blob: 7a0a73d2fcff128850b32af9910a873d6fb384f5 [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
38const unsigned long hugetlb_zero = 0, hugetlb_infinity = ~0UL;
Mel Gorman396faf02007-07-17 04:03:13 -070039unsigned long hugepages_treat_as_movable;
Andi Kleena5516432008-07-23 21:27:41 -070040
Aneesh Kumar K.Vc3f38a32012-07-31 16:42:10 -070041int hugetlb_max_hstate __read_mostly;
Andi Kleene5ff2152008-07-23 21:27:42 -070042unsigned int default_hstate_idx;
43struct hstate hstates[HUGE_MAX_HSTATE];
44
Jon Tollefson53ba51d2008-07-23 21:27:52 -070045__initdata LIST_HEAD(huge_boot_pages);
46
Andi Kleene5ff2152008-07-23 21:27:42 -070047/* for command line parsing */
48static struct hstate * __initdata parsed_hstate;
49static unsigned long __initdata default_hstate_max_huge_pages;
Nick Piggine11bfbf2008-07-23 21:27:52 -070050static unsigned long __initdata default_hstate_size;
Andi Kleene5ff2152008-07-23 21:27:42 -070051
David Gibson3935baa2006-03-22 00:08:53 -080052/*
Naoya Horiguchi31caf662013-09-11 14:21:59 -070053 * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages,
54 * free_huge_pages, and surplus_huge_pages.
David Gibson3935baa2006-03-22 00:08:53 -080055 */
Aneesh Kumar K.Vc3f38a32012-07-31 16:42:10 -070056DEFINE_SPINLOCK(hugetlb_lock);
Eric Paris0bd0f9f2005-11-21 21:32:28 -080057
Davidlohr Bueso8382d912014-04-03 14:47:31 -070058/*
59 * Serializes faults on the same logical page. This is used to
60 * prevent spurious OOMs when the hugepage pool is fully utilized.
61 */
62static int num_fault_mutexes;
63static struct mutex *htlb_fault_mutex_table ____cacheline_aligned_in_smp;
64
David Gibson90481622012-03-21 16:34:12 -070065static inline void unlock_or_release_subpool(struct hugepage_subpool *spool)
66{
67 bool free = (spool->count == 0) && (spool->used_hpages == 0);
68
69 spin_unlock(&spool->lock);
70
71 /* If no pages are used, and no other handles to the subpool
72 * remain, free the subpool the subpool remain */
73 if (free)
74 kfree(spool);
75}
76
77struct hugepage_subpool *hugepage_new_subpool(long nr_blocks)
78{
79 struct hugepage_subpool *spool;
80
81 spool = kmalloc(sizeof(*spool), GFP_KERNEL);
82 if (!spool)
83 return NULL;
84
85 spin_lock_init(&spool->lock);
86 spool->count = 1;
87 spool->max_hpages = nr_blocks;
88 spool->used_hpages = 0;
89
90 return spool;
91}
92
93void hugepage_put_subpool(struct hugepage_subpool *spool)
94{
95 spin_lock(&spool->lock);
96 BUG_ON(!spool->count);
97 spool->count--;
98 unlock_or_release_subpool(spool);
99}
100
101static int hugepage_subpool_get_pages(struct hugepage_subpool *spool,
102 long delta)
103{
104 int ret = 0;
105
106 if (!spool)
107 return 0;
108
109 spin_lock(&spool->lock);
110 if ((spool->used_hpages + delta) <= spool->max_hpages) {
111 spool->used_hpages += delta;
112 } else {
113 ret = -ENOMEM;
114 }
115 spin_unlock(&spool->lock);
116
117 return ret;
118}
119
120static void hugepage_subpool_put_pages(struct hugepage_subpool *spool,
121 long delta)
122{
123 if (!spool)
124 return;
125
126 spin_lock(&spool->lock);
127 spool->used_hpages -= delta;
128 /* If hugetlbfs_put_super couldn't free spool due to
129 * an outstanding quota reference, free it now. */
130 unlock_or_release_subpool(spool);
131}
132
133static inline struct hugepage_subpool *subpool_inode(struct inode *inode)
134{
135 return HUGETLBFS_SB(inode->i_sb)->spool;
136}
137
138static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma)
139{
Al Viro496ad9a2013-01-23 17:07:38 -0500140 return subpool_inode(file_inode(vma->vm_file));
David Gibson90481622012-03-21 16:34:12 -0700141}
142
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700143/*
Andy Whitcroft96822902008-07-23 21:27:29 -0700144 * Region tracking -- allows tracking of reservations and instantiated pages
145 * across the pages in a mapping.
Andy Whitcroft84afd992008-07-23 21:27:32 -0700146 *
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700147 * The region data structures are embedded into a resv_map and
148 * protected by a resv_map's lock
Andy Whitcroft96822902008-07-23 21:27:29 -0700149 */
150struct file_region {
151 struct list_head link;
152 long from;
153 long to;
154};
155
Joonsoo Kim1406ec92014-04-03 14:47:26 -0700156static long region_add(struct resv_map *resv, long f, long t)
Andy Whitcroft96822902008-07-23 21:27:29 -0700157{
Joonsoo Kim1406ec92014-04-03 14:47:26 -0700158 struct list_head *head = &resv->regions;
Andy Whitcroft96822902008-07-23 21:27:29 -0700159 struct file_region *rg, *nrg, *trg;
160
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700161 spin_lock(&resv->lock);
Andy Whitcroft96822902008-07-23 21:27:29 -0700162 /* Locate the region we are either in or before. */
163 list_for_each_entry(rg, head, link)
164 if (f <= rg->to)
165 break;
166
167 /* Round our left edge to the current segment if it encloses us. */
168 if (f > rg->from)
169 f = rg->from;
170
171 /* Check for and consume any regions we now overlap with. */
172 nrg = rg;
173 list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
174 if (&rg->link == head)
175 break;
176 if (rg->from > t)
177 break;
178
179 /* If this area reaches higher then extend our area to
180 * include it completely. If this is not the first area
181 * which we intend to reuse, free it. */
182 if (rg->to > t)
183 t = rg->to;
184 if (rg != nrg) {
185 list_del(&rg->link);
186 kfree(rg);
187 }
188 }
189 nrg->from = f;
190 nrg->to = t;
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700191 spin_unlock(&resv->lock);
Andy Whitcroft96822902008-07-23 21:27:29 -0700192 return 0;
193}
194
Joonsoo Kim1406ec92014-04-03 14:47:26 -0700195static long region_chg(struct resv_map *resv, long f, long t)
Andy Whitcroft96822902008-07-23 21:27:29 -0700196{
Joonsoo Kim1406ec92014-04-03 14:47:26 -0700197 struct list_head *head = &resv->regions;
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700198 struct file_region *rg, *nrg = NULL;
Andy Whitcroft96822902008-07-23 21:27:29 -0700199 long chg = 0;
200
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700201retry:
202 spin_lock(&resv->lock);
Andy Whitcroft96822902008-07-23 21:27:29 -0700203 /* Locate the region we are before or in. */
204 list_for_each_entry(rg, head, link)
205 if (f <= rg->to)
206 break;
207
208 /* If we are below the current region then a new region is required.
209 * Subtle, allocate a new region at the position but make it zero
210 * size such that we can guarantee to record the reservation. */
211 if (&rg->link == head || t < rg->from) {
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700212 if (!nrg) {
213 spin_unlock(&resv->lock);
214 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
215 if (!nrg)
216 return -ENOMEM;
Andy Whitcroft96822902008-07-23 21:27:29 -0700217
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700218 nrg->from = f;
219 nrg->to = f;
220 INIT_LIST_HEAD(&nrg->link);
221 goto retry;
222 }
223
224 list_add(&nrg->link, rg->link.prev);
225 chg = t - f;
226 goto out_nrg;
Andy Whitcroft96822902008-07-23 21:27:29 -0700227 }
228
229 /* Round our left edge to the current segment if it encloses us. */
230 if (f > rg->from)
231 f = rg->from;
232 chg = t - f;
233
234 /* Check for and consume any regions we now overlap with. */
235 list_for_each_entry(rg, rg->link.prev, link) {
236 if (&rg->link == head)
237 break;
238 if (rg->from > t)
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700239 goto out;
Andy Whitcroft96822902008-07-23 21:27:29 -0700240
Lucas De Marchi25985ed2011-03-30 22:57:33 -0300241 /* We overlap with this area, if it extends further than
Andy Whitcroft96822902008-07-23 21:27:29 -0700242 * us then we must extend ourselves. Account for its
243 * existing reservation. */
244 if (rg->to > t) {
245 chg += rg->to - t;
246 t = rg->to;
247 }
248 chg -= rg->to - rg->from;
249 }
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700250
251out:
252 spin_unlock(&resv->lock);
253 /* We already know we raced and no longer need the new region */
254 kfree(nrg);
255 return chg;
256out_nrg:
257 spin_unlock(&resv->lock);
Andy Whitcroft96822902008-07-23 21:27:29 -0700258 return chg;
259}
260
Joonsoo Kim1406ec92014-04-03 14:47:26 -0700261static long region_truncate(struct resv_map *resv, long end)
Andy Whitcroft96822902008-07-23 21:27:29 -0700262{
Joonsoo Kim1406ec92014-04-03 14:47:26 -0700263 struct list_head *head = &resv->regions;
Andy Whitcroft96822902008-07-23 21:27:29 -0700264 struct file_region *rg, *trg;
265 long chg = 0;
266
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700267 spin_lock(&resv->lock);
Andy Whitcroft96822902008-07-23 21:27:29 -0700268 /* Locate the region we are either in or before. */
269 list_for_each_entry(rg, head, link)
270 if (end <= rg->to)
271 break;
272 if (&rg->link == head)
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700273 goto out;
Andy Whitcroft96822902008-07-23 21:27:29 -0700274
275 /* If we are in the middle of a region then adjust it. */
276 if (end > rg->from) {
277 chg = rg->to - end;
278 rg->to = end;
279 rg = list_entry(rg->link.next, typeof(*rg), link);
280 }
281
282 /* Drop any remaining regions. */
283 list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
284 if (&rg->link == head)
285 break;
286 chg += rg->to - rg->from;
287 list_del(&rg->link);
288 kfree(rg);
289 }
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700290
291out:
292 spin_unlock(&resv->lock);
Andy Whitcroft96822902008-07-23 21:27:29 -0700293 return chg;
294}
295
Joonsoo Kim1406ec92014-04-03 14:47:26 -0700296static long region_count(struct resv_map *resv, long f, long t)
Andy Whitcroft84afd992008-07-23 21:27:32 -0700297{
Joonsoo Kim1406ec92014-04-03 14:47:26 -0700298 struct list_head *head = &resv->regions;
Andy Whitcroft84afd992008-07-23 21:27:32 -0700299 struct file_region *rg;
300 long chg = 0;
301
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700302 spin_lock(&resv->lock);
Andy Whitcroft84afd992008-07-23 21:27:32 -0700303 /* Locate each segment we overlap with, and count that overlap. */
304 list_for_each_entry(rg, head, link) {
Wang Sheng-Huif2135a42012-05-29 15:06:17 -0700305 long seg_from;
306 long seg_to;
Andy Whitcroft84afd992008-07-23 21:27:32 -0700307
308 if (rg->to <= f)
309 continue;
310 if (rg->from >= t)
311 break;
312
313 seg_from = max(rg->from, f);
314 seg_to = min(rg->to, t);
315
316 chg += seg_to - seg_from;
317 }
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700318 spin_unlock(&resv->lock);
Andy Whitcroft84afd992008-07-23 21:27:32 -0700319
320 return chg;
321}
322
Andy Whitcroft96822902008-07-23 21:27:29 -0700323/*
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700324 * Convert the address within this vma to the page offset within
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700325 * the mapping, in pagecache page units; huge pages here.
326 */
Andi Kleena5516432008-07-23 21:27:41 -0700327static pgoff_t vma_hugecache_offset(struct hstate *h,
328 struct vm_area_struct *vma, unsigned long address)
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700329{
Andi Kleena5516432008-07-23 21:27:41 -0700330 return ((address - vma->vm_start) >> huge_page_shift(h)) +
331 (vma->vm_pgoff >> huge_page_order(h));
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700332}
333
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +0900334pgoff_t linear_hugepage_index(struct vm_area_struct *vma,
335 unsigned long address)
336{
337 return vma_hugecache_offset(hstate_vma(vma), vma, address);
338}
339
Andy Whitcroft84afd992008-07-23 21:27:32 -0700340/*
Mel Gorman08fba692009-01-06 14:38:53 -0800341 * Return the size of the pages allocated when backing a VMA. In the majority
342 * cases this will be same size as used by the page table entries.
343 */
344unsigned long vma_kernel_pagesize(struct vm_area_struct *vma)
345{
346 struct hstate *hstate;
347
348 if (!is_vm_hugetlb_page(vma))
349 return PAGE_SIZE;
350
351 hstate = hstate_vma(vma);
352
Wanpeng Li2415cf12013-07-03 15:02:43 -0700353 return 1UL << huge_page_shift(hstate);
Mel Gorman08fba692009-01-06 14:38:53 -0800354}
Joerg Roedelf340ca02009-06-19 15:16:22 +0200355EXPORT_SYMBOL_GPL(vma_kernel_pagesize);
Mel Gorman08fba692009-01-06 14:38:53 -0800356
357/*
Mel Gorman33402892009-01-06 14:38:54 -0800358 * Return the page size being used by the MMU to back a VMA. In the majority
359 * of cases, the page size used by the kernel matches the MMU size. On
360 * architectures where it differs, an architecture-specific version of this
361 * function is required.
362 */
363#ifndef vma_mmu_pagesize
364unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
365{
366 return vma_kernel_pagesize(vma);
367}
368#endif
369
370/*
Andy Whitcroft84afd992008-07-23 21:27:32 -0700371 * Flags for MAP_PRIVATE reservations. These are stored in the bottom
372 * bits of the reservation map pointer, which are always clear due to
373 * alignment.
374 */
375#define HPAGE_RESV_OWNER (1UL << 0)
376#define HPAGE_RESV_UNMAPPED (1UL << 1)
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700377#define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
Andy Whitcroft84afd992008-07-23 21:27:32 -0700378
Mel Gormana1e78772008-07-23 21:27:23 -0700379/*
380 * These helpers are used to track how many pages are reserved for
381 * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
382 * is guaranteed to have their future faults succeed.
383 *
384 * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
385 * the reserve counters are updated with the hugetlb_lock held. It is safe
386 * to reset the VMA at fork() time as it is not in use yet and there is no
387 * chance of the global counters getting corrupted as a result of the values.
Andy Whitcroft84afd992008-07-23 21:27:32 -0700388 *
389 * The private mapping reservation is represented in a subtly different
390 * manner to a shared mapping. A shared mapping has a region map associated
391 * with the underlying file, this region map represents the backing file
392 * pages which have ever had a reservation assigned which this persists even
393 * after the page is instantiated. A private mapping has a region map
394 * associated with the original mmap which is attached to all VMAs which
395 * reference it, this region map represents those offsets which have consumed
396 * reservation ie. where pages have been instantiated.
Mel Gormana1e78772008-07-23 21:27:23 -0700397 */
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700398static unsigned long get_vma_private_data(struct vm_area_struct *vma)
399{
400 return (unsigned long)vma->vm_private_data;
401}
402
403static void set_vma_private_data(struct vm_area_struct *vma,
404 unsigned long value)
405{
406 vma->vm_private_data = (void *)value;
407}
408
Joonsoo Kim9119a412014-04-03 14:47:25 -0700409struct resv_map *resv_map_alloc(void)
Andy Whitcroft84afd992008-07-23 21:27:32 -0700410{
411 struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
412 if (!resv_map)
413 return NULL;
414
415 kref_init(&resv_map->refs);
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700416 spin_lock_init(&resv_map->lock);
Andy Whitcroft84afd992008-07-23 21:27:32 -0700417 INIT_LIST_HEAD(&resv_map->regions);
418
419 return resv_map;
420}
421
Joonsoo Kim9119a412014-04-03 14:47:25 -0700422void resv_map_release(struct kref *ref)
Andy Whitcroft84afd992008-07-23 21:27:32 -0700423{
424 struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
425
426 /* Clear out any active regions before we release the map. */
Joonsoo Kim1406ec92014-04-03 14:47:26 -0700427 region_truncate(resv_map, 0);
Andy Whitcroft84afd992008-07-23 21:27:32 -0700428 kfree(resv_map);
429}
430
Joonsoo Kim4e35f482014-04-03 14:47:30 -0700431static inline struct resv_map *inode_resv_map(struct inode *inode)
432{
433 return inode->i_mapping->private_data;
434}
435
Andy Whitcroft84afd992008-07-23 21:27:32 -0700436static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
Mel Gormana1e78772008-07-23 21:27:23 -0700437{
438 VM_BUG_ON(!is_vm_hugetlb_page(vma));
Joonsoo Kim4e35f482014-04-03 14:47:30 -0700439 if (vma->vm_flags & VM_MAYSHARE) {
440 struct address_space *mapping = vma->vm_file->f_mapping;
441 struct inode *inode = mapping->host;
442
443 return inode_resv_map(inode);
444
445 } else {
Andy Whitcroft84afd992008-07-23 21:27:32 -0700446 return (struct resv_map *)(get_vma_private_data(vma) &
447 ~HPAGE_RESV_MASK);
Joonsoo Kim4e35f482014-04-03 14:47:30 -0700448 }
Mel Gormana1e78772008-07-23 21:27:23 -0700449}
450
Andy Whitcroft84afd992008-07-23 21:27:32 -0700451static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
Mel Gormana1e78772008-07-23 21:27:23 -0700452{
453 VM_BUG_ON(!is_vm_hugetlb_page(vma));
Mel Gormanf83a2752009-05-28 14:34:40 -0700454 VM_BUG_ON(vma->vm_flags & VM_MAYSHARE);
Mel Gormana1e78772008-07-23 21:27:23 -0700455
Andy Whitcroft84afd992008-07-23 21:27:32 -0700456 set_vma_private_data(vma, (get_vma_private_data(vma) &
457 HPAGE_RESV_MASK) | (unsigned long)map);
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700458}
459
460static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
461{
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700462 VM_BUG_ON(!is_vm_hugetlb_page(vma));
Mel Gormanf83a2752009-05-28 14:34:40 -0700463 VM_BUG_ON(vma->vm_flags & VM_MAYSHARE);
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700464
465 set_vma_private_data(vma, get_vma_private_data(vma) | flags);
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700466}
467
468static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
469{
470 VM_BUG_ON(!is_vm_hugetlb_page(vma));
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700471
472 return (get_vma_private_data(vma) & flag) != 0;
Mel Gormana1e78772008-07-23 21:27:23 -0700473}
474
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700475/* Reset counters to 0 and clear all HPAGE_RESV_* flags */
Mel Gormana1e78772008-07-23 21:27:23 -0700476void reset_vma_resv_huge_pages(struct vm_area_struct *vma)
477{
478 VM_BUG_ON(!is_vm_hugetlb_page(vma));
Mel Gormanf83a2752009-05-28 14:34:40 -0700479 if (!(vma->vm_flags & VM_MAYSHARE))
Mel Gormana1e78772008-07-23 21:27:23 -0700480 vma->vm_private_data = (void *)0;
481}
482
483/* Returns true if the VMA has associated reserve pages */
Joonsoo Kimaf0ed732013-09-11 14:21:18 -0700484static int vma_has_reserves(struct vm_area_struct *vma, long chg)
Mel Gormana1e78772008-07-23 21:27:23 -0700485{
Joonsoo Kimaf0ed732013-09-11 14:21:18 -0700486 if (vma->vm_flags & VM_NORESERVE) {
487 /*
488 * This address is already reserved by other process(chg == 0),
489 * so, we should decrement reserved count. Without decrementing,
490 * reserve count remains after releasing inode, because this
491 * allocated page will go into page cache and is regarded as
492 * coming from reserved pool in releasing step. Currently, we
493 * don't have any other solution to deal with this situation
494 * properly, so add work-around here.
495 */
496 if (vma->vm_flags & VM_MAYSHARE && chg == 0)
497 return 1;
498 else
499 return 0;
500 }
Joonsoo Kima63884e2013-09-11 14:21:07 -0700501
502 /* Shared mappings always use reserves */
Mel Gormanf83a2752009-05-28 14:34:40 -0700503 if (vma->vm_flags & VM_MAYSHARE)
Mel Gorman7f09ca52008-07-23 21:27:58 -0700504 return 1;
Joonsoo Kima63884e2013-09-11 14:21:07 -0700505
506 /*
507 * Only the process that called mmap() has reserves for
508 * private mappings.
509 */
Mel Gorman7f09ca52008-07-23 21:27:58 -0700510 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER))
511 return 1;
Joonsoo Kima63884e2013-09-11 14:21:07 -0700512
Mel Gorman7f09ca52008-07-23 21:27:58 -0700513 return 0;
Mel Gormana1e78772008-07-23 21:27:23 -0700514}
515
Andi Kleena5516432008-07-23 21:27:41 -0700516static void enqueue_huge_page(struct hstate *h, struct page *page)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700517{
518 int nid = page_to_nid(page);
Aneesh Kumar K.V0edaecf2012-07-31 16:42:07 -0700519 list_move(&page->lru, &h->hugepage_freelists[nid]);
Andi Kleena5516432008-07-23 21:27:41 -0700520 h->free_huge_pages++;
521 h->free_huge_pages_node[nid]++;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700522}
523
Naoya Horiguchibf50bab2010-09-08 10:19:33 +0900524static struct page *dequeue_huge_page_node(struct hstate *h, int nid)
525{
526 struct page *page;
527
Naoya Horiguchic8721bb2013-09-11 14:22:09 -0700528 list_for_each_entry(page, &h->hugepage_freelists[nid], lru)
529 if (!is_migrate_isolate_page(page))
530 break;
531 /*
532 * if 'non-isolated free hugepage' not found on the list,
533 * the allocation fails.
