blob: 4494976c2042fb3321764a9d313531318affb277 [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
2 * Generic hugetlb support.
Nadia Yvette Chambers6d49e352012-12-06 10:39:54 +01003 * (C) Nadia Yvette Chambers, April 2004
Linus Torvalds1da177e2005-04-16 15:20:36 -07004 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07005#include <linux/list.h>
6#include <linux/init.h>
7#include <linux/module.h>
8#include <linux/mm.h>
Alexey Dobriyane1759c22008-10-15 23:50:22 +04009#include <linux/seq_file.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070010#include <linux/sysctl.h>
11#include <linux/highmem.h>
Andrea Arcangelicddb8a52008-07-28 15:46:29 -070012#include <linux/mmu_notifier.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070013#include <linux/nodemask.h>
David Gibson63551ae2005-06-21 17:14:44 -070014#include <linux/pagemap.h>
Christoph Lameter5da7ca82006-01-06 00:10:46 -080015#include <linux/mempolicy.h>
Gideon Israel Dsouza3b321232014-04-07 15:37:26 -070016#include <linux/compiler.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>
Tejun Heo5a0e3ad2010-03-24 17:04:11 +090021#include <linux/slab.h>
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +090022#include <linux/rmap.h>
Naoya Horiguchifd6a03e2010-05-28 09:29:21 +090023#include <linux/swap.h>
24#include <linux/swapops.h>
Naoya Horiguchic8721bb2013-09-11 14:22:09 -070025#include <linux/page-isolation.h>
Davidlohr Bueso8382d912014-04-03 14:47:31 -070026#include <linux/jhash.h>
Linus Torvaldsd6606682008-08-06 12:04:54 -070027
David Gibson63551ae2005-06-21 17:14:44 -070028#include <asm/page.h>
29#include <asm/pgtable.h>
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -070030#include <asm/tlb.h>
David Gibson63551ae2005-06-21 17:14:44 -070031
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -070032#include <linux/io.h>
David Gibson63551ae2005-06-21 17:14:44 -070033#include <linux/hugetlb.h>
Aneesh Kumar K.V9dd540e2012-07-31 16:42:15 -070034#include <linux/hugetlb_cgroup.h>
Lee Schermerhorn9a3052302009-12-14 17:58:25 -080035#include <linux/node.h>
Nick Piggin7835e982006-03-22 00:08:40 -080036#include "internal.h"
Linus Torvalds1da177e2005-04-16 15:20:36 -070037
Andrey Ryabinin753162c2015-02-10 14:11:36 -080038int hugepages_treat_as_movable;
Andi Kleena5516432008-07-23 21:27:41 -070039
Aneesh Kumar K.Vc3f38a32012-07-31 16:42:10 -070040int hugetlb_max_hstate __read_mostly;
Andi Kleene5ff2152008-07-23 21:27:42 -070041unsigned int default_hstate_idx;
42struct hstate hstates[HUGE_MAX_HSTATE];
43
Jon Tollefson53ba51d2008-07-23 21:27:52 -070044__initdata LIST_HEAD(huge_boot_pages);
45
Andi Kleene5ff2152008-07-23 21:27:42 -070046/* for command line parsing */
47static struct hstate * __initdata parsed_hstate;
48static unsigned long __initdata default_hstate_max_huge_pages;
Nick Piggine11bfbf2008-07-23 21:27:52 -070049static unsigned long __initdata default_hstate_size;
Andi Kleene5ff2152008-07-23 21:27:42 -070050
David Gibson3935baa2006-03-22 00:08:53 -080051/*
Naoya Horiguchi31caf662013-09-11 14:21:59 -070052 * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages,
53 * free_huge_pages, and surplus_huge_pages.
David Gibson3935baa2006-03-22 00:08:53 -080054 */
Aneesh Kumar K.Vc3f38a32012-07-31 16:42:10 -070055DEFINE_SPINLOCK(hugetlb_lock);
Eric Paris0bd0f9f2005-11-21 21:32:28 -080056
Davidlohr Bueso8382d912014-04-03 14:47:31 -070057/*
58 * Serializes faults on the same logical page. This is used to
59 * prevent spurious OOMs when the hugepage pool is fully utilized.
60 */
61static int num_fault_mutexes;
62static struct mutex *htlb_fault_mutex_table ____cacheline_aligned_in_smp;
63
David Gibson90481622012-03-21 16:34:12 -070064static inline void unlock_or_release_subpool(struct hugepage_subpool *spool)
65{
66 bool free = (spool->count == 0) && (spool->used_hpages == 0);
67
68 spin_unlock(&spool->lock);
69
70 /* If no pages are used, and no other handles to the subpool
71 * remain, free the subpool the subpool remain */
72 if (free)
73 kfree(spool);
74}
75
76struct hugepage_subpool *hugepage_new_subpool(long nr_blocks)
77{
78 struct hugepage_subpool *spool;
79
Mike Kravetzc6a91822015-04-15 16:13:36 -070080 spool = kzalloc(sizeof(*spool), GFP_KERNEL);
David Gibson90481622012-03-21 16:34:12 -070081 if (!spool)
82 return NULL;
83
84 spin_lock_init(&spool->lock);
85 spool->count = 1;
86 spool->max_hpages = nr_blocks;
David Gibson90481622012-03-21 16:34:12 -070087
88 return spool;
89}
90
91void hugepage_put_subpool(struct hugepage_subpool *spool)
92{
93 spin_lock(&spool->lock);
94 BUG_ON(!spool->count);
95 spool->count--;
96 unlock_or_release_subpool(spool);
97}
98
99static int hugepage_subpool_get_pages(struct hugepage_subpool *spool,
100 long delta)
101{
102 int ret = 0;
103
104 if (!spool)
105 return 0;
106
107 spin_lock(&spool->lock);
108 if ((spool->used_hpages + delta) <= spool->max_hpages) {
109 spool->used_hpages += delta;
110 } else {
111 ret = -ENOMEM;
112 }
113 spin_unlock(&spool->lock);
114
115 return ret;
116}
117
118static void hugepage_subpool_put_pages(struct hugepage_subpool *spool,
119 long delta)
120{
121 if (!spool)
122 return;
123
124 spin_lock(&spool->lock);
125 spool->used_hpages -= delta;
126 /* If hugetlbfs_put_super couldn't free spool due to
127 * an outstanding quota reference, free it now. */
128 unlock_or_release_subpool(spool);
129}
130
131static inline struct hugepage_subpool *subpool_inode(struct inode *inode)
132{
133 return HUGETLBFS_SB(inode->i_sb)->spool;
134}
135
136static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma)
137{
Al Viro496ad9a2013-01-23 17:07:38 -0500138 return subpool_inode(file_inode(vma->vm_file));
David Gibson90481622012-03-21 16:34:12 -0700139}
140
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700141/*
Andy Whitcroft96822902008-07-23 21:27:29 -0700142 * Region tracking -- allows tracking of reservations and instantiated pages
143 * across the pages in a mapping.
Andy Whitcroft84afd992008-07-23 21:27:32 -0700144 *
Davidlohr Bueso7b24d8612014-04-03 14:47:27 -0700145 * The region data structures are embedded into a resv_map and
146 * protected by a resv_map's lock
Andy Whitcroft96822902008-07-23 21:27:29 -0700147 */
148struct file_region {
149 struct list_head link;
150 long from;
151 long to;
152};
153
Joonsoo Kim1406ec92014-04-03 14:47:26 -0700154static long region_add(struct resv_map *resv, long f, long t)
Andy Whitcroft96822902008-07-23 21:27:29 -0700155{
Joonsoo Kim1406ec92014-04-03 14:47:26 -0700156 struct list_head *head = &resv->regions;
Andy Whitcroft96822902008-07-23 21:27:29 -0700157 struct file_region *rg, *nrg, *trg;
158
Davidlohr Bueso7b24d8612014-04-03 14:47:27 -0700159 spin_lock(&resv->lock);
Andy Whitcroft96822902008-07-23 21:27:29 -0700160 /* Locate the region we are either in or before. */
161 list_for_each_entry(rg, head, link)
162 if (f <= rg->to)
163 break;
164
165 /* Round our left edge to the current segment if it encloses us. */
166 if (f > rg->from)
167 f = rg->from;
168
169 /* Check for and consume any regions we now overlap with. */
170 nrg = rg;
171 list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
172 if (&rg->link == head)
173 break;
174 if (rg->from > t)
175 break;
176
177 /* If this area reaches higher then extend our area to
178 * include it completely. If this is not the first area
179 * which we intend to reuse, free it. */
180 if (rg->to > t)
181 t = rg->to;
182 if (rg != nrg) {
183 list_del(&rg->link);
184 kfree(rg);
185 }
186 }
187 nrg->from = f;
188 nrg->to = t;
Davidlohr Bueso7b24d8612014-04-03 14:47:27 -0700189 spin_unlock(&resv->lock);
Andy Whitcroft96822902008-07-23 21:27:29 -0700190 return 0;
191}
192
Joonsoo Kim1406ec92014-04-03 14:47:26 -0700193static long region_chg(struct resv_map *resv, long f, long t)
Andy Whitcroft96822902008-07-23 21:27:29 -0700194{
Joonsoo Kim1406ec92014-04-03 14:47:26 -0700195 struct list_head *head = &resv->regions;
Davidlohr Bueso7b24d8612014-04-03 14:47:27 -0700196 struct file_region *rg, *nrg = NULL;
Andy Whitcroft96822902008-07-23 21:27:29 -0700197 long chg = 0;
198
Davidlohr Bueso7b24d8612014-04-03 14:47:27 -0700199retry:
200 spin_lock(&resv->lock);
Andy Whitcroft96822902008-07-23 21:27:29 -0700201 /* Locate the region we are before or in. */
202 list_for_each_entry(rg, head, link)
203 if (f <= rg->to)
204 break;
205
206 /* If we are below the current region then a new region is required.
207 * Subtle, allocate a new region at the position but make it zero
208 * size such that we can guarantee to record the reservation. */
209 if (&rg->link == head || t < rg->from) {
Davidlohr Bueso7b24d8612014-04-03 14:47:27 -0700210 if (!nrg) {
211 spin_unlock(&resv->lock);
212 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
213 if (!nrg)
214 return -ENOMEM;
Andy Whitcroft96822902008-07-23 21:27:29 -0700215
Davidlohr Bueso7b24d8612014-04-03 14:47:27 -0700216 nrg->from = f;
217 nrg->to = f;
218 INIT_LIST_HEAD(&nrg->link);
219 goto retry;
220 }
221
222 list_add(&nrg->link, rg->link.prev);
223 chg = t - f;
224 goto out_nrg;
Andy Whitcroft96822902008-07-23 21:27:29 -0700225 }
226
227 /* Round our left edge to the current segment if it encloses us. */
228 if (f > rg->from)
229 f = rg->from;
230 chg = t - f;
231
232 /* Check for and consume any regions we now overlap with. */
233 list_for_each_entry(rg, rg->link.prev, link) {
234 if (&rg->link == head)
235 break;
236 if (rg->from > t)
Davidlohr Bueso7b24d8612014-04-03 14:47:27 -0700237 goto out;
Andy Whitcroft96822902008-07-23 21:27:29 -0700238
Lucas De Marchi25985ed2011-03-30 22:57:33 -0300239 /* We overlap with this area, if it extends further than
Andy Whitcroft96822902008-07-23 21:27:29 -0700240 * us then we must extend ourselves. Account for its
241 * existing reservation. */
242 if (rg->to > t) {
243 chg += rg->to - t;
244 t = rg->to;
245 }
246 chg -= rg->to - rg->from;
247 }
Davidlohr Bueso7b24d8612014-04-03 14:47:27 -0700248
249out:
250 spin_unlock(&resv->lock);
251 /* We already know we raced and no longer need the new region */
252 kfree(nrg);
253 return chg;
254out_nrg:
255 spin_unlock(&resv->lock);
Andy Whitcroft96822902008-07-23 21:27:29 -0700256 return chg;
257}
258
Joonsoo Kim1406ec92014-04-03 14:47:26 -0700259static long region_truncate(struct resv_map *resv, long end)
Andy Whitcroft96822902008-07-23 21:27:29 -0700260{
Joonsoo Kim1406ec92014-04-03 14:47:26 -0700261 struct list_head *head = &resv->regions;
Andy Whitcroft96822902008-07-23 21:27:29 -0700262 struct file_region *rg, *trg;
263 long chg = 0;
264
Davidlohr Bueso7b24d8612014-04-03 14:47:27 -0700265 spin_lock(&resv->lock);
Andy Whitcroft96822902008-07-23 21:27:29 -0700266 /* Locate the region we are either in or before. */
267 list_for_each_entry(rg, head, link)
268 if (end <= rg->to)
269 break;
270 if (&rg->link == head)
Davidlohr Bueso7b24d8612014-04-03 14:47:27 -0700271 goto out;
Andy Whitcroft96822902008-07-23 21:27:29 -0700272
273 /* If we are in the middle of a region then adjust it. */
274 if (end > rg->from) {
275 chg = rg->to - end;
276 rg->to = end;
277 rg = list_entry(rg->link.next, typeof(*rg), link);
278 }
279
280 /* Drop any remaining regions. */
281 list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
282 if (&rg->link == head)
283 break;
284 chg += rg->to - rg->from;
285 list_del(&rg->link);
286 kfree(rg);
287 }
Davidlohr Bueso7b24d8612014-04-03 14:47:27 -0700288
289out:
290 spin_unlock(&resv->lock);
Andy Whitcroft96822902008-07-23 21:27:29 -0700291 return chg;
292}
293
Joonsoo Kim1406ec92014-04-03 14:47:26 -0700294static long region_count(struct resv_map *resv, long f, long t)
Andy Whitcroft84afd992008-07-23 21:27:32 -0700295{
Joonsoo Kim1406ec92014-04-03 14:47:26 -0700296 struct list_head *head = &resv->regions;
Andy Whitcroft84afd992008-07-23 21:27:32 -0700297 struct file_region *rg;
298 long chg = 0;
299
Davidlohr Bueso7b24d8612014-04-03 14:47:27 -0700300 spin_lock(&resv->lock);
Andy Whitcroft84afd992008-07-23 21:27:32 -0700301 /* Locate each segment we overlap with, and count that overlap. */
302 list_for_each_entry(rg, head, link) {
Wang Sheng-Huif2135a42012-05-29 15:06:17 -0700303 long seg_from;
304 long seg_to;
Andy Whitcroft84afd992008-07-23 21:27:32 -0700305
306 if (rg->to <= f)
307 continue;
308 if (rg->from >= t)
309 break;
310
311 seg_from = max(rg->from, f);
312 seg_to = min(rg->to, t);
313
314 chg += seg_to - seg_from;
315 }
Davidlohr Bueso7b24d8612014-04-03 14:47:27 -0700316 spin_unlock(&resv->lock);
Andy Whitcroft84afd992008-07-23 21:27:32 -0700317
318 return chg;
319}
320
Andy Whitcroft96822902008-07-23 21:27:29 -0700321/*
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700322 * Convert the address within this vma to the page offset within
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700323 * the mapping, in pagecache page units; huge pages here.
324 */
Andi Kleena5516432008-07-23 21:27:41 -0700325static pgoff_t vma_hugecache_offset(struct hstate *h,
326 struct vm_area_struct *vma, unsigned long address)
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700327{
Andi Kleena5516432008-07-23 21:27:41 -0700328 return ((address - vma->vm_start) >> huge_page_shift(h)) +
329 (vma->vm_pgoff >> huge_page_order(h));
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700330}
331
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +0900332pgoff_t linear_hugepage_index(struct vm_area_struct *vma,
333 unsigned long address)
334{
335 return vma_hugecache_offset(hstate_vma(vma), vma, address);
336}
337
Andy Whitcroft84afd992008-07-23 21:27:32 -0700338/*
Mel Gorman08fba692009-01-06 14:38:53 -0800339 * Return the size of the pages allocated when backing a VMA. In the majority
340 * cases this will be same size as used by the page table entries.
341 */
342unsigned long vma_kernel_pagesize(struct vm_area_struct *vma)
343{
344 struct hstate *hstate;
345
346 if (!is_vm_hugetlb_page(vma))
347 return PAGE_SIZE;
348
349 hstate = hstate_vma(vma);
350
Wanpeng Li2415cf12013-07-03 15:02:43 -0700351 return 1UL << huge_page_shift(hstate);
Mel Gorman08fba692009-01-06 14:38:53 -0800352}
Joerg Roedelf340ca02009-06-19 15:16:22 +0200353EXPORT_SYMBOL_GPL(vma_kernel_pagesize);
Mel Gorman08fba692009-01-06 14:38:53 -0800354
355/*
Mel Gorman33402892009-01-06 14:38:54 -0800356 * Return the page size being used by the MMU to back a VMA. In the majority
357 * of cases, the page size used by the kernel matches the MMU size. On
358 * architectures where it differs, an architecture-specific version of this
359 * function is required.
360 */
361#ifndef vma_mmu_pagesize
362unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
363{
364 return vma_kernel_pagesize(vma);
365}
366#endif
367
368/*
Andy Whitcroft84afd992008-07-23 21:27:32 -0700369 * Flags for MAP_PRIVATE reservations. These are stored in the bottom
370 * bits of the reservation map pointer, which are always clear due to
371 * alignment.
372 */
373#define HPAGE_RESV_OWNER (1UL << 0)
374#define HPAGE_RESV_UNMAPPED (1UL << 1)
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700375#define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
Andy Whitcroft84afd992008-07-23 21:27:32 -0700376
Mel Gormana1e78772008-07-23 21:27:23 -0700377/*
378 * These helpers are used to track how many pages are reserved for
379 * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
380 * is guaranteed to have their future faults succeed.
381 *
382 * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
383 * the reserve counters are updated with the hugetlb_lock held. It is safe
384 * to reset the VMA at fork() time as it is not in use yet and there is no
385 * chance of the global counters getting corrupted as a result of the values.
Andy Whitcroft84afd992008-07-23 21:27:32 -0700386 *
387 * The private mapping reservation is represented in a subtly different
388 * manner to a shared mapping. A shared mapping has a region map associated
389 * with the underlying file, this region map represents the backing file
390 * pages which have ever had a reservation assigned which this persists even
391 * after the page is instantiated. A private mapping has a region map
392 * associated with the original mmap which is attached to all VMAs which
393 * reference it, this region map represents those offsets which have consumed
394 * reservation ie. where pages have been instantiated.
Mel Gormana1e78772008-07-23 21:27:23 -0700395 */
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700396static unsigned long get_vma_private_data(struct vm_area_struct *vma)
397{
398 return (unsigned long)vma->vm_private_data;
399}
400
401static void set_vma_private_data(struct vm_area_struct *vma,
402 unsigned long value)
403{
404 vma->vm_private_data = (void *)value;
405}
406
Joonsoo Kim9119a412014-04-03 14:47:25 -0700407struct resv_map *resv_map_alloc(void)
Andy Whitcroft84afd992008-07-23 21:27:32 -0700408{
409 struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
410 if (!resv_map)
411 return NULL;
412
413 kref_init(&resv_map->refs);
Davidlohr Bueso7b24d8612014-04-03 14:47:27 -0700414 spin_lock_init(&resv_map->lock);
Andy Whitcroft84afd992008-07-23 21:27:32 -0700415 INIT_LIST_HEAD(&resv_map->regions);
416
417 return resv_map;
418}
419
Joonsoo Kim9119a412014-04-03 14:47:25 -0700420void resv_map_release(struct kref *ref)
Andy Whitcroft84afd992008-07-23 21:27:32 -0700421{
422 struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
423
424 /* Clear out any active regions before we release the map. */
Joonsoo Kim1406ec92014-04-03 14:47:26 -0700425 region_truncate(resv_map, 0);
Andy Whitcroft84afd992008-07-23 21:27:32 -0700426 kfree(resv_map);
427}
428
Joonsoo Kim4e35f482014-04-03 14:47:30 -0700429static inline struct resv_map *inode_resv_map(struct inode *inode)
430{
431 return inode->i_mapping->private_data;
432}
433
Andy Whitcroft84afd992008-07-23 21:27:32 -0700434static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
Mel Gormana1e78772008-07-23 21:27:23 -0700435{
Sasha Levin81d1b092014-10-09 15:28:10 -0700436 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
Joonsoo Kim4e35f482014-04-03 14:47:30 -0700437 if (vma->vm_flags & VM_MAYSHARE) {
438 struct address_space *mapping = vma->vm_file->f_mapping;
439 struct inode *inode = mapping->host;
440
441 return inode_resv_map(inode);
442
443 } else {
Andy Whitcroft84afd992008-07-23 21:27:32 -0700444 return (struct resv_map *)(get_vma_private_data(vma) &
445 ~HPAGE_RESV_MASK);
Joonsoo Kim4e35f482014-04-03 14:47:30 -0700446 }
Mel Gormana1e78772008-07-23 21:27:23 -0700447}
448
Andy Whitcroft84afd992008-07-23 21:27:32 -0700449static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
Mel Gormana1e78772008-07-23 21:27:23 -0700450{
Sasha Levin81d1b092014-10-09 15:28:10 -0700451 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
452 VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
Mel Gormana1e78772008-07-23 21:27:23 -0700453
Andy Whitcroft84afd992008-07-23 21:27:32 -0700454 set_vma_private_data(vma, (get_vma_private_data(vma) &
455 HPAGE_RESV_MASK) | (unsigned long)map);
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700456}
457
458static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
459{
Sasha Levin81d1b092014-10-09 15:28:10 -0700460 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
461 VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700462
463 set_vma_private_data(vma, get_vma_private_data(vma) | flags);
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700464}
465
466static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
467{
Sasha Levin81d1b092014-10-09 15:28:10 -0700468 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700469
470 return (get_vma_private_data(vma) & flag) != 0;
Mel Gormana1e78772008-07-23 21:27:23 -0700471}
472
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700473/* Reset counters to 0 and clear all HPAGE_RESV_* flags */
Mel Gormana1e78772008-07-23 21:27:23 -0700474void reset_vma_resv_huge_pages(struct vm_area_struct *vma)
475{
Sasha Levin81d1b092014-10-09 15:28:10 -0700476 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
Mel Gormanf83a2752009-05-28 14:34:40 -0700477 if (!(vma->vm_flags & VM_MAYSHARE))
Mel Gormana1e78772008-07-23 21:27:23 -0700478 vma->vm_private_data = (void *)0;
479}
480
481/* Returns true if the VMA has associated reserve pages */
Joonsoo Kimaf0ed732013-09-11 14:21:18 -0700482static int vma_has_reserves(struct vm_area_struct *vma, long chg)
Mel Gormana1e78772008-07-23 21:27:23 -0700483{
Joonsoo Kimaf0ed732013-09-11 14:21:18 -0700484 if (vma->vm_flags & VM_NORESERVE) {
485 /*
486 * This address is already reserved by other process(chg == 0),
487 * so, we should decrement reserved count. Without decrementing,
488 * reserve count remains after releasing inode, because this
489 * allocated page will go into page cache and is regarded as
490 * coming from reserved pool in releasing step. Currently, we
491 * don't have any other solution to deal with this situation
492 * properly, so add work-around here.
493 */
494 if (vma->vm_flags & VM_MAYSHARE && chg == 0)
495 return 1;
496 else
497 return 0;
498 }
Joonsoo Kima63884e2013-09-11 14:21:07 -0700499
500 /* Shared mappings always use reserves */
Mel Gormanf83a2752009-05-28 14:34:40 -0700501 if (vma->vm_flags & VM_MAYSHARE)
Mel Gorman7f09ca52008-07-23 21:27:58 -0700502 return 1;
Joonsoo Kima63884e2013-09-11 14:21:07 -0700503
504 /*
505 * Only the process that called mmap() has reserves for
506 * private mappings.
507 */
Mel Gorman7f09ca52008-07-23 21:27:58 -0700508 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER))
509 return 1;
Joonsoo Kima63884e2013-09-11 14:21:07 -0700510
Mel Gorman7f09ca52008-07-23 21:27:58 -0700511 return 0;
Mel Gormana1e78772008-07-23 21:27:23 -0700512}
513
Andi Kleena5516432008-07-23 21:27:41 -0700514static void enqueue_huge_page(struct hstate *h, struct page *page)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700515{
516 int nid = page_to_nid(page);
Aneesh Kumar K.V0edaecf2012-07-31 16:42:07 -0700517 list_move(&page->lru, &h->hugepage_freelists[nid]);
Andi Kleena5516432008-07-23 21:27:41 -0700518 h->free_huge_pages++;
519 h->free_huge_pages_node[nid]++;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700520}
521
Naoya Horiguchibf50bab2010-09-08 10:19:33 +0900522static struct page *dequeue_huge_page_node(struct hstate *h, int nid)
523{
524 struct page *page;
525
Naoya Horiguchic8721bb2013-09-11 14:22:09 -0700526 list_for_each_entry(page, &h->hugepage_freelists[nid], lru)
527 if (!is_migrate_isolate_page(page))
528 break;
529 /*
530 * if 'non-isolated free hugepage' not found on the list,
531 * the allocation fails.
