Loading...
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
112static struct kmem_cache *policy_cache;
113static struct kmem_cache *sn_cache;
114
115/* Highest zone. An specific allocation for a zone below that is not
116 policied. */
117enum zone_type policy_zone = 0;
118
119/*
120 * run-time system-wide default policy => local allocation
121 */
122static struct mempolicy default_policy = {
123 .refcnt = ATOMIC_INIT(1), /* never free it */
124 .mode = MPOL_PREFERRED,
125 .flags = MPOL_F_LOCAL,
126};
127
128static struct mempolicy preferred_node_policy[MAX_NUMNODES];
129
130struct 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
149static 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
154static inline int mpol_store_user_nodemask(const struct mempolicy *pol)
155{
156 return pol->flags & MPOL_MODE_FLAGS;
157}
158
159static 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
167static 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
175static 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
186static 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 */
203static 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 */
242static 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. */
287void __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
294static void mpol_rebind_default(struct mempolicy *pol, const nodemask_t *nodes)
295{
296}
297
298static 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
318static 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 */
349static 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
367void 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
378void 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
388static 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
406static int migrate_page_add(struct page *page, struct list_head *pagelist,
407 unsigned long flags);
408
409struct queue_pages {
410 struct list_head *pagelist;
411 unsigned long flags;
412 nodemask_t *nmask;
413 struct vm_area_struct *prev;
414};
415
416/*
417 * Check if the page's nid is in qp->nmask.
418 *
419 * If MPOL_MF_INVERT is set in qp->flags, check if the nid is
420 * in the invert of qp->nmask.
421 */
422static inline bool queue_pages_required(struct page *page,
423 struct queue_pages *qp)
424{
425 int nid = page_to_nid(page);
426 unsigned long flags = qp->flags;
427
428 return node_isset(nid, *qp->nmask) == !(flags & MPOL_MF_INVERT);
429}
430
431/*
432 * queue_pages_pmd() has four possible return values:
433 * 0 - pages are placed on the right node or queued successfully.
434 * 1 - there is unmovable page, and MPOL_MF_MOVE* & MPOL_MF_STRICT were
435 * specified.
436 * 2 - THP was split.
437 * -EIO - is migration entry or only MPOL_MF_STRICT was specified and an
438 * existing page was already on a node that does not follow the
439 * policy.
440 */
441static int queue_pages_pmd(pmd_t *pmd, spinlock_t *ptl, unsigned long addr,
442 unsigned long end, struct mm_walk *walk)
443{
444 int ret = 0;
445 struct page *page;
446 struct queue_pages *qp = walk->private;
447 unsigned long flags;
448
449 if (unlikely(is_pmd_migration_entry(*pmd))) {
450 ret = -EIO;
451 goto unlock;
452 }
453 page = pmd_page(*pmd);
454 if (is_huge_zero_page(page)) {
455 spin_unlock(ptl);
456 __split_huge_pmd(walk->vma, pmd, addr, false, NULL);
457 ret = 2;
458 goto out;
459 }
460 if (!queue_pages_required(page, qp))
461 goto unlock;
462
463 flags = qp->flags;
464 /* go to thp migration */
465 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) {
466 if (!vma_migratable(walk->vma) ||
467 migrate_page_add(page, qp->pagelist, flags)) {
468 ret = 1;
469 goto unlock;
470 }
471 } else
472 ret = -EIO;
473unlock:
474 spin_unlock(ptl);
475out:
476 return ret;
477}
478
479/*
480 * Scan through pages checking if pages follow certain conditions,
481 * and move them to the pagelist if they do.
482 *
483 * queue_pages_pte_range() has three possible return values:
484 * 0 - pages are placed on the right node or queued successfully.
485 * 1 - there is unmovable page, and MPOL_MF_MOVE* & MPOL_MF_STRICT were
486 * specified.
487 * -EIO - only MPOL_MF_STRICT was specified and an existing page was already
488 * on a node that does not follow the policy.
