1 // SPDX-License-Identifier: GPL-2.0-only 2 /* 3 * Simple NUMA memory policy for the Linux kernel. 4 * 5 * Copyright 2003,2004 Andi Kleen, SuSE Labs. 6 * (C) Copyright 2005 Christoph Lameter, Silicon Graphics, Inc. 7 * 8 * NUMA policy allows the user to give hints in which node(s) memory should 9 * be allocated. 10 * 11 * Support four policies per VMA and per process: 12 * 13 * The VMA policy has priority over the process policy for a page fault. 14 * 15 * interleave Allocate memory interleaved over a set of nodes, 16 * with normal fallback if it fails. 17 * For VMA based allocations this interleaves based on the 18 * offset into the backing object or offset into the mapping 19 * for anonymous memory. For process policy an process counter 20 * is used. 21 * 22 * bind Only allocate memory on a specific set of nodes, 23 * no fallback. 24 * FIXME: memory is allocated starting with the first node 25 * to the last. It would be better if bind would truly restrict 26 * the allocation to memory nodes instead 27 * 28 * preferred Try a specific node first before normal fallback. 29 * As a special case NUMA_NO_NODE here means do the allocation 30 * on the local CPU. This is normally identical to default, 31 * but useful to set in a VMA when you have a non default 32 * process policy. 33 * 34 * default Allocate on the local node first, or when on a VMA 35 * use the process policy. This is what Linux always did 36 * in a NUMA aware kernel and still does by, ahem, default. 37 * 38 * The process policy is applied for most non interrupt memory allocations 39 * in that process' context. Interrupts ignore the policies and always 40 * try to allocate on the local CPU. The VMA policy is only applied for memory 41 * allocations for a VMA in the VM. 42 * 43 * Currently there are a few corner cases in swapping where the policy 44 * is not applied, but the majority should be handled. When process policy 45 * is used it is not remembered over swap outs/swap ins. 46 * 47 * Only the highest zone in the zone hierarchy gets policied. Allocations 48 * requesting a lower zone just use default policy. This implies that 49 * on systems with highmem kernel lowmem allocation don't get policied. 50 * Same with GFP_DMA allocations. 51 * 52 * For shmfs/tmpfs/hugetlbfs shared memory the policy is shared between 53 * all users and remembered even when nobody has memory mapped. 54 */ 55 56 /* Notebook: 57 fix mmap readahead to honour policy and enable policy for any page cache 58 object 59 statistics for bigpages 60 global policy for page cache? currently it uses process policy. Requires 61 first item above. 62 handle mremap for shared memory (currently ignored for the policy) 63 grows down? 64 make bind policy root only? It can trigger oom much faster and the 65 kernel is not always grateful with that. 66 */ 67 68 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt 69 70 #include <linux/mempolicy.h> 71 #include <linux/pagewalk.h> 72 #include <linux/highmem.h> 73 #include <linux/hugetlb.h> 74 #include <linux/kernel.h> 75 #include <linux/sched.h> 76 #include <linux/sched/mm.h> 77 #include <linux/sched/numa_balancing.h> 78 #include <linux/sched/task.h> 79 #include <linux/nodemask.h> 80 #include <linux/cpuset.h> 81 #include <linux/slab.h> 82 #include <linux/string.h> 83 #include <linux/export.h> 84 #include <linux/nsproxy.h> 85 #include <linux/interrupt.h> 86 #include <linux/init.h> 87 #include <linux/compat.h> 88 #include <linux/ptrace.h> 89 #include <linux/swap.h> 90 #include <linux/seq_file.h> 91 #include <linux/proc_fs.h> 92 #include <linux/migrate.h> 93 #include <linux/ksm.h> 94 #include <linux/rmap.h> 95 #include <linux/security.h> 96 #include <linux/syscalls.h> 97 #include <linux/ctype.h> 98 #include <linux/mm_inline.h> 99 #include <linux/mmu_notifier.h> 100 #include <linux/printk.h> 101 #include <linux/swapops.h> 102 103 #include <asm/tlbflush.h> 104 #include <linux/uaccess.h> 105 106 #include "internal.h" 107 108 /* Internal flags */ 109 #define MPOL_MF_DISCONTIG_OK (MPOL_MF_INTERNAL << 0) /* Skip checks for continuous vmas */ 110 #define MPOL_MF_INVERT (MPOL_MF_INTERNAL << 1) /* Invert check for nodemask */ 111 112 static struct kmem_cache *policy_cache; 113 static struct kmem_cache *sn_cache; 114 115 /* Highest zone. An specific allocation for a zone below that is not 116 policied. */ 117 enum zone_type policy_zone = 0; 118 119 /* 120 * run-time system-wide default policy => local allocation 121 */ 122 static struct mempolicy default_policy = { 123 .refcnt = ATOMIC_INIT(1), /* never free it */ 124 .mode = MPOL_PREFERRED, 125 .flags = MPOL_F_LOCAL, 126 }; 127 128 static struct mempolicy preferred_node_policy[MAX_NUMNODES]; 129 130 struct mempolicy *get_task_policy(struct task_struct *p) 131 { 132 struct mempolicy *pol = p->mempolicy; 133 int node; 134 135 if (pol) 136 return pol; 137 138 node = numa_node_id(); 139 if (node != NUMA_NO_NODE) { 140 pol = &preferred_node_policy[node]; 141 /* preferred_node_policy is not initialised early in boot */ 142 if (pol->mode) 143 return pol; 144 } 145 146 return &default_policy; 147 } 148 149 static const struct mempolicy_operations { 150 int (*create)(struct mempolicy *pol, const nodemask_t *nodes); 151 void (*rebind)(struct mempolicy *pol, const nodemask_t *nodes); 152 } mpol_ops[MPOL_MAX]; 153 154 static inline int mpol_store_user_nodemask(const struct mempolicy *pol) 155 { 156 return pol->flags & MPOL_MODE_FLAGS; 157 } 158 159 static void mpol_relative_nodemask(nodemask_t *ret, const nodemask_t *orig, 160 const nodemask_t *rel) 161 { 162 nodemask_t tmp; 163 nodes_fold(tmp, *orig, nodes_weight(*rel)); 164 nodes_onto(*ret, tmp, *rel); 165 } 166 167 static int mpol_new_interleave(struct mempolicy *pol, const nodemask_t *nodes) 168 { 169 if (nodes_empty(*nodes)) 170 return -EINVAL; 171 pol->v.nodes = *nodes; 172 return 0; 173 } 174 175 static int mpol_new_preferred(struct mempolicy *pol, const nodemask_t *nodes) 176 { 177 if (!nodes) 178 pol->flags |= MPOL_F_LOCAL; /* local allocation */ 179 else if (nodes_empty(*nodes)) 180 return -EINVAL; /* no allowed nodes */ 181 else 182 pol->v.preferred_node = first_node(*nodes); 183 return 0; 184 } 185 186 static int mpol_new_bind(struct mempolicy *pol, const nodemask_t *nodes) 187 { 188 if (nodes_empty(*nodes)) 189 return -EINVAL; 190 pol->v.nodes = *nodes; 191 return 0; 192 } 193 194 /* 195 * mpol_set_nodemask is called after mpol_new() to set up the nodemask, if 196 * any, for the new policy. mpol_new() has already validated the nodes 197 * parameter with respect to the policy mode and flags. But, we need to 198 * handle an empty nodemask with MPOL_PREFERRED here. 199 * 200 * Must be called holding task's alloc_lock to protect task's mems_allowed 201 * and mempolicy. May also be called holding the mmap_semaphore for write. 202 */ 203 static int mpol_set_nodemask(struct mempolicy *pol, 204 const nodemask_t *nodes, struct nodemask_scratch *nsc) 205 { 206 int ret; 207 208 /* if mode is MPOL_DEFAULT, pol is NULL. This is right. */ 209 if (pol == NULL) 210 return 0; 211 /* Check N_MEMORY */ 212 nodes_and(nsc->mask1, 213 cpuset_current_mems_allowed, node_states[N_MEMORY]); 214 215 VM_BUG_ON(!nodes); 216 if (pol->mode == MPOL_PREFERRED && nodes_empty(*nodes)) 217 nodes = NULL; /* explicit local allocation */ 218 else { 219 if (pol->flags & MPOL_F_RELATIVE_NODES) 220 mpol_relative_nodemask(&nsc->mask2, nodes, &nsc->mask1); 221 else 222 nodes_and(nsc->mask2, *nodes, nsc->mask1); 223 224 if (mpol_store_user_nodemask(pol)) 225 pol->w.user_nodemask = *nodes; 226 else 227 pol->w.cpuset_mems_allowed = 228 cpuset_current_mems_allowed; 229 } 230 231 if (nodes) 232 ret = mpol_ops[pol->mode].create(pol, &nsc->mask2); 233 else 234 ret = mpol_ops[pol->mode].create(pol, NULL); 235 return ret; 236 } 237 238 /* 239 * This function just creates a new policy, does some check and simple 240 * initialization. You must invoke mpol_set_nodemask() to set nodes. 241 */ 242 static struct mempolicy *mpol_new(unsigned short mode, unsigned short flags, 243 nodemask_t *nodes) 244 { 245 struct mempolicy *policy; 246 247 pr_debug("setting mode %d flags %d nodes[0] %lx\n", 248 mode, flags, nodes ? nodes_addr(*nodes)[0] : NUMA_NO_NODE); 249 250 if (mode == MPOL_DEFAULT) { 251 if (nodes && !nodes_empty(*nodes)) 252 return ERR_PTR(-EINVAL); 253 return NULL; 254 } 255 VM_BUG_ON(!nodes); 256 257 /* 258 * MPOL_PREFERRED cannot be used with MPOL_F_STATIC_NODES or 259 * MPOL_F_RELATIVE_NODES if the nodemask is empty (local allocation). 260 * All other modes require a valid pointer to a non-empty nodemask. 261 */ 262 if (mode == MPOL_PREFERRED) { 263 if (nodes_empty(*nodes)) { 264 if (((flags & MPOL_F_STATIC_NODES) || 265 (flags & MPOL_F_RELATIVE_NODES))) 266 return ERR_PTR(-EINVAL); 267 } 268 } else if (mode == MPOL_LOCAL) { 269 if (!nodes_empty(*nodes) || 270 (flags & MPOL_F_STATIC_NODES) || 271 (flags & MPOL_F_RELATIVE_NODES)) 272 return ERR_PTR(-EINVAL); 273 mode = MPOL_PREFERRED; 274 } else if (nodes_empty(*nodes)) 275 return ERR_PTR(-EINVAL); 276 policy = kmem_cache_alloc(policy_cache, GFP_KERNEL); 277 if (!policy) 278 return ERR_PTR(-ENOMEM); 279 atomic_set(&policy->refcnt, 1); 280 policy->mode = mode; 281 policy->flags = flags; 282 283 return policy; 284 } 285 286 /* Slow path of a mpol destructor. */ 287 void __mpol_put(struct mempolicy *p) 288 { 289 if (!atomic_dec_and_test(&p->refcnt)) 290 return; 291 kmem_cache_free(policy_cache, p); 292 } 293 294 static void mpol_rebind_default(struct mempolicy *pol, const nodemask_t *nodes) 295 { 296 } 297 298 static void mpol_rebind_nodemask(struct mempolicy *pol, const nodemask_t *nodes) 299 { 300 nodemask_t tmp; 301 302 if (pol->flags & MPOL_F_STATIC_NODES) 303 nodes_and(tmp, pol->w.user_nodemask, *nodes); 304 else if (pol->flags & MPOL_F_RELATIVE_NODES) 305 mpol_relative_nodemask(&tmp, &pol->w.user_nodemask, nodes); 306 else { 307 nodes_remap(tmp, pol->v.nodes,pol->w.cpuset_mems_allowed, 308 *nodes); 309 pol->w.cpuset_mems_allowed = *nodes; 310 } 311 312 if (nodes_empty(tmp)) 313 tmp = *nodes; 314 315 pol->v.nodes = tmp; 316 } 317 318 static void mpol_rebind_preferred(struct mempolicy *pol, 319 const nodemask_t *nodes) 320 { 321 nodemask_t tmp; 322 323 if (pol->flags & MPOL_F_STATIC_NODES) { 324 int node = first_node(pol->w.user_nodemask); 325 326 if (node_isset(node, *nodes)) { 327 pol->v.preferred_node = node; 328 pol->flags &= ~MPOL_F_LOCAL; 329 } else 330 pol->flags |= MPOL_F_LOCAL; 331 } else if (pol->flags & MPOL_F_RELATIVE_NODES) { 332 mpol_relative_nodemask(&tmp, &pol->w.user_nodemask, nodes); 333 pol->v.preferred_node = first_node(tmp); 334 } else if (!(pol->flags & MPOL_F_LOCAL)) { 335 pol->v.preferred_node = node_remap(pol->v.preferred_node, 336 pol->w.cpuset_mems_allowed, 337 *nodes); 338 pol->w.cpuset_mems_allowed = *nodes; 339 } 340 } 341 342 /* 343 * mpol_rebind_policy - Migrate a policy to a different set of nodes 344 * 345 * Per-vma policies are protected by mmap_sem. Allocations using per-task 346 * policies are protected by task->mems_allowed_seq to prevent a premature 347 * OOM/allocation failure due to parallel nodemask modification. 