blob: 815dbd4a6dcb919f28d4331aceae088b8e5f62d7 [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
2 * Generic hugetlb support.
3 * (C) William Irwin, April 2004
4 */
5#include <linux/gfp.h>
6#include <linux/list.h>
7#include <linux/init.h>
8#include <linux/module.h>
9#include <linux/mm.h>
Alexey Dobriyane1759c22008-10-15 23:50:22 +040010#include <linux/seq_file.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070011#include <linux/sysctl.h>
12#include <linux/highmem.h>
Andrea Arcangelicddb8a52008-07-28 15:46:29 -070013#include <linux/mmu_notifier.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070014#include <linux/nodemask.h>
David Gibson63551ae2005-06-21 17:14:44 -070015#include <linux/pagemap.h>
Christoph Lameter5da7ca82006-01-06 00:10:46 -080016#include <linux/mempolicy.h>
Christoph Lameteraea47ff2006-01-08 01:00:57 -080017#include <linux/cpuset.h>
David Gibson3935baa2006-03-22 00:08:53 -080018#include <linux/mutex.h>
Andi Kleenaa888a72008-07-23 21:27:47 -070019#include <linux/bootmem.h>
Nishanth Aravamudana3437872008-07-23 21:27:44 -070020#include <linux/sysfs.h>
Linus Torvaldsd6606682008-08-06 12:04:54 -070021
David Gibson63551ae2005-06-21 17:14:44 -070022#include <asm/page.h>
23#include <asm/pgtable.h>
Adrian Bunk78a34ae2008-07-28 15:46:30 -070024#include <asm/io.h>
David Gibson63551ae2005-06-21 17:14:44 -070025
26#include <linux/hugetlb.h>
Nick Piggin7835e982006-03-22 00:08:40 -080027#include "internal.h"
Linus Torvalds1da177e2005-04-16 15:20:36 -070028
29const unsigned long hugetlb_zero = 0, hugetlb_infinity = ~0UL;
Mel Gorman396faf02007-07-17 04:03:13 -070030static gfp_t htlb_alloc_mask = GFP_HIGHUSER;
31unsigned long hugepages_treat_as_movable;
Andi Kleena5516432008-07-23 21:27:41 -070032
Andi Kleene5ff2152008-07-23 21:27:42 -070033static int max_hstate;
34unsigned int default_hstate_idx;
35struct hstate hstates[HUGE_MAX_HSTATE];
36
Jon Tollefson53ba51d2008-07-23 21:27:52 -070037__initdata LIST_HEAD(huge_boot_pages);
38
Andi Kleene5ff2152008-07-23 21:27:42 -070039/* for command line parsing */
40static struct hstate * __initdata parsed_hstate;
41static unsigned long __initdata default_hstate_max_huge_pages;
Nick Piggine11bfbf2008-07-23 21:27:52 -070042static unsigned long __initdata default_hstate_size;
Andi Kleene5ff2152008-07-23 21:27:42 -070043
44#define for_each_hstate(h) \
45 for ((h) = hstates; (h) < &hstates[max_hstate]; (h)++)
Mel Gorman396faf02007-07-17 04:03:13 -070046
David Gibson3935baa2006-03-22 00:08:53 -080047/*
48 * Protects updates to hugepage_freelists, nr_huge_pages, and free_huge_pages
49 */
50static DEFINE_SPINLOCK(hugetlb_lock);
Eric Paris0bd0f9f2005-11-21 21:32:28 -080051
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -070052/*
Andy Whitcroft96822902008-07-23 21:27:29 -070053 * Region tracking -- allows tracking of reservations and instantiated pages
54 * across the pages in a mapping.
Andy Whitcroft84afd992008-07-23 21:27:32 -070055 *
56 * The region data structures are protected by a combination of the mmap_sem
57 * and the hugetlb_instantion_mutex. To access or modify a region the caller
58 * must either hold the mmap_sem for write, or the mmap_sem for read and
59 * the hugetlb_instantiation mutex:
60 *
61 * down_write(&mm->mmap_sem);
62 * or
63 * down_read(&mm->mmap_sem);
64 * mutex_lock(&hugetlb_instantiation_mutex);
Andy Whitcroft96822902008-07-23 21:27:29 -070065 */
66struct file_region {
67 struct list_head link;
68 long from;
69 long to;
70};
71
72static long region_add(struct list_head *head, long f, long t)
73{
74 struct file_region *rg, *nrg, *trg;
75
76 /* Locate the region we are either in or before. */
77 list_for_each_entry(rg, head, link)
78 if (f <= rg->to)
79 break;
80
81 /* Round our left edge to the current segment if it encloses us. */
82 if (f > rg->from)
83 f = rg->from;
84
85 /* Check for and consume any regions we now overlap with. */
86 nrg = rg;
87 list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
88 if (&rg->link == head)
89 break;
90 if (rg->from > t)
91 break;
92
93 /* If this area reaches higher then extend our area to
94 * include it completely. If this is not the first area
95 * which we intend to reuse, free it. */
96 if (rg->to > t)
97 t = rg->to;
98 if (rg != nrg) {
99 list_del(&rg->link);
100 kfree(rg);
101 }
102 }
103 nrg->from = f;
104 nrg->to = t;
105 return 0;
106}
107
108static long region_chg(struct list_head *head, long f, long t)
109{
110 struct file_region *rg, *nrg;
111 long chg = 0;
112
113 /* Locate the region we are before or in. */
114 list_for_each_entry(rg, head, link)
115 if (f <= rg->to)
116 break;
117
118 /* If we are below the current region then a new region is required.
119 * Subtle, allocate a new region at the position but make it zero
120 * size such that we can guarantee to record the reservation. */
121 if (&rg->link == head || t < rg->from) {
122 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
123 if (!nrg)
124 return -ENOMEM;
125 nrg->from = f;
126 nrg->to = f;
127 INIT_LIST_HEAD(&nrg->link);
128 list_add(&nrg->link, rg->link.prev);
129
130 return t - f;
131 }
132
133 /* Round our left edge to the current segment if it encloses us. */
134 if (f > rg->from)
135 f = rg->from;
136 chg = t - f;
137
138 /* Check for and consume any regions we now overlap with. */
139 list_for_each_entry(rg, rg->link.prev, link) {
140 if (&rg->link == head)
141 break;
142 if (rg->from > t)
143 return chg;
144
145 /* We overlap with this area, if it extends futher than
146 * us then we must extend ourselves. Account for its
147 * existing reservation. */
148 if (rg->to > t) {
149 chg += rg->to - t;
150 t = rg->to;
151 }
152 chg -= rg->to - rg->from;
153 }
154 return chg;
155}
156
157static long region_truncate(struct list_head *head, long end)
158{
159 struct file_region *rg, *trg;
160 long chg = 0;
161
162 /* Locate the region we are either in or before. */
163 list_for_each_entry(rg, head, link)
164 if (end <= rg->to)
165 break;
166 if (&rg->link == head)
167 return 0;
168
169 /* If we are in the middle of a region then adjust it. */
170 if (end > rg->from) {
171 chg = rg->to - end;
172 rg->to = end;
173 rg = list_entry(rg->link.next, typeof(*rg), link);
174 }
175
176 /* Drop any remaining regions. */
177 list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
178 if (&rg->link == head)
179 break;
180 chg += rg->to - rg->from;
181 list_del(&rg->link);
182 kfree(rg);
183 }
184 return chg;
185}
186
Andy Whitcroft84afd992008-07-23 21:27:32 -0700187static long region_count(struct list_head *head, long f, long t)
188{
189 struct file_region *rg;
190 long chg = 0;
191
192 /* Locate each segment we overlap with, and count that overlap. */
193 list_for_each_entry(rg, head, link) {
194 int seg_from;
195 int seg_to;
196
197 if (rg->to <= f)
198 continue;
199 if (rg->from >= t)
200 break;
201
202 seg_from = max(rg->from, f);
203 seg_to = min(rg->to, t);
204
205 chg += seg_to - seg_from;
206 }
207
208 return chg;
209}
210
Andy Whitcroft96822902008-07-23 21:27:29 -0700211/*
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700212 * Convert the address within this vma to the page offset within
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700213 * the mapping, in pagecache page units; huge pages here.
214 */
Andi Kleena5516432008-07-23 21:27:41 -0700215static pgoff_t vma_hugecache_offset(struct hstate *h,
216 struct vm_area_struct *vma, unsigned long address)
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700217{
Andi Kleena5516432008-07-23 21:27:41 -0700218 return ((address - vma->vm_start) >> huge_page_shift(h)) +
219 (vma->vm_pgoff >> huge_page_order(h));
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700220}
221
Andy Whitcroft84afd992008-07-23 21:27:32 -0700222/*
Mel Gorman08fba692009-01-06 14:38:53 -0800223 * Return the size of the pages allocated when backing a VMA. In the majority
224 * cases this will be same size as used by the page table entries.
225 */
226unsigned long vma_kernel_pagesize(struct vm_area_struct *vma)
227{
228 struct hstate *hstate;
229
230 if (!is_vm_hugetlb_page(vma))
231 return PAGE_SIZE;
232
233 hstate = hstate_vma(vma);
234
235 return 1UL << (hstate->order + PAGE_SHIFT);
236}
Joerg Roedelf340ca02009-06-19 15:16:22 +0200237EXPORT_SYMBOL_GPL(vma_kernel_pagesize);
Mel Gorman08fba692009-01-06 14:38:53 -0800238
239/*
Mel Gorman33402892009-01-06 14:38:54 -0800240 * Return the page size being used by the MMU to back a VMA. In the majority
241 * of cases, the page size used by the kernel matches the MMU size. On
242 * architectures where it differs, an architecture-specific version of this
243 * function is required.
244 */
245#ifndef vma_mmu_pagesize
246unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
247{
248 return vma_kernel_pagesize(vma);
249}
250#endif
251
252/*
Andy Whitcroft84afd992008-07-23 21:27:32 -0700253 * Flags for MAP_PRIVATE reservations. These are stored in the bottom
254 * bits of the reservation map pointer, which are always clear due to
255 * alignment.
256 */
257#define HPAGE_RESV_OWNER (1UL << 0)
258#define HPAGE_RESV_UNMAPPED (1UL << 1)
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700259#define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
Andy Whitcroft84afd992008-07-23 21:27:32 -0700260
Mel Gormana1e78772008-07-23 21:27:23 -0700261/*
262 * These helpers are used to track how many pages are reserved for
263 * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
264 * is guaranteed to have their future faults succeed.
265 *
266 * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
267 * the reserve counters are updated with the hugetlb_lock held. It is safe
268 * to reset the VMA at fork() time as it is not in use yet and there is no
269 * chance of the global counters getting corrupted as a result of the values.
Andy Whitcroft84afd992008-07-23 21:27:32 -0700270 *
271 * The private mapping reservation is represented in a subtly different
272 * manner to a shared mapping. A shared mapping has a region map associated
273 * with the underlying file, this region map represents the backing file
274 * pages which have ever had a reservation assigned which this persists even
275 * after the page is instantiated. A private mapping has a region map
276 * associated with the original mmap which is attached to all VMAs which
277 * reference it, this region map represents those offsets which have consumed
278 * reservation ie. where pages have been instantiated.
Mel Gormana1e78772008-07-23 21:27:23 -0700279 */
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700280static unsigned long get_vma_private_data(struct vm_area_struct *vma)
281{
282 return (unsigned long)vma->vm_private_data;
283}
284
285static void set_vma_private_data(struct vm_area_struct *vma,
286 unsigned long value)
287{
288 vma->vm_private_data = (void *)value;
289}
290
Andy Whitcroft84afd992008-07-23 21:27:32 -0700291struct resv_map {
292 struct kref refs;
293 struct list_head regions;
294};
295
Harvey Harrison2a4b3de2008-10-18 20:27:06 -0700296static struct resv_map *resv_map_alloc(void)
Andy Whitcroft84afd992008-07-23 21:27:32 -0700297{
298 struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
299 if (!resv_map)
300 return NULL;
301
302 kref_init(&resv_map->refs);
303 INIT_LIST_HEAD(&resv_map->regions);
304
305 return resv_map;
306}
307
Harvey Harrison2a4b3de2008-10-18 20:27:06 -0700308static void resv_map_release(struct kref *ref)
Andy Whitcroft84afd992008-07-23 21:27:32 -0700309{
310 struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
311
312 /* Clear out any active regions before we release the map. */
313 region_truncate(&resv_map->regions, 0);
314 kfree(resv_map);
315}
316
317static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
Mel Gormana1e78772008-07-23 21:27:23 -0700318{
319 VM_BUG_ON(!is_vm_hugetlb_page(vma));
Mel Gormanf83a2752009-05-28 14:34:40 -0700320 if (!(vma->vm_flags & VM_MAYSHARE))
Andy Whitcroft84afd992008-07-23 21:27:32 -0700321 return (struct resv_map *)(get_vma_private_data(vma) &
322 ~HPAGE_RESV_MASK);
Harvey Harrison2a4b3de2008-10-18 20:27:06 -0700323 return NULL;
Mel Gormana1e78772008-07-23 21:27:23 -0700324}
325
Andy Whitcroft84afd992008-07-23 21:27:32 -0700326static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
Mel Gormana1e78772008-07-23 21:27:23 -0700327{
328 VM_BUG_ON(!is_vm_hugetlb_page(vma));
Mel Gormanf83a2752009-05-28 14:34:40 -0700329 VM_BUG_ON(vma->vm_flags & VM_MAYSHARE);
Mel Gormana1e78772008-07-23 21:27:23 -0700330
Andy Whitcroft84afd992008-07-23 21:27:32 -0700331 set_vma_private_data(vma, (get_vma_private_data(vma) &
332 HPAGE_RESV_MASK) | (unsigned long)map);
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700333}
334
335static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
336{
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700337 VM_BUG_ON(!is_vm_hugetlb_page(vma));
Mel Gormanf83a2752009-05-28 14:34:40 -0700338 VM_BUG_ON(vma->vm_flags & VM_MAYSHARE);
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700339
340 set_vma_private_data(vma, get_vma_private_data(vma) | flags);
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700341}
342
343static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
344{
345 VM_BUG_ON(!is_vm_hugetlb_page(vma));
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700346
347 return (get_vma_private_data(vma) & flag) != 0;
Mel Gormana1e78772008-07-23 21:27:23 -0700348}
349
350/* Decrement the reserved pages in the hugepage pool by one */
Andi Kleena5516432008-07-23 21:27:41 -0700351static void decrement_hugepage_resv_vma(struct hstate *h,
352 struct vm_area_struct *vma)
Mel Gormana1e78772008-07-23 21:27:23 -0700353{
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -0700354 if (vma->vm_flags & VM_NORESERVE)
355 return;
356
Mel Gormanf83a2752009-05-28 14:34:40 -0700357 if (vma->vm_flags & VM_MAYSHARE) {
Mel Gormana1e78772008-07-23 21:27:23 -0700358 /* Shared mappings always use reserves */
Andi Kleena5516432008-07-23 21:27:41 -0700359 h->resv_huge_pages--;
Andy Whitcroft84afd992008-07-23 21:27:32 -0700360 } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
Mel Gormana1e78772008-07-23 21:27:23 -0700361 /*
362 * Only the process that called mmap() has reserves for
363 * private mappings.
