Commit | Line | Data |
---|---|---|
1da177e4 LT |
1 | /* |
2 | * Generic hugetlb support. | |
6d49e352 | 3 | * (C) Nadia Yvette Chambers, April 2004 |
1da177e4 | 4 | */ |
1da177e4 LT |
5 | #include <linux/list.h> |
6 | #include <linux/init.h> | |
7 | #include <linux/module.h> | |
8 | #include <linux/mm.h> | |
e1759c21 | 9 | #include <linux/seq_file.h> |
1da177e4 LT |
10 | #include <linux/sysctl.h> |
11 | #include <linux/highmem.h> | |
cddb8a5c | 12 | #include <linux/mmu_notifier.h> |
1da177e4 | 13 | #include <linux/nodemask.h> |
63551ae0 | 14 | #include <linux/pagemap.h> |
5da7ca86 | 15 | #include <linux/mempolicy.h> |
3b32123d | 16 | #include <linux/compiler.h> |
aea47ff3 | 17 | #include <linux/cpuset.h> |
3935baa9 | 18 | #include <linux/mutex.h> |
aa888a74 | 19 | #include <linux/bootmem.h> |
a3437870 | 20 | #include <linux/sysfs.h> |
5a0e3ad6 | 21 | #include <linux/slab.h> |
0fe6e20b | 22 | #include <linux/rmap.h> |
fd6a03ed NH |
23 | #include <linux/swap.h> |
24 | #include <linux/swapops.h> | |
c8721bbb | 25 | #include <linux/page-isolation.h> |
8382d914 | 26 | #include <linux/jhash.h> |
d6606683 | 27 | |
63551ae0 DG |
28 | #include <asm/page.h> |
29 | #include <asm/pgtable.h> | |
24669e58 | 30 | #include <asm/tlb.h> |
63551ae0 | 31 | |
24669e58 | 32 | #include <linux/io.h> |
63551ae0 | 33 | #include <linux/hugetlb.h> |
9dd540e2 | 34 | #include <linux/hugetlb_cgroup.h> |
9a305230 | 35 | #include <linux/node.h> |
7835e98b | 36 | #include "internal.h" |
1da177e4 | 37 | |
753162cd | 38 | int hugepages_treat_as_movable; |
a5516438 | 39 | |
c3f38a38 | 40 | int hugetlb_max_hstate __read_mostly; |
e5ff2159 AK |
41 | unsigned int default_hstate_idx; |
42 | struct hstate hstates[HUGE_MAX_HSTATE]; | |
641844f5 NH |
43 | /* |
44 | * Minimum page order among possible hugepage sizes, set to a proper value | |
45 | * at boot time. | |
46 | */ | |
47 | static unsigned int minimum_order __read_mostly = UINT_MAX; | |
e5ff2159 | 48 | |
53ba51d2 JT |
49 | __initdata LIST_HEAD(huge_boot_pages); |
50 | ||
e5ff2159 AK |
51 | /* for command line parsing */ |
52 | static struct hstate * __initdata parsed_hstate; | |
53 | static unsigned long __initdata default_hstate_max_huge_pages; | |
e11bfbfc | 54 | static unsigned long __initdata default_hstate_size; |
e5ff2159 | 55 | |
3935baa9 | 56 | /* |
31caf665 NH |
57 | * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages, |
58 | * free_huge_pages, and surplus_huge_pages. | |
3935baa9 | 59 | */ |
c3f38a38 | 60 | DEFINE_SPINLOCK(hugetlb_lock); |
0bd0f9fb | 61 | |
8382d914 DB |
62 | /* |
63 | * Serializes faults on the same logical page. This is used to | |
64 | * prevent spurious OOMs when the hugepage pool is fully utilized. | |
65 | */ | |
66 | static int num_fault_mutexes; | |
67 | static struct mutex *htlb_fault_mutex_table ____cacheline_aligned_in_smp; | |
68 | ||
7ca02d0a MK |
69 | /* Forward declaration */ |
70 | static int hugetlb_acct_memory(struct hstate *h, long delta); | |
71 | ||
90481622 DG |
72 | static inline void unlock_or_release_subpool(struct hugepage_subpool *spool) |
73 | { | |
74 | bool free = (spool->count == 0) && (spool->used_hpages == 0); | |
75 | ||
76 | spin_unlock(&spool->lock); | |
77 | ||
78 | /* If no pages are used, and no other handles to the subpool | |
7ca02d0a MK |
79 | * remain, give up any reservations mased on minimum size and |
80 | * free the subpool */ | |
81 | if (free) { | |
82 | if (spool->min_hpages != -1) | |
83 | hugetlb_acct_memory(spool->hstate, | |
84 | -spool->min_hpages); | |
90481622 | 85 | kfree(spool); |
7ca02d0a | 86 | } |
90481622 DG |
87 | } |
88 | ||
7ca02d0a MK |
89 | struct hugepage_subpool *hugepage_new_subpool(struct hstate *h, long max_hpages, |
90 | long min_hpages) | |
90481622 DG |
91 | { |
92 | struct hugepage_subpool *spool; | |
93 | ||
c6a91820 | 94 | spool = kzalloc(sizeof(*spool), GFP_KERNEL); |
90481622 DG |
95 | if (!spool) |
96 | return NULL; | |
97 | ||
98 | spin_lock_init(&spool->lock); | |
99 | spool->count = 1; | |
7ca02d0a MK |
100 | spool->max_hpages = max_hpages; |
101 | spool->hstate = h; | |
102 | spool->min_hpages = min_hpages; | |
103 | ||
104 | if (min_hpages != -1 && hugetlb_acct_memory(h, min_hpages)) { | |
105 | kfree(spool); | |
106 | return NULL; | |
107 | } | |
108 | spool->rsv_hpages = min_hpages; | |
90481622 DG |
109 | |
110 | return spool; | |
111 | } | |
112 | ||
113 | void hugepage_put_subpool(struct hugepage_subpool *spool) | |
114 | { | |
115 | spin_lock(&spool->lock); | |
116 | BUG_ON(!spool->count); | |
117 | spool->count--; | |
118 | unlock_or_release_subpool(spool); | |
119 | } | |
120 | ||
1c5ecae3 MK |
121 | /* |
122 | * Subpool accounting for allocating and reserving pages. | |
123 | * Return -ENOMEM if there are not enough resources to satisfy the | |
124 | * the request. Otherwise, return the number of pages by which the | |
125 | * global pools must be adjusted (upward). The returned value may | |
126 | * only be different than the passed value (delta) in the case where | |
127 | * a subpool minimum size must be manitained. | |
128 | */ | |
129 | static long hugepage_subpool_get_pages(struct hugepage_subpool *spool, | |
90481622 DG |
130 | long delta) |
131 | { | |
1c5ecae3 | 132 | long ret = delta; |
90481622 DG |
133 | |
134 | if (!spool) | |
1c5ecae3 | 135 | return ret; |
90481622 DG |
136 | |
137 | spin_lock(&spool->lock); | |
1c5ecae3 MK |
138 | |
139 | if (spool->max_hpages != -1) { /* maximum size accounting */ | |
140 | if ((spool->used_hpages + delta) <= spool->max_hpages) | |
141 | spool->used_hpages += delta; | |
142 | else { | |
143 | ret = -ENOMEM; | |
144 | goto unlock_ret; | |
145 | } | |
90481622 | 146 | } |
90481622 | 147 | |
1c5ecae3 MK |
148 | if (spool->min_hpages != -1) { /* minimum size accounting */ |
149 | if (delta > spool->rsv_hpages) { | |
150 | /* | |
151 | * Asking for more reserves than those already taken on | |
152 | * behalf of subpool. Return difference. | |
153 | */ | |
154 | ret = delta - spool->rsv_hpages; | |
155 | spool->rsv_hpages = 0; | |
156 | } else { | |
157 | ret = 0; /* reserves already accounted for */ | |
158 | spool->rsv_hpages -= delta; | |
159 | } | |
160 | } | |
161 | ||
162 | unlock_ret: | |
163 | spin_unlock(&spool->lock); | |
90481622 DG |
164 | return ret; |
165 | } | |
166 | ||
1c5ecae3 MK |
167 | /* |
168 | * Subpool accounting for freeing and unreserving pages. | |
169 | * Return the number of global page reservations that must be dropped. | |
170 | * The return value may only be different than the passed value (delta) | |
171 | * in the case where a subpool minimum size must be maintained. | |
172 | */ | |
173 | static long hugepage_subpool_put_pages(struct hugepage_subpool *spool, | |
90481622 DG |
174 | long delta) |
175 | { | |
1c5ecae3 MK |
176 | long ret = delta; |
177 | ||
90481622 | 178 | if (!spool) |
1c5ecae3 | 179 | return delta; |
90481622 DG |
180 | |
181 | spin_lock(&spool->lock); | |
1c5ecae3 MK |
182 | |
183 | if (spool->max_hpages != -1) /* maximum size accounting */ | |
184 | spool->used_hpages -= delta; | |
185 | ||
186 | if (spool->min_hpages != -1) { /* minimum size accounting */ | |
187 | if (spool->rsv_hpages + delta <= spool->min_hpages) | |
188 | ret = 0; | |
189 | else | |
190 | ret = spool->rsv_hpages + delta - spool->min_hpages; | |
191 | ||
192 | spool->rsv_hpages += delta; | |
193 | if (spool->rsv_hpages > spool->min_hpages) | |
194 | spool->rsv_hpages = spool->min_hpages; | |
195 | } | |
196 | ||
197 | /* | |
198 | * If hugetlbfs_put_super couldn't free spool due to an outstanding | |
199 | * quota reference, free it now. | |
200 | */ | |
90481622 | 201 | unlock_or_release_subpool(spool); |
1c5ecae3 MK |
202 | |
203 | return ret; | |
90481622 DG |
204 | } |
205 | ||
206 | static inline struct hugepage_subpool *subpool_inode(struct inode *inode) | |
207 | { | |
208 | return HUGETLBFS_SB(inode->i_sb)->spool; | |
209 | } | |
210 | ||
211 | static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma) | |
212 | { | |
496ad9aa | 213 | return subpool_inode(file_inode(vma->vm_file)); |
90481622 DG |
214 | } |
215 | ||
96822904 AW |
216 | /* |
217 | * Region tracking -- allows tracking of reservations and instantiated pages | |
218 | * across the pages in a mapping. | |
84afd99b | 219 | * |
1dd308a7 MK |
220 | * The region data structures are embedded into a resv_map and protected |
221 | * by a resv_map's lock. The set of regions within the resv_map represent | |
222 | * reservations for huge pages, or huge pages that have already been | |
223 | * instantiated within the map. The from and to elements are huge page | |
224 | * indicies into the associated mapping. from indicates the starting index | |
225 | * of the region. to represents the first index past the end of the region. | |
226 | * | |
227 | * For example, a file region structure with from == 0 and to == 4 represents | |
228 | * four huge pages in a mapping. It is important to note that the to element | |
229 | * represents the first element past the end of the region. This is used in | |
230 | * arithmetic as 4(to) - 0(from) = 4 huge pages in the region. | |
231 | * | |
232 | * Interval notation of the form [from, to) will be used to indicate that | |
233 | * the endpoint from is inclusive and to is exclusive. | |
96822904 AW |
234 | */ |
235 | struct file_region { | |
236 | struct list_head link; | |
237 | long from; | |
238 | long to; | |
239 | }; | |
240 | ||
1dd308a7 MK |
241 | /* |
242 | * Add the huge page range represented by [f, t) to the reserve | |
243 | * map. Existing regions will be expanded to accommodate the | |
244 | * specified range. We know only existing regions need to be | |
245 | * expanded, because region_add is only called after region_chg | |
246 | * with the same range. If a new file_region structure must | |
247 | * be allocated, it is done in region_chg. | |
cf3ad20b MK |
248 | * |
249 | * Return the number of new huge pages added to the map. This | |
250 | * number is greater than or equal to zero. | |
1dd308a7 | 251 | */ |
1406ec9b | 252 | static long region_add(struct resv_map *resv, long f, long t) |
96822904 | 253 | { |
1406ec9b | 254 | struct list_head *head = &resv->regions; |
96822904 | 255 | struct file_region *rg, *nrg, *trg; |
cf3ad20b | 256 | long add = 0; |
96822904 | 257 | |
7b24d861 | 258 | spin_lock(&resv->lock); |
96822904 AW |
259 | /* Locate the region we are either in or before. */ |
260 | list_for_each_entry(rg, head, link) | |
261 | if (f <= rg->to) | |
262 | break; | |
263 | ||
264 | /* Round our left edge to the current segment if it encloses us. */ | |
265 | if (f > rg->from) | |
266 | f = rg->from; | |
267 | ||
268 | /* Check for and consume any regions we now overlap with. */ | |
269 | nrg = rg; | |
270 | list_for_each_entry_safe(rg, trg, rg->link.prev, link) { | |
271 | if (&rg->link == head) | |
272 | break; | |
273 | if (rg->from > t) | |
274 | break; | |
275 | ||
276 | /* If this area reaches higher then extend our area to | |
277 | * include it completely. If this is not the first area | |
278 | * which we intend to reuse, free it. */ | |
279 | if (rg->to > t) | |
280 | t = rg->to; | |
281 | if (rg != nrg) { | |
cf3ad20b MK |
282 | /* Decrement return value by the deleted range. |
283 | * Another range will span this area so that by | |
284 | * end of routine add will be >= zero | |
285 | */ | |
286 | add -= (rg->to - rg->from); | |
96822904 AW |
287 | list_del(&rg->link); |
288 | kfree(rg); | |
289 | } | |
290 | } | |
cf3ad20b MK |
291 | |
292 | add += (nrg->from - f); /* Added to beginning of region */ | |
96822904 | 293 | nrg->from = f; |
cf3ad20b | 294 | add += t - nrg->to; /* Added to end of region */ |
96822904 | 295 | nrg->to = t; |
cf3ad20b | 296 | |
7b24d861 | 297 | spin_unlock(&resv->lock); |
cf3ad20b MK |
298 | VM_BUG_ON(add < 0); |
299 | return add; | |
96822904 AW |
300 | } |
301 | ||
1dd308a7 MK |
302 | /* |
303 | * Examine the existing reserve map and determine how many | |
304 | * huge pages in the specified range [f, t) are NOT currently | |
305 | * represented. This routine is called before a subsequent | |
306 | * call to region_add that will actually modify the reserve | |
307 | * map to add the specified range [f, t). region_chg does | |
308 | * not change the number of huge pages represented by the | |
309 | * map. However, if the existing regions in the map can not | |
310 | * be expanded to represent the new range, a new file_region | |
311 | * structure is added to the map as a placeholder. This is | |
312 | * so that the subsequent region_add call will have all the | |
313 | * regions it needs and will not fail. | |
314 | * | |
315 | * Returns the number of huge pages that need to be added | |
316 | * to the existing reservation map for the range [f, t). | |
317 | * This number is greater or equal to zero. -ENOMEM is | |
318 | * returned if a new file_region structure is needed and can | |
319 | * not be allocated. | |
320 | */ | |
1406ec9b | 321 | static long region_chg(struct resv_map *resv, long f, long t) |
96822904 | 322 | { |
1406ec9b | 323 | struct list_head *head = &resv->regions; |
7b24d861 | 324 | struct file_region *rg, *nrg = NULL; |
96822904 AW |
325 | long chg = 0; |
326 | ||
7b24d861 DB |
327 | retry: |
328 | spin_lock(&resv->lock); | |
96822904 AW |
329 | /* Locate the region we are before or in. */ |
330 | list_for_each_entry(rg, head, link) | |
331 | if (f <= rg->to) | |
332 | break; | |
333 | ||
334 | /* If we are below the current region then a new region is required. | |
335 | * Subtle, allocate a new region at the position but make it zero | |
336 | * size such that we can guarantee to record the reservation. */ | |
337 | if (&rg->link == head || t < rg->from) { | |
7b24d861 DB |
338 | if (!nrg) { |
339 | spin_unlock(&resv->lock); | |
340 | nrg = kmalloc(sizeof(*nrg), GFP_KERNEL); | |
341 | if (!nrg) | |
342 | return -ENOMEM; | |
343 | ||
344 | nrg->from = f; | |
345 | nrg->to = f; | |
346 | INIT_LIST_HEAD(&nrg->link); | |
347 | goto retry; | |
348 | } | |
96822904 | 349 | |
7b24d861 DB |
350 | list_add(&nrg->link, rg->link.prev); |
351 | chg = t - f; | |
352 | goto out_nrg; | |
96822904 AW |
353 | } |
354 | ||
355 | /* Round our left edge to the current segment if it encloses us. */ | |
356 | if (f > rg->from) | |
357 | f = rg->from; | |
358 | chg = t - f; | |
359 | ||
360 | /* Check for and consume any regions we now overlap with. */ | |
361 | list_for_each_entry(rg, rg->link.prev, link) { | |
362 | if (&rg->link == head) | |
363 | break; | |
364 | if (rg->from > t) | |
7b24d861 | 365 | goto out; |
96822904 | 366 | |
25985edc | 367 | /* We overlap with this area, if it extends further than |
96822904 AW |
368 | * us then we must extend ourselves. Account for its |
369 | * existing reservation. */ | |
370 | if (rg->to > t) { | |
371 | chg += rg->to - t; | |
372 | t = rg->to; | |
373 | } | |
374 | chg -= rg->to - rg->from; | |
375 | } | |
7b24d861 DB |
376 | |
377 | out: | |
378 | spin_unlock(&resv->lock); | |
379 | /* We already know we raced and no longer need the new region */ | |
380 | kfree(nrg); | |
381 | return chg; | |
382 | out_nrg: | |
383 | spin_unlock(&resv->lock); | |
96822904 AW |
384 | return chg; |
385 | } | |
386 | ||
1dd308a7 MK |
387 | /* |
388 | * Truncate the reserve map at index 'end'. Modify/truncate any | |
389 | * region which contains end. Delete any regions past end. | |
390 | * Return the number of huge pages removed from the map. | |
391 | */ | |
1406ec9b | 392 | static long region_truncate(struct resv_map *resv, long end) |
96822904 | 393 | { |
1406ec9b | 394 | struct list_head *head = &resv->regions; |
96822904 AW |
395 | struct file_region *rg, *trg; |
396 | long chg = 0; | |
397 | ||
7b24d861 | 398 | spin_lock(&resv->lock); |
96822904 AW |
399 | /* Locate the region we are either in or before. */ |
400 | list_for_each_entry(rg, head, link) | |
401 | if (end <= rg->to) | |
402 | break; | |
403 | if (&rg->link == head) | |
7b24d861 | 404 | goto out; |
96822904 AW |
405 | |
406 | /* If we are in the middle of a region then adjust it. */ | |
407 | if (end > rg->from) { | |
408 | chg = rg->to - end; | |
409 | rg->to = end; | |
410 | rg = list_entry(rg->link.next, typeof(*rg), link); | |
411 | } | |
412 | ||
413 | /* Drop any remaining regions. */ | |
414 | list_for_each_entry_safe(rg, trg, rg->link.prev, link) { | |
415 | if (&rg->link == head) | |
416 | break; | |
417 | chg += rg->to - rg->from; | |
418 | list_del(&rg->link); | |
419 | kfree(rg); | |
420 | } | |
7b24d861 DB |
421 | |
422 | out: | |
423 | spin_unlock(&resv->lock); | |
96822904 AW |
424 | return chg; |
425 | } | |
426 | ||
1dd308a7 MK |
427 | /* |
428 | * Count and return the number of huge pages in the reserve map | |
429 | * that intersect with the range [f, t). | |
430 | */ | |
1406ec9b | 431 | static long region_count(struct resv_map *resv, long f, long t) |
84afd99b | 432 | { |
1406ec9b | 433 | struct list_head *head = &resv->regions; |
84afd99b AW |
434 | struct file_region *rg; |
435 | long chg = 0; | |
436 | ||
7b24d861 | 437 | spin_lock(&resv->lock); |
84afd99b AW |
438 | /* Locate each segment we overlap with, and count that overlap. */ |
439 | list_for_each_entry(rg, head, link) { | |
f2135a4a WSH |
440 | long seg_from; |
441 | long seg_to; | |
84afd99b AW |
442 | |
443 | if (rg->to <= f) | |
444 | continue; | |
445 | if (rg->from >= t) | |
446 | break; | |
447 | ||
448 | seg_from = max(rg->from, f); | |
449 | seg_to = min(rg->to, t); | |
450 | ||
451 | chg += seg_to - seg_from; | |
452 | } | |
7b24d861 | 453 | spin_unlock(&resv->lock); |
84afd99b AW |
454 | |
455 | return chg; | |
456 | } | |
457 | ||
e7c4b0bf AW |
458 | /* |
459 | * Convert the address within this vma to the page offset within | |
460 | * the mapping, in pagecache page units; huge pages here. | |
461 | */ | |
a5516438 AK |
462 | static pgoff_t vma_hugecache_offset(struct hstate *h, |
463 | struct vm_area_struct *vma, unsigned long address) | |
e7c4b0bf | 464 | { |
a5516438 AK |
465 | return ((address - vma->vm_start) >> huge_page_shift(h)) + |
466 | (vma->vm_pgoff >> huge_page_order(h)); | |
e7c4b0bf AW |
467 | } |
468 | ||
0fe6e20b NH |
469 | pgoff_t linear_hugepage_index(struct vm_area_struct *vma, |
470 | unsigned long address) | |
471 | { | |
472 | return vma_hugecache_offset(hstate_vma(vma), vma, address); | |
473 | } | |
474 | ||
08fba699 MG |
475 | /* |
476 | * Return the size of the pages allocated when backing a VMA. In the majority | |
477 | * cases this will be same size as used by the page table entries. | |
478 | */ | |
479 | unsigned long vma_kernel_pagesize(struct vm_area_struct *vma) | |
480 | { | |
481 | struct hstate *hstate; | |
482 | ||
483 | if (!is_vm_hugetlb_page(vma)) | |
484 | return PAGE_SIZE; | |
485 | ||
486 | hstate = hstate_vma(vma); | |
487 | ||
2415cf12 | 488 | return 1UL << huge_page_shift(hstate); |
08fba699 | 489 | } |
f340ca0f | 490 | EXPORT_SYMBOL_GPL(vma_kernel_pagesize); |
08fba699 | 491 | |
3340289d MG |
492 | /* |
493 | * Return the page size being used by the MMU to back a VMA. In the majority | |
494 | * of cases, the page size used by the kernel matches the MMU size. On | |
495 | * architectures where it differs, an architecture-specific version of this | |
496 | * function is required. | |
497 | */ | |
498 | #ifndef vma_mmu_pagesize | |
499 | unsigned long vma_mmu_pagesize(struct vm_area_struct *vma) | |
500 | { | |
501 | return vma_kernel_pagesize(vma); | |
502 | } | |
503 | #endif | |
504 | ||
84afd99b AW |
505 | /* |
506 | * Flags for MAP_PRIVATE reservations. These are stored in the bottom | |
507 | * bits of the reservation map pointer, which are always clear due to | |
508 | * alignment. | |
509 | */ | |
510 | #define HPAGE_RESV_OWNER (1UL << 0) | |
511 | #define HPAGE_RESV_UNMAPPED (1UL << 1) | |
04f2cbe3 | 512 | #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED) |
84afd99b | 513 | |
a1e78772 MG |
514 | /* |
515 | * These helpers are used to track how many pages are reserved for | |
516 | * faults in a MAP_PRIVATE mapping. Only the process that called mmap() | |
517 | * is guaranteed to have their future faults succeed. | |