348 */ 349 static void mpol_rebind_policy(struct mempolicy *pol, const nodemask_t *newmask) 350 { 351 if (!pol) 352 return; 353 if (!mpol_store_user_nodemask(pol) && !(pol->flags & MPOL_F_LOCAL) && 354 nodes_equal(pol->w.cpuset_mems_allowed, *newmask)) 355 return; 356 357 mpol_ops[pol->mode].rebind(pol, newmask); 358 } 359 360 /* 361 * Wrapper for mpol_rebind_policy() that just requires task 362 * pointer, and updates task mempolicy. 363 * 364 * Called with task's alloc_lock held. 365 */ 366 367 void mpol_rebind_task(struct task_struct *tsk, const nodemask_t *new) 368 { 369 mpol_rebind_policy(tsk->mempolicy, new); 370 } 371 372 /* 373 * Rebind each vma in mm to new nodemask. 374 * 375 * Call holding a reference to mm. Takes mm->mmap_sem during call. 376 */ 377 378 void mpol_rebind_mm(struct mm_struct *mm, nodemask_t *new) 379 { 380 struct vm_area_struct *vma; 381 382 down_write(&mm->mmap_sem); 383 for (vma = mm->mmap; vma; vma = vma->vm_next) 384 mpol_rebind_policy(vma->vm_policy, new); 385 up_write(&mm->mmap_sem); 386 } 387 388 static const struct mempolicy_operations mpol_ops[MPOL_MAX] = { 389 [MPOL_DEFAULT] = { 390 .rebind = mpol_rebind_default, 391 }, 392 [MPOL_INTERLEAVE] = { 393 .create = mpol_new_interleave, 394 .rebind = mpol_rebind_nodemask, 395 }, 396 [MPOL_PREFERRED] = { 397 .create = mpol_new_preferred, 398 .rebind = mpol_rebind_preferred, 399 }, 400 [MPOL_BIND] = { 401 .create = mpol_new_bind, 402 .rebind = mpol_rebind_nodemask, 403 }, 404 }; 405 406 static int migrate_page_add(struct page *page, struct list_head *pagelist, 407 unsigned long flags); 408 409 struct queue_pages { 410 struct list_head *pagelist; 411 unsigned long flags; 412 nodemask_t *nmask; 413 unsigned long start; 414 unsigned long end; 415 struct vm_area_struct *first; 416 }; 417 418 /* 419 * Check if the page's nid is in qp->nmask. 420 * 421 * If MPOL_MF_INVERT is set in qp->flags, check if the nid is 422 * in the invert of qp->nmask. 423 */ 424 static inline bool queue_pages_required(struct page *page, 425 struct queue_pages *qp) 426 { 427 int nid = page_to_nid(page); 428 unsigned long flags = qp->flags; 429 430 return node_isset(nid, *qp->nmask) == !(flags & MPOL_MF_INVERT); 431 } 432 433 /* 434 * queue_pages_pmd() has four possible return values: 435 * 0 - pages are placed on the right node or queued successfully. 436 * 1 - there is unmovable page, and MPOL_MF_MOVE* & MPOL_MF_STRICT were 437 * specified. 438 * 2 - THP was split. 439 * -EIO - is migration entry or only MPOL_MF_STRICT was specified and an 440 * existing page was already on a node that does not follow the 441 * policy. 442 */ 443 static int queue_pages_pmd(pmd_t *pmd, spinlock_t *ptl, unsigned long addr, 444 unsigned long end, struct mm_walk *walk) 445 __releases(ptl) 446 { 447 int ret = 0; 448 struct page *page; 449 struct queue_pages *qp = walk->private; 450 unsigned long flags; 451 452 if (unlikely(is_pmd_migration_entry(*pmd))) { 453 ret = -EIO; 454 goto unlock; 455 } 456 page = pmd_page(*pmd); 457 if (is_huge_zero_page(page)) { 458 spin_unlock(ptl); 459 __split_huge_pmd(walk->vma, pmd, addr, false, NULL); 460 ret = 2; 461 goto out; 462 } 463 if (!queue_pages_required(page, qp)) 464 goto unlock; 465 466 flags = qp->flags; 467 /* go to thp migration */ 468 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) { 469 if (!vma_migratable(walk->vma) || 470 migrate_page_add(page, qp->pagelist, flags)) { 471 ret = 1; 472 goto unlock; 473 } 474 } else 475 ret = -EIO; 476 unlock: 477 spin_unlock(ptl); 478 out: 479 return ret; 480 } 481 482 /* 483 * Scan through pages checking if pages follow certain conditions, 484 * and move them to the pagelist if they do. 485 * 486 * queue_pages_pte_range() has three possible return values: 487 * 0 - pages are placed on the right node or queued successfully. 488 * 1 - there is unmovable page, and MPOL_MF_MOVE* & MPOL_MF_STRICT were 489 * specified. 490 * -EIO - only MPOL_MF_STRICT was specified and an existing page was already 491 * on a node that does not follow the policy. 492 */ 493 static int queue_pages_pte_range(pmd_t *pmd, unsigned long addr, 494 unsigned long end, struct mm_walk *walk) 495 { 496 struct vm_area_struct *vma = walk->vma; 497 struct page *page; 498 struct queue_pages *qp = walk->private; 499 unsigned long flags = qp->flags; 500 int ret; 501 bool has_unmovable = false; 502 pte_t *pte; 503 spinlock_t *ptl; 504 505 ptl = pmd_trans_huge_lock(pmd, vma); 506 if (ptl) { 507 ret = queue_pages_pmd(pmd, ptl, addr, end, walk); 508 if (ret != 2) 509 return ret; 510 } 511 /* THP was split, fall through to pte walk */ 512 513 if (pmd_trans_unstable(pmd)) 514 return 0; 515 516 pte = pte_offset_map_lock(walk->mm, pmd, addr, &ptl); 517 for (; addr != end; pte++, addr += PAGE_SIZE) { 518 if (!pte_present(*pte)) 519 continue; 520 page = vm_normal_page(vma, addr, *pte); 521 if (!page) 522 continue; 523 /* 524 * vm_normal_page() filters out zero pages, but there might 525 * still be PageReserved pages to skip, perhaps in a VDSO. 526 */ 527 if (PageReserved(page)) 528 continue; 529 if (!queue_pages_required(page, qp)) 530 continue; 531 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) { 532 /* MPOL_MF_STRICT must be specified if we get here */ 533 if (!vma_migratable(vma)) { 534 has_unmovable = true; 535 break; 536 } 537 538 /* 539 * Do not abort immediately since there may be 540 * temporary off LRU pages in the range. Still 541 * need migrate other LRU pages. 542 */ 543 if (migrate_page_add(page, qp->pagelist, flags)) 544 has_unmovable = true; 545 } else 546 break; 547 } 548 pte_unmap_unlock(pte - 1, ptl); 549 cond_resched(); 550 551 if (has_unmovable) 552 return 1; 553 554 return addr != end ? -EIO : 0; 555 } 556 557 static int queue_pages_hugetlb(pte_t *pte, unsigned long hmask, 558 unsigned long addr, unsigned long end, 559 struct mm_walk *walk) 560 { 561 int ret = 0; 562 #ifdef CONFIG_HUGETLB_PAGE 563 struct queue_pages *qp = walk->private; 564 unsigned long flags = (qp->flags & MPOL_MF_VALID); 565 struct page *page; 566 spinlock_t *ptl; 567 pte_t entry; 568 569 ptl = huge_pte_lock(hstate_vma(walk->vma), walk->mm, pte); 570 entry = huge_ptep_get(pte); 571 if (!pte_present(entry)) 572 goto unlock; 573 page = pte_page(entry); 574 if (!queue_pages_required(page, qp)) 575 goto unlock; 576 577 if (flags == MPOL_MF_STRICT) { 578 /* 579 * STRICT alone means only detecting misplaced page and no 580 * need to further check other vma. 581 */ 582 ret = -EIO; 583 goto unlock; 584 } 585 586 if (!vma_migratable(walk->vma)) { 587 /* 588 * Must be STRICT with MOVE*, otherwise .test_walk() have 589 * stopped walking current vma. 590 * Detecting misplaced page but allow migrating pages which 591 * have been queued. 592 */ 593 ret = 1; 594 goto unlock; 595 } 596 597 /* With MPOL_MF_MOVE, we migrate only unshared hugepage. */ 598 if (flags & (MPOL_MF_MOVE_ALL) || 599 (flags & MPOL_MF_MOVE && page_mapcount(page) == 1)) { 600 if (!isolate_huge_page(page, qp->pagelist) && 601 (flags & MPOL_MF_STRICT)) 602 /* 603 * Failed to isolate page but allow migrating pages 604 * which have been queued. 605 */ 606 ret = 1; 607 } 608 unlock: 609 spin_unlock(ptl); 610 #else 611 BUG(); 612 #endif 613 return ret; 614 } 615 616 #ifdef CONFIG_NUMA_BALANCING 617 /* 618 * This is used to mark a range of virtual addresses to be inaccessible. 619 * These are later cleared by a NUMA hinting fault. Depending on these 620 * faults, pages may be migrated for better NUMA placement. 621 * 622 * This is assuming that NUMA faults are handled using PROT_NONE. If 623 * an architecture makes a different choice, it will need further 624 * changes to the core. 625 */ 626 unsigned long change_prot_numa(struct vm_area_struct *vma, 627 unsigned long addr, unsigned long end) 628 { 629 int nr_updated; 630 631 nr_updated = change_protection(vma, addr, end, PAGE_NONE, MM_CP_PROT_NUMA); 632 if (nr_updated) 633 count_vm_numa_events(NUMA_PTE_UPDATES, nr_updated); 634 635 return nr_updated; 636 } 637 #else 638 static unsigned long change_prot_numa(struct vm_area_struct *vma, 639 unsigned long addr, unsigned long end) 640 { 641 return 0; 642 } 643 #endif /* CONFIG_NUMA_BALANCING */ 644 645 static int queue_pages_test_walk(unsigned long start, unsigned long end, 646 struct mm_walk *walk) 647 { 648 struct vm_area_struct *vma = walk->vma; 649 struct queue_pages *qp = walk->private; 650 unsigned long endvma = vma->vm_end; 651 unsigned long flags = qp->flags; 652 653 /* range check first */ 654 VM_BUG_ON_VMA((vma->vm_start > start) || (vma->vm_end < end), vma); 655 656 if (!qp->first) { 657 qp->first = vma; 658 if (!(flags & MPOL_MF_DISCONTIG_OK) && 659 (qp->start < vma->vm_start)) 660 /* hole at head side of range */ 661 return -EFAULT; 662 } 663 if (!(flags & MPOL_MF_DISCONTIG_OK) && 664 ((vma->vm_end < qp->end) && 665 (!vma->vm_next || vma->vm_end < vma->vm_next->vm_start))) 666 /* hole at middle or tail of range */ 667 return -EFAULT; 668 669 /* 670 * Need check MPOL_MF_STRICT to return -EIO if possible 671 * regardless of vma_migratable 672 */ 673 if (!vma_migratable(vma) && 674 !(flags & MPOL_MF_STRICT)) 675 return 1; 676 677 if (endvma > end) 678 endvma = end; 679 680 if (flags & MPOL_MF_LAZY) { 681 /* Similar to task_numa_work, skip inaccessible VMAs */ 682 if (!is_vm_hugetlb_page(vma) && vma_is_accessible(vma) && 683 !(vma->vm_flags & VM_MIXEDMAP)) 684 change_prot_numa(vma, start, endvma); 685 return 1; 686 } 687 688 /* queue pages from current vma */ 689 if (flags & MPOL_MF_VALID) 690 return 0; 691 return 1; 692 } 693 694 static const struct mm_walk_ops queue_pages_walk_ops = { 695 .hugetlb_entry = queue_pages_hugetlb, 696 .pmd_entry = queue_pages_pte_range, 697 .test_walk = queue_pages_test_walk, 698 }; 699 700 /* 701 * Walk through page tables and collect pages to be migrated. 702 * 703 * If pages found in a given range are on a set of nodes (determined by 704 * @nodes and @flags,) it's isolated and queued to the pagelist which is 705 * passed via @private. 706 * 707 * queue_pages_range() has three possible return values: 708 * 1 - there is unmovable page, but MPOL_MF_MOVE* & MPOL_MF_STRICT were 709 * specified. 710 * 0 - queue pages successfully or no misplaced page. 711 * errno - i.e. misplaced pages with MPOL_MF_STRICT specified (-EIO) or 712 * memory range specified by nodemask and maxnode points outside 713 * your accessible address space (-EFAULT) 714 */ 715 static int 716 queue_pages_range(struct mm_struct *mm, unsigned long start, unsigned long end, 717 nodemask_t *nodes, unsigned long flags, 718 struct list_head *pagelist) 719 { 720 int err; 721 struct queue_pages qp = { 722 .pagelist = pagelist, 723 .flags = flags, 724 .nmask = nodes, 725 .start = start, 726 .end = end, 727 .first = NULL, 728 }; 729 730 err = walk_page_range(mm, start, end, &queue_pages_walk_ops, &qp); 731 732 if (!qp.first) 733 /* whole range in hole */ 734 err = -EFAULT; 735 736 return err; 737 } 738 739 /* 740 * Apply policy to a single VMA 741 * This must be called with the mmap_sem held for writing. 