1 // SPDX-License-Identifier: GPL-2.0-only 2 /* 3 * Generic hugetlb support. 4 * (C) Nadia Yvette Chambers, April 2004 5 */ 6 #include <linux/list.h> 7 #include <linux/init.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/compiler.h> 17 #include <linux/cpuset.h> 18 #include <linux/mutex.h> 19 #include <linux/memblock.h> 20 #include <linux/sysfs.h> 21 #include <linux/slab.h> 22 #include <linux/sched/mm.h> 23 #include <linux/mmdebug.h> 24 #include <linux/sched/signal.h> 25 #include <linux/rmap.h> 26 #include <linux/string_helpers.h> 27 #include <linux/swap.h> 28 #include <linux/swapops.h> 29 #include <linux/jhash.h> 30 #include <linux/numa.h> 31 #include <linux/llist.h> 32 #include <linux/cma.h> 33 #include <linux/migrate.h> 34 #include <linux/nospec.h> 35 #include <linux/delayacct.h> 36 #include <linux/memory.h> 37 #include <linux/mm_inline.h> 38 #include <linux/padata.h> 39 40 #include <asm/page.h> 41 #include <asm/pgalloc.h> 42 #include <asm/tlb.h> 43 44 #include <linux/io.h> 45 #include <linux/hugetlb.h> 46 #include <linux/hugetlb_cgroup.h> 47 #include <linux/node.h> 48 #include <linux/page_owner.h> 49 #include "internal.h" 50 #include "hugetlb_vmemmap.h" 51 #include <linux/page-isolation.h> 52 53 int hugetlb_max_hstate __read_mostly; 54 unsigned int default_hstate_idx; 55 struct hstate hstates[HUGE_MAX_HSTATE]; 56 57 #ifdef CONFIG_CMA 58 static struct cma *hugetlb_cma[MAX_NUMNODES]; 59 static unsigned long hugetlb_cma_size_in_node[MAX_NUMNODES] __initdata; 60 #endif 61 static unsigned long hugetlb_cma_size __initdata; 62 63 __initdata struct list_head huge_boot_pages[MAX_NUMNODES]; 64 65 /* for command line parsing */ 66 static struct hstate * __initdata parsed_hstate; 67 static unsigned long __initdata default_hstate_max_huge_pages; 68 static bool __initdata parsed_valid_hugepagesz = true; 69 static bool __initdata parsed_default_hugepagesz; 70 static unsigned int default_hugepages_in_node[MAX_NUMNODES] __initdata; 71 72 /* 73 * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages, 74 * free_huge_pages, and surplus_huge_pages. 75 */ 76 __cacheline_aligned_in_smp DEFINE_SPINLOCK(hugetlb_lock); 77 78 /* 79 * Serializes faults on the same logical page. This is used to 80 * prevent spurious OOMs when the hugepage pool is fully utilized. 81 */ 82 static int num_fault_mutexes __ro_after_init; 83 struct mutex *hugetlb_fault_mutex_table __ro_after_init; 84 85 /* Forward declaration */ 86 static int hugetlb_acct_memory(struct hstate *h, long delta); 87 static void hugetlb_vma_lock_free(struct vm_area_struct *vma); 88 static void hugetlb_vma_lock_alloc(struct vm_area_struct *vma); 89 static void __hugetlb_vma_unlock_write_free(struct vm_area_struct *vma); 90 static void hugetlb_unshare_pmds(struct vm_area_struct *vma, 91 unsigned long start, unsigned long end); 92 static struct resv_map *vma_resv_map(struct vm_area_struct *vma); 93 94 static void hugetlb_free_folio(struct folio *folio) 95 { 96 #ifdef CONFIG_CMA 97 int nid = folio_nid(folio); 98 99 if (cma_free_folio(hugetlb_cma[nid], folio)) 100 return; 101 #endif 102 folio_put(folio); 103 } 104 105 static inline bool subpool_is_free(struct hugepage_subpool *spool) 106 { 107 if (spool->count) 108 return false; 109 if (spool->max_hpages != -1) 110 return spool->used_hpages == 0; 111 if (spool->min_hpages != -1) 112 return spool->rsv_hpages == spool->min_hpages; 113 114 return true; 115 } 116 117 static inline void unlock_or_release_subpool(struct hugepage_subpool *spool, 118 unsigned long irq_flags) 119 { 120 spin_unlock_irqrestore(&spool->lock, irq_flags); 121 122 /* If no pages are used, and no other handles to the subpool 123 * remain, give up any reservations based on minimum size and 124 * free the subpool */ 125 if (subpool_is_free(spool)) { 126 if (spool->min_hpages != -1) 127 hugetlb_acct_memory(spool->hstate, 128 -spool->min_hpages); 129 kfree(spool); 130 } 131 } 132 133 struct hugepage_subpool *hugepage_new_subpool(struct hstate *h, long max_hpages, 134 long min_hpages) 135 { 136 struct hugepage_subpool *spool; 137 138 spool = kzalloc(sizeof(*spool), GFP_KERNEL); 139 if (!spool) 140 return NULL; 141 142 spin_lock_init(&spool->lock); 143 spool->count = 1; 144 spool->max_hpages = max_hpages; 145 spool->hstate = h; 146 spool->min_hpages = min_hpages; 147 148 if (min_hpages != -1 && hugetlb_acct_memory(h, min_hpages)) { 149 kfree(spool); 150 return NULL; 151 } 152 spool->rsv_hpages = min_hpages; 153 154 return spool; 155 } 156 157 void hugepage_put_subpool(struct hugepage_subpool *spool) 158 { 159 unsigned long flags; 160 161 spin_lock_irqsave(&spool->lock, flags); 162 BUG_ON(!spool->count); 163 spool->count--; 164 unlock_or_release_subpool(spool, flags); 165 } 166 167 /* 168 * Subpool accounting for allocating and reserving pages. 169 * Return -ENOMEM if there are not enough resources to satisfy the 170 * request. Otherwise, return the number of pages by which the 171 * global pools must be adjusted (upward). The returned value may 172 * only be different than the passed value (delta) in the case where 173 * a subpool minimum size must be maintained. 174 */ 175 static long hugepage_subpool_get_pages(struct hugepage_subpool *spool, 176 long delta) 177 { 178 long ret = delta; 179 180 if (!spool) 181 return ret; 182 183 spin_lock_irq(&spool->lock); 184 185 if (spool->max_hpages != -1) { /* maximum size accounting */ 186 if ((spool->used_hpages + delta) <= spool->max_hpages) 187 spool->used_hpages += delta; 188 else { 189 ret = -ENOMEM; 190 goto unlock_ret; 191 } 192 } 193 194 /* minimum size accounting */ 195 if (spool->min_hpages != -1 && spool->rsv_hpages) { 196 if (delta > spool->rsv_hpages) { 197 /* 198 * Asking for more reserves than those already taken on 199 * behalf of subpool. Return difference. 200 */ 201 ret = delta - spool->rsv_hpages; 202 spool->rsv_hpages = 0; 203 } else { 204 ret = 0; /* reserves already accounted for */ 205 spool->rsv_hpages -= delta; 206 } 207 } 208 209 unlock_ret: 210 spin_unlock_irq(&spool->lock); 211 return ret; 212 } 213 214 /* 215 * Subpool accounting for freeing and unreserving pages. 216 * Return the number of global page reservations that must be dropped. 217 * The return value may only be different than the passed value (delta) 218 * in the case where a subpool minimum size must be maintained. 219 */ 220 static long hugepage_subpool_put_pages(struct hugepage_subpool *spool, 221 long delta) 222 { 223 long ret = delta; 224 unsigned long flags; 225 226 if (!spool) 227 return delta; 228 229 spin_lock_irqsave(&spool->lock, flags); 230 231 if (spool->max_hpages != -1) /* maximum size accounting */ 232 spool->used_hpages -= delta; 233 234 /* minimum size accounting */ 235 if (spool->min_hpages != -1 && spool->used_hpages < spool->min_hpages) { 236 if (spool->rsv_hpages + delta <= spool->min_hpages) 237 ret = 0; 238 else 239 ret = spool->rsv_hpages + delta - spool->min_hpages; 240 241 spool->rsv_hpages += delta; 242 if (spool->rsv_hpages > spool->min_hpages) 243 spool->rsv_hpages = spool->min_hpages; 244 } 245 246 /* 247 * If hugetlbfs_put_super couldn't free spool due to an outstanding 248 * quota reference, free it now. 249 */ 250 unlock_or_release_subpool(spool, flags); 251 252 return ret; 253 } 254 255 static inline struct hugepage_subpool *subpool_inode(struct inode *inode) 256 { 257 return HUGETLBFS_SB(inode->i_sb)->spool; 258 } 259 260 static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma) 261 { 262 return subpool_inode(file_inode(vma->vm_file)); 263 } 264 265 /* 266 * hugetlb vma_lock helper routines 267 */ 268 void hugetlb_vma_lock_read(struct vm_area_struct *vma) 269 { 270 if (__vma_shareable_lock(vma)) { 271 struct hugetlb_vma_lock *vma_lock = vma->vm_private_data; 272 273 down_read(&vma_lock->rw_sema); 274 } else if (__vma_private_lock(vma)) { 275 struct resv_map *resv_map = vma_resv_map(vma); 276 277 down_read(&resv_map->rw_sema); 278 } 279 } 280 281 void hugetlb_vma_unlock_read(struct vm_area_struct *vma) 282 { 283 if (__vma_shareable_lock(vma)) { 284 struct hugetlb_vma_lock *vma_lock = vma->vm_private_data; 285 286 up_read(&vma_lock->rw_sema); 287 } else if (__vma_private_lock(vma)) { 288 struct resv_map *resv_map = vma_resv_map(vma); 289 290 up_read(&resv_map->rw_sema); 291 } 292 } 293 294 void hugetlb_vma_lock_write(struct vm_area_struct *vma) 295 { 296 if (__vma_shareable_lock(vma)) { 297 struct hugetlb_vma_lock *vma_lock = vma->vm_private_data; 298 299 down_write(&vma_lock->rw_sema); 300 } else if (__vma_private_lock(vma)) { 301 struct resv_map *resv_map = vma_resv_map(vma); 302 303 down_write(&resv_map->rw_sema); 304 } 305 } 306 307 void hugetlb_vma_unlock_write(struct vm_area_struct *vma) 308 { 309 if (__vma_shareable_lock(vma)) { 310 struct hugetlb_vma_lock *vma_lock = vma->vm_private_data; 311 312 up_write(&vma_lock->rw_sema); 313 } else if (__vma_private_lock(vma)) { 314 struct resv_map *resv_map = vma_resv_map(vma); 315 316 up_write(&resv_map->rw_sema); 317 } 318 } 319 320 int hugetlb_vma_trylock_write(struct vm_area_struct *vma) 321 { 322 323 if (__vma_shareable_lock(vma)) { 324 struct hugetlb_vma_lock *vma_lock = vma->vm_private_data; 325 326 return down_write_trylock(&vma_lock->rw_sema); 327 } else if (__vma_private_lock(vma)) { 328 struct resv_map *resv_map = vma_resv_map(vma); 329 330 return down_write_trylock(&resv_map->rw_sema); 331 } 332 333 return 1; 334 } 335 336 void hugetlb_vma_assert_locked(struct vm_area_struct *vma) 337 { 338 if (__vma_shareable_lock(vma)) { 339 struct hugetlb_vma_lock *vma_lock = vma->vm_private_data; 340 341 lockdep_assert_held(&vma_lock->rw_sema); 342 } else if (__vma_private_lock(vma)) { 343 struct resv_map *resv_map = vma_resv_map(vma); 344 345 lockdep_assert_held(&resv_map->rw_sema); 346 } 347 } 348 349 void hugetlb_vma_lock_release(struct kref *kref) 350 { 351 struct hugetlb_vma_lock *vma_lock = container_of(kref, 352 struct hugetlb_vma_lock, refs); 353 354 kfree(vma_lock); 355 } 356 357 static void __hugetlb_vma_unlock_write_put(struct hugetlb_vma_lock *vma_lock) 358 { 359 struct vm_area_struct *vma = vma_lock->vma; 360 361 /* 362 * vma_lock structure may or not be released as a result of put, 363 * it certainly will no longer be attached to vma so clear pointer. 364 * Semaphore synchronizes access to vma_lock->vma field. 365 */ 366 vma_lock->vma = NULL; 367 vma->vm_private_data = NULL; 368 up_write(&vma_lock->rw_sema); 369 kref_put(&vma_lock->refs, hugetlb_vma_lock_release); 370 } 371 372 static void __hugetlb_vma_unlock_write_free(struct vm_area_struct *vma) 373 { 374 if (__vma_shareable_lock(vma)) { 375 struct hugetlb_vma_lock *vma_lock = vma->vm_private_data; 376 377 __hugetlb_vma_unlock_write_put(vma_lock); 378 } else if (__vma_private_lock(vma)) { 379 struct resv_map *resv_map = vma_resv_map(vma); 380 381 /* no free for anon vmas, but still need to unlock */ 382 up_write(&resv_map->rw_sema); 383 } 384 } 385 386 static void hugetlb_vma_lock_free(struct vm_area_struct *vma) 387 { 388 /* 389 * Only present in sharable vmas. 390 */ 391 if (!vma || !__vma_shareable_lock(vma)) 392 return; 393 394 if (vma->vm_private_data) { 395 struct hugetlb_vma_lock *vma_lock = vma->vm_private_data; 396 397 down_write(&vma_lock->rw_sema); 398 __hugetlb_vma_unlock_write_put(vma_lock); 399 } 400 } 401 402 static void hugetlb_vma_lock_alloc(struct vm_area_struct *vma) 403 { 404 struct hugetlb_vma_lock *vma_lock; 405 406 /* Only establish in (flags) sharable vmas */ 407 if (!vma || !(vma->vm_flags & VM_MAYSHARE)) 408 return; 409 410 /* Should never get here with non-NULL vm_private_data */ 411 if (vma->vm_private_data) 412 return; 413 414 vma_lock = kmalloc(sizeof(*vma_lock), GFP_KERNEL); 415 if (!vma_lock) { 416 /* 417 * If we can not allocate structure, then vma can not 418 * participate in pmd sharing. This is only a possible 419 * performance enhancement and memory saving issue. 420 * However, the lock is also used to synchronize page 421 * faults with truncation. If the lock is not present, 422 * unlikely races could leave pages in a file past i_size 423 * until the file is removed. Warn in the unlikely case of 424 * allocation failure. 425 */ 426 pr_warn_once("HugeTLB: unable to allocate vma specific lock\n"); 427 return; 428 } 429 430 kref_init(&vma_lock->refs); 431 init_rwsem(&vma_lock->rw_sema); 432 vma_lock->vma = vma; 433 vma->vm_private_data = vma_lock; 434 } 435 436 /* Helper that removes a struct file_region from the resv_map cache and returns 437 * it for use. 438 */ 439 static struct file_region * 440 get_file_region_entry_from_cache(struct resv_map *resv, long from, long to) 441 { 442 struct file_region *nrg; 443 444 VM_BUG_ON(resv->region_cache_count <= 0); 445 446 resv->region_cache_count--; 447 nrg = list_first_entry(&resv->region_cache, struct file_region, link); 448 list_del(&nrg->link); 449 450 nrg->from = from; 451 nrg->to = to; 452 453 return nrg; 454 } 455 456 static void copy_hugetlb_cgroup_uncharge_info(struct file_region *nrg, 457 struct file_region *rg) 458 { 459 #ifdef CONFIG_CGROUP_HUGETLB 460 nrg->reservation_counter = rg->reservation_counter; 461 nrg->css = rg->css; 462 if (rg->css) 463 css_get(rg->css); 464 #endif 465 } 466 467 /* Helper that records hugetlb_cgroup uncharge info. */ 468 static void record_hugetlb_cgroup_uncharge_info(struct hugetlb_cgroup *h_cg, 469 struct hstate *h, 470 struct resv_map *resv, 471 struct file_region *nrg) 472 { 473 #ifdef CONFIG_CGROUP_HUGETLB 474 if (h_cg) { 475 nrg->reservation_counter = 476 &h_cg->rsvd_hugepage[hstate_index(h)]; 477 nrg->css = &h_cg->css; 478 /* 479 * The caller will hold exactly one h_cg->css reference for the 480 * whole contiguous reservation region. But this area might be 481 * scattered when there are already some file_regions reside in 482 * it. As a result, many file_regions may share only one css 483 * reference. In order to ensure that one file_region must hold 484 * exactly one h_cg->css reference, we should do css_get for 485 * each file_region and leave the reference held by caller 486 * untouched. 487 */ 488 css_get(&h_cg->css); 489 if (!resv->pages_per_hpage) 490 resv->pages_per_hpage = pages_per_huge_page(h); 491 /* pages_per_hpage should be the same for all entries in 492 * a resv_map. 493 */ 494 VM_BUG_ON(resv->pages_per_hpage != pages_per_huge_page(h)); 495 } else { 496 nrg->reservation_counter = NULL; 497 nrg->css = NULL; 498 } 499 #endif 500 } 501 502 static void put_uncharge_info(struct file_region *rg) 503 { 504 #ifdef CONFIG_CGROUP_HUGETLB 505 if (rg->css) 506 css_put(rg->css); 507 #endif 508 } 509 510 static bool has_same_uncharge_info(struct file_region *rg, 511 struct file_region *org) 512 { 513 #ifdef CONFIG_CGROUP_HUGETLB 514 return rg->reservation_counter == org->reservation_counter && 515 rg->css == org->css; 516 517 #else 518 return true; 519 #endif 520 } 521 522 static void coalesce_file_region(struct resv_map *resv, struct file_region *rg) 523 { 524 struct file_region *nrg, *prg; 525 526 prg = list_prev_entry(rg, link); 527 if (&prg->link != &resv->regions && prg->to == rg->from && 528 has_same_uncharge_info(prg, rg)) { 529 prg->to = rg->to; 530 531 list_del(&rg->link); 532 put_uncharge_info(rg); 533 kfree(rg); 534 535 rg = prg; 536 } 537 538 nrg = list_next_entry(rg, link); 539 if (&nrg->link != &resv->regions && nrg->from == rg->to && 540 has_same_uncharge_info(nrg, rg)) { 541 nrg->from = rg->from; 542 543 list_del(&rg->link); 544 put_uncharge_info(rg); 545 kfree(rg); 546 } 547 } 548 549 static inline long 550 hugetlb_resv_map_add(struct resv_map *map, struct list_head *rg, long from, 551 long to, struct hstate *h, struct hugetlb_cgroup *cg, 552 long *regions_needed) 553 { 554 struct file_region *nrg; 555 556 if (!regions_needed) { 557 nrg = get_file_region_entry_from_cache(map, from, to); 558 record_hugetlb_cgroup_uncharge_info(cg, h, map, nrg); 559 list_add(&nrg->link, rg); 560 coalesce_file_region(map, nrg); 561 } else 562 *regions_needed += 1; 563 564 return to - from; 565 } 566 567 /* 568 * Must be called with resv->lock held. 569 * 570 * Calling this with regions_needed != NULL will count the number of pages 571 * to be added but will not modify the linked list. And regions_needed will 572 * indicate the number of file_regions needed in the cache to carry out to add 573 * the regions for this range. 574 */ 575 static long add_reservation_in_range(struct resv_map *resv, long f, long t, 576 struct hugetlb_cgroup *h_cg, 577 struct hstate *h, long *regions_needed) 578 { 579 long add = 0; 580 struct list_head *head = &resv->regions; 581 long last_accounted_offset = f; 582 struct file_region *iter, *trg = NULL; 583 struct list_head *rg = NULL; 584 585 if (regions_needed) 586 *regions_needed = 0; 587 588 /* In this loop, we essentially handle an entry for the range 589 * [last_accounted_offset, iter->from), at every iteration, with some 590 * bounds checking. 591 */ 592 list_for_each_entry_safe(iter, trg, head, link) { 593 /* Skip irrelevant regions that start before our range. */ 594 if (iter->from < f) { 595 /* If this region ends after the last accounted offset, 596 * then we need to update last_accounted_offset. 597 */ 598 if (iter->to > last_accounted_offset) 599 last_accounted_offset = iter->to; 600 continue; 601 } 602 603 /* When we find a region that starts beyond our range, we've 604 * finished. 605 */ 606 if (iter->from >= t) { 607 rg = iter->link.prev; 608 break; 609 } 610 611 /* Add an entry for last_accounted_offset -> iter->from, and 612 * update last_accounted_offset. 613 */ 614 if (iter->from > last_accounted_offset) 615 add += hugetlb_resv_map_add(resv, iter->link.prev, 616 last_accounted_offset, 617 iter->from, h, h_cg, 618 regions_needed); 619 620 last_accounted_offset = iter->to; 621 } 622 623 /* Handle the case where our range extends beyond 624 * last_accounted_offset. 625 */ 626 if (!rg) 627 rg = head->prev; 628 if (last_accounted_offset < t) 629 add += hugetlb_resv_map_add(resv, rg, last_accounted_offset, 630 t, h, h_cg, regions_needed); 631 632 return add; 633 } 634 635 /* Must be called with resv->lock acquired. Will drop lock to allocate entries. 636 */ 637 static int allocate_file_region_entries(struct resv_map *resv, 638 int regions_needed) 639 __must_hold(&resv->lock) 640 { 641 LIST_HEAD(allocated_regions); 642 int to_allocate = 0, i = 0; 643 struct file_region *trg = NULL, *rg = NULL; 644 645 VM_BUG_ON(regions_needed < 0); 646 647 /* 648 * Check for sufficient descriptors in the cache to accommodate 649 * the number of in progress add operations plus regions_needed. 650 * 651 * This is a while loop because when we drop the lock, some other call 652 * to region_add or region_del may have consumed some region_entries, 653 * so we keep looping here until we finally have enough entries for 654 * (adds_in_progress + regions_needed). 655 */ 656 while (resv->region_cache_count < 657 (resv->adds_in_progress + regions_needed)) { 658 to_allocate = resv->adds_in_progress + regions_needed - 659 resv->region_cache_count; 660 661 /* At this point, we should have enough entries in the cache 662 * for all the existing adds_in_progress. We should only be 663 * needing to allocate for regions_needed. 664 */ 665 VM_BUG_ON(resv->region_cache_count < resv->adds_in_progress); 666 667 spin_unlock(&resv->lock); 668 for (i = 0; i < to_allocate; i++) { 669 trg = kmalloc(sizeof(*trg), GFP_KERNEL); 670 if (!trg) 671 goto out_of_memory; 672 list_add(&trg->link, &allocated_regions); 673 } 674 675 spin_lock(&resv->lock); 676 677 list_splice(&allocated_regions, &resv->region_cache); 678 resv->region_cache_count += to_allocate; 679 } 680 681 return 0; 682 683 out_of_memory: 684 list_for_each_entry_safe(rg, trg, &allocated_regions, link) { 685 list_del(&rg->link); 686 kfree(rg); 687 } 688 return -ENOMEM; 689 } 690 691 /* 692 * Add the huge page range represented by [f, t) to the reserve 693 * map. Regions will be taken from the cache to fill in this range. 694 * Sufficient regions should exist in the cache due to the previous 695 * call to region_chg with the same range, but in some cases the cache will not 696 * have sufficient entries due to races with other code doing region_add or 697 * region_del. The extra needed entries will be allocated. 698 * 699 * regions_needed is the out value provided by a previous call to region_chg. 700 * 701 * Return the number of new huge pages added to the map. This number is greater 702 * than or equal to zero. If file_region entries needed to be allocated for 703 * this operation and we were not able to allocate, it returns -ENOMEM. 704 * region_add of regions of length 1 never allocate file_regions and cannot 705 * fail; region_chg will always allocate at least 1 entry and a region_add for 706 * 1 page will only require at most 1 entry. 707 */ 708 static long region_add(struct resv_map *resv, long f, long t, 709 long in_regions_needed, struct hstate *h, 710 struct hugetlb_cgroup *h_cg) 711 { 712 long add = 0, actual_regions_needed = 0; 713 714 spin_lock(&resv->lock); 715 retry: 716 717 /* Count how many regions are actually needed to execute this add. */ 718 add_reservation_in_range(resv, f, t, NULL, NULL, 719 &actual_regions_needed); 720 721 /* 722 * Check for sufficient descriptors in the cache to accommodate 723 * this add operation. Note that actual_regions_needed may be greater 724 * than in_regions_needed, as the resv_map may have been modified since 725 * the region_chg call. In this case, we need to make sure that we 726 * allocate extra entries, such that we have enough for all the 727 * existing adds_in_progress, plus the excess needed for this 728 * operation. 729 */ 730 if (actual_regions_needed > in_regions_needed && 731 resv->region_cache_count < 732 resv->adds_in_progress + 733 (actual_regions_needed - in_regions_needed)) { 734 /* region_add operation of range 1 should never need to 735 * allocate file_region entries. 736 */ 737 VM_BUG_ON(t - f <= 1); 738 739 if (allocate_file_region_entries( 740 resv, actual_regions_needed - in_regions_needed)) { 741 return -ENOMEM; 742 } 743 744 goto retry; 745 } 746 747 add = add_reservation_in_range(resv, f, t, h_cg, h, NULL); 748 749 resv->adds_in_progress -= in_regions_needed; 750 751 spin_unlock(&resv->lock); 752 return add; 753 } 754 755 /* 756 * Examine the existing reserve map and determine how many 757 * huge pages in the specified range [f, t) are NOT currently 758 * represented. This routine is called before a subsequent 759 * call to region_add that will actually modify the reserve 760 * map to add the specified range [f, t). region_chg does 761 * not change the number of huge pages represented by the 762 * map. A number of new file_region structures is added to the cache as a 763 * placeholder, for the subsequent region_add call to use. At least 1 764 * file_region structure is added. 765 * 766 * out_regions_needed is the number of regions added to the 767 * resv->adds_in_progress. This value needs to be provided to a follow up call 768 * to region_add or region_abort for proper accounting. 769 * 770 * Returns the number of huge pages that need to be added to the existing 771 * reservation map for the range [f, t). This number is greater or equal to 772 * zero. -ENOMEM is returned if a new file_region structure or cache entry 773 * is needed and can not be allocated. 774 */ 775 static long region_chg(struct resv_map *resv, long f, long t, 776 long *out_regions_needed) 777 { 778 long chg = 0; 779 780 spin_lock(&resv->lock); 781 782 /* Count how many hugepages in this range are NOT represented. */ 783 chg = add_reservation_in_range(resv, f, t, NULL, NULL, 784 out_regions_needed); 785 786 if (*out_regions_needed == 0) 787 *out_regions_needed = 1; 788 789 if (allocate_file_region_entries(resv, *out_regions_needed)) 790 return -ENOMEM; 791 792 resv->adds_in_progress += *out_regions_needed; 793 794 spin_unlock(&resv->lock); 795 return chg; 796 } 797 798 /* 799 * Abort the in progress add operation. The adds_in_progress field 800 * of the resv_map keeps track of the operations in progress between 801 * calls to region_chg and region_add. Operations are sometimes 802 * aborted after the call to region_chg. In such cases, region_abort 803 * is called to decrement the adds_in_progress counter. regions_needed 804 * is the value returned by the region_chg call, it is used to decrement 805 * the adds_in_progress counter. 806 * 807 * NOTE: The range arguments [f, t) are not needed or used in this 808 * routine. They are kept to make reading the calling code easier as 809 * arguments will match the associated region_chg call. 810 */ 811 static void region_abort(struct resv_map *resv, long f, long t, 812 long regions_needed) 813 { 814 spin_lock(&resv->lock); 815 VM_BUG_ON(!resv->region_cache_count); 816 resv->adds_in_progress -= regions_needed; 817 spin_unlock(&resv->lock); 818 } 819 820 /* 821 * Delete the specified range [f, t) from the reserve map. If the 822 * t parameter is LONG_MAX, this indicates that ALL regions after f 823 * should be deleted. Locate the regions which intersect [f, t) 824 * and either trim, delete or split the existing regions. 825 * 826 * Returns the number of huge pages deleted from the reserve map. 827 * In the normal case, the return value is zero or more. In the 828 * case where a region must be split, a new region descriptor must 829 * be allocated. If the allocation fails, -ENOMEM will be returned. 830 * NOTE: If the parameter t == LONG_MAX, then we will never split 831 * a region and possibly return -ENOMEM. Callers specifying 832 * t == LONG_MAX do not need to check for -ENOMEM error. 833 */ 834 static long region_del(struct resv_map *resv, long f, long t) 835 { 836 struct list_head *head = &resv->regions; 837 struct file_region *rg, *trg; 838 struct file_region *nrg = NULL; 839 long del = 0; 840 841 retry: 842 spin_lock(&resv->lock); 843 list_for_each_entry_safe(rg, trg, head, link) { 844 /* 845 * Skip regions before the range to be deleted. file_region 846 * ranges are normally of the form [from, to). However, there 847 * may be a "placeholder" entry in the map which is of the form 848 * (from, to) with from == to. Check for placeholder entries 849 * at the beginning of the range to be deleted. 850 */ 851 if (rg->to <= f && (rg->to != rg->from || rg->to != f)) 852 continue; 853 854 if (rg->from >= t) 855 break; 856 857 if (f > rg->from && t < rg->to) { /* Must split region */ 858 /* 859 * Check for an entry in the cache before dropping 860 * lock and attempting allocation. 861 */ 862 if (!nrg && 863 resv->region_cache_count > resv->adds_in_progress) { 864 nrg = list_first_entry(&resv->region_cache, 865 struct file_region, 866 link); 867 list_del(&nrg->link); 868 resv->region_cache_count--; 869 } 870 871 if (!nrg) { 872 spin_unlock(&resv->lock); 873 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL); 874 if (!nrg) 875 return -ENOMEM; 876 goto retry; 877 } 878 879 del += t - f; 880 hugetlb_cgroup_uncharge_file_region( 881 resv, rg, t - f, false); 882 883 /* New entry for end of split region */ 884 nrg->from = t; 885 nrg->to = rg->to; 886 887 copy_hugetlb_cgroup_uncharge_info(nrg, rg); 888 889 INIT_LIST_HEAD(&nrg->link); 890 891 /* Original entry is trimmed */ 892 rg->to = f; 893 894 list_add(&nrg->link, &rg->link); 895 nrg = NULL; 896 break; 897 } 898 899 if (f <= rg->from && t >= rg->to) { /* Remove entire region */ 900 del += rg->to - rg->from; 901 hugetlb_cgroup_uncharge_file_region(resv, rg, 902 rg->to - rg->from, true); 903 list_del(&rg->link); 904 kfree(rg); 905 continue; 906 } 907 908 if (f <= rg->from) { /* Trim beginning of region */ 909 hugetlb_cgroup_uncharge_file_region(resv, rg, 910 t - rg->from, false); 911 912 del += t - rg->from; 913 rg->from = t; 914 } else { /* Trim end of region */ 915 hugetlb_cgroup_uncharge_file_region(resv, rg, 916 rg->to - f, false); 917 918 del += rg->to - f; 919 rg->to = f; 920 } 921 } 922 923 spin_unlock(&resv->lock); 924 kfree(nrg); 925 return del; 926 } 927 928 /* 929 * A rare out of memory error was encountered which prevented removal of 930 * the reserve map region for a page. The huge page itself was free'ed 931 * and removed from the page cache. This routine will adjust the subpool 932 * usage count, and the global reserve count if needed. By incrementing 933 * these counts, the reserve map entry which could not be deleted will 934 * appear as a "reserved" entry instead of simply dangling with incorrect 935 * counts. 936 */ 937 void hugetlb_fix_reserve_counts(struct inode *inode) 938 { 939 struct hugepage_subpool *spool = subpool_inode(inode); 940 long rsv_adjust; 941 bool reserved = false; 942 943 rsv_adjust = hugepage_subpool_get_pages(spool, 1); 944 if (rsv_adjust > 0) { 945 struct hstate *h = hstate_inode(inode); 946 947 if (!hugetlb_acct_memory(h, 1)) 948 reserved = true; 949 } else if (!rsv_adjust) { 950 reserved = true; 951 } 952 953 if (!reserved) 954 pr_warn("hugetlb: Huge Page Reserved count may go negative.\n"); 955 } 956 957 /* 958 * Count and return the number of huge pages in the reserve map 959 * that intersect with the range [f, t). 960 */ 961 static long region_count(struct resv_map *resv, long f, long t) 962 { 963 struct list_head *head = &resv->regions; 964 struct file_region *rg; 965 long chg = 0; 966 967 spin_lock(&resv->lock); 968 /* Locate each segment we overlap with, and count that overlap. */ 969 list_for_each_entry(rg, head, link) { 970 long seg_from; 971 long seg_to; 972 973 if (rg->to <= f) 974 continue; 975 if (rg->from >= t) 976 break; 977 978 seg_from = max(rg->from, f); 979 seg_to = min(rg->to, t); 980 981 chg += seg_to - seg_from; 982 } 983 spin_unlock(&resv->lock); 984 985 return chg; 986 } 987 988 /* 989 * Convert the address within this vma to the page offset within 990 * the mapping, huge page units here. 991 */ 992 static pgoff_t vma_hugecache_offset(struct hstate *h, 993 struct vm_area_struct *vma, unsigned long address) 994 { 995 return ((address - vma->vm_start) >> huge_page_shift(h)) + 996 (vma->vm_pgoff >> huge_page_order(h)); 997 } 998 999 /** 1000 * vma_kernel_pagesize - Page size granularity for this VMA. 1001 * @vma: The user mapping. 1002 * 1003 * Folios in this VMA will be aligned to, and at least the size of the 1004 * number of bytes returned by this function. 1005 * 1006 * Return: The default size of the folios allocated when backing a VMA. 1007 */ 1008 unsigned long vma_kernel_pagesize(struct vm_area_struct *vma) 1009 { 1010 if (vma->vm_ops && vma->vm_ops->pagesize) 1011 return vma->vm_ops->pagesize(vma); 1012 return PAGE_SIZE; 1013 } 1014 EXPORT_SYMBOL_GPL(vma_kernel_pagesize); 1015 1016 /* 1017 * Return the page size being used by the MMU to back a VMA. In the majority 1018 * of cases, the page size used by the kernel matches the MMU size. On 1019 * architectures where it differs, an architecture-specific 'strong' 1020 * version of this symbol is required. 