489 */
490static int queue_pages_pte_range(pmd_t *pmd, unsigned long addr,
491 unsigned long end, struct mm_walk *walk)
492{
493 struct vm_area_struct *vma = walk->vma;
494 struct page *page;
495 struct queue_pages *qp = walk->private;
496 unsigned long flags = qp->flags;
497 int ret;
498 bool has_unmovable = false;
499 pte_t *pte;
500 spinlock_t *ptl;
501
502 ptl = pmd_trans_huge_lock(pmd, vma);
503 if (ptl) {
504 ret = queue_pages_pmd(pmd, ptl, addr, end, walk);
505 if (ret != 2)
506 return ret;
507 }
508 /* THP was split, fall through to pte walk */
509
510 if (pmd_trans_unstable(pmd))
511 return 0;
512
513 pte = pte_offset_map_lock(walk->mm, pmd, addr, &ptl);
514 for (; addr != end; pte++, addr += PAGE_SIZE) {
515 if (!pte_present(*pte))
516 continue;
517 page = vm_normal_page(vma, addr, *pte);
518 if (!page)
519 continue;
520 /*
521 * vm_normal_page() filters out zero pages, but there might
522 * still be PageReserved pages to skip, perhaps in a VDSO.
523 */
524 if (PageReserved(page))
525 continue;
526 if (!queue_pages_required(page, qp))
527 continue;
528 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) {
529 /* MPOL_MF_STRICT must be specified if we get here */
530 if (!vma_migratable(vma)) {
531 has_unmovable = true;
532 break;
533 }
534
535 /*
536 * Do not abort immediately since there may be
537 * temporary off LRU pages in the range. Still
538 * need migrate other LRU pages.
539 */
540 if (migrate_page_add(page, qp->pagelist, flags))
541 has_unmovable = true;
542 } else
543 break;
544 }
545 pte_unmap_unlock(pte - 1, ptl);
546 cond_resched();
547
548 if (has_unmovable)
549 return 1;
550
551 return addr != end ? -EIO : 0;
552}
553
554static int queue_pages_hugetlb(pte_t *pte, unsigned long hmask,
555 unsigned long addr, unsigned long end,
556 struct mm_walk *walk)
557{
558#ifdef CONFIG_HUGETLB_PAGE
559 struct queue_pages *qp = walk->private;
560 unsigned long flags = qp->flags;
561 struct page *page;
562 spinlock_t *ptl;
563 pte_t entry;
564
565 ptl = huge_pte_lock(hstate_vma(walk->vma), walk->mm, pte);
566 entry = huge_ptep_get(pte);
567 if (!pte_present(entry))
568 goto unlock;
569 page = pte_page(entry);
570 if (!queue_pages_required(page, qp))
571 goto unlock;
572 /* With MPOL_MF_MOVE, we migrate only unshared hugepage. */
573 if (flags & (MPOL_MF_MOVE_ALL) ||
574 (flags & MPOL_MF_MOVE && page_mapcount(page) == 1))
575 isolate_huge_page(page, qp->pagelist);
576unlock:
577 spin_unlock(ptl);
578#else
579 BUG();
580#endif
581 return 0;
582}
583
584#ifdef CONFIG_NUMA_BALANCING
585/*
586 * This is used to mark a range of virtual addresses to be inaccessible.
587 * These are later cleared by a NUMA hinting fault. Depending on these
588 * faults, pages may be migrated for better NUMA placement.
589 *
590 * This is assuming that NUMA faults are handled using PROT_NONE. If
591 * an architecture makes a different choice, it will need further
592 * changes to the core.
