleena5516432008-07-23 21:27:41 -07001376 while (min_count < persistent_huge_pages(h)) {
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001377 if (!free_pool_huge_page(h, nodes_allowed, 0))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001378 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001379 }
Andi Kleena5516432008-07-23 21:27:41 -07001380 while (count < persistent_huge_pages(h)) {
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001381 if (!adjust_pool_surplus(h, nodes_allowed, 1))
Adam Litke7893d1d2007-10-16 01:26:18 -07001382 break;
1383 }
1384out:
Andi Kleena5516432008-07-23 21:27:41 -07001385 ret = persistent_huge_pages(h);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001386 spin_unlock(&hugetlb_lock);
Adam Litke7893d1d2007-10-16 01:26:18 -07001387 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001388}
1389
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001390#define HSTATE_ATTR_RO(_name) \
1391 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
1392
1393#define HSTATE_ATTR(_name) \
1394 static struct kobj_attribute _name##_attr = \
1395 __ATTR(_name, 0644, _name##_show, _name##_store)
1396
1397static struct kobject *hugepages_kobj;
1398static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
1399
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001400static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
1401
1402static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001403{
1404 int i;
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001405
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001406 for (i = 0; i < HUGE_MAX_HSTATE; i++)
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001407 if (hstate_kobjs[i] == kobj) {
1408 if (nidp)
1409 *nidp = NUMA_NO_NODE;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001410 return &hstates[i];
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001411 }
1412
1413 return kobj_to_node_hstate(kobj, nidp);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001414}
1415
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001416static ssize_t nr_hugepages_show_common(struct kobject *kobj,
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001417 struct kobj_attribute *attr, char *buf)
1418{
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001419 struct hstate *h;
1420 unsigned long nr_huge_pages;
1421 int nid;
1422
1423 h = kobj_to_hstate(kobj, &nid);
1424 if (nid == NUMA_NO_NODE)
1425 nr_huge_pages = h->nr_huge_pages;
1426 else
1427 nr_huge_pages = h->nr_huge_pages_node[nid];
1428
1429 return sprintf(buf, "%lu\n", nr_huge_pages);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001430}
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001431static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
1432 struct kobject *kobj, struct kobj_attribute *attr,
1433 const char *buf, size_t len)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001434{
1435 int err;
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001436 int nid;
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001437 unsigned long count;
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001438 struct hstate *h;
David Rientjesbad44b52009-12-14 17:58:38 -08001439 NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001440
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001441 err = strict_strtoul(buf, 10, &count);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001442 if (err)
1443 return 0;
1444
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001445 h = kobj_to_hstate(kobj, &nid);
1446 if (nid == NUMA_NO_NODE) {
1447 /*
1448 * global hstate attribute
1449 */
1450 if (!(obey_mempolicy &&
1451 init_nodemask_of_mempolicy(nodes_allowed))) {
1452 NODEMASK_FREE(nodes_allowed);
1453 nodes_allowed = &node_states[N_HIGH_MEMORY];
1454 }
1455 } else if (nodes_allowed) {
1456 /*
1457 * per node hstate attribute: adjust count to global,
1458 * but restrict alloc/free to the specified node.
1459 */
1460 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
1461 init_nodemask_of_node(nodes_allowed, nid);
1462 } else
1463 nodes_allowed = &node_states[N_HIGH_MEMORY];
1464
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001465 h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001466
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -08001467 if (nodes_allowed != &node_states[N_HIGH_MEMORY])
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001468 NODEMASK_FREE(nodes_allowed);
1469
1470 return len;
1471}
1472
1473static ssize_t nr_hugepages_show(struct kobject *kobj,
1474 struct kobj_attribute *attr, char *buf)
1475{
1476 return nr_hugepages_show_common(kobj, attr, buf);
1477}
1478
1479static ssize_t nr_hugepages_store(struct kobject *kobj,
1480 struct kobj_attribute *attr, const char *buf, size_t len)
1481{
1482 return nr_hugepages_store_common(false, kobj, attr, buf, len);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001483}
1484HSTATE_ATTR(nr_hugepages);
1485
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001486#ifdef CONFIG_NUMA
1487
1488/*
1489 * hstate attribute for optionally mempolicy-based constraint on persistent
1490 * huge page alloc/free.
1491 */
1492static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
1493 struct kobj_attribute *attr, char *buf)
1494{
1495 return nr_hugepages_show_common(kobj, attr, buf);
1496}
1497
1498static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
1499 struct kobj_attribute *attr, const char *buf, size_t len)
1500{
1501 return nr_hugepages_store_common(true, kobj, attr, buf, len);
1502}
1503HSTATE_ATTR(nr_hugepages_mempolicy);
1504#endif
1505
1506
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001507static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
1508 struct kobj_attribute *attr, char *buf)
1509{
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001510 struct hstate *h = kobj_to_hstate(kobj, NULL);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001511 return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
1512}
1513static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
1514 struct kobj_attribute *attr, const char *buf, size_t count)
1515{
1516 int err;
1517 unsigned long input;
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001518 struct hstate *h = kobj_to_hstate(kobj, NULL);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001519
1520 err = strict_strtoul(buf, 10, &input);
1521 if (err)
1522 return 0;
1523
1524 spin_lock(&hugetlb_lock);
1525 h->nr_overcommit_huge_pages = input;
1526 spin_unlock(&hugetlb_lock);
1527
1528 return count;
1529}
1530HSTATE_ATTR(nr_overcommit_hugepages);
1531
1532static ssize_t free_hugepages_show(struct kobject *kobj,
1533 struct kobj_attribute *attr, char *buf)
1534{
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001535 struct hstate *h;
1536 unsigned long free_huge_pages;
1537 int nid;
1538
1539 h = kobj_to_hstate(kobj, &nid);
1540 if (nid == NUMA_NO_NODE)
1541 free_huge_pages = h->free_huge_pages;
1542 else
1543 free_huge_pages = h->free_huge_pages_node[nid];
1544
1545 return sprintf(buf, "%lu\n", free_huge_pages);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001546}
1547HSTATE_ATTR_RO(free_hugepages);
1548
1549static ssize_t resv_hugepages_show(struct kobject *kobj,
1550 struct kobj_attribute *attr, char *buf)
1551{
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001552 struct hstate *h = kobj_to_hstate(kobj, NULL);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001553 return sprintf(buf, "%lu\n", h->resv_huge_pages);
1554}
1555HSTATE_ATTR_RO(resv_hugepages);
1556
1557static ssize_t surplus_hugepages_show(struct kobject *kobj,
1558 struct kobj_attribute *attr, char *buf)
1559{
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001560 struct hstate *h;
1561 unsigned long surplus_huge_pages;
1562 int nid;
1563
1564 h = kobj_to_hstate(kobj, &nid);
1565 if (nid == NUMA_NO_NODE)
1566 surplus_huge_pages = h->surplus_huge_pages;
1567 else
1568 surplus_huge_pages = h->surplus_huge_pages_node[nid];
1569
1570 return sprintf(buf, "%lu\n", surplus_huge_pages);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001571}
1572HSTATE_ATTR_RO(surplus_hugepages);
1573
1574static struct attribute *hstate_attrs[] = {
1575 &nr_hugepages_attr.attr,
1576 &nr_overcommit_hugepages_attr.attr,
1577 &free_hugepages_attr.attr,
1578 &resv_hugepages_attr.attr,
1579 &surplus_hugepages_attr.attr,
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001580#ifdef CONFIG_NUMA
1581 &nr_hugepages_mempolicy_attr.attr,
1582#endif
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001583 NULL,
1584};
1585
1586static struct attribute_group hstate_attr_group = {
1587 .attrs = hstate_attrs,
1588};
1589
Jeff Mahoney094e9532010-02-02 13:44:14 -08001590static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
1591 struct kobject **hstate_kobjs,
1592 struct attribute_group *hstate_attr_group)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001593{
1594 int retval;
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001595 int hi = h - hstates;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001596
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001597 hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
1598 if (!hstate_kobjs[hi])
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001599 return -ENOMEM;
1600
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001601 retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001602 if (retval)
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001603 kobject_put(hstate_kobjs[hi]);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001604
1605 return retval;
1606}
1607
1608static void __init hugetlb_sysfs_init(void)
1609{
1610 struct hstate *h;
1611 int err;
1612
1613 hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
1614 if (!hugepages_kobj)
1615 return;
1616
1617 for_each_hstate(h) {
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001618 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
1619 hstate_kobjs, &hstate_attr_group);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001620 if (err)
1621 printk(KERN_ERR "Hugetlb: Unable to add hstate %s",
1622 h->name);
1623 }
1624}
1625
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001626#ifdef CONFIG_NUMA
1627
1628/*
1629 * node_hstate/s - associate per node hstate attributes, via their kobjects,
1630 * with node sysdevs in node_devices[] using a parallel array. The array
1631 * index of a node sysdev or _hstate == node id.
1632 * This is here to avoid any static dependency of the node sysdev driver, in
1633 * the base kernel, on the hugetlb module.
1634 */
1635struct node_hstate {
1636 struct kobject *hugepages_kobj;
1637 struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
1638};
1639struct node_hstate node_hstates[MAX_NUMNODES];
1640
1641/*
1642 * A subset of global hstate attributes for node sysdevs
1643 */
1644static struct attribute *per_node_hstate_attrs[] = {
1645 &nr_hugepages_attr.attr,
1646 &free_hugepages_attr.attr,
1647 &surplus_hugepages_attr.attr,
1648 NULL,
1649};
1650
1651static struct attribute_group per_node_hstate_attr_group = {
1652 .attrs = per_node_hstate_attrs,
1653};
1654
1655/*
1656 * kobj_to_node_hstate - lookup global hstate for node sysdev hstate attr kobj.
1657 * Returns node id via non-NULL nidp.
