| 1 | /* |
| 2 | * Generic hugetlb support. |
| 3 | * (C) Nadia Yvette Chambers, April 2004 |
| 4 | */ |
| 5 | #include <linux/list.h> |
| 6 | #include <linux/init.h> |
| 7 | #include <linux/module.h> |
| 8 | #include <linux/mm.h> |
| 9 | #include <linux/seq_file.h> |
| 10 | #include <linux/sysctl.h> |
| 11 | #include <linux/highmem.h> |
| 12 | #include <linux/mmu_notifier.h> |
| 13 | #include <linux/nodemask.h> |
| 14 | #include <linux/pagemap.h> |
| 15 | #include <linux/mempolicy.h> |
| 16 | #include <linux/cpuset.h> |
| 17 | #include <linux/mutex.h> |
| 18 | #include <linux/bootmem.h> |
| 19 | #include <linux/sysfs.h> |
| 20 | #include <linux/slab.h> |
| 21 | #include <linux/rmap.h> |
| 22 | #include <linux/swap.h> |
| 23 | #include <linux/swapops.h> |
| 24 | #include <linux/page-isolation.h> |
| 25 | |
| 26 | #include <asm/page.h> |
| 27 | #include <asm/pgtable.h> |
| 28 | #include <asm/tlb.h> |
| 29 | |
| 30 | #include <linux/io.h> |
| 31 | #include <linux/hugetlb.h> |
| 32 | #include <linux/hugetlb_cgroup.h> |
| 33 | #include <linux/node.h> |
| 34 | #include "internal.h" |
| 35 | |
| 36 | const unsigned long hugetlb_zero = 0, hugetlb_infinity = ~0UL; |
| 37 | unsigned long hugepages_treat_as_movable; |
| 38 | |
| 39 | int hugetlb_max_hstate __read_mostly; |
| 40 | unsigned int default_hstate_idx; |
| 41 | struct hstate hstates[HUGE_MAX_HSTATE]; |
| 42 | |
| 43 | __initdata LIST_HEAD(huge_boot_pages); |
| 44 | |
| 45 | /* for command line parsing */ |
| 46 | static struct hstate * __initdata parsed_hstate; |
| 47 | static unsigned long __initdata default_hstate_max_huge_pages; |
| 48 | static unsigned long __initdata default_hstate_size; |
| 49 | |
| 50 | /* |
| 51 | * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages, |
| 52 | * free_huge_pages, and surplus_huge_pages. |
| 53 | */ |
| 54 | DEFINE_SPINLOCK(hugetlb_lock); |
| 55 | |
| 56 | static inline void unlock_or_release_subpool(struct hugepage_subpool *spool) |
| 57 | { |
| 58 | bool free = (spool->count == 0) && (spool->used_hpages == 0); |
| 59 | |
| 60 | spin_unlock(&spool->lock); |
| 61 | |
| 62 | /* If no pages are used, and no other handles to the subpool |
| 63 | * remain, free the subpool the subpool remain */ |
| 64 | if (free) |
| 65 | kfree(spool); |
| 66 | } |
| 67 | |
| 68 | struct hugepage_subpool *hugepage_new_subpool(long nr_blocks) |
| 69 | { |
| 70 | struct hugepage_subpool *spool; |
| 71 | |
| 72 | spool = kmalloc(sizeof(*spool), GFP_KERNEL); |
| 73 | if (!spool) |
| 74 | return NULL; |
| 75 | |
| 76 | spin_lock_init(&spool->lock); |
| 77 | spool->count = 1; |
| 78 | spool->max_hpages = nr_blocks; |
| 79 | spool->used_hpages = 0; |
| 80 | |
| 81 | return spool; |
| 82 | } |
| 83 | |
| 84 | void hugepage_put_subpool(struct hugepage_subpool *spool) |
| 85 | { |
| 86 | spin_lock(&spool->lock); |
| 87 | BUG_ON(!spool->count); |
| 88 | spool->count--; |
| 89 | unlock_or_release_subpool(spool); |
| 90 | } |
| 91 | |
| 92 | static int hugepage_subpool_get_pages(struct hugepage_subpool *spool, |
| 93 | long delta) |
| 94 | { |
| 95 | int ret = 0; |
| 96 | |
| 97 | if (!spool) |
| 98 | return 0; |
| 99 | |
| 100 | spin_lock(&spool->lock); |
| 101 | if ((spool->used_hpages + delta) <= spool->max_hpages) { |
| 102 | spool->used_hpages += delta; |
| 103 | } else { |
| 104 | ret = -ENOMEM; |
| 105 | } |
| 106 | spin_unlock(&spool->lock); |
| 107 | |
| 108 | return ret; |
| 109 | } |
| 110 | |
| 111 | static void hugepage_subpool_put_pages(struct hugepage_subpool *spool, |
| 112 | long delta) |
| 113 | { |
| 114 | if (!spool) |
| 115 | return; |
| 116 | |
| 117 | spin_lock(&spool->lock); |
| 118 | spool->used_hpages -= delta; |
| 119 | /* If hugetlbfs_put_super couldn't free spool due to |
| 120 | * an outstanding quota reference, free it now. */ |
| 121 | unlock_or_release_subpool(spool); |
| 122 | } |
| 123 | |
| 124 | static inline struct hugepage_subpool *subpool_inode(struct inode *inode) |
| 125 | { |
| 126 | return HUGETLBFS_SB(inode->i_sb)->spool; |
| 127 | } |
| 128 | |
| 129 | static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma) |
| 130 | { |
| 131 | return subpool_inode(file_inode(vma->vm_file)); |
| 132 | } |
| 133 | |
| 134 | /* |
| 135 | * Region tracking -- allows tracking of reservations and instantiated pages |
| 136 | * across the pages in a mapping. |
| 137 | * |
| 138 | * The region data structures are protected by a combination of the mmap_sem |
| 139 | * and the hugetlb_instantiation_mutex. To access or modify a region the caller |
| 140 | * must either hold the mmap_sem for write, or the mmap_sem for read and |
| 141 | * the hugetlb_instantiation_mutex: |
| 142 | * |
| 143 | * down_write(&mm->mmap_sem); |
| 144 | * or |
| 145 | * down_read(&mm->mmap_sem); |
| 146 | * mutex_lock(&hugetlb_instantiation_mutex); |
| 147 | */ |
| 148 | struct file_region { |
| 149 | struct list_head link; |
| 150 | long from; |
| 151 | long to; |
| 152 | }; |
| 153 | |
| 154 | static long region_add(struct list_head *head, long f, long t) |
| 155 | { |
| 156 | struct file_region *rg, *nrg, *trg; |
| 157 | |
| 158 | /* Locate the region we are either in or before. */ |
| 159 | list_for_each_entry(rg, head, link) |
| 160 | if (f <= rg->to) |
| 161 | break; |
| 162 | |
| 163 | /* Round our left edge to the current segment if it encloses us. */ |
| 164 | if (f > rg->from) |
| 165 | f = rg->from; |
| 166 | |
| 167 | /* Check for and consume any regions we now overlap with. */ |
| 168 | nrg = rg; |
| 169 | list_for_each_entry_safe(rg, trg, rg->link.prev, link) { |
| 170 | if (&rg->link == head) |
| 171 | break; |
| 172 | if (rg->from > t) |
| 173 | break; |
| 174 | |
| 175 | /* If this area reaches higher then extend our area to |
| 176 | * include it completely. If this is not the first area |
| 177 | * which we intend to reuse, free it. */ |
| 178 | if (rg->to > t) |
| 179 | t = rg->to; |
| 180 | if (rg != nrg) { |
| 181 | list_del(&rg->link); |
| 182 | kfree(rg); |
| 183 | } |
| 184 | } |
| 185 | nrg->from = f; |
| 186 | nrg->to = t; |
| 187 | return 0; |
| 188 | } |
| 189 | |
| 190 | static long region_chg(struct list_head *head, long f, long t) |
| 191 | { |
| 192 | struct file_region *rg, *nrg; |
| 193 | long chg = 0; |
| 194 | |
| 195 | /* Locate the region we are before or in. */ |
| 196 | list_for_each_entry(rg, head, link) |
| 197 | if (f <= rg->to) |
| 198 | break; |
| 199 | |
| 200 | /* If we are below the current region then a new region is required. |
| 201 | * Subtle, allocate a new region at the position but make it zero |
| 202 | * size such that we can guarantee to record the reservation. */ |
| 203 | if (&rg->link == head || t < rg->from) { |
| 204 | nrg = kmalloc(sizeof(*nrg), GFP_KERNEL); |
| 205 | if (!nrg) |
| 206 | return -ENOMEM; |
| 207 | nrg->from = f; |
| 208 | nrg->to = f; |
| 209 | INIT_LIST_HEAD(&nrg->link); |
| 210 | list_add(&nrg->link, rg->link.prev); |
| 211 | |
| 212 | return t - f; |
| 213 | } |
| 214 | |
| 215 | /* Round our left edge to the current segment if it encloses us. */ |
| 216 | if (f > rg->from) |
| 217 | f = rg->from; |
| 218 | chg = t - f; |
| 219 | |
| 220 | /* Check for and consume any regions we now overlap with. */ |
| 221 | list_for_each_entry(rg, rg->link.prev, link) { |
| 222 | if (&rg->link == head) |
| 223 | break; |
| 224 | if (rg->from > t) |
| 225 | return chg; |
| 226 | |
| 227 | /* We overlap with this area, if it extends further than |
| 228 | * us then we must extend ourselves. Account for its |
| 229 | * existing reservation. */ |
| 230 | if (rg->to > t) { |
| 231 | chg += rg->to - t; |
| 232 | t = rg->to; |
| 233 | } |
| 234 | chg -= rg->to - rg->from; |
| 235 | } |
| 236 | return chg; |
| 237 | } |
| 238 | |
| 239 | static long region_truncate(struct list_head *head, long end) |
| 240 | { |
| 241 | struct file_region *rg, *trg; |
| 242 | long chg = 0; |
| 243 | |
| 244 | /* Locate the region we are either in or before. */ |
| 245 | list_for_each_entry(rg, head, link) |
| 246 | if (end <= rg->to) |
| 247 | break; |
| 248 | if (&rg->link == head) |
| 249 | return 0; |
| 250 | |
| 251 | /* If we are in the middle of a region then adjust it. */ |
| 252 | if (end > rg->from) { |
| 253 | chg = rg->to - end; |
| 254 | rg->to = end; |
| 255 | rg = list_entry(rg->link.next, typeof(*rg), link); |
| 256 | } |
| 257 | |
| 258 | /* Drop any remaining regions. */ |
| 259 | list_for_each_entry_safe(rg, trg, rg->link.prev, link) { |
| 260 | if (&rg->link == head) |
| 261 | break; |
| 262 | chg += rg->to - rg->from; |
| 263 | list_del(&rg->link); |
| 264 | kfree(rg); |
| 265 | } |
| 266 | return chg; |
| 267 | } |
| 268 | |
| 269 | static long region_count(struct list_head *head, long f, long t) |
| 270 | { |
| 271 | struct file_region *rg; |
| 272 | long chg = 0; |
| 273 | |
| 274 | /* Locate each segment we overlap with, and count that overlap. */ |
| 275 | list_for_each_entry(rg, head, link) { |
| 276 | long seg_from; |
| 277 | long seg_to; |
| 278 | |
| 279 | if (rg->to <= f) |
| 280 | continue; |
| 281 | if (rg->from >= t) |
| 282 | break; |
| 283 | |
| 284 | seg_from = max(rg->from, f); |
| 285 | seg_to = min(rg->to, t); |
| 286 | |
| 287 | chg += seg_to - seg_from; |
| 288 | } |
| 289 | |
| 290 | return chg; |
| 291 | } |
| 292 | |
| 293 | /* |
| 294 | * Convert the address within this vma to the page offset within |
| 295 | * the mapping, in pagecache page units; huge pages here. |
| 296 | */ |
| 297 | static pgoff_t vma_hugecache_offset(struct hstate *h, |
| 298 | struct vm_area_struct *vma, unsigned long address) |
| 299 | { |
| 300 | return ((address - vma->vm_start) >> huge_page_shift(h)) + |
| 301 | (vma->vm_pgoff >> huge_page_order(h)); |
| 302 | } |
| 303 | |
| 304 | pgoff_t linear_hugepage_index(struct vm_area_struct *vma, |
| 305 | unsigned long address) |
| 306 | { |
| 307 | return vma_hugecache_offset(hstate_vma(vma), vma, address); |
| 308 | } |
| 309 | |
| 310 | /* |
| 311 | * Return the size of the pages allocated when backing a VMA. In the majority |
| 312 | * cases this will be same size as used by the page table entries. |
| 313 | */ |
| 314 | unsigned long vma_kernel_pagesize(struct vm_area_struct *vma) |
| 315 | { |
| 316 | struct hstate *hstate; |
| 317 | |
| 318 | if (!is_vm_hugetlb_page(vma)) |
| 319 | return PAGE_SIZE; |
| 320 | |
| 321 | hstate = hstate_vma(vma); |
| 322 | |
| 323 | return 1UL << huge_page_shift(hstate); |
| 324 | } |
| 325 | EXPORT_SYMBOL_GPL(vma_kernel_pagesize); |
| 326 | |
| 327 | /* |
| 328 | * Return the page size being used by the MMU to back a VMA. In the majority |
| 329 | * of cases, the page size used by the kernel matches the MMU size. On |
| 330 | * architectures where it differs, an architecture-specific version of this |
| 331 | * function is required. |
| 332 | */ |
| 333 | #ifndef vma_mmu_pagesize |
| 334 | unsigned long vma_mmu_pagesize(struct vm_area_struct *vma) |
| 335 | { |
| 336 | return vma_kernel_pagesize(vma); |
| 337 | } |
| 338 | #endif |
| 339 | |
| 340 | /* |
| 341 | * Flags for MAP_PRIVATE reservations. These are stored in the bottom |
| 342 | * bits of the reservation map pointer, which are always clear due to |
| 343 | * alignment. |
| 344 | */ |
| 345 | #define HPAGE_RESV_OWNER (1UL << 0) |
| 346 | #define HPAGE_RESV_UNMAPPED (1UL << 1) |
| 347 | #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED) |
| 348 | |
| 349 | /* |
| 350 | * These helpers are used to track how many pages are reserved for |
| 351 | * faults in a MAP_PRIVATE mapping. Only the process that called mmap() |
| 352 | * is guaranteed to have their future faults succeed. |
| 353 | * |
| 354 | * With the exception of reset_vma_resv_huge_pages() which is called at fork(), |
| 355 | * the reserve counters are updated with the hugetlb_lock held. It is safe |
| 356 | * to reset the VMA at fork() time as it is not in use yet and there is no |
| 357 | * chance of the global counters getting corrupted as a result of the values. |
| 358 | * |
| 359 | * The private mapping reservation is represented in a subtly different |
| 360 | * manner to a shared mapping. A shared mapping has a region map associated |
| 361 | * with the underlying file, this region map represents the backing file |
| 362 | * pages which have ever had a reservation assigned which this persists even |
| 363 | * after the page is instantiated. A private mapping has a region map |
| 364 | * associated with the original mmap which is attached to all VMAs which |
| 365 | * reference it, this region map represents those offsets which have consumed |
| 366 | * reservation ie. where pages have been instantiated. |
| 367 | */ |
| 368 | static unsigned long get_vma_private_data(struct vm_area_struct *vma) |
| 369 | { |
| 370 | return (unsigned long)vma->vm_private_data; |
| 371 | } |
| 372 | |
| 373 | static void set_vma_private_data(struct vm_area_struct *vma, |
| 374 | unsigned long value) |
| 375 | { |
| 376 | vma->vm_private_data = (void *)value; |
| 377 | } |
| 378 | |
| 379 | struct resv_map { |
| 380 | struct kref refs; |
| 381 | struct list_head regions; |
| 382 | }; |
| 383 | |
| 384 | static struct resv_map *resv_map_alloc(void) |
| 385 | { |
| 386 | struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL); |
| 387 | if (!resv_map) |
| 388 | return NULL; |
| 389 | |
| 390 | kref_init(&resv_map->refs); |
| 391 | INIT_LIST_HEAD(&resv_map->regions); |
| 392 | |
| 393 | return resv_map; |
| 394 | } |
| 395 | |
| 396 | static void resv_map_release(struct kref *ref) |
| 397 | { |
| 398 | struct resv_map *resv_map = container_of(ref, struct resv_map, refs); |
| 399 | |
| 400 | /* Clear out any active regions before we release the map. */ |
| 401 | region_truncate(&resv_map->regions, 0); |
| 402 | kfree(resv_map); |
| 403 | } |
| 404 | |
| 405 | static struct resv_map *vma_resv_map(struct vm_area_struct *vma) |
| 406 | { |
| 407 | VM_BUG_ON(!is_vm_hugetlb_page(vma)); |
| 408 | if (!(vma->vm_flags & VM_MAYSHARE)) |
| 409 | return (struct resv_map *)(get_vma_private_data(vma) & |
| 410 | ~HPAGE_RESV_MASK); |
| 411 | return NULL; |
| 412 | } |
| 413 | |
| 414 | static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map) |
| 415 | { |
| 416 | VM_BUG_ON(!is_vm_hugetlb_page(vma)); |
| 417 | VM_BUG_ON(vma->vm_flags & VM_MAYSHARE); |
| 418 | |
| 419 | set_vma_private_data(vma, (get_vma_private_data(vma) & |
| 420 | HPAGE_RESV_MASK) | (unsigned long)map); |
| 421 | } |
| 422 | |
| 423 | static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags) |
| 424 | { |
| 425 | VM_BUG_ON(!is_vm_hugetlb_page(vma)); |
| 426 | VM_BUG_ON(vma->vm_flags & VM_MAYSHARE); |
| 427 | |
| 428 | set_vma_private_data(vma, get_vma_private_data(vma) | flags); |
| 429 | } |
| 430 | |
| 431 | static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag) |
| 432 | { |
| 433 | VM_BUG_ON(!is_vm_hugetlb_page(vma)); |
| 434 | |
| 435 | return (get_vma_private_data(vma) & flag) != 0; |
| 436 | } |
| 437 | |
| 438 | /* Reset counters to 0 and clear all HPAGE_RESV_* flags */ |
| 439 | void reset_vma_resv_huge_pages(struct vm_area_struct *vma) |
| 440 | { |
| 441 | VM_BUG_ON(!is_vm_hugetlb_page(vma)); |
| 442 | if (!(vma->vm_flags & VM_MAYSHARE)) |
| 443 | vma->vm_private_data = (void *)0; |
| 444 | } |
| 445 | |
| 446 | /* Returns true if the VMA has associated reserve pages */ |
| 447 | static int vma_has_reserves(struct vm_area_struct *vma, long chg) |
| 448 | { |
| 449 | if (vma->vm_flags & VM_NORESERVE) { |
| 450 | /* |
| 451 | * This address is already reserved by other process(chg == 0), |
| 452 | * so, we should decrement reserved count. Without decrementing, |
| 453 | * reserve count remains after releasing inode, because this |
| 454 | * allocated page will go into page cache and is regarded as |
| 455 | * coming from reserved pool in releasing step. Currently, we |