364 */
Andi Kleena5516432008-07-23 21:27:41 -0700365 h->resv_huge_pages--;
Mel Gormana1e78772008-07-23 21:27:23 -0700366 }
367}
368
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700369/* Reset counters to 0 and clear all HPAGE_RESV_* flags */
Mel Gormana1e78772008-07-23 21:27:23 -0700370void reset_vma_resv_huge_pages(struct vm_area_struct *vma)
371{
372 VM_BUG_ON(!is_vm_hugetlb_page(vma));
Mel Gormanf83a2752009-05-28 14:34:40 -0700373 if (!(vma->vm_flags & VM_MAYSHARE))
Mel Gormana1e78772008-07-23 21:27:23 -0700374 vma->vm_private_data = (void *)0;
375}
376
377/* Returns true if the VMA has associated reserve pages */
Mel Gorman7f09ca52008-07-23 21:27:58 -0700378static int vma_has_reserves(struct vm_area_struct *vma)
Mel Gormana1e78772008-07-23 21:27:23 -0700379{
Mel Gormanf83a2752009-05-28 14:34:40 -0700380 if (vma->vm_flags & VM_MAYSHARE)
Mel Gorman7f09ca52008-07-23 21:27:58 -0700381 return 1;
382 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER))
383 return 1;
384 return 0;
Mel Gormana1e78772008-07-23 21:27:23 -0700385}
386
Andy Whitcroft69d177c2008-11-06 12:53:26 -0800387static void clear_gigantic_page(struct page *page,
388 unsigned long addr, unsigned long sz)
389{
390 int i;
391 struct page *p = page;
392
393 might_sleep();
394 for (i = 0; i < sz/PAGE_SIZE; i++, p = mem_map_next(p, page, i)) {
395 cond_resched();
396 clear_user_highpage(p, addr + i * PAGE_SIZE);
397 }
398}
Andi Kleena5516432008-07-23 21:27:41 -0700399static void clear_huge_page(struct page *page,
400 unsigned long addr, unsigned long sz)
David Gibson79ac6ba2006-03-22 00:08:51 -0800401{
402 int i;
403
Hannes Ederebdd4ae2009-01-06 14:39:58 -0800404 if (unlikely(sz > MAX_ORDER_NR_PAGES)) {
405 clear_gigantic_page(page, addr, sz);
406 return;
407 }
Andy Whitcroft69d177c2008-11-06 12:53:26 -0800408
David Gibson79ac6ba2006-03-22 00:08:51 -0800409 might_sleep();
Andi Kleena5516432008-07-23 21:27:41 -0700410 for (i = 0; i < sz/PAGE_SIZE; i++) {
David Gibson79ac6ba2006-03-22 00:08:51 -0800411 cond_resched();
Ralf Baechle281e0e32007-10-01 01:20:10 -0700412 clear_user_highpage(page + i, addr + i * PAGE_SIZE);
David Gibson79ac6ba2006-03-22 00:08:51 -0800413 }
414}
415
Andy Whitcroft69d177c2008-11-06 12:53:26 -0800416static void copy_gigantic_page(struct page *dst, struct page *src,
417 unsigned long addr, struct vm_area_struct *vma)
418{
419 int i;
420 struct hstate *h = hstate_vma(vma);
421 struct page *dst_base = dst;
422 struct page *src_base = src;
423 might_sleep();
424 for (i = 0; i < pages_per_huge_page(h); ) {
425 cond_resched();
426 copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
427
428 i++;
429 dst = mem_map_next(dst, dst_base, i);
430 src = mem_map_next(src, src_base, i);
431 }
432}
David Gibson79ac6ba2006-03-22 00:08:51 -0800433static void copy_huge_page(struct page *dst, struct page *src,
Atsushi Nemoto9de455b2006-12-12 17:14:55 +0000434 unsigned long addr, struct vm_area_struct *vma)
David Gibson79ac6ba2006-03-22 00:08:51 -0800435{
436 int i;
Andi Kleena5516432008-07-23 21:27:41 -0700437 struct hstate *h = hstate_vma(vma);
David Gibson79ac6ba2006-03-22 00:08:51 -0800438
Hannes Ederebdd4ae2009-01-06 14:39:58 -0800439 if (unlikely(pages_per_huge_page(h) > MAX_ORDER_NR_PAGES)) {
440 copy_gigantic_page(dst, src, addr, vma);
441 return;
442 }
Andy Whitcroft69d177c2008-11-06 12:53:26 -0800443
David Gibson79ac6ba2006-03-22 00:08:51 -0800444 might_sleep();
Andi Kleena5516432008-07-23 21:27:41 -0700445 for (i = 0; i < pages_per_huge_page(h); i++) {
David Gibson79ac6ba2006-03-22 00:08:51 -0800446 cond_resched();
Atsushi Nemoto9de455b2006-12-12 17:14:55 +0000447 copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
David Gibson79ac6ba2006-03-22 00:08:51 -0800448 }
449}
450
Andi Kleena5516432008-07-23 21:27:41 -0700451static void enqueue_huge_page(struct hstate *h, struct page *page)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700452{
453 int nid = page_to_nid(page);
Andi Kleena5516432008-07-23 21:27:41 -0700454 list_add(&page->lru, &h->hugepage_freelists[nid]);
455 h->free_huge_pages++;
456 h->free_huge_pages_node[nid]++;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700457}
458
Andi Kleena5516432008-07-23 21:27:41 -0700459static struct page *dequeue_huge_page_vma(struct hstate *h,
460 struct vm_area_struct *vma,
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700461 unsigned long address, int avoid_reserve)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700462{
Nishanth Aravamudan31a5c6e2007-07-15 23:38:02 -0700463 int nid;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700464 struct page *page = NULL;
Lee Schermerhorn480eccf2007-09-18 22:46:47 -0700465 struct mempolicy *mpol;
Mel Gorman19770b32008-04-28 02:12:18 -0700466 nodemask_t *nodemask;
Mel Gorman396faf02007-07-17 04:03:13 -0700467 struct zonelist *zonelist = huge_zonelist(vma, address,
Mel Gorman19770b32008-04-28 02:12:18 -0700468 htlb_alloc_mask, &mpol, &nodemask);
Mel Gormandd1a2392008-04-28 02:12:17 -0700469 struct zone *zone;
470 struct zoneref *z;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700471
Mel Gormana1e78772008-07-23 21:27:23 -0700472 /*
473 * A child process with MAP_PRIVATE mappings created by their parent
474 * have no page reserves. This check ensures that reservations are
475 * not "stolen". The child may still get SIGKILLed
476 */
Mel Gorman7f09ca52008-07-23 21:27:58 -0700477 if (!vma_has_reserves(vma) &&
Andi Kleena5516432008-07-23 21:27:41 -0700478 h->free_huge_pages - h->resv_huge_pages == 0)
Mel Gormana1e78772008-07-23 21:27:23 -0700479 return NULL;
480
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700481 /* If reserves cannot be used, ensure enough pages are in the pool */
Andi Kleena5516432008-07-23 21:27:41 -0700482 if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700483 return NULL;
484
Mel Gorman19770b32008-04-28 02:12:18 -0700485 for_each_zone_zonelist_nodemask(zone, z, zonelist,
486 MAX_NR_ZONES - 1, nodemask) {
Mel Gorman54a6eb52008-04-28 02:12:16 -0700487 nid = zone_to_nid(zone);
488 if (cpuset_zone_allowed_softwall(zone, htlb_alloc_mask) &&
Andi Kleena5516432008-07-23 21:27:41 -0700489 !list_empty(&h->hugepage_freelists[nid])) {
490 page = list_entry(h->hugepage_freelists[nid].next,
Andrew Morton3abf7af2007-07-19 01:49:08 -0700491 struct page, lru);
492 list_del(&page->lru);
Andi Kleena5516432008-07-23 21:27:41 -0700493 h->free_huge_pages--;
494 h->free_huge_pages_node[nid]--;
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700495
496 if (!avoid_reserve)
Andi Kleena5516432008-07-23 21:27:41 -0700497 decrement_hugepage_resv_vma(h, vma);
Mel Gormana1e78772008-07-23 21:27:23 -0700498
Ken Chen5ab3ee72007-07-23 18:44:00 -0700499 break;
Andrew Morton3abf7af2007-07-19 01:49:08 -0700500 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700501 }
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -0700502 mpol_cond_put(mpol);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700503 return page;
504}
505
Andi Kleena5516432008-07-23 21:27:41 -0700506static void update_and_free_page(struct hstate *h, struct page *page)
Adam Litke6af2acb2007-10-16 01:26:16 -0700507{
508 int i;
Andi Kleena5516432008-07-23 21:27:41 -0700509
Andy Whitcroft18229df2008-11-06 12:53:27 -0800510 VM_BUG_ON(h->order >= MAX_ORDER);
511
Andi Kleena5516432008-07-23 21:27:41 -0700512 h->nr_huge_pages--;
513 h->nr_huge_pages_node[page_to_nid(page)]--;
514 for (i = 0; i < pages_per_huge_page(h); i++) {
Adam Litke6af2acb2007-10-16 01:26:16 -0700515 page[i].flags &= ~(1 << PG_locked | 1 << PG_error | 1 << PG_referenced |
516 1 << PG_dirty | 1 << PG_active | 1 << PG_reserved |
517 1 << PG_private | 1<< PG_writeback);
518 }
519 set_compound_page_dtor(page, NULL);
520 set_page_refcounted(page);
Gerald Schaefer7f2e9522008-04-28 02:13:29 -0700521 arch_release_hugepage(page);
Andi Kleena5516432008-07-23 21:27:41 -0700522 __free_pages(page, huge_page_order(h));
Adam Litke6af2acb2007-10-16 01:26:16 -0700523}
524
Andi Kleene5ff2152008-07-23 21:27:42 -0700525struct hstate *size_to_hstate(unsigned long size)
526{
527 struct hstate *h;
528
529 for_each_hstate(h) {
530 if (huge_page_size(h) == size)
531 return h;
532 }
533 return NULL;
534}
535
David Gibson27a85ef2006-03-22 00:08:56 -0800536static void free_huge_page(struct page *page)
537{
Andi Kleena5516432008-07-23 21:27:41 -0700538 /*
539 * Can't pass hstate in here because it is called from the
540 * compound page destructor.
541 */
Andi Kleene5ff2152008-07-23 21:27:42 -0700542 struct hstate *h = page_hstate(page);
Adam Litke7893d1d2007-10-16 01:26:18 -0700543 int nid = page_to_nid(page);
Adam Litkec79fb752007-11-14 16:59:38 -0800544 struct address_space *mapping;
David Gibson27a85ef2006-03-22 00:08:56 -0800545
Adam Litkec79fb752007-11-14 16:59:38 -0800546 mapping = (struct address_space *) page_private(page);
Andy Whitcrofte5df70a2008-02-23 15:23:32 -0800547 set_page_private(page, 0);
Adam Litke7893d1d2007-10-16 01:26:18 -0700548 BUG_ON(page_count(page));
David Gibson27a85ef2006-03-22 00:08:56 -0800549 INIT_LIST_HEAD(&page->lru);
550
551 spin_lock(&hugetlb_lock);
Andi Kleenaa888a72008-07-23 21:27:47 -0700552 if (h->surplus_huge_pages_node[nid] && huge_page_order(h) < MAX_ORDER) {
Andi Kleena5516432008-07-23 21:27:41 -0700553 update_and_free_page(h, page);
554 h->surplus_huge_pages--;
555 h->surplus_huge_pages_node[nid]--;
Adam Litke7893d1d2007-10-16 01:26:18 -0700556 } else {
Andi Kleena5516432008-07-23 21:27:41 -0700557 enqueue_huge_page(h, page);
Adam Litke7893d1d2007-10-16 01:26:18 -0700558 }
David Gibson27a85ef2006-03-22 00:08:56 -0800559 spin_unlock(&hugetlb_lock);
Adam Litkec79fb752007-11-14 16:59:38 -0800560 if (mapping)
Adam Litke9a119c02007-11-14 16:59:41 -0800561 hugetlb_put_quota(mapping, 1);
David Gibson27a85ef2006-03-22 00:08:56 -0800562}
563
Andi Kleena5516432008-07-23 21:27:41 -0700564static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
Andi Kleenb7ba30c2008-07-23 21:27:40 -0700565{
566 set_compound_page_dtor(page, free_huge_page);
567 spin_lock(&hugetlb_lock);
Andi Kleena5516432008-07-23 21:27:41 -0700568 h->nr_huge_pages++;
569 h->nr_huge_pages_node[nid]++;
Andi Kleenb7ba30c2008-07-23 21:27:40 -0700570 spin_unlock(&hugetlb_lock);
571 put_page(page); /* free it into the hugepage allocator */
572}
573
Wu Fengguang20a03072009-06-16 15:32:22 -0700574static void prep_compound_gigantic_page(struct page *page, unsigned long order)
575{
576 int i;
577 int nr_pages = 1 << order;
578 struct page *p = page + 1;
579
580 /* we rely on prep_new_huge_page to set the destructor */
581 set_compound_order(page, order);
582 __SetPageHead(page);
583 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
584 __SetPageTail(p);
585 p->first_page = page;
586 }
587}
588
589int PageHuge(struct page *page)
590{
591 compound_page_dtor *dtor;
592
593 if (!PageCompound(page))
594 return 0;
595
596 page = compound_head(page);
597 dtor = get_compound_page_dtor(page);
598
599 return dtor == free_huge_page;
600}
601
Andi Kleena5516432008-07-23 21:27:41 -0700602static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700603{
Linus Torvalds1da177e2005-04-16 15:20:36 -0700604 struct page *page;
Joe Jinf96efd52007-07-15 23:38:12 -0700605
Andi Kleenaa888a72008-07-23 21:27:47 -0700606 if (h->order >= MAX_ORDER)
607 return NULL;
608
Mel Gorman6484eb32009-06-16 15:31:54 -0700609 page = alloc_pages_exact_node(nid,
Nishanth Aravamudan551883a2008-04-29 00:58:26 -0700610 htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE|
611 __GFP_REPEAT|__GFP_NOWARN,
Andi Kleena5516432008-07-23 21:27:41 -0700612 huge_page_order(h));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700613 if (page) {
Gerald Schaefer7f2e9522008-04-28 02:13:29 -0700614 if (arch_prepare_hugepage(page)) {
Gerald Schaefercaff3a22008-08-12 15:08:38 -0700615 __free_pages(page, huge_page_order(h));
Harvey Harrison7b8ee842008-04-28 14:13:19 -0700616 return NULL;
Gerald Schaefer7f2e9522008-04-28 02:13:29 -0700617 }
Andi Kleena5516432008-07-23 21:27:41 -0700618 prep_new_huge_page(h, page, nid);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700619 }
Nishanth Aravamudan63b46132007-10-16 01:26:24 -0700620
621 return page;
622}
623
Andi Kleen5ced66c2008-07-23 21:27:45 -0700624/*
625 * Use a helper variable to find the next node and then
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700626 * copy it back to next_nid_to_alloc afterwards:
Andi Kleen5ced66c2008-07-23 21:27:45 -0700627 * otherwise there's a window in which a racer might
Mel Gorman6484eb32009-06-16 15:31:54 -0700628 * pass invalid nid MAX_NUMNODES to alloc_pages_exact_node.