518 | * | |
519 | * With the exception of reset_vma_resv_huge_pages() which is called at fork(), | |
520 | * the reserve counters are updated with the hugetlb_lock held. It is safe | |
521 | * to reset the VMA at fork() time as it is not in use yet and there is no | |
522 | * chance of the global counters getting corrupted as a result of the values. | |
84afd99b AW |
523 | * |
524 | * The private mapping reservation is represented in a subtly different | |
525 | * manner to a shared mapping. A shared mapping has a region map associated | |
526 | * with the underlying file, this region map represents the backing file | |
527 | * pages which have ever had a reservation assigned which this persists even | |
528 | * after the page is instantiated. A private mapping has a region map | |
529 | * associated with the original mmap which is attached to all VMAs which | |
530 | * reference it, this region map represents those offsets which have consumed | |
531 | * reservation ie. where pages have been instantiated. | |
a1e78772 | 532 | */ |
e7c4b0bf AW |
533 | static unsigned long get_vma_private_data(struct vm_area_struct *vma) |
534 | { | |
535 | return (unsigned long)vma->vm_private_data; | |
536 | } | |
537 | ||
538 | static void set_vma_private_data(struct vm_area_struct *vma, | |
539 | unsigned long value) | |
540 | { | |
541 | vma->vm_private_data = (void *)value; | |
542 | } | |
543 | ||
9119a41e | 544 | struct resv_map *resv_map_alloc(void) |
84afd99b AW |
545 | { |
546 | struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL); | |
547 | if (!resv_map) | |
548 | return NULL; | |
549 | ||
550 | kref_init(&resv_map->refs); | |
7b24d861 | 551 | spin_lock_init(&resv_map->lock); |
84afd99b AW |
552 | INIT_LIST_HEAD(&resv_map->regions); |
553 | ||
554 | return resv_map; | |
555 | } | |
556 | ||
9119a41e | 557 | void resv_map_release(struct kref *ref) |
84afd99b AW |
558 | { |
559 | struct resv_map *resv_map = container_of(ref, struct resv_map, refs); | |
560 | ||
561 | /* Clear out any active regions before we release the map. */ | |
1406ec9b | 562 | region_truncate(resv_map, 0); |
84afd99b AW |
563 | kfree(resv_map); |
564 | } | |
565 | ||
4e35f483 JK |
566 | static inline struct resv_map *inode_resv_map(struct inode *inode) |
567 | { | |
568 | return inode->i_mapping->private_data; | |
569 | } | |
570 | ||
84afd99b | 571 | static struct resv_map *vma_resv_map(struct vm_area_struct *vma) |
a1e78772 | 572 | { |
81d1b09c | 573 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); |
4e35f483 JK |
574 | if (vma->vm_flags & VM_MAYSHARE) { |
575 | struct address_space *mapping = vma->vm_file->f_mapping; | |
576 | struct inode *inode = mapping->host; | |
577 | ||
578 | return inode_resv_map(inode); | |
579 | ||
580 | } else { | |
84afd99b AW |
581 | return (struct resv_map *)(get_vma_private_data(vma) & |
582 | ~HPAGE_RESV_MASK); | |
4e35f483 | 583 | } |
a1e78772 MG |
584 | } |
585 | ||
84afd99b | 586 | static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map) |
a1e78772 | 587 | { |
81d1b09c SL |
588 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); |
589 | VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma); | |
a1e78772 | 590 | |
84afd99b AW |
591 | set_vma_private_data(vma, (get_vma_private_data(vma) & |
592 | HPAGE_RESV_MASK) | (unsigned long)map); | |
04f2cbe3 MG |
593 | } |
594 | ||
595 | static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags) | |
596 | { | |
81d1b09c SL |
597 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); |
598 | VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma); | |
e7c4b0bf AW |
599 | |
600 | set_vma_private_data(vma, get_vma_private_data(vma) | flags); | |
04f2cbe3 MG |
601 | } |
602 | ||
603 | static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag) | |
604 | { | |
81d1b09c | 605 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); |
e7c4b0bf AW |
606 | |
607 | return (get_vma_private_data(vma) & flag) != 0; | |
a1e78772 MG |
608 | } |
609 | ||
04f2cbe3 | 610 | /* Reset counters to 0 and clear all HPAGE_RESV_* flags */ |
a1e78772 MG |
611 | void reset_vma_resv_huge_pages(struct vm_area_struct *vma) |
612 | { | |
81d1b09c | 613 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); |
f83a275d | 614 | if (!(vma->vm_flags & VM_MAYSHARE)) |
a1e78772 MG |
615 | vma->vm_private_data = (void *)0; |
616 | } | |
617 | ||
618 | /* Returns true if the VMA has associated reserve pages */ | |
af0ed73e | 619 | static int vma_has_reserves(struct vm_area_struct *vma, long chg) |
a1e78772 | 620 | { |
af0ed73e JK |
621 | if (vma->vm_flags & VM_NORESERVE) { |
622 | /* | |
623 | * This address is already reserved by other process(chg == 0), | |
624 | * so, we should decrement reserved count. Without decrementing, | |
625 | * reserve count remains after releasing inode, because this | |
626 | * allocated page will go into page cache and is regarded as | |
627 | * coming from reserved pool in releasing step. Currently, we | |
628 | * don't have any other solution to deal with this situation | |
629 | * properly, so add work-around here. | |
630 | */ | |
631 | if (vma->vm_flags & VM_MAYSHARE && chg == 0) | |
632 | return 1; | |
633 | else | |
634 | return 0; | |
635 | } | |
a63884e9 JK |
636 | |
637 | /* Shared mappings always use reserves */ | |
f83a275d | 638 | if (vma->vm_flags & VM_MAYSHARE) |
7f09ca51 | 639 | return 1; |
a63884e9 JK |
640 | |
641 | /* | |
642 | * Only the process that called mmap() has reserves for | |
643 | * private mappings. | |
644 | */ | |
7f09ca51 MG |
645 | if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) |
646 | return 1; | |
a63884e9 | 647 | |
7f09ca51 | 648 | return 0; |
a1e78772 MG |
649 | } |
650 | ||
a5516438 | 651 | static void enqueue_huge_page(struct hstate *h, struct page *page) |
1da177e4 LT |
652 | { |
653 | int nid = page_to_nid(page); | |
0edaecfa | 654 | list_move(&page->lru, &h->hugepage_freelists[nid]); |
a5516438 AK |
655 | h->free_huge_pages++; |
656 | h->free_huge_pages_node[nid]++; | |
1da177e4 LT |
657 | } |
658 | ||
bf50bab2 NH |
659 | static struct page *dequeue_huge_page_node(struct hstate *h, int nid) |
660 | { | |
661 | struct page *page; | |
662 | ||
c8721bbb NH |
663 | list_for_each_entry(page, &h->hugepage_freelists[nid], lru) |
664 | if (!is_migrate_isolate_page(page)) | |
665 | break; | |
666 | /* | |
667 | * if 'non-isolated free hugepage' not found on the list, | |
668 | * the allocation fails. | |
669 | */ | |
670 | if (&h->hugepage_freelists[nid] == &page->lru) | |
bf50bab2 | 671 | return NULL; |
0edaecfa | 672 | list_move(&page->lru, &h->hugepage_activelist); |
a9869b83 | 673 | set_page_refcounted(page); |
bf50bab2 NH |
674 | h->free_huge_pages--; |
675 | h->free_huge_pages_node[nid]--; | |
676 | return page; | |
677 | } | |
678 | ||
86cdb465 NH |
679 | /* Movability of hugepages depends on migration support. */ |
680 | static inline gfp_t htlb_alloc_mask(struct hstate *h) | |
681 | { | |
100873d7 | 682 | if (hugepages_treat_as_movable || hugepage_migration_supported(h)) |
86cdb465 NH |
683 | return GFP_HIGHUSER_MOVABLE; |
684 | else | |
685 | return GFP_HIGHUSER; | |
686 | } | |
687 | ||
a5516438 AK |
688 | static struct page *dequeue_huge_page_vma(struct hstate *h, |
689 | struct vm_area_struct *vma, | |
af0ed73e JK |
690 | unsigned long address, int avoid_reserve, |
691 | long chg) | |
1da177e4 | 692 | { |
b1c12cbc | 693 | struct page *page = NULL; |
480eccf9 | 694 | struct mempolicy *mpol; |
19770b32 | 695 | nodemask_t *nodemask; |
c0ff7453 | 696 | struct zonelist *zonelist; |
dd1a239f MG |
697 | struct zone *zone; |
698 | struct zoneref *z; | |
cc9a6c87 | 699 | unsigned int cpuset_mems_cookie; |
1da177e4 | 700 | |
a1e78772 MG |
701 | /* |
702 | * A child process with MAP_PRIVATE mappings created by their parent | |
703 | * have no page reserves. This check ensures that reservations are | |
704 | * not "stolen". The child may still get SIGKILLed | |
705 | */ | |
af0ed73e | 706 | if (!vma_has_reserves(vma, chg) && |
a5516438 | 707 | h->free_huge_pages - h->resv_huge_pages == 0) |
c0ff7453 | 708 | goto err; |
a1e78772 | 709 | |
04f2cbe3 | 710 | /* If reserves cannot be used, ensure enough pages are in the pool */ |
a5516438 | 711 | if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0) |
6eab04a8 | 712 | goto err; |
04f2cbe3 | 713 | |
9966c4bb | 714 | retry_cpuset: |
d26914d1 | 715 | cpuset_mems_cookie = read_mems_allowed_begin(); |
9966c4bb | 716 | zonelist = huge_zonelist(vma, address, |
86cdb465 | 717 | htlb_alloc_mask(h), &mpol, &nodemask); |
9966c4bb | 718 | |
19770b32 MG |
719 | for_each_zone_zonelist_nodemask(zone, z, zonelist, |
720 | MAX_NR_ZONES - 1, nodemask) { | |
344736f2 | 721 | if (cpuset_zone_allowed(zone, htlb_alloc_mask(h))) { |
bf50bab2 NH |
722 | page = dequeue_huge_page_node(h, zone_to_nid(zone)); |
723 | if (page) { | |
af0ed73e JK |
724 | if (avoid_reserve) |
725 | break; | |
726 | if (!vma_has_reserves(vma, chg)) | |
727 | break; | |
728 | ||
07443a85 | 729 | SetPagePrivate(page); |
af0ed73e | 730 | h->resv_huge_pages--; |
bf50bab2 NH |
731 | break; |
732 | } | |
3abf7afd | 733 | } |
1da177e4 | 734 | } |
cc9a6c87 | 735 | |
52cd3b07 | 736 | mpol_cond_put(mpol); |
d26914d1 | 737 | if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie))) |
cc9a6c87 | 738 | goto retry_cpuset; |
1da177e4 | 739 | return page; |
cc9a6c87 MG |
740 | |
741 | err: | |
cc9a6c87 | 742 | return NULL; |
1da177e4 LT |
743 | } |
744 | ||
1cac6f2c LC |
745 | /* |
746 | * common helper functions for hstate_next_node_to_{alloc|free}. | |
747 | * We may have allocated or freed a huge page based on a different | |
748 | * nodes_allowed previously, so h->next_node_to_{alloc|free} might | |
749 | * be outside of *nodes_allowed. Ensure that we use an allowed | |
750 | * node for alloc or free. | |
751 | */ | |
752 | static int next_node_allowed(int nid, nodemask_t *nodes_allowed) | |
753 | { | |
754 | nid = next_node(nid, *nodes_allowed); | |
755 | if (nid == MAX_NUMNODES) | |
756 | nid = first_node(*nodes_allowed); | |
757 | VM_BUG_ON(nid >= MAX_NUMNODES); | |
758 | ||
759 | return nid; | |
760 | } | |
761 | ||
762 | static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed) | |
763 | { | |
764 | if (!node_isset(nid, *nodes_allowed)) | |
765 | nid = next_node_allowed(nid, nodes_allowed); | |
766 | return nid; | |
767 | } | |
768 | ||
769 | /* | |
770 | * returns the previously saved node ["this node"] from which to | |
771 | * allocate a persistent huge page for the pool and advance the | |
772 | * next node from which to allocate, handling wrap at end of node | |
773 | * mask. | |
774 | */ | |
775 | static int hstate_next_node_to_alloc(struct hstate *h, | |
776 | nodemask_t *nodes_allowed) | |
777 | { | |
778 | int nid; | |
779 | ||
780 | VM_BUG_ON(!nodes_allowed); | |
781 | ||
782 | nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed); | |
783 | h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed); | |
784 | ||
785 | return nid; | |
786 | } | |
787 | ||
788 | /* | |
789 | * helper for free_pool_huge_page() - return the previously saved | |
790 | * node ["this node"] from which to free a huge page. Advance the | |
791 | * next node id whether or not we find a free huge page to free so | |
792 | * that the next attempt to free addresses the next node. | |
793 | */ | |
794 | static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed) | |
795 | { | |
796 | int nid; | |
797 | ||
798 | VM_BUG_ON(!nodes_allowed); | |
799 | ||
800 | nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed); | |
801 | h->next_nid_to_free = next_node_allowed(nid, nodes_allowed); | |
802 | ||
803 | return nid; | |
804 | } | |
805 | ||
806 | #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \ | |
807 | for (nr_nodes = nodes_weight(*mask); \ | |
808 | nr_nodes > 0 && \ | |
809 | ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \ | |
810 | nr_nodes--) | |
811 | ||
812 | #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \ | |
813 | for (nr_nodes = nodes_weight(*mask); \ | |
814 | nr_nodes > 0 && \ | |
815 | ((node = hstate_next_node_to_free(hs, mask)) || 1); \ | |
816 | nr_nodes--) | |
817 | ||
944d9fec LC |
818 | #if defined(CONFIG_CMA) && defined(CONFIG_X86_64) |
819 | static void destroy_compound_gigantic_page(struct page *page, | |
820 | unsigned long order) | |
821 | { | |
822 | int i; | |
823 | int nr_pages = 1 << order; | |
824 | struct page *p = page + 1; | |
825 | ||
826 | for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { | |
827 | __ClearPageTail(p); | |
828 | set_page_refcounted(p); | |
829 | p->first_page = NULL; | |
830 | } | |
831 | ||
832 | set_compound_order(page, 0); | |
833 | __ClearPageHead(page); | |
834 | } | |
835 | ||
836 | static void free_gigantic_page(struct page *page, unsigned order) | |
837 | { | |
838 | free_contig_range(page_to_pfn(page), 1 << order); | |
839 | } | |
840 | ||
841 | static int __alloc_gigantic_page(unsigned long start_pfn, | |
842 | unsigned long nr_pages) | |
843 | { | |
844 | unsigned long end_pfn = start_pfn + nr_pages; | |
845 | return alloc_contig_range(start_pfn, end_pfn, MIGRATE_MOVABLE); | |
846 | } | |
847 | ||
848 | static bool pfn_range_valid_gigantic(unsigned long start_pfn, | |
849 | unsigned long nr_pages) | |
850 | { | |
851 | unsigned long i, end_pfn = start_pfn + nr_pages; | |
852 | struct page *page; | |
853 | ||
854 | for (i = start_pfn; i < end_pfn; i++) { | |
855 | if (!pfn_valid(i)) | |
856 | return false; | |
857 | ||
858 | page = pfn_to_page(i); | |
859 | ||
860 | if (PageReserved(page)) | |
861 | return false; | |
862 | ||
863 | if (page_count(page) > 0) | |
864 | return false; | |
865 | ||
866 | if (PageHuge(page)) | |
867 | return false; | |
868 | } | |
869 | ||
870 | return true; | |
871 | } | |
872 | ||
873 | static bool zone_spans_last_pfn(const struct zone *zone, | |
874 | unsigned long start_pfn, unsigned long nr_pages) | |
875 | { | |
876 | unsigned long last_pfn = start_pfn + nr_pages - 1; | |
877 | return zone_spans_pfn(zone, last_pfn); | |
878 | } | |
879 | ||
880 | static struct page *alloc_gigantic_page(int nid, unsigned order) | |
881 | { | |
882 | unsigned long nr_pages = 1 << order; | |
883 | unsigned long ret, pfn, flags; | |
884 | struct zone *z; | |
885 | ||
886 | z = NODE_DATA(nid)->node_zones; | |
887 | for (; z - NODE_DATA(nid)->node_zones < MAX_NR_ZONES; z++) { | |
888 | spin_lock_irqsave(&z->lock, flags); | |
889 | ||
890 | pfn = ALIGN(z->zone_start_pfn, nr_pages); | |
891 | while (zone_spans_last_pfn(z, pfn, nr_pages)) { | |
892 | if (pfn_range_valid_gigantic(pfn, nr_pages)) { | |
893 | /* | |
894 | * We release the zone lock here because | |
895 | * alloc_contig_range() will also lock the zone | |
896 | * at some point. If there's an allocation | |
897 | * spinning on this lock, it may win the race | |
898 | * and cause alloc_contig_range() to fail... | |
899 | */ | |
900 | spin_unlock_irqrestore(&z->lock, flags); | |
901 | ret = __alloc_gigantic_page(pfn, nr_pages); | |
902 | if (!ret) | |
903 | return pfn_to_page(pfn); | |
904 | spin_lock_irqsave(&z->lock, flags); | |
905 | } | |
906 | pfn += nr_pages; | |
907 | } | |
908 | ||
909 | spin_unlock_irqrestore(&z->lock, flags); | |
910 | } | |
911 | ||
912 | return NULL; | |
913 | } | |
914 | ||
915 | static void prep_new_huge_page(struct hstate *h, struct page *page, int nid); | |
916 | static void prep_compound_gigantic_page(struct page *page, unsigned long order); | |
917 | ||
918 | static struct page *alloc_fresh_gigantic_page_node(struct hstate *h, int nid) | |
919 | { | |
920 | struct page *page; | |
921 | ||
922 | page = alloc_gigantic_page(nid, huge_page_order(h)); | |
923 | if (page) { | |
924 | prep_compound_gigantic_page(page, huge_page_order(h)); | |
925 | prep_new_huge_page(h, page, nid); | |
926 | } | |
927 | ||
928 | return page; | |
929 | } | |
930 | ||
931 | static int alloc_fresh_gigantic_page(struct hstate *h, | |
932 | nodemask_t *nodes_allowed) | |
933 | { | |
934 | struct page *page = NULL; | |
935 | int nr_nodes, node; | |
936 | ||
937 | for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { | |
938 | page = alloc_fresh_gigantic_page_node(h, node); | |
939 | if (page) | |
940 | return 1; | |
941 | } | |
942 | ||
943 | return 0; | |
944 | } | |
945 | ||
946 | static inline bool gigantic_page_supported(void) { return true; } | |
947 | #else | |
948 | static inline bool gigantic_page_supported(void) { return false; } | |
949 | static inline void free_gigantic_page(struct page *page, unsigned order) { } | |
950 | static inline void destroy_compound_gigantic_page(struct page *page, | |
951 | unsigned long order) { } | |
952 | static inline int alloc_fresh_gigantic_page(struct hstate *h, | |
953 | nodemask_t *nodes_allowed) { return 0; } | |
954 | #endif | |
955 | ||
a5516438 | 956 | static void update_and_free_page(struct hstate *h, struct page *page) |
6af2acb6 AL |
957 | { |
958 | int i; | |
a5516438 | 959 | |
944d9fec LC |
960 | if (hstate_is_gigantic(h) && !gigantic_page_supported()) |
961 | return; | |
18229df5 | 962 | |
a5516438 AK |
963 | h->nr_huge_pages--; |
964 | h->nr_huge_pages_node[page_to_nid(page)]--; | |
965 | for (i = 0; i < pages_per_huge_page(h); i++) { | |
32f84528 CF |
966 | page[i].flags &= ~(1 << PG_locked | 1 << PG_error | |
967 | 1 << PG_referenced | 1 << PG_dirty | | |
a7407a27 LC |
968 | 1 << PG_active | 1 << PG_private | |
969 | 1 << PG_writeback); | |
6af2acb6 | 970 | } |
309381fe | 971 | VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page); |
6af2acb6 AL |
972 | set_compound_page_dtor(page, NULL); |
973 | set_page_refcounted(page); | |
944d9fec LC |
974 | if (hstate_is_gigantic(h)) { |
975 | destroy_compound_gigantic_page(page, huge_page_order(h)); | |
976 | free_gigantic_page(page, huge_page_order(h)); | |
977 | } else { | |
978 | arch_release_hugepage(page); | |
979 | __free_pages(page, huge_page_order(h)); | |
980 | } | |
6af2acb6 AL |
981 | } |
982 | ||
e5ff2159 AK |
983 | struct hstate *size_to_hstate(unsigned long size) |
984 | { | |
985 | struct hstate *h; | |
986 | ||
987 | for_each_hstate(h) { | |
988 | if (huge_page_size(h) == size) | |
989 | return h; | |
990 | } | |
991 | return NULL; | |
992 | } | |
993 | ||
bcc54222 NH |
994 | /* |
995 | * Test to determine whether the hugepage is "active/in-use" (i.e. being linked | |
996 | * to hstate->hugepage_activelist.) | |
997 | * | |
998 | * This function can be called for tail pages, but never returns true for them. | |
999 | */ | |
1000 | bool page_huge_active(struct page *page) | |
1001 | { | |
1002 | VM_BUG_ON_PAGE(!PageHuge(page), page); | |
1003 | return PageHead(page) && PagePrivate(&page[1]); | |
1004 | } | |
1005 | ||
1006 | /* never called for tail page */ | |
1007 | static void set_page_huge_active(struct page *page) | |
1008 | { | |
1009 | VM_BUG_ON_PAGE(!PageHeadHuge(page), page); | |
1010 | SetPagePrivate(&page[1]); | |
1011 | } | |
1012 | ||
1013 | static void clear_page_huge_active(struct page *page) | |
1014 | { | |
1015 | VM_BUG_ON_PAGE(!PageHeadHuge(page), page); | |
1016 | ClearPagePrivate(&page[1]); | |
1017 | } | |
1018 | ||
8f1d26d0 | 1019 | void free_huge_page(struct page *page) |
27a85ef1 | 1020 | { |
a5516438 AK |
1021 | /* |
1022 | * Can't pass hstate in here because it is called from the | |
1023 | * compound page destructor. | |
1024 | */ | |
e5ff2159 | 1025 | struct hstate *h = page_hstate(page); |
7893d1d5 | 1026 | int nid = page_to_nid(page); |
90481622 DG |
1027 | struct hugepage_subpool *spool = |
1028 | (struct hugepage_subpool *)page_private(page); | |
07443a85 | 1029 | bool restore_reserve; |
27a85ef1 | 1030 | |
e5df70ab | 1031 | set_page_private(page, 0); |
23be7468 | 1032 | page->mapping = NULL; |
7893d1d5 | 1033 | BUG_ON(page_count(page)); |
0fe6e20b | 1034 | BUG_ON(page_mapcount(page)); |
07443a85 | 1035 | restore_reserve = PagePrivate(page); |
16c794b4 | 1036 | ClearPagePrivate(page); |
27a85ef1 | 1037 | |
1c5ecae3 MK |
1038 | /* |
1039 | * A return code of zero implies that the subpool will be under its | |
1040 | * minimum size if the reservation is not restored after page is free. | |