742 */ 743 static int vma_replace_policy(struct vm_area_struct *vma, 744 struct mempolicy *pol) 745 { 746 int err; 747 struct mempolicy *old; 748 struct mempolicy *new; 749 750 pr_debug("vma %lx-%lx/%lx vm_ops %p vm_file %p set_policy %p\n", 751 vma->vm_start, vma->vm_end, vma->vm_pgoff, 752 vma->vm_ops, vma->vm_file, 753 vma->vm_ops ? vma->vm_ops->set_policy : NULL); 754 755 new = mpol_dup(pol); 756 if (IS_ERR(new)) 757 return PTR_ERR(new); 758 759 if (vma->vm_ops && vma->vm_ops->set_policy) { 760 err = vma->vm_ops->set_policy(vma, new); 761 if (err) 762 goto err_out; 763 } 764 765 old = vma->vm_policy; 766 vma->vm_policy = new; /* protected by mmap_sem */ 767 mpol_put(old); 768 769 return 0; 770 err_out: 771 mpol_put(new); 772 return err; 773 } 774 775 /* Step 2: apply policy to a range and do splits. */ 776 static int mbind_range(struct mm_struct *mm, unsigned long start, 777 unsigned long end, struct mempolicy *new_pol) 778 { 779 struct vm_area_struct *next; 780 struct vm_area_struct *prev; 781 struct vm_area_struct *vma; 782 int err = 0; 783 pgoff_t pgoff; 784 unsigned long vmstart; 785 unsigned long vmend; 786 787 vma = find_vma(mm, start); 788 VM_BUG_ON(!vma); 789 790 prev = vma->vm_prev; 791 if (start > vma->vm_start) 792 prev = vma; 793 794 for (; vma && vma->vm_start < end; prev = vma, vma = next) { 795 next = vma->vm_next; 796 vmstart = max(start, vma->vm_start); 797 vmend = min(end, vma->vm_end); 798 799 if (mpol_equal(vma_policy(vma), new_pol)) 800 continue; 801 802 pgoff = vma->vm_pgoff + 803 ((vmstart - vma->vm_start) >> PAGE_SHIFT); 804 prev = vma_merge(mm, prev, vmstart, vmend, vma->vm_flags, 805 vma->anon_vma, vma->vm_file, pgoff, 806 new_pol, vma->vm_userfaultfd_ctx); 807 if (prev) { 808 vma = prev; 809 next = vma->vm_next; 810 if (mpol_equal(vma_policy(vma), new_pol)) 811 continue; 812 /* vma_merge() joined vma && vma->next, case 8 */ 813 goto replace; 814 } 815 if (vma->vm_start != vmstart) { 816 err = split_vma(vma->vm_mm, vma, vmstart, 1); 817 if (err) 818 goto out; 819 } 820 if (vma->vm_end != vmend) { 821 err = split_vma(vma->vm_mm, vma, vmend, 0); 822 if (err) 823 goto out; 824 } 825 replace: 826 err = vma_replace_policy(vma, new_pol); 827 if (err) 828 goto out; 829 } 830 831 out: 832 return err; 833 } 834 835 /* Set the process memory policy */ 836 static long do_set_mempolicy(unsigned short mode, unsigned short flags, 837 nodemask_t *nodes) 838 { 839 struct mempolicy *new, *old; 840 NODEMASK_SCRATCH(scratch); 841 int ret; 842 843 if (!scratch) 844 return -ENOMEM; 845 846 new = mpol_new(mode, flags, nodes); 847 if (IS_ERR(new)) { 848 ret = PTR_ERR(new); 849 goto out; 850 } 851 852 task_lock(current); 853 ret = mpol_set_nodemask(new, nodes, scratch); 854 if (ret) { 855 task_unlock(current); 856 mpol_put(new); 857 goto out; 858 } 859 old = current->mempolicy; 860 current->mempolicy = new; 861 if (new && new->mode == MPOL_INTERLEAVE) 862 current->il_prev = MAX_NUMNODES-1; 863 task_unlock(current); 864 mpol_put(old); 865 ret = 0; 866 out: 867 NODEMASK_SCRATCH_FREE(scratch); 868 return ret; 869 } 870 871 /* 872 * Return nodemask for policy for get_mempolicy() query 873 * 874 * Called with task's alloc_lock held 875 */ 876 static void get_policy_nodemask(struct mempolicy *p, nodemask_t *nodes) 877 { 878 nodes_clear(*nodes); 879 if (p == &default_policy) 880 return; 881 882 switch (p->mode) { 883 case MPOL_BIND: 884 case MPOL_INTERLEAVE: 885 *nodes = p->v.nodes; 886 break; 887 case MPOL_PREFERRED: 888 if (!(p->flags & MPOL_F_LOCAL)) 889 node_set(p->v.preferred_node, *nodes); 890 /* else return empty node mask for local allocation */ 891 break; 892 default: 893 BUG(); 894 } 895 } 896 897 static int lookup_node(struct mm_struct *mm, unsigned long addr) 898 { 899 struct page *p; 900 int err; 901 902 int locked = 1; 903 err = get_user_pages_locked(addr & PAGE_MASK, 1, 0, &p, &locked); 904 if (err >= 0) { 905 err = page_to_nid(p); 906 put_page(p); 907 } 908 if (locked) 909 up_read(&mm->mmap_sem); 910 return err; 911 } 912 913 /* Retrieve NUMA policy */ 914 static long do_get_mempolicy(int *policy, nodemask_t *nmask, 915 unsigned long addr, unsigned long flags) 916 { 917 int err; 918 struct mm_struct *mm = current->mm; 919 struct vm_area_struct *vma = NULL; 920 struct mempolicy *pol = current->mempolicy, *pol_refcount = NULL; 921 922 if (flags & 923 ~(unsigned long)(MPOL_F_NODE|MPOL_F_ADDR|MPOL_F_MEMS_ALLOWED)) 924 return -EINVAL; 925 926 if (flags & MPOL_F_MEMS_ALLOWED) { 927 if (flags & (MPOL_F_NODE|MPOL_F_ADDR)) 928 return -EINVAL; 929 *policy = 0; /* just so it's initialized */ 930 task_lock(current); 931 *nmask = cpuset_current_mems_allowed; 932 task_unlock(current); 933 return 0; 934 } 935 936 if (flags & MPOL_F_ADDR) { 937 /* 938 * Do NOT fall back to task policy if the 939 * vma/shared policy at addr is NULL. We 940 * want to return MPOL_DEFAULT in this case. 941 */ 942 down_read(&mm->mmap_sem); 943 vma = find_vma_intersection(mm, addr, addr+1); 944 if (!vma) { 945 up_read(&mm->mmap_sem); 946 return -EFAULT; 947 } 948 if (vma->vm_ops && vma->vm_ops->get_policy) 949 pol = vma->vm_ops->get_policy(vma, addr); 950 else 951 pol = vma->vm_policy; 952 } else if (addr) 953 return -EINVAL; 954 955 if (!pol) 956 pol = &default_policy; /* indicates default behavior */ 957 958 if (flags & MPOL_F_NODE) { 959 if (flags & MPOL_F_ADDR) { 960 /* 961 * Take a refcount on the mpol, lookup_node() 962 * wil drop the mmap_sem, so after calling 963 * lookup_node() only "pol" remains valid, "vma" 964 * is stale. 965 */ 966 pol_refcount = pol; 967 vma = NULL; 968 mpol_get(pol); 969 err = lookup_node(mm, addr); 970 if (err < 0) 971 goto out; 972 *policy = err; 973 } else if (pol == current->mempolicy && 974 pol->mode == MPOL_INTERLEAVE) { 975 *policy = next_node_in(current->il_prev, pol->v.nodes); 976 } else { 977 err = -EINVAL; 978 goto out; 979 } 980 } else { 981 *policy = pol == &default_policy ? MPOL_DEFAULT : 982 pol->mode; 983 /* 984 * Internal mempolicy flags must be masked off before exposing 985 * the policy to userspace. 986 */ 987 *policy |= (pol->flags & MPOL_MODE_FLAGS); 988 } 989 990 err = 0; 991 if (nmask) { 992 if (mpol_store_user_nodemask(pol)) { 993 *nmask = pol->w.user_nodemask; 994 } else { 995 task_lock(current); 996 get_policy_nodemask(pol, nmask); 997 task_unlock(current); 998 } 999 } 1000 1001 out: 1002 mpol_cond_put(pol); 1003 if (vma) 1004 up_read(&mm->mmap_sem); 1005 if (pol_refcount) 1006 mpol_put(pol_refcount); 1007 return err; 1008 } 1009 1010 #ifdef CONFIG_MIGRATION 1011 /* 1012 * page migration, thp tail pages can be passed. 1013 */ 1014 static int migrate_page_add(struct page *page, struct list_head *pagelist, 1015 unsigned long flags) 1016 { 1017 struct page *head = compound_head(page); 1018 /* 1019 * Avoid migrating a page that is shared with others. 1020 */ 1021 if ((flags & MPOL_MF_MOVE_ALL) || page_mapcount(head) == 1) { 1022 if (!isolate_lru_page(head)) { 1023 list_add_tail(&head->lru, pagelist); 1024 mod_node_page_state(page_pgdat(head), 1025 NR_ISOLATED_ANON + page_is_file_lru(head), 1026 hpage_nr_pages(head)); 1027 } else if (flags & MPOL_MF_STRICT) { 1028 /* 1029 * Non-movable page may reach here. And, there may be 1030 * temporary off LRU pages or non-LRU movable pages. 1031 * Treat them as unmovable pages since they can't be 1032 * isolated, so they can't be moved at the moment. It 1033 * should return -EIO for this case too. 1034 */ 1035 return -EIO; 1036 } 1037 } 1038 1039 return 0; 1040 } 1041 1042 /* page allocation callback for NUMA node migration */ 1043 struct page *alloc_new_node_page(struct page *page, unsigned long node) 1044 { 1045 if (PageHuge(page)) 1046 return alloc_huge_page_node(page_hstate(compound_head(page)), 1047 node); 1048 else if (PageTransHuge(page)) { 1049 struct page *thp; 1050 1051 thp = alloc_pages_node(node, 1052 (GFP_TRANSHUGE | __GFP_THISNODE), 1053 HPAGE_PMD_ORDER); 1054 if (!thp) 1055 return NULL; 1056 prep_transhuge_page(thp); 1057 return thp; 1058 } else 1059 return __alloc_pages_node(node, GFP_HIGHUSER_MOVABLE | 1060 __GFP_THISNODE, 0); 1061 } 1062 1063 /* 1064 * Migrate pages from one node to a target node. 1065 * Returns error or the number of pages not migrated. 1066 */ 1067 static int migrate_to_node(struct mm_struct *mm, int source, int dest, 1068 int flags) 1069 { 1070 nodemask_t nmask; 1071 LIST_HEAD(pagelist); 1072 int err = 0; 1073 1074 nodes_clear(nmask); 1075 node_set(source, nmask); 1076 1077 /* 1078 * This does not "check" the range but isolates all pages that 1079 * need migration. Between passing in the full user address 1080 * space range and MPOL_MF_DISCONTIG_OK, this call can not fail. 1081 */ 1082 VM_BUG_ON(!(flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))); 1083 queue_pages_range(mm, mm->mmap->vm_start, mm->task_size, &nmask, 1084 flags | MPOL_MF_DISCONTIG_OK, &pagelist); 1085 1086 if (!list_empty(&pagelist)) { 1087 err = migrate_pages(&pagelist, alloc_new_node_page, NULL, dest, 1088 MIGRATE_SYNC, MR_SYSCALL); 1089 if (err) 1090 putback_movable_pages(&pagelist); 1091 } 1092 1093 return err; 1094 } 1095 1096 /* 1097 * Move pages between the two nodesets so as to preserve the physical 1098 * layout as much as possible. 1099 * 1100 * Returns the number of page that could not be moved. 1101 */ 1102 int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from, 1103 const nodemask_t *to, int flags) 1104 { 1105 int busy = 0; 1106 int err; 1107 nodemask_t tmp; 1108 1109 err = migrate_prep(); 1110 if (err) 1111 return err; 1112 1113 down_read(&mm->mmap_sem); 1114 1115 /* 1116 * Find a 'source' bit set in 'tmp' whose corresponding 'dest' 1117 * bit in 'to' is not also set in 'tmp'. Clear the found 'source' 1118 * bit in 'tmp', and return that <source, dest> pair for migration. 1119 * The pair of nodemasks 'to' and 'from' define the map. 1120 * 1121 * If no pair of bits is found that way, fallback to picking some 1122 * pair of 'source' and 'dest' bits that are not the same. If the 1123 * 'source' and 'dest' bits are the same, this represents a node 1124 * that will be migrating to itself, so no pages need move. 1125 * 1126 * If no bits are left in 'tmp', or if all remaining bits left 1127 * in 'tmp' correspond to the same bit in 'to', return false 1128 * (nothing left to migrate). 