| 456 | * don't have any other solution to deal with this situation |
| 457 | * properly, so add work-around here. |
| 458 | */ |
| 459 | if (vma->vm_flags & VM_MAYSHARE && chg == 0) |
| 460 | return 1; |
| 461 | else |
| 462 | return 0; |
| 463 | } |
| 464 | |
| 465 | /* Shared mappings always use reserves */ |
| 466 | if (vma->vm_flags & VM_MAYSHARE) |
| 467 | return 1; |
| 468 | |
| 469 | /* |
| 470 | * Only the process that called mmap() has reserves for |
| 471 | * private mappings. |
| 472 | */ |
| 473 | if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) |
| 474 | return 1; |
| 475 | |
| 476 | return 0; |
| 477 | } |
| 478 | |
| 479 | static void enqueue_huge_page(struct hstate *h, struct page *page) |
| 480 | { |
| 481 | int nid = page_to_nid(page); |
| 482 | list_move(&page->lru, &h->hugepage_freelists[nid]); |
| 483 | h->free_huge_pages++; |
| 484 | h->free_huge_pages_node[nid]++; |
| 485 | } |
| 486 | |
| 487 | static struct page *dequeue_huge_page_node(struct hstate *h, int nid) |
| 488 | { |
| 489 | struct page *page; |
| 490 | |
| 491 | list_for_each_entry(page, &h->hugepage_freelists[nid], lru) |
| 492 | if (!is_migrate_isolate_page(page)) |
| 493 | break; |
| 494 | /* |
| 495 | * if 'non-isolated free hugepage' not found on the list, |
| 496 | * the allocation fails. |
| 497 | */ |
| 498 | if (&h->hugepage_freelists[nid] == &page->lru) |
| 499 | return NULL; |
| 500 | list_move(&page->lru, &h->hugepage_activelist); |
| 501 | set_page_refcounted(page); |
| 502 | h->free_huge_pages--; |
| 503 | h->free_huge_pages_node[nid]--; |
| 504 | return page; |
| 505 | } |
| 506 | |
| 507 | /* Movability of hugepages depends on migration support. */ |
| 508 | static inline gfp_t htlb_alloc_mask(struct hstate *h) |
| 509 | { |
| 510 | if (hugepages_treat_as_movable || hugepage_migration_support(h)) |
| 511 | return GFP_HIGHUSER_MOVABLE; |
| 512 | else |
| 513 | return GFP_HIGHUSER; |
| 514 | } |
| 515 | |
| 516 | static struct page *dequeue_huge_page_vma(struct hstate *h, |
| 517 | struct vm_area_struct *vma, |
| 518 | unsigned long address, int avoid_reserve, |
| 519 | long chg) |
| 520 | { |
| 521 | struct page *page = NULL; |
| 522 | struct mempolicy *mpol; |
| 523 | nodemask_t *nodemask; |
| 524 | struct zonelist *zonelist; |
| 525 | struct zone *zone; |
| 526 | struct zoneref *z; |
| 527 | unsigned int cpuset_mems_cookie; |
| 528 | |
| 529 | /* |
| 530 | * A child process with MAP_PRIVATE mappings created by their parent |
| 531 | * have no page reserves. This check ensures that reservations are |
| 532 | * not "stolen". The child may still get SIGKILLed |
| 533 | */ |
| 534 | if (!vma_has_reserves(vma, chg) && |
| 535 | h->free_huge_pages - h->resv_huge_pages == 0) |
| 536 | goto err; |
| 537 | |
| 538 | /* If reserves cannot be used, ensure enough pages are in the pool */ |
| 539 | if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0) |
| 540 | goto err; |
| 541 | |
| 542 | retry_cpuset: |
| 543 | cpuset_mems_cookie = get_mems_allowed(); |
| 544 | zonelist = huge_zonelist(vma, address, |
| 545 | htlb_alloc_mask(h), &mpol, &nodemask); |
| 546 | |
| 547 | for_each_zone_zonelist_nodemask(zone, z, zonelist, |
| 548 | MAX_NR_ZONES - 1, nodemask) { |
| 549 | if (cpuset_zone_allowed_softwall(zone, htlb_alloc_mask(h))) { |
| 550 | page = dequeue_huge_page_node(h, zone_to_nid(zone)); |
| 551 | if (page) { |
| 552 | if (avoid_reserve) |
| 553 | break; |
| 554 | if (!vma_has_reserves(vma, chg)) |
| 555 | break; |
| 556 | |
| 557 | SetPagePrivate(page); |
| 558 | h->resv_huge_pages--; |
| 559 | break; |
| 560 | } |
| 561 | } |
| 562 | } |
| 563 | |
| 564 | mpol_cond_put(mpol); |
| 565 | if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page)) |
| 566 | goto retry_cpuset; |
| 567 | return page; |
| 568 | |
| 569 | err: |
| 570 | return NULL; |
| 571 | } |
| 572 | |
| 573 | static void update_and_free_page(struct hstate *h, struct page *page) |
| 574 | { |
| 575 | int i; |
| 576 | |
| 577 | VM_BUG_ON(h->order >= MAX_ORDER); |
| 578 | |
| 579 | h->nr_huge_pages--; |
| 580 | h->nr_huge_pages_node[page_to_nid(page)]--; |
| 581 | for (i = 0; i < pages_per_huge_page(h); i++) { |
| 582 | page[i].flags &= ~(1 << PG_locked | 1 << PG_error | |
| 583 | 1 << PG_referenced | 1 << PG_dirty | |
| 584 | 1 << PG_active | 1 << PG_reserved | |
| 585 | 1 << PG_private | 1 << PG_writeback); |
| 586 | } |
| 587 | VM_BUG_ON(hugetlb_cgroup_from_page(page)); |
| 588 | set_compound_page_dtor(page, NULL); |
| 589 | set_page_refcounted(page); |
| 590 | arch_release_hugepage(page); |
| 591 | __free_pages(page, huge_page_order(h)); |
| 592 | } |
| 593 | |
| 594 | struct hstate *size_to_hstate(unsigned long size) |
| 595 | { |
| 596 | struct hstate *h; |
| 597 | |
| 598 | for_each_hstate(h) { |
| 599 | if (huge_page_size(h) == size) |
| 600 | return h; |
| 601 | } |
| 602 | return NULL; |
| 603 | } |
| 604 | |
| 605 | static void free_huge_page(struct page *page) |
| 606 | { |
| 607 | /* |
| 608 | * Can't pass hstate in here because it is called from the |
| 609 | * compound page destructor. |
| 610 | */ |
| 611 | struct hstate *h = page_hstate(page); |
| 612 | int nid = page_to_nid(page); |
| 613 | struct hugepage_subpool *spool = |
| 614 | (struct hugepage_subpool *)page_private(page); |
| 615 | bool restore_reserve; |
| 616 | |
| 617 | set_page_private(page, 0); |
| 618 | page->mapping = NULL; |
| 619 | BUG_ON(page_count(page)); |
| 620 | BUG_ON(page_mapcount(page)); |
| 621 | restore_reserve = PagePrivate(page); |
| 622 | ClearPagePrivate(page); |
| 623 | |
| 624 | spin_lock(&hugetlb_lock); |
| 625 | hugetlb_cgroup_uncharge_page(hstate_index(h), |
| 626 | pages_per_huge_page(h), page); |
| 627 | if (restore_reserve) |
| 628 | h->resv_huge_pages++; |
| 629 | |
| 630 | if (h->surplus_huge_pages_node[nid] && huge_page_order(h) < MAX_ORDER) { |
| 631 | /* remove the page from active list */ |
| 632 | list_del(&page->lru); |
| 633 | update_and_free_page(h, page); |
| 634 | h->surplus_huge_pages--; |
| 635 | h->surplus_huge_pages_node[nid]--; |
| 636 | } else { |
| 637 | arch_clear_hugepage_flags(page); |
| 638 | enqueue_huge_page(h, page); |
| 639 | } |
| 640 | spin_unlock(&hugetlb_lock); |
| 641 | hugepage_subpool_put_pages(spool, 1); |
| 642 | } |
| 643 | |
| 644 | static void prep_new_huge_page(struct hstate *h, struct page *page, int nid) |
| 645 | { |
| 646 | INIT_LIST_HEAD(&page->lru); |
| 647 | set_compound_page_dtor(page, free_huge_page); |
| 648 | spin_lock(&hugetlb_lock); |
| 649 | set_hugetlb_cgroup(page, NULL); |
| 650 | h->nr_huge_pages++; |
| 651 | h->nr_huge_pages_node[nid]++; |
| 652 | spin_unlock(&hugetlb_lock); |
| 653 | put_page(page); /* free it into the hugepage allocator */ |
| 654 | } |
| 655 | |
| 656 | static void prep_compound_gigantic_page(struct page *page, unsigned long order) |
| 657 | { |
| 658 | int i; |
| 659 | int nr_pages = 1 << order; |
| 660 | struct page *p = page + 1; |
| 661 | |
| 662 | /* we rely on prep_new_huge_page to set the destructor */ |
| 663 | set_compound_order(page, order); |
| 664 | __SetPageHead(page); |
| 665 | __ClearPageReserved(page); |
| 666 | for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { |
| 667 | __SetPageTail(p); |
| 668 | /* |
| 669 | * For gigantic hugepages allocated through bootmem at |
| 670 | * boot, it's safer to be consistent with the not-gigantic |
| 671 | * hugepages and clear the PG_reserved bit from all tail pages |
| 672 | * too. Otherwse drivers using get_user_pages() to access tail |
| 673 | * pages may get the reference counting wrong if they see |
| 674 | * PG_reserved set on a tail page (despite the head page not |
| 675 | * having PG_reserved set). Enforcing this consistency between |
| 676 | * head and tail pages allows drivers to optimize away a check |
| 677 | * on the head page when they need know if put_page() is needed |
| 678 | * after get_user_pages(). |
| 679 | */ |
| 680 | __ClearPageReserved(p); |
| 681 | set_page_count(p, 0); |
| 682 | p->first_page = page; |
| 683 | } |
| 684 | } |
| 685 | |
| 686 | /* |
| 687 | * PageHuge() only returns true for hugetlbfs pages, but not for normal or |
| 688 | * transparent huge pages. See the PageTransHuge() documentation for more |
| 689 | * details. |
| 690 | */ |
| 691 | int PageHuge(struct page *page) |
| 692 | { |
| 693 | if (!PageCompound(page)) |
| 694 | return 0; |
| 695 | |
| 696 | page = compound_head(page); |
| 697 | return get_compound_page_dtor(page) == free_huge_page; |
| 698 | } |
| 699 | EXPORT_SYMBOL_GPL(PageHuge); |
| 700 | |
| 701 | /* |
| 702 | * PageHeadHuge() only returns true for hugetlbfs head page, but not for |
| 703 | * normal or transparent huge pages. |
| 704 | */ |
| 705 | int PageHeadHuge(struct page *page_head) |
| 706 | { |
| 707 | if (!PageHead(page_head)) |
| 708 | return 0; |
| 709 | |
| 710 | return get_compound_page_dtor(page_head) == free_huge_page; |
| 711 | } |
| 712 | |
| 713 | pgoff_t __basepage_index(struct page *page) |
| 714 | { |
| 715 | struct page *page_head = compound_head(page); |
| 716 | pgoff_t index = page_index(page_head); |
| 717 | unsigned long compound_idx; |
| 718 | |
| 719 | if (!PageHuge(page_head)) |
| 720 | return page_index(page); |
| 721 | |
| 722 | if (compound_order(page_head) >= MAX_ORDER) |
| 723 | compound_idx = page_to_pfn(page) - page_to_pfn(page_head); |
| 724 | else |
| 725 | compound_idx = page - page_head; |
| 726 | |
| 727 | return (index << compound_order(page_head)) + compound_idx; |
| 728 | } |
| 729 | |
| 730 | static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid) |
| 731 | { |
| 732 | struct page *page; |
| 733 | |
| 734 | if (h->order >= MAX_ORDER) |
| 735 | return NULL; |
| 736 | |
| 737 | page = alloc_pages_exact_node(nid, |
| 738 | htlb_alloc_mask(h)|__GFP_COMP|__GFP_THISNODE| |
| 739 | __GFP_REPEAT|__GFP_NOWARN, |
| 740 | huge_page_order(h)); |
| 741 | if (page) { |
| 742 | if (arch_prepare_hugepage(page)) { |
| 743 | __free_pages(page, huge_page_order(h)); |
| 744 | return NULL; |
| 745 | } |
| 746 | prep_new_huge_page(h, page, nid); |
| 747 | } |
| 748 | |
| 749 | return page; |
| 750 | } |
| 751 | |
| 752 | /* |
| 753 | * common helper functions for hstate_next_node_to_{alloc|free}. |
| 754 | * We may have allocated or freed a huge page based on a different |
| 755 | * nodes_allowed previously, so h->next_node_to_{alloc|free} might |
| 756 | * be outside of *nodes_allowed. Ensure that we use an allowed |
| 757 | * node for alloc or free. |
| 758 | */ |
| 759 | static int next_node_allowed(int nid, nodemask_t *nodes_allowed) |
| 760 | { |
| 761 | nid = next_node(nid, *nodes_allowed); |
| 762 | if (nid == MAX_NUMNODES) |
| 763 | nid = first_node(*nodes_allowed); |
| 764 | VM_BUG_ON(nid >= MAX_NUMNODES); |
| 765 | |
| 766 | return nid; |
| 767 | } |
| 768 | |
| 769 | static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed) |
| 770 | { |
| 771 | if (!node_isset(nid, *nodes_allowed)) |
| 772 | nid = next_node_allowed(nid, nodes_allowed); |
| 773 | return nid; |
| 774 | } |
| 775 | |
| 776 | /* |
| 777 | * returns the previously saved node ["this node"] from which to |
| 778 | * allocate a persistent huge page for the pool and advance the |
| 779 | * next node from which to allocate, handling wrap at end of node |
| 780 | * mask. |
| 781 | */ |
| 782 | static int hstate_next_node_to_alloc(struct hstate *h, |
| 783 | nodemask_t *nodes_allowed) |
| 784 | { |
| 785 | int nid; |
| 786 | |
| 787 | VM_BUG_ON(!nodes_allowed); |
| 788 | |
| 789 | nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed); |
| 790 | h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed); |
| 791 | |
| 792 | return nid; |
| 793 | } |
| 794 | |
| 795 | /* |
| 796 | * helper for free_pool_huge_page() - return the previously saved |
| 797 | * node ["this node"] from which to free a huge page. Advance the |
| 798 | * next node id whether or not we find a free huge page to free so |
| 799 | * that the next attempt to free addresses the next node. |
| 800 | */ |
| 801 | static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed) |
| 802 | { |
| 803 | int nid; |
| 804 | |
| 805 | VM_BUG_ON(!nodes_allowed); |
| 806 | |
| 807 | nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed); |
| 808 | h->next_nid_to_free = next_node_allowed(nid, nodes_allowed); |
| 809 | |
| 810 | return nid; |
| 811 | } |
| 812 | |
| 813 | #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \ |
| 814 | for (nr_nodes = nodes_weight(*mask); \ |
| 815 | nr_nodes > 0 && \ |
| 816 | ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \ |
| 817 | nr_nodes--) |
| 818 | |
| 819 | #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \ |
| 820 | for (nr_nodes = nodes_weight(*mask); \ |
| 821 | nr_nodes > 0 && \ |
| 822 | ((node = hstate_next_node_to_free(hs, mask)) || 1); \ |
| 823 | nr_nodes--) |
| 824 | |
| 825 | static int alloc_fresh_huge_page(struct hstate *h, nodemask_t *nodes_allowed) |
| 826 | { |
| 827 | struct page *page; |
| 828 | int nr_nodes, node; |
| 829 | int ret = 0; |
| 830 | |
| 831 | for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { |
| 832 | page = alloc_fresh_huge_page_node(h, node); |
| 833 | if (page) { |
| 834 | ret = 1; |
| 835 | break; |
| 836 | } |
| 837 | } |
| 838 | |
| 839 | if (ret) |
| 840 | count_vm_event(HTLB_BUDDY_PGALLOC); |
| 841 | else |
| 842 | count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); |
| 843 | |
| 844 | return ret; |
| 845 | } |
| 846 | |
| 847 | /* |
| 848 | * Free huge page from pool from next node to free. |
| 849 | * Attempt to keep persistent huge pages more or less |
| 850 | * balanced over allowed nodes. |
| 851 | * Called with hugetlb_lock locked. |
| 852 | */ |
| 853 | static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed, |
| 854 | bool acct_surplus) |
| 855 | { |
| 856 | int nr_nodes, node; |
| 857 | int ret = 0; |
| 858 | |
| 859 | for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { |
| 860 | /* |
| 861 | * If we're returning unused surplus pages, only examine |
| 862 | * nodes with surplus pages. |
| 863 | */ |
| 864 | if ((!acct_surplus || h->surplus_huge_pages_node[node]) && |
| 865 | !list_empty(&h->hugepage_freelists[node])) { |
| 866 | struct page *page = |
| 867 | list_entry(h->hugepage_freelists[node].next, |
| 868 | struct page, lru); |
| 869 | list_del(&page->lru); |
| 870 | h->free_huge_pages--; |
| 871 | h->free_huge_pages_node[node]--; |
| 872 | if (acct_surplus) { |
| 873 | h->surplus_huge_pages--; |
| 874 | h->surplus_huge_pages_node[node]--; |
| 875 | } |
| 876 | update_and_free_page(h, page); |
| 877 | ret = 1; |
| 878 | break; |
| 879 | } |
| 880 | } |
| 881 | |
| 882 | return ret; |
| 883 | } |
| 884 | |
| 885 | /* |
| 886 | * Dissolve a given free hugepage into free buddy pages. This function does |
| 887 | * nothing for in-use (including surplus) hugepages. |
| 888 | */ |
| 889 | static void dissolve_free_huge_page(struct page *page) |
| 890 | { |
| 891 | spin_lock(&hugetlb_lock); |
| 892 | if (PageHuge(page) && !page_count(page)) { |
| 893 | struct hstate *h = page_hstate(page); |
| 894 | int nid = page_to_nid(page); |
| 895 | list_del(&page->lru); |
| 896 | h->free_huge_pages--; |
| 897 | h->free_huge_pages_node[nid]--; |
| 898 | update_and_free_page(h, page); |
| 899 | } |
| 900 | spin_unlock(&hugetlb_lock); |
| 901 | } |
| 902 | |
| 903 | /* |
| 904 | * Dissolve free hugepages in a given pfn range. Used by memory hotplug to |
| 905 | * make specified memory blocks removable from the system. |
| 906 | * Note that start_pfn should aligned with (minimum) hugepage size. |
| 907 | */ |
| 908 | void dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn) |
| 909 | { |
| 910 | unsigned int order = 8 * sizeof(void *); |
| 911 | unsigned long pfn; |
| 912 | struct hstate *h; |
| 913 | |
