blob: bbb4a5bbb9582055945237692b8938a731c5492a [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
2 * Generic hugetlb support.
3 * (C) William Irwin, April 2004
4 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07005#include <linux/list.h>
6#include <linux/init.h>
7#include <linux/module.h>
8#include <linux/mm.h>
Alexey Dobriyane1759c22008-10-15 23:50:22 +04009#include <linux/seq_file.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070010#include <linux/sysctl.h>
11#include <linux/highmem.h>
Andrea Arcangelicddb8a52008-07-28 15:46:29 -070012#include <linux/mmu_notifier.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070013#include <linux/nodemask.h>
David Gibson63551ae2005-06-21 17:14:44 -070014#include <linux/pagemap.h>
Christoph Lameter5da7ca82006-01-06 00:10:46 -080015#include <linux/mempolicy.h>
Christoph Lameteraea47ff2006-01-08 01:00:57 -080016#include <linux/cpuset.h>
David Gibson3935baa2006-03-22 00:08:53 -080017#include <linux/mutex.h>
Andi Kleenaa888a72008-07-23 21:27:47 -070018#include <linux/bootmem.h>
Nishanth Aravamudana3437872008-07-23 21:27:44 -070019#include <linux/sysfs.h>
Tejun Heo5a0e3ad2010-03-24 17:04:11 +090020#include <linux/slab.h>
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +090021#include <linux/rmap.h>
Naoya Horiguchifd6a03ed2010-05-28 09:29:21 +090022#include <linux/swap.h>
23#include <linux/swapops.h>
Linus Torvaldsd6606682008-08-06 12:04:54 -070024
David Gibson63551ae2005-06-21 17:14:44 -070025#include <asm/page.h>
26#include <asm/pgtable.h>
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
Lucas De Marchi25985ed2011-03-30 22:57:33 -0300149 /* We overlap with this area, if it extends further than
Andy Whitcroft96822902008-07-23 21:27:29 -0700150 * us then we must extend ourselves. Account for its
151 * existing reservation. */
152 if (rg->to > t) {
153 chg += rg->to - t;
154 t = rg->to;
155 }
156 chg -= rg->to - rg->from;
157 }
158 return chg;
159}
160
161static long region_truncate(struct list_head *head, long end)
162{
163 struct file_region *rg, *trg;
164 long chg = 0;
165
166 /* Locate the region we are either in or before. */
167 list_for_each_entry(rg, head, link)
168 if (end <= rg->to)
169 break;
170 if (&rg->link == head)
171 return 0;
172
173 /* If we are in the middle of a region then adjust it. */
174 if (end > rg->from) {
175 chg = rg->to - end;
176 rg->to = end;
177 rg = list_entry(rg->link.next, typeof(*rg), link);
178 }
179
180 /* Drop any remaining regions. */
181 list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
182 if (&rg->link == head)
183 break;
184 chg += rg->to - rg->from;
185 list_del(&rg->link);
186 kfree(rg);
187 }
188 return chg;
189}
190
Andy Whitcroft84afd992008-07-23 21:27:32 -0700191static long region_count(struct list_head *head, long f, long t)
192{
193 struct file_region *rg;
194 long chg = 0;
195
196 /* Locate each segment we overlap with, and count that overlap. */
197 list_for_each_entry(rg, head, link) {
198 int seg_from;
199 int seg_to;
200
201 if (rg->to <= f)
202 continue;
203 if (rg->from >= t)
204 break;
205
206 seg_from = max(rg->from, f);
207 seg_to = min(rg->to, t);
208
209 chg += seg_to - seg_from;
210 }
211
212 return chg;
213}
214
Andy Whitcroft96822902008-07-23 21:27:29 -0700215/*
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700216 * Convert the address within this vma to the page offset within
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700217 * the mapping, in pagecache page units; huge pages here.
218 */
Andi Kleena5516432008-07-23 21:27:41 -0700219static pgoff_t vma_hugecache_offset(struct hstate *h,
220 struct vm_area_struct *vma, unsigned long address)
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700221{
Andi Kleena5516432008-07-23 21:27:41 -0700222 return ((address - vma->vm_start) >> huge_page_shift(h)) +
223 (vma->vm_pgoff >> huge_page_order(h));
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700224}
225
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +0900226pgoff_t linear_hugepage_index(struct vm_area_struct *vma,
227 unsigned long address)
228{
229 return vma_hugecache_offset(hstate_vma(vma), vma, address);
230}
231
Andy Whitcroft84afd992008-07-23 21:27:32 -0700232/*
Mel Gorman08fba692009-01-06 14:38:53 -0800233 * Return the size of the pages allocated when backing a VMA. In the majority
234 * cases this will be same size as used by the page table entries.
235 */
236unsigned long vma_kernel_pagesize(struct vm_area_struct *vma)
237{
238 struct hstate *hstate;
239
240 if (!is_vm_hugetlb_page(vma))
241 return PAGE_SIZE;
242
243 hstate = hstate_vma(vma);
244
245 return 1UL << (hstate->order + PAGE_SHIFT);
246}
Joerg Roedelf340ca02009-06-19 15:16:22 +0200247EXPORT_SYMBOL_GPL(vma_kernel_pagesize);
Mel Gorman08fba692009-01-06 14:38:53 -0800248
249/*
Mel Gorman33402892009-01-06 14:38:54 -0800250 * Return the page size being used by the MMU to back a VMA. In the majority
251 * of cases, the page size used by the kernel matches the MMU size. On
252 * architectures where it differs, an architecture-specific version of this
253 * function is required.
254 */
255#ifndef vma_mmu_pagesize
256unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
257{
258 return vma_kernel_pagesize(vma);
259}
260#endif
261
262/*
Andy Whitcroft84afd992008-07-23 21:27:32 -0700263 * Flags for MAP_PRIVATE reservations. These are stored in the bottom
264 * bits of the reservation map pointer, which are always clear due to
265 * alignment.
266 */
267#define HPAGE_RESV_OWNER (1UL << 0)
268#define HPAGE_RESV_UNMAPPED (1UL << 1)
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700269#define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
Andy Whitcroft84afd992008-07-23 21:27:32 -0700270
Mel Gormana1e78772008-07-23 21:27:23 -0700271/*
272 * These helpers are used to track how many pages are reserved for
273 * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
274 * is guaranteed to have their future faults succeed.
275 *
276 * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
277 * the reserve counters are updated with the hugetlb_lock held. It is safe
278 * to reset the VMA at fork() time as it is not in use yet and there is no
279 * chance of the global counters getting corrupted as a result of the values.
Andy Whitcroft84afd992008-07-23 21:27:32 -0700280 *
281 * The private mapping reservation is represented in a subtly different
282 * manner to a shared mapping. A shared mapping has a region map associated
283 * with the underlying file, this region map represents the backing file
284 * pages which have ever had a reservation assigned which this persists even
285 * after the page is instantiated. A private mapping has a region map
286 * associated with the original mmap which is attached to all VMAs which
287 * reference it, this region map represents those offsets which have consumed
288 * reservation ie. where pages have been instantiated.
Mel Gormana1e78772008-07-23 21:27:23 -0700289 */
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700290static unsigned long get_vma_private_data(struct vm_area_struct *vma)
291{
292 return (unsigned long)vma->vm_private_data;
293}
294
295static void set_vma_private_data(struct vm_area_struct *vma,
296 unsigned long value)
297{
298 vma->vm_private_data = (void *)value;
299}
300
Andy Whitcroft84afd992008-07-23 21:27:32 -0700301struct resv_map {
302 struct kref refs;
303 struct list_head regions;
304};
305
Harvey Harrison2a4b3de2008-10-18 20:27:06 -0700306static struct resv_map *resv_map_alloc(void)
Andy Whitcroft84afd992008-07-23 21:27:32 -0700307{
308 struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
309 if (!resv_map)
310 return NULL;
311
312 kref_init(&resv_map->refs);
313 INIT_LIST_HEAD(&resv_map->regions);
314
315 return resv_map;
316}
317
Harvey Harrison2a4b3de2008-10-18 20:27:06 -0700318static void resv_map_release(struct kref *ref)
Andy Whitcroft84afd992008-07-23 21:27:32 -0700319{
320 struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
321
322 /* Clear out any active regions before we release the map. */
323 region_truncate(&resv_map->regions, 0);
324 kfree(resv_map);
325}
326
327static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
Mel Gormana1e78772008-07-23 21:27:23 -0700328{
329 VM_BUG_ON(!is_vm_hugetlb_page(vma));
Mel Gormanf83a2752009-05-28 14:34:40 -0700330 if (!(vma->vm_flags & VM_MAYSHARE))
Andy Whitcroft84afd992008-07-23 21:27:32 -0700331 return (struct resv_map *)(get_vma_private_data(vma) &
332 ~HPAGE_RESV_MASK);
Harvey Harrison2a4b3de2008-10-18 20:27:06 -0700333 return NULL;
Mel Gormana1e78772008-07-23 21:27:23 -0700334}
335
Andy Whitcroft84afd992008-07-23 21:27:32 -0700336static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
Mel Gormana1e78772008-07-23 21:27:23 -0700337{
338 VM_BUG_ON(!is_vm_hugetlb_page(vma));
Mel Gormanf83a2752009-05-28 14:34:40 -0700339 VM_BUG_ON(vma->vm_flags & VM_MAYSHARE);
Mel Gormana1e78772008-07-23 21:27:23 -0700340
Andy Whitcroft84afd992008-07-23 21:27:32 -0700341 set_vma_private_data(vma, (get_vma_private_data(vma) &
342 HPAGE_RESV_MASK) | (unsigned long)map);
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700343}
344
345static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
346{
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700347 VM_BUG_ON(!is_vm_hugetlb_page(vma));
Mel Gormanf83a2752009-05-28 14:34:40 -0700348 VM_BUG_ON(vma->vm_flags & VM_MAYSHARE);
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700349
350 set_vma_private_data(vma, get_vma_private_data(vma) | flags);
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700351}
352
353static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
354{
355 VM_BUG_ON(!is_vm_hugetlb_page(vma));
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700356
357 return (get_vma_private_data(vma) & flag) != 0;
Mel Gormana1e78772008-07-23 21:27:23 -0700358}
359
360/* Decrement the reserved pages in the hugepage pool by one */
Andi Kleena5516432008-07-23 21:27:41 -0700361static void decrement_hugepage_resv_vma(struct hstate *h,
362 struct vm_area_struct *vma)
Mel Gormana1e78772008-07-23 21:27:23 -0700363{
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -0700364 if (vma->vm_flags & VM_NORESERVE)
365 return;
366
Mel Gormanf83a2752009-05-28 14:34:40 -0700367 if (vma->vm_flags & VM_MAYSHARE) {
Mel Gormana1e78772008-07-23 21:27:23 -0700368 /* Shared mappings always use reserves */
Andi Kleena5516432008-07-23 21:27:41 -0700369 h->resv_huge_pages--;
Andy Whitcroft84afd992008-07-23 21:27:32 -0700370 } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
Mel Gormana1e78772008-07-23 21:27:23 -0700371 /*
372 * Only the process that called mmap() has reserves for
373 * private mappings.
374 */
Andi Kleena5516432008-07-23 21:27:41 -0700375 h->resv_huge_pages--;
Mel Gormana1e78772008-07-23 21:27:23 -0700376 }
377}
378
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700379/* Reset counters to 0 and clear all HPAGE_RESV_* flags */
Mel Gormana1e78772008-07-23 21:27:23 -0700380void reset_vma_resv_huge_pages(struct vm_area_struct *vma)
381{
382 VM_BUG_ON(!is_vm_hugetlb_page(vma));
Mel Gormanf83a2752009-05-28 14:34:40 -0700383 if (!(vma->vm_flags & VM_MAYSHARE))
Mel Gormana1e78772008-07-23 21:27:23 -0700384 vma->vm_private_data = (void *)0;
385}
386
387/* Returns true if the VMA has associated reserve pages */
Mel Gorman7f09ca52008-07-23 21:27:58 -0700388static int vma_has_reserves(struct vm_area_struct *vma)
Mel Gormana1e78772008-07-23 21:27:23 -0700389{
Mel Gormanf83a2752009-05-28 14:34:40 -0700390 if (vma->vm_flags & VM_MAYSHARE)
Mel Gorman7f09ca52008-07-23 21:27:58 -0700391 return 1;
392 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER))
393 return 1;
394 return 0;
Mel Gormana1e78772008-07-23 21:27:23 -0700395}
396
Naoya Horiguchi0ebabb42010-09-08 10:19:34 +0900397static void copy_gigantic_page(struct page *dst, struct page *src)
398{
399 int i;
400 struct hstate *h = page_hstate(src);
401 struct page *dst_base = dst;
402 struct page *src_base = src;
403
404 for (i = 0; i < pages_per_huge_page(h); ) {
405 cond_resched();
406 copy_highpage(dst, src);
407
408 i++;
409 dst = mem_map_next(dst, dst_base, i);
410 src = mem_map_next(src, src_base, i);
411 }
412}
413
414void copy_huge_page(struct page *dst, struct page *src)
415{
416 int i;
417 struct hstate *h = page_hstate(src);
418
419 if (unlikely(pages_per_huge_page(h) > MAX_ORDER_NR_PAGES)) {
420 copy_gigantic_page(dst, src);
421 return;
422 }
423
424 might_sleep();
425 for (i = 0; i < pages_per_huge_page(h); i++) {
426 cond_resched();
427 copy_highpage(dst + i, src + i);
428 }
429}
430
Andi Kleena5516432008-07-23 21:27:41 -0700431static void enqueue_huge_page(struct hstate *h, struct page *page)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700432{
433 int nid = page_to_nid(page);
Andi Kleena5516432008-07-23 21:27:41 -0700434 list_add(&page->lru, &h->hugepage_freelists[nid]);
435 h->free_huge_pages++;
436 h->free_huge_pages_node[nid]++;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700437}
438
Naoya Horiguchibf50bab2010-09-08 10:19:33 +0900439static struct page *dequeue_huge_page_node(struct hstate *h, int nid)
440{
441 struct page *page;
442
443 if (list_empty(&h->hugepage_freelists[nid]))
444 return NULL;
445 page = list_entry(h->hugepage_freelists[nid].next, struct page, lru);
446 list_del(&page->lru);
Naoya Horiguchia9869b82010-09-08 10:19:37 +0900447 set_page_refcounted(page);
Naoya Horiguchibf50bab2010-09-08 10:19:33 +0900448 h->free_huge_pages--;
449 h->free_huge_pages_node[nid]--;
450 return page;
451}
452
Andi Kleena5516432008-07-23 21:27:41 -0700453static struct page *dequeue_huge_page_vma(struct hstate *h,
454 struct vm_area_struct *vma,
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700455 unsigned long address, int avoid_reserve)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700456{
Linus Torvalds1da177e2005-04-16 15:20:36 -0700457 struct page *page = NULL;
Lee Schermerhorn480eccf2007-09-18 22:46:47 -0700458 struct mempolicy *mpol;
Mel Gorman19770b32008-04-28 02:12:18 -0700459 nodemask_t *nodemask;
Miao Xiec0ff7452010-05-24 14:32:08 -0700460 struct zonelist *zonelist;
Mel Gormandd1a2392008-04-28 02:12:17 -0700461 struct zone *zone;
462 struct zoneref *z;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700463
Miao Xiec0ff7452010-05-24 14:32:08 -0700464 get_mems_allowed();
465 zonelist = huge_zonelist(vma, address,
466 htlb_alloc_mask, &mpol, &nodemask);
Mel Gormana1e78772008-07-23 21:27:23 -0700467 /*
468 * A child process with MAP_PRIVATE mappings created by their parent
469 * have no page reserves. This check ensures that reservations are
470 * not "stolen". The child may still get SIGKILLed
471 */
Mel Gorman7f09ca52008-07-23 21:27:58 -0700472 if (!vma_has_reserves(vma) &&
Andi Kleena5516432008-07-23 21:27:41 -0700473 h->free_huge_pages - h->resv_huge_pages == 0)
Miao Xiec0ff7452010-05-24 14:32:08 -0700474 goto err;
Mel Gormana1e78772008-07-23 21:27:23 -0700475
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700476 /* If reserves cannot be used, ensure enough pages are in the pool */
Andi Kleena5516432008-07-23 21:27:41 -0700477 if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
Justin P. Mattock6eab04a2011-04-08 19:49:08 -0700478 goto err;
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700479
Mel Gorman19770b32008-04-28 02:12:18 -0700480 for_each_zone_zonelist_nodemask(zone, z, zonelist,
481 MAX_NR_ZONES - 1, nodemask) {
Naoya Horiguchibf50bab2010-09-08 10:19:33 +0900482 if (cpuset_zone_allowed_softwall(zone, htlb_alloc_mask)) {
483 page = dequeue_huge_page_node(h, zone_to_nid(zone));
484 if (page) {
485 if (!avoid_reserve)
486 decrement_hugepage_resv_vma(h, vma);
487 break;
488 }
Andrew Morton3abf7af2007-07-19 01:49:08 -0700489 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700490 }
Miao Xiec0ff7452010-05-24 14:32:08 -0700491err:
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -0700492 mpol_cond_put(mpol);
Miao Xiec0ff7452010-05-24 14:32:08 -0700493 put_mems_allowed();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700494 return page;
495}
496
Andi Kleena5516432008-07-23 21:27:41 -0700497static void update_and_free_page(struct hstate *h, struct page *page)
Adam Litke6af2acb2007-10-16 01:26:16 -0700498{
499 int i;
Andi Kleena5516432008-07-23 21:27:41 -0700500
Andy Whitcroft18229df2008-11-06 12:53:27 -0800501 VM_BUG_ON(h->order >= MAX_ORDER);
502
Andi Kleena5516432008-07-23 21:27:41 -0700503 h->nr_huge_pages--;
504 h->nr_huge_pages_node[page_to_nid(page)]--;
505 for (i = 0; i < pages_per_huge_page(h); i++) {
Adam Litke6af2acb2007-10-16 01:26:16 -0700506 page[i].flags &= ~(1 << PG_locked | 1 << PG_error | 1 << PG_referenced |
507 1 << PG_dirty | 1 << PG_active | 1 << PG_reserved |
508 1 << PG_private | 1<< PG_writeback);
509 }
510 set_compound_page_dtor(page, NULL);
511 set_page_refcounted(page);
Gerald Schaefer7f2e9522008-04-28 02:13:29 -0700512 arch_release_hugepage(page);
Andi Kleena5516432008-07-23 21:27:41 -0700513 __free_pages(page, huge_page_order(h));
Adam Litke6af2acb2007-10-16 01:26:16 -0700514}
515
Andi Kleene5ff2152008-07-23 21:27:42 -0700516struct hstate *size_to_hstate(unsigned long size)
517{
518 struct hstate *h;
519
520 for_each_hstate(h) {
521 if (huge_page_size(h) == size)
522 return h;
523 }
524 return NULL;
525}
526
David Gibson27a85ef2006-03-22 00:08:56 -0800527static void free_huge_page(struct page *page)
528{
Andi Kleena5516432008-07-23 21:27:41 -0700529 /*
530 * Can't pass hstate in here because it is called from the
531 * compound page destructor.