1658 */
1659static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
1660{
1661 int nid;
1662
1663 for (nid = 0; nid < nr_node_ids; nid++) {
1664 struct node_hstate *nhs = &node_hstates[nid];
1665 int i;
1666 for (i = 0; i < HUGE_MAX_HSTATE; i++)
1667 if (nhs->hstate_kobjs[i] == kobj) {
1668 if (nidp)
1669 *nidp = nid;
1670 return &hstates[i];
1671 }
1672 }
1673
1674 BUG();
1675 return NULL;
1676}
1677
1678/*
1679 * Unregister hstate attributes from a single node sysdev.
1680 * No-op if no hstate attributes attached.
1681 */
1682void hugetlb_unregister_node(struct node *node)
1683{
1684 struct hstate *h;
1685 struct node_hstate *nhs = &node_hstates[node->sysdev.id];
1686
1687 if (!nhs->hugepages_kobj)
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -08001688 return; /* no hstate attributes */
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001689
1690 for_each_hstate(h)
1691 if (nhs->hstate_kobjs[h - hstates]) {
1692 kobject_put(nhs->hstate_kobjs[h - hstates]);
1693 nhs->hstate_kobjs[h - hstates] = NULL;
1694 }
1695
1696 kobject_put(nhs->hugepages_kobj);
1697 nhs->hugepages_kobj = NULL;
1698}
1699
1700/*
1701 * hugetlb module exit: unregister hstate attributes from node sysdevs
1702 * that have them.
1703 */
1704static void hugetlb_unregister_all_nodes(void)
1705{
1706 int nid;
1707
1708 /*
1709 * disable node sysdev registrations.
1710 */
1711 register_hugetlbfs_with_node(NULL, NULL);
1712
1713 /*
1714 * remove hstate attributes from any nodes that have them.
1715 */
1716 for (nid = 0; nid < nr_node_ids; nid++)
1717 hugetlb_unregister_node(&node_devices[nid]);
1718}
1719
1720/*
1721 * Register hstate attributes for a single node sysdev.
1722 * No-op if attributes already registered.
1723 */
1724void hugetlb_register_node(struct node *node)
1725{
1726 struct hstate *h;
1727 struct node_hstate *nhs = &node_hstates[node->sysdev.id];
1728 int err;
1729
1730 if (nhs->hugepages_kobj)
1731 return; /* already allocated */
1732
1733 nhs->hugepages_kobj = kobject_create_and_add("hugepages",
1734 &node->sysdev.kobj);
1735 if (!nhs->hugepages_kobj)
1736 return;
1737
1738 for_each_hstate(h) {
1739 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
1740 nhs->hstate_kobjs,
1741 &per_node_hstate_attr_group);
1742 if (err) {
1743 printk(KERN_ERR "Hugetlb: Unable to add hstate %s"
1744 " for node %d\n",
1745 h->name, node->sysdev.id);
1746 hugetlb_unregister_node(node);
1747 break;
1748 }
1749 }
1750}
1751
1752/*
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -08001753 * hugetlb init time: register hstate attributes for all registered node
1754 * sysdevs of nodes that have memory. All on-line nodes should have
1755 * registered their associated sysdev by this time.
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001756 */
1757static void hugetlb_register_all_nodes(void)
1758{
1759 int nid;
1760
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -08001761 for_each_node_state(nid, N_HIGH_MEMORY) {
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001762 struct node *node = &node_devices[nid];
1763 if (node->sysdev.id == nid)
1764 hugetlb_register_node(node);
1765 }
1766
1767 /*
1768 * Let the node sysdev driver know we're here so it can
1769 * [un]register hstate attributes on node hotplug.
1770 */
1771 register_hugetlbfs_with_node(hugetlb_register_node,
1772 hugetlb_unregister_node);
1773}
1774#else /* !CONFIG_NUMA */
1775
1776static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
1777{
1778 BUG();
1779 if (nidp)
1780 *nidp = -1;
1781 return NULL;
1782}
1783
1784static void hugetlb_unregister_all_nodes(void) { }
1785
1786static void hugetlb_register_all_nodes(void) { }
1787
1788#endif
1789
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001790static void __exit hugetlb_exit(void)
1791{
1792 struct hstate *h;
1793
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001794 hugetlb_unregister_all_nodes();
1795
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001796 for_each_hstate(h) {
1797 kobject_put(hstate_kobjs[h - hstates]);
1798 }
1799
1800 kobject_put(hugepages_kobj);
1801}
1802module_exit(hugetlb_exit);
1803
1804static int __init hugetlb_init(void)
1805{
Benjamin Herrenschmidt0ef89d22008-07-31 00:07:30 -07001806 /* Some platform decide whether they support huge pages at boot
1807 * time. On these, such as powerpc, HPAGE_SHIFT is set to 0 when
1808 * there is no such support
1809 */
1810 if (HPAGE_SHIFT == 0)
1811 return 0;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001812
Nick Piggine11bfbf2008-07-23 21:27:52 -07001813 if (!size_to_hstate(default_hstate_size)) {
1814 default_hstate_size = HPAGE_SIZE;
1815 if (!size_to_hstate(default_hstate_size))
1816 hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001817 }
Nick Piggine11bfbf2008-07-23 21:27:52 -07001818 default_hstate_idx = size_to_hstate(default_hstate_size) - hstates;
1819 if (default_hstate_max_huge_pages)
1820 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001821
1822 hugetlb_init_hstates();
1823
Andi Kleenaa888a72008-07-23 21:27:47 -07001824 gather_bootmem_prealloc();
1825
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001826 report_hugepages();
1827
1828 hugetlb_sysfs_init();
1829
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001830 hugetlb_register_all_nodes();
1831
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001832 return 0;
1833}
1834module_init(hugetlb_init);
1835
1836/* Should be called on processing a hugepagesz=... option */
1837void __init hugetlb_add_hstate(unsigned order)
1838{
1839 struct hstate *h;
Andi Kleen8faa8b02008-07-23 21:27:48 -07001840 unsigned long i;
1841
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001842 if (size_to_hstate(PAGE_SIZE << order)) {
1843 printk(KERN_WARNING "hugepagesz= specified twice, ignoring\n");
1844 return;
1845 }
1846 BUG_ON(max_hstate >= HUGE_MAX_HSTATE);
1847 BUG_ON(order == 0);
1848 h = &hstates[max_hstate++];
1849 h->order = order;
1850 h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
Andi Kleen8faa8b02008-07-23 21:27:48 -07001851 h->nr_huge_pages = 0;
1852 h->free_huge_pages = 0;
1853 for (i = 0; i < MAX_NUMNODES; ++i)
1854 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -08001855 h->next_nid_to_alloc = first_node(node_states[N_HIGH_MEMORY]);
1856 h->next_nid_to_free = first_node(node_states[N_HIGH_MEMORY]);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001857 snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
1858 huge_page_size(h)/1024);
Andi Kleen8faa8b02008-07-23 21:27:48 -07001859
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001860 parsed_hstate = h;
1861}
1862
Nick Piggine11bfbf2008-07-23 21:27:52 -07001863static int __init hugetlb_nrpages_setup(char *s)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001864{
1865 unsigned long *mhp;
Andi Kleen8faa8b02008-07-23 21:27:48 -07001866 static unsigned long *last_mhp;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001867
1868 /*
1869 * !max_hstate means we haven't parsed a hugepagesz= parameter yet,
1870 * so this hugepages= parameter goes to the "default hstate".
1871 */
1872 if (!max_hstate)
1873 mhp = &default_hstate_max_huge_pages;
1874 else
1875 mhp = &parsed_hstate->max_huge_pages;
1876
Andi Kleen8faa8b02008-07-23 21:27:48 -07001877 if (mhp == last_mhp) {
1878 printk(KERN_WARNING "hugepages= specified twice without "
1879 "interleaving hugepagesz=, ignoring\n");
1880 return 1;
1881 }
1882
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001883 if (sscanf(s, "%lu", mhp) <= 0)
1884 *mhp = 0;
1885
Andi Kleen8faa8b02008-07-23 21:27:48 -07001886 /*
1887 * Global state is always initialized later in hugetlb_init.
1888 * But we need to allocate >= MAX_ORDER hstates here early to still
1889 * use the bootmem allocator.