1021 */ 1022 __weak unsigned long vma_mmu_pagesize(struct vm_area_struct *vma) 1023 { 1024 return vma_kernel_pagesize(vma); 1025 } 1026 1027 /* 1028 * Flags for MAP_PRIVATE reservations. These are stored in the bottom 1029 * bits of the reservation map pointer, which are always clear due to 1030 * alignment. 1031 */ 1032 #define HPAGE_RESV_OWNER (1UL << 0) 1033 #define HPAGE_RESV_UNMAPPED (1UL << 1) 1034 #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED) 1035 1036 /* 1037 * These helpers are used to track how many pages are reserved for 1038 * faults in a MAP_PRIVATE mapping. Only the process that called mmap() 1039 * is guaranteed to have their future faults succeed. 1040 * 1041 * With the exception of hugetlb_dup_vma_private() which is called at fork(), 1042 * the reserve counters are updated with the hugetlb_lock held. It is safe 1043 * to reset the VMA at fork() time as it is not in use yet and there is no 1044 * chance of the global counters getting corrupted as a result of the values. 1045 * 1046 * The private mapping reservation is represented in a subtly different 1047 * manner to a shared mapping. A shared mapping has a region map associated 1048 * with the underlying file, this region map represents the backing file 1049 * pages which have ever had a reservation assigned which this persists even 1050 * after the page is instantiated. A private mapping has a region map 1051 * associated with the original mmap which is attached to all VMAs which 1052 * reference it, this region map represents those offsets which have consumed 1053 * reservation ie. where pages have been instantiated. 1054 */ 1055 static unsigned long get_vma_private_data(struct vm_area_struct *vma) 1056 { 1057 return (unsigned long)vma->vm_private_data; 1058 } 1059 1060 static void set_vma_private_data(struct vm_area_struct *vma, 1061 unsigned long value) 1062 { 1063 vma->vm_private_data = (void *)value; 1064 } 1065 1066 static void 1067 resv_map_set_hugetlb_cgroup_uncharge_info(struct resv_map *resv_map, 1068 struct hugetlb_cgroup *h_cg, 1069 struct hstate *h) 1070 { 1071 #ifdef CONFIG_CGROUP_HUGETLB 1072 if (!h_cg || !h) { 1073 resv_map->reservation_counter = NULL; 1074 resv_map->pages_per_hpage = 0; 1075 resv_map->css = NULL; 1076 } else { 1077 resv_map->reservation_counter = 1078 &h_cg->rsvd_hugepage[hstate_index(h)]; 1079 resv_map->pages_per_hpage = pages_per_huge_page(h); 1080 resv_map->css = &h_cg->css; 1081 } 1082 #endif 1083 } 1084 1085 struct resv_map *resv_map_alloc(void) 1086 { 1087 struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL); 1088 struct file_region *rg = kmalloc(sizeof(*rg), GFP_KERNEL); 1089 1090 if (!resv_map || !rg) { 1091 kfree(resv_map); 1092 kfree(rg); 1093 return NULL; 1094 } 1095 1096 kref_init(&resv_map->refs); 1097 spin_lock_init(&resv_map->lock); 1098 INIT_LIST_HEAD(&resv_map->regions); 1099 init_rwsem(&resv_map->rw_sema); 1100 1101 resv_map->adds_in_progress = 0; 1102 /* 1103 * Initialize these to 0. On shared mappings, 0's here indicate these 1104 * fields don't do cgroup accounting. On private mappings, these will be 1105 * re-initialized to the proper values, to indicate that hugetlb cgroup 1106 * reservations are to be un-charged from here. 1107 */ 1108 resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, NULL, NULL); 1109 1110 INIT_LIST_HEAD(&resv_map->region_cache); 1111 list_add(&rg->link, &resv_map->region_cache); 1112 resv_map->region_cache_count = 1; 1113 1114 return resv_map; 1115 } 1116 1117 void resv_map_release(struct kref *ref) 1118 { 1119 struct resv_map *resv_map = container_of(ref, struct resv_map, refs); 1120 struct list_head *head = &resv_map->region_cache; 1121 struct file_region *rg, *trg; 1122 1123 /* Clear out any active regions before we release the map. */ 1124 region_del(resv_map, 0, LONG_MAX); 1125 1126 /* ... and any entries left in the cache */ 1127 list_for_each_entry_safe(rg, trg, head, link) { 1128 list_del(&rg->link); 1129 kfree(rg); 1130 } 1131 1132 VM_BUG_ON(resv_map->adds_in_progress); 1133 1134 kfree(resv_map); 1135 } 1136 1137 static inline struct resv_map *inode_resv_map(struct inode *inode) 1138 { 1139 /* 1140 * At inode evict time, i_mapping may not point to the original 1141 * address space within the inode. This original address space 1142 * contains the pointer to the resv_map. So, always use the 1143 * address space embedded within the inode. 1144 * The VERY common case is inode->mapping == &inode->i_data but, 1145 * this may not be true for device special inodes. 1146 */ 1147 return (struct resv_map *)(&inode->i_data)->i_private_data; 1148 } 1149 1150 static struct resv_map *vma_resv_map(struct vm_area_struct *vma) 1151 { 1152 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 1153 if (vma->vm_flags & VM_MAYSHARE) { 1154 struct address_space *mapping = vma->vm_file->f_mapping; 1155 struct inode *inode = mapping->host; 1156 1157 return inode_resv_map(inode); 1158 1159 } else { 1160 return (struct resv_map *)(get_vma_private_data(vma) & 1161 ~HPAGE_RESV_MASK); 1162 } 1163 } 1164 1165 static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map) 1166 { 1167 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 1168 VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma); 1169 1170 set_vma_private_data(vma, (unsigned long)map); 1171 } 1172 1173 static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags) 1174 { 1175 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 1176 VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma); 1177 1178 set_vma_private_data(vma, get_vma_private_data(vma) | flags); 1179 } 1180 1181 static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag) 1182 { 1183 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 1184 1185 return (get_vma_private_data(vma) & flag) != 0; 1186 } 1187 1188 bool __vma_private_lock(struct vm_area_struct *vma) 1189 { 1190 return !(vma->vm_flags & VM_MAYSHARE) && 1191 get_vma_private_data(vma) & ~HPAGE_RESV_MASK && 1192 is_vma_resv_set(vma, HPAGE_RESV_OWNER); 1193 } 1194 1195 void hugetlb_dup_vma_private(struct vm_area_struct *vma) 1196 { 1197 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); 1198 /* 1199 * Clear vm_private_data 1200 * - For shared mappings this is a per-vma semaphore that may be 1201 * allocated in a subsequent call to hugetlb_vm_op_open. 1202 * Before clearing, make sure pointer is not associated with vma 1203 * as this will leak the structure. This is the case when called 1204 * via clear_vma_resv_huge_pages() and hugetlb_vm_op_open has already 1205 * been called to allocate a new structure. 1206 * - For MAP_PRIVATE mappings, this is the reserve map which does 1207 * not apply to children. Faults generated by the children are 1208 * not guaranteed to succeed, even if read-only. 1209 */ 1210 if (vma->vm_flags & VM_MAYSHARE) { 1211 struct hugetlb_vma_lock *vma_lock = vma->vm_private_data; 1212 1213 if (vma_lock && vma_lock->vma != vma) 1214 vma->vm_private_data = NULL; 1215 } else 1216 vma->vm_private_data = NULL; 1217 } 1218 1219 /* 1220 * Reset and decrement one ref on hugepage private reservation. 1221 * Called with mm->mmap_lock writer semaphore held. 1222 * This function should be only used by move_vma() and operate on 1223 * same sized vma. It should never come here with last ref on the 1224 * reservation. 1225 */ 1226 void clear_vma_resv_huge_pages(struct vm_area_struct *vma) 1227 { 1228 /* 1229 * Clear the old hugetlb private page reservation. 1230 * It has already been transferred to new_vma. 1231 * 1232 * During a mremap() operation of a hugetlb vma we call move_vma() 1233 * which copies vma into new_vma and unmaps vma. After the copy 1234 * operation both new_vma and vma share a reference to the resv_map 1235 * struct, and at that point vma is about to be unmapped. We don't 1236 * want to return the reservation to the pool at unmap of vma because 1237 * the reservation still lives on in new_vma, so simply decrement the 1238 * ref here and remove the resv_map reference from this vma. 1239 */ 1240 struct resv_map *reservations = vma_resv_map(vma); 1241 1242 if (reservations && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { 1243 resv_map_put_hugetlb_cgroup_uncharge_info(reservations); 1244 kref_put(&reservations->refs, resv_map_release); 1245 } 1246 1247 hugetlb_dup_vma_private(vma); 1248 } 1249 1250 static void enqueue_hugetlb_folio(struct hstate *h, struct folio *folio) 1251 { 1252 int nid = folio_nid(folio); 1253 1254 lockdep_assert_held(&hugetlb_lock); 1255 VM_BUG_ON_FOLIO(folio_ref_count(folio), folio); 1256 1257 list_move(&folio->lru, &h->hugepage_freelists[nid]); 1258 h->free_huge_pages++; 1259 h->free_huge_pages_node[nid]++; 1260 folio_set_hugetlb_freed(folio); 1261 } 1262 1263 static struct folio *dequeue_hugetlb_folio_node_exact(struct hstate *h, 1264 int nid) 1265 { 1266 struct folio *folio; 1267 bool pin = !!(current->flags & PF_MEMALLOC_PIN); 1268 1269 lockdep_assert_held(&hugetlb_lock); 1270 list_for_each_entry(folio, &h->hugepage_freelists[nid], lru) { 1271 if (pin && !folio_is_longterm_pinnable(folio)) 1272 continue; 1273 1274 if (folio_test_hwpoison(folio)) 1275 continue; 1276 1277 if (is_migrate_isolate_page(&folio->page)) 1278 continue; 1279 1280 list_move(&folio->lru, &h->hugepage_activelist); 1281 folio_ref_unfreeze(folio, 1); 1282 folio_clear_hugetlb_freed(folio); 1283 h->free_huge_pages--; 1284 h->free_huge_pages_node[nid]--; 1285 return folio; 1286 } 1287 1288 return NULL; 1289 } 1290 1291 static struct folio *dequeue_hugetlb_folio_nodemask(struct hstate *h, gfp_t gfp_mask, 1292 int nid, nodemask_t *nmask) 1293 { 1294 unsigned int cpuset_mems_cookie; 1295 struct zonelist *zonelist; 1296 struct zone *zone; 1297 struct zoneref *z; 1298 int node = NUMA_NO_NODE; 1299 1300 /* 'nid' should not be NUMA_NO_NODE. Try to catch any misuse of it and rectifiy. */ 1301 if (nid == NUMA_NO_NODE) 1302 nid = numa_node_id(); 1303 1304 zonelist = node_zonelist(nid, gfp_mask); 1305 1306 retry_cpuset: 1307 cpuset_mems_cookie = read_mems_allowed_begin(); 1308 for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), nmask) { 1309 struct folio *folio; 1310 1311 if (!cpuset_zone_allowed(zone, gfp_mask)) 1312 continue; 1313 /* 1314 * no need to ask again on the same node. Pool is node rather than 1315 * zone aware 1316 */ 1317 if (zone_to_nid(zone) == node) 1318 continue; 1319 node = zone_to_nid(zone); 1320 1321 folio = dequeue_hugetlb_folio_node_exact(h, node); 1322 if (folio) 1323 return folio; 1324 } 1325 if (unlikely(read_mems_allowed_retry(cpuset_mems_cookie))) 1326 goto retry_cpuset; 1327 1328 return NULL; 1329 } 1330 1331 static unsigned long available_huge_pages(struct hstate *h) 1332 { 1333 return h->free_huge_pages - h->resv_huge_pages; 1334 } 1335 1336 static struct folio *dequeue_hugetlb_folio_vma(struct hstate *h, 1337 struct vm_area_struct *vma, 1338 unsigned long address, long gbl_chg) 1339 { 1340 struct folio *folio = NULL; 1341 struct mempolicy *mpol; 1342 gfp_t gfp_mask; 1343 nodemask_t *nodemask; 1344 int nid; 1345 1346 /* 1347 * gbl_chg==1 means the allocation requires a new page that was not 1348 * reserved before. Making sure there's at least one free page. 1349 */ 1350 if (gbl_chg && !available_huge_pages(h)) 1351 goto err; 1352 1353 gfp_mask = htlb_alloc_mask(h); 1354 nid = huge_node(vma, address, gfp_mask, &mpol, &nodemask); 1355 1356 if (mpol_is_preferred_many(mpol)) { 1357 folio = dequeue_hugetlb_folio_nodemask(h, gfp_mask, 1358 nid, nodemask); 1359 1360 /* Fallback to all nodes if page==NULL */ 1361 nodemask = NULL; 1362 } 1363 1364 if (!folio) 1365 folio = dequeue_hugetlb_folio_nodemask(h, gfp_mask, 1366 nid, nodemask); 1367 1368 mpol_cond_put(mpol); 1369 return folio; 1370 1371 err: 1372 return NULL; 1373 } 1374 1375 /* 1376 * common helper functions for hstate_next_node_to_{alloc|free}. 1377 * We may have allocated or freed a huge page based on a different 1378 * nodes_allowed previously, so h->next_node_to_{alloc|free} might 1379 * be outside of *nodes_allowed. Ensure that we use an allowed 1380 * node for alloc or free. 1381 */ 1382 static int next_node_allowed(int nid, nodemask_t *nodes_allowed) 1383 { 1384 nid = next_node_in(nid, *nodes_allowed); 1385 VM_BUG_ON(nid >= MAX_NUMNODES); 1386 1387 return nid; 1388 } 1389 1390 static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed) 1391 { 1392 if (!node_isset(nid, *nodes_allowed)) 1393 nid = next_node_allowed(nid, nodes_allowed); 1394 return nid; 1395 } 1396 1397 /* 1398 * returns the previously saved node ["this node"] from which to 1399 * allocate a persistent huge page for the pool and advance the 1400 * next node from which to allocate, handling wrap at end of node 1401 * mask. 1402 */ 1403 static int hstate_next_node_to_alloc(int *next_node, 1404 nodemask_t *nodes_allowed) 1405 { 1406 int nid; 1407 1408 VM_BUG_ON(!nodes_allowed); 1409 1410 nid = get_valid_node_allowed(*next_node, nodes_allowed); 1411 *next_node = next_node_allowed(nid, nodes_allowed); 1412 1413 return nid; 1414 } 1415 1416 /* 1417 * helper for remove_pool_hugetlb_folio() - return the previously saved 1418 * node ["this node"] from which to free a huge page. Advance the 1419 * next node id whether or not we find a free huge page to free so 1420 * that the next attempt to free addresses the next node. 1421 */ 1422 static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed) 1423 { 1424 int nid; 1425 1426 VM_BUG_ON(!nodes_allowed); 1427 1428 nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed); 1429 h->next_nid_to_free = next_node_allowed(nid, nodes_allowed); 1430 1431 return nid; 1432 } 1433 1434 #define for_each_node_mask_to_alloc(next_node, nr_nodes, node, mask) \ 1435 for (nr_nodes = nodes_weight(*mask); \ 1436 nr_nodes > 0 && \ 1437 ((node = hstate_next_node_to_alloc(next_node, mask)) || 1); \ 1438 nr_nodes--) 1439 1440 #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \ 1441 for (nr_nodes = nodes_weight(*mask); \ 1442 nr_nodes > 0 && \ 1443 ((node = hstate_next_node_to_free(hs, mask)) || 1); \ 1444 nr_nodes--) 1445 1446 #ifdef CONFIG_ARCH_HAS_GIGANTIC_PAGE 1447 #ifdef CONFIG_CONTIG_ALLOC 1448 static struct folio *alloc_gigantic_folio(struct hstate *h, gfp_t gfp_mask, 1449 int nid, nodemask_t *nodemask) 1450 { 1451 struct folio *folio; 1452 int order = huge_page_order(h); 1453 bool retried = false; 1454 1455 if (nid == NUMA_NO_NODE) 1456 nid = numa_mem_id(); 1457 retry: 1458 folio = NULL; 1459 #ifdef CONFIG_CMA 1460 { 1461 int node; 1462 1463 if (hugetlb_cma[nid]) 1464 folio = cma_alloc_folio(hugetlb_cma[nid], order, gfp_mask); 1465 1466 if (!folio && !(gfp_mask & __GFP_THISNODE)) { 1467 for_each_node_mask(node, *nodemask) { 1468 if (node == nid || !hugetlb_cma[node]) 1469 continue; 1470 1471 folio = cma_alloc_folio(hugetlb_cma[node], order, gfp_mask); 1472 if (folio) 1473 break; 1474 } 1475 } 1476 } 1477 #endif 1478 if (!folio) { 1479 folio = folio_alloc_gigantic(order, gfp_mask, nid, nodemask); 1480 if (!folio) 1481 return NULL; 1482 } 1483 1484 if (folio_ref_freeze(folio, 1)) 1485 return folio; 1486 1487 pr_warn("HugeTLB: unexpected refcount on PFN %lu\n", folio_pfn(folio)); 1488 hugetlb_free_folio(folio); 1489 if (!retried) { 1490 retried = true; 1491 goto retry; 1492 } 1493 return NULL; 1494 } 1495 1496 #else /* !CONFIG_CONTIG_ALLOC */ 1497 static struct folio *alloc_gigantic_folio(struct hstate *h, gfp_t gfp_mask, 1498 int nid, nodemask_t *nodemask) 1499 { 1500 return NULL; 1501 } 1502 #endif /* CONFIG_CONTIG_ALLOC */ 1503 1504 #else /* !CONFIG_ARCH_HAS_GIGANTIC_PAGE */ 1505 static struct folio *alloc_gigantic_folio(struct hstate *h, gfp_t gfp_mask, 1506 int nid, nodemask_t *nodemask) 1507 { 1508 return NULL; 1509 } 1510 #endif 1511 1512 /* 1513 * Remove hugetlb folio from lists. 1514 * If vmemmap exists for the folio, clear the hugetlb flag so that the 1515 * folio appears as just a compound page. Otherwise, wait until after 1516 * allocating vmemmap to clear the flag. 1517 * 1518 * Must be called with hugetlb lock held. 1519 */ 1520 static void remove_hugetlb_folio(struct hstate *h, struct folio *folio, 1521 bool adjust_surplus) 1522 { 1523 int nid = folio_nid(folio); 1524 1525 VM_BUG_ON_FOLIO(hugetlb_cgroup_from_folio(folio), folio); 1526 VM_BUG_ON_FOLIO(hugetlb_cgroup_from_folio_rsvd(folio), folio); 1527 1528 lockdep_assert_held(&hugetlb_lock); 1529 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) 1530 return; 1531 1532 list_del(&folio->lru); 1533 1534 if (folio_test_hugetlb_freed(folio)) { 1535 folio_clear_hugetlb_freed(folio); 1536 h->free_huge_pages--; 1537 h->free_huge_pages_node[nid]--; 1538 } 1539 if (adjust_surplus) { 1540 h->surplus_huge_pages--; 1541 h->surplus_huge_pages_node[nid]--; 1542 } 1543 1544 /* 1545 * We can only clear the hugetlb flag after allocating vmemmap 1546 * pages. Otherwise, someone (memory error handling) may try to write 1547 * to tail struct pages. 1548 */ 1549 if (!folio_test_hugetlb_vmemmap_optimized(folio)) 1550 __folio_clear_hugetlb(folio); 1551 1552 h->nr_huge_pages--; 1553 h->nr_huge_pages_node[nid]--; 1554 } 1555 1556 static void add_hugetlb_folio(struct hstate *h, struct folio *folio, 1557 bool adjust_surplus) 1558 { 1559 int nid = folio_nid(folio); 1560 1561 VM_BUG_ON_FOLIO(!folio_test_hugetlb_vmemmap_optimized(folio), folio); 1562 1563 lockdep_assert_held(&hugetlb_lock); 1564 1565 INIT_LIST_HEAD(&folio->lru); 1566 h->nr_huge_pages++; 1567 h->nr_huge_pages_node[nid]++; 1568 1569 if (adjust_surplus) { 1570 h->surplus_huge_pages++; 1571 h->surplus_huge_pages_node[nid]++; 1572 } 1573 1574 __folio_set_hugetlb(folio); 1575 folio_change_private(folio, NULL); 1576 /* 1577 * We have to set hugetlb_vmemmap_optimized again as above 1578 * folio_change_private(folio, NULL) cleared it. 1579 */ 1580 folio_set_hugetlb_vmemmap_optimized(folio); 1581 1582 arch_clear_hugetlb_flags(folio); 1583 enqueue_hugetlb_folio(h, folio); 1584 } 1585 1586 static void __update_and_free_hugetlb_folio(struct hstate *h, 1587 struct folio *folio) 1588 { 1589 bool clear_flag = folio_test_hugetlb_vmemmap_optimized(folio); 1590 1591 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) 1592 return; 1593 1594 /* 1595 * If we don't know which subpages are hwpoisoned, we can't free 1596 * the hugepage, so it's leaked intentionally. 1597 */ 1598 if (folio_test_hugetlb_raw_hwp_unreliable(folio)) 1599 return; 1600 1601 /* 1602 * If folio is not vmemmap optimized (!clear_flag), then the folio 1603 * is no longer identified as a hugetlb page. hugetlb_vmemmap_restore_folio 1604 * can only be passed hugetlb pages and will BUG otherwise. 1605 */ 1606 if (clear_flag && hugetlb_vmemmap_restore_folio(h, folio)) { 1607 spin_lock_irq(&hugetlb_lock); 1608 /* 1609 * If we cannot allocate vmemmap pages, just refuse to free the 1610 * page and put the page back on the hugetlb free list and treat 1611 * as a surplus page. 1612 */ 1613 add_hugetlb_folio(h, folio, true); 1614 spin_unlock_irq(&hugetlb_lock); 1615 return; 1616 } 1617 1618 /* 1619 * If vmemmap pages were allocated above, then we need to clear the 1620 * hugetlb flag under the hugetlb lock. 1621 */ 1622 if (folio_test_hugetlb(folio)) { 1623 spin_lock_irq(&hugetlb_lock); 1624 __folio_clear_hugetlb(folio); 1625 spin_unlock_irq(&hugetlb_lock); 1626 } 1627 1628 /* 1629 * Move PageHWPoison flag from head page to the raw error pages, 1630 * which makes any healthy subpages reusable. 1631 */ 1632 if (unlikely(folio_test_hwpoison(folio))) 1633 folio_clear_hugetlb_hwpoison(folio); 1634 1635 folio_ref_unfreeze(folio, 1); 1636 1637 INIT_LIST_HEAD(&folio->_deferred_list); 1638 hugetlb_free_folio(folio); 1639 } 1640 1641 /* 1642 * As update_and_free_hugetlb_folio() can be called under any context, so we cannot 1643 * use GFP_KERNEL to allocate vmemmap pages. However, we can defer the 1644 * actual freeing in a workqueue to prevent from using GFP_ATOMIC to allocate 1645 * the vmemmap pages. 1646 * 1647 * free_hpage_workfn() locklessly retrieves the linked list of pages to be 1648 * freed and frees them one-by-one. As the page->mapping pointer is going 1649 * to be cleared in free_hpage_workfn() anyway, it is reused as the llist_node 1650 * structure of a lockless linked list of huge pages to be freed. 1651 */ 1652 static LLIST_HEAD(hpage_freelist); 1653 1654 static void free_hpage_workfn(struct work_struct *work) 1655 { 1656 struct llist_node *node; 1657 1658 node = llist_del_all(&hpage_freelist); 1659 1660 while (node) { 1661 struct folio *folio; 1662 struct hstate *h; 1663 1664 folio = container_of((struct address_space **)node, 1665 struct folio, mapping); 1666 node = node->next; 1667 folio->mapping = NULL; 1668 /* 1669 * The VM_BUG_ON_FOLIO(!folio_test_hugetlb(folio), folio) in 1670 * folio_hstate() is going to trigger because a previous call to 1671 * remove_hugetlb_folio() will clear the hugetlb bit, so do 1672 * not use folio_hstate() directly. 1673 */ 1674 h = size_to_hstate(folio_size(folio)); 1675 1676 __update_and_free_hugetlb_folio(h, folio); 1677 1678 cond_resched(); 1679 } 1680 } 1681 static DECLARE_WORK(free_hpage_work, free_hpage_workfn); 1682 1683 static inline void flush_free_hpage_work(struct hstate *h) 1684 { 1685 if (hugetlb_vmemmap_optimizable(h)) 1686 flush_work(&free_hpage_work); 1687 } 1688 1689 static void update_and_free_hugetlb_folio(struct hstate *h, struct folio *folio, 1690 bool atomic) 1691 { 1692 if (!folio_test_hugetlb_vmemmap_optimized(folio) || !atomic) { 1693 __update_and_free_hugetlb_folio(h, folio); 1694 return; 1695 } 1696 1697 /* 1698 * Defer freeing to avoid using GFP_ATOMIC to allocate vmemmap pages. 1699 * 1700 * Only call schedule_work() if hpage_freelist is previously 1701 * empty. Otherwise, schedule_work() had been called but the workfn 1702 * hasn't retrieved the list yet. 1703 */ 1704 if (llist_add((struct llist_node *)&folio->mapping, &hpage_freelist)) 1705 schedule_work(&free_hpage_work); 1706 } 1707 1708 static void bulk_vmemmap_restore_error(struct hstate *h, 1709 struct list_head *folio_list, 1710 struct list_head *non_hvo_folios) 1711 { 1712 struct folio *folio, *t_folio; 1713 1714 if (!list_empty(non_hvo_folios)) { 1715 /* 1716 * Free any restored hugetlb pages so that restore of the 1717 * entire list can be retried. 1718 * The idea is that in the common case of ENOMEM errors freeing 1719 * hugetlb pages with vmemmap we will free up memory so that we 1720 * can allocate vmemmap for more hugetlb pages. 1721 */ 1722 list_for_each_entry_safe(folio, t_folio, non_hvo_folios, lru) { 1723 list_del(&folio->lru); 1724 spin_lock_irq(&hugetlb_lock); 1725 __folio_clear_hugetlb(folio); 1726 spin_unlock_irq(&hugetlb_lock); 1727 update_and_free_hugetlb_folio(h, folio, false); 1728 cond_resched(); 1729 } 1730 } else { 1731 /* 1732 * In the case where there are no folios which can be 1733 * immediately freed, we loop through the list trying to restore 1734 * vmemmap individually in the hope that someone elsewhere may 1735 * have done something to cause success (such as freeing some 1736 * memory). If unable to restore a hugetlb page, the hugetlb 1737 * page is made a surplus page and removed from the list. 1738 * If are able to restore vmemmap and free one hugetlb page, we 1739 * quit processing the list to retry the bulk operation. 1740 */ 1741 list_for_each_entry_safe(folio, t_folio, folio_list, lru) 1742 if (hugetlb_vmemmap_restore_folio(h, folio)) { 1743 list_del(&folio->lru); 1744 spin_lock_irq(&hugetlb_lock); 1745 add_hugetlb_folio(h, folio, true); 1746 spin_unlock_irq(&hugetlb_lock); 1747 } else { 1748 list_del(&folio->lru); 1749 spin_lock_irq(&hugetlb_lock); 1750 __folio_clear_hugetlb(folio); 1751 spin_unlock_irq(&hugetlb_lock); 1752 update_and_free_hugetlb_folio(h, folio, false); 1753 cond_resched(); 1754 break; 1755 } 1756 } 1757 } 1758 1759 static void update_and_free_pages_bulk(struct hstate *h, 1760 struct list_head *folio_list) 1761 { 1762 long ret; 1763 struct folio *folio, *t_folio; 1764 LIST_HEAD(non_hvo_folios); 1765 1766 /* 1767 * First allocate required vmemmmap (if necessary) for all folios. 1768 * Carefully handle errors and free up any available hugetlb pages 1769 * in an effort to make forward progress. 1770 */ 1771 retry: 1772 ret = hugetlb_vmemmap_restore_folios(h, folio_list, &non_hvo_folios); 1773 if (ret < 0) { 1774 bulk_vmemmap_restore_error(h, folio_list, &non_hvo_folios); 1775 goto retry; 1776 } 1777 1778 /* 1779 * At this point, list should be empty, ret should be >= 0 and there 1780 * should only be pages on the non_hvo_folios list. 1781 * Do note that the non_hvo_folios list could be empty. 1782 * Without HVO enabled, ret will be 0 and there is no need to call 1783 * __folio_clear_hugetlb as this was done previously. 1784 */ 1785 VM_WARN_ON(!list_empty(folio_list)); 1786 VM_WARN_ON(ret < 0); 1787 if (!list_empty(&non_hvo_folios) && ret) { 1788 spin_lock_irq(&hugetlb_lock); 1789 list_for_each_entry(folio, &non_hvo_folios, lru) 1790 __folio_clear_hugetlb(folio); 1791 spin_unlock_irq(&hugetlb_lock); 1792 } 1793 1794 list_for_each_entry_safe(folio, t_folio, &non_hvo_folios, lru) { 1795 update_and_free_hugetlb_folio(h, folio, false); 1796 cond_resched(); 1797 } 1798 } 1799 1800 struct hstate *size_to_hstate(unsigned long size) 1801 { 1802 struct hstate *h; 1803 1804 for_each_hstate(h) { 1805 if (huge_page_size(h) == size) 1806 return h; 1807 } 1808 return NULL; 1809 } 1810 1811 void free_huge_folio(struct folio *folio) 1812 { 1813 /* 1814 * Can't pass hstate in here because it is called from the 1815 * generic mm code. 1816 */ 1817 struct hstate *h = folio_hstate(folio); 1818 int nid = folio_nid(folio); 1819 struct hugepage_subpool *spool = hugetlb_folio_subpool(folio); 1820 bool restore_reserve; 1821 unsigned long flags; 1822 1823 VM_BUG_ON_FOLIO(folio_ref_count(folio), folio); 1824 VM_BUG_ON_FOLIO(folio_mapcount(folio), folio); 1825 1826 hugetlb_set_folio_subpool(folio, NULL); 1827 if (folio_test_anon(folio)) 1828 __ClearPageAnonExclusive(&folio->page); 1829 folio->mapping = NULL; 1830 restore_reserve = folio_test_hugetlb_restore_reserve(folio); 1831 folio_clear_hugetlb_restore_reserve(folio); 1832 1833 /* 1834 * If HPageRestoreReserve was set on page, page allocation consumed a 1835 * reservation. If the page was associated with a subpool, there 1836 * would have been a page reserved in the subpool before allocation 1837 * via hugepage_subpool_get_pages(). Since we are 'restoring' the 1838 * reservation, do not call hugepage_subpool_put_pages() as this will 1839 * remove the reserved page from the subpool. 1840 */ 1841 if (!restore_reserve) { 1842 /* 1843 * A return code of zero implies that the subpool will be 1844 * under its minimum size if the reservation is not restored 1845 * after page is free. Therefore, force restore_reserve 1846 * operation. 1847 */ 1848 if (hugepage_subpool_put_pages(spool, 1) == 0) 1849 restore_reserve = true; 1850 } 1851 1852 spin_lock_irqsave(&hugetlb_lock, flags); 1853 folio_clear_hugetlb_migratable(folio); 1854 hugetlb_cgroup_uncharge_folio(hstate_index(h), 1855 pages_per_huge_page(h), folio); 1856 hugetlb_cgroup_uncharge_folio_rsvd(hstate_index(h), 1857 pages_per_huge_page(h), folio); 1858 lruvec_stat_mod_folio(folio, NR_HUGETLB, -pages_per_huge_page(h)); 1859 mem_cgroup_uncharge(folio); 1860 if (restore_reserve) 1861 h->resv_huge_pages++; 1862 1863 if (folio_test_hugetlb_temporary(folio)) { 1864 remove_hugetlb_folio(h, folio, false); 1865 spin_unlock_irqrestore(&hugetlb_lock, flags); 1866 update_and_free_hugetlb_folio(h, folio, true); 1867 } else if (h->surplus_huge_pages_node[nid]) { 1868 /* remove the page from active list */ 1869 remove_hugetlb_folio(h, folio, true); 1870 spin_unlock_irqrestore(&hugetlb_lock, flags); 1871 update_and_free_hugetlb_folio(h, folio, true); 1872 } else { 1873 arch_clear_hugetlb_flags(folio); 1874 enqueue_hugetlb_folio(h, folio); 1875 spin_unlock_irqrestore(&hugetlb_lock, flags); 1876 } 1877 } 1878 1879 /* 1880 * Must be called with the hugetlb lock held 1881 */ 1882 static void __prep_account_new_huge_page(struct hstate *h, int nid) 1883 { 1884 lockdep_assert_held(&hugetlb_lock); 1885 h->nr_huge_pages++; 1886 h->nr_huge_pages_node[nid]++; 1887 } 1888 1889 static void init_new_hugetlb_folio(struct hstate *h, struct folio *folio) 1890 { 1891 __folio_set_hugetlb(folio); 1892 INIT_LIST_HEAD(&folio->lru); 1893 hugetlb_set_folio_subpool(folio, NULL); 1894 set_hugetlb_cgroup(folio, NULL); 1895 set_hugetlb_cgroup_rsvd(folio, NULL); 1896 } 1897 1898 static void __prep_new_hugetlb_folio(struct hstate *h, struct folio *folio) 1899 { 1900 init_new_hugetlb_folio(h, folio); 1901 hugetlb_vmemmap_optimize_folio(h, folio); 1902 } 1903 1904 static void prep_new_hugetlb_folio(struct hstate *h, struct folio *folio, int nid) 1905 { 1906 __prep_new_hugetlb_folio(h, folio); 1907 spin_lock_irq(&hugetlb_lock); 1908 __prep_account_new_huge_page(h, nid); 1909 spin_unlock_irq(&hugetlb_lock); 1910 } 1911 1912 /* 1913 * Find and lock address space (mapping) in write mode. 1914 * 1915 * Upon entry, the folio is locked which means that folio_mapping() is 1916 * stable. Due to locking order, we can only trylock_write. If we can 1917 * not get the lock, simply return NULL to caller. 1918 */ 1919 struct address_space *hugetlb_folio_mapping_lock_write(struct folio *folio) 1920 { 1921 struct address_space *mapping = folio_mapping(folio); 1922 1923 if (!mapping) 1924 return mapping; 1925 1926 if (i_mmap_trylock_write(mapping)) 1927 return mapping; 1928 1929 return NULL; 1930 } 1931 1932 static struct folio *alloc_buddy_hugetlb_folio(struct hstate *h, 1933 gfp_t gfp_mask, int nid, nodemask_t *nmask, 1934 nodemask_t *node_alloc_noretry) 1935 { 1936 int order = huge_page_order(h); 1937 struct folio *folio; 1938 bool alloc_try_hard = true; 1939 bool retry = true; 1940 1941 /* 1942 * By default we always try hard to allocate the folio with 1943 * __GFP_RETRY_MAYFAIL flag. However, if we are allocating folios in 1944 * a loop (to adjust global huge page counts) and previous allocation 1945 * failed, do not continue to try hard on the same node. Use the 1946 * node_alloc_noretry bitmap to manage this state information. 