532 */
Andi Kleene5ff2152008-07-23 21:27:42 -0700533 struct hstate *h = page_hstate(page);
Adam Litke7893d1d2007-10-16 01:26:18 -0700534 int nid = page_to_nid(page);
Adam Litkec79fb752007-11-14 16:59:38 -0800535 struct address_space *mapping;
David Gibson27a85ef2006-03-22 00:08:56 -0800536
Adam Litkec79fb752007-11-14 16:59:38 -0800537 mapping = (struct address_space *) page_private(page);
Andy Whitcrofte5df70a2008-02-23 15:23:32 -0800538 set_page_private(page, 0);
Mel Gorman23be7462010-04-23 13:17:56 -0400539 page->mapping = NULL;
Adam Litke7893d1d2007-10-16 01:26:18 -0700540 BUG_ON(page_count(page));
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +0900541 BUG_ON(page_mapcount(page));
David Gibson27a85ef2006-03-22 00:08:56 -0800542 INIT_LIST_HEAD(&page->lru);
543
544 spin_lock(&hugetlb_lock);
Andi Kleenaa888a72008-07-23 21:27:47 -0700545 if (h->surplus_huge_pages_node[nid] && huge_page_order(h) < MAX_ORDER) {
Andi Kleena5516432008-07-23 21:27:41 -0700546 update_and_free_page(h, page);
547 h->surplus_huge_pages--;
548 h->surplus_huge_pages_node[nid]--;
Adam Litke7893d1d2007-10-16 01:26:18 -0700549 } else {
Andi Kleena5516432008-07-23 21:27:41 -0700550 enqueue_huge_page(h, page);
Adam Litke7893d1d2007-10-16 01:26:18 -0700551 }
David Gibson27a85ef2006-03-22 00:08:56 -0800552 spin_unlock(&hugetlb_lock);
Adam Litkec79fb752007-11-14 16:59:38 -0800553 if (mapping)
Adam Litke9a119c02007-11-14 16:59:41 -0800554 hugetlb_put_quota(mapping, 1);
David Gibson27a85ef2006-03-22 00:08:56 -0800555}
556
Andi Kleena5516432008-07-23 21:27:41 -0700557static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
Andi Kleenb7ba30c2008-07-23 21:27:40 -0700558{
559 set_compound_page_dtor(page, free_huge_page);
560 spin_lock(&hugetlb_lock);
Andi Kleena5516432008-07-23 21:27:41 -0700561 h->nr_huge_pages++;
562 h->nr_huge_pages_node[nid]++;
Andi Kleenb7ba30c2008-07-23 21:27:40 -0700563 spin_unlock(&hugetlb_lock);
564 put_page(page); /* free it into the hugepage allocator */
565}
566
Wu Fengguang20a03072009-06-16 15:32:22 -0700567static void prep_compound_gigantic_page(struct page *page, unsigned long order)
568{
569 int i;
570 int nr_pages = 1 << order;
571 struct page *p = page + 1;
572
573 /* we rely on prep_new_huge_page to set the destructor */
574 set_compound_order(page, order);
575 __SetPageHead(page);
576 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
577 __SetPageTail(p);
578 p->first_page = page;
579 }
580}
581
582int PageHuge(struct page *page)
583{
584 compound_page_dtor *dtor;
585
586 if (!PageCompound(page))
587 return 0;
588
589 page = compound_head(page);
590 dtor = get_compound_page_dtor(page);
591
592 return dtor == free_huge_page;
593}
594
Naoya Horiguchi43131e12010-05-28 09:29:22 +0900595EXPORT_SYMBOL_GPL(PageHuge);
596
Andi Kleena5516432008-07-23 21:27:41 -0700597static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700598{
Linus Torvalds1da177e2005-04-16 15:20:36 -0700599 struct page *page;
Joe Jinf96efd52007-07-15 23:38:12 -0700600
Andi Kleenaa888a72008-07-23 21:27:47 -0700601 if (h->order >= MAX_ORDER)
602 return NULL;
603
Mel Gorman6484eb32009-06-16 15:31:54 -0700604 page = alloc_pages_exact_node(nid,
Nishanth Aravamudan551883a2008-04-29 00:58:26 -0700605 htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE|
606 __GFP_REPEAT|__GFP_NOWARN,
Andi Kleena5516432008-07-23 21:27:41 -0700607 huge_page_order(h));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700608 if (page) {
Gerald Schaefer7f2e9522008-04-28 02:13:29 -0700609 if (arch_prepare_hugepage(page)) {
Gerald Schaefercaff3a22008-08-12 15:08:38 -0700610 __free_pages(page, huge_page_order(h));
Harvey Harrison7b8ee842008-04-28 14:13:19 -0700611 return NULL;
Gerald Schaefer7f2e9522008-04-28 02:13:29 -0700612 }
Andi Kleena5516432008-07-23 21:27:41 -0700613 prep_new_huge_page(h, page, nid);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700614 }
Nishanth Aravamudan63b46132007-10-16 01:26:24 -0700615
616 return page;
617}
618
Andi Kleen5ced66c2008-07-23 21:27:45 -0700619/*
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800620 * common helper functions for hstate_next_node_to_{alloc|free}.
621 * We may have allocated or freed a huge page based on a different
622 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
623 * be outside of *nodes_allowed. Ensure that we use an allowed
624 * node for alloc or free.
Lee Schermerhorn9a76db02009-12-14 17:58:15 -0800625 */
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800626static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
Lee Schermerhorn9a76db02009-12-14 17:58:15 -0800627{
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800628 nid = next_node(nid, *nodes_allowed);
Lee Schermerhorn9a76db02009-12-14 17:58:15 -0800629 if (nid == MAX_NUMNODES)
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800630 nid = first_node(*nodes_allowed);
Lee Schermerhorn9a76db02009-12-14 17:58:15 -0800631 VM_BUG_ON(nid >= MAX_NUMNODES);
632
633 return nid;
634}
635
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800636static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
Andi Kleen5ced66c2008-07-23 21:27:45 -0700637{
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800638 if (!node_isset(nid, *nodes_allowed))
639 nid = next_node_allowed(nid, nodes_allowed);
Lee Schermerhorn9a76db02009-12-14 17:58:15 -0800640 return nid;
Andi Kleen5ced66c2008-07-23 21:27:45 -0700641}
642
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800643/*
644 * returns the previously saved node ["this node"] from which to
645 * allocate a persistent huge page for the pool and advance the
646 * next node from which to allocate, handling wrap at end of node
647 * mask.
648 */
649static int hstate_next_node_to_alloc(struct hstate *h,
650 nodemask_t *nodes_allowed)
651{
652 int nid;
653
654 VM_BUG_ON(!nodes_allowed);
655
656 nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
657 h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
658
659 return nid;
660}
661
662static int alloc_fresh_huge_page(struct hstate *h, nodemask_t *nodes_allowed)
Nishanth Aravamudan63b46132007-10-16 01:26:24 -0700663{
664 struct page *page;
665 int start_nid;
666 int next_nid;
667 int ret = 0;
668
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800669 start_nid = hstate_next_node_to_alloc(h, nodes_allowed);
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700670 next_nid = start_nid;
Nishanth Aravamudan63b46132007-10-16 01:26:24 -0700671
672 do {
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700673 page = alloc_fresh_huge_page_node(h, next_nid);
Lee Schermerhorn9a76db02009-12-14 17:58:15 -0800674 if (page) {
Nishanth Aravamudan63b46132007-10-16 01:26:24 -0700675 ret = 1;
Lee Schermerhorn9a76db02009-12-14 17:58:15 -0800676 break;
677 }
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800678 next_nid = hstate_next_node_to_alloc(h, nodes_allowed);
Lee Schermerhorn9a76db02009-12-14 17:58:15 -0800679 } while (next_nid != start_nid);
Nishanth Aravamudan63b46132007-10-16 01:26:24 -0700680
Adam Litke3b116302008-04-28 02:13:06 -0700681 if (ret)
682 count_vm_event(HTLB_BUDDY_PGALLOC);
683 else
684 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
685
Nishanth Aravamudan63b46132007-10-16 01:26:24 -0700686 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700687}
688
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700689/*
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800690 * helper for free_pool_huge_page() - return the previously saved
691 * node ["this node"] from which to free a huge page. Advance the
692 * next node id whether or not we find a free huge page to free so
693 * that the next attempt to free addresses the next node.
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700694 */
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800695static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700696{
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800697 int nid;
Lee Schermerhorn9a76db02009-12-14 17:58:15 -0800698
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800699 VM_BUG_ON(!nodes_allowed);
700
701 nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
702 h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
703
Lee Schermerhorn9a76db02009-12-14 17:58:15 -0800704 return nid;
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700705}
706
707/*
708 * Free huge page from pool from next node to free.
709 * Attempt to keep persistent huge pages more or less
710 * balanced over allowed nodes.
711 * Called with hugetlb_lock locked.
712 */
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800713static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
714 bool acct_surplus)
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700715{
716 int start_nid;
717 int next_nid;
718 int ret = 0;
719
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800720 start_nid = hstate_next_node_to_free(h, nodes_allowed);
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700721 next_nid = start_nid;
722
723 do {
Lee Schermerhorn685f3452009-09-21 17:01:23 -0700724 /*
725 * If we're returning unused surplus pages, only examine
726 * nodes with surplus pages.
727 */
728 if ((!acct_surplus || h->surplus_huge_pages_node[next_nid]) &&
729 !list_empty(&h->hugepage_freelists[next_nid])) {
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700730 struct page *page =
731 list_entry(h->hugepage_freelists[next_nid].next,
732 struct page, lru);
733 list_del(&page->lru);
734 h->free_huge_pages--;
735 h->free_huge_pages_node[next_nid]--;
Lee Schermerhorn685f3452009-09-21 17:01:23 -0700736 if (acct_surplus) {
737 h->surplus_huge_pages--;
738 h->surplus_huge_pages_node[next_nid]--;
739 }
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700740 update_and_free_page(h, page);
741 ret = 1;
Lee Schermerhorn9a76db02009-12-14 17:58:15 -0800742 break;
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700743 }
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -0800744 next_nid = hstate_next_node_to_free(h, nodes_allowed);
Lee Schermerhorn9a76db02009-12-14 17:58:15 -0800745 } while (next_nid != start_nid);
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700746
747 return ret;
748}
749
Naoya Horiguchibf50bab2010-09-08 10:19:33 +0900750static struct page *alloc_buddy_huge_page(struct hstate *h, int nid)
Adam Litke7893d1d2007-10-16 01:26:18 -0700751{
752 struct page *page;
Naoya Horiguchibf50bab2010-09-08 10:19:33 +0900753 unsigned int r_nid;
Adam Litke7893d1d2007-10-16 01:26:18 -0700754
Andi Kleenaa888a72008-07-23 21:27:47 -0700755 if (h->order >= MAX_ORDER)
756 return NULL;
757
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800758 /*
759 * Assume we will successfully allocate the surplus page to
760 * prevent racing processes from causing the surplus to exceed
761 * overcommit
762 *
763 * This however introduces a different race, where a process B
764 * tries to grow the static hugepage pool while alloc_pages() is
765 * called by process A. B will only examine the per-node
766 * counters in determining if surplus huge pages can be
767 * converted to normal huge pages in adjust_pool_surplus(). A
768 * won't be able to increment the per-node counter, until the
769 * lock is dropped by B, but B doesn't drop hugetlb_lock until
770 * no more huge pages can be converted from surplus to normal
771 * state (and doesn't try to convert again). Thus, we have a
772 * case where a surplus huge page exists, the pool is grown, and
773 * the surplus huge page still exists after, even though it
774 * should just have been converted to a normal huge page. This
775 * does not leak memory, though, as the hugepage will be freed
776 * once it is out of use. It also does not allow the counters to
777 * go out of whack in adjust_pool_surplus() as we don't modify
778 * the node values until we've gotten the hugepage and only the
779 * per-node value is checked there.
780 */
781 spin_lock(&hugetlb_lock);
Andi Kleena5516432008-07-23 21:27:41 -0700782 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800783 spin_unlock(&hugetlb_lock);
784 return NULL;
785 } else {
Andi Kleena5516432008-07-23 21:27:41 -0700786 h->nr_huge_pages++;
787 h->surplus_huge_pages++;
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800788 }
789 spin_unlock(&hugetlb_lock);
790
Naoya Horiguchibf50bab2010-09-08 10:19:33 +0900791 if (nid == NUMA_NO_NODE)
792 page = alloc_pages(htlb_alloc_mask|__GFP_COMP|
793 __GFP_REPEAT|__GFP_NOWARN,
794 huge_page_order(h));
795 else
796 page = alloc_pages_exact_node(nid,
797 htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE|
798 __GFP_REPEAT|__GFP_NOWARN, huge_page_order(h));
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800799
Gerald Schaefercaff3a22008-08-12 15:08:38 -0700800 if (page && arch_prepare_hugepage(page)) {
801 __free_pages(page, huge_page_order(h));
802 return NULL;
803 }
804
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800805 spin_lock(&hugetlb_lock);
Adam Litke7893d1d2007-10-16 01:26:18 -0700806 if (page) {
Naoya Horiguchibf50bab2010-09-08 10:19:33 +0900807 r_nid = page_to_nid(page);
Adam Litke7893d1d2007-10-16 01:26:18 -0700808 set_compound_page_dtor(page, free_huge_page);
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800809 /*
810 * We incremented the global counters already
811 */
Naoya Horiguchibf50bab2010-09-08 10:19:33 +0900812 h->nr_huge_pages_node[r_nid]++;
813 h->surplus_huge_pages_node[r_nid]++;
Adam Litke3b116302008-04-28 02:13:06 -0700814 __count_vm_event(HTLB_BUDDY_PGALLOC);
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800815 } else {
Andi Kleena5516432008-07-23 21:27:41 -0700816 h->nr_huge_pages--;
817 h->surplus_huge_pages--;
Adam Litke3b116302008-04-28 02:13:06 -0700818 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
Adam Litke7893d1d2007-10-16 01:26:18 -0700819 }
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800820 spin_unlock(&hugetlb_lock);
Adam Litke7893d1d2007-10-16 01:26:18 -0700821
822 return page;
823}
824
Adam Litkee4e574b2007-10-16 01:26:19 -0700825/*
Naoya Horiguchibf50bab2010-09-08 10:19:33 +0900826 * This allocation function is useful in the context where vma is irrelevant.
827 * E.g. soft-offlining uses this function because it only cares physical
828 * address of error page.
829 */
830struct page *alloc_huge_page_node(struct hstate *h, int nid)
831{
832 struct page *page;
833
834 spin_lock(&hugetlb_lock);
835 page = dequeue_huge_page_node(h, nid);
836 spin_unlock(&hugetlb_lock);
837
838 if (!page)
839 page = alloc_buddy_huge_page(h, nid);
840
841 return page;
842}
843
844/*
Lucas De Marchi25985ed2011-03-30 22:57:33 -0300845 * Increase the hugetlb pool such that it can accommodate a reservation
Adam Litkee4e574b2007-10-16 01:26:19 -0700846 * of size 'delta'.