1041 | * Therefore, force restore_reserve operation. | |
1042 | */ | |
1043 | if (hugepage_subpool_put_pages(spool, 1) == 0) | |
1044 | restore_reserve = true; | |
1045 | ||
27a85ef1 | 1046 | spin_lock(&hugetlb_lock); |
bcc54222 | 1047 | clear_page_huge_active(page); |
6d76dcf4 AK |
1048 | hugetlb_cgroup_uncharge_page(hstate_index(h), |
1049 | pages_per_huge_page(h), page); | |
07443a85 JK |
1050 | if (restore_reserve) |
1051 | h->resv_huge_pages++; | |
1052 | ||
944d9fec | 1053 | if (h->surplus_huge_pages_node[nid]) { |
0edaecfa AK |
1054 | /* remove the page from active list */ |
1055 | list_del(&page->lru); | |
a5516438 AK |
1056 | update_and_free_page(h, page); |
1057 | h->surplus_huge_pages--; | |
1058 | h->surplus_huge_pages_node[nid]--; | |
7893d1d5 | 1059 | } else { |
5d3a551c | 1060 | arch_clear_hugepage_flags(page); |
a5516438 | 1061 | enqueue_huge_page(h, page); |
7893d1d5 | 1062 | } |
27a85ef1 DG |
1063 | spin_unlock(&hugetlb_lock); |
1064 | } | |
1065 | ||
a5516438 | 1066 | static void prep_new_huge_page(struct hstate *h, struct page *page, int nid) |
b7ba30c6 | 1067 | { |
0edaecfa | 1068 | INIT_LIST_HEAD(&page->lru); |
b7ba30c6 AK |
1069 | set_compound_page_dtor(page, free_huge_page); |
1070 | spin_lock(&hugetlb_lock); | |
9dd540e2 | 1071 | set_hugetlb_cgroup(page, NULL); |
a5516438 AK |
1072 | h->nr_huge_pages++; |
1073 | h->nr_huge_pages_node[nid]++; | |
b7ba30c6 AK |
1074 | spin_unlock(&hugetlb_lock); |
1075 | put_page(page); /* free it into the hugepage allocator */ | |
1076 | } | |
1077 | ||
2906dd52 | 1078 | static void prep_compound_gigantic_page(struct page *page, unsigned long order) |
20a0307c WF |
1079 | { |
1080 | int i; | |
1081 | int nr_pages = 1 << order; | |
1082 | struct page *p = page + 1; | |
1083 | ||
1084 | /* we rely on prep_new_huge_page to set the destructor */ | |
1085 | set_compound_order(page, order); | |
1086 | __SetPageHead(page); | |
ef5a22be | 1087 | __ClearPageReserved(page); |
20a0307c | 1088 | for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { |
ef5a22be AA |
1089 | /* |
1090 | * For gigantic hugepages allocated through bootmem at | |
1091 | * boot, it's safer to be consistent with the not-gigantic | |
1092 | * hugepages and clear the PG_reserved bit from all tail pages | |
1093 | * too. Otherwse drivers using get_user_pages() to access tail | |
1094 | * pages may get the reference counting wrong if they see | |
1095 | * PG_reserved set on a tail page (despite the head page not | |
1096 | * having PG_reserved set). Enforcing this consistency between | |
1097 | * head and tail pages allows drivers to optimize away a check | |
1098 | * on the head page when they need know if put_page() is needed | |
1099 | * after get_user_pages(). | |
1100 | */ | |
1101 | __ClearPageReserved(p); | |
58a84aa9 | 1102 | set_page_count(p, 0); |
20a0307c | 1103 | p->first_page = page; |
44fc8057 DR |
1104 | /* Make sure p->first_page is always valid for PageTail() */ |
1105 | smp_wmb(); | |
1106 | __SetPageTail(p); | |
20a0307c WF |
1107 | } |
1108 | } | |
1109 | ||
7795912c AM |
1110 | /* |
1111 | * PageHuge() only returns true for hugetlbfs pages, but not for normal or | |
1112 | * transparent huge pages. See the PageTransHuge() documentation for more | |
1113 | * details. | |
1114 | */ | |
20a0307c WF |
1115 | int PageHuge(struct page *page) |
1116 | { | |
20a0307c WF |
1117 | if (!PageCompound(page)) |
1118 | return 0; | |
1119 | ||
1120 | page = compound_head(page); | |
758f66a2 | 1121 | return get_compound_page_dtor(page) == free_huge_page; |
20a0307c | 1122 | } |
43131e14 NH |
1123 | EXPORT_SYMBOL_GPL(PageHuge); |
1124 | ||
27c73ae7 AA |
1125 | /* |
1126 | * PageHeadHuge() only returns true for hugetlbfs head page, but not for | |
1127 | * normal or transparent huge pages. | |
1128 | */ | |
1129 | int PageHeadHuge(struct page *page_head) | |
1130 | { | |
27c73ae7 AA |
1131 | if (!PageHead(page_head)) |
1132 | return 0; | |
1133 | ||
758f66a2 | 1134 | return get_compound_page_dtor(page_head) == free_huge_page; |
27c73ae7 | 1135 | } |
27c73ae7 | 1136 | |
13d60f4b ZY |
1137 | pgoff_t __basepage_index(struct page *page) |
1138 | { | |
1139 | struct page *page_head = compound_head(page); | |
1140 | pgoff_t index = page_index(page_head); | |
1141 | unsigned long compound_idx; | |
1142 | ||
1143 | if (!PageHuge(page_head)) | |
1144 | return page_index(page); | |
1145 | ||
1146 | if (compound_order(page_head) >= MAX_ORDER) | |
1147 | compound_idx = page_to_pfn(page) - page_to_pfn(page_head); | |
1148 | else | |
1149 | compound_idx = page - page_head; | |
1150 | ||
1151 | return (index << compound_order(page_head)) + compound_idx; | |
1152 | } | |
1153 | ||
a5516438 | 1154 | static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid) |
1da177e4 | 1155 | { |
1da177e4 | 1156 | struct page *page; |
f96efd58 | 1157 | |
6484eb3e | 1158 | page = alloc_pages_exact_node(nid, |
86cdb465 | 1159 | htlb_alloc_mask(h)|__GFP_COMP|__GFP_THISNODE| |
551883ae | 1160 | __GFP_REPEAT|__GFP_NOWARN, |
a5516438 | 1161 | huge_page_order(h)); |
1da177e4 | 1162 | if (page) { |
7f2e9525 | 1163 | if (arch_prepare_hugepage(page)) { |
caff3a2c | 1164 | __free_pages(page, huge_page_order(h)); |
7b8ee84d | 1165 | return NULL; |
7f2e9525 | 1166 | } |
a5516438 | 1167 | prep_new_huge_page(h, page, nid); |
1da177e4 | 1168 | } |
63b4613c NA |
1169 | |
1170 | return page; | |
1171 | } | |
1172 | ||
b2261026 JK |
1173 | static int alloc_fresh_huge_page(struct hstate *h, nodemask_t *nodes_allowed) |
1174 | { | |
1175 | struct page *page; | |
1176 | int nr_nodes, node; | |
1177 | int ret = 0; | |
1178 | ||
1179 | for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { | |
1180 | page = alloc_fresh_huge_page_node(h, node); | |
1181 | if (page) { | |
1182 | ret = 1; | |
1183 | break; | |
1184 | } | |
1185 | } | |
1186 | ||
1187 | if (ret) | |
1188 | count_vm_event(HTLB_BUDDY_PGALLOC); | |
1189 | else | |
1190 | count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); | |
1191 | ||
1192 | return ret; | |
1193 | } | |
1194 | ||
e8c5c824 LS |
1195 | /* |
1196 | * Free huge page from pool from next node to free. | |
1197 | * Attempt to keep persistent huge pages more or less | |
1198 | * balanced over allowed nodes. | |
1199 | * Called with hugetlb_lock locked. | |
1200 | */ | |
6ae11b27 LS |
1201 | static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed, |
1202 | bool acct_surplus) | |
e8c5c824 | 1203 | { |
b2261026 | 1204 | int nr_nodes, node; |
e8c5c824 LS |
1205 | int ret = 0; |
1206 | ||
b2261026 | 1207 | for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { |
685f3457 LS |
1208 | /* |
1209 | * If we're returning unused surplus pages, only examine | |
1210 | * nodes with surplus pages. | |
1211 | */ | |
b2261026 JK |
1212 | if ((!acct_surplus || h->surplus_huge_pages_node[node]) && |
1213 | !list_empty(&h->hugepage_freelists[node])) { | |
e8c5c824 | 1214 | struct page *page = |
b2261026 | 1215 | list_entry(h->hugepage_freelists[node].next, |
e8c5c824 LS |
1216 | struct page, lru); |
1217 | list_del(&page->lru); | |
1218 | h->free_huge_pages--; | |
b2261026 | 1219 | h->free_huge_pages_node[node]--; |
685f3457 LS |
1220 | if (acct_surplus) { |
1221 | h->surplus_huge_pages--; | |
b2261026 | 1222 | h->surplus_huge_pages_node[node]--; |
685f3457 | 1223 | } |
e8c5c824 LS |
1224 | update_and_free_page(h, page); |
1225 | ret = 1; | |
9a76db09 | 1226 | break; |
e8c5c824 | 1227 | } |
b2261026 | 1228 | } |
e8c5c824 LS |
1229 | |
1230 | return ret; | |
1231 | } | |
1232 | ||
c8721bbb NH |
1233 | /* |
1234 | * Dissolve a given free hugepage into free buddy pages. This function does | |
1235 | * nothing for in-use (including surplus) hugepages. | |
1236 | */ | |
1237 | static void dissolve_free_huge_page(struct page *page) | |
1238 | { | |
1239 | spin_lock(&hugetlb_lock); | |
1240 | if (PageHuge(page) && !page_count(page)) { | |
1241 | struct hstate *h = page_hstate(page); | |
1242 | int nid = page_to_nid(page); | |
1243 | list_del(&page->lru); | |
1244 | h->free_huge_pages--; | |
1245 | h->free_huge_pages_node[nid]--; | |
1246 | update_and_free_page(h, page); | |
1247 | } | |
1248 | spin_unlock(&hugetlb_lock); | |
1249 | } | |
1250 | ||
1251 | /* | |
1252 | * Dissolve free hugepages in a given pfn range. Used by memory hotplug to | |
1253 | * make specified memory blocks removable from the system. | |
1254 | * Note that start_pfn should aligned with (minimum) hugepage size. | |
1255 | */ | |
1256 | void dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn) | |
1257 | { | |
c8721bbb | 1258 | unsigned long pfn; |
c8721bbb | 1259 | |
d0177639 LZ |
1260 | if (!hugepages_supported()) |
1261 | return; | |
1262 | ||
641844f5 NH |
1263 | VM_BUG_ON(!IS_ALIGNED(start_pfn, 1 << minimum_order)); |
1264 | for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << minimum_order) | |
c8721bbb NH |
1265 | dissolve_free_huge_page(pfn_to_page(pfn)); |
1266 | } | |
1267 | ||
bf50bab2 | 1268 | static struct page *alloc_buddy_huge_page(struct hstate *h, int nid) |
7893d1d5 AL |
1269 | { |
1270 | struct page *page; | |
bf50bab2 | 1271 | unsigned int r_nid; |
7893d1d5 | 1272 | |
bae7f4ae | 1273 | if (hstate_is_gigantic(h)) |
aa888a74 AK |
1274 | return NULL; |
1275 | ||
d1c3fb1f NA |
1276 | /* |
1277 | * Assume we will successfully allocate the surplus page to | |
1278 | * prevent racing processes from causing the surplus to exceed | |
1279 | * overcommit | |
1280 | * | |
1281 | * This however introduces a different race, where a process B | |
1282 | * tries to grow the static hugepage pool while alloc_pages() is | |
1283 | * called by process A. B will only examine the per-node | |
1284 | * counters in determining if surplus huge pages can be | |
1285 | * converted to normal huge pages in adjust_pool_surplus(). A | |
1286 | * won't be able to increment the per-node counter, until the | |
1287 | * lock is dropped by B, but B doesn't drop hugetlb_lock until | |
1288 | * no more huge pages can be converted from surplus to normal | |
1289 | * state (and doesn't try to convert again). Thus, we have a | |
1290 | * case where a surplus huge page exists, the pool is grown, and | |
1291 | * the surplus huge page still exists after, even though it | |
1292 | * should just have been converted to a normal huge page. This | |
1293 | * does not leak memory, though, as the hugepage will be freed | |
1294 | * once it is out of use. It also does not allow the counters to | |
1295 | * go out of whack in adjust_pool_surplus() as we don't modify | |
1296 | * the node values until we've gotten the hugepage and only the | |
1297 | * per-node value is checked there. | |
1298 | */ | |
1299 | spin_lock(&hugetlb_lock); | |
a5516438 | 1300 | if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) { |
d1c3fb1f NA |
1301 | spin_unlock(&hugetlb_lock); |
1302 | return NULL; | |
1303 | } else { | |
a5516438 AK |
1304 | h->nr_huge_pages++; |
1305 | h->surplus_huge_pages++; | |
d1c3fb1f NA |
1306 | } |
1307 | spin_unlock(&hugetlb_lock); | |
1308 | ||
bf50bab2 | 1309 | if (nid == NUMA_NO_NODE) |
86cdb465 | 1310 | page = alloc_pages(htlb_alloc_mask(h)|__GFP_COMP| |
bf50bab2 NH |
1311 | __GFP_REPEAT|__GFP_NOWARN, |
1312 | huge_page_order(h)); | |
1313 | else | |
1314 | page = alloc_pages_exact_node(nid, | |
86cdb465 | 1315 | htlb_alloc_mask(h)|__GFP_COMP|__GFP_THISNODE| |
bf50bab2 | 1316 | __GFP_REPEAT|__GFP_NOWARN, huge_page_order(h)); |
d1c3fb1f | 1317 | |
caff3a2c GS |
1318 | if (page && arch_prepare_hugepage(page)) { |
1319 | __free_pages(page, huge_page_order(h)); | |
ea5768c7 | 1320 | page = NULL; |
caff3a2c GS |
1321 | } |
1322 | ||
d1c3fb1f | 1323 | spin_lock(&hugetlb_lock); |
7893d1d5 | 1324 | if (page) { |
0edaecfa | 1325 | INIT_LIST_HEAD(&page->lru); |
bf50bab2 | 1326 | r_nid = page_to_nid(page); |
7893d1d5 | 1327 | set_compound_page_dtor(page, free_huge_page); |
9dd540e2 | 1328 | set_hugetlb_cgroup(page, NULL); |
d1c3fb1f NA |
1329 | /* |
1330 | * We incremented the global counters already | |
1331 | */ | |
bf50bab2 NH |
1332 | h->nr_huge_pages_node[r_nid]++; |
1333 | h->surplus_huge_pages_node[r_nid]++; | |
3b116300 | 1334 | __count_vm_event(HTLB_BUDDY_PGALLOC); |
d1c3fb1f | 1335 | } else { |
a5516438 AK |
1336 | h->nr_huge_pages--; |
1337 | h->surplus_huge_pages--; | |
3b116300 | 1338 | __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); |
7893d1d5 | 1339 | } |
d1c3fb1f | 1340 | spin_unlock(&hugetlb_lock); |
7893d1d5 AL |
1341 | |
1342 | return page; | |
1343 | } | |
1344 | ||
bf50bab2 NH |
1345 | /* |
1346 | * This allocation function is useful in the context where vma is irrelevant. | |
1347 | * E.g. soft-offlining uses this function because it only cares physical | |
1348 | * address of error page. | |
1349 | */ | |
1350 | struct page *alloc_huge_page_node(struct hstate *h, int nid) | |
1351 | { | |
4ef91848 | 1352 | struct page *page = NULL; |
bf50bab2 NH |
1353 | |
1354 | spin_lock(&hugetlb_lock); | |
4ef91848 JK |
1355 | if (h->free_huge_pages - h->resv_huge_pages > 0) |
1356 | page = dequeue_huge_page_node(h, nid); | |
bf50bab2 NH |
1357 | spin_unlock(&hugetlb_lock); |
1358 | ||
94ae8ba7 | 1359 | if (!page) |
bf50bab2 NH |
1360 | page = alloc_buddy_huge_page(h, nid); |
1361 | ||
1362 | return page; | |
1363 | } | |
1364 | ||
e4e574b7 | 1365 | /* |
25985edc | 1366 | * Increase the hugetlb pool such that it can accommodate a reservation |
e4e574b7 AL |
1367 | * of size 'delta'. |
1368 | */ | |
a5516438 | 1369 | static int gather_surplus_pages(struct hstate *h, int delta) |
e4e574b7 AL |
1370 | { |
1371 | struct list_head surplus_list; | |
1372 | struct page *page, *tmp; | |
1373 | int ret, i; | |
1374 | int needed, allocated; | |
28073b02 | 1375 | bool alloc_ok = true; |
e4e574b7 | 1376 | |
a5516438 | 1377 | needed = (h->resv_huge_pages + delta) - h->free_huge_pages; |
ac09b3a1 | 1378 | if (needed <= 0) { |
a5516438 | 1379 | h->resv_huge_pages += delta; |
e4e574b7 | 1380 | return 0; |
ac09b3a1 | 1381 | } |
e4e574b7 AL |
1382 | |
1383 | allocated = 0; | |
1384 | INIT_LIST_HEAD(&surplus_list); | |
1385 | ||
1386 | ret = -ENOMEM; | |
1387 | retry: | |
1388 | spin_unlock(&hugetlb_lock); | |
1389 | for (i = 0; i < needed; i++) { | |
bf50bab2 | 1390 | page = alloc_buddy_huge_page(h, NUMA_NO_NODE); |
28073b02 HD |
1391 | if (!page) { |
1392 | alloc_ok = false; | |
1393 | break; | |
1394 | } | |
e4e574b7 AL |
1395 | list_add(&page->lru, &surplus_list); |
1396 | } | |
28073b02 | 1397 | allocated += i; |
e4e574b7 AL |
1398 | |
1399 | /* | |
1400 | * After retaking hugetlb_lock, we need to recalculate 'needed' | |
1401 | * because either resv_huge_pages or free_huge_pages may have changed. | |
1402 | */ | |
1403 | spin_lock(&hugetlb_lock); | |
a5516438 AK |
1404 | needed = (h->resv_huge_pages + delta) - |
1405 | (h->free_huge_pages + allocated); | |
28073b02 HD |
1406 | if (needed > 0) { |
1407 | if (alloc_ok) | |
1408 | goto retry; | |
1409 | /* | |
1410 | * We were not able to allocate enough pages to | |
1411 | * satisfy the entire reservation so we free what | |
1412 | * we've allocated so far. | |
1413 | */ | |
1414 | goto free; | |
1415 | } | |
e4e574b7 AL |
1416 | /* |
1417 | * The surplus_list now contains _at_least_ the number of extra pages | |
25985edc | 1418 | * needed to accommodate the reservation. Add the appropriate number |
e4e574b7 | 1419 | * of pages to the hugetlb pool and free the extras back to the buddy |
ac09b3a1 AL |
1420 | * allocator. Commit the entire reservation here to prevent another |
1421 | * process from stealing the pages as they are added to the pool but | |
1422 | * before they are reserved. | |
e4e574b7 AL |
1423 | */ |
1424 | needed += allocated; | |
a5516438 | 1425 | h->resv_huge_pages += delta; |
e4e574b7 | 1426 | ret = 0; |
a9869b83 | 1427 | |
19fc3f0a | 1428 | /* Free the needed pages to the hugetlb pool */ |
e4e574b7 | 1429 | list_for_each_entry_safe(page, tmp, &surplus_list, lru) { |
19fc3f0a AL |
1430 | if ((--needed) < 0) |
1431 | break; | |
a9869b83 NH |
1432 | /* |
1433 | * This page is now managed by the hugetlb allocator and has | |
1434 | * no users -- drop the buddy allocator's reference. | |
1435 | */ | |
1436 | put_page_testzero(page); | |
309381fe | 1437 | VM_BUG_ON_PAGE(page_count(page), page); |
a5516438 | 1438 | enqueue_huge_page(h, page); |
19fc3f0a | 1439 | } |
28073b02 | 1440 | free: |
b0365c8d | 1441 | spin_unlock(&hugetlb_lock); |
19fc3f0a AL |
1442 | |
1443 | /* Free unnecessary surplus pages to the buddy allocator */ | |
c0d934ba JK |
1444 | list_for_each_entry_safe(page, tmp, &surplus_list, lru) |
1445 | put_page(page); | |
a9869b83 | 1446 | spin_lock(&hugetlb_lock); |
e4e574b7 AL |
1447 | |
1448 | return ret; | |
1449 | } | |
1450 | ||
1451 | /* | |
1452 | * When releasing a hugetlb pool reservation, any surplus pages that were | |
1453 | * allocated to satisfy the reservation must be explicitly freed if they were | |
1454 | * never used. | |
685f3457 | 1455 | * Called with hugetlb_lock held. |
e4e574b7 | 1456 | */ |
a5516438 AK |
1457 | static void return_unused_surplus_pages(struct hstate *h, |
1458 | unsigned long unused_resv_pages) | |
e4e574b7 | 1459 | { |
e4e574b7 AL |
1460 | unsigned long nr_pages; |
1461 | ||
ac09b3a1 | 1462 | /* Uncommit the reservation */ |
a5516438 | 1463 | h->resv_huge_pages -= unused_resv_pages; |
ac09b3a1 | 1464 | |
aa888a74 | 1465 | /* Cannot return gigantic pages currently */ |
bae7f4ae | 1466 | if (hstate_is_gigantic(h)) |
aa888a74 AK |
1467 | return; |
1468 | ||
a5516438 | 1469 | nr_pages = min(unused_resv_pages, h->surplus_huge_pages); |
e4e574b7 | 1470 | |
685f3457 LS |
1471 | /* |
1472 | * We want to release as many surplus pages as possible, spread | |
9b5e5d0f LS |
1473 | * evenly across all nodes with memory. Iterate across these nodes |
1474 | * until we can no longer free unreserved surplus pages. This occurs | |
1475 | * when the nodes with surplus pages have no free pages. | |
1476 | * free_pool_huge_page() will balance the the freed pages across the | |
1477 | * on-line nodes with memory and will handle the hstate accounting. | |
685f3457 LS |
1478 | */ |
1479 | while (nr_pages--) { | |
8cebfcd0 | 1480 | if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1)) |
685f3457 | 1481 | break; |
7848a4bf | 1482 | cond_resched_lock(&hugetlb_lock); |
e4e574b7 AL |
1483 | } |
1484 | } | |
1485 | ||
c37f9fb1 | 1486 | /* |
cf3ad20b MK |
1487 | * vma_needs_reservation and vma_commit_reservation are used by the huge |
1488 | * page allocation routines to manage reservations. | |
1489 | * | |
1490 | * vma_needs_reservation is called to determine if the huge page at addr | |
1491 | * within the vma has an associated reservation. If a reservation is | |
1492 | * needed, the value 1 is returned. The caller is then responsible for | |
1493 | * managing the global reservation and subpool usage counts. After | |
1494 | * the huge page has been allocated, vma_commit_reservation is called | |
1495 | * to add the page to the reservation map. | |
1496 | * | |
1497 | * In the normal case, vma_commit_reservation returns the same value | |
1498 | * as the preceding vma_needs_reservation call. The only time this | |
1499 | * is not the case is if a reserve map was changed between calls. It | |
1500 | * is the responsibility of the caller to notice the difference and | |
1501 | * take appropriate action. | |
c37f9fb1 | 1502 | */ |
cf3ad20b MK |
1503 | static long __vma_reservation_common(struct hstate *h, |
1504 | struct vm_area_struct *vma, unsigned long addr, | |
1505 | bool commit) | |
c37f9fb1 | 1506 | { |
4e35f483 JK |
1507 | struct resv_map *resv; |
1508 | pgoff_t idx; | |
cf3ad20b | 1509 | long ret; |
c37f9fb1 | 1510 | |
4e35f483 JK |
1511 | resv = vma_resv_map(vma); |
1512 | if (!resv) | |
84afd99b | 1513 | return 1; |
c37f9fb1 | 1514 | |
4e35f483 | 1515 | idx = vma_hugecache_offset(h, vma, addr); |
cf3ad20b MK |
1516 | if (commit) |
1517 | ret = region_add(resv, idx, idx + 1); | |
1518 | else | |
1519 | ret = region_chg(resv, idx, idx + 1); | |
84afd99b | 1520 | |
4e35f483 | 1521 | if (vma->vm_flags & VM_MAYSHARE) |
cf3ad20b | 1522 | return ret; |