Andi Kleen5ced66c2008-07-23 21:27:45 -0700629 * But we don't need to use a spin_lock here: it really
630 * doesn't matter if occasionally a racer chooses the
631 * same nid as we do. Move nid forward in the mask even
632 * if we just successfully allocated a hugepage so that
633 * the next caller gets hugepages on the next node.
634 */
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700635static int hstate_next_node_to_alloc(struct hstate *h)
Andi Kleen5ced66c2008-07-23 21:27:45 -0700636{
637 int next_nid;
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700638 next_nid = next_node(h->next_nid_to_alloc, node_online_map);
Andi Kleen5ced66c2008-07-23 21:27:45 -0700639 if (next_nid == MAX_NUMNODES)
640 next_nid = first_node(node_online_map);
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700641 h->next_nid_to_alloc = next_nid;
Andi Kleen5ced66c2008-07-23 21:27:45 -0700642 return next_nid;
643}
644
Andi Kleena5516432008-07-23 21:27:41 -0700645static int alloc_fresh_huge_page(struct hstate *h)
Nishanth Aravamudan63b46132007-10-16 01:26:24 -0700646{
647 struct page *page;
648 int start_nid;
649 int next_nid;
650 int ret = 0;
651
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700652 start_nid = h->next_nid_to_alloc;
653 next_nid = start_nid;
Nishanth Aravamudan63b46132007-10-16 01:26:24 -0700654
655 do {
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700656 page = alloc_fresh_huge_page_node(h, next_nid);
Nishanth Aravamudan63b46132007-10-16 01:26:24 -0700657 if (page)
658 ret = 1;
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700659 next_nid = hstate_next_node_to_alloc(h);
660 } while (!page && next_nid != start_nid);
Nishanth Aravamudan63b46132007-10-16 01:26:24 -0700661
Adam Litke3b116302008-04-28 02:13:06 -0700662 if (ret)
663 count_vm_event(HTLB_BUDDY_PGALLOC);
664 else
665 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
666
Nishanth Aravamudan63b46132007-10-16 01:26:24 -0700667 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700668}
669
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700670/*
671 * helper for free_pool_huge_page() - find next node
672 * from which to free a huge page
673 */
674static int hstate_next_node_to_free(struct hstate *h)
675{
676 int next_nid;
677 next_nid = next_node(h->next_nid_to_free, node_online_map);
678 if (next_nid == MAX_NUMNODES)
679 next_nid = first_node(node_online_map);
680 h->next_nid_to_free = next_nid;
681 return next_nid;
682}
683
684/*
685 * Free huge page from pool from next node to free.
686 * Attempt to keep persistent huge pages more or less
687 * balanced over allowed nodes.
688 * Called with hugetlb_lock locked.
689 */
Lee Schermerhorn685f3452009-09-21 17:01:23 -0700690static int free_pool_huge_page(struct hstate *h, bool acct_surplus)
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700691{
692 int start_nid;
693 int next_nid;
694 int ret = 0;
695
696 start_nid = h->next_nid_to_free;
697 next_nid = start_nid;
698
699 do {
Lee Schermerhorn685f3452009-09-21 17:01:23 -0700700 /*
701 * If we're returning unused surplus pages, only examine
702 * nodes with surplus pages.
703 */
704 if ((!acct_surplus || h->surplus_huge_pages_node[next_nid]) &&
705 !list_empty(&h->hugepage_freelists[next_nid])) {
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700706 struct page *page =
707 list_entry(h->hugepage_freelists[next_nid].next,
708 struct page, lru);
709 list_del(&page->lru);
710 h->free_huge_pages--;
711 h->free_huge_pages_node[next_nid]--;
Lee Schermerhorn685f3452009-09-21 17:01:23 -0700712 if (acct_surplus) {
713 h->surplus_huge_pages--;
714 h->surplus_huge_pages_node[next_nid]--;
715 }
Lee Schermerhorne8c5c822009-09-21 17:01:22 -0700716 update_and_free_page(h, page);
717 ret = 1;
718 }
719 next_nid = hstate_next_node_to_free(h);
720 } while (!ret && next_nid != start_nid);
721
722 return ret;
723}
724
Andi Kleena5516432008-07-23 21:27:41 -0700725static struct page *alloc_buddy_huge_page(struct hstate *h,
726 struct vm_area_struct *vma, unsigned long address)
Adam Litke7893d1d2007-10-16 01:26:18 -0700727{
728 struct page *page;
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800729 unsigned int nid;
Adam Litke7893d1d2007-10-16 01:26:18 -0700730
Andi Kleenaa888a72008-07-23 21:27:47 -0700731 if (h->order >= MAX_ORDER)
732 return NULL;
733
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800734 /*
735 * Assume we will successfully allocate the surplus page to
736 * prevent racing processes from causing the surplus to exceed
737 * overcommit
738 *
739 * This however introduces a different race, where a process B
740 * tries to grow the static hugepage pool while alloc_pages() is
741 * called by process A. B will only examine the per-node
742 * counters in determining if surplus huge pages can be
743 * converted to normal huge pages in adjust_pool_surplus(). A
744 * won't be able to increment the per-node counter, until the
745 * lock is dropped by B, but B doesn't drop hugetlb_lock until
746 * no more huge pages can be converted from surplus to normal
747 * state (and doesn't try to convert again). Thus, we have a
748 * case where a surplus huge page exists, the pool is grown, and
749 * the surplus huge page still exists after, even though it
750 * should just have been converted to a normal huge page. This
751 * does not leak memory, though, as the hugepage will be freed
752 * once it is out of use. It also does not allow the counters to
753 * go out of whack in adjust_pool_surplus() as we don't modify
754 * the node values until we've gotten the hugepage and only the
755 * per-node value is checked there.
756 */
757 spin_lock(&hugetlb_lock);
Andi Kleena5516432008-07-23 21:27:41 -0700758 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800759 spin_unlock(&hugetlb_lock);
760 return NULL;
761 } else {
Andi Kleena5516432008-07-23 21:27:41 -0700762 h->nr_huge_pages++;
763 h->surplus_huge_pages++;
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800764 }
765 spin_unlock(&hugetlb_lock);
766
Nishanth Aravamudan551883a2008-04-29 00:58:26 -0700767 page = alloc_pages(htlb_alloc_mask|__GFP_COMP|
768 __GFP_REPEAT|__GFP_NOWARN,
Andi Kleena5516432008-07-23 21:27:41 -0700769 huge_page_order(h));
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800770
Gerald Schaefercaff3a22008-08-12 15:08:38 -0700771 if (page && arch_prepare_hugepage(page)) {
772 __free_pages(page, huge_page_order(h));
773 return NULL;
774 }
775
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800776 spin_lock(&hugetlb_lock);
Adam Litke7893d1d2007-10-16 01:26:18 -0700777 if (page) {
Adam Litke2668db92008-03-10 11:43:50 -0700778 /*
779 * This page is now managed by the hugetlb allocator and has
780 * no users -- drop the buddy allocator's reference.
781 */
782 put_page_testzero(page);
783 VM_BUG_ON(page_count(page));
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800784 nid = page_to_nid(page);
Adam Litke7893d1d2007-10-16 01:26:18 -0700785 set_compound_page_dtor(page, free_huge_page);
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800786 /*
787 * We incremented the global counters already
788 */
Andi Kleena5516432008-07-23 21:27:41 -0700789 h->nr_huge_pages_node[nid]++;
790 h->surplus_huge_pages_node[nid]++;
Adam Litke3b116302008-04-28 02:13:06 -0700791 __count_vm_event(HTLB_BUDDY_PGALLOC);
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800792 } else {
Andi Kleena5516432008-07-23 21:27:41 -0700793 h->nr_huge_pages--;
794 h->surplus_huge_pages--;
Adam Litke3b116302008-04-28 02:13:06 -0700795 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
Adam Litke7893d1d2007-10-16 01:26:18 -0700796 }
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -0800797 spin_unlock(&hugetlb_lock);
Adam Litke7893d1d2007-10-16 01:26:18 -0700798
799 return page;
800}
801
Adam Litkee4e574b2007-10-16 01:26:19 -0700802/*
803 * Increase the hugetlb pool such that it can accomodate a reservation
804 * of size 'delta'.
805 */
Andi Kleena5516432008-07-23 21:27:41 -0700806static int gather_surplus_pages(struct hstate *h, int delta)
Adam Litkee4e574b2007-10-16 01:26:19 -0700807{
808 struct list_head surplus_list;
809 struct page *page, *tmp;
810 int ret, i;
811 int needed, allocated;
812
Andi Kleena5516432008-07-23 21:27:41 -0700813 needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
Adam Litkeac09b3a2008-03-04 14:29:38 -0800814 if (needed <= 0) {
Andi Kleena5516432008-07-23 21:27:41 -0700815 h->resv_huge_pages += delta;
Adam Litkee4e574b2007-10-16 01:26:19 -0700816 return 0;
Adam Litkeac09b3a2008-03-04 14:29:38 -0800817 }
Adam Litkee4e574b2007-10-16 01:26:19 -0700818
819 allocated = 0;
820 INIT_LIST_HEAD(&surplus_list);
821
822 ret = -ENOMEM;
823retry:
824 spin_unlock(&hugetlb_lock);
825 for (i = 0; i < needed; i++) {
Andi Kleena5516432008-07-23 21:27:41 -0700826 page = alloc_buddy_huge_page(h, NULL, 0);
Adam Litkee4e574b2007-10-16 01:26:19 -0700827 if (!page) {
828 /*
829 * We were not able to allocate enough pages to
830 * satisfy the entire reservation so we free what
831 * we've allocated so far.
832 */
833 spin_lock(&hugetlb_lock);
834 needed = 0;
835 goto free;
836 }
837
838 list_add(&page->lru, &surplus_list);
839 }
840 allocated += needed;
841
842 /*
843 * After retaking hugetlb_lock, we need to recalculate 'needed'
844 * because either resv_huge_pages or free_huge_pages may have changed.
845 */
846 spin_lock(&hugetlb_lock);
Andi Kleena5516432008-07-23 21:27:41 -0700847 needed = (h->resv_huge_pages + delta) -
848 (h->free_huge_pages + allocated);
Adam Litkee4e574b2007-10-16 01:26:19 -0700849 if (needed > 0)
850 goto retry;
851
852 /*
853 * The surplus_list now contains _at_least_ the number of extra pages
854 * needed to accomodate the reservation. Add the appropriate number
855 * of pages to the hugetlb pool and free the extras back to the buddy
Adam Litkeac09b3a2008-03-04 14:29:38 -0800856 * allocator. Commit the entire reservation here to prevent another
857 * process from stealing the pages as they are added to the pool but
858 * before they are reserved.
Adam Litkee4e574b2007-10-16 01:26:19 -0700859 */
860 needed += allocated;
Andi Kleena5516432008-07-23 21:27:41 -0700861 h->resv_huge_pages += delta;
Adam Litkee4e574b2007-10-16 01:26:19 -0700862 ret = 0;
863free:
Adam Litke19fc3f02008-04-28 02:12:20 -0700864 /* Free the needed pages to the hugetlb pool */
Adam Litkee4e574b2007-10-16 01:26:19 -0700865 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
Adam Litke19fc3f02008-04-28 02:12:20 -0700866 if ((--needed) < 0)
867 break;
Adam Litkee4e574b2007-10-16 01:26:19 -0700868 list_del(&page->lru);
Andi Kleena5516432008-07-23 21:27:41 -0700869 enqueue_huge_page(h, page);
Adam Litke19fc3f02008-04-28 02:12:20 -0700870 }
871
872 /* Free unnecessary surplus pages to the buddy allocator */
873 if (!list_empty(&surplus_list)) {
874 spin_unlock(&hugetlb_lock);
875 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
876 list_del(&page->lru);
Adam Litkeaf767cb2007-10-16 01:26:25 -0700877 /*
Adam Litke2668db92008-03-10 11:43:50 -0700878 * The page has a reference count of zero already, so
879 * call free_huge_page directly instead of using
880 * put_page. This must be done with hugetlb_lock
Adam Litkeaf767cb2007-10-16 01:26:25 -0700881 * unlocked which is safe because free_huge_page takes
882 * hugetlb_lock before deciding how to free the page.
883 */
Adam Litke2668db92008-03-10 11:43:50 -0700884 free_huge_page(page);
Adam Litkeaf767cb2007-10-16 01:26:25 -0700885 }
Adam Litke19fc3f02008-04-28 02:12:20 -0700886 spin_lock(&hugetlb_lock);
Adam Litkee4e574b2007-10-16 01:26:19 -0700887 }
888
889 return ret;
890}
891
892/*
893 * When releasing a hugetlb pool reservation, any surplus pages that were
894 * allocated to satisfy the reservation must be explicitly freed if they were
895 * never used.
Lee Schermerhorn685f3452009-09-21 17:01:23 -0700896 * Called with hugetlb_lock held.
Adam Litkee4e574b2007-10-16 01:26:19 -0700897 */
Andi Kleena5516432008-07-23 21:27:41 -0700898static void return_unused_surplus_pages(struct hstate *h,
899 unsigned long unused_resv_pages)
Adam Litkee4e574b2007-10-16 01:26:19 -0700900{
Adam Litkee4e574b2007-10-16 01:26:19 -0700901 unsigned long nr_pages;
902
Adam Litkeac09b3a2008-03-04 14:29:38 -0800903 /* Uncommit the reservation */
Andi Kleena5516432008-07-23 21:27:41 -0700904 h->resv_huge_pages -= unused_resv_pages;
Adam Litkeac09b3a2008-03-04 14:29:38 -0800905
Andi Kleenaa888a72008-07-23 21:27:47 -0700906 /* Cannot return gigantic pages currently */
907 if (h->order >= MAX_ORDER)
908 return;
909
Andi Kleena5516432008-07-23 21:27:41 -0700910 nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
Adam Litkee4e574b2007-10-16 01:26:19 -0700911
Lee Schermerhorn685f3452009-09-21 17:01:23 -0700912 /*
913 * We want to release as many surplus pages as possible, spread
914 * evenly across all nodes. Iterate across all nodes until we
915 * can no longer free unreserved surplus pages. This occurs when
916 * the nodes with surplus pages have no free pages.
917 * free_pool_huge_page() will balance the the frees across the
918 * on-line nodes for us and will handle the hstate accounting.
919 */
920 while (nr_pages--) {
921 if (!free_pool_huge_page(h, 1))
922 break;
Adam Litkee4e574b2007-10-16 01:26:19 -0700923 }
924}
925
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -0700926/*
927 * Determine if the huge page at addr within the vma has an associated
928 * reservation. Where it does not we will need to logically increase
929 * reservation and actually increase quota before an allocation can occur.
930 * Where any new reservation would be required the reservation change is
931 * prepared, but not committed. Once the page has been quota'd allocated
932 * an instantiated the change should be committed via vma_commit_reservation.
933 * No action is required on failure.