534 */
535 if (&h->hugepage_freelists[nid] == &page->lru)
Naoya Horiguchibf50bab2010-09-08 10:19:33 +0900536 return NULL;
Aneesh Kumar K.V0edaecf2012-07-31 16:42:07 -0700537 list_move(&page->lru, &h->hugepage_activelist);
Naoya Horiguchia9869b82010-09-08 10:19:37 +0900538 set_page_refcounted(page);
Naoya Horiguchibf50bab2010-09-08 10:19:33 +0900539 h->free_huge_pages--;
540 h->free_huge_pages_node[nid]--;
541 return page;
542}
543
Naoya Horiguchi86cdb462013-09-11 14:22:13 -0700544/* Movability of hugepages depends on migration support. */
545static inline gfp_t htlb_alloc_mask(struct hstate *h)
546{
Naoya Horiguchi100873d2014-06-04 16:10:56 -0700547 if (hugepages_treat_as_movable || hugepage_migration_supported(h))
Naoya Horiguchi86cdb462013-09-11 14:22:13 -0700548 return GFP_HIGHUSER_MOVABLE;
549 else
550 return GFP_HIGHUSER;
551}
552
Andi Kleena5516432008-07-23 21:27:41 -0700553static struct page *dequeue_huge_page_vma(struct hstate *h,
554 struct vm_area_struct *vma,
Joonsoo Kimaf0ed732013-09-11 14:21:18 -0700555 unsigned long address, int avoid_reserve,
556 long chg)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700557{
Konstantin Khlebnikovb1c12cb2012-04-25 16:01:46 -0700558 struct page *page = NULL;
Lee Schermerhorn480eccf2007-09-18 22:46:47 -0700559 struct mempolicy *mpol;
Mel Gorman19770b32008-04-28 02:12:18 -0700560 nodemask_t *nodemask;
Miao Xiec0ff7452010-05-24 14:32:08 -0700561 struct zonelist *zonelist;
Mel Gormandd1a2392008-04-28 02:12:17 -0700562 struct zone *zone;
563 struct zoneref *z;
Mel Gormancc9a6c82012-03-21 16:34:11 -0700564 unsigned int cpuset_mems_cookie;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700565
Mel Gormana1e78772008-07-23 21:27:23 -0700566 /*
567 * A child process with MAP_PRIVATE mappings created by their parent
568 * have no page reserves. This check ensures that reservations are
569 * not "stolen". The child may still get SIGKILLed
570 */
Joonsoo Kimaf0ed732013-09-11 14:21:18 -0700571 if (!vma_has_reserves(vma, chg) &&
Andi Kleena5516432008-07-23 21:27:41 -0700572 h->free_huge_pages - h->resv_huge_pages == 0)
Miao Xiec0ff7452010-05-24 14:32:08 -0700573 goto err;
Mel Gormana1e78772008-07-23 21:27:23 -0700574
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700575 /* If reserves cannot be used, ensure enough pages are in the pool */
Andi Kleena5516432008-07-23 21:27:41 -0700576 if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
Justin P. Mattock6eab04a2011-04-08 19:49:08 -0700577 goto err;
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700578
Joonsoo Kim9966c4b2013-09-11 14:20:50 -0700579retry_cpuset:
Mel Gormand26914d2014-04-03 14:47:24 -0700580 cpuset_mems_cookie = read_mems_allowed_begin();
Joonsoo Kim9966c4b2013-09-11 14:20:50 -0700581 zonelist = huge_zonelist(vma, address,
Naoya Horiguchi86cdb462013-09-11 14:22:13 -0700582 htlb_alloc_mask(h), &mpol, &nodemask);
Joonsoo Kim9966c4b2013-09-11 14:20:50 -0700583
Mel Gorman19770b32008-04-28 02:12:18 -0700584 for_each_zone_zonelist_nodemask(zone, z, zonelist,
585 MAX_NR_ZONES - 1, nodemask) {
Naoya Horiguchi86cdb462013-09-11 14:22:13 -0700586 if (cpuset_zone_allowed_softwall(zone, htlb_alloc_mask(h))) {
Naoya Horiguchibf50bab2010-09-08 10:19:33 +0900587 page = dequeue_huge_page_node(h, zone_to_nid(zone));
588 if (page) {
Joonsoo Kimaf0ed732013-09-11 14:21:18 -0700589 if (avoid_reserve)
590 break;
591 if (!vma_has_reserves(vma, chg))
592 break;
593
Joonsoo Kim07443a82013-09-11 14:21:58 -0700594 SetPagePrivate(page);
Joonsoo Kimaf0ed732013-09-11 14:21:18 -0700595 h->resv_huge_pages--;
Naoya Horiguchibf50bab2010-09-08 10:19:33 +0900596 break;
597 }
Andrew Morton3abf7af2007-07-19 01:49:08 -0700598 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700599 }
Mel Gormancc9a6c82012-03-21 16:34:11 -0700600
601 mpol_cond_put(mpol);
Mel Gormand26914d2014-04-03 14:47:24 -0700602 if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie)))
Mel Gormancc9a6c82012-03-21 16:34:11 -0700603 goto retry_cpuset;
604 return page;
605
Miao Xiec0ff7452010-05-24 14:32:08 -0700606err:
Mel Gormancc9a6c82012-03-21 16:34:11 -0700607 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700608}
609
Luiz Capitulino1cac6f22014-06-04 16:07:11 -0700610/*
611 * common helper functions for hstate_next_node_to_{alloc|free}.
612 * We may have allocated or freed a huge page based on a different
613 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
614 * be outside of *nodes_allowed. Ensure that we use an allowed
615 * node for alloc or free.
616 */
617static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
618{
619 nid = next_node(nid, *nodes_allowed);
620 if (nid == MAX_NUMNODES)
621 nid = first_node(*nodes_allowed);
622 VM_BUG_ON(nid >= MAX_NUMNODES);
623
624 return nid;
625}
626
627static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
628{
629 if (!node_isset(nid, *nodes_allowed))
630 nid = next_node_allowed(nid, nodes_allowed);
631 return nid;
632}
633
634/*
635 * returns the previously saved node ["this node"] from which to
636 * allocate a persistent huge page for the pool and advance the
637 * next node from which to allocate, handling wrap at end of node
638 * mask.
639 */
640static int hstate_next_node_to_alloc(struct hstate *h,
641 nodemask_t *nodes_allowed)
642{
643 int nid;
644
645 VM_BUG_ON(!nodes_allowed);
646
647 nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
648 h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
649
650 return nid;
651}
652
653/*
654 * helper for free_pool_huge_page() - return the previously saved
655 * node ["this node"] from which to free a huge page. Advance the
656 * next node id whether or not we find a free huge page to free so
657 * that the next attempt to free addresses the next node.
658 */
659static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
660{
661 int nid;
662
663 VM_BUG_ON(!nodes_allowed);
664
665 nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
666 h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
667
668 return nid;
669}
670
671#define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \
672 for (nr_nodes = nodes_weight(*mask); \
673 nr_nodes > 0 && \
674 ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \
675 nr_nodes--)
676
677#define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \
678 for (nr_nodes = nodes_weight(*mask); \
679 nr_nodes > 0 && \
680 ((node = hstate_next_node_to_free(hs, mask)) || 1); \
681 nr_nodes--)
682
Luiz Capitulino944d9fe2014-06-04 16:07:13 -0700683#if defined(CONFIG_CMA) && defined(CONFIG_X86_64)
684static void destroy_compound_gigantic_page(struct page *page,
685 unsigned long order)
686{
687 int i;
688 int nr_pages = 1 << order;
689 struct page *p = page + 1;
690
691 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
692 __ClearPageTail(p);
693 set_page_refcounted(p);
694 p->first_page = NULL;
695 }
696
697 set_compound_order(page, 0);
698 __ClearPageHead(page);
699}
700
701static void free_gigantic_page(struct page *page, unsigned order)
702{
703 free_contig_range(page_to_pfn(page), 1 << order);
704}
705
706static int __alloc_gigantic_page(unsigned long start_pfn,
707 unsigned long nr_pages)
708{
709 unsigned long end_pfn = start_pfn + nr_pages;
710 return alloc_contig_range(start_pfn, end_pfn, MIGRATE_MOVABLE);
711}
712
713static bool pfn_range_valid_gigantic(unsigned long start_pfn,
714 unsigned long nr_pages)
715{
716 unsigned long i, end_pfn = start_pfn + nr_pages;
717 struct page *page;
718
719 for (i = start_pfn; i < end_pfn; i++) {
720 if (!pfn_valid(i))
721 return false;
722
723 page = pfn_to_page(i);
724
725 if (PageReserved(page))
726 return false;
727
728 if (page_count(page) > 0)
729 return false;
730
731 if (PageHuge(page))
732 return false;
733 }
734
735 return true;
736}
737
738static bool zone_spans_last_pfn(const struct zone *zone,
739 unsigned long start_pfn, unsigned long nr_pages)
740{
741 unsigned long last_pfn = start_pfn + nr_pages - 1;
742 return zone_spans_pfn(zone, last_pfn);
743}
744
745static struct page *alloc_gigantic_page(int nid, unsigned order)
746{
747 unsigned long nr_pages = 1 << order;
748 unsigned long ret, pfn, flags;
749 struct zone *z;
750
751 z = NODE_DATA(nid)->node_zones;
752 for (; z - NODE_DATA(nid)->node_zones < MAX_NR_ZONES; z++) {
753 spin_lock_irqsave(&z->lock, flags);
754
755 pfn = ALIGN(z->zone_start_pfn, nr_pages);
756 while (zone_spans_last_pfn(z, pfn, nr_pages)) {
757 if (pfn_range_valid_gigantic(pfn, nr_pages)) {
758 /*
759 * We release the zone lock here because
760 * alloc_contig_range() will also lock the zone
761 * at some point. If there's an allocation
762 * spinning on this lock, it may win the race
763 * and cause alloc_contig_range() to fail...
764 */
765 spin_unlock_irqrestore(&z->lock, flags);
766 ret = __alloc_gigantic_page(pfn, nr_pages);
767 if (!ret)
768 return pfn_to_page(pfn);
769 spin_lock_irqsave(&z->lock, flags);
770 }
771 pfn += nr_pages;
772 }
773
774 spin_unlock_irqrestore(&z->lock, flags);
775 }
776
777 return NULL;
778}
779
780static void prep_new_huge_page(struct hstate *h, struct page *page, int nid);
781static void prep_compound_gigantic_page(struct page *page, unsigned long order);
782
783static struct page *alloc_fresh_gigantic_page_node(struct hstate *h, int nid)
784{
785 struct page *page;
786
787 page = alloc_gigantic_page(nid, huge_page_order(h));
788 if (page) {
789 prep_compound_gigantic_page(page, huge_page_order(h));
790 prep_new_huge_page(h, page, nid);
791 }
792
793 return page;
794}
795
796static int alloc_fresh_gigantic_page(struct hstate *h,
797 nodemask_t *nodes_allowed)
798{
799 struct page *page = NULL;
800 int nr_nodes, node;
801
802 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
803 page = alloc_fresh_gigantic_page_node(h, node);
804 if (page)
805 return 1;
806 }
807
808 return 0;
809}
810
811static inline bool gigantic_page_supported(void) { return true; }
812#else
813static inline bool gigantic_page_supported(void) { return false; }
814static inline void free_gigantic_page(struct page *page, unsigned order) { }
815static inline void destroy_compound_gigantic_page(struct page *page,
816 unsigned long order) { }
817static inline int alloc_fresh_gigantic_page(struct hstate *h,
818 nodemask_t *nodes_allowed) { return 0; }
819#endif
820
Andi Kleena5516432008-07-23 21:27:41 -0700821static void update_and_free_page(struct hstate *h, struct page *page)
Adam Litke6af2acb2007-10-16 01:26:16 -0700822{
823 int i;
Andi Kleena5516432008-07-23 21:27:41 -0700824
Luiz Capitulino944d9fe2014-06-04 16:07:13 -0700825 if (hstate_is_gigantic(h) && !gigantic_page_supported())
826 return;
Andy Whitcroft18229df2008-11-06 12:53:27 -0800827
Andi Kleena5516432008-07-23 21:27:41 -0700828 h->nr_huge_pages--;
829 h->nr_huge_pages_node[page_to_nid(page)]--;
830 for (i = 0; i < pages_per_huge_page(h); i++) {
Chris Forbes32f84522011-07-25 17:12:14 -0700831 page[i].flags &= ~(1 << PG_locked | 1 << PG_error |
832 1 << PG_referenced | 1 << PG_dirty |
Luiz Capitulinoa7407a22014-06-04 16:07:09 -0700833 1 << PG_active | 1 << PG_private |
834 1 << PG_writeback);
Adam Litke6af2acb2007-10-16 01:26:16 -0700835 }
Sasha Levin309381fea2014-01-23 15:52:54 -0800836 VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page);
Adam Litke6af2acb2007-10-16 01:26:16 -0700837 set_compound_page_dtor(page, NULL);
838 set_page_refcounted(page);
Luiz Capitulino944d9fe2014-06-04 16:07:13 -0700839 if (hstate_is_gigantic(h)) {
840 destroy_compound_gigantic_page(page, huge_page_order(h));
841 free_gigantic_page(page, huge_page_order(h));
842 } else {
843 arch_release_hugepage(page);
844 __free_pages(page, huge_page_order(h));
845 }
Adam Litke6af2acb2007-10-16 01:26:16 -0700846}
847
Andi Kleene5ff2152008-07-23 21:27:42 -0700848struct hstate *size_to_hstate(unsigned long size)
849{
850 struct hstate *h;
851
852 for_each_hstate(h) {
853 if (huge_page_size(h) == size)
854 return h;
855 }
856 return NULL;
857}
858
Atsushi Kumagai8f1d26d2014-07-30 16:08:39 -0700859void free_huge_page(struct page *page)
David Gibson27a85ef2006-03-22 00:08:56 -0800860{
Andi Kleena5516432008-07-23 21:27:41 -0700861 /*
862 * Can't pass hstate in here because it is called from the
863 * compound page destructor.
864 */
Andi Kleene5ff2152008-07-23 21:27:42 -0700865 struct hstate *h = page_hstate(page);
Adam Litke7893d1d2007-10-16 01:26:18 -0700866 int nid = page_to_nid(page);
David Gibson90481622012-03-21 16:34:12 -0700867 struct hugepage_subpool *spool =
868 (struct hugepage_subpool *)page_private(page);
Joonsoo Kim07443a82013-09-11 14:21:58 -0700869 bool restore_reserve;
David Gibson27a85ef2006-03-22 00:08:56 -0800870
Andy Whitcrofte5df70a2008-02-23 15:23:32 -0800871 set_page_private(page, 0);
Mel Gorman23be7462010-04-23 13:17:56 -0400872 page->mapping = NULL;
Adam Litke7893d1d2007-10-16 01:26:18 -0700873 BUG_ON(page_count(page));
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +0900874 BUG_ON(page_mapcount(page));
Joonsoo Kim07443a82013-09-11 14:21:58 -0700875 restore_reserve = PagePrivate(page);
Joonsoo Kim16c794b2013-10-16 13:46:48 -0700876 ClearPagePrivate(page);
David Gibson27a85ef2006-03-22 00:08:56 -0800877
878 spin_lock(&hugetlb_lock);
Aneesh Kumar K.V6d76dcf2012-07-31 16:42:18 -0700879 hugetlb_cgroup_uncharge_page(hstate_index(h),
880 pages_per_huge_page(h), page);
Joonsoo Kim07443a82013-09-11 14:21:58 -0700881 if (restore_reserve)
882 h->resv_huge_pages++;
883
Luiz Capitulino944d9fe2014-06-04 16:07:13 -0700884 if (h->surplus_huge_pages_node[nid]) {
Aneesh Kumar K.V0edaecf2012-07-31 16:42:07 -0700885 /* remove the page from active list */
886 list_del(&page->lru);
Andi Kleena5516432008-07-23 21:27:41 -0700887 update_and_free_page(h, page);
888 h->surplus_huge_pages--;
889 h->surplus_huge_pages_node[nid]--;
Adam Litke7893d1d2007-10-16 01:26:18 -0700890 } else {
Will Deacon5d3a5512012-10-08 16:29:32 -0700891 arch_clear_hugepage_flags(page);
Andi Kleena5516432008-07-23 21:27:41 -0700892 enqueue_huge_page(h, page);
Adam Litke7893d1d2007-10-16 01:26:18 -0700893 }
David Gibson27a85ef2006-03-22 00:08:56 -0800894 spin_unlock(&hugetlb_lock);
David Gibson90481622012-03-21 16:34:12 -0700895 hugepage_subpool_put_pages(spool, 1);
David Gibson27a85ef2006-03-22 00:08:56 -0800896}
897
Andi Kleena5516432008-07-23 21:27:41 -0700898static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
Andi Kleenb7ba30c2008-07-23 21:27:40 -0700899{
Aneesh Kumar K.V0edaecf2012-07-31 16:42:07 -0700900 INIT_LIST_HEAD(&page->lru);
Andi Kleenb7ba30c2008-07-23 21:27:40 -0700901 set_compound_page_dtor(page, free_huge_page);
902 spin_lock(&hugetlb_lock);
Aneesh Kumar K.V9dd540e2012-07-31 16:42:15 -0700903 set_hugetlb_cgroup(page, NULL);
Andi Kleena5516432008-07-23 21:27:41 -0700904 h->nr_huge_pages++;
905 h->nr_huge_pages_node[nid]++;
Andi Kleenb7ba30c2008-07-23 21:27:40 -0700906 spin_unlock(&hugetlb_lock);
907 put_page(page); /* free it into the hugepage allocator */
908}
909
Luiz Capitulino2906dd52014-06-04 16:07:06 -0700910static void prep_compound_gigantic_page(struct page *page, unsigned long order)
Wu Fengguang20a03072009-06-16 15:32:22 -0700911{
912 int i;
913 int nr_pages = 1 << order;
914 struct page *p = page + 1;
915
916 /* we rely on prep_new_huge_page to set the destructor */
917 set_compound_order(page, order);
918 __SetPageHead(page);
Andrea Arcangelief5a22b2013-10-16 13:46:56 -0700919 __ClearPageReserved(page);
Wu Fengguang20a03072009-06-16 15:32:22 -0700920 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
921 __SetPageTail(p);
Andrea Arcangelief5a22b2013-10-16 13:46:56 -0700922 /*
923 * For gigantic hugepages allocated through bootmem at
924 * boot, it's safer to be consistent with the not-gigantic
925 * hugepages and clear the PG_reserved bit from all tail pages
926 * too. Otherwse drivers using get_user_pages() to access tail
927 * pages may get the reference counting wrong if they see
928 * PG_reserved set on a tail page (despite the head page not
929 * having PG_reserved set). Enforcing this consistency between
930 * head and tail pages allows drivers to optimize away a check
931 * on the head page when they need know if put_page() is needed
932 * after get_user_pages().
933 */
934 __ClearPageReserved(p);
Youquan Song58a84aa2011-12-08 14:34:18 -0800935 set_page_count(p, 0);
Wu Fengguang20a03072009-06-16 15:32:22 -0700936 p->first_page = page;
937 }
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
1092 /* Set scan step to minimum hugepage size */
1093 for_each_hstate(h)
1094 if (order > huge_page_order(h))
1095 order = huge_page_order(h);
1096 VM_BUG_ON(!IS_ALIGNED(start_pfn, 1 << order));
1097 for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << order)
1098 dissolve_free_huge_page(pfn_to_page(pfn));
1099}
1100
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001101static struct page *alloc_buddy_huge_page(struct hstate *h, int nid)
Adam Litke7893d1d2007-10-16 01:26:18 -07001102{
1103 struct page *page;
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001104 unsigned int r_nid;
Adam Litke7893d1d2007-10-16 01:26:18 -07001105
Luiz Capitulinobae7f4a2014-06-04 16:07:08 -07001106 if (hstate_is_gigantic(h))
Andi Kleenaa888a72008-07-23 21:27:47 -07001107 return NULL;
1108
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001109 /*
1110 * Assume we will successfully allocate the surplus page to
1111 * prevent racing processes from causing the surplus to exceed
1112 * overcommit
1113 *
1114 * This however introduces a different race, where a process B
1115 * tries to grow the static hugepage pool while alloc_pages() is
1116 * called by process A. B will only examine the per-node
1117 * counters in determining if surplus huge pages can be
1118 * converted to normal huge pages in adjust_pool_surplus(). A
1119 * won't be able to increment the per-node counter, until the
1120 * lock is dropped by B, but B doesn't drop hugetlb_lock until
1121 * no more huge pages can be converted from surplus to normal
1122 * state (and doesn't try to convert again). Thus, we have a
1123 * case where a surplus huge page exists, the pool is grown, and
1124 * the surplus huge page still exists after, even though it
1125 * should just have been converted to a normal huge page. This
1126 * does not leak memory, though, as the hugepage will be freed
1127 * once it is out of use. It also does not allow the counters to
1128 * go out of whack in adjust_pool_surplus() as we don't modify
1129 * the node values until we've gotten the hugepage and only the
1130 * per-node value is checked there.
1131 */
1132 spin_lock(&hugetlb_lock);
Andi Kleena5516432008-07-23 21:27:41 -07001133 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001134 spin_unlock(&hugetlb_lock);
1135 return NULL;
1136 } else {
Andi Kleena5516432008-07-23 21:27:41 -07001137 h->nr_huge_pages++;
1138 h->surplus_huge_pages++;
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001139 }
1140 spin_unlock(&hugetlb_lock);
1141
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001142 if (nid == NUMA_NO_NODE)
Naoya Horiguchi86cdb462013-09-11 14:22:13 -07001143 page = alloc_pages(htlb_alloc_mask(h)|__GFP_COMP|
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001144 __GFP_REPEAT|__GFP_NOWARN,
1145 huge_page_order(h));
1146 else
1147 page = alloc_pages_exact_node(nid,
Naoya Horiguchi86cdb462013-09-11 14:22:13 -07001148 htlb_alloc_mask(h)|__GFP_COMP|__GFP_THISNODE|
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001149 __GFP_REPEAT|__GFP_NOWARN, huge_page_order(h));
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001150
Gerald Schaefercaff3a22008-08-12 15:08:38 -07001151 if (page && arch_prepare_hugepage(page)) {
1152 __free_pages(page, huge_page_order(h));
Hillf Dantonea5768c2012-01-10 15:08:30 -08001153 page = NULL;
Gerald Schaefercaff3a22008-08-12 15:08:38 -07001154 }
1155
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001156 spin_lock(&hugetlb_lock);
Adam Litke7893d1d2007-10-16 01:26:18 -07001157 if (page) {
Aneesh Kumar K.V0edaecf2012-07-31 16:42:07 -07001158 INIT_LIST_HEAD(&page->lru);
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001159 r_nid = page_to_nid(page);
Adam Litke7893d1d2007-10-16 01:26:18 -07001160 set_compound_page_dtor(page, free_huge_page);
Aneesh Kumar K.V9dd540e2012-07-31 16:42:15 -07001161 set_hugetlb_cgroup(page, NULL);
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001162 /*
1163 * We incremented the global counters already
1164 */
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001165 h->nr_huge_pages_node[r_nid]++;
1166 h->surplus_huge_pages_node[r_nid]++;
Adam Litke3b116302008-04-28 02:13:06 -07001167 __count_vm_event(HTLB_BUDDY_PGALLOC);
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001168 } else {
Andi Kleena5516432008-07-23 21:27:41 -07001169 h->nr_huge_pages--;
1170 h->surplus_huge_pages--;
Adam Litke3b116302008-04-28 02:13:06 -07001171 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
Adam Litke7893d1d2007-10-16 01:26:18 -07001172 }
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001173 spin_unlock(&hugetlb_lock);
Adam Litke7893d1d2007-10-16 01:26:18 -07001174
1175 return page;
1176}
1177
Adam Litkee4e574b2007-10-16 01:26:19 -07001178/*
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001179 * This allocation function is useful in the context where vma is irrelevant.