| 914 | /* Set scan step to minimum hugepage size */ |
| 915 | for_each_hstate(h) |
| 916 | if (order > huge_page_order(h)) |
| 917 | order = huge_page_order(h); |
| 918 | VM_BUG_ON(!IS_ALIGNED(start_pfn, 1 << order)); |
| 919 | for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << order) |
| 920 | dissolve_free_huge_page(pfn_to_page(pfn)); |
| 921 | } |
| 922 | |
| 923 | static struct page *alloc_buddy_huge_page(struct hstate *h, int nid) |
| 924 | { |
| 925 | struct page *page; |
| 926 | unsigned int r_nid; |
| 927 | |
| 928 | if (h->order >= MAX_ORDER) |
| 929 | return NULL; |
| 930 | |
| 931 | /* |
| 932 | * Assume we will successfully allocate the surplus page to |
| 933 | * prevent racing processes from causing the surplus to exceed |
| 934 | * overcommit |
| 935 | * |
| 936 | * This however introduces a different race, where a process B |
| 937 | * tries to grow the static hugepage pool while alloc_pages() is |
| 938 | * called by process A. B will only examine the per-node |
| 939 | * counters in determining if surplus huge pages can be |
| 940 | * converted to normal huge pages in adjust_pool_surplus(). A |
| 941 | * won't be able to increment the per-node counter, until the |
| 942 | * lock is dropped by B, but B doesn't drop hugetlb_lock until |
| 943 | * no more huge pages can be converted from surplus to normal |
| 944 | * state (and doesn't try to convert again). Thus, we have a |
| 945 | * case where a surplus huge page exists, the pool is grown, and |
| 946 | * the surplus huge page still exists after, even though it |
| 947 | * should just have been converted to a normal huge page. This |
| 948 | * does not leak memory, though, as the hugepage will be freed |
| 949 | * once it is out of use. It also does not allow the counters to |
| 950 | * go out of whack in adjust_pool_surplus() as we don't modify |
| 951 | * the node values until we've gotten the hugepage and only the |
| 952 | * per-node value is checked there. |
| 953 | */ |
| 954 | spin_lock(&hugetlb_lock); |
| 955 | if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) { |
| 956 | spin_unlock(&hugetlb_lock); |
| 957 | return NULL; |
| 958 | } else { |
| 959 | h->nr_huge_pages++; |
| 960 | h->surplus_huge_pages++; |
| 961 | } |
| 962 | spin_unlock(&hugetlb_lock); |
| 963 | |
| 964 | if (nid == NUMA_NO_NODE) |
| 965 | page = alloc_pages(htlb_alloc_mask(h)|__GFP_COMP| |
| 966 | __GFP_REPEAT|__GFP_NOWARN, |
| 967 | huge_page_order(h)); |
| 968 | else |
| 969 | page = alloc_pages_exact_node(nid, |
| 970 | htlb_alloc_mask(h)|__GFP_COMP|__GFP_THISNODE| |
| 971 | __GFP_REPEAT|__GFP_NOWARN, huge_page_order(h)); |
| 972 | |
| 973 | if (page && arch_prepare_hugepage(page)) { |
| 974 | __free_pages(page, huge_page_order(h)); |
| 975 | page = NULL; |
| 976 | } |
| 977 | |
| 978 | spin_lock(&hugetlb_lock); |
| 979 | if (page) { |
| 980 | INIT_LIST_HEAD(&page->lru); |
| 981 | r_nid = page_to_nid(page); |
| 982 | set_compound_page_dtor(page, free_huge_page); |
| 983 | set_hugetlb_cgroup(page, NULL); |
| 984 | /* |
| 985 | * We incremented the global counters already |
| 986 | */ |
| 987 | h->nr_huge_pages_node[r_nid]++; |
| 988 | h->surplus_huge_pages_node[r_nid]++; |
| 989 | __count_vm_event(HTLB_BUDDY_PGALLOC); |
| 990 | } else { |
| 991 | h->nr_huge_pages--; |
| 992 | h->surplus_huge_pages--; |
| 993 | __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); |
| 994 | } |
| 995 | spin_unlock(&hugetlb_lock); |
| 996 | |
| 997 | return page; |
| 998 | } |
| 999 | |
| 1000 | /* |
| 1001 | * This allocation function is useful in the context where vma is irrelevant. |
| 1002 | * E.g. soft-offlining uses this function because it only cares physical |
| 1003 | * address of error page. |
| 1004 | */ |
| 1005 | struct page *alloc_huge_page_node(struct hstate *h, int nid) |
| 1006 | { |
| 1007 | struct page *page = NULL; |
| 1008 | |
| 1009 | spin_lock(&hugetlb_lock); |
| 1010 | if (h->free_huge_pages - h->resv_huge_pages > 0) |
| 1011 | page = dequeue_huge_page_node(h, nid); |
| 1012 | spin_unlock(&hugetlb_lock); |
| 1013 | |
| 1014 | if (!page) |
| 1015 | page = alloc_buddy_huge_page(h, nid); |
| 1016 | |
| 1017 | return page; |
| 1018 | } |
| 1019 | |
| 1020 | /* |
| 1021 | * Increase the hugetlb pool such that it can accommodate a reservation |
| 1022 | * of size 'delta'. |
| 1023 | */ |
| 1024 | static int gather_surplus_pages(struct hstate *h, int delta) |
| 1025 | { |
| 1026 | struct list_head surplus_list; |
| 1027 | struct page *page, *tmp; |
| 1028 | int ret, i; |
| 1029 | int needed, allocated; |
| 1030 | bool alloc_ok = true; |
| 1031 | |
| 1032 | needed = (h->resv_huge_pages + delta) - h->free_huge_pages; |
| 1033 | if (needed <= 0) { |
| 1034 | h->resv_huge_pages += delta; |
| 1035 | return 0; |
| 1036 | } |
| 1037 | |
| 1038 | allocated = 0; |
| 1039 | INIT_LIST_HEAD(&surplus_list); |
| 1040 | |
| 1041 | ret = -ENOMEM; |
| 1042 | retry: |
| 1043 | spin_unlock(&hugetlb_lock); |
| 1044 | for (i = 0; i < needed; i++) { |
| 1045 | page = alloc_buddy_huge_page(h, NUMA_NO_NODE); |
| 1046 | if (!page) { |
| 1047 | alloc_ok = false; |
| 1048 | break; |
| 1049 | } |
| 1050 | list_add(&page->lru, &surplus_list); |
| 1051 | } |
| 1052 | allocated += i; |
| 1053 | |
| 1054 | /* |
| 1055 | * After retaking hugetlb_lock, we need to recalculate 'needed' |
| 1056 | * because either resv_huge_pages or free_huge_pages may have changed. |
| 1057 | */ |
| 1058 | spin_lock(&hugetlb_lock); |
| 1059 | needed = (h->resv_huge_pages + delta) - |
| 1060 | (h->free_huge_pages + allocated); |
| 1061 | if (needed > 0) { |
| 1062 | if (alloc_ok) |
| 1063 | goto retry; |
| 1064 | /* |
| 1065 | * We were not able to allocate enough pages to |
| 1066 | * satisfy the entire reservation so we free what |
| 1067 | * we've allocated so far. |
| 1068 | */ |
| 1069 | goto free; |
| 1070 | } |
| 1071 | /* |
| 1072 | * The surplus_list now contains _at_least_ the number of extra pages |
| 1073 | * needed to accommodate the reservation. Add the appropriate number |
| 1074 | * of pages to the hugetlb pool and free the extras back to the buddy |
| 1075 | * allocator. Commit the entire reservation here to prevent another |
| 1076 | * process from stealing the pages as they are added to the pool but |
| 1077 | * before they are reserved. |
| 1078 | */ |
| 1079 | needed += allocated; |
| 1080 | h->resv_huge_pages += delta; |
| 1081 | ret = 0; |
| 1082 | |
| 1083 | /* Free the needed pages to the hugetlb pool */ |
| 1084 | list_for_each_entry_safe(page, tmp, &surplus_list, lru) { |
| 1085 | if ((--needed) < 0) |
| 1086 | break; |
| 1087 | /* |
| 1088 | * This page is now managed by the hugetlb allocator and has |
| 1089 | * no users -- drop the buddy allocator's reference. |
| 1090 | */ |
| 1091 | put_page_testzero(page); |
| 1092 | VM_BUG_ON(page_count(page)); |
| 1093 | enqueue_huge_page(h, page); |
| 1094 | } |
| 1095 | free: |
| 1096 | spin_unlock(&hugetlb_lock); |
| 1097 | |
| 1098 | /* Free unnecessary surplus pages to the buddy allocator */ |
| 1099 | list_for_each_entry_safe(page, tmp, &surplus_list, lru) |
| 1100 | put_page(page); |
| 1101 | spin_lock(&hugetlb_lock); |
| 1102 | |
| 1103 | return ret; |
| 1104 | } |
| 1105 | |
| 1106 | /* |
| 1107 | * When releasing a hugetlb pool reservation, any surplus pages that were |
| 1108 | * allocated to satisfy the reservation must be explicitly freed if they were |
| 1109 | * never used. |
| 1110 | * Called with hugetlb_lock held. |
| 1111 | */ |
| 1112 | static void return_unused_surplus_pages(struct hstate *h, |
| 1113 | unsigned long unused_resv_pages) |
| 1114 | { |
| 1115 | unsigned long nr_pages; |
| 1116 | |
| 1117 | /* Uncommit the reservation */ |
| 1118 | h->resv_huge_pages -= unused_resv_pages; |
| 1119 | |
| 1120 | /* Cannot return gigantic pages currently */ |
| 1121 | if (h->order >= MAX_ORDER) |
| 1122 | return; |
| 1123 | |
| 1124 | nr_pages = min(unused_resv_pages, h->surplus_huge_pages); |
| 1125 | |
| 1126 | /* |
| 1127 | * We want to release as many surplus pages as possible, spread |
| 1128 | * evenly across all nodes with memory. Iterate across these nodes |
| 1129 | * until we can no longer free unreserved surplus pages. This occurs |
| 1130 | * when the nodes with surplus pages have no free pages. |
| 1131 | * free_pool_huge_page() will balance the the freed pages across the |
| 1132 | * on-line nodes with memory and will handle the hstate accounting. |
| 1133 | */ |
| 1134 | while (nr_pages--) { |
| 1135 | if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1)) |
| 1136 | break; |
| 1137 | } |
| 1138 | } |
| 1139 | |
| 1140 | /* |
| 1141 | * Determine if the huge page at addr within the vma has an associated |
| 1142 | * reservation. Where it does not we will need to logically increase |
| 1143 | * reservation and actually increase subpool usage before an allocation |
| 1144 | * can occur. Where any new reservation would be required the |
| 1145 | * reservation change is prepared, but not committed. Once the page |
| 1146 | * has been allocated from the subpool and instantiated the change should |
| 1147 | * be committed via vma_commit_reservation. No action is required on |
| 1148 | * failure. |
| 1149 | */ |
| 1150 | static long vma_needs_reservation(struct hstate *h, |
| 1151 | struct vm_area_struct *vma, unsigned long addr) |
| 1152 | { |
| 1153 | struct address_space *mapping = vma->vm_file->f_mapping; |
| 1154 | struct inode *inode = mapping->host; |
| 1155 | |
| 1156 | if (vma->vm_flags & VM_MAYSHARE) { |
| 1157 | pgoff_t idx = vma_hugecache_offset(h, vma, addr); |
| 1158 | return region_chg(&inode->i_mapping->private_list, |
| 1159 | idx, idx + 1); |
| 1160 | |
| 1161 | } else if (!is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { |
| 1162 | return 1; |
| 1163 | |
| 1164 | } else { |
| 1165 | long err; |
| 1166 | pgoff_t idx = vma_hugecache_offset(h, vma, addr); |
| 1167 | struct resv_map *resv = vma_resv_map(vma); |
| 1168 | |
| 1169 | err = region_chg(&resv->regions, idx, idx + 1); |
| 1170 | if (err < 0) |
| 1171 | return err; |
| 1172 | return 0; |
| 1173 | } |
| 1174 | } |
| 1175 | static void vma_commit_reservation(struct hstate *h, |
| 1176 | struct vm_area_struct *vma, unsigned long addr) |
| 1177 | { |
| 1178 | struct address_space *mapping = vma->vm_file->f_mapping; |
| 1179 | struct inode *inode = mapping->host; |
| 1180 | |
| 1181 | if (vma->vm_flags & VM_MAYSHARE) { |
| 1182 | pgoff_t idx = vma_hugecache_offset(h, vma, addr); |
| 1183 | region_add(&inode->i_mapping->private_list, idx, idx + 1); |
| 1184 | |
| 1185 | } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { |
| 1186 | pgoff_t idx = vma_hugecache_offset(h, vma, addr); |
| 1187 | struct resv_map *resv = vma_resv_map(vma); |
| 1188 | |
| 1189 | /* Mark this page used in the map. */ |
| 1190 | region_add(&resv->regions, idx, idx + 1); |
| 1191 | } |
| 1192 | } |
| 1193 | |
| 1194 | static struct page *alloc_huge_page(struct vm_area_struct *vma, |
| 1195 | unsigned long addr, int avoid_reserve) |
| 1196 | { |
| 1197 | struct hugepage_subpool *spool = subpool_vma(vma); |
| 1198 | struct hstate *h = hstate_vma(vma); |
| 1199 | struct page *page; |
| 1200 | long chg; |
| 1201 | int ret, idx; |
| 1202 | struct hugetlb_cgroup *h_cg; |
| 1203 | |
| 1204 | idx = hstate_index(h); |
| 1205 | /* |
| 1206 | * Processes that did not create the mapping will have no |
| 1207 | * reserves and will not have accounted against subpool |
| 1208 | * limit. Check that the subpool limit can be made before |
| 1209 | * satisfying the allocation MAP_NORESERVE mappings may also |
| 1210 | * need pages and subpool limit allocated allocated if no reserve |
| 1211 | * mapping overlaps. |
| 1212 | */ |
| 1213 | chg = vma_needs_reservation(h, vma, addr); |
| 1214 | if (chg < 0) |
| 1215 | return ERR_PTR(-ENOMEM); |
| 1216 | if (chg || avoid_reserve) |
| 1217 | if (hugepage_subpool_get_pages(spool, 1)) |
| 1218 | return ERR_PTR(-ENOSPC); |
| 1219 | |
| 1220 | ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg); |
| 1221 | if (ret) { |
| 1222 | if (chg || avoid_reserve) |
| 1223 | hugepage_subpool_put_pages(spool, 1); |
| 1224 | return ERR_PTR(-ENOSPC); |
| 1225 | } |
| 1226 | spin_lock(&hugetlb_lock); |
| 1227 | page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, chg); |
| 1228 | if (!page) { |
| 1229 | spin_unlock(&hugetlb_lock); |
| 1230 | page = alloc_buddy_huge_page(h, NUMA_NO_NODE); |
| 1231 | if (!page) { |
| 1232 | hugetlb_cgroup_uncharge_cgroup(idx, |
| 1233 | pages_per_huge_page(h), |
| 1234 | h_cg); |
| 1235 | if (chg || avoid_reserve) |
| 1236 | hugepage_subpool_put_pages(spool, 1); |
| 1237 | return ERR_PTR(-ENOSPC); |
| 1238 | } |
| 1239 | spin_lock(&hugetlb_lock); |
| 1240 | list_move(&page->lru, &h->hugepage_activelist); |
| 1241 | /* Fall through */ |
| 1242 | } |
| 1243 | hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page); |
| 1244 | spin_unlock(&hugetlb_lock); |
| 1245 | |
| 1246 | set_page_private(page, (unsigned long)spool); |
| 1247 | |
| 1248 | vma_commit_reservation(h, vma, addr); |
| 1249 | return page; |
| 1250 | } |
| 1251 | |
| 1252 | /* |
| 1253 | * alloc_huge_page()'s wrapper which simply returns the page if allocation |
| 1254 | * succeeds, otherwise NULL. This function is called from new_vma_page(), |
| 1255 | * where no ERR_VALUE is expected to be returned. |
| 1256 | */ |
| 1257 | struct page *alloc_huge_page_noerr(struct vm_area_struct *vma, |
| 1258 | unsigned long addr, int avoid_reserve) |
| 1259 | { |
| 1260 | struct page *page = alloc_huge_page(vma, addr, avoid_reserve); |
| 1261 | if (IS_ERR(page)) |
| 1262 | page = NULL; |
| 1263 | return page; |
| 1264 | } |
| 1265 | |
| 1266 | int __weak alloc_bootmem_huge_page(struct hstate *h) |
| 1267 | { |
| 1268 | struct huge_bootmem_page *m; |
| 1269 | int nr_nodes, node; |
| 1270 | |
| 1271 | for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) { |
| 1272 | void *addr; |
| 1273 | |
| 1274 | addr = __alloc_bootmem_node_nopanic(NODE_DATA(node), |
| 1275 | huge_page_size(h), huge_page_size(h), 0); |
| 1276 | |
| 1277 | if (addr) { |
| 1278 | /* |
| 1279 | * Use the beginning of the huge page to store the |
| 1280 | * huge_bootmem_page struct (until gather_bootmem |
| 1281 | * puts them into the mem_map). |
| 1282 | */ |
| 1283 | m = addr; |
| 1284 | goto found; |
| 1285 | } |
| 1286 | } |
| 1287 | return 0; |
| 1288 | |
| 1289 | found: |
| 1290 | BUG_ON((unsigned long)virt_to_phys(m) & (huge_page_size(h) - 1)); |
| 1291 | /* Put them into a private list first because mem_map is not up yet */ |
| 1292 | list_add(&m->list, &huge_boot_pages); |
| 1293 | m->hstate = h; |
| 1294 | return 1; |
| 1295 | } |
| 1296 | |
| 1297 | static void prep_compound_huge_page(struct page *page, int order) |
| 1298 | { |
| 1299 | if (unlikely(order > (MAX_ORDER - 1))) |
| 1300 | prep_compound_gigantic_page(page, order); |
| 1301 | else |
| 1302 | prep_compound_page(page, order); |
| 1303 | } |
| 1304 | |
| 1305 | /* Put bootmem huge pages into the standard lists after mem_map is up */ |
| 1306 | static void __init gather_bootmem_prealloc(void) |
| 1307 | { |
| 1308 | struct huge_bootmem_page *m; |
| 1309 | |
| 1310 | list_for_each_entry(m, &huge_boot_pages, list) { |
| 1311 | struct hstate *h = m->hstate; |
| 1312 | struct page *page; |
| 1313 | |
| 1314 | #ifdef CONFIG_HIGHMEM |
| 1315 | page = pfn_to_page(m->phys >> PAGE_SHIFT); |
| 1316 | free_bootmem_late((unsigned long)m, |
| 1317 | sizeof(struct huge_bootmem_page)); |
| 1318 | #else |
| 1319 | page = virt_to_page(m); |
| 1320 | #endif |
| 1321 | WARN_ON(page_count(page) != 1); |
| 1322 | prep_compound_huge_page(page, h->order); |
| 1323 | WARN_ON(PageReserved(page)); |
| 1324 | prep_new_huge_page(h, page, page_to_nid(page)); |
| 1325 | /* |
| 1326 | * If we had gigantic hugepages allocated at boot time, we need |