934 */
Roel Kluine2f17d92009-03-31 15:23:15 -0700935static long vma_needs_reservation(struct hstate *h,
Andi Kleena5516432008-07-23 21:27:41 -0700936 struct vm_area_struct *vma, unsigned long addr)
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -0700937{
938 struct address_space *mapping = vma->vm_file->f_mapping;
939 struct inode *inode = mapping->host;
940
Mel Gormanf83a2752009-05-28 14:34:40 -0700941 if (vma->vm_flags & VM_MAYSHARE) {
Andi Kleena5516432008-07-23 21:27:41 -0700942 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -0700943 return region_chg(&inode->i_mapping->private_list,
944 idx, idx + 1);
945
Andy Whitcroft84afd992008-07-23 21:27:32 -0700946 } else if (!is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
947 return 1;
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -0700948
Andy Whitcroft84afd992008-07-23 21:27:32 -0700949 } else {
Roel Kluine2f17d92009-03-31 15:23:15 -0700950 long err;
Andi Kleena5516432008-07-23 21:27:41 -0700951 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
Andy Whitcroft84afd992008-07-23 21:27:32 -0700952 struct resv_map *reservations = vma_resv_map(vma);
953
954 err = region_chg(&reservations->regions, idx, idx + 1);
955 if (err < 0)
956 return err;
957 return 0;
958 }
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -0700959}
Andi Kleena5516432008-07-23 21:27:41 -0700960static void vma_commit_reservation(struct hstate *h,
961 struct vm_area_struct *vma, unsigned long addr)
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -0700962{
963 struct address_space *mapping = vma->vm_file->f_mapping;
964 struct inode *inode = mapping->host;
965
Mel Gormanf83a2752009-05-28 14:34:40 -0700966 if (vma->vm_flags & VM_MAYSHARE) {
Andi Kleena5516432008-07-23 21:27:41 -0700967 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -0700968 region_add(&inode->i_mapping->private_list, idx, idx + 1);
Andy Whitcroft84afd992008-07-23 21:27:32 -0700969
970 } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
Andi Kleena5516432008-07-23 21:27:41 -0700971 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
Andy Whitcroft84afd992008-07-23 21:27:32 -0700972 struct resv_map *reservations = vma_resv_map(vma);
973
974 /* Mark this page used in the map. */
975 region_add(&reservations->regions, idx, idx + 1);
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -0700976 }
977}
978
David Gibson27a85ef2006-03-22 00:08:56 -0800979static struct page *alloc_huge_page(struct vm_area_struct *vma,
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700980 unsigned long addr, int avoid_reserve)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700981{
Andi Kleena5516432008-07-23 21:27:41 -0700982 struct hstate *h = hstate_vma(vma);
Adam Litke348ea202007-11-14 16:59:37 -0800983 struct page *page;
Adam Litke2fc39ce2007-11-14 16:59:39 -0800984 struct address_space *mapping = vma->vm_file->f_mapping;
Mel Gormana1e78772008-07-23 21:27:23 -0700985 struct inode *inode = mapping->host;
Roel Kluine2f17d92009-03-31 15:23:15 -0700986 long chg;
Adam Litke2fc39ce2007-11-14 16:59:39 -0800987
Mel Gormana1e78772008-07-23 21:27:23 -0700988 /*
989 * Processes that did not create the mapping will have no reserves and
990 * will not have accounted against quota. Check that the quota can be
991 * made before satisfying the allocation
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -0700992 * MAP_NORESERVE mappings may also need pages and quota allocated
993 * if no reserve mapping overlaps.
Mel Gormana1e78772008-07-23 21:27:23 -0700994 */
Andi Kleena5516432008-07-23 21:27:41 -0700995 chg = vma_needs_reservation(h, vma, addr);
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -0700996 if (chg < 0)
997 return ERR_PTR(chg);
998 if (chg)
Mel Gormana1e78772008-07-23 21:27:23 -0700999 if (hugetlb_get_quota(inode->i_mapping, chg))
1000 return ERR_PTR(-ENOSPC);
Mel Gormana1e78772008-07-23 21:27:23 -07001001
1002 spin_lock(&hugetlb_lock);
Andi Kleena5516432008-07-23 21:27:41 -07001003 page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve);
Mel Gormana1e78772008-07-23 21:27:23 -07001004 spin_unlock(&hugetlb_lock);
1005
1006 if (!page) {
Andi Kleena5516432008-07-23 21:27:41 -07001007 page = alloc_buddy_huge_page(h, vma, addr);
Mel Gormana1e78772008-07-23 21:27:23 -07001008 if (!page) {
1009 hugetlb_put_quota(inode->i_mapping, chg);
1010 return ERR_PTR(-VM_FAULT_OOM);
1011 }
1012 }
1013
1014 set_page_refcounted(page);
1015 set_page_private(page, (unsigned long) mapping);
1016
Andi Kleena5516432008-07-23 21:27:41 -07001017 vma_commit_reservation(h, vma, addr);
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001018
Adam Litke90d8b7e2007-11-14 16:59:42 -08001019 return page;
David Gibsonb45b5bd2006-03-22 00:08:55 -08001020}
1021
Cyrill Gorcunov91f47662009-01-06 14:40:33 -08001022int __weak alloc_bootmem_huge_page(struct hstate *h)
Andi Kleenaa888a72008-07-23 21:27:47 -07001023{
1024 struct huge_bootmem_page *m;
1025 int nr_nodes = nodes_weight(node_online_map);
1026
1027 while (nr_nodes) {
1028 void *addr;
1029
1030 addr = __alloc_bootmem_node_nopanic(
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001031 NODE_DATA(h->next_nid_to_alloc),
Andi Kleenaa888a72008-07-23 21:27:47 -07001032 huge_page_size(h), huge_page_size(h), 0);
1033
Lee Schermerhorn57dd28f2009-09-21 17:01:25 -07001034 hstate_next_node_to_alloc(h);
Andi Kleenaa888a72008-07-23 21:27:47 -07001035 if (addr) {
1036 /*
1037 * Use the beginning of the huge page to store the
1038 * huge_bootmem_page struct (until gather_bootmem
1039 * puts them into the mem_map).
1040 */
1041 m = addr;
Cyrill Gorcunov91f47662009-01-06 14:40:33 -08001042 goto found;
Andi Kleenaa888a72008-07-23 21:27:47 -07001043 }
Andi Kleenaa888a72008-07-23 21:27:47 -07001044 nr_nodes--;
1045 }
1046 return 0;
1047
1048found:
1049 BUG_ON((unsigned long)virt_to_phys(m) & (huge_page_size(h) - 1));
1050 /* Put them into a private list first because mem_map is not up yet */
1051 list_add(&m->list, &huge_boot_pages);
1052 m->hstate = h;
1053 return 1;
1054}
1055
Andy Whitcroft18229df2008-11-06 12:53:27 -08001056static void prep_compound_huge_page(struct page *page, int order)
1057{
1058 if (unlikely(order > (MAX_ORDER - 1)))
1059 prep_compound_gigantic_page(page, order);
1060 else
1061 prep_compound_page(page, order);
1062}
1063
Andi Kleenaa888a72008-07-23 21:27:47 -07001064/* Put bootmem huge pages into the standard lists after mem_map is up */
1065static void __init gather_bootmem_prealloc(void)
1066{
1067 struct huge_bootmem_page *m;
1068
1069 list_for_each_entry(m, &huge_boot_pages, list) {
1070 struct page *page = virt_to_page(m);
1071 struct hstate *h = m->hstate;
1072 __ClearPageReserved(page);
1073 WARN_ON(page_count(page) != 1);
Andy Whitcroft18229df2008-11-06 12:53:27 -08001074 prep_compound_huge_page(page, h->order);
Andi Kleenaa888a72008-07-23 21:27:47 -07001075 prep_new_huge_page(h, page, page_to_nid(page));
1076 }
1077}
1078
Andi Kleen8faa8b02008-07-23 21:27:48 -07001079static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001080{
1081 unsigned long i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001082
Andi Kleene5ff2152008-07-23 21:27:42 -07001083 for (i = 0; i < h->max_huge_pages; ++i) {
Andi Kleenaa888a72008-07-23 21:27:47 -07001084 if (h->order >= MAX_ORDER) {
1085 if (!alloc_bootmem_huge_page(h))
1086 break;
1087 } else if (!alloc_fresh_huge_page(h))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001088 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001089 }
Andi Kleen8faa8b02008-07-23 21:27:48 -07001090 h->max_huge_pages = i;
Andi Kleene5ff2152008-07-23 21:27:42 -07001091}
1092
1093static void __init hugetlb_init_hstates(void)
1094{
1095 struct hstate *h;
1096
1097 for_each_hstate(h) {
Andi Kleen8faa8b02008-07-23 21:27:48 -07001098 /* oversize hugepages were init'ed in early boot */
1099 if (h->order < MAX_ORDER)
1100 hugetlb_hstate_alloc_pages(h);
Andi Kleene5ff2152008-07-23 21:27:42 -07001101 }
1102}
1103
Andi Kleen4abd32d2008-07-23 21:27:49 -07001104static char * __init memfmt(char *buf, unsigned long n)
1105{
1106 if (n >= (1UL << 30))
1107 sprintf(buf, "%lu GB", n >> 30);
1108 else if (n >= (1UL << 20))
1109 sprintf(buf, "%lu MB", n >> 20);
1110 else
1111 sprintf(buf, "%lu KB", n >> 10);
1112 return buf;
1113}
1114
Andi Kleene5ff2152008-07-23 21:27:42 -07001115static void __init report_hugepages(void)
1116{
1117 struct hstate *h;
1118
1119 for_each_hstate(h) {
Andi Kleen4abd32d2008-07-23 21:27:49 -07001120 char buf[32];
1121 printk(KERN_INFO "HugeTLB registered %s page size, "
1122 "pre-allocated %ld pages\n",
1123 memfmt(buf, huge_page_size(h)),
1124 h->free_huge_pages);
Andi Kleene5ff2152008-07-23 21:27:42 -07001125 }
1126}
1127
Linus Torvalds1da177e2005-04-16 15:20:36 -07001128#ifdef CONFIG_HIGHMEM
Andi Kleena5516432008-07-23 21:27:41 -07001129static void try_to_free_low(struct hstate *h, unsigned long count)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001130{
Christoph Lameter4415cc82006-09-25 23:31:55 -07001131 int i;
1132
Andi Kleenaa888a72008-07-23 21:27:47 -07001133 if (h->order >= MAX_ORDER)
1134 return;
1135
Linus Torvalds1da177e2005-04-16 15:20:36 -07001136 for (i = 0; i < MAX_NUMNODES; ++i) {
1137 struct page *page, *next;
Andi Kleena5516432008-07-23 21:27:41 -07001138 struct list_head *freel = &h->hugepage_freelists[i];
1139 list_for_each_entry_safe(page, next, freel, lru) {
1140 if (count >= h->nr_huge_pages)
Adam Litke6b0c8802007-10-16 01:26:23 -07001141 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001142 if (PageHighMem(page))
1143 continue;
1144 list_del(&page->lru);
Andi Kleene5ff2152008-07-23 21:27:42 -07001145 update_and_free_page(h, page);
Andi Kleena5516432008-07-23 21:27:41 -07001146 h->free_huge_pages--;
1147 h->free_huge_pages_node[page_to_nid(page)]--;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001148 }
1149 }
1150}
1151#else
Andi Kleena5516432008-07-23 21:27:41 -07001152static inline void try_to_free_low(struct hstate *h, unsigned long count)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001153{
1154}
1155#endif
1156
Wu Fengguang20a03072009-06-16 15:32:22 -07001157/*
1158 * Increment or decrement surplus_huge_pages. Keep node-specific counters
1159 * balanced by operating on them in a round-robin fashion.
1160 * Returns 1 if an adjustment was made.
1161 */
1162static int adjust_pool_surplus(struct hstate *h, int delta)
1163{
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001164 int start_nid, next_nid;
Wu Fengguang20a03072009-06-16 15:32:22 -07001165 int ret = 0;
1166
1167 VM_BUG_ON(delta != -1 && delta != 1);
Wu Fengguang20a03072009-06-16 15:32:22 -07001168
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001169 if (delta < 0)
1170 start_nid = h->next_nid_to_alloc;
1171 else
1172 start_nid = h->next_nid_to_free;
1173 next_nid = start_nid;
1174
1175 do {
1176 int nid = next_nid;
1177 if (delta < 0) {
1178 next_nid = hstate_next_node_to_alloc(h);
1179 /*
1180 * To shrink on this node, there must be a surplus page
1181 */
1182 if (!h->surplus_huge_pages_node[nid])
1183 continue;
1184 }
1185 if (delta > 0) {
1186 next_nid = hstate_next_node_to_free(h);
1187 /*
1188 * Surplus cannot exceed the total number of pages
1189 */
1190 if (h->surplus_huge_pages_node[nid] >=
Wu Fengguang20a03072009-06-16 15:32:22 -07001191 h->nr_huge_pages_node[nid])
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001192 continue;
1193 }
Wu Fengguang20a03072009-06-16 15:32:22 -07001194
1195 h->surplus_huge_pages += delta;
1196 h->surplus_huge_pages_node[nid] += delta;
1197 ret = 1;
1198 break;
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001199 } while (next_nid != start_nid);
Wu Fengguang20a03072009-06-16 15:32:22 -07001200
Wu Fengguang20a03072009-06-16 15:32:22 -07001201 return ret;
1202}
1203
Andi Kleena5516432008-07-23 21:27:41 -07001204#define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
Andi Kleene5ff2152008-07-23 21:27:42 -07001205static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001206{
Adam Litke7893d1d2007-10-16 01:26:18 -07001207 unsigned long min_count, ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001208
Andi Kleenaa888a72008-07-23 21:27:47 -07001209 if (h->order >= MAX_ORDER)
1210 return h->max_huge_pages;
1211
Adam Litke7893d1d2007-10-16 01:26:18 -07001212 /*
1213 * Increase the pool size
1214 * First take pages out of surplus state. Then make up the
1215 * remaining difference by allocating fresh huge pages.
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001216 *
1217 * We might race with alloc_buddy_huge_page() here and be unable
1218 * to convert a surplus huge page to a normal huge page. That is
1219 * not critical, though, it just means the overall size of the
1220 * pool might be one hugepage larger than it needs to be, but
1221 * within all the constraints specified by the sysctls.
Adam Litke7893d1d2007-10-16 01:26:18 -07001222 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001223 spin_lock(&hugetlb_lock);
Andi Kleena5516432008-07-23 21:27:41 -07001224 while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
1225 if (!adjust_pool_surplus(h, -1))
Adam Litke7893d1d2007-10-16 01:26:18 -07001226 break;
1227 }
1228
Andi Kleena5516432008-07-23 21:27:41 -07001229 while (count > persistent_huge_pages(h)) {
Adam Litke7893d1d2007-10-16 01:26:18 -07001230 /*
1231 * If this allocation races such that we no longer need the
1232 * page, free_huge_page will handle it by freeing the page
1233 * and reducing the surplus.
1234 */
1235 spin_unlock(&hugetlb_lock);
Andi Kleena5516432008-07-23 21:27:41 -07001236 ret = alloc_fresh_huge_page(h);
Adam Litke7893d1d2007-10-16 01:26:18 -07001237 spin_lock(&hugetlb_lock);
1238 if (!ret)
1239 goto out;
1240
1241 }
Adam Litke7893d1d2007-10-16 01:26:18 -07001242
1243 /*
1244 * Decrease the pool size
1245 * First return free pages to the buddy allocator (being careful
1246 * to keep enough around to satisfy reservations). Then place
1247 * pages into surplus state as needed so the pool will shrink
1248 * to the desired size as pages become free.
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001249 *
1250 * By placing pages into the surplus state independent of the
1251 * overcommit value, we are allowing the surplus pool size to
1252 * exceed overcommit. There are few sane options here. Since
1253 * alloc_buddy_huge_page() is checking the global counter,
1254 * though, we'll note that we're not allowed to exceed surplus
1255 * and won't grow the pool anywhere else. Not until one of the
1256 * sysctls are changed, or the surplus pages go out of use.