1180 * E.g. soft-offlining uses this function because it only cares physical
1181 * address of error page.
1182 */
1183struct page *alloc_huge_page_node(struct hstate *h, int nid)
1184{
Joonsoo Kim4ef91842013-09-11 14:21:51 -07001185 struct page *page = NULL;
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001186
1187 spin_lock(&hugetlb_lock);
Joonsoo Kim4ef91842013-09-11 14:21:51 -07001188 if (h->free_huge_pages - h->resv_huge_pages > 0)
1189 page = dequeue_huge_page_node(h, nid);
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001190 spin_unlock(&hugetlb_lock);
1191
Aneesh Kumar K.V94ae8ba2012-07-31 16:42:35 -07001192 if (!page)
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001193 page = alloc_buddy_huge_page(h, nid);
1194
1195 return page;
1196}
1197
1198/*
Lucas De Marchi25985ed2011-03-30 22:57:33 -03001199 * Increase the hugetlb pool such that it can accommodate a reservation
Adam Litkee4e574b2007-10-16 01:26:19 -07001200 * of size 'delta'.
1201 */
Andi Kleena5516432008-07-23 21:27:41 -07001202static int gather_surplus_pages(struct hstate *h, int delta)
Adam Litkee4e574b2007-10-16 01:26:19 -07001203{
1204 struct list_head surplus_list;
1205 struct page *page, *tmp;
1206 int ret, i;
1207 int needed, allocated;
Hillf Danton28073b02012-03-21 16:34:00 -07001208 bool alloc_ok = true;
Adam Litkee4e574b2007-10-16 01:26:19 -07001209
Andi Kleena5516432008-07-23 21:27:41 -07001210 needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
Adam Litkeac09b3a2008-03-04 14:29:38 -08001211 if (needed <= 0) {
Andi Kleena5516432008-07-23 21:27:41 -07001212 h->resv_huge_pages += delta;
Adam Litkee4e574b2007-10-16 01:26:19 -07001213 return 0;
Adam Litkeac09b3a2008-03-04 14:29:38 -08001214 }
Adam Litkee4e574b2007-10-16 01:26:19 -07001215
1216 allocated = 0;
1217 INIT_LIST_HEAD(&surplus_list);
1218
1219 ret = -ENOMEM;
1220retry:
1221 spin_unlock(&hugetlb_lock);
1222 for (i = 0; i < needed; i++) {
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001223 page = alloc_buddy_huge_page(h, NUMA_NO_NODE);
Hillf Danton28073b02012-03-21 16:34:00 -07001224 if (!page) {
1225 alloc_ok = false;
1226 break;
1227 }
Adam Litkee4e574b2007-10-16 01:26:19 -07001228 list_add(&page->lru, &surplus_list);
1229 }
Hillf Danton28073b02012-03-21 16:34:00 -07001230 allocated += i;
Adam Litkee4e574b2007-10-16 01:26:19 -07001231
1232 /*
1233 * After retaking hugetlb_lock, we need to recalculate 'needed'
1234 * because either resv_huge_pages or free_huge_pages may have changed.
1235 */
1236 spin_lock(&hugetlb_lock);
Andi Kleena5516432008-07-23 21:27:41 -07001237 needed = (h->resv_huge_pages + delta) -
1238 (h->free_huge_pages + allocated);
Hillf Danton28073b02012-03-21 16:34:00 -07001239 if (needed > 0) {
1240 if (alloc_ok)
1241 goto retry;
1242 /*
1243 * We were not able to allocate enough pages to
1244 * satisfy the entire reservation so we free what
1245 * we've allocated so far.
1246 */
1247 goto free;
1248 }
Adam Litkee4e574b2007-10-16 01:26:19 -07001249 /*
1250 * The surplus_list now contains _at_least_ the number of extra pages
Lucas De Marchi25985ed2011-03-30 22:57:33 -03001251 * needed to accommodate the reservation. Add the appropriate number
Adam Litkee4e574b2007-10-16 01:26:19 -07001252 * of pages to the hugetlb pool and free the extras back to the buddy
Adam Litkeac09b3a2008-03-04 14:29:38 -08001253 * allocator. Commit the entire reservation here to prevent another
1254 * process from stealing the pages as they are added to the pool but
1255 * before they are reserved.
Adam Litkee4e574b2007-10-16 01:26:19 -07001256 */
1257 needed += allocated;
Andi Kleena5516432008-07-23 21:27:41 -07001258 h->resv_huge_pages += delta;
Adam Litkee4e574b2007-10-16 01:26:19 -07001259 ret = 0;
Naoya Horiguchia9869b82010-09-08 10:19:37 +09001260
Adam Litke19fc3f02008-04-28 02:12:20 -07001261 /* Free the needed pages to the hugetlb pool */
Adam Litkee4e574b2007-10-16 01:26:19 -07001262 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
Adam Litke19fc3f02008-04-28 02:12:20 -07001263 if ((--needed) < 0)
1264 break;
Naoya Horiguchia9869b82010-09-08 10:19:37 +09001265 /*
1266 * This page is now managed by the hugetlb allocator and has
1267 * no users -- drop the buddy allocator's reference.
1268 */
1269 put_page_testzero(page);
Sasha Levin309381fea2014-01-23 15:52:54 -08001270 VM_BUG_ON_PAGE(page_count(page), page);
Andi Kleena5516432008-07-23 21:27:41 -07001271 enqueue_huge_page(h, page);
Adam Litke19fc3f02008-04-28 02:12:20 -07001272 }
Hillf Danton28073b02012-03-21 16:34:00 -07001273free:
Hillf Dantonb0365c82011-12-28 15:57:16 -08001274 spin_unlock(&hugetlb_lock);
Adam Litke19fc3f02008-04-28 02:12:20 -07001275
1276 /* Free unnecessary surplus pages to the buddy allocator */
Joonsoo Kimc0d934b2013-09-11 14:21:02 -07001277 list_for_each_entry_safe(page, tmp, &surplus_list, lru)
1278 put_page(page);
Naoya Horiguchia9869b82010-09-08 10:19:37 +09001279 spin_lock(&hugetlb_lock);
Adam Litkee4e574b2007-10-16 01:26:19 -07001280
1281 return ret;
1282}
1283
1284/*
1285 * When releasing a hugetlb pool reservation, any surplus pages that were
1286 * allocated to satisfy the reservation must be explicitly freed if they were
1287 * never used.
Lee Schermerhorn685f3452009-09-21 17:01:23 -07001288 * Called with hugetlb_lock held.
Adam Litkee4e574b2007-10-16 01:26:19 -07001289 */
Andi Kleena5516432008-07-23 21:27:41 -07001290static void return_unused_surplus_pages(struct hstate *h,
1291 unsigned long unused_resv_pages)
Adam Litkee4e574b2007-10-16 01:26:19 -07001292{
Adam Litkee4e574b2007-10-16 01:26:19 -07001293 unsigned long nr_pages;
1294
Adam Litkeac09b3a2008-03-04 14:29:38 -08001295 /* Uncommit the reservation */
Andi Kleena5516432008-07-23 21:27:41 -07001296 h->resv_huge_pages -= unused_resv_pages;
Adam Litkeac09b3a2008-03-04 14:29:38 -08001297
Andi Kleenaa888a72008-07-23 21:27:47 -07001298 /* Cannot return gigantic pages currently */
Luiz Capitulinobae7f4a2014-06-04 16:07:08 -07001299 if (hstate_is_gigantic(h))
Andi Kleenaa888a72008-07-23 21:27:47 -07001300 return;
1301
Andi Kleena5516432008-07-23 21:27:41 -07001302 nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
Adam Litkee4e574b2007-10-16 01:26:19 -07001303
Lee Schermerhorn685f3452009-09-21 17:01:23 -07001304 /*
1305 * We want to release as many surplus pages as possible, spread
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -08001306 * evenly across all nodes with memory. Iterate across these nodes
1307 * until we can no longer free unreserved surplus pages. This occurs
1308 * when the nodes with surplus pages have no free pages.
1309 * free_pool_huge_page() will balance the the freed pages across the
1310 * on-line nodes with memory and will handle the hstate accounting.
Lee Schermerhorn685f3452009-09-21 17:01:23 -07001311 */
1312 while (nr_pages--) {
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08001313 if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1))
Lee Schermerhorn685f3452009-09-21 17:01:23 -07001314 break;
Mizuma, Masayoshi7848a4b2014-04-18 15:07:18 -07001315 cond_resched_lock(&hugetlb_lock);
Adam Litkee4e574b2007-10-16 01:26:19 -07001316 }
1317}
1318
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001319/*
1320 * Determine if the huge page at addr within the vma has an associated
1321 * reservation. Where it does not we will need to logically increase
David Gibson90481622012-03-21 16:34:12 -07001322 * reservation and actually increase subpool usage before an allocation
1323 * can occur. Where any new reservation would be required the
1324 * reservation change is prepared, but not committed. Once the page
1325 * has been allocated from the subpool and instantiated the change should
1326 * be committed via vma_commit_reservation. No action is required on
1327 * failure.
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001328 */
Roel Kluine2f17d92009-03-31 15:23:15 -07001329static long vma_needs_reservation(struct hstate *h,
Andi Kleena5516432008-07-23 21:27:41 -07001330 struct vm_area_struct *vma, unsigned long addr)
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001331{
Joonsoo Kim4e35f482014-04-03 14:47:30 -07001332 struct resv_map *resv;
1333 pgoff_t idx;
1334 long chg;
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001335
Joonsoo Kim4e35f482014-04-03 14:47:30 -07001336 resv = vma_resv_map(vma);
1337 if (!resv)
Andy Whitcroft84afd992008-07-23 21:27:32 -07001338 return 1;
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001339
Joonsoo Kim4e35f482014-04-03 14:47:30 -07001340 idx = vma_hugecache_offset(h, vma, addr);
1341 chg = region_chg(resv, idx, idx + 1);
Andy Whitcroft84afd992008-07-23 21:27:32 -07001342
Joonsoo Kim4e35f482014-04-03 14:47:30 -07001343 if (vma->vm_flags & VM_MAYSHARE)
1344 return chg;
1345 else
1346 return chg < 0 ? chg : 0;
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001347}
Andi Kleena5516432008-07-23 21:27:41 -07001348static void vma_commit_reservation(struct hstate *h,
1349 struct vm_area_struct *vma, unsigned long addr)
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001350{
Joonsoo Kim4e35f482014-04-03 14:47:30 -07001351 struct resv_map *resv;
1352 pgoff_t idx;
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001353
Joonsoo Kim4e35f482014-04-03 14:47:30 -07001354 resv = vma_resv_map(vma);
1355 if (!resv)
1356 return;
Joonsoo Kim9119a412014-04-03 14:47:25 -07001357
Joonsoo Kim4e35f482014-04-03 14:47:30 -07001358 idx = vma_hugecache_offset(h, vma, addr);
1359 region_add(resv, idx, idx + 1);
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001360}
1361
David Gibson27a85ef2006-03-22 00:08:56 -08001362static struct page *alloc_huge_page(struct vm_area_struct *vma,
Mel Gorman04f2cbe2008-07-23 21:27:25 -07001363 unsigned long addr, int avoid_reserve)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001364{
David Gibson90481622012-03-21 16:34:12 -07001365 struct hugepage_subpool *spool = subpool_vma(vma);
Andi Kleena5516432008-07-23 21:27:41 -07001366 struct hstate *h = hstate_vma(vma);
Adam Litke348ea202007-11-14 16:59:37 -08001367 struct page *page;
Roel Kluine2f17d92009-03-31 15:23:15 -07001368 long chg;
Aneesh Kumar K.V6d76dcf2012-07-31 16:42:18 -07001369 int ret, idx;
1370 struct hugetlb_cgroup *h_cg;
Adam Litke2fc39ce2007-11-14 16:59:39 -08001371
Aneesh Kumar K.V6d76dcf2012-07-31 16:42:18 -07001372 idx = hstate_index(h);
Mel Gormana1e78772008-07-23 21:27:23 -07001373 /*
David Gibson90481622012-03-21 16:34:12 -07001374 * Processes that did not create the mapping will have no
1375 * reserves and will not have accounted against subpool
1376 * limit. Check that the subpool limit can be made before
1377 * satisfying the allocation MAP_NORESERVE mappings may also
1378 * need pages and subpool limit allocated allocated if no reserve
1379 * mapping overlaps.
Mel Gormana1e78772008-07-23 21:27:23 -07001380 */
Andi Kleena5516432008-07-23 21:27:41 -07001381 chg = vma_needs_reservation(h, vma, addr);
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07001382 if (chg < 0)
Aneesh Kumar K.V76dcee72012-07-31 16:41:57 -07001383 return ERR_PTR(-ENOMEM);
Joonsoo Kim8bb3f122013-09-11 14:21:54 -07001384 if (chg || avoid_reserve)
1385 if (hugepage_subpool_get_pages(spool, 1))
Aneesh Kumar K.V76dcee72012-07-31 16:41:57 -07001386 return ERR_PTR(-ENOSPC);
Mel Gormana1e78772008-07-23 21:27:23 -07001387
Aneesh Kumar K.V6d76dcf2012-07-31 16:42:18 -07001388 ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg);
Jianyu Zhan8f34af62014-06-04 16:10:36 -07001389 if (ret)
1390 goto out_subpool_put;
1391
Mel Gormana1e78772008-07-23 21:27:23 -07001392 spin_lock(&hugetlb_lock);
Joonsoo Kimaf0ed732013-09-11 14:21:18 -07001393 page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, chg);
Joonsoo Kim81a6fca2013-09-11 14:20:58 -07001394 if (!page) {
Aneesh Kumar K.V94ae8ba2012-07-31 16:42:35 -07001395 spin_unlock(&hugetlb_lock);
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001396 page = alloc_buddy_huge_page(h, NUMA_NO_NODE);
Jianyu Zhan8f34af62014-06-04 16:10:36 -07001397 if (!page)
1398 goto out_uncharge_cgroup;
1399
Aneesh Kumar K.V79dbb232012-07-31 16:42:32 -07001400 spin_lock(&hugetlb_lock);
1401 list_move(&page->lru, &h->hugepage_activelist);
Joonsoo Kim81a6fca2013-09-11 14:20:58 -07001402 /* Fall through */
Mel Gormana1e78772008-07-23 21:27:23 -07001403 }
Joonsoo Kim81a6fca2013-09-11 14:20:58 -07001404 hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page);
1405 spin_unlock(&hugetlb_lock);
Mel Gormana1e78772008-07-23 21:27:23 -07001406
David Gibson90481622012-03-21 16:34:12 -07001407 set_page_private(page, (unsigned long)spool);
Mel Gormana1e78772008-07-23 21:27:23 -07001408
Andi Kleena5516432008-07-23 21:27:41 -07001409 vma_commit_reservation(h, vma, addr);
Adam Litke90d8b7e2007-11-14 16:59:42 -08001410 return page;
Jianyu Zhan8f34af62014-06-04 16:10:36 -07001411
1412out_uncharge_cgroup:
1413 hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg);
1414out_subpool_put:
1415 if (chg || avoid_reserve)
1416 hugepage_subpool_put_pages(spool, 1);
1417 return ERR_PTR(-ENOSPC);
David Gibsonb45b5bd2006-03-22 00:08:55 -08001418}
1419
Naoya Horiguchi74060e42013-09-11 14:22:06 -07001420/*
1421 * alloc_huge_page()'s wrapper which simply returns the page if allocation
1422 * succeeds, otherwise NULL. This function is called from new_vma_page(),
1423 * where no ERR_VALUE is expected to be returned.
1424 */
1425struct page *alloc_huge_page_noerr(struct vm_area_struct *vma,
1426 unsigned long addr, int avoid_reserve)
1427{
1428 struct page *page = alloc_huge_page(vma, addr, avoid_reserve);
1429 if (IS_ERR(page))
1430 page = NULL;
1431 return page;
1432}
1433
Cyrill Gorcunov91f47662009-01-06 14:40:33 -08001434int __weak alloc_bootmem_huge_page(struct hstate *h)
Andi Kleenaa888a72008-07-23 21:27:47 -07001435{
1436 struct huge_bootmem_page *m;
Joonsoo Kimb2261022013-09-11 14:21:00 -07001437 int nr_nodes, node;
Andi Kleenaa888a72008-07-23 21:27:47 -07001438
Joonsoo Kimb2261022013-09-11 14:21:00 -07001439 for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) {
Andi Kleenaa888a72008-07-23 21:27:47 -07001440 void *addr;
1441
Grygorii Strashko8b89a112014-01-21 15:50:36 -08001442 addr = memblock_virt_alloc_try_nid_nopanic(
1443 huge_page_size(h), huge_page_size(h),
1444 0, BOOTMEM_ALLOC_ACCESSIBLE, node);
Andi Kleenaa888a72008-07-23 21:27:47 -07001445 if (addr) {
1446 /*
1447 * Use the beginning of the huge page to store the
1448 * huge_bootmem_page struct (until gather_bootmem
1449 * puts them into the mem_map).
1450 */
1451 m = addr;
Cyrill Gorcunov91f47662009-01-06 14:40:33 -08001452 goto found;
Andi Kleenaa888a72008-07-23 21:27:47 -07001453 }
Andi Kleenaa888a72008-07-23 21:27:47 -07001454 }
1455 return 0;
1456
1457found:
1458 BUG_ON((unsigned long)virt_to_phys(m) & (huge_page_size(h) - 1));
1459 /* Put them into a private list first because mem_map is not up yet */
1460 list_add(&m->list, &huge_boot_pages);
1461 m->hstate = h;
1462 return 1;
1463}
1464
David Rientjesf412c972014-04-03 14:47:59 -07001465static void __init prep_compound_huge_page(struct page *page, int order)
Andy Whitcroft18229df2008-11-06 12:53:27 -08001466{
1467 if (unlikely(order > (MAX_ORDER - 1)))
1468 prep_compound_gigantic_page(page, order);
1469 else
1470 prep_compound_page(page, order);
1471}
1472
Andi Kleenaa888a72008-07-23 21:27:47 -07001473/* Put bootmem huge pages into the standard lists after mem_map is up */
1474static void __init gather_bootmem_prealloc(void)
1475{
1476 struct huge_bootmem_page *m;
1477
1478 list_for_each_entry(m, &huge_boot_pages, list) {
Andi Kleenaa888a72008-07-23 21:27:47 -07001479 struct hstate *h = m->hstate;
Becky Bruceee8f2482011-07-25 17:11:50 -07001480 struct page *page;
1481
1482#ifdef CONFIG_HIGHMEM
1483 page = pfn_to_page(m->phys >> PAGE_SHIFT);
Grygorii Strashko8b89a112014-01-21 15:50:36 -08001484 memblock_free_late(__pa(m),
1485 sizeof(struct huge_bootmem_page));
Becky Bruceee8f2482011-07-25 17:11:50 -07001486#else
1487 page = virt_to_page(m);
1488#endif
Andi Kleenaa888a72008-07-23 21:27:47 -07001489 WARN_ON(page_count(page) != 1);
Andy Whitcroft18229df2008-11-06 12:53:27 -08001490 prep_compound_huge_page(page, h->order);
Andrea Arcangelief5a22b2013-10-16 13:46:56 -07001491 WARN_ON(PageReserved(page));
Andi Kleenaa888a72008-07-23 21:27:47 -07001492 prep_new_huge_page(h, page, page_to_nid(page));
Rafael Aquinib0320c72011-06-15 15:08:39 -07001493 /*
1494 * If we had gigantic hugepages allocated at boot time, we need
1495 * to restore the 'stolen' pages to totalram_pages in order to
1496 * fix confusing memory reports from free(1) and another
1497 * side-effects, like CommitLimit going negative.