847 */
Andi Kleena5516432008-07-23 21:27:41 -0700848static int gather_surplus_pages(struct hstate *h, int delta)
Adam Litkee4e574b2007-10-16 01:26:19 -0700849{
850 struct list_head surplus_list;
851 struct page *page, *tmp;
852 int ret, i;
853 int needed, allocated;
854
Andi Kleena5516432008-07-23 21:27:41 -0700855 needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
Adam Litkeac09b3a2008-03-04 14:29:38 -0800856 if (needed <= 0) {
Andi Kleena5516432008-07-23 21:27:41 -0700857 h->resv_huge_pages += delta;
Adam Litkee4e574b2007-10-16 01:26:19 -0700858 return 0;
Adam Litkeac09b3a2008-03-04 14:29:38 -0800859 }
Adam Litkee4e574b2007-10-16 01:26:19 -0700860
861 allocated = 0;
862 INIT_LIST_HEAD(&surplus_list);
863
864 ret = -ENOMEM;
865retry:
866 spin_unlock(&hugetlb_lock);
867 for (i = 0; i < needed; i++) {
Naoya Horiguchibf50bab2010-09-08 10:19:33 +0900868 page = alloc_buddy_huge_page(h, NUMA_NO_NODE);
Naoya Horiguchia9869b82010-09-08 10:19:37 +0900869 if (!page)
Adam Litkee4e574b2007-10-16 01:26:19 -0700870 /*
871 * We were not able to allocate enough pages to
872 * satisfy the entire reservation so we free what
873 * we've allocated so far.
874 */
Adam Litkee4e574b2007-10-16 01:26:19 -0700875 goto free;
Adam Litkee4e574b2007-10-16 01:26:19 -0700876
877 list_add(&page->lru, &surplus_list);
878 }
879 allocated += needed;
880
881 /*
882 * After retaking hugetlb_lock, we need to recalculate 'needed'
883 * because either resv_huge_pages or free_huge_pages may have changed.
884 */
885 spin_lock(&hugetlb_lock);
Andi Kleena5516432008-07-23 21:27:41 -0700886 needed = (h->resv_huge_pages + delta) -
887 (h->free_huge_pages + allocated);
Adam Litkee4e574b2007-10-16 01:26:19 -0700888 if (needed > 0)
889 goto retry;
890
891 /*
892 * The surplus_list now contains _at_least_ the number of extra pages
Lucas De Marchi25985ed2011-03-30 22:57:33 -0300893 * needed to accommodate the reservation. Add the appropriate number
Adam Litkee4e574b2007-10-16 01:26:19 -0700894 * of pages to the hugetlb pool and free the extras back to the buddy
Adam Litkeac09b3a2008-03-04 14:29:38 -0800895 * allocator. Commit the entire reservation here to prevent another
896 * process from stealing the pages as they are added to the pool but
897 * before they are reserved.
Adam Litkee4e574b2007-10-16 01:26:19 -0700898 */
899 needed += allocated;
Andi Kleena5516432008-07-23 21:27:41 -0700900 h->resv_huge_pages += delta;
Adam Litkee4e574b2007-10-16 01:26:19 -0700901 ret = 0;
Naoya Horiguchia9869b82010-09-08 10:19:37 +0900902
903 spin_unlock(&hugetlb_lock);
Adam Litke19fc3f02008-04-28 02:12:20 -0700904 /* Free the needed pages to the hugetlb pool */
Adam Litkee4e574b2007-10-16 01:26:19 -0700905 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
Adam Litke19fc3f02008-04-28 02:12:20 -0700906 if ((--needed) < 0)
907 break;
Adam Litkee4e574b2007-10-16 01:26:19 -0700908 list_del(&page->lru);
Naoya Horiguchia9869b82010-09-08 10:19:37 +0900909 /*
910 * This page is now managed by the hugetlb allocator and has
911 * no users -- drop the buddy allocator's reference.
912 */
913 put_page_testzero(page);
914 VM_BUG_ON(page_count(page));
Andi Kleena5516432008-07-23 21:27:41 -0700915 enqueue_huge_page(h, page);
Adam Litke19fc3f02008-04-28 02:12:20 -0700916 }
917
918 /* Free unnecessary surplus pages to the buddy allocator */
Naoya Horiguchia9869b82010-09-08 10:19:37 +0900919free:
Adam Litke19fc3f02008-04-28 02:12:20 -0700920 if (!list_empty(&surplus_list)) {
Adam Litke19fc3f02008-04-28 02:12:20 -0700921 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
922 list_del(&page->lru);
Naoya Horiguchia9869b82010-09-08 10:19:37 +0900923 put_page(page);
Adam Litkeaf767cb2007-10-16 01:26:25 -0700924 }
Adam Litkee4e574b2007-10-16 01:26:19 -0700925 }
Naoya Horiguchia9869b82010-09-08 10:19:37 +0900926 spin_lock(&hugetlb_lock);
Adam Litkee4e574b2007-10-16 01:26:19 -0700927
928 return ret;
929}
930
931/*
932 * When releasing a hugetlb pool reservation, any surplus pages that were
933 * allocated to satisfy the reservation must be explicitly freed if they were
934 * never used.
Lee Schermerhorn685f3452009-09-21 17:01:23 -0700935 * Called with hugetlb_lock held.
Adam Litkee4e574b2007-10-16 01:26:19 -0700936 */
Andi Kleena5516432008-07-23 21:27:41 -0700937static void return_unused_surplus_pages(struct hstate *h,
938 unsigned long unused_resv_pages)
Adam Litkee4e574b2007-10-16 01:26:19 -0700939{
Adam Litkee4e574b2007-10-16 01:26:19 -0700940 unsigned long nr_pages;
941
Adam Litkeac09b3a2008-03-04 14:29:38 -0800942 /* Uncommit the reservation */
Andi Kleena5516432008-07-23 21:27:41 -0700943 h->resv_huge_pages -= unused_resv_pages;
Adam Litkeac09b3a2008-03-04 14:29:38 -0800944
Andi Kleenaa888a72008-07-23 21:27:47 -0700945 /* Cannot return gigantic pages currently */
946 if (h->order >= MAX_ORDER)
947 return;
948
Andi Kleena5516432008-07-23 21:27:41 -0700949 nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
Adam Litkee4e574b2007-10-16 01:26:19 -0700950
Lee Schermerhorn685f3452009-09-21 17:01:23 -0700951 /*
952 * We want to release as many surplus pages as possible, spread
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -0800953 * evenly across all nodes with memory. Iterate across these nodes
954 * until we can no longer free unreserved surplus pages. This occurs
955 * when the nodes with surplus pages have no free pages.
956 * free_pool_huge_page() will balance the the freed pages across the
957 * on-line nodes with memory and will handle the hstate accounting.
Lee Schermerhorn685f3452009-09-21 17:01:23 -0700958 */
959 while (nr_pages--) {
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -0800960 if (!free_pool_huge_page(h, &node_states[N_HIGH_MEMORY], 1))
Lee Schermerhorn685f3452009-09-21 17:01:23 -0700961 break;
Adam Litkee4e574b2007-10-16 01:26:19 -0700962 }
963}
964
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -0700965/*
966 * Determine if the huge page at addr within the vma has an associated
967 * reservation. Where it does not we will need to logically increase
968 * reservation and actually increase quota before an allocation can occur.
969 * Where any new reservation would be required the reservation change is
970 * prepared, but not committed. Once the page has been quota'd allocated
971 * an instantiated the change should be committed via vma_commit_reservation.
972 * No action is required on failure.
973 */
Roel Kluine2f17d92009-03-31 15:23:15 -0700974static long vma_needs_reservation(struct hstate *h,
Andi Kleena5516432008-07-23 21:27:41 -0700975 struct vm_area_struct *vma, unsigned long addr)
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -0700976{
977 struct address_space *mapping = vma->vm_file->f_mapping;
978 struct inode *inode = mapping->host;
979
Mel Gormanf83a2752009-05-28 14:34:40 -0700980 if (vma->vm_flags & VM_MAYSHARE) {
Andi Kleena5516432008-07-23 21:27:41 -0700981 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -0700982 return region_chg(&inode->i_mapping->private_list,
983 idx, idx + 1);
984
Andy Whitcroft84afd992008-07-23 21:27:32 -0700985 } else if (!is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
986 return 1;
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -0700987
Andy Whitcroft84afd992008-07-23 21:27:32 -0700988 } else {
Roel Kluine2f17d92009-03-31 15:23:15 -0700989 long err;
Andi Kleena5516432008-07-23 21:27:41 -0700990 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
Andy Whitcroft84afd992008-07-23 21:27:32 -0700991 struct resv_map *reservations = vma_resv_map(vma);
992
993 err = region_chg(&reservations->regions, idx, idx + 1);
994 if (err < 0)
995 return err;
996 return 0;
997 }
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -0700998}
Andi Kleena5516432008-07-23 21:27:41 -0700999static void vma_commit_reservation(struct hstate *h,
1000 struct vm_area_struct *vma, unsigned long addr)
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001001{
1002 struct address_space *mapping = vma->vm_file->f_mapping;
1003 struct inode *inode = mapping->host;
1004
Mel Gormanf83a2752009-05-28 14:34:40 -07001005 if (vma->vm_flags & VM_MAYSHARE) {
Andi Kleena5516432008-07-23 21:27:41 -07001006 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001007 region_add(&inode->i_mapping->private_list, idx, idx + 1);
Andy Whitcroft84afd992008-07-23 21:27:32 -07001008
1009 } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
Andi Kleena5516432008-07-23 21:27:41 -07001010 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
Andy Whitcroft84afd992008-07-23 21:27:32 -07001011 struct resv_map *reservations = vma_resv_map(vma);
1012
1013 /* Mark this page used in the map. */
1014 region_add(&reservations->regions, idx, idx + 1);
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001015 }
1016}
1017
David Gibson27a85ef2006-03-22 00:08:56 -08001018static struct page *alloc_huge_page(struct vm_area_struct *vma,
Mel Gorman04f2cbe2008-07-23 21:27:25 -07001019 unsigned long addr, int avoid_reserve)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001020{
Andi Kleena5516432008-07-23 21:27:41 -07001021 struct hstate *h = hstate_vma(vma);
Adam Litke348ea202007-11-14 16:59:37 -08001022 struct page *page;
Adam Litke2fc39ce2007-11-14 16:59:39 -08001023 struct address_space *mapping = vma->vm_file->f_mapping;
Mel Gormana1e78772008-07-23 21:27:23 -07001024 struct inode *inode = mapping->host;
Roel Kluine2f17d92009-03-31 15:23:15 -07001025 long chg;
Adam Litke2fc39ce2007-11-14 16:59:39 -08001026
Mel Gormana1e78772008-07-23 21:27:23 -07001027 /*
1028 * Processes that did not create the mapping will have no reserves and
1029 * will not have accounted against quota. Check that the quota can be
1030 * made before satisfying the allocation
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001031 * MAP_NORESERVE mappings may also need pages and quota allocated
1032 * if no reserve mapping overlaps.
Mel Gormana1e78772008-07-23 21:27:23 -07001033 */
Andi Kleena5516432008-07-23 21:27:41 -07001034 chg = vma_needs_reservation(h, vma, addr);
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001035 if (chg < 0)
1036 return ERR_PTR(chg);
1037 if (chg)
Mel Gormana1e78772008-07-23 21:27:23 -07001038 if (hugetlb_get_quota(inode->i_mapping, chg))
1039 return ERR_PTR(-ENOSPC);
Mel Gormana1e78772008-07-23 21:27:23 -07001040
1041 spin_lock(&hugetlb_lock);
Andi Kleena5516432008-07-23 21:27:41 -07001042 page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve);
Mel Gormana1e78772008-07-23 21:27:23 -07001043 spin_unlock(&hugetlb_lock);
1044
1045 if (!page) {
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001046 page = alloc_buddy_huge_page(h, NUMA_NO_NODE);
Mel Gormana1e78772008-07-23 21:27:23 -07001047 if (!page) {
1048 hugetlb_put_quota(inode->i_mapping, chg);
Mel Gorman4a6018f2010-05-11 14:06:53 -07001049 return ERR_PTR(-VM_FAULT_SIGBUS);
Mel Gormana1e78772008-07-23 21:27:23 -07001050 }
1051 }
1052
Mel Gormana1e78772008-07-23 21:27:23 -07001053 set_page_private(page, (unsigned long) mapping);
1054
Andi Kleena5516432008-07-23 21:27:41 -07001055 vma_commit_reservation(h, vma, addr);
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001056
Adam Litke90d8b7e2007-11-14 16:59:42 -08001057 return page;
David Gibsonb45b5bd2006-03-22 00:08:55 -08001058}
1059
Cyrill Gorcunov91f47662009-01-06 14:40:33 -08001060int __weak alloc_bootmem_huge_page(struct hstate *h)
Andi Kleenaa888a72008-07-23 21:27:47 -07001061{
1062 struct huge_bootmem_page *m;
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -08001063 int nr_nodes = nodes_weight(node_states[N_HIGH_MEMORY]);
Andi Kleenaa888a72008-07-23 21:27:47 -07001064
1065 while (nr_nodes) {
1066 void *addr;
1067
1068 addr = __alloc_bootmem_node_nopanic(
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001069 NODE_DATA(hstate_next_node_to_alloc(h,
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -08001070 &node_states[N_HIGH_MEMORY])),
Andi Kleenaa888a72008-07-23 21:27:47 -07001071 huge_page_size(h), huge_page_size(h), 0);
1072
1073 if (addr) {
1074 /*
1075 * Use the beginning of the huge page to store the
1076 * huge_bootmem_page struct (until gather_bootmem
1077 * puts them into the mem_map).
1078 */
1079 m = addr;
Cyrill Gorcunov91f47662009-01-06 14:40:33 -08001080 goto found;
Andi Kleenaa888a72008-07-23 21:27:47 -07001081 }
Andi Kleenaa888a72008-07-23 21:27:47 -07001082 nr_nodes--;
1083 }
1084 return 0;
1085
1086found:
1087 BUG_ON((unsigned long)virt_to_phys(m) & (huge_page_size(h) - 1));
1088 /* Put them into a private list first because mem_map is not up yet */
1089 list_add(&m->list, &huge_boot_pages);
1090 m->hstate = h;
1091 return 1;
1092}
1093
Andy Whitcroft18229df2008-11-06 12:53:27 -08001094static void prep_compound_huge_page(struct page *page, int order)
1095{
1096 if (unlikely(order > (MAX_ORDER - 1)))
1097 prep_compound_gigantic_page(page, order);
1098 else
1099 prep_compound_page(page, order);
1100}
1101
Andi Kleenaa888a72008-07-23 21:27:47 -07001102/* Put bootmem huge pages into the standard lists after mem_map is up */
1103static void __init gather_bootmem_prealloc(void)
1104{
1105 struct huge_bootmem_page *m;
1106
1107 list_for_each_entry(m, &huge_boot_pages, list) {
1108 struct page *page = virt_to_page(m);
1109 struct hstate *h = m->hstate;
1110 __ClearPageReserved(page);
1111 WARN_ON(page_count(page) != 1);
Andy Whitcroft18229df2008-11-06 12:53:27 -08001112 prep_compound_huge_page(page, h->order);
Andi Kleenaa888a72008-07-23 21:27:47 -07001113 prep_new_huge_page(h, page, page_to_nid(page));
1114 }
1115}
1116
Andi Kleen8faa8b02008-07-23 21:27:48 -07001117static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001118{
1119 unsigned long i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001120
Andi Kleene5ff2152008-07-23 21:27:42 -07001121 for (i = 0; i < h->max_huge_pages; ++i) {
Andi Kleenaa888a72008-07-23 21:27:47 -07001122 if (h->order >= MAX_ORDER) {
1123 if (!alloc_bootmem_huge_page(h))
1124 break;
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -08001125 } else if (!alloc_fresh_huge_page(h,
1126 &node_states[N_HIGH_MEMORY]))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001127 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001128 }
Andi Kleen8faa8b02008-07-23 21:27:48 -07001129 h->max_huge_pages = i;
Andi Kleene5ff2152008-07-23 21:27:42 -07001130}
1131
1132static void __init hugetlb_init_hstates(void)
1133{
1134 struct hstate *h;
1135
1136 for_each_hstate(h) {
Andi Kleen8faa8b02008-07-23 21:27:48 -07001137 /* oversize hugepages were init'ed in early boot */
1138 if (h->order < MAX_ORDER)
1139 hugetlb_hstate_alloc_pages(h);
Andi Kleene5ff2152008-07-23 21:27:42 -07001140 }
1141}
1142
Andi Kleen4abd32d2008-07-23 21:27:49 -07001143static char * __init memfmt(char *buf, unsigned long n)
1144{
1145 if (n >= (1UL << 30))
1146 sprintf(buf, "%lu GB", n >> 30);
1147 else if (n >= (1UL << 20))
1148 sprintf(buf, "%lu MB", n >> 20);
1149 else
1150 sprintf(buf, "%lu KB", n >> 10);
1151 return buf;
1152}
1153
Andi Kleene5ff2152008-07-23 21:27:42 -07001154static void __init report_hugepages(void)
1155{
1156 struct hstate *h;
1157
1158 for_each_hstate(h) {
Andi Kleen4abd32d2008-07-23 21:27:49 -07001159 char buf[32];
1160 printk(KERN_INFO "HugeTLB registered %s page size, "
1161 "pre-allocated %ld pages\n",
1162 memfmt(buf, huge_page_size(h)),
1163 h->free_huge_pages);
Andi Kleene5ff2152008-07-23 21:27:42 -07001164 }
1165}
1166
Linus Torvalds1da177e2005-04-16 15:20:36 -07001167#ifdef CONFIG_HIGHMEM
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001168static void try_to_free_low(struct hstate *h, unsigned long count,
1169 nodemask_t *nodes_allowed)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001170{
Christoph Lameter4415cc82006-09-25 23:31:55 -07001171 int i;
1172
Andi Kleenaa888a72008-07-23 21:27:47 -07001173 if (h->order >= MAX_ORDER)
1174 return;
1175
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001176 for_each_node_mask(i, *nodes_allowed) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001177 struct page *page, *next;
Andi Kleena5516432008-07-23 21:27:41 -07001178 struct list_head *freel = &h->hugepage_freelists[i];
1179 list_for_each_entry_safe(page, next, freel, lru) {
1180 if (count >= h->nr_huge_pages)
Adam Litke6b0c8802007-10-16 01:26:23 -07001181 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001182 if (PageHighMem(page))
1183 continue;
1184 list_del(&page->lru);
Andi Kleene5ff2152008-07-23 21:27:42 -07001185 update_and_free_page(h, page);
Andi Kleena5516432008-07-23 21:27:41 -07001186 h->free_huge_pages--;
1187 h->free_huge_pages_node[page_to_nid(page)]--;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001188 }
1189 }
1190}
1191#else
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001192static inline void try_to_free_low(struct hstate *h, unsigned long count,
1193 nodemask_t *nodes_allowed)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001194{
1195}
1196#endif
1197
Wu Fengguang20a03072009-06-16 15:32:22 -07001198/*
1199 * Increment or decrement surplus_huge_pages. Keep node-specific counters
1200 * balanced by operating on them in a round-robin fashion.