Adam Litke7893d1d2007-10-16 01:26:18 -07001257 */
Andi Kleena5516432008-07-23 21:27:41 -07001258 min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
Adam Litke6b0c8802007-10-16 01:26:23 -07001259 min_count = max(count, min_count);
Andi Kleena5516432008-07-23 21:27:41 -07001260 try_to_free_low(h, min_count);
1261 while (min_count < persistent_huge_pages(h)) {
Lee Schermerhorn685f3452009-09-21 17:01:23 -07001262 if (!free_pool_huge_page(h, 0))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001263 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001264 }
Andi Kleena5516432008-07-23 21:27:41 -07001265 while (count < persistent_huge_pages(h)) {
1266 if (!adjust_pool_surplus(h, 1))
Adam Litke7893d1d2007-10-16 01:26:18 -07001267 break;
1268 }
1269out:
Andi Kleena5516432008-07-23 21:27:41 -07001270 ret = persistent_huge_pages(h);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001271 spin_unlock(&hugetlb_lock);
Adam Litke7893d1d2007-10-16 01:26:18 -07001272 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001273}
1274
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001275#define HSTATE_ATTR_RO(_name) \
1276 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
1277
1278#define HSTATE_ATTR(_name) \
1279 static struct kobj_attribute _name##_attr = \
1280 __ATTR(_name, 0644, _name##_show, _name##_store)
1281
1282static struct kobject *hugepages_kobj;
1283static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
1284
1285static struct hstate *kobj_to_hstate(struct kobject *kobj)
1286{
1287 int i;
1288 for (i = 0; i < HUGE_MAX_HSTATE; i++)
1289 if (hstate_kobjs[i] == kobj)
1290 return &hstates[i];
1291 BUG();
1292 return NULL;
1293}
1294
1295static ssize_t nr_hugepages_show(struct kobject *kobj,
1296 struct kobj_attribute *attr, char *buf)
1297{
1298 struct hstate *h = kobj_to_hstate(kobj);
1299 return sprintf(buf, "%lu\n", h->nr_huge_pages);
1300}
1301static ssize_t nr_hugepages_store(struct kobject *kobj,
1302 struct kobj_attribute *attr, const char *buf, size_t count)
1303{
1304 int err;
1305 unsigned long input;
1306 struct hstate *h = kobj_to_hstate(kobj);
1307
1308 err = strict_strtoul(buf, 10, &input);
1309 if (err)
1310 return 0;
1311
1312 h->max_huge_pages = set_max_huge_pages(h, input);
1313
1314 return count;
1315}
1316HSTATE_ATTR(nr_hugepages);
1317
1318static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
1319 struct kobj_attribute *attr, char *buf)
1320{
1321 struct hstate *h = kobj_to_hstate(kobj);
1322 return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
1323}
1324static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
1325 struct kobj_attribute *attr, const char *buf, size_t count)
1326{
1327 int err;
1328 unsigned long input;
1329 struct hstate *h = kobj_to_hstate(kobj);
1330
1331 err = strict_strtoul(buf, 10, &input);
1332 if (err)
1333 return 0;
1334
1335 spin_lock(&hugetlb_lock);
1336 h->nr_overcommit_huge_pages = input;
1337 spin_unlock(&hugetlb_lock);
1338
1339 return count;
1340}
1341HSTATE_ATTR(nr_overcommit_hugepages);
1342
1343static ssize_t free_hugepages_show(struct kobject *kobj,
1344 struct kobj_attribute *attr, char *buf)
1345{
1346 struct hstate *h = kobj_to_hstate(kobj);
1347 return sprintf(buf, "%lu\n", h->free_huge_pages);
1348}
1349HSTATE_ATTR_RO(free_hugepages);
1350
1351static ssize_t resv_hugepages_show(struct kobject *kobj,
1352 struct kobj_attribute *attr, char *buf)
1353{
1354 struct hstate *h = kobj_to_hstate(kobj);
1355 return sprintf(buf, "%lu\n", h->resv_huge_pages);
1356}
1357HSTATE_ATTR_RO(resv_hugepages);
1358
1359static ssize_t surplus_hugepages_show(struct kobject *kobj,
1360 struct kobj_attribute *attr, char *buf)
1361{
1362 struct hstate *h = kobj_to_hstate(kobj);
1363 return sprintf(buf, "%lu\n", h->surplus_huge_pages);
1364}
1365HSTATE_ATTR_RO(surplus_hugepages);
1366
1367static struct attribute *hstate_attrs[] = {
1368 &nr_hugepages_attr.attr,
1369 &nr_overcommit_hugepages_attr.attr,
1370 &free_hugepages_attr.attr,
1371 &resv_hugepages_attr.attr,
1372 &surplus_hugepages_attr.attr,
1373 NULL,
1374};
1375
1376static struct attribute_group hstate_attr_group = {
1377 .attrs = hstate_attrs,
1378};
1379
1380static int __init hugetlb_sysfs_add_hstate(struct hstate *h)
1381{
1382 int retval;
1383
1384 hstate_kobjs[h - hstates] = kobject_create_and_add(h->name,
1385 hugepages_kobj);
1386 if (!hstate_kobjs[h - hstates])
1387 return -ENOMEM;
1388
1389 retval = sysfs_create_group(hstate_kobjs[h - hstates],
1390 &hstate_attr_group);
1391 if (retval)
1392 kobject_put(hstate_kobjs[h - hstates]);
1393
1394 return retval;
1395}
1396
1397static void __init hugetlb_sysfs_init(void)
1398{
1399 struct hstate *h;
1400 int err;
1401
1402 hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
1403 if (!hugepages_kobj)
1404 return;
1405
1406 for_each_hstate(h) {
1407 err = hugetlb_sysfs_add_hstate(h);
1408 if (err)
1409 printk(KERN_ERR "Hugetlb: Unable to add hstate %s",
1410 h->name);
1411 }
1412}
1413
1414static void __exit hugetlb_exit(void)
1415{
1416 struct hstate *h;
1417
1418 for_each_hstate(h) {
1419 kobject_put(hstate_kobjs[h - hstates]);
1420 }
1421
1422 kobject_put(hugepages_kobj);
1423}
1424module_exit(hugetlb_exit);
1425
1426static int __init hugetlb_init(void)
1427{
Benjamin Herrenschmidt0ef89d22008-07-31 00:07:30 -07001428 /* Some platform decide whether they support huge pages at boot
1429 * time. On these, such as powerpc, HPAGE_SHIFT is set to 0 when
1430 * there is no such support
1431 */
1432 if (HPAGE_SHIFT == 0)
1433 return 0;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001434
Nick Piggine11bfbf2008-07-23 21:27:52 -07001435 if (!size_to_hstate(default_hstate_size)) {
1436 default_hstate_size = HPAGE_SIZE;
1437 if (!size_to_hstate(default_hstate_size))
1438 hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001439 }
Nick Piggine11bfbf2008-07-23 21:27:52 -07001440 default_hstate_idx = size_to_hstate(default_hstate_size) - hstates;
1441 if (default_hstate_max_huge_pages)
1442 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001443
1444 hugetlb_init_hstates();
1445
Andi Kleenaa888a72008-07-23 21:27:47 -07001446 gather_bootmem_prealloc();
1447
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001448 report_hugepages();
1449
1450 hugetlb_sysfs_init();
1451
1452 return 0;
1453}
1454module_init(hugetlb_init);
1455
1456/* Should be called on processing a hugepagesz=... option */
1457void __init hugetlb_add_hstate(unsigned order)
1458{
1459 struct hstate *h;
Andi Kleen8faa8b02008-07-23 21:27:48 -07001460 unsigned long i;
1461
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001462 if (size_to_hstate(PAGE_SIZE << order)) {
1463 printk(KERN_WARNING "hugepagesz= specified twice, ignoring\n");
1464 return;
1465 }
1466 BUG_ON(max_hstate >= HUGE_MAX_HSTATE);
1467 BUG_ON(order == 0);
1468 h = &hstates[max_hstate++];
1469 h->order = order;
1470 h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
Andi Kleen8faa8b02008-07-23 21:27:48 -07001471 h->nr_huge_pages = 0;
1472 h->free_huge_pages = 0;
1473 for (i = 0; i < MAX_NUMNODES; ++i)
1474 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001475 h->next_nid_to_alloc = first_node(node_online_map);
1476 h->next_nid_to_free = first_node(node_online_map);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001477 snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
1478 huge_page_size(h)/1024);
Andi Kleen8faa8b02008-07-23 21:27:48 -07001479
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001480 parsed_hstate = h;
1481}
1482
Nick Piggine11bfbf2008-07-23 21:27:52 -07001483static int __init hugetlb_nrpages_setup(char *s)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001484{
1485 unsigned long *mhp;
Andi Kleen8faa8b02008-07-23 21:27:48 -07001486 static unsigned long *last_mhp;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001487
1488 /*
1489 * !max_hstate means we haven't parsed a hugepagesz= parameter yet,
1490 * so this hugepages= parameter goes to the "default hstate".
1491 */
1492 if (!max_hstate)
1493 mhp = &default_hstate_max_huge_pages;
1494 else
1495 mhp = &parsed_hstate->max_huge_pages;
1496
Andi Kleen8faa8b02008-07-23 21:27:48 -07001497 if (mhp == last_mhp) {
1498 printk(KERN_WARNING "hugepages= specified twice without "
1499 "interleaving hugepagesz=, ignoring\n");
1500 return 1;
1501 }
1502
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001503 if (sscanf(s, "%lu", mhp) <= 0)
1504 *mhp = 0;
1505
Andi Kleen8faa8b02008-07-23 21:27:48 -07001506 /*
1507 * Global state is always initialized later in hugetlb_init.
1508 * But we need to allocate >= MAX_ORDER hstates here early to still
1509 * use the bootmem allocator.
1510 */
1511 if (max_hstate && parsed_hstate->order >= MAX_ORDER)
1512 hugetlb_hstate_alloc_pages(parsed_hstate);
1513
1514 last_mhp = mhp;
1515
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001516 return 1;
1517}
Nick Piggine11bfbf2008-07-23 21:27:52 -07001518__setup("hugepages=", hugetlb_nrpages_setup);
1519
1520static int __init hugetlb_default_setup(char *s)
1521{
1522 default_hstate_size = memparse(s, &s);
1523 return 1;
1524}
1525__setup("default_hugepagesz=", hugetlb_default_setup);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001526
Nishanth Aravamudan8a213462008-07-25 19:44:37 -07001527static unsigned int cpuset_mems_nr(unsigned int *array)
1528{
1529 int node;
1530 unsigned int nr = 0;
1531
1532 for_each_node_mask(node, cpuset_current_mems_allowed)
1533 nr += array[node];
1534
1535 return nr;
1536}
1537
1538#ifdef CONFIG_SYSCTL
Linus Torvalds1da177e2005-04-16 15:20:36 -07001539int hugetlb_sysctl_handler(struct ctl_table *table, int write,
1540 struct file *file, void __user *buffer,
1541 size_t *length, loff_t *ppos)
1542{
Andi Kleene5ff2152008-07-23 21:27:42 -07001543 struct hstate *h = &default_hstate;
1544 unsigned long tmp;
1545
1546 if (!write)
1547 tmp = h->max_huge_pages;
1548
1549 table->data = &tmp;
1550 table->maxlen = sizeof(unsigned long);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001551 proc_doulongvec_minmax(table, write, file, buffer, length, ppos);
Andi Kleene5ff2152008-07-23 21:27:42 -07001552
1553 if (write)
1554 h->max_huge_pages = set_max_huge_pages(h, tmp);
1555
Linus Torvalds1da177e2005-04-16 15:20:36 -07001556 return 0;
1557}
Mel Gorman396faf02007-07-17 04:03:13 -07001558
1559int hugetlb_treat_movable_handler(struct ctl_table *table, int write,
1560 struct file *file, void __user *buffer,
1561 size_t *length, loff_t *ppos)
1562{
1563 proc_dointvec(table, write, file, buffer, length, ppos);
1564 if (hugepages_treat_as_movable)
1565 htlb_alloc_mask = GFP_HIGHUSER_MOVABLE;
1566 else
1567 htlb_alloc_mask = GFP_HIGHUSER;
1568 return 0;
1569}
1570
Nishanth Aravamudana3d0c6a2008-02-08 04:18:18 -08001571int hugetlb_overcommit_handler(struct ctl_table *table, int write,
1572 struct file *file, void __user *buffer,
1573 size_t *length, loff_t *ppos)
1574{
Andi Kleena5516432008-07-23 21:27:41 -07001575 struct hstate *h = &default_hstate;
Andi Kleene5ff2152008-07-23 21:27:42 -07001576 unsigned long tmp;
1577
1578 if (!write)
1579 tmp = h->nr_overcommit_huge_pages;
1580
1581 table->data = &tmp;
1582 table->maxlen = sizeof(unsigned long);
Nishanth Aravamudana3d0c6a2008-02-08 04:18:18 -08001583 proc_doulongvec_minmax(table, write, file, buffer, length, ppos);
Andi Kleene5ff2152008-07-23 21:27:42 -07001584
1585 if (write) {
1586 spin_lock(&hugetlb_lock);
1587 h->nr_overcommit_huge_pages = tmp;
1588 spin_unlock(&hugetlb_lock);
1589 }
1590
Nishanth Aravamudana3d0c6a2008-02-08 04:18:18 -08001591 return 0;
1592}
1593
Linus Torvalds1da177e2005-04-16 15:20:36 -07001594#endif /* CONFIG_SYSCTL */
1595
Alexey Dobriyane1759c22008-10-15 23:50:22 +04001596void hugetlb_report_meminfo(struct seq_file *m)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001597{
Andi Kleena5516432008-07-23 21:27:41 -07001598 struct hstate *h = &default_hstate;
Alexey Dobriyane1759c22008-10-15 23:50:22 +04001599 seq_printf(m,
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001600 "HugePages_Total: %5lu\n"
1601 "HugePages_Free: %5lu\n"
1602 "HugePages_Rsvd: %5lu\n"
1603 "HugePages_Surp: %5lu\n"
1604 "Hugepagesize: %8lu kB\n",
Andi Kleena5516432008-07-23 21:27:41 -07001605 h->nr_huge_pages,
1606 h->free_huge_pages,
1607 h->resv_huge_pages,
1608 h->surplus_huge_pages,
1609 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001610}
1611
1612int hugetlb_report_node_meminfo(int nid, char *buf)
1613{
Andi Kleena5516432008-07-23 21:27:41 -07001614 struct hstate *h = &default_hstate;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001615 return sprintf(buf,
1616 "Node %d HugePages_Total: %5u\n"
Nishanth Aravamudana1de0912008-03-26 14:37:53 -07001617 "Node %d HugePages_Free: %5u\n"
1618 "Node %d HugePages_Surp: %5u\n",
Andi Kleena5516432008-07-23 21:27:41 -07001619 nid, h->nr_huge_pages_node[nid],
1620 nid, h->free_huge_pages_node[nid],
1621 nid, h->surplus_huge_pages_node[nid]);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001622}
1623
Linus Torvalds1da177e2005-04-16 15:20:36 -07001624/* Return the number pages of memory we physically have, in PAGE_SIZE units. */
1625unsigned long hugetlb_total_pages(void)
1626{
Andi Kleena5516432008-07-23 21:27:41 -07001627 struct hstate *h = &default_hstate;
1628 return h->nr_huge_pages * pages_per_huge_page(h);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001629}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001630
Andi Kleena5516432008-07-23 21:27:41 -07001631static int hugetlb_acct_memory(struct hstate *h, long delta)
Mel Gormanfc1b8a72008-07-23 21:27:22 -07001632{
1633 int ret = -ENOMEM;
1634
1635 spin_lock(&hugetlb_lock);
1636 /*
1637 * When cpuset is configured, it breaks the strict hugetlb page
1638 * reservation as the accounting is done on a global variable. Such
1639 * reservation is completely rubbish in the presence of cpuset because
1640 * the reservation is not checked against page availability for the
1641 * current cpuset. Application can still potentially OOM'ed by kernel
1642 * with lack of free htlb page in cpuset that the task is in.