1498 */
Luiz Capitulinobae7f4a2014-06-04 16:07:08 -07001499 if (hstate_is_gigantic(h))
Jiang Liu3dcc0572013-07-03 15:03:21 -07001500 adjust_managed_page_count(page, 1 << h->order);
Andi Kleenaa888a72008-07-23 21:27:47 -07001501 }
1502}
1503
Andi Kleen8faa8b02008-07-23 21:27:48 -07001504static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001505{
1506 unsigned long i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001507
Andi Kleene5ff2152008-07-23 21:27:42 -07001508 for (i = 0; i < h->max_huge_pages; ++i) {
Luiz Capitulinobae7f4a2014-06-04 16:07:08 -07001509 if (hstate_is_gigantic(h)) {
Andi Kleenaa888a72008-07-23 21:27:47 -07001510 if (!alloc_bootmem_huge_page(h))
1511 break;
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -08001512 } else if (!alloc_fresh_huge_page(h,
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08001513 &node_states[N_MEMORY]))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001514 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001515 }
Andi Kleen8faa8b02008-07-23 21:27:48 -07001516 h->max_huge_pages = i;
Andi Kleene5ff2152008-07-23 21:27:42 -07001517}
1518
1519static void __init hugetlb_init_hstates(void)
1520{
1521 struct hstate *h;
1522
1523 for_each_hstate(h) {
Andi Kleen8faa8b02008-07-23 21:27:48 -07001524 /* oversize hugepages were init'ed in early boot */
Luiz Capitulinobae7f4a2014-06-04 16:07:08 -07001525 if (!hstate_is_gigantic(h))
Andi Kleen8faa8b02008-07-23 21:27:48 -07001526 hugetlb_hstate_alloc_pages(h);
Andi Kleene5ff2152008-07-23 21:27:42 -07001527 }
1528}
1529
Andi Kleen4abd32d2008-07-23 21:27:49 -07001530static char * __init memfmt(char *buf, unsigned long n)
1531{
1532 if (n >= (1UL << 30))
1533 sprintf(buf, "%lu GB", n >> 30);
1534 else if (n >= (1UL << 20))
1535 sprintf(buf, "%lu MB", n >> 20);
1536 else
1537 sprintf(buf, "%lu KB", n >> 10);
1538 return buf;
1539}
1540
Andi Kleene5ff2152008-07-23 21:27:42 -07001541static void __init report_hugepages(void)
1542{
1543 struct hstate *h;
1544
1545 for_each_hstate(h) {
Andi Kleen4abd32d2008-07-23 21:27:49 -07001546 char buf[32];
Andrew Mortonffb22af2013-02-22 16:32:08 -08001547 pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n",
Andi Kleen4abd32d2008-07-23 21:27:49 -07001548 memfmt(buf, huge_page_size(h)),
1549 h->free_huge_pages);
Andi Kleene5ff2152008-07-23 21:27:42 -07001550 }
1551}
1552
Linus Torvalds1da177e2005-04-16 15:20:36 -07001553#ifdef CONFIG_HIGHMEM
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001554static void try_to_free_low(struct hstate *h, unsigned long count,
1555 nodemask_t *nodes_allowed)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001556{
Christoph Lameter4415cc82006-09-25 23:31:55 -07001557 int i;
1558
Luiz Capitulinobae7f4a2014-06-04 16:07:08 -07001559 if (hstate_is_gigantic(h))
Andi Kleenaa888a72008-07-23 21:27:47 -07001560 return;
1561
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001562 for_each_node_mask(i, *nodes_allowed) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001563 struct page *page, *next;
Andi Kleena5516432008-07-23 21:27:41 -07001564 struct list_head *freel = &h->hugepage_freelists[i];
1565 list_for_each_entry_safe(page, next, freel, lru) {
1566 if (count >= h->nr_huge_pages)
Adam Litke6b0c8802007-10-16 01:26:23 -07001567 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001568 if (PageHighMem(page))
1569 continue;
1570 list_del(&page->lru);
Andi Kleene5ff2152008-07-23 21:27:42 -07001571 update_and_free_page(h, page);
Andi Kleena5516432008-07-23 21:27:41 -07001572 h->free_huge_pages--;
1573 h->free_huge_pages_node[page_to_nid(page)]--;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001574 }
1575 }
1576}
1577#else
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001578static inline void try_to_free_low(struct hstate *h, unsigned long count,
1579 nodemask_t *nodes_allowed)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001580{
1581}
1582#endif
1583
Wu Fengguang20a03072009-06-16 15:32:22 -07001584/*
1585 * Increment or decrement surplus_huge_pages. Keep node-specific counters
1586 * balanced by operating on them in a round-robin fashion.
1587 * Returns 1 if an adjustment was made.
1588 */
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001589static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
1590 int delta)
Wu Fengguang20a03072009-06-16 15:32:22 -07001591{
Joonsoo Kimb2261022013-09-11 14:21:00 -07001592 int nr_nodes, node;
Wu Fengguang20a03072009-06-16 15:32:22 -07001593
1594 VM_BUG_ON(delta != -1 && delta != 1);
Wu Fengguang20a03072009-06-16 15:32:22 -07001595
Joonsoo Kimb2261022013-09-11 14:21:00 -07001596 if (delta < 0) {
1597 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
1598 if (h->surplus_huge_pages_node[node])
1599 goto found;
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001600 }
Joonsoo Kimb2261022013-09-11 14:21:00 -07001601 } else {
1602 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
1603 if (h->surplus_huge_pages_node[node] <
1604 h->nr_huge_pages_node[node])
1605 goto found;
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001606 }
Joonsoo Kimb2261022013-09-11 14:21:00 -07001607 }
1608 return 0;
Wu Fengguang20a03072009-06-16 15:32:22 -07001609
Joonsoo Kimb2261022013-09-11 14:21:00 -07001610found:
1611 h->surplus_huge_pages += delta;
1612 h->surplus_huge_pages_node[node] += delta;
1613 return 1;
Wu Fengguang20a03072009-06-16 15:32:22 -07001614}
1615
Andi Kleena5516432008-07-23 21:27:41 -07001616#define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001617static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count,
1618 nodemask_t *nodes_allowed)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001619{
Adam Litke7893d1d2007-10-16 01:26:18 -07001620 unsigned long min_count, ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001621
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001622 if (hstate_is_gigantic(h) && !gigantic_page_supported())
Andi Kleenaa888a72008-07-23 21:27:47 -07001623 return h->max_huge_pages;
1624
Adam Litke7893d1d2007-10-16 01:26:18 -07001625 /*
1626 * Increase the pool size
1627 * First take pages out of surplus state. Then make up the
1628 * remaining difference by allocating fresh huge pages.
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001629 *
1630 * We might race with alloc_buddy_huge_page() here and be unable
1631 * to convert a surplus huge page to a normal huge page. That is
1632 * not critical, though, it just means the overall size of the
1633 * pool might be one hugepage larger than it needs to be, but
1634 * within all the constraints specified by the sysctls.
Adam Litke7893d1d2007-10-16 01:26:18 -07001635 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001636 spin_lock(&hugetlb_lock);
Andi Kleena5516432008-07-23 21:27:41 -07001637 while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001638 if (!adjust_pool_surplus(h, nodes_allowed, -1))
Adam Litke7893d1d2007-10-16 01:26:18 -07001639 break;
1640 }
1641
Andi Kleena5516432008-07-23 21:27:41 -07001642 while (count > persistent_huge_pages(h)) {
Adam Litke7893d1d2007-10-16 01:26:18 -07001643 /*
1644 * If this allocation races such that we no longer need the
1645 * page, free_huge_page will handle it by freeing the page
1646 * and reducing the surplus.
1647 */
1648 spin_unlock(&hugetlb_lock);
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001649 if (hstate_is_gigantic(h))
1650 ret = alloc_fresh_gigantic_page(h, nodes_allowed);
1651 else
1652 ret = alloc_fresh_huge_page(h, nodes_allowed);
Adam Litke7893d1d2007-10-16 01:26:18 -07001653 spin_lock(&hugetlb_lock);
1654 if (!ret)
1655 goto out;
1656
Mel Gorman536240f22009-12-14 17:59:56 -08001657 /* Bail for signals. Probably ctrl-c from user */
1658 if (signal_pending(current))
1659 goto out;
Adam Litke7893d1d2007-10-16 01:26:18 -07001660 }
Adam Litke7893d1d2007-10-16 01:26:18 -07001661
1662 /*
1663 * Decrease the pool size
1664 * First return free pages to the buddy allocator (being careful
1665 * to keep enough around to satisfy reservations). Then place
1666 * pages into surplus state as needed so the pool will shrink
1667 * to the desired size as pages become free.
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001668 *
1669 * By placing pages into the surplus state independent of the
1670 * overcommit value, we are allowing the surplus pool size to
1671 * exceed overcommit. There are few sane options here. Since
1672 * alloc_buddy_huge_page() is checking the global counter,
1673 * though, we'll note that we're not allowed to exceed surplus
1674 * and won't grow the pool anywhere else. Not until one of the
1675 * sysctls are changed, or the surplus pages go out of use.
Adam Litke7893d1d2007-10-16 01:26:18 -07001676 */
Andi Kleena5516432008-07-23 21:27:41 -07001677 min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
Adam Litke6b0c8802007-10-16 01:26:23 -07001678 min_count = max(count, min_count);
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001679 try_to_free_low(h, min_count, nodes_allowed);
Andi Kleena5516432008-07-23 21:27:41 -07001680 while (min_count < persistent_huge_pages(h)) {
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001681 if (!free_pool_huge_page(h, nodes_allowed, 0))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001682 break;
Mizuma, Masayoshi55f67142014-04-07 15:37:54 -07001683 cond_resched_lock(&hugetlb_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001684 }
Andi Kleena5516432008-07-23 21:27:41 -07001685 while (count < persistent_huge_pages(h)) {
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001686 if (!adjust_pool_surplus(h, nodes_allowed, 1))
Adam Litke7893d1d2007-10-16 01:26:18 -07001687 break;
1688 }
1689out:
Andi Kleena5516432008-07-23 21:27:41 -07001690 ret = persistent_huge_pages(h);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001691 spin_unlock(&hugetlb_lock);
Adam Litke7893d1d2007-10-16 01:26:18 -07001692 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001693}
1694
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001695#define HSTATE_ATTR_RO(_name) \
1696 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
1697
1698#define HSTATE_ATTR(_name) \
1699 static struct kobj_attribute _name##_attr = \
1700 __ATTR(_name, 0644, _name##_show, _name##_store)
1701
1702static struct kobject *hugepages_kobj;
1703static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
1704
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001705static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
1706
1707static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001708{
1709 int i;
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001710
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001711 for (i = 0; i < HUGE_MAX_HSTATE; i++)
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001712 if (hstate_kobjs[i] == kobj) {
1713 if (nidp)
1714 *nidp = NUMA_NO_NODE;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001715 return &hstates[i];
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001716 }
1717
1718 return kobj_to_node_hstate(kobj, nidp);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001719}
1720
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001721static ssize_t nr_hugepages_show_common(struct kobject *kobj,
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001722 struct kobj_attribute *attr, char *buf)
1723{
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001724 struct hstate *h;
1725 unsigned long nr_huge_pages;
1726 int nid;
1727
1728 h = kobj_to_hstate(kobj, &nid);
1729 if (nid == NUMA_NO_NODE)
1730 nr_huge_pages = h->nr_huge_pages;
1731 else
1732 nr_huge_pages = h->nr_huge_pages_node[nid];
1733
1734 return sprintf(buf, "%lu\n", nr_huge_pages);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001735}
Eric B Munsonadbe8722011-01-13 15:47:27 -08001736
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001737static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
1738 struct kobject *kobj, struct kobj_attribute *attr,
1739 const char *buf, size_t len)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001740{
1741 int err;
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001742 int nid;
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001743 unsigned long count;
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001744 struct hstate *h;
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
Jingoo Han3dbb95f2013-09-11 14:20:25 -07001747 err = kstrtoul(buf, 10, &count);
Eric B Munson73ae31e2011-01-13 15:47:28 -08001748 if (err)
Eric B Munsonadbe8722011-01-13 15:47:27 -08001749 goto out;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001750
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001751 h = kobj_to_hstate(kobj, &nid);
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001752 if (hstate_is_gigantic(h) && !gigantic_page_supported()) {
Eric B Munsonadbe8722011-01-13 15:47:27 -08001753 err = -EINVAL;
1754 goto out;
1755 }
1756
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001757 if (nid == NUMA_NO_NODE) {
1758 /*
1759 * global hstate attribute
1760 */
1761 if (!(obey_mempolicy &&
1762 init_nodemask_of_mempolicy(nodes_allowed))) {
1763 NODEMASK_FREE(nodes_allowed);
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08001764 nodes_allowed = &node_states[N_MEMORY];
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001765 }
1766 } else if (nodes_allowed) {
1767 /*
1768 * per node hstate attribute: adjust count to global,
1769 * but restrict alloc/free to the specified node.
1770 */
1771 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
1772 init_nodemask_of_node(nodes_allowed, nid);
1773 } else
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08001774 nodes_allowed = &node_states[N_MEMORY];
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001775
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001776 h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001777
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08001778 if (nodes_allowed != &node_states[N_MEMORY])
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001779 NODEMASK_FREE(nodes_allowed);
1780
1781 return len;
Eric B Munsonadbe8722011-01-13 15:47:27 -08001782out:
1783 NODEMASK_FREE(nodes_allowed);
1784 return err;
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001785}
1786
1787static ssize_t nr_hugepages_show(struct kobject *kobj,
1788 struct kobj_attribute *attr, char *buf)
1789{
1790 return nr_hugepages_show_common(kobj, attr, buf);
1791}
1792
1793static ssize_t nr_hugepages_store(struct kobject *kobj,
1794 struct kobj_attribute *attr, const char *buf, size_t len)
1795{
1796 return nr_hugepages_store_common(false, kobj, attr, buf, len);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001797}
1798HSTATE_ATTR(nr_hugepages);
1799
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001800#ifdef CONFIG_NUMA
1801
1802/*
1803 * hstate attribute for optionally mempolicy-based constraint on persistent
1804 * huge page alloc/free.
1805 */
1806static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
1807 struct kobj_attribute *attr, char *buf)
1808{
1809 return nr_hugepages_show_common(kobj, attr, buf);
1810}
1811
1812static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
1813 struct kobj_attribute *attr, const char *buf, size_t len)
1814{
1815 return nr_hugepages_store_common(true, kobj, attr, buf, len);
1816}
1817HSTATE_ATTR(nr_hugepages_mempolicy);
1818#endif
1819
1820
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001821static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
1822 struct kobj_attribute *attr, char *buf)
1823{
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001824 struct hstate *h = kobj_to_hstate(kobj, NULL);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001825 return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
1826}
Eric B Munsonadbe8722011-01-13 15:47:27 -08001827
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001828static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
1829 struct kobj_attribute *attr, const char *buf, size_t count)
1830{
1831 int err;
1832 unsigned long input;
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001833 struct hstate *h = kobj_to_hstate(kobj, NULL);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001834
Luiz Capitulinobae7f4a2014-06-04 16:07:08 -07001835 if (hstate_is_gigantic(h))
Eric B Munsonadbe8722011-01-13 15:47:27 -08001836 return -EINVAL;
1837
Jingoo Han3dbb95f2013-09-11 14:20:25 -07001838 err = kstrtoul(buf, 10, &input);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001839 if (err)
Eric B Munson73ae31e2011-01-13 15:47:28 -08001840 return err;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001841
1842 spin_lock(&hugetlb_lock);
1843 h->nr_overcommit_huge_pages = input;
1844 spin_unlock(&hugetlb_lock);
1845
1846 return count;
1847}
1848HSTATE_ATTR(nr_overcommit_hugepages);
1849
1850static ssize_t free_hugepages_show(struct kobject *kobj,
1851 struct kobj_attribute *attr, char *buf)
1852{
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001853 struct hstate *h;
1854 unsigned long free_huge_pages;
1855 int nid;
1856
1857 h = kobj_to_hstate(kobj, &nid);
1858 if (nid == NUMA_NO_NODE)
1859 free_huge_pages = h->free_huge_pages;
1860 else
1861 free_huge_pages = h->free_huge_pages_node[nid];
1862
1863 return sprintf(buf, "%lu\n", free_huge_pages);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001864}
1865HSTATE_ATTR_RO(free_hugepages);
1866
1867static ssize_t resv_hugepages_show(struct kobject *kobj,
1868 struct kobj_attribute *attr, char *buf)
1869{
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001870 struct hstate *h = kobj_to_hstate(kobj, NULL);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001871 return sprintf(buf, "%lu\n", h->resv_huge_pages);
1872}
1873HSTATE_ATTR_RO(resv_hugepages);
1874
1875static ssize_t surplus_hugepages_show(struct kobject *kobj,
1876 struct kobj_attribute *attr, char *buf)
1877{
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001878 struct hstate *h;
1879 unsigned long surplus_huge_pages;
1880 int nid;
1881
1882 h = kobj_to_hstate(kobj, &nid);
1883 if (nid == NUMA_NO_NODE)
1884 surplus_huge_pages = h->surplus_huge_pages;
1885 else
1886 surplus_huge_pages = h->surplus_huge_pages_node[nid];
1887
1888 return sprintf(buf, "%lu\n", surplus_huge_pages);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001889}
1890HSTATE_ATTR_RO(surplus_hugepages);
1891
1892static struct attribute *hstate_attrs[] = {
1893 &nr_hugepages_attr.attr,
1894 &nr_overcommit_hugepages_attr.attr,
1895 &free_hugepages_attr.attr,
1896 &resv_hugepages_attr.attr,
1897 &surplus_hugepages_attr.attr,
Lee Schermerhorn06808b02009-12-14 17:58:21 -08001898#ifdef CONFIG_NUMA
1899 &nr_hugepages_mempolicy_attr.attr,
1900#endif
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001901 NULL,
1902};
1903
1904static struct attribute_group hstate_attr_group = {
1905 .attrs = hstate_attrs,
1906};
1907
Jeff Mahoney094e9532010-02-02 13:44:14 -08001908static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
1909 struct kobject **hstate_kobjs,
1910 struct attribute_group *hstate_attr_group)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001911{
1912 int retval;
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07001913 int hi = hstate_index(h);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001914
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001915 hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
1916 if (!hstate_kobjs[hi])
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001917 return -ENOMEM;
1918
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001919 retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001920 if (retval)
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001921 kobject_put(hstate_kobjs[hi]);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001922
1923 return retval;
1924}
1925
1926static void __init hugetlb_sysfs_init(void)
1927{
1928 struct hstate *h;
1929 int err;
1930
1931 hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
1932 if (!hugepages_kobj)
1933 return;
1934
1935 for_each_hstate(h) {
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001936 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
1937 hstate_kobjs, &hstate_attr_group);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001938 if (err)
Andrew Mortonffb22af2013-02-22 16:32:08 -08001939 pr_err("Hugetlb: Unable to add hstate %s", h->name);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07001940 }
1941}
1942
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001943#ifdef CONFIG_NUMA
1944
1945/*
1946 * node_hstate/s - associate per node hstate attributes, via their kobjects,
Kay Sievers10fbcf42011-12-21 14:48:43 -08001947 * with node devices in node_devices[] using a parallel array. The array
1948 * index of a node device or _hstate == node id.
1949 * This is here to avoid any static dependency of the node device driver, in
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001950 * the base kernel, on the hugetlb module.
1951 */
1952struct node_hstate {
1953 struct kobject *hugepages_kobj;
1954 struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
1955};
1956struct node_hstate node_hstates[MAX_NUMNODES];
1957
1958/*
Kay Sievers10fbcf42011-12-21 14:48:43 -08001959 * A subset of global hstate attributes for node devices
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001960 */
1961static struct attribute *per_node_hstate_attrs[] = {
1962 &nr_hugepages_attr.attr,
1963 &free_hugepages_attr.attr,
1964 &surplus_hugepages_attr.attr,
1965 NULL,
1966};
1967
1968static struct attribute_group per_node_hstate_attr_group = {
1969 .attrs = per_node_hstate_attrs,
1970};
1971
1972/*
Kay Sievers10fbcf42011-12-21 14:48:43 -08001973 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001974 * Returns node id via non-NULL nidp.
1975 */
1976static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
1977{
1978 int nid;
1979
1980 for (nid = 0; nid < nr_node_ids; nid++) {
1981 struct node_hstate *nhs = &node_hstates[nid];
1982 int i;
1983 for (i = 0; i < HUGE_MAX_HSTATE; i++)
1984 if (nhs->hstate_kobjs[i] == kobj) {
1985 if (nidp)
1986 *nidp = nid;
1987 return &hstates[i];
1988 }
1989 }
1990
1991 BUG();
1992 return NULL;
1993}
1994
1995/*
Kay Sievers10fbcf42011-12-21 14:48:43 -08001996 * Unregister hstate attributes from a single node device.
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08001997 * No-op if no hstate attributes attached.
1998 */
Claudiu Ghioc3cd8b442013-03-04 12:46:15 +02001999static void hugetlb_unregister_node(struct node *node)
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002000{
2001 struct hstate *h;
Kay Sievers10fbcf42011-12-21 14:48:43 -08002002 struct node_hstate *nhs = &node_hstates[node->dev.id];
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002003
2004 if (!nhs->hugepages_kobj)
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -08002005 return; /* no hstate attributes */
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002006
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07002007 for_each_hstate(h) {
2008 int idx = hstate_index(h);
2009 if (nhs->hstate_kobjs[idx]) {
2010 kobject_put(nhs->hstate_kobjs[idx]);
2011 nhs->hstate_kobjs[idx] = NULL;
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002012 }
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07002013 }
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002014
2015 kobject_put(nhs->hugepages_kobj);
2016 nhs->hugepages_kobj = NULL;
2017}
2018
2019/*
Kay Sievers10fbcf42011-12-21 14:48:43 -08002020 * hugetlb module exit: unregister hstate attributes from node devices
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002021 * that have them.
2022 */
2023static void hugetlb_unregister_all_nodes(void)
2024{
2025 int nid;
2026
2027 /*
Kay Sievers10fbcf42011-12-21 14:48:43 -08002028 * disable node device registrations.
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002029 */
2030 register_hugetlbfs_with_node(NULL, NULL);
2031
2032 /*
2033 * remove hstate attributes from any nodes that have them.
2034 */
2035 for (nid = 0; nid < nr_node_ids; nid++)
Wen Congyang87327942012-12-11 16:00:56 -08002036 hugetlb_unregister_node(node_devices[nid]);
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002037}
2038
2039/*
Kay Sievers10fbcf42011-12-21 14:48:43 -08002040 * Register hstate attributes for a single node device.
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002041 * No-op if attributes already registered.
2042 */
Claudiu Ghioc3cd8b442013-03-04 12:46:15 +02002043static void hugetlb_register_node(struct node *node)
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002044{
2045 struct hstate *h;
Kay Sievers10fbcf42011-12-21 14:48:43 -08002046 struct node_hstate *nhs = &node_hstates[node->dev.id];
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002047 int err;
2048
2049 if (nhs->hugepages_kobj)
2050 return; /* already allocated */
2051
2052 nhs->hugepages_kobj = kobject_create_and_add("hugepages",
Kay Sievers10fbcf42011-12-21 14:48:43 -08002053 &node->dev.kobj);
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002054 if (!nhs->hugepages_kobj)
2055 return;
2056
2057 for_each_hstate(h) {
2058 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
2059 nhs->hstate_kobjs,
2060 &per_node_hstate_attr_group);
2061 if (err) {
Andrew Mortonffb22af2013-02-22 16:32:08 -08002062 pr_err("Hugetlb: Unable to add hstate %s for node %d\n",
2063 h->name, node->dev.id);
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002064 hugetlb_unregister_node(node);
2065 break;
2066 }
2067 }
2068}
2069
2070/*
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -08002071 * hugetlb init time: register hstate attributes for all registered node
Kay Sievers10fbcf42011-12-21 14:48:43 -08002072 * devices of nodes that have memory. All on-line nodes should have
2073 * registered their associated device by this time.