1129 * 1130 * This lets us pick a pair of nodes to migrate between, such that 1131 * if possible the dest node is not already occupied by some other 1132 * source node, minimizing the risk of overloading the memory on a 1133 * node that would happen if we migrated incoming memory to a node 1134 * before migrating outgoing memory source that same node. 1135 * 1136 * A single scan of tmp is sufficient. As we go, we remember the 1137 * most recent <s, d> pair that moved (s != d). If we find a pair 1138 * that not only moved, but what's better, moved to an empty slot 1139 * (d is not set in tmp), then we break out then, with that pair. 1140 * Otherwise when we finish scanning from_tmp, we at least have the 1141 * most recent <s, d> pair that moved. If we get all the way through 1142 * the scan of tmp without finding any node that moved, much less 1143 * moved to an empty node, then there is nothing left worth migrating. 1144 */ 1145 1146 tmp = *from; 1147 while (!nodes_empty(tmp)) { 1148 int s,d; 1149 int source = NUMA_NO_NODE; 1150 int dest = 0; 1151 1152 for_each_node_mask(s, tmp) { 1153 1154 /* 1155 * do_migrate_pages() tries to maintain the relative 1156 * node relationship of the pages established between 1157 * threads and memory areas. 1158 * 1159 * However if the number of source nodes is not equal to 1160 * the number of destination nodes we can not preserve 1161 * this node relative relationship. In that case, skip 1162 * copying memory from a node that is in the destination 1163 * mask. 1164 * 1165 * Example: [2,3,4] -> [3,4,5] moves everything. 1166 * [0-7] - > [3,4,5] moves only 0,1,2,6,7. 1167 */ 1168 1169 if ((nodes_weight(*from) != nodes_weight(*to)) && 1170 (node_isset(s, *to))) 1171 continue; 1172 1173 d = node_remap(s, *from, *to); 1174 if (s == d) 1175 continue; 1176 1177 source = s; /* Node moved. Memorize */ 1178 dest = d; 1179 1180 /* dest not in remaining from nodes? */ 1181 if (!node_isset(dest, tmp)) 1182 break; 1183 } 1184 if (source == NUMA_NO_NODE) 1185 break; 1186 1187 node_clear(source, tmp); 1188 err = migrate_to_node(mm, source, dest, flags); 1189 if (err > 0) 1190 busy += err; 1191 if (err < 0) 1192 break; 1193 } 1194 up_read(&mm->mmap_sem); 1195 if (err < 0) 1196 return err; 1197 return busy; 1198 1199 } 1200 1201 /* 1202 * Allocate a new page for page migration based on vma policy. 1203 * Start by assuming the page is mapped by the same vma as contains @start. 1204 * Search forward from there, if not. N.B., this assumes that the 1205 * list of pages handed to migrate_pages()--which is how we get here-- 1206 * is in virtual address order. 1207 */ 1208 static struct page *new_page(struct page *page, unsigned long start) 1209 { 1210 struct vm_area_struct *vma; 1211 unsigned long uninitialized_var(address); 1212 1213 vma = find_vma(current->mm, start); 1214 while (vma) { 1215 address = page_address_in_vma(page, vma); 1216 if (address != -EFAULT) 1217 break; 1218 vma = vma->vm_next; 1219 } 1220 1221 if (PageHuge(page)) { 1222 return alloc_huge_page_vma(page_hstate(compound_head(page)), 1223 vma, address); 1224 } else if (PageTransHuge(page)) { 1225 struct page *thp; 1226 1227 thp = alloc_hugepage_vma(GFP_TRANSHUGE, vma, address, 1228 HPAGE_PMD_ORDER); 1229 if (!thp) 1230 return NULL; 1231 prep_transhuge_page(thp); 1232 return thp; 1233 } 1234 /* 1235 * if !vma, alloc_page_vma() will use task or system default policy 1236 */ 1237 return alloc_page_vma(GFP_HIGHUSER_MOVABLE | __GFP_RETRY_MAYFAIL, 1238 vma, address); 1239 } 1240 #else 1241 1242 static int migrate_page_add(struct page *page, struct list_head *pagelist, 1243 unsigned long flags) 1244 { 1245 return -EIO; 1246 } 1247 1248 int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from, 1249 const nodemask_t *to, int flags) 1250 { 1251 return -ENOSYS; 1252 } 1253 1254 static struct page *new_page(struct page *page, unsigned long start) 1255 { 1256 return NULL; 1257 } 1258 #endif 1259 1260 static long do_mbind(unsigned long start, unsigned long len, 1261 unsigned short mode, unsigned short mode_flags, 1262 nodemask_t *nmask, unsigned long flags) 1263 { 1264 struct mm_struct *mm = current->mm; 1265 struct mempolicy *new; 1266 unsigned long end; 1267 int err; 1268 int ret; 1269 LIST_HEAD(pagelist); 1270 1271 if (flags & ~(unsigned long)MPOL_MF_VALID) 1272 return -EINVAL; 1273 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE)) 1274 return -EPERM; 1275 1276 if (start & ~PAGE_MASK) 1277 return -EINVAL; 1278 1279 if (mode == MPOL_DEFAULT) 1280 flags &= ~MPOL_MF_STRICT; 1281 1282 len = (len + PAGE_SIZE - 1) & PAGE_MASK; 1283 end = start + len; 1284 1285 if (end < start) 1286 return -EINVAL; 1287 if (end == start) 1288 return 0; 1289 1290 new = mpol_new(mode, mode_flags, nmask); 1291 if (IS_ERR(new)) 1292 return PTR_ERR(new); 1293 1294 if (flags & MPOL_MF_LAZY) 1295 new->flags |= MPOL_F_MOF; 1296 1297 /* 1298 * If we are using the default policy then operation 1299 * on discontinuous address spaces is okay after all 1300 */ 1301 if (!new) 1302 flags |= MPOL_MF_DISCONTIG_OK; 1303 1304 pr_debug("mbind %lx-%lx mode:%d flags:%d nodes:%lx\n", 1305 start, start + len, mode, mode_flags, 1306 nmask ? nodes_addr(*nmask)[0] : NUMA_NO_NODE); 1307 1308 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) { 1309 1310 err = migrate_prep(); 1311 if (err) 1312 goto mpol_out; 1313 } 1314 { 1315 NODEMASK_SCRATCH(scratch); 1316 if (scratch) { 1317 down_write(&mm->mmap_sem); 1318 task_lock(current); 1319 err = mpol_set_nodemask(new, nmask, scratch); 1320 task_unlock(current); 1321 if (err) 1322 up_write(&mm->mmap_sem); 1323 } else 1324 err = -ENOMEM; 1325 NODEMASK_SCRATCH_FREE(scratch); 1326 } 1327 if (err) 1328 goto mpol_out; 1329 1330 ret = queue_pages_range(mm, start, end, nmask, 1331 flags | MPOL_MF_INVERT, &pagelist); 1332 1333 if (ret < 0) { 1334 err = ret; 1335 goto up_out; 1336 } 1337 1338 err = mbind_range(mm, start, end, new); 1339 1340 if (!err) { 1341 int nr_failed = 0; 1342 1343 if (!list_empty(&pagelist)) { 1344 WARN_ON_ONCE(flags & MPOL_MF_LAZY); 1345 nr_failed = migrate_pages(&pagelist, new_page, NULL, 1346 start, MIGRATE_SYNC, MR_MEMPOLICY_MBIND); 1347 if (nr_failed) 1348 putback_movable_pages(&pagelist); 1349 } 1350 1351 if ((ret > 0) || (nr_failed && (flags & MPOL_MF_STRICT))) 1352 err = -EIO; 1353 } else { 1354 up_out: 1355 if (!list_empty(&pagelist)) 1356 putback_movable_pages(&pagelist); 1357 } 1358 1359 up_write(&mm->mmap_sem); 1360 mpol_out: 1361 mpol_put(new); 1362 return err; 1363 } 1364 1365 /* 1366 * User space interface with variable sized bitmaps for nodelists. 1367 */ 1368 1369 /* Copy a node mask from user space. */ 1370 static int get_nodes(nodemask_t *nodes, const unsigned long __user *nmask, 1371 unsigned long maxnode) 1372 { 1373 unsigned long k; 1374 unsigned long t; 1375 unsigned long nlongs; 1376 unsigned long endmask; 1377 1378 --maxnode; 1379 nodes_clear(*nodes); 1380 if (maxnode == 0 || !nmask) 1381 return 0; 1382 if (maxnode > PAGE_SIZE*BITS_PER_BYTE) 1383 return -EINVAL; 1384 1385 nlongs = BITS_TO_LONGS(maxnode); 1386 if ((maxnode % BITS_PER_LONG) == 0) 1387 endmask = ~0UL; 1388 else 1389 endmask = (1UL << (maxnode % BITS_PER_LONG)) - 1; 1390 1391 /* 1392 * When the user specified more nodes than supported just check 1393 * if the non supported part is all zero. 1394 * 1395 * If maxnode have more longs than MAX_NUMNODES, check 1396 * the bits in that area first. And then go through to 1397 * check the rest bits which equal or bigger than MAX_NUMNODES. 1398 * Otherwise, just check bits [MAX_NUMNODES, maxnode). 1399 */ 1400 if (nlongs > BITS_TO_LONGS(MAX_NUMNODES)) { 1401 for (k = BITS_TO_LONGS(MAX_NUMNODES); k < nlongs; k++) { 1402 if (get_user(t, nmask + k)) 1403 return -EFAULT; 1404 if (k == nlongs - 1) { 1405 if (t & endmask) 1406 return -EINVAL; 1407 } else if (t) 1408 return -EINVAL; 1409 } 1410 nlongs = BITS_TO_LONGS(MAX_NUMNODES); 1411 endmask = ~0UL; 1412 } 1413 1414 if (maxnode > MAX_NUMNODES && MAX_NUMNODES % BITS_PER_LONG != 0) { 1415 unsigned long valid_mask = endmask; 1416 1417 valid_mask &= ~((1UL << (MAX_NUMNODES % BITS_PER_LONG)) - 1); 1418 if (get_user(t, nmask + nlongs - 1)) 1419 return -EFAULT; 1420 if (t & valid_mask) 1421 return -EINVAL; 1422 } 1423 1424 if (copy_from_user(nodes_addr(*nodes), nmask, nlongs*sizeof(unsigned long))) 1425 return -EFAULT; 1426 nodes_addr(*nodes)[nlongs-1] &= endmask; 1427 return 0; 1428 } 1429 1430 /* Copy a kernel node mask to user space */ 1431 static int copy_nodes_to_user(unsigned long __user *mask, unsigned long maxnode, 1432 nodemask_t *nodes) 1433 { 1434 unsigned long copy = ALIGN(maxnode-1, 64) / 8; 1435 unsigned int nbytes = BITS_TO_LONGS(nr_node_ids) * sizeof(long); 1436 1437 if (copy > nbytes) { 1438 if (copy > PAGE_SIZE) 1439 return -EINVAL; 1440 if (clear_user((char __user *)mask + nbytes, copy - nbytes)) 1441 return -EFAULT; 1442 copy = nbytes; 1443 } 1444 return copy_to_user(mask, nodes_addr(*nodes), copy) ? -EFAULT : 0; 1445 } 1446 1447 static long kernel_mbind(unsigned long start, unsigned long len, 1448 unsigned long mode, const unsigned long __user *nmask, 1449 unsigned long maxnode, unsigned int flags) 1450 { 1451 nodemask_t nodes; 1452 int err; 1453 unsigned short mode_flags; 1454 1455 start = untagged_addr(start); 1456 mode_flags = mode & MPOL_MODE_FLAGS; 1457 mode &= ~MPOL_MODE_FLAGS; 1458 if (mode >= MPOL_MAX) 1459 return -EINVAL; 1460 if ((mode_flags & MPOL_F_STATIC_NODES) && 1461 (mode_flags & MPOL_F_RELATIVE_NODES)) 1462 return -EINVAL; 1463 err = get_nodes(&nodes, nmask, maxnode); 1464 if (err) 1465 return err; 1466 return do_mbind(start, len, mode, mode_flags, &nodes, flags); 1467 } 1468 1469 SYSCALL_DEFINE6(mbind, unsigned long, start, unsigned long, len, 1470 unsigned long, mode, const unsigned long __user *, nmask, 1471 unsigned long, maxnode, unsigned int, flags) 1472 { 1473 return kernel_mbind(start, len, mode, nmask, maxnode, flags); 1474 } 1475 1476 /* Set the process memory policy */ 1477 static long kernel_set_mempolicy(int mode, const unsigned long __user *nmask, 1478 unsigned long maxnode) 1479 { 1480 int err; 1481 nodemask_t nodes; 1482 unsigned short flags; 1483 1484 flags = mode & MPOL_MODE_FLAGS; 1485 mode &= ~MPOL_MODE_FLAGS; 1486 if ((unsigned int)mode >= MPOL_MAX) 1487 return -EINVAL; 1488 if ((flags & MPOL_F_STATIC_NODES) && (flags & MPOL_F_RELATIVE_NODES)) 1489 return -EINVAL; 1490 err = get_nodes(&nodes, nmask, maxnode); 1491 if (err) 1492 return err; 1493 return do_set_mempolicy(mode, flags, &nodes); 1494 } 1495 1496 SYSCALL_DEFINE3(set_mempolicy, int, mode, const unsigned long __user *, nmask, 1497 unsigned long, maxnode) 1498 { 1499 return kernel_set_mempolicy(mode, nmask, maxnode); 1500 } 1501 1502 static int kernel_migrate_pages(pid_t pid, unsigned long maxnode, 1503 const unsigned long __user *old_nodes, 1504 const unsigned long __user *new_nodes) 1505 { 1506 struct mm_struct *mm = NULL; 1507 struct task_struct *task; 1508 nodemask_t task_nodes; 1509 int err; 1510 nodemask_t *old; 1511 nodemask_t *new; 1512 NODEMASK_SCRATCH(scratch); 1513 1514 if (!scratch) 1515 return -ENOMEM; 1516 1517 old = &scratch->mask1; 1518 new = &scratch->mask2; 1519 1520 err = get_nodes(old, old_nodes, maxnode); 1521 if (err) 1522 goto out; 1523 1524 err = get_nodes(new, new_nodes, maxnode); 1525 if (err) 1526 goto out; 1527 1528 /* Find the mm_struct */ 1529 rcu_read_lock(); 1530 task = pid ? find_task_by_vpid(pid) : current; 1531 if (!task) { 1532 rcu_read_unlock(); 1533 err = -ESRCH; 1534 goto out; 1535 } 1536 get_task_struct(task); 1537 1538 err = -EINVAL; 1539 1540 /* 1541 * Check if this process has the right to modify the specified process. 