4e35f483 | 1523 | else |
cf3ad20b | 1524 | return ret < 0 ? ret : 0; |
c37f9fb1 | 1525 | } |
cf3ad20b MK |
1526 | |
1527 | static long vma_needs_reservation(struct hstate *h, | |
a5516438 | 1528 | struct vm_area_struct *vma, unsigned long addr) |
c37f9fb1 | 1529 | { |
cf3ad20b MK |
1530 | return __vma_reservation_common(h, vma, addr, false); |
1531 | } | |
84afd99b | 1532 | |
cf3ad20b MK |
1533 | static long vma_commit_reservation(struct hstate *h, |
1534 | struct vm_area_struct *vma, unsigned long addr) | |
1535 | { | |
1536 | return __vma_reservation_common(h, vma, addr, true); | |
c37f9fb1 AW |
1537 | } |
1538 | ||
a1e78772 | 1539 | static struct page *alloc_huge_page(struct vm_area_struct *vma, |
04f2cbe3 | 1540 | unsigned long addr, int avoid_reserve) |
1da177e4 | 1541 | { |
90481622 | 1542 | struct hugepage_subpool *spool = subpool_vma(vma); |
a5516438 | 1543 | struct hstate *h = hstate_vma(vma); |
348ea204 | 1544 | struct page *page; |
e2f17d94 | 1545 | long chg; |
6d76dcf4 AK |
1546 | int ret, idx; |
1547 | struct hugetlb_cgroup *h_cg; | |
a1e78772 | 1548 | |
6d76dcf4 | 1549 | idx = hstate_index(h); |
a1e78772 | 1550 | /* |
90481622 DG |
1551 | * Processes that did not create the mapping will have no |
1552 | * reserves and will not have accounted against subpool | |
1553 | * limit. Check that the subpool limit can be made before | |
1554 | * satisfying the allocation MAP_NORESERVE mappings may also | |
1555 | * need pages and subpool limit allocated allocated if no reserve | |
1556 | * mapping overlaps. | |
a1e78772 | 1557 | */ |
a5516438 | 1558 | chg = vma_needs_reservation(h, vma, addr); |
c37f9fb1 | 1559 | if (chg < 0) |
76dcee75 | 1560 | return ERR_PTR(-ENOMEM); |
8bb3f12e | 1561 | if (chg || avoid_reserve) |
1c5ecae3 | 1562 | if (hugepage_subpool_get_pages(spool, 1) < 0) |
76dcee75 | 1563 | return ERR_PTR(-ENOSPC); |
1da177e4 | 1564 | |
6d76dcf4 | 1565 | ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg); |
8f34af6f JZ |
1566 | if (ret) |
1567 | goto out_subpool_put; | |
1568 | ||
1da177e4 | 1569 | spin_lock(&hugetlb_lock); |
af0ed73e | 1570 | page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, chg); |
81a6fcae | 1571 | if (!page) { |
94ae8ba7 | 1572 | spin_unlock(&hugetlb_lock); |
bf50bab2 | 1573 | page = alloc_buddy_huge_page(h, NUMA_NO_NODE); |
8f34af6f JZ |
1574 | if (!page) |
1575 | goto out_uncharge_cgroup; | |
1576 | ||
79dbb236 AK |
1577 | spin_lock(&hugetlb_lock); |
1578 | list_move(&page->lru, &h->hugepage_activelist); | |
81a6fcae | 1579 | /* Fall through */ |
68842c9b | 1580 | } |
81a6fcae JK |
1581 | hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page); |
1582 | spin_unlock(&hugetlb_lock); | |
348ea204 | 1583 | |
90481622 | 1584 | set_page_private(page, (unsigned long)spool); |
90d8b7e6 | 1585 | |
a5516438 | 1586 | vma_commit_reservation(h, vma, addr); |
90d8b7e6 | 1587 | return page; |
8f34af6f JZ |
1588 | |
1589 | out_uncharge_cgroup: | |
1590 | hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg); | |
1591 | out_subpool_put: | |
1592 | if (chg || avoid_reserve) | |
1593 | hugepage_subpool_put_pages(spool, 1); | |
1594 | return ERR_PTR(-ENOSPC); | |
b45b5bd6 DG |
1595 | } |
1596 | ||
74060e4d NH |
1597 | /* |
1598 | * alloc_huge_page()'s wrapper which simply returns the page if allocation | |
1599 | * succeeds, otherwise NULL. This function is called from new_vma_page(), | |
1600 | * where no ERR_VALUE is expected to be returned. | |
1601 | */ | |
1602 | struct page *alloc_huge_page_noerr(struct vm_area_struct *vma, | |
1603 | unsigned long addr, int avoid_reserve) | |
1604 | { | |
1605 | struct page *page = alloc_huge_page(vma, addr, avoid_reserve); | |
1606 | if (IS_ERR(page)) | |
1607 | page = NULL; | |
1608 | return page; | |
1609 | } | |
1610 | ||
91f47662 | 1611 | int __weak alloc_bootmem_huge_page(struct hstate *h) |
aa888a74 AK |
1612 | { |
1613 | struct huge_bootmem_page *m; | |
b2261026 | 1614 | int nr_nodes, node; |
aa888a74 | 1615 | |
b2261026 | 1616 | for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) { |
aa888a74 AK |
1617 | void *addr; |
1618 | ||
8b89a116 GS |
1619 | addr = memblock_virt_alloc_try_nid_nopanic( |
1620 | huge_page_size(h), huge_page_size(h), | |
1621 | 0, BOOTMEM_ALLOC_ACCESSIBLE, node); | |
aa888a74 AK |
1622 | if (addr) { |
1623 | /* | |
1624 | * Use the beginning of the huge page to store the | |
1625 | * huge_bootmem_page struct (until gather_bootmem | |
1626 | * puts them into the mem_map). | |
1627 | */ | |
1628 | m = addr; | |
91f47662 | 1629 | goto found; |
aa888a74 | 1630 | } |
aa888a74 AK |
1631 | } |
1632 | return 0; | |
1633 | ||
1634 | found: | |
df994ead | 1635 | BUG_ON(!IS_ALIGNED(virt_to_phys(m), huge_page_size(h))); |
aa888a74 AK |
1636 | /* Put them into a private list first because mem_map is not up yet */ |
1637 | list_add(&m->list, &huge_boot_pages); | |
1638 | m->hstate = h; | |
1639 | return 1; | |
1640 | } | |
1641 | ||
f412c97a | 1642 | static void __init prep_compound_huge_page(struct page *page, int order) |
18229df5 AW |
1643 | { |
1644 | if (unlikely(order > (MAX_ORDER - 1))) | |
1645 | prep_compound_gigantic_page(page, order); | |
1646 | else | |
1647 | prep_compound_page(page, order); | |
1648 | } | |
1649 | ||
aa888a74 AK |
1650 | /* Put bootmem huge pages into the standard lists after mem_map is up */ |
1651 | static void __init gather_bootmem_prealloc(void) | |
1652 | { | |
1653 | struct huge_bootmem_page *m; | |
1654 | ||
1655 | list_for_each_entry(m, &huge_boot_pages, list) { | |
aa888a74 | 1656 | struct hstate *h = m->hstate; |
ee8f248d BB |
1657 | struct page *page; |
1658 | ||
1659 | #ifdef CONFIG_HIGHMEM | |
1660 | page = pfn_to_page(m->phys >> PAGE_SHIFT); | |
8b89a116 GS |
1661 | memblock_free_late(__pa(m), |
1662 | sizeof(struct huge_bootmem_page)); | |
ee8f248d BB |
1663 | #else |
1664 | page = virt_to_page(m); | |
1665 | #endif | |
aa888a74 | 1666 | WARN_ON(page_count(page) != 1); |
18229df5 | 1667 | prep_compound_huge_page(page, h->order); |
ef5a22be | 1668 | WARN_ON(PageReserved(page)); |
aa888a74 | 1669 | prep_new_huge_page(h, page, page_to_nid(page)); |
b0320c7b RA |
1670 | /* |
1671 | * If we had gigantic hugepages allocated at boot time, we need | |
1672 | * to restore the 'stolen' pages to totalram_pages in order to | |
1673 | * fix confusing memory reports from free(1) and another | |
1674 | * side-effects, like CommitLimit going negative. | |
1675 | */ | |
bae7f4ae | 1676 | if (hstate_is_gigantic(h)) |
3dcc0571 | 1677 | adjust_managed_page_count(page, 1 << h->order); |
aa888a74 AK |
1678 | } |
1679 | } | |
1680 | ||
8faa8b07 | 1681 | static void __init hugetlb_hstate_alloc_pages(struct hstate *h) |
1da177e4 LT |
1682 | { |
1683 | unsigned long i; | |
a5516438 | 1684 | |
e5ff2159 | 1685 | for (i = 0; i < h->max_huge_pages; ++i) { |
bae7f4ae | 1686 | if (hstate_is_gigantic(h)) { |
aa888a74 AK |
1687 | if (!alloc_bootmem_huge_page(h)) |
1688 | break; | |
9b5e5d0f | 1689 | } else if (!alloc_fresh_huge_page(h, |
8cebfcd0 | 1690 | &node_states[N_MEMORY])) |
1da177e4 | 1691 | break; |
1da177e4 | 1692 | } |
8faa8b07 | 1693 | h->max_huge_pages = i; |
e5ff2159 AK |
1694 | } |
1695 | ||
1696 | static void __init hugetlb_init_hstates(void) | |
1697 | { | |
1698 | struct hstate *h; | |
1699 | ||
1700 | for_each_hstate(h) { | |
641844f5 NH |
1701 | if (minimum_order > huge_page_order(h)) |
1702 | minimum_order = huge_page_order(h); | |
1703 | ||
8faa8b07 | 1704 | /* oversize hugepages were init'ed in early boot */ |
bae7f4ae | 1705 | if (!hstate_is_gigantic(h)) |
8faa8b07 | 1706 | hugetlb_hstate_alloc_pages(h); |
e5ff2159 | 1707 | } |
641844f5 | 1708 | VM_BUG_ON(minimum_order == UINT_MAX); |
e5ff2159 AK |
1709 | } |
1710 | ||
4abd32db AK |
1711 | static char * __init memfmt(char *buf, unsigned long n) |
1712 | { | |
1713 | if (n >= (1UL << 30)) | |
1714 | sprintf(buf, "%lu GB", n >> 30); | |
1715 | else if (n >= (1UL << 20)) | |
1716 | sprintf(buf, "%lu MB", n >> 20); | |
1717 | else | |
1718 | sprintf(buf, "%lu KB", n >> 10); | |
1719 | return buf; | |
1720 | } | |
1721 | ||
e5ff2159 AK |
1722 | static void __init report_hugepages(void) |
1723 | { | |
1724 | struct hstate *h; | |
1725 | ||
1726 | for_each_hstate(h) { | |
4abd32db | 1727 | char buf[32]; |
ffb22af5 | 1728 | pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n", |
4abd32db AK |
1729 | memfmt(buf, huge_page_size(h)), |
1730 | h->free_huge_pages); | |
e5ff2159 AK |
1731 | } |
1732 | } | |
1733 | ||
1da177e4 | 1734 | #ifdef CONFIG_HIGHMEM |
6ae11b27 LS |
1735 | static void try_to_free_low(struct hstate *h, unsigned long count, |
1736 | nodemask_t *nodes_allowed) | |
1da177e4 | 1737 | { |
4415cc8d CL |
1738 | int i; |
1739 | ||
bae7f4ae | 1740 | if (hstate_is_gigantic(h)) |
aa888a74 AK |
1741 | return; |
1742 | ||
6ae11b27 | 1743 | for_each_node_mask(i, *nodes_allowed) { |
1da177e4 | 1744 | struct page *page, *next; |
a5516438 AK |
1745 | struct list_head *freel = &h->hugepage_freelists[i]; |
1746 | list_for_each_entry_safe(page, next, freel, lru) { | |
1747 | if (count >= h->nr_huge_pages) | |
6b0c880d | 1748 | return; |
1da177e4 LT |
1749 | if (PageHighMem(page)) |
1750 | continue; | |
1751 | list_del(&page->lru); | |
e5ff2159 | 1752 | update_and_free_page(h, page); |
a5516438 AK |
1753 | h->free_huge_pages--; |
1754 | h->free_huge_pages_node[page_to_nid(page)]--; | |
1da177e4 LT |
1755 | } |
1756 | } | |
1757 | } | |
1758 | #else | |
6ae11b27 LS |
1759 | static inline void try_to_free_low(struct hstate *h, unsigned long count, |
1760 | nodemask_t *nodes_allowed) | |
1da177e4 LT |
1761 | { |
1762 | } | |
1763 | #endif | |
1764 | ||
20a0307c WF |
1765 | /* |
1766 | * Increment or decrement surplus_huge_pages. Keep node-specific counters | |
1767 | * balanced by operating on them in a round-robin fashion. | |
1768 | * Returns 1 if an adjustment was made. | |
1769 | */ | |
6ae11b27 LS |
1770 | static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed, |
1771 | int delta) | |
20a0307c | 1772 | { |
b2261026 | 1773 | int nr_nodes, node; |
20a0307c WF |
1774 | |
1775 | VM_BUG_ON(delta != -1 && delta != 1); | |
20a0307c | 1776 | |
b2261026 JK |
1777 | if (delta < 0) { |
1778 | for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { | |
1779 | if (h->surplus_huge_pages_node[node]) | |
1780 | goto found; | |
e8c5c824 | 1781 | } |
b2261026 JK |
1782 | } else { |
1783 | for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { | |
1784 | if (h->surplus_huge_pages_node[node] < | |
1785 | h->nr_huge_pages_node[node]) | |
1786 | goto found; | |
e8c5c824 | 1787 | } |
b2261026 JK |
1788 | } |
1789 | return 0; | |
20a0307c | 1790 | |
b2261026 JK |
1791 | found: |
1792 | h->surplus_huge_pages += delta; | |
1793 | h->surplus_huge_pages_node[node] += delta; | |
1794 | return 1; | |
20a0307c WF |
1795 | } |
1796 | ||
a5516438 | 1797 | #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages) |
6ae11b27 LS |
1798 | static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count, |
1799 | nodemask_t *nodes_allowed) | |
1da177e4 | 1800 | { |
7893d1d5 | 1801 | unsigned long min_count, ret; |
1da177e4 | 1802 | |
944d9fec | 1803 | if (hstate_is_gigantic(h) && !gigantic_page_supported()) |
aa888a74 AK |
1804 | return h->max_huge_pages; |
1805 | ||
7893d1d5 AL |
1806 | /* |
1807 | * Increase the pool size | |
1808 | * First take pages out of surplus state. Then make up the | |
1809 | * remaining difference by allocating fresh huge pages. | |
d1c3fb1f NA |
1810 | * |
1811 | * We might race with alloc_buddy_huge_page() here and be unable | |
1812 | * to convert a surplus huge page to a normal huge page. That is | |
1813 | * not critical, though, it just means the overall size of the | |
1814 | * pool might be one hugepage larger than it needs to be, but | |
1815 | * within all the constraints specified by the sysctls. | |
7893d1d5 | 1816 | */ |
1da177e4 | 1817 | spin_lock(&hugetlb_lock); |
a5516438 | 1818 | while (h->surplus_huge_pages && count > persistent_huge_pages(h)) { |
6ae11b27 | 1819 | if (!adjust_pool_surplus(h, nodes_allowed, -1)) |
7893d1d5 AL |
1820 | break; |
1821 | } | |
1822 | ||
a5516438 | 1823 | while (count > persistent_huge_pages(h)) { |
7893d1d5 AL |
1824 | /* |
1825 | * If this allocation races such that we no longer need the | |
1826 | * page, free_huge_page will handle it by freeing the page | |
1827 | * and reducing the surplus. | |
1828 | */ | |
1829 | spin_unlock(&hugetlb_lock); | |
944d9fec LC |
1830 | if (hstate_is_gigantic(h)) |
1831 | ret = alloc_fresh_gigantic_page(h, nodes_allowed); | |
1832 | else | |
1833 | ret = alloc_fresh_huge_page(h, nodes_allowed); | |
7893d1d5 AL |
1834 | spin_lock(&hugetlb_lock); |
1835 | if (!ret) | |
1836 | goto out; | |
1837 | ||
536240f2 MG |
1838 | /* Bail for signals. Probably ctrl-c from user */ |
1839 | if (signal_pending(current)) | |
1840 | goto out; | |
7893d1d5 | 1841 | } |
7893d1d5 AL |
1842 | |
1843 | /* | |
1844 | * Decrease the pool size | |
1845 | * First return free pages to the buddy allocator (being careful | |
1846 | * to keep enough around to satisfy reservations). Then place | |
1847 | * pages into surplus state as needed so the pool will shrink | |
1848 | * to the desired size as pages become free. | |
d1c3fb1f NA |
1849 | * |
1850 | * By placing pages into the surplus state independent of the | |
1851 | * overcommit value, we are allowing the surplus pool size to | |
1852 | * exceed overcommit. There are few sane options here. Since | |
1853 | * alloc_buddy_huge_page() is checking the global counter, | |
1854 | * though, we'll note that we're not allowed to exceed surplus | |
1855 | * and won't grow the pool anywhere else. Not until one of the | |
1856 | * sysctls are changed, or the surplus pages go out of use. | |
7893d1d5 | 1857 | */ |
a5516438 | 1858 | min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages; |
6b0c880d | 1859 | min_count = max(count, min_count); |
6ae11b27 | 1860 | try_to_free_low(h, min_count, nodes_allowed); |
a5516438 | 1861 | while (min_count < persistent_huge_pages(h)) { |
6ae11b27 | 1862 | if (!free_pool_huge_page(h, nodes_allowed, 0)) |
1da177e4 | 1863 | break; |
55f67141 | 1864 | cond_resched_lock(&hugetlb_lock); |
1da177e4 | 1865 | } |
a5516438 | 1866 | while (count < persistent_huge_pages(h)) { |
6ae11b27 | 1867 | if (!adjust_pool_surplus(h, nodes_allowed, 1)) |
7893d1d5 AL |
1868 | break; |
1869 | } | |
1870 | out: | |
a5516438 | 1871 | ret = persistent_huge_pages(h); |
1da177e4 | 1872 | spin_unlock(&hugetlb_lock); |
7893d1d5 | 1873 | return ret; |
1da177e4 LT |
1874 | } |
1875 | ||
a3437870 NA |
1876 | #define HSTATE_ATTR_RO(_name) \ |
1877 | static struct kobj_attribute _name##_attr = __ATTR_RO(_name) | |
1878 | ||
1879 | #define HSTATE_ATTR(_name) \ | |
1880 | static struct kobj_attribute _name##_attr = \ | |
1881 | __ATTR(_name, 0644, _name##_show, _name##_store) | |
1882 | ||
1883 | static struct kobject *hugepages_kobj; | |
1884 | static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; | |
1885 | ||
9a305230 LS |
1886 | static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp); |
1887 | ||
1888 | static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp) | |
a3437870 NA |
1889 | { |
1890 | int i; | |
9a305230 | 1891 | |
a3437870 | 1892 | for (i = 0; i < HUGE_MAX_HSTATE; i++) |
9a305230 LS |
1893 | if (hstate_kobjs[i] == kobj) { |
1894 | if (nidp) | |
1895 | *nidp = NUMA_NO_NODE; | |
a3437870 | 1896 | return &hstates[i]; |
9a305230 LS |
1897 | } |
1898 | ||
1899 | return kobj_to_node_hstate(kobj, nidp); | |
a3437870 NA |
1900 | } |
1901 | ||
06808b08 | 1902 | static ssize_t nr_hugepages_show_common(struct kobject *kobj, |
a3437870 NA |
1903 | struct kobj_attribute *attr, char *buf) |
1904 | { | |
9a305230 LS |
1905 | struct hstate *h; |
1906 | unsigned long nr_huge_pages; | |
1907 | int nid; | |
1908 | ||
1909 | h = kobj_to_hstate(kobj, &nid); | |
1910 | if (nid == NUMA_NO_NODE) | |
1911 | nr_huge_pages = h->nr_huge_pages; | |
1912 | else | |
1913 | nr_huge_pages = h->nr_huge_pages_node[nid]; | |
1914 | ||
1915 | return sprintf(buf, "%lu\n", nr_huge_pages); | |
a3437870 | 1916 | } |
adbe8726 | 1917 | |
238d3c13 DR |
1918 | static ssize_t __nr_hugepages_store_common(bool obey_mempolicy, |
1919 | struct hstate *h, int nid, | |
1920 | unsigned long count, size_t len) | |
a3437870 NA |
1921 | { |
1922 | int err; | |
bad44b5b | 1923 | NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY); |
a3437870 | 1924 | |
944d9fec | 1925 | if (hstate_is_gigantic(h) && !gigantic_page_supported()) { |
adbe8726 EM |
1926 | err = -EINVAL; |
1927 | goto out; | |
1928 | } | |
1929 | ||
9a305230 LS |
1930 | if (nid == NUMA_NO_NODE) { |
1931 | /* | |
1932 | * global hstate attribute | |
1933 | */ | |
1934 | if (!(obey_mempolicy && | |
1935 | init_nodemask_of_mempolicy(nodes_allowed))) { | |
1936 | NODEMASK_FREE(nodes_allowed); | |
8cebfcd0 | 1937 | nodes_allowed = &node_states[N_MEMORY]; |
9a305230 LS |
1938 | } |
1939 | } else if (nodes_allowed) { | |
1940 | /* | |
1941 | * per node hstate attribute: adjust count to global, | |
1942 | * but restrict alloc/free to the specified node. | |
1943 | */ | |
1944 | count += h->nr_huge_pages - h->nr_huge_pages_node[nid]; | |
1945 | init_nodemask_of_node(nodes_allowed, nid); | |
1946 | } else | |
8cebfcd0 | 1947 | nodes_allowed = &node_states[N_MEMORY]; |
9a305230 | 1948 | |
06808b08 | 1949 | h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed); |
a3437870 | 1950 | |
8cebfcd0 | 1951 | if (nodes_allowed != &node_states[N_MEMORY]) |
06808b08 LS |
1952 | NODEMASK_FREE(nodes_allowed); |
1953 | ||
1954 | return len; | |
adbe8726 EM |
1955 | out: |
1956 | NODEMASK_FREE(nodes_allowed); | |
1957 | return err; | |
06808b08 LS |
1958 | } |
1959 | ||
238d3c13 DR |
1960 | static ssize_t nr_hugepages_store_common(bool obey_mempolicy, |
1961 | struct kobject *kobj, const char *buf, | |
1962 | size_t len) | |
1963 | { | |
1964 | struct hstate *h; | |
1965 | unsigned long count; | |
1966 | int nid; | |
1967 | int err; | |
1968 | ||
1969 | err = kstrtoul(buf, 10, &count); | |
1970 | if (err) | |
1971 | return err; | |
1972 | ||
1973 | h = kobj_to_hstate(kobj, &nid); | |
1974 | return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len); | |
1975 | } | |
1976 | ||
06808b08 LS |
1977 | static ssize_t nr_hugepages_show(struct kobject *kobj, |
1978 | struct kobj_attribute *attr, char *buf) | |
1979 | { | |
1980 | return nr_hugepages_show_common(kobj, attr, buf); | |
1981 | } | |
1982 | ||
1983 | static ssize_t nr_hugepages_store(struct kobject *kobj, | |
1984 | struct kobj_attribute *attr, const char *buf, size_t len) | |
1985 | { | |
238d3c13 | 1986 | return nr_hugepages_store_common(false, kobj, buf, len); |
a3437870 NA |
1987 | } |
1988 | HSTATE_ATTR(nr_hugepages); | |
1989 | ||
06808b08 LS |
1990 | #ifdef CONFIG_NUMA |
1991 | ||
1992 | /* | |
1993 | * hstate attribute for optionally mempolicy-based constraint on persistent | |
1994 | * huge page alloc/free. | |
1995 | */ | |
1996 | static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj, | |
1997 | struct kobj_attribute *attr, char *buf) | |
1998 | { | |
1999 | return nr_hugepages_show_common(kobj, attr, buf); | |
2000 | } | |
2001 | ||
2002 | static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj, | |
2003 | struct kobj_attribute *attr, const char *buf, size_t len) | |
2004 | { | |
238d3c13 | 2005 | return nr_hugepages_store_common(true, kobj, buf, len); |
06808b08 LS |
2006 | } |
2007 | HSTATE_ATTR(nr_hugepages_mempolicy); | |
2008 | #endif | |
2009 | ||
2010 | ||
a3437870 NA |
2011 | static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj, |
2012 | struct kobj_attribute *attr, char *buf) | |
2013 | { | |
9a305230 | 2014 | struct hstate *h = kobj_to_hstate(kobj, NULL); |
a3437870 NA |
2015 | return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages); |
2016 | } | |
adbe8726 | 2017 | |
a3437870 NA |
2018 | static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj, |
2019 | struct kobj_attribute *attr, const char *buf, size_t count) | |
2020 | { | |
2021 | int err; | |
2022 | unsigned long input; | |