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002074 */
2075static void hugetlb_register_all_nodes(void)
2076{
2077 int nid;
2078
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08002079 for_each_node_state(nid, N_MEMORY) {
Wen Congyang87327942012-12-11 16:00:56 -08002080 struct node *node = node_devices[nid];
Kay Sievers10fbcf42011-12-21 14:48:43 -08002081 if (node->dev.id == nid)
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002082 hugetlb_register_node(node);
2083 }
2084
2085 /*
Kay Sievers10fbcf42011-12-21 14:48:43 -08002086 * Let the node device driver know we're here so it can
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002087 * [un]register hstate attributes on node hotplug.
2088 */
2089 register_hugetlbfs_with_node(hugetlb_register_node,
2090 hugetlb_unregister_node);
2091}
2092#else /* !CONFIG_NUMA */
2093
2094static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
2095{
2096 BUG();
2097 if (nidp)
2098 *nidp = -1;
2099 return NULL;
2100}
2101
2102static void hugetlb_unregister_all_nodes(void) { }
2103
2104static void hugetlb_register_all_nodes(void) { }
2105
2106#endif
2107
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002108static void __exit hugetlb_exit(void)
2109{
2110 struct hstate *h;
2111
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002112 hugetlb_unregister_all_nodes();
2113
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002114 for_each_hstate(h) {
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07002115 kobject_put(hstate_kobjs[hstate_index(h)]);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002116 }
2117
2118 kobject_put(hugepages_kobj);
Davidlohr Bueso8382d912014-04-03 14:47:31 -07002119 kfree(htlb_fault_mutex_table);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002120}
2121module_exit(hugetlb_exit);
2122
2123static int __init hugetlb_init(void)
2124{
Davidlohr Bueso8382d912014-04-03 14:47:31 -07002125 int i;
2126
Nishanth Aravamudan457c1b22014-05-06 12:50:00 -07002127 if (!hugepages_supported())
Benjamin Herrenschmidt0ef89d22008-07-31 00:07:30 -07002128 return 0;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002129
Nick Piggine11bfbf2008-07-23 21:27:52 -07002130 if (!size_to_hstate(default_hstate_size)) {
2131 default_hstate_size = HPAGE_SIZE;
2132 if (!size_to_hstate(default_hstate_size))
2133 hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002134 }
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07002135 default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size));
Nick Piggine11bfbf2008-07-23 21:27:52 -07002136 if (default_hstate_max_huge_pages)
2137 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002138
2139 hugetlb_init_hstates();
Andi Kleenaa888a72008-07-23 21:27:47 -07002140 gather_bootmem_prealloc();
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002141 report_hugepages();
2142
2143 hugetlb_sysfs_init();
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002144 hugetlb_register_all_nodes();
Jianguo Wu7179e7b2012-12-18 14:23:19 -08002145 hugetlb_cgroup_file_init();
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002146
Davidlohr Bueso8382d912014-04-03 14:47:31 -07002147#ifdef CONFIG_SMP
2148 num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus());
2149#else
2150 num_fault_mutexes = 1;
2151#endif
2152 htlb_fault_mutex_table =
2153 kmalloc(sizeof(struct mutex) * num_fault_mutexes, GFP_KERNEL);
2154 BUG_ON(!htlb_fault_mutex_table);
2155
2156 for (i = 0; i < num_fault_mutexes; i++)
2157 mutex_init(&htlb_fault_mutex_table[i]);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002158 return 0;
2159}
2160module_init(hugetlb_init);
2161
2162/* Should be called on processing a hugepagesz=... option */
2163void __init hugetlb_add_hstate(unsigned order)
2164{
2165 struct hstate *h;
Andi Kleen8faa8b02008-07-23 21:27:48 -07002166 unsigned long i;
2167
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002168 if (size_to_hstate(PAGE_SIZE << order)) {
Andrew Mortonffb22af2013-02-22 16:32:08 -08002169 pr_warning("hugepagesz= specified twice, ignoring\n");
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002170 return;
2171 }
Aneesh Kumar K.V47d38342012-07-31 16:41:54 -07002172 BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002173 BUG_ON(order == 0);
Aneesh Kumar K.V47d38342012-07-31 16:41:54 -07002174 h = &hstates[hugetlb_max_hstate++];
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002175 h->order = order;
2176 h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
Andi Kleen8faa8b02008-07-23 21:27:48 -07002177 h->nr_huge_pages = 0;
2178 h->free_huge_pages = 0;
2179 for (i = 0; i < MAX_NUMNODES; ++i)
2180 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
Aneesh Kumar K.V0edaecf2012-07-31 16:42:07 -07002181 INIT_LIST_HEAD(&h->hugepage_activelist);
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08002182 h->next_nid_to_alloc = first_node(node_states[N_MEMORY]);
2183 h->next_nid_to_free = first_node(node_states[N_MEMORY]);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002184 snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
2185 huge_page_size(h)/1024);
Andi Kleen8faa8b02008-07-23 21:27:48 -07002186
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002187 parsed_hstate = h;
2188}
2189
Nick Piggine11bfbf2008-07-23 21:27:52 -07002190static int __init hugetlb_nrpages_setup(char *s)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002191{
2192 unsigned long *mhp;
Andi Kleen8faa8b02008-07-23 21:27:48 -07002193 static unsigned long *last_mhp;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002194
2195 /*
Aneesh Kumar K.V47d38342012-07-31 16:41:54 -07002196 * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet,
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002197 * so this hugepages= parameter goes to the "default hstate".
2198 */
Aneesh Kumar K.V47d38342012-07-31 16:41:54 -07002199 if (!hugetlb_max_hstate)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002200 mhp = &default_hstate_max_huge_pages;
2201 else
2202 mhp = &parsed_hstate->max_huge_pages;
2203
Andi Kleen8faa8b02008-07-23 21:27:48 -07002204 if (mhp == last_mhp) {
Andrew Mortonffb22af2013-02-22 16:32:08 -08002205 pr_warning("hugepages= specified twice without "
2206 "interleaving hugepagesz=, ignoring\n");
Andi Kleen8faa8b02008-07-23 21:27:48 -07002207 return 1;
2208 }
2209
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002210 if (sscanf(s, "%lu", mhp) <= 0)
2211 *mhp = 0;
2212
Andi Kleen8faa8b02008-07-23 21:27:48 -07002213 /*
2214 * Global state is always initialized later in hugetlb_init.
2215 * But we need to allocate >= MAX_ORDER hstates here early to still
2216 * use the bootmem allocator.
2217 */
Aneesh Kumar K.V47d38342012-07-31 16:41:54 -07002218 if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER)
Andi Kleen8faa8b02008-07-23 21:27:48 -07002219 hugetlb_hstate_alloc_pages(parsed_hstate);
2220
2221 last_mhp = mhp;
2222
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002223 return 1;
2224}
Nick Piggine11bfbf2008-07-23 21:27:52 -07002225__setup("hugepages=", hugetlb_nrpages_setup);
2226
2227static int __init hugetlb_default_setup(char *s)
2228{
2229 default_hstate_size = memparse(s, &s);
2230 return 1;
2231}
2232__setup("default_hugepagesz=", hugetlb_default_setup);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002233
Nishanth Aravamudan8a213462008-07-25 19:44:37 -07002234static unsigned int cpuset_mems_nr(unsigned int *array)
2235{
2236 int node;
2237 unsigned int nr = 0;
2238
2239 for_each_node_mask(node, cpuset_current_mems_allowed)
2240 nr += array[node];
2241
2242 return nr;
2243}
2244
2245#ifdef CONFIG_SYSCTL
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002246static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
2247 struct ctl_table *table, int write,
2248 void __user *buffer, size_t *length, loff_t *ppos)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002249{
Andi Kleene5ff2152008-07-23 21:27:42 -07002250 struct hstate *h = &default_hstate;
2251 unsigned long tmp;
Michal Hocko08d4a242011-01-13 15:47:26 -08002252 int ret;
Andi Kleene5ff2152008-07-23 21:27:42 -07002253
Nishanth Aravamudan457c1b22014-05-06 12:50:00 -07002254 if (!hugepages_supported())
2255 return -ENOTSUPP;
2256
Petr Holasekc033a932011-03-22 16:33:05 -07002257 tmp = h->max_huge_pages;
Andi Kleene5ff2152008-07-23 21:27:42 -07002258
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07002259 if (write && hstate_is_gigantic(h) && !gigantic_page_supported())
Eric B Munsonadbe8722011-01-13 15:47:27 -08002260 return -EINVAL;
2261
Andi Kleene5ff2152008-07-23 21:27:42 -07002262 table->data = &tmp;
2263 table->maxlen = sizeof(unsigned long);
Michal Hocko08d4a242011-01-13 15:47:26 -08002264 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2265 if (ret)
2266 goto out;
Andi Kleene5ff2152008-07-23 21:27:42 -07002267
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002268 if (write) {
David Rientjesbad44b52009-12-14 17:58:38 -08002269 NODEMASK_ALLOC(nodemask_t, nodes_allowed,
2270 GFP_KERNEL | __GFP_NORETRY);
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002271 if (!(obey_mempolicy &&
2272 init_nodemask_of_mempolicy(nodes_allowed))) {
2273 NODEMASK_FREE(nodes_allowed);
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08002274 nodes_allowed = &node_states[N_MEMORY];
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002275 }
2276 h->max_huge_pages = set_max_huge_pages(h, tmp, nodes_allowed);
2277
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08002278 if (nodes_allowed != &node_states[N_MEMORY])
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002279 NODEMASK_FREE(nodes_allowed);
2280 }
Michal Hocko08d4a242011-01-13 15:47:26 -08002281out:
2282 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002283}
Mel Gorman396faf02007-07-17 04:03:13 -07002284
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002285int hugetlb_sysctl_handler(struct ctl_table *table, int write,
2286 void __user *buffer, size_t *length, loff_t *ppos)
2287{
2288
2289 return hugetlb_sysctl_handler_common(false, table, write,
2290 buffer, length, ppos);
2291}
2292
2293#ifdef CONFIG_NUMA
2294int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
2295 void __user *buffer, size_t *length, loff_t *ppos)
2296{
2297 return hugetlb_sysctl_handler_common(true, table, write,
2298 buffer, length, ppos);
2299}
2300#endif /* CONFIG_NUMA */
2301
Nishanth Aravamudana3d0c6a2008-02-08 04:18:18 -08002302int hugetlb_overcommit_handler(struct ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07002303 void __user *buffer,
Nishanth Aravamudana3d0c6a2008-02-08 04:18:18 -08002304 size_t *length, loff_t *ppos)
2305{
Andi Kleena5516432008-07-23 21:27:41 -07002306 struct hstate *h = &default_hstate;
Andi Kleene5ff2152008-07-23 21:27:42 -07002307 unsigned long tmp;
Michal Hocko08d4a242011-01-13 15:47:26 -08002308 int ret;
Andi Kleene5ff2152008-07-23 21:27:42 -07002309
Nishanth Aravamudan457c1b22014-05-06 12:50:00 -07002310 if (!hugepages_supported())
2311 return -ENOTSUPP;
2312
Petr Holasekc033a932011-03-22 16:33:05 -07002313 tmp = h->nr_overcommit_huge_pages;
Andi Kleene5ff2152008-07-23 21:27:42 -07002314
Luiz Capitulinobae7f4a2014-06-04 16:07:08 -07002315 if (write && hstate_is_gigantic(h))
Eric B Munsonadbe8722011-01-13 15:47:27 -08002316 return -EINVAL;
2317
Andi Kleene5ff2152008-07-23 21:27:42 -07002318 table->data = &tmp;
2319 table->maxlen = sizeof(unsigned long);
Michal Hocko08d4a242011-01-13 15:47:26 -08002320 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2321 if (ret)
2322 goto out;
Andi Kleene5ff2152008-07-23 21:27:42 -07002323
2324 if (write) {
2325 spin_lock(&hugetlb_lock);
2326 h->nr_overcommit_huge_pages = tmp;
2327 spin_unlock(&hugetlb_lock);
2328 }
Michal Hocko08d4a242011-01-13 15:47:26 -08002329out:
2330 return ret;
Nishanth Aravamudana3d0c6a2008-02-08 04:18:18 -08002331}
2332
Linus Torvalds1da177e2005-04-16 15:20:36 -07002333#endif /* CONFIG_SYSCTL */
2334
Alexey Dobriyane1759c22008-10-15 23:50:22 +04002335void hugetlb_report_meminfo(struct seq_file *m)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002336{
Andi Kleena5516432008-07-23 21:27:41 -07002337 struct hstate *h = &default_hstate;
Nishanth Aravamudan457c1b22014-05-06 12:50:00 -07002338 if (!hugepages_supported())
2339 return;
Alexey Dobriyane1759c22008-10-15 23:50:22 +04002340 seq_printf(m,
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002341 "HugePages_Total: %5lu\n"
2342 "HugePages_Free: %5lu\n"
2343 "HugePages_Rsvd: %5lu\n"
2344 "HugePages_Surp: %5lu\n"
2345 "Hugepagesize: %8lu kB\n",
Andi Kleena5516432008-07-23 21:27:41 -07002346 h->nr_huge_pages,
2347 h->free_huge_pages,
2348 h->resv_huge_pages,
2349 h->surplus_huge_pages,
2350 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002351}
2352
2353int hugetlb_report_node_meminfo(int nid, char *buf)
2354{
Andi Kleena5516432008-07-23 21:27:41 -07002355 struct hstate *h = &default_hstate;
Nishanth Aravamudan457c1b22014-05-06 12:50:00 -07002356 if (!hugepages_supported())
2357 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002358 return sprintf(buf,
2359 "Node %d HugePages_Total: %5u\n"
Nishanth Aravamudana1de0912008-03-26 14:37:53 -07002360 "Node %d HugePages_Free: %5u\n"
2361 "Node %d HugePages_Surp: %5u\n",
Andi Kleena5516432008-07-23 21:27:41 -07002362 nid, h->nr_huge_pages_node[nid],
2363 nid, h->free_huge_pages_node[nid],
2364 nid, h->surplus_huge_pages_node[nid]);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002365}
2366
David Rientjes949f7ec2013-04-29 15:07:48 -07002367void hugetlb_show_meminfo(void)
2368{
2369 struct hstate *h;
2370 int nid;
2371
Nishanth Aravamudan457c1b22014-05-06 12:50:00 -07002372 if (!hugepages_supported())
2373 return;
2374
David Rientjes949f7ec2013-04-29 15:07:48 -07002375 for_each_node_state(nid, N_MEMORY)
2376 for_each_hstate(h)
2377 pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
2378 nid,
2379 h->nr_huge_pages_node[nid],
2380 h->free_huge_pages_node[nid],
2381 h->surplus_huge_pages_node[nid],
2382 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
2383}
2384
Linus Torvalds1da177e2005-04-16 15:20:36 -07002385/* Return the number pages of memory we physically have, in PAGE_SIZE units. */
2386unsigned long hugetlb_total_pages(void)
2387{
Wanpeng Lid0028582013-03-22 15:04:40 -07002388 struct hstate *h;
2389 unsigned long nr_total_pages = 0;
2390
2391 for_each_hstate(h)
2392 nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h);
2393 return nr_total_pages;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002394}
Linus Torvalds1da177e2005-04-16 15:20:36 -07002395
Andi Kleena5516432008-07-23 21:27:41 -07002396static int hugetlb_acct_memory(struct hstate *h, long delta)
Mel Gormanfc1b8a72008-07-23 21:27:22 -07002397{
2398 int ret = -ENOMEM;
2399
2400 spin_lock(&hugetlb_lock);
2401 /*
2402 * When cpuset is configured, it breaks the strict hugetlb page
2403 * reservation as the accounting is done on a global variable. Such
2404 * reservation is completely rubbish in the presence of cpuset because
2405 * the reservation is not checked against page availability for the
2406 * current cpuset. Application can still potentially OOM'ed by kernel
2407 * with lack of free htlb page in cpuset that the task is in.
2408 * Attempt to enforce strict accounting with cpuset is almost
2409 * impossible (or too ugly) because cpuset is too fluid that
2410 * task or memory node can be dynamically moved between cpusets.
2411 *
2412 * The change of semantics for shared hugetlb mapping with cpuset is
2413 * undesirable. However, in order to preserve some of the semantics,
2414 * we fall back to check against current free page availability as
2415 * a best attempt and hopefully to minimize the impact of changing
2416 * semantics that cpuset has.
2417 */
2418 if (delta > 0) {
Andi Kleena5516432008-07-23 21:27:41 -07002419 if (gather_surplus_pages(h, delta) < 0)
Mel Gormanfc1b8a72008-07-23 21:27:22 -07002420 goto out;
2421
Andi Kleena5516432008-07-23 21:27:41 -07002422 if (delta > cpuset_mems_nr(h->free_huge_pages_node)) {
2423 return_unused_surplus_pages(h, delta);
Mel Gormanfc1b8a72008-07-23 21:27:22 -07002424 goto out;
2425 }
2426 }
2427
2428 ret = 0;
2429 if (delta < 0)
Andi Kleena5516432008-07-23 21:27:41 -07002430 return_unused_surplus_pages(h, (unsigned long) -delta);
Mel Gormanfc1b8a72008-07-23 21:27:22 -07002431
2432out:
2433 spin_unlock(&hugetlb_lock);
2434 return ret;
2435}
2436
Andy Whitcroft84afd992008-07-23 21:27:32 -07002437static void hugetlb_vm_op_open(struct vm_area_struct *vma)
2438{
Joonsoo Kimf522c3a2013-09-11 14:21:53 -07002439 struct resv_map *resv = vma_resv_map(vma);
Andy Whitcroft84afd992008-07-23 21:27:32 -07002440
2441 /*
2442 * This new VMA should share its siblings reservation map if present.
2443 * The VMA will only ever have a valid reservation map pointer where
2444 * it is being copied for another still existing VMA. As that VMA
Lucas De Marchi25985ed2011-03-30 22:57:33 -03002445 * has a reference to the reservation map it cannot disappear until
Andy Whitcroft84afd992008-07-23 21:27:32 -07002446 * after this open call completes. It is therefore safe to take a
2447 * new reference here without additional locking.
2448 */
Joonsoo Kim4e35f482014-04-03 14:47:30 -07002449 if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
Joonsoo Kimf522c3a2013-09-11 14:21:53 -07002450 kref_get(&resv->refs);
Andy Whitcroft84afd992008-07-23 21:27:32 -07002451}
2452
Mel Gormana1e78772008-07-23 21:27:23 -07002453static void hugetlb_vm_op_close(struct vm_area_struct *vma)
2454{
Andi Kleena5516432008-07-23 21:27:41 -07002455 struct hstate *h = hstate_vma(vma);
Joonsoo Kimf522c3a2013-09-11 14:21:53 -07002456 struct resv_map *resv = vma_resv_map(vma);
David Gibson90481622012-03-21 16:34:12 -07002457 struct hugepage_subpool *spool = subpool_vma(vma);
Joonsoo Kim4e35f482014-04-03 14:47:30 -07002458 unsigned long reserve, start, end;
Andy Whitcroft84afd992008-07-23 21:27:32 -07002459
Joonsoo Kim4e35f482014-04-03 14:47:30 -07002460 if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER))
2461 return;
Andy Whitcroft84afd992008-07-23 21:27:32 -07002462
Joonsoo Kim4e35f482014-04-03 14:47:30 -07002463 start = vma_hugecache_offset(h, vma, vma->vm_start);
2464 end = vma_hugecache_offset(h, vma, vma->vm_end);
Andy Whitcroft84afd992008-07-23 21:27:32 -07002465
Joonsoo Kim4e35f482014-04-03 14:47:30 -07002466 reserve = (end - start) - region_count(resv, start, end);
Andy Whitcroft84afd992008-07-23 21:27:32 -07002467
Joonsoo Kim4e35f482014-04-03 14:47:30 -07002468 kref_put(&resv->refs, resv_map_release);
2469
2470 if (reserve) {
2471 hugetlb_acct_memory(h, -reserve);
2472 hugepage_subpool_put_pages(spool, reserve);
Andy Whitcroft84afd992008-07-23 21:27:32 -07002473 }
Mel Gormana1e78772008-07-23 21:27:23 -07002474}
2475
Linus Torvalds1da177e2005-04-16 15:20:36 -07002476/*
2477 * We cannot handle pagefaults against hugetlb pages at all. They cause
2478 * handle_mm_fault() to try to instantiate regular-sized pages in the
2479 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
2480 * this far.