1542 * Use the regular "ptrace_may_access()" checks. 1543 */ 1544 if (!ptrace_may_access(task, PTRACE_MODE_READ_REALCREDS)) { 1545 rcu_read_unlock(); 1546 err = -EPERM; 1547 goto out_put; 1548 } 1549 rcu_read_unlock(); 1550 1551 task_nodes = cpuset_mems_allowed(task); 1552 /* Is the user allowed to access the target nodes? */ 1553 if (!nodes_subset(*new, task_nodes) && !capable(CAP_SYS_NICE)) { 1554 err = -EPERM; 1555 goto out_put; 1556 } 1557 1558 task_nodes = cpuset_mems_allowed(current); 1559 nodes_and(*new, *new, task_nodes); 1560 if (nodes_empty(*new)) 1561 goto out_put; 1562 1563 err = security_task_movememory(task); 1564 if (err) 1565 goto out_put; 1566 1567 mm = get_task_mm(task); 1568 put_task_struct(task); 1569 1570 if (!mm) { 1571 err = -EINVAL; 1572 goto out; 1573 } 1574 1575 err = do_migrate_pages(mm, old, new, 1576 capable(CAP_SYS_NICE) ? MPOL_MF_MOVE_ALL : MPOL_MF_MOVE); 1577 1578 mmput(mm); 1579 out: 1580 NODEMASK_SCRATCH_FREE(scratch); 1581 1582 return err; 1583 1584 out_put: 1585 put_task_struct(task); 1586 goto out; 1587 1588 } 1589 1590 SYSCALL_DEFINE4(migrate_pages, pid_t, pid, unsigned long, maxnode, 1591 const unsigned long __user *, old_nodes, 1592 const unsigned long __user *, new_nodes) 1593 { 1594 return kernel_migrate_pages(pid, maxnode, old_nodes, new_nodes); 1595 } 1596 1597 1598 /* Retrieve NUMA policy */ 1599 static int kernel_get_mempolicy(int __user *policy, 1600 unsigned long __user *nmask, 1601 unsigned long maxnode, 1602 unsigned long addr, 1603 unsigned long flags) 1604 { 1605 int err; 1606 int uninitialized_var(pval); 1607 nodemask_t nodes; 1608 1609 addr = untagged_addr(addr); 1610 1611 if (nmask != NULL && maxnode < nr_node_ids) 1612 return -EINVAL; 1613 1614 err = do_get_mempolicy(&pval, &nodes, addr, flags); 1615 1616 if (err) 1617 return err; 1618 1619 if (policy && put_user(pval, policy)) 1620 return -EFAULT; 1621 1622 if (nmask) 1623 err = copy_nodes_to_user(nmask, maxnode, &nodes); 1624 1625 return err; 1626 } 1627 1628 SYSCALL_DEFINE5(get_mempolicy, int __user *, policy, 1629 unsigned long __user *, nmask, unsigned long, maxnode, 1630 unsigned long, addr, unsigned long, flags) 1631 { 1632 return kernel_get_mempolicy(policy, nmask, maxnode, addr, flags); 1633 } 1634 1635 #ifdef CONFIG_COMPAT 1636 1637 COMPAT_SYSCALL_DEFINE5(get_mempolicy, int __user *, policy, 1638 compat_ulong_t __user *, nmask, 1639 compat_ulong_t, maxnode, 1640 compat_ulong_t, addr, compat_ulong_t, flags) 1641 { 1642 long err; 1643 unsigned long __user *nm = NULL; 1644 unsigned long nr_bits, alloc_size; 1645 DECLARE_BITMAP(bm, MAX_NUMNODES); 1646 1647 nr_bits = min_t(unsigned long, maxnode-1, nr_node_ids); 1648 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8; 1649 1650 if (nmask) 1651 nm = compat_alloc_user_space(alloc_size); 1652 1653 err = kernel_get_mempolicy(policy, nm, nr_bits+1, addr, flags); 1654 1655 if (!err && nmask) { 1656 unsigned long copy_size; 1657 copy_size = min_t(unsigned long, sizeof(bm), alloc_size); 1658 err = copy_from_user(bm, nm, copy_size); 1659 /* ensure entire bitmap is zeroed */ 1660 err |= clear_user(nmask, ALIGN(maxnode-1, 8) / 8); 1661 err |= compat_put_bitmap(nmask, bm, nr_bits); 1662 } 1663 1664 return err; 1665 } 1666 1667 COMPAT_SYSCALL_DEFINE3(set_mempolicy, int, mode, compat_ulong_t __user *, nmask, 1668 compat_ulong_t, maxnode) 1669 { 1670 unsigned long __user *nm = NULL; 1671 unsigned long nr_bits, alloc_size; 1672 DECLARE_BITMAP(bm, MAX_NUMNODES); 1673 1674 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES); 1675 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8; 1676 1677 if (nmask) { 1678 if (compat_get_bitmap(bm, nmask, nr_bits)) 1679 return -EFAULT; 1680 nm = compat_alloc_user_space(alloc_size); 1681 if (copy_to_user(nm, bm, alloc_size)) 1682 return -EFAULT; 1683 } 1684 1685 return kernel_set_mempolicy(mode, nm, nr_bits+1); 1686 } 1687 1688 COMPAT_SYSCALL_DEFINE6(mbind, compat_ulong_t, start, compat_ulong_t, len, 1689 compat_ulong_t, mode, compat_ulong_t __user *, nmask, 1690 compat_ulong_t, maxnode, compat_ulong_t, flags) 1691 { 1692 unsigned long __user *nm = NULL; 1693 unsigned long nr_bits, alloc_size; 1694 nodemask_t bm; 1695 1696 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES); 1697 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8; 1698 1699 if (nmask) { 1700 if (compat_get_bitmap(nodes_addr(bm), nmask, nr_bits)) 1701 return -EFAULT; 1702 nm = compat_alloc_user_space(alloc_size); 1703 if (copy_to_user(nm, nodes_addr(bm), alloc_size)) 1704 return -EFAULT; 1705 } 1706 1707 return kernel_mbind(start, len, mode, nm, nr_bits+1, flags); 1708 } 1709 1710 COMPAT_SYSCALL_DEFINE4(migrate_pages, compat_pid_t, pid, 1711 compat_ulong_t, maxnode, 1712 const compat_ulong_t __user *, old_nodes, 1713 const compat_ulong_t __user *, new_nodes) 1714 { 1715 unsigned long __user *old = NULL; 1716 unsigned long __user *new = NULL; 1717 nodemask_t tmp_mask; 1718 unsigned long nr_bits; 1719 unsigned long size; 1720 1721 nr_bits = min_t(unsigned long, maxnode - 1, MAX_NUMNODES); 1722 size = ALIGN(nr_bits, BITS_PER_LONG) / 8; 1723 if (old_nodes) { 1724 if (compat_get_bitmap(nodes_addr(tmp_mask), old_nodes, nr_bits)) 1725 return -EFAULT; 1726 old = compat_alloc_user_space(new_nodes ? size * 2 : size); 1727 if (new_nodes) 1728 new = old + size / sizeof(unsigned long); 1729 if (copy_to_user(old, nodes_addr(tmp_mask), size)) 1730 return -EFAULT; 1731 } 1732 if (new_nodes) { 1733 if (compat_get_bitmap(nodes_addr(tmp_mask), new_nodes, nr_bits)) 1734 return -EFAULT; 1735 if (new == NULL) 1736 new = compat_alloc_user_space(size); 1737 if (copy_to_user(new, nodes_addr(tmp_mask), size)) 1738 return -EFAULT; 1739 } 1740 return kernel_migrate_pages(pid, nr_bits + 1, old, new); 1741 } 1742 1743 #endif /* CONFIG_COMPAT */ 1744 1745 bool vma_migratable(struct vm_area_struct *vma) 1746 { 1747 if (vma->vm_flags & (VM_IO | VM_PFNMAP)) 1748 return false; 1749 1750 /* 1751 * DAX device mappings require predictable access latency, so avoid 1752 * incurring periodic faults. 1753 */ 1754 if (vma_is_dax(vma)) 1755 return false; 1756 1757 if (is_vm_hugetlb_page(vma) && 1758 !hugepage_migration_supported(hstate_vma(vma))) 1759 return false; 1760 1761 /* 1762 * Migration allocates pages in the highest zone. If we cannot 1763 * do so then migration (at least from node to node) is not 1764 * possible. 1765 */ 1766 if (vma->vm_file && 1767 gfp_zone(mapping_gfp_mask(vma->vm_file->f_mapping)) 1768 < policy_zone) 1769 return false; 1770 return true; 1771 } 1772 1773 struct mempolicy *__get_vma_policy(struct vm_area_struct *vma, 1774 unsigned long addr) 1775 { 1776 struct mempolicy *pol = NULL; 1777 1778 if (vma) { 1779 if (vma->vm_ops && vma->vm_ops->get_policy) { 1780 pol = vma->vm_ops->get_policy(vma, addr); 1781 } else if (vma->vm_policy) { 1782 pol = vma->vm_policy; 1783 1784 /* 1785 * shmem_alloc_page() passes MPOL_F_SHARED policy with 1786 * a pseudo vma whose vma->vm_ops=NULL. Take a reference 1787 * count on these policies which will be dropped by 1788 * mpol_cond_put() later 1789 */ 1790 if (mpol_needs_cond_ref(pol)) 1791 mpol_get(pol); 1792 } 1793 } 1794 1795 return pol; 1796 } 1797 1798 /* 1799 * get_vma_policy(@vma, @addr) 1800 * @vma: virtual memory area whose policy is sought 1801 * @addr: address in @vma for shared policy lookup 1802 * 1803 * Returns effective policy for a VMA at specified address. 1804 * Falls back to current->mempolicy or system default policy, as necessary. 1805 * Shared policies [those marked as MPOL_F_SHARED] require an extra reference 1806 * count--added by the get_policy() vm_op, as appropriate--to protect against 1807 * freeing by another task. It is the caller's responsibility to free the 1808 * extra reference for shared policies. 1809 */ 1810 static struct mempolicy *get_vma_policy(struct vm_area_struct *vma, 1811 unsigned long addr) 1812 { 1813 struct mempolicy *pol = __get_vma_policy(vma, addr); 1814 1815 if (!pol) 1816 pol = get_task_policy(current); 1817 1818 return pol; 1819 } 1820 1821 bool vma_policy_mof(struct vm_area_struct *vma) 1822 { 1823 struct mempolicy *pol; 1824 1825 if (vma->vm_ops && vma->vm_ops->get_policy) { 1826 bool ret = false; 1827 1828 pol = vma->vm_ops->get_policy(vma, vma->vm_start); 1829 if (pol && (pol->flags & MPOL_F_MOF)) 1830 ret = true; 1831 mpol_cond_put(pol); 1832 1833 return ret; 1834 } 1835 1836 pol = vma->vm_policy; 1837 if (!pol) 1838 pol = get_task_policy(current); 1839 1840 return pol->flags & MPOL_F_MOF; 1841 } 1842 1843 static int apply_policy_zone(struct mempolicy *policy, enum zone_type zone) 1844 { 1845 enum zone_type dynamic_policy_zone = policy_zone; 1846 1847 BUG_ON(dynamic_policy_zone == ZONE_MOVABLE); 1848 1849 /* 1850 * if policy->v.nodes has movable memory only, 1851 * we apply policy when gfp_zone(gfp) = ZONE_MOVABLE only. 1852 * 1853 * policy->v.nodes is intersect with node_states[N_MEMORY]. 1854 * so if the following test faile, it implies 1855 * policy->v.nodes has movable memory only. 1856 */ 1857 if (!nodes_intersects(policy->v.nodes, node_states[N_HIGH_MEMORY])) 1858 dynamic_policy_zone = ZONE_MOVABLE; 1859 1860 return zone >= dynamic_policy_zone; 1861 } 1862 1863 /* 1864 * Return a nodemask representing a mempolicy for filtering nodes for 1865 * page allocation 1866 */ 1867 static nodemask_t *policy_nodemask(gfp_t gfp, struct mempolicy *policy) 1868 { 1869 /* Lower zones don't get a nodemask applied for MPOL_BIND */ 1870 if (unlikely(policy->mode == MPOL_BIND) && 1871 apply_policy_zone(policy, gfp_zone(gfp)) && 1872 cpuset_nodemask_valid_mems_allowed(&policy->v.nodes)) 1873 return &policy->v.nodes; 1874 1875 return NULL; 1876 } 1877 1878 /* Return the node id preferred by the given mempolicy, or the given id */ 1879 static int policy_node(gfp_t gfp, struct mempolicy *policy, 1880 int nd) 1881 { 1882 if (policy->mode == MPOL_PREFERRED && !(policy->flags & MPOL_F_LOCAL)) 1883 nd = policy->v.preferred_node; 1884 else { 1885 /* 1886 * __GFP_THISNODE shouldn't even be used with the bind policy 1887 * because we might easily break the expectation to stay on the 1888 * requested node and not break the policy. 1889 */ 1890 WARN_ON_ONCE(policy->mode == MPOL_BIND && (gfp & __GFP_THISNODE)); 1891 } 1892 1893 return nd; 1894 } 1895 1896 /* Do dynamic interleaving for a process */ 1897 static unsigned interleave_nodes(struct mempolicy *policy) 1898 { 1899 unsigned next; 1900 struct task_struct *me = current; 1901 1902 next = next_node_in(me->il_prev, policy->v.nodes); 1903 if (next < MAX_NUMNODES) 1904 me->il_prev = next; 1905 return next; 1906 } 1907 1908 /* 1909 * Depending on the memory policy provide a node from which to allocate the 1910 * next slab entry. 