532 */
533 if (&h->hugepage_freelists[nid] == &page->lru)
Naoya Horiguchibf50bab2010-09-08 10:19:33 +0900534 return NULL;
Aneesh Kumar K.V0edaecf2012-07-31 16:42:07 -0700535 list_move(&page->lru, &h->hugepage_activelist);
Naoya Horiguchia9869b82010-09-08 10:19:37 +0900536 set_page_refcounted(page);
Naoya Horiguchibf50bab2010-09-08 10:19:33 +0900537 h->free_huge_pages--;
538 h->free_huge_pages_node[nid]--;
539 return page;
540}
541
Naoya Horiguchi86cdb462013-09-11 14:22:13 -0700542/* Movability of hugepages depends on migration support. */
543static inline gfp_t htlb_alloc_mask(struct hstate *h)
544{
Naoya Horiguchi100873d2014-06-04 16:10:56 -0700545 if (hugepages_treat_as_movable || hugepage_migration_supported(h))
Naoya Horiguchi86cdb462013-09-11 14:22:13 -0700546 return GFP_HIGHUSER_MOVABLE;
547 else
548 return GFP_HIGHUSER;
549}
550
Andi Kleena5516432008-07-23 21:27:41 -0700551static struct page *dequeue_huge_page_vma(struct hstate *h,
552 struct vm_area_struct *vma,
Joonsoo Kimaf0ed732013-09-11 14:21:18 -0700553 unsigned long address, int avoid_reserve,
554 long chg)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700555{
Konstantin Khlebnikovb1c12cb2012-04-25 16:01:46 -0700556 struct page *page = NULL;
Lee Schermerhorn480eccf2007-09-18 22:46:47 -0700557 struct mempolicy *mpol;
Mel Gorman19770b32008-04-28 02:12:18 -0700558 nodemask_t *nodemask;
Miao Xiec0ff7452010-05-24 14:32:08 -0700559 struct zonelist *zonelist;
Mel Gormandd1a2392008-04-28 02:12:17 -0700560 struct zone *zone;
561 struct zoneref *z;
Mel Gormancc9a6c82012-03-21 16:34:11 -0700562 unsigned int cpuset_mems_cookie;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700563
Mel Gormana1e78772008-07-23 21:27:23 -0700564 /*
565 * A child process with MAP_PRIVATE mappings created by their parent
566 * have no page reserves. This check ensures that reservations are
567 * not "stolen". The child may still get SIGKILLed
568 */
Joonsoo Kimaf0ed732013-09-11 14:21:18 -0700569 if (!vma_has_reserves(vma, chg) &&
Andi Kleena5516432008-07-23 21:27:41 -0700570 h->free_huge_pages - h->resv_huge_pages == 0)
Miao Xiec0ff7452010-05-24 14:32:08 -0700571 goto err;
Mel Gormana1e78772008-07-23 21:27:23 -0700572
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700573 /* If reserves cannot be used, ensure enough pages are in the pool */
Andi Kleena5516432008-07-23 21:27:41 -0700574 if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
Justin P. Mattock6eab04a2011-04-08 19:49:08 -0700575 goto err;
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700576
Joonsoo Kim9966c4bb2013-09-11 14:20:50 -0700577retry_cpuset:
Mel Gormand26914d2014-04-03 14:47:24 -0700578 cpuset_mems_cookie = read_mems_allowed_begin();
Joonsoo Kim9966c4bb2013-09-11 14:20:50 -0700579 zonelist = huge_zonelist(vma, address,
Naoya Horiguchi86cdb462013-09-11 14:22:13 -0700580 htlb_alloc_mask(h), &mpol, &nodemask);
Joonsoo Kim9966c4bb2013-09-11 14:20:50 -0700581
Mel Gorman19770b32008-04-28 02:12:18 -0700582 for_each_zone_zonelist_nodemask(zone, z, zonelist,
583 MAX_NR_ZONES - 1, nodemask) {
Vladimir Davydov344736f2014-10-20 15:50:30 +0400584 if (cpuset_zone_allowed(zone, htlb_alloc_mask(h))) {
Naoya Horiguchibf50bab2010-09-08 10:19:33 +0900585 page = dequeue_huge_page_node(h, zone_to_nid(zone));
586 if (page) {
Joonsoo Kimaf0ed732013-09-11 14:21:18 -0700587 if (avoid_reserve)
588 break;
589 if (!vma_has_reserves(vma, chg))
590 break;
591
Joonsoo Kim07443a82013-09-11 14:21:58 -0700592 SetPagePrivate(page);
Joonsoo Kimaf0ed732013-09-11 14:21:18 -0700593 h->resv_huge_pages--;
Naoya Horiguchibf50bab2010-09-08 10:19:33 +0900594 break;
595 }
Andrew Morton3abf7af2007-07-19 01:49:08 -0700596 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700597 }
Mel Gormancc9a6c82012-03-21 16:34:11 -0700598
599 mpol_cond_put(mpol);
Mel Gormand26914d2014-04-03 14:47:24 -0700600 if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie)))
Mel Gormancc9a6c82012-03-21 16:34:11 -0700601 goto retry_cpuset;
602 return page;
603
Miao Xiec0ff7452010-05-24 14:32:08 -0700604err:
Mel Gormancc9a6c82012-03-21 16:34:11 -0700605 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700606}
607
Luiz Capitulino1cac6f22014-06-04 16:07:11 -0700608/*
609 * common helper functions for hstate_next_node_to_{alloc|free}.
610 * We may have allocated or freed a huge page based on a different
611 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
612 * be outside of *nodes_allowed. Ensure that we use an allowed
613 * node for alloc or free.
614 */
615static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
616{
617 nid = next_node(nid, *nodes_allowed);
618 if (nid == MAX_NUMNODES)
619 nid = first_node(*nodes_allowed);
620 VM_BUG_ON(nid >= MAX_NUMNODES);
621
622 return nid;
623}
624
625static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
626{
627 if (!node_isset(nid, *nodes_allowed))
628 nid = next_node_allowed(nid, nodes_allowed);
629 return nid;
630}
631
632/*
633 * returns the previously saved node ["this node"] from which to
634 * allocate a persistent huge page for the pool and advance the
635 * next node from which to allocate, handling wrap at end of node
636 * mask.
637 */
638static int hstate_next_node_to_alloc(struct hstate *h,
639 nodemask_t *nodes_allowed)
640{
641 int nid;
642
643 VM_BUG_ON(!nodes_allowed);
644
645 nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
646 h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
647
648 return nid;
649}
650
651/*
652 * helper for free_pool_huge_page() - return the previously saved
653 * node ["this node"] from which to free a huge page. Advance the
654 * next node id whether or not we find a free huge page to free so
655 * that the next attempt to free addresses the next node.
656 */
657static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
658{
659 int nid;
660
661 VM_BUG_ON(!nodes_allowed);
662
663 nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
664 h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
665
666 return nid;
667}
668
669#define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \
670 for (nr_nodes = nodes_weight(*mask); \
671 nr_nodes > 0 && \
672 ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \
673 nr_nodes--)
674
675#define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \
676 for (nr_nodes = nodes_weight(*mask); \
677 nr_nodes > 0 && \
678 ((node = hstate_next_node_to_free(hs, mask)) || 1); \
679 nr_nodes--)
680
Luiz Capitulino944d9fe2014-06-04 16:07:13 -0700681#if defined(CONFIG_CMA) && defined(CONFIG_X86_64)
682static void destroy_compound_gigantic_page(struct page *page,
683 unsigned long order)
684{
685 int i;
686 int nr_pages = 1 << order;
687 struct page *p = page + 1;
688
689 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
690 __ClearPageTail(p);
691 set_page_refcounted(p);
692 p->first_page = NULL;
693 }
694
695 set_compound_order(page, 0);
696 __ClearPageHead(page);
697}
698
699static void free_gigantic_page(struct page *page, unsigned order)
700{
701 free_contig_range(page_to_pfn(page), 1 << order);
702}
703
704static int __alloc_gigantic_page(unsigned long start_pfn,
705 unsigned long nr_pages)
706{
707 unsigned long end_pfn = start_pfn + nr_pages;
708 return alloc_contig_range(start_pfn, end_pfn, MIGRATE_MOVABLE);
709}
710
711static bool pfn_range_valid_gigantic(unsigned long start_pfn,
712 unsigned long nr_pages)
713{
714 unsigned long i, end_pfn = start_pfn + nr_pages;
715 struct page *page;
716
717 for (i = start_pfn; i < end_pfn; i++) {
718 if (!pfn_valid(i))
719 return false;
720
721 page = pfn_to_page(i);
722
723 if (PageReserved(page))
724 return false;
725
726 if (page_count(page) > 0)
727 return false;
728
729 if (PageHuge(page))
730 return false;
731 }
732
733 return true;
734}
735
736static bool zone_spans_last_pfn(const struct zone *zone,
737 unsigned long start_pfn, unsigned long nr_pages)
738{
739 unsigned long last_pfn = start_pfn + nr_pages - 1;
740 return zone_spans_pfn(zone, last_pfn);
741}
742
743static struct page *alloc_gigantic_page(int nid, unsigned order)
744{
745 unsigned long nr_pages = 1 << order;
746 unsigned long ret, pfn, flags;
747 struct zone *z;
748
749 z = NODE_DATA(nid)->node_zones;
750 for (; z - NODE_DATA(nid)->node_zones < MAX_NR_ZONES; z++) {
751 spin_lock_irqsave(&z->lock, flags);
752
753 pfn = ALIGN(z->zone_start_pfn, nr_pages);
754 while (zone_spans_last_pfn(z, pfn, nr_pages)) {
755 if (pfn_range_valid_gigantic(pfn, nr_pages)) {
756 /*
757 * We release the zone lock here because
758 * alloc_contig_range() will also lock the zone
759 * at some point. If there's an allocation
760 * spinning on this lock, it may win the race
761 * and cause alloc_contig_range() to fail...
762 */
763 spin_unlock_irqrestore(&z->lock, flags);
764 ret = __alloc_gigantic_page(pfn, nr_pages);
765 if (!ret)
766 return pfn_to_page(pfn);
767 spin_lock_irqsave(&z->lock, flags);
768 }
769 pfn += nr_pages;
770 }
771
772 spin_unlock_irqrestore(&z->lock, flags);
773 }
774
775 return NULL;
776}
777
778static void prep_new_huge_page(struct hstate *h, struct page *page, int nid);
779static void prep_compound_gigantic_page(struct page *page, unsigned long order);
780
781static struct page *alloc_fresh_gigantic_page_node(struct hstate *h, int nid)
782{
783 struct page *page;
784
785 page = alloc_gigantic_page(nid, huge_page_order(h));
786 if (page) {
787 prep_compound_gigantic_page(page, huge_page_order(h));
788 prep_new_huge_page(h, page, nid);
789 }
790
791 return page;
792}
793
794static int alloc_fresh_gigantic_page(struct hstate *h,
795 nodemask_t *nodes_allowed)
796{
797 struct page *page = NULL;
798 int nr_nodes, node;
799
800 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
801 page = alloc_fresh_gigantic_page_node(h, node);
802 if (page)
803 return 1;
804 }
805
806 return 0;
807}
808
809static inline bool gigantic_page_supported(void) { return true; }
810#else
811static inline bool gigantic_page_supported(void) { return false; }
812static inline void free_gigantic_page(struct page *page, unsigned order) { }
813static inline void destroy_compound_gigantic_page(struct page *page,
814 unsigned long order) { }
815static inline int alloc_fresh_gigantic_page(struct hstate *h,
816 nodemask_t *nodes_allowed) { return 0; }
817#endif
818
Andi Kleena5516432008-07-23 21:27:41 -0700819static void update_and_free_page(struct hstate *h, struct page *page)
Adam Litke6af2acb2007-10-16 01:26:16 -0700820{
821 int i;
Andi Kleena5516432008-07-23 21:27:41 -0700822
Luiz Capitulino944d9fe2014-06-04 16:07:13 -0700823 if (hstate_is_gigantic(h) && !gigantic_page_supported())
824 return;
Andy Whitcroft18229df2008-11-06 12:53:27 -0800825
Andi Kleena5516432008-07-23 21:27:41 -0700826 h->nr_huge_pages--;
827 h->nr_huge_pages_node[page_to_nid(page)]--;
828 for (i = 0; i < pages_per_huge_page(h); i++) {
Chris Forbes32f84522011-07-25 17:12:14 -0700829 page[i].flags &= ~(1 << PG_locked | 1 << PG_error |
830 1 << PG_referenced | 1 << PG_dirty |
Luiz Capitulinoa7407a22014-06-04 16:07:09 -0700831 1 << PG_active | 1 << PG_private |
832 1 << PG_writeback);
Adam Litke6af2acb2007-10-16 01:26:16 -0700833 }
Sasha Levin309381fea2014-01-23 15:52:54 -0800834 VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page);
Adam Litke6af2acb2007-10-16 01:26:16 -0700835 set_compound_page_dtor(page, NULL);
836 set_page_refcounted(page);
Luiz Capitulino944d9fe2014-06-04 16:07:13 -0700837 if (hstate_is_gigantic(h)) {
838 destroy_compound_gigantic_page(page, huge_page_order(h));
839 free_gigantic_page(page, huge_page_order(h));
840 } else {
841 arch_release_hugepage(page);
842 __free_pages(page, huge_page_order(h));
843 }
Adam Litke6af2acb2007-10-16 01:26:16 -0700844}
845
Andi Kleene5ff2152008-07-23 21:27:42 -0700846struct hstate *size_to_hstate(unsigned long size)
847{
848 struct hstate *h;
849
850 for_each_hstate(h) {
851 if (huge_page_size(h) == size)
852 return h;
853 }
854 return NULL;
855}
856
Atsushi Kumagai8f1d26d2014-07-30 16:08:39 -0700857void free_huge_page(struct page *page)
David Gibson27a85ef2006-03-22 00:08:56 -0800858{
Andi Kleena5516432008-07-23 21:27:41 -0700859 /*
860 * Can't pass hstate in here because it is called from the
861 * compound page destructor.
862 */
Andi Kleene5ff2152008-07-23 21:27:42 -0700863 struct hstate *h = page_hstate(page);
Adam Litke7893d1d2007-10-16 01:26:18 -0700864 int nid = page_to_nid(page);
David Gibson90481622012-03-21 16:34:12 -0700865 struct hugepage_subpool *spool =
866 (struct hugepage_subpool *)page_private(page);
Joonsoo Kim07443a82013-09-11 14:21:58 -0700867 bool restore_reserve;
David Gibson27a85ef2006-03-22 00:08:56 -0800868
Andy Whitcrofte5df70a2008-02-23 15:23:32 -0800869 set_page_private(page, 0);
Mel Gorman23be7462010-04-23 13:17:56 -0400870 page->mapping = NULL;
Adam Litke7893d1d2007-10-16 01:26:18 -0700871 BUG_ON(page_count(page));
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +0900872 BUG_ON(page_mapcount(page));
Joonsoo Kim07443a82013-09-11 14:21:58 -0700873 restore_reserve = PagePrivate(page);
Joonsoo Kim16c794b2013-10-16 13:46:48 -0700874 ClearPagePrivate(page);
David Gibson27a85ef2006-03-22 00:08:56 -0800875
876 spin_lock(&hugetlb_lock);
Aneesh Kumar K.V6d76dcf2012-07-31 16:42:18 -0700877 hugetlb_cgroup_uncharge_page(hstate_index(h),
878 pages_per_huge_page(h), page);
Joonsoo Kim07443a82013-09-11 14:21:58 -0700879 if (restore_reserve)
880 h->resv_huge_pages++;
881
Luiz Capitulino944d9fe2014-06-04 16:07:13 -0700882 if (h->surplus_huge_pages_node[nid]) {
Aneesh Kumar K.V0edaecf2012-07-31 16:42:07 -0700883 /* remove the page from active list */
884 list_del(&page->lru);
Andi Kleena5516432008-07-23 21:27:41 -0700885 update_and_free_page(h, page);
886 h->surplus_huge_pages--;
887 h->surplus_huge_pages_node[nid]--;
Adam Litke7893d1d2007-10-16 01:26:18 -0700888 } else {
Will Deacon5d3a5512012-10-08 16:29:32 -0700889 arch_clear_hugepage_flags(page);
Andi Kleena5516432008-07-23 21:27:41 -0700890 enqueue_huge_page(h, page);
Adam Litke7893d1d2007-10-16 01:26:18 -0700891 }
David Gibson27a85ef2006-03-22 00:08:56 -0800892 spin_unlock(&hugetlb_lock);
David Gibson90481622012-03-21 16:34:12 -0700893 hugepage_subpool_put_pages(spool, 1);
David Gibson27a85ef2006-03-22 00:08:56 -0800894}
895
Andi Kleena5516432008-07-23 21:27:41 -0700896static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
Andi Kleenb7ba30c2008-07-23 21:27:40 -0700897{
Aneesh Kumar K.V0edaecf2012-07-31 16:42:07 -0700898 INIT_LIST_HEAD(&page->lru);
Andi Kleenb7ba30c2008-07-23 21:27:40 -0700899 set_compound_page_dtor(page, free_huge_page);
900 spin_lock(&hugetlb_lock);
Aneesh Kumar K.V9dd540e2012-07-31 16:42:15 -0700901 set_hugetlb_cgroup(page, NULL);
Andi Kleena5516432008-07-23 21:27:41 -0700902 h->nr_huge_pages++;
903 h->nr_huge_pages_node[nid]++;
Andi Kleenb7ba30c2008-07-23 21:27:40 -0700904 spin_unlock(&hugetlb_lock);
905 put_page(page); /* free it into the hugepage allocator */
906}
907
Luiz Capitulino2906dd52014-06-04 16:07:06 -0700908static void prep_compound_gigantic_page(struct page *page, unsigned long order)
Wu Fengguang20a03072009-06-16 15:32:22 -0700909{
910 int i;
911 int nr_pages = 1 << order;
912 struct page *p = page + 1;
913
914 /* we rely on prep_new_huge_page to set the destructor */
915 set_compound_order(page, order);
916 __SetPageHead(page);
Andrea Arcangelief5a22b2013-10-16 13:46:56 -0700917 __ClearPageReserved(page);
Wu Fengguang20a03072009-06-16 15:32:22 -0700918 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
Andrea Arcangelief5a22b2013-10-16 13:46:56 -0700919 /*
920 * For gigantic hugepages allocated through bootmem at
921 * boot, it's safer to be consistent with the not-gigantic
922 * hugepages and clear the PG_reserved bit from all tail pages
923 * too. Otherwse drivers using get_user_pages() to access tail
924 * pages may get the reference counting wrong if they see
925 * PG_reserved set on a tail page (despite the head page not
926 * having PG_reserved set). Enforcing this consistency between
927 * head and tail pages allows drivers to optimize away a check
928 * on the head page when they need know if put_page() is needed
929 * after get_user_pages().
930 */
931 __ClearPageReserved(p);
Youquan Song58a84aa2011-12-08 14:34:18 -0800932 set_page_count(p, 0);
Wu Fengguang20a03072009-06-16 15:32:22 -0700933 p->first_page = page;
David Rientjes44fc8052015-03-12 16:25:54 -0700934 /* Make sure p->first_page is always valid for PageTail() */
935 smp_wmb();
936 __SetPageTail(p);
Wu Fengguang20a03072009-06-16 15:32:22 -0700937 }
938}
939
Andrew Morton77959122012-10-08 16:34:11 -0700940/*
941 * PageHuge() only returns true for hugetlbfs pages, but not for normal or
942 * transparent huge pages. See the PageTransHuge() documentation for more
943 * details.
944 */
Wu Fengguang20a03072009-06-16 15:32:22 -0700945int PageHuge(struct page *page)
946{
Wu Fengguang20a03072009-06-16 15:32:22 -0700947 if (!PageCompound(page))
948 return 0;
949
950 page = compound_head(page);
Andrew Morton758f66a2014-01-21 15:48:57 -0800951 return get_compound_page_dtor(page) == free_huge_page;
Wu Fengguang20a03072009-06-16 15:32:22 -0700952}
Naoya Horiguchi43131e12010-05-28 09:29:22 +0900953EXPORT_SYMBOL_GPL(PageHuge);
954
Andrea Arcangeli27c73ae2013-11-21 14:32:02 -0800955/*
956 * PageHeadHuge() only returns true for hugetlbfs head page, but not for
957 * normal or transparent huge pages.
958 */
959int PageHeadHuge(struct page *page_head)
960{
Andrea Arcangeli27c73ae2013-11-21 14:32:02 -0800961 if (!PageHead(page_head))
962 return 0;
963
Andrew Morton758f66a2014-01-21 15:48:57 -0800964 return get_compound_page_dtor(page_head) == free_huge_page;
Andrea Arcangeli27c73ae2013-11-21 14:32:02 -0800965}
Andrea Arcangeli27c73ae2013-11-21 14:32:02 -0800966
Zhang Yi13d60f42013-06-25 21:19:31 +0800967pgoff_t __basepage_index(struct page *page)
968{
969 struct page *page_head = compound_head(page);
970 pgoff_t index = page_index(page_head);
971 unsigned long compound_idx;
972
973 if (!PageHuge(page_head))
974 return page_index(page);
975
976 if (compound_order(page_head) >= MAX_ORDER)
977 compound_idx = page_to_pfn(page) - page_to_pfn(page_head);
978 else
979 compound_idx = page - page_head;
980
981 return (index << compound_order(page_head)) + compound_idx;
982}
983
Andi Kleena5516432008-07-23 21:27:41 -0700984static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700985{
Linus Torvalds1da177e2005-04-16 15:20:36 -0700986 struct page *page;
Joe Jinf96efd52007-07-15 23:38:12 -0700987
Mel Gorman6484eb32009-06-16 15:31:54 -0700988 page = alloc_pages_exact_node(nid,
Naoya Horiguchi86cdb462013-09-11 14:22:13 -0700989 htlb_alloc_mask(h)|__GFP_COMP|__GFP_THISNODE|
Nishanth Aravamudan551883a2008-04-29 00:58:26 -0700990 __GFP_REPEAT|__GFP_NOWARN,
Andi Kleena5516432008-07-23 21:27:41 -0700991 huge_page_order(h));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700992 if (page) {
Gerald Schaefer7f2e9522008-04-28 02:13:29 -0700993 if (arch_prepare_hugepage(page)) {
Gerald Schaefercaff3a22008-08-12 15:08:38 -0700994 __free_pages(page, huge_page_order(h));
Harvey Harrison7b8ee842008-04-28 14:13:19 -0700995 return NULL;
Gerald Schaefer7f2e9522008-04-28 02:13:29 -0700996 }
Andi Kleena5516432008-07-23 21:27:41 -0700997 prep_new_huge_page(h, page, nid);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700998 }
Nishanth Aravamudan63b46132007-10-16 01:26:24 -0700999
1000 return page;
1001}
1002
Joonsoo Kimb2261022013-09-11 14:21:00 -07001003static int alloc_fresh_huge_page(struct hstate *h, nodemask_t *nodes_allowed)
1004{
1005 struct page *page;
1006 int nr_nodes, node;
1007 int ret = 0;
1008
1009 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
1010 page = alloc_fresh_huge_page_node(h, node);
1011 if (page) {
1012 ret = 1;
1013 break;
1014 }
1015 }
1016
1017 if (ret)
1018 count_vm_event(HTLB_BUDDY_PGALLOC);
1019 else
1020 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
1021
1022 return ret;
1023}
1024
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001025/*
1026 * Free huge page from pool from next node to free.
1027 * Attempt to keep persistent huge pages more or less
1028 * balanced over allowed nodes.
1029 * Called with hugetlb_lock locked.
1030 */
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001031static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
1032 bool acct_surplus)
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001033{
Joonsoo Kimb2261022013-09-11 14:21:00 -07001034 int nr_nodes, node;
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001035 int ret = 0;
1036
Joonsoo Kimb2261022013-09-11 14:21:00 -07001037 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
Lee Schermerhorn685f3452009-09-21 17:01:23 -07001038 /*
1039 * If we're returning unused surplus pages, only examine
1040 * nodes with surplus pages.