1947 */ 1948 if (node_alloc_noretry && node_isset(nid, *node_alloc_noretry)) 1949 alloc_try_hard = false; 1950 if (alloc_try_hard) 1951 gfp_mask |= __GFP_RETRY_MAYFAIL; 1952 if (nid == NUMA_NO_NODE) 1953 nid = numa_mem_id(); 1954 retry: 1955 folio = __folio_alloc(gfp_mask, order, nid, nmask); 1956 /* Ensure hugetlb folio won't have large_rmappable flag set. */ 1957 if (folio) 1958 folio_clear_large_rmappable(folio); 1959 1960 if (folio && !folio_ref_freeze(folio, 1)) { 1961 folio_put(folio); 1962 if (retry) { /* retry once */ 1963 retry = false; 1964 goto retry; 1965 } 1966 /* WOW! twice in a row. */ 1967 pr_warn("HugeTLB unexpected inflated folio ref count\n"); 1968 folio = NULL; 1969 } 1970 1971 /* 1972 * If we did not specify __GFP_RETRY_MAYFAIL, but still got a 1973 * folio this indicates an overall state change. Clear bit so 1974 * that we resume normal 'try hard' allocations. 1975 */ 1976 if (node_alloc_noretry && folio && !alloc_try_hard) 1977 node_clear(nid, *node_alloc_noretry); 1978 1979 /* 1980 * If we tried hard to get a folio but failed, set bit so that 1981 * subsequent attempts will not try as hard until there is an 1982 * overall state change. 1983 */ 1984 if (node_alloc_noretry && !folio && alloc_try_hard) 1985 node_set(nid, *node_alloc_noretry); 1986 1987 if (!folio) { 1988 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); 1989 return NULL; 1990 } 1991 1992 __count_vm_event(HTLB_BUDDY_PGALLOC); 1993 return folio; 1994 } 1995 1996 static struct folio *only_alloc_fresh_hugetlb_folio(struct hstate *h, 1997 gfp_t gfp_mask, int nid, nodemask_t *nmask, 1998 nodemask_t *node_alloc_noretry) 1999 { 2000 struct folio *folio; 2001 2002 if (hstate_is_gigantic(h)) 2003 folio = alloc_gigantic_folio(h, gfp_mask, nid, nmask); 2004 else 2005 folio = alloc_buddy_hugetlb_folio(h, gfp_mask, nid, nmask, node_alloc_noretry); 2006 if (folio) 2007 init_new_hugetlb_folio(h, folio); 2008 return folio; 2009 } 2010 2011 /* 2012 * Common helper to allocate a fresh hugetlb page. All specific allocators 2013 * should use this function to get new hugetlb pages 2014 * 2015 * Note that returned page is 'frozen': ref count of head page and all tail 2016 * pages is zero. 2017 */ 2018 static struct folio *alloc_fresh_hugetlb_folio(struct hstate *h, 2019 gfp_t gfp_mask, int nid, nodemask_t *nmask) 2020 { 2021 struct folio *folio; 2022 2023 if (hstate_is_gigantic(h)) 2024 folio = alloc_gigantic_folio(h, gfp_mask, nid, nmask); 2025 else 2026 folio = alloc_buddy_hugetlb_folio(h, gfp_mask, nid, nmask, NULL); 2027 if (!folio) 2028 return NULL; 2029 2030 prep_new_hugetlb_folio(h, folio, folio_nid(folio)); 2031 return folio; 2032 } 2033 2034 static void prep_and_add_allocated_folios(struct hstate *h, 2035 struct list_head *folio_list) 2036 { 2037 unsigned long flags; 2038 struct folio *folio, *tmp_f; 2039 2040 /* Send list for bulk vmemmap optimization processing */ 2041 hugetlb_vmemmap_optimize_folios(h, folio_list); 2042 2043 /* Add all new pool pages to free lists in one lock cycle */ 2044 spin_lock_irqsave(&hugetlb_lock, flags); 2045 list_for_each_entry_safe(folio, tmp_f, folio_list, lru) { 2046 __prep_account_new_huge_page(h, folio_nid(folio)); 2047 enqueue_hugetlb_folio(h, folio); 2048 } 2049 spin_unlock_irqrestore(&hugetlb_lock, flags); 2050 } 2051 2052 /* 2053 * Allocates a fresh hugetlb page in a node interleaved manner. The page 2054 * will later be added to the appropriate hugetlb pool. 2055 */ 2056 static struct folio *alloc_pool_huge_folio(struct hstate *h, 2057 nodemask_t *nodes_allowed, 2058 nodemask_t *node_alloc_noretry, 2059 int *next_node) 2060 { 2061 gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE; 2062 int nr_nodes, node; 2063 2064 for_each_node_mask_to_alloc(next_node, nr_nodes, node, nodes_allowed) { 2065 struct folio *folio; 2066 2067 folio = only_alloc_fresh_hugetlb_folio(h, gfp_mask, node, 2068 nodes_allowed, node_alloc_noretry); 2069 if (folio) 2070 return folio; 2071 } 2072 2073 return NULL; 2074 } 2075 2076 /* 2077 * Remove huge page from pool from next node to free. Attempt to keep 2078 * persistent huge pages more or less balanced over allowed nodes. 2079 * This routine only 'removes' the hugetlb page. The caller must make 2080 * an additional call to free the page to low level allocators. 2081 * Called with hugetlb_lock locked. 2082 */ 2083 static struct folio *remove_pool_hugetlb_folio(struct hstate *h, 2084 nodemask_t *nodes_allowed, bool acct_surplus) 2085 { 2086 int nr_nodes, node; 2087 struct folio *folio = NULL; 2088 2089 lockdep_assert_held(&hugetlb_lock); 2090 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { 2091 /* 2092 * If we're returning unused surplus pages, only examine 2093 * nodes with surplus pages. 2094 */ 2095 if ((!acct_surplus || h->surplus_huge_pages_node[node]) && 2096 !list_empty(&h->hugepage_freelists[node])) { 2097 folio = list_entry(h->hugepage_freelists[node].next, 2098 struct folio, lru); 2099 remove_hugetlb_folio(h, folio, acct_surplus); 2100 break; 2101 } 2102 } 2103 2104 return folio; 2105 } 2106 2107 /* 2108 * Dissolve a given free hugetlb folio into free buddy pages. This function 2109 * does nothing for in-use hugetlb folios and non-hugetlb folios. 2110 * This function returns values like below: 2111 * 2112 * -ENOMEM: failed to allocate vmemmap pages to free the freed hugepages 2113 * when the system is under memory pressure and the feature of 2114 * freeing unused vmemmap pages associated with each hugetlb page 2115 * is enabled. 2116 * -EBUSY: failed to dissolved free hugepages or the hugepage is in-use 2117 * (allocated or reserved.) 2118 * 0: successfully dissolved free hugepages or the page is not a 2119 * hugepage (considered as already dissolved) 2120 */ 2121 int dissolve_free_hugetlb_folio(struct folio *folio) 2122 { 2123 int rc = -EBUSY; 2124 2125 retry: 2126 /* Not to disrupt normal path by vainly holding hugetlb_lock */ 2127 if (!folio_test_hugetlb(folio)) 2128 return 0; 2129 2130 spin_lock_irq(&hugetlb_lock); 2131 if (!folio_test_hugetlb(folio)) { 2132 rc = 0; 2133 goto out; 2134 } 2135 2136 if (!folio_ref_count(folio)) { 2137 struct hstate *h = folio_hstate(folio); 2138 if (!available_huge_pages(h)) 2139 goto out; 2140 2141 /* 2142 * We should make sure that the page is already on the free list 2143 * when it is dissolved. 2144 */ 2145 if (unlikely(!folio_test_hugetlb_freed(folio))) { 2146 spin_unlock_irq(&hugetlb_lock); 2147 cond_resched(); 2148 2149 /* 2150 * Theoretically, we should return -EBUSY when we 2151 * encounter this race. In fact, we have a chance 2152 * to successfully dissolve the page if we do a 2153 * retry. Because the race window is quite small. 2154 * If we seize this opportunity, it is an optimization 2155 * for increasing the success rate of dissolving page. 2156 */ 2157 goto retry; 2158 } 2159 2160 remove_hugetlb_folio(h, folio, false); 2161 h->max_huge_pages--; 2162 spin_unlock_irq(&hugetlb_lock); 2163 2164 /* 2165 * Normally update_and_free_hugtlb_folio will allocate required vmemmmap 2166 * before freeing the page. update_and_free_hugtlb_folio will fail to 2167 * free the page if it can not allocate required vmemmap. We 2168 * need to adjust max_huge_pages if the page is not freed. 2169 * Attempt to allocate vmemmmap here so that we can take 2170 * appropriate action on failure. 2171 * 2172 * The folio_test_hugetlb check here is because 2173 * remove_hugetlb_folio will clear hugetlb folio flag for 2174 * non-vmemmap optimized hugetlb folios. 2175 */ 2176 if (folio_test_hugetlb(folio)) { 2177 rc = hugetlb_vmemmap_restore_folio(h, folio); 2178 if (rc) { 2179 spin_lock_irq(&hugetlb_lock); 2180 add_hugetlb_folio(h, folio, false); 2181 h->max_huge_pages++; 2182 goto out; 2183 } 2184 } else 2185 rc = 0; 2186 2187 update_and_free_hugetlb_folio(h, folio, false); 2188 return rc; 2189 } 2190 out: 2191 spin_unlock_irq(&hugetlb_lock); 2192 return rc; 2193 } 2194 2195 /* 2196 * Dissolve free hugepages in a given pfn range. Used by memory hotplug to 2197 * make specified memory blocks removable from the system. 2198 * Note that this will dissolve a free gigantic hugepage completely, if any 2199 * part of it lies within the given range. 2200 * Also note that if dissolve_free_hugetlb_folio() returns with an error, all 2201 * free hugetlb folios that were dissolved before that error are lost. 2202 */ 2203 int dissolve_free_hugetlb_folios(unsigned long start_pfn, unsigned long end_pfn) 2204 { 2205 unsigned long pfn; 2206 struct folio *folio; 2207 int rc = 0; 2208 unsigned int order; 2209 struct hstate *h; 2210 2211 if (!hugepages_supported()) 2212 return rc; 2213 2214 order = huge_page_order(&default_hstate); 2215 for_each_hstate(h) 2216 order = min(order, huge_page_order(h)); 2217 2218 for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << order) { 2219 folio = pfn_folio(pfn); 2220 rc = dissolve_free_hugetlb_folio(folio); 2221 if (rc) 2222 break; 2223 } 2224 2225 return rc; 2226 } 2227 2228 /* 2229 * Allocates a fresh surplus page from the page allocator. 2230 */ 2231 static struct folio *alloc_surplus_hugetlb_folio(struct hstate *h, 2232 gfp_t gfp_mask, int nid, nodemask_t *nmask) 2233 { 2234 struct folio *folio = NULL; 2235 2236 if (hstate_is_gigantic(h)) 2237 return NULL; 2238 2239 spin_lock_irq(&hugetlb_lock); 2240 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) 2241 goto out_unlock; 2242 spin_unlock_irq(&hugetlb_lock); 2243 2244 folio = alloc_fresh_hugetlb_folio(h, gfp_mask, nid, nmask); 2245 if (!folio) 2246 return NULL; 2247 2248 spin_lock_irq(&hugetlb_lock); 2249 /* 2250 * We could have raced with the pool size change. 2251 * Double check that and simply deallocate the new page 2252 * if we would end up overcommiting the surpluses. Abuse 2253 * temporary page to workaround the nasty free_huge_folio 2254 * codeflow 2255 */ 2256 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) { 2257 folio_set_hugetlb_temporary(folio); 2258 spin_unlock_irq(&hugetlb_lock); 2259 free_huge_folio(folio); 2260 return NULL; 2261 } 2262 2263 h->surplus_huge_pages++; 2264 h->surplus_huge_pages_node[folio_nid(folio)]++; 2265 2266 out_unlock: 2267 spin_unlock_irq(&hugetlb_lock); 2268 2269 return folio; 2270 } 2271 2272 static struct folio *alloc_migrate_hugetlb_folio(struct hstate *h, gfp_t gfp_mask, 2273 int nid, nodemask_t *nmask) 2274 { 2275 struct folio *folio; 2276 2277 if (hstate_is_gigantic(h)) 2278 return NULL; 2279 2280 folio = alloc_fresh_hugetlb_folio(h, gfp_mask, nid, nmask); 2281 if (!folio) 2282 return NULL; 2283 2284 /* fresh huge pages are frozen */ 2285 folio_ref_unfreeze(folio, 1); 2286 /* 2287 * We do not account these pages as surplus because they are only 2288 * temporary and will be released properly on the last reference 2289 */ 2290 folio_set_hugetlb_temporary(folio); 2291 2292 return folio; 2293 } 2294 2295 /* 2296 * Use the VMA's mpolicy to allocate a huge page from the buddy. 2297 */ 2298 static 2299 struct folio *alloc_buddy_hugetlb_folio_with_mpol(struct hstate *h, 2300 struct vm_area_struct *vma, unsigned long addr) 2301 { 2302 struct folio *folio = NULL; 2303 struct mempolicy *mpol; 2304 gfp_t gfp_mask = htlb_alloc_mask(h); 2305 int nid; 2306 nodemask_t *nodemask; 2307 2308 nid = huge_node(vma, addr, gfp_mask, &mpol, &nodemask); 2309 if (mpol_is_preferred_many(mpol)) { 2310 gfp_t gfp = gfp_mask & ~(__GFP_DIRECT_RECLAIM | __GFP_NOFAIL); 2311 2312 folio = alloc_surplus_hugetlb_folio(h, gfp, nid, nodemask); 2313 2314 /* Fallback to all nodes if page==NULL */ 2315 nodemask = NULL; 2316 } 2317 2318 if (!folio) 2319 folio = alloc_surplus_hugetlb_folio(h, gfp_mask, nid, nodemask); 2320 mpol_cond_put(mpol); 2321 return folio; 2322 } 2323 2324 struct folio *alloc_hugetlb_folio_reserve(struct hstate *h, int preferred_nid, 2325 nodemask_t *nmask, gfp_t gfp_mask) 2326 { 2327 struct folio *folio; 2328 2329 spin_lock_irq(&hugetlb_lock); 2330 folio = dequeue_hugetlb_folio_nodemask(h, gfp_mask, preferred_nid, 2331 nmask); 2332 if (folio) { 2333 VM_BUG_ON(!h->resv_huge_pages); 2334 h->resv_huge_pages--; 2335 } 2336 2337 spin_unlock_irq(&hugetlb_lock); 2338 return folio; 2339 } 2340 2341 /* folio migration callback function */ 2342 struct folio *alloc_hugetlb_folio_nodemask(struct hstate *h, int preferred_nid, 2343 nodemask_t *nmask, gfp_t gfp_mask, bool allow_alloc_fallback) 2344 { 2345 spin_lock_irq(&hugetlb_lock); 2346 if (available_huge_pages(h)) { 2347 struct folio *folio; 2348 2349 folio = dequeue_hugetlb_folio_nodemask(h, gfp_mask, 2350 preferred_nid, nmask); 2351 if (folio) { 2352 spin_unlock_irq(&hugetlb_lock); 2353 return folio; 2354 } 2355 } 2356 spin_unlock_irq(&hugetlb_lock); 2357 2358 /* We cannot fallback to other nodes, as we could break the per-node pool. */ 2359 if (!allow_alloc_fallback) 2360 gfp_mask |= __GFP_THISNODE; 2361 2362 return alloc_migrate_hugetlb_folio(h, gfp_mask, preferred_nid, nmask); 2363 } 2364 2365 static nodemask_t *policy_mbind_nodemask(gfp_t gfp) 2366 { 2367 #ifdef CONFIG_NUMA 2368 struct mempolicy *mpol = get_task_policy(current); 2369 2370 /* 2371 * Only enforce MPOL_BIND policy which overlaps with cpuset policy 2372 * (from policy_nodemask) specifically for hugetlb case 2373 */ 2374 if (mpol->mode == MPOL_BIND && 2375 (apply_policy_zone(mpol, gfp_zone(gfp)) && 2376 cpuset_nodemask_valid_mems_allowed(&mpol->nodes))) 2377 return &mpol->nodes; 2378 #endif 2379 return NULL; 2380 } 2381 2382 /* 2383 * Increase the hugetlb pool such that it can accommodate a reservation 2384 * of size 'delta'. 2385 */ 2386 static int gather_surplus_pages(struct hstate *h, long delta) 2387 __must_hold(&hugetlb_lock) 2388 { 2389 LIST_HEAD(surplus_list); 2390 struct folio *folio, *tmp; 2391 int ret; 2392 long i; 2393 long needed, allocated; 2394 bool alloc_ok = true; 2395 int node; 2396 nodemask_t *mbind_nodemask, alloc_nodemask; 2397 2398 mbind_nodemask = policy_mbind_nodemask(htlb_alloc_mask(h)); 2399 if (mbind_nodemask) 2400 nodes_and(alloc_nodemask, *mbind_nodemask, cpuset_current_mems_allowed); 2401 else 2402 alloc_nodemask = cpuset_current_mems_allowed; 2403 2404 lockdep_assert_held(&hugetlb_lock); 2405 needed = (h->resv_huge_pages + delta) - h->free_huge_pages; 2406 if (needed <= 0) { 2407 h->resv_huge_pages += delta; 2408 return 0; 2409 } 2410 2411 allocated = 0; 2412 2413 ret = -ENOMEM; 2414 retry: 2415 spin_unlock_irq(&hugetlb_lock); 2416 for (i = 0; i < needed; i++) { 2417 folio = NULL; 2418 2419 /* Prioritize current node */ 2420 if (node_isset(numa_mem_id(), alloc_nodemask)) 2421 folio = alloc_surplus_hugetlb_folio(h, htlb_alloc_mask(h), 2422 numa_mem_id(), NULL); 2423 2424 if (!folio) { 2425 for_each_node_mask(node, alloc_nodemask) { 2426 if (node == numa_mem_id()) 2427 continue; 2428 folio = alloc_surplus_hugetlb_folio(h, htlb_alloc_mask(h), 2429 node, NULL); 2430 if (folio) 2431 break; 2432 } 2433 } 2434 if (!folio) { 2435 alloc_ok = false; 2436 break; 2437 } 2438 list_add(&folio->lru, &surplus_list); 2439 cond_resched(); 2440 } 2441 allocated += i; 2442 2443 /* 2444 * After retaking hugetlb_lock, we need to recalculate 'needed' 2445 * because either resv_huge_pages or free_huge_pages may have changed. 2446 */ 2447 spin_lock_irq(&hugetlb_lock); 2448 needed = (h->resv_huge_pages + delta) - 2449 (h->free_huge_pages + allocated); 2450 if (needed > 0) { 2451 if (alloc_ok) 2452 goto retry; 2453 /* 2454 * We were not able to allocate enough pages to 2455 * satisfy the entire reservation so we free what 2456 * we've allocated so far. 2457 */ 2458 goto free; 2459 } 2460 /* 2461 * The surplus_list now contains _at_least_ the number of extra pages 2462 * needed to accommodate the reservation. Add the appropriate number 2463 * of pages to the hugetlb pool and free the extras back to the buddy 2464 * allocator. Commit the entire reservation here to prevent another 2465 * process from stealing the pages as they are added to the pool but 2466 * before they are reserved. 2467 */ 2468 needed += allocated; 2469 h->resv_huge_pages += delta; 2470 ret = 0; 2471 2472 /* Free the needed pages to the hugetlb pool */ 2473 list_for_each_entry_safe(folio, tmp, &surplus_list, lru) { 2474 if ((--needed) < 0) 2475 break; 2476 /* Add the page to the hugetlb allocator */ 2477 enqueue_hugetlb_folio(h, folio); 2478 } 2479 free: 2480 spin_unlock_irq(&hugetlb_lock); 2481 2482 /* 2483 * Free unnecessary surplus pages to the buddy allocator. 2484 * Pages have no ref count, call free_huge_folio directly. 2485 */ 2486 list_for_each_entry_safe(folio, tmp, &surplus_list, lru) 2487 free_huge_folio(folio); 2488 spin_lock_irq(&hugetlb_lock); 2489 2490 return ret; 2491 } 2492 2493 /* 2494 * This routine has two main purposes: 2495 * 1) Decrement the reservation count (resv_huge_pages) by the value passed 2496 * in unused_resv_pages. This corresponds to the prior adjustments made 2497 * to the associated reservation map. 2498 * 2) Free any unused surplus pages that may have been allocated to satisfy 2499 * the reservation. As many as unused_resv_pages may be freed. 2500 */ 2501 static void return_unused_surplus_pages(struct hstate *h, 2502 unsigned long unused_resv_pages) 2503 { 2504 unsigned long nr_pages; 2505 LIST_HEAD(page_list); 2506 2507 lockdep_assert_held(&hugetlb_lock); 2508 /* Uncommit the reservation */ 2509 h->resv_huge_pages -= unused_resv_pages; 2510 2511 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) 2512 goto out; 2513 2514 /* 2515 * Part (or even all) of the reservation could have been backed 2516 * by pre-allocated pages. Only free surplus pages. 2517 */ 2518 nr_pages = min(unused_resv_pages, h->surplus_huge_pages); 2519 2520 /* 2521 * We want to release as many surplus pages as possible, spread 2522 * evenly across all nodes with memory. Iterate across these nodes 2523 * until we can no longer free unreserved surplus pages. This occurs 2524 * when the nodes with surplus pages have no free pages. 2525 * remove_pool_hugetlb_folio() will balance the freed pages across the 2526 * on-line nodes with memory and will handle the hstate accounting. 2527 */ 2528 while (nr_pages--) { 2529 struct folio *folio; 2530 2531 folio = remove_pool_hugetlb_folio(h, &node_states[N_MEMORY], 1); 2532 if (!folio) 2533 goto out; 2534 2535 list_add(&folio->lru, &page_list); 2536 } 2537 2538 out: 2539 spin_unlock_irq(&hugetlb_lock); 2540 update_and_free_pages_bulk(h, &page_list); 2541 spin_lock_irq(&hugetlb_lock); 2542 } 2543 2544 2545 /* 2546 * vma_needs_reservation, vma_commit_reservation and vma_end_reservation 2547 * are used by the huge page allocation routines to manage reservations. 2548 * 2549 * vma_needs_reservation is called to determine if the huge page at addr 2550 * within the vma has an associated reservation. If a reservation is 2551 * needed, the value 1 is returned. The caller is then responsible for 2552 * managing the global reservation and subpool usage counts. After 2553 * the huge page has been allocated, vma_commit_reservation is called 2554 * to add the page to the reservation map. If the page allocation fails, 2555 * the reservation must be ended instead of committed. vma_end_reservation 2556 * is called in such cases. 2557 * 2558 * In the normal case, vma_commit_reservation returns the same value 2559 * as the preceding vma_needs_reservation call. The only time this 2560 * is not the case is if a reserve map was changed between calls. It 2561 * is the responsibility of the caller to notice the difference and 2562 * take appropriate action. 2563 * 2564 * vma_add_reservation is used in error paths where a reservation must 2565 * be restored when a newly allocated huge page must be freed. It is 2566 * to be called after calling vma_needs_reservation to determine if a 2567 * reservation exists. 2568 * 2569 * vma_del_reservation is used in error paths where an entry in the reserve 2570 * map was created during huge page allocation and must be removed. It is to 2571 * be called after calling vma_needs_reservation to determine if a reservation 2572 * exists. 2573 */ 2574 enum vma_resv_mode { 2575 VMA_NEEDS_RESV, 2576 VMA_COMMIT_RESV, 2577 VMA_END_RESV, 2578 VMA_ADD_RESV, 2579 VMA_DEL_RESV, 2580 }; 2581 static long __vma_reservation_common(struct hstate *h, 2582 struct vm_area_struct *vma, unsigned long addr, 2583 enum vma_resv_mode mode) 2584 { 2585 struct resv_map *resv; 2586 pgoff_t idx; 2587 long ret; 2588 long dummy_out_regions_needed; 2589 2590 resv = vma_resv_map(vma); 2591 if (!resv) 2592 return 1; 2593 2594 idx = vma_hugecache_offset(h, vma, addr); 2595 switch (mode) { 2596 case VMA_NEEDS_RESV: 2597 ret = region_chg(resv, idx, idx + 1, &dummy_out_regions_needed); 2598 /* We assume that vma_reservation_* routines always operate on 2599 * 1 page, and that adding to resv map a 1 page entry can only 2600 * ever require 1 region. 2601 */ 2602 VM_BUG_ON(dummy_out_regions_needed != 1); 2603 break; 2604 case VMA_COMMIT_RESV: 2605 ret = region_add(resv, idx, idx + 1, 1, NULL, NULL); 2606 /* region_add calls of range 1 should never fail. */ 2607 VM_BUG_ON(ret < 0); 2608 break; 2609 case VMA_END_RESV: 2610 region_abort(resv, idx, idx + 1, 1); 2611 ret = 0; 2612 break; 2613 case VMA_ADD_RESV: 2614 if (vma->vm_flags & VM_MAYSHARE) { 2615 ret = region_add(resv, idx, idx + 1, 1, NULL, NULL); 2616 /* region_add calls of range 1 should never fail. */ 2617 VM_BUG_ON(ret < 0); 2618 } else { 2619 region_abort(resv, idx, idx + 1, 1); 2620 ret = region_del(resv, idx, idx + 1); 2621 } 2622 break; 2623 case VMA_DEL_RESV: 2624 if (vma->vm_flags & VM_MAYSHARE) { 2625 region_abort(resv, idx, idx + 1, 1); 2626 ret = region_del(resv, idx, idx + 1); 2627 } else { 2628 ret = region_add(resv, idx, idx + 1, 1, NULL, NULL); 2629 /* region_add calls of range 1 should never fail. */ 2630 VM_BUG_ON(ret < 0); 2631 } 2632 break; 2633 default: 2634 BUG(); 2635 } 2636 2637 if (vma->vm_flags & VM_MAYSHARE || mode == VMA_DEL_RESV) 2638 return ret; 2639 /* 2640 * We know private mapping must have HPAGE_RESV_OWNER set. 2641 * 2642 * In most cases, reserves always exist for private mappings. 2643 * However, a file associated with mapping could have been 2644 * hole punched or truncated after reserves were consumed. 2645 * As subsequent fault on such a range will not use reserves. 2646 * Subtle - The reserve map for private mappings has the 2647 * opposite meaning than that of shared mappings. If NO 2648 * entry is in the reserve map, it means a reservation exists. 2649 * If an entry exists in the reserve map, it means the 2650 * reservation has already been consumed. As a result, the 2651 * return value of this routine is the opposite of the 2652 * value returned from reserve map manipulation routines above. 2653 */ 2654 if (ret > 0) 2655 return 0; 2656 if (ret == 0) 2657 return 1; 2658 return ret; 2659 } 2660 2661 static long vma_needs_reservation(struct hstate *h, 2662 struct vm_area_struct *vma, unsigned long addr) 2663 { 2664 return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV); 2665 } 2666 2667 static long vma_commit_reservation(struct hstate *h, 2668 struct vm_area_struct *vma, unsigned long addr) 2669 { 2670 return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV); 2671 } 2672 2673 static void vma_end_reservation(struct hstate *h, 2674 struct vm_area_struct *vma, unsigned long addr) 2675 { 2676 (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV); 2677 } 2678 2679 static long vma_add_reservation(struct hstate *h, 2680 struct vm_area_struct *vma, unsigned long addr) 2681 { 2682 return __vma_reservation_common(h, vma, addr, VMA_ADD_RESV); 2683 } 2684 2685 static long vma_del_reservation(struct hstate *h, 2686 struct vm_area_struct *vma, unsigned long addr) 2687 { 2688 return __vma_reservation_common(h, vma, addr, VMA_DEL_RESV); 2689 } 2690 2691 /* 2692 * This routine is called to restore reservation information on error paths. 2693 * It should ONLY be called for folios allocated via alloc_hugetlb_folio(), 2694 * and the hugetlb mutex should remain held when calling this routine. 2695 * 2696 * It handles two specific cases: 2697 * 1) A reservation was in place and the folio consumed the reservation. 2698 * hugetlb_restore_reserve is set in the folio. 2699 * 2) No reservation was in place for the page, so hugetlb_restore_reserve is 2700 * not set. However, alloc_hugetlb_folio always updates the reserve map. 2701 * 2702 * In case 1, free_huge_folio later in the error path will increment the 2703 * global reserve count. But, free_huge_folio does not have enough context 2704 * to adjust the reservation map. This case deals primarily with private 2705 * mappings. Adjust the reserve map here to be consistent with global 2706 * reserve count adjustments to be made by free_huge_folio. Make sure the 2707 * reserve map indicates there is a reservation present. 2708 * 2709 * In case 2, simply undo reserve map modifications done by alloc_hugetlb_folio. 2710 */ 2711 void restore_reserve_on_error(struct hstate *h, struct vm_area_struct *vma, 2712 unsigned long address, struct folio *folio) 2713 { 2714 long rc = vma_needs_reservation(h, vma, address); 2715 2716 if (folio_test_hugetlb_restore_reserve(folio)) { 2717 if (unlikely(rc < 0)) 2718 /* 2719 * Rare out of memory condition in reserve map 2720 * manipulation. Clear hugetlb_restore_reserve so 2721 * that global reserve count will not be incremented 2722 * by free_huge_folio. This will make it appear 2723 * as though the reservation for this folio was 2724 * consumed. This may prevent the task from 2725 * faulting in the folio at a later time. This 2726 * is better than inconsistent global huge page 2727 * accounting of reserve counts. 2728 */ 2729 folio_clear_hugetlb_restore_reserve(folio); 2730 else if (rc) 2731 (void)vma_add_reservation(h, vma, address); 2732 else 2733 vma_end_reservation(h, vma, address); 2734 } else { 2735 if (!rc) { 2736 /* 2737 * This indicates there is an entry in the reserve map 2738 * not added by alloc_hugetlb_folio. We know it was added 2739 * before the alloc_hugetlb_folio call, otherwise 2740 * hugetlb_restore_reserve would be set on the folio. 2741 * Remove the entry so that a subsequent allocation 2742 * does not consume a reservation. 2743 */ 2744 rc = vma_del_reservation(h, vma, address); 2745 if (rc < 0) 2746 /* 2747 * VERY rare out of memory condition. Since 2748 * we can not delete the entry, set 2749 * hugetlb_restore_reserve so that the reserve 2750 * count will be incremented when the folio 2751 * is freed. This reserve will be consumed 2752 * on a subsequent allocation. 2753 */ 2754 folio_set_hugetlb_restore_reserve(folio); 2755 } else if (rc < 0) { 2756 /* 2757 * Rare out of memory condition from 2758 * vma_needs_reservation call. Memory allocation is 2759 * only attempted if a new entry is needed. Therefore, 2760 * this implies there is not an entry in the 2761 * reserve map. 2762 * 2763 * For shared mappings, no entry in the map indicates 2764 * no reservation. We are done. 2765 */ 2766 if (!(vma->vm_flags & VM_MAYSHARE)) 2767 /* 2768 * For private mappings, no entry indicates 2769 * a reservation is present. Since we can 2770 * not add an entry, set hugetlb_restore_reserve 2771 * on the folio so reserve count will be 2772 * incremented when freed. This reserve will 2773 * be consumed on a subsequent allocation. 2774 */ 2775 folio_set_hugetlb_restore_reserve(folio); 2776 } else 2777 /* 2778 * No reservation present, do nothing 2779 */ 2780 vma_end_reservation(h, vma, address); 2781 } 2782 } 2783 2784 /* 2785 * alloc_and_dissolve_hugetlb_folio - Allocate a new folio and dissolve 2786 * the old one 2787 * @h: struct hstate old page belongs to 2788 * @old_folio: Old folio to dissolve 2789 * @list: List to isolate the page in case we need to 2790 * Returns 0 on success, otherwise negated error. 2791 */ 2792 static int alloc_and_dissolve_hugetlb_folio(struct hstate *h, 2793 struct folio *old_folio, struct list_head *list) 2794 { 2795 gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE; 2796 int nid = folio_nid(old_folio); 2797 struct folio *new_folio = NULL; 2798 int ret = 0; 2799 2800 retry: 2801 spin_lock_irq(&hugetlb_lock); 2802 if (!folio_test_hugetlb(old_folio)) { 2803 /* 2804 * Freed from under us. Drop new_folio too. 2805 */ 2806 goto free_new; 2807 } else if (folio_ref_count(old_folio)) { 2808 bool isolated; 2809 2810 /* 2811 * Someone has grabbed the folio, try to isolate it here. 2812 * Fail with -EBUSY if not possible. 2813 */ 2814 spin_unlock_irq(&hugetlb_lock); 2815 isolated = folio_isolate_hugetlb(old_folio, list); 2816 ret = isolated ? 0 : -EBUSY; 2817 spin_lock_irq(&hugetlb_lock); 2818 goto free_new; 2819 } else if (!folio_test_hugetlb_freed(old_folio)) { 2820 /* 2821 * Folio's refcount is 0 but it has not been enqueued in the 2822 * freelist yet. Race window is small, so we can succeed here if 2823 * we retry. 2824 */ 2825 spin_unlock_irq(&hugetlb_lock); 2826 cond_resched(); 2827 goto retry; 2828 } else { 2829 if (!new_folio) { 2830 spin_unlock_irq(&hugetlb_lock); 2831 new_folio = alloc_buddy_hugetlb_folio(h, gfp_mask, nid, 2832 NULL, NULL); 2833 if (!new_folio) 2834 return -ENOMEM; 2835 __prep_new_hugetlb_folio(h, new_folio); 2836 goto retry; 2837 } 2838 2839 /* 2840 * Ok, old_folio is still a genuine free hugepage. Remove it from 2841 * the freelist and decrease the counters. These will be 2842 * incremented again when calling __prep_account_new_huge_page() 2843 * and enqueue_hugetlb_folio() for new_folio. The counters will 2844 * remain stable since this happens under the lock. 2845 */ 2846 remove_hugetlb_folio(h, old_folio, false); 2847 2848 /* 2849 * Ref count on new_folio is already zero as it was dropped 2850 * earlier. It can be directly added to the pool free list. 2851 */ 2852 __prep_account_new_huge_page(h, nid); 2853 enqueue_hugetlb_folio(h, new_folio); 2854 2855 /* 2856 * Folio has been replaced, we can safely free the old one. 2857 */ 2858 spin_unlock_irq(&hugetlb_lock); 2859 update_and_free_hugetlb_folio(h, old_folio, false); 2860 } 2861 2862 return ret; 2863 2864 free_new: 2865 spin_unlock_irq(&hugetlb_lock); 2866 if (new_folio) 2867 update_and_free_hugetlb_folio(h, new_folio, false); 2868 2869 return ret; 2870 } 2871 2872 int isolate_or_dissolve_huge_page(struct page *page, struct list_head *list) 2873 { 2874 struct hstate *h; 2875 struct folio *folio = page_folio(page); 2876 int ret = -EBUSY; 2877 2878 /* 2879 * The page might have been dissolved from under our feet, so make sure 2880 * to carefully check the state under the lock. 2881 * Return success when racing as if we dissolved the page ourselves. 