| 1327 | * to restore the 'stolen' pages to totalram_pages in order to |
| 1328 | * fix confusing memory reports from free(1) and another |
| 1329 | * side-effects, like CommitLimit going negative. |
| 1330 | */ |
| 1331 | if (h->order > (MAX_ORDER - 1)) |
| 1332 | adjust_managed_page_count(page, 1 << h->order); |
| 1333 | } |
| 1334 | } |
| 1335 | |
| 1336 | static void __init hugetlb_hstate_alloc_pages(struct hstate *h) |
| 1337 | { |
| 1338 | unsigned long i; |
| 1339 | |
| 1340 | for (i = 0; i < h->max_huge_pages; ++i) { |
| 1341 | if (h->order >= MAX_ORDER) { |
| 1342 | if (!alloc_bootmem_huge_page(h)) |
| 1343 | break; |
| 1344 | } else if (!alloc_fresh_huge_page(h, |
| 1345 | &node_states[N_MEMORY])) |
| 1346 | break; |
| 1347 | } |
| 1348 | h->max_huge_pages = i; |
| 1349 | } |
| 1350 | |
| 1351 | static void __init hugetlb_init_hstates(void) |
| 1352 | { |
| 1353 | struct hstate *h; |
| 1354 | |
| 1355 | for_each_hstate(h) { |
| 1356 | /* oversize hugepages were init'ed in early boot */ |
| 1357 | if (h->order < MAX_ORDER) |
| 1358 | hugetlb_hstate_alloc_pages(h); |
| 1359 | } |
| 1360 | } |
| 1361 | |
| 1362 | static char * __init memfmt(char *buf, unsigned long n) |
| 1363 | { |
| 1364 | if (n >= (1UL << 30)) |
| 1365 | sprintf(buf, "%lu GB", n >> 30); |
| 1366 | else if (n >= (1UL << 20)) |
| 1367 | sprintf(buf, "%lu MB", n >> 20); |
| 1368 | else |
| 1369 | sprintf(buf, "%lu KB", n >> 10); |
| 1370 | return buf; |
| 1371 | } |
| 1372 | |
| 1373 | static void __init report_hugepages(void) |
| 1374 | { |
| 1375 | struct hstate *h; |
| 1376 | |
| 1377 | for_each_hstate(h) { |
| 1378 | char buf[32]; |
| 1379 | pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n", |
| 1380 | memfmt(buf, huge_page_size(h)), |
| 1381 | h->free_huge_pages); |
| 1382 | } |
| 1383 | } |
| 1384 | |
| 1385 | #ifdef CONFIG_HIGHMEM |
| 1386 | static void try_to_free_low(struct hstate *h, unsigned long count, |
| 1387 | nodemask_t *nodes_allowed) |
| 1388 | { |
| 1389 | int i; |
| 1390 | |
| 1391 | if (h->order >= MAX_ORDER) |
| 1392 | return; |
| 1393 | |
| 1394 | for_each_node_mask(i, *nodes_allowed) { |
| 1395 | struct page *page, *next; |
| 1396 | struct list_head *freel = &h->hugepage_freelists[i]; |
| 1397 | list_for_each_entry_safe(page, next, freel, lru) { |
| 1398 | if (count >= h->nr_huge_pages) |
| 1399 | return; |
| 1400 | if (PageHighMem(page)) |
| 1401 | continue; |
| 1402 | list_del(&page->lru); |
| 1403 | update_and_free_page(h, page); |
| 1404 | h->free_huge_pages--; |
| 1405 | h->free_huge_pages_node[page_to_nid(page)]--; |
| 1406 | } |
| 1407 | } |
| 1408 | } |
| 1409 | #else |
| 1410 | static inline void try_to_free_low(struct hstate *h, unsigned long count, |
| 1411 | nodemask_t *nodes_allowed) |
| 1412 | { |
| 1413 | } |
| 1414 | #endif |
| 1415 | |
| 1416 | /* |
| 1417 | * Increment or decrement surplus_huge_pages. Keep node-specific counters |
| 1418 | * balanced by operating on them in a round-robin fashion. |
| 1419 | * Returns 1 if an adjustment was made. |
| 1420 | */ |
| 1421 | static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed, |
| 1422 | int delta) |
| 1423 | { |
| 1424 | int nr_nodes, node; |
| 1425 | |
| 1426 | VM_BUG_ON(delta != -1 && delta != 1); |
| 1427 | |
| 1428 | if (delta < 0) { |
| 1429 | for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { |
| 1430 | if (h->surplus_huge_pages_node[node]) |
| 1431 | goto found; |
| 1432 | } |
| 1433 | } else { |
| 1434 | for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { |
| 1435 | if (h->surplus_huge_pages_node[node] < |
| 1436 | h->nr_huge_pages_node[node]) |
| 1437 | goto found; |
| 1438 | } |
| 1439 | } |
| 1440 | return 0; |
| 1441 | |
| 1442 | found: |
| 1443 | h->surplus_huge_pages += delta; |
| 1444 | h->surplus_huge_pages_node[node] += delta; |
| 1445 | return 1; |
| 1446 | } |
| 1447 | |
| 1448 | #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages) |
| 1449 | static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count, |
| 1450 | nodemask_t *nodes_allowed) |
| 1451 | { |
| 1452 | unsigned long min_count, ret; |
| 1453 | |
| 1454 | if (h->order >= MAX_ORDER) |
| 1455 | return h->max_huge_pages; |
| 1456 | |
| 1457 | /* |
| 1458 | * Increase the pool size |
| 1459 | * First take pages out of surplus state. Then make up the |
| 1460 | * remaining difference by allocating fresh huge pages. |
| 1461 | * |
| 1462 | * We might race with alloc_buddy_huge_page() here and be unable |
| 1463 | * to convert a surplus huge page to a normal huge page. That is |
| 1464 | * not critical, though, it just means the overall size of the |
| 1465 | * pool might be one hugepage larger than it needs to be, but |
| 1466 | * within all the constraints specified by the sysctls. |
| 1467 | */ |
| 1468 | spin_lock(&hugetlb_lock); |
| 1469 | while (h->surplus_huge_pages && count > persistent_huge_pages(h)) { |
| 1470 | if (!adjust_pool_surplus(h, nodes_allowed, -1)) |
| 1471 | break; |
| 1472 | } |
| 1473 | |
| 1474 | while (count > persistent_huge_pages(h)) { |
| 1475 | /* |
| 1476 | * If this allocation races such that we no longer need the |
| 1477 | * page, free_huge_page will handle it by freeing the page |
| 1478 | * and reducing the surplus. |
| 1479 | */ |
| 1480 | spin_unlock(&hugetlb_lock); |
| 1481 | ret = alloc_fresh_huge_page(h, nodes_allowed); |
| 1482 | spin_lock(&hugetlb_lock); |
| 1483 | if (!ret) |
| 1484 | goto out; |
| 1485 | |
| 1486 | /* Bail for signals. Probably ctrl-c from user */ |
| 1487 | if (signal_pending(current)) |
| 1488 | goto out; |
| 1489 | } |
| 1490 | |
| 1491 | /* |
| 1492 | * Decrease the pool size |
| 1493 | * First return free pages to the buddy allocator (being careful |
| 1494 | * to keep enough around to satisfy reservations). Then place |
| 1495 | * pages into surplus state as needed so the pool will shrink |
| 1496 | * to the desired size as pages become free. |
| 1497 | * |
| 1498 | * By placing pages into the surplus state independent of the |
| 1499 | * overcommit value, we are allowing the surplus pool size to |
| 1500 | * exceed overcommit. There are few sane options here. Since |
| 1501 | * alloc_buddy_huge_page() is checking the global counter, |
| 1502 | * though, we'll note that we're not allowed to exceed surplus |
| 1503 | * and won't grow the pool anywhere else. Not until one of the |
| 1504 | * sysctls are changed, or the surplus pages go out of use. |
| 1505 | */ |
| 1506 | min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages; |
| 1507 | min_count = max(count, min_count); |
| 1508 | try_to_free_low(h, min_count, nodes_allowed); |
| 1509 | while (min_count < persistent_huge_pages(h)) { |
| 1510 | if (!free_pool_huge_page(h, nodes_allowed, 0)) |
| 1511 | break; |
| 1512 | } |
| 1513 | while (count < persistent_huge_pages(h)) { |
| 1514 | if (!adjust_pool_surplus(h, nodes_allowed, 1)) |
| 1515 | break; |
| 1516 | } |
| 1517 | out: |
| 1518 | ret = persistent_huge_pages(h); |
| 1519 | spin_unlock(&hugetlb_lock); |
| 1520 | return ret; |
| 1521 | } |
| 1522 | |
| 1523 | #define HSTATE_ATTR_RO(_name) \ |
| 1524 | static struct kobj_attribute _name##_attr = __ATTR_RO(_name) |
| 1525 | |
| 1526 | #define HSTATE_ATTR(_name) \ |
| 1527 | static struct kobj_attribute _name##_attr = \ |
| 1528 | __ATTR(_name, 0644, _name##_show, _name##_store) |
| 1529 | |
| 1530 | static struct kobject *hugepages_kobj; |
| 1531 | static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; |
| 1532 | |
| 1533 | static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp); |
| 1534 | |
| 1535 | static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp) |
| 1536 | { |
| 1537 | int i; |
| 1538 | |
| 1539 | for (i = 0; i < HUGE_MAX_HSTATE; i++) |
| 1540 | if (hstate_kobjs[i] == kobj) { |
| 1541 | if (nidp) |
| 1542 | *nidp = NUMA_NO_NODE; |
| 1543 | return &hstates[i]; |
| 1544 | } |
| 1545 | |
| 1546 | return kobj_to_node_hstate(kobj, nidp); |
| 1547 | } |
| 1548 | |
| 1549 | static ssize_t nr_hugepages_show_common(struct kobject *kobj, |
| 1550 | struct kobj_attribute *attr, char *buf) |
| 1551 | { |
| 1552 | struct hstate *h; |
| 1553 | unsigned long nr_huge_pages; |
| 1554 | int nid; |
| 1555 | |
| 1556 | h = kobj_to_hstate(kobj, &nid); |
| 1557 | if (nid == NUMA_NO_NODE) |
| 1558 | nr_huge_pages = h->nr_huge_pages; |
| 1559 | else |
| 1560 | nr_huge_pages = h->nr_huge_pages_node[nid]; |
| 1561 | |
| 1562 | return sprintf(buf, "%lu\n", nr_huge_pages); |
| 1563 | } |
| 1564 | |
| 1565 | static ssize_t nr_hugepages_store_common(bool obey_mempolicy, |
| 1566 | struct kobject *kobj, struct kobj_attribute *attr, |
| 1567 | const char *buf, size_t len) |
| 1568 | { |
| 1569 | int err; |
| 1570 | int nid; |
| 1571 | unsigned long count; |
| 1572 | struct hstate *h; |
| 1573 | NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY); |
| 1574 | |
| 1575 | err = kstrtoul(buf, 10, &count); |
| 1576 | if (err) |
| 1577 | goto out; |
| 1578 | |
| 1579 | h = kobj_to_hstate(kobj, &nid); |
| 1580 | if (h->order >= MAX_ORDER) { |
| 1581 | err = -EINVAL; |
| 1582 | goto out; |
| 1583 | } |
| 1584 | |
| 1585 | if (nid == NUMA_NO_NODE) { |
| 1586 | /* |
| 1587 | * global hstate attribute |
| 1588 | */ |
| 1589 | if (!(obey_mempolicy && |
| 1590 | init_nodemask_of_mempolicy(nodes_allowed))) { |
| 1591 | NODEMASK_FREE(nodes_allowed); |
| 1592 | nodes_allowed = &node_states[N_MEMORY]; |
| 1593 | } |
| 1594 | } else if (nodes_allowed) { |
| 1595 | /* |
| 1596 | * per node hstate attribute: adjust count to global, |
| 1597 | * but restrict alloc/free to the specified node. |
| 1598 | */ |
| 1599 | count += h->nr_huge_pages - h->nr_huge_pages_node[nid]; |
| 1600 | init_nodemask_of_node(nodes_allowed, nid); |
| 1601 | } else |
| 1602 | nodes_allowed = &node_states[N_MEMORY]; |
| 1603 | |
| 1604 | h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed); |
| 1605 | |
| 1606 | if (nodes_allowed != &node_states[N_MEMORY]) |
| 1607 | NODEMASK_FREE(nodes_allowed); |
| 1608 | |
| 1609 | return len; |
| 1610 | out: |
| 1611 | NODEMASK_FREE(nodes_allowed); |
| 1612 | return err; |
| 1613 | } |
| 1614 | |
| 1615 | static ssize_t nr_hugepages_show(struct kobject *kobj, |
| 1616 | struct kobj_attribute *attr, char *buf) |
| 1617 | { |
| 1618 | return nr_hugepages_show_common(kobj, attr, buf); |
| 1619 | } |
| 1620 | |
| 1621 | static ssize_t nr_hugepages_store(struct kobject *kobj, |
| 1622 | struct kobj_attribute *attr, const char *buf, size_t len) |
| 1623 | { |
| 1624 | return nr_hugepages_store_common(false, kobj, attr, buf, len); |
| 1625 | } |
| 1626 | HSTATE_ATTR(nr_hugepages); |
| 1627 | |
| 1628 | #ifdef CONFIG_NUMA |
| 1629 | |
| 1630 | /* |
| 1631 | * hstate attribute for optionally mempolicy-based constraint on persistent |
| 1632 | * huge page alloc/free. |
| 1633 | */ |
| 1634 | static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj, |
| 1635 | struct kobj_attribute *attr, char *buf) |
| 1636 | { |
| 1637 | return nr_hugepages_show_common(kobj, attr, buf); |
| 1638 | } |
| 1639 | |
| 1640 | static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj, |
| 1641 | struct kobj_attribute *attr, const char *buf, size_t len) |
| 1642 | { |
| 1643 | return nr_hugepages_store_common(true, kobj, attr, buf, len); |
| 1644 | } |
| 1645 | HSTATE_ATTR(nr_hugepages_mempolicy); |
| 1646 | #endif |
| 1647 | |
| 1648 | |
| 1649 | static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj, |
| 1650 | struct kobj_attribute *attr, char *buf) |
| 1651 | { |
| 1652 | struct hstate *h = kobj_to_hstate(kobj, NULL); |
| 1653 | return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages); |
| 1654 | } |
| 1655 | |
| 1656 | static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj, |
| 1657 | struct kobj_attribute *attr, const char *buf, size_t count) |
| 1658 | { |
| 1659 | int err; |
| 1660 | unsigned long input; |
| 1661 | struct hstate *h = kobj_to_hstate(kobj, NULL); |
| 1662 | |
| 1663 | if (h->order >= MAX_ORDER) |
| 1664 | return -EINVAL; |
| 1665 | |
| 1666 | err = kstrtoul(buf, 10, &input); |
| 1667 | if (err) |
| 1668 | return err; |
| 1669 | |
| 1670 | spin_lock(&hugetlb_lock); |
| 1671 | h->nr_overcommit_huge_pages = input; |
| 1672 | spin_unlock(&hugetlb_lock); |
| 1673 | |
| 1674 | return count; |
| 1675 | } |
| 1676 | HSTATE_ATTR(nr_overcommit_hugepages); |
| 1677 | |
| 1678 | static ssize_t free_hugepages_show(struct kobject *kobj, |
| 1679 | struct kobj_attribute *attr, char *buf) |
| 1680 | { |
| 1681 | struct hstate *h; |
| 1682 | unsigned long free_huge_pages; |
| 1683 | int nid; |
| 1684 | |
| 1685 | h = kobj_to_hstate(kobj, &nid); |
| 1686 | if (nid == NUMA_NO_NODE) |
| 1687 | free_huge_pages = h->free_huge_pages; |
| 1688 | else |
| 1689 | free_huge_pages = h->free_huge_pages_node[nid]; |
| 1690 | |
| 1691 | return sprintf(buf, "%lu\n", free_huge_pages); |
| 1692 | } |
| 1693 | HSTATE_ATTR_RO(free_hugepages); |
| 1694 | |
| 1695 | static ssize_t resv_hugepages_show(struct kobject *kobj, |
| 1696 | struct kobj_attribute *attr, char *buf) |
| 1697 | { |
| 1698 | struct hstate *h = kobj_to_hstate(kobj, NULL); |
| 1699 | return sprintf(buf, "%lu\n", h->resv_huge_pages); |
| 1700 | } |
| 1701 | HSTATE_ATTR_RO(resv_hugepages); |
| 1702 | |
| 1703 | static ssize_t surplus_hugepages_show(struct kobject *kobj, |
| 1704 | struct kobj_attribute *attr, char *buf) |
| 1705 | { |
| 1706 | struct hstate *h; |
| 1707 | unsigned long surplus_huge_pages; |
| 1708 | int nid; |
| 1709 | |
| 1710 | h = kobj_to_hstate(kobj, &nid); |
| 1711 | if (nid == NUMA_NO_NODE) |
| 1712 | surplus_huge_pages = h->surplus_huge_pages; |
| 1713 | else |
| 1714 | surplus_huge_pages = h->surplus_huge_pages_node[nid]; |
| 1715 | |
| 1716 | return sprintf(buf, "%lu\n", surplus_huge_pages); |
| 1717 | } |
| 1718 | HSTATE_ATTR_RO(surplus_hugepages); |
| 1719 | |
| 1720 | static struct attribute *hstate_attrs[] = { |
| 1721 | &nr_hugepages_attr.attr, |
| 1722 | &nr_overcommit_hugepages_attr.attr, |
| 1723 | &free_hugepages_attr.attr, |
| 1724 | &resv_hugepages_attr.attr, |
| 1725 | &surplus_hugepages_attr.attr, |
| 1726 | #ifdef CONFIG_NUMA |
| 1727 | &nr_hugepages_mempolicy_attr.attr, |
| 1728 | #endif |
| 1729 | NULL, |
| 1730 | }; |
| 1731 | |
| 1732 | static struct attribute_group hstate_attr_group = { |
| 1733 | .attrs = hstate_attrs, |
| 1734 | }; |
| 1735 | |
| 1736 | static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent, |
| 1737 | struct kobject **hstate_kobjs, |
| 1738 | struct attribute_group *hstate_attr_group) |
| 1739 | { |
| 1740 | int retval; |
| 1741 | int hi = hstate_index(h); |
| 1742 | |
| 1743 | hstate_kobjs[hi] = kobject_create_and_add(h->name, parent); |
| 1744 | if (!hstate_kobjs[hi]) |
| 1745 | return -ENOMEM; |
| 1746 | |
| 1747 | retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group); |
| 1748 | if (retval) |
| 1749 | kobject_put(hstate_kobjs[hi]); |
| 1750 | |
| 1751 | return retval; |
| 1752 | } |
| 1753 | |
| 1754 | static void __init hugetlb_sysfs_init(void) |
| 1755 | { |
| 1756 | struct hstate *h; |
| 1757 | int err; |
| 1758 | |
| 1759 | hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj); |
| 1760 | if (!hugepages_kobj) |
| 1761 | return; |
| 1762 | |
| 1763 | for_each_hstate(h) { |
| 1764 | err = hugetlb_sysfs_add_hstate(h, hugepages_kobj, |
| 1765 | hstate_kobjs, &hstate_attr_group); |
| 1766 | if (err) |
| 1767 | pr_err("Hugetlb: Unable to add hstate %s", h->name); |
| 1768 | } |
| 1769 | } |
| 1770 | |
| 1771 | #ifdef CONFIG_NUMA |
| 1772 | |
| 1773 | /* |
| 1774 | * node_hstate/s - associate per node hstate attributes, via their kobjects, |
| 1775 | * with node devices in node_devices[] using a parallel array. The array |
| 1776 | * index of a node device or _hstate == node id. |