1643 * Attempt to enforce strict accounting with cpuset is almost
1644 * impossible (or too ugly) because cpuset is too fluid that
1645 * task or memory node can be dynamically moved between cpusets.
1646 *
1647 * The change of semantics for shared hugetlb mapping with cpuset is
1648 * undesirable. However, in order to preserve some of the semantics,
1649 * we fall back to check against current free page availability as
1650 * a best attempt and hopefully to minimize the impact of changing
1651 * semantics that cpuset has.
1652 */
1653 if (delta > 0) {
Andi Kleena5516432008-07-23 21:27:41 -07001654 if (gather_surplus_pages(h, delta) < 0)
Mel Gormanfc1b8a72008-07-23 21:27:22 -07001655 goto out;
1656
Andi Kleena5516432008-07-23 21:27:41 -07001657 if (delta > cpuset_mems_nr(h->free_huge_pages_node)) {
1658 return_unused_surplus_pages(h, delta);
Mel Gormanfc1b8a72008-07-23 21:27:22 -07001659 goto out;
1660 }
1661 }
1662
1663 ret = 0;
1664 if (delta < 0)
Andi Kleena5516432008-07-23 21:27:41 -07001665 return_unused_surplus_pages(h, (unsigned long) -delta);
Mel Gormanfc1b8a72008-07-23 21:27:22 -07001666
1667out:
1668 spin_unlock(&hugetlb_lock);
1669 return ret;
1670}
1671
Andy Whitcroft84afd992008-07-23 21:27:32 -07001672static void hugetlb_vm_op_open(struct vm_area_struct *vma)
1673{
1674 struct resv_map *reservations = vma_resv_map(vma);
1675
1676 /*
1677 * This new VMA should share its siblings reservation map if present.
1678 * The VMA will only ever have a valid reservation map pointer where
1679 * it is being copied for another still existing VMA. As that VMA
1680 * has a reference to the reservation map it cannot dissappear until
1681 * after this open call completes. It is therefore safe to take a
1682 * new reference here without additional locking.
1683 */
1684 if (reservations)
1685 kref_get(&reservations->refs);
1686}
1687
Mel Gormana1e78772008-07-23 21:27:23 -07001688static void hugetlb_vm_op_close(struct vm_area_struct *vma)
1689{
Andi Kleena5516432008-07-23 21:27:41 -07001690 struct hstate *h = hstate_vma(vma);
Andy Whitcroft84afd992008-07-23 21:27:32 -07001691 struct resv_map *reservations = vma_resv_map(vma);
1692 unsigned long reserve;
1693 unsigned long start;
1694 unsigned long end;
1695
1696 if (reservations) {
Andi Kleena5516432008-07-23 21:27:41 -07001697 start = vma_hugecache_offset(h, vma, vma->vm_start);
1698 end = vma_hugecache_offset(h, vma, vma->vm_end);
Andy Whitcroft84afd992008-07-23 21:27:32 -07001699
1700 reserve = (end - start) -
1701 region_count(&reservations->regions, start, end);
1702
1703 kref_put(&reservations->refs, resv_map_release);
1704
Adam Litke7251ff782008-07-23 21:27:59 -07001705 if (reserve) {
Andi Kleena5516432008-07-23 21:27:41 -07001706 hugetlb_acct_memory(h, -reserve);
Adam Litke7251ff782008-07-23 21:27:59 -07001707 hugetlb_put_quota(vma->vm_file->f_mapping, reserve);
1708 }
Andy Whitcroft84afd992008-07-23 21:27:32 -07001709 }
Mel Gormana1e78772008-07-23 21:27:23 -07001710}
1711
Linus Torvalds1da177e2005-04-16 15:20:36 -07001712/*
1713 * We cannot handle pagefaults against hugetlb pages at all. They cause
1714 * handle_mm_fault() to try to instantiate regular-sized pages in the
1715 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
1716 * this far.
1717 */
Nick Piggind0217ac2007-07-19 01:47:03 -07001718static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001719{
1720 BUG();
Nick Piggind0217ac2007-07-19 01:47:03 -07001721 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001722}
1723
1724struct vm_operations_struct hugetlb_vm_ops = {
Nick Piggind0217ac2007-07-19 01:47:03 -07001725 .fault = hugetlb_vm_op_fault,
Andy Whitcroft84afd992008-07-23 21:27:32 -07001726 .open = hugetlb_vm_op_open,
Mel Gormana1e78772008-07-23 21:27:23 -07001727 .close = hugetlb_vm_op_close,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001728};
1729
David Gibson1e8f8892006-01-06 00:10:44 -08001730static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
1731 int writable)
David Gibson63551ae2005-06-21 17:14:44 -07001732{
1733 pte_t entry;
1734
David Gibson1e8f8892006-01-06 00:10:44 -08001735 if (writable) {
David Gibson63551ae2005-06-21 17:14:44 -07001736 entry =
1737 pte_mkwrite(pte_mkdirty(mk_pte(page, vma->vm_page_prot)));
1738 } else {
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07001739 entry = huge_pte_wrprotect(mk_pte(page, vma->vm_page_prot));
David Gibson63551ae2005-06-21 17:14:44 -07001740 }
1741 entry = pte_mkyoung(entry);
1742 entry = pte_mkhuge(entry);
1743
1744 return entry;
1745}
1746
David Gibson1e8f8892006-01-06 00:10:44 -08001747static void set_huge_ptep_writable(struct vm_area_struct *vma,
1748 unsigned long address, pte_t *ptep)
1749{
1750 pte_t entry;
1751
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07001752 entry = pte_mkwrite(pte_mkdirty(huge_ptep_get(ptep)));
1753 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) {
Benjamin Herrenschmidt8dab5242007-06-16 10:16:12 -07001754 update_mmu_cache(vma, address, entry);
Benjamin Herrenschmidt8dab5242007-06-16 10:16:12 -07001755 }
David Gibson1e8f8892006-01-06 00:10:44 -08001756}
1757
1758
David Gibson63551ae2005-06-21 17:14:44 -07001759int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
1760 struct vm_area_struct *vma)
1761{
1762 pte_t *src_pte, *dst_pte, entry;
1763 struct page *ptepage;
Hugh Dickins1c598272005-10-19 21:23:43 -07001764 unsigned long addr;
David Gibson1e8f8892006-01-06 00:10:44 -08001765 int cow;
Andi Kleena5516432008-07-23 21:27:41 -07001766 struct hstate *h = hstate_vma(vma);
1767 unsigned long sz = huge_page_size(h);
David Gibson1e8f8892006-01-06 00:10:44 -08001768
1769 cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
David Gibson63551ae2005-06-21 17:14:44 -07001770
Andi Kleena5516432008-07-23 21:27:41 -07001771 for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
Hugh Dickinsc74df322005-10-29 18:16:23 -07001772 src_pte = huge_pte_offset(src, addr);
1773 if (!src_pte)
1774 continue;
Andi Kleena5516432008-07-23 21:27:41 -07001775 dst_pte = huge_pte_alloc(dst, addr, sz);
David Gibson63551ae2005-06-21 17:14:44 -07001776 if (!dst_pte)
1777 goto nomem;
Larry Woodmanc5c99422008-01-24 05:49:25 -08001778
1779 /* If the pagetables are shared don't copy or take references */
1780 if (dst_pte == src_pte)
1781 continue;
1782
Hugh Dickinsc74df322005-10-29 18:16:23 -07001783 spin_lock(&dst->page_table_lock);
Nick Piggin46478752008-06-05 22:45:57 -07001784 spin_lock_nested(&src->page_table_lock, SINGLE_DEPTH_NESTING);
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07001785 if (!huge_pte_none(huge_ptep_get(src_pte))) {
David Gibson1e8f8892006-01-06 00:10:44 -08001786 if (cow)
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07001787 huge_ptep_set_wrprotect(src, addr, src_pte);
1788 entry = huge_ptep_get(src_pte);
Hugh Dickins1c598272005-10-19 21:23:43 -07001789 ptepage = pte_page(entry);
1790 get_page(ptepage);
Hugh Dickins1c598272005-10-19 21:23:43 -07001791 set_huge_pte_at(dst, addr, dst_pte, entry);
1792 }
1793 spin_unlock(&src->page_table_lock);
Hugh Dickinsc74df322005-10-29 18:16:23 -07001794 spin_unlock(&dst->page_table_lock);
David Gibson63551ae2005-06-21 17:14:44 -07001795 }
1796 return 0;
1797
1798nomem:
1799 return -ENOMEM;
1800}
1801
Chen, Kenneth W502717f2006-10-11 01:20:46 -07001802void __unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
Mel Gorman04f2cbe2008-07-23 21:27:25 -07001803 unsigned long end, struct page *ref_page)
David Gibson63551ae2005-06-21 17:14:44 -07001804{
1805 struct mm_struct *mm = vma->vm_mm;
1806 unsigned long address;
David Gibsonc7546f82005-08-05 11:59:35 -07001807 pte_t *ptep;
David Gibson63551ae2005-06-21 17:14:44 -07001808 pte_t pte;
1809 struct page *page;
Chen, Kenneth Wfe1668a2006-10-04 02:15:24 -07001810 struct page *tmp;
Andi Kleena5516432008-07-23 21:27:41 -07001811 struct hstate *h = hstate_vma(vma);
1812 unsigned long sz = huge_page_size(h);
1813
Chen, Kenneth Wc0a499c2006-12-06 20:31:39 -08001814 /*
1815 * A page gathering list, protected by per file i_mmap_lock. The
1816 * lock is used to avoid list corruption from multiple unmapping
1817 * of the same page since we are using page->lru.
1818 */
Chen, Kenneth Wfe1668a2006-10-04 02:15:24 -07001819 LIST_HEAD(page_list);
David Gibson63551ae2005-06-21 17:14:44 -07001820
1821 WARN_ON(!is_vm_hugetlb_page(vma));
Andi Kleena5516432008-07-23 21:27:41 -07001822 BUG_ON(start & ~huge_page_mask(h));
1823 BUG_ON(end & ~huge_page_mask(h));
David Gibson63551ae2005-06-21 17:14:44 -07001824
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001825 mmu_notifier_invalidate_range_start(mm, start, end);
Hugh Dickins508034a2005-10-29 18:16:30 -07001826 spin_lock(&mm->page_table_lock);
Andi Kleena5516432008-07-23 21:27:41 -07001827 for (address = start; address < end; address += sz) {
David Gibsonc7546f82005-08-05 11:59:35 -07001828 ptep = huge_pte_offset(mm, address);
Adam Litke4c887262005-10-29 18:16:46 -07001829 if (!ptep)
David Gibsonc7546f82005-08-05 11:59:35 -07001830 continue;
1831
Chen, Kenneth W39dde652006-12-06 20:32:03 -08001832 if (huge_pmd_unshare(mm, &address, ptep))
1833 continue;
1834
Mel Gorman04f2cbe2008-07-23 21:27:25 -07001835 /*
1836 * If a reference page is supplied, it is because a specific
1837 * page is being unmapped, not a range. Ensure the page we
1838 * are about to unmap is the actual page of interest.
1839 */
1840 if (ref_page) {
1841 pte = huge_ptep_get(ptep);
1842 if (huge_pte_none(pte))
1843 continue;
1844 page = pte_page(pte);
1845 if (page != ref_page)
1846 continue;
1847
1848 /*
1849 * Mark the VMA as having unmapped its page so that
1850 * future faults in this VMA will fail rather than
1851 * looking like data was lost
1852 */
1853 set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
1854 }
1855
David Gibsonc7546f82005-08-05 11:59:35 -07001856 pte = huge_ptep_get_and_clear(mm, address, ptep);
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07001857 if (huge_pte_none(pte))
David Gibson63551ae2005-06-21 17:14:44 -07001858 continue;
David Gibsonc7546f82005-08-05 11:59:35 -07001859
David Gibson63551ae2005-06-21 17:14:44 -07001860 page = pte_page(pte);
Ken Chen6649a382007-02-08 14:20:27 -08001861 if (pte_dirty(pte))
1862 set_page_dirty(page);
Chen, Kenneth Wfe1668a2006-10-04 02:15:24 -07001863 list_add(&page->lru, &page_list);
David Gibson63551ae2005-06-21 17:14:44 -07001864 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001865 spin_unlock(&mm->page_table_lock);
Hugh Dickins508034a2005-10-29 18:16:30 -07001866 flush_tlb_range(vma, start, end);
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001867 mmu_notifier_invalidate_range_end(mm, start, end);
Chen, Kenneth Wfe1668a2006-10-04 02:15:24 -07001868 list_for_each_entry_safe(page, tmp, &page_list, lru) {
1869 list_del(&page->lru);
1870 put_page(page);
1871 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001872}
David Gibson63551ae2005-06-21 17:14:44 -07001873
Chen, Kenneth W502717f2006-10-11 01:20:46 -07001874void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
Mel Gorman04f2cbe2008-07-23 21:27:25 -07001875 unsigned long end, struct page *ref_page)
Chen, Kenneth W502717f2006-10-11 01:20:46 -07001876{
Andi Kleena137e1c2008-07-23 21:27:43 -07001877 spin_lock(&vma->vm_file->f_mapping->i_mmap_lock);
1878 __unmap_hugepage_range(vma, start, end, ref_page);
1879 spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock);
Chen, Kenneth W502717f2006-10-11 01:20:46 -07001880}
1881
Mel Gorman04f2cbe2008-07-23 21:27:25 -07001882/*
1883 * This is called when the original mapper is failing to COW a MAP_PRIVATE
1884 * mappping it owns the reserve page for. The intention is to unmap the page
1885 * from other VMAs and let the children be SIGKILLed if they are faulting the
1886 * same region.
1887 */
Harvey Harrison2a4b3de2008-10-18 20:27:06 -07001888static int unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
1889 struct page *page, unsigned long address)
Mel Gorman04f2cbe2008-07-23 21:27:25 -07001890{
Adam Litke75266742008-11-12 13:24:56 -08001891 struct hstate *h = hstate_vma(vma);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07001892 struct vm_area_struct *iter_vma;
1893 struct address_space *mapping;
1894 struct prio_tree_iter iter;
1895 pgoff_t pgoff;
1896
1897 /*
1898 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
1899 * from page cache lookup which is in HPAGE_SIZE units.