2882 */ 2883 spin_lock_irq(&hugetlb_lock); 2884 if (folio_test_hugetlb(folio)) { 2885 h = folio_hstate(folio); 2886 } else { 2887 spin_unlock_irq(&hugetlb_lock); 2888 return 0; 2889 } 2890 spin_unlock_irq(&hugetlb_lock); 2891 2892 /* 2893 * Fence off gigantic pages as there is a cyclic dependency between 2894 * alloc_contig_range and them. Return -ENOMEM as this has the effect 2895 * of bailing out right away without further retrying. 2896 */ 2897 if (hstate_is_gigantic(h)) 2898 return -ENOMEM; 2899 2900 if (folio_ref_count(folio) && folio_isolate_hugetlb(folio, list)) 2901 ret = 0; 2902 else if (!folio_ref_count(folio)) 2903 ret = alloc_and_dissolve_hugetlb_folio(h, folio, list); 2904 2905 return ret; 2906 } 2907 2908 /* 2909 * replace_free_hugepage_folios - Replace free hugepage folios in a given pfn 2910 * range with new folios. 2911 * @start_pfn: start pfn of the given pfn range 2912 * @end_pfn: end pfn of the given pfn range 2913 * Returns 0 on success, otherwise negated error. 2914 */ 2915 int replace_free_hugepage_folios(unsigned long start_pfn, unsigned long end_pfn) 2916 { 2917 struct hstate *h; 2918 struct folio *folio; 2919 int ret = 0; 2920 2921 LIST_HEAD(isolate_list); 2922 2923 while (start_pfn < end_pfn) { 2924 folio = pfn_folio(start_pfn); 2925 if (folio_test_hugetlb(folio)) { 2926 h = folio_hstate(folio); 2927 } else { 2928 start_pfn++; 2929 continue; 2930 } 2931 2932 if (!folio_ref_count(folio)) { 2933 ret = alloc_and_dissolve_hugetlb_folio(h, folio, 2934 &isolate_list); 2935 if (ret) 2936 break; 2937 2938 putback_movable_pages(&isolate_list); 2939 } 2940 start_pfn++; 2941 } 2942 2943 return ret; 2944 } 2945 2946 void wait_for_freed_hugetlb_folios(void) 2947 { 2948 if (llist_empty(&hpage_freelist)) 2949 return; 2950 2951 flush_work(&free_hpage_work); 2952 } 2953 2954 typedef enum { 2955 /* 2956 * For either 0/1: we checked the per-vma resv map, and one resv 2957 * count either can be reused (0), or an extra needed (1). 2958 */ 2959 MAP_CHG_REUSE = 0, 2960 MAP_CHG_NEEDED = 1, 2961 /* 2962 * Cannot use per-vma resv count can be used, hence a new resv 2963 * count is enforced. 2964 * 2965 * NOTE: This is mostly identical to MAP_CHG_NEEDED, except 2966 * that currently vma_needs_reservation() has an unwanted side 2967 * effect to either use end() or commit() to complete the 2968 * transaction. Hence it needs to differenciate from NEEDED. 2969 */ 2970 MAP_CHG_ENFORCED = 2, 2971 } map_chg_state; 2972 2973 /* 2974 * NOTE! "cow_from_owner" represents a very hacky usage only used in CoW 2975 * faults of hugetlb private mappings on top of a non-page-cache folio (in 2976 * which case even if there's a private vma resv map it won't cover such 2977 * allocation). New call sites should (probably) never set it to true!! 2978 * When it's set, the allocation will bypass all vma level reservations. 2979 */ 2980 struct folio *alloc_hugetlb_folio(struct vm_area_struct *vma, 2981 unsigned long addr, bool cow_from_owner) 2982 { 2983 struct hugepage_subpool *spool = subpool_vma(vma); 2984 struct hstate *h = hstate_vma(vma); 2985 struct folio *folio; 2986 long retval, gbl_chg; 2987 map_chg_state map_chg; 2988 int ret, idx; 2989 struct hugetlb_cgroup *h_cg = NULL; 2990 gfp_t gfp = htlb_alloc_mask(h) | __GFP_RETRY_MAYFAIL; 2991 2992 idx = hstate_index(h); 2993 2994 /* Whether we need a separate per-vma reservation? */ 2995 if (cow_from_owner) { 2996 /* 2997 * Special case! Since it's a CoW on top of a reserved 2998 * page, the private resv map doesn't count. So it cannot 2999 * consume the per-vma resv map even if it's reserved. 3000 */ 3001 map_chg = MAP_CHG_ENFORCED; 3002 } else { 3003 /* 3004 * Examine the region/reserve map to determine if the process 3005 * has a reservation for the page to be allocated. A return 3006 * code of zero indicates a reservation exists (no change). 3007 */ 3008 retval = vma_needs_reservation(h, vma, addr); 3009 if (retval < 0) 3010 return ERR_PTR(-ENOMEM); 3011 map_chg = retval ? MAP_CHG_NEEDED : MAP_CHG_REUSE; 3012 } 3013 3014 /* 3015 * Whether we need a separate global reservation? 3016 * 3017 * Processes that did not create the mapping will have no 3018 * reserves as indicated by the region/reserve map. Check 3019 * that the allocation will not exceed the subpool limit. 3020 * Or if it can get one from the pool reservation directly. 3021 */ 3022 if (map_chg) { 3023 gbl_chg = hugepage_subpool_get_pages(spool, 1); 3024 if (gbl_chg < 0) 3025 goto out_end_reservation; 3026 } else { 3027 /* 3028 * If we have the vma reservation ready, no need for extra 3029 * global reservation. 3030 */ 3031 gbl_chg = 0; 3032 } 3033 3034 /* 3035 * If this allocation is not consuming a per-vma reservation, 3036 * charge the hugetlb cgroup now. 3037 */ 3038 if (map_chg) { 3039 ret = hugetlb_cgroup_charge_cgroup_rsvd( 3040 idx, pages_per_huge_page(h), &h_cg); 3041 if (ret) 3042 goto out_subpool_put; 3043 } 3044 3045 ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg); 3046 if (ret) 3047 goto out_uncharge_cgroup_reservation; 3048 3049 spin_lock_irq(&hugetlb_lock); 3050 /* 3051 * glb_chg is passed to indicate whether or not a page must be taken 3052 * from the global free pool (global change). gbl_chg == 0 indicates 3053 * a reservation exists for the allocation. 3054 */ 3055 folio = dequeue_hugetlb_folio_vma(h, vma, addr, gbl_chg); 3056 if (!folio) { 3057 spin_unlock_irq(&hugetlb_lock); 3058 folio = alloc_buddy_hugetlb_folio_with_mpol(h, vma, addr); 3059 if (!folio) 3060 goto out_uncharge_cgroup; 3061 spin_lock_irq(&hugetlb_lock); 3062 list_add(&folio->lru, &h->hugepage_activelist); 3063 folio_ref_unfreeze(folio, 1); 3064 /* Fall through */ 3065 } 3066 3067 /* 3068 * Either dequeued or buddy-allocated folio needs to add special 3069 * mark to the folio when it consumes a global reservation. 3070 */ 3071 if (!gbl_chg) { 3072 folio_set_hugetlb_restore_reserve(folio); 3073 h->resv_huge_pages--; 3074 } 3075 3076 hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, folio); 3077 /* If allocation is not consuming a reservation, also store the 3078 * hugetlb_cgroup pointer on the page. 3079 */ 3080 if (map_chg) { 3081 hugetlb_cgroup_commit_charge_rsvd(idx, pages_per_huge_page(h), 3082 h_cg, folio); 3083 } 3084 3085 spin_unlock_irq(&hugetlb_lock); 3086 3087 hugetlb_set_folio_subpool(folio, spool); 3088 3089 if (map_chg != MAP_CHG_ENFORCED) { 3090 /* commit() is only needed if the map_chg is not enforced */ 3091 retval = vma_commit_reservation(h, vma, addr); 3092 /* 3093 * Check for possible race conditions. When it happens.. 3094 * The page was added to the reservation map between 3095 * vma_needs_reservation and vma_commit_reservation. 3096 * This indicates a race with hugetlb_reserve_pages. 3097 * Adjust for the subpool count incremented above AND 3098 * in hugetlb_reserve_pages for the same page. Also, 3099 * the reservation count added in hugetlb_reserve_pages 3100 * no longer applies. 3101 */ 3102 if (unlikely(map_chg == MAP_CHG_NEEDED && retval == 0)) { 3103 long rsv_adjust; 3104 3105 rsv_adjust = hugepage_subpool_put_pages(spool, 1); 3106 hugetlb_acct_memory(h, -rsv_adjust); 3107 if (map_chg) { 3108 spin_lock_irq(&hugetlb_lock); 3109 hugetlb_cgroup_uncharge_folio_rsvd( 3110 hstate_index(h), pages_per_huge_page(h), 3111 folio); 3112 spin_unlock_irq(&hugetlb_lock); 3113 } 3114 } 3115 } 3116 3117 ret = mem_cgroup_charge_hugetlb(folio, gfp); 3118 /* 3119 * Unconditionally increment NR_HUGETLB here. If it turns out that 3120 * mem_cgroup_charge_hugetlb failed, then immediately free the page and 3121 * decrement NR_HUGETLB. 3122 */ 3123 lruvec_stat_mod_folio(folio, NR_HUGETLB, pages_per_huge_page(h)); 3124 3125 if (ret == -ENOMEM) { 3126 free_huge_folio(folio); 3127 return ERR_PTR(-ENOMEM); 3128 } 3129 3130 return folio; 3131 3132 out_uncharge_cgroup: 3133 hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg); 3134 out_uncharge_cgroup_reservation: 3135 if (map_chg) 3136 hugetlb_cgroup_uncharge_cgroup_rsvd(idx, pages_per_huge_page(h), 3137 h_cg); 3138 out_subpool_put: 3139 if (map_chg) 3140 hugepage_subpool_put_pages(spool, 1); 3141 out_end_reservation: 3142 if (map_chg != MAP_CHG_ENFORCED) 3143 vma_end_reservation(h, vma, addr); 3144 return ERR_PTR(-ENOSPC); 3145 } 3146 3147 int alloc_bootmem_huge_page(struct hstate *h, int nid) 3148 __attribute__ ((weak, alias("__alloc_bootmem_huge_page"))); 3149 int __alloc_bootmem_huge_page(struct hstate *h, int nid) 3150 { 3151 struct huge_bootmem_page *m = NULL; /* initialize for clang */ 3152 int nr_nodes, node = nid; 3153 3154 /* do node specific alloc */ 3155 if (nid != NUMA_NO_NODE) { 3156 m = memblock_alloc_exact_nid_raw(huge_page_size(h), huge_page_size(h), 3157 0, MEMBLOCK_ALLOC_ACCESSIBLE, nid); 3158 if (!m) 3159 return 0; 3160 goto found; 3161 } 3162 /* allocate from next node when distributing huge pages */ 3163 for_each_node_mask_to_alloc(&h->next_nid_to_alloc, nr_nodes, node, &node_states[N_MEMORY]) { 3164 m = memblock_alloc_try_nid_raw( 3165 huge_page_size(h), huge_page_size(h), 3166 0, MEMBLOCK_ALLOC_ACCESSIBLE, node); 3167 /* 3168 * Use the beginning of the huge page to store the 3169 * huge_bootmem_page struct (until gather_bootmem 3170 * puts them into the mem_map). 3171 */ 3172 if (!m) 3173 return 0; 3174 goto found; 3175 } 3176 3177 found: 3178 3179 /* 3180 * Only initialize the head struct page in memmap_init_reserved_pages, 3181 * rest of the struct pages will be initialized by the HugeTLB 3182 * subsystem itself. 3183 * The head struct page is used to get folio information by the HugeTLB 3184 * subsystem like zone id and node id. 3185 */ 3186 memblock_reserved_mark_noinit(virt_to_phys((void *)m + PAGE_SIZE), 3187 huge_page_size(h) - PAGE_SIZE); 3188 /* Put them into a private list first because mem_map is not up yet */ 3189 INIT_LIST_HEAD(&m->list); 3190 list_add(&m->list, &huge_boot_pages[node]); 3191 m->hstate = h; 3192 return 1; 3193 } 3194 3195 /* Initialize [start_page:end_page_number] tail struct pages of a hugepage */ 3196 static void __init hugetlb_folio_init_tail_vmemmap(struct folio *folio, 3197 unsigned long start_page_number, 3198 unsigned long end_page_number) 3199 { 3200 enum zone_type zone = zone_idx(folio_zone(folio)); 3201 int nid = folio_nid(folio); 3202 unsigned long head_pfn = folio_pfn(folio); 3203 unsigned long pfn, end_pfn = head_pfn + end_page_number; 3204 int ret; 3205 3206 for (pfn = head_pfn + start_page_number; pfn < end_pfn; pfn++) { 3207 struct page *page = pfn_to_page(pfn); 3208 3209 __ClearPageReserved(folio_page(folio, pfn - head_pfn)); 3210 __init_single_page(page, pfn, zone, nid); 3211 prep_compound_tail((struct page *)folio, pfn - head_pfn); 3212 ret = page_ref_freeze(page, 1); 3213 VM_BUG_ON(!ret); 3214 } 3215 } 3216 3217 static void __init hugetlb_folio_init_vmemmap(struct folio *folio, 3218 struct hstate *h, 3219 unsigned long nr_pages) 3220 { 3221 int ret; 3222 3223 /* Prepare folio head */ 3224 __folio_clear_reserved(folio); 3225 __folio_set_head(folio); 3226 ret = folio_ref_freeze(folio, 1); 3227 VM_BUG_ON(!ret); 3228 /* Initialize the necessary tail struct pages */ 3229 hugetlb_folio_init_tail_vmemmap(folio, 1, nr_pages); 3230 prep_compound_head((struct page *)folio, huge_page_order(h)); 3231 } 3232 3233 static void __init prep_and_add_bootmem_folios(struct hstate *h, 3234 struct list_head *folio_list) 3235 { 3236 unsigned long flags; 3237 struct folio *folio, *tmp_f; 3238 3239 /* Send list for bulk vmemmap optimization processing */ 3240 hugetlb_vmemmap_optimize_folios(h, folio_list); 3241 3242 list_for_each_entry_safe(folio, tmp_f, folio_list, lru) { 3243 if (!folio_test_hugetlb_vmemmap_optimized(folio)) { 3244 /* 3245 * If HVO fails, initialize all tail struct pages 3246 * We do not worry about potential long lock hold 3247 * time as this is early in boot and there should 3248 * be no contention. 3249 */ 3250 hugetlb_folio_init_tail_vmemmap(folio, 3251 HUGETLB_VMEMMAP_RESERVE_PAGES, 3252 pages_per_huge_page(h)); 3253 } 3254 /* Subdivide locks to achieve better parallel performance */ 3255 spin_lock_irqsave(&hugetlb_lock, flags); 3256 __prep_account_new_huge_page(h, folio_nid(folio)); 3257 enqueue_hugetlb_folio(h, folio); 3258 spin_unlock_irqrestore(&hugetlb_lock, flags); 3259 } 3260 } 3261 3262 /* 3263 * Put bootmem huge pages into the standard lists after mem_map is up. 3264 * Note: This only applies to gigantic (order > MAX_PAGE_ORDER) pages. 3265 */ 3266 static void __init gather_bootmem_prealloc_node(unsigned long nid) 3267 { 3268 LIST_HEAD(folio_list); 3269 struct huge_bootmem_page *m; 3270 struct hstate *h = NULL, *prev_h = NULL; 3271 3272 list_for_each_entry(m, &huge_boot_pages[nid], list) { 3273 struct page *page = virt_to_page(m); 3274 struct folio *folio = (void *)page; 3275 3276 h = m->hstate; 3277 /* 3278 * It is possible to have multiple huge page sizes (hstates) 3279 * in this list. If so, process each size separately. 3280 */ 3281 if (h != prev_h && prev_h != NULL) 3282 prep_and_add_bootmem_folios(prev_h, &folio_list); 3283 prev_h = h; 3284 3285 VM_BUG_ON(!hstate_is_gigantic(h)); 3286 WARN_ON(folio_ref_count(folio) != 1); 3287 3288 hugetlb_folio_init_vmemmap(folio, h, 3289 HUGETLB_VMEMMAP_RESERVE_PAGES); 3290 init_new_hugetlb_folio(h, folio); 3291 list_add(&folio->lru, &folio_list); 3292 3293 /* 3294 * We need to restore the 'stolen' pages to totalram_pages 3295 * in order to fix confusing memory reports from free(1) and 3296 * other side-effects, like CommitLimit going negative. 3297 */ 3298 adjust_managed_page_count(page, pages_per_huge_page(h)); 3299 cond_resched(); 3300 } 3301 3302 prep_and_add_bootmem_folios(h, &folio_list); 3303 } 3304 3305 static void __init gather_bootmem_prealloc_parallel(unsigned long start, 3306 unsigned long end, void *arg) 3307 { 3308 int nid; 3309 3310 for (nid = start; nid < end; nid++) 3311 gather_bootmem_prealloc_node(nid); 3312 } 3313 3314 static void __init gather_bootmem_prealloc(void) 3315 { 3316 struct padata_mt_job job = { 3317 .thread_fn = gather_bootmem_prealloc_parallel, 3318 .fn_arg = NULL, 3319 .start = 0, 3320 .size = nr_node_ids, 3321 .align = 1, 3322 .min_chunk = 1, 3323 .max_threads = num_node_state(N_MEMORY), 3324 .numa_aware = true, 3325 }; 3326 3327 padata_do_multithreaded(&job); 3328 } 3329 3330 static void __init hugetlb_hstate_alloc_pages_onenode(struct hstate *h, int nid) 3331 { 3332 unsigned long i; 3333 char buf[32]; 3334 LIST_HEAD(folio_list); 3335 3336 for (i = 0; i < h->max_huge_pages_node[nid]; ++i) { 3337 if (hstate_is_gigantic(h)) { 3338 if (!alloc_bootmem_huge_page(h, nid)) 3339 break; 3340 } else { 3341 struct folio *folio; 3342 gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE; 3343 3344 folio = only_alloc_fresh_hugetlb_folio(h, gfp_mask, nid, 3345 &node_states[N_MEMORY], NULL); 3346 if (!folio) 3347 break; 3348 list_add(&folio->lru, &folio_list); 3349 } 3350 cond_resched(); 3351 } 3352 3353 if (!list_empty(&folio_list)) 3354 prep_and_add_allocated_folios(h, &folio_list); 3355 3356 if (i == h->max_huge_pages_node[nid]) 3357 return; 3358 3359 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32); 3360 pr_warn("HugeTLB: allocating %u of page size %s failed node%d. Only allocated %lu hugepages.\n", 3361 h->max_huge_pages_node[nid], buf, nid, i); 3362 h->max_huge_pages -= (h->max_huge_pages_node[nid] - i); 3363 h->max_huge_pages_node[nid] = i; 3364 } 3365 3366 static bool __init hugetlb_hstate_alloc_pages_specific_nodes(struct hstate *h) 3367 { 3368 int i; 3369 bool node_specific_alloc = false; 3370 3371 for_each_online_node(i) { 3372 if (h->max_huge_pages_node[i] > 0) { 3373 hugetlb_hstate_alloc_pages_onenode(h, i); 3374 node_specific_alloc = true; 3375 } 3376 } 3377 3378 return node_specific_alloc; 3379 } 3380 3381 static void __init hugetlb_hstate_alloc_pages_errcheck(unsigned long allocated, struct hstate *h) 3382 { 3383 if (allocated < h->max_huge_pages) { 3384 char buf[32]; 3385 3386 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32); 3387 pr_warn("HugeTLB: allocating %lu of page size %s failed. Only allocated %lu hugepages.\n", 3388 h->max_huge_pages, buf, allocated); 3389 h->max_huge_pages = allocated; 3390 } 3391 } 3392 3393 static void __init hugetlb_pages_alloc_boot_node(unsigned long start, unsigned long end, void *arg) 3394 { 3395 struct hstate *h = (struct hstate *)arg; 3396 int i, num = end - start; 3397 nodemask_t node_alloc_noretry; 3398 LIST_HEAD(folio_list); 3399 int next_node = first_online_node; 3400 3401 /* Bit mask controlling how hard we retry per-node allocations.*/ 3402 nodes_clear(node_alloc_noretry); 3403 3404 for (i = 0; i < num; ++i) { 3405 struct folio *folio = alloc_pool_huge_folio(h, &node_states[N_MEMORY], 3406 &node_alloc_noretry, &next_node); 3407 if (!folio) 3408 break; 3409 3410 list_move(&folio->lru, &folio_list); 3411 cond_resched(); 3412 } 3413 3414 prep_and_add_allocated_folios(h, &folio_list); 3415 } 3416 3417 static unsigned long __init hugetlb_gigantic_pages_alloc_boot(struct hstate *h) 3418 { 3419 unsigned long i; 3420 3421 for (i = 0; i < h->max_huge_pages; ++i) { 3422 if (!alloc_bootmem_huge_page(h, NUMA_NO_NODE)) 3423 break; 3424 cond_resched(); 3425 } 3426 3427 return i; 3428 } 3429 3430 static unsigned long __init hugetlb_pages_alloc_boot(struct hstate *h) 3431 { 3432 struct padata_mt_job job = { 3433 .fn_arg = h, 3434 .align = 1, 3435 .numa_aware = true 3436 }; 3437 3438 job.thread_fn = hugetlb_pages_alloc_boot_node; 3439 job.start = 0; 3440 job.size = h->max_huge_pages; 3441 3442 /* 3443 * job.max_threads is twice the num_node_state(N_MEMORY), 3444 * 3445 * Tests below indicate that a multiplier of 2 significantly improves 3446 * performance, and although larger values also provide improvements, 3447 * the gains are marginal. 3448 * 3449 * Therefore, choosing 2 as the multiplier strikes a good balance between 3450 * enhancing parallel processing capabilities and maintaining efficient 3451 * resource management. 3452 * 3453 * +------------+-------+-------+-------+-------+-------+ 3454 * | multiplier | 1 | 2 | 3 | 4 | 5 | 3455 * +------------+-------+-------+-------+-------+-------+ 3456 * | 256G 2node | 358ms | 215ms | 157ms | 134ms | 126ms | 3457 * | 2T 4node | 979ms | 679ms | 543ms | 489ms | 481ms | 3458 * | 50G 2node | 71ms | 44ms | 37ms | 30ms | 31ms | 3459 * +------------+-------+-------+-------+-------+-------+ 3460 */ 3461 job.max_threads = num_node_state(N_MEMORY) * 2; 3462 job.min_chunk = h->max_huge_pages / num_node_state(N_MEMORY) / 2; 3463 padata_do_multithreaded(&job); 3464 3465 return h->nr_huge_pages; 3466 } 3467 3468 /* 3469 * NOTE: this routine is called in different contexts for gigantic and 3470 * non-gigantic pages. 3471 * - For gigantic pages, this is called early in the boot process and 3472 * pages are allocated from memblock allocated or something similar. 3473 * Gigantic pages are actually added to pools later with the routine 3474 * gather_bootmem_prealloc. 3475 * - For non-gigantic pages, this is called later in the boot process after 3476 * all of mm is up and functional. Pages are allocated from buddy and 3477 * then added to hugetlb pools. 3478 */ 3479 static void __init hugetlb_hstate_alloc_pages(struct hstate *h) 3480 { 3481 unsigned long allocated; 3482 static bool initialized __initdata; 3483 3484 /* skip gigantic hugepages allocation if hugetlb_cma enabled */ 3485 if (hstate_is_gigantic(h) && hugetlb_cma_size) { 3486 pr_warn_once("HugeTLB: hugetlb_cma is enabled, skip boot time allocation\n"); 3487 return; 3488 } 3489 3490 /* hugetlb_hstate_alloc_pages will be called many times, initialize huge_boot_pages once */ 3491 if (!initialized) { 3492 int i = 0; 3493 3494 for (i = 0; i < MAX_NUMNODES; i++) 3495 INIT_LIST_HEAD(&huge_boot_pages[i]); 3496 initialized = true; 3497 } 3498 3499 /* do node specific alloc */ 3500 if (hugetlb_hstate_alloc_pages_specific_nodes(h)) 3501 return; 3502 3503 /* below will do all node balanced alloc */ 3504 if (hstate_is_gigantic(h)) 3505 allocated = hugetlb_gigantic_pages_alloc_boot(h); 3506 else 3507 allocated = hugetlb_pages_alloc_boot(h); 3508 3509 hugetlb_hstate_alloc_pages_errcheck(allocated, h); 3510 } 3511 3512 static void __init hugetlb_init_hstates(void) 3513 { 3514 struct hstate *h, *h2; 3515 3516 for_each_hstate(h) { 3517 /* oversize hugepages were init'ed in early boot */ 3518 if (!hstate_is_gigantic(h)) 3519 hugetlb_hstate_alloc_pages(h); 3520 3521 /* 3522 * Set demote order for each hstate. Note that 3523 * h->demote_order is initially 0. 3524 * - We can not demote gigantic pages if runtime freeing 3525 * is not supported, so skip this. 3526 * - If CMA allocation is possible, we can not demote 3527 * HUGETLB_PAGE_ORDER or smaller size pages. 3528 */ 3529 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) 3530 continue; 3531 if (hugetlb_cma_size && h->order <= HUGETLB_PAGE_ORDER) 3532 continue; 3533 for_each_hstate(h2) { 3534 if (h2 == h) 3535 continue; 3536 if (h2->order < h->order && 3537 h2->order > h->demote_order) 3538 h->demote_order = h2->order; 3539 } 3540 } 3541 } 3542 3543 static void __init report_hugepages(void) 3544 { 3545 struct hstate *h; 3546 3547 for_each_hstate(h) { 3548 char buf[32]; 3549 3550 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32); 3551 pr_info("HugeTLB: registered %s page size, pre-allocated %ld pages\n", 3552 buf, h->free_huge_pages); 3553 pr_info("HugeTLB: %d KiB vmemmap can be freed for a %s page\n", 3554 hugetlb_vmemmap_optimizable_size(h) / SZ_1K, buf); 3555 } 3556 } 3557 3558 #ifdef CONFIG_HIGHMEM 3559 static void try_to_free_low(struct hstate *h, unsigned long count, 3560 nodemask_t *nodes_allowed) 3561 { 3562 int i; 3563 LIST_HEAD(page_list); 3564 3565 lockdep_assert_held(&hugetlb_lock); 3566 if (hstate_is_gigantic(h)) 3567 return; 3568 3569 /* 3570 * Collect pages to be freed on a list, and free after dropping lock 3571 */ 3572 for_each_node_mask(i, *nodes_allowed) { 3573 struct folio *folio, *next; 3574 struct list_head *freel = &h->hugepage_freelists[i]; 3575 list_for_each_entry_safe(folio, next, freel, lru) { 3576 if (count >= h->nr_huge_pages) 3577 goto out; 3578 if (folio_test_highmem(folio)) 3579 continue; 3580 remove_hugetlb_folio(h, folio, false); 3581 list_add(&folio->lru, &page_list); 3582 } 3583 } 3584 3585 out: 3586 spin_unlock_irq(&hugetlb_lock); 3587 update_and_free_pages_bulk(h, &page_list); 3588 spin_lock_irq(&hugetlb_lock); 3589 } 3590 #else 3591 static inline void try_to_free_low(struct hstate *h, unsigned long count, 3592 nodemask_t *nodes_allowed) 3593 { 3594 } 3595 #endif 3596 3597 /* 3598 * Increment or decrement surplus_huge_pages. Keep node-specific counters 3599 * balanced by operating on them in a round-robin fashion. 3600 * Returns 1 if an adjustment was made. 3601 */ 3602 static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed, 3603 int delta) 3604 { 3605 int nr_nodes, node; 3606 3607 lockdep_assert_held(&hugetlb_lock); 3608 VM_BUG_ON(delta != -1 && delta != 1); 3609 3610 if (delta < 0) { 3611 for_each_node_mask_to_alloc(&h->next_nid_to_alloc, nr_nodes, node, nodes_allowed) { 3612 if (h->surplus_huge_pages_node[node]) 3613 goto found; 3614 } 3615 } else { 3616 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { 3617 if (h->surplus_huge_pages_node[node] < 3618 h->nr_huge_pages_node[node]) 3619 goto found; 3620 } 3621 } 3622 return 0; 3623 3624 found: 3625 h->surplus_huge_pages += delta; 3626 h->surplus_huge_pages_node[node] += delta; 3627 return 1; 3628 } 3629 3630 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages) 3631 static int set_max_huge_pages(struct hstate *h, unsigned long count, int nid, 3632 nodemask_t *nodes_allowed) 3633 { 3634 unsigned long min_count; 3635 unsigned long allocated; 3636 struct folio *folio; 3637 LIST_HEAD(page_list); 3638 NODEMASK_ALLOC(nodemask_t, node_alloc_noretry, GFP_KERNEL); 3639 3640 /* 3641 * Bit mask controlling how hard we retry per-node allocations. 3642 * If we can not allocate the bit mask, do not attempt to allocate 3643 * the requested huge pages. 3644 */ 3645 if (node_alloc_noretry) 3646 nodes_clear(*node_alloc_noretry); 3647 else 3648 return -ENOMEM; 3649 3650 /* 3651 * resize_lock mutex prevents concurrent adjustments to number of 3652 * pages in hstate via the proc/sysfs interfaces. 3653 */ 3654 mutex_lock(&h->resize_lock); 3655 flush_free_hpage_work(h); 3656 spin_lock_irq(&hugetlb_lock); 3657 3658 /* 3659 * Check for a node specific request. 3660 * Changing node specific huge page count may require a corresponding 3661 * change to the global count. In any case, the passed node mask 3662 * (nodes_allowed) will restrict alloc/free to the specified node. 3663 */ 3664 if (nid != NUMA_NO_NODE) { 3665 unsigned long old_count = count; 3666 3667 count += persistent_huge_pages(h) - 3668 (h->nr_huge_pages_node[nid] - 3669 h->surplus_huge_pages_node[nid]); 3670 /* 3671 * User may have specified a large count value which caused the 3672 * above calculation to overflow. In this case, they wanted 3673 * to allocate as many huge pages as possible. Set count to 3674 * largest possible value to align with their intention. 3675 */ 3676 if (count < old_count) 3677 count = ULONG_MAX; 3678 } 3679 3680 /* 3681 * Gigantic pages runtime allocation depend on the capability for large 3682 * page range allocation. 3683 * If the system does not provide this feature, return an error when 3684 * the user tries to allocate gigantic pages but let the user free the 3685 * boottime allocated gigantic pages. 3686 */ 3687 if (hstate_is_gigantic(h) && !IS_ENABLED(CONFIG_CONTIG_ALLOC)) { 3688 if (count > persistent_huge_pages(h)) { 3689 spin_unlock_irq(&hugetlb_lock); 3690 mutex_unlock(&h->resize_lock); 3691 NODEMASK_FREE(node_alloc_noretry); 3692 return -EINVAL; 3693 } 3694 /* Fall through to decrease pool */ 3695 } 3696 3697 /* 3698 * Increase the pool size 3699 * First take pages out of surplus state. Then make up the 3700 * remaining difference by allocating fresh huge pages. 3701 * 3702 * We might race with alloc_surplus_hugetlb_folio() here and be unable 3703 * to convert a surplus huge page to a normal huge page. That is 3704 * not critical, though, it just means the overall size of the 3705 * pool might be one hugepage larger than it needs to be, but 3706 * within all the constraints specified by the sysctls. 3707 */ 3708 while (h->surplus_huge_pages && count > persistent_huge_pages(h)) { 3709 if (!adjust_pool_surplus(h, nodes_allowed, -1)) 3710 break; 3711 } 3712 3713 allocated = 0; 3714 while (count > (persistent_huge_pages(h) + allocated)) { 3715 /* 3716 * If this allocation races such that we no longer need the 3717 * page, free_huge_folio will handle it by freeing the page 3718 * and reducing the surplus. 3719 */ 3720 spin_unlock_irq(&hugetlb_lock); 3721 3722 /* yield cpu to avoid soft lockup */ 3723 cond_resched(); 3724 3725 folio = alloc_pool_huge_folio(h, nodes_allowed, 3726 node_alloc_noretry, 3727 &h->next_nid_to_alloc); 3728 if (!folio) { 3729 prep_and_add_allocated_folios(h, &page_list); 3730 spin_lock_irq(&hugetlb_lock); 3731 goto out; 3732 } 3733 3734 list_add(&folio->lru, &page_list); 3735 allocated++; 3736 3737 /* Bail for signals. Probably ctrl-c from user */ 3738 if (signal_pending(current)) { 3739 prep_and_add_allocated_folios(h, &page_list); 3740 spin_lock_irq(&hugetlb_lock); 3741 goto out; 3742 } 3743 3744 spin_lock_irq(&hugetlb_lock); 3745 } 3746 3747 /* Add allocated pages to the pool */ 3748 if (!list_empty(&page_list)) { 3749 spin_unlock_irq(&hugetlb_lock); 3750 prep_and_add_allocated_folios(h, &page_list); 3751 spin_lock_irq(&hugetlb_lock); 3752 } 3753 3754 /* 3755 * Decrease the pool size 3756 * First return free pages to the buddy allocator (being careful 3757 * to keep enough around to satisfy reservations). Then place 3758 * pages into surplus state as needed so the pool will shrink 3759 * to the desired size as pages become free. 3760 * 3761 * By placing pages into the surplus state independent of the 3762 * overcommit value, we are allowing the surplus pool size to 3763 * exceed overcommit. There are few sane options here. Since 3764 * alloc_surplus_hugetlb_folio() is checking the global counter, 3765 * though, we'll note that we're not allowed to exceed surplus 3766 * and won't grow the pool anywhere else. Not until one of the 3767 * sysctls are changed, or the surplus pages go out of use. 3768 */ 3769 min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages; 3770 min_count = max(count, min_count); 3771 try_to_free_low(h, min_count, nodes_allowed); 3772 3773 /* 3774 * Collect pages to be removed on list without dropping lock 3775 */ 3776 while (min_count < persistent_huge_pages(h)) { 3777 folio = remove_pool_hugetlb_folio(h, nodes_allowed, 0); 3778 if (!folio) 3779 break; 3780 3781 list_add(&folio->lru, &page_list); 3782 } 3783 /* free the pages after dropping lock */ 3784 spin_unlock_irq(&hugetlb_lock); 3785 update_and_free_pages_bulk(h, &page_list); 3786 flush_free_hpage_work(h); 3787 spin_lock_irq(&hugetlb_lock); 3788 3789 while (count < persistent_huge_pages(h)) { 3790 if (!adjust_pool_surplus(h, nodes_allowed, 1)) 3791 break; 3792 } 3793 out: 3794 h->max_huge_pages = persistent_huge_pages(h); 3795 spin_unlock_irq(&hugetlb_lock); 3796 mutex_unlock(&h->resize_lock); 3797 3798 NODEMASK_FREE(node_alloc_noretry); 3799 3800 return 0; 3801 } 3802 3803 static long demote_free_hugetlb_folios(struct hstate *src, struct hstate *dst, 3804 struct list_head *src_list) 3805 { 3806 long rc; 3807 struct folio *folio, *next; 3808 LIST_HEAD(dst_list); 3809 LIST_HEAD(ret_list); 3810 3811 rc = hugetlb_vmemmap_restore_folios(src, src_list, &ret_list); 3812 list_splice_init(&ret_list, src_list); 3813 3814 /* 3815 * Taking target hstate mutex synchronizes with set_max_huge_pages. 3816 * Without the mutex, pages added to target hstate could be marked 3817 * as surplus. 3818 * 3819 * Note that we already hold src->resize_lock. To prevent deadlock, 3820 * use the convention of always taking larger size hstate mutex first. 3821 */ 3822 mutex_lock(&dst->resize_lock); 3823 3824 list_for_each_entry_safe(folio, next, src_list, lru) { 3825 int i; 3826 3827 if (folio_test_hugetlb_vmemmap_optimized(folio)) 3828 continue; 3829 3830 list_del(&folio->lru); 3831 3832 split_page_owner(&folio->page, huge_page_order(src), huge_page_order(dst)); 3833 pgalloc_tag_split(folio, huge_page_order(src), huge_page_order(dst)); 3834 3835 for (i = 0; i < pages_per_huge_page(src); i += pages_per_huge_page(dst)) { 3836 struct page *page = folio_page(folio, i); 3837 /* Careful: see __split_huge_page_tail() */ 3838 struct folio *new_folio = (struct folio *)page; 3839 3840 clear_compound_head(page); 3841 prep_compound_page(page, dst->order); 3842 3843 new_folio->mapping = NULL; 3844 init_new_hugetlb_folio(dst, new_folio); 3845 list_add(&new_folio->lru, &dst_list); 3846 } 3847 } 3848 3849 prep_and_add_allocated_folios(dst, &dst_list); 3850 3851 mutex_unlock(&dst->resize_lock); 3852 3853 return rc; 3854 } 3855 3856 static long demote_pool_huge_page(struct hstate *src, nodemask_t *nodes_allowed, 3857 unsigned long nr_to_demote) 3858 __must_hold(&hugetlb_lock) 3859 { 3860 int nr_nodes, node; 3861 struct hstate *dst; 3862 long rc = 0; 3863 long nr_demoted = 0; 3864 3865 lockdep_assert_held(&hugetlb_lock); 3866 3867 /* We should never get here if no demote order */ 3868 if (!src->demote_order) { 3869 pr_warn("HugeTLB: NULL demote order passed to demote_pool_huge_page.