9a305230 | 2023 | struct hstate *h = kobj_to_hstate(kobj, NULL); |
a3437870 | 2024 | |
bae7f4ae | 2025 | if (hstate_is_gigantic(h)) |
adbe8726 EM |
2026 | return -EINVAL; |
2027 | ||
3dbb95f7 | 2028 | err = kstrtoul(buf, 10, &input); |
a3437870 | 2029 | if (err) |
73ae31e5 | 2030 | return err; |
a3437870 NA |
2031 | |
2032 | spin_lock(&hugetlb_lock); | |
2033 | h->nr_overcommit_huge_pages = input; | |
2034 | spin_unlock(&hugetlb_lock); | |
2035 | ||
2036 | return count; | |
2037 | } | |
2038 | HSTATE_ATTR(nr_overcommit_hugepages); | |
2039 | ||
2040 | static ssize_t free_hugepages_show(struct kobject *kobj, | |
2041 | struct kobj_attribute *attr, char *buf) | |
2042 | { | |
9a305230 LS |
2043 | struct hstate *h; |
2044 | unsigned long free_huge_pages; | |
2045 | int nid; | |
2046 | ||
2047 | h = kobj_to_hstate(kobj, &nid); | |
2048 | if (nid == NUMA_NO_NODE) | |
2049 | free_huge_pages = h->free_huge_pages; | |
2050 | else | |
2051 | free_huge_pages = h->free_huge_pages_node[nid]; | |
2052 | ||
2053 | return sprintf(buf, "%lu\n", free_huge_pages); | |
a3437870 NA |
2054 | } |
2055 | HSTATE_ATTR_RO(free_hugepages); | |
2056 | ||
2057 | static ssize_t resv_hugepages_show(struct kobject *kobj, | |
2058 | struct kobj_attribute *attr, char *buf) | |
2059 | { | |
9a305230 | 2060 | struct hstate *h = kobj_to_hstate(kobj, NULL); |
a3437870 NA |
2061 | return sprintf(buf, "%lu\n", h->resv_huge_pages); |
2062 | } | |
2063 | HSTATE_ATTR_RO(resv_hugepages); | |
2064 | ||
2065 | static ssize_t surplus_hugepages_show(struct kobject *kobj, | |
2066 | struct kobj_attribute *attr, char *buf) | |
2067 | { | |
9a305230 LS |
2068 | struct hstate *h; |
2069 | unsigned long surplus_huge_pages; | |
2070 | int nid; | |
2071 | ||
2072 | h = kobj_to_hstate(kobj, &nid); | |
2073 | if (nid == NUMA_NO_NODE) | |
2074 | surplus_huge_pages = h->surplus_huge_pages; | |
2075 | else | |
2076 | surplus_huge_pages = h->surplus_huge_pages_node[nid]; | |
2077 | ||
2078 | return sprintf(buf, "%lu\n", surplus_huge_pages); | |
a3437870 NA |
2079 | } |
2080 | HSTATE_ATTR_RO(surplus_hugepages); | |
2081 | ||
2082 | static struct attribute *hstate_attrs[] = { | |
2083 | &nr_hugepages_attr.attr, | |
2084 | &nr_overcommit_hugepages_attr.attr, | |
2085 | &free_hugepages_attr.attr, | |
2086 | &resv_hugepages_attr.attr, | |
2087 | &surplus_hugepages_attr.attr, | |
06808b08 LS |
2088 | #ifdef CONFIG_NUMA |
2089 | &nr_hugepages_mempolicy_attr.attr, | |
2090 | #endif | |
a3437870 NA |
2091 | NULL, |
2092 | }; | |
2093 | ||
2094 | static struct attribute_group hstate_attr_group = { | |
2095 | .attrs = hstate_attrs, | |
2096 | }; | |
2097 | ||
094e9539 JM |
2098 | static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent, |
2099 | struct kobject **hstate_kobjs, | |
2100 | struct attribute_group *hstate_attr_group) | |
a3437870 NA |
2101 | { |
2102 | int retval; | |
972dc4de | 2103 | int hi = hstate_index(h); |
a3437870 | 2104 | |
9a305230 LS |
2105 | hstate_kobjs[hi] = kobject_create_and_add(h->name, parent); |
2106 | if (!hstate_kobjs[hi]) | |
a3437870 NA |
2107 | return -ENOMEM; |
2108 | ||
9a305230 | 2109 | retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group); |
a3437870 | 2110 | if (retval) |
9a305230 | 2111 | kobject_put(hstate_kobjs[hi]); |
a3437870 NA |
2112 | |
2113 | return retval; | |
2114 | } | |
2115 | ||
2116 | static void __init hugetlb_sysfs_init(void) | |
2117 | { | |
2118 | struct hstate *h; | |
2119 | int err; | |
2120 | ||
2121 | hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj); | |
2122 | if (!hugepages_kobj) | |
2123 | return; | |
2124 | ||
2125 | for_each_hstate(h) { | |
9a305230 LS |
2126 | err = hugetlb_sysfs_add_hstate(h, hugepages_kobj, |
2127 | hstate_kobjs, &hstate_attr_group); | |
a3437870 | 2128 | if (err) |
ffb22af5 | 2129 | pr_err("Hugetlb: Unable to add hstate %s", h->name); |
a3437870 NA |
2130 | } |
2131 | } | |
2132 | ||
9a305230 LS |
2133 | #ifdef CONFIG_NUMA |
2134 | ||
2135 | /* | |
2136 | * node_hstate/s - associate per node hstate attributes, via their kobjects, | |
10fbcf4c KS |
2137 | * with node devices in node_devices[] using a parallel array. The array |
2138 | * index of a node device or _hstate == node id. | |
2139 | * This is here to avoid any static dependency of the node device driver, in | |
9a305230 LS |
2140 | * the base kernel, on the hugetlb module. |
2141 | */ | |
2142 | struct node_hstate { | |
2143 | struct kobject *hugepages_kobj; | |
2144 | struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; | |
2145 | }; | |
2146 | struct node_hstate node_hstates[MAX_NUMNODES]; | |
2147 | ||
2148 | /* | |
10fbcf4c | 2149 | * A subset of global hstate attributes for node devices |
9a305230 LS |
2150 | */ |
2151 | static struct attribute *per_node_hstate_attrs[] = { | |
2152 | &nr_hugepages_attr.attr, | |
2153 | &free_hugepages_attr.attr, | |
2154 | &surplus_hugepages_attr.attr, | |
2155 | NULL, | |
2156 | }; | |
2157 | ||
2158 | static struct attribute_group per_node_hstate_attr_group = { | |
2159 | .attrs = per_node_hstate_attrs, | |
2160 | }; | |
2161 | ||
2162 | /* | |
10fbcf4c | 2163 | * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj. |
9a305230 LS |
2164 | * Returns node id via non-NULL nidp. |
2165 | */ | |
2166 | static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) | |
2167 | { | |
2168 | int nid; | |
2169 | ||
2170 | for (nid = 0; nid < nr_node_ids; nid++) { | |
2171 | struct node_hstate *nhs = &node_hstates[nid]; | |
2172 | int i; | |
2173 | for (i = 0; i < HUGE_MAX_HSTATE; i++) | |
2174 | if (nhs->hstate_kobjs[i] == kobj) { | |
2175 | if (nidp) | |
2176 | *nidp = nid; | |
2177 | return &hstates[i]; | |
2178 | } | |
2179 | } | |
2180 | ||
2181 | BUG(); | |
2182 | return NULL; | |
2183 | } | |
2184 | ||
2185 | /* | |
10fbcf4c | 2186 | * Unregister hstate attributes from a single node device. |
9a305230 LS |
2187 | * No-op if no hstate attributes attached. |
2188 | */ | |
3cd8b44f | 2189 | static void hugetlb_unregister_node(struct node *node) |
9a305230 LS |
2190 | { |
2191 | struct hstate *h; | |
10fbcf4c | 2192 | struct node_hstate *nhs = &node_hstates[node->dev.id]; |
9a305230 LS |
2193 | |
2194 | if (!nhs->hugepages_kobj) | |
9b5e5d0f | 2195 | return; /* no hstate attributes */ |
9a305230 | 2196 | |
972dc4de AK |
2197 | for_each_hstate(h) { |
2198 | int idx = hstate_index(h); | |
2199 | if (nhs->hstate_kobjs[idx]) { | |
2200 | kobject_put(nhs->hstate_kobjs[idx]); | |
2201 | nhs->hstate_kobjs[idx] = NULL; | |
9a305230 | 2202 | } |
972dc4de | 2203 | } |
9a305230 LS |
2204 | |
2205 | kobject_put(nhs->hugepages_kobj); | |
2206 | nhs->hugepages_kobj = NULL; | |
2207 | } | |
2208 | ||
2209 | /* | |
10fbcf4c | 2210 | * hugetlb module exit: unregister hstate attributes from node devices |
9a305230 LS |
2211 | * that have them. |
2212 | */ | |
2213 | static void hugetlb_unregister_all_nodes(void) | |
2214 | { | |
2215 | int nid; | |
2216 | ||
2217 | /* | |
10fbcf4c | 2218 | * disable node device registrations. |
9a305230 LS |
2219 | */ |
2220 | register_hugetlbfs_with_node(NULL, NULL); | |
2221 | ||
2222 | /* | |
2223 | * remove hstate attributes from any nodes that have them. | |
2224 | */ | |
2225 | for (nid = 0; nid < nr_node_ids; nid++) | |
8732794b | 2226 | hugetlb_unregister_node(node_devices[nid]); |
9a305230 LS |
2227 | } |
2228 | ||
2229 | /* | |
10fbcf4c | 2230 | * Register hstate attributes for a single node device. |
9a305230 LS |
2231 | * No-op if attributes already registered. |
2232 | */ | |
3cd8b44f | 2233 | static void hugetlb_register_node(struct node *node) |
9a305230 LS |
2234 | { |
2235 | struct hstate *h; | |
10fbcf4c | 2236 | struct node_hstate *nhs = &node_hstates[node->dev.id]; |
9a305230 LS |
2237 | int err; |
2238 | ||
2239 | if (nhs->hugepages_kobj) | |
2240 | return; /* already allocated */ | |
2241 | ||
2242 | nhs->hugepages_kobj = kobject_create_and_add("hugepages", | |
10fbcf4c | 2243 | &node->dev.kobj); |
9a305230 LS |
2244 | if (!nhs->hugepages_kobj) |
2245 | return; | |
2246 | ||
2247 | for_each_hstate(h) { | |
2248 | err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj, | |
2249 | nhs->hstate_kobjs, | |
2250 | &per_node_hstate_attr_group); | |
2251 | if (err) { | |
ffb22af5 AM |
2252 | pr_err("Hugetlb: Unable to add hstate %s for node %d\n", |
2253 | h->name, node->dev.id); | |
9a305230 LS |
2254 | hugetlb_unregister_node(node); |
2255 | break; | |
2256 | } | |
2257 | } | |
2258 | } | |
2259 | ||
2260 | /* | |
9b5e5d0f | 2261 | * hugetlb init time: register hstate attributes for all registered node |
10fbcf4c KS |
2262 | * devices of nodes that have memory. All on-line nodes should have |
2263 | * registered their associated device by this time. | |
9a305230 | 2264 | */ |
7d9ca000 | 2265 | static void __init hugetlb_register_all_nodes(void) |
9a305230 LS |
2266 | { |
2267 | int nid; | |
2268 | ||
8cebfcd0 | 2269 | for_each_node_state(nid, N_MEMORY) { |
8732794b | 2270 | struct node *node = node_devices[nid]; |
10fbcf4c | 2271 | if (node->dev.id == nid) |
9a305230 LS |
2272 | hugetlb_register_node(node); |
2273 | } | |
2274 | ||
2275 | /* | |
10fbcf4c | 2276 | * Let the node device driver know we're here so it can |
9a305230 LS |
2277 | * [un]register hstate attributes on node hotplug. |
2278 | */ | |
2279 | register_hugetlbfs_with_node(hugetlb_register_node, | |
2280 | hugetlb_unregister_node); | |
2281 | } | |
2282 | #else /* !CONFIG_NUMA */ | |
2283 | ||
2284 | static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) | |
2285 | { | |
2286 | BUG(); | |
2287 | if (nidp) | |
2288 | *nidp = -1; | |
2289 | return NULL; | |
2290 | } | |
2291 | ||
2292 | static void hugetlb_unregister_all_nodes(void) { } | |
2293 | ||
2294 | static void hugetlb_register_all_nodes(void) { } | |
2295 | ||
2296 | #endif | |
2297 | ||
a3437870 NA |
2298 | static void __exit hugetlb_exit(void) |
2299 | { | |
2300 | struct hstate *h; | |
2301 | ||
9a305230 LS |
2302 | hugetlb_unregister_all_nodes(); |
2303 | ||
a3437870 | 2304 | for_each_hstate(h) { |
972dc4de | 2305 | kobject_put(hstate_kobjs[hstate_index(h)]); |
a3437870 NA |
2306 | } |
2307 | ||
2308 | kobject_put(hugepages_kobj); | |
8382d914 | 2309 | kfree(htlb_fault_mutex_table); |
a3437870 NA |
2310 | } |
2311 | module_exit(hugetlb_exit); | |
2312 | ||
2313 | static int __init hugetlb_init(void) | |
2314 | { | |
8382d914 DB |
2315 | int i; |
2316 | ||
457c1b27 | 2317 | if (!hugepages_supported()) |
0ef89d25 | 2318 | return 0; |
a3437870 | 2319 | |
e11bfbfc NP |
2320 | if (!size_to_hstate(default_hstate_size)) { |
2321 | default_hstate_size = HPAGE_SIZE; | |
2322 | if (!size_to_hstate(default_hstate_size)) | |
2323 | hugetlb_add_hstate(HUGETLB_PAGE_ORDER); | |
a3437870 | 2324 | } |
972dc4de | 2325 | default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size)); |
e11bfbfc NP |
2326 | if (default_hstate_max_huge_pages) |
2327 | default_hstate.max_huge_pages = default_hstate_max_huge_pages; | |
a3437870 NA |
2328 | |
2329 | hugetlb_init_hstates(); | |
aa888a74 | 2330 | gather_bootmem_prealloc(); |
a3437870 NA |
2331 | report_hugepages(); |
2332 | ||
2333 | hugetlb_sysfs_init(); | |
9a305230 | 2334 | hugetlb_register_all_nodes(); |
7179e7bf | 2335 | hugetlb_cgroup_file_init(); |
9a305230 | 2336 | |
8382d914 DB |
2337 | #ifdef CONFIG_SMP |
2338 | num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus()); | |
2339 | #else | |
2340 | num_fault_mutexes = 1; | |
2341 | #endif | |
2342 | htlb_fault_mutex_table = | |
2343 | kmalloc(sizeof(struct mutex) * num_fault_mutexes, GFP_KERNEL); | |
2344 | BUG_ON(!htlb_fault_mutex_table); | |
2345 | ||
2346 | for (i = 0; i < num_fault_mutexes; i++) | |
2347 | mutex_init(&htlb_fault_mutex_table[i]); | |
a3437870 NA |
2348 | return 0; |
2349 | } | |
2350 | module_init(hugetlb_init); | |
2351 | ||
2352 | /* Should be called on processing a hugepagesz=... option */ | |
2353 | void __init hugetlb_add_hstate(unsigned order) | |
2354 | { | |
2355 | struct hstate *h; | |
8faa8b07 AK |
2356 | unsigned long i; |
2357 | ||
a3437870 | 2358 | if (size_to_hstate(PAGE_SIZE << order)) { |
ffb22af5 | 2359 | pr_warning("hugepagesz= specified twice, ignoring\n"); |
a3437870 NA |
2360 | return; |
2361 | } | |
47d38344 | 2362 | BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE); |
a3437870 | 2363 | BUG_ON(order == 0); |
47d38344 | 2364 | h = &hstates[hugetlb_max_hstate++]; |
a3437870 NA |
2365 | h->order = order; |
2366 | h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1); | |
8faa8b07 AK |
2367 | h->nr_huge_pages = 0; |
2368 | h->free_huge_pages = 0; | |
2369 | for (i = 0; i < MAX_NUMNODES; ++i) | |
2370 | INIT_LIST_HEAD(&h->hugepage_freelists[i]); | |
0edaecfa | 2371 | INIT_LIST_HEAD(&h->hugepage_activelist); |
8cebfcd0 LJ |
2372 | h->next_nid_to_alloc = first_node(node_states[N_MEMORY]); |
2373 | h->next_nid_to_free = first_node(node_states[N_MEMORY]); | |
a3437870 NA |
2374 | snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB", |
2375 | huge_page_size(h)/1024); | |
8faa8b07 | 2376 | |
a3437870 NA |
2377 | parsed_hstate = h; |
2378 | } | |
2379 | ||
e11bfbfc | 2380 | static int __init hugetlb_nrpages_setup(char *s) |
a3437870 NA |
2381 | { |
2382 | unsigned long *mhp; | |
8faa8b07 | 2383 | static unsigned long *last_mhp; |
a3437870 NA |
2384 | |
2385 | /* | |
47d38344 | 2386 | * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet, |
a3437870 NA |
2387 | * so this hugepages= parameter goes to the "default hstate". |
2388 | */ | |
47d38344 | 2389 | if (!hugetlb_max_hstate) |
a3437870 NA |
2390 | mhp = &default_hstate_max_huge_pages; |
2391 | else | |
2392 | mhp = &parsed_hstate->max_huge_pages; | |
2393 | ||
8faa8b07 | 2394 | if (mhp == last_mhp) { |
ffb22af5 AM |
2395 | pr_warning("hugepages= specified twice without " |
2396 | "interleaving hugepagesz=, ignoring\n"); | |
8faa8b07 AK |
2397 | return 1; |
2398 | } | |
2399 | ||
a3437870 NA |
2400 | if (sscanf(s, "%lu", mhp) <= 0) |
2401 | *mhp = 0; | |
2402 | ||
8faa8b07 AK |
2403 | /* |
2404 | * Global state is always initialized later in hugetlb_init. | |
2405 | * But we need to allocate >= MAX_ORDER hstates here early to still | |
2406 | * use the bootmem allocator. | |
2407 | */ | |
47d38344 | 2408 | if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER) |
8faa8b07 AK |
2409 | hugetlb_hstate_alloc_pages(parsed_hstate); |
2410 | ||
2411 | last_mhp = mhp; | |
2412 | ||
a3437870 NA |
2413 | return 1; |
2414 | } | |
e11bfbfc NP |
2415 | __setup("hugepages=", hugetlb_nrpages_setup); |
2416 | ||
2417 | static int __init hugetlb_default_setup(char *s) | |
2418 | { | |
2419 | default_hstate_size = memparse(s, &s); | |
2420 | return 1; | |
2421 | } | |
2422 | __setup("default_hugepagesz=", hugetlb_default_setup); | |
a3437870 | 2423 | |
8a213460 NA |
2424 | static unsigned int cpuset_mems_nr(unsigned int *array) |
2425 | { | |
2426 | int node; | |
2427 | unsigned int nr = 0; | |
2428 | ||
2429 | for_each_node_mask(node, cpuset_current_mems_allowed) | |
2430 | nr += array[node]; | |
2431 | ||
2432 | return nr; | |
2433 | } | |
2434 | ||
2435 | #ifdef CONFIG_SYSCTL | |
06808b08 LS |
2436 | static int hugetlb_sysctl_handler_common(bool obey_mempolicy, |
2437 | struct ctl_table *table, int write, | |
2438 | void __user *buffer, size_t *length, loff_t *ppos) | |
1da177e4 | 2439 | { |
e5ff2159 | 2440 | struct hstate *h = &default_hstate; |
238d3c13 | 2441 | unsigned long tmp = h->max_huge_pages; |
08d4a246 | 2442 | int ret; |
e5ff2159 | 2443 | |
457c1b27 NA |
2444 | if (!hugepages_supported()) |
2445 | return -ENOTSUPP; | |
2446 | ||
e5ff2159 AK |
2447 | table->data = &tmp; |
2448 | table->maxlen = sizeof(unsigned long); | |
08d4a246 MH |
2449 | ret = proc_doulongvec_minmax(table, write, buffer, length, ppos); |
2450 | if (ret) | |
2451 | goto out; | |
e5ff2159 | 2452 | |
238d3c13 DR |
2453 | if (write) |
2454 | ret = __nr_hugepages_store_common(obey_mempolicy, h, | |
2455 | NUMA_NO_NODE, tmp, *length); | |
08d4a246 MH |
2456 | out: |
2457 | return ret; | |
1da177e4 | 2458 | } |
396faf03 | 2459 | |
06808b08 LS |
2460 | int hugetlb_sysctl_handler(struct ctl_table *table, int write, |
2461 | void __user *buffer, size_t *length, loff_t *ppos) | |
2462 | { | |
2463 | ||
2464 | return hugetlb_sysctl_handler_common(false, table, write, | |
2465 | buffer, length, ppos); | |
2466 | } | |
2467 | ||
2468 | #ifdef CONFIG_NUMA | |
2469 | int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write, | |
2470 | void __user *buffer, size_t *length, loff_t *ppos) | |
2471 | { | |
2472 | return hugetlb_sysctl_handler_common(true, table, write, | |
2473 | buffer, length, ppos); | |
2474 | } | |
2475 | #endif /* CONFIG_NUMA */ | |
2476 | ||
a3d0c6aa | 2477 | int hugetlb_overcommit_handler(struct ctl_table *table, int write, |
8d65af78 | 2478 | void __user *buffer, |
a3d0c6aa NA |
2479 | size_t *length, loff_t *ppos) |
2480 | { | |
a5516438 | 2481 | struct hstate *h = &default_hstate; |
e5ff2159 | 2482 | unsigned long tmp; |
08d4a246 | 2483 | int ret; |
e5ff2159 | 2484 | |
457c1b27 NA |
2485 | if (!hugepages_supported()) |
2486 | return -ENOTSUPP; | |
2487 | ||
c033a93c | 2488 | tmp = h->nr_overcommit_huge_pages; |
e5ff2159 | 2489 | |
bae7f4ae | 2490 | if (write && hstate_is_gigantic(h)) |
adbe8726 EM |
2491 | return -EINVAL; |
2492 | ||
e5ff2159 AK |
2493 | table->data = &tmp; |
2494 | table->maxlen = sizeof(unsigned long); | |
08d4a246 MH |
2495 | ret = proc_doulongvec_minmax(table, write, buffer, length, ppos); |
2496 | if (ret) | |
2497 | goto out; | |
e5ff2159 AK |
2498 | |
2499 | if (write) { | |
2500 | spin_lock(&hugetlb_lock); | |
2501 | h->nr_overcommit_huge_pages = tmp; | |
2502 | spin_unlock(&hugetlb_lock); | |
2503 | } | |
08d4a246 MH |
2504 | out: |
2505 | return ret; | |
a3d0c6aa NA |
2506 | } |
2507 | ||
1da177e4 LT |
2508 | #endif /* CONFIG_SYSCTL */ |
2509 | ||
e1759c21 | 2510 | void hugetlb_report_meminfo(struct seq_file *m) |
1da177e4 | 2511 | { |
a5516438 | 2512 | struct hstate *h = &default_hstate; |
457c1b27 NA |
2513 | if (!hugepages_supported()) |
2514 | return; | |
e1759c21 | 2515 | seq_printf(m, |
4f98a2fe RR |
2516 | "HugePages_Total: %5lu\n" |
2517 | "HugePages_Free: %5lu\n" | |
2518 | "HugePages_Rsvd: %5lu\n" | |
2519 | "HugePages_Surp: %5lu\n" | |
2520 | "Hugepagesize: %8lu kB\n", | |
a5516438 AK |
2521 | h->nr_huge_pages, |
2522 | h->free_huge_pages, | |
2523 | h->resv_huge_pages, | |
2524 | h->surplus_huge_pages, | |
2525 | 1UL << (huge_page_order(h) + PAGE_SHIFT - 10)); | |
1da177e4 LT |
2526 | } |
2527 | ||
2528 | int hugetlb_report_node_meminfo(int nid, char *buf) | |
2529 | { | |
a5516438 | 2530 | struct hstate *h = &default_hstate; |
457c1b27 NA |
2531 | if (!hugepages_supported()) |
2532 | return 0; | |
1da177e4 LT |
2533 | return sprintf(buf, |
2534 | "Node %d HugePages_Total: %5u\n" | |
a1de0919 NA |
2535 | "Node %d HugePages_Free: %5u\n" |
2536 | "Node %d HugePages_Surp: %5u\n", | |
a5516438 AK |
2537 | nid, h->nr_huge_pages_node[nid], |
2538 | nid, h->free_huge_pages_node[nid], | |
2539 | nid, h->surplus_huge_pages_node[nid]); | |
1da177e4 LT |
2540 | } |
2541 | ||
949f7ec5 DR |
2542 | void hugetlb_show_meminfo(void) |
2543 | { | |
2544 | struct hstate *h; | |
2545 | int nid; | |
2546 | ||
457c1b27 NA |
2547 | if (!hugepages_supported()) |
2548 | return; | |
2549 | ||
949f7ec5 DR |
2550 | for_each_node_state(nid, N_MEMORY) |
2551 | for_each_hstate(h) | |
2552 | pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n", | |
2553 | nid, | |
2554 | h->nr_huge_pages_node[nid], | |
2555 | h->free_huge_pages_node[nid], | |
2556 | h->surplus_huge_pages_node[nid], | |
2557 | 1UL << (huge_page_order(h) + PAGE_SHIFT - 10)); | |
2558 | } | |
2559 | ||
1da177e4 LT |
2560 | /* Return the number pages of memory we physically have, in PAGE_SIZE units. */ |
2561 | unsigned long hugetlb_total_pages(void) | |
2562 | { | |
d0028588 WL |
2563 | struct hstate *h; |
2564 | unsigned long nr_total_pages = 0; | |
2565 | ||
2566 | for_each_hstate(h) | |
2567 | nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h); | |
2568 | return nr_total_pages; | |
1da177e4 | 2569 | } |
1da177e4 | 2570 | |
a5516438 | 2571 | static int hugetlb_acct_memory(struct hstate *h, long delta) |