593 */
594unsigned long change_prot_numa(struct vm_area_struct *vma,
595 unsigned long addr, unsigned long end)
596{
597 int nr_updated;
598
599 nr_updated = change_protection(vma, addr, end, PAGE_NONE, 0, 1);
600 if (nr_updated)
601 count_vm_numa_events(NUMA_PTE_UPDATES, nr_updated);
602
603 return nr_updated;
604}
605#else
606static unsigned long change_prot_numa(struct vm_area_struct *vma,
607 unsigned long addr, unsigned long end)
608{
609 return 0;
610}
611#endif /* CONFIG_NUMA_BALANCING */
612
613static int queue_pages_test_walk(unsigned long start, unsigned long end,
614 struct mm_walk *walk)
615{
616 struct vm_area_struct *vma = walk->vma;
617 struct queue_pages *qp = walk->private;
618 unsigned long endvma = vma->vm_end;
619 unsigned long flags = qp->flags;
620
621 /*
622 * Need check MPOL_MF_STRICT to return -EIO if possible
623 * regardless of vma_migratable
624 */
625 if (!vma_migratable(vma) &&
626 !(flags & MPOL_MF_STRICT))
627 return 1;
628
629 if (endvma > end)
630 endvma = end;
631 if (vma->vm_start > start)
632 start = vma->vm_start;
633
634 if (!(flags & MPOL_MF_DISCONTIG_OK)) {
635 if (!vma->vm_next && vma->vm_end < end)
636 return -EFAULT;
637 if (qp->prev && qp->prev->vm_end < vma->vm_start)
638 return -EFAULT;
639 }
640
641 qp->prev = vma;
642
643 if (flags & MPOL_MF_LAZY) {
644 /* Similar to task_numa_work, skip inaccessible VMAs */
645 if (!is_vm_hugetlb_page(vma) &&
646 (vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)) &&
647 !(vma->vm_flags & VM_MIXEDMAP))
648 change_prot_numa(vma, start, endvma);
649 return 1;
650 }
651
652 /* queue pages from current vma */
653 if (flags & MPOL_MF_VALID)
654 return 0;
655 return 1;
656}
657
658static const struct mm_walk_ops queue_pages_walk_ops = {
659 .hugetlb_entry = queue_pages_hugetlb,
660 .pmd_entry = queue_pages_pte_range,
661 .test_walk = queue_pages_test_walk,
662};
663
664/*
665 * Walk through page tables and collect pages to be migrated.
666 *
667 * If pages found in a given range are on a set of nodes (determined by
668 * @nodes and @flags,) it's isolated and queued to the pagelist which is
669 * passed via @private.
670 *
671 * queue_pages_range() has three possible return values:
672 * 1 - there is unmovable page, but MPOL_MF_MOVE* & MPOL_MF_STRICT were
673 * specified.
674 * 0 - queue pages successfully or no misplaced page.
675 * errno - i.e. misplaced pages with MPOL_MF_STRICT specified (-EIO) or
676 * memory range specified by nodemask and maxnode points outside
677 * your accessible address space (-EFAULT)
678 */
679static int
680queue_pages_range(struct mm_struct *mm, unsigned long start, unsigned long end,
681 nodemask_t *nodes, unsigned long flags,
682 struct list_head *pagelist)
683{
684 struct queue_pages qp = {
685 .pagelist = pagelist,
686 .flags = flags,
687 .nmask = nodes,
688 .prev = NULL,
689 };
690
691 return walk_page_range(mm, start, end, &queue_pages_walk_ops, &qp);
692}
693
694/*
695 * Apply policy to a single VMA
696 * This must be called with the mmap_sem held for writing.
697 */
698static int vma_replace_policy(struct vm_area_struct *vma,
699 struct mempolicy *pol)
700{
701 int err;
702 struct mempolicy *old;
703 struct mempolicy *new;
704
705 pr_debug("vma %lx-%lx/%lx vm_ops %p vm_file %p set_policy %p\n",
706 vma->vm_start, vma->vm_end, vma->vm_pgoff,
707 vma->vm_ops, vma->vm_file,
708 vma->vm_ops ? vma->vm_ops->set_policy : NULL);
709
710 new = mpol_dup(pol);
711 if (IS_ERR(new))
712 return PTR_ERR(new);
713
714 if (vma->vm_ops && vma->vm_ops->set_policy) {
715 err = vma->vm_ops->set_policy(vma, new);
716 if (err)
717 goto err_out;
718 }
719
720 old = vma->vm_policy;
721 vma->vm_policy = new; /* protected by mmap_sem */
722 mpol_put(old);
723
724 return 0;
725 err_out:
726 mpol_put(new);
727 return err;
728}
729
730/* Step 2: apply policy to a range and do splits. */
731static int mbind_range(struct mm_struct *mm, unsigned long start,
732 unsigned long end, struct mempolicy *new_pol)
733{