\n"); 3870 return -EINVAL; /* internal error */ 3871 } 3872 dst = size_to_hstate(PAGE_SIZE << src->demote_order); 3873 3874 for_each_node_mask_to_free(src, nr_nodes, node, nodes_allowed) { 3875 LIST_HEAD(list); 3876 struct folio *folio, *next; 3877 3878 list_for_each_entry_safe(folio, next, &src->hugepage_freelists[node], lru) { 3879 if (folio_test_hwpoison(folio)) 3880 continue; 3881 3882 remove_hugetlb_folio(src, folio, false); 3883 list_add(&folio->lru, &list); 3884 3885 if (++nr_demoted == nr_to_demote) 3886 break; 3887 } 3888 3889 spin_unlock_irq(&hugetlb_lock); 3890 3891 rc = demote_free_hugetlb_folios(src, dst, &list); 3892 3893 spin_lock_irq(&hugetlb_lock); 3894 3895 list_for_each_entry_safe(folio, next, &list, lru) { 3896 list_del(&folio->lru); 3897 add_hugetlb_folio(src, folio, false); 3898 3899 nr_demoted--; 3900 } 3901 3902 if (rc < 0 || nr_demoted == nr_to_demote) 3903 break; 3904 } 3905 3906 /* 3907 * Not absolutely necessary, but for consistency update max_huge_pages 3908 * based on pool changes for the demoted page. 3909 */ 3910 src->max_huge_pages -= nr_demoted; 3911 dst->max_huge_pages += nr_demoted << (huge_page_order(src) - huge_page_order(dst)); 3912 3913 if (rc < 0) 3914 return rc; 3915 3916 if (nr_demoted) 3917 return nr_demoted; 3918 /* 3919 * Only way to get here is if all pages on free lists are poisoned. 3920 * Return -EBUSY so that caller will not retry. 3921 */ 3922 return -EBUSY; 3923 } 3924 3925 #define HSTATE_ATTR_RO(_name) \ 3926 static struct kobj_attribute _name##_attr = __ATTR_RO(_name) 3927 3928 #define HSTATE_ATTR_WO(_name) \ 3929 static struct kobj_attribute _name##_attr = __ATTR_WO(_name) 3930 3931 #define HSTATE_ATTR(_name) \ 3932 static struct kobj_attribute _name##_attr = __ATTR_RW(_name) 3933 3934 static struct kobject *hugepages_kobj; 3935 static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; 3936 3937 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp); 3938 3939 static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp) 3940 { 3941 int i; 3942 3943 for (i = 0; i < HUGE_MAX_HSTATE; i++) 3944 if (hstate_kobjs[i] == kobj) { 3945 if (nidp) 3946 *nidp = NUMA_NO_NODE; 3947 return &hstates[i]; 3948 } 3949 3950 return kobj_to_node_hstate(kobj, nidp); 3951 } 3952 3953 static ssize_t nr_hugepages_show_common(struct kobject *kobj, 3954 struct kobj_attribute *attr, char *buf) 3955 { 3956 struct hstate *h; 3957 unsigned long nr_huge_pages; 3958 int nid; 3959 3960 h = kobj_to_hstate(kobj, &nid); 3961 if (nid == NUMA_NO_NODE) 3962 nr_huge_pages = h->nr_huge_pages; 3963 else 3964 nr_huge_pages = h->nr_huge_pages_node[nid]; 3965 3966 return sysfs_emit(buf, "%lu\n", nr_huge_pages); 3967 } 3968 3969 static ssize_t __nr_hugepages_store_common(bool obey_mempolicy, 3970 struct hstate *h, int nid, 3971 unsigned long count, size_t len) 3972 { 3973 int err; 3974 nodemask_t nodes_allowed, *n_mask; 3975 3976 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) 3977 return -EINVAL; 3978 3979 if (nid == NUMA_NO_NODE) { 3980 /* 3981 * global hstate attribute 3982 */ 3983 if (!(obey_mempolicy && 3984 init_nodemask_of_mempolicy(&nodes_allowed))) 3985 n_mask = &node_states[N_MEMORY]; 3986 else 3987 n_mask = &nodes_allowed; 3988 } else { 3989 /* 3990 * Node specific request. count adjustment happens in 3991 * set_max_huge_pages() after acquiring hugetlb_lock. 3992 */ 3993 init_nodemask_of_node(&nodes_allowed, nid); 3994 n_mask = &nodes_allowed; 3995 } 3996 3997 err = set_max_huge_pages(h, count, nid, n_mask); 3998 3999 return err ? err : len; 4000 } 4001 4002 static ssize_t nr_hugepages_store_common(bool obey_mempolicy, 4003 struct kobject *kobj, const char *buf, 4004 size_t len) 4005 { 4006 struct hstate *h; 4007 unsigned long count; 4008 int nid; 4009 int err; 4010 4011 err = kstrtoul(buf, 10, &count); 4012 if (err) 4013 return err; 4014 4015 h = kobj_to_hstate(kobj, &nid); 4016 return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len); 4017 } 4018 4019 static ssize_t nr_hugepages_show(struct kobject *kobj, 4020 struct kobj_attribute *attr, char *buf) 4021 { 4022 return nr_hugepages_show_common(kobj, attr, buf); 4023 } 4024 4025 static ssize_t nr_hugepages_store(struct kobject *kobj, 4026 struct kobj_attribute *attr, const char *buf, size_t len) 4027 { 4028 return nr_hugepages_store_common(false, kobj, buf, len); 4029 } 4030 HSTATE_ATTR(nr_hugepages); 4031 4032 #ifdef CONFIG_NUMA 4033 4034 /* 4035 * hstate attribute for optionally mempolicy-based constraint on persistent 4036 * huge page alloc/free. 4037 */ 4038 static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj, 4039 struct kobj_attribute *attr, 4040 char *buf) 4041 { 4042 return nr_hugepages_show_common(kobj, attr, buf); 4043 } 4044 4045 static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj, 4046 struct kobj_attribute *attr, const char *buf, size_t len) 4047 { 4048 return nr_hugepages_store_common(true, kobj, buf, len); 4049 } 4050 HSTATE_ATTR(nr_hugepages_mempolicy); 4051 #endif 4052 4053 4054 static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj, 4055 struct kobj_attribute *attr, char *buf) 4056 { 4057 struct hstate *h = kobj_to_hstate(kobj, NULL); 4058 return sysfs_emit(buf, "%lu\n", h->nr_overcommit_huge_pages); 4059 } 4060 4061 static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj, 4062 struct kobj_attribute *attr, const char *buf, size_t count) 4063 { 4064 int err; 4065 unsigned long input; 4066 struct hstate *h = kobj_to_hstate(kobj, NULL); 4067 4068 if (hstate_is_gigantic(h)) 4069 return -EINVAL; 4070 4071 err = kstrtoul(buf, 10, &input); 4072 if (err) 4073 return err; 4074 4075 spin_lock_irq(&hugetlb_lock); 4076 h->nr_overcommit_huge_pages = input; 4077 spin_unlock_irq(&hugetlb_lock); 4078 4079 return count; 4080 } 4081 HSTATE_ATTR(nr_overcommit_hugepages); 4082 4083 static ssize_t free_hugepages_show(struct kobject *kobj, 4084 struct kobj_attribute *attr, char *buf) 4085 { 4086 struct hstate *h; 4087 unsigned long free_huge_pages; 4088 int nid; 4089 4090 h = kobj_to_hstate(kobj, &nid); 4091 if (nid == NUMA_NO_NODE) 4092 free_huge_pages = h->free_huge_pages; 4093 else 4094 free_huge_pages = h->free_huge_pages_node[nid]; 4095 4096 return sysfs_emit(buf, "%lu\n", free_huge_pages); 4097 } 4098 HSTATE_ATTR_RO(free_hugepages); 4099 4100 static ssize_t resv_hugepages_show(struct kobject *kobj, 4101 struct kobj_attribute *attr, char *buf) 4102 { 4103 struct hstate *h = kobj_to_hstate(kobj, NULL); 4104 return sysfs_emit(buf, "%lu\n", h->resv_huge_pages); 4105 } 4106 HSTATE_ATTR_RO(resv_hugepages); 4107 4108 static ssize_t surplus_hugepages_show(struct kobject *kobj, 4109 struct kobj_attribute *attr, char *buf) 4110 { 4111 struct hstate *h; 4112 unsigned long surplus_huge_pages; 4113 int nid; 4114 4115 h = kobj_to_hstate(kobj, &nid); 4116 if (nid == NUMA_NO_NODE) 4117 surplus_huge_pages = h->surplus_huge_pages; 4118 else 4119 surplus_huge_pages = h->surplus_huge_pages_node[nid]; 4120 4121 return sysfs_emit(buf, "%lu\n", surplus_huge_pages); 4122 } 4123 HSTATE_ATTR_RO(surplus_hugepages); 4124 4125 static ssize_t demote_store(struct kobject *kobj, 4126 struct kobj_attribute *attr, const char *buf, size_t len) 4127 { 4128 unsigned long nr_demote; 4129 unsigned long nr_available; 4130 nodemask_t nodes_allowed, *n_mask; 4131 struct hstate *h; 4132 int err; 4133 int nid; 4134 4135 err = kstrtoul(buf, 10, &nr_demote); 4136 if (err) 4137 return err; 4138 h = kobj_to_hstate(kobj, &nid); 4139 4140 if (nid != NUMA_NO_NODE) { 4141 init_nodemask_of_node(&nodes_allowed, nid); 4142 n_mask = &nodes_allowed; 4143 } else { 4144 n_mask = &node_states[N_MEMORY]; 4145 } 4146 4147 /* Synchronize with other sysfs operations modifying huge pages */ 4148 mutex_lock(&h->resize_lock); 4149 spin_lock_irq(&hugetlb_lock); 4150 4151 while (nr_demote) { 4152 long rc; 4153 4154 /* 4155 * Check for available pages to demote each time thorough the 4156 * loop as demote_pool_huge_page will drop hugetlb_lock. 4157 */ 4158 if (nid != NUMA_NO_NODE) 4159 nr_available = h->free_huge_pages_node[nid]; 4160 else 4161 nr_available = h->free_huge_pages; 4162 nr_available -= h->resv_huge_pages; 4163 if (!nr_available) 4164 break; 4165 4166 rc = demote_pool_huge_page(h, n_mask, nr_demote); 4167 if (rc < 0) { 4168 err = rc; 4169 break; 4170 } 4171 4172 nr_demote -= rc; 4173 } 4174 4175 spin_unlock_irq(&hugetlb_lock); 4176 mutex_unlock(&h->resize_lock); 4177 4178 if (err) 4179 return err; 4180 return len; 4181 } 4182 HSTATE_ATTR_WO(demote); 4183 4184 static ssize_t demote_size_show(struct kobject *kobj, 4185 struct kobj_attribute *attr, char *buf) 4186 { 4187 struct hstate *h = kobj_to_hstate(kobj, NULL); 4188 unsigned long demote_size = (PAGE_SIZE << h->demote_order) / SZ_1K; 4189 4190 return sysfs_emit(buf, "%lukB\n", demote_size); 4191 } 4192 4193 static ssize_t demote_size_store(struct kobject *kobj, 4194 struct kobj_attribute *attr, 4195 const char *buf, size_t count) 4196 { 4197 struct hstate *h, *demote_hstate; 4198 unsigned long demote_size; 4199 unsigned int demote_order; 4200 4201 demote_size = (unsigned long)memparse(buf, NULL); 4202 4203 demote_hstate = size_to_hstate(demote_size); 4204 if (!demote_hstate) 4205 return -EINVAL; 4206 demote_order = demote_hstate->order; 4207 if (demote_order < HUGETLB_PAGE_ORDER) 4208 return -EINVAL; 4209 4210 /* demote order must be smaller than hstate order */ 4211 h = kobj_to_hstate(kobj, NULL); 4212 if (demote_order >= h->order) 4213 return -EINVAL; 4214 4215 /* resize_lock synchronizes access to demote size and writes */ 4216 mutex_lock(&h->resize_lock); 4217 h->demote_order = demote_order; 4218 mutex_unlock(&h->resize_lock); 4219 4220 return count; 4221 } 4222 HSTATE_ATTR(demote_size); 4223 4224 static struct attribute *hstate_attrs[] = { 4225 &nr_hugepages_attr.attr, 4226 &nr_overcommit_hugepages_attr.attr, 4227 &free_hugepages_attr.attr, 4228 &resv_hugepages_attr.attr, 4229 &surplus_hugepages_attr.attr, 4230 #ifdef CONFIG_NUMA 4231 &nr_hugepages_mempolicy_attr.attr, 4232 #endif 4233 NULL, 4234 }; 4235 4236 static const struct attribute_group hstate_attr_group = { 4237 .attrs = hstate_attrs, 4238 }; 4239 4240 static struct attribute *hstate_demote_attrs[] = { 4241 &demote_size_attr.attr, 4242 &demote_attr.attr, 4243 NULL, 4244 }; 4245 4246 static const struct attribute_group hstate_demote_attr_group = { 4247 .attrs = hstate_demote_attrs, 4248 }; 4249 4250 static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent, 4251 struct kobject **hstate_kobjs, 4252 const struct attribute_group *hstate_attr_group) 4253 { 4254 int retval; 4255 int hi = hstate_index(h); 4256 4257 hstate_kobjs[hi] = kobject_create_and_add(h->name, parent); 4258 if (!hstate_kobjs[hi]) 4259 return -ENOMEM; 4260 4261 retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group); 4262 if (retval) { 4263 kobject_put(hstate_kobjs[hi]); 4264 hstate_kobjs[hi] = NULL; 4265 return retval; 4266 } 4267 4268 if (h->demote_order) { 4269 retval = sysfs_create_group(hstate_kobjs[hi], 4270 &hstate_demote_attr_group); 4271 if (retval) { 4272 pr_warn("HugeTLB unable to create demote interfaces for %s\n", h->name); 4273 sysfs_remove_group(hstate_kobjs[hi], hstate_attr_group); 4274 kobject_put(hstate_kobjs[hi]); 4275 hstate_kobjs[hi] = NULL; 4276 return retval; 4277 } 4278 } 4279 4280 return 0; 4281 } 4282 4283 #ifdef CONFIG_NUMA 4284 static bool hugetlb_sysfs_initialized __ro_after_init; 4285 4286 /* 4287 * node_hstate/s - associate per node hstate attributes, via their kobjects, 4288 * with node devices in node_devices[] using a parallel array. The array 4289 * index of a node device or _hstate == node id. 4290 * This is here to avoid any static dependency of the node device driver, in 4291 * the base kernel, on the hugetlb module. 4292 */ 4293 struct node_hstate { 4294 struct kobject *hugepages_kobj; 4295 struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; 4296 }; 4297 static struct node_hstate node_hstates[MAX_NUMNODES]; 4298 4299 /* 4300 * A subset of global hstate attributes for node devices 4301 */ 4302 static struct attribute *per_node_hstate_attrs[] = { 4303 &nr_hugepages_attr.attr, 4304 &free_hugepages_attr.attr, 4305 &surplus_hugepages_attr.attr, 4306 NULL, 4307 }; 4308 4309 static const struct attribute_group per_node_hstate_attr_group = { 4310 .attrs = per_node_hstate_attrs, 4311 }; 4312 4313 /* 4314 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj. 4315 * Returns node id via non-NULL nidp. 4316 */ 4317 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) 4318 { 4319 int nid; 4320 4321 for (nid = 0; nid < nr_node_ids; nid++) { 4322 struct node_hstate *nhs = &node_hstates[nid]; 4323 int i; 4324 for (i = 0; i < HUGE_MAX_HSTATE; i++) 4325 if (nhs->hstate_kobjs[i] == kobj) { 4326 if (nidp) 4327 *nidp = nid; 4328 return &hstates[i]; 4329 } 4330 } 4331 4332 BUG(); 4333 return NULL; 4334 } 4335 4336 /* 4337 * Unregister hstate attributes from a single node device. 4338 * No-op if no hstate attributes attached. 4339 */ 4340 void hugetlb_unregister_node(struct node *node) 4341 { 4342 struct hstate *h; 4343 struct node_hstate *nhs = &node_hstates[node->dev.id]; 4344 4345 if (!nhs->hugepages_kobj) 4346 return; /* no hstate attributes */ 4347 4348 for_each_hstate(h) { 4349 int idx = hstate_index(h); 4350 struct kobject *hstate_kobj = nhs->hstate_kobjs[idx]; 4351 4352 if (!hstate_kobj) 4353 continue; 4354 if (h->demote_order) 4355 sysfs_remove_group(hstate_kobj, &hstate_demote_attr_group); 4356 sysfs_remove_group(hstate_kobj, &per_node_hstate_attr_group); 4357 kobject_put(hstate_kobj); 4358 nhs->hstate_kobjs[idx] = NULL; 4359 } 4360 4361 kobject_put(nhs->hugepages_kobj); 4362 nhs->hugepages_kobj = NULL; 4363 } 4364 4365 4366 /* 4367 * Register hstate attributes for a single node device. 4368 * No-op if attributes already registered. 4369 */ 4370 void hugetlb_register_node(struct node *node) 4371 { 4372 struct hstate *h; 4373 struct node_hstate *nhs = &node_hstates[node->dev.id]; 4374 int err; 4375 4376 if (!hugetlb_sysfs_initialized) 4377 return; 4378 4379 if (nhs->hugepages_kobj) 4380 return; /* already allocated */ 4381 4382 nhs->hugepages_kobj = kobject_create_and_add("hugepages", 4383 &node->dev.kobj); 4384 if (!nhs->hugepages_kobj) 4385 return; 4386 4387 for_each_hstate(h) { 4388 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj, 4389 nhs->hstate_kobjs, 4390 &per_node_hstate_attr_group); 4391 if (err) { 4392 pr_err("HugeTLB: Unable to add hstate %s for node %d\n", 4393 h->name, node->dev.id); 4394 hugetlb_unregister_node(node); 4395 break; 4396 } 4397 } 4398 } 4399 4400 /* 4401 * hugetlb init time: register hstate attributes for all registered node 4402 * devices of nodes that have memory. All on-line nodes should have 4403 * registered their associated device by this time. 4404 */ 4405 static void __init hugetlb_register_all_nodes(void) 4406 { 4407 int nid; 4408 4409 for_each_online_node(nid) 4410 hugetlb_register_node(node_devices[nid]); 4411 } 4412 #else /* !CONFIG_NUMA */ 4413 4414 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) 4415 { 4416 BUG(); 4417 if (nidp) 4418 *nidp = -1; 4419 return NULL; 4420 } 4421 4422 static void hugetlb_register_all_nodes(void) { } 4423 4424 #endif 4425 4426 #ifdef CONFIG_CMA 4427 static void __init hugetlb_cma_check(void); 4428 #else 4429 static inline __init void hugetlb_cma_check(void) 4430 { 4431 } 4432 #endif 4433 4434 static void __init hugetlb_sysfs_init(void) 4435 { 4436 struct hstate *h; 4437 int err; 4438 4439 hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj); 4440 if (!hugepages_kobj) 4441 return; 4442 4443 for_each_hstate(h) { 4444 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj, 4445 hstate_kobjs, &hstate_attr_group); 4446 if (err) 4447 pr_err("HugeTLB: Unable to add hstate %s", h->name); 4448 } 4449 4450 #ifdef CONFIG_NUMA 4451 hugetlb_sysfs_initialized = true; 4452 #endif 4453 hugetlb_register_all_nodes(); 4454 } 4455 4456 #ifdef CONFIG_SYSCTL 4457 static void hugetlb_sysctl_init(void); 4458 #else 4459 static inline void hugetlb_sysctl_init(void) { } 4460 #endif 4461 4462 static int __init hugetlb_init(void) 4463 { 4464 int i; 4465 4466 BUILD_BUG_ON(sizeof_field(struct page, private) * BITS_PER_BYTE < 4467 __NR_HPAGEFLAGS); 4468 4469 if (!hugepages_supported()) { 4470 if (hugetlb_max_hstate || default_hstate_max_huge_pages) 4471 pr_warn("HugeTLB: huge pages not supported, ignoring associated command-line parameters\n"); 4472 return 0; 4473 } 4474 4475 /* 4476 * Make sure HPAGE_SIZE (HUGETLB_PAGE_ORDER) hstate exists. Some 4477 * architectures depend on setup being done here. 4478 */ 4479 hugetlb_add_hstate(HUGETLB_PAGE_ORDER); 4480 if (!parsed_default_hugepagesz) { 4481 /* 4482 * If we did not parse a default huge page size, set 4483 * default_hstate_idx to HPAGE_SIZE hstate. And, if the 4484 * number of huge pages for this default size was implicitly 4485 * specified, set that here as well. 4486 * Note that the implicit setting will overwrite an explicit 4487 * setting. A warning will be printed in this case. 4488 */ 4489 default_hstate_idx = hstate_index(size_to_hstate(HPAGE_SIZE)); 4490 if (default_hstate_max_huge_pages) { 4491 if (default_hstate.max_huge_pages) { 4492 char buf[32]; 4493 4494 string_get_size(huge_page_size(&default_hstate), 4495 1, STRING_UNITS_2, buf, 32); 4496 pr_warn("HugeTLB: Ignoring hugepages=%lu associated with %s page size\n", 4497 default_hstate.max_huge_pages, buf); 4498 pr_warn("HugeTLB: Using hugepages=%lu for number of default huge pages\n", 4499 default_hstate_max_huge_pages); 4500 } 4501 default_hstate.max_huge_pages = 4502 default_hstate_max_huge_pages; 4503 4504 for_each_online_node(i) 4505 default_hstate.max_huge_pages_node[i] = 4506 default_hugepages_in_node[i]; 4507 } 4508 } 4509 4510 hugetlb_cma_check(); 4511 hugetlb_init_hstates(); 4512 gather_bootmem_prealloc(); 4513 report_hugepages(); 4514 4515 hugetlb_sysfs_init(); 4516 hugetlb_cgroup_file_init(); 4517 hugetlb_sysctl_init(); 4518 4519 #ifdef CONFIG_SMP 4520 num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus()); 4521 #else 4522 num_fault_mutexes = 1; 4523 #endif 4524 hugetlb_fault_mutex_table = 4525 kmalloc_array(num_fault_mutexes, sizeof(struct mutex), 4526 GFP_KERNEL); 4527 BUG_ON(!hugetlb_fault_mutex_table); 4528 4529 for (i = 0; i < num_fault_mutexes; i++) 4530 mutex_init(&hugetlb_fault_mutex_table[i]); 4531 return 0; 4532 } 4533 subsys_initcall(hugetlb_init); 4534 4535 /* Overwritten by architectures with more huge page sizes */ 4536 bool __init __attribute((weak)) arch_hugetlb_valid_size(unsigned long size) 4537 { 4538 return size == HPAGE_SIZE; 4539 } 4540 4541 void __init hugetlb_add_hstate(unsigned int order) 4542 { 4543 struct hstate *h; 4544 unsigned long i; 4545 4546 if (size_to_hstate(PAGE_SIZE << order)) { 4547 return; 4548 } 4549 BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE); 4550 BUG_ON(order < order_base_2(__NR_USED_SUBPAGE)); 4551 h = &hstates[hugetlb_max_hstate++]; 4552 __mutex_init(&h->resize_lock, "resize mutex", &h->resize_key); 4553 h->order = order; 4554 h->mask = ~(huge_page_size(h) - 1); 4555 for (i = 0; i < MAX_NUMNODES; ++i) 4556 INIT_LIST_HEAD(&h->hugepage_freelists[i]); 4557 INIT_LIST_HEAD(&h->hugepage_activelist); 4558 h->next_nid_to_alloc = first_memory_node; 4559 h->next_nid_to_free = first_memory_node; 4560 snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB", 4561 huge_page_size(h)/SZ_1K); 4562 4563 parsed_hstate = h; 4564 } 4565 4566 bool __init __weak hugetlb_node_alloc_supported(void) 4567 { 4568 return true; 4569 } 4570 4571 static void __init hugepages_clear_pages_in_node(void) 4572 { 4573 if (!hugetlb_max_hstate) { 4574 default_hstate_max_huge_pages = 0; 4575 memset(default_hugepages_in_node, 0, 4576 sizeof(default_hugepages_in_node)); 4577 } else { 4578 parsed_hstate->max_huge_pages = 0; 4579 memset(parsed_hstate->max_huge_pages_node, 0, 4580 sizeof(parsed_hstate->max_huge_pages_node)); 4581 } 4582 } 4583 4584 /* 4585 * hugepages command line processing 4586 * hugepages normally follows a valid hugepagsz or default_hugepagsz 4587 * specification. If not, ignore the hugepages value. hugepages can also 4588 * be the first huge page command line option in which case it implicitly 4589 * specifies the number of huge pages for the default size. 4590 */ 4591 static int __init hugepages_setup(char *s) 4592 { 4593 unsigned long *mhp; 4594 static unsigned long *last_mhp; 4595 int node = NUMA_NO_NODE; 4596 int count; 4597 unsigned long tmp; 4598 char *p = s; 4599 4600 if (!parsed_valid_hugepagesz) { 4601 pr_warn("HugeTLB: hugepages=%s does not follow a valid hugepagesz, ignoring\n", s); 4602 parsed_valid_hugepagesz = true; 4603 return 1; 4604 } 4605 4606 /* 4607 * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter 4608 * yet, so this hugepages= parameter goes to the "default hstate". 4609 * Otherwise, it goes with the previously parsed hugepagesz or 4610 * default_hugepagesz. 4611 */ 4612 else if (!hugetlb_max_hstate) 4613 mhp = &default_hstate_max_huge_pages; 4614 else 4615 mhp = &parsed_hstate->max_huge_pages; 4616 4617 if (mhp == last_mhp) { 4618 pr_warn("HugeTLB: hugepages= specified twice without interleaving hugepagesz=, ignoring hugepages=%s\n", s); 4619 return 1; 4620 } 4621 4622 while (*p) { 4623 count = 0; 4624 if (sscanf(p, "%lu%n", &tmp, &count) != 1) 4625 goto invalid; 4626 /* Parameter is node format */ 4627 if (p[count] == ':') { 4628 if (!hugetlb_node_alloc_supported()) { 4629 pr_warn("HugeTLB: architecture can't support node specific alloc, ignoring!\n"); 4630 return 1; 4631 } 4632 if (tmp >= MAX_NUMNODES || !node_online(tmp)) 4633 goto invalid; 4634 node = array_index_nospec(tmp, MAX_NUMNODES); 4635 p += count + 1; 4636 /* Parse hugepages */ 4637 if (sscanf(p, "%lu%n", &tmp, &count) != 1) 4638 goto invalid; 4639 if (!hugetlb_max_hstate) 4640 default_hugepages_in_node[node] = tmp; 4641 else 4642 parsed_hstate->max_huge_pages_node[node] = tmp; 4643 *mhp += tmp; 4644 /* Go to parse next node*/ 4645 if (p[count] == ',') 4646 p += count + 1; 4647 else 4648 break; 4649 } else { 4650 if (p != s) 4651 goto invalid; 4652 *mhp = tmp; 4653 break; 4654 } 4655 } 4656 4657 /* 4658 * Global state is always initialized later in hugetlb_init. 4659 * But we need to allocate gigantic hstates here early to still 4660 * use the bootmem allocator. 4661 */ 4662 if (hugetlb_max_hstate && hstate_is_gigantic(parsed_hstate)) 4663 hugetlb_hstate_alloc_pages(parsed_hstate); 4664 4665 last_mhp = mhp; 4666 4667 return 1; 4668 4669 invalid: 4670 pr_warn("HugeTLB: Invalid hugepages parameter %s\n", p); 4671 hugepages_clear_pages_in_node(); 4672 return 1; 4673 } 4674 __setup("hugepages=", hugepages_setup); 4675 4676 /* 4677 * hugepagesz command line processing 4678 * A specific huge page size can only be specified once with hugepagesz. 4679 * hugepagesz is followed by hugepages on the command line. The global 4680 * variable 'parsed_valid_hugepagesz' is used to determine if prior 4681 * hugepagesz argument was valid. 4682 */ 4683 static int __init hugepagesz_setup(char *s) 4684 { 4685 unsigned long size; 4686 struct hstate *h; 4687 4688 parsed_valid_hugepagesz = false; 4689 size = (unsigned long)memparse(s, NULL); 4690 4691 if (!arch_hugetlb_valid_size(size)) { 4692 pr_err("HugeTLB: unsupported hugepagesz=%s\n", s); 4693 return 1; 4694 } 4695 4696 h = size_to_hstate(size); 4697 if (h) { 4698 /* 4699 * hstate for this size already exists. This is normally 4700 * an error, but is allowed if the existing hstate is the 4701 * default hstate. More specifically, it is only allowed if 4702 * the number of huge pages for the default hstate was not 4703 * previously specified. 4704 */ 4705 if (!parsed_default_hugepagesz || h != &default_hstate || 4706 default_hstate.max_huge_pages) { 4707 pr_warn("HugeTLB: hugepagesz=%s specified twice, ignoring\n", s); 4708 return 1; 4709 } 4710 4711 /* 4712 * No need to call hugetlb_add_hstate() as hstate already 4713 * exists. But, do set parsed_hstate so that a following 4714 * hugepages= parameter will be applied to this hstate. 4715 */ 4716 parsed_hstate = h; 4717 parsed_valid_hugepagesz = true; 4718 return 1; 4719 } 4720 4721 hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT); 4722 parsed_valid_hugepagesz = true; 4723 return 1; 4724 } 4725 __setup("hugepagesz=", hugepagesz_setup); 4726 4727 /* 4728 * default_hugepagesz command line input 4729 * Only one instance of default_hugepagesz allowed on command line. 4730 */ 4731 static int __init default_hugepagesz_setup(char *s) 4732 { 4733 unsigned long size; 4734 int i; 4735 4736 parsed_valid_hugepagesz = false; 4737 if (parsed_default_hugepagesz) { 4738 pr_err("HugeTLB: default_hugepagesz previously specified, ignoring %s\n", s); 4739 return 1; 4740 } 4741 4742 size = (unsigned long)memparse(s, NULL); 4743 4744 if (!arch_hugetlb_valid_size(size)) { 4745 pr_err("HugeTLB: unsupported default_hugepagesz=%s\n", s); 4746 return 1; 4747 } 4748 4749 hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT); 4750 parsed_valid_hugepagesz = true; 4751 parsed_default_hugepagesz = true; 4752 default_hstate_idx = hstate_index(size_to_hstate(size)); 4753 4754 /* 4755 * The number of default huge pages (for this size) could have been 4756 * specified as the first hugetlb parameter: hugepages=X. If so, 4757 * then default_hstate_max_huge_pages is set. If the default huge 4758 * page size is gigantic (> MAX_PAGE_ORDER), then the pages must be 4759 * allocated here from bootmem allocator. 4760 */ 4761 if (default_hstate_max_huge_pages) { 4762 default_hstate.max_huge_pages = default_hstate_max_huge_pages; 4763 for_each_online_node(i) 4764 default_hstate.max_huge_pages_node[i] = 4765 default_hugepages_in_node[i]; 4766 if (hstate_is_gigantic(&default_hstate)) 4767 hugetlb_hstate_alloc_pages(&default_hstate); 4768 default_hstate_max_huge_pages = 0; 4769 } 4770 4771 return 1; 4772 } 4773 __setup("default_hugepagesz=", default_hugepagesz_setup); 4774 4775 static unsigned int allowed_mems_nr(struct hstate *h) 4776 { 4777 int node; 4778 unsigned int nr = 0; 4779 nodemask_t *mbind_nodemask; 4780 unsigned int *array = h->free_huge_pages_node; 4781 gfp_t gfp_mask = htlb_alloc_mask(h); 4782 4783 mbind_nodemask = policy_mbind_nodemask(gfp_mask); 4784 for_each_node_mask(node, cpuset_current_mems_allowed) { 4785 if (!mbind_nodemask || node_isset(node, *mbind_nodemask)) 4786 nr += array[node]; 4787 } 4788 4789 return nr; 4790 } 4791 4792 #ifdef CONFIG_SYSCTL 4793 static int proc_hugetlb_doulongvec_minmax(const struct ctl_table *table, int write, 4794 void *buffer, size_t *length, 4795 loff_t *ppos, unsigned long *out) 4796 { 4797 struct ctl_table dup_table; 4798 4799 /* 4800 * In order to avoid races with __do_proc_doulongvec_minmax(), we 4801 * can duplicate the @table and alter the duplicate of it. 4802 */ 4803 dup_table = *table; 4804 dup_table.data = out; 4805 4806 return proc_doulongvec_minmax(&dup_table, write, buffer, length, ppos); 4807 } 4808 4809 static int hugetlb_sysctl_handler_common(bool obey_mempolicy, 4810 const struct ctl_table *table, int write, 4811 void *buffer, size_t *length, loff_t *ppos) 4812 { 4813 struct hstate *h = &default_hstate; 4814 unsigned long tmp = h->max_huge_pages; 4815 int ret; 4816 4817 if (!hugepages_supported()) 4818 return -EOPNOTSUPP; 4819 4820 ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos, 4821 &tmp); 4822 if (ret) 4823 goto out; 4824 4825 if (write) 4826 ret = __nr_hugepages_store_common(obey_mempolicy, h, 4827 NUMA_NO_NODE, tmp, *length); 4828 out: 4829 return ret; 4830 } 4831 4832 static int hugetlb_sysctl_handler(const struct ctl_table *table, int write, 4833 void *buffer, size_t *length, loff_t *ppos) 4834 { 4835 4836 return hugetlb_sysctl_handler_common(false, table, write, 4837 buffer, length, ppos); 4838 } 4839 4840 #ifdef CONFIG_NUMA 4841 static int hugetlb_mempolicy_sysctl_handler(const struct ctl_table *table, int write, 4842 void *buffer, size_t *length, loff_t *ppos) 4843 { 4844 return hugetlb_sysctl_handler_common(true, table, write, 4845 buffer, length, ppos); 4846 } 4847 #endif /* CONFIG_NUMA */ 4848 4849 static int hugetlb_overcommit_handler(const struct ctl_table *table, int write, 4850 void *buffer, size_t *length, loff_t *ppos) 4851 { 4852 struct hstate *h = &default_hstate; 4853 unsigned long tmp; 4854 int ret; 4855 4856 if (!hugepages_supported()) 4857 return -EOPNOTSUPP; 4858 4859 tmp = h->nr_overcommit_huge_pages; 4860 4861 if (write && hstate_is_gigantic(h)) 4862 return -EINVAL; 4863 4864 ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos, 4865 &tmp); 4866 if (ret) 4867 goto out; 4868 4869 if (write) { 4870 spin_lock_irq(&hugetlb_lock); 4871 h->nr_overcommit_huge_pages = tmp; 4872 spin_unlock_irq(&hugetlb_lock); 4873 } 4874 out: 4875 return ret; 4876 } 4877 4878 static const struct ctl_table hugetlb_table[] = { 4879 { 4880 .procname = "nr_hugepages", 4881 .data = NULL, 4882 .maxlen = sizeof(unsigned long), 4883 .mode = 0644, 4884 .proc_handler = hugetlb_sysctl_handler, 4885 }, 4886 #ifdef CONFIG_NUMA 4887 { 4888 .procname = "nr_hugepages_mempolicy", 4889 .data = NULL, 4890 .maxlen = sizeof(unsigned long), 4891 .mode = 0644, 4892 .proc_handler = &hugetlb_mempolicy_sysctl_handler, 4893 }, 4894 #endif 4895 { 4896 .procname = "hugetlb_shm_group", 4897 .data = &sysctl_hugetlb_shm_group, 4898 .maxlen = sizeof(gid_t), 4899 .mode = 0644, 4900 .proc_handler = proc_dointvec, 4901 }, 4902 { 4903 .procname = "nr_overcommit_hugepages", 4904 .data = NULL, 4905 .maxlen = sizeof(unsigned long), 4906 .mode = 0644, 4907 .proc_handler = hugetlb_overcommit_handler, 4908 }, 4909 }; 4910 4911 static void hugetlb_sysctl_init(void) 4912 { 4913 register_sysctl_init("vm", hugetlb_table); 4914 } 4915 #endif /* CONFIG_SYSCTL */ 4916 4917 void hugetlb_report_meminfo(struct seq_file *m) 4918 { 4919 struct hstate *h; 4920 unsigned long total = 0; 4921 4922 if (!hugepages_supported()) 4923 return; 4924 4925 for_each_hstate(h) { 4926 unsigned long count = h->nr_huge_pages; 4927 4928 total += huge_page_size(h) * count; 4929 4930 if (h == &default_hstate) 4931 seq_printf(m, 4932 "HugePages_Total: %5lu\n" 4933 "HugePages_Free: %5lu\n" 4934 "HugePages_Rsvd: %5lu\n" 4935 "HugePages_Surp: %5lu\n" 4936 "Hugepagesize: %8lu kB\n", 4937 count, 4938 h->free_huge_pages, 4939 h->resv_huge_pages, 4940 h->surplus_huge_pages, 4941 huge_page_size(h) / SZ_1K); 4942 } 4943 4944 seq_printf(m, "Hugetlb: %8lu kB\n", total / SZ_1K); 4945 } 4946 4947 int hugetlb_report_node_meminfo(char *buf, int len, int nid) 4948 { 4949 struct hstate *h = &default_hstate; 4950 4951 if (!hugepages_supported()) 4952 return 0; 4953 4954 return sysfs_emit_at(buf, len, 4955 "Node %d HugePages_Total: %5u\n" 4956 "Node %d HugePages_Free: %5u\n" 4957 "Node %d HugePages_Surp: %5u\n", 4958 nid, h->nr_huge_pages_node[nid], 4959 nid, h->free_huge_pages_node[nid], 4960 nid, h->surplus_huge_pages_node[nid]); 4961 } 4962 4963 void hugetlb_show_meminfo_node(int nid) 4964 { 4965 struct hstate *h; 4966 4967 if (!hugepages_supported()) 4968 return; 4969 4970 for_each_hstate(h) 4971 printk("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n", 4972 nid, 4973 h->nr_huge_pages_node[nid], 4974 h->free_huge_pages_node[nid], 4975 h->surplus_huge_pages_node[nid], 4976 huge_page_size(h) / SZ_1K); 4977 } 4978 4979 void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm) 4980 { 4981 seq_printf(m, "HugetlbPages:\t%8lu kB\n", 4982 K(atomic_long_read(&mm->hugetlb_usage))); 4983 } 4984 4985 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */ 4986 unsigned long hugetlb_total_pages(void) 4987 { 4988 struct hstate *h; 4989 unsigned long nr_total_pages = 0; 4990 4991 for_each_hstate(h) 4992 nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h); 4993 return nr_total_pages; 4994 } 4995 4996 static int hugetlb_acct_memory(struct hstate *h, long delta) 4997 { 4998 int ret = -ENOMEM; 4999 5000 if (!delta) 5001 return 0; 5002 5003 spin_lock_irq(&hugetlb_lock); 5004 /* 5005 * When cpuset is configured, it breaks the strict hugetlb page 5006 * reservation as the accounting is done on a global variable. Such 5007 * reservation is completely rubbish in the presence of cpuset because 5008 * the reservation is not checked against page availability for the 5009 * current cpuset. Application can still potentially OOM'ed by kernel 5010 * with lack of free htlb page in cpuset that the task is in. 5011 * Attempt to enforce strict accounting with cpuset is almost 5012 * impossible (or too ugly) because cpuset is too fluid that 5013 * task or memory node can be dynamically moved between cpusets. 