1890 */
1891 if (max_hstate && parsed_hstate->order >= MAX_ORDER)
1892 hugetlb_hstate_alloc_pages(parsed_hstate);
1893
1894 last_mhp = mhp;
1895
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001896 return 1;
1897}
Nick Piggine11bfbf2008-07-23 21:27:52 -07001898__setup("hugepages=", hugetlb_nrpages_setup);
1899
1900static int __init hugetlb_default_setup(char *s)
1901{
1902 default_hstate_size = memparse(s, &s);
1903 return 1;
1904}
1905__setup("default_hugepagesz=", hugetlb_default_setup);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001906
Nishanth Aravamudan8a213462008-07-25 19:44:37 -07001907static unsigned int cpuset_mems_nr(unsigned int *array)
1908{
1909 int node;
1910 unsigned int nr = 0;
1911
1912 for_each_node_mask(node, cpuset_current_mems_allowed)
1913 nr += array[node];
1914
1915 return nr;
1916}
1917
1918#ifdef CONFIG_SYSCTL
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001919static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
1920 struct ctl_table *table, int write,
1921 void __user *buffer, size_t *length, loff_t *ppos)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001922{
Andi Kleene5ff2152008-07-23 21:27:42 -07001923 struct hstate *h = &default_hstate;
1924 unsigned long tmp;
1925
1926 if (!write)
1927 tmp = h->max_huge_pages;
1928
1929 table->data = &tmp;
1930 table->maxlen = sizeof(unsigned long);
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07001931 proc_doulongvec_minmax(table, write, buffer, length, ppos);
Andi Kleene5ff2152008-07-23 21:27:42 -07001932
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001933 if (write) {
David Rientjesbad44b52009-12-14 17:58:38 -08001934 NODEMASK_ALLOC(nodemask_t, nodes_allowed,
1935 GFP_KERNEL | __GFP_NORETRY);
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001936 if (!(obey_mempolicy &&
1937 init_nodemask_of_mempolicy(nodes_allowed))) {
1938 NODEMASK_FREE(nodes_allowed);
1939 nodes_allowed = &node_states[N_HIGH_MEMORY];
1940 }
1941 h->max_huge_pages = set_max_huge_pages(h, tmp, nodes_allowed);
1942
1943 if (nodes_allowed != &node_states[N_HIGH_MEMORY])
1944 NODEMASK_FREE(nodes_allowed);
1945 }
Andi Kleene5ff2152008-07-23 21:27:42 -07001946
Linus Torvalds1da177e2005-04-16 15:20:36 -07001947 return 0;
1948}
Mel Gorman396faf02007-07-17 04:03:13 -07001949
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001950int hugetlb_sysctl_handler(struct ctl_table *table, int write,
1951 void __user *buffer, size_t *length, loff_t *ppos)
1952{
1953
1954 return hugetlb_sysctl_handler_common(false, table, write,
1955 buffer, length, ppos);
1956}
1957
1958#ifdef CONFIG_NUMA
1959int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
1960 void __user *buffer, size_t *length, loff_t *ppos)
1961{
1962 return hugetlb_sysctl_handler_common(true, table, write,
1963 buffer, length, ppos);
1964}
1965#endif /* CONFIG_NUMA */
1966
Mel Gorman396faf02007-07-17 04:03:13 -07001967int hugetlb_treat_movable_handler(struct ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07001968 void __user *buffer,
Mel Gorman396faf02007-07-17 04:03:13 -07001969 size_t *length, loff_t *ppos)
1970{
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07001971 proc_dointvec(table, write, buffer, length, ppos);
Mel Gorman396faf02007-07-17 04:03:13 -07001972 if (hugepages_treat_as_movable)
1973 htlb_alloc_mask = GFP_HIGHUSER_MOVABLE;
1974 else
1975 htlb_alloc_mask = GFP_HIGHUSER;
1976 return 0;
1977}
1978
Nishanth Aravamudana3d0c6a2008-02-08 04:18:18 -08001979int hugetlb_overcommit_handler(struct ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07001980 void __user *buffer,
Nishanth Aravamudana3d0c6a2008-02-08 04:18:18 -08001981 size_t *length, loff_t *ppos)
1982{
Andi Kleena5516432008-07-23 21:27:41 -07001983 struct hstate *h = &default_hstate;
Andi Kleene5ff2152008-07-23 21:27:42 -07001984 unsigned long tmp;
1985
1986 if (!write)
1987 tmp = h->nr_overcommit_huge_pages;
1988
1989 table->data = &tmp;
1990 table->maxlen = sizeof(unsigned long);
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07001991 proc_doulongvec_minmax(table, write, buffer, length, ppos);
Andi Kleene5ff2152008-07-23 21:27:42 -07001992
1993 if (write) {
1994 spin_lock(&hugetlb_lock);
1995 h->nr_overcommit_huge_pages = tmp;
1996 spin_unlock(&hugetlb_lock);
1997 }
1998
Nishanth Aravamudana3d0c6a2008-02-08 04:18:18 -08001999 return 0;
2000}
2001
Linus Torvalds1da177e2005-04-16 15:20:36 -07002002#endif /* CONFIG_SYSCTL */
2003
Alexey Dobriyane1759c22008-10-15 23:50:22 +04002004void hugetlb_report_meminfo(struct seq_file *m)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002005{
Andi Kleena5516432008-07-23 21:27:41 -07002006 struct hstate *h = &default_hstate;
Alexey Dobriyane1759c22008-10-15 23:50:22 +04002007 seq_printf(m,
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002008 "HugePages_Total: %5lu\n"
2009 "HugePages_Free: %5lu\n"
2010 "HugePages_Rsvd: %5lu\n"
2011 "HugePages_Surp: %5lu\n"
2012 "Hugepagesize: %8lu kB\n",
Andi Kleena5516432008-07-23 21:27:41 -07002013 h->nr_huge_pages,
2014 h->free_huge_pages,
2015 h->resv_huge_pages,
2016 h->surplus_huge_pages,
2017 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002018}
2019
2020int hugetlb_report_node_meminfo(int nid, char *buf)
2021{
Andi Kleena5516432008-07-23 21:27:41 -07002022 struct hstate *h = &default_hstate;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002023 return sprintf(buf,
2024 "Node %d HugePages_Total: %5u\n"
Nishanth Aravamudana1de0912008-03-26 14:37:53 -07002025 "Node %d HugePages_Free: %5u\n"
2026 "Node %d HugePages_Surp: %5u\n",
Andi Kleena5516432008-07-23 21:27:41 -07002027 nid, h->nr_huge_pages_node[nid],
2028 nid, h->free_huge_pages_node[nid],
2029 nid, h->surplus_huge_pages_node[nid]);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002030}
2031
Linus Torvalds1da177e2005-04-16 15:20:36 -07002032/* Return the number pages of memory we physically have, in PAGE_SIZE units. */
2033unsigned long hugetlb_total_pages(void)
2034{
Andi Kleena5516432008-07-23 21:27:41 -07002035 struct hstate *h = &default_hstate;
2036 return h->nr_huge_pages * pages_per_huge_page(h);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002037}
Linus Torvalds1da177e2005-04-16 15:20:36 -07002038
Andi Kleena5516432008-07-23 21:27:41 -07002039static int hugetlb_acct_memory(struct hstate *h, long delta)
Mel Gormanfc1b8a72008-07-23 21:27:22 -07002040{
2041 int ret = -ENOMEM;
2042
2043 spin_lock(&hugetlb_lock);
2044 /*
2045 * When cpuset is configured, it breaks the strict hugetlb page
2046 * reservation as the accounting is done on a global variable. Such
2047 * reservation is completely rubbish in the presence of cpuset because
2048 * the reservation is not checked against page availability for the
2049 * current cpuset. Application can still potentially OOM'ed by kernel
2050 * with lack of free htlb page in cpuset that the task is in.
2051 * Attempt to enforce strict accounting with cpuset is almost
2052 * impossible (or too ugly) because cpuset is too fluid that
2053 * task or memory node can be dynamically moved between cpusets.
2054 *
2055 * The change of semantics for shared hugetlb mapping with cpuset is
2056 * undesirable. However, in order to preserve some of the semantics,
2057 * we fall back to check against current free page availability as
2058 * a best attempt and hopefully to minimize the impact of changing
2059 * semantics that cpuset has.
2060 */
2061 if (delta > 0) {
Andi Kleena5516432008-07-23 21:27:41 -07002062 if (gather_surplus_pages(h, delta) < 0)
Mel Gormanfc1b8a72008-07-23 21:27:22 -07002063 goto out;
2064
Andi Kleena5516432008-07-23 21:27:41 -07002065 if (delta > cpuset_mems_nr(h->free_huge_pages_node)) {
2066 return_unused_surplus_pages(h, delta);
Mel Gormanfc1b8a72008-07-23 21:27:22 -07002067 goto out;
2068 }
2069 }
2070
2071 ret = 0;
2072 if (delta < 0)
Andi Kleena5516432008-07-23 21:27:41 -07002073 return_unused_surplus_pages(h, (unsigned long) -delta);
Mel Gormanfc1b8a72008-07-23 21:27:22 -07002074
2075out:
2076 spin_unlock(&hugetlb_lock);
2077 return ret;
2078}
2079
Andy Whitcroft84afd992008-07-23 21:27:32 -07002080static void hugetlb_vm_op_open(struct vm_area_struct *vma)
2081{
2082 struct resv_map *reservations = vma_resv_map(vma);
2083
2084 /*
2085 * This new VMA should share its siblings reservation map if present.
2086 * The VMA will only ever have a valid reservation map pointer where
2087 * it is being copied for another still existing VMA. As that VMA
2088 * has a reference to the reservation map it cannot dissappear until
2089 * after this open call completes. It is therefore safe to take a
2090 * new reference here without additional locking.
2091 */
2092 if (reservations)
2093 kref_get(&reservations->refs);
2094}
2095
Mel Gormana1e78772008-07-23 21:27:23 -07002096static void hugetlb_vm_op_close(struct vm_area_struct *vma)
2097{
Andi Kleena5516432008-07-23 21:27:41 -07002098 struct hstate *h = hstate_vma(vma);
Andy Whitcroft84afd992008-07-23 21:27:32 -07002099 struct resv_map *reservations = vma_resv_map(vma);
2100 unsigned long reserve;
2101 unsigned long start;
2102 unsigned long end;
2103
2104 if (reservations) {
Andi Kleena5516432008-07-23 21:27:41 -07002105 start = vma_hugecache_offset(h, vma, vma->vm_start);
2106 end = vma_hugecache_offset(h, vma, vma->vm_end);
Andy Whitcroft84afd992008-07-23 21:27:32 -07002107
2108 reserve = (end - start) -
2109 region_count(&reservations->regions, start, end);
2110
2111 kref_put(&reservations->refs, resv_map_release);
2112
Adam Litke7251ff782008-07-23 21:27:59 -07002113 if (reserve) {
Andi Kleena5516432008-07-23 21:27:41 -07002114 hugetlb_acct_memory(h, -reserve);
Adam Litke7251ff782008-07-23 21:27:59 -07002115 hugetlb_put_quota(vma->vm_file->f_mapping, reserve);
2116 }
Andy Whitcroft84afd992008-07-23 21:27:32 -07002117 }
Mel Gormana1e78772008-07-23 21:27:23 -07002118}
2119
Linus Torvalds1da177e2005-04-16 15:20:36 -07002120/*
2121 * We cannot handle pagefaults against hugetlb pages at all. They cause
2122 * handle_mm_fault() to try to instantiate regular-sized pages in the
2123 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
2124 * this far.