1041 */
Joonsoo Kimb2261022013-09-11 14:21:00 -07001042 if ((!acct_surplus || h->surplus_huge_pages_node[node]) &&
1043 !list_empty(&h->hugepage_freelists[node])) {
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001044 struct page *page =
Joonsoo Kimb2261022013-09-11 14:21:00 -07001045 list_entry(h->hugepage_freelists[node].next,
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001046 struct page, lru);
1047 list_del(&page->lru);
1048 h->free_huge_pages--;
Joonsoo Kimb2261022013-09-11 14:21:00 -07001049 h->free_huge_pages_node[node]--;
Lee Schermerhorn685f3452009-09-21 17:01:23 -07001050 if (acct_surplus) {
1051 h->surplus_huge_pages--;
Joonsoo Kimb2261022013-09-11 14:21:00 -07001052 h->surplus_huge_pages_node[node]--;
Lee Schermerhorn685f3452009-09-21 17:01:23 -07001053 }
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001054 update_and_free_page(h, page);
1055 ret = 1;
Lee Schermerhorn9a76db02009-12-14 17:58:15 -08001056 break;
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001057 }
Joonsoo Kimb2261022013-09-11 14:21:00 -07001058 }
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001059
1060 return ret;
1061}
1062
Naoya Horiguchic8721bb2013-09-11 14:22:09 -07001063/*
1064 * Dissolve a given free hugepage into free buddy pages. This function does
1065 * nothing for in-use (including surplus) hugepages.
1066 */
1067static void dissolve_free_huge_page(struct page *page)
1068{
1069 spin_lock(&hugetlb_lock);
1070 if (PageHuge(page) && !page_count(page)) {
1071 struct hstate *h = page_hstate(page);
1072 int nid = page_to_nid(page);
1073 list_del(&page->lru);
1074 h->free_huge_pages--;
1075 h->free_huge_pages_node[nid]--;
1076 update_and_free_page(h, page);
1077 }
1078 spin_unlock(&hugetlb_lock);
1079}
1080
1081/*
1082 * Dissolve free hugepages in a given pfn range. Used by memory hotplug to
1083 * make specified memory blocks removable from the system.
1084 * Note that start_pfn should aligned with (minimum) hugepage size.
1085 */
1086void dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn)
1087{
1088 unsigned int order = 8 * sizeof(void *);
1089 unsigned long pfn;
1090 struct hstate *h;
1091
Li Zhongd0177632014-08-06 16:07:56 -07001092 if (!hugepages_supported())
1093 return;
1094
Naoya Horiguchic8721bb2013-09-11 14:22:09 -07001095 /* Set scan step to minimum hugepage size */
1096 for_each_hstate(h)
1097 if (order > huge_page_order(h))
1098 order = huge_page_order(h);
1099 VM_BUG_ON(!IS_ALIGNED(start_pfn, 1 << order));
1100 for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << order)
1101 dissolve_free_huge_page(pfn_to_page(pfn));
1102}
1103
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001104static struct page *alloc_buddy_huge_page(struct hstate *h, int nid)
Adam Litke7893d1d2007-10-16 01:26:18 -07001105{
1106 struct page *page;
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001107 unsigned int r_nid;
Adam Litke7893d1d2007-10-16 01:26:18 -07001108
Luiz Capitulinobae7f4a2014-06-04 16:07:08 -07001109 if (hstate_is_gigantic(h))
Andi Kleenaa888a72008-07-23 21:27:47 -07001110 return NULL;
1111
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001112 /*
1113 * Assume we will successfully allocate the surplus page to
1114 * prevent racing processes from causing the surplus to exceed
1115 * overcommit
1116 *
1117 * This however introduces a different race, where a process B
1118 * tries to grow the static hugepage pool while alloc_pages() is
1119 * called by process A. B will only examine the per-node
1120 * counters in determining if surplus huge pages can be
1121 * converted to normal huge pages in adjust_pool_surplus(). A
1122 * won't be able to increment the per-node counter, until the
1123 * lock is dropped by B, but B doesn't drop hugetlb_lock until
1124 * no more huge pages can be converted from surplus to normal
1125 * state (and doesn't try to convert again). Thus, we have a
1126 * case where a surplus huge page exists, the pool is grown, and
1127 * the surplus huge page still exists after, even though it
1128 * should just have been converted to a normal huge page. This
1129 * does not leak memory, though, as the hugepage will be freed
1130 * once it is out of use. It also does not allow the counters to
1131 * go out of whack in adjust_pool_surplus() as we don't modify
1132 * the node values until we've gotten the hugepage and only the
1133 * per-node value is checked there.
1134 */
1135 spin_lock(&hugetlb_lock);
Andi Kleena5516432008-07-23 21:27:41 -07001136 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001137 spin_unlock(&hugetlb_lock);
1138 return NULL;
1139 } else {
Andi Kleena5516432008-07-23 21:27:41 -07001140 h->nr_huge_pages++;
1141 h->surplus_huge_pages++;
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001142 }
1143 spin_unlock(&hugetlb_lock);
1144
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001145 if (nid == NUMA_NO_NODE)
Naoya Horiguchi86cdb462013-09-11 14:22:13 -07001146 page = alloc_pages(htlb_alloc_mask(h)|__GFP_COMP|
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001147 __GFP_REPEAT|__GFP_NOWARN,
1148 huge_page_order(h));
1149 else
1150 page = alloc_pages_exact_node(nid,
Naoya Horiguchi86cdb462013-09-11 14:22:13 -07001151 htlb_alloc_mask(h)|__GFP_COMP|__GFP_THISNODE|
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001152 __GFP_REPEAT|__GFP_NOWARN, huge_page_order(h));
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001153
Gerald Schaefercaff3a22008-08-12 15:08:38 -07001154 if (page && arch_prepare_hugepage(page)) {
1155 __free_pages(page, huge_page_order(h));
Hillf Dantonea5768c2012-01-10 15:08:30 -08001156 page = NULL;
Gerald Schaefercaff3a22008-08-12 15:08:38 -07001157 }
1158
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001159 spin_lock(&hugetlb_lock);
Adam Litke7893d1d2007-10-16 01:26:18 -07001160 if (page) {
Aneesh Kumar K.V0edaecf2012-07-31 16:42:07 -07001161 INIT_LIST_HEAD(&page->lru);
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001162 r_nid = page_to_nid(page);
Adam Litke7893d1d2007-10-16 01:26:18 -07001163 set_compound_page_dtor(page, free_huge_page);
Aneesh Kumar K.V9dd540e2012-07-31 16:42:15 -07001164 set_hugetlb_cgroup(page, NULL);
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001165 /*
1166 * We incremented the global counters already
1167 */
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001168 h->nr_huge_pages_node[r_nid]++;
1169 h->surplus_huge_pages_node[r_nid]++;
Adam Litke3b116302008-04-28 02:13:06 -07001170 __count_vm_event(HTLB_BUDDY_PGALLOC);
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001171 } else {
Andi Kleena5516432008-07-23 21:27:41 -07001172 h->nr_huge_pages--;
1173 h->surplus_huge_pages--;
Adam Litke3b116302008-04-28 02:13:06 -07001174 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
Adam Litke7893d1d2007-10-16 01:26:18 -07001175 }
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001176 spin_unlock(&hugetlb_lock);
Adam Litke7893d1d2007-10-16 01:26:18 -07001177
1178 return page;
1179}
1180
Adam Litkee4e574b2007-10-16 01:26:19 -07001181/*
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001182 * This allocation function is useful in the context where vma is irrelevant.
1183 * E.g. soft-offlining uses this function because it only cares physical
1184 * address of error page.
1185 */
1186struct page *alloc_huge_page_node(struct hstate *h, int nid)
1187{
Joonsoo Kim4ef91842013-09-11 14:21:51 -07001188 struct page *page = NULL;
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001189
1190 spin_lock(&hugetlb_lock);
Joonsoo Kim4ef91842013-09-11 14:21:51 -07001191 if (h->free_huge_pages - h->resv_huge_pages > 0)
1192 page = dequeue_huge_page_node(h, nid);
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001193 spin_unlock(&hugetlb_lock);
1194
Aneesh Kumar K.V94ae8ba2012-07-31 16:42:35 -07001195 if (!page)
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001196 page = alloc_buddy_huge_page(h, nid);
1197
1198 return page;
1199}
1200
1201/*
Lucas De Marchi25985ed2011-03-30 22:57:33 -03001202 * Increase the hugetlb pool such that it can accommodate a reservation
Adam Litkee4e574b2007-10-16 01:26:19 -07001203 * of size 'delta'.
1204 */
Andi Kleena5516432008-07-23 21:27:41 -07001205static int gather_surplus_pages(struct hstate *h, int delta)
Adam Litkee4e574b2007-10-16 01:26:19 -07001206{
1207 struct list_head surplus_list;
1208 struct page *page, *tmp;
1209 int ret, i;
1210 int needed, allocated;
Hillf Danton28073b02012-03-21 16:34:00 -07001211 bool alloc_ok = true;
Adam Litkee4e574b2007-10-16 01:26:19 -07001212
Andi Kleena5516432008-07-23 21:27:41 -07001213 needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
Adam Litkeac09b3a2008-03-04 14:29:38 -08001214 if (needed <= 0) {
Andi Kleena5516432008-07-23 21:27:41 -07001215 h->resv_huge_pages += delta;
Adam Litkee4e574b2007-10-16 01:26:19 -07001216 return 0;
Adam Litkeac09b3a2008-03-04 14:29:38 -08001217 }
Adam Litkee4e574b2007-10-16 01:26:19 -07001218
1219 allocated = 0;
1220 INIT_LIST_HEAD(&surplus_list);
1221
1222 ret = -ENOMEM;
1223retry:
1224 spin_unlock(&hugetlb_lock);
1225 for (i = 0; i < needed; i++) {
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001226 page = alloc_buddy_huge_page(h, NUMA_NO_NODE);
Hillf Danton28073b02012-03-21 16:34:00 -07001227 if (!page) {
1228 alloc_ok = false;
1229 break;
1230 }
Adam Litkee4e574b2007-10-16 01:26:19 -07001231 list_add(&page->lru, &surplus_list);
1232 }
Hillf Danton28073b02012-03-21 16:34:00 -07001233 allocated += i;
Adam Litkee4e574b2007-10-16 01:26:19 -07001234
1235 /*
1236 * After retaking hugetlb_lock, we need to recalculate 'needed'
1237 * because either resv_huge_pages or free_huge_pages may have changed.
1238 */
1239 spin_lock(&hugetlb_lock);
Andi Kleena5516432008-07-23 21:27:41 -07001240 needed = (h->resv_huge_pages + delta) -
1241 (h->free_huge_pages + allocated);
Hillf Danton28073b02012-03-21 16:34:00 -07001242 if (needed > 0) {
1243 if (alloc_ok)
1244 goto retry;
1245 /*
1246 * We were not able to allocate enough pages to
1247 * satisfy the entire reservation so we free what
1248 * we've allocated so far.
1249 */
1250 goto free;
1251 }
Adam Litkee4e574b2007-10-16 01:26:19 -07001252 /*
1253 * The surplus_list now contains _at_least_ the number of extra pages
Lucas De Marchi25985ed2011-03-30 22:57:33 -03001254 * needed to accommodate the reservation. Add the appropriate number
Adam Litkee4e574b2007-10-16 01:26:19 -07001255 * of pages to the hugetlb pool and free the extras back to the buddy
Adam Litkeac09b3a2008-03-04 14:29:38 -08001256 * allocator. Commit the entire reservation here to prevent another
1257 * process from stealing the pages as they are added to the pool but
1258 * before they are reserved.
Adam Litkee4e574b2007-10-16 01:26:19 -07001259 */
1260 needed += allocated;
Andi Kleena5516432008-07-23 21:27:41 -07001261 h->resv_huge_pages += delta;
Adam Litkee4e574b2007-10-16 01:26:19 -07001262 ret = 0;
Naoya Horiguchia9869b82010-09-08 10:19:37 +09001263
Adam Litke19fc3f02008-04-28 02:12:20 -07001264 /* Free the needed pages to the hugetlb pool */
Adam Litkee4e574b2007-10-16 01:26:19 -07001265 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
Adam Litke19fc3f02008-04-28 02:12:20 -07001266 if ((--needed) < 0)
1267 break;
Naoya Horiguchia9869b82010-09-08 10:19:37 +09001268 /*
1269 * This page is now managed by the hugetlb allocator and has
1270 * no users -- drop the buddy allocator's reference.
1271 */
1272 put_page_testzero(page);
Sasha Levin309381fea2014-01-23 15:52:54 -08001273 VM_BUG_ON_PAGE(page_count(page), page);
Andi Kleena5516432008-07-23 21:27:41 -07001274 enqueue_huge_page(h, page);
Adam Litke19fc3f02008-04-28 02:12:20 -07001275 }
Hillf Danton28073b02012-03-21 16:34:00 -07001276free:
Hillf Dantonb0365c82011-12-28 15:57:16 -08001277 spin_unlock(&hugetlb_lock);
Adam Litke19fc3f02008-04-28 02:12:20 -07001278
1279 /* Free unnecessary surplus pages to the buddy allocator */
Joonsoo Kimc0d934b2013-09-11 14:21:02 -07001280 list_for_each_entry_safe(page, tmp, &surplus_list, lru)
1281 put_page(page);
Naoya Horiguchia9869b82010-09-08 10:19:37 +09001282 spin_lock(&hugetlb_lock);
Adam Litkee4e574b2007-10-16 01:26:19 -07001283
1284 return ret;
1285}
1286
1287/*
1288 * When releasing a hugetlb pool reservation, any surplus pages that were
1289 * allocated to satisfy the reservation must be explicitly freed if they were
1290 * never used.
Lee Schermerhorn685f3452009-09-21 17:01:23 -07001291 * Called with hugetlb_lock held.
Adam Litkee4e574b2007-10-16 01:26:19 -07001292 */
Andi Kleena5516432008-07-23 21:27:41 -07001293static void return_unused_surplus_pages(struct hstate *h,
1294 unsigned long unused_resv_pages)
Adam Litkee4e574b2007-10-16 01:26:19 -07001295{
Adam Litkee4e574b2007-10-16 01:26:19 -07001296 unsigned long nr_pages;
1297
Adam Litkeac09b3a2008-03-04 14:29:38 -08001298 /* Uncommit the reservation */
Andi Kleena5516432008-07-23 21:27:41 -07001299 h->resv_huge_pages -= unused_resv_pages;
Adam Litkeac09b3a2008-03-04 14:29:38 -08001300
Andi Kleenaa888a72008-07-23 21:27:47 -07001301 /* Cannot return gigantic pages currently */
Luiz Capitulinobae7f4a2014-06-04 16:07:08 -07001302 if (hstate_is_gigantic(h))
Andi Kleenaa888a72008-07-23 21:27:47 -07001303 return;
1304
Andi Kleena5516432008-07-23 21:27:41 -07001305 nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
Adam Litkee4e574b2007-10-16 01:26:19 -07001306
Lee Schermerhorn685f3452009-09-21 17:01:23 -07001307 /*
1308 * We want to release as many surplus pages as possible, spread
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -08001309 * evenly across all nodes with memory. Iterate across these nodes
1310 * until we can no longer free unreserved surplus pages. This occurs
1311 * when the nodes with surplus pages have no free pages.
1312 * free_pool_huge_page() will balance the the freed pages across the
1313 * on-line nodes with memory and will handle the hstate accounting.
Lee Schermerhorn685f3452009-09-21 17:01:23 -07001314 */
1315 while (nr_pages--) {
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08001316 if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1))
Lee Schermerhorn685f3452009-09-21 17:01:23 -07001317 break;
Mizuma, Masayoshi7848a4b2014-04-18 15:07:18 -07001318 cond_resched_lock(&hugetlb_lock);
Adam Litkee4e574b2007-10-16 01:26:19 -07001319 }
1320}
1321
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001322/*
1323 * Determine if the huge page at addr within the vma has an associated
1324 * reservation. Where it does not we will need to logically increase
David Gibson90481622012-03-21 16:34:12 -07001325 * reservation and actually increase subpool usage before an allocation
1326 * can occur. Where any new reservation would be required the
1327 * reservation change is prepared, but not committed. Once the page
1328 * has been allocated from the subpool and instantiated the change should
1329 * be committed via vma_commit_reservation. No action is required on
1330 * failure.
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001331 */
Roel Kluine2f17d92009-03-31 15:23:15 -07001332static long vma_needs_reservation(struct hstate *h,
Andi Kleena5516432008-07-23 21:27:41 -07001333 struct vm_area_struct *vma, unsigned long addr)
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001334{
Joonsoo Kim4e35f482014-04-03 14:47:30 -07001335 struct resv_map *resv;
1336 pgoff_t idx;
1337 long chg;
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001338
Joonsoo Kim4e35f482014-04-03 14:47:30 -07001339 resv = vma_resv_map(vma);
1340 if (!resv)
Andy Whitcroft84afd992008-07-23 21:27:32 -07001341 return 1;
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001342
Joonsoo Kim4e35f482014-04-03 14:47:30 -07001343 idx = vma_hugecache_offset(h, vma, addr);
1344 chg = region_chg(resv, idx, idx + 1);
Andy Whitcroft84afd992008-07-23 21:27:32 -07001345
Joonsoo Kim4e35f482014-04-03 14:47:30 -07001346 if (vma->vm_flags & VM_MAYSHARE)
1347 return chg;
1348 else
1349 return chg < 0 ? chg : 0;
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001350}
Andi Kleena5516432008-07-23 21:27:41 -07001351static void vma_commit_reservation(struct hstate *h,
1352 struct vm_area_struct *vma, unsigned long addr)
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001353{
Joonsoo Kim4e35f482014-04-03 14:47:30 -07001354 struct resv_map *resv;
1355 pgoff_t idx;
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001356
Joonsoo Kim4e35f482014-04-03 14:47:30 -07001357 resv = vma_resv_map(vma);
1358 if (!resv)
1359 return;
Joonsoo Kim9119a412014-04-03 14:47:25 -07001360
Joonsoo Kim4e35f482014-04-03 14:47:30 -07001361 idx = vma_hugecache_offset(h, vma, addr);
1362 region_add(resv, idx, idx + 1);
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001363}
1364
David Gibson27a85ef2006-03-22 00:08:56 -08001365static struct page *alloc_huge_page(struct vm_area_struct *vma,
Mel Gorman04f2cbe2008-07-23 21:27:25 -07001366 unsigned long addr, int avoid_reserve)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001367{
David Gibson90481622012-03-21 16:34:12 -07001368 struct hugepage_subpool *spool = subpool_vma(vma);
Andi Kleena5516432008-07-23 21:27:41 -07001369 struct hstate *h = hstate_vma(vma);
Adam Litke348ea202007-11-14 16:59:37 -08001370 struct page *page;
Roel Kluine2f17d92009-03-31 15:23:15 -07001371 long chg;
Aneesh Kumar K.V6d76dcf2012-07-31 16:42:18 -07001372 int ret, idx;
1373 struct hugetlb_cgroup *h_cg;
Adam Litke2fc39ce2007-11-14 16:59:39 -08001374
Aneesh Kumar K.V6d76dcf2012-07-31 16:42:18 -07001375 idx = hstate_index(h);
Mel Gormana1e78772008-07-23 21:27:23 -07001376 /*
David Gibson90481622012-03-21 16:34:12 -07001377 * Processes that did not create the mapping will have no
1378 * reserves and will not have accounted against subpool
1379 * limit. Check that the subpool limit can be made before
1380 * satisfying the allocation MAP_NORESERVE mappings may also
1381 * need pages and subpool limit allocated allocated if no reserve
1382 * mapping overlaps.
Mel Gormana1e78772008-07-23 21:27:23 -07001383 */
Andi Kleena5516432008-07-23 21:27:41 -07001384 chg = vma_needs_reservation(h, vma, addr);
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001385 if (chg < 0)
Aneesh Kumar K.V76dcee72012-07-31 16:41:57 -07001386 return ERR_PTR(-ENOMEM);
Joonsoo Kim8bb3f122013-09-11 14:21:54 -07001387 if (chg || avoid_reserve)
1388 if (hugepage_subpool_get_pages(spool, 1))
Aneesh Kumar K.V76dcee72012-07-31 16:41:57 -07001389 return ERR_PTR(-ENOSPC);
Mel Gormana1e78772008-07-23 21:27:23 -07001390
Aneesh Kumar K.V6d76dcf2012-07-31 16:42:18 -07001391 ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg);
Jianyu Zhan8f34af62014-06-04 16:10:36 -07001392 if (ret)
1393 goto out_subpool_put;
1394
Mel Gormana1e78772008-07-23 21:27:23 -07001395 spin_lock(&hugetlb_lock);
Joonsoo Kimaf0ed732013-09-11 14:21:18 -07001396 page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, chg);
Joonsoo Kim81a6fca2013-09-11 14:20:58 -07001397 if (!page) {
Aneesh Kumar K.V94ae8ba2012-07-31 16:42:35 -07001398 spin_unlock(&hugetlb_lock);
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001399 page = alloc_buddy_huge_page(h, NUMA_NO_NODE);
Jianyu Zhan8f34af62014-06-04 16:10:36 -07001400 if (!page)
1401 goto out_uncharge_cgroup;
1402
Aneesh Kumar K.V79dbb232012-07-31 16:42:32 -07001403 spin_lock(&hugetlb_lock);
1404 list_move(&page->lru, &h->hugepage_activelist);
Joonsoo Kim81a6fca2013-09-11 14:20:58 -07001405 /* Fall through */
Mel Gormana1e78772008-07-23 21:27:23 -07001406 }
Joonsoo Kim81a6fca2013-09-11 14:20:58 -07001407 hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page);
1408 spin_unlock(&hugetlb_lock);
Mel Gormana1e78772008-07-23 21:27:23 -07001409
David Gibson90481622012-03-21 16:34:12 -07001410 set_page_private(page, (unsigned long)spool);
Mel Gormana1e78772008-07-23 21:27:23 -07001411
Andi Kleena5516432008-07-23 21:27:41 -07001412 vma_commit_reservation(h, vma, addr);
Adam Litke90d8b7e2007-11-14 16:59:42 -08001413 return page;
Jianyu Zhan8f34af62014-06-04 16:10:36 -07001414
1415out_uncharge_cgroup:
1416 hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg);
1417out_subpool_put:
1418 if (chg || avoid_reserve)
1419 hugepage_subpool_put_pages(spool, 1);
1420 return ERR_PTR(-ENOSPC);
David Gibsonb45b5bd2006-03-22 00:08:55 -08001421}
1422
Naoya Horiguchi74060e42013-09-11 14:22:06 -07001423/*
1424 * alloc_huge_page()'s wrapper which simply returns the page if allocation
1425 * succeeds, otherwise NULL. This function is called from new_vma_page(),
1426 * where no ERR_VALUE is expected to be returned.
1427 */
1428struct page *alloc_huge_page_noerr(struct vm_area_struct *vma,
1429 unsigned long addr, int avoid_reserve)
1430{
1431 struct page *page = alloc_huge_page(vma, addr, avoid_reserve);
1432 if (IS_ERR(page))
1433 page = NULL;
1434 return page;
1435}
1436
Cyrill Gorcunov91f47662009-01-06 14:40:33 -08001437int __weak alloc_bootmem_huge_page(struct hstate *h)
Andi Kleenaa888a72008-07-23 21:27:47 -07001438{
1439 struct huge_bootmem_page *m;
Joonsoo Kimb2261022013-09-11 14:21:00 -07001440 int nr_nodes, node;
Andi Kleenaa888a72008-07-23 21:27:47 -07001441
Joonsoo Kimb2261022013-09-11 14:21:00 -07001442 for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) {
Andi Kleenaa888a72008-07-23 21:27:47 -07001443 void *addr;
1444
Grygorii Strashko8b89a112014-01-21 15:50:36 -08001445 addr = memblock_virt_alloc_try_nid_nopanic(
1446 huge_page_size(h), huge_page_size(h),
1447 0, BOOTMEM_ALLOC_ACCESSIBLE, node);
Andi Kleenaa888a72008-07-23 21:27:47 -07001448 if (addr) {
1449 /*
1450 * Use the beginning of the huge page to store the
1451 * huge_bootmem_page struct (until gather_bootmem
1452 * puts them into the mem_map).