| 1777 | * This is here to avoid any static dependency of the node device driver, in |
| 1778 | * the base kernel, on the hugetlb module. |
| 1779 | */ |
| 1780 | struct node_hstate { |
| 1781 | struct kobject *hugepages_kobj; |
| 1782 | struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; |
| 1783 | }; |
| 1784 | struct node_hstate node_hstates[MAX_NUMNODES]; |
| 1785 | |
| 1786 | /* |
| 1787 | * A subset of global hstate attributes for node devices |
| 1788 | */ |
| 1789 | static struct attribute *per_node_hstate_attrs[] = { |
| 1790 | &nr_hugepages_attr.attr, |
| 1791 | &free_hugepages_attr.attr, |
| 1792 | &surplus_hugepages_attr.attr, |
| 1793 | NULL, |
| 1794 | }; |
| 1795 | |
| 1796 | static struct attribute_group per_node_hstate_attr_group = { |
| 1797 | .attrs = per_node_hstate_attrs, |
| 1798 | }; |
| 1799 | |
| 1800 | /* |
| 1801 | * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj. |
| 1802 | * Returns node id via non-NULL nidp. |
| 1803 | */ |
| 1804 | static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) |
| 1805 | { |
| 1806 | int nid; |
| 1807 | |
| 1808 | for (nid = 0; nid < nr_node_ids; nid++) { |
| 1809 | struct node_hstate *nhs = &node_hstates[nid]; |
| 1810 | int i; |
| 1811 | for (i = 0; i < HUGE_MAX_HSTATE; i++) |
| 1812 | if (nhs->hstate_kobjs[i] == kobj) { |
| 1813 | if (nidp) |
| 1814 | *nidp = nid; |
| 1815 | return &hstates[i]; |
| 1816 | } |
| 1817 | } |
| 1818 | |
| 1819 | BUG(); |
| 1820 | return NULL; |
| 1821 | } |
| 1822 | |
| 1823 | /* |
| 1824 | * Unregister hstate attributes from a single node device. |
| 1825 | * No-op if no hstate attributes attached. |
| 1826 | */ |
| 1827 | static void hugetlb_unregister_node(struct node *node) |
| 1828 | { |
| 1829 | struct hstate *h; |
| 1830 | struct node_hstate *nhs = &node_hstates[node->dev.id]; |
| 1831 | |
| 1832 | if (!nhs->hugepages_kobj) |
| 1833 | return; /* no hstate attributes */ |
| 1834 | |
| 1835 | for_each_hstate(h) { |
| 1836 | int idx = hstate_index(h); |
| 1837 | if (nhs->hstate_kobjs[idx]) { |
| 1838 | kobject_put(nhs->hstate_kobjs[idx]); |
| 1839 | nhs->hstate_kobjs[idx] = NULL; |
| 1840 | } |
| 1841 | } |
| 1842 | |
| 1843 | kobject_put(nhs->hugepages_kobj); |
| 1844 | nhs->hugepages_kobj = NULL; |
| 1845 | } |
| 1846 | |
| 1847 | /* |
| 1848 | * hugetlb module exit: unregister hstate attributes from node devices |
| 1849 | * that have them. |
| 1850 | */ |
| 1851 | static void hugetlb_unregister_all_nodes(void) |
| 1852 | { |
| 1853 | int nid; |
| 1854 | |
| 1855 | /* |
| 1856 | * disable node device registrations. |
| 1857 | */ |
| 1858 | register_hugetlbfs_with_node(NULL, NULL); |
| 1859 | |
| 1860 | /* |
| 1861 | * remove hstate attributes from any nodes that have them. |
| 1862 | */ |
| 1863 | for (nid = 0; nid < nr_node_ids; nid++) |
| 1864 | hugetlb_unregister_node(node_devices[nid]); |
| 1865 | } |
| 1866 | |
| 1867 | /* |
| 1868 | * Register hstate attributes for a single node device. |
| 1869 | * No-op if attributes already registered. |
| 1870 | */ |
| 1871 | static void hugetlb_register_node(struct node *node) |
| 1872 | { |
| 1873 | struct hstate *h; |
| 1874 | struct node_hstate *nhs = &node_hstates[node->dev.id]; |
| 1875 | int err; |
| 1876 | |
| 1877 | if (nhs->hugepages_kobj) |
| 1878 | return; /* already allocated */ |
| 1879 | |
| 1880 | nhs->hugepages_kobj = kobject_create_and_add("hugepages", |
| 1881 | &node->dev.kobj); |
| 1882 | if (!nhs->hugepages_kobj) |
| 1883 | return; |
| 1884 | |
| 1885 | for_each_hstate(h) { |
| 1886 | err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj, |
| 1887 | nhs->hstate_kobjs, |
| 1888 | &per_node_hstate_attr_group); |
| 1889 | if (err) { |
| 1890 | pr_err("Hugetlb: Unable to add hstate %s for node %d\n", |
| 1891 | h->name, node->dev.id); |
| 1892 | hugetlb_unregister_node(node); |
| 1893 | break; |
| 1894 | } |
| 1895 | } |
| 1896 | } |
| 1897 | |
| 1898 | /* |
| 1899 | * hugetlb init time: register hstate attributes for all registered node |
| 1900 | * devices of nodes that have memory. All on-line nodes should have |
| 1901 | * registered their associated device by this time. |
| 1902 | */ |
| 1903 | static void hugetlb_register_all_nodes(void) |
| 1904 | { |
| 1905 | int nid; |
| 1906 | |
| 1907 | for_each_node_state(nid, N_MEMORY) { |
| 1908 | struct node *node = node_devices[nid]; |
| 1909 | if (node->dev.id == nid) |
| 1910 | hugetlb_register_node(node); |
| 1911 | } |
| 1912 | |
| 1913 | /* |
| 1914 | * Let the node device driver know we're here so it can |
| 1915 | * [un]register hstate attributes on node hotplug. |
| 1916 | */ |
| 1917 | register_hugetlbfs_with_node(hugetlb_register_node, |
| 1918 | hugetlb_unregister_node); |
| 1919 | } |
| 1920 | #else /* !CONFIG_NUMA */ |
| 1921 | |
| 1922 | static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) |
| 1923 | { |
| 1924 | BUG(); |
| 1925 | if (nidp) |
| 1926 | *nidp = -1; |
| 1927 | return NULL; |
| 1928 | } |
| 1929 | |
| 1930 | static void hugetlb_unregister_all_nodes(void) { } |
| 1931 | |
| 1932 | static void hugetlb_register_all_nodes(void) { } |
| 1933 | |
| 1934 | #endif |
| 1935 | |
| 1936 | static void __exit hugetlb_exit(void) |
| 1937 | { |
| 1938 | struct hstate *h; |
| 1939 | |
| 1940 | hugetlb_unregister_all_nodes(); |
| 1941 | |
| 1942 | for_each_hstate(h) { |
| 1943 | kobject_put(hstate_kobjs[hstate_index(h)]); |
| 1944 | } |
| 1945 | |
| 1946 | kobject_put(hugepages_kobj); |
| 1947 | } |
| 1948 | module_exit(hugetlb_exit); |
| 1949 | |
| 1950 | static int __init hugetlb_init(void) |
| 1951 | { |
| 1952 | /* Some platform decide whether they support huge pages at boot |
| 1953 | * time. On these, such as powerpc, HPAGE_SHIFT is set to 0 when |
| 1954 | * there is no such support |
| 1955 | */ |
| 1956 | if (HPAGE_SHIFT == 0) |
| 1957 | return 0; |
| 1958 | |
| 1959 | if (!size_to_hstate(default_hstate_size)) { |
| 1960 | default_hstate_size = HPAGE_SIZE; |
| 1961 | if (!size_to_hstate(default_hstate_size)) |
| 1962 | hugetlb_add_hstate(HUGETLB_PAGE_ORDER); |
| 1963 | } |
| 1964 | default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size)); |
| 1965 | if (default_hstate_max_huge_pages) |
| 1966 | default_hstate.max_huge_pages = default_hstate_max_huge_pages; |
| 1967 | |
| 1968 | hugetlb_init_hstates(); |
| 1969 | gather_bootmem_prealloc(); |
| 1970 | report_hugepages(); |
| 1971 | |
| 1972 | hugetlb_sysfs_init(); |
| 1973 | hugetlb_register_all_nodes(); |
| 1974 | hugetlb_cgroup_file_init(); |
| 1975 | |
| 1976 | return 0; |
| 1977 | } |
| 1978 | module_init(hugetlb_init); |
| 1979 | |
| 1980 | /* Should be called on processing a hugepagesz=... option */ |
| 1981 | void __init hugetlb_add_hstate(unsigned order) |
| 1982 | { |
| 1983 | struct hstate *h; |
| 1984 | unsigned long i; |
| 1985 | |
| 1986 | if (size_to_hstate(PAGE_SIZE << order)) { |
| 1987 | pr_warning("hugepagesz= specified twice, ignoring\n"); |
| 1988 | return; |
| 1989 | } |
| 1990 | BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE); |
| 1991 | BUG_ON(order == 0); |
| 1992 | h = &hstates[hugetlb_max_hstate++]; |
| 1993 | h->order = order; |
| 1994 | h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1); |
| 1995 | h->nr_huge_pages = 0; |
| 1996 | h->free_huge_pages = 0; |
| 1997 | for (i = 0; i < MAX_NUMNODES; ++i) |
| 1998 | INIT_LIST_HEAD(&h->hugepage_freelists[i]); |
| 1999 | INIT_LIST_HEAD(&h->hugepage_activelist); |
| 2000 | h->next_nid_to_alloc = first_node(node_states[N_MEMORY]); |
| 2001 | h->next_nid_to_free = first_node(node_states[N_MEMORY]); |
| 2002 | snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB", |
| 2003 | huge_page_size(h)/1024); |
| 2004 | |
| 2005 | parsed_hstate = h; |
| 2006 | } |
| 2007 | |
| 2008 | static int __init hugetlb_nrpages_setup(char *s) |
| 2009 | { |
| 2010 | unsigned long *mhp; |
| 2011 | static unsigned long *last_mhp; |
| 2012 | |
| 2013 | /* |
| 2014 | * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet, |
| 2015 | * so this hugepages= parameter goes to the "default hstate". |
| 2016 | */ |
| 2017 | if (!hugetlb_max_hstate) |
| 2018 | mhp = &default_hstate_max_huge_pages; |
| 2019 | else |
| 2020 | mhp = &parsed_hstate->max_huge_pages; |
| 2021 | |
| 2022 | if (mhp == last_mhp) { |
| 2023 | pr_warning("hugepages= specified twice without " |
| 2024 | "interleaving hugepagesz=, ignoring\n"); |
| 2025 | return 1; |
| 2026 | } |
| 2027 | |
| 2028 | if (sscanf(s, "%lu", mhp) <= 0) |
| 2029 | *mhp = 0; |
| 2030 | |
| 2031 | /* |
| 2032 | * Global state is always initialized later in hugetlb_init. |
| 2033 | * But we need to allocate >= MAX_ORDER hstates here early to still |
| 2034 | * use the bootmem allocator. |
| 2035 | */ |
| 2036 | if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER) |
| 2037 | hugetlb_hstate_alloc_pages(parsed_hstate); |
| 2038 | |
| 2039 | last_mhp = mhp; |
| 2040 | |
| 2041 | return 1; |
| 2042 | } |
| 2043 | __setup("hugepages=", hugetlb_nrpages_setup); |
| 2044 | |
| 2045 | static int __init hugetlb_default_setup(char *s) |
| 2046 | { |
| 2047 | default_hstate_size = memparse(s, &s); |
| 2048 | return 1; |
| 2049 | } |
| 2050 | __setup("default_hugepagesz=", hugetlb_default_setup); |
| 2051 | |
| 2052 | static unsigned int cpuset_mems_nr(unsigned int *array) |
| 2053 | { |
| 2054 | int node; |
| 2055 | unsigned int nr = 0; |
| 2056 | |
| 2057 | for_each_node_mask(node, cpuset_current_mems_allowed) |
| 2058 | nr += array[node]; |
| 2059 | |
| 2060 | return nr; |
| 2061 | } |
| 2062 | |
| 2063 | #ifdef CONFIG_SYSCTL |
| 2064 | static int hugetlb_sysctl_handler_common(bool obey_mempolicy, |
| 2065 | struct ctl_table *table, int write, |
| 2066 | void __user *buffer, size_t *length, loff_t *ppos) |
| 2067 | { |
| 2068 | struct hstate *h = &default_hstate; |
| 2069 | unsigned long tmp; |
| 2070 | int ret; |
| 2071 | |
| 2072 | tmp = h->max_huge_pages; |
| 2073 | |
| 2074 | if (write && h->order >= MAX_ORDER) |
| 2075 | return -EINVAL; |
| 2076 | |
| 2077 | table->data = &tmp; |
| 2078 | table->maxlen = sizeof(unsigned long); |
| 2079 | ret = proc_doulongvec_minmax(table, write, buffer, length, ppos); |
| 2080 | if (ret) |
| 2081 | goto out; |
| 2082 | |
| 2083 | if (write) { |
| 2084 | NODEMASK_ALLOC(nodemask_t, nodes_allowed, |
| 2085 | GFP_KERNEL | __GFP_NORETRY); |
| 2086 | if (!(obey_mempolicy && |
| 2087 | init_nodemask_of_mempolicy(nodes_allowed))) { |
| 2088 | NODEMASK_FREE(nodes_allowed); |
| 2089 | nodes_allowed = &node_states[N_MEMORY]; |
| 2090 | } |
| 2091 | h->max_huge_pages = set_max_huge_pages(h, tmp, nodes_allowed); |
| 2092 | |
| 2093 | if (nodes_allowed != &node_states[N_MEMORY]) |
| 2094 | NODEMASK_FREE(nodes_allowed); |
| 2095 | } |
| 2096 | out: |
| 2097 | return ret; |
| 2098 | } |
| 2099 | |
| 2100 | int hugetlb_sysctl_handler(struct ctl_table *table, int write, |
| 2101 | void __user *buffer, size_t *length, loff_t *ppos) |
| 2102 | { |
| 2103 | |
| 2104 | return hugetlb_sysctl_handler_common(false, table, write, |
| 2105 | buffer, length, ppos); |
| 2106 | } |
| 2107 | |
| 2108 | #ifdef CONFIG_NUMA |
| 2109 | int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write, |
| 2110 | void __user *buffer, size_t *length, loff_t *ppos) |
| 2111 | { |
| 2112 | return hugetlb_sysctl_handler_common(true, table, write, |
| 2113 | buffer, length, ppos); |
| 2114 | } |
| 2115 | #endif /* CONFIG_NUMA */ |
| 2116 | |
| 2117 | int hugetlb_overcommit_handler(struct ctl_table *table, int write, |
| 2118 | void __user *buffer, |
| 2119 | size_t *length, loff_t *ppos) |
| 2120 | { |
| 2121 | struct hstate *h = &default_hstate; |
| 2122 | unsigned long tmp; |
| 2123 | int ret; |
| 2124 | |
| 2125 | tmp = h->nr_overcommit_huge_pages; |
| 2126 | |
| 2127 | if (write && h->order >= MAX_ORDER) |
| 2128 | return -EINVAL; |
| 2129 | |
| 2130 | table->data = &tmp; |
| 2131 | table->maxlen = sizeof(unsigned long); |
| 2132 | ret = proc_doulongvec_minmax(table, write, buffer, length, ppos); |
| 2133 | if (ret) |
| 2134 | goto out; |
| 2135 | |
| 2136 | if (write) { |
| 2137 | spin_lock(&hugetlb_lock); |
| 2138 | h->nr_overcommit_huge_pages = tmp; |
| 2139 | spin_unlock(&hugetlb_lock); |
| 2140 | } |
| 2141 | out: |
| 2142 | return ret; |
| 2143 | } |
| 2144 | |
| 2145 | #endif /* CONFIG_SYSCTL */ |
| 2146 | |
| 2147 | void hugetlb_report_meminfo(struct seq_file *m) |
| 2148 | { |
| 2149 | struct hstate *h = &default_hstate; |
| 2150 | seq_printf(m, |
| 2151 | "HugePages_Total: %5lu\n" |
| 2152 | "HugePages_Free: %5lu\n" |
| 2153 | "HugePages_Rsvd: %5lu\n" |
| 2154 | "HugePages_Surp: %5lu\n" |
| 2155 | "Hugepagesize: %8lu kB\n", |
| 2156 | h->nr_huge_pages, |
| 2157 | h->free_huge_pages, |
| 2158 | h->resv_huge_pages, |
| 2159 | h->surplus_huge_pages, |
| 2160 | 1UL << (huge_page_order(h) + PAGE_SHIFT - 10)); |
| 2161 | } |
| 2162 | |
| 2163 | int hugetlb_report_node_meminfo(int nid, char *buf) |
| 2164 | { |
| 2165 | struct hstate *h = &default_hstate; |
| 2166 | return sprintf(buf, |
| 2167 | "Node %d HugePages_Total: %5u\n" |
| 2168 | "Node %d HugePages_Free: %5u\n" |
| 2169 | "Node %d HugePages_Surp: %5u\n", |
| 2170 | nid, h->nr_huge_pages_node[nid], |
| 2171 | nid, h->free_huge_pages_node[nid], |
| 2172 | nid, h->surplus_huge_pages_node[nid]); |
| 2173 | } |
| 2174 | |
| 2175 | void hugetlb_show_meminfo(void) |
| 2176 | { |
| 2177 | struct hstate *h; |
| 2178 | int nid; |
| 2179 | |
| 2180 | for_each_node_state(nid, N_MEMORY) |
| 2181 | for_each_hstate(h) |
| 2182 | pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n", |
| 2183 | nid, |
| 2184 | h->nr_huge_pages_node[nid], |
| 2185 | h->free_huge_pages_node[nid], |
| 2186 | h->surplus_huge_pages_node[nid], |
| 2187 | 1UL << (huge_page_order(h) + PAGE_SHIFT - 10)); |
| 2188 | } |
| 2189 | |
| 2190 | /* Return the number pages of memory we physically have, in PAGE_SIZE units. */ |
| 2191 | unsigned long hugetlb_total_pages(void) |
| 2192 | { |
| 2193 | struct hstate *h; |
| 2194 | unsigned long nr_total_pages = 0; |
| 2195 | |
| 2196 | for_each_hstate(h) |
| 2197 | nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h); |
| 2198 | return nr_total_pages; |
| 2199 | } |
| 2200 | |
| 2201 | static int hugetlb_acct_memory(struct hstate *h, long delta) |
| 2202 | { |
| 2203 | int ret = -ENOMEM; |
| 2204 | |
| 2205 | spin_lock(&hugetlb_lock); |
| 2206 | /* |
| 2207 | * When cpuset is configured, it breaks the strict hugetlb page |
| 2208 | * reservation as the accounting is done on a global variable. Such |
| 2209 | * reservation is completely rubbish in the presence of cpuset because |
| 2210 | * the reservation is not checked against page availability for the |
| 2211 | * current cpuset. Application can still potentially OOM'ed by kernel |
| 2212 | * with lack of free htlb page in cpuset that the task is in. |
| 2213 | * Attempt to enforce strict accounting with cpuset is almost |
| 2214 | * impossible (or too ugly) because cpuset is too fluid that |
| 2215 | * task or memory node can be dynamically moved between cpusets. |
| 2216 | * |
| 2217 | * The change of semantics for shared hugetlb mapping with cpuset is |
| 2218 | * undesirable. However, in order to preserve some of the semantics, |
| 2219 | * we fall back to check against current free page availability as |
| 2220 | * a best attempt and hopefully to minimize the impact of changing |
| 2221 | * semantics that cpuset has. |
| 2222 | */ |
| 2223 | if (delta > 0) { |
| 2224 | if (gather_surplus_pages(h, delta) < 0) |
| 2225 | goto out; |
| 2226 | |
| 2227 | if (delta > cpuset_mems_nr(h->free_huge_pages_node)) { |
| 2228 | return_unused_surplus_pages(h, delta); |
| 2229 | goto out; |
| 2230 | } |
| 2231 | } |
| 2232 | |
| 2233 | ret = 0; |
| 2234 | if (delta < 0) |
| 2235 | return_unused_surplus_pages(h, (unsigned long) -delta); |
| 2236 | |
| 2237 | out: |
| 2238 | spin_unlock(&hugetlb_lock); |