734 struct vm_area_struct *next;
735 struct vm_area_struct *prev;
736 struct vm_area_struct *vma;
737 int err = 0;
738 pgoff_t pgoff;
739 unsigned long vmstart;
740 unsigned long vmend;
741
742 vma = find_vma(mm, start);
743 if (!vma || vma->vm_start > start)
744 return -EFAULT;
745
746 prev = vma->vm_prev;
747 if (start > vma->vm_start)
748 prev = vma;
749
750 for (; vma && vma->vm_start < end; prev = vma, vma = next) {
751 next = vma->vm_next;
752 vmstart = max(start, vma->vm_start);
753 vmend = min(end, vma->vm_end);
754
755 if (mpol_equal(vma_policy(vma), new_pol))
756 continue;
757
758 pgoff = vma->vm_pgoff +
759 ((vmstart - vma->vm_start) >> PAGE_SHIFT);
760 prev = vma_merge(mm, prev, vmstart, vmend, vma->vm_flags,
761 vma->anon_vma, vma->vm_file, pgoff,
762 new_pol, vma->vm_userfaultfd_ctx);
763 if (prev) {
764 vma = prev;
765 next = vma->vm_next;
766 if (mpol_equal(vma_policy(vma), new_pol))
767 continue;
768 /* vma_merge() joined vma && vma->next, case 8 */
769 goto replace;
770 }
771 if (vma->vm_start != vmstart) {
772 err = split_vma(vma->vm_mm, vma, vmstart, 1);
773 if (err)
774 goto out;
775 }
776 if (vma->vm_end != vmend) {
777 err = split_vma(vma->vm_mm, vma, vmend, 0);
778 if (err)
779 goto out;
780 }
781 replace:
782 err = vma_replace_policy(vma, new_pol);
783 if (err)
784 goto out;
785 }
786
787 out:
788 return err;
789}
790
791/* Set the process memory policy */
792static long do_set_mempolicy(unsigned short mode, unsigned short flags,
793 nodemask_t *nodes)
794{
795 struct mempolicy *new, *old;
796 NODEMASK_SCRATCH(scratch);
797 int ret;
798
799 if (!scratch)
800 return -ENOMEM;
801
802 new = mpol_new(mode, flags, nodes);
803 if (IS_ERR(new)) {
804 ret = PTR_ERR(new);
805 goto out;
806 }
807
808 task_lock(current);
809 ret = mpol_set_nodemask(new, nodes, scratch);
810 if (ret) {
811 task_unlock(current);
812 mpol_put(new);
813 goto out;
814 }
815 old = current->mempolicy;
816 current->mempolicy = new;
817 if (new && new->mode == MPOL_INTERLEAVE)
818 current->il_prev = MAX_NUMNODES-1;
819 task_unlock(current);
820 mpol_put(old);
821 ret = 0;
822out:
823 NODEMASK_SCRATCH_FREE(scratch);
824 return ret;
825}
826
827/*
828 * Return nodemask for policy for get_mempolicy() query
829 *
830 * Called with task's alloc_lock held
831 */
832static void get_policy_nodemask(struct mempolicy *p, nodemask_t *nodes)
833{
834 nodes_clear(*nodes);
835 if (p == &default_policy)
836 return;
837
838 switch (p->mode) {
839 case MPOL_BIND:
840 /* Fall through */
841 case MPOL_INTERLEAVE:
842 *nodes = p->v.nodes;
843 break;
844 case MPOL_PREFERRED:
845 if (!(p->flags & MPOL_F_LOCAL))
846 node_set(p->v.preferred_node, *nodes);
847 /* else return empty node mask for local allocation */
848 break;
849 default:
850 BUG();
851 }
852}
853
854static int lookup_node(struct mm_struct *mm, unsigned long addr)
855{
856 struct page *p;
857 int err;
858
859 int locked = 1;
860 err = get_user_pages_locked(addr & PAGE_MASK, 1, 0, &p, &locked);
861 if (err >= 0) {
862 err = page_to_nid(p);
863 put_page(p);
864 }
865 if (locked)
866 up_read(&mm->mmap_sem);
867 return err;
868}
869
870/* Retrieve NUMA policy */
871static long do_get_mempolicy(int *policy, nodemask_t *nmask,
872 unsigned long addr, unsigned long flags)
873{
874 int err;
875 struct mm_struct *mm = current->mm;
876 struct vm_area_struct *vma = NULL;
877 struct mempolicy *pol = current->mempolicy, *pol_refcount = NULL;
878
879 if (flags &
880 ~(unsigned long)(MPOL_F_NODE|MPOL_F_ADDR|MPOL_F_MEMS_ALLOWED))
881 return -EINVAL;
882
883 if (flags & MPOL_F_MEMS_ALLOWED) {
884 if (flags & (MPOL_F_NODE|MPOL_F_ADDR))
885 return -EINVAL;
886 *policy = 0; /* just so it's initialized */
887 task_lock(current);
888 *nmask = cpuset_current_mems_allowed;
889 task_unlock(current);
890 return 0;
891 }
892
893 if (flags & MPOL_F_ADDR) {
894 /*
895 * Do NOT fall back to task policy if the
896 * vma/shared policy at addr is NULL. We
897 * want to return MPOL_DEFAULT in this case.