1453 */
1454 m = addr;
Cyrill Gorcunov91f47662009-01-06 14:40:33 -08001455 goto found;
Andi Kleenaa888a72008-07-23 21:27:47 -07001456 }
Andi Kleenaa888a72008-07-23 21:27:47 -07001457 }
1458 return 0;
1459
1460found:
Luiz Capitulinodf994ea2014-12-12 16:55:21 -08001461 BUG_ON(!IS_ALIGNED(virt_to_phys(m), huge_page_size(h)));
Andi Kleenaa888a72008-07-23 21:27:47 -07001462 /* Put them into a private list first because mem_map is not up yet */
1463 list_add(&m->list, &huge_boot_pages);
1464 m->hstate = h;
1465 return 1;
1466}
1467
David Rientjesf412c972014-04-03 14:47:59 -07001468static void __init prep_compound_huge_page(struct page *page, int order)
Andy Whitcroft18229df2008-11-06 12:53:27 -08001469{
1470 if (unlikely(order > (MAX_ORDER - 1)))
1471 prep_compound_gigantic_page(page, order);
1472 else
1473 prep_compound_page(page, order);
1474}
1475
Andi Kleenaa888a72008-07-23 21:27:47 -07001476/* Put bootmem huge pages into the standard lists after mem_map is up */
1477static void __init gather_bootmem_prealloc(void)
1478{
1479 struct huge_bootmem_page *m;
1480
1481 list_for_each_entry(m, &huge_boot_pages, list) {
Andi Kleenaa888a72008-07-23 21:27:47 -07001482 struct hstate *h = m->hstate;
Becky Bruceee8f2482011-07-25 17:11:50 -07001483 struct page *page;
1484
1485#ifdef CONFIG_HIGHMEM
1486 page = pfn_to_page(m->phys >> PAGE_SHIFT);
Grygorii Strashko8b89a112014-01-21 15:50:36 -08001487 memblock_free_late(__pa(m),
1488 sizeof(struct huge_bootmem_page));
Becky Bruceee8f2482011-07-25 17:11:50 -07001489#else
1490 page = virt_to_page(m);
1491#endif
Andi Kleenaa888a72008-07-23 21:27:47 -07001492 WARN_ON(page_count(page) != 1);
Andy Whitcroft18229df2008-11-06 12:53:27 -08001493 prep_compound_huge_page(page, h->order);
Andrea Arcangelief5a22b2013-10-16 13:46:56 -07001494 WARN_ON(PageReserved(page));
Andi Kleenaa888a72008-07-23 21:27:47 -07001495 prep_new_huge_page(h, page, page_to_nid(page));
Rafael Aquinib0320c72011-06-15 15:08:39 -07001496 /*
1497 * If we had gigantic hugepages allocated at boot time, we need
1498 * to restore the 'stolen' pages to totalram_pages in order to
1499 * fix confusing memory reports from free(1) and another
1500 * side-effects, like CommitLimit going negative.
1501 */
Luiz Capitulinobae7f4a2014-06-04 16:07:08 -07001502 if (hstate_is_gigantic(h))
Jiang Liu3dcc0572013-07-03 15:03:21 -07001503 adjust_managed_page_count(page, 1 << h->order);
Andi Kleenaa888a72008-07-23 21:27:47 -07001504 }
1505}
1506
Andi Kleen8faa8b02008-07-23 21:27:48 -07001507static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001508{
1509 unsigned long i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001510
Andi Kleene5ff2152008-07-23 21:27:42 -07001511 for (i = 0; i < h->max_huge_pages; ++i) {
Luiz Capitulinobae7f4a2014-06-04 16:07:08 -07001512 if (hstate_is_gigantic(h)) {
Andi Kleenaa888a72008-07-23 21:27:47 -07001513 if (!alloc_bootmem_huge_page(h))
1514 break;
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -08001515 } else if (!alloc_fresh_huge_page(h,
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08001516 &node_states[N_MEMORY]))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001517 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001518 }
Andi Kleen8faa8b02008-07-23 21:27:48 -07001519 h->max_huge_pages = i;
Andi Kleene5ff2152008-07-23 21:27:42 -07001520}
1521
1522static void __init hugetlb_init_hstates(void)
1523{
1524 struct hstate *h;
1525
1526 for_each_hstate(h) {
Andi Kleen8faa8b02008-07-23 21:27:48 -07001527 /* oversize hugepages were init'ed in early boot */
Luiz Capitulinobae7f4a2014-06-04 16:07:08 -07001528 if (!hstate_is_gigantic(h))
Andi Kleen8faa8b02008-07-23 21:27:48 -07001529 hugetlb_hstate_alloc_pages(h);
Andi Kleene5ff2152008-07-23 21:27:42 -07001530 }
1531}
1532
Andi Kleen4abd32d2008-07-23 21:27:49 -07001533static char * __init memfmt(char *buf, unsigned long n)
1534{
1535 if (n >= (1UL << 30))
1536 sprintf(buf, "%lu GB", n >> 30);
1537 else if (n >= (1UL << 20))
1538 sprintf(buf, "%lu MB", n >> 20);
1539 else
1540 sprintf(buf, "%lu KB", n >> 10);
1541 return buf;
1542}
1543
Andi Kleene5ff2152008-07-23 21:27:42 -07001544static void __init report_hugepages(void)
1545{
1546 struct hstate *h;
1547
1548 for_each_hstate(h) {
Andi Kleen4abd32d2008-07-23 21:27:49 -07001549 char buf[32];
Andrew Mortonffb22af2013-02-22 16:32:08 -08001550 pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n",
Andi Kleen4abd32d2008-07-23 21:27:49 -07001551 memfmt(buf, huge_page_size(h)),
1552 h->free_huge_pages);
Andi Kleene5ff2152008-07-23 21:27:42 -07001553 }
1554}
1555
Linus Torvalds1da177e2005-04-16 15:20:36 -07001556#ifdef CONFIG_HIGHMEM
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001557static void try_to_free_low(struct hstate *h, unsigned long count,
1558 nodemask_t *nodes_allowed)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001559{
Christoph Lameter4415cc82006-09-25 23:31:55 -07001560 int i;
1561
Luiz Capitulinobae7f4a2014-06-04 16:07:08 -07001562 if (hstate_is_gigantic(h))
Andi Kleenaa888a72008-07-23 21:27:47 -07001563 return;
1564
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001565 for_each_node_mask(i, *nodes_allowed) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001566 struct page *page, *next;
Andi Kleena5516432008-07-23 21:27:41 -07001567 struct list_head *freel = &h->hugepage_freelists[i];
1568 list_for_each_entry_safe(page, next, freel, lru) {
1569 if (count >= h->nr_huge_pages)
Adam Litke6b0c8802007-10-16 01:26:23 -07001570 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001571 if (PageHighMem(page))
1572 continue;
1573 list_del(&page->lru);
Andi Kleene5ff2152008-07-23 21:27:42 -07001574 update_and_free_page(h, page);
Andi Kleena5516432008-07-23 21:27:41 -07001575 h->free_huge_pages--;
1576 h->free_huge_pages_node[page_to_nid(page)]--;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001577 }
1578 }
1579}
1580#else
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001581static inline void try_to_free_low(struct hstate *h, unsigned long count,
1582 nodemask_t *nodes_allowed)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001583{
1584}
1585#endif
1586
Wu Fengguang20a03072009-06-16 15:32:22 -07001587/*
1588 * Increment or decrement surplus_huge_pages. Keep node-specific counters
1589 * balanced by operating on them in a round-robin fashion.
1590 * Returns 1 if an adjustment was made.
1591 */
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001592static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
1593 int delta)
Wu Fengguang20a03072009-06-16 15:32:22 -07001594{
Joonsoo Kimb2261022013-09-11 14:21:00 -07001595 int nr_nodes, node;
Wu Fengguang20a03072009-06-16 15:32:22 -07001596
1597 VM_BUG_ON(delta != -1 && delta != 1);
Wu Fengguang20a03072009-06-16 15:32:22 -07001598
Joonsoo Kimb2261022013-09-11 14:21:00 -07001599 if (delta < 0) {
1600 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
1601 if (h->surplus_huge_pages_node[node])
1602 goto found;
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001603 }
Joonsoo Kimb2261022013-09-11 14:21:00 -07001604 } else {
1605 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
1606 if (h->surplus_huge_pages_node[node] <
1607 h->nr_huge_pages_node[node])
1608 goto found;
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001609 }
Joonsoo Kimb2261022013-09-11 14:21:00 -07001610 }
1611 return 0;
Wu Fengguang20a03072009-06-16 15:32:22 -07001612
Joonsoo Kimb2261022013-09-11 14:21:00 -07001613found:
1614 h->surplus_huge_pages += delta;
1615 h->surplus_huge_pages_node[node] += delta;
1616 return 1;
Wu Fengguang20a03072009-06-16 15:32:22 -07001617}
1618
Andi Kleena5516432008-07-23 21:27:41 -07001619#define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001620static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count,
1621 nodemask_t *nodes_allowed)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001622{
Adam Litke7893d1d2007-10-16 01:26:18 -07001623 unsigned long min_count, ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001624
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001625 if (hstate_is_gigantic(h) && !gigantic_page_supported())
Andi Kleenaa888a72008-07-23 21:27:47 -07001626 return h->max_huge_pages;
1627
Adam Litke7893d1d2007-10-16 01:26:18 -07001628 /*
1629 * Increase the pool size
1630 * First take pages out of surplus state. Then make up the
1631 * remaining difference by allocating fresh huge pages.
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001632 *
1633 * We might race with alloc_buddy_huge_page() here and be unable
1634 * to convert a surplus huge page to a normal huge page. That is
1635 * not critical, though, it just means the overall size of the
1636 * pool might be one hugepage larger than it needs to be, but
1637 * within all the constraints specified by the sysctls.
Adam Litke7893d1d2007-10-16 01:26:18 -07001638 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001639 spin_lock(&hugetlb_lock);
Andi Kleena5516432008-07-23 21:27:41 -07001640 while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001641 if (!adjust_pool_surplus(h, nodes_allowed, -1))
Adam Litke7893d1d2007-10-16 01:26:18 -07001642 break;
1643 }
1644
Andi Kleena5516432008-07-23 21:27:41 -07001645 while (count > persistent_huge_pages(h)) {
Adam Litke7893d1d2007-10-16 01:26:18 -07001646 /*
1647 * If this allocation races such that we no longer need the
1648 * page, free_huge_page will handle it by freeing the page
1649 * and reducing the surplus.
1650 */
1651 spin_unlock(&hugetlb_lock);
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001652 if (hstate_is_gigantic(h))
1653 ret = alloc_fresh_gigantic_page(h, nodes_allowed);
1654 else
1655 ret = alloc_fresh_huge_page(h, nodes_allowed);
Adam Litke7893d1d2007-10-16 01:26:18 -07001656 spin_lock(&hugetlb_lock);
1657 if (!ret)
1658 goto out;
1659
Mel Gorman536240f22009-12-14 17:59:56 -08001660 /* Bail for signals. Probably ctrl-c from user */
1661 if (signal_pending(current))
1662 goto out;
Adam Litke7893d1d2007-10-16 01:26:18 -07001663 }
Adam Litke7893d1d2007-10-16 01:26:18 -07001664
1665 /*
1666 * Decrease the pool size
1667 * First return free pages to the buddy allocator (being careful
1668 * to keep enough around to satisfy reservations). Then place
1669 * pages into surplus state as needed so the pool will shrink
1670 * to the desired size as pages become free.
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001671 *
1672 * By placing pages into the surplus state independent of the
1673 * overcommit value, we are allowing the surplus pool size to
1674 * exceed overcommit. There are few sane options here. Since
1675 * alloc_buddy_huge_page() is checking the global counter,
1676 * though, we'll note that we're not allowed to exceed surplus
1677 * and won't grow the pool anywhere else. Not until one of the
1678 * sysctls are changed, or the surplus pages go out of use.
Adam Litke7893d1d2007-10-16 01:26:18 -07001679 */
Andi Kleena5516432008-07-23 21:27:41 -07001680 min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
Adam Litke6b0c8802007-10-16 01:26:23 -07001681 min_count = max(count, min_count);
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001682 try_to_free_low(h, min_count, nodes_allowed);
Andi Kleena5516432008-07-23 21:27:41 -07001683 while (min_count < persistent_huge_pages(h)) {
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001684 if (!free_pool_huge_page(h, nodes_allowed, 0))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001685 break;
Mizuma, Masayoshi55f67142014-04-07 15:37:54 -07001686 cond_resched_lock(&hugetlb_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001687 }
Andi Kleena5516432008-07-23 21:27:41 -07001688 while (count < persistent_huge_pages(h)) {
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001689 if (!adjust_pool_surplus(h, nodes_allowed, 1))
Adam Litke7893d1d2007-10-16 01:26:18 -07001690 break;
1691 }
1692out:
Andi Kleena5516432008-07-23 21:27:41 -07001693 ret = persistent_huge_pages(h);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001694 spin_unlock(&hugetlb_lock);
Adam Litke7893d1d2007-10-16 01:26:18 -07001695 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001696}
1697
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001698#define HSTATE_ATTR_RO(_name) \
1699 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
1700
1701#define HSTATE_ATTR(_name) \
1702 static struct kobj_attribute _name##_attr = \
1703 __ATTR(_name, 0644, _name##_show, _name##_store)
1704
1705static struct kobject *hugepages_kobj;
1706static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
1707
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001708static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
1709
1710static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001711{
1712 int i;
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001713
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001714 for (i = 0; i < HUGE_MAX_HSTATE; i++)
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001715 if (hstate_kobjs[i] == kobj) {
1716 if (nidp)
1717 *nidp = NUMA_NO_NODE;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001718 return &hstates[i];
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001719 }
1720
1721 return kobj_to_node_hstate(kobj, nidp);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001722}
1723
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001724static ssize_t nr_hugepages_show_common(struct kobject *kobj,
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001725 struct kobj_attribute *attr, char *buf)
1726{
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001727 struct hstate *h;
1728 unsigned long nr_huge_pages;
1729 int nid;
1730
1731 h = kobj_to_hstate(kobj, &nid);
1732 if (nid == NUMA_NO_NODE)
1733 nr_huge_pages = h->nr_huge_pages;
1734 else
1735 nr_huge_pages = h->nr_huge_pages_node[nid];
1736
1737 return sprintf(buf, "%lu\n", nr_huge_pages);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001738}
Eric B Munsonadbe8722011-01-13 15:47:27 -08001739
David Rientjes238d3c12014-08-06 16:06:51 -07001740static ssize_t __nr_hugepages_store_common(bool obey_mempolicy,
1741 struct hstate *h, int nid,
1742 unsigned long count, size_t len)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001743{
1744 int err;
David Rientjesbad44b52009-12-14 17:58:38 -08001745 NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001746
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001747 if (hstate_is_gigantic(h) && !gigantic_page_supported()) {
Eric B Munsonadbe8722011-01-13 15:47:27 -08001748 err = -EINVAL;
1749 goto out;
1750 }
1751
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001752 if (nid == NUMA_NO_NODE) {
1753 /*
1754 * global hstate attribute
1755 */
1756 if (!(obey_mempolicy &&
1757 init_nodemask_of_mempolicy(nodes_allowed))) {
1758 NODEMASK_FREE(nodes_allowed);
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08001759 nodes_allowed = &node_states[N_MEMORY];
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001760 }
1761 } else if (nodes_allowed) {
1762 /*
1763 * per node hstate attribute: adjust count to global,
1764 * but restrict alloc/free to the specified node.
1765 */
1766 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
1767 init_nodemask_of_node(nodes_allowed, nid);
1768 } else
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08001769 nodes_allowed = &node_states[N_MEMORY];
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001770
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001771 h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001772
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08001773 if (nodes_allowed != &node_states[N_MEMORY])
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001774 NODEMASK_FREE(nodes_allowed);
1775
1776 return len;
Eric B Munsonadbe8722011-01-13 15:47:27 -08001777out:
1778 NODEMASK_FREE(nodes_allowed);
1779 return err;
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001780}
1781
David Rientjes238d3c12014-08-06 16:06:51 -07001782static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
1783 struct kobject *kobj, const char *buf,
1784 size_t len)
1785{
1786 struct hstate *h;
1787 unsigned long count;
1788 int nid;
1789 int err;
1790
1791 err = kstrtoul(buf, 10, &count);
1792 if (err)
1793 return err;
1794
1795 h = kobj_to_hstate(kobj, &nid);
1796 return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len);
1797}
1798
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001799static ssize_t nr_hugepages_show(struct kobject *kobj,
1800 struct kobj_attribute *attr, char *buf)
1801{
1802 return nr_hugepages_show_common(kobj, attr, buf);
1803}
1804
1805static ssize_t nr_hugepages_store(struct kobject *kobj,
1806 struct kobj_attribute *attr, const char *buf, size_t len)
1807{
David Rientjes238d3c12014-08-06 16:06:51 -07001808 return nr_hugepages_store_common(false, kobj, buf, len);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001809}
1810HSTATE_ATTR(nr_hugepages);
1811
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001812#ifdef CONFIG_NUMA
1813
1814/*
1815 * hstate attribute for optionally mempolicy-based constraint on persistent
1816 * huge page alloc/free.
1817 */
1818static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
1819 struct kobj_attribute *attr, char *buf)
1820{
1821 return nr_hugepages_show_common(kobj, attr, buf);
1822}
1823
1824static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
1825 struct kobj_attribute *attr, const char *buf, size_t len)
1826{
David Rientjes238d3c12014-08-06 16:06:51 -07001827 return nr_hugepages_store_common(true, kobj, buf, len);
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001828}
1829HSTATE_ATTR(nr_hugepages_mempolicy);
1830#endif
1831
1832
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001833static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
1834 struct kobj_attribute *attr, char *buf)
1835{
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001836 struct hstate *h = kobj_to_hstate(kobj, NULL);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001837 return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
1838}
Eric B Munsonadbe8722011-01-13 15:47:27 -08001839
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001840static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
1841 struct kobj_attribute *attr, const char *buf, size_t count)
1842{
1843 int err;
1844 unsigned long input;
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001845 struct hstate *h = kobj_to_hstate(kobj, NULL);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001846
Luiz Capitulinobae7f4a2014-06-04 16:07:08 -07001847 if (hstate_is_gigantic(h))
Eric B Munsonadbe8722011-01-13 15:47:27 -08001848 return -EINVAL;
1849
Jingoo Han3dbb95f2013-09-11 14:20:25 -07001850 err = kstrtoul(buf, 10, &input);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001851 if (err)
Eric B Munson73ae31e2011-01-13 15:47:28 -08001852 return err;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001853
1854 spin_lock(&hugetlb_lock);
1855 h->nr_overcommit_huge_pages = input;
1856 spin_unlock(&hugetlb_lock);
1857
1858 return count;
1859}
1860HSTATE_ATTR(nr_overcommit_hugepages);
1861
1862static ssize_t free_hugepages_show(struct kobject *kobj,
1863 struct kobj_attribute *attr, char *buf)
1864{
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001865 struct hstate *h;
1866 unsigned long free_huge_pages;
1867 int nid;
1868
1869 h = kobj_to_hstate(kobj, &nid);
1870 if (nid == NUMA_NO_NODE)
1871 free_huge_pages = h->free_huge_pages;
1872 else
1873 free_huge_pages = h->free_huge_pages_node[nid];
1874
1875 return sprintf(buf, "%lu\n", free_huge_pages);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001876}
1877HSTATE_ATTR_RO(free_hugepages);
1878
1879static ssize_t resv_hugepages_show(struct kobject *kobj,
1880 struct kobj_attribute *attr, char *buf)
1881{
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001882 struct hstate *h = kobj_to_hstate(kobj, NULL);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001883 return sprintf(buf, "%lu\n", h->resv_huge_pages);
1884}
1885HSTATE_ATTR_RO(resv_hugepages);
1886
1887static ssize_t surplus_hugepages_show(struct kobject *kobj,
1888 struct kobj_attribute *attr, char *buf)
1889{
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001890 struct hstate *h;
1891 unsigned long surplus_huge_pages;
1892 int nid;
1893
1894 h = kobj_to_hstate(kobj, &nid);
1895 if (nid == NUMA_NO_NODE)
1896 surplus_huge_pages = h->surplus_huge_pages;
1897 else
1898 surplus_huge_pages = h->surplus_huge_pages_node[nid];
1899
1900 return sprintf(buf, "%lu\n", surplus_huge_pages);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001901}
1902HSTATE_ATTR_RO(surplus_hugepages);
1903
1904static struct attribute *hstate_attrs[] = {
1905 &nr_hugepages_attr.attr,
1906 &nr_overcommit_hugepages_attr.attr,
1907 &free_hugepages_attr.attr,
1908 &resv_hugepages_attr.attr,
1909 &surplus_hugepages_attr.attr,
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001910#ifdef CONFIG_NUMA
1911 &nr_hugepages_mempolicy_attr.attr,
1912#endif
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001913 NULL,
1914};
1915
1916static struct attribute_group hstate_attr_group = {
1917 .attrs = hstate_attrs,
1918};
1919
Jeff Mahoney094e9532010-02-02 13:44:14 -08001920static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
1921 struct kobject **hstate_kobjs,
1922 struct attribute_group *hstate_attr_group)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001923{
1924 int retval;
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07001925 int hi = hstate_index(h);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001926
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001927 hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
1928 if (!hstate_kobjs[hi])
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001929 return -ENOMEM;
1930
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001931 retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001932 if (retval)
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001933 kobject_put(hstate_kobjs[hi]);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001934
1935 return retval;
1936}
1937
1938static void __init hugetlb_sysfs_init(void)
1939{
1940 struct hstate *h;
1941 int err;
1942
1943 hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
1944 if (!hugepages_kobj)
1945 return;
1946
1947 for_each_hstate(h) {
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001948 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
1949 hstate_kobjs, &hstate_attr_group);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001950 if (err)
Andrew Mortonffb22af2013-02-22 16:32:08 -08001951 pr_err("Hugetlb: Unable to add hstate %s", h->name);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001952 }
1953}
1954
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001955#ifdef CONFIG_NUMA
1956
1957/*
1958 * node_hstate/s - associate per node hstate attributes, via their kobjects,
Kay Sievers10fbcf42011-12-21 14:48:43 -08001959 * with node devices in node_devices[] using a parallel array. The array
1960 * index of a node device or _hstate == node id.
1961 * This is here to avoid any static dependency of the node device driver, in
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001962 * the base kernel, on the hugetlb module.
1963 */
1964struct node_hstate {
1965 struct kobject *hugepages_kobj;
1966 struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
1967};
1968struct node_hstate node_hstates[MAX_NUMNODES];
1969
1970/*
Kay Sievers10fbcf42011-12-21 14:48:43 -08001971 * A subset of global hstate attributes for node devices
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001972 */
1973static struct attribute *per_node_hstate_attrs[] = {
1974 &nr_hugepages_attr.attr,
1975 &free_hugepages_attr.attr,
1976 &surplus_hugepages_attr.attr,
1977 NULL,
1978};
1979
1980static struct attribute_group per_node_hstate_attr_group = {
1981 .attrs = per_node_hstate_attrs,
1982};
1983
1984/*
Kay Sievers10fbcf42011-12-21 14:48:43 -08001985 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001986 * Returns node id via non-NULL nidp.
1987 */
1988static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
1989{
1990 int nid;
1991
1992 for (nid = 0; nid < nr_node_ids; nid++) {
1993 struct node_hstate *nhs = &node_hstates[nid];
1994 int i;
1995 for (i = 0; i < HUGE_MAX_HSTATE; i++)
1996 if (nhs->hstate_kobjs[i] == kobj) {
1997 if (nidp)
1998 *nidp = nid;
1999 return &hstates[i];
2000 }
2001 }
2002
2003 BUG();
2004 return NULL;
2005}
2006
2007/*
Kay Sievers10fbcf42011-12-21 14:48:43 -08002008 * Unregister hstate attributes from a single node device.
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002009 * No-op if no hstate attributes attached.
2010 */
Claudiu Ghioc3cd8b442013-03-04 12:46:15 +02002011static void hugetlb_unregister_node(struct node *node)
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002012{
2013 struct hstate *h;
Kay Sievers10fbcf42011-12-21 14:48:43 -08002014 struct node_hstate *nhs = &node_hstates[node->dev.id];
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002015
2016 if (!nhs->hugepages_kobj)
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -08002017 return; /* no hstate attributes */
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002018
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07002019 for_each_hstate(h) {
2020 int idx = hstate_index(h);
2021 if (nhs->hstate_kobjs[idx]) {
2022 kobject_put(nhs->hstate_kobjs[idx]);
2023 nhs->hstate_kobjs[idx] = NULL;
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002024 }
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07002025 }
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002026
2027 kobject_put(nhs->hugepages_kobj);
2028 nhs->hugepages_kobj = NULL;
2029}
2030
2031/*
Kay Sievers10fbcf42011-12-21 14:48:43 -08002032 * hugetlb module exit: unregister hstate attributes from node devices
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002033 * that have them.