1900 */
Adam Litke75266742008-11-12 13:24:56 -08001901 address = address & huge_page_mask(h);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07001902 pgoff = ((address - vma->vm_start) >> PAGE_SHIFT)
1903 + (vma->vm_pgoff >> PAGE_SHIFT);
1904 mapping = (struct address_space *)page_private(page);
1905
1906 vma_prio_tree_foreach(iter_vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
1907 /* Do not unmap the current VMA */
1908 if (iter_vma == vma)
1909 continue;
1910
1911 /*
1912 * Unmap the page from other VMAs without their own reserves.
1913 * They get marked to be SIGKILLed if they fault in these
1914 * areas. This is because a future no-page fault on this VMA
1915 * could insert a zeroed page instead of the data existing
1916 * from the time of fork. This would look like data corruption
1917 */
1918 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
1919 unmap_hugepage_range(iter_vma,
Adam Litke75266742008-11-12 13:24:56 -08001920 address, address + huge_page_size(h),
Mel Gorman04f2cbe2008-07-23 21:27:25 -07001921 page);
1922 }
1923
1924 return 1;
1925}
1926
David Gibson1e8f8892006-01-06 00:10:44 -08001927static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
Mel Gorman04f2cbe2008-07-23 21:27:25 -07001928 unsigned long address, pte_t *ptep, pte_t pte,
1929 struct page *pagecache_page)
David Gibson1e8f8892006-01-06 00:10:44 -08001930{
Andi Kleena5516432008-07-23 21:27:41 -07001931 struct hstate *h = hstate_vma(vma);
David Gibson1e8f8892006-01-06 00:10:44 -08001932 struct page *old_page, *new_page;
David Gibson79ac6ba2006-03-22 00:08:51 -08001933 int avoidcopy;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07001934 int outside_reserve = 0;
David Gibson1e8f8892006-01-06 00:10:44 -08001935
1936 old_page = pte_page(pte);
1937
Mel Gorman04f2cbe2008-07-23 21:27:25 -07001938retry_avoidcopy:
David Gibson1e8f8892006-01-06 00:10:44 -08001939 /* If no-one else is actually using this page, avoid the copy
1940 * and just make the page writable */
1941 avoidcopy = (page_count(old_page) == 1);
1942 if (avoidcopy) {
1943 set_huge_ptep_writable(vma, address, ptep);
Nick Piggin83c54072007-07-19 01:47:05 -07001944 return 0;
David Gibson1e8f8892006-01-06 00:10:44 -08001945 }
1946
Mel Gorman04f2cbe2008-07-23 21:27:25 -07001947 /*
1948 * If the process that created a MAP_PRIVATE mapping is about to
1949 * perform a COW due to a shared page count, attempt to satisfy
1950 * the allocation without using the existing reserves. The pagecache
1951 * page is used to determine if the reserve at this address was
1952 * consumed or not. If reserves were used, a partial faulted mapping
1953 * at the time of fork() could consume its reserves on COW instead
1954 * of the full address range.
1955 */
Mel Gormanf83a2752009-05-28 14:34:40 -07001956 if (!(vma->vm_flags & VM_MAYSHARE) &&
Mel Gorman04f2cbe2008-07-23 21:27:25 -07001957 is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
1958 old_page != pagecache_page)
1959 outside_reserve = 1;
1960
David Gibson1e8f8892006-01-06 00:10:44 -08001961 page_cache_get(old_page);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07001962 new_page = alloc_huge_page(vma, address, outside_reserve);
David Gibson1e8f8892006-01-06 00:10:44 -08001963
Adam Litke2fc39ce2007-11-14 16:59:39 -08001964 if (IS_ERR(new_page)) {
David Gibson1e8f8892006-01-06 00:10:44 -08001965 page_cache_release(old_page);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07001966
1967 /*
1968 * If a process owning a MAP_PRIVATE mapping fails to COW,
1969 * it is due to references held by a child and an insufficient
1970 * huge page pool. To guarantee the original mappers
1971 * reliability, unmap the page from child processes. The child
1972 * may get SIGKILLed if it later faults.
1973 */
1974 if (outside_reserve) {
1975 BUG_ON(huge_pte_none(pte));
1976 if (unmap_ref_private(mm, vma, old_page, address)) {
1977 BUG_ON(page_count(old_page) != 1);
1978 BUG_ON(huge_pte_none(pte));
1979 goto retry_avoidcopy;
1980 }
1981 WARN_ON_ONCE(1);
1982 }
1983
Adam Litke2fc39ce2007-11-14 16:59:39 -08001984 return -PTR_ERR(new_page);
David Gibson1e8f8892006-01-06 00:10:44 -08001985 }
1986
1987 spin_unlock(&mm->page_table_lock);
Atsushi Nemoto9de455b2006-12-12 17:14:55 +00001988 copy_huge_page(new_page, old_page, address, vma);
Nick Piggin0ed361d2008-02-04 22:29:34 -08001989 __SetPageUptodate(new_page);
David Gibson1e8f8892006-01-06 00:10:44 -08001990 spin_lock(&mm->page_table_lock);
1991
Andi Kleena5516432008-07-23 21:27:41 -07001992 ptep = huge_pte_offset(mm, address & huge_page_mask(h));
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07001993 if (likely(pte_same(huge_ptep_get(ptep), pte))) {
David Gibson1e8f8892006-01-06 00:10:44 -08001994 /* Break COW */
Gerald Schaefer8fe627e2008-04-28 02:13:28 -07001995 huge_ptep_clear_flush(vma, address, ptep);
David Gibson1e8f8892006-01-06 00:10:44 -08001996 set_huge_pte_at(mm, address, ptep,
1997 make_huge_pte(vma, new_page, 1));
1998 /* Make the old page be freed below */
1999 new_page = old_page;
2000 }
2001 page_cache_release(new_page);
2002 page_cache_release(old_page);
Nick Piggin83c54072007-07-19 01:47:05 -07002003 return 0;
David Gibson1e8f8892006-01-06 00:10:44 -08002004}
2005
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002006/* Return the pagecache page at a given address within a VMA */
Andi Kleena5516432008-07-23 21:27:41 -07002007static struct page *hugetlbfs_pagecache_page(struct hstate *h,
2008 struct vm_area_struct *vma, unsigned long address)
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002009{
2010 struct address_space *mapping;
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -07002011 pgoff_t idx;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002012
2013 mapping = vma->vm_file->f_mapping;
Andi Kleena5516432008-07-23 21:27:41 -07002014 idx = vma_hugecache_offset(h, vma, address);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002015
2016 return find_lock_page(mapping, idx);
2017}
2018
Hugh Dickins3ae77f42009-09-21 17:03:33 -07002019/*
2020 * Return whether there is a pagecache page to back given address within VMA.
2021 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
2022 */
2023static bool hugetlbfs_pagecache_present(struct hstate *h,
Hugh Dickins2a15efc2009-09-21 17:03:27 -07002024 struct vm_area_struct *vma, unsigned long address)
2025{
2026 struct address_space *mapping;
2027 pgoff_t idx;
2028 struct page *page;
2029
2030 mapping = vma->vm_file->f_mapping;
2031 idx = vma_hugecache_offset(h, vma, address);
2032
2033 page = find_get_page(mapping, idx);
2034 if (page)
2035 put_page(page);
2036 return page != NULL;
2037}
2038
Robert P. J. Daya1ed3dd2007-07-17 04:03:33 -07002039static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
Hugh Dickins788c7df2009-06-23 13:49:05 +01002040 unsigned long address, pte_t *ptep, unsigned int flags)
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01002041{
Andi Kleena5516432008-07-23 21:27:41 -07002042 struct hstate *h = hstate_vma(vma);
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01002043 int ret = VM_FAULT_SIGBUS;
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -07002044 pgoff_t idx;
Adam Litke4c887262005-10-29 18:16:46 -07002045 unsigned long size;
Adam Litke4c887262005-10-29 18:16:46 -07002046 struct page *page;
2047 struct address_space *mapping;
David Gibson1e8f8892006-01-06 00:10:44 -08002048 pte_t new_pte;
Adam Litke4c887262005-10-29 18:16:46 -07002049
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002050 /*
2051 * Currently, we are forced to kill the process in the event the
2052 * original mapper has unmapped pages from the child due to a failed
2053 * COW. Warn that such a situation has occured as it may not be obvious
2054 */
2055 if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
2056 printk(KERN_WARNING
2057 "PID %d killed due to inadequate hugepage pool\n",
2058 current->pid);
2059 return ret;
2060 }
2061
Adam Litke4c887262005-10-29 18:16:46 -07002062 mapping = vma->vm_file->f_mapping;
Andi Kleena5516432008-07-23 21:27:41 -07002063 idx = vma_hugecache_offset(h, vma, address);
Adam Litke4c887262005-10-29 18:16:46 -07002064
2065 /*
2066 * Use page lock to guard against racing truncation
2067 * before we get page_table_lock.
2068 */
Christoph Lameter6bda6662006-01-06 00:10:49 -08002069retry:
2070 page = find_lock_page(mapping, idx);
2071 if (!page) {
Andi Kleena5516432008-07-23 21:27:41 -07002072 size = i_size_read(mapping->host) >> huge_page_shift(h);
Hugh Dickinsebed4bf2006-10-28 10:38:43 -07002073 if (idx >= size)
2074 goto out;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002075 page = alloc_huge_page(vma, address, 0);
Adam Litke2fc39ce2007-11-14 16:59:39 -08002076 if (IS_ERR(page)) {
2077 ret = -PTR_ERR(page);
Christoph Lameter6bda6662006-01-06 00:10:49 -08002078 goto out;
2079 }
Andi Kleena5516432008-07-23 21:27:41 -07002080 clear_huge_page(page, address, huge_page_size(h));
Nick Piggin0ed361d2008-02-04 22:29:34 -08002081 __SetPageUptodate(page);
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01002082
Mel Gormanf83a2752009-05-28 14:34:40 -07002083 if (vma->vm_flags & VM_MAYSHARE) {
Christoph Lameter6bda6662006-01-06 00:10:49 -08002084 int err;
Ken Chen45c682a2007-11-14 16:59:44 -08002085 struct inode *inode = mapping->host;
Christoph Lameter6bda6662006-01-06 00:10:49 -08002086
2087 err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
2088 if (err) {
2089 put_page(page);
Christoph Lameter6bda6662006-01-06 00:10:49 -08002090 if (err == -EEXIST)
2091 goto retry;
2092 goto out;
2093 }
Ken Chen45c682a2007-11-14 16:59:44 -08002094
2095 spin_lock(&inode->i_lock);
Andi Kleena5516432008-07-23 21:27:41 -07002096 inode->i_blocks += blocks_per_huge_page(h);
Ken Chen45c682a2007-11-14 16:59:44 -08002097 spin_unlock(&inode->i_lock);
Christoph Lameter6bda6662006-01-06 00:10:49 -08002098 } else
2099 lock_page(page);
2100 }
David Gibson1e8f8892006-01-06 00:10:44 -08002101
Andy Whitcroft57303d82008-08-12 15:08:47 -07002102 /*
2103 * If we are going to COW a private mapping later, we examine the
2104 * pending reservations for this page now. This will ensure that
2105 * any allocations necessary to record that reservation occur outside
2106 * the spinlock.
2107 */
Hugh Dickins788c7df2009-06-23 13:49:05 +01002108 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED))
Andy Whitcroft2b267362008-08-12 15:08:49 -07002109 if (vma_needs_reservation(h, vma, address) < 0) {
2110 ret = VM_FAULT_OOM;
2111 goto backout_unlocked;
2112 }
Andy Whitcroft57303d82008-08-12 15:08:47 -07002113
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01002114 spin_lock(&mm->page_table_lock);
Andi Kleena5516432008-07-23 21:27:41 -07002115 size = i_size_read(mapping->host) >> huge_page_shift(h);
Adam Litke4c887262005-10-29 18:16:46 -07002116 if (idx >= size)
2117 goto backout;
2118
Nick Piggin83c54072007-07-19 01:47:05 -07002119 ret = 0;
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07002120 if (!huge_pte_none(huge_ptep_get(ptep)))
Adam Litke4c887262005-10-29 18:16:46 -07002121 goto backout;
2122
David Gibson1e8f8892006-01-06 00:10:44 -08002123 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
2124 && (vma->vm_flags & VM_SHARED)));
2125 set_huge_pte_at(mm, address, ptep, new_pte);
2126
Hugh Dickins788c7df2009-06-23 13:49:05 +01002127 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
David Gibson1e8f8892006-01-06 00:10:44 -08002128 /* Optimization, do the COW without a second fault */
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002129 ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page);
David Gibson1e8f8892006-01-06 00:10:44 -08002130 }
2131
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01002132 spin_unlock(&mm->page_table_lock);
Adam Litke4c887262005-10-29 18:16:46 -07002133 unlock_page(page);
2134out:
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01002135 return ret;
Adam Litke4c887262005-10-29 18:16:46 -07002136
2137backout:
2138 spin_unlock(&mm->page_table_lock);
Andy Whitcroft2b267362008-08-12 15:08:49 -07002139backout_unlocked:
Adam Litke4c887262005-10-29 18:16:46 -07002140 unlock_page(page);
2141 put_page(page);
2142 goto out;
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01002143}
2144
Adam Litke86e52162006-01-06 00:10:43 -08002145int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
Hugh Dickins788c7df2009-06-23 13:49:05 +01002146 unsigned long address, unsigned int flags)
Adam Litke86e52162006-01-06 00:10:43 -08002147{
2148 pte_t *ptep;
2149 pte_t entry;
David Gibson1e8f8892006-01-06 00:10:44 -08002150 int ret;
Andy Whitcroft57303d82008-08-12 15:08:47 -07002151 struct page *pagecache_page = NULL;
David Gibson3935baa2006-03-22 00:08:53 -08002152 static DEFINE_MUTEX(hugetlb_instantiation_mutex);
Andi Kleena5516432008-07-23 21:27:41 -07002153 struct hstate *h = hstate_vma(vma);
Adam Litke86e52162006-01-06 00:10:43 -08002154
Andi Kleena5516432008-07-23 21:27:41 -07002155 ptep = huge_pte_alloc(mm, address, huge_page_size(h));
Adam Litke86e52162006-01-06 00:10:43 -08002156 if (!ptep)
2157 return VM_FAULT_OOM;
2158
David Gibson3935baa2006-03-22 00:08:53 -08002159 /*
2160 * Serialize hugepage allocation and instantiation, so that we don't
2161 * get spurious allocation failures if two CPUs race to instantiate
2162 * the same page in the page cache.
2163 */
2164 mutex_lock(&hugetlb_instantiation_mutex);
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07002165 entry = huge_ptep_get(ptep);
2166 if (huge_pte_none(entry)) {
Hugh Dickins788c7df2009-06-23 13:49:05 +01002167 ret = hugetlb_no_page(mm, vma, address, ptep, flags);
David Gibsonb4d1d992008-10-15 22:01:11 -07002168 goto out_mutex;
David Gibson3935baa2006-03-22 00:08:53 -08002169 }
Adam Litke86e52162006-01-06 00:10:43 -08002170
Nick Piggin83c54072007-07-19 01:47:05 -07002171 ret = 0;
David Gibson1e8f8892006-01-06 00:10:44 -08002172
Andy Whitcroft57303d82008-08-12 15:08:47 -07002173 /*
2174 * If we are going to COW the mapping later, we examine the pending
2175 * reservations for this page now. This will ensure that any
2176 * allocations necessary to record that reservation occur outside the
2177 * spinlock. For private mappings, we also lookup the pagecache
2178 * page now as it is used to determine if a reservation has been
2179 * consumed.