2481 */
Nick Piggind0217ac2007-07-19 01:47:03 -07002482static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002483{
2484 BUG();
Nick Piggind0217ac2007-07-19 01:47:03 -07002485 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002486}
2487
Alexey Dobriyanf0f37e22009-09-27 22:29:37 +04002488const struct vm_operations_struct hugetlb_vm_ops = {
Nick Piggind0217ac2007-07-19 01:47:03 -07002489 .fault = hugetlb_vm_op_fault,
Andy Whitcroft84afd992008-07-23 21:27:32 -07002490 .open = hugetlb_vm_op_open,
Mel Gormana1e78772008-07-23 21:27:23 -07002491 .close = hugetlb_vm_op_close,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002492};
2493
David Gibson1e8f8892006-01-06 00:10:44 -08002494static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
2495 int writable)
David Gibson63551ae2005-06-21 17:14:44 -07002496{
2497 pte_t entry;
2498
David Gibson1e8f8892006-01-06 00:10:44 -08002499 if (writable) {
Gerald Schaefer106c9922013-04-29 15:07:23 -07002500 entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page,
2501 vma->vm_page_prot)));
David Gibson63551ae2005-06-21 17:14:44 -07002502 } else {
Gerald Schaefer106c9922013-04-29 15:07:23 -07002503 entry = huge_pte_wrprotect(mk_huge_pte(page,
2504 vma->vm_page_prot));
David Gibson63551ae2005-06-21 17:14:44 -07002505 }
2506 entry = pte_mkyoung(entry);
2507 entry = pte_mkhuge(entry);
Chris Metcalfd9ed9fa2012-04-01 14:01:34 -04002508 entry = arch_make_huge_pte(entry, vma, page, writable);
David Gibson63551ae2005-06-21 17:14:44 -07002509
2510 return entry;
2511}
2512
David Gibson1e8f8892006-01-06 00:10:44 -08002513static void set_huge_ptep_writable(struct vm_area_struct *vma,
2514 unsigned long address, pte_t *ptep)
2515{
2516 pte_t entry;
2517
Gerald Schaefer106c9922013-04-29 15:07:23 -07002518 entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep)));
Chris Forbes32f84522011-07-25 17:12:14 -07002519 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
Russell King4b3073e2009-12-18 16:40:18 +00002520 update_mmu_cache(vma, address, ptep);
David Gibson1e8f8892006-01-06 00:10:44 -08002521}
2522
Naoya Horiguchi4a705fe2014-06-23 13:22:03 -07002523static int is_hugetlb_entry_migration(pte_t pte)
2524{
2525 swp_entry_t swp;
2526
2527 if (huge_pte_none(pte) || pte_present(pte))
2528 return 0;
2529 swp = pte_to_swp_entry(pte);
2530 if (non_swap_entry(swp) && is_migration_entry(swp))
2531 return 1;
2532 else
2533 return 0;
2534}
2535
2536static int is_hugetlb_entry_hwpoisoned(pte_t pte)
2537{
2538 swp_entry_t swp;
2539
2540 if (huge_pte_none(pte) || pte_present(pte))
2541 return 0;
2542 swp = pte_to_swp_entry(pte);
2543 if (non_swap_entry(swp) && is_hwpoison_entry(swp))
2544 return 1;
2545 else
2546 return 0;
2547}
David Gibson1e8f8892006-01-06 00:10:44 -08002548
David Gibson63551ae2005-06-21 17:14:44 -07002549int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
2550 struct vm_area_struct *vma)
2551{
2552 pte_t *src_pte, *dst_pte, entry;
2553 struct page *ptepage;
Hugh Dickins1c598272005-10-19 21:23:43 -07002554 unsigned long addr;
David Gibson1e8f8892006-01-06 00:10:44 -08002555 int cow;
Andi Kleena5516432008-07-23 21:27:41 -07002556 struct hstate *h = hstate_vma(vma);
2557 unsigned long sz = huge_page_size(h);
Andreas Sandberge8569dd2014-01-21 15:49:09 -08002558 unsigned long mmun_start; /* For mmu_notifiers */
2559 unsigned long mmun_end; /* For mmu_notifiers */
2560 int ret = 0;
David Gibson1e8f8892006-01-06 00:10:44 -08002561
2562 cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
David Gibson63551ae2005-06-21 17:14:44 -07002563
Andreas Sandberge8569dd2014-01-21 15:49:09 -08002564 mmun_start = vma->vm_start;
2565 mmun_end = vma->vm_end;
2566 if (cow)
2567 mmu_notifier_invalidate_range_start(src, mmun_start, mmun_end);
2568
Andi Kleena5516432008-07-23 21:27:41 -07002569 for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002570 spinlock_t *src_ptl, *dst_ptl;
Hugh Dickinsc74df322005-10-29 18:16:23 -07002571 src_pte = huge_pte_offset(src, addr);
2572 if (!src_pte)
2573 continue;
Andi Kleena5516432008-07-23 21:27:41 -07002574 dst_pte = huge_pte_alloc(dst, addr, sz);
Andreas Sandberge8569dd2014-01-21 15:49:09 -08002575 if (!dst_pte) {
2576 ret = -ENOMEM;
2577 break;
2578 }
Larry Woodmanc5c99422008-01-24 05:49:25 -08002579
2580 /* If the pagetables are shared don't copy or take references */
2581 if (dst_pte == src_pte)
2582 continue;
2583
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002584 dst_ptl = huge_pte_lock(h, dst, dst_pte);
2585 src_ptl = huge_pte_lockptr(h, src, src_pte);
2586 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
Naoya Horiguchi4a705fe2014-06-23 13:22:03 -07002587 entry = huge_ptep_get(src_pte);
2588 if (huge_pte_none(entry)) { /* skip none entry */
2589 ;
2590 } else if (unlikely(is_hugetlb_entry_migration(entry) ||
2591 is_hugetlb_entry_hwpoisoned(entry))) {
2592 swp_entry_t swp_entry = pte_to_swp_entry(entry);
2593
2594 if (is_write_migration_entry(swp_entry) && cow) {
2595 /*
2596 * COW mappings require pages in both
2597 * parent and child to be set to read.
2598 */
2599 make_migration_entry_read(&swp_entry);
2600 entry = swp_entry_to_pte(swp_entry);
2601 set_huge_pte_at(src, addr, src_pte, entry);
2602 }
2603 set_huge_pte_at(dst, addr, dst_pte, entry);
2604 } else {
David Gibson1e8f8892006-01-06 00:10:44 -08002605 if (cow)
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07002606 huge_ptep_set_wrprotect(src, addr, src_pte);
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);
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07002645again:
Andi Kleena5516432008-07-23 21:27:41 -07002646 for (address = start; address < end; address += sz) {
David Gibsonc7546f82005-08-05 11:59:35 -07002647 ptep = huge_pte_offset(mm, address);
Adam Litke4c887262005-10-29 18:16:46 -07002648 if (!ptep)
David Gibsonc7546f82005-08-05 11:59:35 -07002649 continue;
2650
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002651 ptl = huge_pte_lock(h, mm, ptep);
Chen, Kenneth W39dde652006-12-06 20:32:03 -08002652 if (huge_pmd_unshare(mm, &address, ptep))
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002653 goto unlock;
Chen, Kenneth W39dde652006-12-06 20:32:03 -08002654
Hillf Danton66293262012-03-23 15:01:48 -07002655 pte = huge_ptep_get(ptep);
2656 if (huge_pte_none(pte))
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002657 goto unlock;
Hillf Danton66293262012-03-23 15:01:48 -07002658
2659 /*
2660 * HWPoisoned hugepage is already unmapped and dropped reference
2661 */
Naoya Horiguchi8c4894c2012-12-12 13:52:28 -08002662 if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) {
Gerald Schaefer106c9922013-04-29 15:07:23 -07002663 huge_pte_clear(mm, address, ptep);
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002664 goto unlock;
Naoya Horiguchi8c4894c2012-12-12 13:52:28 -08002665 }
Hillf Danton66293262012-03-23 15:01:48 -07002666
2667 page = pte_page(pte);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002668 /*
2669 * If a reference page is supplied, it is because a specific
2670 * page is being unmapped, not a range. Ensure the page we
2671 * are about to unmap is the actual page of interest.
2672 */
2673 if (ref_page) {
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002674 if (page != ref_page)
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002675 goto unlock;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002676
2677 /*
2678 * Mark the VMA as having unmapped its page so that
2679 * future faults in this VMA will fail rather than
2680 * looking like data was lost
2681 */
2682 set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
2683 }
2684
David Gibsonc7546f82005-08-05 11:59:35 -07002685 pte = huge_ptep_get_and_clear(mm, address, ptep);
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07002686 tlb_remove_tlb_entry(tlb, ptep, address);
Gerald Schaefer106c9922013-04-29 15:07:23 -07002687 if (huge_pte_dirty(pte))
Ken Chen6649a382007-02-08 14:20:27 -08002688 set_page_dirty(page);
Hillf Danton9e811302012-03-21 16:34:03 -07002689
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07002690 page_remove_rmap(page);
2691 force_flush = !__tlb_remove_page(tlb, page);
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002692 if (force_flush) {
2693 spin_unlock(ptl);
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07002694 break;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002695 }
Hillf Danton9e811302012-03-21 16:34:03 -07002696 /* Bail out after unmapping reference page if supplied */
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002697 if (ref_page) {
2698 spin_unlock(ptl);
Hillf Danton9e811302012-03-21 16:34:03 -07002699 break;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002700 }
2701unlock:
2702 spin_unlock(ptl);
David Gibson63551ae2005-06-21 17:14:44 -07002703 }
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07002704 /*
2705 * mmu_gather ran out of room to batch pages, we break out of
2706 * the PTE lock to avoid doing the potential expensive TLB invalidate
2707 * and page-free while holding it.
2708 */
2709 if (force_flush) {
2710 force_flush = 0;
2711 tlb_flush_mmu(tlb);
2712 if (address < end && !ref_page)
2713 goto again;
Chen, Kenneth Wfe1668a2006-10-04 02:15:24 -07002714 }
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07002715 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07002716 tlb_end_vma(tlb, vma);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002717}
David Gibson63551ae2005-06-21 17:14:44 -07002718
Mel Gormand8333522012-07-31 16:46:20 -07002719void __unmap_hugepage_range_final(struct mmu_gather *tlb,
2720 struct vm_area_struct *vma, unsigned long start,
2721 unsigned long end, struct page *ref_page)
2722{
2723 __unmap_hugepage_range(tlb, vma, start, end, ref_page);
2724
2725 /*
2726 * Clear this flag so that x86's huge_pmd_share page_table_shareable
2727 * test will fail on a vma being torn down, and not grab a page table
2728 * on its way out. We're lucky that the flag has such an appropriate
2729 * name, and can in fact be safely cleared here. We could clear it
2730 * before the __unmap_hugepage_range above, but all that's necessary
2731 * is to clear it before releasing the i_mmap_mutex. This works
2732 * because in the context this is called, the VMA is about to be
2733 * destroyed and the i_mmap_mutex is held.
2734 */
2735 vma->vm_flags &= ~VM_MAYSHARE;
2736}
2737
Chen, Kenneth W502717f2006-10-11 01:20:46 -07002738void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002739 unsigned long end, struct page *ref_page)
Chen, Kenneth W502717f2006-10-11 01:20:46 -07002740{
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07002741 struct mm_struct *mm;
2742 struct mmu_gather tlb;
2743
2744 mm = vma->vm_mm;
2745
Linus Torvalds2b047252013-08-15 11:42:25 -07002746 tlb_gather_mmu(&tlb, mm, start, end);
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07002747 __unmap_hugepage_range(&tlb, vma, start, end, ref_page);
2748 tlb_finish_mmu(&tlb, start, end);
Chen, Kenneth W502717f2006-10-11 01:20:46 -07002749}
2750
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002751/*
2752 * This is called when the original mapper is failing to COW a MAP_PRIVATE
2753 * mappping it owns the reserve page for. The intention is to unmap the page
2754 * from other VMAs and let the children be SIGKILLed if they are faulting the
2755 * same region.
2756 */
Harvey Harrison2a4b3de2008-10-18 20:27:06 -07002757static int unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
2758 struct page *page, unsigned long address)
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002759{
Adam Litke75266742008-11-12 13:24:56 -08002760 struct hstate *h = hstate_vma(vma);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002761 struct vm_area_struct *iter_vma;
2762 struct address_space *mapping;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002763 pgoff_t pgoff;
2764
2765 /*
2766 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
2767 * from page cache lookup which is in HPAGE_SIZE units.
2768 */
Adam Litke75266742008-11-12 13:24:56 -08002769 address = address & huge_page_mask(h);
Michal Hocko36e4f202012-10-08 16:33:31 -07002770 pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) +
2771 vma->vm_pgoff;
Al Viro496ad9a2013-01-23 17:07:38 -05002772 mapping = file_inode(vma->vm_file)->i_mapping;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002773
Mel Gorman4eb2b1d2009-12-14 17:59:53 -08002774 /*
2775 * Take the mapping lock for the duration of the table walk. As
2776 * this mapping should be shared between all the VMAs,
2777 * __unmap_hugepage_range() is called as the lock is already held
2778 */
Peter Zijlstra3d48ae42011-05-24 17:12:06 -07002779 mutex_lock(&mapping->i_mmap_mutex);
Michel Lespinasse6b2dbba2012-10-08 16:31:25 -07002780 vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) {
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002781 /* Do not unmap the current VMA */
2782 if (iter_vma == vma)
2783 continue;
2784
2785 /*
2786 * Unmap the page from other VMAs without their own reserves.
2787 * They get marked to be SIGKILLed if they fault in these
2788 * areas. This is because a future no-page fault on this VMA
2789 * could insert a zeroed page instead of the data existing
2790 * from the time of fork. This would look like data corruption
2791 */
2792 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07002793 unmap_hugepage_range(iter_vma, address,
2794 address + huge_page_size(h), page);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002795 }
Peter Zijlstra3d48ae42011-05-24 17:12:06 -07002796 mutex_unlock(&mapping->i_mmap_mutex);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002797
2798 return 1;
2799}
2800
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002801/*
2802 * Hugetlb_cow() should be called with page lock of the original hugepage held.
Michal Hockoef009b22012-01-10 15:07:21 -08002803 * Called with hugetlb_instantiation_mutex held and pte_page locked so we
2804 * cannot race with other handlers or page migration.
2805 * Keep the pte_same checks anyway to make transition from the mutex easier.
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002806 */
David Gibson1e8f8892006-01-06 00:10:44 -08002807static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002808 unsigned long address, pte_t *ptep, pte_t pte,
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002809 struct page *pagecache_page, spinlock_t *ptl)
David Gibson1e8f8892006-01-06 00:10:44 -08002810{
Andi Kleena5516432008-07-23 21:27:41 -07002811 struct hstate *h = hstate_vma(vma);
David Gibson1e8f8892006-01-06 00:10:44 -08002812 struct page *old_page, *new_page;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002813 int outside_reserve = 0;
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07002814 unsigned long mmun_start; /* For mmu_notifiers */
2815 unsigned long mmun_end; /* For mmu_notifiers */
David Gibson1e8f8892006-01-06 00:10:44 -08002816
2817 old_page = pte_page(pte);
2818
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002819retry_avoidcopy:
David Gibson1e8f8892006-01-06 00:10:44 -08002820 /* If no-one else is actually using this page, avoid the copy
2821 * and just make the page writable */
Joonsoo Kim37a21402013-09-11 14:21:04 -07002822 if (page_mapcount(old_page) == 1 && PageAnon(old_page)) {
2823 page_move_anon_rmap(old_page, vma, address);
David Gibson1e8f8892006-01-06 00:10:44 -08002824 set_huge_ptep_writable(vma, address, ptep);
Nick Piggin83c54072007-07-19 01:47:05 -07002825 return 0;
David Gibson1e8f8892006-01-06 00:10:44 -08002826 }
2827
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002828 /*
2829 * If the process that created a MAP_PRIVATE mapping is about to
2830 * perform a COW due to a shared page count, attempt to satisfy
2831 * the allocation without using the existing reserves. The pagecache
2832 * page is used to determine if the reserve at this address was
2833 * consumed or not. If reserves were used, a partial faulted mapping
2834 * at the time of fork() could consume its reserves on COW instead
2835 * of the full address range.
2836 */
Joonsoo Kim5944d012013-09-11 14:21:55 -07002837 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002838 old_page != pagecache_page)
2839 outside_reserve = 1;
2840
David Gibson1e8f8892006-01-06 00:10:44 -08002841 page_cache_get(old_page);
Larry Woodmanb76c8cf2009-12-14 17:59:37 -08002842
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002843 /* Drop page table lock as buddy allocator may be called */
2844 spin_unlock(ptl);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002845 new_page = alloc_huge_page(vma, address, outside_reserve);
David Gibson1e8f8892006-01-06 00:10:44 -08002846
Adam Litke2fc39ce2007-11-14 16:59:39 -08002847 if (IS_ERR(new_page)) {
Aneesh Kumar K.V76dcee72012-07-31 16:41:57 -07002848 long err = PTR_ERR(new_page);
David Gibson1e8f8892006-01-06 00:10:44 -08002849 page_cache_release(old_page);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002850
2851 /*
2852 * If a process owning a MAP_PRIVATE mapping fails to COW,
2853 * it is due to references held by a child and an insufficient
2854 * huge page pool. To guarantee the original mappers
2855 * reliability, unmap the page from child processes. The child
2856 * may get SIGKILLed if it later faults.
2857 */
2858 if (outside_reserve) {
2859 BUG_ON(huge_pte_none(pte));
2860 if (unmap_ref_private(mm, vma, old_page, address)) {
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002861 BUG_ON(huge_pte_none(pte));
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002862 spin_lock(ptl);
Hillf Dantona734bcc2012-01-10 15:07:20 -08002863 ptep = huge_pte_offset(mm, address & huge_page_mask(h));
Naoya Horiguchia9af0c52014-04-07 15:36:54 -07002864 if (likely(ptep &&
2865 pte_same(huge_ptep_get(ptep), pte)))
Hillf Dantona734bcc2012-01-10 15:07:20 -08002866 goto retry_avoidcopy;
2867 /*
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002868 * race occurs while re-acquiring page table
2869 * lock, and our job is done.
Hillf Dantona734bcc2012-01-10 15:07:20 -08002870 */
2871 return 0;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002872 }
2873 WARN_ON_ONCE(1);
2874 }
2875
Larry Woodmanb76c8cf2009-12-14 17:59:37 -08002876 /* Caller expects lock to be held */
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002877 spin_lock(ptl);
Aneesh Kumar K.V76dcee72012-07-31 16:41:57 -07002878 if (err == -ENOMEM)
2879 return VM_FAULT_OOM;
2880 else
2881 return VM_FAULT_SIGBUS;
David Gibson1e8f8892006-01-06 00:10:44 -08002882 }
2883
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002884 /*
2885 * When the original hugepage is shared one, it does not have
2886 * anon_vma prepared.
2887 */
Dean Nelson44e2aa92010-10-26 14:22:08 -07002888 if (unlikely(anon_vma_prepare(vma))) {
Hillf Dantonea4039a2011-11-15 14:36:12 -08002889 page_cache_release(new_page);
2890 page_cache_release(old_page);
Dean Nelson44e2aa92010-10-26 14:22:08 -07002891 /* Caller expects lock to be held */
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002892 spin_lock(ptl);
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002893 return VM_FAULT_OOM;
Dean Nelson44e2aa92010-10-26 14:22:08 -07002894 }
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002895
Andrea Arcangeli47ad8472011-01-13 15:46:47 -08002896 copy_user_huge_page(new_page, old_page, address, vma,
2897 pages_per_huge_page(h));
Nick Piggin0ed361d2008-02-04 22:29:34 -08002898 __SetPageUptodate(new_page);
David Gibson1e8f8892006-01-06 00:10:44 -08002899
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07002900 mmun_start = address & huge_page_mask(h);
2901 mmun_end = mmun_start + huge_page_size(h);
2902 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
Larry Woodmanb76c8cf2009-12-14 17:59:37 -08002903 /*
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002904 * Retake the page table lock to check for racing updates
Larry Woodmanb76c8cf2009-12-14 17:59:37 -08002905 * before the page tables are altered
2906 */
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002907 spin_lock(ptl);
Andi Kleena5516432008-07-23 21:27:41 -07002908 ptep = huge_pte_offset(mm, address & huge_page_mask(h));
Naoya Horiguchia9af0c52014-04-07 15:36:54 -07002909 if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) {
Joonsoo Kim07443a82013-09-11 14:21:58 -07002910 ClearPagePrivate(new_page);
2911
David Gibson1e8f8892006-01-06 00:10:44 -08002912 /* Break COW */
Gerald Schaefer8fe627e2008-04-28 02:13:28 -07002913 huge_ptep_clear_flush(vma, address, ptep);
David Gibson1e8f8892006-01-06 00:10:44 -08002914 set_huge_pte_at(mm, address, ptep,
2915 make_huge_pte(vma, new_page, 1));
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09002916 page_remove_rmap(old_page);
Naoya Horiguchicd67f0d2010-09-10 13:23:04 +09002917 hugepage_add_new_anon_rmap(new_page, vma, address);
David Gibson1e8f8892006-01-06 00:10:44 -08002918 /* Make the old page be freed below */
2919 new_page = old_page;
2920 }
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002921 spin_unlock(ptl);
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07002922 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
David Gibson1e8f8892006-01-06 00:10:44 -08002923 page_cache_release(new_page);
2924 page_cache_release(old_page);
Joonsoo Kim83120342013-09-11 14:21:57 -07002925
2926 /* Caller expects lock to be held */
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002927 spin_lock(ptl);
Nick Piggin83c54072007-07-19 01:47:05 -07002928 return 0;
David Gibson1e8f8892006-01-06 00:10:44 -08002929}
2930
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002931/* Return the pagecache page at a given address within a VMA */
Andi Kleena5516432008-07-23 21:27:41 -07002932static struct page *hugetlbfs_pagecache_page(struct hstate *h,
2933 struct vm_area_struct *vma, unsigned long address)
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002934{
2935 struct address_space *mapping;
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -07002936 pgoff_t idx;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002937
2938 mapping = vma->vm_file->f_mapping;
Andi Kleena5516432008-07-23 21:27:41 -07002939 idx = vma_hugecache_offset(h, vma, address);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002940
2941 return find_lock_page(mapping, idx);
2942}
2943
Hugh Dickins3ae77f42009-09-21 17:03:33 -07002944/*
2945 * Return whether there is a pagecache page to back given address within VMA.