2034 */
2035static void hugetlb_unregister_all_nodes(void)
2036{
2037 int nid;
2038
2039 /*
Kay Sievers10fbcf42011-12-21 14:48:43 -08002040 * disable node device registrations.
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002041 */
2042 register_hugetlbfs_with_node(NULL, NULL);
2043
2044 /*
2045 * remove hstate attributes from any nodes that have them.
2046 */
2047 for (nid = 0; nid < nr_node_ids; nid++)
Wen Congyang87327942012-12-11 16:00:56 -08002048 hugetlb_unregister_node(node_devices[nid]);
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002049}
2050
2051/*
Kay Sievers10fbcf42011-12-21 14:48:43 -08002052 * Register hstate attributes for a single node device.
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002053 * No-op if attributes already registered.
2054 */
Claudiu Ghioc3cd8b442013-03-04 12:46:15 +02002055static void hugetlb_register_node(struct node *node)
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002056{
2057 struct hstate *h;
Kay Sievers10fbcf42011-12-21 14:48:43 -08002058 struct node_hstate *nhs = &node_hstates[node->dev.id];
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002059 int err;
2060
2061 if (nhs->hugepages_kobj)
2062 return; /* already allocated */
2063
2064 nhs->hugepages_kobj = kobject_create_and_add("hugepages",
Kay Sievers10fbcf42011-12-21 14:48:43 -08002065 &node->dev.kobj);
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002066 if (!nhs->hugepages_kobj)
2067 return;
2068
2069 for_each_hstate(h) {
2070 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
2071 nhs->hstate_kobjs,
2072 &per_node_hstate_attr_group);
2073 if (err) {
Andrew Mortonffb22af2013-02-22 16:32:08 -08002074 pr_err("Hugetlb: Unable to add hstate %s for node %d\n",
2075 h->name, node->dev.id);
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002076 hugetlb_unregister_node(node);
2077 break;
2078 }
2079 }
2080}
2081
2082/*
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -08002083 * hugetlb init time: register hstate attributes for all registered node
Kay Sievers10fbcf42011-12-21 14:48:43 -08002084 * devices of nodes that have memory. All on-line nodes should have
2085 * registered their associated device by this time.
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002086 */
Luiz Capitulino7d9ca002014-12-12 16:55:24 -08002087static void __init hugetlb_register_all_nodes(void)
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002088{
2089 int nid;
2090
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08002091 for_each_node_state(nid, N_MEMORY) {
Wen Congyang87327942012-12-11 16:00:56 -08002092 struct node *node = node_devices[nid];
Kay Sievers10fbcf42011-12-21 14:48:43 -08002093 if (node->dev.id == nid)
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002094 hugetlb_register_node(node);
2095 }
2096
2097 /*
Kay Sievers10fbcf42011-12-21 14:48:43 -08002098 * Let the node device driver know we're here so it can
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002099 * [un]register hstate attributes on node hotplug.
2100 */
2101 register_hugetlbfs_with_node(hugetlb_register_node,
2102 hugetlb_unregister_node);
2103}
2104#else /* !CONFIG_NUMA */
2105
2106static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
2107{
2108 BUG();
2109 if (nidp)
2110 *nidp = -1;
2111 return NULL;
2112}
2113
2114static void hugetlb_unregister_all_nodes(void) { }
2115
2116static void hugetlb_register_all_nodes(void) { }
2117
2118#endif
2119
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002120static void __exit hugetlb_exit(void)
2121{
2122 struct hstate *h;
2123
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002124 hugetlb_unregister_all_nodes();
2125
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002126 for_each_hstate(h) {
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07002127 kobject_put(hstate_kobjs[hstate_index(h)]);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002128 }
2129
2130 kobject_put(hugepages_kobj);
Davidlohr Bueso8382d912014-04-03 14:47:31 -07002131 kfree(htlb_fault_mutex_table);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002132}
2133module_exit(hugetlb_exit);
2134
2135static int __init hugetlb_init(void)
2136{
Davidlohr Bueso8382d912014-04-03 14:47:31 -07002137 int i;
2138
Nishanth Aravamudan457c1b22014-05-06 12:50:00 -07002139 if (!hugepages_supported())
Benjamin Herrenschmidt0ef89d22008-07-31 00:07:30 -07002140 return 0;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002141
Nick Piggine11bfbf2008-07-23 21:27:52 -07002142 if (!size_to_hstate(default_hstate_size)) {
2143 default_hstate_size = HPAGE_SIZE;
2144 if (!size_to_hstate(default_hstate_size))
2145 hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002146 }
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07002147 default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size));
Nick Piggine11bfbf2008-07-23 21:27:52 -07002148 if (default_hstate_max_huge_pages)
2149 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002150
2151 hugetlb_init_hstates();
Andi Kleenaa888a72008-07-23 21:27:47 -07002152 gather_bootmem_prealloc();
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002153 report_hugepages();
2154
2155 hugetlb_sysfs_init();
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002156 hugetlb_register_all_nodes();
Jianguo Wu7179e7b2012-12-18 14:23:19 -08002157 hugetlb_cgroup_file_init();
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002158
Davidlohr Bueso8382d912014-04-03 14:47:31 -07002159#ifdef CONFIG_SMP
2160 num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus());
2161#else
2162 num_fault_mutexes = 1;
2163#endif
2164 htlb_fault_mutex_table =
2165 kmalloc(sizeof(struct mutex) * num_fault_mutexes, GFP_KERNEL);
2166 BUG_ON(!htlb_fault_mutex_table);
2167
2168 for (i = 0; i < num_fault_mutexes; i++)
2169 mutex_init(&htlb_fault_mutex_table[i]);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002170 return 0;
2171}
2172module_init(hugetlb_init);
2173
2174/* Should be called on processing a hugepagesz=... option */
2175void __init hugetlb_add_hstate(unsigned order)
2176{
2177 struct hstate *h;
Andi Kleen8faa8b02008-07-23 21:27:48 -07002178 unsigned long i;
2179
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002180 if (size_to_hstate(PAGE_SIZE << order)) {
Andrew Mortonffb22af2013-02-22 16:32:08 -08002181 pr_warning("hugepagesz= specified twice, ignoring\n");
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002182 return;
2183 }
Aneesh Kumar K.V47d38342012-07-31 16:41:54 -07002184 BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002185 BUG_ON(order == 0);
Aneesh Kumar K.V47d38342012-07-31 16:41:54 -07002186 h = &hstates[hugetlb_max_hstate++];
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002187 h->order = order;
2188 h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
Andi Kleen8faa8b02008-07-23 21:27:48 -07002189 h->nr_huge_pages = 0;
2190 h->free_huge_pages = 0;
2191 for (i = 0; i < MAX_NUMNODES; ++i)
2192 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
Aneesh Kumar K.V0edaecf2012-07-31 16:42:07 -07002193 INIT_LIST_HEAD(&h->hugepage_activelist);
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08002194 h->next_nid_to_alloc = first_node(node_states[N_MEMORY]);
2195 h->next_nid_to_free = first_node(node_states[N_MEMORY]);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002196 snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
2197 huge_page_size(h)/1024);
Andi Kleen8faa8b02008-07-23 21:27:48 -07002198
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002199 parsed_hstate = h;
2200}
2201
Nick Piggine11bfbf2008-07-23 21:27:52 -07002202static int __init hugetlb_nrpages_setup(char *s)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002203{
2204 unsigned long *mhp;
Andi Kleen8faa8b02008-07-23 21:27:48 -07002205 static unsigned long *last_mhp;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002206
2207 /*
Aneesh Kumar K.V47d38342012-07-31 16:41:54 -07002208 * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet,
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002209 * so this hugepages= parameter goes to the "default hstate".
2210 */
Aneesh Kumar K.V47d38342012-07-31 16:41:54 -07002211 if (!hugetlb_max_hstate)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002212 mhp = &default_hstate_max_huge_pages;
2213 else
2214 mhp = &parsed_hstate->max_huge_pages;
2215
Andi Kleen8faa8b02008-07-23 21:27:48 -07002216 if (mhp == last_mhp) {
Andrew Mortonffb22af2013-02-22 16:32:08 -08002217 pr_warning("hugepages= specified twice without "
2218 "interleaving hugepagesz=, ignoring\n");
Andi Kleen8faa8b02008-07-23 21:27:48 -07002219 return 1;
2220 }
2221
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002222 if (sscanf(s, "%lu", mhp) <= 0)
2223 *mhp = 0;
2224
Andi Kleen8faa8b02008-07-23 21:27:48 -07002225 /*
2226 * Global state is always initialized later in hugetlb_init.
2227 * But we need to allocate >= MAX_ORDER hstates here early to still
2228 * use the bootmem allocator.
2229 */
Aneesh Kumar K.V47d38342012-07-31 16:41:54 -07002230 if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER)
Andi Kleen8faa8b02008-07-23 21:27:48 -07002231 hugetlb_hstate_alloc_pages(parsed_hstate);
2232
2233 last_mhp = mhp;
2234
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002235 return 1;
2236}
Nick Piggine11bfbf2008-07-23 21:27:52 -07002237__setup("hugepages=", hugetlb_nrpages_setup);
2238
2239static int __init hugetlb_default_setup(char *s)
2240{
2241 default_hstate_size = memparse(s, &s);
2242 return 1;
2243}
2244__setup("default_hugepagesz=", hugetlb_default_setup);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002245
Nishanth Aravamudan8a213462008-07-25 19:44:37 -07002246static unsigned int cpuset_mems_nr(unsigned int *array)
2247{
2248 int node;
2249 unsigned int nr = 0;
2250
2251 for_each_node_mask(node, cpuset_current_mems_allowed)
2252 nr += array[node];
2253
2254 return nr;
2255}
2256
2257#ifdef CONFIG_SYSCTL
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002258static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
2259 struct ctl_table *table, int write,
2260 void __user *buffer, size_t *length, loff_t *ppos)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002261{
Andi Kleene5ff2152008-07-23 21:27:42 -07002262 struct hstate *h = &default_hstate;
David Rientjes238d3c12014-08-06 16:06:51 -07002263 unsigned long tmp = h->max_huge_pages;
Michal Hocko08d4a242011-01-13 15:47:26 -08002264 int ret;
Andi Kleene5ff2152008-07-23 21:27:42 -07002265
Nishanth Aravamudan457c1b22014-05-06 12:50:00 -07002266 if (!hugepages_supported())
2267 return -ENOTSUPP;
2268
Andi Kleene5ff2152008-07-23 21:27:42 -07002269 table->data = &tmp;
2270 table->maxlen = sizeof(unsigned long);
Michal Hocko08d4a242011-01-13 15:47:26 -08002271 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2272 if (ret)
2273 goto out;
Andi Kleene5ff2152008-07-23 21:27:42 -07002274
David Rientjes238d3c12014-08-06 16:06:51 -07002275 if (write)
2276 ret = __nr_hugepages_store_common(obey_mempolicy, h,
2277 NUMA_NO_NODE, tmp, *length);
Michal Hocko08d4a242011-01-13 15:47:26 -08002278out:
2279 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002280}
Mel Gorman396faf02007-07-17 04:03:13 -07002281
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002282int hugetlb_sysctl_handler(struct ctl_table *table, int write,
2283 void __user *buffer, size_t *length, loff_t *ppos)
2284{
2285
2286 return hugetlb_sysctl_handler_common(false, table, write,
2287 buffer, length, ppos);
2288}
2289
2290#ifdef CONFIG_NUMA
2291int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
2292 void __user *buffer, size_t *length, loff_t *ppos)
2293{
2294 return hugetlb_sysctl_handler_common(true, table, write,
2295 buffer, length, ppos);
2296}
2297#endif /* CONFIG_NUMA */
2298
Nishanth Aravamudana3d0c6a2008-02-08 04:18:18 -08002299int hugetlb_overcommit_handler(struct ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07002300 void __user *buffer,
Nishanth Aravamudana3d0c6a2008-02-08 04:18:18 -08002301 size_t *length, loff_t *ppos)
2302{
Andi Kleena5516432008-07-23 21:27:41 -07002303 struct hstate *h = &default_hstate;
Andi Kleene5ff2152008-07-23 21:27:42 -07002304 unsigned long tmp;
Michal Hocko08d4a242011-01-13 15:47:26 -08002305 int ret;
Andi Kleene5ff2152008-07-23 21:27:42 -07002306
Nishanth Aravamudan457c1b22014-05-06 12:50:00 -07002307 if (!hugepages_supported())
2308 return -ENOTSUPP;
2309
Petr Holasekc033a932011-03-22 16:33:05 -07002310 tmp = h->nr_overcommit_huge_pages;
Andi Kleene5ff2152008-07-23 21:27:42 -07002311
Luiz Capitulinobae7f4a2014-06-04 16:07:08 -07002312 if (write && hstate_is_gigantic(h))
Eric B Munsonadbe8722011-01-13 15:47:27 -08002313 return -EINVAL;
2314
Andi Kleene5ff2152008-07-23 21:27:42 -07002315 table->data = &tmp;
2316 table->maxlen = sizeof(unsigned long);
Michal Hocko08d4a242011-01-13 15:47:26 -08002317 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2318 if (ret)
2319 goto out;
Andi Kleene5ff2152008-07-23 21:27:42 -07002320
2321 if (write) {
2322 spin_lock(&hugetlb_lock);
2323 h->nr_overcommit_huge_pages = tmp;
2324 spin_unlock(&hugetlb_lock);
2325 }
Michal Hocko08d4a242011-01-13 15:47:26 -08002326out:
2327 return ret;
Nishanth Aravamudana3d0c6a2008-02-08 04:18:18 -08002328}
2329
Linus Torvalds1da177e2005-04-16 15:20:36 -07002330#endif /* CONFIG_SYSCTL */
2331
Alexey Dobriyane1759c22008-10-15 23:50:22 +04002332void hugetlb_report_meminfo(struct seq_file *m)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002333{
Andi Kleena5516432008-07-23 21:27:41 -07002334 struct hstate *h = &default_hstate;
Nishanth Aravamudan457c1b22014-05-06 12:50:00 -07002335 if (!hugepages_supported())
2336 return;
Alexey Dobriyane1759c22008-10-15 23:50:22 +04002337 seq_printf(m,
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002338 "HugePages_Total: %5lu\n"
2339 "HugePages_Free: %5lu\n"
2340 "HugePages_Rsvd: %5lu\n"
2341 "HugePages_Surp: %5lu\n"
2342 "Hugepagesize: %8lu kB\n",
Andi Kleena5516432008-07-23 21:27:41 -07002343 h->nr_huge_pages,
2344 h->free_huge_pages,
2345 h->resv_huge_pages,
2346 h->surplus_huge_pages,
2347 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002348}
2349
2350int hugetlb_report_node_meminfo(int nid, char *buf)
2351{
Andi Kleena5516432008-07-23 21:27:41 -07002352 struct hstate *h = &default_hstate;
Nishanth Aravamudan457c1b22014-05-06 12:50:00 -07002353 if (!hugepages_supported())
2354 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002355 return sprintf(buf,
2356 "Node %d HugePages_Total: %5u\n"
Nishanth Aravamudana1de0912008-03-26 14:37:53 -07002357 "Node %d HugePages_Free: %5u\n"
2358 "Node %d HugePages_Surp: %5u\n",
Andi Kleena5516432008-07-23 21:27:41 -07002359 nid, h->nr_huge_pages_node[nid],
2360 nid, h->free_huge_pages_node[nid],
2361 nid, h->surplus_huge_pages_node[nid]);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002362}
2363
David Rientjes949f7ec2013-04-29 15:07:48 -07002364void hugetlb_show_meminfo(void)
2365{
2366 struct hstate *h;
2367 int nid;
2368
Nishanth Aravamudan457c1b22014-05-06 12:50:00 -07002369 if (!hugepages_supported())
2370 return;
2371
David Rientjes949f7ec2013-04-29 15:07:48 -07002372 for_each_node_state(nid, N_MEMORY)
2373 for_each_hstate(h)
2374 pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
2375 nid,
2376 h->nr_huge_pages_node[nid],
2377 h->free_huge_pages_node[nid],
2378 h->surplus_huge_pages_node[nid],
2379 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
2380}
2381
Linus Torvalds1da177e2005-04-16 15:20:36 -07002382/* Return the number pages of memory we physically have, in PAGE_SIZE units. */
2383unsigned long hugetlb_total_pages(void)
2384{
Wanpeng Lid0028582013-03-22 15:04:40 -07002385 struct hstate *h;
2386 unsigned long nr_total_pages = 0;
2387
2388 for_each_hstate(h)
2389 nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h);
2390 return nr_total_pages;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002391}
Linus Torvalds1da177e2005-04-16 15:20:36 -07002392
Andi Kleena5516432008-07-23 21:27:41 -07002393static int hugetlb_acct_memory(struct hstate *h, long delta)
Mel Gormanfc1b8a72008-07-23 21:27:22 -07002394{
2395 int ret = -ENOMEM;
2396
2397 spin_lock(&hugetlb_lock);
2398 /*
2399 * When cpuset is configured, it breaks the strict hugetlb page
2400 * reservation as the accounting is done on a global variable. Such
2401 * reservation is completely rubbish in the presence of cpuset because
2402 * the reservation is not checked against page availability for the
2403 * current cpuset. Application can still potentially OOM'ed by kernel
2404 * with lack of free htlb page in cpuset that the task is in.
2405 * Attempt to enforce strict accounting with cpuset is almost
2406 * impossible (or too ugly) because cpuset is too fluid that
2407 * task or memory node can be dynamically moved between cpusets.
2408 *
2409 * The change of semantics for shared hugetlb mapping with cpuset is
2410 * undesirable. However, in order to preserve some of the semantics,
2411 * we fall back to check against current free page availability as
2412 * a best attempt and hopefully to minimize the impact of changing
2413 * semantics that cpuset has.
2414 */
2415 if (delta > 0) {
Andi Kleena5516432008-07-23 21:27:41 -07002416 if (gather_surplus_pages(h, delta) < 0)
Mel Gormanfc1b8a72008-07-23 21:27:22 -07002417 goto out;
2418
Andi Kleena5516432008-07-23 21:27:41 -07002419 if (delta > cpuset_mems_nr(h->free_huge_pages_node)) {
2420 return_unused_surplus_pages(h, delta);
Mel Gormanfc1b8a72008-07-23 21:27:22 -07002421 goto out;
2422 }
2423 }
2424
2425 ret = 0;
2426 if (delta < 0)
Andi Kleena5516432008-07-23 21:27:41 -07002427 return_unused_surplus_pages(h, (unsigned long) -delta);
Mel Gormanfc1b8a72008-07-23 21:27:22 -07002428
2429out:
2430 spin_unlock(&hugetlb_lock);
2431 return ret;
2432}
2433
Andy Whitcroft84afd992008-07-23 21:27:32 -07002434static void hugetlb_vm_op_open(struct vm_area_struct *vma)
2435{
Joonsoo Kimf522c3a2013-09-11 14:21:53 -07002436 struct resv_map *resv = vma_resv_map(vma);
Andy Whitcroft84afd992008-07-23 21:27:32 -07002437
2438 /*
2439 * This new VMA should share its siblings reservation map if present.
2440 * The VMA will only ever have a valid reservation map pointer where
2441 * it is being copied for another still existing VMA. As that VMA
Lucas De Marchi25985ed2011-03-30 22:57:33 -03002442 * has a reference to the reservation map it cannot disappear until
Andy Whitcroft84afd992008-07-23 21:27:32 -07002443 * after this open call completes. It is therefore safe to take a
2444 * new reference here without additional locking.
2445 */
Joonsoo Kim4e35f482014-04-03 14:47:30 -07002446 if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
Joonsoo Kimf522c3a2013-09-11 14:21:53 -07002447 kref_get(&resv->refs);
Andy Whitcroft84afd992008-07-23 21:27:32 -07002448}
2449
Mel Gormana1e78772008-07-23 21:27:23 -07002450static void hugetlb_vm_op_close(struct vm_area_struct *vma)
2451{
Andi Kleena5516432008-07-23 21:27:41 -07002452 struct hstate *h = hstate_vma(vma);
Joonsoo Kimf522c3a2013-09-11 14:21:53 -07002453 struct resv_map *resv = vma_resv_map(vma);
David Gibson90481622012-03-21 16:34:12 -07002454 struct hugepage_subpool *spool = subpool_vma(vma);
Joonsoo Kim4e35f482014-04-03 14:47:30 -07002455 unsigned long reserve, start, end;
Andy Whitcroft84afd992008-07-23 21:27:32 -07002456
Joonsoo Kim4e35f482014-04-03 14:47:30 -07002457 if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER))
2458 return;
Andy Whitcroft84afd992008-07-23 21:27:32 -07002459
Joonsoo Kim4e35f482014-04-03 14:47:30 -07002460 start = vma_hugecache_offset(h, vma, vma->vm_start);
2461 end = vma_hugecache_offset(h, vma, vma->vm_end);
Andy Whitcroft84afd992008-07-23 21:27:32 -07002462
Joonsoo Kim4e35f482014-04-03 14:47:30 -07002463 reserve = (end - start) - region_count(resv, start, end);
Andy Whitcroft84afd992008-07-23 21:27:32 -07002464
Joonsoo Kim4e35f482014-04-03 14:47:30 -07002465 kref_put(&resv->refs, resv_map_release);
2466
2467 if (reserve) {
2468 hugetlb_acct_memory(h, -reserve);
2469 hugepage_subpool_put_pages(spool, reserve);
Andy Whitcroft84afd992008-07-23 21:27:32 -07002470 }
Mel Gormana1e78772008-07-23 21:27:23 -07002471}
2472
Linus Torvalds1da177e2005-04-16 15:20:36 -07002473/*
2474 * We cannot handle pagefaults against hugetlb pages at all. They cause
2475 * handle_mm_fault() to try to instantiate regular-sized pages in the
2476 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
2477 * this far.