2180 */
Hugh Dickins788c7df2009-06-23 13:49:05 +01002181 if ((flags & FAULT_FLAG_WRITE) && !pte_write(entry)) {
Andy Whitcroft2b267362008-08-12 15:08:49 -07002182 if (vma_needs_reservation(h, vma, address) < 0) {
2183 ret = VM_FAULT_OOM;
David Gibsonb4d1d992008-10-15 22:01:11 -07002184 goto out_mutex;
Andy Whitcroft2b267362008-08-12 15:08:49 -07002185 }
Andy Whitcroft57303d82008-08-12 15:08:47 -07002186
Mel Gormanf83a2752009-05-28 14:34:40 -07002187 if (!(vma->vm_flags & VM_MAYSHARE))
Andy Whitcroft57303d82008-08-12 15:08:47 -07002188 pagecache_page = hugetlbfs_pagecache_page(h,
2189 vma, address);
2190 }
2191
David Gibson1e8f8892006-01-06 00:10:44 -08002192 spin_lock(&mm->page_table_lock);
2193 /* Check for a racing update before calling hugetlb_cow */
David Gibsonb4d1d992008-10-15 22:01:11 -07002194 if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
2195 goto out_page_table_lock;
2196
2197
Hugh Dickins788c7df2009-06-23 13:49:05 +01002198 if (flags & FAULT_FLAG_WRITE) {
David Gibsonb4d1d992008-10-15 22:01:11 -07002199 if (!pte_write(entry)) {
Andy Whitcroft57303d82008-08-12 15:08:47 -07002200 ret = hugetlb_cow(mm, vma, address, ptep, entry,
2201 pagecache_page);
David Gibsonb4d1d992008-10-15 22:01:11 -07002202 goto out_page_table_lock;
2203 }
2204 entry = pte_mkdirty(entry);
2205 }
2206 entry = pte_mkyoung(entry);
Hugh Dickins788c7df2009-06-23 13:49:05 +01002207 if (huge_ptep_set_access_flags(vma, address, ptep, entry,
2208 flags & FAULT_FLAG_WRITE))
David Gibsonb4d1d992008-10-15 22:01:11 -07002209 update_mmu_cache(vma, address, entry);
2210
2211out_page_table_lock:
David Gibson1e8f8892006-01-06 00:10:44 -08002212 spin_unlock(&mm->page_table_lock);
Andy Whitcroft57303d82008-08-12 15:08:47 -07002213
2214 if (pagecache_page) {
2215 unlock_page(pagecache_page);
2216 put_page(pagecache_page);
2217 }
2218
David Gibsonb4d1d992008-10-15 22:01:11 -07002219out_mutex:
David Gibson3935baa2006-03-22 00:08:53 -08002220 mutex_unlock(&hugetlb_instantiation_mutex);
David Gibson1e8f8892006-01-06 00:10:44 -08002221
2222 return ret;
Adam Litke86e52162006-01-06 00:10:43 -08002223}
2224
Andi Kleenceb86872008-07-23 21:27:50 -07002225/* Can be overriden by architectures */
2226__attribute__((weak)) struct page *
2227follow_huge_pud(struct mm_struct *mm, unsigned long address,
2228 pud_t *pud, int write)
2229{
2230 BUG();
2231 return NULL;
2232}
2233
David Gibson63551ae2005-06-21 17:14:44 -07002234int follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
2235 struct page **pages, struct vm_area_struct **vmas,
Adam Litke5b23dbe2007-11-14 16:59:33 -08002236 unsigned long *position, int *length, int i,
Hugh Dickins2a15efc2009-09-21 17:03:27 -07002237 unsigned int flags)
David Gibson63551ae2005-06-21 17:14:44 -07002238{
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08002239 unsigned long pfn_offset;
2240 unsigned long vaddr = *position;
David Gibson63551ae2005-06-21 17:14:44 -07002241 int remainder = *length;
Andi Kleena5516432008-07-23 21:27:41 -07002242 struct hstate *h = hstate_vma(vma);
David Gibson63551ae2005-06-21 17:14:44 -07002243
Hugh Dickins1c598272005-10-19 21:23:43 -07002244 spin_lock(&mm->page_table_lock);
David Gibson63551ae2005-06-21 17:14:44 -07002245 while (vaddr < vma->vm_end && remainder) {
Adam Litke4c887262005-10-29 18:16:46 -07002246 pte_t *pte;
Hugh Dickins2a15efc2009-09-21 17:03:27 -07002247 int absent;
Adam Litke4c887262005-10-29 18:16:46 -07002248 struct page *page;
2249
2250 /*
2251 * Some archs (sparc64, sh*) have multiple pte_ts to
Hugh Dickins2a15efc2009-09-21 17:03:27 -07002252 * each hugepage. We have to make sure we get the
Adam Litke4c887262005-10-29 18:16:46 -07002253 * first, for the page indexing below to work.
2254 */
Andi Kleena5516432008-07-23 21:27:41 -07002255 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h));
Hugh Dickins2a15efc2009-09-21 17:03:27 -07002256 absent = !pte || huge_pte_none(huge_ptep_get(pte));
Adam Litke4c887262005-10-29 18:16:46 -07002257
Hugh Dickins2a15efc2009-09-21 17:03:27 -07002258 /*
2259 * When coredumping, it suits get_dump_page if we just return
Hugh Dickins3ae77f42009-09-21 17:03:33 -07002260 * an error where there's an empty slot with no huge pagecache
2261 * to back it. This way, we avoid allocating a hugepage, and
2262 * the sparse dumpfile avoids allocating disk blocks, but its
2263 * huge holes still show up with zeroes where they need to be.
Hugh Dickins2a15efc2009-09-21 17:03:27 -07002264 */
Hugh Dickins3ae77f42009-09-21 17:03:33 -07002265 if (absent && (flags & FOLL_DUMP) &&
2266 !hugetlbfs_pagecache_present(h, vma, vaddr)) {
Hugh Dickins2a15efc2009-09-21 17:03:27 -07002267 remainder = 0;
2268 break;
2269 }
2270
2271 if (absent ||
2272 ((flags & FOLL_WRITE) && !pte_write(huge_ptep_get(pte)))) {
Adam Litke4c887262005-10-29 18:16:46 -07002273 int ret;
2274
2275 spin_unlock(&mm->page_table_lock);
Hugh Dickins2a15efc2009-09-21 17:03:27 -07002276 ret = hugetlb_fault(mm, vma, vaddr,
2277 (flags & FOLL_WRITE) ? FAULT_FLAG_WRITE : 0);
Adam Litke4c887262005-10-29 18:16:46 -07002278 spin_lock(&mm->page_table_lock);
Adam Litkea89182c2007-08-22 14:01:51 -07002279 if (!(ret & VM_FAULT_ERROR))
Adam Litke4c887262005-10-29 18:16:46 -07002280 continue;
2281
2282 remainder = 0;
Adam Litke4c887262005-10-29 18:16:46 -07002283 break;
2284 }
David Gibson63551ae2005-06-21 17:14:44 -07002285
Andi Kleena5516432008-07-23 21:27:41 -07002286 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07002287 page = pte_page(huge_ptep_get(pte));
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08002288same_page:
Chen, Kenneth Wd6692182006-03-31 02:29:57 -08002289 if (pages) {
Hugh Dickins2a15efc2009-09-21 17:03:27 -07002290 pages[i] = mem_map_offset(page, pfn_offset);
KOSAKI Motohiro4b2e38a2008-10-18 20:27:10 -07002291 get_page(pages[i]);
Chen, Kenneth Wd6692182006-03-31 02:29:57 -08002292 }
David Gibson63551ae2005-06-21 17:14:44 -07002293
2294 if (vmas)
2295 vmas[i] = vma;
2296
2297 vaddr += PAGE_SIZE;
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08002298 ++pfn_offset;
David Gibson63551ae2005-06-21 17:14:44 -07002299 --remainder;
2300 ++i;
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08002301 if (vaddr < vma->vm_end && remainder &&
Andi Kleena5516432008-07-23 21:27:41 -07002302 pfn_offset < pages_per_huge_page(h)) {
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08002303 /*
2304 * We use pfn_offset to avoid touching the pageframes
2305 * of this compound page.
2306 */
2307 goto same_page;
2308 }
David Gibson63551ae2005-06-21 17:14:44 -07002309 }
Hugh Dickins1c598272005-10-19 21:23:43 -07002310 spin_unlock(&mm->page_table_lock);
David Gibson63551ae2005-06-21 17:14:44 -07002311 *length = remainder;
2312 *position = vaddr;
2313
Hugh Dickins2a15efc2009-09-21 17:03:27 -07002314 return i ? i : -EFAULT;
David Gibson63551ae2005-06-21 17:14:44 -07002315}
Zhang, Yanmin8f860592006-03-22 00:08:50 -08002316
2317void hugetlb_change_protection(struct vm_area_struct *vma,
2318 unsigned long address, unsigned long end, pgprot_t newprot)
2319{
2320 struct mm_struct *mm = vma->vm_mm;
2321 unsigned long start = address;
2322 pte_t *ptep;
2323 pte_t pte;
Andi Kleena5516432008-07-23 21:27:41 -07002324 struct hstate *h = hstate_vma(vma);
Zhang, Yanmin8f860592006-03-22 00:08:50 -08002325
2326 BUG_ON(address >= end);
2327 flush_cache_range(vma, address, end);
2328
Chen, Kenneth W39dde652006-12-06 20:32:03 -08002329 spin_lock(&vma->vm_file->f_mapping->i_mmap_lock);
Zhang, Yanmin8f860592006-03-22 00:08:50 -08002330 spin_lock(&mm->page_table_lock);
Andi Kleena5516432008-07-23 21:27:41 -07002331 for (; address < end; address += huge_page_size(h)) {
Zhang, Yanmin8f860592006-03-22 00:08:50 -08002332 ptep = huge_pte_offset(mm, address);
2333 if (!ptep)
2334 continue;
Chen, Kenneth W39dde652006-12-06 20:32:03 -08002335 if (huge_pmd_unshare(mm, &address, ptep))
2336 continue;
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07002337 if (!huge_pte_none(huge_ptep_get(ptep))) {
Zhang, Yanmin8f860592006-03-22 00:08:50 -08002338 pte = huge_ptep_get_and_clear(mm, address, ptep);
2339 pte = pte_mkhuge(pte_modify(pte, newprot));
2340 set_huge_pte_at(mm, address, ptep, pte);
Zhang, Yanmin8f860592006-03-22 00:08:50 -08002341 }
2342 }
2343 spin_unlock(&mm->page_table_lock);
Chen, Kenneth W39dde652006-12-06 20:32:03 -08002344 spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock);
Zhang, Yanmin8f860592006-03-22 00:08:50 -08002345
2346 flush_tlb_range(vma, start, end);
2347}
2348
Mel Gormana1e78772008-07-23 21:27:23 -07002349int hugetlb_reserve_pages(struct inode *inode,
2350 long from, long to,
Mel Gorman5a6fe122009-02-10 14:02:27 +00002351 struct vm_area_struct *vma,
2352 int acctflag)
Adam Litkee4e574b2007-10-16 01:26:19 -07002353{
Mel Gorman17c9d122009-02-11 16:34:16 +00002354 long ret, chg;
Andi Kleena5516432008-07-23 21:27:41 -07002355 struct hstate *h = hstate_inode(inode);
Adam Litkee4e574b2007-10-16 01:26:19 -07002356
Mel Gormana1e78772008-07-23 21:27:23 -07002357 /*
Mel Gorman17c9d122009-02-11 16:34:16 +00002358 * Only apply hugepage reservation if asked. At fault time, an
2359 * attempt will be made for VM_NORESERVE to allocate a page
2360 * and filesystem quota without using reserves
2361 */
2362 if (acctflag & VM_NORESERVE)
2363 return 0;
2364
2365 /*
Mel Gormana1e78772008-07-23 21:27:23 -07002366 * Shared mappings base their reservation on the number of pages that
2367 * are already allocated on behalf of the file. Private mappings need
2368 * to reserve the full area even if read-only as mprotect() may be
2369 * called to make the mapping read-write. Assume !vma is a shm mapping
2370 */
Mel Gormanf83a2752009-05-28 14:34:40 -07002371 if (!vma || vma->vm_flags & VM_MAYSHARE)
Mel Gormana1e78772008-07-23 21:27:23 -07002372 chg = region_chg(&inode->i_mapping->private_list, from, to);
Mel Gorman5a6fe122009-02-10 14:02:27 +00002373 else {
2374 struct resv_map *resv_map = resv_map_alloc();
Mel Gorman5a6fe122009-02-10 14:02:27 +00002375 if (!resv_map)
2376 return -ENOMEM;
2377
Mel Gorman17c9d122009-02-11 16:34:16 +00002378 chg = to - from;
2379
Mel Gorman5a6fe122009-02-10 14:02:27 +00002380 set_vma_resv_map(vma, resv_map);
2381 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
2382 }
2383
Mel Gorman17c9d122009-02-11 16:34:16 +00002384 if (chg < 0)
2385 return chg;
2386
2387 /* There must be enough filesystem quota for the mapping */
2388 if (hugetlb_get_quota(inode->i_mapping, chg))
2389 return -ENOSPC;
2390
2391 /*
2392 * Check enough hugepages are available for the reservation.
2393 * Hand back the quota if there are not
2394 */
2395 ret = hugetlb_acct_memory(h, chg);
2396 if (ret < 0) {
2397 hugetlb_put_quota(inode->i_mapping, chg);
2398 return ret;
2399 }
2400
2401 /*
2402 * Account for the reservations made. Shared mappings record regions
2403 * that have reservations as they are shared by multiple VMAs.
2404 * When the last VMA disappears, the region map says how much
2405 * the reservation was and the page cache tells how much of
2406 * the reservation was consumed. Private mappings are per-VMA and
2407 * only the consumed reservations are tracked. When the VMA
2408 * disappears, the original reservation is the VMA size and the
2409 * consumed reservations are stored in the map. Hence, nothing
2410 * else has to be done for private mappings here
2411 */
Mel Gormanf83a2752009-05-28 14:34:40 -07002412 if (!vma || vma->vm_flags & VM_MAYSHARE)
Mel Gorman17c9d122009-02-11 16:34:16 +00002413 region_add(&inode->i_mapping->private_list, from, to);
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -07002414 return 0;
2415}
2416
2417void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed)
2418{
Andi Kleena5516432008-07-23 21:27:41 -07002419 struct hstate *h = hstate_inode(inode);
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -07002420 long chg = region_truncate(&inode->i_mapping->private_list, offset);
Ken Chen45c682a2007-11-14 16:59:44 -08002421
2422 spin_lock(&inode->i_lock);
Eric Sandeene4c6f8b2009-07-29 15:02:16 -07002423 inode->i_blocks -= (blocks_per_huge_page(h) * freed);
Ken Chen45c682a2007-11-14 16:59:44 -08002424 spin_unlock(&inode->i_lock);
2425
Adam Litke90d8b7e2007-11-14 16:59:42 -08002426 hugetlb_put_quota(inode->i_mapping, (chg - freed));
Andi Kleena5516432008-07-23 21:27:41 -07002427 hugetlb_acct_memory(h, -(chg - freed));
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -07002428}