1911 */ 1912 unsigned int mempolicy_slab_node(void) 1913 { 1914 struct mempolicy *policy; 1915 int node = numa_mem_id(); 1916 1917 if (in_interrupt()) 1918 return node; 1919 1920 policy = current->mempolicy; 1921 if (!policy || policy->flags & MPOL_F_LOCAL) 1922 return node; 1923 1924 switch (policy->mode) { 1925 case MPOL_PREFERRED: 1926 /* 1927 * handled MPOL_F_LOCAL above 1928 */ 1929 return policy->v.preferred_node; 1930 1931 case MPOL_INTERLEAVE: 1932 return interleave_nodes(policy); 1933 1934 case MPOL_BIND: { 1935 struct zoneref *z; 1936 1937 /* 1938 * Follow bind policy behavior and start allocation at the 1939 * first node. 1940 */ 1941 struct zonelist *zonelist; 1942 enum zone_type highest_zoneidx = gfp_zone(GFP_KERNEL); 1943 zonelist = &NODE_DATA(node)->node_zonelists[ZONELIST_FALLBACK]; 1944 z = first_zones_zonelist(zonelist, highest_zoneidx, 1945 &policy->v.nodes); 1946 return z->zone ? zone_to_nid(z->zone) : node; 1947 } 1948 1949 default: 1950 BUG(); 1951 } 1952 } 1953 1954 /* 1955 * Do static interleaving for a VMA with known offset @n. Returns the n'th 1956 * node in pol->v.nodes (starting from n=0), wrapping around if n exceeds the 1957 * number of present nodes. 1958 */ 1959 static unsigned offset_il_node(struct mempolicy *pol, unsigned long n) 1960 { 1961 unsigned nnodes = nodes_weight(pol->v.nodes); 1962 unsigned target; 1963 int i; 1964 int nid; 1965 1966 if (!nnodes) 1967 return numa_node_id(); 1968 target = (unsigned int)n % nnodes; 1969 nid = first_node(pol->v.nodes); 1970 for (i = 0; i < target; i++) 1971 nid = next_node(nid, pol->v.nodes); 1972 return nid; 1973 } 1974 1975 /* Determine a node number for interleave */ 1976 static inline unsigned interleave_nid(struct mempolicy *pol, 1977 struct vm_area_struct *vma, unsigned long addr, int shift) 1978 { 1979 if (vma) { 1980 unsigned long off; 1981 1982 /* 1983 * for small pages, there is no difference between 1984 * shift and PAGE_SHIFT, so the bit-shift is safe. 1985 * for huge pages, since vm_pgoff is in units of small 1986 * pages, we need to shift off the always 0 bits to get 1987 * a useful offset. 1988 */ 1989 BUG_ON(shift < PAGE_SHIFT); 1990 off = vma->vm_pgoff >> (shift - PAGE_SHIFT); 1991 off += (addr - vma->vm_start) >> shift; 1992 return offset_il_node(pol, off); 1993 } else 1994 return interleave_nodes(pol); 1995 } 1996 1997 #ifdef CONFIG_HUGETLBFS 1998 /* 1999 * huge_node(@vma, @addr, @gfp_flags, @mpol) 2000 * @vma: virtual memory area whose policy is sought 2001 * @addr: address in @vma for shared policy lookup and interleave policy 2002 * @gfp_flags: for requested zone 2003 * @mpol: pointer to mempolicy pointer for reference counted mempolicy 2004 * @nodemask: pointer to nodemask pointer for MPOL_BIND nodemask 2005 * 2006 * Returns a nid suitable for a huge page allocation and a pointer 2007 * to the struct mempolicy for conditional unref after allocation. 2008 * If the effective policy is 'BIND, returns a pointer to the mempolicy's 2009 * @nodemask for filtering the zonelist. 2010 * 2011 * Must be protected by read_mems_allowed_begin() 2012 */ 2013 int huge_node(struct vm_area_struct *vma, unsigned long addr, gfp_t gfp_flags, 2014 struct mempolicy **mpol, nodemask_t **nodemask) 2015 { 2016 int nid; 2017 2018 *mpol = get_vma_policy(vma, addr); 2019 *nodemask = NULL; /* assume !MPOL_BIND */ 2020 2021 if (unlikely((*mpol)->mode == MPOL_INTERLEAVE)) { 2022 nid = interleave_nid(*mpol, vma, addr, 2023 huge_page_shift(hstate_vma(vma))); 2024 } else { 2025 nid = policy_node(gfp_flags, *mpol, numa_node_id()); 2026 if ((*mpol)->mode == MPOL_BIND) 2027 *nodemask = &(*mpol)->v.nodes; 2028 } 2029 return nid; 2030 } 2031 2032 /* 2033 * init_nodemask_of_mempolicy 2034 * 2035 * If the current task's mempolicy is "default" [NULL], return 'false' 2036 * to indicate default policy. Otherwise, extract the policy nodemask 2037 * for 'bind' or 'interleave' policy into the argument nodemask, or 2038 * initialize the argument nodemask to contain the single node for 2039 * 'preferred' or 'local' policy and return 'true' to indicate presence 2040 * of non-default mempolicy. 2041 * 2042 * We don't bother with reference counting the mempolicy [mpol_get/put] 2043 * because the current task is examining it's own mempolicy and a task's 2044 * mempolicy is only ever changed by the task itself. 2045 * 2046 * N.B., it is the caller's responsibility to free a returned nodemask. 2047 */ 2048 bool init_nodemask_of_mempolicy(nodemask_t *mask) 2049 { 2050 struct mempolicy *mempolicy; 2051 int nid; 2052 2053 if (!(mask && current->mempolicy)) 2054 return false; 2055 2056 task_lock(current); 2057 mempolicy = current->mempolicy; 2058 switch (mempolicy->mode) { 2059 case MPOL_PREFERRED: 2060 if (mempolicy->flags & MPOL_F_LOCAL) 2061 nid = numa_node_id(); 2062 else 2063 nid = mempolicy->v.preferred_node; 2064 init_nodemask_of_node(mask, nid); 2065 break; 2066 2067 case MPOL_BIND: 2068 case MPOL_INTERLEAVE: 2069 *mask = mempolicy->v.nodes; 2070 break; 2071 2072 default: 2073 BUG(); 2074 } 2075 task_unlock(current); 2076 2077 return true; 2078 } 2079 #endif 2080 2081 /* 2082 * mempolicy_nodemask_intersects 2083 * 2084 * If tsk's mempolicy is "default" [NULL], return 'true' to indicate default 2085 * policy. Otherwise, check for intersection between mask and the policy 2086 * nodemask for 'bind' or 'interleave' policy. For 'perferred' or 'local' 2087 * policy, always return true since it may allocate elsewhere on fallback. 2088 * 2089 * Takes task_lock(tsk) to prevent freeing of its mempolicy. 2090 */ 2091 bool mempolicy_nodemask_intersects(struct task_struct *tsk, 2092 const nodemask_t *mask) 2093 { 2094 struct mempolicy *mempolicy; 2095 bool ret = true; 2096 2097 if (!mask) 2098 return ret; 2099 task_lock(tsk); 2100 mempolicy = tsk->mempolicy; 2101 if (!mempolicy) 2102 goto out; 2103 2104 switch (mempolicy->mode) { 2105 case MPOL_PREFERRED: 2106 /* 2107 * MPOL_PREFERRED and MPOL_F_LOCAL are only preferred nodes to 2108 * allocate from, they may fallback to other nodes when oom. 2109 * Thus, it's possible for tsk to have allocated memory from 2110 * nodes in mask. 2111 */ 2112 break; 2113 case MPOL_BIND: 2114 case MPOL_INTERLEAVE: 2115 ret = nodes_intersects(mempolicy->v.nodes, *mask); 2116 break; 2117 default: 2118 BUG(); 2119 } 2120 out: 2121 task_unlock(tsk); 2122 return ret; 2123 } 2124 2125 /* Allocate a page in interleaved policy. 2126 Own path because it needs to do special accounting. */ 2127 static struct page *alloc_page_interleave(gfp_t gfp, unsigned order, 2128 unsigned nid) 2129 { 2130 struct page *page; 2131 2132 page = __alloc_pages(gfp, order, nid); 2133 /* skip NUMA_INTERLEAVE_HIT counter update if numa stats is disabled */ 2134 if (!static_branch_likely(&vm_numa_stat_key)) 2135 return page; 2136 if (page && page_to_nid(page) == nid) { 2137 preempt_disable(); 2138 __inc_numa_state(page_zone(page), NUMA_INTERLEAVE_HIT); 2139 preempt_enable(); 2140 } 2141 return page; 2142 } 2143 2144 /** 2145 * alloc_pages_vma - Allocate a page for a VMA. 2146 * 2147 * @gfp: 2148 * %GFP_USER user allocation. 2149 * %GFP_KERNEL kernel allocations, 2150 * %GFP_HIGHMEM highmem/user allocations, 2151 * %GFP_FS allocation should not call back into a file system. 2152 * %GFP_ATOMIC don't sleep. 2153 * 2154 * @order:Order of the GFP allocation. 2155 * @vma: Pointer to VMA or NULL if not available. 2156 * @addr: Virtual Address of the allocation. Must be inside the VMA. 2157 * @node: Which node to prefer for allocation (modulo policy). 2158 * @hugepage: for hugepages try only the preferred node if possible 2159 * 2160 * This function allocates a page from the kernel page pool and applies 2161 * a NUMA policy associated with the VMA or the current process. 2162 * When VMA is not NULL caller must hold down_read on the mmap_sem of the 2163 * mm_struct of the VMA to prevent it from going away. Should be used for 2164 * all allocations for pages that will be mapped into user space. Returns 2165 * NULL when no page can be allocated. 2166 */ 2167 struct page * 2168 alloc_pages_vma(gfp_t gfp, int order, struct vm_area_struct *vma, 2169 unsigned long addr, int node, bool hugepage) 2170 { 2171 struct mempolicy *pol; 2172 struct page *page; 2173 int preferred_nid; 2174 nodemask_t *nmask; 2175 2176 pol = get_vma_policy(vma, addr); 2177 2178 if (pol->mode == MPOL_INTERLEAVE) { 2179 unsigned nid; 2180 2181 nid = interleave_nid(pol, vma, addr, PAGE_SHIFT + order); 2182 mpol_cond_put(pol); 2183 page = alloc_page_interleave(gfp, order, nid); 2184 goto out; 2185 } 2186 2187 if (unlikely(IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE) && hugepage)) { 2188 int hpage_node = node; 2189 2190 /* 2191 * For hugepage allocation and non-interleave policy which 2192 * allows the current node (or other explicitly preferred 2193 * node) we only try to allocate from the current/preferred 2194 * node and don't fall back to other nodes, as the cost of 2195 * remote accesses would likely offset THP benefits. 2196 * 2197 * If the policy is interleave, or does not allow the current 2198 * node in its nodemask, we allocate the standard way. 2199 */ 2200 if (pol->mode == MPOL_PREFERRED && !(pol->flags & MPOL_F_LOCAL)) 2201 hpage_node = pol->v.preferred_node; 2202 2203 nmask = policy_nodemask(gfp, pol); 2204 if (!nmask || node_isset(hpage_node, *nmask)) { 2205 mpol_cond_put(pol); 2206 /* 2207 * First, try to allocate THP only on local node, but 2208 * don't reclaim unnecessarily, just compact. 2209 */ 2210 page = __alloc_pages_node(hpage_node, 2211 gfp | __GFP_THISNODE | __GFP_NORETRY, order); 2212 2213 /* 2214 * If hugepage allocations are configured to always 2215 * synchronous compact or the vma has been madvised 2216 * to prefer hugepage backing, retry allowing remote 2217 * memory with both reclaim and compact as well. 2218 */ 2219 if (!page && (gfp & __GFP_DIRECT_RECLAIM)) 2220 page = __alloc_pages_node(hpage_node, 2221 gfp, order); 2222 2223 goto out; 2224 } 2225 } 2226 2227 nmask = policy_nodemask(gfp, pol); 2228 preferred_nid = policy_node(gfp, pol, node); 2229 page = __alloc_pages_nodemask(gfp, order, preferred_nid, nmask); 2230 mpol_cond_put(pol); 2231 out: 2232 return page; 2233 } 2234 EXPORT_SYMBOL(alloc_pages_vma); 2235 2236 /** 2237 * alloc_pages_current - Allocate pages. 2238 * 2239 * @gfp: 2240 * %GFP_USER user allocation, 2241 * %GFP_KERNEL kernel allocation, 2242 * %GFP_HIGHMEM highmem allocation, 2243 * %GFP_FS don't call back into a file system. 2244 * %GFP_ATOMIC don't sleep. 2245 * @order: Power of two of allocation size in pages. 0 is a single page. 2246 * 2247 * Allocate a page from the kernel page pool. When not in 2248 * interrupt context and apply the current process NUMA policy. 