5014 * 5015 * The change of semantics for shared hugetlb mapping with cpuset is 5016 * undesirable. However, in order to preserve some of the semantics, 5017 * we fall back to check against current free page availability as 5018 * a best attempt and hopefully to minimize the impact of changing 5019 * semantics that cpuset has. 5020 * 5021 * Apart from cpuset, we also have memory policy mechanism that 5022 * also determines from which node the kernel will allocate memory 5023 * in a NUMA system. So similar to cpuset, we also should consider 5024 * the memory policy of the current task. Similar to the description 5025 * above. 5026 */ 5027 if (delta > 0) { 5028 if (gather_surplus_pages(h, delta) < 0) 5029 goto out; 5030 5031 if (delta > allowed_mems_nr(h)) { 5032 return_unused_surplus_pages(h, delta); 5033 goto out; 5034 } 5035 } 5036 5037 ret = 0; 5038 if (delta < 0) 5039 return_unused_surplus_pages(h, (unsigned long) -delta); 5040 5041 out: 5042 spin_unlock_irq(&hugetlb_lock); 5043 return ret; 5044 } 5045 5046 static void hugetlb_vm_op_open(struct vm_area_struct *vma) 5047 { 5048 struct resv_map *resv = vma_resv_map(vma); 5049 5050 /* 5051 * HPAGE_RESV_OWNER indicates a private mapping. 5052 * This new VMA should share its siblings reservation map if present. 5053 * The VMA will only ever have a valid reservation map pointer where 5054 * it is being copied for another still existing VMA. As that VMA 5055 * has a reference to the reservation map it cannot disappear until 5056 * after this open call completes. It is therefore safe to take a 5057 * new reference here without additional locking. 5058 */ 5059 if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { 5060 resv_map_dup_hugetlb_cgroup_uncharge_info(resv); 5061 kref_get(&resv->refs); 5062 } 5063 5064 /* 5065 * vma_lock structure for sharable mappings is vma specific. 5066 * Clear old pointer (if copied via vm_area_dup) and allocate 5067 * new structure. Before clearing, make sure vma_lock is not 5068 * for this vma. 5069 */ 5070 if (vma->vm_flags & VM_MAYSHARE) { 5071 struct hugetlb_vma_lock *vma_lock = vma->vm_private_data; 5072 5073 if (vma_lock) { 5074 if (vma_lock->vma != vma) { 5075 vma->vm_private_data = NULL; 5076 hugetlb_vma_lock_alloc(vma); 5077 } else 5078 pr_warn("HugeTLB: vma_lock already exists in %s.\n", __func__); 5079 } else 5080 hugetlb_vma_lock_alloc(vma); 5081 } 5082 } 5083 5084 static void hugetlb_vm_op_close(struct vm_area_struct *vma) 5085 { 5086 struct hstate *h = hstate_vma(vma); 5087 struct resv_map *resv; 5088 struct hugepage_subpool *spool = subpool_vma(vma); 5089 unsigned long reserve, start, end; 5090 long gbl_reserve; 5091 5092 hugetlb_vma_lock_free(vma); 5093 5094 resv = vma_resv_map(vma); 5095 if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER)) 5096 return; 5097 5098 start = vma_hugecache_offset(h, vma, vma->vm_start); 5099 end = vma_hugecache_offset(h, vma, vma->vm_end); 5100 5101 reserve = (end - start) - region_count(resv, start, end); 5102 hugetlb_cgroup_uncharge_counter(resv, start, end); 5103 if (reserve) { 5104 /* 5105 * Decrement reserve counts. The global reserve count may be 5106 * adjusted if the subpool has a minimum size. 5107 */ 5108 gbl_reserve = hugepage_subpool_put_pages(spool, reserve); 5109 hugetlb_acct_memory(h, -gbl_reserve); 5110 } 5111 5112 kref_put(&resv->refs, resv_map_release); 5113 } 5114 5115 static int hugetlb_vm_op_split(struct vm_area_struct *vma, unsigned long addr) 5116 { 5117 if (addr & ~(huge_page_mask(hstate_vma(vma)))) 5118 return -EINVAL; 5119 5120 /* 5121 * PMD sharing is only possible for PUD_SIZE-aligned address ranges 5122 * in HugeTLB VMAs. If we will lose PUD_SIZE alignment due to this 5123 * split, unshare PMDs in the PUD_SIZE interval surrounding addr now. 5124 */ 5125 if (addr & ~PUD_MASK) { 5126 /* 5127 * hugetlb_vm_op_split is called right before we attempt to 5128 * split the VMA. We will need to unshare PMDs in the old and 5129 * new VMAs, so let's unshare before we split. 5130 */ 5131 unsigned long floor = addr & PUD_MASK; 5132 unsigned long ceil = floor + PUD_SIZE; 5133 5134 if (floor >= vma->vm_start && ceil <= vma->vm_end) 5135 hugetlb_unshare_pmds(vma, floor, ceil); 5136 } 5137 5138 return 0; 5139 } 5140 5141 static unsigned long hugetlb_vm_op_pagesize(struct vm_area_struct *vma) 5142 { 5143 return huge_page_size(hstate_vma(vma)); 5144 } 5145 5146 /* 5147 * We cannot handle pagefaults against hugetlb pages at all. They cause 5148 * handle_mm_fault() to try to instantiate regular-sized pages in the 5149 * hugepage VMA. do_page_fault() is supposed to trap this, so BUG is we get 5150 * this far. 5151 */ 5152 static vm_fault_t hugetlb_vm_op_fault(struct vm_fault *vmf) 5153 { 5154 BUG(); 5155 return 0; 5156 } 5157 5158 /* 5159 * When a new function is introduced to vm_operations_struct and added 5160 * to hugetlb_vm_ops, please consider adding the function to shm_vm_ops. 5161 * This is because under System V memory model, mappings created via 5162 * shmget/shmat with "huge page" specified are backed by hugetlbfs files, 5163 * their original vm_ops are overwritten with shm_vm_ops. 5164 */ 5165 const struct vm_operations_struct hugetlb_vm_ops = { 5166 .fault = hugetlb_vm_op_fault, 5167 .open = hugetlb_vm_op_open, 5168 .close = hugetlb_vm_op_close, 5169 .may_split = hugetlb_vm_op_split, 5170 .pagesize = hugetlb_vm_op_pagesize, 5171 }; 5172 5173 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page, 5174 bool try_mkwrite) 5175 { 5176 pte_t entry; 5177 unsigned int shift = huge_page_shift(hstate_vma(vma)); 5178 5179 if (try_mkwrite && (vma->vm_flags & VM_WRITE)) { 5180 entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page, 5181 vma->vm_page_prot))); 5182 } else { 5183 entry = huge_pte_wrprotect(mk_huge_pte(page, 5184 vma->vm_page_prot)); 5185 } 5186 entry = pte_mkyoung(entry); 5187 entry = arch_make_huge_pte(entry, shift, vma->vm_flags); 5188 5189 return entry; 5190 } 5191 5192 static void set_huge_ptep_writable(struct vm_area_struct *vma, 5193 unsigned long address, pte_t *ptep) 5194 { 5195 pte_t entry; 5196 5197 entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(vma->vm_mm, address, ptep))); 5198 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) 5199 update_mmu_cache(vma, address, ptep); 5200 } 5201 5202 static void set_huge_ptep_maybe_writable(struct vm_area_struct *vma, 5203 unsigned long address, pte_t *ptep) 5204 { 5205 if (vma->vm_flags & VM_WRITE) 5206 set_huge_ptep_writable(vma, address, ptep); 5207 } 5208 5209 bool is_hugetlb_entry_migration(pte_t pte) 5210 { 5211 swp_entry_t swp; 5212 5213 if (huge_pte_none(pte) || pte_present(pte)) 5214 return false; 5215 swp = pte_to_swp_entry(pte); 5216 if (is_migration_entry(swp)) 5217 return true; 5218 else 5219 return false; 5220 } 5221 5222 bool is_hugetlb_entry_hwpoisoned(pte_t pte) 5223 { 5224 swp_entry_t swp; 5225 5226 if (huge_pte_none(pte) || pte_present(pte)) 5227 return false; 5228 swp = pte_to_swp_entry(pte); 5229 if (is_hwpoison_entry(swp)) 5230 return true; 5231 else 5232 return false; 5233 } 5234 5235 static void 5236 hugetlb_install_folio(struct vm_area_struct *vma, pte_t *ptep, unsigned long addr, 5237 struct folio *new_folio, pte_t old, unsigned long sz) 5238 { 5239 pte_t newpte = make_huge_pte(vma, &new_folio->page, true); 5240 5241 __folio_mark_uptodate(new_folio); 5242 hugetlb_add_new_anon_rmap(new_folio, vma, addr); 5243 if (userfaultfd_wp(vma) && huge_pte_uffd_wp(old)) 5244 newpte = huge_pte_mkuffd_wp(newpte); 5245 set_huge_pte_at(vma->vm_mm, addr, ptep, newpte, sz); 5246 hugetlb_count_add(pages_per_huge_page(hstate_vma(vma)), vma->vm_mm); 5247 folio_set_hugetlb_migratable(new_folio); 5248 } 5249 5250 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src, 5251 struct vm_area_struct *dst_vma, 5252 struct vm_area_struct *src_vma) 5253 { 5254 pte_t *src_pte, *dst_pte, entry; 5255 struct folio *pte_folio; 5256 unsigned long addr; 5257 bool cow = is_cow_mapping(src_vma->vm_flags); 5258 struct hstate *h = hstate_vma(src_vma); 5259 unsigned long sz = huge_page_size(h); 5260 unsigned long npages = pages_per_huge_page(h); 5261 struct mmu_notifier_range range; 5262 unsigned long last_addr_mask; 5263 int ret = 0; 5264 5265 if (cow) { 5266 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, src, 5267 src_vma->vm_start, 5268 src_vma->vm_end); 5269 mmu_notifier_invalidate_range_start(&range); 5270 vma_assert_write_locked(src_vma); 5271 raw_write_seqcount_begin(&src->write_protect_seq); 5272 } else { 5273 /* 5274 * For shared mappings the vma lock must be held before 5275 * calling hugetlb_walk() in the src vma. Otherwise, the 5276 * returned ptep could go away if part of a shared pmd and 5277 * another thread calls huge_pmd_unshare. 5278 */ 5279 hugetlb_vma_lock_read(src_vma); 5280 } 5281 5282 last_addr_mask = hugetlb_mask_last_page(h); 5283 for (addr = src_vma->vm_start; addr < src_vma->vm_end; addr += sz) { 5284 spinlock_t *src_ptl, *dst_ptl; 5285 src_pte = hugetlb_walk(src_vma, addr, sz); 5286 if (!src_pte) { 5287 addr |= last_addr_mask; 5288 continue; 5289 } 5290 dst_pte = huge_pte_alloc(dst, dst_vma, addr, sz); 5291 if (!dst_pte) { 5292 ret = -ENOMEM; 5293 break; 5294 } 5295 5296 /* 5297 * If the pagetables are shared don't copy or take references. 5298 * 5299 * dst_pte == src_pte is the common case of src/dest sharing. 5300 * However, src could have 'unshared' and dst shares with 5301 * another vma. So page_count of ptep page is checked instead 5302 * to reliably determine whether pte is shared. 5303 */ 5304 if (page_count(virt_to_page(dst_pte)) > 1) { 5305 addr |= last_addr_mask; 5306 continue; 5307 } 5308 5309 dst_ptl = huge_pte_lock(h, dst, dst_pte); 5310 src_ptl = huge_pte_lockptr(h, src, src_pte); 5311 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); 5312 entry = huge_ptep_get(src_vma->vm_mm, addr, src_pte); 5313 again: 5314 if (huge_pte_none(entry)) { 5315 /* 5316 * Skip if src entry none. 5317 */ 5318 ; 5319 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry))) { 5320 if (!userfaultfd_wp(dst_vma)) 5321 entry = huge_pte_clear_uffd_wp(entry); 5322 set_huge_pte_at(dst, addr, dst_pte, entry, sz); 5323 } else if (unlikely(is_hugetlb_entry_migration(entry))) { 5324 swp_entry_t swp_entry = pte_to_swp_entry(entry); 5325 bool uffd_wp = pte_swp_uffd_wp(entry); 5326 5327 if (!is_readable_migration_entry(swp_entry) && cow) { 5328 /* 5329 * COW mappings require pages in both 5330 * parent and child to be set to read. 5331 */ 5332 swp_entry = make_readable_migration_entry( 5333 swp_offset(swp_entry)); 5334 entry = swp_entry_to_pte(swp_entry); 5335 if (userfaultfd_wp(src_vma) && uffd_wp) 5336 entry = pte_swp_mkuffd_wp(entry); 5337 set_huge_pte_at(src, addr, src_pte, entry, sz); 5338 } 5339 if (!userfaultfd_wp(dst_vma)) 5340 entry = huge_pte_clear_uffd_wp(entry); 5341 set_huge_pte_at(dst, addr, dst_pte, entry, sz); 5342 } else if (unlikely(is_pte_marker(entry))) { 5343 pte_marker marker = copy_pte_marker( 5344 pte_to_swp_entry(entry), dst_vma); 5345 5346 if (marker) 5347 set_huge_pte_at(dst, addr, dst_pte, 5348 make_pte_marker(marker), sz); 5349 } else { 5350 entry = huge_ptep_get(src_vma->vm_mm, addr, src_pte); 5351 pte_folio = page_folio(pte_page(entry)); 5352 folio_get(pte_folio); 5353 5354 /* 5355 * Failing to duplicate the anon rmap is a rare case 5356 * where we see pinned hugetlb pages while they're 5357 * prone to COW. We need to do the COW earlier during 5358 * fork. 5359 * 5360 * When pre-allocating the page or copying data, we 5361 * need to be without the pgtable locks since we could 5362 * sleep during the process. 5363 */ 5364 if (!folio_test_anon(pte_folio)) { 5365 hugetlb_add_file_rmap(pte_folio); 5366 } else if (hugetlb_try_dup_anon_rmap(pte_folio, src_vma)) { 5367 pte_t src_pte_old = entry; 5368 struct folio *new_folio; 5369 5370 spin_unlock(src_ptl); 5371 spin_unlock(dst_ptl); 5372 /* Do not use reserve as it's private owned */ 5373 new_folio = alloc_hugetlb_folio(dst_vma, addr, false); 5374 if (IS_ERR(new_folio)) { 5375 folio_put(pte_folio); 5376 ret = PTR_ERR(new_folio); 5377 break; 5378 } 5379 ret = copy_user_large_folio(new_folio, pte_folio, 5380 addr, dst_vma); 5381 folio_put(pte_folio); 5382 if (ret) { 5383 folio_put(new_folio); 5384 break; 5385 } 5386 5387 /* Install the new hugetlb folio if src pte stable */ 5388 dst_ptl = huge_pte_lock(h, dst, dst_pte); 5389 src_ptl = huge_pte_lockptr(h, src, src_pte); 5390 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); 5391 entry = huge_ptep_get(src_vma->vm_mm, addr, src_pte); 5392 if (!pte_same(src_pte_old, entry)) { 5393 restore_reserve_on_error(h, dst_vma, addr, 5394 new_folio); 5395 folio_put(new_folio); 5396 /* huge_ptep of dst_pte won't change as in child */ 5397 goto again; 5398 } 5399 hugetlb_install_folio(dst_vma, dst_pte, addr, 5400 new_folio, src_pte_old, sz); 5401 spin_unlock(src_ptl); 5402 spin_unlock(dst_ptl); 5403 continue; 5404 } 5405 5406 if (cow) { 5407 /* 5408 * No need to notify as we are downgrading page 5409 * table protection not changing it to point 5410 * to a new page. 5411 * 5412 * See Documentation/mm/mmu_notifier.rst 5413 */ 5414 huge_ptep_set_wrprotect(src, addr, src_pte); 5415 entry = huge_pte_wrprotect(entry); 5416 } 5417 5418 if (!userfaultfd_wp(dst_vma)) 5419 entry = huge_pte_clear_uffd_wp(entry); 5420 5421 set_huge_pte_at(dst, addr, dst_pte, entry, sz); 5422 hugetlb_count_add(npages, dst); 5423 } 5424 spin_unlock(src_ptl); 5425 spin_unlock(dst_ptl); 5426 } 5427 5428 if (cow) { 5429 raw_write_seqcount_end(&src->write_protect_seq); 5430 mmu_notifier_invalidate_range_end(&range); 5431 } else { 5432 hugetlb_vma_unlock_read(src_vma); 5433 } 5434 5435 return ret; 5436 } 5437 5438 static void move_huge_pte(struct vm_area_struct *vma, unsigned long old_addr, 5439 unsigned long new_addr, pte_t *src_pte, pte_t *dst_pte, 5440 unsigned long sz) 5441 { 5442 bool need_clear_uffd_wp = vma_has_uffd_without_event_remap(vma); 5443 struct hstate *h = hstate_vma(vma); 5444 struct mm_struct *mm = vma->vm_mm; 5445 spinlock_t *src_ptl, *dst_ptl; 5446 pte_t pte; 5447 5448 dst_ptl = huge_pte_lock(h, mm, dst_pte); 5449 src_ptl = huge_pte_lockptr(h, mm, src_pte); 5450 5451 /* 5452 * We don't have to worry about the ordering of src and dst ptlocks 5453 * because exclusive mmap_lock (or the i_mmap_lock) prevents deadlock. 5454 */ 5455 if (src_ptl != dst_ptl) 5456 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); 5457 5458 pte = huge_ptep_get_and_clear(mm, old_addr, src_pte, sz); 5459 5460 if (need_clear_uffd_wp && pte_marker_uffd_wp(pte)) 5461 huge_pte_clear(mm, new_addr, dst_pte, sz); 5462 else { 5463 if (need_clear_uffd_wp) { 5464 if (pte_present(pte)) 5465 pte = huge_pte_clear_uffd_wp(pte); 5466 else if (is_swap_pte(pte)) 5467 pte = pte_swp_clear_uffd_wp(pte); 5468 } 5469 set_huge_pte_at(mm, new_addr, dst_pte, pte, sz); 5470 } 5471 5472 if (src_ptl != dst_ptl) 5473 spin_unlock(src_ptl); 5474 spin_unlock(dst_ptl); 5475 } 5476 5477 int move_hugetlb_page_tables(struct vm_area_struct *vma, 5478 struct vm_area_struct *new_vma, 5479 unsigned long old_addr, unsigned long new_addr, 5480 unsigned long len) 5481 { 5482 struct hstate *h = hstate_vma(vma); 5483 struct address_space *mapping = vma->vm_file->f_mapping; 5484 unsigned long sz = huge_page_size(h); 5485 struct mm_struct *mm = vma->vm_mm; 5486 unsigned long old_end = old_addr + len; 5487 unsigned long last_addr_mask; 5488 pte_t *src_pte, *dst_pte; 5489 struct mmu_notifier_range range; 5490 bool shared_pmd = false; 5491 5492 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, mm, old_addr, 5493 old_end); 5494 adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end); 5495 /* 5496 * In case of shared PMDs, we should cover the maximum possible 5497 * range. 5498 */ 5499 flush_cache_range(vma, range.start, range.end); 5500 5501 mmu_notifier_invalidate_range_start(&range); 5502 last_addr_mask = hugetlb_mask_last_page(h); 5503 /* Prevent race with file truncation */ 5504 hugetlb_vma_lock_write(vma); 5505 i_mmap_lock_write(mapping); 5506 for (; old_addr < old_end; old_addr += sz, new_addr += sz) { 5507 src_pte = hugetlb_walk(vma, old_addr, sz); 5508 if (!src_pte) { 5509 old_addr |= last_addr_mask; 5510 new_addr |= last_addr_mask; 5511 continue; 5512 } 5513 if (huge_pte_none(huge_ptep_get(mm, old_addr, src_pte))) 5514 continue; 5515 5516 if (huge_pmd_unshare(mm, vma, old_addr, src_pte)) { 5517 shared_pmd = true; 5518 old_addr |= last_addr_mask; 5519 new_addr |= last_addr_mask; 5520 continue; 5521 } 5522 5523 dst_pte = huge_pte_alloc(mm, new_vma, new_addr, sz); 5524 if (!dst_pte) 5525 break; 5526 5527 move_huge_pte(vma, old_addr, new_addr, src_pte, dst_pte, sz); 5528 } 5529 5530 if (shared_pmd) 5531 flush_hugetlb_tlb_range(vma, range.start, range.end); 5532 else 5533 flush_hugetlb_tlb_range(vma, old_end - len, old_end); 5534 mmu_notifier_invalidate_range_end(&range); 5535 i_mmap_unlock_write(mapping); 5536 hugetlb_vma_unlock_write(vma); 5537 5538 return len + old_addr - old_end; 5539 } 5540 5541 void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma, 5542 unsigned long start, unsigned long end, 5543 struct page *ref_page, zap_flags_t zap_flags) 5544 { 5545 struct mm_struct *mm = vma->vm_mm; 5546 unsigned long address; 5547 pte_t *ptep; 5548 pte_t pte; 5549 spinlock_t *ptl; 5550 struct page *page; 5551 struct hstate *h = hstate_vma(vma); 5552 unsigned long sz = huge_page_size(h); 5553 bool adjust_reservation = false; 5554 unsigned long last_addr_mask; 5555 bool force_flush = false; 5556 5557 WARN_ON(!is_vm_hugetlb_page(vma)); 5558 BUG_ON(start & ~huge_page_mask(h)); 5559 BUG_ON(end & ~huge_page_mask(h)); 5560 5561 /* 5562 * This is a hugetlb vma, all the pte entries should point 5563 * to huge page. 5564 */ 5565 tlb_change_page_size(tlb, sz); 5566 tlb_start_vma(tlb, vma); 5567 5568 last_addr_mask = hugetlb_mask_last_page(h); 5569 address = start; 5570 for (; address < end; address += sz) { 5571 ptep = hugetlb_walk(vma, address, sz); 5572 if (!ptep) { 5573 address |= last_addr_mask; 5574 continue; 5575 } 5576 5577 ptl = huge_pte_lock(h, mm, ptep); 5578 if (huge_pmd_unshare(mm, vma, address, ptep)) { 5579 spin_unlock(ptl); 5580 tlb_flush_pmd_range(tlb, address & PUD_MASK, PUD_SIZE); 5581 force_flush = true; 5582 address |= last_addr_mask; 5583 continue; 5584 } 5585 5586 pte = huge_ptep_get(mm, address, ptep); 5587 if (huge_pte_none(pte)) { 5588 spin_unlock(ptl); 5589 continue; 5590 } 5591 5592 /* 5593 * Migrating hugepage or HWPoisoned hugepage is already 5594 * unmapped and its refcount is dropped, so just clear pte here. 5595 */ 5596 if (unlikely(!pte_present(pte))) { 5597 /* 5598 * If the pte was wr-protected by uffd-wp in any of the 5599 * swap forms, meanwhile the caller does not want to 5600 * drop the uffd-wp bit in this zap, then replace the 5601 * pte with a marker. 5602 */ 5603 if (pte_swp_uffd_wp_any(pte) && 5604 !(zap_flags & ZAP_FLAG_DROP_MARKER)) 5605 set_huge_pte_at(mm, address, ptep, 5606 make_pte_marker(PTE_MARKER_UFFD_WP), 5607 sz); 5608 else 5609 huge_pte_clear(mm, address, ptep, sz); 5610 spin_unlock(ptl); 5611 continue; 5612 } 5613 5614 page = pte_page(pte); 5615 /* 5616 * If a reference page is supplied, it is because a specific 5617 * page is being unmapped, not a range. Ensure the page we 5618 * are about to unmap is the actual page of interest. 5619 */ 5620 if (ref_page) { 5621 if (page != ref_page) { 5622 spin_unlock(ptl); 5623 continue; 5624 } 5625 /* 5626 * Mark the VMA as having unmapped its page so that 5627 * future faults in this VMA will fail rather than 5628 * looking like data was lost 5629 */ 5630 set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED); 5631 } 5632 5633 pte = huge_ptep_get_and_clear(mm, address, ptep, sz); 5634 tlb_remove_huge_tlb_entry(h, tlb, ptep, address); 5635 if (huge_pte_dirty(pte)) 5636 set_page_dirty(page); 5637 /* Leave a uffd-wp pte marker if needed */ 5638 if (huge_pte_uffd_wp(pte) && 5639 !(zap_flags & ZAP_FLAG_DROP_MARKER)) 5640 set_huge_pte_at(mm, address, ptep, 5641 make_pte_marker(PTE_MARKER_UFFD_WP), 5642 sz); 5643 hugetlb_count_sub(pages_per_huge_page(h), mm); 5644 hugetlb_remove_rmap(page_folio(page)); 5645 5646 /* 5647 * Restore the reservation for anonymous page, otherwise the 5648 * backing page could be stolen by someone. 5649 * If there we are freeing a surplus, do not set the restore 5650 * reservation bit. 5651 */ 5652 if (!h->surplus_huge_pages && __vma_private_lock(vma) && 5653 folio_test_anon(page_folio(page))) { 5654 folio_set_hugetlb_restore_reserve(page_folio(page)); 5655 /* Reservation to be adjusted after the spin lock */ 5656 adjust_reservation = true; 5657 } 5658 5659 spin_unlock(ptl); 5660 5661 /* 5662 * Adjust the reservation for the region that will have the 5663 * reserve restored. Keep in mind that vma_needs_reservation() changes 5664 * resv->adds_in_progress if it succeeds. If this is not done, 5665 * do_exit() will not see it, and will keep the reservation 5666 * forever. 5667 */ 5668 if (adjust_reservation) { 5669 int rc = vma_needs_reservation(h, vma, address); 5670 5671 if (rc < 0) 5672 /* Pressumably allocate_file_region_entries failed 5673 * to allocate a file_region struct. Clear 5674 * hugetlb_restore_reserve so that global reserve 5675 * count will not be incremented by free_huge_folio. 5676 * Act as if we consumed the reservation. 5677 */ 5678 folio_clear_hugetlb_restore_reserve(page_folio(page)); 5679 else if (rc) 5680 vma_add_reservation(h, vma, address); 5681 } 5682 5683 tlb_remove_page_size(tlb, page, huge_page_size(h)); 5684 /* 5685 * Bail out after unmapping reference page if supplied 5686 */ 5687 if (ref_page) 5688 break; 5689 } 5690 tlb_end_vma(tlb, vma); 5691 5692 /* 5693 * If we unshared PMDs, the TLB flush was not recorded in mmu_gather. We 5694 * could defer the flush until now, since by holding i_mmap_rwsem we 5695 * guaranteed that the last refernece would not be dropped. But we must 5696 * do the flushing before we return, as otherwise i_mmap_rwsem will be 5697 * dropped and the last reference to the shared PMDs page might be 5698 * dropped as well. 5699 * 5700 * In theory we could defer the freeing of the PMD pages as well, but 5701 * huge_pmd_unshare() relies on the exact page_count for the PMD page to 5702 * detect sharing, so we cannot defer the release of the page either. 5703 * Instead, do flush now. 5704 */ 5705 if (force_flush) 5706 tlb_flush_mmu_tlbonly(tlb); 5707 } 5708 5709 void __hugetlb_zap_begin(struct vm_area_struct *vma, 5710 unsigned long *start, unsigned long *end) 5711 { 5712 if (!vma->vm_file) /* hugetlbfs_file_mmap error */ 5713 return; 5714 5715 adjust_range_if_pmd_sharing_possible(vma, start, end); 5716 hugetlb_vma_lock_write(vma); 5717 if (vma->vm_file) 5718 i_mmap_lock_write(vma->vm_file->f_mapping); 5719 } 5720 5721 void __hugetlb_zap_end(struct vm_area_struct *vma, 5722 struct zap_details *details) 5723 { 5724 zap_flags_t zap_flags = details ? details->zap_flags : 0; 5725 5726 if (!vma->vm_file) /* hugetlbfs_file_mmap error */ 5727 return; 5728 5729 if (zap_flags & ZAP_FLAG_UNMAP) { /* final unmap */ 5730 /* 5731 * Unlock and free the vma lock before releasing i_mmap_rwsem. 5732 * When the vma_lock is freed, this makes the vma ineligible 5733 * for pmd sharing. And, i_mmap_rwsem is required to set up 5734 * pmd sharing. This is important as page tables for this 5735 * unmapped range will be asynchrously deleted. If the page 5736 * tables are shared, there will be issues when accessed by 5737 * someone else. 5738 */ 5739 __hugetlb_vma_unlock_write_free(vma); 5740 } else { 5741 hugetlb_vma_unlock_write(vma); 5742 } 5743 5744 if (vma->vm_file) 5745 i_mmap_unlock_write(vma->vm_file->f_mapping); 5746 } 5747 5748 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start, 5749 unsigned long end, struct page *ref_page, 5750 zap_flags_t zap_flags) 5751 { 5752 struct mmu_notifier_range range; 5753 struct mmu_gather tlb; 5754 5755 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma->vm_mm, 5756 start, end); 5757 adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end); 5758 mmu_notifier_invalidate_range_start(&range); 5759 tlb_gather_mmu(&tlb, vma->vm_mm); 5760 5761 __unmap_hugepage_range(&tlb, vma, start, end, ref_page, zap_flags); 5762 5763 mmu_notifier_invalidate_range_end(&range); 5764 tlb_finish_mmu(&tlb); 5765 } 5766 5767 /* 5768 * This is called when the original mapper is failing to COW a MAP_PRIVATE 5769 * mapping it owns the reserve page for. The intention is to unmap the page 5770 * from other VMAs and let the children be SIGKILLed if they are faulting the 5771 * same region. 5772 */ 5773 static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma, 5774 struct page *page, unsigned long address) 5775 { 5776 struct hstate *h = hstate_vma(vma); 5777 struct vm_area_struct *iter_vma; 5778 struct address_space *mapping; 5779 pgoff_t pgoff; 5780 5781 /* 5782 * vm_pgoff is in PAGE_SIZE units, hence the different calculation 5783 * from page cache lookup which is in HPAGE_SIZE units. 5784 */ 5785 address = address & huge_page_mask(h); 5786 pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) + 5787 vma->vm_pgoff; 5788 mapping = vma->vm_file->f_mapping; 5789 5790 /* 5791 * Take the mapping lock for the duration of the table walk. As 5792 * this mapping should be shared between all the VMAs, 5793 * __unmap_hugepage_range() is called as the lock is already held 5794 */ 5795 i_mmap_lock_write(mapping); 5796 vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) { 5797 /* Do not unmap the current VMA */ 5798 if (iter_vma == vma) 5799 continue; 5800 5801 /* 5802 * Shared VMAs have their own reserves and do not affect 5803 * MAP_PRIVATE accounting but it is possible that a shared 5804 * VMA is using the same page so check and skip such VMAs. 5805 */ 5806 if (iter_vma->vm_flags & VM_MAYSHARE) 5807 continue; 5808 5809 /* 5810 * Unmap the page from other VMAs without their own reserves. 5811 * They get marked to be SIGKILLed if they fault in these 5812 * areas. This is because a future no-page fault on this VMA 5813 * could insert a zeroed page instead of the data existing 5814 * from the time of fork. This would look like data corruption 5815 */ 5816 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER)) 5817 unmap_hugepage_range(iter_vma, address, 5818 address + huge_page_size(h), page, 0); 5819 } 5820 i_mmap_unlock_write(mapping); 5821 } 5822 5823 /* 5824 * hugetlb_wp() should be called with page lock of the original hugepage held. 5825 * Called with hugetlb_fault_mutex_table held and pte_page locked so we 5826 * cannot race with other handlers or page migration. 5827 * Keep the pte_same checks anyway to make transition from the mutex easier. 5828 */ 5829 static vm_fault_t hugetlb_wp(struct folio *pagecache_folio, 5830 struct vm_fault *vmf) 5831 { 5832 struct vm_area_struct *vma = vmf->vma; 5833 struct mm_struct *mm = vma->vm_mm; 5834 const bool unshare = vmf->flags & FAULT_FLAG_UNSHARE; 5835 pte_t pte = huge_ptep_get(mm, vmf->address, vmf->pte); 5836 struct hstate *h = hstate_vma(vma); 5837 struct folio *old_folio; 5838 struct folio *new_folio; 5839 bool cow_from_owner = 0; 5840 vm_fault_t ret = 0; 5841 struct mmu_notifier_range range; 5842 5843 /* 5844 * Never handle CoW for uffd-wp protected pages. It should be only 5845 * handled when the uffd-wp protection is removed. 5846 * 5847 * Note that only the CoW optimization path (in hugetlb_no_page()) 5848 * can trigger this, because hugetlb_fault() will always resolve 5849 * uffd-wp bit first. 