898 */
899 down_read(&mm->mmap_sem);
900 vma = find_vma_intersection(mm, addr, addr+1);
901 if (!vma) {
902 up_read(&mm->mmap_sem);
903 return -EFAULT;
904 }
905 if (vma->vm_ops && vma->vm_ops->get_policy)
906 pol = vma->vm_ops->get_policy(vma, addr);
907 else
908 pol = vma->vm_policy;
909 } else if (addr)
910 return -EINVAL;
911
912 if (!pol)
913 pol = &default_policy; /* indicates default behavior */
914
915 if (flags & MPOL_F_NODE) {
916 if (flags & MPOL_F_ADDR) {
917 /*
918 * Take a refcount on the mpol, lookup_node()
919 * wil drop the mmap_sem, so after calling
920 * lookup_node() only "pol" remains valid, "vma"
921 * is stale.
922 */
923 pol_refcount = pol;
924 vma = NULL;
925 mpol_get(pol);
926 err = lookup_node(mm, addr);
927 if (err < 0)
928 goto out;
929 *policy = err;
930 } else if (pol == current->mempolicy &&
931 pol->mode == MPOL_INTERLEAVE) {
932 *policy = next_node_in(current->il_prev, pol->v.nodes);
933 } else {
934 err = -EINVAL;
935 goto out;
936 }
937 } else {
938 *policy = pol == &default_policy ? MPOL_DEFAULT :
939 pol->mode;
940 /*
941 * Internal mempolicy flags must be masked off before exposing
942 * the policy to userspace.
943 */
944 *policy |= (pol->flags & MPOL_MODE_FLAGS);
945 }
946
947 err = 0;
948 if (nmask) {
949 if (mpol_store_user_nodemask(pol)) {
950 *nmask = pol->w.user_nodemask;
951 } else {
952 task_lock(current);
953 get_policy_nodemask(pol, nmask);
954 task_unlock(current);
955 }
956 }
957
958 out:
959 mpol_cond_put(pol);
960 if (vma)
961 up_read(&mm->mmap_sem);
962 if (pol_refcount)
963 mpol_put(pol_refcount);
964 return err;
965}
966
967#ifdef CONFIG_MIGRATION
968/*
969 * page migration, thp tail pages can be passed.
970 */
971static int migrate_page_add(struct page *page, struct list_head *pagelist,
972 unsigned long flags)
973{
974 struct page *head = compound_head(page);
975 /*
976 * Avoid migrating a page that is shared with others.
977 */
978 if ((flags & MPOL_MF_MOVE_ALL) || page_mapcount(head) == 1) {
979 if (!isolate_lru_page(head)) {
980 list_add_tail(&head->lru, pagelist);
981 mod_node_page_state(page_pgdat(head),
982 NR_ISOLATED_ANON + page_is_file_cache(head),
983 hpage_nr_pages(head));
984 } else if (flags & MPOL_MF_STRICT) {
985 /*
986 * Non-movable page may reach here. And, there may be
987 * temporary off LRU pages or non-LRU movable pages.
988 * Treat them as unmovable pages since they can't be
989 * isolated, so they can't be moved at the moment. It
990 * should return -EIO for this case too.