2946 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
2947 */
2948static bool hugetlbfs_pagecache_present(struct hstate *h,
Hugh Dickins2a15efc2009-09-21 17:03:27 -07002949 struct vm_area_struct *vma, unsigned long address)
2950{
2951 struct address_space *mapping;
2952 pgoff_t idx;
2953 struct page *page;
2954
2955 mapping = vma->vm_file->f_mapping;
2956 idx = vma_hugecache_offset(h, vma, address);
2957
2958 page = find_get_page(mapping, idx);
2959 if (page)
2960 put_page(page);
2961 return page != NULL;
2962}
2963
Robert P. J. Daya1ed3dd2007-07-17 04:03:33 -07002964static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
Davidlohr Bueso8382d912014-04-03 14:47:31 -07002965 struct address_space *mapping, pgoff_t idx,
2966 unsigned long address, pte_t *ptep, unsigned int flags)
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01002967{
Andi Kleena5516432008-07-23 21:27:41 -07002968 struct hstate *h = hstate_vma(vma);
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01002969 int ret = VM_FAULT_SIGBUS;
Hillf Danton409eb8c2012-01-20 14:34:13 -08002970 int anon_rmap = 0;
Adam Litke4c887262005-10-29 18:16:46 -07002971 unsigned long size;
Adam Litke4c887262005-10-29 18:16:46 -07002972 struct page *page;
David Gibson1e8f8892006-01-06 00:10:44 -08002973 pte_t new_pte;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08002974 spinlock_t *ptl;
Adam Litke4c887262005-10-29 18:16:46 -07002975
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002976 /*
2977 * Currently, we are forced to kill the process in the event the
2978 * original mapper has unmapped pages from the child due to a failed
Lucas De Marchi25985ed2011-03-30 22:57:33 -03002979 * COW. Warn that such a situation has occurred as it may not be obvious
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002980 */
2981 if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
Andrew Mortonffb22af2013-02-22 16:32:08 -08002982 pr_warning("PID %d killed due to inadequate hugepage pool\n",
2983 current->pid);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002984 return ret;
2985 }
2986
Adam Litke4c887262005-10-29 18:16:46 -07002987 /*
2988 * Use page lock to guard against racing truncation
2989 * before we get page_table_lock.
2990 */
Christoph Lameter6bda6662006-01-06 00:10:49 -08002991retry:
2992 page = find_lock_page(mapping, idx);
2993 if (!page) {
Andi Kleena5516432008-07-23 21:27:41 -07002994 size = i_size_read(mapping->host) >> huge_page_shift(h);
Hugh Dickinsebed4bf2006-10-28 10:38:43 -07002995 if (idx >= size)
2996 goto out;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002997 page = alloc_huge_page(vma, address, 0);
Adam Litke2fc39ce2007-11-14 16:59:39 -08002998 if (IS_ERR(page)) {
Aneesh Kumar K.V76dcee72012-07-31 16:41:57 -07002999 ret = PTR_ERR(page);
3000 if (ret == -ENOMEM)
3001 ret = VM_FAULT_OOM;
3002 else
3003 ret = VM_FAULT_SIGBUS;
Christoph Lameter6bda6662006-01-06 00:10:49 -08003004 goto out;
3005 }
Andrea Arcangeli47ad8472011-01-13 15:46:47 -08003006 clear_huge_page(page, address, pages_per_huge_page(h));
Nick Piggin0ed361d2008-02-04 22:29:34 -08003007 __SetPageUptodate(page);
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01003008
Mel Gormanf83a2752009-05-28 14:34:40 -07003009 if (vma->vm_flags & VM_MAYSHARE) {
Christoph Lameter6bda6662006-01-06 00:10:49 -08003010 int err;
Ken Chen45c682a2007-11-14 16:59:44 -08003011 struct inode *inode = mapping->host;
Christoph Lameter6bda6662006-01-06 00:10:49 -08003012
3013 err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
3014 if (err) {
3015 put_page(page);
Christoph Lameter6bda6662006-01-06 00:10:49 -08003016 if (err == -EEXIST)
3017 goto retry;
3018 goto out;
3019 }
Joonsoo Kim07443a82013-09-11 14:21:58 -07003020 ClearPagePrivate(page);
Ken Chen45c682a2007-11-14 16:59:44 -08003021
3022 spin_lock(&inode->i_lock);
Andi Kleena5516432008-07-23 21:27:41 -07003023 inode->i_blocks += blocks_per_huge_page(h);
Ken Chen45c682a2007-11-14 16:59:44 -08003024 spin_unlock(&inode->i_lock);
Mel Gorman23be7462010-04-23 13:17:56 -04003025 } else {
Christoph Lameter6bda6662006-01-06 00:10:49 -08003026 lock_page(page);
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09003027 if (unlikely(anon_vma_prepare(vma))) {
3028 ret = VM_FAULT_OOM;
3029 goto backout_unlocked;
3030 }
Hillf Danton409eb8c2012-01-20 14:34:13 -08003031 anon_rmap = 1;
Mel Gorman23be7462010-04-23 13:17:56 -04003032 }
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09003033 } else {
Naoya Horiguchi998b4382010-09-08 10:19:32 +09003034 /*
3035 * If memory error occurs between mmap() and fault, some process
3036 * don't have hwpoisoned swap entry for errored virtual address.
3037 * So we need to block hugepage fault by PG_hwpoison bit check.
3038 */
3039 if (unlikely(PageHWPoison(page))) {
Chris Forbes32f84522011-07-25 17:12:14 -07003040 ret = VM_FAULT_HWPOISON |
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07003041 VM_FAULT_SET_HINDEX(hstate_index(h));
Naoya Horiguchi998b4382010-09-08 10:19:32 +09003042 goto backout_unlocked;
3043 }
Christoph Lameter6bda6662006-01-06 00:10:49 -08003044 }
David Gibson1e8f8892006-01-06 00:10:44 -08003045
Andy Whitcroft57303d82008-08-12 15:08:47 -07003046 /*
3047 * If we are going to COW a private mapping later, we examine the
3048 * pending reservations for this page now. This will ensure that
3049 * any allocations necessary to record that reservation occur outside
3050 * the spinlock.
3051 */
Hugh Dickins788c7df2009-06-23 13:49:05 +01003052 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED))
Andy Whitcroft2b267362008-08-12 15:08:49 -07003053 if (vma_needs_reservation(h, vma, address) < 0) {
3054 ret = VM_FAULT_OOM;
3055 goto backout_unlocked;
3056 }
Andy Whitcroft57303d82008-08-12 15:08:47 -07003057
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003058 ptl = huge_pte_lockptr(h, mm, ptep);
3059 spin_lock(ptl);
Andi Kleena5516432008-07-23 21:27:41 -07003060 size = i_size_read(mapping->host) >> huge_page_shift(h);
Adam Litke4c887262005-10-29 18:16:46 -07003061 if (idx >= size)
3062 goto backout;
3063
Nick Piggin83c54072007-07-19 01:47:05 -07003064 ret = 0;
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07003065 if (!huge_pte_none(huge_ptep_get(ptep)))
Adam Litke4c887262005-10-29 18:16:46 -07003066 goto backout;
3067
Joonsoo Kim07443a82013-09-11 14:21:58 -07003068 if (anon_rmap) {
3069 ClearPagePrivate(page);
Hillf Danton409eb8c2012-01-20 14:34:13 -08003070 hugepage_add_new_anon_rmap(page, vma, address);
Choi Gi-yongac714902014-04-07 15:37:36 -07003071 } else
Hillf Danton409eb8c2012-01-20 14:34:13 -08003072 page_dup_rmap(page);
David Gibson1e8f8892006-01-06 00:10:44 -08003073 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
3074 && (vma->vm_flags & VM_SHARED)));
3075 set_huge_pte_at(mm, address, ptep, new_pte);
3076
Hugh Dickins788c7df2009-06-23 13:49:05 +01003077 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
David Gibson1e8f8892006-01-06 00:10:44 -08003078 /* Optimization, do the COW without a second fault */
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003079 ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page, ptl);
David Gibson1e8f8892006-01-06 00:10:44 -08003080 }
3081
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003082 spin_unlock(ptl);
Adam Litke4c887262005-10-29 18:16:46 -07003083 unlock_page(page);
3084out:
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01003085 return ret;
Adam Litke4c887262005-10-29 18:16:46 -07003086
3087backout:
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003088 spin_unlock(ptl);
Andy Whitcroft2b267362008-08-12 15:08:49 -07003089backout_unlocked:
Adam Litke4c887262005-10-29 18:16:46 -07003090 unlock_page(page);
3091 put_page(page);
3092 goto out;
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01003093}
3094
Davidlohr Bueso8382d912014-04-03 14:47:31 -07003095#ifdef CONFIG_SMP
3096static u32 fault_mutex_hash(struct hstate *h, struct mm_struct *mm,
3097 struct vm_area_struct *vma,
3098 struct address_space *mapping,
3099 pgoff_t idx, unsigned long address)
3100{
3101 unsigned long key[2];
3102 u32 hash;
3103
3104 if (vma->vm_flags & VM_SHARED) {
3105 key[0] = (unsigned long) mapping;
3106 key[1] = idx;
3107 } else {
3108 key[0] = (unsigned long) mm;
3109 key[1] = address >> huge_page_shift(h);
3110 }
3111
3112 hash = jhash2((u32 *)&key, sizeof(key)/sizeof(u32), 0);
3113
3114 return hash & (num_fault_mutexes - 1);
3115}
3116#else
3117/*
3118 * For uniprocesor systems we always use a single mutex, so just
3119 * return 0 and avoid the hashing overhead.
3120 */
3121static u32 fault_mutex_hash(struct hstate *h, struct mm_struct *mm,
3122 struct vm_area_struct *vma,
3123 struct address_space *mapping,
3124 pgoff_t idx, unsigned long address)
3125{
3126 return 0;
3127}
3128#endif
3129
Adam Litke86e52162006-01-06 00:10:43 -08003130int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
Hugh Dickins788c7df2009-06-23 13:49:05 +01003131 unsigned long address, unsigned int flags)
Adam Litke86e52162006-01-06 00:10:43 -08003132{
Davidlohr Bueso8382d912014-04-03 14:47:31 -07003133 pte_t *ptep, entry;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003134 spinlock_t *ptl;
David Gibson1e8f8892006-01-06 00:10:44 -08003135 int ret;
Davidlohr Bueso8382d912014-04-03 14:47:31 -07003136 u32 hash;
3137 pgoff_t idx;
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09003138 struct page *page = NULL;
Andy Whitcroft57303d82008-08-12 15:08:47 -07003139 struct page *pagecache_page = NULL;
Andi Kleena5516432008-07-23 21:27:41 -07003140 struct hstate *h = hstate_vma(vma);
Davidlohr Bueso8382d912014-04-03 14:47:31 -07003141 struct address_space *mapping;
Adam Litke86e52162006-01-06 00:10:43 -08003142
KAMEZAWA Hiroyuki1e16a532012-01-10 15:07:22 -08003143 address &= huge_page_mask(h);
3144
Naoya Horiguchifd6a03e2010-05-28 09:29:21 +09003145 ptep = huge_pte_offset(mm, address);
3146 if (ptep) {
3147 entry = huge_ptep_get(ptep);
Naoya Horiguchi290408d2010-09-08 10:19:35 +09003148 if (unlikely(is_hugetlb_entry_migration(entry))) {
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003149 migration_entry_wait_huge(vma, mm, ptep);
Naoya Horiguchi290408d2010-09-08 10:19:35 +09003150 return 0;
3151 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
Chris Forbes32f84522011-07-25 17:12:14 -07003152 return VM_FAULT_HWPOISON_LARGE |
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07003153 VM_FAULT_SET_HINDEX(hstate_index(h));
Naoya Horiguchifd6a03e2010-05-28 09:29:21 +09003154 }
3155
Andi Kleena5516432008-07-23 21:27:41 -07003156 ptep = huge_pte_alloc(mm, address, huge_page_size(h));
Adam Litke86e52162006-01-06 00:10:43 -08003157 if (!ptep)
3158 return VM_FAULT_OOM;
3159
Davidlohr Bueso8382d912014-04-03 14:47:31 -07003160 mapping = vma->vm_file->f_mapping;
3161 idx = vma_hugecache_offset(h, vma, address);
3162
David Gibson3935baa2006-03-22 00:08:53 -08003163 /*
3164 * Serialize hugepage allocation and instantiation, so that we don't
3165 * get spurious allocation failures if two CPUs race to instantiate
3166 * the same page in the page cache.
3167 */
Davidlohr Bueso8382d912014-04-03 14:47:31 -07003168 hash = fault_mutex_hash(h, mm, vma, mapping, idx, address);
3169 mutex_lock(&htlb_fault_mutex_table[hash]);
3170
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07003171 entry = huge_ptep_get(ptep);
3172 if (huge_pte_none(entry)) {
Davidlohr Bueso8382d912014-04-03 14:47:31 -07003173 ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags);
David Gibsonb4d1d992008-10-15 22:01:11 -07003174 goto out_mutex;
David Gibson3935baa2006-03-22 00:08:53 -08003175 }
Adam Litke86e52162006-01-06 00:10:43 -08003176
Nick Piggin83c54072007-07-19 01:47:05 -07003177 ret = 0;
David Gibson1e8f8892006-01-06 00:10:44 -08003178
Andy Whitcroft57303d82008-08-12 15:08:47 -07003179 /*
3180 * If we are going to COW the mapping later, we examine the pending
3181 * reservations for this page now. This will ensure that any
3182 * allocations necessary to record that reservation occur outside the
3183 * spinlock. For private mappings, we also lookup the pagecache
3184 * page now as it is used to determine if a reservation has been
3185 * consumed.
3186 */
Gerald Schaefer106c9922013-04-29 15:07:23 -07003187 if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) {
Andy Whitcroft2b267362008-08-12 15:08:49 -07003188 if (vma_needs_reservation(h, vma, address) < 0) {
3189 ret = VM_FAULT_OOM;
David Gibsonb4d1d992008-10-15 22:01:11 -07003190 goto out_mutex;
Andy Whitcroft2b267362008-08-12 15:08:49 -07003191 }
Andy Whitcroft57303d82008-08-12 15:08:47 -07003192
Mel Gormanf83a2752009-05-28 14:34:40 -07003193 if (!(vma->vm_flags & VM_MAYSHARE))
Andy Whitcroft57303d82008-08-12 15:08:47 -07003194 pagecache_page = hugetlbfs_pagecache_page(h,
3195 vma, address);
3196 }
3197
Naoya Horiguchi56c9cfb2010-09-10 13:23:04 +09003198 /*
3199 * hugetlb_cow() requires page locks of pte_page(entry) and
3200 * pagecache_page, so here we need take the former one
3201 * when page != pagecache_page or !pagecache_page.
3202 * Note that locking order is always pagecache_page -> page,
3203 * so no worry about deadlock.
3204 */
3205 page = pte_page(entry);
Chris Metcalf66aebce2012-04-12 12:49:15 -07003206 get_page(page);
Naoya Horiguchi56c9cfb2010-09-10 13:23:04 +09003207 if (page != pagecache_page)
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09003208 lock_page(page);
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09003209
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003210 ptl = huge_pte_lockptr(h, mm, ptep);
3211 spin_lock(ptl);
David Gibson1e8f8892006-01-06 00:10:44 -08003212 /* Check for a racing update before calling hugetlb_cow */
David Gibsonb4d1d992008-10-15 22:01:11 -07003213 if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003214 goto out_ptl;
David Gibsonb4d1d992008-10-15 22:01:11 -07003215
3216
Hugh Dickins788c7df2009-06-23 13:49:05 +01003217 if (flags & FAULT_FLAG_WRITE) {
Gerald Schaefer106c9922013-04-29 15:07:23 -07003218 if (!huge_pte_write(entry)) {
Andy Whitcroft57303d82008-08-12 15:08:47 -07003219 ret = hugetlb_cow(mm, vma, address, ptep, entry,
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003220 pagecache_page, ptl);
3221 goto out_ptl;
David Gibsonb4d1d992008-10-15 22:01:11 -07003222 }
Gerald Schaefer106c9922013-04-29 15:07:23 -07003223 entry = huge_pte_mkdirty(entry);
David Gibsonb4d1d992008-10-15 22:01:11 -07003224 }
3225 entry = pte_mkyoung(entry);
Hugh Dickins788c7df2009-06-23 13:49:05 +01003226 if (huge_ptep_set_access_flags(vma, address, ptep, entry,
3227 flags & FAULT_FLAG_WRITE))
Russell King4b3073e2009-12-18 16:40:18 +00003228 update_mmu_cache(vma, address, ptep);
David Gibsonb4d1d992008-10-15 22:01:11 -07003229
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003230out_ptl:
3231 spin_unlock(ptl);
Andy Whitcroft57303d82008-08-12 15:08:47 -07003232
3233 if (pagecache_page) {
3234 unlock_page(pagecache_page);
3235 put_page(pagecache_page);
3236 }
Dean Nelson1f64d692010-12-02 14:31:12 -08003237 if (page != pagecache_page)
3238 unlock_page(page);
Chris Metcalf66aebce2012-04-12 12:49:15 -07003239 put_page(page);
Andy Whitcroft57303d82008-08-12 15:08:47 -07003240
David Gibsonb4d1d992008-10-15 22:01:11 -07003241out_mutex:
Davidlohr Bueso8382d912014-04-03 14:47:31 -07003242 mutex_unlock(&htlb_fault_mutex_table[hash]);
David Gibson1e8f8892006-01-06 00:10:44 -08003243 return ret;
Adam Litke86e52162006-01-06 00:10:43 -08003244}
3245
Michel Lespinasse28a35712013-02-22 16:35:55 -08003246long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
3247 struct page **pages, struct vm_area_struct **vmas,
3248 unsigned long *position, unsigned long *nr_pages,
3249 long i, unsigned int flags)
David Gibson63551ae2005-06-21 17:14:44 -07003250{
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08003251 unsigned long pfn_offset;
3252 unsigned long vaddr = *position;
Michel Lespinasse28a35712013-02-22 16:35:55 -08003253 unsigned long remainder = *nr_pages;
Andi Kleena5516432008-07-23 21:27:41 -07003254 struct hstate *h = hstate_vma(vma);
David Gibson63551ae2005-06-21 17:14:44 -07003255
David Gibson63551ae2005-06-21 17:14:44 -07003256 while (vaddr < vma->vm_end && remainder) {
Adam Litke4c887262005-10-29 18:16:46 -07003257 pte_t *pte;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003258 spinlock_t *ptl = NULL;
Hugh Dickins2a15efc2009-09-21 17:03:27 -07003259 int absent;
Adam Litke4c887262005-10-29 18:16:46 -07003260 struct page *page;
3261
3262 /*
3263 * Some archs (sparc64, sh*) have multiple pte_ts to
Hugh Dickins2a15efc2009-09-21 17:03:27 -07003264 * each hugepage. We have to make sure we get the
Adam Litke4c887262005-10-29 18:16:46 -07003265 * first, for the page indexing below to work.
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003266 *
3267 * Note that page table lock is not held when pte is null.
Adam Litke4c887262005-10-29 18:16:46 -07003268 */
Andi Kleena5516432008-07-23 21:27:41 -07003269 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h));
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003270 if (pte)
3271 ptl = huge_pte_lock(h, mm, pte);
Hugh Dickins2a15efc2009-09-21 17:03:27 -07003272 absent = !pte || huge_pte_none(huge_ptep_get(pte));
Adam Litke4c887262005-10-29 18:16:46 -07003273
Hugh Dickins2a15efc2009-09-21 17:03:27 -07003274 /*
3275 * When coredumping, it suits get_dump_page if we just return
Hugh Dickins3ae77f42009-09-21 17:03:33 -07003276 * an error where there's an empty slot with no huge pagecache
3277 * to back it. This way, we avoid allocating a hugepage, and
3278 * the sparse dumpfile avoids allocating disk blocks, but its
3279 * huge holes still show up with zeroes where they need to be.
Hugh Dickins2a15efc2009-09-21 17:03:27 -07003280 */
Hugh Dickins3ae77f42009-09-21 17:03:33 -07003281 if (absent && (flags & FOLL_DUMP) &&
3282 !hugetlbfs_pagecache_present(h, vma, vaddr)) {
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003283 if (pte)
3284 spin_unlock(ptl);
Hugh Dickins2a15efc2009-09-21 17:03:27 -07003285 remainder = 0;
3286 break;
3287 }
3288
Naoya Horiguchi9cc3a5b2013-04-17 15:58:30 -07003289 /*
3290 * We need call hugetlb_fault for both hugepages under migration
3291 * (in which case hugetlb_fault waits for the migration,) and
3292 * hwpoisoned hugepages (in which case we need to prevent the
3293 * caller from accessing to them.) In order to do this, we use
3294 * here is_swap_pte instead of is_hugetlb_entry_migration and
3295 * is_hugetlb_entry_hwpoisoned. This is because it simply covers
3296 * both cases, and because we can't follow correct pages
3297 * directly from any kind of swap entries.
3298 */
3299 if (absent || is_swap_pte(huge_ptep_get(pte)) ||
Gerald Schaefer106c9922013-04-29 15:07:23 -07003300 ((flags & FOLL_WRITE) &&
3301 !huge_pte_write(huge_ptep_get(pte)))) {
Adam Litke4c887262005-10-29 18:16:46 -07003302 int ret;
3303
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003304 if (pte)
3305 spin_unlock(ptl);
Hugh Dickins2a15efc2009-09-21 17:03:27 -07003306 ret = hugetlb_fault(mm, vma, vaddr,
3307 (flags & FOLL_WRITE) ? FAULT_FLAG_WRITE : 0);
Adam Litkea89182c2007-08-22 14:01:51 -07003308 if (!(ret & VM_FAULT_ERROR))
Adam Litke4c887262005-10-29 18:16:46 -07003309 continue;
3310
3311 remainder = 0;
Adam Litke4c887262005-10-29 18:16:46 -07003312 break;
3313 }
David Gibson63551ae2005-06-21 17:14:44 -07003314
Andi Kleena5516432008-07-23 21:27:41 -07003315 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07003316 page = pte_page(huge_ptep_get(pte));
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08003317same_page:
Chen, Kenneth Wd6692182006-03-31 02:29:57 -08003318 if (pages) {
Hugh Dickins2a15efc2009-09-21 17:03:27 -07003319 pages[i] = mem_map_offset(page, pfn_offset);
Andrea Arcangelia0368d42014-01-21 15:48:49 -08003320 get_page_foll(pages[i]);
Chen, Kenneth Wd6692182006-03-31 02:29:57 -08003321 }
David Gibson63551ae2005-06-21 17:14:44 -07003322
3323 if (vmas)
3324 vmas[i] = vma;
3325
3326 vaddr += PAGE_SIZE;
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08003327 ++pfn_offset;
David Gibson63551ae2005-06-21 17:14:44 -07003328 --remainder;
3329 ++i;
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08003330 if (vaddr < vma->vm_end && remainder &&
Andi Kleena5516432008-07-23 21:27:41 -07003331 pfn_offset < pages_per_huge_page(h)) {
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08003332 /*
3333 * We use pfn_offset to avoid touching the pageframes
3334 * of this compound page.