2478 */
Nick Piggind0217ac2007-07-19 01:47:03 -07002479static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002480{
2481 BUG();
Nick Piggind0217ac2007-07-19 01:47:03 -07002482 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002483}
2484
Alexey Dobriyanf0f37e2f2009-09-27 22:29:37 +04002485const struct vm_operations_struct hugetlb_vm_ops = {
Nick Piggind0217ac2007-07-19 01:47:03 -07002486 .fault = hugetlb_vm_op_fault,
Andy Whitcroft84afd992008-07-23 21:27:32 -07002487 .open = hugetlb_vm_op_open,
Mel Gormana1e78772008-07-23 21:27:23 -07002488 .close = hugetlb_vm_op_close,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002489};
2490
David Gibson1e8f8892006-01-06 00:10:44 -08002491static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
2492 int writable)
David Gibson63551ae2005-06-21 17:14:44 -07002493{
2494 pte_t entry;
2495
David Gibson1e8f8892006-01-06 00:10:44 -08002496 if (writable) {
Gerald Schaefer106c9922013-04-29 15:07:23 -07002497 entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page,
2498 vma->vm_page_prot)));
David Gibson63551ae2005-06-21 17:14:44 -07002499 } else {
Gerald Schaefer106c9922013-04-29 15:07:23 -07002500 entry = huge_pte_wrprotect(mk_huge_pte(page,
2501 vma->vm_page_prot));
David Gibson63551ae2005-06-21 17:14:44 -07002502 }
2503 entry = pte_mkyoung(entry);
2504 entry = pte_mkhuge(entry);
Chris Metcalfd9ed9fa2012-04-01 14:01:34 -04002505 entry = arch_make_huge_pte(entry, vma, page, writable);
David Gibson63551ae2005-06-21 17:14:44 -07002506
2507 return entry;
2508}
2509
David Gibson1e8f8892006-01-06 00:10:44 -08002510static void set_huge_ptep_writable(struct vm_area_struct *vma,
2511 unsigned long address, pte_t *ptep)
2512{
2513 pte_t entry;
2514
Gerald Schaefer106c9922013-04-29 15:07:23 -07002515 entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep)));
Chris Forbes32f84522011-07-25 17:12:14 -07002516 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
Russell King4b3073e2009-12-18 16:40:18 +00002517 update_mmu_cache(vma, address, ptep);
David Gibson1e8f8892006-01-06 00:10:44 -08002518}
2519
Naoya Horiguchi4a705fe2014-06-23 13:22:03 -07002520static int is_hugetlb_entry_migration(pte_t pte)
2521{
2522 swp_entry_t swp;
2523
2524 if (huge_pte_none(pte) || pte_present(pte))
2525 return 0;
2526 swp = pte_to_swp_entry(pte);
2527 if (non_swap_entry(swp) && is_migration_entry(swp))
2528 return 1;
2529 else
2530 return 0;
2531}
2532
2533static int is_hugetlb_entry_hwpoisoned(pte_t pte)
2534{
2535 swp_entry_t swp;
2536
2537 if (huge_pte_none(pte) || pte_present(pte))
2538 return 0;
2539 swp = pte_to_swp_entry(pte);
2540 if (non_swap_entry(swp) && is_hwpoison_entry(swp))
2541 return 1;
2542 else
2543 return 0;
2544}
David Gibson1e8f8892006-01-06 00:10:44 -08002545
David Gibson63551ae2005-06-21 17:14:44 -07002546int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
2547 struct vm_area_struct *vma)
2548{
2549 pte_t *src_pte, *dst_pte, entry;
2550 struct page *ptepage;
Hugh Dickins1c598272005-10-19 21:23:43 -07002551 unsigned long addr;
David Gibson1e8f8892006-01-06 00:10:44 -08002552 int cow;
Andi Kleena5516432008-07-23 21:27:41 -07002553 struct hstate *h = hstate_vma(vma);
2554 unsigned long sz = huge_page_size(h);
Andreas Sandberge8569dd2014-01-21 15:49:09 -08002555 unsigned long mmun_start; /* For mmu_notifiers */
2556 unsigned long mmun_end; /* For mmu_notifiers */
2557 int ret = 0;
David Gibson1e8f8892006-01-06 00:10:44 -08002558
2559 cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
David Gibson63551ae2005-06-21 17:14:44 -07002560
Andreas Sandberge8569dd2014-01-21 15:49:09 -08002561 mmun_start = vma->vm_start;
2562 mmun_end = vma->vm_end;
2563 if (cow)
2564 mmu_notifier_invalidate_range_start(src, mmun_start, mmun_end);
2565
Andi Kleena5516432008-07-23 21:27:41 -07002566 for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002567 spinlock_t *src_ptl, *dst_ptl;
Hugh Dickinsc74df322005-10-29 18:16:23 -07002568 src_pte = huge_pte_offset(src, addr);
2569 if (!src_pte)
2570 continue;
Andi Kleena5516432008-07-23 21:27:41 -07002571 dst_pte = huge_pte_alloc(dst, addr, sz);
Andreas Sandberge8569dd2014-01-21 15:49:09 -08002572 if (!dst_pte) {
2573 ret = -ENOMEM;
2574 break;
2575 }
Larry Woodmanc5c99422008-01-24 05:49:25 -08002576
2577 /* If the pagetables are shared don't copy or take references */
2578 if (dst_pte == src_pte)
2579 continue;
2580
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002581 dst_ptl = huge_pte_lock(h, dst, dst_pte);
2582 src_ptl = huge_pte_lockptr(h, src, src_pte);
2583 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
Naoya Horiguchi4a705fe2014-06-23 13:22:03 -07002584 entry = huge_ptep_get(src_pte);
2585 if (huge_pte_none(entry)) { /* skip none entry */
2586 ;
2587 } else if (unlikely(is_hugetlb_entry_migration(entry) ||
2588 is_hugetlb_entry_hwpoisoned(entry))) {
2589 swp_entry_t swp_entry = pte_to_swp_entry(entry);
2590
2591 if (is_write_migration_entry(swp_entry) && cow) {
2592 /*
2593 * COW mappings require pages in both
2594 * parent and child to be set to read.
2595 */
2596 make_migration_entry_read(&swp_entry);
2597 entry = swp_entry_to_pte(swp_entry);
2598 set_huge_pte_at(src, addr, src_pte, entry);
2599 }
2600 set_huge_pte_at(dst, addr, dst_pte, entry);
2601 } else {
Joerg Roedel34ee6452014-11-13 13:46:09 +11002602 if (cow) {
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07002603 huge_ptep_set_wrprotect(src, addr, src_pte);
Joerg Roedel34ee6452014-11-13 13:46:09 +11002604 mmu_notifier_invalidate_range(src, mmun_start,
2605 mmun_end);
2606 }
Naoya Horiguchi0253d632014-07-23 14:00:19 -07002607 entry = huge_ptep_get(src_pte);
Hugh Dickins1c598272005-10-19 21:23:43 -07002608 ptepage = pte_page(entry);
2609 get_page(ptepage);
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002610 page_dup_rmap(ptepage);
Hugh Dickins1c598272005-10-19 21:23:43 -07002611 set_huge_pte_at(dst, addr, dst_pte, entry);
2612 }
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002613 spin_unlock(src_ptl);
2614 spin_unlock(dst_ptl);
David Gibson63551ae2005-06-21 17:14:44 -07002615 }
David Gibson63551ae2005-06-21 17:14:44 -07002616
Andreas Sandberge8569dd2014-01-21 15:49:09 -08002617 if (cow)
2618 mmu_notifier_invalidate_range_end(src, mmun_start, mmun_end);
2619
2620 return ret;
David Gibson63551ae2005-06-21 17:14:44 -07002621}
2622
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07002623void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
2624 unsigned long start, unsigned long end,
2625 struct page *ref_page)
David Gibson63551ae2005-06-21 17:14:44 -07002626{
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07002627 int force_flush = 0;
David Gibson63551ae2005-06-21 17:14:44 -07002628 struct mm_struct *mm = vma->vm_mm;
2629 unsigned long address;
David Gibsonc7546f82005-08-05 11:59:35 -07002630 pte_t *ptep;
David Gibson63551ae2005-06-21 17:14:44 -07002631 pte_t pte;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002632 spinlock_t *ptl;
David Gibson63551ae2005-06-21 17:14:44 -07002633 struct page *page;
Andi Kleena5516432008-07-23 21:27:41 -07002634 struct hstate *h = hstate_vma(vma);
2635 unsigned long sz = huge_page_size(h);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07002636 const unsigned long mmun_start = start; /* For mmu_notifiers */
2637 const unsigned long mmun_end = end; /* For mmu_notifiers */
Andi Kleena5516432008-07-23 21:27:41 -07002638
David Gibson63551ae2005-06-21 17:14:44 -07002639 WARN_ON(!is_vm_hugetlb_page(vma));
Andi Kleena5516432008-07-23 21:27:41 -07002640 BUG_ON(start & ~huge_page_mask(h));
2641 BUG_ON(end & ~huge_page_mask(h));
David Gibson63551ae2005-06-21 17:14:44 -07002642
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07002643 tlb_start_vma(tlb, vma);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07002644 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
Hillf Danton569f48b82014-12-10 15:44:41 -08002645 address = start;
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07002646again:
Hillf Danton569f48b82014-12-10 15:44:41 -08002647 for (; address < end; address += sz) {
David Gibsonc7546f82005-08-05 11:59:35 -07002648 ptep = huge_pte_offset(mm, address);
Adam Litke4c887262005-10-29 18:16:46 -07002649 if (!ptep)
David Gibsonc7546f82005-08-05 11:59:35 -07002650 continue;
2651
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002652 ptl = huge_pte_lock(h, mm, ptep);
Chen, Kenneth W39dde652006-12-06 20:32:03 -08002653 if (huge_pmd_unshare(mm, &address, ptep))
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002654 goto unlock;
Chen, Kenneth W39dde652006-12-06 20:32:03 -08002655
Hillf Danton66293262012-03-23 15:01:48 -07002656 pte = huge_ptep_get(ptep);
2657 if (huge_pte_none(pte))
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002658 goto unlock;
Hillf Danton66293262012-03-23 15:01:48 -07002659
2660 /*
Naoya Horiguchi9fbc1f62015-02-11 15:25:32 -08002661 * Migrating hugepage or HWPoisoned hugepage is already
2662 * unmapped and its refcount is dropped, so just clear pte here.
Hillf Danton66293262012-03-23 15:01:48 -07002663 */
Naoya Horiguchi9fbc1f62015-02-11 15:25:32 -08002664 if (unlikely(!pte_present(pte))) {
Gerald Schaefer106c9922013-04-29 15:07:23 -07002665 huge_pte_clear(mm, address, ptep);
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002666 goto unlock;
Naoya Horiguchi8c4894c2012-12-12 13:52:28 -08002667 }
Hillf Danton66293262012-03-23 15:01:48 -07002668
2669 page = pte_page(pte);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002670 /*
2671 * If a reference page is supplied, it is because a specific
2672 * page is being unmapped, not a range. Ensure the page we
2673 * are about to unmap is the actual page of interest.
2674 */
2675 if (ref_page) {
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002676 if (page != ref_page)
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002677 goto unlock;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002678
2679 /*
2680 * Mark the VMA as having unmapped its page so that
2681 * future faults in this VMA will fail rather than
2682 * looking like data was lost
2683 */
2684 set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
2685 }
2686
David Gibsonc7546f82005-08-05 11:59:35 -07002687 pte = huge_ptep_get_and_clear(mm, address, ptep);
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07002688 tlb_remove_tlb_entry(tlb, ptep, address);
Gerald Schaefer106c9922013-04-29 15:07:23 -07002689 if (huge_pte_dirty(pte))
Ken Chen6649a382007-02-08 14:20:27 -08002690 set_page_dirty(page);
Hillf Danton9e811302012-03-21 16:34:03 -07002691
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07002692 page_remove_rmap(page);
2693 force_flush = !__tlb_remove_page(tlb, page);
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002694 if (force_flush) {
Hillf Danton569f48b82014-12-10 15:44:41 -08002695 address += sz;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002696 spin_unlock(ptl);
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07002697 break;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002698 }
Hillf Danton9e811302012-03-21 16:34:03 -07002699 /* Bail out after unmapping reference page if supplied */
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002700 if (ref_page) {
2701 spin_unlock(ptl);
Hillf Danton9e811302012-03-21 16:34:03 -07002702 break;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002703 }
2704unlock:
2705 spin_unlock(ptl);
David Gibson63551ae2005-06-21 17:14:44 -07002706 }
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07002707 /*
2708 * mmu_gather ran out of room to batch pages, we break out of
2709 * the PTE lock to avoid doing the potential expensive TLB invalidate
2710 * and page-free while holding it.
2711 */
2712 if (force_flush) {
2713 force_flush = 0;
2714 tlb_flush_mmu(tlb);
2715 if (address < end && !ref_page)
2716 goto again;
Chen, Kenneth Wfe1668a2006-10-04 02:15:24 -07002717 }
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07002718 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07002719 tlb_end_vma(tlb, vma);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002720}
David Gibson63551ae2005-06-21 17:14:44 -07002721
Mel Gormand8333522012-07-31 16:46:20 -07002722void __unmap_hugepage_range_final(struct mmu_gather *tlb,
2723 struct vm_area_struct *vma, unsigned long start,
2724 unsigned long end, struct page *ref_page)
2725{
2726 __unmap_hugepage_range(tlb, vma, start, end, ref_page);
2727
2728 /*
2729 * Clear this flag so that x86's huge_pmd_share page_table_shareable
2730 * test will fail on a vma being torn down, and not grab a page table
2731 * on its way out. We're lucky that the flag has such an appropriate
2732 * name, and can in fact be safely cleared here. We could clear it
2733 * before the __unmap_hugepage_range above, but all that's necessary
Davidlohr Buesoc8c06ef2014-12-12 16:54:24 -08002734 * is to clear it before releasing the i_mmap_rwsem. This works
Mel Gormand8333522012-07-31 16:46:20 -07002735 * because in the context this is called, the VMA is about to be
Davidlohr Buesoc8c06ef2014-12-12 16:54:24 -08002736 * destroyed and the i_mmap_rwsem is held.
Mel Gormand8333522012-07-31 16:46:20 -07002737 */
2738 vma->vm_flags &= ~VM_MAYSHARE;
2739}
2740
Chen, Kenneth W502717f2006-10-11 01:20:46 -07002741void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002742 unsigned long end, struct page *ref_page)
Chen, Kenneth W502717f2006-10-11 01:20:46 -07002743{
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07002744 struct mm_struct *mm;
2745 struct mmu_gather tlb;
2746
2747 mm = vma->vm_mm;
2748
Linus Torvalds2b047252013-08-15 11:42:25 -07002749 tlb_gather_mmu(&tlb, mm, start, end);
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07002750 __unmap_hugepage_range(&tlb, vma, start, end, ref_page);
2751 tlb_finish_mmu(&tlb, start, end);
Chen, Kenneth W502717f2006-10-11 01:20:46 -07002752}
2753
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002754/*
2755 * This is called when the original mapper is failing to COW a MAP_PRIVATE
2756 * mappping it owns the reserve page for. The intention is to unmap the page
2757 * from other VMAs and let the children be SIGKILLed if they are faulting the
2758 * same region.
2759 */
Davidlohr Bueso2f4612a2014-08-06 16:06:45 -07002760static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
2761 struct page *page, unsigned long address)
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002762{
Adam Litke75266742008-11-12 13:24:56 -08002763 struct hstate *h = hstate_vma(vma);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002764 struct vm_area_struct *iter_vma;
2765 struct address_space *mapping;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002766 pgoff_t pgoff;
2767
2768 /*
2769 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
2770 * from page cache lookup which is in HPAGE_SIZE units.
2771 */
Adam Litke75266742008-11-12 13:24:56 -08002772 address = address & huge_page_mask(h);
Michal Hocko36e4f202012-10-08 16:33:31 -07002773 pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) +
2774 vma->vm_pgoff;
Al Viro496ad9a2013-01-23 17:07:38 -05002775 mapping = file_inode(vma->vm_file)->i_mapping;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002776
Mel Gorman4eb2b1d2009-12-14 17:59:53 -08002777 /*
2778 * Take the mapping lock for the duration of the table walk. As
2779 * this mapping should be shared between all the VMAs,
2780 * __unmap_hugepage_range() is called as the lock is already held
2781 */
Davidlohr Bueso83cde9e2014-12-12 16:54:21 -08002782 i_mmap_lock_write(mapping);
Michel Lespinasse6b2dbba2012-10-08 16:31:25 -07002783 vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) {
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002784 /* Do not unmap the current VMA */
2785 if (iter_vma == vma)
2786 continue;
2787
2788 /*
2789 * Unmap the page from other VMAs without their own reserves.
2790 * They get marked to be SIGKILLed if they fault in these
2791 * areas. This is because a future no-page fault on this VMA
2792 * could insert a zeroed page instead of the data existing
2793 * from the time of fork. This would look like data corruption
2794 */
2795 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07002796 unmap_hugepage_range(iter_vma, address,
2797 address + huge_page_size(h), page);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002798 }
Davidlohr Bueso83cde9e2014-12-12 16:54:21 -08002799 i_mmap_unlock_write(mapping);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002800}
2801
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002802/*
2803 * Hugetlb_cow() should be called with page lock of the original hugepage held.
Michal Hockoef009b22012-01-10 15:07:21 -08002804 * Called with hugetlb_instantiation_mutex held and pte_page locked so we
2805 * cannot race with other handlers or page migration.
2806 * Keep the pte_same checks anyway to make transition from the mutex easier.
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002807 */
David Gibson1e8f8892006-01-06 00:10:44 -08002808static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002809 unsigned long address, pte_t *ptep, pte_t pte,
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002810 struct page *pagecache_page, spinlock_t *ptl)
David Gibson1e8f8892006-01-06 00:10:44 -08002811{
Andi Kleena5516432008-07-23 21:27:41 -07002812 struct hstate *h = hstate_vma(vma);
David Gibson1e8f8892006-01-06 00:10:44 -08002813 struct page *old_page, *new_page;
Davidlohr Buesoad4404a2014-08-06 16:06:47 -07002814 int ret = 0, outside_reserve = 0;
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07002815 unsigned long mmun_start; /* For mmu_notifiers */
2816 unsigned long mmun_end; /* For mmu_notifiers */
David Gibson1e8f8892006-01-06 00:10:44 -08002817
2818 old_page = pte_page(pte);
2819
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002820retry_avoidcopy:
David Gibson1e8f8892006-01-06 00:10:44 -08002821 /* If no-one else is actually using this page, avoid the copy
2822 * and just make the page writable */
Joonsoo Kim37a21402013-09-11 14:21:04 -07002823 if (page_mapcount(old_page) == 1 && PageAnon(old_page)) {
2824 page_move_anon_rmap(old_page, vma, address);
David Gibson1e8f8892006-01-06 00:10:44 -08002825 set_huge_ptep_writable(vma, address, ptep);
Nick Piggin83c54072007-07-19 01:47:05 -07002826 return 0;
David Gibson1e8f8892006-01-06 00:10:44 -08002827 }
2828
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002829 /*
2830 * If the process that created a MAP_PRIVATE mapping is about to
2831 * perform a COW due to a shared page count, attempt to satisfy
2832 * the allocation without using the existing reserves. The pagecache
2833 * page is used to determine if the reserve at this address was
2834 * consumed or not. If reserves were used, a partial faulted mapping
2835 * at the time of fork() could consume its reserves on COW instead
2836 * of the full address range.
2837 */
Joonsoo Kim5944d012013-09-11 14:21:55 -07002838 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002839 old_page != pagecache_page)
2840 outside_reserve = 1;
2841
David Gibson1e8f8892006-01-06 00:10:44 -08002842 page_cache_get(old_page);
Larry Woodmanb76c8cf2009-12-14 17:59:37 -08002843
Davidlohr Buesoad4404a2014-08-06 16:06:47 -07002844 /*
2845 * Drop page table lock as buddy allocator may be called. It will
2846 * be acquired again before returning to the caller, as expected.
2847 */
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002848 spin_unlock(ptl);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002849 new_page = alloc_huge_page(vma, address, outside_reserve);
David Gibson1e8f8892006-01-06 00:10:44 -08002850
Adam Litke2fc39ce2007-11-14 16:59:39 -08002851 if (IS_ERR(new_page)) {
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002852 /*
2853 * If a process owning a MAP_PRIVATE mapping fails to COW,
2854 * it is due to references held by a child and an insufficient
2855 * huge page pool. To guarantee the original mappers
2856 * reliability, unmap the page from child processes. The child
2857 * may get SIGKILLed if it later faults.
2858 */
2859 if (outside_reserve) {
Davidlohr Buesoad4404a2014-08-06 16:06:47 -07002860 page_cache_release(old_page);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002861 BUG_ON(huge_pte_none(pte));
Davidlohr Bueso2f4612a2014-08-06 16:06:45 -07002862 unmap_ref_private(mm, vma, old_page, address);
2863 BUG_ON(huge_pte_none(pte));
2864 spin_lock(ptl);
2865 ptep = huge_pte_offset(mm, address & huge_page_mask(h));
2866 if (likely(ptep &&
2867 pte_same(huge_ptep_get(ptep), pte)))
2868 goto retry_avoidcopy;
2869 /*
2870 * race occurs while re-acquiring page table
2871 * lock, and our job is done.
2872 */
2873 return 0;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002874 }
2875
Davidlohr Buesoad4404a2014-08-06 16:06:47 -07002876 ret = (PTR_ERR(new_page) == -ENOMEM) ?
2877 VM_FAULT_OOM : VM_FAULT_SIGBUS;
2878 goto out_release_old;
David Gibson1e8f8892006-01-06 00:10:44 -08002879 }
2880
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002881 /*
2882 * When the original hugepage is shared one, it does not have
2883 * anon_vma prepared.
2884 */
Dean Nelson44e2aa92010-10-26 14:22:08 -07002885 if (unlikely(anon_vma_prepare(vma))) {
Davidlohr Buesoad4404a2014-08-06 16:06:47 -07002886 ret = VM_FAULT_OOM;
2887 goto out_release_all;
Dean Nelson44e2aa92010-10-26 14:22:08 -07002888 }
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002889
Andrea Arcangeli47ad8472011-01-13 15:46:47 -08002890 copy_user_huge_page(new_page, old_page, address, vma,
2891 pages_per_huge_page(h));
Nick Piggin0ed361d2008-02-04 22:29:34 -08002892 __SetPageUptodate(new_page);
David Gibson1e8f8892006-01-06 00:10:44 -08002893
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07002894 mmun_start = address & huge_page_mask(h);
2895 mmun_end = mmun_start + huge_page_size(h);
2896 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
Davidlohr Buesoad4404a2014-08-06 16:06:47 -07002897
Larry Woodmanb76c8cf2009-12-14 17:59:37 -08002898 /*
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002899 * Retake the page table lock to check for racing updates
Larry Woodmanb76c8cf2009-12-14 17:59:37 -08002900 * before the page tables are altered
2901 */
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002902 spin_lock(ptl);
Andi Kleena5516432008-07-23 21:27:41 -07002903 ptep = huge_pte_offset(mm, address & huge_page_mask(h));
Naoya Horiguchia9af0c52014-04-07 15:36:54 -07002904 if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) {
Joonsoo Kim07443a82013-09-11 14:21:58 -07002905 ClearPagePrivate(new_page);
2906
David Gibson1e8f8892006-01-06 00:10:44 -08002907 /* Break COW */
Gerald Schaefer8fe627e2008-04-28 02:13:28 -07002908 huge_ptep_clear_flush(vma, address, ptep);
Joerg Roedel34ee6452014-11-13 13:46:09 +11002909 mmu_notifier_invalidate_range(mm, mmun_start, mmun_end);
David Gibson1e8f8892006-01-06 00:10:44 -08002910 set_huge_pte_at(mm, address, ptep,
2911 make_huge_pte(vma, new_page, 1));
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002912 page_remove_rmap(old_page);
Naoya Horiguchicd67f0d2010-09-10 13:23:04 +09002913 hugepage_add_new_anon_rmap(new_page, vma, address);
David Gibson1e8f8892006-01-06 00:10:44 -08002914 /* Make the old page be freed below */
2915 new_page = old_page;
2916 }
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002917 spin_unlock(ptl);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07002918 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
Davidlohr Buesoad4404a2014-08-06 16:06:47 -07002919out_release_all:
David Gibson1e8f8892006-01-06 00:10:44 -08002920 page_cache_release(new_page);
Davidlohr Buesoad4404a2014-08-06 16:06:47 -07002921out_release_old:
David Gibson1e8f8892006-01-06 00:10:44 -08002922 page_cache_release(old_page);
Joonsoo Kim83120342013-09-11 14:21:57 -07002923
Davidlohr Buesoad4404a2014-08-06 16:06:47 -07002924 spin_lock(ptl); /* Caller expects lock to be held */
2925 return ret;
David Gibson1e8f8892006-01-06 00:10:44 -08002926}
2927
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002928/* Return the pagecache page at a given address within a VMA */
Andi Kleena5516432008-07-23 21:27:41 -07002929static struct page *hugetlbfs_pagecache_page(struct hstate *h,
2930 struct vm_area_struct *vma, unsigned long address)
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002931{
2932 struct address_space *mapping;
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -07002933 pgoff_t idx;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002934
2935 mapping = vma->vm_file->f_mapping;
Andi Kleena5516432008-07-23 21:27:41 -07002936 idx = vma_hugecache_offset(h, vma, address);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002937
2938 return find_lock_page(mapping, idx);
2939}
2940
Hugh Dickins3ae77f42009-09-21 17:03:33 -07002941/*
2942 * Return whether there is a pagecache page to back given address within VMA.
2943 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
2944 */
2945static bool hugetlbfs_pagecache_present(struct hstate *h,
Hugh Dickins2a15efc2009-09-21 17:03:27 -07002946 struct vm_area_struct *vma, unsigned long address)
2947{
2948 struct address_space *mapping;
2949 pgoff_t idx;
2950 struct page *page;
2951
2952 mapping = vma->vm_file->f_mapping;
2953 idx = vma_hugecache_offset(h, vma, address);
2954
2955 page = find_get_page(mapping, idx);
2956 if (page)
2957 put_page(page);
2958 return page != NULL;
2959}
2960
Robert P. J. Daya1ed3dd2007-07-17 04:03:33 -07002961static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
Davidlohr Bueso8382d912014-04-03 14:47:31 -07002962 struct address_space *mapping, pgoff_t idx,
2963 unsigned long address, pte_t *ptep, unsigned int flags)
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01002964{
Andi Kleena5516432008-07-23 21:27:41 -07002965 struct hstate *h = hstate_vma(vma);
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01002966 int ret = VM_FAULT_SIGBUS;
Hillf Danton409eb8c2012-01-20 14:34:13 -08002967 int anon_rmap = 0;
Adam Litke4c887262005-10-29 18:16:46 -07002968 unsigned long size;
Adam Litke4c887262005-10-29 18:16:46 -07002969 struct page *page;
David Gibson1e8f8892006-01-06 00:10:44 -08002970 pte_t new_pte;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002971 spinlock_t *ptl;
Adam Litke4c887262005-10-29 18:16:46 -07002972
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002973 /*
2974 * Currently, we are forced to kill the process in the event the
2975 * original mapper has unmapped pages from the child due to a failed
Lucas De Marchi25985ed2011-03-30 22:57:33 -03002976 * COW. Warn that such a situation has occurred as it may not be obvious
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002977 */
2978 if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
Andrew Mortonffb22af2013-02-22 16:32:08 -08002979 pr_warning("PID %d killed due to inadequate hugepage pool\n",
2980 current->pid);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002981 return ret;
2982 }
2983
Adam Litke4c887262005-10-29 18:16:46 -07002984 /*
2985 * Use page lock to guard against racing truncation
2986 * before we get page_table_lock.