991 */
992 return -EIO;
993 }
994 }
995
996 return 0;
997}
998
999/* page allocation callback for NUMA node migration */
1000struct page *alloc_new_node_page(struct page *page, unsigned long node)
1001{
1002 if (PageHuge(page))
1003 return alloc_huge_page_node(page_hstate(compound_head(page)),
1004 node);
1005 else if (PageTransHuge(page)) {
1006 struct page *thp;
1007
1008 thp = alloc_pages_node(node,
1009 (GFP_TRANSHUGE | __GFP_THISNODE),
1010 HPAGE_PMD_ORDER);
1011 if (!thp)
1012 return NULL;
1013 prep_transhuge_page(thp);
1014 return thp;
1015 } else
1016 return __alloc_pages_node(node, GFP_HIGHUSER_MOVABLE |
1017 __GFP_THISNODE, 0);
1018}
1019
1020/*
1021 * Migrate pages from one node to a target node.
1022 * Returns error or the number of pages not migrated.
1023 */
1024static int migrate_to_node(struct mm_struct *mm, int source, int dest,
1025 int flags)
1026{
1027 nodemask_t nmask;
1028 LIST_HEAD(pagelist);
1029 int err = 0;
1030
1031 nodes_clear(nmask);
1032 node_set(source, nmask);
1033
1034 /*
1035 * This does not "check" the range but isolates all pages that
1036 * need migration. Between passing in the full user address
1037 * space range and MPOL_MF_DISCONTIG_OK, this call can not fail.
1038 */
1039 VM_BUG_ON(!(flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)));
1040 queue_pages_range(mm, mm->mmap->vm_start, mm->task_size, &nmask,
1041 flags | MPOL_MF_DISCONTIG_OK, &pagelist);
1042
1043 if (!list_empty(&pagelist)) {
1044 err = migrate_pages(&pagelist, alloc_new_node_page, NULL, dest,
1045 MIGRATE_SYNC, MR_SYSCALL);
1046 if (err)
1047 putback_movable_pages(&pagelist);
1048 }
1049
1050 return err;
1051}
1052
1053/*
1054 * Move pages between the two nodesets so as to preserve the physical
1055 * layout as much as possible.
1056 *
1057 * Returns the number of page that could not be moved.
1058 */
1059int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
1060 const nodemask_t *to, int flags)
1061{
1062 int busy = 0;
1063 int err;
1064 nodemask_t tmp;
1065
1066 err = migrate_prep();
1067 if (err)
1068 return err;
1069
1070 down_read(&mm->mmap_sem);
1071
1072 /*
1073 * Find a 'source' bit set in 'tmp' whose corresponding 'dest'
1074 * bit in 'to' is not also set in 'tmp'. Clear the found 'source'
1075 * bit in 'tmp', and return that <source, dest> pair for migration.
1076 * The pair of nodemasks 'to' and 'from' define the map.
1077 *
1078 * If no pair of bits is found that way, fallback to picking some
1079 * pair of 'source' and 'dest' bits that are not the same. If the
1080 * 'source' and 'dest' bits are the same, this represents a node
1081 * that will be migrating to itself, so no pages need move.
1082 *
1083 * If no bits are left in 'tmp', or if all remaining bits left
1084 * in 'tmp' correspond to the same bit in 'to', return false
1085 * (nothing left to migrate).
1086 *
1087 * This lets us pick a pair of nodes to migrate between, such that
1088 * if possible the dest node is not already occupied by some other
1089 * source node, minimizing the risk of overloading the memory on a
1090 * node that would happen if we migrated incoming memory to a node
1091 * before migrating outgoing memory source that same node.
1092 *
1093 * A single scan of tmp is sufficient. As we go, we remember the
1094 * most recent <s, d> pair that moved (s != d). If we find a pair
1095 * that not only moved, but what's better, moved to an empty slot
1096 * (d is not set in tmp), then we break out then, with that pair.
1097 * Otherwise when we finish scanning from_tmp, we at least have the
1098 * most recent <s, d> pair that moved. If we get all the way through
1099 * the scan of tmp without finding any node that moved, much less
1100 * moved to an empty node, then there is nothing left worth migrating.
1101 */
1102
1103 tmp = *from;
1104 while (!nodes_empty(tmp)) {
1105 int s,d;
1106 int source = NUMA_NO_NODE;
1107 int dest = 0;
1108
1109 for_each_node_mask(s, tmp) {
1110
1111 /*
1112 * do_migrate_pages() tries to maintain the relative
1113 * node relationship of the pages established between
1114 * threads and memory areas.