2125 */
Nick Piggind0217ac2007-07-19 01:47:03 -07002126static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002127{
2128 BUG();
Nick Piggind0217ac2007-07-19 01:47:03 -07002129 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002130}
2131
Alexey Dobriyanf0f37e22009-09-27 22:29:37 +04002132const struct vm_operations_struct hugetlb_vm_ops = {
Nick Piggind0217ac2007-07-19 01:47:03 -07002133 .fault = hugetlb_vm_op_fault,
Andy Whitcroft84afd992008-07-23 21:27:32 -07002134 .open = hugetlb_vm_op_open,
Mel Gormana1e78772008-07-23 21:27:23 -07002135 .close = hugetlb_vm_op_close,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002136};
2137
David Gibson1e8f8892006-01-06 00:10:44 -08002138static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
2139 int writable)
David Gibson63551ae2005-06-21 17:14:44 -07002140{
2141 pte_t entry;
2142
David Gibson1e8f8892006-01-06 00:10:44 -08002143 if (writable) {
David Gibson63551ae2005-06-21 17:14:44 -07002144 entry =
2145 pte_mkwrite(pte_mkdirty(mk_pte(page, vma->vm_page_prot)));
2146 } else {
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07002147 entry = huge_pte_wrprotect(mk_pte(page, vma->vm_page_prot));
David Gibson63551ae2005-06-21 17:14:44 -07002148 }
2149 entry = pte_mkyoung(entry);
2150 entry = pte_mkhuge(entry);
2151
2152 return entry;
2153}
2154
David Gibson1e8f8892006-01-06 00:10:44 -08002155static void set_huge_ptep_writable(struct vm_area_struct *vma,
2156 unsigned long address, pte_t *ptep)
2157{
2158 pte_t entry;
2159
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07002160 entry = pte_mkwrite(pte_mkdirty(huge_ptep_get(ptep)));
2161 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) {
Russell King4b3073e2009-12-18 16:40:18 +00002162 update_mmu_cache(vma, address, ptep);
Benjamin Herrenschmidt8dab5242007-06-16 10:16:12 -07002163 }
David Gibson1e8f8892006-01-06 00:10:44 -08002164}
2165
2166
David Gibson63551ae2005-06-21 17:14:44 -07002167int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
2168 struct vm_area_struct *vma)
2169{
2170 pte_t *src_pte, *dst_pte, entry;
2171 struct page *ptepage;
Hugh Dickins1c598272005-10-19 21:23:43 -07002172 unsigned long addr;
David Gibson1e8f8892006-01-06 00:10:44 -08002173 int cow;
Andi Kleena5516432008-07-23 21:27:41 -07002174 struct hstate *h = hstate_vma(vma);
2175 unsigned long sz = huge_page_size(h);
David Gibson1e8f8892006-01-06 00:10:44 -08002176
2177 cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
David Gibson63551ae2005-06-21 17:14:44 -07002178
Andi Kleena5516432008-07-23 21:27:41 -07002179 for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
Hugh Dickinsc74df322005-10-29 18:16:23 -07002180 src_pte = huge_pte_offset(src, addr);
2181 if (!src_pte)
2182 continue;
Andi Kleena5516432008-07-23 21:27:41 -07002183 dst_pte = huge_pte_alloc(dst, addr, sz);
David Gibson63551ae2005-06-21 17:14:44 -07002184 if (!dst_pte)
2185 goto nomem;
Larry Woodmanc5c99422008-01-24 05:49:25 -08002186
2187 /* If the pagetables are shared don't copy or take references */
2188 if (dst_pte == src_pte)
2189 continue;
2190
Hugh Dickinsc74df322005-10-29 18:16:23 -07002191 spin_lock(&dst->page_table_lock);
Nick Piggin46478752008-06-05 22:45:57 -07002192 spin_lock_nested(&src->page_table_lock, SINGLE_DEPTH_NESTING);
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07002193 if (!huge_pte_none(huge_ptep_get(src_pte))) {
David Gibson1e8f8892006-01-06 00:10:44 -08002194 if (cow)
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07002195 huge_ptep_set_wrprotect(src, addr, src_pte);
2196 entry = huge_ptep_get(src_pte);
Hugh Dickins1c598272005-10-19 21:23:43 -07002197 ptepage = pte_page(entry);
2198 get_page(ptepage);
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002199 page_dup_rmap(ptepage);
Hugh Dickins1c598272005-10-19 21:23:43 -07002200 set_huge_pte_at(dst, addr, dst_pte, entry);
2201 }
2202 spin_unlock(&src->page_table_lock);
Hugh Dickinsc74df322005-10-29 18:16:23 -07002203 spin_unlock(&dst->page_table_lock);
David Gibson63551ae2005-06-21 17:14:44 -07002204 }
2205 return 0;
2206
2207nomem:
2208 return -ENOMEM;
2209}
2210
Naoya Horiguchi290408d2010-09-08 10:19:35 +09002211static int is_hugetlb_entry_migration(pte_t pte)
2212{
2213 swp_entry_t swp;
2214
2215 if (huge_pte_none(pte) || pte_present(pte))
2216 return 0;
2217 swp = pte_to_swp_entry(pte);
2218 if (non_swap_entry(swp) && is_migration_entry(swp)) {
2219 return 1;
2220 } else
2221 return 0;
2222}
2223
Naoya Horiguchifd6a03e2010-05-28 09:29:21 +09002224static int is_hugetlb_entry_hwpoisoned(pte_t pte)
2225{
2226 swp_entry_t swp;
2227
2228 if (huge_pte_none(pte) || pte_present(pte))
2229 return 0;
2230 swp = pte_to_swp_entry(pte);
2231 if (non_swap_entry(swp) && is_hwpoison_entry(swp)) {
2232 return 1;
2233 } else
2234 return 0;
2235}
2236
Chen, Kenneth W502717f2006-10-11 01:20:46 -07002237void __unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002238 unsigned long end, struct page *ref_page)
David Gibson63551ae2005-06-21 17:14:44 -07002239{
2240 struct mm_struct *mm = vma->vm_mm;
2241 unsigned long address;
David Gibsonc7546f82005-08-05 11:59:35 -07002242 pte_t *ptep;
David Gibson63551ae2005-06-21 17:14:44 -07002243 pte_t pte;
2244 struct page *page;
Chen, Kenneth Wfe1668a2006-10-04 02:15:24 -07002245 struct page *tmp;
Andi Kleena5516432008-07-23 21:27:41 -07002246 struct hstate *h = hstate_vma(vma);
2247 unsigned long sz = huge_page_size(h);
2248
Chen, Kenneth Wc0a499c2006-12-06 20:31:39 -08002249 /*
2250 * A page gathering list, protected by per file i_mmap_lock. The
2251 * lock is used to avoid list corruption from multiple unmapping
2252 * of the same page since we are using page->lru.
2253 */
Chen, Kenneth Wfe1668a2006-10-04 02:15:24 -07002254 LIST_HEAD(page_list);
David Gibson63551ae2005-06-21 17:14:44 -07002255
2256 WARN_ON(!is_vm_hugetlb_page(vma));
Andi Kleena5516432008-07-23 21:27:41 -07002257 BUG_ON(start & ~huge_page_mask(h));
2258 BUG_ON(end & ~huge_page_mask(h));
David Gibson63551ae2005-06-21 17:14:44 -07002259
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07002260 mmu_notifier_invalidate_range_start(mm, start, end);
Hugh Dickins508034a2005-10-29 18:16:30 -07002261 spin_lock(&mm->page_table_lock);
Andi Kleena5516432008-07-23 21:27:41 -07002262 for (address = start; address < end; address += sz) {
David Gibsonc7546f82005-08-05 11:59:35 -07002263 ptep = huge_pte_offset(mm, address);
Adam Litke4c887262005-10-29 18:16:46 -07002264 if (!ptep)
David Gibsonc7546f82005-08-05 11:59:35 -07002265 continue;
2266
Chen, Kenneth W39dde652006-12-06 20:32:03 -08002267 if (huge_pmd_unshare(mm, &address, ptep))
2268 continue;
2269
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002270 /*
2271 * If a reference page is supplied, it is because a specific
2272 * page is being unmapped, not a range. Ensure the page we
2273 * are about to unmap is the actual page of interest.
2274 */
2275 if (ref_page) {
2276 pte = huge_ptep_get(ptep);
2277 if (huge_pte_none(pte))
2278 continue;
2279 page = pte_page(pte);
2280 if (page != ref_page)
2281 continue;
2282
2283 /*
2284 * Mark the VMA as having unmapped its page so that
2285 * future faults in this VMA will fail rather than
2286 * looking like data was lost
2287 */
2288 set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
2289 }
2290
David Gibsonc7546f82005-08-05 11:59:35 -07002291 pte = huge_ptep_get_and_clear(mm, address, ptep);
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07002292 if (huge_pte_none(pte))
David Gibson63551ae2005-06-21 17:14:44 -07002293 continue;
David Gibsonc7546f82005-08-05 11:59:35 -07002294
Naoya Horiguchifd6a03e2010-05-28 09:29:21 +09002295 /*
2296 * HWPoisoned hugepage is already unmapped and dropped reference
2297 */
2298 if (unlikely(is_hugetlb_entry_hwpoisoned(pte)))
2299 continue;
2300
David Gibson63551ae2005-06-21 17:14:44 -07002301 page = pte_page(pte);
Ken Chen6649a382007-02-08 14:20:27 -08002302 if (pte_dirty(pte))
2303 set_page_dirty(page);
Chen, Kenneth Wfe1668a2006-10-04 02:15:24 -07002304 list_add(&page->lru, &page_list);
David Gibson63551ae2005-06-21 17:14:44 -07002305 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002306 spin_unlock(&mm->page_table_lock);
Hugh Dickins508034a2005-10-29 18:16:30 -07002307 flush_tlb_range(vma, start, end);
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07002308 mmu_notifier_invalidate_range_end(mm, start, end);
Chen, Kenneth Wfe1668a2006-10-04 02:15:24 -07002309 list_for_each_entry_safe(page, tmp, &page_list, lru) {
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002310 page_remove_rmap(page);
Chen, Kenneth Wfe1668a2006-10-04 02:15:24 -07002311 list_del(&page->lru);
2312 put_page(page);
2313 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002314}
David Gibson63551ae2005-06-21 17:14:44 -07002315
Chen, Kenneth W502717f2006-10-11 01:20:46 -07002316void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002317 unsigned long end, struct page *ref_page)
Chen, Kenneth W502717f2006-10-11 01:20:46 -07002318{
Andi Kleena137e1c2008-07-23 21:27:43 -07002319 spin_lock(&vma->vm_file->f_mapping->i_mmap_lock);
2320 __unmap_hugepage_range(vma, start, end, ref_page);
2321 spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock);
Chen, Kenneth W502717f2006-10-11 01:20:46 -07002322}
2323
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002324/*
2325 * This is called when the original mapper is failing to COW a MAP_PRIVATE
2326 * mappping it owns the reserve page for. The intention is to unmap the page
2327 * from other VMAs and let the children be SIGKILLed if they are faulting the
2328 * same region.