5850 */ 5851 if (!unshare && huge_pte_uffd_wp(pte)) 5852 return 0; 5853 5854 /* Let's take out MAP_SHARED mappings first. */ 5855 if (vma->vm_flags & VM_MAYSHARE) { 5856 set_huge_ptep_writable(vma, vmf->address, vmf->pte); 5857 return 0; 5858 } 5859 5860 old_folio = page_folio(pte_page(pte)); 5861 5862 delayacct_wpcopy_start(); 5863 5864 retry_avoidcopy: 5865 /* 5866 * If no-one else is actually using this page, we're the exclusive 5867 * owner and can reuse this page. 5868 * 5869 * Note that we don't rely on the (safer) folio refcount here, because 5870 * copying the hugetlb folio when there are unexpected (temporary) 5871 * folio references could harm simple fork()+exit() users when 5872 * we run out of free hugetlb folios: we would have to kill processes 5873 * in scenarios that used to work. As a side effect, there can still 5874 * be leaks between processes, for example, with FOLL_GET users. 5875 */ 5876 if (folio_mapcount(old_folio) == 1 && folio_test_anon(old_folio)) { 5877 if (!PageAnonExclusive(&old_folio->page)) { 5878 folio_move_anon_rmap(old_folio, vma); 5879 SetPageAnonExclusive(&old_folio->page); 5880 } 5881 if (likely(!unshare)) 5882 set_huge_ptep_maybe_writable(vma, vmf->address, 5883 vmf->pte); 5884 5885 delayacct_wpcopy_end(); 5886 return 0; 5887 } 5888 VM_BUG_ON_PAGE(folio_test_anon(old_folio) && 5889 PageAnonExclusive(&old_folio->page), &old_folio->page); 5890 5891 /* 5892 * If the process that created a MAP_PRIVATE mapping is about to 5893 * perform a COW due to a shared page count, attempt to satisfy 5894 * the allocation without using the existing reserves. The pagecache 5895 * page is used to determine if the reserve at this address was 5896 * consumed or not. If reserves were used, a partial faulted mapping 5897 * at the time of fork() could consume its reserves on COW instead 5898 * of the full address range. 5899 */ 5900 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && 5901 old_folio != pagecache_folio) 5902 cow_from_owner = true; 5903 5904 folio_get(old_folio); 5905 5906 /* 5907 * Drop page table lock as buddy allocator may be called. It will 5908 * be acquired again before returning to the caller, as expected. 5909 */ 5910 spin_unlock(vmf->ptl); 5911 new_folio = alloc_hugetlb_folio(vma, vmf->address, cow_from_owner); 5912 5913 if (IS_ERR(new_folio)) { 5914 /* 5915 * If a process owning a MAP_PRIVATE mapping fails to COW, 5916 * it is due to references held by a child and an insufficient 5917 * huge page pool. To guarantee the original mappers 5918 * reliability, unmap the page from child processes. The child 5919 * may get SIGKILLed if it later faults. 5920 */ 5921 if (cow_from_owner) { 5922 struct address_space *mapping = vma->vm_file->f_mapping; 5923 pgoff_t idx; 5924 u32 hash; 5925 5926 folio_put(old_folio); 5927 /* 5928 * Drop hugetlb_fault_mutex and vma_lock before 5929 * unmapping. unmapping needs to hold vma_lock 5930 * in write mode. Dropping vma_lock in read mode 5931 * here is OK as COW mappings do not interact with 5932 * PMD sharing. 5933 * 5934 * Reacquire both after unmap operation. 5935 */ 5936 idx = vma_hugecache_offset(h, vma, vmf->address); 5937 hash = hugetlb_fault_mutex_hash(mapping, idx); 5938 hugetlb_vma_unlock_read(vma); 5939 mutex_unlock(&hugetlb_fault_mutex_table[hash]); 5940 5941 unmap_ref_private(mm, vma, &old_folio->page, 5942 vmf->address); 5943 5944 mutex_lock(&hugetlb_fault_mutex_table[hash]); 5945 hugetlb_vma_lock_read(vma); 5946 spin_lock(vmf->ptl); 5947 vmf->pte = hugetlb_walk(vma, vmf->address, 5948 huge_page_size(h)); 5949 if (likely(vmf->pte && 5950 pte_same(huge_ptep_get(mm, vmf->address, vmf->pte), pte))) 5951 goto retry_avoidcopy; 5952 /* 5953 * race occurs while re-acquiring page table 5954 * lock, and our job is done. 5955 */ 5956 delayacct_wpcopy_end(); 5957 return 0; 5958 } 5959 5960 ret = vmf_error(PTR_ERR(new_folio)); 5961 goto out_release_old; 5962 } 5963 5964 /* 5965 * When the original hugepage is shared one, it does not have 5966 * anon_vma prepared. 5967 */ 5968 ret = __vmf_anon_prepare(vmf); 5969 if (unlikely(ret)) 5970 goto out_release_all; 5971 5972 if (copy_user_large_folio(new_folio, old_folio, vmf->real_address, vma)) { 5973 ret = VM_FAULT_HWPOISON_LARGE | VM_FAULT_SET_HINDEX(hstate_index(h)); 5974 goto out_release_all; 5975 } 5976 __folio_mark_uptodate(new_folio); 5977 5978 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, mm, vmf->address, 5979 vmf->address + huge_page_size(h)); 5980 mmu_notifier_invalidate_range_start(&range); 5981 5982 /* 5983 * Retake the page table lock to check for racing updates 5984 * before the page tables are altered 5985 */ 5986 spin_lock(vmf->ptl); 5987 vmf->pte = hugetlb_walk(vma, vmf->address, huge_page_size(h)); 5988 if (likely(vmf->pte && pte_same(huge_ptep_get(mm, vmf->address, vmf->pte), pte))) { 5989 pte_t newpte = make_huge_pte(vma, &new_folio->page, !unshare); 5990 5991 /* Break COW or unshare */ 5992 huge_ptep_clear_flush(vma, vmf->address, vmf->pte); 5993 hugetlb_remove_rmap(old_folio); 5994 hugetlb_add_new_anon_rmap(new_folio, vma, vmf->address); 5995 if (huge_pte_uffd_wp(pte)) 5996 newpte = huge_pte_mkuffd_wp(newpte); 5997 set_huge_pte_at(mm, vmf->address, vmf->pte, newpte, 5998 huge_page_size(h)); 5999 folio_set_hugetlb_migratable(new_folio); 6000 /* Make the old page be freed below */ 6001 new_folio = old_folio; 6002 } 6003 spin_unlock(vmf->ptl); 6004 mmu_notifier_invalidate_range_end(&range); 6005 out_release_all: 6006 /* 6007 * No restore in case of successful pagetable update (Break COW or 6008 * unshare) 6009 */ 6010 if (new_folio != old_folio) 6011 restore_reserve_on_error(h, vma, vmf->address, new_folio); 6012 folio_put(new_folio); 6013 out_release_old: 6014 folio_put(old_folio); 6015 6016 spin_lock(vmf->ptl); /* Caller expects lock to be held */ 6017 6018 delayacct_wpcopy_end(); 6019 return ret; 6020 } 6021 6022 /* 6023 * Return whether there is a pagecache page to back given address within VMA. 6024 */ 6025 bool hugetlbfs_pagecache_present(struct hstate *h, 6026 struct vm_area_struct *vma, unsigned long address) 6027 { 6028 struct address_space *mapping = vma->vm_file->f_mapping; 6029 pgoff_t idx = linear_page_index(vma, address); 6030 struct folio *folio; 6031 6032 folio = filemap_get_folio(mapping, idx); 6033 if (IS_ERR(folio)) 6034 return false; 6035 folio_put(folio); 6036 return true; 6037 } 6038 6039 int hugetlb_add_to_page_cache(struct folio *folio, struct address_space *mapping, 6040 pgoff_t idx) 6041 { 6042 struct inode *inode = mapping->host; 6043 struct hstate *h = hstate_inode(inode); 6044 int err; 6045 6046 idx <<= huge_page_order(h); 6047 __folio_set_locked(folio); 6048 err = __filemap_add_folio(mapping, folio, idx, GFP_KERNEL, NULL); 6049 6050 if (unlikely(err)) { 6051 __folio_clear_locked(folio); 6052 return err; 6053 } 6054 folio_clear_hugetlb_restore_reserve(folio); 6055 6056 /* 6057 * mark folio dirty so that it will not be removed from cache/file 6058 * by non-hugetlbfs specific code paths. 6059 */ 6060 folio_mark_dirty(folio); 6061 6062 spin_lock(&inode->i_lock); 6063 inode->i_blocks += blocks_per_huge_page(h); 6064 spin_unlock(&inode->i_lock); 6065 return 0; 6066 } 6067 6068 static inline vm_fault_t hugetlb_handle_userfault(struct vm_fault *vmf, 6069 struct address_space *mapping, 6070 unsigned long reason) 6071 { 6072 u32 hash; 6073 6074 /* 6075 * vma_lock and hugetlb_fault_mutex must be dropped before handling 6076 * userfault. Also mmap_lock could be dropped due to handling 6077 * userfault, any vma operation should be careful from here. 6078 */ 6079 hugetlb_vma_unlock_read(vmf->vma); 6080 hash = hugetlb_fault_mutex_hash(mapping, vmf->pgoff); 6081 mutex_unlock(&hugetlb_fault_mutex_table[hash]); 6082 return handle_userfault(vmf, reason); 6083 } 6084 6085 /* 6086 * Recheck pte with pgtable lock. Returns true if pte didn't change, or 6087 * false if pte changed or is changing. 6088 */ 6089 static bool hugetlb_pte_stable(struct hstate *h, struct mm_struct *mm, unsigned long addr, 6090 pte_t *ptep, pte_t old_pte) 6091 { 6092 spinlock_t *ptl; 6093 bool same; 6094 6095 ptl = huge_pte_lock(h, mm, ptep); 6096 same = pte_same(huge_ptep_get(mm, addr, ptep), old_pte); 6097 spin_unlock(ptl); 6098 6099 return same; 6100 } 6101 6102 static vm_fault_t hugetlb_no_page(struct address_space *mapping, 6103 struct vm_fault *vmf) 6104 { 6105 struct vm_area_struct *vma = vmf->vma; 6106 struct mm_struct *mm = vma->vm_mm; 6107 struct hstate *h = hstate_vma(vma); 6108 vm_fault_t ret = VM_FAULT_SIGBUS; 6109 int anon_rmap = 0; 6110 unsigned long size; 6111 struct folio *folio; 6112 pte_t new_pte; 6113 bool new_folio, new_pagecache_folio = false; 6114 u32 hash = hugetlb_fault_mutex_hash(mapping, vmf->pgoff); 6115 6116 /* 6117 * Currently, we are forced to kill the process in the event the 6118 * original mapper has unmapped pages from the child due to a failed 6119 * COW/unsharing. Warn that such a situation has occurred as it may not 6120 * be obvious. 6121 */ 6122 if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) { 6123 pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n", 6124 current->pid); 6125 goto out; 6126 } 6127 6128 /* 6129 * Use page lock to guard against racing truncation 6130 * before we get page_table_lock. 6131 */ 6132 new_folio = false; 6133 folio = filemap_lock_hugetlb_folio(h, mapping, vmf->pgoff); 6134 if (IS_ERR(folio)) { 6135 size = i_size_read(mapping->host) >> huge_page_shift(h); 6136 if (vmf->pgoff >= size) 6137 goto out; 6138 /* Check for page in userfault range */ 6139 if (userfaultfd_missing(vma)) { 6140 /* 6141 * Since hugetlb_no_page() was examining pte 6142 * without pgtable lock, we need to re-test under 6143 * lock because the pte may not be stable and could 6144 * have changed from under us. Try to detect 6145 * either changed or during-changing ptes and retry 6146 * properly when needed. 6147 * 6148 * Note that userfaultfd is actually fine with 6149 * false positives (e.g. caused by pte changed), 6150 * but not wrong logical events (e.g. caused by 6151 * reading a pte during changing). The latter can 6152 * confuse the userspace, so the strictness is very 6153 * much preferred. E.g., MISSING event should 6154 * never happen on the page after UFFDIO_COPY has 6155 * correctly installed the page and returned. 6156 */ 6157 if (!hugetlb_pte_stable(h, mm, vmf->address, vmf->pte, vmf->orig_pte)) { 6158 ret = 0; 6159 goto out; 6160 } 6161 6162 return hugetlb_handle_userfault(vmf, mapping, 6163 VM_UFFD_MISSING); 6164 } 6165 6166 if (!(vma->vm_flags & VM_MAYSHARE)) { 6167 ret = __vmf_anon_prepare(vmf); 6168 if (unlikely(ret)) 6169 goto out; 6170 } 6171 6172 folio = alloc_hugetlb_folio(vma, vmf->address, false); 6173 if (IS_ERR(folio)) { 6174 /* 6175 * Returning error will result in faulting task being 6176 * sent SIGBUS. The hugetlb fault mutex prevents two 6177 * tasks from racing to fault in the same page which 6178 * could result in false unable to allocate errors. 6179 * Page migration does not take the fault mutex, but 6180 * does a clear then write of pte's under page table 6181 * lock. Page fault code could race with migration, 6182 * notice the clear pte and try to allocate a page 6183 * here. Before returning error, get ptl and make 6184 * sure there really is no pte entry. 6185 */ 6186 if (hugetlb_pte_stable(h, mm, vmf->address, vmf->pte, vmf->orig_pte)) 6187 ret = vmf_error(PTR_ERR(folio)); 6188 else 6189 ret = 0; 6190 goto out; 6191 } 6192 folio_zero_user(folio, vmf->real_address); 6193 __folio_mark_uptodate(folio); 6194 new_folio = true; 6195 6196 if (vma->vm_flags & VM_MAYSHARE) { 6197 int err = hugetlb_add_to_page_cache(folio, mapping, 6198 vmf->pgoff); 6199 if (err) { 6200 /* 6201 * err can't be -EEXIST which implies someone 6202 * else consumed the reservation since hugetlb 6203 * fault mutex is held when add a hugetlb page 6204 * to the page cache. So it's safe to call 6205 * restore_reserve_on_error() here. 6206 */ 6207 restore_reserve_on_error(h, vma, vmf->address, 6208 folio); 6209 folio_put(folio); 6210 ret = VM_FAULT_SIGBUS; 6211 goto out; 6212 } 6213 new_pagecache_folio = true; 6214 } else { 6215 folio_lock(folio); 6216 anon_rmap = 1; 6217 } 6218 } else { 6219 /* 6220 * If memory error occurs between mmap() and fault, some process 6221 * don't have hwpoisoned swap entry for errored virtual address. 6222 * So we need to block hugepage fault by PG_hwpoison bit check. 6223 */ 6224 if (unlikely(folio_test_hwpoison(folio))) { 6225 ret = VM_FAULT_HWPOISON_LARGE | 6226 VM_FAULT_SET_HINDEX(hstate_index(h)); 6227 goto backout_unlocked; 6228 } 6229 6230 /* Check for page in userfault range. */ 6231 if (userfaultfd_minor(vma)) { 6232 folio_unlock(folio); 6233 folio_put(folio); 6234 /* See comment in userfaultfd_missing() block above */ 6235 if (!hugetlb_pte_stable(h, mm, vmf->address, vmf->pte, vmf->orig_pte)) { 6236 ret = 0; 6237 goto out; 6238 } 6239 return hugetlb_handle_userfault(vmf, mapping, 6240 VM_UFFD_MINOR); 6241 } 6242 } 6243 6244 /* 6245 * If we are going to COW a private mapping later, we examine the 6246 * pending reservations for this page now. This will ensure that 6247 * any allocations necessary to record that reservation occur outside 6248 * the spinlock. 6249 */ 6250 if ((vmf->flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { 6251 if (vma_needs_reservation(h, vma, vmf->address) < 0) { 6252 ret = VM_FAULT_OOM; 6253 goto backout_unlocked; 6254 } 6255 /* Just decrements count, does not deallocate */ 6256 vma_end_reservation(h, vma, vmf->address); 6257 } 6258 6259 vmf->ptl = huge_pte_lock(h, mm, vmf->pte); 6260 ret = 0; 6261 /* If pte changed from under us, retry */ 6262 if (!pte_same(huge_ptep_get(mm, vmf->address, vmf->pte), vmf->orig_pte)) 6263 goto backout; 6264 6265 if (anon_rmap) 6266 hugetlb_add_new_anon_rmap(folio, vma, vmf->address); 6267 else 6268 hugetlb_add_file_rmap(folio); 6269 new_pte = make_huge_pte(vma, &folio->page, vma->vm_flags & VM_SHARED); 6270 /* 6271 * If this pte was previously wr-protected, keep it wr-protected even 6272 * if populated. 6273 */ 6274 if (unlikely(pte_marker_uffd_wp(vmf->orig_pte))) 6275 new_pte = huge_pte_mkuffd_wp(new_pte); 6276 set_huge_pte_at(mm, vmf->address, vmf->pte, new_pte, huge_page_size(h)); 6277 6278 hugetlb_count_add(pages_per_huge_page(h), mm); 6279 if ((vmf->flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { 6280 /* Optimization, do the COW without a second fault */ 6281 ret = hugetlb_wp(folio, vmf); 6282 } 6283 6284 spin_unlock(vmf->ptl); 6285 6286 /* 6287 * Only set hugetlb_migratable in newly allocated pages. Existing pages 6288 * found in the pagecache may not have hugetlb_migratable if they have 6289 * been isolated for migration. 6290 */ 6291 if (new_folio) 6292 folio_set_hugetlb_migratable(folio); 6293 6294 folio_unlock(folio); 6295 out: 6296 hugetlb_vma_unlock_read(vma); 6297 6298 /* 6299 * We must check to release the per-VMA lock. __vmf_anon_prepare() is 6300 * the only way ret can be set to VM_FAULT_RETRY. 6301 */ 6302 if (unlikely(ret & VM_FAULT_RETRY)) 6303 vma_end_read(vma); 6304 6305 mutex_unlock(&hugetlb_fault_mutex_table[hash]); 6306 return ret; 6307 6308 backout: 6309 spin_unlock(vmf->ptl); 6310 backout_unlocked: 6311 if (new_folio && !new_pagecache_folio) 6312 restore_reserve_on_error(h, vma, vmf->address, folio); 6313 6314 folio_unlock(folio); 6315 folio_put(folio); 6316 goto out; 6317 } 6318 6319 #ifdef CONFIG_SMP 6320 u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx) 6321 { 6322 unsigned long key[2]; 6323 u32 hash; 6324 6325 key[0] = (unsigned long) mapping; 6326 key[1] = idx; 6327 6328 hash = jhash2((u32 *)&key, sizeof(key)/(sizeof(u32)), 0); 6329 6330 return hash & (num_fault_mutexes - 1); 6331 } 6332 #else 6333 /* 6334 * For uniprocessor systems we always use a single mutex, so just 6335 * return 0 and avoid the hashing overhead. 6336 */ 6337 u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx) 6338 { 6339 return 0; 6340 } 6341 #endif 6342 6343 vm_fault_t hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma, 6344 unsigned long address, unsigned int flags) 6345 { 6346 vm_fault_t ret; 6347 u32 hash; 6348 struct folio *folio = NULL; 6349 struct folio *pagecache_folio = NULL; 6350 struct hstate *h = hstate_vma(vma); 6351 struct address_space *mapping; 6352 int need_wait_lock = 0; 6353 struct vm_fault vmf = { 6354 .vma = vma, 6355 .address = address & huge_page_mask(h), 6356 .real_address = address, 6357 .flags = flags, 6358 .pgoff = vma_hugecache_offset(h, vma, 6359 address & huge_page_mask(h)), 6360 /* TODO: Track hugetlb faults using vm_fault */ 6361 6362 /* 6363 * Some fields may not be initialized, be careful as it may 6364 * be hard to debug if called functions make assumptions 6365 */ 6366 }; 6367 6368 /* 6369 * Serialize hugepage allocation and instantiation, so that we don't 6370 * get spurious allocation failures if two CPUs race to instantiate 6371 * the same page in the page cache. 6372 */ 6373 mapping = vma->vm_file->f_mapping; 6374 hash = hugetlb_fault_mutex_hash(mapping, vmf.pgoff); 6375 mutex_lock(&hugetlb_fault_mutex_table[hash]); 6376 6377 /* 6378 * Acquire vma lock before calling huge_pte_alloc and hold 6379 * until finished with vmf.pte. This prevents huge_pmd_unshare from 6380 * being called elsewhere and making the vmf.pte no longer valid. 6381 */ 6382 hugetlb_vma_lock_read(vma); 6383 vmf.pte = huge_pte_alloc(mm, vma, vmf.address, huge_page_size(h)); 6384 if (!vmf.pte) { 6385 hugetlb_vma_unlock_read(vma); 6386 mutex_unlock(&hugetlb_fault_mutex_table[hash]); 6387 return VM_FAULT_OOM; 6388 } 6389 6390 vmf.orig_pte = huge_ptep_get(mm, vmf.address, vmf.pte); 6391 if (huge_pte_none_mostly(vmf.orig_pte)) { 6392 if (is_pte_marker(vmf.orig_pte)) { 6393 pte_marker marker = 6394 pte_marker_get(pte_to_swp_entry(vmf.orig_pte)); 6395 6396 if (marker & PTE_MARKER_POISONED) { 6397 ret = VM_FAULT_HWPOISON_LARGE | 6398 VM_FAULT_SET_HINDEX(hstate_index(h)); 6399 goto out_mutex; 6400 } else if (WARN_ON_ONCE(marker & PTE_MARKER_GUARD)) { 6401 /* This isn't supported in hugetlb. */ 6402 ret = VM_FAULT_SIGSEGV; 6403 goto out_mutex; 6404 } 6405 } 6406 6407 /* 6408 * Other PTE markers should be handled the same way as none PTE. 6409 * 6410 * hugetlb_no_page will drop vma lock and hugetlb fault 6411 * mutex internally, which make us return immediately. 6412 */ 6413 return hugetlb_no_page(mapping, &vmf); 6414 } 6415 6416 ret = 0; 6417 6418 /* 6419 * vmf.orig_pte could be a migration/hwpoison vmf.orig_pte at this 6420 * point, so this check prevents the kernel from going below assuming 6421 * that we have an active hugepage in pagecache. This goto expects 6422 * the 2nd page fault, and is_hugetlb_entry_(migration|hwpoisoned) 6423 * check will properly handle it. 6424 */ 6425 if (!pte_present(vmf.orig_pte)) { 6426 if (unlikely(is_hugetlb_entry_migration(vmf.orig_pte))) { 6427 /* 6428 * Release the hugetlb fault lock now, but retain 6429 * the vma lock, because it is needed to guard the 6430 * huge_pte_lockptr() later in 6431 * migration_entry_wait_huge(). The vma lock will 6432 * be released there. 6433 */ 6434 mutex_unlock(&hugetlb_fault_mutex_table[hash]); 6435 migration_entry_wait_huge(vma, vmf.address, vmf.pte); 6436 return 0; 6437 } else if (unlikely(is_hugetlb_entry_hwpoisoned(vmf.orig_pte))) 6438 ret = VM_FAULT_HWPOISON_LARGE | 6439 VM_FAULT_SET_HINDEX(hstate_index(h)); 6440 goto out_mutex; 6441 } 6442 6443 /* 6444 * If we are going to COW/unshare the mapping later, we examine the 6445 * pending reservations for this page now. This will ensure that any 6446 * allocations necessary to record that reservation occur outside the 6447 * spinlock. Also lookup the pagecache page now as it is used to 6448 * determine if a reservation has been consumed. 6449 */ 6450 if ((flags & (FAULT_FLAG_WRITE|FAULT_FLAG_UNSHARE)) && 6451 !(vma->vm_flags & VM_MAYSHARE) && !huge_pte_write(vmf.orig_pte)) { 6452 if (vma_needs_reservation(h, vma, vmf.address) < 0) { 6453 ret = VM_FAULT_OOM; 6454 goto out_mutex; 6455 } 6456 /* Just decrements count, does not deallocate */ 6457 vma_end_reservation(h, vma, vmf.address); 6458 6459 pagecache_folio = filemap_lock_hugetlb_folio(h, mapping, 6460 vmf.pgoff); 6461 if (IS_ERR(pagecache_folio)) 6462 pagecache_folio = NULL; 6463 } 6464 6465 vmf.ptl = huge_pte_lock(h, mm, vmf.pte); 6466 6467 /* Check for a racing update before calling hugetlb_wp() */ 6468 if (unlikely(!pte_same(vmf.orig_pte, huge_ptep_get(mm, vmf.address, vmf.pte)))) 6469 goto out_ptl; 6470 6471 /* Handle userfault-wp first, before trying to lock more pages */ 6472 if (userfaultfd_wp(vma) && huge_pte_uffd_wp(huge_ptep_get(mm, vmf.address, vmf.pte)) && 6473 (flags & FAULT_FLAG_WRITE) && !huge_pte_write(vmf.orig_pte)) { 6474 if (!userfaultfd_wp_async(vma)) { 6475 spin_unlock(vmf.ptl); 6476 if (pagecache_folio) { 6477 folio_unlock(pagecache_folio); 6478 folio_put(pagecache_folio); 6479 } 6480 hugetlb_vma_unlock_read(vma); 6481 mutex_unlock(&hugetlb_fault_mutex_table[hash]); 6482 return handle_userfault(&vmf, VM_UFFD_WP); 6483 } 6484 6485 vmf.orig_pte = huge_pte_clear_uffd_wp(vmf.orig_pte); 6486 set_huge_pte_at(mm, vmf.address, vmf.pte, vmf.orig_pte, 6487 huge_page_size(hstate_vma(vma))); 6488 /* Fallthrough to CoW */ 6489 } 6490 6491 /* 6492 * hugetlb_wp() requires page locks of pte_page(vmf.orig_pte) and 6493 * pagecache_folio, so here we need take the former one 6494 * when folio != pagecache_folio or !pagecache_folio. 6495 */ 6496 folio = page_folio(pte_page(vmf.orig_pte)); 6497 if (folio != pagecache_folio) 6498 if (!folio_trylock(folio)) { 6499 need_wait_lock = 1; 6500 goto out_ptl; 6501 } 6502 6503 folio_get(folio); 6504 6505 if (flags & (FAULT_FLAG_WRITE|FAULT_FLAG_UNSHARE)) { 6506 if (!huge_pte_write(vmf.orig_pte)) { 6507 ret = hugetlb_wp(pagecache_folio, &vmf); 6508 goto out_put_page; 6509 } else if (likely(flags & FAULT_FLAG_WRITE)) { 6510 vmf.orig_pte = huge_pte_mkdirty(vmf.orig_pte); 6511 } 6512 } 6513 vmf.orig_pte = pte_mkyoung(vmf.orig_pte); 6514 if (huge_ptep_set_access_flags(vma, vmf.address, vmf.pte, vmf.orig_pte, 6515 flags & FAULT_FLAG_WRITE)) 6516 update_mmu_cache(vma, vmf.address, vmf.pte); 6517 out_put_page: 6518 if (folio != pagecache_folio) 6519 folio_unlock(folio); 6520 folio_put(folio); 6521 out_ptl: 6522 spin_unlock(vmf.ptl); 6523 6524 if (pagecache_folio) { 6525 folio_unlock(pagecache_folio); 6526 folio_put(pagecache_folio); 6527 } 6528 out_mutex: 6529 hugetlb_vma_unlock_read(vma); 6530 6531 /* 6532 * We must check to release the per-VMA lock. __vmf_anon_prepare() in 6533 * hugetlb_wp() is the only way ret can be set to VM_FAULT_RETRY. 6534 */ 6535 if (unlikely(ret & VM_FAULT_RETRY)) 6536 vma_end_read(vma); 6537 6538 mutex_unlock(&hugetlb_fault_mutex_table[hash]); 6539 /* 6540 * Generally it's safe to hold refcount during waiting page lock. But 6541 * here we just wait to defer the next page fault to avoid busy loop and 6542 * the page is not used after unlocked before returning from the current 6543 * page fault. So we are safe from accessing freed page, even if we wait 6544 * here without taking refcount. 6545 */ 6546 if (need_wait_lock) 6547 folio_wait_locked(folio); 6548 return ret; 6549 } 6550 6551 #ifdef CONFIG_USERFAULTFD 6552 /* 6553 * Can probably be eliminated, but still used by hugetlb_mfill_atomic_pte(). 6554 */ 6555 static struct folio *alloc_hugetlb_folio_vma(struct hstate *h, 6556 struct vm_area_struct *vma, unsigned long address) 6557 { 6558 struct mempolicy *mpol; 6559 nodemask_t *nodemask; 6560 struct folio *folio; 6561 gfp_t gfp_mask; 6562 int node; 6563 6564 gfp_mask = htlb_alloc_mask(h); 6565 node = huge_node(vma, address, gfp_mask, &mpol, &nodemask); 6566 /* 6567 * This is used to allocate a temporary hugetlb to hold the copied 6568 * content, which will then be copied again to the final hugetlb 6569 * consuming a reservation. Set the alloc_fallback to false to indicate 6570 * that breaking the per-node hugetlb pool is not allowed in this case. 6571 */ 6572 folio = alloc_hugetlb_folio_nodemask(h, node, nodemask, gfp_mask, false); 6573 mpol_cond_put(mpol); 6574 6575 return folio; 6576 } 6577 6578 /* 6579 * Used by userfaultfd UFFDIO_* ioctls. Based on userfaultfd's mfill_atomic_pte 6580 * with modifications for hugetlb pages. 6581 */ 6582 int hugetlb_mfill_atomic_pte(pte_t *dst_pte, 6583 struct vm_area_struct *dst_vma, 6584 unsigned long dst_addr, 6585 unsigned long src_addr, 6586 uffd_flags_t flags, 6587 struct folio **foliop) 6588 { 6589 struct mm_struct *dst_mm = dst_vma->vm_mm; 6590 bool is_continue = uffd_flags_mode_is(flags, MFILL_ATOMIC_CONTINUE); 6591 bool wp_enabled = (flags & MFILL_ATOMIC_WP); 6592 struct hstate *h = hstate_vma(dst_vma); 6593 struct address_space *mapping = dst_vma->vm_file->f_mapping; 6594 pgoff_t idx = vma_hugecache_offset(h, dst_vma, dst_addr); 6595 unsigned long size = huge_page_size(h); 6596 int vm_shared = dst_vma->vm_flags & VM_SHARED; 6597 pte_t _dst_pte; 6598 spinlock_t *ptl; 6599 int ret = -ENOMEM; 6600 struct folio *folio; 6601 bool folio_in_pagecache = false; 6602 6603 if (uffd_flags_mode_is(flags, MFILL_ATOMIC_POISON)) { 6604 ptl = huge_pte_lock(h, dst_mm, dst_pte); 6605 6606 /* Don't overwrite any existing PTEs (even markers) */ 6607 if (!huge_pte_none(huge_ptep_get(dst_mm, dst_addr, dst_pte))) { 6608 spin_unlock(ptl); 6609 return -EEXIST; 6610 } 6611 6612 _dst_pte = make_pte_marker(PTE_MARKER_POISONED); 6613 set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte, size); 6614 6615 /* No need to invalidate - it was non-present before */ 6616 update_mmu_cache(dst_vma, dst_addr, dst_pte); 6617 6618 spin_unlock(ptl); 6619 return 0; 6620 } 6621 6622 if (is_continue) { 6623 ret = -EFAULT; 6624 folio = filemap_lock_hugetlb_folio(h, mapping, idx); 6625 if (IS_ERR(folio)) 6626 goto out; 6627 folio_in_pagecache = true; 6628 } else if (!*foliop) { 6629 /* If a folio already exists, then it's UFFDIO_COPY for 6630 * a non-missing case. Return -EEXIST. 6631 */ 6632 if (vm_shared && 6633 hugetlbfs_pagecache_present(h, dst_vma, dst_addr)) { 6634 ret = -EEXIST; 6635 goto out; 6636 } 6637 6638 folio = alloc_hugetlb_folio(dst_vma, dst_addr, false); 6639 if (IS_ERR(folio)) { 6640 ret = -ENOMEM; 6641 goto out; 6642 } 6643 6644 ret = copy_folio_from_user(folio, (const void __user *) src_addr, 6645 false); 6646 6647 /* fallback to copy_from_user outside mmap_lock */ 6648 if (unlikely(ret)) { 6649 ret = -ENOENT; 6650 /* Free the allocated folio which may have 6651 * consumed a reservation. 6652 */ 6653 restore_reserve_on_error(h, dst_vma, dst_addr, folio); 6654 folio_put(folio); 6655 6656 /* Allocate a temporary folio to hold the copied 6657 * contents. 6658 */ 6659 folio = alloc_hugetlb_folio_vma(h, dst_vma, dst_addr); 6660 if (!folio) { 6661 ret = -ENOMEM; 6662 goto out; 6663 } 6664 *foliop = folio; 6665 /* Set the outparam foliop and return to the caller to 6666 * copy the contents outside the lock. Don't free the 6667 * folio. 6668 */ 6669 goto out; 6670 } 6671 } else { 6672 if (vm_shared && 6673 hugetlbfs_pagecache_present(h, dst_vma, dst_addr)) { 6674 folio_put(*foliop); 6675 ret = -EEXIST; 6676 *foliop = NULL; 6677 goto out; 6678 } 6679 6680 folio = alloc_hugetlb_folio(dst_vma, dst_addr, false); 6681 if (IS_ERR(folio)) { 6682 folio_put(*foliop); 6683 ret = -ENOMEM; 6684 *foliop = NULL; 6685 goto out; 6686 } 6687 ret = copy_user_large_folio(folio, *foliop, dst_addr, dst_vma); 6688 folio_put(*foliop); 6689 *foliop = NULL; 6690 if (ret) { 6691 folio_put(folio); 6692 goto out; 6693 } 6694 } 6695 6696 /* 6697 * If we just allocated a new page, we need a memory barrier to ensure 6698 * that preceding stores to the page become visible before the 6699 * set_pte_at() write. The memory barrier inside __folio_mark_uptodate 6700 * is what we need. 6701 * 6702 * In the case where we have not allocated a new page (is_continue), 6703 * the page must already be uptodate. UFFDIO_CONTINUE already includes 6704 * an earlier smp_wmb() to ensure that prior stores will be visible 6705 * before the set_pte_at() write. 6706 */ 6707 if (!is_continue) 6708 __folio_mark_uptodate(folio); 6709 else 6710 WARN_ON_ONCE(!folio_test_uptodate(folio)); 6711 6712 /* Add shared, newly allocated pages to the page cache. */ 6713 if (vm_shared && !is_continue) { 6714 ret = -EFAULT; 6715 if (idx >= (i_size_read(mapping->host) >> huge_page_shift(h))) 6716 goto out_release_nounlock; 6717 6718 /* 6719 * Serialization between remove_inode_hugepages() and 6720 * hugetlb_add_to_page_cache() below happens through the 6721 * hugetlb_fault_mutex_table that here must be hold by 6722 * the caller. 6723 */ 6724 ret = hugetlb_add_to_page_cache(folio, mapping, idx); 6725 if (ret) 6726 goto out_release_nounlock; 6727 folio_in_pagecache = true; 6728 } 6729 6730 ptl = huge_pte_lock(h, dst_mm, dst_pte); 6731 6732 ret = -EIO; 6733 if (folio_test_hwpoison(folio)) 6734 goto out_release_unlock; 6735 6736 /* 6737 * We allow to overwrite a pte marker: consider when both MISSING|WP 6738 * registered, we firstly wr-protect a none pte which has no page cache 6739 * page backing it, then access the page. 6740 */ 6741 ret = -EEXIST; 6742 if (!huge_pte_none_mostly(huge_ptep_get(dst_mm, dst_addr, dst_pte))) 6743 goto out_release_unlock; 6744 6745 if (folio_in_pagecache) 6746 hugetlb_add_file_rmap(folio); 6747 else 6748 hugetlb_add_new_anon_rmap(folio, dst_vma, dst_addr); 6749 6750 /* 6751 * For either: (1) CONTINUE on a non-shared VMA, or (2) UFFDIO_COPY 6752 * with wp flag set, don't set pte write bit. 6753 */ 6754 _dst_pte = make_huge_pte(dst_vma, &folio->page, 6755 !wp_enabled && !(is_continue && !vm_shared)); 6756 /* 6757 * Always mark UFFDIO_COPY page dirty; note that this may not be 6758 * extremely important for hugetlbfs for now since swapping is not 6759 * supported, but we should still be clear in that this page cannot be 6760 * thrown away at will, even if write bit not set. 6761 */ 6762 _dst_pte = huge_pte_mkdirty(_dst_pte); 6763 _dst_pte = pte_mkyoung(_dst_pte); 6764 6765 if (wp_enabled) 6766 _dst_pte = huge_pte_mkuffd_wp(_dst_pte); 6767 6768 set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte, size); 6769 6770 hugetlb_count_add(pages_per_huge_page(h), dst_mm); 6771 6772 /* No need to invalidate - it was non-present before */ 6773 update_mmu_cache(dst_vma, dst_addr, dst_pte); 6774 6775 spin_unlock(ptl); 6776 if (!is_continue) 6777 folio_set_hugetlb_migratable(folio); 6778 if (vm_shared || is_continue) 6779 folio_unlock(folio); 6780 ret = 0; 6781 out: 6782 return ret; 6783 out_release_unlock: 6784 spin_unlock(ptl); 6785 if (vm_shared || is_continue) 6786 folio_unlock(folio); 6787 out_release_nounlock: 6788 if (!folio_in_pagecache) 6789 restore_reserve_on_error(h, dst_vma, dst_addr, folio); 6790 folio_put(folio); 6791 goto out; 6792 } 6793 #endif /* CONFIG_USERFAULTFD */ 6794 6795 long hugetlb_change_protection(struct vm_area_struct *vma, 6796 unsigned long address, unsigned long end, 6797 pgprot_t newprot, unsigned long cp_flags) 6798 { 6799 struct mm_struct *mm = vma->vm_mm; 6800 unsigned long start = address; 6801 pte_t *ptep; 6802 pte_t pte; 6803 struct hstate *h = hstate_vma(vma); 6804 long pages = 0, psize = huge_page_size(h); 6805 bool shared_pmd = false; 6806 struct mmu_notifier_range range; 6807 unsigned long last_addr_mask; 6808 bool uffd_wp = cp_flags & MM_CP_UFFD_WP; 6809 bool uffd_wp_resolve = cp_flags & MM_CP_UFFD_WP_RESOLVE; 6810 6811 /* 6812 * In the case of shared PMDs, the area to flush could be beyond 6813 * start/end. Set range.start/range.end to cover the maximum possible 6814 * range if PMD sharing is possible. 