| 2239 | return ret; |
| 2240 | } |
| 2241 | |
| 2242 | static void hugetlb_vm_op_open(struct vm_area_struct *vma) |
| 2243 | { |
| 2244 | struct resv_map *resv = vma_resv_map(vma); |
| 2245 | |
| 2246 | /* |
| 2247 | * This new VMA should share its siblings reservation map if present. |
| 2248 | * The VMA will only ever have a valid reservation map pointer where |
| 2249 | * it is being copied for another still existing VMA. As that VMA |
| 2250 | * has a reference to the reservation map it cannot disappear until |
| 2251 | * after this open call completes. It is therefore safe to take a |
| 2252 | * new reference here without additional locking. |
| 2253 | */ |
| 2254 | if (resv) |
| 2255 | kref_get(&resv->refs); |
| 2256 | } |
| 2257 | |
| 2258 | static void resv_map_put(struct vm_area_struct *vma) |
| 2259 | { |
| 2260 | struct resv_map *resv = vma_resv_map(vma); |
| 2261 | |
| 2262 | if (!resv) |
| 2263 | return; |
| 2264 | kref_put(&resv->refs, resv_map_release); |
| 2265 | } |
| 2266 | |
| 2267 | static void hugetlb_vm_op_close(struct vm_area_struct *vma) |
| 2268 | { |
| 2269 | struct hstate *h = hstate_vma(vma); |
| 2270 | struct resv_map *resv = vma_resv_map(vma); |
| 2271 | struct hugepage_subpool *spool = subpool_vma(vma); |
| 2272 | unsigned long reserve; |
| 2273 | unsigned long start; |
| 2274 | unsigned long end; |
| 2275 | |
| 2276 | if (resv) { |
| 2277 | start = vma_hugecache_offset(h, vma, vma->vm_start); |
| 2278 | end = vma_hugecache_offset(h, vma, vma->vm_end); |
| 2279 | |
| 2280 | reserve = (end - start) - |
| 2281 | region_count(&resv->regions, start, end); |
| 2282 | |
| 2283 | resv_map_put(vma); |
| 2284 | |
| 2285 | if (reserve) { |
| 2286 | hugetlb_acct_memory(h, -reserve); |
| 2287 | hugepage_subpool_put_pages(spool, reserve); |
| 2288 | } |
| 2289 | } |
| 2290 | } |
| 2291 | |
| 2292 | /* |
| 2293 | * We cannot handle pagefaults against hugetlb pages at all. They cause |
| 2294 | * handle_mm_fault() to try to instantiate regular-sized pages in the |
| 2295 | * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get |
| 2296 | * this far. |
| 2297 | */ |
| 2298 | static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf) |
| 2299 | { |
| 2300 | BUG(); |
| 2301 | return 0; |
| 2302 | } |
| 2303 | |
| 2304 | const struct vm_operations_struct hugetlb_vm_ops = { |
| 2305 | .fault = hugetlb_vm_op_fault, |
| 2306 | .open = hugetlb_vm_op_open, |
| 2307 | .close = hugetlb_vm_op_close, |
| 2308 | }; |
| 2309 | |
| 2310 | static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page, |
| 2311 | int writable) |
| 2312 | { |
| 2313 | pte_t entry; |
| 2314 | |
| 2315 | if (writable) { |
| 2316 | entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page, |
| 2317 | vma->vm_page_prot))); |
| 2318 | } else { |
| 2319 | entry = huge_pte_wrprotect(mk_huge_pte(page, |
| 2320 | vma->vm_page_prot)); |
| 2321 | } |
| 2322 | entry = pte_mkyoung(entry); |
| 2323 | entry = pte_mkhuge(entry); |
| 2324 | entry = arch_make_huge_pte(entry, vma, page, writable); |
| 2325 | |
| 2326 | return entry; |
| 2327 | } |
| 2328 | |
| 2329 | static void set_huge_ptep_writable(struct vm_area_struct *vma, |
| 2330 | unsigned long address, pte_t *ptep) |
| 2331 | { |
| 2332 | pte_t entry; |
| 2333 | |
| 2334 | entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep))); |
| 2335 | if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) |
| 2336 | update_mmu_cache(vma, address, ptep); |
| 2337 | } |
| 2338 | |
| 2339 | |
| 2340 | int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src, |
| 2341 | struct vm_area_struct *vma) |
| 2342 | { |
| 2343 | pte_t *src_pte, *dst_pte, entry; |
| 2344 | struct page *ptepage; |
| 2345 | unsigned long addr; |
| 2346 | int cow; |
| 2347 | struct hstate *h = hstate_vma(vma); |
| 2348 | unsigned long sz = huge_page_size(h); |
| 2349 | |
| 2350 | cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE; |
| 2351 | |
| 2352 | for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) { |
| 2353 | spinlock_t *src_ptl, *dst_ptl; |
| 2354 | src_pte = huge_pte_offset(src, addr); |
| 2355 | if (!src_pte) |
| 2356 | continue; |
| 2357 | dst_pte = huge_pte_alloc(dst, addr, sz); |
| 2358 | if (!dst_pte) |
| 2359 | goto nomem; |
| 2360 | |
| 2361 | /* If the pagetables are shared don't copy or take references */ |
| 2362 | if (dst_pte == src_pte) |
| 2363 | continue; |
| 2364 | |
| 2365 | dst_ptl = huge_pte_lock(h, dst, dst_pte); |
| 2366 | src_ptl = huge_pte_lockptr(h, src, src_pte); |
| 2367 | spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); |
| 2368 | if (!huge_pte_none(huge_ptep_get(src_pte))) { |
| 2369 | if (cow) |
| 2370 | huge_ptep_set_wrprotect(src, addr, src_pte); |
| 2371 | entry = huge_ptep_get(src_pte); |
| 2372 | ptepage = pte_page(entry); |
| 2373 | get_page(ptepage); |
| 2374 | page_dup_rmap(ptepage); |
| 2375 | set_huge_pte_at(dst, addr, dst_pte, entry); |
| 2376 | } |
| 2377 | spin_unlock(src_ptl); |
| 2378 | spin_unlock(dst_ptl); |
| 2379 | } |
| 2380 | return 0; |
| 2381 | |
| 2382 | nomem: |
| 2383 | return -ENOMEM; |
| 2384 | } |
| 2385 | |
| 2386 | static int is_hugetlb_entry_migration(pte_t pte) |
| 2387 | { |
| 2388 | swp_entry_t swp; |
| 2389 | |
| 2390 | if (huge_pte_none(pte) || pte_present(pte)) |
| 2391 | return 0; |
| 2392 | swp = pte_to_swp_entry(pte); |
| 2393 | if (non_swap_entry(swp) && is_migration_entry(swp)) |
| 2394 | return 1; |
| 2395 | else |
| 2396 | return 0; |
| 2397 | } |
| 2398 | |
| 2399 | static int is_hugetlb_entry_hwpoisoned(pte_t pte) |
| 2400 | { |
| 2401 | swp_entry_t swp; |
| 2402 | |
| 2403 | if (huge_pte_none(pte) || pte_present(pte)) |
| 2404 | return 0; |
| 2405 | swp = pte_to_swp_entry(pte); |
| 2406 | if (non_swap_entry(swp) && is_hwpoison_entry(swp)) |
| 2407 | return 1; |
| 2408 | else |
| 2409 | return 0; |
| 2410 | } |
| 2411 | |
| 2412 | void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma, |
| 2413 | unsigned long start, unsigned long end, |
| 2414 | struct page *ref_page) |
| 2415 | { |
| 2416 | int force_flush = 0; |
| 2417 | struct mm_struct *mm = vma->vm_mm; |
| 2418 | unsigned long address; |
| 2419 | pte_t *ptep; |
| 2420 | pte_t pte; |
| 2421 | spinlock_t *ptl; |
| 2422 | struct page *page; |
| 2423 | struct hstate *h = hstate_vma(vma); |
| 2424 | unsigned long sz = huge_page_size(h); |
| 2425 | const unsigned long mmun_start = start; /* For mmu_notifiers */ |
| 2426 | const unsigned long mmun_end = end; /* For mmu_notifiers */ |
| 2427 | |
| 2428 | WARN_ON(!is_vm_hugetlb_page(vma)); |
| 2429 | BUG_ON(start & ~huge_page_mask(h)); |
| 2430 | BUG_ON(end & ~huge_page_mask(h)); |
| 2431 | |
| 2432 | tlb_start_vma(tlb, vma); |
| 2433 | mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); |
| 2434 | again: |
| 2435 | for (address = start; address < end; address += sz) { |
| 2436 | ptep = huge_pte_offset(mm, address); |
| 2437 | if (!ptep) |
| 2438 | continue; |
| 2439 | |
| 2440 | ptl = huge_pte_lock(h, mm, ptep); |
| 2441 | if (huge_pmd_unshare(mm, &address, ptep)) |
| 2442 | goto unlock; |
| 2443 | |
| 2444 | pte = huge_ptep_get(ptep); |
| 2445 | if (huge_pte_none(pte)) |
| 2446 | goto unlock; |
| 2447 | |
| 2448 | /* |
| 2449 | * HWPoisoned hugepage is already unmapped and dropped reference |
| 2450 | */ |
| 2451 | if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) { |
| 2452 | huge_pte_clear(mm, address, ptep); |
| 2453 | goto unlock; |
| 2454 | } |
| 2455 | |
| 2456 | page = pte_page(pte); |
| 2457 | /* |
| 2458 | * If a reference page is supplied, it is because a specific |
| 2459 | * page is being unmapped, not a range. Ensure the page we |
| 2460 | * are about to unmap is the actual page of interest. |
| 2461 | */ |
| 2462 | if (ref_page) { |
| 2463 | if (page != ref_page) |
| 2464 | goto unlock; |
| 2465 | |
| 2466 | /* |
| 2467 | * Mark the VMA as having unmapped its page so that |
| 2468 | * future faults in this VMA will fail rather than |
| 2469 | * looking like data was lost |
| 2470 | */ |
| 2471 | set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED); |
| 2472 | } |
| 2473 | |
| 2474 | pte = huge_ptep_get_and_clear(mm, address, ptep); |
| 2475 | tlb_remove_tlb_entry(tlb, ptep, address); |
| 2476 | if (huge_pte_dirty(pte)) |
| 2477 | set_page_dirty(page); |
| 2478 | |
| 2479 | page_remove_rmap(page); |
| 2480 | force_flush = !__tlb_remove_page(tlb, page); |
| 2481 | if (force_flush) { |
| 2482 | spin_unlock(ptl); |
| 2483 | break; |
| 2484 | } |
| 2485 | /* Bail out after unmapping reference page if supplied */ |
| 2486 | if (ref_page) { |
| 2487 | spin_unlock(ptl); |
| 2488 | break; |
| 2489 | } |
| 2490 | unlock: |
| 2491 | spin_unlock(ptl); |
| 2492 | } |
| 2493 | /* |
| 2494 | * mmu_gather ran out of room to batch pages, we break out of |
| 2495 | * the PTE lock to avoid doing the potential expensive TLB invalidate |
| 2496 | * and page-free while holding it. |
| 2497 | */ |
| 2498 | if (force_flush) { |
| 2499 | force_flush = 0; |
| 2500 | tlb_flush_mmu(tlb); |
| 2501 | if (address < end && !ref_page) |
| 2502 | goto again; |
| 2503 | } |
| 2504 | mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); |
| 2505 | tlb_end_vma(tlb, vma); |
| 2506 | } |
| 2507 | |
| 2508 | void __unmap_hugepage_range_final(struct mmu_gather *tlb, |
| 2509 | struct vm_area_struct *vma, unsigned long start, |
| 2510 | unsigned long end, struct page *ref_page) |
| 2511 | { |
| 2512 | __unmap_hugepage_range(tlb, vma, start, end, ref_page); |
| 2513 | |
| 2514 | /* |
| 2515 | * Clear this flag so that x86's huge_pmd_share page_table_shareable |
| 2516 | * test will fail on a vma being torn down, and not grab a page table |
| 2517 | * on its way out. We're lucky that the flag has such an appropriate |
| 2518 | * name, and can in fact be safely cleared here. We could clear it |
| 2519 | * before the __unmap_hugepage_range above, but all that's necessary |
| 2520 | * is to clear it before releasing the i_mmap_mutex. This works |
| 2521 | * because in the context this is called, the VMA is about to be |
| 2522 | * destroyed and the i_mmap_mutex is held. |
| 2523 | */ |
| 2524 | vma->vm_flags &= ~VM_MAYSHARE; |
| 2525 | } |
| 2526 | |
| 2527 | void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start, |
| 2528 | unsigned long end, struct page *ref_page) |
| 2529 | { |
| 2530 | struct mm_struct *mm; |
| 2531 | struct mmu_gather tlb; |
| 2532 | |
| 2533 | mm = vma->vm_mm; |
| 2534 | |
| 2535 | tlb_gather_mmu(&tlb, mm, start, end); |
| 2536 | __unmap_hugepage_range(&tlb, vma, start, end, ref_page); |
| 2537 | tlb_finish_mmu(&tlb, start, end); |
| 2538 | } |
| 2539 | |
| 2540 | /* |
| 2541 | * This is called when the original mapper is failing to COW a MAP_PRIVATE |
| 2542 | * mappping it owns the reserve page for. The intention is to unmap the page |
| 2543 | * from other VMAs and let the children be SIGKILLed if they are faulting the |
| 2544 | * same region. |
| 2545 | */ |
| 2546 | static int unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma, |
| 2547 | struct page *page, unsigned long address) |
| 2548 | { |
| 2549 | struct hstate *h = hstate_vma(vma); |
| 2550 | struct vm_area_struct *iter_vma; |
| 2551 | struct address_space *mapping; |
| 2552 | pgoff_t pgoff; |
| 2553 | |
| 2554 | /* |
| 2555 | * vm_pgoff is in PAGE_SIZE units, hence the different calculation |
| 2556 | * from page cache lookup which is in HPAGE_SIZE units. |
| 2557 | */ |
| 2558 | address = address & huge_page_mask(h); |
| 2559 | pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) + |
| 2560 | vma->vm_pgoff; |
| 2561 | mapping = file_inode(vma->vm_file)->i_mapping; |
| 2562 | |
| 2563 | /* |
| 2564 | * Take the mapping lock for the duration of the table walk. As |
| 2565 | * this mapping should be shared between all the VMAs, |
| 2566 | * __unmap_hugepage_range() is called as the lock is already held |
| 2567 | */ |
| 2568 | mutex_lock(&mapping->i_mmap_mutex); |
| 2569 | vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) { |
| 2570 | /* Do not unmap the current VMA */ |
| 2571 | if (iter_vma == vma) |
| 2572 | continue; |
| 2573 | |
| 2574 | /* |
| 2575 | * Unmap the page from other VMAs without their own reserves. |
| 2576 | * They get marked to be SIGKILLed if they fault in these |
| 2577 | * areas. This is because a future no-page fault on this VMA |
| 2578 | * could insert a zeroed page instead of the data existing |
| 2579 | * from the time of fork. This would look like data corruption |
| 2580 | */ |
| 2581 | if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER)) |
| 2582 | unmap_hugepage_range(iter_vma, address, |
| 2583 | address + huge_page_size(h), page); |
| 2584 | } |
| 2585 | mutex_unlock(&mapping->i_mmap_mutex); |
| 2586 | |
| 2587 | return 1; |
| 2588 | } |
| 2589 | |
| 2590 | /* |
| 2591 | * Hugetlb_cow() should be called with page lock of the original hugepage held. |
| 2592 | * Called with hugetlb_instantiation_mutex held and pte_page locked so we |
| 2593 | * cannot race with other handlers or page migration. |
| 2594 | * Keep the pte_same checks anyway to make transition from the mutex easier. |
| 2595 | */ |
| 2596 | static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma, |
| 2597 | unsigned long address, pte_t *ptep, pte_t pte, |
| 2598 | struct page *pagecache_page, spinlock_t *ptl) |
| 2599 | { |
| 2600 | struct hstate *h = hstate_vma(vma); |
| 2601 | struct page *old_page, *new_page; |
| 2602 | int outside_reserve = 0; |
| 2603 | unsigned long mmun_start; /* For mmu_notifiers */ |
| 2604 | unsigned long mmun_end; /* For mmu_notifiers */ |
| 2605 | |
| 2606 | old_page = pte_page(pte); |
| 2607 | |
| 2608 | retry_avoidcopy: |
| 2609 | /* If no-one else is actually using this page, avoid the copy |
| 2610 | * and just make the page writable */ |
| 2611 | if (page_mapcount(old_page) == 1 && PageAnon(old_page)) { |
| 2612 | page_move_anon_rmap(old_page, vma, address); |
| 2613 | set_huge_ptep_writable(vma, address, ptep); |
| 2614 | return 0; |
| 2615 | } |
| 2616 | |
| 2617 | /* |
| 2618 | * If the process that created a MAP_PRIVATE mapping is about to |
| 2619 | * perform a COW due to a shared page count, attempt to satisfy |
| 2620 | * the allocation without using the existing reserves. The pagecache |
| 2621 | * page is used to determine if the reserve at this address was |
| 2622 | * consumed or not. If reserves were used, a partial faulted mapping |
| 2623 | * at the time of fork() could consume its reserves on COW instead |
| 2624 | * of the full address range. |
| 2625 | */ |
| 2626 | if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && |
| 2627 | old_page != pagecache_page) |
| 2628 | outside_reserve = 1; |
| 2629 | |
| 2630 | page_cache_get(old_page); |
| 2631 | |
| 2632 | /* Drop page table lock as buddy allocator may be called */ |
| 2633 | spin_unlock(ptl); |
| 2634 | new_page = alloc_huge_page(vma, address, outside_reserve); |
| 2635 | |
| 2636 | if (IS_ERR(new_page)) { |
| 2637 | long err = PTR_ERR(new_page); |
| 2638 | page_cache_release(old_page); |
| 2639 | |
| 2640 | /* |
| 2641 | * If a process owning a MAP_PRIVATE mapping fails to COW, |
| 2642 | * it is due to references held by a child and an insufficient |
| 2643 | * huge page pool. To guarantee the original mappers |
| 2644 | * reliability, unmap the page from child processes. The child |
| 2645 | * may get SIGKILLed if it later faults. |
| 2646 | */ |
| 2647 | if (outside_reserve) { |
| 2648 | BUG_ON(huge_pte_none(pte)); |
| 2649 | if (unmap_ref_private(mm, vma, old_page, address)) { |
| 2650 | BUG_ON(huge_pte_none(pte)); |
| 2651 | spin_lock(ptl); |
| 2652 | ptep = huge_pte_offset(mm, address & huge_page_mask(h)); |
| 2653 | if (likely(pte_same(huge_ptep_get(ptep), pte))) |
| 2654 | goto retry_avoidcopy; |
| 2655 | /* |
| 2656 | * race occurs while re-acquiring page table |