1201 * Returns 1 if an adjustment was made.
1202 */
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001203static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
1204 int delta)
Wu Fengguang20a03072009-06-16 15:32:22 -07001205{
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001206 int start_nid, next_nid;
Wu Fengguang20a03072009-06-16 15:32:22 -07001207 int ret = 0;
1208
1209 VM_BUG_ON(delta != -1 && delta != 1);
Wu Fengguang20a03072009-06-16 15:32:22 -07001210
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001211 if (delta < 0)
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001212 start_nid = hstate_next_node_to_alloc(h, nodes_allowed);
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001213 else
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001214 start_nid = hstate_next_node_to_free(h, nodes_allowed);
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001215 next_nid = start_nid;
1216
1217 do {
1218 int nid = next_nid;
1219 if (delta < 0) {
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001220 /*
1221 * To shrink on this node, there must be a surplus page
1222 */
Lee Schermerhorn9a76db02009-12-14 17:58:15 -08001223 if (!h->surplus_huge_pages_node[nid]) {
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001224 next_nid = hstate_next_node_to_alloc(h,
1225 nodes_allowed);
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001226 continue;
Lee Schermerhorn9a76db02009-12-14 17:58:15 -08001227 }
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001228 }
1229 if (delta > 0) {
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001230 /*
1231 * Surplus cannot exceed the total number of pages
1232 */
1233 if (h->surplus_huge_pages_node[nid] >=
Lee Schermerhorn9a76db02009-12-14 17:58:15 -08001234 h->nr_huge_pages_node[nid]) {
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001235 next_nid = hstate_next_node_to_free(h,
1236 nodes_allowed);
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001237 continue;
Lee Schermerhorn9a76db02009-12-14 17:58:15 -08001238 }
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001239 }
Wu Fengguang20a03072009-06-16 15:32:22 -07001240
1241 h->surplus_huge_pages += delta;
1242 h->surplus_huge_pages_node[nid] += delta;
1243 ret = 1;
1244 break;
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001245 } while (next_nid != start_nid);
Wu Fengguang20a03072009-06-16 15:32:22 -07001246
Wu Fengguang20a03072009-06-16 15:32:22 -07001247 return ret;
1248}
1249
Andi Kleena5516432008-07-23 21:27:41 -07001250#define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001251static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count,
1252 nodemask_t *nodes_allowed)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001253{
Adam Litke7893d1d2007-10-16 01:26:18 -07001254 unsigned long min_count, ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001255
Andi Kleenaa888a72008-07-23 21:27:47 -07001256 if (h->order >= MAX_ORDER)
1257 return h->max_huge_pages;
1258
Adam Litke7893d1d2007-10-16 01:26:18 -07001259 /*
1260 * Increase the pool size
1261 * First take pages out of surplus state. Then make up the
1262 * remaining difference by allocating fresh huge pages.
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001263 *
1264 * We might race with alloc_buddy_huge_page() here and be unable
1265 * to convert a surplus huge page to a normal huge page. That is
1266 * not critical, though, it just means the overall size of the
1267 * pool might be one hugepage larger than it needs to be, but
1268 * within all the constraints specified by the sysctls.
Adam Litke7893d1d2007-10-16 01:26:18 -07001269 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001270 spin_lock(&hugetlb_lock);
Andi Kleena5516432008-07-23 21:27:41 -07001271 while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001272 if (!adjust_pool_surplus(h, nodes_allowed, -1))
Adam Litke7893d1d2007-10-16 01:26:18 -07001273 break;
1274 }
1275
Andi Kleena5516432008-07-23 21:27:41 -07001276 while (count > persistent_huge_pages(h)) {
Adam Litke7893d1d2007-10-16 01:26:18 -07001277 /*
1278 * If this allocation races such that we no longer need the
1279 * page, free_huge_page will handle it by freeing the page
1280 * and reducing the surplus.
1281 */
1282 spin_unlock(&hugetlb_lock);
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001283 ret = alloc_fresh_huge_page(h, nodes_allowed);
Adam Litke7893d1d2007-10-16 01:26:18 -07001284 spin_lock(&hugetlb_lock);
1285 if (!ret)
1286 goto out;
1287
Mel Gorman536240f22009-12-14 17:59:56 -08001288 /* Bail for signals. Probably ctrl-c from user */
1289 if (signal_pending(current))
1290 goto out;
Adam Litke7893d1d2007-10-16 01:26:18 -07001291 }
Adam Litke7893d1d2007-10-16 01:26:18 -07001292
1293 /*
1294 * Decrease the pool size
1295 * First return free pages to the buddy allocator (being careful
1296 * to keep enough around to satisfy reservations). Then place
1297 * pages into surplus state as needed so the pool will shrink
1298 * to the desired size as pages become free.
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001299 *
1300 * By placing pages into the surplus state independent of the
1301 * overcommit value, we are allowing the surplus pool size to
1302 * exceed overcommit. There are few sane options here. Since
1303 * alloc_buddy_huge_page() is checking the global counter,
1304 * though, we'll note that we're not allowed to exceed surplus
1305 * and won't grow the pool anywhere else. Not until one of the
1306 * sysctls are changed, or the surplus pages go out of use.
Adam Litke7893d1d2007-10-16 01:26:18 -07001307 */
Andi Kleena5516432008-07-23 21:27:41 -07001308 min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
Adam Litke6b0c8802007-10-16 01:26:23 -07001309 min_count = max(count, min_count);
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001310 try_to_free_low(h, min_count, nodes_allowed);
Andi Kleena5516432008-07-23 21:27:41 -07001311 while (min_count < persistent_huge_pages(h)) {
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001312 if (!free_pool_huge_page(h, nodes_allowed, 0))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001313 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001314 }
Andi Kleena5516432008-07-23 21:27:41 -07001315 while (count < persistent_huge_pages(h)) {
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001316 if (!adjust_pool_surplus(h, nodes_allowed, 1))
Adam Litke7893d1d2007-10-16 01:26:18 -07001317 break;
1318 }
1319out:
Andi Kleena5516432008-07-23 21:27:41 -07001320 ret = persistent_huge_pages(h);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001321 spin_unlock(&hugetlb_lock);
Adam Litke7893d1d2007-10-16 01:26:18 -07001322 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001323}
1324
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001325#define HSTATE_ATTR_RO(_name) \
1326 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
1327
1328#define HSTATE_ATTR(_name) \
1329 static struct kobj_attribute _name##_attr = \
1330 __ATTR(_name, 0644, _name##_show, _name##_store)
1331
1332static struct kobject *hugepages_kobj;
1333static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
1334
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001335static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
1336
1337static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001338{
1339 int i;
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001340
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001341 for (i = 0; i < HUGE_MAX_HSTATE; i++)
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001342 if (hstate_kobjs[i] == kobj) {
1343 if (nidp)
1344 *nidp = NUMA_NO_NODE;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001345 return &hstates[i];
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001346 }
1347
1348 return kobj_to_node_hstate(kobj, nidp);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001349}
1350
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001351static ssize_t nr_hugepages_show_common(struct kobject *kobj,
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001352 struct kobj_attribute *attr, char *buf)
1353{
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001354 struct hstate *h;
1355 unsigned long nr_huge_pages;
1356 int nid;
1357
1358 h = kobj_to_hstate(kobj, &nid);
1359 if (nid == NUMA_NO_NODE)
1360 nr_huge_pages = h->nr_huge_pages;
1361 else
1362 nr_huge_pages = h->nr_huge_pages_node[nid];
1363
1364 return sprintf(buf, "%lu\n", nr_huge_pages);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001365}
Eric B Munsonadbe8722011-01-13 15:47:27 -08001366
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001367static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
1368 struct kobject *kobj, struct kobj_attribute *attr,
1369 const char *buf, size_t len)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001370{
1371 int err;
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001372 int nid;
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001373 unsigned long count;
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001374 struct hstate *h;
David Rientjesbad44b52009-12-14 17:58:38 -08001375 NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001376
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001377 err = strict_strtoul(buf, 10, &count);
Eric B Munson73ae31e2011-01-13 15:47:28 -08001378 if (err)
Eric B Munsonadbe8722011-01-13 15:47:27 -08001379 goto out;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001380
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001381 h = kobj_to_hstate(kobj, &nid);
Eric B Munsonadbe8722011-01-13 15:47:27 -08001382 if (h->order >= MAX_ORDER) {
1383 err = -EINVAL;
1384 goto out;
1385 }
1386
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001387 if (nid == NUMA_NO_NODE) {
1388 /*
1389 * global hstate attribute
1390 */
1391 if (!(obey_mempolicy &&
1392 init_nodemask_of_mempolicy(nodes_allowed))) {
1393 NODEMASK_FREE(nodes_allowed);
1394 nodes_allowed = &node_states[N_HIGH_MEMORY];
1395 }
1396 } else if (nodes_allowed) {
1397 /*
1398 * per node hstate attribute: adjust count to global,
1399 * but restrict alloc/free to the specified node.
1400 */
1401 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
1402 init_nodemask_of_node(nodes_allowed, nid);
1403 } else
1404 nodes_allowed = &node_states[N_HIGH_MEMORY];
1405
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001406 h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001407
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -08001408 if (nodes_allowed != &node_states[N_HIGH_MEMORY])
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001409 NODEMASK_FREE(nodes_allowed);
1410
1411 return len;
Eric B Munsonadbe8722011-01-13 15:47:27 -08001412out:
1413 NODEMASK_FREE(nodes_allowed);
1414 return err;
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001415}
1416
1417static ssize_t nr_hugepages_show(struct kobject *kobj,
1418 struct kobj_attribute *attr, char *buf)
1419{
1420 return nr_hugepages_show_common(kobj, attr, buf);
1421}
1422
1423static ssize_t nr_hugepages_store(struct kobject *kobj,
1424 struct kobj_attribute *attr, const char *buf, size_t len)
1425{
1426 return nr_hugepages_store_common(false, kobj, attr, buf, len);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001427}
1428HSTATE_ATTR(nr_hugepages);
1429
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001430#ifdef CONFIG_NUMA
1431
1432/*
1433 * hstate attribute for optionally mempolicy-based constraint on persistent
1434 * huge page alloc/free.
1435 */
1436static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
1437 struct kobj_attribute *attr, char *buf)
1438{
1439 return nr_hugepages_show_common(kobj, attr, buf);
1440}
1441
1442static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
1443 struct kobj_attribute *attr, const char *buf, size_t len)
1444{
1445 return nr_hugepages_store_common(true, kobj, attr, buf, len);
1446}
1447HSTATE_ATTR(nr_hugepages_mempolicy);
1448#endif
1449
1450
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001451static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
1452 struct kobj_attribute *attr, char *buf)
1453{
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001454 struct hstate *h = kobj_to_hstate(kobj, NULL);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001455 return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
1456}
Eric B Munsonadbe8722011-01-13 15:47:27 -08001457
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001458static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
1459 struct kobj_attribute *attr, const char *buf, size_t count)
1460{
1461 int err;
1462 unsigned long input;
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001463 struct hstate *h = kobj_to_hstate(kobj, NULL);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001464
Eric B Munsonadbe8722011-01-13 15:47:27 -08001465 if (h->order >= MAX_ORDER)
1466 return -EINVAL;
1467
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001468 err = strict_strtoul(buf, 10, &input);
1469 if (err)
Eric B Munson73ae31e2011-01-13 15:47:28 -08001470 return err;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001471
1472 spin_lock(&hugetlb_lock);
1473 h->nr_overcommit_huge_pages = input;
1474 spin_unlock(&hugetlb_lock);
1475
1476 return count;
1477}
1478HSTATE_ATTR(nr_overcommit_hugepages);
1479
1480static ssize_t free_hugepages_show(struct kobject *kobj,
1481 struct kobj_attribute *attr, char *buf)
1482{
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001483 struct hstate *h;
1484 unsigned long free_huge_pages;
1485 int nid;
1486
1487 h = kobj_to_hstate(kobj, &nid);
1488 if (nid == NUMA_NO_NODE)
1489 free_huge_pages = h->free_huge_pages;
1490 else
1491 free_huge_pages = h->free_huge_pages_node[nid];
1492
1493 return sprintf(buf, "%lu\n", free_huge_pages);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001494}
1495HSTATE_ATTR_RO(free_hugepages);
1496
1497static ssize_t resv_hugepages_show(struct kobject *kobj,
1498 struct kobj_attribute *attr, char *buf)
1499{
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001500 struct hstate *h = kobj_to_hstate(kobj, NULL);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001501 return sprintf(buf, "%lu\n", h->resv_huge_pages);
1502}
1503HSTATE_ATTR_RO(resv_hugepages);
1504
1505static ssize_t surplus_hugepages_show(struct kobject *kobj,
1506 struct kobj_attribute *attr, char *buf)
1507{
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001508 struct hstate *h;
1509 unsigned long surplus_huge_pages;
1510 int nid;
1511
1512 h = kobj_to_hstate(kobj, &nid);
1513 if (nid == NUMA_NO_NODE)
1514 surplus_huge_pages = h->surplus_huge_pages;
1515 else
1516 surplus_huge_pages = h->surplus_huge_pages_node[nid];
1517
1518 return sprintf(buf, "%lu\n", surplus_huge_pages);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001519}
1520HSTATE_ATTR_RO(surplus_hugepages);
1521
1522static struct attribute *hstate_attrs[] = {
1523 &nr_hugepages_attr.attr,
1524 &nr_overcommit_hugepages_attr.attr,
1525 &free_hugepages_attr.attr,
1526 &resv_hugepages_attr.attr,
1527 &surplus_hugepages_attr.attr,
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001528#ifdef CONFIG_NUMA
1529 &nr_hugepages_mempolicy_attr.attr,
1530#endif
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001531 NULL,
1532};
1533
1534static struct attribute_group hstate_attr_group = {
1535 .attrs = hstate_attrs,
1536};
1537
Jeff Mahoney094e9532010-02-02 13:44:14 -08001538static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
1539 struct kobject **hstate_kobjs,
1540 struct attribute_group *hstate_attr_group)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001541{
1542 int retval;
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001543 int hi = h - hstates;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001544
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001545 hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
1546 if (!hstate_kobjs[hi])
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001547 return -ENOMEM;
1548
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001549 retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001550 if (retval)
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001551 kobject_put(hstate_kobjs[hi]);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001552
1553 return retval;
1554}
1555
1556static void __init hugetlb_sysfs_init(void)
1557{
1558 struct hstate *h;
1559 int err;
1560
1561 hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
1562 if (!hugepages_kobj)
1563 return;
1564
1565 for_each_hstate(h) {
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001566 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
1567 hstate_kobjs, &hstate_attr_group);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001568 if (err)
1569 printk(KERN_ERR "Hugetlb: Unable to add hstate %s",
1570 h->name);
1571 }
1572}
1573
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001574#ifdef CONFIG_NUMA
1575
1576/*
1577 * node_hstate/s - associate per node hstate attributes, via their kobjects,
1578 * with node sysdevs in node_devices[] using a parallel array. The array
1579 * index of a node sysdev or _hstate == node id.
1580 * This is here to avoid any static dependency of the node sysdev driver, in
1581 * the base kernel, on the hugetlb module.
1582 */
1583struct node_hstate {
1584 struct kobject *hugepages_kobj;
1585 struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
1586};
1587struct node_hstate node_hstates[MAX_NUMNODES];
1588
1589/*
1590 * A subset of global hstate attributes for node sysdevs
1591 */
1592static struct attribute *per_node_hstate_attrs[] = {
1593 &nr_hugepages_attr.attr,
1594 &free_hugepages_attr.attr,
1595 &surplus_hugepages_attr.attr,
1596 NULL,
1597};
1598
1599static struct attribute_group per_node_hstate_attr_group = {
1600 .attrs = per_node_hstate_attrs,
1601};
1602
1603/*
1604 * kobj_to_node_hstate - lookup global hstate for node sysdev hstate attr kobj.
1605 * Returns node id via non-NULL nidp.
1606 */
1607static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
1608{
1609 int nid;
1610
1611 for (nid = 0; nid < nr_node_ids; nid++) {
1612 struct node_hstate *nhs = &node_hstates[nid];
1613 int i;
1614 for (i = 0; i < HUGE_MAX_HSTATE; i++)
1615 if (nhs->hstate_kobjs[i] == kobj) {
1616 if (nidp)
1617 *nidp = nid;
1618 return &hstates[i];
1619 }
1620 }
1621
1622 BUG();
1623 return NULL;
1624}
1625
1626/*
1627 * Unregister hstate attributes from a single node sysdev.