2987 */
Christoph Lameter6bda6662006-01-06 00:10:49 -08002988retry:
2989 page = find_lock_page(mapping, idx);
2990 if (!page) {
Andi Kleena5516432008-07-23 21:27:41 -07002991 size = i_size_read(mapping->host) >> huge_page_shift(h);
Hugh Dickinsebed4bf2006-10-28 10:38:43 -07002992 if (idx >= size)
2993 goto out;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002994 page = alloc_huge_page(vma, address, 0);
Adam Litke2fc39ce2007-11-14 16:59:39 -08002995 if (IS_ERR(page)) {
Aneesh Kumar K.V76dcee72012-07-31 16:41:57 -07002996 ret = PTR_ERR(page);
2997 if (ret == -ENOMEM)
2998 ret = VM_FAULT_OOM;
2999 else
3000 ret = VM_FAULT_SIGBUS;
Christoph Lameter6bda6662006-01-06 00:10:49 -08003001 goto out;
3002 }
Andrea Arcangeli47ad8472011-01-13 15:46:47 -08003003 clear_huge_page(page, address, pages_per_huge_page(h));
Nick Piggin0ed361d2008-02-04 22:29:34 -08003004 __SetPageUptodate(page);
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01003005
Mel Gormanf83a2752009-05-28 14:34:40 -07003006 if (vma->vm_flags & VM_MAYSHARE) {
Christoph Lameter6bda6662006-01-06 00:10:49 -08003007 int err;
Ken Chen45c682a2007-11-14 16:59:44 -08003008 struct inode *inode = mapping->host;
Christoph Lameter6bda6662006-01-06 00:10:49 -08003009
3010 err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
3011 if (err) {
3012 put_page(page);
Christoph Lameter6bda6662006-01-06 00:10:49 -08003013 if (err == -EEXIST)
3014 goto retry;
3015 goto out;
3016 }
Joonsoo Kim07443a82013-09-11 14:21:58 -07003017 ClearPagePrivate(page);
Ken Chen45c682a2007-11-14 16:59:44 -08003018
3019 spin_lock(&inode->i_lock);
Andi Kleena5516432008-07-23 21:27:41 -07003020 inode->i_blocks += blocks_per_huge_page(h);
Ken Chen45c682a2007-11-14 16:59:44 -08003021 spin_unlock(&inode->i_lock);
Mel Gorman23be7462010-04-23 13:17:56 -04003022 } else {
Christoph Lameter6bda6662006-01-06 00:10:49 -08003023 lock_page(page);
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09003024 if (unlikely(anon_vma_prepare(vma))) {
3025 ret = VM_FAULT_OOM;
3026 goto backout_unlocked;
3027 }
Hillf Danton409eb8c2012-01-20 14:34:13 -08003028 anon_rmap = 1;
Mel Gorman23be7462010-04-23 13:17:56 -04003029 }
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09003030 } else {
Naoya Horiguchi998b4382010-09-08 10:19:32 +09003031 /*
3032 * If memory error occurs between mmap() and fault, some process
3033 * don't have hwpoisoned swap entry for errored virtual address.
3034 * So we need to block hugepage fault by PG_hwpoison bit check.
3035 */
3036 if (unlikely(PageHWPoison(page))) {
Chris Forbes32f84522011-07-25 17:12:14 -07003037 ret = VM_FAULT_HWPOISON |
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07003038 VM_FAULT_SET_HINDEX(hstate_index(h));
Naoya Horiguchi998b4382010-09-08 10:19:32 +09003039 goto backout_unlocked;
3040 }
Christoph Lameter6bda6662006-01-06 00:10:49 -08003041 }
David Gibson1e8f8892006-01-06 00:10:44 -08003042
Andy Whitcroft57303d82008-08-12 15:08:47 -07003043 /*
3044 * If we are going to COW a private mapping later, we examine the
3045 * pending reservations for this page now. This will ensure that
3046 * any allocations necessary to record that reservation occur outside
3047 * the spinlock.
3048 */
Hugh Dickins788c7df2009-06-23 13:49:05 +01003049 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED))
Andy Whitcroft2b267362008-08-12 15:08:49 -07003050 if (vma_needs_reservation(h, vma, address) < 0) {
3051 ret = VM_FAULT_OOM;
3052 goto backout_unlocked;
3053 }
Andy Whitcroft57303d82008-08-12 15:08:47 -07003054
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003055 ptl = huge_pte_lockptr(h, mm, ptep);
3056 spin_lock(ptl);
Andi Kleena5516432008-07-23 21:27:41 -07003057 size = i_size_read(mapping->host) >> huge_page_shift(h);
Adam Litke4c887262005-10-29 18:16:46 -07003058 if (idx >= size)
3059 goto backout;
3060
Nick Piggin83c54072007-07-19 01:47:05 -07003061 ret = 0;
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07003062 if (!huge_pte_none(huge_ptep_get(ptep)))
Adam Litke4c887262005-10-29 18:16:46 -07003063 goto backout;
3064
Joonsoo Kim07443a82013-09-11 14:21:58 -07003065 if (anon_rmap) {
3066 ClearPagePrivate(page);
Hillf Danton409eb8c2012-01-20 14:34:13 -08003067 hugepage_add_new_anon_rmap(page, vma, address);
Choi Gi-yongac714902014-04-07 15:37:36 -07003068 } else
Hillf Danton409eb8c2012-01-20 14:34:13 -08003069 page_dup_rmap(page);
David Gibson1e8f8892006-01-06 00:10:44 -08003070 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
3071 && (vma->vm_flags & VM_SHARED)));
3072 set_huge_pte_at(mm, address, ptep, new_pte);
3073
Hugh Dickins788c7df2009-06-23 13:49:05 +01003074 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
David Gibson1e8f8892006-01-06 00:10:44 -08003075 /* Optimization, do the COW without a second fault */
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003076 ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page, ptl);
David Gibson1e8f8892006-01-06 00:10:44 -08003077 }
3078
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003079 spin_unlock(ptl);
Adam Litke4c887262005-10-29 18:16:46 -07003080 unlock_page(page);
3081out:
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01003082 return ret;
Adam Litke4c887262005-10-29 18:16:46 -07003083
3084backout:
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003085 spin_unlock(ptl);
Andy Whitcroft2b267362008-08-12 15:08:49 -07003086backout_unlocked:
Adam Litke4c887262005-10-29 18:16:46 -07003087 unlock_page(page);
3088 put_page(page);
3089 goto out;
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01003090}
3091
Davidlohr Bueso8382d912014-04-03 14:47:31 -07003092#ifdef CONFIG_SMP
3093static u32 fault_mutex_hash(struct hstate *h, struct mm_struct *mm,
3094 struct vm_area_struct *vma,
3095 struct address_space *mapping,
3096 pgoff_t idx, unsigned long address)
3097{
3098 unsigned long key[2];
3099 u32 hash;
3100
3101 if (vma->vm_flags & VM_SHARED) {
3102 key[0] = (unsigned long) mapping;
3103 key[1] = idx;
3104 } else {
3105 key[0] = (unsigned long) mm;
3106 key[1] = address >> huge_page_shift(h);
3107 }
3108
3109 hash = jhash2((u32 *)&key, sizeof(key)/sizeof(u32), 0);
3110
3111 return hash & (num_fault_mutexes - 1);
3112}
3113#else
3114/*
3115 * For uniprocesor systems we always use a single mutex, so just
3116 * return 0 and avoid the hashing overhead.
3117 */
3118static u32 fault_mutex_hash(struct hstate *h, struct mm_struct *mm,
3119 struct vm_area_struct *vma,
3120 struct address_space *mapping,
3121 pgoff_t idx, unsigned long address)
3122{
3123 return 0;
3124}
3125#endif
3126
Adam Litke86e52162006-01-06 00:10:43 -08003127int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
Hugh Dickins788c7df2009-06-23 13:49:05 +01003128 unsigned long address, unsigned int flags)
Adam Litke86e52162006-01-06 00:10:43 -08003129{
Davidlohr Bueso8382d912014-04-03 14:47:31 -07003130 pte_t *ptep, entry;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003131 spinlock_t *ptl;
David Gibson1e8f8892006-01-06 00:10:44 -08003132 int ret;
Davidlohr Bueso8382d912014-04-03 14:47:31 -07003133 u32 hash;
3134 pgoff_t idx;
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09003135 struct page *page = NULL;
Andy Whitcroft57303d82008-08-12 15:08:47 -07003136 struct page *pagecache_page = NULL;
Andi Kleena5516432008-07-23 21:27:41 -07003137 struct hstate *h = hstate_vma(vma);
Davidlohr Bueso8382d912014-04-03 14:47:31 -07003138 struct address_space *mapping;
Naoya Horiguchi0f792cf2015-02-11 15:25:25 -08003139 int need_wait_lock = 0;
Adam Litke86e52162006-01-06 00:10:43 -08003140
KAMEZAWA Hiroyuki1e16a532012-01-10 15:07:22 -08003141 address &= huge_page_mask(h);
3142
Naoya Horiguchifd6a03e2010-05-28 09:29:21 +09003143 ptep = huge_pte_offset(mm, address);
3144 if (ptep) {
3145 entry = huge_ptep_get(ptep);
Naoya Horiguchi290408d2010-09-08 10:19:35 +09003146 if (unlikely(is_hugetlb_entry_migration(entry))) {
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003147 migration_entry_wait_huge(vma, mm, ptep);
Naoya Horiguchi290408d2010-09-08 10:19:35 +09003148 return 0;
3149 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
Chris Forbes32f84522011-07-25 17:12:14 -07003150 return VM_FAULT_HWPOISON_LARGE |
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07003151 VM_FAULT_SET_HINDEX(hstate_index(h));
Naoya Horiguchifd6a03e2010-05-28 09:29:21 +09003152 }
3153
Andi Kleena5516432008-07-23 21:27:41 -07003154 ptep = huge_pte_alloc(mm, address, huge_page_size(h));
Adam Litke86e52162006-01-06 00:10:43 -08003155 if (!ptep)
3156 return VM_FAULT_OOM;
3157
Davidlohr Bueso8382d912014-04-03 14:47:31 -07003158 mapping = vma->vm_file->f_mapping;
3159 idx = vma_hugecache_offset(h, vma, address);
3160
David Gibson3935baa2006-03-22 00:08:53 -08003161 /*
3162 * Serialize hugepage allocation and instantiation, so that we don't
3163 * get spurious allocation failures if two CPUs race to instantiate
3164 * the same page in the page cache.
3165 */
Davidlohr Bueso8382d912014-04-03 14:47:31 -07003166 hash = fault_mutex_hash(h, mm, vma, mapping, idx, address);
3167 mutex_lock(&htlb_fault_mutex_table[hash]);
3168
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07003169 entry = huge_ptep_get(ptep);
3170 if (huge_pte_none(entry)) {
Davidlohr Bueso8382d912014-04-03 14:47:31 -07003171 ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags);
David Gibsonb4d1d992008-10-15 22:01:11 -07003172 goto out_mutex;
David Gibson3935baa2006-03-22 00:08:53 -08003173 }
Adam Litke86e52162006-01-06 00:10:43 -08003174
Nick Piggin83c54072007-07-19 01:47:05 -07003175 ret = 0;
David Gibson1e8f8892006-01-06 00:10:44 -08003176
Andy Whitcroft57303d82008-08-12 15:08:47 -07003177 /*
Naoya Horiguchi0f792cf2015-02-11 15:25:25 -08003178 * entry could be a migration/hwpoison entry at this point, so this
3179 * check prevents the kernel from going below assuming that we have
3180 * a active hugepage in pagecache. This goto expects the 2nd page fault,
3181 * and is_hugetlb_entry_(migration|hwpoisoned) check will properly
3182 * handle it.
3183 */
3184 if (!pte_present(entry))
3185 goto out_mutex;
3186
3187 /*
Andy Whitcroft57303d82008-08-12 15:08:47 -07003188 * If we are going to COW the mapping later, we examine the pending
3189 * reservations for this page now. This will ensure that any
3190 * allocations necessary to record that reservation occur outside the
3191 * spinlock. For private mappings, we also lookup the pagecache
3192 * page now as it is used to determine if a reservation has been
3193 * consumed.
3194 */
Gerald Schaefer106c9922013-04-29 15:07:23 -07003195 if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) {
Andy Whitcroft2b267362008-08-12 15:08:49 -07003196 if (vma_needs_reservation(h, vma, address) < 0) {
3197 ret = VM_FAULT_OOM;
David Gibsonb4d1d992008-10-15 22:01:11 -07003198 goto out_mutex;
Andy Whitcroft2b267362008-08-12 15:08:49 -07003199 }
Andy Whitcroft57303d82008-08-12 15:08:47 -07003200
Mel Gormanf83a2752009-05-28 14:34:40 -07003201 if (!(vma->vm_flags & VM_MAYSHARE))
Andy Whitcroft57303d82008-08-12 15:08:47 -07003202 pagecache_page = hugetlbfs_pagecache_page(h,
3203 vma, address);
3204 }
3205
Naoya Horiguchi0f792cf2015-02-11 15:25:25 -08003206 ptl = huge_pte_lock(h, mm, ptep);
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09003207
David Gibson1e8f8892006-01-06 00:10:44 -08003208 /* Check for a racing update before calling hugetlb_cow */
David Gibsonb4d1d992008-10-15 22:01:11 -07003209 if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003210 goto out_ptl;
David Gibsonb4d1d992008-10-15 22:01:11 -07003211
Naoya Horiguchi0f792cf2015-02-11 15:25:25 -08003212 /*
3213 * hugetlb_cow() requires page locks of pte_page(entry) and
3214 * pagecache_page, so here we need take the former one
3215 * when page != pagecache_page or !pagecache_page.
3216 */
3217 page = pte_page(entry);
3218 if (page != pagecache_page)
3219 if (!trylock_page(page)) {
3220 need_wait_lock = 1;
3221 goto out_ptl;
3222 }
3223
3224 get_page(page);
David Gibsonb4d1d992008-10-15 22:01:11 -07003225
Hugh Dickins788c7df2009-06-23 13:49:05 +01003226 if (flags & FAULT_FLAG_WRITE) {
Gerald Schaefer106c9922013-04-29 15:07:23 -07003227 if (!huge_pte_write(entry)) {
Andy Whitcroft57303d82008-08-12 15:08:47 -07003228 ret = hugetlb_cow(mm, vma, address, ptep, entry,
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003229 pagecache_page, ptl);
Naoya Horiguchi0f792cf2015-02-11 15:25:25 -08003230 goto out_put_page;
David Gibsonb4d1d992008-10-15 22:01:11 -07003231 }
Gerald Schaefer106c9922013-04-29 15:07:23 -07003232 entry = huge_pte_mkdirty(entry);
David Gibsonb4d1d992008-10-15 22:01:11 -07003233 }
3234 entry = pte_mkyoung(entry);
Hugh Dickins788c7df2009-06-23 13:49:05 +01003235 if (huge_ptep_set_access_flags(vma, address, ptep, entry,
3236 flags & FAULT_FLAG_WRITE))
Russell King4b3073e2009-12-18 16:40:18 +00003237 update_mmu_cache(vma, address, ptep);
Naoya Horiguchi0f792cf2015-02-11 15:25:25 -08003238out_put_page:
3239 if (page != pagecache_page)
3240 unlock_page(page);
3241 put_page(page);
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003242out_ptl:
3243 spin_unlock(ptl);
Andy Whitcroft57303d82008-08-12 15:08:47 -07003244
3245 if (pagecache_page) {
3246 unlock_page(pagecache_page);
3247 put_page(pagecache_page);
3248 }
David Gibsonb4d1d992008-10-15 22:01:11 -07003249out_mutex:
Davidlohr Bueso8382d912014-04-03 14:47:31 -07003250 mutex_unlock(&htlb_fault_mutex_table[hash]);
Naoya Horiguchi0f792cf2015-02-11 15:25:25 -08003251 /*
3252 * Generally it's safe to hold refcount during waiting page lock. But
3253 * here we just wait to defer the next page fault to avoid busy loop and
3254 * the page is not used after unlocked before returning from the current
3255 * page fault. So we are safe from accessing freed page, even if we wait
3256 * here without taking refcount.
3257 */
3258 if (need_wait_lock)
3259 wait_on_page_locked(page);
David Gibson1e8f8892006-01-06 00:10:44 -08003260 return ret;
Adam Litke86e52162006-01-06 00:10:43 -08003261}
3262
Michel Lespinasse28a35712013-02-22 16:35:55 -08003263long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
3264 struct page **pages, struct vm_area_struct **vmas,
3265 unsigned long *position, unsigned long *nr_pages,
3266 long i, unsigned int flags)
David Gibson63551ae2005-06-21 17:14:44 -07003267{
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08003268 unsigned long pfn_offset;
3269 unsigned long vaddr = *position;
Michel Lespinasse28a35712013-02-22 16:35:55 -08003270 unsigned long remainder = *nr_pages;
Andi Kleena5516432008-07-23 21:27:41 -07003271 struct hstate *h = hstate_vma(vma);
David Gibson63551ae2005-06-21 17:14:44 -07003272
David Gibson63551ae2005-06-21 17:14:44 -07003273 while (vaddr < vma->vm_end && remainder) {
Adam Litke4c887262005-10-29 18:16:46 -07003274 pte_t *pte;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003275 spinlock_t *ptl = NULL;
Hugh Dickins2a15efc2009-09-21 17:03:27 -07003276 int absent;
Adam Litke4c887262005-10-29 18:16:46 -07003277 struct page *page;
3278
3279 /*
David Rientjes02057962015-04-14 15:48:24 -07003280 * If we have a pending SIGKILL, don't keep faulting pages and
3281 * potentially allocating memory.
3282 */
3283 if (unlikely(fatal_signal_pending(current))) {
3284 remainder = 0;
3285 break;
3286 }
3287
3288 /*
Adam Litke4c887262005-10-29 18:16:46 -07003289 * Some archs (sparc64, sh*) have multiple pte_ts to
Hugh Dickins2a15efc2009-09-21 17:03:27 -07003290 * each hugepage. We have to make sure we get the
Adam Litke4c887262005-10-29 18:16:46 -07003291 * first, for the page indexing below to work.
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003292 *
3293 * Note that page table lock is not held when pte is null.
Adam Litke4c887262005-10-29 18:16:46 -07003294 */
Andi Kleena5516432008-07-23 21:27:41 -07003295 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h));
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003296 if (pte)
3297 ptl = huge_pte_lock(h, mm, pte);
Hugh Dickins2a15efc2009-09-21 17:03:27 -07003298 absent = !pte || huge_pte_none(huge_ptep_get(pte));
Adam Litke4c887262005-10-29 18:16:46 -07003299
Hugh Dickins2a15efc2009-09-21 17:03:27 -07003300 /*
3301 * When coredumping, it suits get_dump_page if we just return
Hugh Dickins3ae77f42009-09-21 17:03:33 -07003302 * an error where there's an empty slot with no huge pagecache
3303 * to back it. This way, we avoid allocating a hugepage, and
3304 * the sparse dumpfile avoids allocating disk blocks, but its
3305 * huge holes still show up with zeroes where they need to be.
Hugh Dickins2a15efc2009-09-21 17:03:27 -07003306 */
Hugh Dickins3ae77f42009-09-21 17:03:33 -07003307 if (absent && (flags & FOLL_DUMP) &&
3308 !hugetlbfs_pagecache_present(h, vma, vaddr)) {
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003309 if (pte)
3310 spin_unlock(ptl);
Hugh Dickins2a15efc2009-09-21 17:03:27 -07003311 remainder = 0;
3312 break;
3313 }
3314
Naoya Horiguchi9cc3a5b2013-04-17 15:58:30 -07003315 /*
3316 * We need call hugetlb_fault for both hugepages under migration
3317 * (in which case hugetlb_fault waits for the migration,) and
3318 * hwpoisoned hugepages (in which case we need to prevent the
3319 * caller from accessing to them.) In order to do this, we use
3320 * here is_swap_pte instead of is_hugetlb_entry_migration and
3321 * is_hugetlb_entry_hwpoisoned. This is because it simply covers
3322 * both cases, and because we can't follow correct pages
3323 * directly from any kind of swap entries.
3324 */
3325 if (absent || is_swap_pte(huge_ptep_get(pte)) ||
Gerald Schaefer106c9922013-04-29 15:07:23 -07003326 ((flags & FOLL_WRITE) &&
3327 !huge_pte_write(huge_ptep_get(pte)))) {
Adam Litke4c887262005-10-29 18:16:46 -07003328 int ret;
3329
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003330 if (pte)
3331 spin_unlock(ptl);
Hugh Dickins2a15efc2009-09-21 17:03:27 -07003332 ret = hugetlb_fault(mm, vma, vaddr,
3333 (flags & FOLL_WRITE) ? FAULT_FLAG_WRITE : 0);
Adam Litkea89182c2007-08-22 14:01:51 -07003334 if (!(ret & VM_FAULT_ERROR))
Adam Litke4c887262005-10-29 18:16:46 -07003335 continue;
3336
3337 remainder = 0;
Adam Litke4c887262005-10-29 18:16:46 -07003338 break;
3339 }
David Gibson63551ae2005-06-21 17:14:44 -07003340
Andi Kleena5516432008-07-23 21:27:41 -07003341 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07003342 page = pte_page(huge_ptep_get(pte));
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08003343same_page:
Chen, Kenneth Wd6692182006-03-31 02:29:57 -08003344 if (pages) {
Hugh Dickins2a15efc2009-09-21 17:03:27 -07003345 pages[i] = mem_map_offset(page, pfn_offset);
Andrea Arcangelia0368d42014-01-21 15:48:49 -08003346 get_page_foll(pages[i]);
Chen, Kenneth Wd6692182006-03-31 02:29:57 -08003347 }
David Gibson63551ae2005-06-21 17:14:44 -07003348
3349 if (vmas)
3350 vmas[i] = vma;
3351
3352 vaddr += PAGE_SIZE;
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08003353 ++pfn_offset;
David Gibson63551ae2005-06-21 17:14:44 -07003354 --remainder;
3355 ++i;
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08003356 if (vaddr < vma->vm_end && remainder &&
Andi Kleena5516432008-07-23 21:27:41 -07003357 pfn_offset < pages_per_huge_page(h)) {
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08003358 /*
3359 * We use pfn_offset to avoid touching the pageframes
3360 * of this compound page.