1115 *
1116 * However if the number of source nodes is not equal to
1117 * the number of destination nodes we can not preserve
1118 * this node relative relationship. In that case, skip
1119 * copying memory from a node that is in the destination
1120 * mask.
1121 *
1122 * Example: [2,3,4] -> [3,4,5] moves everything.
1123 * [0-7] - > [3,4,5] moves only 0,1,2,6,7.
1124 */
1125
1126 if ((nodes_weight(*from) != nodes_weight(*to)) &&
1127 (node_isset(s, *to)))
1128 continue;
1129
1130 d = node_remap(s, *from, *to);
1131 if (s == d)
1132 continue;
1133
1134 source = s; /* Node moved. Memorize */
1135 dest = d;
1136
1137 /* dest not in remaining from nodes? */
1138 if (!node_isset(dest, tmp))
1139 break;
1140 }
1141 if (source == NUMA_NO_NODE)
1142 break;
1143
1144 node_clear(source, tmp);
1145 err = migrate_to_node(mm, source, dest, flags);
1146 if (err > 0)
1147 busy += err;
1148 if (err < 0)
1149 break;
1150 }
1151 up_read(&mm->mmap_sem);
1152 if (err < 0)
1153 return err;
1154 return busy;
1155
1156}
1157
1158/*
1159 * Allocate a new page for page migration based on vma policy.
1160 * Start by assuming the page is mapped by the same vma as contains @start.
1161 * Search forward from there, if not. N.B., this assumes that the
1162 * list of pages handed to migrate_pages()--which is how we get here--
1163 * is in virtual address order.
1164 */
1165static struct page *new_page(struct page *page, unsigned long start)
1166{
1167 struct vm_area_struct *vma;
1168 unsigned long uninitialized_var(address);
1169
1170 vma = find_vma(current->mm, start);
1171 while (vma) {
1172 address = page_address_in_vma(page, vma);
1173 if (address != -EFAULT)
1174 break;
1175 vma = vma->vm_next;
1176 }
1177
1178 if (PageHuge(page)) {
1179 return alloc_huge_page_vma(page_hstate(compound_head(page)),
1180 vma, address);
1181 } else if (PageTransHuge(page)) {
1182 struct page *thp;
1183
1184 thp = alloc_hugepage_vma(GFP_TRANSHUGE, vma, address,
1185 HPAGE_PMD_ORDER);
1186 if (!thp)
1187 return NULL;
1188 prep_transhuge_page(thp);
1189 return thp;
1190 }
1191 /*
1192 * if !vma, alloc_page_vma() will use task or system default policy
1193 */
1194 return alloc_page_vma(GFP_HIGHUSER_MOVABLE | __GFP_RETRY_MAYFAIL,
1195 vma, address);
1196}
1197#else
1198
1199static int migrate_page_add(struct page *page, struct list_head *pagelist,
1200 unsigned long flags)
1201{
1202 return -EIO;
1203}
1204
1205int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
1206 const nodemask_t *to, int flags)
1207{
1208 return -ENOSYS;
1209}
1210
1211static struct page *new_page(struct page *page, unsigned long start)
1212{
1213 return NULL;
1214}
1215#endif
1216
1217static long do_mbind(unsigned long start, unsigned long len,
1218 unsigned short mode, unsigned short mode_flags,
1219 nodemask_t *nmask, unsigned long flags)
1220{
1221 struct mm_struct *mm = current->mm;
1222 struct mempolicy *new;
1223 unsigned long end;
1224 int err;
1225 int ret;
1226 LIST_HEAD(pagelist);
1227
1228 if (flags & ~(unsigned long)MPOL_MF_VALID)
1229 return -EINVAL;
1230 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1231 return -EPERM;
1232
1233 if (start & ~PAGE_MASK)
1234 return -EINVAL;
1235
1236 if (mode == MPOL_DEFAULT)
1237 flags &= ~MPOL_MF_STRICT;
1238