2249 * Returns NULL when no page can be allocated. 2250 */ 2251 struct page *alloc_pages_current(gfp_t gfp, unsigned order) 2252 { 2253 struct mempolicy *pol = &default_policy; 2254 struct page *page; 2255 2256 if (!in_interrupt() && !(gfp & __GFP_THISNODE)) 2257 pol = get_task_policy(current); 2258 2259 /* 2260 * No reference counting needed for current->mempolicy 2261 * nor system default_policy 2262 */ 2263 if (pol->mode == MPOL_INTERLEAVE) 2264 page = alloc_page_interleave(gfp, order, interleave_nodes(pol)); 2265 else 2266 page = __alloc_pages_nodemask(gfp, order, 2267 policy_node(gfp, pol, numa_node_id()), 2268 policy_nodemask(gfp, pol)); 2269 2270 return page; 2271 } 2272 EXPORT_SYMBOL(alloc_pages_current); 2273 2274 int vma_dup_policy(struct vm_area_struct *src, struct vm_area_struct *dst) 2275 { 2276 struct mempolicy *pol = mpol_dup(vma_policy(src)); 2277 2278 if (IS_ERR(pol)) 2279 return PTR_ERR(pol); 2280 dst->vm_policy = pol; 2281 return 0; 2282 } 2283 2284 /* 2285 * If mpol_dup() sees current->cpuset == cpuset_being_rebound, then it 2286 * rebinds the mempolicy its copying by calling mpol_rebind_policy() 2287 * with the mems_allowed returned by cpuset_mems_allowed(). This 2288 * keeps mempolicies cpuset relative after its cpuset moves. See 2289 * further kernel/cpuset.c update_nodemask(). 2290 * 2291 * current's mempolicy may be rebinded by the other task(the task that changes 2292 * cpuset's mems), so we needn't do rebind work for current task. 2293 */ 2294 2295 /* Slow path of a mempolicy duplicate */ 2296 struct mempolicy *__mpol_dup(struct mempolicy *old) 2297 { 2298 struct mempolicy *new = kmem_cache_alloc(policy_cache, GFP_KERNEL); 2299 2300 if (!new) 2301 return ERR_PTR(-ENOMEM); 2302 2303 /* task's mempolicy is protected by alloc_lock */ 2304 if (old == current->mempolicy) { 2305 task_lock(current); 2306 *new = *old; 2307 task_unlock(current); 2308 } else 2309 *new = *old; 2310 2311 if (current_cpuset_is_being_rebound()) { 2312 nodemask_t mems = cpuset_mems_allowed(current); 2313 mpol_rebind_policy(new, &mems); 2314 } 2315 atomic_set(&new->refcnt, 1); 2316 return new; 2317 } 2318 2319 /* Slow path of a mempolicy comparison */ 2320 bool __mpol_equal(struct mempolicy *a, struct mempolicy *b) 2321 { 2322 if (!a || !b) 2323 return false; 2324 if (a->mode != b->mode) 2325 return false; 2326 if (a->flags != b->flags) 2327 return false; 2328 if (mpol_store_user_nodemask(a)) 2329 if (!nodes_equal(a->w.user_nodemask, b->w.user_nodemask)) 2330 return false; 2331 2332 switch (a->mode) { 2333 case MPOL_BIND: 2334 case MPOL_INTERLEAVE: 2335 return !!nodes_equal(a->v.nodes, b->v.nodes); 2336 case MPOL_PREFERRED: 2337 /* a's ->flags is the same as b's */ 2338 if (a->flags & MPOL_F_LOCAL) 2339 return true; 2340 return a->v.preferred_node == b->v.preferred_node; 2341 default: 2342 BUG(); 2343 return false; 2344 } 2345 } 2346 2347 /* 2348 * Shared memory backing store policy support. 2349 * 2350 * Remember policies even when nobody has shared memory mapped. 2351 * The policies are kept in Red-Black tree linked from the inode. 2352 * They are protected by the sp->lock rwlock, which should be held 2353 * for any accesses to the tree. 2354 */ 2355 2356 /* 2357 * lookup first element intersecting start-end. Caller holds sp->lock for 2358 * reading or for writing 2359 */ 2360 static struct sp_node * 2361 sp_lookup(struct shared_policy *sp, unsigned long start, unsigned long end) 2362 { 2363 struct rb_node *n = sp->root.rb_node; 2364 2365 while (n) { 2366 struct sp_node *p = rb_entry(n, struct sp_node, nd); 2367 2368 if (start >= p->end) 2369 n = n->rb_right; 2370 else if (end <= p->start) 2371 n = n->rb_left; 2372 else 2373 break; 2374 } 2375 if (!n) 2376 return NULL; 2377 for (;;) { 2378 struct sp_node *w = NULL; 2379 struct rb_node *prev = rb_prev(n); 2380 if (!prev) 2381 break; 2382 w = rb_entry(prev, struct sp_node, nd); 2383 if (w->end <= start) 2384 break; 2385 n = prev; 2386 } 2387 return rb_entry(n, struct sp_node, nd); 2388 } 2389 2390 /* 2391 * Insert a new shared policy into the list. Caller holds sp->lock for 2392 * writing. 2393 */ 2394 static void sp_insert(struct shared_policy *sp, struct sp_node *new) 2395 { 2396 struct rb_node **p = &sp->root.rb_node; 2397 struct rb_node *parent = NULL; 2398 struct sp_node *nd; 2399 2400 while (*p) { 2401 parent = *p; 2402 nd = rb_entry(parent, struct sp_node, nd); 2403 if (new->start < nd->start) 2404 p = &(*p)->rb_left; 2405 else if (new->end > nd->end) 2406 p = &(*p)->rb_right; 2407 else 2408 BUG(); 2409 } 2410 rb_link_node(&new->nd, parent, p); 2411 rb_insert_color(&new->nd, &sp->root); 2412 pr_debug("inserting %lx-%lx: %d\n", new->start, new->end, 2413 new->policy ? new->policy->mode : 0); 2414 } 2415 2416 /* Find shared policy intersecting idx */ 2417 struct mempolicy * 2418 mpol_shared_policy_lookup(struct shared_policy *sp, unsigned long idx) 2419 { 2420 struct mempolicy *pol = NULL; 2421 struct sp_node *sn; 2422 2423 if (!sp->root.rb_node) 2424 return NULL; 2425 read_lock(&sp->lock); 2426 sn = sp_lookup(sp, idx, idx+1); 2427 if (sn) { 2428 mpol_get(sn->policy); 2429 pol = sn->policy; 2430 } 2431 read_unlock(&sp->lock); 2432 return pol; 2433 } 2434 2435 static void sp_free(struct sp_node *n) 2436 { 2437 mpol_put(n->policy); 2438 kmem_cache_free(sn_cache, n); 2439 } 2440 2441 /** 2442 * mpol_misplaced - check whether current page node is valid in policy 2443 * 2444 * @page: page to be checked 2445 * @vma: vm area where page mapped 2446 * @addr: virtual address where page mapped 2447 * 2448 * Lookup current policy node id for vma,addr and "compare to" page's 2449 * node id. 2450 * 2451 * Returns: 2452 * -1 - not misplaced, page is in the right node 2453 * node - node id where the page should be 2454 * 2455 * Policy determination "mimics" alloc_page_vma(). 2456 * Called from fault path where we know the vma and faulting address. 2457 */ 2458 int mpol_misplaced(struct page *page, struct vm_area_struct *vma, unsigned long addr) 2459 { 2460 struct mempolicy *pol; 2461 struct zoneref *z; 2462 int curnid = page_to_nid(page); 2463 unsigned long pgoff; 2464 int thiscpu = raw_smp_processor_id(); 2465 int thisnid = cpu_to_node(thiscpu); 2466 int polnid = NUMA_NO_NODE; 2467 int ret = -1; 2468 2469 pol = get_vma_policy(vma, addr); 2470 if (!(pol->flags & MPOL_F_MOF)) 2471 goto out; 2472 2473 switch (pol->mode) { 2474 case MPOL_INTERLEAVE: 2475 pgoff = vma->vm_pgoff; 2476 pgoff += (addr - vma->vm_start) >> PAGE_SHIFT; 2477 polnid = offset_il_node(pol, pgoff); 2478 break; 2479 2480 case MPOL_PREFERRED: 2481 if (pol->flags & MPOL_F_LOCAL) 2482 polnid = numa_node_id(); 2483 else 2484 polnid = pol->v.preferred_node; 2485 break; 2486 2487 case MPOL_BIND: 2488 2489 /* 2490 * allows binding to multiple nodes. 2491 * use current page if in policy nodemask, 2492 * else select nearest allowed node, if any. 2493 * If no allowed nodes, use current [!misplaced]. 2494 */ 2495 if (node_isset(curnid, pol->v.nodes)) 2496 goto out; 2497 z = first_zones_zonelist( 2498 node_zonelist(numa_node_id(), GFP_HIGHUSER), 2499 gfp_zone(GFP_HIGHUSER), 2500 &pol->v.nodes); 2501 polnid = zone_to_nid(z->zone); 2502 break; 2503 2504 default: 2505 BUG(); 2506 } 2507 2508 /* Migrate the page towards the node whose CPU is referencing it */ 2509 if (pol->flags & MPOL_F_MORON) { 2510 polnid = thisnid; 2511 2512 if (!should_numa_migrate_memory(current, page, curnid, thiscpu)) 2513 goto out; 2514 } 2515 2516 if (curnid != polnid) 2517 ret = polnid; 2518 out: 2519 mpol_cond_put(pol); 2520 2521 return ret; 2522 } 2523 2524 /* 2525 * Drop the (possibly final) reference to task->mempolicy. It needs to be 2526 * dropped after task->mempolicy is set to NULL so that any allocation done as 2527 * part of its kmem_cache_free(), such as by KASAN, doesn't reference a freed 2528 * policy. 2529 */ 2530 void mpol_put_task_policy(struct task_struct *task) 2531 { 2532 struct mempolicy *pol; 2533 2534 task_lock(task); 2535 pol = task->mempolicy; 2536 task->mempolicy = NULL; 2537 task_unlock(task); 2538 mpol_put(pol); 2539 } 2540 2541 static void sp_delete(struct shared_policy *sp, struct sp_node *n) 2542 { 2543 pr_debug("deleting %lx-l%lx\n", n->start, n->end); 2544 rb_erase(&n->nd, &sp->root); 2545 sp_free(n); 2546 } 2547 2548 static void sp_node_init(struct sp_node *node, unsigned long start, 2549 unsigned long end, struct mempolicy *pol) 2550 { 2551 node->start = start; 2552 node->end = end; 2553 node->policy = pol; 2554 } 2555 2556 static struct sp_node *sp_alloc(unsigned long start, unsigned long end, 2557 struct mempolicy *pol) 2558 { 2559 struct sp_node *n; 2560 struct mempolicy *newpol; 2561 2562 n = kmem_cache_alloc(sn_cache, GFP_KERNEL); 2563 if (!n) 2564 return NULL; 2565 2566 newpol = mpol_dup(pol); 2567 if (IS_ERR(newpol)) { 2568 kmem_cache_free(sn_cache, n); 2569 return NULL; 2570 } 2571 newpol->flags |= MPOL_F_SHARED; 2572 sp_node_init(n, start, end, newpol); 2573 2574 return n; 2575 } 2576 2577 /* Replace a policy range. */ 2578 static int shared_policy_replace(struct shared_policy *sp, unsigned long start, 2579 unsigned long end, struct sp_node *new) 2580 { 2581 struct sp_node *n; 2582 struct sp_node *n_new = NULL; 2583 struct mempolicy *mpol_new = NULL; 2584 int ret = 0; 2585 2586 restart: 2587 write_lock(&sp->lock); 2588 n = sp_lookup(sp, start, end); 2589 /* Take care of old policies in the same range. */ 2590 while (n && n->start < end) { 2591 struct rb_node *next = rb_next(&n->nd); 2592 if (n->start >= start) { 2593 if (n->end <= end) 2594 sp_delete(sp, n); 2595 else 2596 n->start = end; 2597 } else { 2598 /* Old policy spanning whole new range. */ 2599 if (n->end > end) { 2600 if (!n_new) 2601 goto alloc_new; 2602 2603 *mpol_new = *n->policy; 2604 atomic_set(&mpol_new->refcnt, 1); 2605 sp_node_init(n_new, end, n->end, mpol_new); 2606 n->end = start; 2607 sp_insert(sp, n_new); 2608 n_new = NULL; 2609 mpol_new = NULL; 2610 break; 2611 } else 2612 n->end = start; 2613 } 2614 if (!next) 2615 break; 2616 n = rb_entry(next, struct sp_node, nd); 2617 } 2618 if (new) 2619 sp_insert(sp, new); 2620 write_unlock(&sp->lock); 2621 ret = 0; 2622 2623 err_out: 2624 if (mpol_new) 2625 mpol_put(mpol_new); 2626 if (n_new) 2627 kmem_cache_free(sn_cache, n_new); 2628 2629 return ret; 2630 2631 alloc_new: 2632 write_unlock(&sp->lock); 2633 ret = -ENOMEM; 2634 n_new = kmem_cache_alloc(sn_cache, GFP_KERNEL); 2635 if (!n_new) 2636 goto err_out; 2637 mpol_new = kmem_cache_alloc(policy_cache, GFP_KERNEL); 2638 if (!mpol_new) 2639 goto err_out; 2640 goto restart; 2641 } 2642 2643 /** 2644 * mpol_shared_policy_init - initialize shared policy for inode 2645 * @sp: pointer to inode shared policy 2646 * @mpol: struct mempolicy to install 2647 * 2648 * Install non-NULL @mpol in inode's shared policy rb-tree. 2649 * On entry, the current task has a reference on a non-NULL @mpol. 2650 * This must be released on exit. 2651 * This is called at get_inode() calls and we can use GFP_KERNEL. 