3361 */
3362 goto same_page;
3363 }
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003364 spin_unlock(ptl);
David Gibson63551ae2005-06-21 17:14:44 -07003365 }
Michel Lespinasse28a35712013-02-22 16:35:55 -08003366 *nr_pages = remainder;
David Gibson63551ae2005-06-21 17:14:44 -07003367 *position = vaddr;
3368
Hugh Dickins2a15efc2009-09-21 17:03:27 -07003369 return i ? i : -EFAULT;
David Gibson63551ae2005-06-21 17:14:44 -07003370}
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003371
Peter Zijlstra7da4d642012-11-19 03:14:23 +01003372unsigned long hugetlb_change_protection(struct vm_area_struct *vma,
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003373 unsigned long address, unsigned long end, pgprot_t newprot)
3374{
3375 struct mm_struct *mm = vma->vm_mm;
3376 unsigned long start = address;
3377 pte_t *ptep;
3378 pte_t pte;
Andi Kleena5516432008-07-23 21:27:41 -07003379 struct hstate *h = hstate_vma(vma);
Peter Zijlstra7da4d642012-11-19 03:14:23 +01003380 unsigned long pages = 0;
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003381
3382 BUG_ON(address >= end);
3383 flush_cache_range(vma, address, end);
3384
Rik van Riela5338092014-04-07 15:36:57 -07003385 mmu_notifier_invalidate_range_start(mm, start, end);
Davidlohr Bueso83cde9e2014-12-12 16:54:21 -08003386 i_mmap_lock_write(vma->vm_file->f_mapping);
Andi Kleena5516432008-07-23 21:27:41 -07003387 for (; address < end; address += huge_page_size(h)) {
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003388 spinlock_t *ptl;
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003389 ptep = huge_pte_offset(mm, address);
3390 if (!ptep)
3391 continue;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003392 ptl = huge_pte_lock(h, mm, ptep);
Peter Zijlstra7da4d642012-11-19 03:14:23 +01003393 if (huge_pmd_unshare(mm, &address, ptep)) {
3394 pages++;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003395 spin_unlock(ptl);
Chen, Kenneth W39dde652006-12-06 20:32:03 -08003396 continue;
Peter Zijlstra7da4d642012-11-19 03:14:23 +01003397 }
Naoya Horiguchia8bda282015-02-11 15:25:28 -08003398 pte = huge_ptep_get(ptep);
3399 if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) {
3400 spin_unlock(ptl);
3401 continue;
3402 }
3403 if (unlikely(is_hugetlb_entry_migration(pte))) {
3404 swp_entry_t entry = pte_to_swp_entry(pte);
3405
3406 if (is_write_migration_entry(entry)) {
3407 pte_t newpte;
3408
3409 make_migration_entry_read(&entry);
3410 newpte = swp_entry_to_pte(entry);
3411 set_huge_pte_at(mm, address, ptep, newpte);
3412 pages++;
3413 }
3414 spin_unlock(ptl);
3415 continue;
3416 }
3417 if (!huge_pte_none(pte)) {
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003418 pte = huge_ptep_get_and_clear(mm, address, ptep);
Gerald Schaefer106c9922013-04-29 15:07:23 -07003419 pte = pte_mkhuge(huge_pte_modify(pte, newprot));
Tony Lube7517d2013-02-04 14:28:46 -08003420 pte = arch_make_huge_pte(pte, vma, NULL, 0);
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003421 set_huge_pte_at(mm, address, ptep, pte);
Peter Zijlstra7da4d642012-11-19 03:14:23 +01003422 pages++;
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003423 }
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003424 spin_unlock(ptl);
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003425 }
Mel Gormand8333522012-07-31 16:46:20 -07003426 /*
Davidlohr Buesoc8c06ef2014-12-12 16:54:24 -08003427 * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare
Mel Gormand8333522012-07-31 16:46:20 -07003428 * may have cleared our pud entry and done put_page on the page table:
Davidlohr Buesoc8c06ef2014-12-12 16:54:24 -08003429 * once we release i_mmap_rwsem, another task can do the final put_page
Mel Gormand8333522012-07-31 16:46:20 -07003430 * and that page table be reused and filled with junk.
3431 */
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003432 flush_tlb_range(vma, start, end);
Joerg Roedel34ee6452014-11-13 13:46:09 +11003433 mmu_notifier_invalidate_range(mm, start, end);
Davidlohr Bueso83cde9e2014-12-12 16:54:21 -08003434 i_mmap_unlock_write(vma->vm_file->f_mapping);
Rik van Riela5338092014-04-07 15:36:57 -07003435 mmu_notifier_invalidate_range_end(mm, start, end);
Peter Zijlstra7da4d642012-11-19 03:14:23 +01003436
3437 return pages << h->order;
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003438}
3439
Mel Gormana1e78772008-07-23 21:27:23 -07003440int hugetlb_reserve_pages(struct inode *inode,
3441 long from, long to,
Mel Gorman5a6fe122009-02-10 14:02:27 +00003442 struct vm_area_struct *vma,
KOSAKI Motohiroca16d142011-05-26 19:16:19 +09003443 vm_flags_t vm_flags)
Adam Litkee4e574b2007-10-16 01:26:19 -07003444{
Mel Gorman17c9d122009-02-11 16:34:16 +00003445 long ret, chg;
Andi Kleena5516432008-07-23 21:27:41 -07003446 struct hstate *h = hstate_inode(inode);
David Gibson90481622012-03-21 16:34:12 -07003447 struct hugepage_subpool *spool = subpool_inode(inode);
Joonsoo Kim9119a412014-04-03 14:47:25 -07003448 struct resv_map *resv_map;
Adam Litkee4e574b2007-10-16 01:26:19 -07003449
Mel Gormana1e78772008-07-23 21:27:23 -07003450 /*
Mel Gorman17c9d122009-02-11 16:34:16 +00003451 * Only apply hugepage reservation if asked. At fault time, an
3452 * attempt will be made for VM_NORESERVE to allocate a page
David Gibson90481622012-03-21 16:34:12 -07003453 * without using reserves
Mel Gorman17c9d122009-02-11 16:34:16 +00003454 */
KOSAKI Motohiroca16d142011-05-26 19:16:19 +09003455 if (vm_flags & VM_NORESERVE)
Mel Gorman17c9d122009-02-11 16:34:16 +00003456 return 0;
3457
3458 /*
Mel Gormana1e78772008-07-23 21:27:23 -07003459 * Shared mappings base their reservation on the number of pages that
3460 * are already allocated on behalf of the file. Private mappings need
3461 * to reserve the full area even if read-only as mprotect() may be
3462 * called to make the mapping read-write. Assume !vma is a shm mapping
3463 */
Joonsoo Kim9119a412014-04-03 14:47:25 -07003464 if (!vma || vma->vm_flags & VM_MAYSHARE) {
Joonsoo Kim4e35f482014-04-03 14:47:30 -07003465 resv_map = inode_resv_map(inode);
Joonsoo Kim9119a412014-04-03 14:47:25 -07003466
Joonsoo Kim1406ec92014-04-03 14:47:26 -07003467 chg = region_chg(resv_map, from, to);
Joonsoo Kim9119a412014-04-03 14:47:25 -07003468
3469 } else {
3470 resv_map = resv_map_alloc();
Mel Gorman5a6fe122009-02-10 14:02:27 +00003471 if (!resv_map)
3472 return -ENOMEM;
3473
Mel Gorman17c9d122009-02-11 16:34:16 +00003474 chg = to - from;
3475
Mel Gorman5a6fe122009-02-10 14:02:27 +00003476 set_vma_resv_map(vma, resv_map);
3477 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
3478 }
3479
Dave Hansenc50ac052012-05-29 15:06:46 -07003480 if (chg < 0) {
3481 ret = chg;
3482 goto out_err;
3483 }
Mel Gorman17c9d122009-02-11 16:34:16 +00003484
David Gibson90481622012-03-21 16:34:12 -07003485 /* There must be enough pages in the subpool for the mapping */
Dave Hansenc50ac052012-05-29 15:06:46 -07003486 if (hugepage_subpool_get_pages(spool, chg)) {
3487 ret = -ENOSPC;
3488 goto out_err;
3489 }
Mel Gorman17c9d122009-02-11 16:34:16 +00003490
3491 /*
3492 * Check enough hugepages are available for the reservation.
David Gibson90481622012-03-21 16:34:12 -07003493 * Hand the pages back to the subpool if there are not
Mel Gorman17c9d122009-02-11 16:34:16 +00003494 */
3495 ret = hugetlb_acct_memory(h, chg);
3496 if (ret < 0) {
David Gibson90481622012-03-21 16:34:12 -07003497 hugepage_subpool_put_pages(spool, chg);
Dave Hansenc50ac052012-05-29 15:06:46 -07003498 goto out_err;
Mel Gorman17c9d122009-02-11 16:34:16 +00003499 }
3500
3501 /*
3502 * Account for the reservations made. Shared mappings record regions
3503 * that have reservations as they are shared by multiple VMAs.
3504 * When the last VMA disappears, the region map says how much
3505 * the reservation was and the page cache tells how much of
3506 * the reservation was consumed. Private mappings are per-VMA and
3507 * only the consumed reservations are tracked. When the VMA
3508 * disappears, the original reservation is the VMA size and the
3509 * consumed reservations are stored in the map. Hence, nothing
3510 * else has to be done for private mappings here
3511 */
Mel Gormanf83a2752009-05-28 14:34:40 -07003512 if (!vma || vma->vm_flags & VM_MAYSHARE)
Joonsoo Kim1406ec92014-04-03 14:47:26 -07003513 region_add(resv_map, from, to);
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -07003514 return 0;
Dave Hansenc50ac052012-05-29 15:06:46 -07003515out_err:
Joonsoo Kimf031dd22014-04-03 14:47:28 -07003516 if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
3517 kref_put(&resv_map->refs, resv_map_release);
Dave Hansenc50ac052012-05-29 15:06:46 -07003518 return ret;
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -07003519}
3520
3521void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed)
3522{
Andi Kleena5516432008-07-23 21:27:41 -07003523 struct hstate *h = hstate_inode(inode);
Joonsoo Kim4e35f482014-04-03 14:47:30 -07003524 struct resv_map *resv_map = inode_resv_map(inode);
Joonsoo Kim9119a412014-04-03 14:47:25 -07003525 long chg = 0;
David Gibson90481622012-03-21 16:34:12 -07003526 struct hugepage_subpool *spool = subpool_inode(inode);
Ken Chen45c682a2007-11-14 16:59:44 -08003527
Joonsoo Kim9119a412014-04-03 14:47:25 -07003528 if (resv_map)
Joonsoo Kim1406ec92014-04-03 14:47:26 -07003529 chg = region_truncate(resv_map, offset);
Ken Chen45c682a2007-11-14 16:59:44 -08003530 spin_lock(&inode->i_lock);
Eric Sandeene4c6f8b2009-07-29 15:02:16 -07003531 inode->i_blocks -= (blocks_per_huge_page(h) * freed);
Ken Chen45c682a2007-11-14 16:59:44 -08003532 spin_unlock(&inode->i_lock);
3533
David Gibson90481622012-03-21 16:34:12 -07003534 hugepage_subpool_put_pages(spool, (chg - freed));
Andi Kleena5516432008-07-23 21:27:41 -07003535 hugetlb_acct_memory(h, -(chg - freed));
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -07003536}
Naoya Horiguchi93f70f92010-05-28 09:29:20 +09003537
Steve Capper3212b532013-04-23 12:35:02 +01003538#ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
3539static unsigned long page_table_shareable(struct vm_area_struct *svma,
3540 struct vm_area_struct *vma,
3541 unsigned long addr, pgoff_t idx)
3542{
3543 unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) +
3544 svma->vm_start;
3545 unsigned long sbase = saddr & PUD_MASK;
3546 unsigned long s_end = sbase + PUD_SIZE;
3547
3548 /* Allow segments to share if only one is marked locked */
3549 unsigned long vm_flags = vma->vm_flags & ~VM_LOCKED;
3550 unsigned long svm_flags = svma->vm_flags & ~VM_LOCKED;
3551
3552 /*
3553 * match the virtual addresses, permission and the alignment of the
3554 * page table page.
3555 */
3556 if (pmd_index(addr) != pmd_index(saddr) ||
3557 vm_flags != svm_flags ||
3558 sbase < svma->vm_start || svma->vm_end < s_end)
3559 return 0;
3560
3561 return saddr;
3562}
3563
3564static int vma_shareable(struct vm_area_struct *vma, unsigned long addr)
3565{
3566 unsigned long base = addr & PUD_MASK;
3567 unsigned long end = base + PUD_SIZE;
3568
3569 /*
3570 * check on proper vm_flags and page table alignment
3571 */
3572 if (vma->vm_flags & VM_MAYSHARE &&
3573 vma->vm_start <= base && end <= vma->vm_end)
3574 return 1;
3575 return 0;
3576}
3577
3578/*
3579 * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
3580 * and returns the corresponding pte. While this is not necessary for the
3581 * !shared pmd case because we can allocate the pmd later as well, it makes the
3582 * code much cleaner. pmd allocation is essential for the shared case because
Davidlohr Buesoc8c06ef2014-12-12 16:54:24 -08003583 * pud has to be populated inside the same i_mmap_rwsem section - otherwise
Steve Capper3212b532013-04-23 12:35:02 +01003584 * racing tasks could either miss the sharing (see huge_pte_offset) or select a
3585 * bad pmd for sharing.
3586 */
3587pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
3588{
3589 struct vm_area_struct *vma = find_vma(mm, addr);
3590 struct address_space *mapping = vma->vm_file->f_mapping;
3591 pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) +
3592 vma->vm_pgoff;
3593 struct vm_area_struct *svma;
3594 unsigned long saddr;
3595 pte_t *spte = NULL;
3596 pte_t *pte;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003597 spinlock_t *ptl;
Steve Capper3212b532013-04-23 12:35:02 +01003598
3599 if (!vma_shareable(vma, addr))
3600 return (pte_t *)pmd_alloc(mm, pud, addr);
3601
Davidlohr Bueso83cde9e2014-12-12 16:54:21 -08003602 i_mmap_lock_write(mapping);
Steve Capper3212b532013-04-23 12:35:02 +01003603 vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) {
3604 if (svma == vma)
3605 continue;
3606
3607 saddr = page_table_shareable(svma, vma, addr, idx);
3608 if (saddr) {
3609 spte = huge_pte_offset(svma->vm_mm, saddr);
3610 if (spte) {
Kirill A. Shutemovdc6c9a32015-02-11 15:26:50 -08003611 mm_inc_nr_pmds(mm);
Steve Capper3212b532013-04-23 12:35:02 +01003612 get_page(virt_to_page(spte));
3613 break;
3614 }
3615 }
3616 }
3617
3618 if (!spte)
3619 goto out;
3620
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003621 ptl = huge_pte_lockptr(hstate_vma(vma), mm, spte);
3622 spin_lock(ptl);
Kirill A. Shutemovdc6c9a32015-02-11 15:26:50 -08003623 if (pud_none(*pud)) {
Steve Capper3212b532013-04-23 12:35:02 +01003624 pud_populate(mm, pud,
3625 (pmd_t *)((unsigned long)spte & PAGE_MASK));
Kirill A. Shutemovdc6c9a32015-02-11 15:26:50 -08003626 } else {
Steve Capper3212b532013-04-23 12:35:02 +01003627 put_page(virt_to_page(spte));
Kirill A. Shutemovdc6c9a32015-02-11 15:26:50 -08003628 mm_inc_nr_pmds(mm);
3629 }
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003630 spin_unlock(ptl);
Steve Capper3212b532013-04-23 12:35:02 +01003631out:
3632 pte = (pte_t *)pmd_alloc(mm, pud, addr);
Davidlohr Bueso83cde9e2014-12-12 16:54:21 -08003633 i_mmap_unlock_write(mapping);
Steve Capper3212b532013-04-23 12:35:02 +01003634 return pte;
3635}
3636
3637/*
3638 * unmap huge page backed by shared pte.
3639 *
3640 * Hugetlb pte page is ref counted at the time of mapping. If pte is shared
3641 * indicated by page_count > 1, unmap is achieved by clearing pud and
3642 * decrementing the ref count. If count == 1, the pte page is not shared.
3643 *
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003644 * called with page table lock held.
Steve Capper3212b532013-04-23 12:35:02 +01003645 *
3646 * returns: 1 successfully unmapped a shared pte page
3647 * 0 the underlying pte page is not shared, or it is the last user
3648 */
3649int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
3650{
3651 pgd_t *pgd = pgd_offset(mm, *addr);
3652 pud_t *pud = pud_offset(pgd, *addr);
3653
3654 BUG_ON(page_count(virt_to_page(ptep)) == 0);
3655 if (page_count(virt_to_page(ptep)) == 1)
3656 return 0;
3657
3658 pud_clear(pud);
3659 put_page(virt_to_page(ptep));
Kirill A. Shutemovdc6c9a32015-02-11 15:26:50 -08003660 mm_dec_nr_pmds(mm);
Steve Capper3212b532013-04-23 12:35:02 +01003661 *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE;
3662 return 1;
3663}
Steve Capper9e5fc742013-04-30 08:02:03 +01003664#define want_pmd_share() (1)
3665#else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
3666pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
3667{
3668 return NULL;
3669}
3670#define want_pmd_share() (0)
Steve Capper3212b532013-04-23 12:35:02 +01003671#endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
3672
Steve Capper9e5fc742013-04-30 08:02:03 +01003673#ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
3674pte_t *huge_pte_alloc(struct mm_struct *mm,
3675 unsigned long addr, unsigned long sz)
3676{
3677 pgd_t *pgd;
3678 pud_t *pud;
3679 pte_t *pte = NULL;
3680
3681 pgd = pgd_offset(mm, addr);
3682 pud = pud_alloc(mm, pgd, addr);
3683 if (pud) {
3684 if (sz == PUD_SIZE) {
3685 pte = (pte_t *)pud;
3686 } else {
3687 BUG_ON(sz != PMD_SIZE);
3688 if (want_pmd_share() && pud_none(*pud))
3689 pte = huge_pmd_share(mm, addr, pud);
3690 else
3691 pte = (pte_t *)pmd_alloc(mm, pud, addr);
3692 }
3693 }
3694 BUG_ON(pte && !pte_none(*pte) && !pte_huge(*pte));
3695
3696 return pte;
3697}
3698
3699pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr)
3700{
3701 pgd_t *pgd;
3702 pud_t *pud;
3703 pmd_t *pmd = NULL;
3704
3705 pgd = pgd_offset(mm, addr);
3706 if (pgd_present(*pgd)) {
3707 pud = pud_offset(pgd, addr);
3708 if (pud_present(*pud)) {
3709 if (pud_huge(*pud))
3710 return (pte_t *)pud;
3711 pmd = pmd_offset(pud, addr);
3712 }
3713 }
3714 return (pte_t *) pmd;
3715}
3716
Naoya Horiguchi61f77ed2015-02-11 15:25:15 -08003717#endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
3718
3719/*
3720 * These functions are overwritable if your architecture needs its own
3721 * behavior.
3722 */
3723struct page * __weak
3724follow_huge_addr(struct mm_struct *mm, unsigned long address,
3725 int write)
3726{
3727 return ERR_PTR(-EINVAL);
3728}
3729
3730struct page * __weak
Steve Capper9e5fc742013-04-30 08:02:03 +01003731follow_huge_pmd(struct mm_struct *mm, unsigned long address,
Naoya Horiguchie66f17f2015-02-11 15:25:22 -08003732 pmd_t *pmd, int flags)
Steve Capper9e5fc742013-04-30 08:02:03 +01003733{
Naoya Horiguchie66f17f2015-02-11 15:25:22 -08003734 struct page *page = NULL;
3735 spinlock_t *ptl;
3736retry:
3737 ptl = pmd_lockptr(mm, pmd);
3738 spin_lock(ptl);
3739 /*
3740 * make sure that the address range covered by this pmd is not
3741 * unmapped from other threads.
3742 */
3743 if (!pmd_huge(*pmd))
3744 goto out;
3745 if (pmd_present(*pmd)) {
Gerald Schaefer97534122015-04-14 15:42:30 -07003746 page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT);
Naoya Horiguchie66f17f2015-02-11 15:25:22 -08003747 if (flags & FOLL_GET)
3748 get_page(page);
3749 } else {
3750 if (is_hugetlb_entry_migration(huge_ptep_get((pte_t *)pmd))) {
3751 spin_unlock(ptl);
3752 __migration_entry_wait(mm, (pte_t *)pmd, ptl);
3753 goto retry;
3754 }
3755 /*
3756 * hwpoisoned entry is treated as no_page_table in
3757 * follow_page_mask().
3758 */
3759 }
3760out:
3761 spin_unlock(ptl);
Steve Capper9e5fc742013-04-30 08:02:03 +01003762 return page;
3763}
3764
Naoya Horiguchi61f77ed2015-02-11 15:25:15 -08003765struct page * __weak
Steve Capper9e5fc742013-04-30 08:02:03 +01003766follow_huge_pud(struct mm_struct *mm, unsigned long address,
Naoya Horiguchie66f17f2015-02-11 15:25:22 -08003767 pud_t *pud, int flags)
Steve Capper9e5fc742013-04-30 08:02:03 +01003768{
Naoya Horiguchie66f17f2015-02-11 15:25:22 -08003769 if (flags & FOLL_GET)
3770 return NULL;
Steve Capper9e5fc742013-04-30 08:02:03 +01003771
Naoya Horiguchie66f17f2015-02-11 15:25:22 -08003772 return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT);
Steve Capper9e5fc742013-04-30 08:02:03 +01003773}
3774
Andi Kleend5bd9102010-09-27 09:00:12 +02003775#ifdef CONFIG_MEMORY_FAILURE
3776
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +09003777/* Should be called in hugetlb_lock */
3778static int is_hugepage_on_freelist(struct page *hpage)
3779{
3780 struct page *page;
3781 struct page *tmp;
3782 struct hstate *h = page_hstate(hpage);
3783 int nid = page_to_nid(hpage);
3784
3785 list_for_each_entry_safe(page, tmp, &h->hugepage_freelists[nid], lru)
3786 if (page == hpage)
3787 return 1;
3788 return 0;
3789}
3790
Naoya Horiguchi93f70f92010-05-28 09:29:20 +09003791/*
3792 * This function is called from memory failure code.
3793 * Assume the caller holds page lock of the head page.
3794 */
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +09003795int dequeue_hwpoisoned_huge_page(struct page *hpage)
Naoya Horiguchi93f70f92010-05-28 09:29:20 +09003796{
3797 struct hstate *h = page_hstate(hpage);
3798 int nid = page_to_nid(hpage);
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +09003799 int ret = -EBUSY;
Naoya Horiguchi93f70f92010-05-28 09:29:20 +09003800
3801 spin_lock(&hugetlb_lock);
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +09003802 if (is_hugepage_on_freelist(hpage)) {
Naoya Horiguchi56f2fb12012-12-12 13:52:33 -08003803 /*
3804 * Hwpoisoned hugepage isn't linked to activelist or freelist,
3805 * but dangling hpage->lru can trigger list-debug warnings
3806 * (this happens when we call unpoison_memory() on it),
3807 * so let it point to itself with list_del_init().
3808 */
3809 list_del_init(&hpage->lru);
Naoya Horiguchi8c6c2ec2010-09-08 10:19:38 +09003810 set_page_refcounted(hpage);
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +09003811 h->free_huge_pages--;
3812 h->free_huge_pages_node[nid]--;
3813 ret = 0;
3814 }
Naoya Horiguchi93f70f92010-05-28 09:29:20 +09003815 spin_unlock(&hugetlb_lock);
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +09003816 return ret;
Naoya Horiguchi93f70f92010-05-28 09:29:20 +09003817}
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +09003818#endif
Naoya Horiguchi31caf662013-09-11 14:21:59 -07003819
3820bool isolate_huge_page(struct page *page, struct list_head *list)
3821{
Sasha Levin309381fea2014-01-23 15:52:54 -08003822 VM_BUG_ON_PAGE(!PageHead(page), page);
Naoya Horiguchi31caf662013-09-11 14:21:59 -07003823 if (!get_page_unless_zero(page))
3824 return false;
3825 spin_lock(&hugetlb_lock);
3826 list_move_tail(&page->lru, list);
3827 spin_unlock(&hugetlb_lock);
3828 return true;
3829}
3830
3831void putback_active_hugepage(struct page *page)
3832{
Sasha Levin309381fea2014-01-23 15:52:54 -08003833 VM_BUG_ON_PAGE(!PageHead(page), page);
Naoya Horiguchi31caf662013-09-11 14:21:59 -07003834 spin_lock(&hugetlb_lock);
3835 list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist);
3836 spin_unlock(&hugetlb_lock);
3837 put_page(page);
3838}
Naoya Horiguchic8721bb2013-09-11 14:22:09 -07003839
3840bool is_hugepage_active(struct page *page)
3841{
Sasha Levin309381fea2014-01-23 15:52:54 -08003842 VM_BUG_ON_PAGE(!PageHuge(page), page);
Naoya Horiguchic8721bb2013-09-11 14:22:09 -07003843 /*
3844 * This function can be called for a tail page because the caller,
3845 * scan_movable_pages, scans through a given pfn-range which typically
3846 * covers one memory block. In systems using gigantic hugepage (1GB
3847 * for x86_64,) a hugepage is larger than a memory block, and we don't
3848 * support migrating such large hugepages for now, so return false
3849 * when called for tail pages.
3850 */
3851 if (PageTail(page))
3852 return false;
3853 /*
3854 * Refcount of a hwpoisoned hugepages is 1, but they are not active,
3855 * so we should return false for them.
3856 */
3857 if (unlikely(PageHWPoison(page)))
3858 return false;
3859 return page_count(page) > 0;
3860}