| 2657 | * lock, and our job is done. |
| 2658 | */ |
| 2659 | return 0; |
| 2660 | } |
| 2661 | WARN_ON_ONCE(1); |
| 2662 | } |
| 2663 | |
| 2664 | /* Caller expects lock to be held */ |
| 2665 | spin_lock(ptl); |
| 2666 | if (err == -ENOMEM) |
| 2667 | return VM_FAULT_OOM; |
| 2668 | else |
| 2669 | return VM_FAULT_SIGBUS; |
| 2670 | } |
| 2671 | |
| 2672 | /* |
| 2673 | * When the original hugepage is shared one, it does not have |
| 2674 | * anon_vma prepared. |
| 2675 | */ |
| 2676 | if (unlikely(anon_vma_prepare(vma))) { |
| 2677 | page_cache_release(new_page); |
| 2678 | page_cache_release(old_page); |
| 2679 | /* Caller expects lock to be held */ |
| 2680 | spin_lock(ptl); |
| 2681 | return VM_FAULT_OOM; |
| 2682 | } |
| 2683 | |
| 2684 | copy_user_huge_page(new_page, old_page, address, vma, |
| 2685 | pages_per_huge_page(h)); |
| 2686 | __SetPageUptodate(new_page); |
| 2687 | |
| 2688 | mmun_start = address & huge_page_mask(h); |
| 2689 | mmun_end = mmun_start + huge_page_size(h); |
| 2690 | mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); |
| 2691 | /* |
| 2692 | * Retake the page table lock to check for racing updates |
| 2693 | * before the page tables are altered |
| 2694 | */ |
| 2695 | spin_lock(ptl); |
| 2696 | ptep = huge_pte_offset(mm, address & huge_page_mask(h)); |
| 2697 | if (likely(pte_same(huge_ptep_get(ptep), pte))) { |
| 2698 | ClearPagePrivate(new_page); |
| 2699 | |
| 2700 | /* Break COW */ |
| 2701 | huge_ptep_clear_flush(vma, address, ptep); |
| 2702 | set_huge_pte_at(mm, address, ptep, |
| 2703 | make_huge_pte(vma, new_page, 1)); |
| 2704 | page_remove_rmap(old_page); |
| 2705 | hugepage_add_new_anon_rmap(new_page, vma, address); |
| 2706 | /* Make the old page be freed below */ |
| 2707 | new_page = old_page; |
| 2708 | } |
| 2709 | spin_unlock(ptl); |
| 2710 | mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); |
| 2711 | page_cache_release(new_page); |
| 2712 | page_cache_release(old_page); |
| 2713 | |
| 2714 | /* Caller expects lock to be held */ |
| 2715 | spin_lock(ptl); |
| 2716 | return 0; |
| 2717 | } |
| 2718 | |
| 2719 | /* Return the pagecache page at a given address within a VMA */ |
| 2720 | static struct page *hugetlbfs_pagecache_page(struct hstate *h, |
| 2721 | struct vm_area_struct *vma, unsigned long address) |
| 2722 | { |
| 2723 | struct address_space *mapping; |
| 2724 | pgoff_t idx; |
| 2725 | |
| 2726 | mapping = vma->vm_file->f_mapping; |
| 2727 | idx = vma_hugecache_offset(h, vma, address); |
| 2728 | |
| 2729 | return find_lock_page(mapping, idx); |
| 2730 | } |
| 2731 | |
| 2732 | /* |
| 2733 | * Return whether there is a pagecache page to back given address within VMA. |
| 2734 | * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page. |
| 2735 | */ |
| 2736 | static bool hugetlbfs_pagecache_present(struct hstate *h, |
| 2737 | struct vm_area_struct *vma, unsigned long address) |
| 2738 | { |
| 2739 | struct address_space *mapping; |
| 2740 | pgoff_t idx; |
| 2741 | struct page *page; |
| 2742 | |
| 2743 | mapping = vma->vm_file->f_mapping; |
| 2744 | idx = vma_hugecache_offset(h, vma, address); |
| 2745 | |
| 2746 | page = find_get_page(mapping, idx); |
| 2747 | if (page) |
| 2748 | put_page(page); |
| 2749 | return page != NULL; |
| 2750 | } |
| 2751 | |
| 2752 | static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma, |
| 2753 | unsigned long address, pte_t *ptep, unsigned int flags) |
| 2754 | { |
| 2755 | struct hstate *h = hstate_vma(vma); |
| 2756 | int ret = VM_FAULT_SIGBUS; |
| 2757 | int anon_rmap = 0; |
| 2758 | pgoff_t idx; |
| 2759 | unsigned long size; |
| 2760 | struct page *page; |
| 2761 | struct address_space *mapping; |
| 2762 | pte_t new_pte; |
| 2763 | spinlock_t *ptl; |
| 2764 | |
| 2765 | /* |
| 2766 | * Currently, we are forced to kill the process in the event the |
| 2767 | * original mapper has unmapped pages from the child due to a failed |
| 2768 | * COW. Warn that such a situation has occurred as it may not be obvious |
| 2769 | */ |
| 2770 | if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) { |
| 2771 | pr_warning("PID %d killed due to inadequate hugepage pool\n", |
| 2772 | current->pid); |
| 2773 | return ret; |
| 2774 | } |
| 2775 | |
| 2776 | mapping = vma->vm_file->f_mapping; |
| 2777 | idx = vma_hugecache_offset(h, vma, address); |
| 2778 | |
| 2779 | /* |
| 2780 | * Use page lock to guard against racing truncation |
| 2781 | * before we get page_table_lock. |
| 2782 | */ |
| 2783 | retry: |
| 2784 | page = find_lock_page(mapping, idx); |
| 2785 | if (!page) { |
| 2786 | size = i_size_read(mapping->host) >> huge_page_shift(h); |
| 2787 | if (idx >= size) |
| 2788 | goto out; |
| 2789 | page = alloc_huge_page(vma, address, 0); |
| 2790 | if (IS_ERR(page)) { |
| 2791 | ret = PTR_ERR(page); |
| 2792 | if (ret == -ENOMEM) |
| 2793 | ret = VM_FAULT_OOM; |
| 2794 | else |
| 2795 | ret = VM_FAULT_SIGBUS; |
| 2796 | goto out; |
| 2797 | } |
| 2798 | clear_huge_page(page, address, pages_per_huge_page(h)); |
| 2799 | __SetPageUptodate(page); |
| 2800 | |
| 2801 | if (vma->vm_flags & VM_MAYSHARE) { |
| 2802 | int err; |
| 2803 | struct inode *inode = mapping->host; |
| 2804 | |
| 2805 | err = add_to_page_cache(page, mapping, idx, GFP_KERNEL); |
| 2806 | if (err) { |
| 2807 | put_page(page); |
| 2808 | if (err == -EEXIST) |
| 2809 | goto retry; |
| 2810 | goto out; |
| 2811 | } |
| 2812 | ClearPagePrivate(page); |
| 2813 | |
| 2814 | spin_lock(&inode->i_lock); |
| 2815 | inode->i_blocks += blocks_per_huge_page(h); |
| 2816 | spin_unlock(&inode->i_lock); |
| 2817 | } else { |
| 2818 | lock_page(page); |
| 2819 | if (unlikely(anon_vma_prepare(vma))) { |
| 2820 | ret = VM_FAULT_OOM; |
| 2821 | goto backout_unlocked; |
| 2822 | } |
| 2823 | anon_rmap = 1; |
| 2824 | } |
| 2825 | } else { |
| 2826 | /* |
| 2827 | * If memory error occurs between mmap() and fault, some process |
| 2828 | * don't have hwpoisoned swap entry for errored virtual address. |
| 2829 | * So we need to block hugepage fault by PG_hwpoison bit check. |
| 2830 | */ |
| 2831 | if (unlikely(PageHWPoison(page))) { |
| 2832 | ret = VM_FAULT_HWPOISON | |
| 2833 | VM_FAULT_SET_HINDEX(hstate_index(h)); |
| 2834 | goto backout_unlocked; |
| 2835 | } |
| 2836 | } |
| 2837 | |
| 2838 | /* |
| 2839 | * If we are going to COW a private mapping later, we examine the |
| 2840 | * pending reservations for this page now. This will ensure that |
| 2841 | * any allocations necessary to record that reservation occur outside |
| 2842 | * the spinlock. |
| 2843 | */ |
| 2844 | if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) |
| 2845 | if (vma_needs_reservation(h, vma, address) < 0) { |
| 2846 | ret = VM_FAULT_OOM; |
| 2847 | goto backout_unlocked; |
| 2848 | } |
| 2849 | |
| 2850 | ptl = huge_pte_lockptr(h, mm, ptep); |
| 2851 | spin_lock(ptl); |
| 2852 | size = i_size_read(mapping->host) >> huge_page_shift(h); |
| 2853 | if (idx >= size) |
| 2854 | goto backout; |
| 2855 | |
| 2856 | ret = 0; |
| 2857 | if (!huge_pte_none(huge_ptep_get(ptep))) |
| 2858 | goto backout; |
| 2859 | |
| 2860 | if (anon_rmap) { |
| 2861 | ClearPagePrivate(page); |
| 2862 | hugepage_add_new_anon_rmap(page, vma, address); |
| 2863 | } |
| 2864 | else |
| 2865 | page_dup_rmap(page); |
| 2866 | new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE) |
| 2867 | && (vma->vm_flags & VM_SHARED))); |
| 2868 | set_huge_pte_at(mm, address, ptep, new_pte); |
| 2869 | |
| 2870 | if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { |
| 2871 | /* Optimization, do the COW without a second fault */ |
| 2872 | ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page, ptl); |
| 2873 | } |
| 2874 | |
| 2875 | spin_unlock(ptl); |
| 2876 | unlock_page(page); |
| 2877 | out: |
| 2878 | return ret; |
| 2879 | |
| 2880 | backout: |
| 2881 | spin_unlock(ptl); |
| 2882 | backout_unlocked: |
| 2883 | unlock_page(page); |
| 2884 | put_page(page); |
| 2885 | goto out; |
| 2886 | } |
| 2887 | |
| 2888 | int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma, |
| 2889 | unsigned long address, unsigned int flags) |
| 2890 | { |
| 2891 | pte_t *ptep; |
| 2892 | pte_t entry; |
| 2893 | spinlock_t *ptl; |
| 2894 | int ret; |
| 2895 | struct page *page = NULL; |
| 2896 | struct page *pagecache_page = NULL; |
| 2897 | static DEFINE_MUTEX(hugetlb_instantiation_mutex); |
| 2898 | struct hstate *h = hstate_vma(vma); |
| 2899 | |
| 2900 | address &= huge_page_mask(h); |
| 2901 | |
| 2902 | ptep = huge_pte_offset(mm, address); |
| 2903 | if (ptep) { |
| 2904 | entry = huge_ptep_get(ptep); |
| 2905 | if (unlikely(is_hugetlb_entry_migration(entry))) { |
| 2906 | migration_entry_wait_huge(vma, mm, ptep); |
| 2907 | return 0; |
| 2908 | } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry))) |
| 2909 | return VM_FAULT_HWPOISON_LARGE | |
| 2910 | VM_FAULT_SET_HINDEX(hstate_index(h)); |
| 2911 | } |
| 2912 | |
| 2913 | ptep = huge_pte_alloc(mm, address, huge_page_size(h)); |
| 2914 | if (!ptep) |
| 2915 | return VM_FAULT_OOM; |
| 2916 | |
| 2917 | /* |
| 2918 | * Serialize hugepage allocation and instantiation, so that we don't |
| 2919 | * get spurious allocation failures if two CPUs race to instantiate |
| 2920 | * the same page in the page cache. |
| 2921 | */ |
| 2922 | mutex_lock(&hugetlb_instantiation_mutex); |
| 2923 | entry = huge_ptep_get(ptep); |
| 2924 | if (huge_pte_none(entry)) { |
| 2925 | ret = hugetlb_no_page(mm, vma, address, ptep, flags); |
| 2926 | goto out_mutex; |
| 2927 | } |
| 2928 | |
| 2929 | ret = 0; |
| 2930 | |
| 2931 | /* |
| 2932 | * If we are going to COW the mapping later, we examine the pending |
| 2933 | * reservations for this page now. This will ensure that any |
| 2934 | * allocations necessary to record that reservation occur outside the |
| 2935 | * spinlock. For private mappings, we also lookup the pagecache |
| 2936 | * page now as it is used to determine if a reservation has been |
| 2937 | * consumed. |
| 2938 | */ |
| 2939 | if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) { |
| 2940 | if (vma_needs_reservation(h, vma, address) < 0) { |
| 2941 | ret = VM_FAULT_OOM; |
| 2942 | goto out_mutex; |
| 2943 | } |
| 2944 | |
| 2945 | if (!(vma->vm_flags & VM_MAYSHARE)) |
| 2946 | pagecache_page = hugetlbfs_pagecache_page(h, |
| 2947 | vma, address); |
| 2948 | } |
| 2949 | |
| 2950 | /* |
| 2951 | * hugetlb_cow() requires page locks of pte_page(entry) and |
| 2952 | * pagecache_page, so here we need take the former one |
| 2953 | * when page != pagecache_page or !pagecache_page. |
| 2954 | * Note that locking order is always pagecache_page -> page, |
| 2955 | * so no worry about deadlock. |
| 2956 | */ |
| 2957 | page = pte_page(entry); |
| 2958 | get_page(page); |
| 2959 | if (page != pagecache_page) |
| 2960 | lock_page(page); |
| 2961 | |
| 2962 | ptl = huge_pte_lockptr(h, mm, ptep); |
| 2963 | spin_lock(ptl); |
| 2964 | /* Check for a racing update before calling hugetlb_cow */ |
| 2965 | if (unlikely(!pte_same(entry, huge_ptep_get(ptep)))) |
| 2966 | goto out_ptl; |
| 2967 | |
| 2968 | |
| 2969 | if (flags & FAULT_FLAG_WRITE) { |
| 2970 | if (!huge_pte_write(entry)) { |
| 2971 | ret = hugetlb_cow(mm, vma, address, ptep, entry, |
| 2972 | pagecache_page, ptl); |
| 2973 | goto out_ptl; |
| 2974 | } |
| 2975 | entry = huge_pte_mkdirty(entry); |
| 2976 | } |
| 2977 | entry = pte_mkyoung(entry); |
| 2978 | if (huge_ptep_set_access_flags(vma, address, ptep, entry, |
| 2979 | flags & FAULT_FLAG_WRITE)) |
| 2980 | update_mmu_cache(vma, address, ptep); |
| 2981 | |
| 2982 | out_ptl: |
| 2983 | spin_unlock(ptl); |
| 2984 | |
| 2985 | if (pagecache_page) { |
| 2986 | unlock_page(pagecache_page); |
| 2987 | put_page(pagecache_page); |
| 2988 | } |
| 2989 | if (page != pagecache_page) |
| 2990 | unlock_page(page); |
| 2991 | put_page(page); |
| 2992 | |
| 2993 | out_mutex: |
| 2994 | mutex_unlock(&hugetlb_instantiation_mutex); |
| 2995 | |
| 2996 | return ret; |
| 2997 | } |
| 2998 | |
| 2999 | long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma, |
| 3000 | struct page **pages, struct vm_area_struct **vmas, |
| 3001 | unsigned long *position, unsigned long *nr_pages, |
| 3002 | long i, unsigned int flags) |
| 3003 | { |
| 3004 | unsigned long pfn_offset; |
| 3005 | unsigned long vaddr = *position; |
| 3006 | unsigned long remainder = *nr_pages; |
| 3007 | struct hstate *h = hstate_vma(vma); |
| 3008 | |
| 3009 | while (vaddr < vma->vm_end && remainder) { |
| 3010 | pte_t *pte; |
| 3011 | spinlock_t *ptl = NULL; |
| 3012 | int absent; |
| 3013 | struct page *page; |
| 3014 | |
| 3015 | /* |
| 3016 | * Some archs (sparc64, sh*) have multiple pte_ts to |
| 3017 | * each hugepage. We have to make sure we get the |
| 3018 | * first, for the page indexing below to work. |
| 3019 | * |
| 3020 | * Note that page table lock is not held when pte is null. |
| 3021 | */ |
| 3022 | pte = huge_pte_offset(mm, vaddr & huge_page_mask(h)); |
| 3023 | if (pte) |
| 3024 | ptl = huge_pte_lock(h, mm, pte); |
| 3025 | absent = !pte || huge_pte_none(huge_ptep_get(pte)); |
| 3026 | |
| 3027 | /* |
| 3028 | * When coredumping, it suits get_dump_page if we just return |
| 3029 | * an error where there's an empty slot with no huge pagecache |
| 3030 | * to back it. This way, we avoid allocating a hugepage, and |
| 3031 | * the sparse dumpfile avoids allocating disk blocks, but its |
| 3032 | * huge holes still show up with zeroes where they need to be. |
| 3033 | */ |
| 3034 | if (absent && (flags & FOLL_DUMP) && |
| 3035 | !hugetlbfs_pagecache_present(h, vma, vaddr)) { |
| 3036 | if (pte) |
| 3037 | spin_unlock(ptl); |
| 3038 | remainder = 0; |
| 3039 | break; |
| 3040 | } |
| 3041 | |
| 3042 | /* |
| 3043 | * We need call hugetlb_fault for both hugepages under migration |
| 3044 | * (in which case hugetlb_fault waits for the migration,) and |
| 3045 | * hwpoisoned hugepages (in which case we need to prevent the |
| 3046 | * caller from accessing to them.) In order to do this, we use |
| 3047 | * here is_swap_pte instead of is_hugetlb_entry_migration and |
| 3048 | * is_hugetlb_entry_hwpoisoned. This is because it simply covers |
| 3049 | * both cases, and because we can't follow correct pages |
| 3050 | * directly from any kind of swap entries. |
| 3051 | */ |
| 3052 | if (absent || is_swap_pte(huge_ptep_get(pte)) || |
| 3053 | ((flags & FOLL_WRITE) && |
| 3054 | !huge_pte_write(huge_ptep_get(pte)))) { |
| 3055 | int ret; |
| 3056 | |
| 3057 | if (pte) |
| 3058 | spin_unlock(ptl); |
| 3059 | ret = hugetlb_fault(mm, vma, vaddr, |
| 3060 | (flags & FOLL_WRITE) ? FAULT_FLAG_WRITE : 0); |
| 3061 | if (!(ret & VM_FAULT_ERROR)) |
| 3062 | continue; |
| 3063 | |
| 3064 | remainder = 0; |
| 3065 | break; |
| 3066 | } |
| 3067 | |
| 3068 | pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT; |
| 3069 | page = pte_page(huge_ptep_get(pte)); |
| 3070 | same_page: |
| 3071 | if (pages) { |
| 3072 | pages[i] = mem_map_offset(page, pfn_offset); |
| 3073 | get_page_foll(pages[i]); |
| 3074 | } |
| 3075 | |
| 3076 | if (vmas) |
| 3077 | vmas[i] = vma; |
| 3078 | |
| 3079 | vaddr += PAGE_SIZE; |
| 3080 | ++pfn_offset; |
| 3081 | --remainder; |
| 3082 | ++i; |
| 3083 | if (vaddr < vma->vm_end && remainder && |
| 3084 | pfn_offset < pages_per_huge_page(h)) { |
| 3085 | /* |
| 3086 | * We use pfn_offset to avoid touching the pageframes |
| 3087 | * of this compound page. |
| 3088 | */ |
| 3089 | goto same_page; |
| 3090 | } |
| 3091 | spin_unlock(ptl); |
| 3092 | } |
| 3093 | *nr_pages = remainder; |
| 3094 | *position = vaddr; |
| 3095 | |
| 3096 | return i ? i : -EFAULT; |
| 3097 | } |
| 3098 | |
| 3099 | unsigned long hugetlb_change_protection(struct vm_area_struct *vma, |
| 3100 | unsigned long address, unsigned long end, pgprot_t newprot) |