1628 * No-op if no hstate attributes attached.
1629 */
1630void hugetlb_unregister_node(struct node *node)
1631{
1632 struct hstate *h;
1633 struct node_hstate *nhs = &node_hstates[node->sysdev.id];
1634
1635 if (!nhs->hugepages_kobj)
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -08001636 return; /* no hstate attributes */
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001637
1638 for_each_hstate(h)
1639 if (nhs->hstate_kobjs[h - hstates]) {
1640 kobject_put(nhs->hstate_kobjs[h - hstates]);
1641 nhs->hstate_kobjs[h - hstates] = NULL;
1642 }
1643
1644 kobject_put(nhs->hugepages_kobj);
1645 nhs->hugepages_kobj = NULL;
1646}
1647
1648/*
1649 * hugetlb module exit: unregister hstate attributes from node sysdevs
1650 * that have them.
1651 */
1652static void hugetlb_unregister_all_nodes(void)
1653{
1654 int nid;
1655
1656 /*
1657 * disable node sysdev registrations.
1658 */
1659 register_hugetlbfs_with_node(NULL, NULL);
1660
1661 /*
1662 * remove hstate attributes from any nodes that have them.
1663 */
1664 for (nid = 0; nid < nr_node_ids; nid++)
1665 hugetlb_unregister_node(&node_devices[nid]);
1666}
1667
1668/*
1669 * Register hstate attributes for a single node sysdev.
1670 * No-op if attributes already registered.
1671 */
1672void hugetlb_register_node(struct node *node)
1673{
1674 struct hstate *h;
1675 struct node_hstate *nhs = &node_hstates[node->sysdev.id];
1676 int err;
1677
1678 if (nhs->hugepages_kobj)
1679 return; /* already allocated */
1680
1681 nhs->hugepages_kobj = kobject_create_and_add("hugepages",
1682 &node->sysdev.kobj);
1683 if (!nhs->hugepages_kobj)
1684 return;
1685
1686 for_each_hstate(h) {
1687 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
1688 nhs->hstate_kobjs,
1689 &per_node_hstate_attr_group);
1690 if (err) {
1691 printk(KERN_ERR "Hugetlb: Unable to add hstate %s"
1692 " for node %d\n",
1693 h->name, node->sysdev.id);
1694 hugetlb_unregister_node(node);
1695 break;
1696 }
1697 }
1698}
1699
1700/*
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -08001701 * hugetlb init time: register hstate attributes for all registered node
1702 * sysdevs of nodes that have memory. All on-line nodes should have
1703 * registered their associated sysdev by this time.
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001704 */
1705static void hugetlb_register_all_nodes(void)
1706{
1707 int nid;
1708
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -08001709 for_each_node_state(nid, N_HIGH_MEMORY) {
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001710 struct node *node = &node_devices[nid];
1711 if (node->sysdev.id == nid)
1712 hugetlb_register_node(node);
1713 }
1714
1715 /*
1716 * Let the node sysdev driver know we're here so it can
1717 * [un]register hstate attributes on node hotplug.
1718 */
1719 register_hugetlbfs_with_node(hugetlb_register_node,
1720 hugetlb_unregister_node);
1721}
1722#else /* !CONFIG_NUMA */
1723
1724static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
1725{
1726 BUG();
1727 if (nidp)
1728 *nidp = -1;
1729 return NULL;
1730}
1731
1732static void hugetlb_unregister_all_nodes(void) { }
1733
1734static void hugetlb_register_all_nodes(void) { }
1735
1736#endif
1737
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001738static void __exit hugetlb_exit(void)
1739{
1740 struct hstate *h;
1741
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001742 hugetlb_unregister_all_nodes();
1743
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001744 for_each_hstate(h) {
1745 kobject_put(hstate_kobjs[h - hstates]);
1746 }
1747
1748 kobject_put(hugepages_kobj);
1749}
1750module_exit(hugetlb_exit);
1751
1752static int __init hugetlb_init(void)
1753{
Benjamin Herrenschmidt0ef89d22008-07-31 00:07:30 -07001754 /* Some platform decide whether they support huge pages at boot
1755 * time. On these, such as powerpc, HPAGE_SHIFT is set to 0 when
1756 * there is no such support
1757 */
1758 if (HPAGE_SHIFT == 0)
1759 return 0;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001760
Nick Piggine11bfbf2008-07-23 21:27:52 -07001761 if (!size_to_hstate(default_hstate_size)) {
1762 default_hstate_size = HPAGE_SIZE;
1763 if (!size_to_hstate(default_hstate_size))
1764 hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001765 }
Nick Piggine11bfbf2008-07-23 21:27:52 -07001766 default_hstate_idx = size_to_hstate(default_hstate_size) - hstates;
1767 if (default_hstate_max_huge_pages)
1768 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001769
1770 hugetlb_init_hstates();
1771
Andi Kleenaa888a72008-07-23 21:27:47 -07001772 gather_bootmem_prealloc();
1773
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001774 report_hugepages();
1775
1776 hugetlb_sysfs_init();
1777
Lee Schermerhorn9a305232009-12-14 17:58:25 -08001778 hugetlb_register_all_nodes();
1779
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001780 return 0;
1781}
1782module_init(hugetlb_init);
1783
1784/* Should be called on processing a hugepagesz=... option */
1785void __init hugetlb_add_hstate(unsigned order)
1786{
1787 struct hstate *h;
Andi Kleen8faa8b02008-07-23 21:27:48 -07001788 unsigned long i;
1789
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001790 if (size_to_hstate(PAGE_SIZE << order)) {
1791 printk(KERN_WARNING "hugepagesz= specified twice, ignoring\n");
1792 return;
1793 }
1794 BUG_ON(max_hstate >= HUGE_MAX_HSTATE);
1795 BUG_ON(order == 0);
1796 h = &hstates[max_hstate++];
1797 h->order = order;
1798 h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
Andi Kleen8faa8b02008-07-23 21:27:48 -07001799 h->nr_huge_pages = 0;
1800 h->free_huge_pages = 0;
1801 for (i = 0; i < MAX_NUMNODES; ++i)
1802 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -08001803 h->next_nid_to_alloc = first_node(node_states[N_HIGH_MEMORY]);
1804 h->next_nid_to_free = first_node(node_states[N_HIGH_MEMORY]);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001805 snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
1806 huge_page_size(h)/1024);
Andi Kleen8faa8b02008-07-23 21:27:48 -07001807
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001808 parsed_hstate = h;
1809}
1810
Nick Piggine11bfbf2008-07-23 21:27:52 -07001811static int __init hugetlb_nrpages_setup(char *s)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001812{
1813 unsigned long *mhp;
Andi Kleen8faa8b02008-07-23 21:27:48 -07001814 static unsigned long *last_mhp;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001815
1816 /*
1817 * !max_hstate means we haven't parsed a hugepagesz= parameter yet,
1818 * so this hugepages= parameter goes to the "default hstate".
1819 */
1820 if (!max_hstate)
1821 mhp = &default_hstate_max_huge_pages;
1822 else
1823 mhp = &parsed_hstate->max_huge_pages;
1824
Andi Kleen8faa8b02008-07-23 21:27:48 -07001825 if (mhp == last_mhp) {
1826 printk(KERN_WARNING "hugepages= specified twice without "
1827 "interleaving hugepagesz=, ignoring\n");
1828 return 1;
1829 }
1830
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001831 if (sscanf(s, "%lu", mhp) <= 0)
1832 *mhp = 0;
1833
Andi Kleen8faa8b02008-07-23 21:27:48 -07001834 /*
1835 * Global state is always initialized later in hugetlb_init.
1836 * But we need to allocate >= MAX_ORDER hstates here early to still
1837 * use the bootmem allocator.
1838 */
1839 if (max_hstate && parsed_hstate->order >= MAX_ORDER)
1840 hugetlb_hstate_alloc_pages(parsed_hstate);
1841
1842 last_mhp = mhp;
1843
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001844 return 1;
1845}
Nick Piggine11bfbf2008-07-23 21:27:52 -07001846__setup("hugepages=", hugetlb_nrpages_setup);
1847
1848static int __init hugetlb_default_setup(char *s)
1849{
1850 default_hstate_size = memparse(s, &s);
1851 return 1;
1852}
1853__setup("default_hugepagesz=", hugetlb_default_setup);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001854
Nishanth Aravamudan8a213462008-07-25 19:44:37 -07001855static unsigned int cpuset_mems_nr(unsigned int *array)
1856{
1857 int node;
1858 unsigned int nr = 0;
1859
1860 for_each_node_mask(node, cpuset_current_mems_allowed)
1861 nr += array[node];
1862
1863 return nr;
1864}
1865
1866#ifdef CONFIG_SYSCTL
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001867static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
1868 struct ctl_table *table, int write,
1869 void __user *buffer, size_t *length, loff_t *ppos)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001870{
Andi Kleene5ff2152008-07-23 21:27:42 -07001871 struct hstate *h = &default_hstate;
1872 unsigned long tmp;
Michal Hocko08d4a242011-01-13 15:47:26 -08001873 int ret;
Andi Kleene5ff2152008-07-23 21:27:42 -07001874
Petr Holasekc033a932011-03-22 16:33:05 -07001875 tmp = h->max_huge_pages;
Andi Kleene5ff2152008-07-23 21:27:42 -07001876
Eric B Munsonadbe8722011-01-13 15:47:27 -08001877 if (write && h->order >= MAX_ORDER)
1878 return -EINVAL;
1879
Andi Kleene5ff2152008-07-23 21:27:42 -07001880 table->data = &tmp;
1881 table->maxlen = sizeof(unsigned long);
Michal Hocko08d4a242011-01-13 15:47:26 -08001882 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
1883 if (ret)
1884 goto out;
Andi Kleene5ff2152008-07-23 21:27:42 -07001885
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001886 if (write) {
David Rientjesbad44b52009-12-14 17:58:38 -08001887 NODEMASK_ALLOC(nodemask_t, nodes_allowed,
1888 GFP_KERNEL | __GFP_NORETRY);
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001889 if (!(obey_mempolicy &&
1890 init_nodemask_of_mempolicy(nodes_allowed))) {
1891 NODEMASK_FREE(nodes_allowed);
1892 nodes_allowed = &node_states[N_HIGH_MEMORY];
1893 }
1894 h->max_huge_pages = set_max_huge_pages(h, tmp, nodes_allowed);
1895
1896 if (nodes_allowed != &node_states[N_HIGH_MEMORY])
1897 NODEMASK_FREE(nodes_allowed);
1898 }
Michal Hocko08d4a242011-01-13 15:47:26 -08001899out:
1900 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001901}
Mel Gorman396faf02007-07-17 04:03:13 -07001902
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001903int hugetlb_sysctl_handler(struct ctl_table *table, int write,
1904 void __user *buffer, size_t *length, loff_t *ppos)
1905{
1906
1907 return hugetlb_sysctl_handler_common(false, table, write,
1908 buffer, length, ppos);
1909}
1910
1911#ifdef CONFIG_NUMA
1912int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
1913 void __user *buffer, size_t *length, loff_t *ppos)
1914{
1915 return hugetlb_sysctl_handler_common(true, table, write,
1916 buffer, length, ppos);
1917}
1918#endif /* CONFIG_NUMA */
1919
Mel Gorman396faf02007-07-17 04:03:13 -07001920int hugetlb_treat_movable_handler(struct ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07001921 void __user *buffer,
Mel Gorman396faf02007-07-17 04:03:13 -07001922 size_t *length, loff_t *ppos)
1923{
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07001924 proc_dointvec(table, write, buffer, length, ppos);
Mel Gorman396faf02007-07-17 04:03:13 -07001925 if (hugepages_treat_as_movable)
1926 htlb_alloc_mask = GFP_HIGHUSER_MOVABLE;
1927 else
1928 htlb_alloc_mask = GFP_HIGHUSER;
1929 return 0;
1930}
1931
Nishanth Aravamudana3d0c6a2008-02-08 04:18:18 -08001932int hugetlb_overcommit_handler(struct ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07001933 void __user *buffer,
Nishanth Aravamudana3d0c6a2008-02-08 04:18:18 -08001934 size_t *length, loff_t *ppos)
1935{
Andi Kleena5516432008-07-23 21:27:41 -07001936 struct hstate *h = &default_hstate;
Andi Kleene5ff2152008-07-23 21:27:42 -07001937 unsigned long tmp;
Michal Hocko08d4a242011-01-13 15:47:26 -08001938 int ret;
Andi Kleene5ff2152008-07-23 21:27:42 -07001939
Petr Holasekc033a932011-03-22 16:33:05 -07001940 tmp = h->nr_overcommit_huge_pages;
Andi Kleene5ff2152008-07-23 21:27:42 -07001941
Eric B Munsonadbe8722011-01-13 15:47:27 -08001942 if (write && h->order >= MAX_ORDER)
1943 return -EINVAL;
1944
Andi Kleene5ff2152008-07-23 21:27:42 -07001945 table->data = &tmp;
1946 table->maxlen = sizeof(unsigned long);
Michal Hocko08d4a242011-01-13 15:47:26 -08001947 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
1948 if (ret)
1949 goto out;
Andi Kleene5ff2152008-07-23 21:27:42 -07001950
1951 if (write) {
1952 spin_lock(&hugetlb_lock);
1953 h->nr_overcommit_huge_pages = tmp;
1954 spin_unlock(&hugetlb_lock);
1955 }
Michal Hocko08d4a242011-01-13 15:47:26 -08001956out:
1957 return ret;
Nishanth Aravamudana3d0c6a2008-02-08 04:18:18 -08001958}
1959
Linus Torvalds1da177e2005-04-16 15:20:36 -07001960#endif /* CONFIG_SYSCTL */
1961
Alexey Dobriyane1759c22008-10-15 23:50:22 +04001962void hugetlb_report_meminfo(struct seq_file *m)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001963{
Andi Kleena5516432008-07-23 21:27:41 -07001964 struct hstate *h = &default_hstate;
Alexey Dobriyane1759c22008-10-15 23:50:22 +04001965 seq_printf(m,
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001966 "HugePages_Total: %5lu\n"
1967 "HugePages_Free: %5lu\n"
1968 "HugePages_Rsvd: %5lu\n"
1969 "HugePages_Surp: %5lu\n"
1970 "Hugepagesize: %8lu kB\n",
Andi Kleena5516432008-07-23 21:27:41 -07001971 h->nr_huge_pages,
1972 h->free_huge_pages,
1973 h->resv_huge_pages,
1974 h->surplus_huge_pages,
1975 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001976}
1977
1978int hugetlb_report_node_meminfo(int nid, char *buf)
1979{
Andi Kleena5516432008-07-23 21:27:41 -07001980 struct hstate *h = &default_hstate;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001981 return sprintf(buf,
1982 "Node %d HugePages_Total: %5u\n"
Nishanth Aravamudana1de0912008-03-26 14:37:53 -07001983 "Node %d HugePages_Free: %5u\n"
1984 "Node %d HugePages_Surp: %5u\n",
Andi Kleena5516432008-07-23 21:27:41 -07001985 nid, h->nr_huge_pages_node[nid],
1986 nid, h->free_huge_pages_node[nid],
1987 nid, h->surplus_huge_pages_node[nid]);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001988}
1989
Linus Torvalds1da177e2005-04-16 15:20:36 -07001990/* Return the number pages of memory we physically have, in PAGE_SIZE units. */
1991unsigned long hugetlb_total_pages(void)
1992{
Andi Kleena5516432008-07-23 21:27:41 -07001993 struct hstate *h = &default_hstate;
1994 return h->nr_huge_pages * pages_per_huge_page(h);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001995}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001996
Andi Kleena5516432008-07-23 21:27:41 -07001997static int hugetlb_acct_memory(struct hstate *h, long delta)
Mel Gormanfc1b8a72008-07-23 21:27:22 -07001998{
1999 int ret = -ENOMEM;
2000
2001 spin_lock(&hugetlb_lock);
2002 /*
2003 * When cpuset is configured, it breaks the strict hugetlb page
2004 * reservation as the accounting is done on a global variable. Such
2005 * reservation is completely rubbish in the presence of cpuset because
2006 * the reservation is not checked against page availability for the
2007 * current cpuset. Application can still potentially OOM'ed by kernel
2008 * with lack of free htlb page in cpuset that the task is in.
2009 * Attempt to enforce strict accounting with cpuset is almost
2010 * impossible (or too ugly) because cpuset is too fluid that
2011 * task or memory node can be dynamically moved between cpusets.
2012 *
2013 * The change of semantics for shared hugetlb mapping with cpuset is
2014 * undesirable. However, in order to preserve some of the semantics,
2015 * we fall back to check against current free page availability as
2016 * a best attempt and hopefully to minimize the impact of changing
2017 * semantics that cpuset has.