2329 */
Harvey Harrison2a4b3de2008-10-18 20:27:06 -07002330static int unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
2331 struct page *page, unsigned long address)
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002332{
Adam Litke75266742008-11-12 13:24:56 -08002333 struct hstate *h = hstate_vma(vma);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002334 struct vm_area_struct *iter_vma;
2335 struct address_space *mapping;
2336 struct prio_tree_iter iter;
2337 pgoff_t pgoff;
2338
2339 /*
2340 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
2341 * from page cache lookup which is in HPAGE_SIZE units.
2342 */
Adam Litke75266742008-11-12 13:24:56 -08002343 address = address & huge_page_mask(h);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002344 pgoff = ((address - vma->vm_start) >> PAGE_SHIFT)
2345 + (vma->vm_pgoff >> PAGE_SHIFT);
2346 mapping = (struct address_space *)page_private(page);
2347
Mel Gorman4eb2b1d2009-12-14 17:59:53 -08002348 /*
2349 * Take the mapping lock for the duration of the table walk. As
2350 * this mapping should be shared between all the VMAs,
2351 * __unmap_hugepage_range() is called as the lock is already held
2352 */
2353 spin_lock(&mapping->i_mmap_lock);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002354 vma_prio_tree_foreach(iter_vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
2355 /* Do not unmap the current VMA */
2356 if (iter_vma == vma)
2357 continue;
2358
2359 /*
2360 * Unmap the page from other VMAs without their own reserves.
2361 * They get marked to be SIGKILLed if they fault in these
2362 * areas. This is because a future no-page fault on this VMA
2363 * could insert a zeroed page instead of the data existing
2364 * from the time of fork. This would look like data corruption
2365 */
2366 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
Mel Gorman4eb2b1d2009-12-14 17:59:53 -08002367 __unmap_hugepage_range(iter_vma,
Adam Litke75266742008-11-12 13:24:56 -08002368 address, address + huge_page_size(h),
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002369 page);
2370 }
Mel Gorman4eb2b1d2009-12-14 17:59:53 -08002371 spin_unlock(&mapping->i_mmap_lock);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002372
2373 return 1;
2374}
2375
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002376/*
2377 * Hugetlb_cow() should be called with page lock of the original hugepage held.
2378 */
David Gibson1e8f8892006-01-06 00:10:44 -08002379static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002380 unsigned long address, pte_t *ptep, pte_t pte,
2381 struct page *pagecache_page)
David Gibson1e8f8892006-01-06 00:10:44 -08002382{
Andi Kleena5516432008-07-23 21:27:41 -07002383 struct hstate *h = hstate_vma(vma);
David Gibson1e8f8892006-01-06 00:10:44 -08002384 struct page *old_page, *new_page;
David Gibson79ac6ba2006-03-22 00:08:51 -08002385 int avoidcopy;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002386 int outside_reserve = 0;
David Gibson1e8f8892006-01-06 00:10:44 -08002387
2388 old_page = pte_page(pte);
2389
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002390retry_avoidcopy:
David Gibson1e8f8892006-01-06 00:10:44 -08002391 /* If no-one else is actually using this page, avoid the copy
2392 * and just make the page writable */
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002393 avoidcopy = (page_mapcount(old_page) == 1);
David Gibson1e8f8892006-01-06 00:10:44 -08002394 if (avoidcopy) {
Naoya Horiguchi56c9cfb2010-09-10 13:23:04 +09002395 if (PageAnon(old_page))
2396 page_move_anon_rmap(old_page, vma, address);
David Gibson1e8f8892006-01-06 00:10:44 -08002397 set_huge_ptep_writable(vma, address, ptep);
Nick Piggin83c54072007-07-19 01:47:05 -07002398 return 0;
David Gibson1e8f8892006-01-06 00:10:44 -08002399 }
2400
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002401 /*
2402 * If the process that created a MAP_PRIVATE mapping is about to
2403 * perform a COW due to a shared page count, attempt to satisfy
2404 * the allocation without using the existing reserves. The pagecache
2405 * page is used to determine if the reserve at this address was
2406 * consumed or not. If reserves were used, a partial faulted mapping
2407 * at the time of fork() could consume its reserves on COW instead
2408 * of the full address range.
2409 */
Mel Gormanf83a2752009-05-28 14:34:40 -07002410 if (!(vma->vm_flags & VM_MAYSHARE) &&
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002411 is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
2412 old_page != pagecache_page)
2413 outside_reserve = 1;
2414
David Gibson1e8f8892006-01-06 00:10:44 -08002415 page_cache_get(old_page);
Larry Woodmanb76c8cf2009-12-14 17:59:37 -08002416
2417 /* Drop page_table_lock as buddy allocator may be called */
2418 spin_unlock(&mm->page_table_lock);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002419 new_page = alloc_huge_page(vma, address, outside_reserve);
David Gibson1e8f8892006-01-06 00:10:44 -08002420
Adam Litke2fc39ce2007-11-14 16:59:39 -08002421 if (IS_ERR(new_page)) {
David Gibson1e8f8892006-01-06 00:10:44 -08002422 page_cache_release(old_page);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002423
2424 /*
2425 * If a process owning a MAP_PRIVATE mapping fails to COW,
2426 * it is due to references held by a child and an insufficient
2427 * huge page pool. To guarantee the original mappers
2428 * reliability, unmap the page from child processes. The child
2429 * may get SIGKILLed if it later faults.
2430 */
2431 if (outside_reserve) {
2432 BUG_ON(huge_pte_none(pte));
2433 if (unmap_ref_private(mm, vma, old_page, address)) {
2434 BUG_ON(page_count(old_page) != 1);
2435 BUG_ON(huge_pte_none(pte));
Larry Woodmanb76c8cf2009-12-14 17:59:37 -08002436 spin_lock(&mm->page_table_lock);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002437 goto retry_avoidcopy;
2438 }
2439 WARN_ON_ONCE(1);
2440 }
2441
Larry Woodmanb76c8cf2009-12-14 17:59:37 -08002442 /* Caller expects lock to be held */
2443 spin_lock(&mm->page_table_lock);
Adam Litke2fc39ce2007-11-14 16:59:39 -08002444 return -PTR_ERR(new_page);
David Gibson1e8f8892006-01-06 00:10:44 -08002445 }
2446
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002447 /*
2448 * When the original hugepage is shared one, it does not have
2449 * anon_vma prepared.
2450 */
Dean Nelson44e2aa92010-10-26 14:22:08 -07002451 if (unlikely(anon_vma_prepare(vma))) {
2452 /* Caller expects lock to be held */
2453 spin_lock(&mm->page_table_lock);
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002454 return VM_FAULT_OOM;
Dean Nelson44e2aa92010-10-26 14:22:08 -07002455 }
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002456
Naoya Horiguchi0ebabb42010-09-08 10:19:34 +09002457 copy_user_huge_page(new_page, old_page, address, vma);
Nick Piggin0ed361d2008-02-04 22:29:34 -08002458 __SetPageUptodate(new_page);
David Gibson1e8f8892006-01-06 00:10:44 -08002459
Larry Woodmanb76c8cf2009-12-14 17:59:37 -08002460 /*
2461 * Retake the page_table_lock to check for racing updates
2462 * before the page tables are altered
2463 */
2464 spin_lock(&mm->page_table_lock);
Andi Kleena5516432008-07-23 21:27:41 -07002465 ptep = huge_pte_offset(mm, address & huge_page_mask(h));
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07002466 if (likely(pte_same(huge_ptep_get(ptep), pte))) {
David Gibson1e8f8892006-01-06 00:10:44 -08002467 /* Break COW */
Doug Doan3edd4fc2010-08-09 17:18:30 -07002468 mmu_notifier_invalidate_range_start(mm,
2469 address & huge_page_mask(h),
2470 (address & huge_page_mask(h)) + huge_page_size(h));
Gerald Schaefer8fe627e2008-04-28 02:13:28 -07002471 huge_ptep_clear_flush(vma, address, ptep);
David Gibson1e8f8892006-01-06 00:10:44 -08002472 set_huge_pte_at(mm, address, ptep,
2473 make_huge_pte(vma, new_page, 1));
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002474 page_remove_rmap(old_page);
Naoya Horiguchicd67f0d2010-09-10 13:23:04 +09002475 hugepage_add_new_anon_rmap(new_page, vma, address);
David Gibson1e8f8892006-01-06 00:10:44 -08002476 /* Make the old page be freed below */
2477 new_page = old_page;
Doug Doan3edd4fc2010-08-09 17:18:30 -07002478 mmu_notifier_invalidate_range_end(mm,
2479 address & huge_page_mask(h),
2480 (address & huge_page_mask(h)) + huge_page_size(h));
David Gibson1e8f8892006-01-06 00:10:44 -08002481 }
2482 page_cache_release(new_page);
2483 page_cache_release(old_page);
Nick Piggin83c54072007-07-19 01:47:05 -07002484 return 0;
David Gibson1e8f8892006-01-06 00:10:44 -08002485}
2486
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002487/* Return the pagecache page at a given address within a VMA */
Andi Kleena5516432008-07-23 21:27:41 -07002488static struct page *hugetlbfs_pagecache_page(struct hstate *h,
2489 struct vm_area_struct *vma, unsigned long address)
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002490{
2491 struct address_space *mapping;
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -07002492 pgoff_t idx;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002493
2494 mapping = vma->vm_file->f_mapping;
Andi Kleena5516432008-07-23 21:27:41 -07002495 idx = vma_hugecache_offset(h, vma, address);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002496
2497 return find_lock_page(mapping, idx);
2498}
2499
Hugh Dickins3ae77f42009-09-21 17:03:33 -07002500/*
2501 * Return whether there is a pagecache page to back given address within VMA.