3335 */
3336 goto same_page;
3337 }
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003338 spin_unlock(ptl);
David Gibson63551ae2005-06-21 17:14:44 -07003339 }
Michel Lespinasse28a35712013-02-22 16:35:55 -08003340 *nr_pages = remainder;
David Gibson63551ae2005-06-21 17:14:44 -07003341 *position = vaddr;
3342
Hugh Dickins2a15efc2009-09-21 17:03:27 -07003343 return i ? i : -EFAULT;
David Gibson63551ae2005-06-21 17:14:44 -07003344}
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003345
Peter Zijlstra7da4d642012-11-19 03:14:23 +01003346unsigned long hugetlb_change_protection(struct vm_area_struct *vma,
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003347 unsigned long address, unsigned long end, pgprot_t newprot)
3348{
3349 struct mm_struct *mm = vma->vm_mm;
3350 unsigned long start = address;
3351 pte_t *ptep;
3352 pte_t pte;
Andi Kleena5516432008-07-23 21:27:41 -07003353 struct hstate *h = hstate_vma(vma);
Peter Zijlstra7da4d642012-11-19 03:14:23 +01003354 unsigned long pages = 0;
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003355
3356 BUG_ON(address >= end);
3357 flush_cache_range(vma, address, end);
3358
Rik van Riela5338092014-04-07 15:36:57 -07003359 mmu_notifier_invalidate_range_start(mm, start, end);
Peter Zijlstra3d48ae42011-05-24 17:12:06 -07003360 mutex_lock(&vma->vm_file->f_mapping->i_mmap_mutex);
Andi Kleena5516432008-07-23 21:27:41 -07003361 for (; address < end; address += huge_page_size(h)) {
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003362 spinlock_t *ptl;
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003363 ptep = huge_pte_offset(mm, address);
3364 if (!ptep)
3365 continue;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003366 ptl = huge_pte_lock(h, mm, ptep);
Peter Zijlstra7da4d642012-11-19 03:14:23 +01003367 if (huge_pmd_unshare(mm, &address, ptep)) {
3368 pages++;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003369 spin_unlock(ptl);
Chen, Kenneth W39dde652006-12-06 20:32:03 -08003370 continue;
Peter Zijlstra7da4d642012-11-19 03:14:23 +01003371 }
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07003372 if (!huge_pte_none(huge_ptep_get(ptep))) {
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003373 pte = huge_ptep_get_and_clear(mm, address, ptep);
Gerald Schaefer106c9922013-04-29 15:07:23 -07003374 pte = pte_mkhuge(huge_pte_modify(pte, newprot));
Tony Lube7517d2013-02-04 14:28:46 -08003375 pte = arch_make_huge_pte(pte, vma, NULL, 0);
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003376 set_huge_pte_at(mm, address, ptep, pte);
Peter Zijlstra7da4d642012-11-19 03:14:23 +01003377 pages++;
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003378 }
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003379 spin_unlock(ptl);
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003380 }
Mel Gormand8333522012-07-31 16:46:20 -07003381 /*
3382 * Must flush TLB before releasing i_mmap_mutex: x86's huge_pmd_unshare
3383 * may have cleared our pud entry and done put_page on the page table:
3384 * once we release i_mmap_mutex, another task can do the final put_page
3385 * and that page table be reused and filled with junk.
3386 */
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003387 flush_tlb_range(vma, start, end);
Mel Gormand8333522012-07-31 16:46:20 -07003388 mutex_unlock(&vma->vm_file->f_mapping->i_mmap_mutex);
Rik van Riela5338092014-04-07 15:36:57 -07003389 mmu_notifier_invalidate_range_end(mm, start, end);
Peter Zijlstra7da4d642012-11-19 03:14:23 +01003390
3391 return pages << h->order;
Zhang, Yanmin8f860592006-03-22 00:08:50 -08003392}
3393
Mel Gormana1e78772008-07-23 21:27:23 -07003394int hugetlb_reserve_pages(struct inode *inode,
3395 long from, long to,
Mel Gorman5a6fe122009-02-10 14:02:27 +00003396 struct vm_area_struct *vma,
KOSAKI Motohiroca16d142011-05-26 19:16:19 +09003397 vm_flags_t vm_flags)
Adam Litkee4e574b2007-10-16 01:26:19 -07003398{
Mel Gorman17c9d122009-02-11 16:34:16 +00003399 long ret, chg;
Andi Kleena5516432008-07-23 21:27:41 -07003400 struct hstate *h = hstate_inode(inode);
David Gibson90481622012-03-21 16:34:12 -07003401 struct hugepage_subpool *spool = subpool_inode(inode);
Joonsoo Kim9119a412014-04-03 14:47:25 -07003402 struct resv_map *resv_map;
Adam Litkee4e574b2007-10-16 01:26:19 -07003403
Mel Gormana1e78772008-07-23 21:27:23 -07003404 /*
Mel Gorman17c9d122009-02-11 16:34:16 +00003405 * Only apply hugepage reservation if asked. At fault time, an
3406 * attempt will be made for VM_NORESERVE to allocate a page
David Gibson90481622012-03-21 16:34:12 -07003407 * without using reserves
Mel Gorman17c9d122009-02-11 16:34:16 +00003408 */
KOSAKI Motohiroca16d142011-05-26 19:16:19 +09003409 if (vm_flags & VM_NORESERVE)
Mel Gorman17c9d122009-02-11 16:34:16 +00003410 return 0;
3411
3412 /*
Mel Gormana1e78772008-07-23 21:27:23 -07003413 * Shared mappings base their reservation on the number of pages that
3414 * are already allocated on behalf of the file. Private mappings need
3415 * to reserve the full area even if read-only as mprotect() may be
3416 * called to make the mapping read-write. Assume !vma is a shm mapping
3417 */
Joonsoo Kim9119a412014-04-03 14:47:25 -07003418 if (!vma || vma->vm_flags & VM_MAYSHARE) {
Joonsoo Kim4e35f482014-04-03 14:47:30 -07003419 resv_map = inode_resv_map(inode);
Joonsoo Kim9119a412014-04-03 14:47:25 -07003420
Joonsoo Kim1406ec92014-04-03 14:47:26 -07003421 chg = region_chg(resv_map, from, to);
Joonsoo Kim9119a412014-04-03 14:47:25 -07003422
3423 } else {
3424 resv_map = resv_map_alloc();
Mel Gorman5a6fe122009-02-10 14:02:27 +00003425 if (!resv_map)
3426 return -ENOMEM;
3427
Mel Gorman17c9d122009-02-11 16:34:16 +00003428 chg = to - from;
3429
Mel Gorman5a6fe122009-02-10 14:02:27 +00003430 set_vma_resv_map(vma, resv_map);
3431 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
3432 }
3433
Dave Hansenc50ac052012-05-29 15:06:46 -07003434 if (chg < 0) {
3435 ret = chg;
3436 goto out_err;
3437 }
Mel Gorman17c9d122009-02-11 16:34:16 +00003438
David Gibson90481622012-03-21 16:34:12 -07003439 /* There must be enough pages in the subpool for the mapping */
Dave Hansenc50ac052012-05-29 15:06:46 -07003440 if (hugepage_subpool_get_pages(spool, chg)) {
3441 ret = -ENOSPC;
3442 goto out_err;
3443 }
Mel Gorman17c9d122009-02-11 16:34:16 +00003444
3445 /*
3446 * Check enough hugepages are available for the reservation.
David Gibson90481622012-03-21 16:34:12 -07003447 * Hand the pages back to the subpool if there are not
Mel Gorman17c9d122009-02-11 16:34:16 +00003448 */
3449 ret = hugetlb_acct_memory(h, chg);
3450 if (ret < 0) {
David Gibson90481622012-03-21 16:34:12 -07003451 hugepage_subpool_put_pages(spool, chg);
Dave Hansenc50ac052012-05-29 15:06:46 -07003452 goto out_err;
Mel Gorman17c9d122009-02-11 16:34:16 +00003453 }
3454
3455 /*
3456 * Account for the reservations made. Shared mappings record regions
3457 * that have reservations as they are shared by multiple VMAs.
3458 * When the last VMA disappears, the region map says how much
3459 * the reservation was and the page cache tells how much of
3460 * the reservation was consumed. Private mappings are per-VMA and
3461 * only the consumed reservations are tracked. When the VMA
3462 * disappears, the original reservation is the VMA size and the
3463 * consumed reservations are stored in the map. Hence, nothing
3464 * else has to be done for private mappings here
3465 */
Mel Gormanf83a2752009-05-28 14:34:40 -07003466 if (!vma || vma->vm_flags & VM_MAYSHARE)
Joonsoo Kim1406ec92014-04-03 14:47:26 -07003467 region_add(resv_map, from, to);
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -07003468 return 0;
Dave Hansenc50ac052012-05-29 15:06:46 -07003469out_err:
Joonsoo Kimf031dd22014-04-03 14:47:28 -07003470 if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
3471 kref_put(&resv_map->refs, resv_map_release);
Dave Hansenc50ac052012-05-29 15:06:46 -07003472 return ret;
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -07003473}
3474
3475void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed)
3476{
Andi Kleena5516432008-07-23 21:27:41 -07003477 struct hstate *h = hstate_inode(inode);
Joonsoo Kim4e35f482014-04-03 14:47:30 -07003478 struct resv_map *resv_map = inode_resv_map(inode);
Joonsoo Kim9119a412014-04-03 14:47:25 -07003479 long chg = 0;
David Gibson90481622012-03-21 16:34:12 -07003480 struct hugepage_subpool *spool = subpool_inode(inode);
Ken Chen45c682a2007-11-14 16:59:44 -08003481
Joonsoo Kim9119a412014-04-03 14:47:25 -07003482 if (resv_map)
Joonsoo Kim1406ec92014-04-03 14:47:26 -07003483 chg = region_truncate(resv_map, offset);
Ken Chen45c682a2007-11-14 16:59:44 -08003484 spin_lock(&inode->i_lock);
Eric Sandeene4c6f8b2009-07-29 15:02:16 -07003485 inode->i_blocks -= (blocks_per_huge_page(h) * freed);
Ken Chen45c682a2007-11-14 16:59:44 -08003486 spin_unlock(&inode->i_lock);
3487
David Gibson90481622012-03-21 16:34:12 -07003488 hugepage_subpool_put_pages(spool, (chg - freed));
Andi Kleena5516432008-07-23 21:27:41 -07003489 hugetlb_acct_memory(h, -(chg - freed));
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -07003490}
Naoya Horiguchi93f70f92010-05-28 09:29:20 +09003491
Steve Capper3212b532013-04-23 12:35:02 +01003492#ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
3493static unsigned long page_table_shareable(struct vm_area_struct *svma,
3494 struct vm_area_struct *vma,
3495 unsigned long addr, pgoff_t idx)
3496{
3497 unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) +
3498 svma->vm_start;
3499 unsigned long sbase = saddr & PUD_MASK;
3500 unsigned long s_end = sbase + PUD_SIZE;
3501
3502 /* Allow segments to share if only one is marked locked */
3503 unsigned long vm_flags = vma->vm_flags & ~VM_LOCKED;
3504 unsigned long svm_flags = svma->vm_flags & ~VM_LOCKED;
3505
3506 /*
3507 * match the virtual addresses, permission and the alignment of the
3508 * page table page.
3509 */
3510 if (pmd_index(addr) != pmd_index(saddr) ||
3511 vm_flags != svm_flags ||
3512 sbase < svma->vm_start || svma->vm_end < s_end)
3513 return 0;
3514
3515 return saddr;
3516}
3517
3518static int vma_shareable(struct vm_area_struct *vma, unsigned long addr)
3519{
3520 unsigned long base = addr & PUD_MASK;
3521 unsigned long end = base + PUD_SIZE;
3522
3523 /*
3524 * check on proper vm_flags and page table alignment
3525 */
3526 if (vma->vm_flags & VM_MAYSHARE &&
3527 vma->vm_start <= base && end <= vma->vm_end)
3528 return 1;
3529 return 0;
3530}
3531
3532/*
3533 * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
3534 * and returns the corresponding pte. While this is not necessary for the
3535 * !shared pmd case because we can allocate the pmd later as well, it makes the
3536 * code much cleaner. pmd allocation is essential for the shared case because
3537 * pud has to be populated inside the same i_mmap_mutex section - otherwise
3538 * racing tasks could either miss the sharing (see huge_pte_offset) or select a
3539 * bad pmd for sharing.
3540 */
3541pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
3542{
3543 struct vm_area_struct *vma = find_vma(mm, addr);
3544 struct address_space *mapping = vma->vm_file->f_mapping;
3545 pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) +
3546 vma->vm_pgoff;
3547 struct vm_area_struct *svma;
3548 unsigned long saddr;
3549 pte_t *spte = NULL;
3550 pte_t *pte;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003551 spinlock_t *ptl;
Steve Capper3212b532013-04-23 12:35:02 +01003552
3553 if (!vma_shareable(vma, addr))
3554 return (pte_t *)pmd_alloc(mm, pud, addr);
3555
3556 mutex_lock(&mapping->i_mmap_mutex);
3557 vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) {
3558 if (svma == vma)
3559 continue;
3560
3561 saddr = page_table_shareable(svma, vma, addr, idx);
3562 if (saddr) {
3563 spte = huge_pte_offset(svma->vm_mm, saddr);
3564 if (spte) {
3565 get_page(virt_to_page(spte));
3566 break;
3567 }
3568 }
3569 }
3570
3571 if (!spte)
3572 goto out;
3573
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003574 ptl = huge_pte_lockptr(hstate_vma(vma), mm, spte);
3575 spin_lock(ptl);
Steve Capper3212b532013-04-23 12:35:02 +01003576 if (pud_none(*pud))
3577 pud_populate(mm, pud,
3578 (pmd_t *)((unsigned long)spte & PAGE_MASK));
3579 else
3580 put_page(virt_to_page(spte));
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003581 spin_unlock(ptl);
Steve Capper3212b532013-04-23 12:35:02 +01003582out:
3583 pte = (pte_t *)pmd_alloc(mm, pud, addr);
3584 mutex_unlock(&mapping->i_mmap_mutex);
3585 return pte;
3586}
3587
3588/*
3589 * unmap huge page backed by shared pte.
3590 *
3591 * Hugetlb pte page is ref counted at the time of mapping. If pte is shared
3592 * indicated by page_count > 1, unmap is achieved by clearing pud and
3593 * decrementing the ref count. If count == 1, the pte page is not shared.
3594 *
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003595 * called with page table lock held.
Steve Capper3212b532013-04-23 12:35:02 +01003596 *
3597 * returns: 1 successfully unmapped a shared pte page
3598 * 0 the underlying pte page is not shared, or it is the last user
3599 */
3600int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
3601{
3602 pgd_t *pgd = pgd_offset(mm, *addr);
3603 pud_t *pud = pud_offset(pgd, *addr);
3604
3605 BUG_ON(page_count(virt_to_page(ptep)) == 0);
3606 if (page_count(virt_to_page(ptep)) == 1)
3607 return 0;
3608
3609 pud_clear(pud);
3610 put_page(virt_to_page(ptep));
3611 *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE;
3612 return 1;
3613}
Steve Capper9e5fc742013-04-30 08:02:03 +01003614#define want_pmd_share() (1)
3615#else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
3616pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
3617{
3618 return NULL;
3619}
3620#define want_pmd_share() (0)
Steve Capper3212b532013-04-23 12:35:02 +01003621#endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
3622
Steve Capper9e5fc742013-04-30 08:02:03 +01003623#ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
3624pte_t *huge_pte_alloc(struct mm_struct *mm,
3625 unsigned long addr, unsigned long sz)
3626{
3627 pgd_t *pgd;
3628 pud_t *pud;
3629 pte_t *pte = NULL;
3630
3631 pgd = pgd_offset(mm, addr);
3632 pud = pud_alloc(mm, pgd, addr);
3633 if (pud) {
3634 if (sz == PUD_SIZE) {
3635 pte = (pte_t *)pud;
3636 } else {
3637 BUG_ON(sz != PMD_SIZE);
3638 if (want_pmd_share() && pud_none(*pud))
3639 pte = huge_pmd_share(mm, addr, pud);
3640 else
3641 pte = (pte_t *)pmd_alloc(mm, pud, addr);
3642 }
3643 }
3644 BUG_ON(pte && !pte_none(*pte) && !pte_huge(*pte));
3645
3646 return pte;
3647}
3648
3649pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr)
3650{
3651 pgd_t *pgd;
3652 pud_t *pud;
3653 pmd_t *pmd = NULL;
3654
3655 pgd = pgd_offset(mm, addr);
3656 if (pgd_present(*pgd)) {
3657 pud = pud_offset(pgd, addr);
3658 if (pud_present(*pud)) {
3659 if (pud_huge(*pud))
3660 return (pte_t *)pud;
3661 pmd = pmd_offset(pud, addr);
3662 }
3663 }
3664 return (pte_t *) pmd;
3665}
3666
3667struct page *
3668follow_huge_pmd(struct mm_struct *mm, unsigned long address,
3669 pmd_t *pmd, int write)
3670{
3671 struct page *page;
3672
3673 page = pte_page(*(pte_t *)pmd);
3674 if (page)
3675 page += ((address & ~PMD_MASK) >> PAGE_SHIFT);
3676 return page;
3677}
3678
3679struct page *
3680follow_huge_pud(struct mm_struct *mm, unsigned long address,
3681 pud_t *pud, int write)
3682{
3683 struct page *page;
3684
3685 page = pte_page(*(pte_t *)pud);
3686 if (page)
3687 page += ((address & ~PUD_MASK) >> PAGE_SHIFT);
3688 return page;
3689}
3690
3691#else /* !CONFIG_ARCH_WANT_GENERAL_HUGETLB */
3692
3693/* Can be overriden by architectures */
Gideon Israel Dsouza3b321232014-04-07 15:37:26 -07003694struct page * __weak
Steve Capper9e5fc742013-04-30 08:02:03 +01003695follow_huge_pud(struct mm_struct *mm, unsigned long address,
3696 pud_t *pud, int write)
3697{
3698 BUG();
3699 return NULL;
3700}
3701
3702#endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
3703
Andi Kleend5bd9102010-09-27 09:00:12 +02003704#ifdef CONFIG_MEMORY_FAILURE
3705
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +09003706/* Should be called in hugetlb_lock */
3707static int is_hugepage_on_freelist(struct page *hpage)
3708{
3709 struct page *page;
3710 struct page *tmp;
3711 struct hstate *h = page_hstate(hpage);
3712 int nid = page_to_nid(hpage);
3713
3714 list_for_each_entry_safe(page, tmp, &h->hugepage_freelists[nid], lru)
3715 if (page == hpage)
3716 return 1;
3717 return 0;
3718}
3719
Naoya Horiguchi93f70f92010-05-28 09:29:20 +09003720/*
3721 * This function is called from memory failure code.
3722 * Assume the caller holds page lock of the head page.
3723 */
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +09003724int dequeue_hwpoisoned_huge_page(struct page *hpage)
Naoya Horiguchi93f70f92010-05-28 09:29:20 +09003725{
3726 struct hstate *h = page_hstate(hpage);
3727 int nid = page_to_nid(hpage);
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +09003728 int ret = -EBUSY;
Naoya Horiguchi93f70f92010-05-28 09:29:20 +09003729
3730 spin_lock(&hugetlb_lock);
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +09003731 if (is_hugepage_on_freelist(hpage)) {
Naoya Horiguchi56f2fb12012-12-12 13:52:33 -08003732 /*
3733 * Hwpoisoned hugepage isn't linked to activelist or freelist,
3734 * but dangling hpage->lru can trigger list-debug warnings
3735 * (this happens when we call unpoison_memory() on it),
3736 * so let it point to itself with list_del_init().
3737 */
3738 list_del_init(&hpage->lru);
Naoya Horiguchi8c6c2ec2010-09-08 10:19:38 +09003739 set_page_refcounted(hpage);
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +09003740 h->free_huge_pages--;
3741 h->free_huge_pages_node[nid]--;
3742 ret = 0;
3743 }
Naoya Horiguchi93f70f92010-05-28 09:29:20 +09003744 spin_unlock(&hugetlb_lock);
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +09003745 return ret;
Naoya Horiguchi93f70f92010-05-28 09:29:20 +09003746}
Naoya Horiguchi6de2b1a2010-09-08 10:19:36 +09003747#endif
Naoya Horiguchi31caf662013-09-11 14:21:59 -07003748
3749bool isolate_huge_page(struct page *page, struct list_head *list)
3750{
Sasha Levin309381fea2014-01-23 15:52:54 -08003751 VM_BUG_ON_PAGE(!PageHead(page), page);
Naoya Horiguchi31caf662013-09-11 14:21:59 -07003752 if (!get_page_unless_zero(page))
3753 return false;
3754 spin_lock(&hugetlb_lock);
3755 list_move_tail(&page->lru, list);
3756 spin_unlock(&hugetlb_lock);
3757 return true;
3758}
3759
3760void putback_active_hugepage(struct page *page)
3761{
Sasha Levin309381fea2014-01-23 15:52:54 -08003762 VM_BUG_ON_PAGE(!PageHead(page), page);
Naoya Horiguchi31caf662013-09-11 14:21:59 -07003763 spin_lock(&hugetlb_lock);
3764 list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist);
3765 spin_unlock(&hugetlb_lock);
3766 put_page(page);
3767}
Naoya Horiguchic8721bb2013-09-11 14:22:09 -07003768
3769bool is_hugepage_active(struct page *page)
3770{
Sasha Levin309381fea2014-01-23 15:52:54 -08003771 VM_BUG_ON_PAGE(!PageHuge(page), page);
Naoya Horiguchic8721bb2013-09-11 14:22:09 -07003772 /*
3773 * This function can be called for a tail page because the caller,
3774 * scan_movable_pages, scans through a given pfn-range which typically
3775 * covers one memory block. In systems using gigantic hugepage (1GB
3776 * for x86_64,) a hugepage is larger than a memory block, and we don't
3777 * support migrating such large hugepages for now, so return false
3778 * when called for tail pages.
3779 */
3780 if (PageTail(page))
3781 return false;
3782 /*
3783 * Refcount of a hwpoisoned hugepages is 1, but they are not active,
3784 * so we should return false for them.
3785 */
3786 if (unlikely(PageHWPoison(page)))
3787 return false;
3788 return page_count(page) > 0;
3789}