6815 */ 6816 mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_VMA, 6817 0, mm, start, end); 6818 adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end); 6819 6820 BUG_ON(address >= end); 6821 flush_cache_range(vma, range.start, range.end); 6822 6823 mmu_notifier_invalidate_range_start(&range); 6824 hugetlb_vma_lock_write(vma); 6825 i_mmap_lock_write(vma->vm_file->f_mapping); 6826 last_addr_mask = hugetlb_mask_last_page(h); 6827 for (; address < end; address += psize) { 6828 spinlock_t *ptl; 6829 ptep = hugetlb_walk(vma, address, psize); 6830 if (!ptep) { 6831 if (!uffd_wp) { 6832 address |= last_addr_mask; 6833 continue; 6834 } 6835 /* 6836 * Userfaultfd wr-protect requires pgtable 6837 * pre-allocations to install pte markers. 6838 */ 6839 ptep = huge_pte_alloc(mm, vma, address, psize); 6840 if (!ptep) { 6841 pages = -ENOMEM; 6842 break; 6843 } 6844 } 6845 ptl = huge_pte_lock(h, mm, ptep); 6846 if (huge_pmd_unshare(mm, vma, address, ptep)) { 6847 /* 6848 * When uffd-wp is enabled on the vma, unshare 6849 * shouldn't happen at all. Warn about it if it 6850 * happened due to some reason. 6851 */ 6852 WARN_ON_ONCE(uffd_wp || uffd_wp_resolve); 6853 pages++; 6854 spin_unlock(ptl); 6855 shared_pmd = true; 6856 address |= last_addr_mask; 6857 continue; 6858 } 6859 pte = huge_ptep_get(mm, address, ptep); 6860 if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) { 6861 /* Nothing to do. */ 6862 } else if (unlikely(is_hugetlb_entry_migration(pte))) { 6863 swp_entry_t entry = pte_to_swp_entry(pte); 6864 struct page *page = pfn_swap_entry_to_page(entry); 6865 pte_t newpte = pte; 6866 6867 if (is_writable_migration_entry(entry)) { 6868 if (PageAnon(page)) 6869 entry = make_readable_exclusive_migration_entry( 6870 swp_offset(entry)); 6871 else 6872 entry = make_readable_migration_entry( 6873 swp_offset(entry)); 6874 newpte = swp_entry_to_pte(entry); 6875 pages++; 6876 } 6877 6878 if (uffd_wp) 6879 newpte = pte_swp_mkuffd_wp(newpte); 6880 else if (uffd_wp_resolve) 6881 newpte = pte_swp_clear_uffd_wp(newpte); 6882 if (!pte_same(pte, newpte)) 6883 set_huge_pte_at(mm, address, ptep, newpte, psize); 6884 } else if (unlikely(is_pte_marker(pte))) { 6885 /* 6886 * Do nothing on a poison marker; page is 6887 * corrupted, permissons do not apply. Here 6888 * pte_marker_uffd_wp()==true implies !poison 6889 * because they're mutual exclusive. 6890 */ 6891 if (pte_marker_uffd_wp(pte) && uffd_wp_resolve) 6892 /* Safe to modify directly (non-present->none). */ 6893 huge_pte_clear(mm, address, ptep, psize); 6894 } else if (!huge_pte_none(pte)) { 6895 pte_t old_pte; 6896 unsigned int shift = huge_page_shift(hstate_vma(vma)); 6897 6898 old_pte = huge_ptep_modify_prot_start(vma, address, ptep); 6899 pte = huge_pte_modify(old_pte, newprot); 6900 pte = arch_make_huge_pte(pte, shift, vma->vm_flags); 6901 if (uffd_wp) 6902 pte = huge_pte_mkuffd_wp(pte); 6903 else if (uffd_wp_resolve) 6904 pte = huge_pte_clear_uffd_wp(pte); 6905 huge_ptep_modify_prot_commit(vma, address, ptep, old_pte, pte); 6906 pages++; 6907 } else { 6908 /* None pte */ 6909 if (unlikely(uffd_wp)) 6910 /* Safe to modify directly (none->non-present). */ 6911 set_huge_pte_at(mm, address, ptep, 6912 make_pte_marker(PTE_MARKER_UFFD_WP), 6913 psize); 6914 } 6915 spin_unlock(ptl); 6916 } 6917 /* 6918 * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare 6919 * may have cleared our pud entry and done put_page on the page table: 6920 * once we release i_mmap_rwsem, another task can do the final put_page 6921 * and that page table be reused and filled with junk. If we actually 6922 * did unshare a page of pmds, flush the range corresponding to the pud. 6923 */ 6924 if (shared_pmd) 6925 flush_hugetlb_tlb_range(vma, range.start, range.end); 6926 else 6927 flush_hugetlb_tlb_range(vma, start, end); 6928 /* 6929 * No need to call mmu_notifier_arch_invalidate_secondary_tlbs() we are 6930 * downgrading page table protection not changing it to point to a new 6931 * page. 6932 * 6933 * See Documentation/mm/mmu_notifier.rst 6934 */ 6935 i_mmap_unlock_write(vma->vm_file->f_mapping); 6936 hugetlb_vma_unlock_write(vma); 6937 mmu_notifier_invalidate_range_end(&range); 6938 6939 return pages > 0 ? (pages << h->order) : pages; 6940 } 6941 6942 /* Return true if reservation was successful, false otherwise. */ 6943 bool hugetlb_reserve_pages(struct inode *inode, 6944 long from, long to, 6945 struct vm_area_struct *vma, 6946 vm_flags_t vm_flags) 6947 { 6948 long chg = -1, add = -1; 6949 struct hstate *h = hstate_inode(inode); 6950 struct hugepage_subpool *spool = subpool_inode(inode); 6951 struct resv_map *resv_map; 6952 struct hugetlb_cgroup *h_cg = NULL; 6953 long gbl_reserve, regions_needed = 0; 6954 6955 /* This should never happen */ 6956 if (from > to) { 6957 VM_WARN(1, "%s called with a negative range\n", __func__); 6958 return false; 6959 } 6960 6961 /* 6962 * vma specific semaphore used for pmd sharing and fault/truncation 6963 * synchronization 6964 */ 6965 hugetlb_vma_lock_alloc(vma); 6966 6967 /* 6968 * Only apply hugepage reservation if asked. At fault time, an 6969 * attempt will be made for VM_NORESERVE to allocate a page 6970 * without using reserves 6971 */ 6972 if (vm_flags & VM_NORESERVE) 6973 return true; 6974 6975 /* 6976 * Shared mappings base their reservation on the number of pages that 6977 * are already allocated on behalf of the file. Private mappings need 6978 * to reserve the full area even if read-only as mprotect() may be 6979 * called to make the mapping read-write. Assume !vma is a shm mapping 6980 */ 6981 if (!vma || vma->vm_flags & VM_MAYSHARE) { 6982 /* 6983 * resv_map can not be NULL as hugetlb_reserve_pages is only 6984 * called for inodes for which resv_maps were created (see 6985 * hugetlbfs_get_inode). 6986 */ 6987 resv_map = inode_resv_map(inode); 6988 6989 chg = region_chg(resv_map, from, to, ®ions_needed); 6990 } else { 6991 /* Private mapping. */ 6992 resv_map = resv_map_alloc(); 6993 if (!resv_map) 6994 goto out_err; 6995 6996 chg = to - from; 6997 6998 set_vma_resv_map(vma, resv_map); 6999 set_vma_resv_flags(vma, HPAGE_RESV_OWNER); 7000 } 7001 7002 if (chg < 0) 7003 goto out_err; 7004 7005 if (hugetlb_cgroup_charge_cgroup_rsvd(hstate_index(h), 7006 chg * pages_per_huge_page(h), &h_cg) < 0) 7007 goto out_err; 7008 7009 if (vma && !(vma->vm_flags & VM_MAYSHARE) && h_cg) { 7010 /* For private mappings, the hugetlb_cgroup uncharge info hangs 7011 * of the resv_map. 7012 */ 7013 resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, h_cg, h); 7014 } 7015 7016 /* 7017 * There must be enough pages in the subpool for the mapping. If 7018 * the subpool has a minimum size, there may be some global 7019 * reservations already in place (gbl_reserve). 7020 */ 7021 gbl_reserve = hugepage_subpool_get_pages(spool, chg); 7022 if (gbl_reserve < 0) 7023 goto out_uncharge_cgroup; 7024 7025 /* 7026 * Check enough hugepages are available for the reservation. 7027 * Hand the pages back to the subpool if there are not 7028 */ 7029 if (hugetlb_acct_memory(h, gbl_reserve) < 0) 7030 goto out_put_pages; 7031 7032 /* 7033 * Account for the reservations made. Shared mappings record regions 7034 * that have reservations as they are shared by multiple VMAs. 7035 * When the last VMA disappears, the region map says how much 7036 * the reservation was and the page cache tells how much of 7037 * the reservation was consumed. Private mappings are per-VMA and 7038 * only the consumed reservations are tracked. When the VMA 7039 * disappears, the original reservation is the VMA size and the 7040 * consumed reservations are stored in the map. Hence, nothing 7041 * else has to be done for private mappings here 7042 */ 7043 if (!vma || vma->vm_flags & VM_MAYSHARE) { 7044 add = region_add(resv_map, from, to, regions_needed, h, h_cg); 7045 7046 if (unlikely(add < 0)) { 7047 hugetlb_acct_memory(h, -gbl_reserve); 7048 goto out_put_pages; 7049 } else if (unlikely(chg > add)) { 7050 /* 7051 * pages in this range were added to the reserve 7052 * map between region_chg and region_add. This 7053 * indicates a race with alloc_hugetlb_folio. Adjust 7054 * the subpool and reserve counts modified above 7055 * based on the difference. 7056 */ 7057 long rsv_adjust; 7058 7059 /* 7060 * hugetlb_cgroup_uncharge_cgroup_rsvd() will put the 7061 * reference to h_cg->css. See comment below for detail. 7062 */ 7063 hugetlb_cgroup_uncharge_cgroup_rsvd( 7064 hstate_index(h), 7065 (chg - add) * pages_per_huge_page(h), h_cg); 7066 7067 rsv_adjust = hugepage_subpool_put_pages(spool, 7068 chg - add); 7069 hugetlb_acct_memory(h, -rsv_adjust); 7070 } else if (h_cg) { 7071 /* 7072 * The file_regions will hold their own reference to 7073 * h_cg->css. So we should release the reference held 7074 * via hugetlb_cgroup_charge_cgroup_rsvd() when we are 7075 * done. 7076 */ 7077 hugetlb_cgroup_put_rsvd_cgroup(h_cg); 7078 } 7079 } 7080 return true; 7081 7082 out_put_pages: 7083 /* put back original number of pages, chg */ 7084 (void)hugepage_subpool_put_pages(spool, chg); 7085 out_uncharge_cgroup: 7086 hugetlb_cgroup_uncharge_cgroup_rsvd(hstate_index(h), 7087 chg * pages_per_huge_page(h), h_cg); 7088 out_err: 7089 hugetlb_vma_lock_free(vma); 7090 if (!vma || vma->vm_flags & VM_MAYSHARE) 7091 /* Only call region_abort if the region_chg succeeded but the 7092 * region_add failed or didn't run. 7093 */ 7094 if (chg >= 0 && add < 0) 7095 region_abort(resv_map, from, to, regions_needed); 7096 if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { 7097 kref_put(&resv_map->refs, resv_map_release); 7098 set_vma_resv_map(vma, NULL); 7099 } 7100 return false; 7101 } 7102 7103 long hugetlb_unreserve_pages(struct inode *inode, long start, long end, 7104 long freed) 7105 { 7106 struct hstate *h = hstate_inode(inode); 7107 struct resv_map *resv_map = inode_resv_map(inode); 7108 long chg = 0; 7109 struct hugepage_subpool *spool = subpool_inode(inode); 7110 long gbl_reserve; 7111 7112 /* 7113 * Since this routine can be called in the evict inode path for all 7114 * hugetlbfs inodes, resv_map could be NULL. 7115 */ 7116 if (resv_map) { 7117 chg = region_del(resv_map, start, end); 7118 /* 7119 * region_del() can fail in the rare case where a region 7120 * must be split and another region descriptor can not be 7121 * allocated. If end == LONG_MAX, it will not fail. 7122 */ 7123 if (chg < 0) 7124 return chg; 7125 } 7126 7127 spin_lock(&inode->i_lock); 7128 inode->i_blocks -= (blocks_per_huge_page(h) * freed); 7129 spin_unlock(&inode->i_lock); 7130 7131 /* 7132 * If the subpool has a minimum size, the number of global 7133 * reservations to be released may be adjusted. 7134 * 7135 * Note that !resv_map implies freed == 0. So (chg - freed) 7136 * won't go negative. 7137 */ 7138 gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed)); 7139 hugetlb_acct_memory(h, -gbl_reserve); 7140 7141 return 0; 7142 } 7143 7144 #ifdef CONFIG_HUGETLB_PMD_PAGE_TABLE_SHARING 7145 static unsigned long page_table_shareable(struct vm_area_struct *svma, 7146 struct vm_area_struct *vma, 7147 unsigned long addr, pgoff_t idx) 7148 { 7149 unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) + 7150 svma->vm_start; 7151 unsigned long sbase = saddr & PUD_MASK; 7152 unsigned long s_end = sbase + PUD_SIZE; 7153 7154 /* Allow segments to share if only one is marked locked */ 7155 unsigned long vm_flags = vma->vm_flags & ~VM_LOCKED_MASK; 7156 unsigned long svm_flags = svma->vm_flags & ~VM_LOCKED_MASK; 7157 7158 /* 7159 * match the virtual addresses, permission and the alignment of the 7160 * page table page. 7161 * 7162 * Also, vma_lock (vm_private_data) is required for sharing. 7163 */ 7164 if (pmd_index(addr) != pmd_index(saddr) || 7165 vm_flags != svm_flags || 7166 !range_in_vma(svma, sbase, s_end) || 7167 !svma->vm_private_data) 7168 return 0; 7169 7170 return saddr; 7171 } 7172 7173 bool want_pmd_share(struct vm_area_struct *vma, unsigned long addr) 7174 { 7175 unsigned long start = addr & PUD_MASK; 7176 unsigned long end = start + PUD_SIZE; 7177 7178 #ifdef CONFIG_USERFAULTFD 7179 if (uffd_disable_huge_pmd_share(vma)) 7180 return false; 7181 #endif 7182 /* 7183 * check on proper vm_flags and page table alignment 7184 */ 7185 if (!(vma->vm_flags & VM_MAYSHARE)) 7186 return false; 7187 if (!vma->vm_private_data) /* vma lock required for sharing */ 7188 return false; 7189 if (!range_in_vma(vma, start, end)) 7190 return false; 7191 return true; 7192 } 7193 7194 /* 7195 * Determine if start,end range within vma could be mapped by shared pmd. 7196 * If yes, adjust start and end to cover range associated with possible 7197 * shared pmd mappings. 7198 */ 7199 void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma, 7200 unsigned long *start, unsigned long *end) 7201 { 7202 unsigned long v_start = ALIGN(vma->vm_start, PUD_SIZE), 7203 v_end = ALIGN_DOWN(vma->vm_end, PUD_SIZE); 7204 7205 /* 7206 * vma needs to span at least one aligned PUD size, and the range 7207 * must be at least partially within in. 7208 */ 7209 if (!(vma->vm_flags & VM_MAYSHARE) || !(v_end > v_start) || 7210 (*end <= v_start) || (*start >= v_end)) 7211 return; 7212 7213 /* Extend the range to be PUD aligned for a worst case scenario */ 7214 if (*start > v_start) 7215 *start = ALIGN_DOWN(*start, PUD_SIZE); 7216 7217 if (*end < v_end) 7218 *end = ALIGN(*end, PUD_SIZE); 7219 } 7220 7221 /* 7222 * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc() 7223 * and returns the corresponding pte. While this is not necessary for the 7224 * !shared pmd case because we can allocate the pmd later as well, it makes the 7225 * code much cleaner. pmd allocation is essential for the shared case because 7226 * pud has to be populated inside the same i_mmap_rwsem section - otherwise 7227 * racing tasks could either miss the sharing (see huge_pte_offset) or select a 7228 * bad pmd for sharing. 7229 */ 7230 pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct *vma, 7231 unsigned long addr, pud_t *pud) 7232 { 7233 struct address_space *mapping = vma->vm_file->f_mapping; 7234 pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) + 7235 vma->vm_pgoff; 7236 struct vm_area_struct *svma; 7237 unsigned long saddr; 7238 pte_t *spte = NULL; 7239 pte_t *pte; 7240 7241 i_mmap_lock_read(mapping); 7242 vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) { 7243 if (svma == vma) 7244 continue; 7245 7246 saddr = page_table_shareable(svma, vma, addr, idx); 7247 if (saddr) { 7248 spte = hugetlb_walk(svma, saddr, 7249 vma_mmu_pagesize(svma)); 7250 if (spte) { 7251 ptdesc_pmd_pts_inc(virt_to_ptdesc(spte)); 7252 break; 7253 } 7254 } 7255 } 7256 7257 if (!spte) 7258 goto out; 7259 7260 spin_lock(&mm->page_table_lock); 7261 if (pud_none(*pud)) { 7262 pud_populate(mm, pud, 7263 (pmd_t *)((unsigned long)spte & PAGE_MASK)); 7264 mm_inc_nr_pmds(mm); 7265 } else { 7266 ptdesc_pmd_pts_dec(virt_to_ptdesc(spte)); 7267 } 7268 spin_unlock(&mm->page_table_lock); 7269 out: 7270 pte = (pte_t *)pmd_alloc(mm, pud, addr); 7271 i_mmap_unlock_read(mapping); 7272 return pte; 7273 } 7274 7275 /* 7276 * unmap huge page backed by shared pte. 7277 * 7278 * Called with page table lock held. 7279 * 7280 * returns: 1 successfully unmapped a shared pte page 7281 * 0 the underlying pte page is not shared, or it is the last user 7282 */ 7283 int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma, 7284 unsigned long addr, pte_t *ptep) 7285 { 7286 unsigned long sz = huge_page_size(hstate_vma(vma)); 7287 pgd_t *pgd = pgd_offset(mm, addr); 7288 p4d_t *p4d = p4d_offset(pgd, addr); 7289 pud_t *pud = pud_offset(p4d, addr); 7290 7291 i_mmap_assert_write_locked(vma->vm_file->f_mapping); 7292 hugetlb_vma_assert_locked(vma); 7293 if (sz != PMD_SIZE) 7294 return 0; 7295 if (!ptdesc_pmd_pts_count(virt_to_ptdesc(ptep))) 7296 return 0; 7297 7298 pud_clear(pud); 7299 ptdesc_pmd_pts_dec(virt_to_ptdesc(ptep)); 7300 mm_dec_nr_pmds(mm); 7301 return 1; 7302 } 7303 7304 #else /* !CONFIG_HUGETLB_PMD_PAGE_TABLE_SHARING */ 7305 7306 pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct *vma, 7307 unsigned long addr, pud_t *pud) 7308 { 7309 return NULL; 7310 } 7311 7312 int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma, 7313 unsigned long addr, pte_t *ptep) 7314 { 7315 return 0; 7316 } 7317 7318 void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma, 7319 unsigned long *start, unsigned long *end) 7320 { 7321 } 7322 7323 bool want_pmd_share(struct vm_area_struct *vma, unsigned long addr) 7324 { 7325 return false; 7326 } 7327 #endif /* CONFIG_HUGETLB_PMD_PAGE_TABLE_SHARING */ 7328 7329 #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB 7330 pte_t *huge_pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma, 7331 unsigned long addr, unsigned long sz) 7332 { 7333 pgd_t *pgd; 7334 p4d_t *p4d; 7335 pud_t *pud; 7336 pte_t *pte = NULL; 7337 7338 pgd = pgd_offset(mm, addr); 7339 p4d = p4d_alloc(mm, pgd, addr); 7340 if (!p4d) 7341 return NULL; 7342 pud = pud_alloc(mm, p4d, addr); 7343 if (pud) { 7344 if (sz == PUD_SIZE) { 7345 pte = (pte_t *)pud; 7346 } else { 7347 BUG_ON(sz != PMD_SIZE); 7348 if (want_pmd_share(vma, addr) && pud_none(*pud)) 7349 pte = huge_pmd_share(mm, vma, addr, pud); 7350 else 7351 pte = (pte_t *)pmd_alloc(mm, pud, addr); 7352 } 7353 } 7354 7355 if (pte) { 7356 pte_t pteval = ptep_get_lockless(pte); 7357 7358 BUG_ON(pte_present(pteval) && !pte_huge(pteval)); 7359 } 7360 7361 return pte; 7362 } 7363 7364 /* 7365 * huge_pte_offset() - Walk the page table to resolve the hugepage 7366 * entry at address @addr 7367 * 7368 * Return: Pointer to page table entry (PUD or PMD) for 7369 * address @addr, or NULL if a !p*d_present() entry is encountered and the 7370 * size @sz doesn't match the hugepage size at this level of the page 7371 * table. 7372 */ 7373 pte_t *huge_pte_offset(struct mm_struct *mm, 7374 unsigned long addr, unsigned long sz) 7375 { 7376 pgd_t *pgd; 7377 p4d_t *p4d; 7378 pud_t *pud; 7379 pmd_t *pmd; 7380 7381 pgd = pgd_offset(mm, addr); 7382 if (!pgd_present(*pgd)) 7383 return NULL; 7384 p4d = p4d_offset(pgd, addr); 7385 if (!p4d_present(*p4d)) 7386 return NULL; 7387 7388 pud = pud_offset(p4d, addr); 7389 if (sz == PUD_SIZE) 7390 /* must be pud huge, non-present or none */ 7391 return (pte_t *)pud; 7392 if (!pud_present(*pud)) 7393 return NULL; 7394 /* must have a valid entry and size to go further */ 7395 7396 pmd = pmd_offset(pud, addr); 7397 /* must be pmd huge, non-present or none */ 7398 return (pte_t *)pmd; 7399 } 7400 7401 /* 7402 * Return a mask that can be used to update an address to the last huge 7403 * page in a page table page mapping size. Used to skip non-present 7404 * page table entries when linearly scanning address ranges. Architectures 7405 * with unique huge page to page table relationships can define their own 7406 * version of this routine. 7407 */ 7408 unsigned long hugetlb_mask_last_page(struct hstate *h) 7409 { 7410 unsigned long hp_size = huge_page_size(h); 7411 7412 if (hp_size == PUD_SIZE) 7413 return P4D_SIZE - PUD_SIZE; 7414 else if (hp_size == PMD_SIZE) 7415 return PUD_SIZE - PMD_SIZE; 7416 else 7417 return 0UL; 7418 } 7419 7420 #else 7421 7422 /* See description above. Architectures can provide their own version. */ 7423 __weak unsigned long hugetlb_mask_last_page(struct hstate *h) 7424 { 7425 #ifdef CONFIG_HUGETLB_PMD_PAGE_TABLE_SHARING 7426 if (huge_page_size(h) == PMD_SIZE) 7427 return PUD_SIZE - PMD_SIZE; 7428 #endif 7429 return 0UL; 7430 } 7431 7432 #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */ 7433 7434 /** 7435 * folio_isolate_hugetlb - try to isolate an allocated hugetlb folio 7436 * @folio: the folio to isolate 7437 * @list: the list to add the folio to on success 7438 * 7439 * Isolate an allocated (refcount > 0) hugetlb folio, marking it as 7440 * isolated/non-migratable, and moving it from the active list to the 7441 * given list. 7442 * 7443 * Isolation will fail if @folio is not an allocated hugetlb folio, or if 7444 * it is already isolated/non-migratable. 7445 * 7446 * On success, an additional folio reference is taken that must be dropped 7447 * using folio_putback_hugetlb() to undo the isolation. 7448 * 7449 * Return: True if isolation worked, otherwise False. 7450 */ 7451 bool folio_isolate_hugetlb(struct folio *folio, struct list_head *list) 7452 { 7453 bool ret = true; 7454 7455 spin_lock_irq(&hugetlb_lock); 7456 if (!folio_test_hugetlb(folio) || 7457 !folio_test_hugetlb_migratable(folio) || 7458 !folio_try_get(folio)) { 7459 ret = false; 7460 goto unlock; 7461 } 7462 folio_clear_hugetlb_migratable(folio); 7463 list_move_tail(&folio->lru, list); 7464 unlock: 7465 spin_unlock_irq(&hugetlb_lock); 7466 return ret; 7467 } 7468 7469 int get_hwpoison_hugetlb_folio(struct folio *folio, bool *hugetlb, bool unpoison) 7470 { 7471 int ret = 0; 7472 7473 *hugetlb = false; 7474 spin_lock_irq(&hugetlb_lock); 7475 if (folio_test_hugetlb(folio)) { 7476 *hugetlb = true; 7477 if (folio_test_hugetlb_freed(folio)) 7478 ret = 0; 7479 else if (folio_test_hugetlb_migratable(folio) || unpoison) 7480 ret = folio_try_get(folio); 7481 else 7482 ret = -EBUSY; 7483 } 7484 spin_unlock_irq(&hugetlb_lock); 7485 return ret; 7486 } 7487 7488 int get_huge_page_for_hwpoison(unsigned long pfn, int flags, 7489 bool *migratable_cleared) 7490 { 7491 int ret; 7492 7493 spin_lock_irq(&hugetlb_lock); 7494 ret = __get_huge_page_for_hwpoison(pfn, flags, migratable_cleared); 7495 spin_unlock_irq(&hugetlb_lock); 7496 return ret; 7497 } 7498 7499 /** 7500 * folio_putback_hugetlb - unisolate a hugetlb folio 7501 * @folio: the isolated hugetlb folio 7502 * 7503 * Putback/un-isolate the hugetlb folio that was previous isolated using 7504 * folio_isolate_hugetlb(): marking it non-isolated/migratable and putting it 7505 * back onto the active list. 7506 * 7507 * Will drop the additional folio reference obtained through 7508 * folio_isolate_hugetlb(). 7509 */ 7510 void folio_putback_hugetlb(struct folio *folio) 7511 { 7512 spin_lock_irq(&hugetlb_lock); 7513 folio_set_hugetlb_migratable(folio); 7514 list_move_tail(&folio->lru, &(folio_hstate(folio))->hugepage_activelist); 7515 spin_unlock_irq(&hugetlb_lock); 7516 folio_put(folio); 7517 } 7518 7519 void move_hugetlb_state(struct folio *old_folio, struct folio *new_folio, int reason) 7520 { 7521 struct hstate *h = folio_hstate(old_folio); 7522 7523 hugetlb_cgroup_migrate(old_folio, new_folio); 7524 set_page_owner_migrate_reason(&new_folio->page, reason); 7525 7526 /* 7527 * transfer temporary state of the new hugetlb folio. This is 7528 * reverse to other transitions because the newpage is going to 7529 * be final while the old one will be freed so it takes over 7530 * the temporary status. 7531 * 7532 * Also note that we have to transfer the per-node surplus state 7533 * here as well otherwise the global surplus count will not match 7534 * the per-node's. 7535 */ 7536 if (folio_test_hugetlb_temporary(new_folio)) { 7537 int old_nid = folio_nid(old_folio); 7538 int new_nid = folio_nid(new_folio); 7539 7540 folio_set_hugetlb_temporary(old_folio); 7541 folio_clear_hugetlb_temporary(new_folio); 7542 7543 7544 /* 7545 * There is no need to transfer the per-node surplus state 7546 * when we do not cross the node. 7547 */ 7548 if (new_nid == old_nid) 7549 return; 7550 spin_lock_irq(&hugetlb_lock); 7551 if (h->surplus_huge_pages_node[old_nid]) { 7552 h->surplus_huge_pages_node[old_nid]--; 7553 h->surplus_huge_pages_node[new_nid]++; 7554 } 7555 spin_unlock_irq(&hugetlb_lock); 7556 } 7557 7558 /* 7559 * Our old folio is isolated and has "migratable" cleared until it 7560 * is putback. As migration succeeded, set the new folio "migratable" 7561 * and add it to the active list. 7562 */ 7563 spin_lock_irq(&hugetlb_lock); 7564 folio_set_hugetlb_migratable(new_folio); 7565 list_move_tail(&new_folio->lru, &(folio_hstate(new_folio))->hugepage_activelist); 7566 spin_unlock_irq(&hugetlb_lock); 7567 } 7568 7569 static void hugetlb_unshare_pmds(struct vm_area_struct *vma, 7570 unsigned long start, 7571 unsigned long end) 7572 { 7573 struct hstate *h = hstate_vma(vma); 7574 unsigned long sz = huge_page_size(h); 7575 struct mm_struct *mm = vma->vm_mm; 7576 struct mmu_notifier_range range; 7577 unsigned long address; 7578 spinlock_t *ptl; 7579 pte_t *ptep; 7580 7581 if (!(vma->vm_flags & VM_MAYSHARE)) 7582 return; 7583 7584 if (start >= end) 7585 return; 7586 7587 flush_cache_range(vma, start, end); 7588 /* 7589 * No need to call adjust_range_if_pmd_sharing_possible(), because 7590 * we have already done the PUD_SIZE alignment. 7591 */ 7592 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, mm, 7593 start, end); 7594 mmu_notifier_invalidate_range_start(&range); 7595 hugetlb_vma_lock_write(vma); 7596 i_mmap_lock_write(vma->vm_file->f_mapping); 7597 for (address = start; address < end; address += PUD_SIZE) { 7598 ptep = hugetlb_walk(vma, address, sz); 7599 if (!ptep) 7600 continue; 7601 ptl = huge_pte_lock(h, mm, ptep); 7602 huge_pmd_unshare(mm, vma, address, ptep); 7603 spin_unlock(ptl); 7604 } 7605 flush_hugetlb_tlb_range(vma, start, end); 7606 i_mmap_unlock_write(vma->vm_file->f_mapping); 7607 hugetlb_vma_unlock_write(vma); 7608 /* 7609 * No need to call mmu_notifier_arch_invalidate_secondary_tlbs(), see 7610 * Documentation/mm/mmu_notifier.rst. 7611 */ 7612 mmu_notifier_invalidate_range_end(&range); 7613 } 7614 7615 /* 7616 * This function will unconditionally remove all the shared pmd pgtable entries 7617 * within the specific vma for a hugetlbfs memory range. 7618 */ 7619 void hugetlb_unshare_all_pmds(struct vm_area_struct *vma) 7620 { 7621 hugetlb_unshare_pmds(vma, ALIGN(vma->vm_start, PUD_SIZE), 7622 ALIGN_DOWN(vma->vm_end, PUD_SIZE)); 7623 } 7624 7625 #ifdef CONFIG_CMA 7626 static bool cma_reserve_called __initdata; 7627 7628 static int __init cmdline_parse_hugetlb_cma(char *p) 7629 { 7630 int nid, count = 0; 7631 unsigned long tmp; 7632 char *s = p; 7633 7634 while (*s) { 7635 if (sscanf(s, "%lu%n", &tmp, &count) != 1) 7636 break; 7637 7638 if (s[count] == ':') { 7639 if (tmp >= MAX_NUMNODES) 7640 break; 7641 nid = array_index_nospec(tmp, MAX_NUMNODES); 7642 7643 s += count + 1; 7644 tmp = memparse(s, &s); 7645 hugetlb_cma_size_in_node[nid] = tmp; 7646 hugetlb_cma_size += tmp; 7647 7648 /* 7649 * Skip the separator if have one, otherwise 7650 * break the parsing. 7651 */ 7652 if (*s == ',') 7653 s++; 7654 else 7655 break; 7656 } else { 7657 hugetlb_cma_size = memparse(p, &p); 7658 break; 7659 } 7660 } 7661 7662 return 0; 7663 } 7664 7665 early_param("hugetlb_cma", cmdline_parse_hugetlb_cma); 7666 7667 void __init hugetlb_cma_reserve(int order) 7668 { 7669 unsigned long size, reserved, per_node; 7670 bool node_specific_cma_alloc = false; 7671 int nid; 7672 7673 /* 7674 * HugeTLB CMA reservation is required for gigantic 7675 * huge pages which could not be allocated via the 7676 * page allocator. Just warn if there is any change 7677 * breaking this assumption. 7678 */ 7679 VM_WARN_ON(order <= MAX_PAGE_ORDER); 7680 cma_reserve_called = true; 7681 7682 if (!hugetlb_cma_size) 7683 return; 7684 7685 for (nid = 0; nid < MAX_NUMNODES; nid++) { 7686 if (hugetlb_cma_size_in_node[nid] == 0) 7687 continue; 7688 7689 if (!node_online(nid)) { 7690 pr_warn("hugetlb_cma: invalid node %d specified\n", nid); 7691 hugetlb_cma_size -= hugetlb_cma_size_in_node[nid]; 7692 hugetlb_cma_size_in_node[nid] = 0; 7693 continue; 7694 } 7695 7696 if (hugetlb_cma_size_in_node[nid] < (PAGE_SIZE << order)) { 7697 pr_warn("hugetlb_cma: cma area of node %d should be at least %lu MiB\n", 7698 nid, (PAGE_SIZE << order) / SZ_1M); 7699 hugetlb_cma_size -= hugetlb_cma_size_in_node[nid]; 7700 hugetlb_cma_size_in_node[nid] = 0; 7701 } else { 7702 node_specific_cma_alloc = true; 7703 } 7704 } 7705 7706 /* Validate the CMA size again in case some invalid nodes specified. */ 7707 if (!hugetlb_cma_size) 7708 return; 7709 7710 if (hugetlb_cma_size < (PAGE_SIZE << order)) { 7711 pr_warn("hugetlb_cma: cma area should be at least %lu MiB\n", 7712 (PAGE_SIZE << order) / SZ_1M); 7713 hugetlb_cma_size = 0; 7714 return; 7715 } 7716 7717 if (!node_specific_cma_alloc) { 7718 /* 7719 * If 3 GB area is requested on a machine with 4 numa nodes, 7720 * let's allocate 1 GB on first three nodes and ignore the last one. 7721 */ 7722 per_node = DIV_ROUND_UP(hugetlb_cma_size, nr_online_nodes); 7723 pr_info("hugetlb_cma: reserve %lu MiB, up to %lu MiB per node\n", 7724 hugetlb_cma_size / SZ_1M, per_node / SZ_1M); 7725 } 7726 7727 reserved = 0; 7728 for_each_online_node(nid) { 7729 int res; 7730 char name[CMA_MAX_NAME]; 7731 7732 if (node_specific_cma_alloc) { 7733 if (hugetlb_cma_size_in_node[nid] == 0) 7734 continue; 7735 7736 size = hugetlb_cma_size_in_node[nid]; 7737 } else { 7738 size = min(per_node, hugetlb_cma_size - reserved); 7739 } 7740 7741 size = round_up(size, PAGE_SIZE << order); 7742 7743 snprintf(name, sizeof(name), "hugetlb%d", nid); 7744 /* 7745 * Note that 'order per bit' is based on smallest size that 7746 * may be returned to CMA allocator in the case of 7747 * huge page demotion. 7748 */ 7749 res = cma_declare_contiguous_nid(0, size, 0, 7750 PAGE_SIZE << order, 7751 HUGETLB_PAGE_ORDER, false, name, 7752 &hugetlb_cma[nid], nid); 7753 if (res) { 7754 pr_warn("hugetlb_cma: reservation failed: err %d, node %d", 7755 res, nid); 7756 continue; 7757 } 7758 7759 reserved += size; 7760 pr_info("hugetlb_cma: reserved %lu MiB on node %d\n", 7761 size / SZ_1M, nid); 7762 7763 if (reserved >= hugetlb_cma_size) 7764 break; 7765 } 7766 7767 if (!reserved) 7768 /* 7769 * hugetlb_cma_size is used to determine if allocations from 7770 * cma are possible. Set to zero if no cma regions are set up. 7771 */ 7772 hugetlb_cma_size = 0; 7773 } 7774 7775 static void __init hugetlb_cma_check(void) 7776 { 7777 if (!hugetlb_cma_size || cma_reserve_called) 7778 return; 7779 7780 pr_warn("hugetlb_cma: the option isn't supported by current arch\n"); 7781 } 7782 7783 #endif /* CONFIG_CMA */ 7784