2502 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
2503 */
2504static bool hugetlbfs_pagecache_present(struct hstate *h,
Hugh Dickins2a15efc2009-09-21 17:03:27 -07002505 struct vm_area_struct *vma, unsigned long address)
2506{
2507 struct address_space *mapping;
2508 pgoff_t idx;
2509 struct page *page;
2510
2511 mapping = vma->vm_file->f_mapping;
2512 idx = vma_hugecache_offset(h, vma, address);
2513
2514 page = find_get_page(mapping, idx);
2515 if (page)
2516 put_page(page);
2517 return page != NULL;
2518}
2519
Robert P. J. Daya1ed3dd2007-07-17 04:03:33 -07002520static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
Hugh Dickins788c7df2009-06-23 13:49:05 +01002521 unsigned long address, pte_t *ptep, unsigned int flags)
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01002522{
Andi Kleena5516432008-07-23 21:27:41 -07002523 struct hstate *h = hstate_vma(vma);
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01002524 int ret = VM_FAULT_SIGBUS;
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -07002525 pgoff_t idx;
Adam Litke4c887262005-10-29 18:16:46 -07002526 unsigned long size;
Adam Litke4c887262005-10-29 18:16:46 -07002527 struct page *page;
2528 struct address_space *mapping;
David Gibson1e8f8892006-01-06 00:10:44 -08002529 pte_t new_pte;
Adam Litke4c887262005-10-29 18:16:46 -07002530
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002531 /*
2532 * Currently, we are forced to kill the process in the event the
2533 * original mapper has unmapped pages from the child due to a failed
2534 * COW. Warn that such a situation has occured as it may not be obvious
2535 */
2536 if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
2537 printk(KERN_WARNING
2538 "PID %d killed due to inadequate hugepage pool\n",
2539 current->pid);
2540 return ret;
2541 }
2542
Adam Litke4c887262005-10-29 18:16:46 -07002543 mapping = vma->vm_file->f_mapping;
Andi Kleena5516432008-07-23 21:27:41 -07002544 idx = vma_hugecache_offset(h, vma, address);
Adam Litke4c887262005-10-29 18:16:46 -07002545
2546 /*
2547 * Use page lock to guard against racing truncation
2548 * before we get page_table_lock.
2549 */
Christoph Lameter6bda6662006-01-06 00:10:49 -08002550retry:
2551 page = find_lock_page(mapping, idx);
2552 if (!page) {
Andi Kleena5516432008-07-23 21:27:41 -07002553 size = i_size_read(mapping->host) >> huge_page_shift(h);
Hugh Dickinsebed4bf2006-10-28 10:38:43 -07002554 if (idx >= size)
2555 goto out;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002556 page = alloc_huge_page(vma, address, 0);
Adam Litke2fc39ce2007-11-14 16:59:39 -08002557 if (IS_ERR(page)) {
2558 ret = -PTR_ERR(page);
Christoph Lameter6bda6662006-01-06 00:10:49 -08002559 goto out;
2560 }
Andi Kleena5516432008-07-23 21:27:41 -07002561 clear_huge_page(page, address, huge_page_size(h));
Nick Piggin0ed361d2008-02-04 22:29:34 -08002562 __SetPageUptodate(page);
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01002563
Mel Gormanf83a2752009-05-28 14:34:40 -07002564 if (vma->vm_flags & VM_MAYSHARE) {
Christoph Lameter6bda6662006-01-06 00:10:49 -08002565 int err;
Ken Chen45c682a2007-11-14 16:59:44 -08002566 struct inode *inode = mapping->host;
Christoph Lameter6bda6662006-01-06 00:10:49 -08002567
2568 err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
2569 if (err) {
2570 put_page(page);
Christoph Lameter6bda6662006-01-06 00:10:49 -08002571 if (err == -EEXIST)
2572 goto retry;
2573 goto out;
2574 }
Ken Chen45c682a2007-11-14 16:59:44 -08002575
2576 spin_lock(&inode->i_lock);
Andi Kleena5516432008-07-23 21:27:41 -07002577 inode->i_blocks += blocks_per_huge_page(h);
Ken Chen45c682a2007-11-14 16:59:44 -08002578 spin_unlock(&inode->i_lock);
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002579 page_dup_rmap(page);
Mel Gorman23be7462010-04-23 13:17:56 -04002580 } else {
Christoph Lameter6bda6662006-01-06 00:10:49 -08002581 lock_page(page);
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002582 if (unlikely(anon_vma_prepare(vma))) {
2583 ret = VM_FAULT_OOM;
2584 goto backout_unlocked;
2585 }
2586 hugepage_add_new_anon_rmap(page, vma, address);
Mel Gorman23be7462010-04-23 13:17:56 -04002587 }
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002588 } else {
Naoya Horiguchi998b4382010-09-08 10:19:32 +09002589 /*
2590 * If memory error occurs between mmap() and fault, some process
2591 * don't have hwpoisoned swap entry for errored virtual address.
2592 * So we need to block hugepage fault by PG_hwpoison bit check.
2593 */
2594 if (unlikely(PageHWPoison(page))) {
Andi Kleenaa50d3a2010-10-06 21:45:00 +02002595 ret = VM_FAULT_HWPOISON |
2596 VM_FAULT_SET_HINDEX(h - hstates);
Naoya Horiguchi998b4382010-09-08 10:19:32 +09002597 goto backout_unlocked;
2598 }
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002599 page_dup_rmap(page);
Christoph Lameter6bda6662006-01-06 00:10:49 -08002600 }
David Gibson1e8f8892006-01-06 00:10:44 -08002601
Andy Whitcroft57303d82008-08-12 15:08:47 -07002602 /*
2603 * If we are going to COW a private mapping later, we examine the
2604 * pending reservations for this page now. This will ensure that
2605 * any allocations necessary to record that reservation occur outside
2606 * the spinlock.
2607 */
Hugh Dickins788c7df2009-06-23 13:49:05 +01002608 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED))
Andy Whitcroft2b267362008-08-12 15:08:49 -07002609 if (vma_needs_reservation(h, vma, address) < 0) {
2610 ret = VM_FAULT_OOM;
2611 goto backout_unlocked;
2612 }
Andy Whitcroft57303d82008-08-12 15:08:47 -07002613
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01002614 spin_lock(&mm->page_table_lock);
Andi Kleena5516432008-07-23 21:27:41 -07002615 size = i_size_read(mapping->host) >> huge_page_shift(h);
Adam Litke4c887262005-10-29 18:16:46 -07002616 if (idx >= size)
2617 goto backout;
2618
Nick Piggin83c54072007-07-19 01:47:05 -07002619 ret = 0;
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07002620 if (!huge_pte_none(huge_ptep_get(ptep)))
Adam Litke4c887262005-10-29 18:16:46 -07002621 goto backout;
2622
David Gibson1e8f8892006-01-06 00:10:44 -08002623 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
2624 && (vma->vm_flags & VM_SHARED)));
2625 set_huge_pte_at(mm, address, ptep, new_pte);
2626
Hugh Dickins788c7df2009-06-23 13:49:05 +01002627 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
David Gibson1e8f8892006-01-06 00:10:44 -08002628 /* Optimization, do the COW without a second fault */
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002629 ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page);
David Gibson1e8f8892006-01-06 00:10:44 -08002630 }
2631
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01002632 spin_unlock(&mm->page_table_lock);
Adam Litke4c887262005-10-29 18:16:46 -07002633 unlock_page(page);
2634out:
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01002635 return ret;
Adam Litke4c887262005-10-29 18:16:46 -07002636
2637backout:
2638 spin_unlock(&mm->page_table_lock);
Andy Whitcroft2b267362008-08-12 15:08:49 -07002639backout_unlocked:
Adam Litke4c887262005-10-29 18:16:46 -07002640 unlock_page(page);
2641 put_page(page);
2642 goto out;
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01002643}
2644
Adam Litke86e52162006-01-06 00:10:43 -08002645int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
Hugh Dickins788c7df2009-06-23 13:49:05 +01002646 unsigned long address, unsigned int flags)
Adam Litke86e52162006-01-06 00:10:43 -08002647{
2648 pte_t *ptep;
2649 pte_t entry;
David Gibson1e8f8892006-01-06 00:10:44 -08002650 int ret;
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002651 struct page *page = NULL;
Andy Whitcroft57303d82008-08-12 15:08:47 -07002652 struct page *pagecache_page = NULL;
David Gibson3935baa2006-03-22 00:08:53 -08002653 static DEFINE_MUTEX(hugetlb_instantiation_mutex);
Andi Kleena5516432008-07-23 21:27:41 -07002654 struct hstate *h = hstate_vma(vma);
Adam Litke86e52162006-01-06 00:10:43 -08002655
Naoya Horiguchifd6a03e2010-05-28 09:29:21 +09002656 ptep = huge_pte_offset(mm, address);
2657 if (ptep) {
2658 entry = huge_ptep_get(ptep);
Naoya Horiguchi290408d2010-09-08 10:19:35 +09002659 if (unlikely(is_hugetlb_entry_migration(entry))) {
2660 migration_entry_wait(mm, (pmd_t *)ptep, address);
2661 return 0;
2662 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
Andi Kleenaa50d3a2010-10-06 21:45:00 +02002663 return VM_FAULT_HWPOISON_LARGE |
2664 VM_FAULT_SET_HINDEX(h - hstates);
Naoya Horiguchifd6a03e2010-05-28 09:29:21 +09002665 }
2666
Andi Kleena5516432008-07-23 21:27:41 -07002667 ptep = huge_pte_alloc(mm, address, huge_page_size(h));
Adam Litke86e52162006-01-06 00:10:43 -08002668 if (!ptep)
2669 return VM_FAULT_OOM;
2670
David Gibson3935baa2006-03-22 00:08:53 -08002671 /*
2672 * Serialize hugepage allocation and instantiation, so that we don't
2673 * get spurious allocation failures if two CPUs race to instantiate
2674 * the same page in the page cache.