2652 */ 2653 void mpol_shared_policy_init(struct shared_policy *sp, struct mempolicy *mpol) 2654 { 2655 int ret; 2656 2657 sp->root = RB_ROOT; /* empty tree == default mempolicy */ 2658 rwlock_init(&sp->lock); 2659 2660 if (mpol) { 2661 struct vm_area_struct pvma; 2662 struct mempolicy *new; 2663 NODEMASK_SCRATCH(scratch); 2664 2665 if (!scratch) 2666 goto put_mpol; 2667 /* contextualize the tmpfs mount point mempolicy */ 2668 new = mpol_new(mpol->mode, mpol->flags, &mpol->w.user_nodemask); 2669 if (IS_ERR(new)) 2670 goto free_scratch; /* no valid nodemask intersection */ 2671 2672 task_lock(current); 2673 ret = mpol_set_nodemask(new, &mpol->w.user_nodemask, scratch); 2674 task_unlock(current); 2675 if (ret) 2676 goto put_new; 2677 2678 /* Create pseudo-vma that contains just the policy */ 2679 vma_init(&pvma, NULL); 2680 pvma.vm_end = TASK_SIZE; /* policy covers entire file */ 2681 mpol_set_shared_policy(sp, &pvma, new); /* adds ref */ 2682 2683 put_new: 2684 mpol_put(new); /* drop initial ref */ 2685 free_scratch: 2686 NODEMASK_SCRATCH_FREE(scratch); 2687 put_mpol: 2688 mpol_put(mpol); /* drop our incoming ref on sb mpol */ 2689 } 2690 } 2691 2692 int mpol_set_shared_policy(struct shared_policy *info, 2693 struct vm_area_struct *vma, struct mempolicy *npol) 2694 { 2695 int err; 2696 struct sp_node *new = NULL; 2697 unsigned long sz = vma_pages(vma); 2698 2699 pr_debug("set_shared_policy %lx sz %lu %d %d %lx\n", 2700 vma->vm_pgoff, 2701 sz, npol ? npol->mode : -1, 2702 npol ? npol->flags : -1, 2703 npol ? nodes_addr(npol->v.nodes)[0] : NUMA_NO_NODE); 2704 2705 if (npol) { 2706 new = sp_alloc(vma->vm_pgoff, vma->vm_pgoff + sz, npol); 2707 if (!new) 2708 return -ENOMEM; 2709 } 2710 err = shared_policy_replace(info, vma->vm_pgoff, vma->vm_pgoff+sz, new); 2711 if (err && new) 2712 sp_free(new); 2713 return err; 2714 } 2715 2716 /* Free a backing policy store on inode delete. */ 2717 void mpol_free_shared_policy(struct shared_policy *p) 2718 { 2719 struct sp_node *n; 2720 struct rb_node *next; 2721 2722 if (!p->root.rb_node) 2723 return; 2724 write_lock(&p->lock); 2725 next = rb_first(&p->root); 2726 while (next) { 2727 n = rb_entry(next, struct sp_node, nd); 2728 next = rb_next(&n->nd); 2729 sp_delete(p, n); 2730 } 2731 write_unlock(&p->lock); 2732 } 2733 2734 #ifdef CONFIG_NUMA_BALANCING 2735 static int __initdata numabalancing_override; 2736 2737 static void __init check_numabalancing_enable(void) 2738 { 2739 bool numabalancing_default = false; 2740 2741 if (IS_ENABLED(CONFIG_NUMA_BALANCING_DEFAULT_ENABLED)) 2742 numabalancing_default = true; 2743 2744 /* Parsed by setup_numabalancing. override == 1 enables, -1 disables */ 2745 if (numabalancing_override) 2746 set_numabalancing_state(numabalancing_override == 1); 2747 2748 if (num_online_nodes() > 1 && !numabalancing_override) { 2749 pr_info("%s automatic NUMA balancing. Configure with numa_balancing= or the kernel.numa_balancing sysctl\n", 2750 numabalancing_default ? "Enabling" : "Disabling"); 2751 set_numabalancing_state(numabalancing_default); 2752 } 2753 } 2754 2755 static int __init setup_numabalancing(char *str) 2756 { 2757 int ret = 0; 2758 if (!str) 2759 goto out; 2760 2761 if (!strcmp(str, "enable")) { 2762 numabalancing_override = 1; 2763 ret = 1; 2764 } else if (!strcmp(str, "disable")) { 2765 numabalancing_override = -1; 2766 ret = 1; 2767 } 2768 out: 2769 if (!ret) 2770 pr_warn("Unable to parse numa_balancing=\n"); 2771 2772 return ret; 2773 } 2774 __setup("numa_balancing=", setup_numabalancing); 2775 #else 2776 static inline void __init check_numabalancing_enable(void) 2777 { 2778 } 2779 #endif /* CONFIG_NUMA_BALANCING */ 2780 2781 /* assumes fs == KERNEL_DS */ 2782 void __init numa_policy_init(void) 2783 { 2784 nodemask_t interleave_nodes; 2785 unsigned long largest = 0; 2786 int nid, prefer = 0; 2787 2788 policy_cache = kmem_cache_create("numa_policy", 2789 sizeof(struct mempolicy), 2790 0, SLAB_PANIC, NULL); 2791 2792 sn_cache = kmem_cache_create("shared_policy_node", 2793 sizeof(struct sp_node), 2794 0, SLAB_PANIC, NULL); 2795 2796 for_each_node(nid) { 2797 preferred_node_policy[nid] = (struct mempolicy) { 2798 .refcnt = ATOMIC_INIT(1), 2799 .mode = MPOL_PREFERRED, 2800 .flags = MPOL_F_MOF | MPOL_F_MORON, 2801 .v = { .preferred_node = nid, }, 2802 }; 2803 } 2804 2805 /* 2806 * Set interleaving policy for system init. Interleaving is only 2807 * enabled across suitably sized nodes (default is >= 16MB), or 2808 * fall back to the largest node if they're all smaller. 2809 */ 2810 nodes_clear(interleave_nodes); 2811 for_each_node_state(nid, N_MEMORY) { 2812 unsigned long total_pages = node_present_pages(nid); 2813 2814 /* Preserve the largest node */ 2815 if (largest < total_pages) { 2816 largest = total_pages; 2817 prefer = nid; 2818 } 2819 2820 /* Interleave this node? */ 2821 if ((total_pages << PAGE_SHIFT) >= (16 << 20)) 2822 node_set(nid, interleave_nodes); 2823 } 2824 2825 /* All too small, use the largest */ 2826 if (unlikely(nodes_empty(interleave_nodes))) 2827 node_set(prefer, interleave_nodes); 2828 2829 if (do_set_mempolicy(MPOL_INTERLEAVE, 0, &interleave_nodes)) 2830 pr_err("%s: interleaving failed\n", __func__); 2831 2832 check_numabalancing_enable(); 2833 } 2834 2835 /* Reset policy of current process to default */ 2836 void numa_default_policy(void) 2837 { 2838 do_set_mempolicy(MPOL_DEFAULT, 0, NULL); 2839 } 2840 2841 /* 2842 * Parse and format mempolicy from/to strings 2843 */ 2844 2845 /* 2846 * "local" is implemented internally by MPOL_PREFERRED with MPOL_F_LOCAL flag. 2847 */ 2848 static const char * const policy_modes[] = 2849 { 2850 [MPOL_DEFAULT] = "default", 2851 [MPOL_PREFERRED] = "prefer", 2852 [MPOL_BIND] = "bind", 2853 [MPOL_INTERLEAVE] = "interleave", 2854 [MPOL_LOCAL] = "local", 2855 }; 2856 2857 2858 #ifdef CONFIG_TMPFS 2859 /** 2860 * mpol_parse_str - parse string to mempolicy, for tmpfs mpol mount option. 2861 * @str: string containing mempolicy to parse 2862 * @mpol: pointer to struct mempolicy pointer, returned on success. 2863 * 2864 * Format of input: 2865 * <mode>[=<flags>][:<nodelist>] 2866 * 2867 * On success, returns 0, else 1 2868 */ 2869 int mpol_parse_str(char *str, struct mempolicy **mpol) 2870 { 2871 struct mempolicy *new = NULL; 2872 unsigned short mode_flags; 2873 nodemask_t nodes; 2874 char *nodelist = strchr(str, ':'); 2875 char *flags = strchr(str, '='); 2876 int err = 1, mode; 2877 2878 if (flags) 2879 *flags++ = '\0'; /* terminate mode string */ 2880 2881 if (nodelist) { 2882 /* NUL-terminate mode or flags string */ 2883 *nodelist++ = '\0'; 2884 if (nodelist_parse(nodelist, nodes)) 2885 goto out; 2886 if (!nodes_subset(nodes, node_states[N_MEMORY])) 2887 goto out; 2888 } else 2889 nodes_clear(nodes); 2890 2891 mode = match_string(policy_modes, MPOL_MAX, str); 2892 if (mode < 0) 2893 goto out; 2894 2895 switch (mode) { 2896 case MPOL_PREFERRED: 2897 /* 2898 * Insist on a nodelist of one node only, although later 2899 * we use first_node(nodes) to grab a single node, so here 2900 * nodelist (or nodes) cannot be empty. 2901 */ 2902 if (nodelist) { 2903 char *rest = nodelist; 2904 while (isdigit(*rest)) 2905 rest++; 2906 if (*rest) 2907 goto out; 2908 if (nodes_empty(nodes)) 2909 goto out; 2910 } 2911 break; 2912 case MPOL_INTERLEAVE: 2913 /* 2914 * Default to online nodes with memory if no nodelist 2915 */ 2916 if (!nodelist) 2917 nodes = node_states[N_MEMORY]; 2918 break; 2919 case MPOL_LOCAL: 2920 /* 2921 * Don't allow a nodelist; mpol_new() checks flags 2922 */ 2923 if (nodelist) 2924 goto out; 2925 mode = MPOL_PREFERRED; 2926 break; 2927 case MPOL_DEFAULT: 2928 /* 2929 * Insist on a empty nodelist 2930 */ 2931 if (!nodelist) 2932 err = 0; 2933 goto out; 2934 case MPOL_BIND: 2935 /* 2936 * Insist on a nodelist 2937 */ 2938 if (!nodelist) 2939 goto out; 2940 } 2941 2942 mode_flags = 0; 2943 if (flags) { 2944 /* 2945 * Currently, we only support two mutually exclusive 2946 * mode flags. 2947 */ 2948 if (!strcmp(flags, "static")) 2949 mode_flags |= MPOL_F_STATIC_NODES; 2950 else if (!strcmp(flags, "relative")) 2951 mode_flags |= MPOL_F_RELATIVE_NODES; 2952 else 2953 goto out; 2954 } 2955 2956 new = mpol_new(mode, mode_flags, &nodes); 2957 if (IS_ERR(new)) 2958 goto out; 2959 2960 /* 2961 * Save nodes for mpol_to_str() to show the tmpfs mount options 2962 * for /proc/mounts, /proc/pid/mounts and /proc/pid/mountinfo. 2963 */ 2964 if (mode != MPOL_PREFERRED) 2965 new->v.nodes = nodes; 2966 else if (nodelist) 2967 new->v.preferred_node = first_node(nodes); 2968 else 2969 new->flags |= MPOL_F_LOCAL; 2970 2971 /* 2972 * Save nodes for contextualization: this will be used to "clone" 2973 * the mempolicy in a specific context [cpuset] at a later time. 2974 */ 2975 new->w.user_nodemask = nodes; 2976 2977 err = 0; 2978 2979 out: 2980 /* Restore string for error message */ 2981 if (nodelist) 2982 *--nodelist = ':'; 2983 if (flags) 2984 *--flags = '='; 2985 if (!err) 2986 *mpol = new; 2987 return err; 2988 } 2989 #endif /* CONFIG_TMPFS */ 2990 2991 /** 2992 * mpol_to_str - format a mempolicy structure for printing 2993 * @buffer: to contain formatted mempolicy string 2994 * @maxlen: length of @buffer 2995 * @pol: pointer to mempolicy to be formatted 2996 * 2997 * Convert @pol into a string. If @buffer is too short, truncate the string. 2998 * Recommend a @maxlen of at least 32 for the longest mode, "interleave", the 2999 * longest flag, "relative", and to display at least a few node ids. 3000 */ 3001 void mpol_to_str(char *buffer, int maxlen, struct mempolicy *pol) 3002 { 3003 char *p = buffer; 3004 nodemask_t nodes = NODE_MASK_NONE; 3005 unsigned short mode = MPOL_DEFAULT; 3006 unsigned short flags = 0; 3007 3008 if (pol && pol != &default_policy && !(pol->flags & MPOL_F_MORON)) { 3009 mode = pol->mode; 3010 flags = pol->flags; 3011 } 3012 3013 switch (mode) { 3014 case MPOL_DEFAULT: 3015 break; 3016 case MPOL_PREFERRED: 3017 if (flags & MPOL_F_LOCAL) 3018 mode = MPOL_LOCAL; 3019 else 3020 node_set(pol->v.preferred_node, nodes); 3021 break; 3022 case MPOL_BIND: 3023 case MPOL_INTERLEAVE: 3024 nodes = pol->v.nodes; 3025 break; 3026 default: 3027 WARN_ON_ONCE(1); 3028 snprintf(p, maxlen, "unknown"); 3029 return; 3030 } 3031 3032 p += snprintf(p, maxlen, "%s", policy_modes[mode]); 3033 3034 if (flags & MPOL_MODE_FLAGS) { 3035 p += snprintf(p, buffer + maxlen - p, "="); 3036 3037 /* 3038 * Currently, the only defined flags are mutually exclusive 3039 */ 3040 if (flags & MPOL_F_STATIC_NODES) 3041 p += snprintf(p, buffer + maxlen - p, "static"); 3042 else if (flags & MPOL_F_RELATIVE_NODES) 3043 p += snprintf(p, buffer + maxlen - p, "relative"); 3044 } 3045 3046 if (!nodes_empty(nodes)) 3047 p += scnprintf(p, buffer + maxlen - p, ":%*pbl", 3048 nodemask_pr_args(&nodes)); 3049 } 3050