| 3101 | { |
| 3102 | struct mm_struct *mm = vma->vm_mm; |
| 3103 | unsigned long start = address; |
| 3104 | pte_t *ptep; |
| 3105 | pte_t pte; |
| 3106 | struct hstate *h = hstate_vma(vma); |
| 3107 | unsigned long pages = 0; |
| 3108 | |
| 3109 | BUG_ON(address >= end); |
| 3110 | flush_cache_range(vma, address, end); |
| 3111 | |
| 3112 | mutex_lock(&vma->vm_file->f_mapping->i_mmap_mutex); |
| 3113 | for (; address < end; address += huge_page_size(h)) { |
| 3114 | spinlock_t *ptl; |
| 3115 | ptep = huge_pte_offset(mm, address); |
| 3116 | if (!ptep) |
| 3117 | continue; |
| 3118 | ptl = huge_pte_lock(h, mm, ptep); |
| 3119 | if (huge_pmd_unshare(mm, &address, ptep)) { |
| 3120 | pages++; |
| 3121 | spin_unlock(ptl); |
| 3122 | continue; |
| 3123 | } |
| 3124 | if (!huge_pte_none(huge_ptep_get(ptep))) { |
| 3125 | pte = huge_ptep_get_and_clear(mm, address, ptep); |
| 3126 | pte = pte_mkhuge(huge_pte_modify(pte, newprot)); |
| 3127 | pte = arch_make_huge_pte(pte, vma, NULL, 0); |
| 3128 | set_huge_pte_at(mm, address, ptep, pte); |
| 3129 | pages++; |
| 3130 | } |
| 3131 | spin_unlock(ptl); |
| 3132 | } |
| 3133 | /* |
| 3134 | * Must flush TLB before releasing i_mmap_mutex: x86's huge_pmd_unshare |
| 3135 | * may have cleared our pud entry and done put_page on the page table: |
| 3136 | * once we release i_mmap_mutex, another task can do the final put_page |
| 3137 | * and that page table be reused and filled with junk. |
| 3138 | */ |
| 3139 | flush_tlb_range(vma, start, end); |
| 3140 | mutex_unlock(&vma->vm_file->f_mapping->i_mmap_mutex); |
| 3141 | |
| 3142 | return pages << h->order; |
| 3143 | } |
| 3144 | |
| 3145 | int hugetlb_reserve_pages(struct inode *inode, |
| 3146 | long from, long to, |
| 3147 | struct vm_area_struct *vma, |
| 3148 | vm_flags_t vm_flags) |
| 3149 | { |
| 3150 | long ret, chg; |
| 3151 | struct hstate *h = hstate_inode(inode); |
| 3152 | struct hugepage_subpool *spool = subpool_inode(inode); |
| 3153 | |
| 3154 | /* |
| 3155 | * Only apply hugepage reservation if asked. At fault time, an |
| 3156 | * attempt will be made for VM_NORESERVE to allocate a page |
| 3157 | * without using reserves |
| 3158 | */ |
| 3159 | if (vm_flags & VM_NORESERVE) |
| 3160 | return 0; |
| 3161 | |
| 3162 | /* |
| 3163 | * Shared mappings base their reservation on the number of pages that |
| 3164 | * are already allocated on behalf of the file. Private mappings need |
| 3165 | * to reserve the full area even if read-only as mprotect() may be |
| 3166 | * called to make the mapping read-write. Assume !vma is a shm mapping |
| 3167 | */ |
| 3168 | if (!vma || vma->vm_flags & VM_MAYSHARE) |
| 3169 | chg = region_chg(&inode->i_mapping->private_list, from, to); |
| 3170 | else { |
| 3171 | struct resv_map *resv_map = resv_map_alloc(); |
| 3172 | if (!resv_map) |
| 3173 | return -ENOMEM; |
| 3174 | |
| 3175 | chg = to - from; |
| 3176 | |
| 3177 | set_vma_resv_map(vma, resv_map); |
| 3178 | set_vma_resv_flags(vma, HPAGE_RESV_OWNER); |
| 3179 | } |
| 3180 | |
| 3181 | if (chg < 0) { |
| 3182 | ret = chg; |
| 3183 | goto out_err; |
| 3184 | } |
| 3185 | |
| 3186 | /* There must be enough pages in the subpool for the mapping */ |
| 3187 | if (hugepage_subpool_get_pages(spool, chg)) { |
| 3188 | ret = -ENOSPC; |
| 3189 | goto out_err; |
| 3190 | } |
| 3191 | |
| 3192 | /* |
| 3193 | * Check enough hugepages are available for the reservation. |
| 3194 | * Hand the pages back to the subpool if there are not |
| 3195 | */ |
| 3196 | ret = hugetlb_acct_memory(h, chg); |
| 3197 | if (ret < 0) { |
| 3198 | hugepage_subpool_put_pages(spool, chg); |
| 3199 | goto out_err; |
| 3200 | } |
| 3201 | |
| 3202 | /* |
| 3203 | * Account for the reservations made. Shared mappings record regions |
| 3204 | * that have reservations as they are shared by multiple VMAs. |
| 3205 | * When the last VMA disappears, the region map says how much |
| 3206 | * the reservation was and the page cache tells how much of |
| 3207 | * the reservation was consumed. Private mappings are per-VMA and |
| 3208 | * only the consumed reservations are tracked. When the VMA |
| 3209 | * disappears, the original reservation is the VMA size and the |
| 3210 | * consumed reservations are stored in the map. Hence, nothing |
| 3211 | * else has to be done for private mappings here |
| 3212 | */ |
| 3213 | if (!vma || vma->vm_flags & VM_MAYSHARE) |
| 3214 | region_add(&inode->i_mapping->private_list, from, to); |
| 3215 | return 0; |
| 3216 | out_err: |
| 3217 | if (vma) |
| 3218 | resv_map_put(vma); |
| 3219 | return ret; |
| 3220 | } |
| 3221 | |
| 3222 | void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed) |
| 3223 | { |
| 3224 | struct hstate *h = hstate_inode(inode); |
| 3225 | long chg = region_truncate(&inode->i_mapping->private_list, offset); |
| 3226 | struct hugepage_subpool *spool = subpool_inode(inode); |
| 3227 | |
| 3228 | spin_lock(&inode->i_lock); |
| 3229 | inode->i_blocks -= (blocks_per_huge_page(h) * freed); |
| 3230 | spin_unlock(&inode->i_lock); |
| 3231 | |
| 3232 | hugepage_subpool_put_pages(spool, (chg - freed)); |
| 3233 | hugetlb_acct_memory(h, -(chg - freed)); |
| 3234 | } |
| 3235 | |
| 3236 | #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE |
| 3237 | static unsigned long page_table_shareable(struct vm_area_struct *svma, |
| 3238 | struct vm_area_struct *vma, |
| 3239 | unsigned long addr, pgoff_t idx) |
| 3240 | { |
| 3241 | unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) + |
| 3242 | svma->vm_start; |
| 3243 | unsigned long sbase = saddr & PUD_MASK; |
| 3244 | unsigned long s_end = sbase + PUD_SIZE; |
| 3245 | |
| 3246 | /* Allow segments to share if only one is marked locked */ |
| 3247 | unsigned long vm_flags = vma->vm_flags & ~VM_LOCKED; |
| 3248 | unsigned long svm_flags = svma->vm_flags & ~VM_LOCKED; |
| 3249 | |
| 3250 | /* |
| 3251 | * match the virtual addresses, permission and the alignment of the |
| 3252 | * page table page. |
| 3253 | */ |
| 3254 | if (pmd_index(addr) != pmd_index(saddr) || |
| 3255 | vm_flags != svm_flags || |
| 3256 | sbase < svma->vm_start || svma->vm_end < s_end) |
| 3257 | return 0; |
| 3258 | |
| 3259 | return saddr; |
| 3260 | } |
| 3261 | |
| 3262 | static int vma_shareable(struct vm_area_struct *vma, unsigned long addr) |
| 3263 | { |
| 3264 | unsigned long base = addr & PUD_MASK; |
| 3265 | unsigned long end = base + PUD_SIZE; |
| 3266 | |
| 3267 | /* |
| 3268 | * check on proper vm_flags and page table alignment |
| 3269 | */ |
| 3270 | if (vma->vm_flags & VM_MAYSHARE && |
| 3271 | vma->vm_start <= base && end <= vma->vm_end) |
| 3272 | return 1; |
| 3273 | return 0; |
| 3274 | } |
| 3275 | |
| 3276 | /* |
| 3277 | * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc() |
| 3278 | * and returns the corresponding pte. While this is not necessary for the |
| 3279 | * !shared pmd case because we can allocate the pmd later as well, it makes the |
| 3280 | * code much cleaner. pmd allocation is essential for the shared case because |
| 3281 | * pud has to be populated inside the same i_mmap_mutex section - otherwise |
| 3282 | * racing tasks could either miss the sharing (see huge_pte_offset) or select a |
| 3283 | * bad pmd for sharing. |
| 3284 | */ |
| 3285 | pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud) |
| 3286 | { |
| 3287 | struct vm_area_struct *vma = find_vma(mm, addr); |
| 3288 | struct address_space *mapping = vma->vm_file->f_mapping; |
| 3289 | pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) + |
| 3290 | vma->vm_pgoff; |
| 3291 | struct vm_area_struct *svma; |
| 3292 | unsigned long saddr; |
| 3293 | pte_t *spte = NULL; |
| 3294 | pte_t *pte; |
| 3295 | spinlock_t *ptl; |
| 3296 | |
| 3297 | if (!vma_shareable(vma, addr)) |
| 3298 | return (pte_t *)pmd_alloc(mm, pud, addr); |
| 3299 | |
| 3300 | mutex_lock(&mapping->i_mmap_mutex); |
| 3301 | vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) { |
| 3302 | if (svma == vma) |
| 3303 | continue; |
| 3304 | |
| 3305 | saddr = page_table_shareable(svma, vma, addr, idx); |
| 3306 | if (saddr) { |
| 3307 | spte = huge_pte_offset(svma->vm_mm, saddr); |
| 3308 | if (spte) { |
| 3309 | get_page(virt_to_page(spte)); |
| 3310 | break; |
| 3311 | } |
| 3312 | } |
| 3313 | } |
| 3314 | |
| 3315 | if (!spte) |
| 3316 | goto out; |
| 3317 | |
| 3318 | ptl = huge_pte_lockptr(hstate_vma(vma), mm, spte); |
| 3319 | spin_lock(ptl); |
| 3320 | if (pud_none(*pud)) |
| 3321 | pud_populate(mm, pud, |
| 3322 | (pmd_t *)((unsigned long)spte & PAGE_MASK)); |
| 3323 | else |
| 3324 | put_page(virt_to_page(spte)); |
| 3325 | spin_unlock(ptl); |
| 3326 | out: |
| 3327 | pte = (pte_t *)pmd_alloc(mm, pud, addr); |
| 3328 | mutex_unlock(&mapping->i_mmap_mutex); |
| 3329 | return pte; |
| 3330 | } |
| 3331 | |
| 3332 | /* |
| 3333 | * unmap huge page backed by shared pte. |
| 3334 | * |
| 3335 | * Hugetlb pte page is ref counted at the time of mapping. If pte is shared |
| 3336 | * indicated by page_count > 1, unmap is achieved by clearing pud and |
| 3337 | * decrementing the ref count. If count == 1, the pte page is not shared. |
| 3338 | * |
| 3339 | * called with page table lock held. |
| 3340 | * |
| 3341 | * returns: 1 successfully unmapped a shared pte page |
| 3342 | * 0 the underlying pte page is not shared, or it is the last user |
| 3343 | */ |
| 3344 | int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep) |
| 3345 | { |
| 3346 | pgd_t *pgd = pgd_offset(mm, *addr); |
| 3347 | pud_t *pud = pud_offset(pgd, *addr); |
| 3348 | |
| 3349 | BUG_ON(page_count(virt_to_page(ptep)) == 0); |
| 3350 | if (page_count(virt_to_page(ptep)) == 1) |
| 3351 | return 0; |
| 3352 | |
| 3353 | pud_clear(pud); |
| 3354 | put_page(virt_to_page(ptep)); |
| 3355 | *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE; |
| 3356 | return 1; |
| 3357 | } |
| 3358 | #define want_pmd_share() (1) |
| 3359 | #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ |
| 3360 | pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud) |
| 3361 | { |
| 3362 | return NULL; |
| 3363 | } |
| 3364 | #define want_pmd_share() (0) |
| 3365 | #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ |
| 3366 | |
| 3367 | #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB |
| 3368 | pte_t *huge_pte_alloc(struct mm_struct *mm, |
| 3369 | unsigned long addr, unsigned long sz) |
| 3370 | { |
| 3371 | pgd_t *pgd; |
| 3372 | pud_t *pud; |
| 3373 | pte_t *pte = NULL; |
| 3374 | |
| 3375 | pgd = pgd_offset(mm, addr); |
| 3376 | pud = pud_alloc(mm, pgd, addr); |
| 3377 | if (pud) { |
| 3378 | if (sz == PUD_SIZE) { |
| 3379 | pte = (pte_t *)pud; |
| 3380 | } else { |
| 3381 | BUG_ON(sz != PMD_SIZE); |
| 3382 | if (want_pmd_share() && pud_none(*pud)) |
| 3383 | pte = huge_pmd_share(mm, addr, pud); |
| 3384 | else |
| 3385 | pte = (pte_t *)pmd_alloc(mm, pud, addr); |
| 3386 | } |
| 3387 | } |
| 3388 | BUG_ON(pte && !pte_none(*pte) && !pte_huge(*pte)); |
| 3389 | |
| 3390 | return pte; |
| 3391 | } |
| 3392 | |
| 3393 | pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr) |
| 3394 | { |
| 3395 | pgd_t *pgd; |
| 3396 | pud_t *pud; |
| 3397 | pmd_t *pmd = NULL; |
| 3398 | |
| 3399 | pgd = pgd_offset(mm, addr); |
| 3400 | if (pgd_present(*pgd)) { |
| 3401 | pud = pud_offset(pgd, addr); |
| 3402 | if (pud_present(*pud)) { |
| 3403 | if (pud_huge(*pud)) |
| 3404 | return (pte_t *)pud; |
| 3405 | pmd = pmd_offset(pud, addr); |
| 3406 | } |
| 3407 | } |
| 3408 | return (pte_t *) pmd; |
| 3409 | } |
| 3410 | |
| 3411 | struct page * |
| 3412 | follow_huge_pmd(struct mm_struct *mm, unsigned long address, |
| 3413 | pmd_t *pmd, int write) |
| 3414 | { |
| 3415 | struct page *page; |
| 3416 | |
| 3417 | page = pte_page(*(pte_t *)pmd); |
| 3418 | if (page) |
| 3419 | page += ((address & ~PMD_MASK) >> PAGE_SHIFT); |
| 3420 | return page; |
| 3421 | } |
| 3422 | |
| 3423 | struct page * |
| 3424 | follow_huge_pud(struct mm_struct *mm, unsigned long address, |
| 3425 | pud_t *pud, int write) |
| 3426 | { |
| 3427 | struct page *page; |
| 3428 | |
| 3429 | page = pte_page(*(pte_t *)pud); |
| 3430 | if (page) |
| 3431 | page += ((address & ~PUD_MASK) >> PAGE_SHIFT); |
| 3432 | return page; |
| 3433 | } |
| 3434 | |
| 3435 | #else /* !CONFIG_ARCH_WANT_GENERAL_HUGETLB */ |
| 3436 | |
| 3437 | /* Can be overriden by architectures */ |
| 3438 | __attribute__((weak)) struct page * |
| 3439 | follow_huge_pud(struct mm_struct *mm, unsigned long address, |
| 3440 | pud_t *pud, int write) |
| 3441 | { |
| 3442 | BUG(); |
| 3443 | return NULL; |
| 3444 | } |
| 3445 | |
| 3446 | #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */ |
| 3447 | |
| 3448 | #ifdef CONFIG_MEMORY_FAILURE |
| 3449 | |
| 3450 | /* Should be called in hugetlb_lock */ |
| 3451 | static int is_hugepage_on_freelist(struct page *hpage) |
| 3452 | { |
| 3453 | struct page *page; |
| 3454 | struct page *tmp; |
| 3455 | struct hstate *h = page_hstate(hpage); |
| 3456 | int nid = page_to_nid(hpage); |
| 3457 | |
| 3458 | list_for_each_entry_safe(page, tmp, &h->hugepage_freelists[nid], lru) |
| 3459 | if (page == hpage) |
| 3460 | return 1; |
| 3461 | return 0; |
| 3462 | } |
| 3463 | |
| 3464 | /* |
| 3465 | * This function is called from memory failure code. |
| 3466 | * Assume the caller holds page lock of the head page. |
| 3467 | */ |
| 3468 | int dequeue_hwpoisoned_huge_page(struct page *hpage) |
| 3469 | { |
| 3470 | struct hstate *h = page_hstate(hpage); |
| 3471 | int nid = page_to_nid(hpage); |
| 3472 | int ret = -EBUSY; |
| 3473 | |
| 3474 | spin_lock(&hugetlb_lock); |
| 3475 | if (is_hugepage_on_freelist(hpage)) { |
| 3476 | /* |
| 3477 | * Hwpoisoned hugepage isn't linked to activelist or freelist, |
| 3478 | * but dangling hpage->lru can trigger list-debug warnings |
| 3479 | * (this happens when we call unpoison_memory() on it), |
| 3480 | * so let it point to itself with list_del_init(). |
| 3481 | */ |
| 3482 | list_del_init(&hpage->lru); |
| 3483 | set_page_refcounted(hpage); |
| 3484 | h->free_huge_pages--; |
| 3485 | h->free_huge_pages_node[nid]--; |
| 3486 | ret = 0; |
| 3487 | } |
| 3488 | spin_unlock(&hugetlb_lock); |
| 3489 | return ret; |
| 3490 | } |
| 3491 | #endif |
| 3492 | |
| 3493 | bool isolate_huge_page(struct page *page, struct list_head *list) |
| 3494 | { |
| 3495 | VM_BUG_ON(!PageHead(page)); |
| 3496 | if (!get_page_unless_zero(page)) |
| 3497 | return false; |
| 3498 | spin_lock(&hugetlb_lock); |
| 3499 | list_move_tail(&page->lru, list); |
| 3500 | spin_unlock(&hugetlb_lock); |
| 3501 | return true; |
| 3502 | } |
| 3503 | |
| 3504 | void putback_active_hugepage(struct page *page) |
| 3505 | { |
| 3506 | VM_BUG_ON(!PageHead(page)); |
| 3507 | spin_lock(&hugetlb_lock); |
| 3508 | list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist); |
| 3509 | spin_unlock(&hugetlb_lock); |
| 3510 | put_page(page); |
| 3511 | } |
| 3512 | |
| 3513 | bool is_hugepage_active(struct page *page) |
| 3514 | { |
| 3515 | VM_BUG_ON(!PageHuge(page)); |
| 3516 | /* |
| 3517 | * This function can be called for a tail page because the caller, |
| 3518 | * scan_movable_pages, scans through a given pfn-range which typically |
| 3519 | * covers one memory block. In systems using gigantic hugepage (1GB |
| 3520 | * for x86_64,) a hugepage is larger than a memory block, and we don't |
| 3521 | * support migrating such large hugepages for now, so return false |
| 3522 | * when called for tail pages. |
| 3523 | */ |
| 3524 | if (PageTail(page)) |
| 3525 | return false; |
| 3526 | /* |
| 3527 | * Refcount of a hwpoisoned hugepages is 1, but they are not active, |
| 3528 | * so we should return false for them. |
| 3529 | */ |
| 3530 | if (unlikely(PageHWPoison(page))) |
| 3531 | return false; |
| 3532 | return page_count(page) > 0; |
| 3533 | } |