2018 */
2019 if (delta > 0) {
Andi Kleena5516432008-07-23 21:27:41 -07002020 if (gather_surplus_pages(h, delta) < 0)
Mel Gormanfc1b8a72008-07-23 21:27:22 -07002021 goto out;
2022
Andi Kleena5516432008-07-23 21:27:41 -07002023 if (delta > cpuset_mems_nr(h->free_huge_pages_node)) {
2024 return_unused_surplus_pages(h, delta);
Mel Gormanfc1b8a72008-07-23 21:27:22 -07002025 goto out;
2026 }
2027 }
2028
2029 ret = 0;
2030 if (delta < 0)
Andi Kleena5516432008-07-23 21:27:41 -07002031 return_unused_surplus_pages(h, (unsigned long) -delta);
Mel Gormanfc1b8a72008-07-23 21:27:22 -07002032
2033out:
2034 spin_unlock(&hugetlb_lock);
2035 return ret;
2036}
2037
Andy Whitcroft84afd992008-07-23 21:27:32 -07002038static void hugetlb_vm_op_open(struct vm_area_struct *vma)
2039{
2040 struct resv_map *reservations = vma_resv_map(vma);
2041
2042 /*
2043 * This new VMA should share its siblings reservation map if present.
2044 * The VMA will only ever have a valid reservation map pointer where
2045 * it is being copied for another still existing VMA. As that VMA
Lucas De Marchi25985ed2011-03-30 22:57:33 -03002046 * has a reference to the reservation map it cannot disappear until
Andy Whitcroft84afd992008-07-23 21:27:32 -07002047 * after this open call completes. It is therefore safe to take a
2048 * new reference here without additional locking.
2049 */
2050 if (reservations)
2051 kref_get(&reservations->refs);
2052}
2053
Mel Gormana1e78772008-07-23 21:27:23 -07002054static void hugetlb_vm_op_close(struct vm_area_struct *vma)
2055{
Andi Kleena5516432008-07-23 21:27:41 -07002056 struct hstate *h = hstate_vma(vma);
Andy Whitcroft84afd992008-07-23 21:27:32 -07002057 struct resv_map *reservations = vma_resv_map(vma);
2058 unsigned long reserve;
2059 unsigned long start;
2060 unsigned long end;
2061
2062 if (reservations) {
Andi Kleena5516432008-07-23 21:27:41 -07002063 start = vma_hugecache_offset(h, vma, vma->vm_start);
2064 end = vma_hugecache_offset(h, vma, vma->vm_end);
Andy Whitcroft84afd992008-07-23 21:27:32 -07002065
2066 reserve = (end - start) -
2067 region_count(&reservations->regions, start, end);
2068
2069 kref_put(&reservations->refs, resv_map_release);
2070
Adam Litke7251ff72008-07-23 21:27:59 -07002071 if (reserve) {
Andi Kleena5516432008-07-23 21:27:41 -07002072 hugetlb_acct_memory(h, -reserve);
Adam Litke7251ff72008-07-23 21:27:59 -07002073 hugetlb_put_quota(vma->vm_file->f_mapping, reserve);
2074 }
Andy Whitcroft84afd992008-07-23 21:27:32 -07002075 }
Mel Gormana1e78772008-07-23 21:27:23 -07002076}
2077
Linus Torvalds1da177e2005-04-16 15:20:36 -07002078/*
2079 * We cannot handle pagefaults against hugetlb pages at all. They cause
2080 * handle_mm_fault() to try to instantiate regular-sized pages in the
2081 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
2082 * this far.
2083 */
Nick Piggind0217ac2007-07-19 01:47:03 -07002084static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002085{
2086 BUG();
Nick Piggind0217ac2007-07-19 01:47:03 -07002087 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002088}
2089
Alexey Dobriyanf0f37e22009-09-27 22:29:37 +04002090const struct vm_operations_struct hugetlb_vm_ops = {
Nick Piggind0217ac2007-07-19 01:47:03 -07002091 .fault = hugetlb_vm_op_fault,
Andy Whitcroft84afd992008-07-23 21:27:32 -07002092 .open = hugetlb_vm_op_open,
Mel Gormana1e78772008-07-23 21:27:23 -07002093 .close = hugetlb_vm_op_close,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002094};
2095
David Gibson1e8f8892006-01-06 00:10:44 -08002096static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
2097 int writable)
David Gibson63551ae2005-06-21 17:14:44 -07002098{
2099 pte_t entry;
2100
David Gibson1e8f8892006-01-06 00:10:44 -08002101 if (writable) {
David Gibson63551ae2005-06-21 17:14:44 -07002102 entry =
2103 pte_mkwrite(pte_mkdirty(mk_pte(page, vma->vm_page_prot)));
2104 } else {
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07002105 entry = huge_pte_wrprotect(mk_pte(page, vma->vm_page_prot));
David Gibson63551ae2005-06-21 17:14:44 -07002106 }
2107 entry = pte_mkyoung(entry);
2108 entry = pte_mkhuge(entry);
2109
2110 return entry;
2111}
2112
David Gibson1e8f8892006-01-06 00:10:44 -08002113static void set_huge_ptep_writable(struct vm_area_struct *vma,
2114 unsigned long address, pte_t *ptep)
2115{
2116 pte_t entry;
2117
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07002118 entry = pte_mkwrite(pte_mkdirty(huge_ptep_get(ptep)));
2119 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) {
Russell King4b3073e2009-12-18 16:40:18 +00002120 update_mmu_cache(vma, address, ptep);
Benjamin Herrenschmidt8dab5242007-06-16 10:16:12 -07002121 }
David Gibson1e8f8892006-01-06 00:10:44 -08002122}
2123
2124
David Gibson63551ae2005-06-21 17:14:44 -07002125int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
2126 struct vm_area_struct *vma)
2127{
2128 pte_t *src_pte, *dst_pte, entry;
2129 struct page *ptepage;
Hugh Dickins1c598272005-10-19 21:23:43 -07002130 unsigned long addr;
David Gibson1e8f8892006-01-06 00:10:44 -08002131 int cow;
Andi Kleena5516432008-07-23 21:27:41 -07002132 struct hstate *h = hstate_vma(vma);
2133 unsigned long sz = huge_page_size(h);
David Gibson1e8f8892006-01-06 00:10:44 -08002134
2135 cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
David Gibson63551ae2005-06-21 17:14:44 -07002136
Andi Kleena5516432008-07-23 21:27:41 -07002137 for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
Hugh Dickinsc74df322005-10-29 18:16:23 -07002138 src_pte = huge_pte_offset(src, addr);
2139 if (!src_pte)
2140 continue;
Andi Kleena5516432008-07-23 21:27:41 -07002141 dst_pte = huge_pte_alloc(dst, addr, sz);
David Gibson63551ae2005-06-21 17:14:44 -07002142 if (!dst_pte)
2143 goto nomem;
Larry Woodmanc5c99422008-01-24 05:49:25 -08002144
2145 /* If the pagetables are shared don't copy or take references */
2146 if (dst_pte == src_pte)
2147 continue;
2148
Hugh Dickinsc74df322005-10-29 18:16:23 -07002149 spin_lock(&dst->page_table_lock);
Nick Piggin46478752008-06-05 22:45:57 -07002150 spin_lock_nested(&src->page_table_lock, SINGLE_DEPTH_NESTING);
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07002151 if (!huge_pte_none(huge_ptep_get(src_pte))) {
David Gibson1e8f8892006-01-06 00:10:44 -08002152 if (cow)
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07002153 huge_ptep_set_wrprotect(src, addr, src_pte);
2154 entry = huge_ptep_get(src_pte);
Hugh Dickins1c598272005-10-19 21:23:43 -07002155 ptepage = pte_page(entry);
2156 get_page(ptepage);
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002157 page_dup_rmap(ptepage);
Hugh Dickins1c598272005-10-19 21:23:43 -07002158 set_huge_pte_at(dst, addr, dst_pte, entry);
2159 }
2160 spin_unlock(&src->page_table_lock);
Hugh Dickinsc74df322005-10-29 18:16:23 -07002161 spin_unlock(&dst->page_table_lock);
David Gibson63551ae2005-06-21 17:14:44 -07002162 }
2163 return 0;
2164
2165nomem:
2166 return -ENOMEM;
2167}
2168
Naoya Horiguchi290408d2010-09-08 10:19:35 +09002169static int is_hugetlb_entry_migration(pte_t pte)
2170{
2171 swp_entry_t swp;
2172
2173 if (huge_pte_none(pte) || pte_present(pte))
2174 return 0;
2175 swp = pte_to_swp_entry(pte);
2176 if (non_swap_entry(swp) && is_migration_entry(swp)) {
2177 return 1;
2178 } else
2179 return 0;
2180}
2181
Naoya Horiguchifd6a03ed2010-05-28 09:29:21 +09002182static int is_hugetlb_entry_hwpoisoned(pte_t pte)
2183{
2184 swp_entry_t swp;
2185
2186 if (huge_pte_none(pte) || pte_present(pte))
2187 return 0;
2188 swp = pte_to_swp_entry(pte);
2189 if (non_swap_entry(swp) && is_hwpoison_entry(swp)) {
2190 return 1;
2191 } else
2192 return 0;
2193}
2194
Chen, Kenneth W502717f2006-10-11 01:20:46 -07002195void __unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002196 unsigned long end, struct page *ref_page)
David Gibson63551ae2005-06-21 17:14:44 -07002197{
2198 struct mm_struct *mm = vma->vm_mm;
2199 unsigned long address;
David Gibsonc7546f82005-08-05 11:59:35 -07002200 pte_t *ptep;
David Gibson63551ae2005-06-21 17:14:44 -07002201 pte_t pte;
2202 struct page *page;
Chen, Kenneth Wfe1668a2006-10-04 02:15:24 -07002203 struct page *tmp;
Andi Kleena5516432008-07-23 21:27:41 -07002204 struct hstate *h = hstate_vma(vma);
2205 unsigned long sz = huge_page_size(h);
2206
Chen, Kenneth Wc0a499c2006-12-06 20:31:39 -08002207 /*
2208 * A page gathering list, protected by per file i_mmap_lock. The
2209 * lock is used to avoid list corruption from multiple unmapping
2210 * of the same page since we are using page->lru.
2211 */
Chen, Kenneth Wfe1668a2006-10-04 02:15:24 -07002212 LIST_HEAD(page_list);
David Gibson63551ae2005-06-21 17:14:44 -07002213
2214 WARN_ON(!is_vm_hugetlb_page(vma));
Andi Kleena5516432008-07-23 21:27:41 -07002215 BUG_ON(start & ~huge_page_mask(h));
2216 BUG_ON(end & ~huge_page_mask(h));
David Gibson63551ae2005-06-21 17:14:44 -07002217
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07002218 mmu_notifier_invalidate_range_start(mm, start, end);
Hugh Dickins508034a2005-10-29 18:16:30 -07002219 spin_lock(&mm->page_table_lock);
Andi Kleena5516432008-07-23 21:27:41 -07002220 for (address = start; address < end; address += sz) {
David Gibsonc7546f82005-08-05 11:59:35 -07002221 ptep = huge_pte_offset(mm, address);
Adam Litke4c887262005-10-29 18:16:46 -07002222 if (!ptep)
David Gibsonc7546f82005-08-05 11:59:35 -07002223 continue;
2224
Chen, Kenneth W39dde652006-12-06 20:32:03 -08002225 if (huge_pmd_unshare(mm, &address, ptep))
2226 continue;
2227
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002228 /*
2229 * If a reference page is supplied, it is because a specific
2230 * page is being unmapped, not a range. Ensure the page we
2231 * are about to unmap is the actual page of interest.
2232 */
2233 if (ref_page) {
2234 pte = huge_ptep_get(ptep);
2235 if (huge_pte_none(pte))
2236 continue;
2237 page = pte_page(pte);
2238 if (page != ref_page)
2239 continue;
2240
2241 /*
2242 * Mark the VMA as having unmapped its page so that
2243 * future faults in this VMA will fail rather than
2244 * looking like data was lost
2245 */
2246 set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
2247 }
2248
David Gibsonc7546f82005-08-05 11:59:35 -07002249 pte = huge_ptep_get_and_clear(mm, address, ptep);
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07002250 if (huge_pte_none(pte))
David Gibson63551ae2005-06-21 17:14:44 -07002251 continue;
David Gibsonc7546f82005-08-05 11:59:35 -07002252
Naoya Horiguchifd6a03ed2010-05-28 09:29:21 +09002253 /*
2254 * HWPoisoned hugepage is already unmapped and dropped reference
2255 */
2256 if (unlikely(is_hugetlb_entry_hwpoisoned(pte)))
2257 continue;
2258
David Gibson63551ae2005-06-21 17:14:44 -07002259 page = pte_page(pte);
Ken Chen6649a382007-02-08 14:20:27 -08002260 if (pte_dirty(pte))
2261 set_page_dirty(page);
Chen, Kenneth Wfe1668a2006-10-04 02:15:24 -07002262 list_add(&page->lru, &page_list);
David Gibson63551ae2005-06-21 17:14:44 -07002263 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002264 spin_unlock(&mm->page_table_lock);
Hugh Dickins508034a2005-10-29 18:16:30 -07002265 flush_tlb_range(vma, start, end);
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07002266 mmu_notifier_invalidate_range_end(mm, start, end);
Chen, Kenneth Wfe1668a2006-10-04 02:15:24 -07002267 list_for_each_entry_safe(page, tmp, &page_list, lru) {
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002268 page_remove_rmap(page);
Chen, Kenneth Wfe1668a2006-10-04 02:15:24 -07002269 list_del(&page->lru);
2270 put_page(page);
2271 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002272}
David Gibson63551ae2005-06-21 17:14:44 -07002273
Chen, Kenneth W502717f2006-10-11 01:20:46 -07002274void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002275 unsigned long end, struct page *ref_page)
Chen, Kenneth W502717f2006-10-11 01:20:46 -07002276{
Andi Kleena137e1c2008-07-23 21:27:43 -07002277 spin_lock(&vma->vm_file->f_mapping->i_mmap_lock);
2278 __unmap_hugepage_range(vma, start, end, ref_page);
2279 spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock);
Chen, Kenneth W502717f2006-10-11 01:20:46 -07002280}
2281
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002282/*
2283 * This is called when the original mapper is failing to COW a MAP_PRIVATE
2284 * mappping it owns the reserve page for. The intention is to unmap the page
2285 * from other VMAs and let the children be SIGKILLed if they are faulting the
2286 * same region.
2287 */
Harvey Harrison2a4b3de2008-10-18 20:27:06 -07002288static int unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
2289 struct page *page, unsigned long address)
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002290{
Adam Litke75266742008-11-12 13:24:56 -08002291 struct hstate *h = hstate_vma(vma);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002292 struct vm_area_struct *iter_vma;
2293 struct address_space *mapping;
2294 struct prio_tree_iter iter;
2295 pgoff_t pgoff;
2296
2297 /*
2298 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
2299 * from page cache lookup which is in HPAGE_SIZE units.
2300 */
Adam Litke75266742008-11-12 13:24:56 -08002301 address = address & huge_page_mask(h);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002302 pgoff = ((address - vma->vm_start) >> PAGE_SHIFT)
2303 + (vma->vm_pgoff >> PAGE_SHIFT);
2304 mapping = (struct address_space *)page_private(page);
2305
Mel Gorman4eb2b1d2009-12-14 17:59:53 -08002306 /*
2307 * Take the mapping lock for the duration of the table walk. As
2308 * this mapping should be shared between all the VMAs,
2309 * __unmap_hugepage_range() is called as the lock is already held
2310 */
2311 spin_lock(&mapping->i_mmap_lock);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002312 vma_prio_tree_foreach(iter_vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
2313 /* Do not unmap the current VMA */
2314 if (iter_vma == vma)
2315 continue;
2316
2317 /*
2318 * Unmap the page from other VMAs without their own reserves.
2319 * They get marked to be SIGKILLed if they fault in these
2320 * areas. This is because a future no-page fault on this VMA
2321 * could insert a zeroed page instead of the data existing
2322 * from the time of fork. This would look like data corruption
2323 */
2324 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
Mel Gorman4eb2b1d2009-12-14 17:59:53 -08002325 __unmap_hugepage_range(iter_vma,
Adam Litke75266742008-11-12 13:24:56 -08002326 address, address + huge_page_size(h),
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002327 page);
2328 }
Mel Gorman4eb2b1d2009-12-14 17:59:53 -08002329 spin_unlock(&mapping->i_mmap_lock);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002330
2331 return 1;
2332}
2333
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002334/*
2335 * Hugetlb_cow() should be called with page lock of the original hugepage held.
2336 */
David Gibson1e8f8892006-01-06 00:10:44 -08002337static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002338 unsigned long address, pte_t *ptep, pte_t pte,
2339 struct page *pagecache_page)
David Gibson1e8f8892006-01-06 00:10:44 -08002340{
Andi Kleena5516432008-07-23 21:27:41 -07002341 struct hstate *h = hstate_vma(vma);
David Gibson1e8f8892006-01-06 00:10:44 -08002342 struct page *old_page, *new_page;
David Gibson79ac6ba2006-03-22 00:08:51 -08002343 int avoidcopy;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002344 int outside_reserve = 0;
David Gibson1e8f8892006-01-06 00:10:44 -08002345
2346 old_page = pte_page(pte);
2347
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002348retry_avoidcopy:
David Gibson1e8f8892006-01-06 00:10:44 -08002349 /* If no-one else is actually using this page, avoid the copy
2350 * and just make the page writable */
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002351 avoidcopy = (page_mapcount(old_page) == 1);
David Gibson1e8f8892006-01-06 00:10:44 -08002352 if (avoidcopy) {
Naoya Horiguchi56c9cfb2010-09-10 13:23:04 +09002353 if (PageAnon(old_page))
2354 page_move_anon_rmap(old_page, vma, address);
David Gibson1e8f8892006-01-06 00:10:44 -08002355 set_huge_ptep_writable(vma, address, ptep);
Nick Piggin83c54072007-07-19 01:47:05 -07002356 return 0;
David Gibson1e8f8892006-01-06 00:10:44 -08002357 }
2358
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002359 /*
2360 * If the process that created a MAP_PRIVATE mapping is about to
2361 * perform a COW due to a shared page count, attempt to satisfy
2362 * the allocation without using the existing reserves. The pagecache
2363 * page is used to determine if the reserve at this address was
2364 * consumed or not. If reserves were used, a partial faulted mapping
2365 * at the time of fork() could consume its reserves on COW instead
2366 * of the full address range.