2675 */
2676 mutex_lock(&hugetlb_instantiation_mutex);
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07002677 entry = huge_ptep_get(ptep);
2678 if (huge_pte_none(entry)) {
Hugh Dickins788c7df2009-06-23 13:49:05 +01002679 ret = hugetlb_no_page(mm, vma, address, ptep, flags);
David Gibsonb4d1d992008-10-15 22:01:11 -07002680 goto out_mutex;
David Gibson3935baa2006-03-22 00:08:53 -08002681 }
Adam Litke86e52162006-01-06 00:10:43 -08002682
Nick Piggin83c54072007-07-19 01:47:05 -07002683 ret = 0;
David Gibson1e8f8892006-01-06 00:10:44 -08002684
Andy Whitcroft57303d82008-08-12 15:08:47 -07002685 /*
2686 * If we are going to COW the mapping later, we examine the pending
2687 * reservations for this page now. This will ensure that any
2688 * allocations necessary to record that reservation occur outside the
2689 * spinlock. For private mappings, we also lookup the pagecache
2690 * page now as it is used to determine if a reservation has been
2691 * consumed.
2692 */
Hugh Dickins788c7df2009-06-23 13:49:05 +01002693 if ((flags & FAULT_FLAG_WRITE) && !pte_write(entry)) {
Andy Whitcroft2b267362008-08-12 15:08:49 -07002694 if (vma_needs_reservation(h, vma, address) < 0) {
2695 ret = VM_FAULT_OOM;
David Gibsonb4d1d992008-10-15 22:01:11 -07002696 goto out_mutex;
Andy Whitcroft2b267362008-08-12 15:08:49 -07002697 }
Andy Whitcroft57303d82008-08-12 15:08:47 -07002698
Mel Gormanf83a2752009-05-28 14:34:40 -07002699 if (!(vma->vm_flags & VM_MAYSHARE))
Andy Whitcroft57303d82008-08-12 15:08:47 -07002700 pagecache_page = hugetlbfs_pagecache_page(h,
2701 vma, address);
2702 }
2703
Naoya Horiguchi56c9cfb2010-09-10 13:23:04 +09002704 /*
2705 * hugetlb_cow() requires page locks of pte_page(entry) and
2706 * pagecache_page, so here we need take the former one
2707 * when page != pagecache_page or !pagecache_page.
2708 * Note that locking order is always pagecache_page -> page,
2709 * so no worry about deadlock.
2710 */
2711 page = pte_page(entry);
2712 if (page != pagecache_page)
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002713 lock_page(page);
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002714
David Gibson1e8f8892006-01-06 00:10:44 -08002715 spin_lock(&mm->page_table_lock);
2716 /* Check for a racing update before calling hugetlb_cow */
David Gibsonb4d1d992008-10-15 22:01:11 -07002717 if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
2718 goto out_page_table_lock;
2719
2720
Hugh Dickins788c7df2009-06-23 13:49:05 +01002721 if (flags & FAULT_FLAG_WRITE) {
David Gibsonb4d1d992008-10-15 22:01:11 -07002722 if (!pte_write(entry)) {
Andy Whitcroft57303d82008-08-12 15:08:47 -07002723 ret = hugetlb_cow(mm, vma, address, ptep, entry,
2724 pagecache_page);
David Gibsonb4d1d992008-10-15 22:01:11 -07002725 goto out_page_table_lock;
2726 }
2727 entry = pte_mkdirty(entry);
2728 }
2729 entry = pte_mkyoung(entry);
Hugh Dickins788c7df2009-06-23 13:49:05 +01002730 if (huge_ptep_set_access_flags(vma, address, ptep, entry,
2731 flags & FAULT_FLAG_WRITE))
Russell King4b3073e2009-12-18 16:40:18 +00002732 update_mmu_cache(vma, address, ptep);
David Gibsonb4d1d992008-10-15 22:01:11 -07002733
2734out_page_table_lock:
David Gibson1e8f8892006-01-06 00:10:44 -08002735 spin_unlock(&mm->page_table_lock);
Andy Whitcroft57303d82008-08-12 15:08:47 -07002736
2737 if (pagecache_page) {
2738 unlock_page(pagecache_page);
2739 put_page(pagecache_page);
2740 }
Dean Nelson1f64d692010-12-02 14:31:12 -08002741 if (page != pagecache_page)
2742 unlock_page(page);
Andy Whitcroft57303d82008-08-12 15:08:47 -07002743
David Gibsonb4d1d992008-10-15 22:01:11 -07002744out_mutex:
David Gibson3935baa2006-03-22 00:08:53 -08002745 mutex_unlock(&hugetlb_instantiation_mutex);
David Gibson1e8f8892006-01-06 00:10:44 -08002746
2747 return ret;
Adam Litke86e52162006-01-06 00:10:43 -08002748}
2749
Andi Kleenceb86872008-07-23 21:27:50 -07002750/* Can be overriden by architectures */
2751__attribute__((weak)) struct page *
2752follow_huge_pud(struct mm_struct *mm, unsigned long address,
2753 pud_t *pud, int write)
2754{
2755 BUG();
2756 return NULL;
2757}
2758
David Gibson63551ae2005-06-21 17:14:44 -07002759int follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
2760 struct page **pages, struct vm_area_struct **vmas,
Adam Litke5b23dbe2007-11-14 16:59:33 -08002761 unsigned long *position, int *length, int i,
Hugh Dickins2a15efc2009-09-21 17:03:27 -07002762 unsigned int flags)
David Gibson63551ae2005-06-21 17:14:44 -07002763{
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08002764 unsigned long pfn_offset;
2765 unsigned long vaddr = *position;
David Gibson63551ae2005-06-21 17:14:44 -07002766 int remainder = *length;
Andi Kleena5516432008-07-23 21:27:41 -07002767 struct hstate *h = hstate_vma(vma);
David Gibson63551ae2005-06-21 17:14:44 -07002768
Hugh Dickins1c598272005-10-19 21:23:43 -07002769 spin_lock(&mm->page_table_lock);
David Gibson63551ae2005-06-21 17:14:44 -07002770 while (vaddr < vma->vm_end && remainder) {
Adam Litke4c887262005-10-29 18:16:46 -07002771 pte_t *pte;
Hugh Dickins2a15efc2009-09-21 17:03:27 -07002772 int absent;
Adam Litke4c887262005-10-29 18:16:46 -07002773 struct page *page;
2774
2775 /*
2776 * Some archs (sparc64, sh*) have multiple pte_ts to
Hugh Dickins2a15efc2009-09-21 17:03:27 -07002777 * each hugepage. We have to make sure we get the
Adam Litke4c887262005-10-29 18:16:46 -07002778 * first, for the page indexing below to work.
2779 */
Andi Kleena5516432008-07-23 21:27:41 -07002780 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h));
Hugh Dickins2a15efc2009-09-21 17:03:27 -07002781 absent = !pte || huge_pte_none(huge_ptep_get(pte));
Adam Litke4c887262005-10-29 18:16:46 -07002782
Hugh Dickins2a15efc2009-09-21 17:03:27 -07002783 /*
2784 * When coredumping, it suits get_dump_page if we just return
Hugh Dickins3ae77f42009-09-21 17:03:33 -07002785 * an error where there's an empty slot with no huge pagecache
2786 * to back it. This way, we avoid allocating a hugepage, and
2787 * the sparse dumpfile avoids allocating disk blocks, but its
2788 * huge holes still show up with zeroes where they need to be.
Hugh Dickins2a15efc2009-09-21 17:03:27 -07002789 */
Hugh Dickins3ae77f42009-09-21 17:03:33 -07002790 if (absent && (flags & FOLL_DUMP) &&
2791 !hugetlbfs_pagecache_present(h, vma, vaddr)) {
Hugh Dickins2a15efc2009-09-21 17:03:27 -07002792 remainder = 0;
2793 break;
2794 }
2795
2796 if (absent ||
2797 ((flags & FOLL_WRITE) && !pte_write(huge_ptep_get(pte)))) {
Adam Litke4c887262005-10-29 18:16:46 -07002798 int ret;
2799
2800 spin_unlock(&mm->page_table_lock);
Hugh Dickins2a15efc2009-09-21 17:03:27 -07002801 ret = hugetlb_fault(mm, vma, vaddr,
2802 (flags & FOLL_WRITE) ? FAULT_FLAG_WRITE : 0);
Adam Litke4c887262005-10-29 18:16:46 -07002803 spin_lock(&mm->page_table_lock);
Adam Litkea89182c2007-08-22 14:01:51 -07002804 if (!(ret & VM_FAULT_ERROR))
Adam Litke4c887262005-10-29 18:16:46 -07002805 continue;
2806
2807 remainder = 0;
Adam Litke4c887262005-10-29 18:16:46 -07002808 break;
2809 }
David Gibson63551ae2005-06-21 17:14:44 -07002810
Andi Kleena5516432008-07-23 21:27:41 -07002811 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07002812 page = pte_page(huge_ptep_get(pte));
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08002813same_page:
Chen, Kenneth Wd6692182006-03-31 02:29:57 -08002814 if (pages) {
Hugh Dickins2a15efc2009-09-21 17:03:27 -07002815 pages[i] = mem_map_offset(page, pfn_offset);
KOSAKI Motohiro4b2e38a2008-10-18 20:27:10 -07002816 get_page(pages[i]);
Chen, Kenneth Wd6692182006-03-31 02:29:57 -08002817 }
David Gibson63551ae2005-06-21 17:14:44 -07002818
2819 if (vmas)
2820 vmas[i] = vma;
2821
2822 vaddr += PAGE_SIZE;
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08002823 ++pfn_offset;
David Gibson63551ae2005-06-21 17:14:44 -07002824 --remainder;
2825 ++i;
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08002826 if (vaddr < vma->vm_end && remainder &&
Andi Kleena5516432008-07-23 21:27:41 -07002827 pfn_offset < pages_per_huge_page(h)) {
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08002828 /*
2829 * We use pfn_offset to avoid touching the pageframes
2830 * of this compound page.