fc1b8a73 MG |
2572 | { |
2573 | int ret = -ENOMEM; | |
2574 | ||
2575 | spin_lock(&hugetlb_lock); | |
2576 | /* | |
2577 | * When cpuset is configured, it breaks the strict hugetlb page | |
2578 | * reservation as the accounting is done on a global variable. Such | |
2579 | * reservation is completely rubbish in the presence of cpuset because | |
2580 | * the reservation is not checked against page availability for the | |
2581 | * current cpuset. Application can still potentially OOM'ed by kernel | |
2582 | * with lack of free htlb page in cpuset that the task is in. | |
2583 | * Attempt to enforce strict accounting with cpuset is almost | |
2584 | * impossible (or too ugly) because cpuset is too fluid that | |
2585 | * task or memory node can be dynamically moved between cpusets. | |
2586 | * | |
2587 | * The change of semantics for shared hugetlb mapping with cpuset is | |
2588 | * undesirable. However, in order to preserve some of the semantics, | |
2589 | * we fall back to check against current free page availability as | |
2590 | * a best attempt and hopefully to minimize the impact of changing | |
2591 | * semantics that cpuset has. | |
2592 | */ | |
2593 | if (delta > 0) { | |
a5516438 | 2594 | if (gather_surplus_pages(h, delta) < 0) |
fc1b8a73 MG |
2595 | goto out; |
2596 | ||
a5516438 AK |
2597 | if (delta > cpuset_mems_nr(h->free_huge_pages_node)) { |
2598 | return_unused_surplus_pages(h, delta); | |
fc1b8a73 MG |
2599 | goto out; |
2600 | } | |
2601 | } | |
2602 | ||
2603 | ret = 0; | |
2604 | if (delta < 0) | |
a5516438 | 2605 | return_unused_surplus_pages(h, (unsigned long) -delta); |
fc1b8a73 MG |
2606 | |
2607 | out: | |
2608 | spin_unlock(&hugetlb_lock); | |
2609 | return ret; | |
2610 | } | |
2611 | ||
84afd99b AW |
2612 | static void hugetlb_vm_op_open(struct vm_area_struct *vma) |
2613 | { | |
f522c3ac | 2614 | struct resv_map *resv = vma_resv_map(vma); |
84afd99b AW |
2615 | |
2616 | /* | |
2617 | * This new VMA should share its siblings reservation map if present. | |
2618 | * The VMA will only ever have a valid reservation map pointer where | |
2619 | * it is being copied for another still existing VMA. As that VMA | |
25985edc | 2620 | * has a reference to the reservation map it cannot disappear until |
84afd99b AW |
2621 | * after this open call completes. It is therefore safe to take a |
2622 | * new reference here without additional locking. | |
2623 | */ | |
4e35f483 | 2624 | if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) |
f522c3ac | 2625 | kref_get(&resv->refs); |
84afd99b AW |
2626 | } |
2627 | ||
a1e78772 MG |
2628 | static void hugetlb_vm_op_close(struct vm_area_struct *vma) |
2629 | { | |
a5516438 | 2630 | struct hstate *h = hstate_vma(vma); |
f522c3ac | 2631 | struct resv_map *resv = vma_resv_map(vma); |
90481622 | 2632 | struct hugepage_subpool *spool = subpool_vma(vma); |
4e35f483 | 2633 | unsigned long reserve, start, end; |
1c5ecae3 | 2634 | long gbl_reserve; |
84afd99b | 2635 | |
4e35f483 JK |
2636 | if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER)) |
2637 | return; | |
84afd99b | 2638 | |
4e35f483 JK |
2639 | start = vma_hugecache_offset(h, vma, vma->vm_start); |
2640 | end = vma_hugecache_offset(h, vma, vma->vm_end); | |
84afd99b | 2641 | |
4e35f483 | 2642 | reserve = (end - start) - region_count(resv, start, end); |
84afd99b | 2643 | |
4e35f483 JK |
2644 | kref_put(&resv->refs, resv_map_release); |
2645 | ||
2646 | if (reserve) { | |
1c5ecae3 MK |
2647 | /* |
2648 | * Decrement reserve counts. The global reserve count may be | |
2649 | * adjusted if the subpool has a minimum size. | |
2650 | */ | |
2651 | gbl_reserve = hugepage_subpool_put_pages(spool, reserve); | |
2652 | hugetlb_acct_memory(h, -gbl_reserve); | |
84afd99b | 2653 | } |
a1e78772 MG |
2654 | } |
2655 | ||
1da177e4 LT |
2656 | /* |
2657 | * We cannot handle pagefaults against hugetlb pages at all. They cause | |
2658 | * handle_mm_fault() to try to instantiate regular-sized pages in the | |
2659 | * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get | |
2660 | * this far. | |
2661 | */ | |
d0217ac0 | 2662 | static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf) |
1da177e4 LT |
2663 | { |
2664 | BUG(); | |
d0217ac0 | 2665 | return 0; |
1da177e4 LT |
2666 | } |
2667 | ||
f0f37e2f | 2668 | const struct vm_operations_struct hugetlb_vm_ops = { |
d0217ac0 | 2669 | .fault = hugetlb_vm_op_fault, |
84afd99b | 2670 | .open = hugetlb_vm_op_open, |
a1e78772 | 2671 | .close = hugetlb_vm_op_close, |
1da177e4 LT |
2672 | }; |
2673 | ||
1e8f889b DG |
2674 | static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page, |
2675 | int writable) | |
63551ae0 DG |
2676 | { |
2677 | pte_t entry; | |
2678 | ||
1e8f889b | 2679 | if (writable) { |
106c992a GS |
2680 | entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page, |
2681 | vma->vm_page_prot))); | |
63551ae0 | 2682 | } else { |
106c992a GS |
2683 | entry = huge_pte_wrprotect(mk_huge_pte(page, |
2684 | vma->vm_page_prot)); | |
63551ae0 DG |
2685 | } |
2686 | entry = pte_mkyoung(entry); | |
2687 | entry = pte_mkhuge(entry); | |
d9ed9faa | 2688 | entry = arch_make_huge_pte(entry, vma, page, writable); |
63551ae0 DG |
2689 | |
2690 | return entry; | |
2691 | } | |
2692 | ||
1e8f889b DG |
2693 | static void set_huge_ptep_writable(struct vm_area_struct *vma, |
2694 | unsigned long address, pte_t *ptep) | |
2695 | { | |
2696 | pte_t entry; | |
2697 | ||
106c992a | 2698 | entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep))); |
32f84528 | 2699 | if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) |
4b3073e1 | 2700 | update_mmu_cache(vma, address, ptep); |
1e8f889b DG |
2701 | } |
2702 | ||
4a705fef NH |
2703 | static int is_hugetlb_entry_migration(pte_t pte) |
2704 | { | |
2705 | swp_entry_t swp; | |
2706 | ||
2707 | if (huge_pte_none(pte) || pte_present(pte)) | |
2708 | return 0; | |
2709 | swp = pte_to_swp_entry(pte); | |
2710 | if (non_swap_entry(swp) && is_migration_entry(swp)) | |
2711 | return 1; | |
2712 | else | |
2713 | return 0; | |
2714 | } | |
2715 | ||
2716 | static int is_hugetlb_entry_hwpoisoned(pte_t pte) | |
2717 | { | |
2718 | swp_entry_t swp; | |
2719 | ||
2720 | if (huge_pte_none(pte) || pte_present(pte)) | |
2721 | return 0; | |
2722 | swp = pte_to_swp_entry(pte); | |
2723 | if (non_swap_entry(swp) && is_hwpoison_entry(swp)) | |
2724 | return 1; | |
2725 | else | |
2726 | return 0; | |
2727 | } | |
1e8f889b | 2728 | |
63551ae0 DG |
2729 | int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src, |
2730 | struct vm_area_struct *vma) | |
2731 | { | |
2732 | pte_t *src_pte, *dst_pte, entry; | |
2733 | struct page *ptepage; | |
1c59827d | 2734 | unsigned long addr; |
1e8f889b | 2735 | int cow; |
a5516438 AK |
2736 | struct hstate *h = hstate_vma(vma); |
2737 | unsigned long sz = huge_page_size(h); | |
e8569dd2 AS |
2738 | unsigned long mmun_start; /* For mmu_notifiers */ |
2739 | unsigned long mmun_end; /* For mmu_notifiers */ | |
2740 | int ret = 0; | |
1e8f889b DG |
2741 | |
2742 | cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE; | |
63551ae0 | 2743 | |
e8569dd2 AS |
2744 | mmun_start = vma->vm_start; |
2745 | mmun_end = vma->vm_end; | |
2746 | if (cow) | |
2747 | mmu_notifier_invalidate_range_start(src, mmun_start, mmun_end); | |
2748 | ||
a5516438 | 2749 | for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) { |
cb900f41 | 2750 | spinlock_t *src_ptl, *dst_ptl; |
c74df32c HD |
2751 | src_pte = huge_pte_offset(src, addr); |
2752 | if (!src_pte) | |
2753 | continue; | |
a5516438 | 2754 | dst_pte = huge_pte_alloc(dst, addr, sz); |
e8569dd2 AS |
2755 | if (!dst_pte) { |
2756 | ret = -ENOMEM; | |
2757 | break; | |
2758 | } | |
c5c99429 LW |
2759 | |
2760 | /* If the pagetables are shared don't copy or take references */ | |
2761 | if (dst_pte == src_pte) | |
2762 | continue; | |
2763 | ||
cb900f41 KS |
2764 | dst_ptl = huge_pte_lock(h, dst, dst_pte); |
2765 | src_ptl = huge_pte_lockptr(h, src, src_pte); | |
2766 | spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); | |
4a705fef NH |
2767 | entry = huge_ptep_get(src_pte); |
2768 | if (huge_pte_none(entry)) { /* skip none entry */ | |
2769 | ; | |
2770 | } else if (unlikely(is_hugetlb_entry_migration(entry) || | |
2771 | is_hugetlb_entry_hwpoisoned(entry))) { | |
2772 | swp_entry_t swp_entry = pte_to_swp_entry(entry); | |
2773 | ||
2774 | if (is_write_migration_entry(swp_entry) && cow) { | |
2775 | /* | |
2776 | * COW mappings require pages in both | |
2777 | * parent and child to be set to read. | |
2778 | */ | |
2779 | make_migration_entry_read(&swp_entry); | |
2780 | entry = swp_entry_to_pte(swp_entry); | |
2781 | set_huge_pte_at(src, addr, src_pte, entry); | |
2782 | } | |
2783 | set_huge_pte_at(dst, addr, dst_pte, entry); | |
2784 | } else { | |
34ee645e | 2785 | if (cow) { |
7f2e9525 | 2786 | huge_ptep_set_wrprotect(src, addr, src_pte); |
34ee645e JR |
2787 | mmu_notifier_invalidate_range(src, mmun_start, |
2788 | mmun_end); | |
2789 | } | |
0253d634 | 2790 | entry = huge_ptep_get(src_pte); |
1c59827d HD |
2791 | ptepage = pte_page(entry); |
2792 | get_page(ptepage); | |
0fe6e20b | 2793 | page_dup_rmap(ptepage); |
1c59827d HD |
2794 | set_huge_pte_at(dst, addr, dst_pte, entry); |
2795 | } | |
cb900f41 KS |
2796 | spin_unlock(src_ptl); |
2797 | spin_unlock(dst_ptl); | |
63551ae0 | 2798 | } |
63551ae0 | 2799 | |
e8569dd2 AS |
2800 | if (cow) |
2801 | mmu_notifier_invalidate_range_end(src, mmun_start, mmun_end); | |
2802 | ||
2803 | return ret; | |
63551ae0 DG |
2804 | } |
2805 | ||
24669e58 AK |
2806 | void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma, |
2807 | unsigned long start, unsigned long end, | |
2808 | struct page *ref_page) | |
63551ae0 | 2809 | { |
24669e58 | 2810 | int force_flush = 0; |
63551ae0 DG |
2811 | struct mm_struct *mm = vma->vm_mm; |
2812 | unsigned long address; | |
c7546f8f | 2813 | pte_t *ptep; |
63551ae0 | 2814 | pte_t pte; |
cb900f41 | 2815 | spinlock_t *ptl; |
63551ae0 | 2816 | struct page *page; |
a5516438 AK |
2817 | struct hstate *h = hstate_vma(vma); |
2818 | unsigned long sz = huge_page_size(h); | |
2ec74c3e SG |
2819 | const unsigned long mmun_start = start; /* For mmu_notifiers */ |
2820 | const unsigned long mmun_end = end; /* For mmu_notifiers */ | |
a5516438 | 2821 | |
63551ae0 | 2822 | WARN_ON(!is_vm_hugetlb_page(vma)); |
a5516438 AK |
2823 | BUG_ON(start & ~huge_page_mask(h)); |
2824 | BUG_ON(end & ~huge_page_mask(h)); | |
63551ae0 | 2825 | |
24669e58 | 2826 | tlb_start_vma(tlb, vma); |
2ec74c3e | 2827 | mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); |
569f48b8 | 2828 | address = start; |
24669e58 | 2829 | again: |
569f48b8 | 2830 | for (; address < end; address += sz) { |
c7546f8f | 2831 | ptep = huge_pte_offset(mm, address); |
4c887265 | 2832 | if (!ptep) |
c7546f8f DG |
2833 | continue; |
2834 | ||
cb900f41 | 2835 | ptl = huge_pte_lock(h, mm, ptep); |
39dde65c | 2836 | if (huge_pmd_unshare(mm, &address, ptep)) |
cb900f41 | 2837 | goto unlock; |
39dde65c | 2838 | |
6629326b HD |
2839 | pte = huge_ptep_get(ptep); |
2840 | if (huge_pte_none(pte)) | |
cb900f41 | 2841 | goto unlock; |
6629326b HD |
2842 | |
2843 | /* | |
9fbc1f63 NH |
2844 | * Migrating hugepage or HWPoisoned hugepage is already |
2845 | * unmapped and its refcount is dropped, so just clear pte here. | |
6629326b | 2846 | */ |
9fbc1f63 | 2847 | if (unlikely(!pte_present(pte))) { |
106c992a | 2848 | huge_pte_clear(mm, address, ptep); |
cb900f41 | 2849 | goto unlock; |
8c4894c6 | 2850 | } |
6629326b HD |
2851 | |
2852 | page = pte_page(pte); | |
04f2cbe3 MG |
2853 | /* |
2854 | * If a reference page is supplied, it is because a specific | |
2855 | * page is being unmapped, not a range. Ensure the page we | |
2856 | * are about to unmap is the actual page of interest. | |
2857 | */ | |
2858 | if (ref_page) { | |
04f2cbe3 | 2859 | if (page != ref_page) |
cb900f41 | 2860 | goto unlock; |
04f2cbe3 MG |
2861 | |
2862 | /* | |
2863 | * Mark the VMA as having unmapped its page so that | |
2864 | * future faults in this VMA will fail rather than | |
2865 | * looking like data was lost | |
2866 | */ | |
2867 | set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED); | |
2868 | } | |
2869 | ||
c7546f8f | 2870 | pte = huge_ptep_get_and_clear(mm, address, ptep); |
24669e58 | 2871 | tlb_remove_tlb_entry(tlb, ptep, address); |
106c992a | 2872 | if (huge_pte_dirty(pte)) |
6649a386 | 2873 | set_page_dirty(page); |
9e81130b | 2874 | |
24669e58 AK |
2875 | page_remove_rmap(page); |
2876 | force_flush = !__tlb_remove_page(tlb, page); | |
cb900f41 | 2877 | if (force_flush) { |
569f48b8 | 2878 | address += sz; |
cb900f41 | 2879 | spin_unlock(ptl); |
24669e58 | 2880 | break; |
cb900f41 | 2881 | } |
9e81130b | 2882 | /* Bail out after unmapping reference page if supplied */ |
cb900f41 KS |
2883 | if (ref_page) { |
2884 | spin_unlock(ptl); | |
9e81130b | 2885 | break; |
cb900f41 KS |
2886 | } |
2887 | unlock: | |
2888 | spin_unlock(ptl); | |
63551ae0 | 2889 | } |
24669e58 AK |
2890 | /* |
2891 | * mmu_gather ran out of room to batch pages, we break out of | |
2892 | * the PTE lock to avoid doing the potential expensive TLB invalidate | |
2893 | * and page-free while holding it. | |
2894 | */ | |
2895 | if (force_flush) { | |
2896 | force_flush = 0; | |
2897 | tlb_flush_mmu(tlb); | |
2898 | if (address < end && !ref_page) | |
2899 | goto again; | |
fe1668ae | 2900 | } |
2ec74c3e | 2901 | mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); |
24669e58 | 2902 | tlb_end_vma(tlb, vma); |
1da177e4 | 2903 | } |
63551ae0 | 2904 | |
d833352a MG |
2905 | void __unmap_hugepage_range_final(struct mmu_gather *tlb, |
2906 | struct vm_area_struct *vma, unsigned long start, | |
2907 | unsigned long end, struct page *ref_page) | |
2908 | { | |
2909 | __unmap_hugepage_range(tlb, vma, start, end, ref_page); | |
2910 | ||
2911 | /* | |
2912 | * Clear this flag so that x86's huge_pmd_share page_table_shareable | |
2913 | * test will fail on a vma being torn down, and not grab a page table | |
2914 | * on its way out. We're lucky that the flag has such an appropriate | |
2915 | * name, and can in fact be safely cleared here. We could clear it | |
2916 | * before the __unmap_hugepage_range above, but all that's necessary | |
c8c06efa | 2917 | * is to clear it before releasing the i_mmap_rwsem. This works |
d833352a | 2918 | * because in the context this is called, the VMA is about to be |
c8c06efa | 2919 | * destroyed and the i_mmap_rwsem is held. |
d833352a MG |
2920 | */ |
2921 | vma->vm_flags &= ~VM_MAYSHARE; | |
2922 | } | |
2923 | ||
502717f4 | 2924 | void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start, |
04f2cbe3 | 2925 | unsigned long end, struct page *ref_page) |
502717f4 | 2926 | { |
24669e58 AK |
2927 | struct mm_struct *mm; |
2928 | struct mmu_gather tlb; | |
2929 | ||
2930 | mm = vma->vm_mm; | |
2931 | ||
2b047252 | 2932 | tlb_gather_mmu(&tlb, mm, start, end); |
24669e58 AK |
2933 | __unmap_hugepage_range(&tlb, vma, start, end, ref_page); |
2934 | tlb_finish_mmu(&tlb, start, end); | |
502717f4 CK |
2935 | } |
2936 | ||
04f2cbe3 MG |
2937 | /* |
2938 | * This is called when the original mapper is failing to COW a MAP_PRIVATE | |
2939 | * mappping it owns the reserve page for. The intention is to unmap the page | |
2940 | * from other VMAs and let the children be SIGKILLed if they are faulting the | |
2941 | * same region. | |
2942 | */ | |
2f4612af DB |
2943 | static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma, |
2944 | struct page *page, unsigned long address) | |
04f2cbe3 | 2945 | { |
7526674d | 2946 | struct hstate *h = hstate_vma(vma); |
04f2cbe3 MG |
2947 | struct vm_area_struct *iter_vma; |
2948 | struct address_space *mapping; | |
04f2cbe3 MG |
2949 | pgoff_t pgoff; |
2950 | ||
2951 | /* | |
2952 | * vm_pgoff is in PAGE_SIZE units, hence the different calculation | |
2953 | * from page cache lookup which is in HPAGE_SIZE units. | |
2954 | */ | |
7526674d | 2955 | address = address & huge_page_mask(h); |
36e4f20a MH |
2956 | pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) + |
2957 | vma->vm_pgoff; | |
496ad9aa | 2958 | mapping = file_inode(vma->vm_file)->i_mapping; |
04f2cbe3 | 2959 | |
4eb2b1dc MG |
2960 | /* |
2961 | * Take the mapping lock for the duration of the table walk. As | |
2962 | * this mapping should be shared between all the VMAs, | |
2963 | * __unmap_hugepage_range() is called as the lock is already held | |
2964 | */ | |
83cde9e8 | 2965 | i_mmap_lock_write(mapping); |
6b2dbba8 | 2966 | vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) { |
04f2cbe3 MG |
2967 | /* Do not unmap the current VMA */ |
2968 | if (iter_vma == vma) | |
2969 | continue; | |
2970 | ||
2971 | /* | |
2972 | * Unmap the page from other VMAs without their own reserves. | |
2973 | * They get marked to be SIGKILLed if they fault in these | |
2974 | * areas. This is because a future no-page fault on this VMA | |
2975 | * could insert a zeroed page instead of the data existing | |
2976 | * from the time of fork. This would look like data corruption | |
2977 | */ | |
2978 | if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER)) | |
24669e58 AK |
2979 | unmap_hugepage_range(iter_vma, address, |
2980 | address + huge_page_size(h), page); | |
04f2cbe3 | 2981 | } |
83cde9e8 | 2982 | i_mmap_unlock_write(mapping); |
04f2cbe3 MG |
2983 | } |
2984 | ||
0fe6e20b NH |
2985 | /* |
2986 | * Hugetlb_cow() should be called with page lock of the original hugepage held. | |
ef009b25 MH |
2987 | * Called with hugetlb_instantiation_mutex held and pte_page locked so we |
2988 | * cannot race with other handlers or page migration. | |
2989 | * Keep the pte_same checks anyway to make transition from the mutex easier. | |
0fe6e20b | 2990 | */ |
1e8f889b | 2991 | static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma, |
04f2cbe3 | 2992 | unsigned long address, pte_t *ptep, pte_t pte, |
cb900f41 | 2993 | struct page *pagecache_page, spinlock_t *ptl) |
1e8f889b | 2994 | { |
a5516438 | 2995 | struct hstate *h = hstate_vma(vma); |
1e8f889b | 2996 | struct page *old_page, *new_page; |
ad4404a2 | 2997 | int ret = 0, outside_reserve = 0; |
2ec74c3e SG |
2998 | unsigned long mmun_start; /* For mmu_notifiers */ |
2999 | unsigned long mmun_end; /* For mmu_notifiers */ | |
1e8f889b DG |
3000 | |
3001 | old_page = pte_page(pte); | |
3002 | ||
04f2cbe3 | 3003 | retry_avoidcopy: |
1e8f889b DG |
3004 | /* If no-one else is actually using this page, avoid the copy |
3005 | * and just make the page writable */ | |
37a2140d JK |
3006 | if (page_mapcount(old_page) == 1 && PageAnon(old_page)) { |
3007 | page_move_anon_rmap(old_page, vma, address); | |
1e8f889b | 3008 | set_huge_ptep_writable(vma, address, ptep); |
83c54070 | 3009 | return 0; |
1e8f889b DG |
3010 | } |
3011 | ||
04f2cbe3 MG |
3012 | /* |
3013 | * If the process that created a MAP_PRIVATE mapping is about to | |
3014 | * perform a COW due to a shared page count, attempt to satisfy | |
3015 | * the allocation without using the existing reserves. The pagecache | |
3016 | * page is used to determine if the reserve at this address was | |
3017 | * consumed or not. If reserves were used, a partial faulted mapping | |
3018 | * at the time of fork() could consume its reserves on COW instead | |
3019 | * of the full address range. | |
3020 | */ | |
5944d011 | 3021 | if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && |
04f2cbe3 MG |
3022 | old_page != pagecache_page) |
3023 | outside_reserve = 1; | |
3024 | ||
1e8f889b | 3025 | page_cache_get(old_page); |
b76c8cfb | 3026 | |
ad4404a2 DB |
3027 | /* |
3028 | * Drop page table lock as buddy allocator may be called. It will | |
3029 | * be acquired again before returning to the caller, as expected. | |
3030 | */ | |
cb900f41 | 3031 | spin_unlock(ptl); |
04f2cbe3 | 3032 | new_page = alloc_huge_page(vma, address, outside_reserve); |
1e8f889b | 3033 | |
2fc39cec | 3034 | if (IS_ERR(new_page)) { |
04f2cbe3 MG |
3035 | /* |
3036 | * If a process owning a MAP_PRIVATE mapping fails to COW, | |