2367 */
Mel Gormanf83a2752009-05-28 14:34:40 -07002368 if (!(vma->vm_flags & VM_MAYSHARE) &&
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002369 is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
2370 old_page != pagecache_page)
2371 outside_reserve = 1;
2372
David Gibson1e8f8892006-01-06 00:10:44 -08002373 page_cache_get(old_page);
Larry Woodmanb76c8cf2009-12-14 17:59:37 -08002374
2375 /* Drop page_table_lock as buddy allocator may be called */
2376 spin_unlock(&mm->page_table_lock);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002377 new_page = alloc_huge_page(vma, address, outside_reserve);
David Gibson1e8f8892006-01-06 00:10:44 -08002378
Adam Litke2fc39ce2007-11-14 16:59:39 -08002379 if (IS_ERR(new_page)) {
David Gibson1e8f8892006-01-06 00:10:44 -08002380 page_cache_release(old_page);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002381
2382 /*
2383 * If a process owning a MAP_PRIVATE mapping fails to COW,
2384 * it is due to references held by a child and an insufficient
2385 * huge page pool. To guarantee the original mappers
2386 * reliability, unmap the page from child processes. The child
2387 * may get SIGKILLed if it later faults.
2388 */
2389 if (outside_reserve) {
2390 BUG_ON(huge_pte_none(pte));
2391 if (unmap_ref_private(mm, vma, old_page, address)) {
2392 BUG_ON(page_count(old_page) != 1);
2393 BUG_ON(huge_pte_none(pte));
Larry Woodmanb76c8cf2009-12-14 17:59:37 -08002394 spin_lock(&mm->page_table_lock);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002395 goto retry_avoidcopy;
2396 }
2397 WARN_ON_ONCE(1);
2398 }
2399
Larry Woodmanb76c8cf2009-12-14 17:59:37 -08002400 /* Caller expects lock to be held */
2401 spin_lock(&mm->page_table_lock);
Adam Litke2fc39ce2007-11-14 16:59:39 -08002402 return -PTR_ERR(new_page);
David Gibson1e8f8892006-01-06 00:10:44 -08002403 }
2404
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002405 /*
2406 * When the original hugepage is shared one, it does not have
2407 * anon_vma prepared.
2408 */
Dean Nelson44e2aa92010-10-26 14:22:08 -07002409 if (unlikely(anon_vma_prepare(vma))) {
2410 /* Caller expects lock to be held */
2411 spin_lock(&mm->page_table_lock);
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002412 return VM_FAULT_OOM;
Dean Nelson44e2aa92010-10-26 14:22:08 -07002413 }
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002414
Andrea Arcangeli47ad8472011-01-13 15:46:47 -08002415 copy_user_huge_page(new_page, old_page, address, vma,
2416 pages_per_huge_page(h));
Nick Piggin0ed361d2008-02-04 22:29:34 -08002417 __SetPageUptodate(new_page);
David Gibson1e8f8892006-01-06 00:10:44 -08002418
Larry Woodmanb76c8cf2009-12-14 17:59:37 -08002419 /*
2420 * Retake the page_table_lock to check for racing updates
2421 * before the page tables are altered
2422 */
2423 spin_lock(&mm->page_table_lock);
Andi Kleena5516432008-07-23 21:27:41 -07002424 ptep = huge_pte_offset(mm, address & huge_page_mask(h));
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07002425 if (likely(pte_same(huge_ptep_get(ptep), pte))) {
David Gibson1e8f8892006-01-06 00:10:44 -08002426 /* Break COW */
Doug Doan3edd4fc2010-08-09 17:18:30 -07002427 mmu_notifier_invalidate_range_start(mm,
2428 address & huge_page_mask(h),
2429 (address & huge_page_mask(h)) + huge_page_size(h));
Gerald Schaefer8fe627e2008-04-28 02:13:28 -07002430 huge_ptep_clear_flush(vma, address, ptep);
David Gibson1e8f8892006-01-06 00:10:44 -08002431 set_huge_pte_at(mm, address, ptep,
2432 make_huge_pte(vma, new_page, 1));
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002433 page_remove_rmap(old_page);
Naoya Horiguchicd67f0d2010-09-10 13:23:04 +09002434 hugepage_add_new_anon_rmap(new_page, vma, address);
David Gibson1e8f8892006-01-06 00:10:44 -08002435 /* Make the old page be freed below */
2436 new_page = old_page;
Doug Doan3edd4fc2010-08-09 17:18:30 -07002437 mmu_notifier_invalidate_range_end(mm,
2438 address & huge_page_mask(h),
2439 (address & huge_page_mask(h)) + huge_page_size(h));
David Gibson1e8f8892006-01-06 00:10:44 -08002440 }
2441 page_cache_release(new_page);
2442 page_cache_release(old_page);
Nick Piggin83c54072007-07-19 01:47:05 -07002443 return 0;
David Gibson1e8f8892006-01-06 00:10:44 -08002444}
2445
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002446/* Return the pagecache page at a given address within a VMA */
Andi Kleena5516432008-07-23 21:27:41 -07002447static struct page *hugetlbfs_pagecache_page(struct hstate *h,
2448 struct vm_area_struct *vma, unsigned long address)
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002449{
2450 struct address_space *mapping;
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -07002451 pgoff_t idx;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002452
2453 mapping = vma->vm_file->f_mapping;
Andi Kleena5516432008-07-23 21:27:41 -07002454 idx = vma_hugecache_offset(h, vma, address);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002455
2456 return find_lock_page(mapping, idx);
2457}
2458
Hugh Dickins3ae77f42009-09-21 17:03:33 -07002459/*
2460 * Return whether there is a pagecache page to back given address within VMA.
2461 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
2462 */
2463static bool hugetlbfs_pagecache_present(struct hstate *h,
Hugh Dickins2a15efc2009-09-21 17:03:27 -07002464 struct vm_area_struct *vma, unsigned long address)
2465{
2466 struct address_space *mapping;
2467 pgoff_t idx;
2468 struct page *page;
2469
2470 mapping = vma->vm_file->f_mapping;
2471 idx = vma_hugecache_offset(h, vma, address);
2472
2473 page = find_get_page(mapping, idx);
2474 if (page)
2475 put_page(page);
2476 return page != NULL;
2477}
2478
Robert P. J. Daya1ed3dd2007-07-17 04:03:33 -07002479static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
Hugh Dickins788c7df2009-06-23 13:49:05 +01002480 unsigned long address, pte_t *ptep, unsigned int flags)
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01002481{
Andi Kleena5516432008-07-23 21:27:41 -07002482 struct hstate *h = hstate_vma(vma);
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01002483 int ret = VM_FAULT_SIGBUS;
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -07002484 pgoff_t idx;
Adam Litke4c887262005-10-29 18:16:46 -07002485 unsigned long size;
Adam Litke4c887262005-10-29 18:16:46 -07002486 struct page *page;
2487 struct address_space *mapping;
David Gibson1e8f8892006-01-06 00:10:44 -08002488 pte_t new_pte;
Adam Litke4c887262005-10-29 18:16:46 -07002489
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002490 /*
2491 * Currently, we are forced to kill the process in the event the
2492 * original mapper has unmapped pages from the child due to a failed
Lucas De Marchi25985ed2011-03-30 22:57:33 -03002493 * COW. Warn that such a situation has occurred as it may not be obvious
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002494 */
2495 if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
2496 printk(KERN_WARNING
2497 "PID %d killed due to inadequate hugepage pool\n",
2498 current->pid);
2499 return ret;
2500 }
2501
Adam Litke4c887262005-10-29 18:16:46 -07002502 mapping = vma->vm_file->f_mapping;
Andi Kleena5516432008-07-23 21:27:41 -07002503 idx = vma_hugecache_offset(h, vma, address);
Adam Litke4c887262005-10-29 18:16:46 -07002504
2505 /*
2506 * Use page lock to guard against racing truncation
2507 * before we get page_table_lock.
2508 */
Christoph Lameter6bda6662006-01-06 00:10:49 -08002509retry:
2510 page = find_lock_page(mapping, idx);
2511 if (!page) {
Andi Kleena5516432008-07-23 21:27:41 -07002512 size = i_size_read(mapping->host) >> huge_page_shift(h);
Hugh Dickinsebed4bf2006-10-28 10:38:43 -07002513 if (idx >= size)
2514 goto out;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002515 page = alloc_huge_page(vma, address, 0);
Adam Litke2fc39ce2007-11-14 16:59:39 -08002516 if (IS_ERR(page)) {
2517 ret = -PTR_ERR(page);
Christoph Lameter6bda6662006-01-06 00:10:49 -08002518 goto out;
2519 }
Andrea Arcangeli47ad8472011-01-13 15:46:47 -08002520 clear_huge_page(page, address, pages_per_huge_page(h));
Nick Piggin0ed361d2008-02-04 22:29:34 -08002521 __SetPageUptodate(page);
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01002522
Mel Gormanf83a2752009-05-28 14:34:40 -07002523 if (vma->vm_flags & VM_MAYSHARE) {
Christoph Lameter6bda6662006-01-06 00:10:49 -08002524 int err;
Ken Chen45c682a2007-11-14 16:59:44 -08002525 struct inode *inode = mapping->host;
Christoph Lameter6bda6662006-01-06 00:10:49 -08002526
2527 err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
2528 if (err) {
2529 put_page(page);
Christoph Lameter6bda6662006-01-06 00:10:49 -08002530 if (err == -EEXIST)
2531 goto retry;
2532 goto out;
2533 }
Ken Chen45c682a2007-11-14 16:59:44 -08002534
2535 spin_lock(&inode->i_lock);
Andi Kleena5516432008-07-23 21:27:41 -07002536 inode->i_blocks += blocks_per_huge_page(h);
Ken Chen45c682a2007-11-14 16:59:44 -08002537 spin_unlock(&inode->i_lock);
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002538 page_dup_rmap(page);
Mel Gorman23be7462010-04-23 13:17:56 -04002539 } else {
Christoph Lameter6bda6662006-01-06 00:10:49 -08002540 lock_page(page);
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002541 if (unlikely(anon_vma_prepare(vma))) {
2542 ret = VM_FAULT_OOM;
2543 goto backout_unlocked;
2544 }
2545 hugepage_add_new_anon_rmap(page, vma, address);
Mel Gorman23be7462010-04-23 13:17:56 -04002546 }
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002547 } else {
Naoya Horiguchi998b4382010-09-08 10:19:32 +09002548 /*
2549 * If memory error occurs between mmap() and fault, some process
2550 * don't have hwpoisoned swap entry for errored virtual address.
2551 * So we need to block hugepage fault by PG_hwpoison bit check.
2552 */
2553 if (unlikely(PageHWPoison(page))) {
Andi Kleenaa50d3a2010-10-06 21:45:00 +02002554 ret = VM_FAULT_HWPOISON |
2555 VM_FAULT_SET_HINDEX(h - hstates);
Naoya Horiguchi998b4382010-09-08 10:19:32 +09002556 goto backout_unlocked;
2557 }
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002558 page_dup_rmap(page);
Christoph Lameter6bda6662006-01-06 00:10:49 -08002559 }
David Gibson1e8f8892006-01-06 00:10:44 -08002560
Andy Whitcroft57303d82008-08-12 15:08:47 -07002561 /*
2562 * If we are going to COW a private mapping later, we examine the
2563 * pending reservations for this page now. This will ensure that
2564 * any allocations necessary to record that reservation occur outside
2565 * the spinlock.
2566 */
Hugh Dickins788c7df2009-06-23 13:49:05 +01002567 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED))
Andy Whitcroft2b267362008-08-12 15:08:49 -07002568 if (vma_needs_reservation(h, vma, address) < 0) {
2569 ret = VM_FAULT_OOM;
2570 goto backout_unlocked;
2571 }
Andy Whitcroft57303d82008-08-12 15:08:47 -07002572
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01002573 spin_lock(&mm->page_table_lock);
Andi Kleena5516432008-07-23 21:27:41 -07002574 size = i_size_read(mapping->host) >> huge_page_shift(h);
Adam Litke4c887262005-10-29 18:16:46 -07002575 if (idx >= size)
2576 goto backout;
2577
Nick Piggin83c54072007-07-19 01:47:05 -07002578 ret = 0;
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07002579 if (!huge_pte_none(huge_ptep_get(ptep)))
Adam Litke4c887262005-10-29 18:16:46 -07002580 goto backout;
2581
David Gibson1e8f8892006-01-06 00:10:44 -08002582 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
2583 && (vma->vm_flags & VM_SHARED)));
2584 set_huge_pte_at(mm, address, ptep, new_pte);
2585
Hugh Dickins788c7df2009-06-23 13:49:05 +01002586 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
David Gibson1e8f8892006-01-06 00:10:44 -08002587 /* Optimization, do the COW without a second fault */
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002588 ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page);
David Gibson1e8f8892006-01-06 00:10:44 -08002589 }
2590
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01002591 spin_unlock(&mm->page_table_lock);
Adam Litke4c887262005-10-29 18:16:46 -07002592 unlock_page(page);
2593out:
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01002594 return ret;
Adam Litke4c887262005-10-29 18:16:46 -07002595
2596backout:
2597 spin_unlock(&mm->page_table_lock);
Andy Whitcroft2b267362008-08-12 15:08:49 -07002598backout_unlocked:
Adam Litke4c887262005-10-29 18:16:46 -07002599 unlock_page(page);
2600 put_page(page);
2601 goto out;
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01002602}
2603
Adam Litke86e52162006-01-06 00:10:43 -08002604int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
Hugh Dickins788c7df2009-06-23 13:49:05 +01002605 unsigned long address, unsigned int flags)
Adam Litke86e52162006-01-06 00:10:43 -08002606{
2607 pte_t *ptep;
2608 pte_t entry;
David Gibson1e8f8892006-01-06 00:10:44 -08002609 int ret;
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002610 struct page *page = NULL;
Andy Whitcroft57303d82008-08-12 15:08:47 -07002611 struct page *pagecache_page = NULL;
David Gibson3935baa2006-03-22 00:08:53 -08002612 static DEFINE_MUTEX(hugetlb_instantiation_mutex);
Andi Kleena5516432008-07-23 21:27:41 -07002613 struct hstate *h = hstate_vma(vma);
Adam Litke86e52162006-01-06 00:10:43 -08002614
Naoya Horiguchifd6a03ed2010-05-28 09:29:21 +09002615 ptep = huge_pte_offset(mm, address);
2616 if (ptep) {
2617 entry = huge_ptep_get(ptep);
Naoya Horiguchi290408d2010-09-08 10:19:35 +09002618 if (unlikely(is_hugetlb_entry_migration(entry))) {
2619 migration_entry_wait(mm, (pmd_t *)ptep, address);
2620 return 0;
2621 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
Andi Kleenaa50d3a2010-10-06 21:45:00 +02002622 return VM_FAULT_HWPOISON_LARGE |
2623 VM_FAULT_SET_HINDEX(h - hstates);
Naoya Horiguchifd6a03ed2010-05-28 09:29:21 +09002624 }
2625
Andi Kleena5516432008-07-23 21:27:41 -07002626 ptep = huge_pte_alloc(mm, address, huge_page_size(h));
Adam Litke86e52162006-01-06 00:10:43 -08002627 if (!ptep)
2628 return VM_FAULT_OOM;
2629
David Gibson3935baa2006-03-22 00:08:53 -08002630 /*
2631 * Serialize hugepage allocation and instantiation, so that we don't
2632 * get spurious allocation failures if two CPUs race to instantiate
2633 * the same page in the page cache.
2634 */
2635 mutex_lock(&hugetlb_instantiation_mutex);
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07002636 entry = huge_ptep_get(ptep);
2637 if (huge_pte_none(entry)) {
Hugh Dickins788c7df2009-06-23 13:49:05 +01002638 ret = hugetlb_no_page(mm, vma, address, ptep, flags);
David Gibsonb4d1d992008-10-15 22:01:11 -07002639 goto out_mutex;
David Gibson3935baa2006-03-22 00:08:53 -08002640 }
Adam Litke86e52162006-01-06 00:10:43 -08002641
Nick Piggin83c54072007-07-19 01:47:05 -07002642 ret = 0;
David Gibson1e8f8892006-01-06 00:10:44 -08002643
Andy Whitcroft57303d82008-08-12 15:08:47 -07002644 /*
2645 * If we are going to COW the mapping later, we examine the pending
2646 * reservations for this page now. This will ensure that any
2647 * allocations necessary to record that reservation occur outside the
2648 * spinlock. For private mappings, we also lookup the pagecache
2649 * page now as it is used to determine if a reservation has been
2650 * consumed.