2831 */
2832 goto same_page;
2833 }
David Gibson63551ae2005-06-21 17:14:44 -07002834 }
Hugh Dickins1c598272005-10-19 21:23:43 -07002835 spin_unlock(&mm->page_table_lock);
David Gibson63551ae2005-06-21 17:14:44 -07002836 *length = remainder;
2837 *position = vaddr;
2838
Hugh Dickins2a15efc2009-09-21 17:03:27 -07002839 return i ? i : -EFAULT;
David Gibson63551ae2005-06-21 17:14:44 -07002840}
Zhang, Yanmin8f860592006-03-22 00:08:50 -08002841
2842void hugetlb_change_protection(struct vm_area_struct *vma,
2843 unsigned long address, unsigned long end, pgprot_t newprot)
2844{
2845 struct mm_struct *mm = vma->vm_mm;
2846 unsigned long start = address;
2847 pte_t *ptep;
2848 pte_t pte;
Andi Kleena5516432008-07-23 21:27:41 -07002849 struct hstate *h = hstate_vma(vma);
Zhang, Yanmin8f860592006-03-22 00:08:50 -08002850
2851 BUG_ON(address >= end);
2852 flush_cache_range(vma, address, end);
2853
Chen, Kenneth W39dde652006-12-06 20:32:03 -08002854 spin_lock(&vma->vm_file->f_mapping->i_mmap_lock);
Zhang, Yanmin8f860592006-03-22 00:08:50 -08002855 spin_lock(&mm->page_table_lock);
Andi Kleena5516432008-07-23 21:27:41 -07002856 for (; address < end; address += huge_page_size(h)) {
Zhang, Yanmin8f860592006-03-22 00:08:50 -08002857 ptep = huge_pte_offset(mm, address);
2858 if (!ptep)
2859 continue;
Chen, Kenneth W39dde652006-12-06 20:32:03 -08002860 if (huge_pmd_unshare(mm, &address, ptep))
2861 continue;
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07002862 if (!huge_pte_none(huge_ptep_get(ptep))) {
Zhang, Yanmin8f860592006-03-22 00:08:50 -08002863 pte = huge_ptep_get_and_clear(mm, address, ptep);
2864 pte = pte_mkhuge(pte_modify(pte, newprot));
2865 set_huge_pte_at(mm, address, ptep, pte);
Zhang, Yanmin8f860592006-03-22 00:08:50 -08002866 }
2867 }
2868 spin_unlock(&mm->page_table_lock);
Chen, Kenneth W39dde652006-12-06 20:32:03 -08002869 spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock);
Zhang, Yanmin8f860592006-03-22 00:08:50 -08002870
2871 flush_tlb_range(vma, start, end);
2872}
2873
Mel Gormana1e78772008-07-23 21:27:23 -07002874int hugetlb_reserve_pages(struct inode *inode,
2875 long from, long to,
Mel Gorman5a6fe122009-02-10 14:02:27 +00002876 struct vm_area_struct *vma,
2877 int acctflag)
Adam Litkee4e574b2007-10-16 01:26:19 -07002878{
Mel Gorman17c9d122009-02-11 16:34:16 +00002879 long ret, chg;
Andi Kleena5516432008-07-23 21:27:41 -07002880 struct hstate *h = hstate_inode(inode);
Adam Litkee4e574b2007-10-16 01:26:19 -07002881
Mel Gormana1e78772008-07-23 21:27:23 -07002882 /*
Mel Gorman17c9d122009-02-11 16:34:16 +00002883 * Only apply hugepage reservation if asked. At fault time, an
2884 * attempt will be made for VM_NORESERVE to allocate a page
2885 * and filesystem quota without using reserves
2886 */
2887 if (acctflag & VM_NORESERVE)
2888 return 0;
2889
2890 /*
Mel Gormana1e78772008-07-23 21:27:23 -07002891 * Shared mappings base their reservation on the number of pages that
2892 * are already allocated on behalf of the file. Private mappings need
2893 * to reserve the full area even if read-only as mprotect() may be
2894 * called to make the mapping read-write. Assume !vma is a shm mapping
2895 */
Mel Gormanf83a2752009-05-28 14:34:40 -07002896 if (!vma || vma->vm_flags & VM_MAYSHARE)
Mel Gormana1e78772008-07-23 21:27:23 -07002897 chg = region_chg(&inode->i_mapping->private_list, from, to);
Mel Gorman5a6fe122009-02-10 14:02:27 +00002898 else {
2899 struct resv_map *resv_map = resv_map_alloc();
Mel Gorman5a6fe122009-02-10 14:02:27 +00002900 if (!resv_map)
2901 return -ENOMEM;
2902
Mel Gorman17c9d122009-02-11 16:34:16 +00002903 chg = to - from;
2904
Mel Gorman5a6fe122009-02-10 14:02:27 +00002905 set_vma_resv_map(vma, resv_map);
2906 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
2907 }
2908
Mel Gorman17c9d122009-02-11 16:34:16 +00002909 if (chg < 0)
2910 return chg;
2911
2912 /* There must be enough filesystem quota for the mapping */
2913 if (hugetlb_get_quota(inode->i_mapping, chg))
2914 return -ENOSPC;
2915
2916 /*
2917 * Check enough hugepages are available for the reservation.
2918 * Hand back the quota if there are not
2919 */
2920 ret = hugetlb_acct_memory(h, chg);
2921 if (ret < 0) {
2922 hugetlb_put_quota(inode->i_mapping, chg);
2923 return ret;
2924 }
2925
2926 /*
2927 * Account for the reservations made. Shared mappings record regions
2928 * that have reservations as they are shared by multiple VMAs.
2929 * When the last VMA disappears, the region map says how much
2930 * the reservation was and the page cache tells how much of
2931 * the reservation was consumed. Private mappings are per-VMA and
2932 * only the consumed reservations are tracked. When the VMA
2933 * disappears, the original reservation is the VMA size and the
2934 * consumed reservations are stored in the map. Hence, nothing
2935 * else has to be done for private mappings here
2936 */
Mel Gormanf83a2752009-05-28 14:34:40 -07002937 if (!vma || vma->vm_flags & VM_MAYSHARE)
Mel Gorman17c9d122009-02-11 16:34:16 +00002938 region_add(&inode->i_mapping->private_list, from, to);
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -07002939 return 0;
2940}
2941
2942void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed)
2943{
Andi Kleena5516432008-07-23 21:27:41 -07002944 struct hstate *h = hstate_inode(inode);
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -07002945 long chg = region_truncate(&inode->i_mapping->private_list, offset);
Ken Chen45c682a2007-11-14 16:59:44 -08002946
2947 spin_lock(&inode->i_lock);
Eric Sandeene4c6f8b2009-07-29 15:02:16 -07002948 inode->i_blocks -= (blocks_per_huge_page(h) * freed);
Ken Chen45c682a2007-11-14 16:59:44 -08002949 spin_unlock(&inode->i_lock);
2950
Adam Litke90d8b7e2007-11-14 16:59:42 -08002951 hugetlb_put_quota(inode->i_mapping, (chg - freed));
Andi Kleena5516432008-07-23 21:27:41 -07002952 hugetlb_acct_memory(h, -(chg - freed));
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -07002953}
Naoya Horiguchi93f70f92010-05-28 09:29:20 +09002954
Andi Kleend5bd9102010-09-27 09:00:12 +02002955#ifdef CONFIG_MEMORY_FAILURE
2956
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +09002957/* Should be called in hugetlb_lock */
2958static int is_hugepage_on_freelist(struct page *hpage)
2959{
2960 struct page *page;
2961 struct page *tmp;
2962 struct hstate *h = page_hstate(hpage);
2963 int nid = page_to_nid(hpage);
2964
2965 list_for_each_entry_safe(page, tmp, &h->hugepage_freelists[nid], lru)
2966 if (page == hpage)
2967 return 1;
2968 return 0;
2969}
2970
Naoya Horiguchi93f70f92010-05-28 09:29:20 +09002971/*
2972 * This function is called from memory failure code.
2973 * Assume the caller holds page lock of the head page.
2974 */
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +09002975int dequeue_hwpoisoned_huge_page(struct page *hpage)
Naoya Horiguchi93f70f92010-05-28 09:29:20 +09002976{
2977 struct hstate *h = page_hstate(hpage);
2978 int nid = page_to_nid(hpage);
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +09002979 int ret = -EBUSY;
Naoya Horiguchi93f70f92010-05-28 09:29:20 +09002980
2981 spin_lock(&hugetlb_lock);
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +09002982 if (is_hugepage_on_freelist(hpage)) {
2983 list_del(&hpage->lru);
Naoya Horiguchi8c6c2ec2010-09-08 10:19:38 +09002984 set_page_refcounted(hpage);
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +09002985 h->free_huge_pages--;
2986 h->free_huge_pages_node[nid]--;
2987 ret = 0;
2988 }
Naoya Horiguchi93f70f92010-05-28 09:29:20 +09002989 spin_unlock(&hugetlb_lock);
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +09002990 return ret;
Naoya Horiguchi93f70f92010-05-28 09:29:20 +09002991}
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +09002992#endif