3037 | * it is due to references held by a child and an insufficient | |
3038 | * huge page pool. To guarantee the original mappers | |
3039 | * reliability, unmap the page from child processes. The child | |
3040 | * may get SIGKILLed if it later faults. | |
3041 | */ | |
3042 | if (outside_reserve) { | |
ad4404a2 | 3043 | page_cache_release(old_page); |
04f2cbe3 | 3044 | BUG_ON(huge_pte_none(pte)); |
2f4612af DB |
3045 | unmap_ref_private(mm, vma, old_page, address); |
3046 | BUG_ON(huge_pte_none(pte)); | |
3047 | spin_lock(ptl); | |
3048 | ptep = huge_pte_offset(mm, address & huge_page_mask(h)); | |
3049 | if (likely(ptep && | |
3050 | pte_same(huge_ptep_get(ptep), pte))) | |
3051 | goto retry_avoidcopy; | |
3052 | /* | |
3053 | * race occurs while re-acquiring page table | |
3054 | * lock, and our job is done. | |
3055 | */ | |
3056 | return 0; | |
04f2cbe3 MG |
3057 | } |
3058 | ||
ad4404a2 DB |
3059 | ret = (PTR_ERR(new_page) == -ENOMEM) ? |
3060 | VM_FAULT_OOM : VM_FAULT_SIGBUS; | |
3061 | goto out_release_old; | |
1e8f889b DG |
3062 | } |
3063 | ||
0fe6e20b NH |
3064 | /* |
3065 | * When the original hugepage is shared one, it does not have | |
3066 | * anon_vma prepared. | |
3067 | */ | |
44e2aa93 | 3068 | if (unlikely(anon_vma_prepare(vma))) { |
ad4404a2 DB |
3069 | ret = VM_FAULT_OOM; |
3070 | goto out_release_all; | |
44e2aa93 | 3071 | } |
0fe6e20b | 3072 | |
47ad8475 AA |
3073 | copy_user_huge_page(new_page, old_page, address, vma, |
3074 | pages_per_huge_page(h)); | |
0ed361de | 3075 | __SetPageUptodate(new_page); |
bcc54222 | 3076 | set_page_huge_active(new_page); |
1e8f889b | 3077 | |
2ec74c3e SG |
3078 | mmun_start = address & huge_page_mask(h); |
3079 | mmun_end = mmun_start + huge_page_size(h); | |
3080 | mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); | |
ad4404a2 | 3081 | |
b76c8cfb | 3082 | /* |
cb900f41 | 3083 | * Retake the page table lock to check for racing updates |
b76c8cfb LW |
3084 | * before the page tables are altered |
3085 | */ | |
cb900f41 | 3086 | spin_lock(ptl); |
a5516438 | 3087 | ptep = huge_pte_offset(mm, address & huge_page_mask(h)); |
a9af0c5d | 3088 | if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) { |
07443a85 JK |
3089 | ClearPagePrivate(new_page); |
3090 | ||
1e8f889b | 3091 | /* Break COW */ |
8fe627ec | 3092 | huge_ptep_clear_flush(vma, address, ptep); |
34ee645e | 3093 | mmu_notifier_invalidate_range(mm, mmun_start, mmun_end); |
1e8f889b DG |
3094 | set_huge_pte_at(mm, address, ptep, |
3095 | make_huge_pte(vma, new_page, 1)); | |
0fe6e20b | 3096 | page_remove_rmap(old_page); |
cd67f0d2 | 3097 | hugepage_add_new_anon_rmap(new_page, vma, address); |
1e8f889b DG |
3098 | /* Make the old page be freed below */ |
3099 | new_page = old_page; | |
3100 | } | |
cb900f41 | 3101 | spin_unlock(ptl); |
2ec74c3e | 3102 | mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); |
ad4404a2 | 3103 | out_release_all: |
1e8f889b | 3104 | page_cache_release(new_page); |
ad4404a2 | 3105 | out_release_old: |
1e8f889b | 3106 | page_cache_release(old_page); |
8312034f | 3107 | |
ad4404a2 DB |
3108 | spin_lock(ptl); /* Caller expects lock to be held */ |
3109 | return ret; | |
1e8f889b DG |
3110 | } |
3111 | ||
04f2cbe3 | 3112 | /* Return the pagecache page at a given address within a VMA */ |
a5516438 AK |
3113 | static struct page *hugetlbfs_pagecache_page(struct hstate *h, |
3114 | struct vm_area_struct *vma, unsigned long address) | |
04f2cbe3 MG |
3115 | { |
3116 | struct address_space *mapping; | |
e7c4b0bf | 3117 | pgoff_t idx; |
04f2cbe3 MG |
3118 | |
3119 | mapping = vma->vm_file->f_mapping; | |
a5516438 | 3120 | idx = vma_hugecache_offset(h, vma, address); |
04f2cbe3 MG |
3121 | |
3122 | return find_lock_page(mapping, idx); | |
3123 | } | |
3124 | ||
3ae77f43 HD |
3125 | /* |
3126 | * Return whether there is a pagecache page to back given address within VMA. | |
3127 | * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page. | |
3128 | */ | |
3129 | static bool hugetlbfs_pagecache_present(struct hstate *h, | |
2a15efc9 HD |
3130 | struct vm_area_struct *vma, unsigned long address) |
3131 | { | |
3132 | struct address_space *mapping; | |
3133 | pgoff_t idx; | |
3134 | struct page *page; | |
3135 | ||
3136 | mapping = vma->vm_file->f_mapping; | |
3137 | idx = vma_hugecache_offset(h, vma, address); | |
3138 | ||
3139 | page = find_get_page(mapping, idx); | |
3140 | if (page) | |
3141 | put_page(page); | |
3142 | return page != NULL; | |
3143 | } | |
3144 | ||
a1ed3dda | 3145 | static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma, |
8382d914 DB |
3146 | struct address_space *mapping, pgoff_t idx, |
3147 | unsigned long address, pte_t *ptep, unsigned int flags) | |
ac9b9c66 | 3148 | { |
a5516438 | 3149 | struct hstate *h = hstate_vma(vma); |
ac9b9c66 | 3150 | int ret = VM_FAULT_SIGBUS; |
409eb8c2 | 3151 | int anon_rmap = 0; |
4c887265 | 3152 | unsigned long size; |
4c887265 | 3153 | struct page *page; |
1e8f889b | 3154 | pte_t new_pte; |
cb900f41 | 3155 | spinlock_t *ptl; |
4c887265 | 3156 | |
04f2cbe3 MG |
3157 | /* |
3158 | * Currently, we are forced to kill the process in the event the | |
3159 | * original mapper has unmapped pages from the child due to a failed | |
25985edc | 3160 | * COW. Warn that such a situation has occurred as it may not be obvious |
04f2cbe3 MG |
3161 | */ |
3162 | if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) { | |
ffb22af5 AM |
3163 | pr_warning("PID %d killed due to inadequate hugepage pool\n", |
3164 | current->pid); | |
04f2cbe3 MG |
3165 | return ret; |
3166 | } | |
3167 | ||
4c887265 AL |
3168 | /* |
3169 | * Use page lock to guard against racing truncation | |
3170 | * before we get page_table_lock. | |
3171 | */ | |
6bda666a CL |
3172 | retry: |
3173 | page = find_lock_page(mapping, idx); | |
3174 | if (!page) { | |
a5516438 | 3175 | size = i_size_read(mapping->host) >> huge_page_shift(h); |
ebed4bfc HD |
3176 | if (idx >= size) |
3177 | goto out; | |
04f2cbe3 | 3178 | page = alloc_huge_page(vma, address, 0); |
2fc39cec | 3179 | if (IS_ERR(page)) { |
76dcee75 AK |
3180 | ret = PTR_ERR(page); |
3181 | if (ret == -ENOMEM) | |
3182 | ret = VM_FAULT_OOM; | |
3183 | else | |
3184 | ret = VM_FAULT_SIGBUS; | |
6bda666a CL |
3185 | goto out; |
3186 | } | |
47ad8475 | 3187 | clear_huge_page(page, address, pages_per_huge_page(h)); |
0ed361de | 3188 | __SetPageUptodate(page); |
bcc54222 | 3189 | set_page_huge_active(page); |
ac9b9c66 | 3190 | |
f83a275d | 3191 | if (vma->vm_flags & VM_MAYSHARE) { |
6bda666a | 3192 | int err; |
45c682a6 | 3193 | struct inode *inode = mapping->host; |
6bda666a CL |
3194 | |
3195 | err = add_to_page_cache(page, mapping, idx, GFP_KERNEL); | |
3196 | if (err) { | |
3197 | put_page(page); | |
6bda666a CL |
3198 | if (err == -EEXIST) |
3199 | goto retry; | |
3200 | goto out; | |
3201 | } | |
07443a85 | 3202 | ClearPagePrivate(page); |
45c682a6 KC |
3203 | |
3204 | spin_lock(&inode->i_lock); | |
a5516438 | 3205 | inode->i_blocks += blocks_per_huge_page(h); |
45c682a6 | 3206 | spin_unlock(&inode->i_lock); |
23be7468 | 3207 | } else { |
6bda666a | 3208 | lock_page(page); |
0fe6e20b NH |
3209 | if (unlikely(anon_vma_prepare(vma))) { |
3210 | ret = VM_FAULT_OOM; | |
3211 | goto backout_unlocked; | |
3212 | } | |
409eb8c2 | 3213 | anon_rmap = 1; |
23be7468 | 3214 | } |
0fe6e20b | 3215 | } else { |
998b4382 NH |
3216 | /* |
3217 | * If memory error occurs between mmap() and fault, some process | |
3218 | * don't have hwpoisoned swap entry for errored virtual address. | |
3219 | * So we need to block hugepage fault by PG_hwpoison bit check. | |
3220 | */ | |
3221 | if (unlikely(PageHWPoison(page))) { | |
32f84528 | 3222 | ret = VM_FAULT_HWPOISON | |
972dc4de | 3223 | VM_FAULT_SET_HINDEX(hstate_index(h)); |
998b4382 NH |
3224 | goto backout_unlocked; |
3225 | } | |
6bda666a | 3226 | } |
1e8f889b | 3227 | |
57303d80 AW |
3228 | /* |
3229 | * If we are going to COW a private mapping later, we examine the | |
3230 | * pending reservations for this page now. This will ensure that | |
3231 | * any allocations necessary to record that reservation occur outside | |
3232 | * the spinlock. | |
3233 | */ | |
788c7df4 | 3234 | if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) |
2b26736c AW |
3235 | if (vma_needs_reservation(h, vma, address) < 0) { |
3236 | ret = VM_FAULT_OOM; | |
3237 | goto backout_unlocked; | |
3238 | } | |
57303d80 | 3239 | |
cb900f41 KS |
3240 | ptl = huge_pte_lockptr(h, mm, ptep); |
3241 | spin_lock(ptl); | |
a5516438 | 3242 | size = i_size_read(mapping->host) >> huge_page_shift(h); |
4c887265 AL |
3243 | if (idx >= size) |
3244 | goto backout; | |
3245 | ||
83c54070 | 3246 | ret = 0; |
7f2e9525 | 3247 | if (!huge_pte_none(huge_ptep_get(ptep))) |
4c887265 AL |
3248 | goto backout; |
3249 | ||
07443a85 JK |
3250 | if (anon_rmap) { |
3251 | ClearPagePrivate(page); | |
409eb8c2 | 3252 | hugepage_add_new_anon_rmap(page, vma, address); |
ac714904 | 3253 | } else |
409eb8c2 | 3254 | page_dup_rmap(page); |
1e8f889b DG |
3255 | new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE) |
3256 | && (vma->vm_flags & VM_SHARED))); | |
3257 | set_huge_pte_at(mm, address, ptep, new_pte); | |
3258 | ||
788c7df4 | 3259 | if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { |
1e8f889b | 3260 | /* Optimization, do the COW without a second fault */ |
cb900f41 | 3261 | ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page, ptl); |
1e8f889b DG |
3262 | } |
3263 | ||
cb900f41 | 3264 | spin_unlock(ptl); |
4c887265 AL |
3265 | unlock_page(page); |
3266 | out: | |
ac9b9c66 | 3267 | return ret; |
4c887265 AL |
3268 | |
3269 | backout: | |
cb900f41 | 3270 | spin_unlock(ptl); |
2b26736c | 3271 | backout_unlocked: |
4c887265 AL |
3272 | unlock_page(page); |
3273 | put_page(page); | |
3274 | goto out; | |
ac9b9c66 HD |
3275 | } |
3276 | ||
8382d914 DB |
3277 | #ifdef CONFIG_SMP |
3278 | static u32 fault_mutex_hash(struct hstate *h, struct mm_struct *mm, | |
3279 | struct vm_area_struct *vma, | |
3280 | struct address_space *mapping, | |
3281 | pgoff_t idx, unsigned long address) | |
3282 | { | |
3283 | unsigned long key[2]; | |
3284 | u32 hash; | |
3285 | ||
3286 | if (vma->vm_flags & VM_SHARED) { | |
3287 | key[0] = (unsigned long) mapping; | |
3288 | key[1] = idx; | |
3289 | } else { | |
3290 | key[0] = (unsigned long) mm; | |
3291 | key[1] = address >> huge_page_shift(h); | |
3292 | } | |
3293 | ||
3294 | hash = jhash2((u32 *)&key, sizeof(key)/sizeof(u32), 0); | |
3295 | ||
3296 | return hash & (num_fault_mutexes - 1); | |
3297 | } | |
3298 | #else | |
3299 | /* | |
3300 | * For uniprocesor systems we always use a single mutex, so just | |
3301 | * return 0 and avoid the hashing overhead. | |
3302 | */ | |
3303 | static u32 fault_mutex_hash(struct hstate *h, struct mm_struct *mm, | |
3304 | struct vm_area_struct *vma, | |
3305 | struct address_space *mapping, | |
3306 | pgoff_t idx, unsigned long address) | |
3307 | { | |
3308 | return 0; | |
3309 | } | |
3310 | #endif | |
3311 | ||
86e5216f | 3312 | int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma, |
788c7df4 | 3313 | unsigned long address, unsigned int flags) |
86e5216f | 3314 | { |
8382d914 | 3315 | pte_t *ptep, entry; |
cb900f41 | 3316 | spinlock_t *ptl; |
1e8f889b | 3317 | int ret; |
8382d914 DB |
3318 | u32 hash; |
3319 | pgoff_t idx; | |
0fe6e20b | 3320 | struct page *page = NULL; |
57303d80 | 3321 | struct page *pagecache_page = NULL; |
a5516438 | 3322 | struct hstate *h = hstate_vma(vma); |
8382d914 | 3323 | struct address_space *mapping; |
0f792cf9 | 3324 | int need_wait_lock = 0; |
86e5216f | 3325 | |
1e16a539 KH |
3326 | address &= huge_page_mask(h); |
3327 | ||
fd6a03ed NH |
3328 | ptep = huge_pte_offset(mm, address); |
3329 | if (ptep) { | |
3330 | entry = huge_ptep_get(ptep); | |
290408d4 | 3331 | if (unlikely(is_hugetlb_entry_migration(entry))) { |
cb900f41 | 3332 | migration_entry_wait_huge(vma, mm, ptep); |
290408d4 NH |
3333 | return 0; |
3334 | } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry))) | |
32f84528 | 3335 | return VM_FAULT_HWPOISON_LARGE | |
972dc4de | 3336 | VM_FAULT_SET_HINDEX(hstate_index(h)); |
fd6a03ed NH |
3337 | } |
3338 | ||
a5516438 | 3339 | ptep = huge_pte_alloc(mm, address, huge_page_size(h)); |
86e5216f AL |
3340 | if (!ptep) |
3341 | return VM_FAULT_OOM; | |
3342 | ||
8382d914 DB |
3343 | mapping = vma->vm_file->f_mapping; |
3344 | idx = vma_hugecache_offset(h, vma, address); | |
3345 | ||
3935baa9 DG |
3346 | /* |
3347 | * Serialize hugepage allocation and instantiation, so that we don't | |
3348 | * get spurious allocation failures if two CPUs race to instantiate | |
3349 | * the same page in the page cache. | |
3350 | */ | |
8382d914 DB |
3351 | hash = fault_mutex_hash(h, mm, vma, mapping, idx, address); |
3352 | mutex_lock(&htlb_fault_mutex_table[hash]); | |
3353 | ||
7f2e9525 GS |
3354 | entry = huge_ptep_get(ptep); |
3355 | if (huge_pte_none(entry)) { | |
8382d914 | 3356 | ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags); |
b4d1d99f | 3357 | goto out_mutex; |
3935baa9 | 3358 | } |
86e5216f | 3359 | |
83c54070 | 3360 | ret = 0; |
1e8f889b | 3361 | |
0f792cf9 NH |
3362 | /* |
3363 | * entry could be a migration/hwpoison entry at this point, so this | |
3364 | * check prevents the kernel from going below assuming that we have | |
3365 | * a active hugepage in pagecache. This goto expects the 2nd page fault, | |
3366 | * and is_hugetlb_entry_(migration|hwpoisoned) check will properly | |
3367 | * handle it. | |
3368 | */ | |
3369 | if (!pte_present(entry)) | |
3370 | goto out_mutex; | |
3371 | ||
57303d80 AW |
3372 | /* |
3373 | * If we are going to COW the mapping later, we examine the pending | |
3374 | * reservations for this page now. This will ensure that any | |
3375 | * allocations necessary to record that reservation occur outside the | |
3376 | * spinlock. For private mappings, we also lookup the pagecache | |
3377 | * page now as it is used to determine if a reservation has been | |
3378 | * consumed. | |
3379 | */ | |
106c992a | 3380 | if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) { |
2b26736c AW |
3381 | if (vma_needs_reservation(h, vma, address) < 0) { |
3382 | ret = VM_FAULT_OOM; | |
b4d1d99f | 3383 | goto out_mutex; |
2b26736c | 3384 | } |
57303d80 | 3385 | |
f83a275d | 3386 | if (!(vma->vm_flags & VM_MAYSHARE)) |
57303d80 AW |
3387 | pagecache_page = hugetlbfs_pagecache_page(h, |
3388 | vma, address); | |
3389 | } | |
3390 | ||
0f792cf9 NH |
3391 | ptl = huge_pte_lock(h, mm, ptep); |
3392 | ||
3393 | /* Check for a racing update before calling hugetlb_cow */ | |
3394 | if (unlikely(!pte_same(entry, huge_ptep_get(ptep)))) | |
3395 | goto out_ptl; | |
3396 | ||
56c9cfb1 NH |
3397 | /* |
3398 | * hugetlb_cow() requires page locks of pte_page(entry) and | |
3399 | * pagecache_page, so here we need take the former one | |
3400 | * when page != pagecache_page or !pagecache_page. | |
56c9cfb1 NH |
3401 | */ |
3402 | page = pte_page(entry); | |
3403 | if (page != pagecache_page) | |
0f792cf9 NH |
3404 | if (!trylock_page(page)) { |
3405 | need_wait_lock = 1; | |
3406 | goto out_ptl; | |
3407 | } | |
b4d1d99f | 3408 | |
0f792cf9 | 3409 | get_page(page); |
b4d1d99f | 3410 | |
788c7df4 | 3411 | if (flags & FAULT_FLAG_WRITE) { |
106c992a | 3412 | if (!huge_pte_write(entry)) { |
57303d80 | 3413 | ret = hugetlb_cow(mm, vma, address, ptep, entry, |
cb900f41 | 3414 | pagecache_page, ptl); |
0f792cf9 | 3415 | goto out_put_page; |
b4d1d99f | 3416 | } |
106c992a | 3417 | entry = huge_pte_mkdirty(entry); |
b4d1d99f DG |
3418 | } |
3419 | entry = pte_mkyoung(entry); | |
788c7df4 HD |
3420 | if (huge_ptep_set_access_flags(vma, address, ptep, entry, |
3421 | flags & FAULT_FLAG_WRITE)) | |
4b3073e1 | 3422 | update_mmu_cache(vma, address, ptep); |
0f792cf9 NH |
3423 | out_put_page: |
3424 | if (page != pagecache_page) | |
3425 | unlock_page(page); | |
3426 | put_page(page); | |
cb900f41 KS |
3427 | out_ptl: |
3428 | spin_unlock(ptl); | |
57303d80 AW |
3429 | |
3430 | if (pagecache_page) { | |
3431 | unlock_page(pagecache_page); | |
3432 | put_page(pagecache_page); | |
3433 | } | |
b4d1d99f | 3434 | out_mutex: |
8382d914 | 3435 | mutex_unlock(&htlb_fault_mutex_table[hash]); |
0f792cf9 NH |
3436 | /* |
3437 | * Generally it's safe to hold refcount during waiting page lock. But | |
3438 | * here we just wait to defer the next page fault to avoid busy loop and | |
3439 | * the page is not used after unlocked before returning from the current | |
3440 | * page fault. So we are safe from accessing freed page, even if we wait | |
3441 | * here without taking refcount. | |
3442 | */ | |
3443 | if (need_wait_lock) | |
3444 | wait_on_page_locked(page); | |
1e8f889b | 3445 | return ret; |
86e5216f AL |
3446 | } |
3447 | ||
28a35716 ML |
3448 | long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma, |
3449 | struct page **pages, struct vm_area_struct **vmas, | |
3450 | unsigned long *position, unsigned long *nr_pages, | |
3451 | long i, unsigned int flags) | |
63551ae0 | 3452 | { |
d5d4b0aa CK |
3453 | unsigned long pfn_offset; |
3454 | unsigned long vaddr = *position; | |
28a35716 | 3455 | unsigned long remainder = *nr_pages; |
a5516438 | 3456 | struct hstate *h = hstate_vma(vma); |
63551ae0 | 3457 | |
63551ae0 | 3458 | while (vaddr < vma->vm_end && remainder) { |
4c887265 | 3459 | pte_t *pte; |
cb900f41 | 3460 | spinlock_t *ptl = NULL; |
2a15efc9 | 3461 | int absent; |
4c887265 | 3462 | struct page *page; |
63551ae0 | 3463 | |
02057967 DR |
3464 | /* |
3465 | * If we have a pending SIGKILL, don't keep faulting pages and | |
3466 | * potentially allocating memory. | |
3467 | */ | |
3468 | if (unlikely(fatal_signal_pending(current))) { | |
3469 | remainder = 0; | |
3470 | break; | |
3471 | } | |
3472 | ||
4c887265 AL |
3473 | /* |
3474 | * Some archs (sparc64, sh*) have multiple pte_ts to | |
2a15efc9 | 3475 | * each hugepage. We have to make sure we get the |
4c887265 | 3476 | * first, for the page indexing below to work. |
cb900f41 KS |
3477 | * |
3478 | * Note that page table lock is not held when pte is null. | |
4c887265 | 3479 | */ |
a5516438 | 3480 | pte = huge_pte_offset(mm, vaddr & huge_page_mask(h)); |
cb900f41 KS |
3481 | if (pte) |
3482 | ptl = huge_pte_lock(h, mm, pte); | |
2a15efc9 HD |
3483 | absent = !pte || huge_pte_none(huge_ptep_get(pte)); |
3484 | ||
3485 | /* | |
3486 | * When coredumping, it suits get_dump_page if we just return | |
3ae77f43 HD |
3487 | * an error where there's an empty slot with no huge pagecache |
3488 | * to back it. This way, we avoid allocating a hugepage, and | |
3489 | * the sparse dumpfile avoids allocating disk blocks, but its | |
3490 | * huge holes still show up with zeroes where they need to be. | |
2a15efc9 | 3491 | */ |
3ae77f43 HD |
3492 | if (absent && (flags & FOLL_DUMP) && |
3493 | !hugetlbfs_pagecache_present(h, vma, vaddr)) { | |
cb900f41 KS |
3494 | if (pte) |
3495 | spin_unlock(ptl); | |
2a15efc9 HD |
3496 | remainder = 0; |
3497 | break; | |
3498 | } | |
63551ae0 | 3499 | |
9cc3a5bd NH |
3500 | /* |
3501 | * We need call hugetlb_fault for both hugepages under migration | |
3502 | * (in which case hugetlb_fault waits for the migration,) and | |
3503 | * hwpoisoned hugepages (in which case we need to prevent the | |
3504 | * caller from accessing to them.) In order to do this, we use | |
3505 | * here is_swap_pte instead of is_hugetlb_entry_migration and | |
3506 | * is_hugetlb_entry_hwpoisoned. This is because it simply covers | |
3507 | * both cases, and because we can't follow correct pages | |
3508 | * directly from any kind of swap entries. | |
3509 | */ | |
3510 | if (absent || is_swap_pte(huge_ptep_get(pte)) || | |
106c992a GS |
3511 | ((flags & FOLL_WRITE) && |
3512 | !huge_pte_write(huge_ptep_get(pte)))) { | |
4c887265 | 3513 | int ret; |
63551ae0 | 3514 | |
cb900f41 KS |
3515 | if (pte) |
3516 | spin_unlock(ptl); | |
2a15efc9 HD |
3517 | ret = hugetlb_fault(mm, vma, vaddr, |
3518 | (flags & FOLL_WRITE) ? FAULT_FLAG_WRITE : 0); | |
a89182c7 | 3519 | if (!(ret & VM_FAULT_ERROR)) |
4c887265 | 3520 | continue; |
63551ae0 | 3521 | |
4c887265 | 3522 | remainder = 0; |