2651 */
Hugh Dickins788c7df2009-06-23 13:49:05 +01002652 if ((flags & FAULT_FLAG_WRITE) && !pte_write(entry)) {
Andy Whitcroft2b267362008-08-12 15:08:49 -07002653 if (vma_needs_reservation(h, vma, address) < 0) {
2654 ret = VM_FAULT_OOM;
David Gibsonb4d1d992008-10-15 22:01:11 -07002655 goto out_mutex;
Andy Whitcroft2b267362008-08-12 15:08:49 -07002656 }
Andy Whitcroft57303d82008-08-12 15:08:47 -07002657
Mel Gormanf83a2752009-05-28 14:34:40 -07002658 if (!(vma->vm_flags & VM_MAYSHARE))
Andy Whitcroft57303d82008-08-12 15:08:47 -07002659 pagecache_page = hugetlbfs_pagecache_page(h,
2660 vma, address);
2661 }
2662
Naoya Horiguchi56c9cfb2010-09-10 13:23:04 +09002663 /*
2664 * hugetlb_cow() requires page locks of pte_page(entry) and
2665 * pagecache_page, so here we need take the former one
2666 * when page != pagecache_page or !pagecache_page.
2667 * Note that locking order is always pagecache_page -> page,
2668 * so no worry about deadlock.
2669 */
2670 page = pte_page(entry);
2671 if (page != pagecache_page)
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002672 lock_page(page);
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002673
David Gibson1e8f8892006-01-06 00:10:44 -08002674 spin_lock(&mm->page_table_lock);
2675 /* Check for a racing update before calling hugetlb_cow */
David Gibsonb4d1d992008-10-15 22:01:11 -07002676 if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
2677 goto out_page_table_lock;
2678
2679
Hugh Dickins788c7df2009-06-23 13:49:05 +01002680 if (flags & FAULT_FLAG_WRITE) {
David Gibsonb4d1d992008-10-15 22:01:11 -07002681 if (!pte_write(entry)) {
Andy Whitcroft57303d82008-08-12 15:08:47 -07002682 ret = hugetlb_cow(mm, vma, address, ptep, entry,
2683 pagecache_page);
David Gibsonb4d1d992008-10-15 22:01:11 -07002684 goto out_page_table_lock;
2685 }
2686 entry = pte_mkdirty(entry);
2687 }
2688 entry = pte_mkyoung(entry);
Hugh Dickins788c7df2009-06-23 13:49:05 +01002689 if (huge_ptep_set_access_flags(vma, address, ptep, entry,
2690 flags & FAULT_FLAG_WRITE))
Russell King4b3073e2009-12-18 16:40:18 +00002691 update_mmu_cache(vma, address, ptep);
David Gibsonb4d1d992008-10-15 22:01:11 -07002692
2693out_page_table_lock:
David Gibson1e8f8892006-01-06 00:10:44 -08002694 spin_unlock(&mm->page_table_lock);
Andy Whitcroft57303d82008-08-12 15:08:47 -07002695
2696 if (pagecache_page) {
2697 unlock_page(pagecache_page);
2698 put_page(pagecache_page);
2699 }
Dean Nelson1f64d692010-12-02 14:31:12 -08002700 if (page != pagecache_page)
2701 unlock_page(page);
Andy Whitcroft57303d82008-08-12 15:08:47 -07002702
David Gibsonb4d1d992008-10-15 22:01:11 -07002703out_mutex:
David Gibson3935baa2006-03-22 00:08:53 -08002704 mutex_unlock(&hugetlb_instantiation_mutex);
David Gibson1e8f8892006-01-06 00:10:44 -08002705
2706 return ret;
Adam Litke86e52162006-01-06 00:10:43 -08002707}
2708
Andi Kleenceb86872008-07-23 21:27:50 -07002709/* Can be overriden by architectures */
2710__attribute__((weak)) struct page *
2711follow_huge_pud(struct mm_struct *mm, unsigned long address,
2712 pud_t *pud, int write)
2713{
2714 BUG();
2715 return NULL;
2716}
2717
David Gibson63551ae2005-06-21 17:14:44 -07002718int follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
2719 struct page **pages, struct vm_area_struct **vmas,
Adam Litke5b23dbe2007-11-14 16:59:33 -08002720 unsigned long *position, int *length, int i,
Hugh Dickins2a15efc2009-09-21 17:03:27 -07002721 unsigned int flags)
David Gibson63551ae2005-06-21 17:14:44 -07002722{
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08002723 unsigned long pfn_offset;
2724 unsigned long vaddr = *position;
David Gibson63551ae2005-06-21 17:14:44 -07002725 int remainder = *length;
Andi Kleena5516432008-07-23 21:27:41 -07002726 struct hstate *h = hstate_vma(vma);
David Gibson63551ae2005-06-21 17:14:44 -07002727
Hugh Dickins1c598272005-10-19 21:23:43 -07002728 spin_lock(&mm->page_table_lock);
David Gibson63551ae2005-06-21 17:14:44 -07002729 while (vaddr < vma->vm_end && remainder) {
Adam Litke4c887262005-10-29 18:16:46 -07002730 pte_t *pte;
Hugh Dickins2a15efc2009-09-21 17:03:27 -07002731 int absent;
Adam Litke4c887262005-10-29 18:16:46 -07002732 struct page *page;
2733
2734 /*
2735 * Some archs (sparc64, sh*) have multiple pte_ts to
Hugh Dickins2a15efc2009-09-21 17:03:27 -07002736 * each hugepage. We have to make sure we get the
Adam Litke4c887262005-10-29 18:16:46 -07002737 * first, for the page indexing below to work.
2738 */
Andi Kleena5516432008-07-23 21:27:41 -07002739 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h));
Hugh Dickins2a15efc2009-09-21 17:03:27 -07002740 absent = !pte || huge_pte_none(huge_ptep_get(pte));
Adam Litke4c887262005-10-29 18:16:46 -07002741
Hugh Dickins2a15efc2009-09-21 17:03:27 -07002742 /*
2743 * When coredumping, it suits get_dump_page if we just return
Hugh Dickins3ae77f42009-09-21 17:03:33 -07002744 * an error where there's an empty slot with no huge pagecache
2745 * to back it. This way, we avoid allocating a hugepage, and
2746 * the sparse dumpfile avoids allocating disk blocks, but its
2747 * huge holes still show up with zeroes where they need to be.
Hugh Dickins2a15efc2009-09-21 17:03:27 -07002748 */
Hugh Dickins3ae77f42009-09-21 17:03:33 -07002749 if (absent && (flags & FOLL_DUMP) &&
2750 !hugetlbfs_pagecache_present(h, vma, vaddr)) {
Hugh Dickins2a15efc2009-09-21 17:03:27 -07002751 remainder = 0;
2752 break;
2753 }
2754
2755 if (absent ||
2756 ((flags & FOLL_WRITE) && !pte_write(huge_ptep_get(pte)))) {
Adam Litke4c887262005-10-29 18:16:46 -07002757 int ret;
2758
2759 spin_unlock(&mm->page_table_lock);
Hugh Dickins2a15efc2009-09-21 17:03:27 -07002760 ret = hugetlb_fault(mm, vma, vaddr,
2761 (flags & FOLL_WRITE) ? FAULT_FLAG_WRITE : 0);
Adam Litke4c887262005-10-29 18:16:46 -07002762 spin_lock(&mm->page_table_lock);
Adam Litkea89182c2007-08-22 14:01:51 -07002763 if (!(ret & VM_FAULT_ERROR))
Adam Litke4c887262005-10-29 18:16:46 -07002764 continue;
2765
2766 remainder = 0;
Adam Litke4c887262005-10-29 18:16:46 -07002767 break;
2768 }
David Gibson63551ae2005-06-21 17:14:44 -07002769
Andi Kleena5516432008-07-23 21:27:41 -07002770 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07002771 page = pte_page(huge_ptep_get(pte));
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08002772same_page:
Chen, Kenneth Wd6692182006-03-31 02:29:57 -08002773 if (pages) {
Hugh Dickins2a15efc2009-09-21 17:03:27 -07002774 pages[i] = mem_map_offset(page, pfn_offset);
KOSAKI Motohiro4b2e38a2008-10-18 20:27:10 -07002775 get_page(pages[i]);
Chen, Kenneth Wd6692182006-03-31 02:29:57 -08002776 }
David Gibson63551ae2005-06-21 17:14:44 -07002777
2778 if (vmas)
2779 vmas[i] = vma;
2780
2781 vaddr += PAGE_SIZE;
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08002782 ++pfn_offset;
David Gibson63551ae2005-06-21 17:14:44 -07002783 --remainder;
2784 ++i;
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08002785 if (vaddr < vma->vm_end && remainder &&
Andi Kleena5516432008-07-23 21:27:41 -07002786 pfn_offset < pages_per_huge_page(h)) {
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08002787 /*
2788 * We use pfn_offset to avoid touching the pageframes
2789 * of this compound page.
2790 */
2791 goto same_page;
2792 }
David Gibson63551ae2005-06-21 17:14:44 -07002793 }
Hugh Dickins1c598272005-10-19 21:23:43 -07002794 spin_unlock(&mm->page_table_lock);
David Gibson63551ae2005-06-21 17:14:44 -07002795 *length = remainder;
2796 *position = vaddr;
2797
Hugh Dickins2a15efc2009-09-21 17:03:27 -07002798 return i ? i : -EFAULT;
David Gibson63551ae2005-06-21 17:14:44 -07002799}
Zhang, Yanmin8f860592006-03-22 00:08:50 -08002800
2801void hugetlb_change_protection(struct vm_area_struct *vma,
2802 unsigned long address, unsigned long end, pgprot_t newprot)
2803{
2804 struct mm_struct *mm = vma->vm_mm;
2805 unsigned long start = address;
2806 pte_t *ptep;
2807 pte_t pte;
Andi Kleena5516432008-07-23 21:27:41 -07002808 struct hstate *h = hstate_vma(vma);
Zhang, Yanmin8f860592006-03-22 00:08:50 -08002809
2810 BUG_ON(address >= end);
2811 flush_cache_range(vma, address, end);
2812
Chen, Kenneth W39dde652006-12-06 20:32:03 -08002813 spin_lock(&vma->vm_file->f_mapping->i_mmap_lock);
Zhang, Yanmin8f860592006-03-22 00:08:50 -08002814 spin_lock(&mm->page_table_lock);
Andi Kleena5516432008-07-23 21:27:41 -07002815 for (; address < end; address += huge_page_size(h)) {
Zhang, Yanmin8f860592006-03-22 00:08:50 -08002816 ptep = huge_pte_offset(mm, address);
2817 if (!ptep)
2818 continue;
Chen, Kenneth W39dde652006-12-06 20:32:03 -08002819 if (huge_pmd_unshare(mm, &address, ptep))
2820 continue;
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07002821 if (!huge_pte_none(huge_ptep_get(ptep))) {
Zhang, Yanmin8f860592006-03-22 00:08:50 -08002822 pte = huge_ptep_get_and_clear(mm, address, ptep);
2823 pte = pte_mkhuge(pte_modify(pte, newprot));
2824 set_huge_pte_at(mm, address, ptep, pte);
Zhang, Yanmin8f860592006-03-22 00:08:50 -08002825 }
2826 }
2827 spin_unlock(&mm->page_table_lock);
Chen, Kenneth W39dde652006-12-06 20:32:03 -08002828 spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock);
Zhang, Yanmin8f860592006-03-22 00:08:50 -08002829
2830 flush_tlb_range(vma, start, end);
2831}
2832
Mel Gormana1e78772008-07-23 21:27:23 -07002833int hugetlb_reserve_pages(struct inode *inode,
2834 long from, long to,
Mel Gorman5a6fe122009-02-10 14:02:27 +00002835 struct vm_area_struct *vma,
2836 int acctflag)
Adam Litkee4e574b2007-10-16 01:26:19 -07002837{
Mel Gorman17c9d122009-02-11 16:34:16 +00002838 long ret, chg;
Andi Kleena5516432008-07-23 21:27:41 -07002839 struct hstate *h = hstate_inode(inode);
Adam Litkee4e574b2007-10-16 01:26:19 -07002840
Mel Gormana1e78772008-07-23 21:27:23 -07002841 /*
Mel Gorman17c9d122009-02-11 16:34:16 +00002842 * Only apply hugepage reservation if asked. At fault time, an
2843 * attempt will be made for VM_NORESERVE to allocate a page
2844 * and filesystem quota without using reserves
2845 */
2846 if (acctflag & VM_NORESERVE)
2847 return 0;
2848
2849 /*
Mel Gormana1e78772008-07-23 21:27:23 -07002850 * Shared mappings base their reservation on the number of pages that
2851 * are already allocated on behalf of the file. Private mappings need
2852 * to reserve the full area even if read-only as mprotect() may be
2853 * called to make the mapping read-write. Assume !vma is a shm mapping
2854 */
Mel Gormanf83a2752009-05-28 14:34:40 -07002855 if (!vma || vma->vm_flags & VM_MAYSHARE)
Mel Gormana1e78772008-07-23 21:27:23 -07002856 chg = region_chg(&inode->i_mapping->private_list, from, to);
Mel Gorman5a6fe122009-02-10 14:02:27 +00002857 else {
2858 struct resv_map *resv_map = resv_map_alloc();
Mel Gorman5a6fe122009-02-10 14:02:27 +00002859 if (!resv_map)
2860 return -ENOMEM;
2861
Mel Gorman17c9d122009-02-11 16:34:16 +00002862 chg = to - from;
2863
Mel Gorman5a6fe122009-02-10 14:02:27 +00002864 set_vma_resv_map(vma, resv_map);
2865 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
2866 }
2867
Mel Gorman17c9d122009-02-11 16:34:16 +00002868 if (chg < 0)
2869 return chg;
2870
2871 /* There must be enough filesystem quota for the mapping */
2872 if (hugetlb_get_quota(inode->i_mapping, chg))
2873 return -ENOSPC;
2874
2875 /*
2876 * Check enough hugepages are available for the reservation.
2877 * Hand back the quota if there are not
2878 */
2879 ret = hugetlb_acct_memory(h, chg);
2880 if (ret < 0) {
2881 hugetlb_put_quota(inode->i_mapping, chg);
2882 return ret;
2883 }
2884
2885 /*
2886 * Account for the reservations made. Shared mappings record regions
2887 * that have reservations as they are shared by multiple VMAs.
2888 * When the last VMA disappears, the region map says how much
2889 * the reservation was and the page cache tells how much of
2890 * the reservation was consumed. Private mappings are per-VMA and
2891 * only the consumed reservations are tracked. When the VMA
2892 * disappears, the original reservation is the VMA size and the
2893 * consumed reservations are stored in the map. Hence, nothing
2894 * else has to be done for private mappings here
2895 */
Mel Gormanf83a2752009-05-28 14:34:40 -07002896 if (!vma || vma->vm_flags & VM_MAYSHARE)
Mel Gorman17c9d122009-02-11 16:34:16 +00002897 region_add(&inode->i_mapping->private_list, from, to);
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -07002898 return 0;
2899}
2900
2901void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed)
2902{
Andi Kleena5516432008-07-23 21:27:41 -07002903 struct hstate *h = hstate_inode(inode);
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -07002904 long chg = region_truncate(&inode->i_mapping->private_list, offset);
Ken Chen45c682a2007-11-14 16:59:44 -08002905
2906 spin_lock(&inode->i_lock);
Eric Sandeene4c6f8b2009-07-29 15:02:16 -07002907 inode->i_blocks -= (blocks_per_huge_page(h) * freed);
Ken Chen45c682a2007-11-14 16:59:44 -08002908 spin_unlock(&inode->i_lock);
2909
Adam Litke90d8b7e2007-11-14 16:59:42 -08002910 hugetlb_put_quota(inode->i_mapping, (chg - freed));
Andi Kleena5516432008-07-23 21:27:41 -07002911 hugetlb_acct_memory(h, -(chg - freed));
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -07002912}
Naoya Horiguchi93f70f92010-05-28 09:29:20 +09002913
Andi Kleend5bd9102010-09-27 09:00:12 +02002914#ifdef CONFIG_MEMORY_FAILURE
2915
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +09002916/* Should be called in hugetlb_lock */
2917static int is_hugepage_on_freelist(struct page *hpage)
2918{
2919 struct page *page;
2920 struct page *tmp;
2921 struct hstate *h = page_hstate(hpage);
2922 int nid = page_to_nid(hpage);
2923
2924 list_for_each_entry_safe(page, tmp, &h->hugepage_freelists[nid], lru)
2925 if (page == hpage)
2926 return 1;
2927 return 0;
2928}
2929
Naoya Horiguchi93f70f92010-05-28 09:29:20 +09002930/*
2931 * This function is called from memory failure code.
2932 * Assume the caller holds page lock of the head page.
2933 */
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +09002934int dequeue_hwpoisoned_huge_page(struct page *hpage)
Naoya Horiguchi93f70f92010-05-28 09:29:20 +09002935{
2936 struct hstate *h = page_hstate(hpage);
2937 int nid = page_to_nid(hpage);
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +09002938 int ret = -EBUSY;
Naoya Horiguchi93f70f92010-05-28 09:29:20 +09002939
2940 spin_lock(&hugetlb_lock);
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +09002941 if (is_hugepage_on_freelist(hpage)) {
2942 list_del(&hpage->lru);
Naoya Horiguchi8c6c2ec2010-09-08 10:19:38 +09002943 set_page_refcounted(hpage);
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +09002944 h->free_huge_pages--;
2945 h->free_huge_pages_node[nid]--;
2946 ret = 0;
2947 }
Naoya Horiguchi93f70f92010-05-28 09:29:20 +09002948 spin_unlock(&hugetlb_lock);
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +09002949 return ret;
Naoya Horiguchi93f70f92010-05-28 09:29:20 +09002950}
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +09002951#endif