4c887265 AL |
3523 | break; |
3524 | } | |
3525 | ||
a5516438 | 3526 | pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT; |
7f2e9525 | 3527 | page = pte_page(huge_ptep_get(pte)); |
d5d4b0aa | 3528 | same_page: |
d6692183 | 3529 | if (pages) { |
2a15efc9 | 3530 | pages[i] = mem_map_offset(page, pfn_offset); |
a0368d4e | 3531 | get_page_foll(pages[i]); |
d6692183 | 3532 | } |
63551ae0 DG |
3533 | |
3534 | if (vmas) | |
3535 | vmas[i] = vma; | |
3536 | ||
3537 | vaddr += PAGE_SIZE; | |
d5d4b0aa | 3538 | ++pfn_offset; |
63551ae0 DG |
3539 | --remainder; |
3540 | ++i; | |
d5d4b0aa | 3541 | if (vaddr < vma->vm_end && remainder && |
a5516438 | 3542 | pfn_offset < pages_per_huge_page(h)) { |
d5d4b0aa CK |
3543 | /* |
3544 | * We use pfn_offset to avoid touching the pageframes | |
3545 | * of this compound page. | |
3546 | */ | |
3547 | goto same_page; | |
3548 | } | |
cb900f41 | 3549 | spin_unlock(ptl); |
63551ae0 | 3550 | } |
28a35716 | 3551 | *nr_pages = remainder; |
63551ae0 DG |
3552 | *position = vaddr; |
3553 | ||
2a15efc9 | 3554 | return i ? i : -EFAULT; |
63551ae0 | 3555 | } |
8f860591 | 3556 | |
7da4d641 | 3557 | unsigned long hugetlb_change_protection(struct vm_area_struct *vma, |
8f860591 ZY |
3558 | unsigned long address, unsigned long end, pgprot_t newprot) |
3559 | { | |
3560 | struct mm_struct *mm = vma->vm_mm; | |
3561 | unsigned long start = address; | |
3562 | pte_t *ptep; | |
3563 | pte_t pte; | |
a5516438 | 3564 | struct hstate *h = hstate_vma(vma); |
7da4d641 | 3565 | unsigned long pages = 0; |
8f860591 ZY |
3566 | |
3567 | BUG_ON(address >= end); | |
3568 | flush_cache_range(vma, address, end); | |
3569 | ||
a5338093 | 3570 | mmu_notifier_invalidate_range_start(mm, start, end); |
83cde9e8 | 3571 | i_mmap_lock_write(vma->vm_file->f_mapping); |
a5516438 | 3572 | for (; address < end; address += huge_page_size(h)) { |
cb900f41 | 3573 | spinlock_t *ptl; |
8f860591 ZY |
3574 | ptep = huge_pte_offset(mm, address); |
3575 | if (!ptep) | |
3576 | continue; | |
cb900f41 | 3577 | ptl = huge_pte_lock(h, mm, ptep); |
7da4d641 PZ |
3578 | if (huge_pmd_unshare(mm, &address, ptep)) { |
3579 | pages++; | |
cb900f41 | 3580 | spin_unlock(ptl); |
39dde65c | 3581 | continue; |
7da4d641 | 3582 | } |
a8bda28d NH |
3583 | pte = huge_ptep_get(ptep); |
3584 | if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) { | |
3585 | spin_unlock(ptl); | |
3586 | continue; | |
3587 | } | |
3588 | if (unlikely(is_hugetlb_entry_migration(pte))) { | |
3589 | swp_entry_t entry = pte_to_swp_entry(pte); | |
3590 | ||
3591 | if (is_write_migration_entry(entry)) { | |
3592 | pte_t newpte; | |
3593 | ||
3594 | make_migration_entry_read(&entry); | |
3595 | newpte = swp_entry_to_pte(entry); | |
3596 | set_huge_pte_at(mm, address, ptep, newpte); | |
3597 | pages++; | |
3598 | } | |
3599 | spin_unlock(ptl); | |
3600 | continue; | |
3601 | } | |
3602 | if (!huge_pte_none(pte)) { | |
8f860591 | 3603 | pte = huge_ptep_get_and_clear(mm, address, ptep); |
106c992a | 3604 | pte = pte_mkhuge(huge_pte_modify(pte, newprot)); |
be7517d6 | 3605 | pte = arch_make_huge_pte(pte, vma, NULL, 0); |
8f860591 | 3606 | set_huge_pte_at(mm, address, ptep, pte); |
7da4d641 | 3607 | pages++; |
8f860591 | 3608 | } |
cb900f41 | 3609 | spin_unlock(ptl); |
8f860591 | 3610 | } |
d833352a | 3611 | /* |
c8c06efa | 3612 | * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare |
d833352a | 3613 | * may have cleared our pud entry and done put_page on the page table: |
c8c06efa | 3614 | * once we release i_mmap_rwsem, another task can do the final put_page |
d833352a MG |
3615 | * and that page table be reused and filled with junk. |
3616 | */ | |
8f860591 | 3617 | flush_tlb_range(vma, start, end); |
34ee645e | 3618 | mmu_notifier_invalidate_range(mm, start, end); |
83cde9e8 | 3619 | i_mmap_unlock_write(vma->vm_file->f_mapping); |
a5338093 | 3620 | mmu_notifier_invalidate_range_end(mm, start, end); |
7da4d641 PZ |
3621 | |
3622 | return pages << h->order; | |
8f860591 ZY |
3623 | } |
3624 | ||
a1e78772 MG |
3625 | int hugetlb_reserve_pages(struct inode *inode, |
3626 | long from, long to, | |
5a6fe125 | 3627 | struct vm_area_struct *vma, |
ca16d140 | 3628 | vm_flags_t vm_flags) |
e4e574b7 | 3629 | { |
17c9d12e | 3630 | long ret, chg; |
a5516438 | 3631 | struct hstate *h = hstate_inode(inode); |
90481622 | 3632 | struct hugepage_subpool *spool = subpool_inode(inode); |
9119a41e | 3633 | struct resv_map *resv_map; |
1c5ecae3 | 3634 | long gbl_reserve; |
e4e574b7 | 3635 | |
17c9d12e MG |
3636 | /* |
3637 | * Only apply hugepage reservation if asked. At fault time, an | |
3638 | * attempt will be made for VM_NORESERVE to allocate a page | |
90481622 | 3639 | * without using reserves |
17c9d12e | 3640 | */ |
ca16d140 | 3641 | if (vm_flags & VM_NORESERVE) |
17c9d12e MG |
3642 | return 0; |
3643 | ||
a1e78772 MG |
3644 | /* |
3645 | * Shared mappings base their reservation on the number of pages that | |
3646 | * are already allocated on behalf of the file. Private mappings need | |
3647 | * to reserve the full area even if read-only as mprotect() may be | |
3648 | * called to make the mapping read-write. Assume !vma is a shm mapping | |
3649 | */ | |
9119a41e | 3650 | if (!vma || vma->vm_flags & VM_MAYSHARE) { |
4e35f483 | 3651 | resv_map = inode_resv_map(inode); |
9119a41e | 3652 | |
1406ec9b | 3653 | chg = region_chg(resv_map, from, to); |
9119a41e JK |
3654 | |
3655 | } else { | |
3656 | resv_map = resv_map_alloc(); | |
17c9d12e MG |
3657 | if (!resv_map) |
3658 | return -ENOMEM; | |
3659 | ||
a1e78772 | 3660 | chg = to - from; |
84afd99b | 3661 | |
17c9d12e MG |
3662 | set_vma_resv_map(vma, resv_map); |
3663 | set_vma_resv_flags(vma, HPAGE_RESV_OWNER); | |
3664 | } | |
3665 | ||
c50ac050 DH |
3666 | if (chg < 0) { |
3667 | ret = chg; | |
3668 | goto out_err; | |
3669 | } | |
8a630112 | 3670 | |
1c5ecae3 MK |
3671 | /* |
3672 | * There must be enough pages in the subpool for the mapping. If | |
3673 | * the subpool has a minimum size, there may be some global | |
3674 | * reservations already in place (gbl_reserve). | |
3675 | */ | |
3676 | gbl_reserve = hugepage_subpool_get_pages(spool, chg); | |
3677 | if (gbl_reserve < 0) { | |
c50ac050 DH |
3678 | ret = -ENOSPC; |
3679 | goto out_err; | |
3680 | } | |
5a6fe125 MG |
3681 | |
3682 | /* | |
17c9d12e | 3683 | * Check enough hugepages are available for the reservation. |
90481622 | 3684 | * Hand the pages back to the subpool if there are not |
5a6fe125 | 3685 | */ |
1c5ecae3 | 3686 | ret = hugetlb_acct_memory(h, gbl_reserve); |
68842c9b | 3687 | if (ret < 0) { |
1c5ecae3 MK |
3688 | /* put back original number of pages, chg */ |
3689 | (void)hugepage_subpool_put_pages(spool, chg); | |
c50ac050 | 3690 | goto out_err; |
68842c9b | 3691 | } |
17c9d12e MG |
3692 | |
3693 | /* | |
3694 | * Account for the reservations made. Shared mappings record regions | |
3695 | * that have reservations as they are shared by multiple VMAs. | |
3696 | * When the last VMA disappears, the region map says how much | |
3697 | * the reservation was and the page cache tells how much of | |
3698 | * the reservation was consumed. Private mappings are per-VMA and | |
3699 | * only the consumed reservations are tracked. When the VMA | |
3700 | * disappears, the original reservation is the VMA size and the | |
3701 | * consumed reservations are stored in the map. Hence, nothing | |
3702 | * else has to be done for private mappings here | |
3703 | */ | |
f83a275d | 3704 | if (!vma || vma->vm_flags & VM_MAYSHARE) |
1406ec9b | 3705 | region_add(resv_map, from, to); |
a43a8c39 | 3706 | return 0; |
c50ac050 | 3707 | out_err: |
f031dd27 JK |
3708 | if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) |
3709 | kref_put(&resv_map->refs, resv_map_release); | |
c50ac050 | 3710 | return ret; |
a43a8c39 CK |
3711 | } |
3712 | ||
3713 | void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed) | |
3714 | { | |
a5516438 | 3715 | struct hstate *h = hstate_inode(inode); |
4e35f483 | 3716 | struct resv_map *resv_map = inode_resv_map(inode); |
9119a41e | 3717 | long chg = 0; |
90481622 | 3718 | struct hugepage_subpool *spool = subpool_inode(inode); |
1c5ecae3 | 3719 | long gbl_reserve; |
45c682a6 | 3720 | |
9119a41e | 3721 | if (resv_map) |
1406ec9b | 3722 | chg = region_truncate(resv_map, offset); |
45c682a6 | 3723 | spin_lock(&inode->i_lock); |
e4c6f8be | 3724 | inode->i_blocks -= (blocks_per_huge_page(h) * freed); |
45c682a6 KC |
3725 | spin_unlock(&inode->i_lock); |
3726 | ||
1c5ecae3 MK |
3727 | /* |
3728 | * If the subpool has a minimum size, the number of global | |
3729 | * reservations to be released may be adjusted. | |
3730 | */ | |
3731 | gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed)); | |
3732 | hugetlb_acct_memory(h, -gbl_reserve); | |
a43a8c39 | 3733 | } |
93f70f90 | 3734 | |
3212b535 SC |
3735 | #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE |
3736 | static unsigned long page_table_shareable(struct vm_area_struct *svma, | |
3737 | struct vm_area_struct *vma, | |
3738 | unsigned long addr, pgoff_t idx) | |
3739 | { | |
3740 | unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) + | |
3741 | svma->vm_start; | |
3742 | unsigned long sbase = saddr & PUD_MASK; | |
3743 | unsigned long s_end = sbase + PUD_SIZE; | |
3744 | ||
3745 | /* Allow segments to share if only one is marked locked */ | |
3746 | unsigned long vm_flags = vma->vm_flags & ~VM_LOCKED; | |
3747 | unsigned long svm_flags = svma->vm_flags & ~VM_LOCKED; | |
3748 | ||
3749 | /* | |
3750 | * match the virtual addresses, permission and the alignment of the | |
3751 | * page table page. | |
3752 | */ | |
3753 | if (pmd_index(addr) != pmd_index(saddr) || | |
3754 | vm_flags != svm_flags || | |
3755 | sbase < svma->vm_start || svma->vm_end < s_end) | |
3756 | return 0; | |
3757 | ||
3758 | return saddr; | |
3759 | } | |
3760 | ||
3761 | static int vma_shareable(struct vm_area_struct *vma, unsigned long addr) | |
3762 | { | |
3763 | unsigned long base = addr & PUD_MASK; | |
3764 | unsigned long end = base + PUD_SIZE; | |
3765 | ||
3766 | /* | |
3767 | * check on proper vm_flags and page table alignment | |
3768 | */ | |
3769 | if (vma->vm_flags & VM_MAYSHARE && | |
3770 | vma->vm_start <= base && end <= vma->vm_end) | |
3771 | return 1; | |
3772 | return 0; | |
3773 | } | |
3774 | ||
3775 | /* | |
3776 | * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc() | |
3777 | * and returns the corresponding pte. While this is not necessary for the | |
3778 | * !shared pmd case because we can allocate the pmd later as well, it makes the | |
3779 | * code much cleaner. pmd allocation is essential for the shared case because | |
c8c06efa | 3780 | * pud has to be populated inside the same i_mmap_rwsem section - otherwise |
3212b535 SC |
3781 | * racing tasks could either miss the sharing (see huge_pte_offset) or select a |
3782 | * bad pmd for sharing. | |
3783 | */ | |
3784 | pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud) | |
3785 | { | |
3786 | struct vm_area_struct *vma = find_vma(mm, addr); | |
3787 | struct address_space *mapping = vma->vm_file->f_mapping; | |
3788 | pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) + | |
3789 | vma->vm_pgoff; | |
3790 | struct vm_area_struct *svma; | |
3791 | unsigned long saddr; | |
3792 | pte_t *spte = NULL; | |
3793 | pte_t *pte; | |
cb900f41 | 3794 | spinlock_t *ptl; |
3212b535 SC |
3795 | |
3796 | if (!vma_shareable(vma, addr)) | |
3797 | return (pte_t *)pmd_alloc(mm, pud, addr); | |
3798 | ||
83cde9e8 | 3799 | i_mmap_lock_write(mapping); |
3212b535 SC |
3800 | vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) { |
3801 | if (svma == vma) | |
3802 | continue; | |
3803 | ||
3804 | saddr = page_table_shareable(svma, vma, addr, idx); | |
3805 | if (saddr) { | |
3806 | spte = huge_pte_offset(svma->vm_mm, saddr); | |
3807 | if (spte) { | |
dc6c9a35 | 3808 | mm_inc_nr_pmds(mm); |
3212b535 SC |
3809 | get_page(virt_to_page(spte)); |
3810 | break; | |
3811 | } | |
3812 | } | |
3813 | } | |
3814 | ||
3815 | if (!spte) | |
3816 | goto out; | |
3817 | ||
cb900f41 KS |
3818 | ptl = huge_pte_lockptr(hstate_vma(vma), mm, spte); |
3819 | spin_lock(ptl); | |
dc6c9a35 | 3820 | if (pud_none(*pud)) { |
3212b535 SC |
3821 | pud_populate(mm, pud, |
3822 | (pmd_t *)((unsigned long)spte & PAGE_MASK)); | |
dc6c9a35 | 3823 | } else { |
3212b535 | 3824 | put_page(virt_to_page(spte)); |
dc6c9a35 KS |
3825 | mm_inc_nr_pmds(mm); |
3826 | } | |
cb900f41 | 3827 | spin_unlock(ptl); |
3212b535 SC |
3828 | out: |
3829 | pte = (pte_t *)pmd_alloc(mm, pud, addr); | |
83cde9e8 | 3830 | i_mmap_unlock_write(mapping); |
3212b535 SC |
3831 | return pte; |
3832 | } | |
3833 | ||
3834 | /* | |
3835 | * unmap huge page backed by shared pte. | |
3836 | * | |
3837 | * Hugetlb pte page is ref counted at the time of mapping. If pte is shared | |
3838 | * indicated by page_count > 1, unmap is achieved by clearing pud and | |
3839 | * decrementing the ref count. If count == 1, the pte page is not shared. | |
3840 | * | |
cb900f41 | 3841 | * called with page table lock held. |
3212b535 SC |
3842 | * |
3843 | * returns: 1 successfully unmapped a shared pte page | |
3844 | * 0 the underlying pte page is not shared, or it is the last user | |
3845 | */ | |
3846 | int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep) | |
3847 | { | |
3848 | pgd_t *pgd = pgd_offset(mm, *addr); | |
3849 | pud_t *pud = pud_offset(pgd, *addr); | |
3850 | ||
3851 | BUG_ON(page_count(virt_to_page(ptep)) == 0); | |
3852 | if (page_count(virt_to_page(ptep)) == 1) | |
3853 | return 0; | |
3854 | ||
3855 | pud_clear(pud); | |
3856 | put_page(virt_to_page(ptep)); | |
dc6c9a35 | 3857 | mm_dec_nr_pmds(mm); |
3212b535 SC |
3858 | *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE; |
3859 | return 1; | |
3860 | } | |
9e5fc74c SC |
3861 | #define want_pmd_share() (1) |
3862 | #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ | |
3863 | pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud) | |
3864 | { | |
3865 | return NULL; | |
3866 | } | |
e81f2d22 ZZ |
3867 | |
3868 | int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep) | |
3869 | { | |
3870 | return 0; | |
3871 | } | |
9e5fc74c | 3872 | #define want_pmd_share() (0) |
3212b535 SC |
3873 | #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ |
3874 | ||
9e5fc74c SC |
3875 | #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB |
3876 | pte_t *huge_pte_alloc(struct mm_struct *mm, | |
3877 | unsigned long addr, unsigned long sz) | |
3878 | { | |
3879 | pgd_t *pgd; | |
3880 | pud_t *pud; | |
3881 | pte_t *pte = NULL; | |
3882 | ||
3883 | pgd = pgd_offset(mm, addr); | |
3884 | pud = pud_alloc(mm, pgd, addr); | |
3885 | if (pud) { | |
3886 | if (sz == PUD_SIZE) { | |
3887 | pte = (pte_t *)pud; | |
3888 | } else { | |
3889 | BUG_ON(sz != PMD_SIZE); | |
3890 | if (want_pmd_share() && pud_none(*pud)) | |
3891 | pte = huge_pmd_share(mm, addr, pud); | |
3892 | else | |
3893 | pte = (pte_t *)pmd_alloc(mm, pud, addr); | |
3894 | } | |
3895 | } | |
3896 | BUG_ON(pte && !pte_none(*pte) && !pte_huge(*pte)); | |
3897 | ||
3898 | return pte; | |
3899 | } | |
3900 | ||
3901 | pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr) | |
3902 | { | |
3903 | pgd_t *pgd; | |
3904 | pud_t *pud; | |
3905 | pmd_t *pmd = NULL; | |
3906 | ||
3907 | pgd = pgd_offset(mm, addr); | |
3908 | if (pgd_present(*pgd)) { | |
3909 | pud = pud_offset(pgd, addr); | |
3910 | if (pud_present(*pud)) { | |
3911 | if (pud_huge(*pud)) | |
3912 | return (pte_t *)pud; | |
3913 | pmd = pmd_offset(pud, addr); | |
3914 | } | |
3915 | } | |
3916 | return (pte_t *) pmd; | |
3917 | } | |
3918 | ||
61f77eda NH |
3919 | #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */ |
3920 | ||
3921 | /* | |
3922 | * These functions are overwritable if your architecture needs its own | |
3923 | * behavior. | |
3924 | */ | |
3925 | struct page * __weak | |
3926 | follow_huge_addr(struct mm_struct *mm, unsigned long address, | |
3927 | int write) | |
3928 | { | |
3929 | return ERR_PTR(-EINVAL); | |
3930 | } | |
3931 | ||
3932 | struct page * __weak | |
9e5fc74c | 3933 | follow_huge_pmd(struct mm_struct *mm, unsigned long address, |
e66f17ff | 3934 | pmd_t *pmd, int flags) |
9e5fc74c | 3935 | { |
e66f17ff NH |
3936 | struct page *page = NULL; |
3937 | spinlock_t *ptl; | |
3938 | retry: | |
3939 | ptl = pmd_lockptr(mm, pmd); | |
3940 | spin_lock(ptl); | |
3941 | /* | |
3942 | * make sure that the address range covered by this pmd is not | |
3943 | * unmapped from other threads. | |
3944 | */ | |
3945 | if (!pmd_huge(*pmd)) | |
3946 | goto out; | |
3947 | if (pmd_present(*pmd)) { | |
97534127 | 3948 | page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT); |
e66f17ff NH |
3949 | if (flags & FOLL_GET) |
3950 | get_page(page); | |
3951 | } else { | |
3952 | if (is_hugetlb_entry_migration(huge_ptep_get((pte_t *)pmd))) { | |
3953 | spin_unlock(ptl); | |
3954 | __migration_entry_wait(mm, (pte_t *)pmd, ptl); | |
3955 | goto retry; | |
3956 | } | |
3957 | /* | |
3958 | * hwpoisoned entry is treated as no_page_table in | |
3959 | * follow_page_mask(). | |
3960 | */ | |
3961 | } | |
3962 | out: | |
3963 | spin_unlock(ptl); | |
9e5fc74c SC |
3964 | return page; |
3965 | } | |
3966 | ||
61f77eda | 3967 | struct page * __weak |
9e5fc74c | 3968 | follow_huge_pud(struct mm_struct *mm, unsigned long address, |
e66f17ff | 3969 | pud_t *pud, int flags) |
9e5fc74c | 3970 | { |
e66f17ff NH |
3971 | if (flags & FOLL_GET) |
3972 | return NULL; | |
9e5fc74c | 3973 | |
e66f17ff | 3974 | return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT); |
9e5fc74c SC |
3975 | } |
3976 | ||
d5bd9106 AK |
3977 | #ifdef CONFIG_MEMORY_FAILURE |
3978 | ||
93f70f90 NH |
3979 | /* |
3980 | * This function is called from memory failure code. | |
3981 | * Assume the caller holds page lock of the head page. | |
3982 | */ | |
6de2b1aa | 3983 | int dequeue_hwpoisoned_huge_page(struct page *hpage) |
93f70f90 NH |
3984 | { |
3985 | struct hstate *h = page_hstate(hpage); | |
3986 | int nid = page_to_nid(hpage); | |
6de2b1aa | 3987 | int ret = -EBUSY; |
93f70f90 NH |
3988 | |
3989 | spin_lock(&hugetlb_lock); | |
7e1f049e NH |
3990 | /* |
3991 | * Just checking !page_huge_active is not enough, because that could be | |
3992 | * an isolated/hwpoisoned hugepage (which have >0 refcount). | |
3993 | */ | |
3994 | if (!page_huge_active(hpage) && !page_count(hpage)) { | |
56f2fb14 NH |
3995 | /* |
3996 | * Hwpoisoned hugepage isn't linked to activelist or freelist, | |
3997 | * but dangling hpage->lru can trigger list-debug warnings | |
3998 | * (this happens when we call unpoison_memory() on it), | |
3999 | * so let it point to itself with list_del_init(). | |
4000 | */ | |
4001 | list_del_init(&hpage->lru); | |
8c6c2ecb | 4002 | set_page_refcounted(hpage); |
6de2b1aa NH |
4003 | h->free_huge_pages--; |
4004 | h->free_huge_pages_node[nid]--; | |
4005 | ret = 0; | |
4006 | } | |
93f70f90 | 4007 | spin_unlock(&hugetlb_lock); |
6de2b1aa | 4008 | return ret; |
93f70f90 | 4009 | } |
6de2b1aa | 4010 | #endif |
31caf665 NH |
4011 | |
4012 | bool isolate_huge_page(struct page *page, struct list_head *list) | |
4013 | { | |
bcc54222 NH |
4014 | bool ret = true; |
4015 | ||
309381fe | 4016 | VM_BUG_ON_PAGE(!PageHead(page), page); |
31caf665 | 4017 | spin_lock(&hugetlb_lock); |
bcc54222 NH |
4018 | if (!page_huge_active(page) || !get_page_unless_zero(page)) { |
4019 | ret = false; | |
4020 | goto unlock; | |
4021 | } | |
4022 | clear_page_huge_active(page); | |
31caf665 | 4023 | list_move_tail(&page->lru, list); |
bcc54222 | 4024 | unlock: |
31caf665 | 4025 | spin_unlock(&hugetlb_lock); |
bcc54222 | 4026 | return ret; |
31caf665 NH |
4027 | } |
4028 | ||
4029 | void putback_active_hugepage(struct page *page) | |
4030 | { | |
309381fe | 4031 | VM_BUG_ON_PAGE(!PageHead(page), page); |
31caf665 | 4032 | spin_lock(&hugetlb_lock); |
bcc54222 | 4033 | set_page_huge_active(page); |
31caf665 NH |
4034 | list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist); |
4035 | spin_unlock(&hugetlb_lock); | |
4036 | put_page(page); | |
4037 | } |