1239 len = (len + PAGE_SIZE - 1) & PAGE_MASK;
1240 end = start + len;
1241
1242 if (end < start)
1243 return -EINVAL;
1244 if (end == start)
1245 return 0;
1246
1247 new = mpol_new(mode, mode_flags, nmask);
1248 if (IS_ERR(new))
1249 return PTR_ERR(new);
1250
1251 if (flags & MPOL_MF_LAZY)
1252 new->flags |= MPOL_F_MOF;
1253
1254 /*
1255 * If we are using the default policy then operation
1256 * on discontinuous address spaces is okay after all
1257 */
1258 if (!new)
1259 flags |= MPOL_MF_DISCONTIG_OK;
1260
1261 pr_debug("mbind %lx-%lx mode:%d flags:%d nodes:%lx\n",
1262 start, start + len, mode, mode_flags,
1263 nmask ? nodes_addr(*nmask)[0] : NUMA_NO_NODE);
1264
1265 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) {
1266
1267 err = migrate_prep();
1268 if (err)
1269 goto mpol_out;
1270 }
1271 {
1272 NODEMASK_SCRATCH(scratch);
1273 if (scratch) {
1274 down_write(&mm->mmap_sem);
1275 task_lock(current);
1276 err = mpol_set_nodemask(new, nmask, scratch);
1277 task_unlock(current);
1278 if (err)
1279 up_write(&mm->mmap_sem);
1280 } else
1281 err = -ENOMEM;
1282 NODEMASK_SCRATCH_FREE(scratch);
1283 }
1284 if (err)
1285 goto mpol_out;
1286
1287 ret = queue_pages_range(mm, start, end, nmask,
1288 flags | MPOL_MF_INVERT, &pagelist);
1289
1290 if (ret < 0) {
1291 err = ret;
1292 goto up_out;
1293 }
1294
1295 err = mbind_range(mm, start, end, new);
1296
1297 if (!err) {
1298 int nr_failed = 0;
1299
1300 if (!list_empty(&pagelist)) {
1301 WARN_ON_ONCE(flags & MPOL_MF_LAZY);
1302 nr_failed = migrate_pages(&pagelist, new_page, NULL,
1303 start, MIGRATE_SYNC, MR_MEMPOLICY_MBIND);
1304 if (nr_failed)
1305 putback_movable_pages(&pagelist);
1306 }
1307
1308 if ((ret > 0) || (nr_failed && (flags & MPOL_MF_STRICT)))
1309 err = -EIO;
1310 } else {
1311up_out:
1312 if (!list_empty(&pagelist))
1313 putback_movable_pages(&pagelist);
1314 }
1315
1316 up_write(&mm->mmap_sem);
1317mpol_out:
1318 mpol_put(new);
1319 return err;
1320}
1321
1322/*
1323 * User space interface with variable sized bitmaps for nodelists.
1324 */
1325
1326/* Copy a node mask from user space. */
1327static int get_nodes(nodemask_t *nodes, const unsigned long __user *nmask,
1328 unsigned long maxnode)
1329{
1330 unsigned long k;
1331 unsigned long t;
1332 unsigned long nlongs;
1333 unsigned long endmask;
1334
1335 --maxnode;
1336 nodes_clear(*nodes);
1337 if (maxnode == 0 || !nmask)
1338 return 0;
1339 if (maxnode > PAGE_SIZE*BITS_PER_BYTE)
1340 return -EINVAL;
1341
1342 nlongs = BITS_TO_LONGS(maxnode);
1343 if ((maxnode % BITS_PER_LONG) == 0)
1344 endmask = ~0UL;
1345 else
1346 endmask = (1UL << (maxnode % BITS_PER_LONG)) - 1;
1347
1348 /*
1349 * When the user specified more nodes than supported just check
1350 * if the non supported part is all zero.
1351 *
1352 * If maxnode have more longs than MAX_NUMNODES, check
1353 * the bits in that area first. And then go through to
1354 * check the rest bits which equal or bigger than MAX_NUMNODES.
1355 * Otherwise, just check bits [MAX_NUMNODES, maxnode).
1356 */
1357 if (nlongs > BITS_TO_LONGS(MAX_NUMNODES)) {
1358 for (k = BITS_TO_LONGS(MAX_NUMNODES); k < nlongs; k++) {
1359 if (get_user(t, nmask + k))
1360 return -EFAULT;
1361 if (k == nlongs - 1) {
1362 if (t & endmask)
1363 return -EINVAL;
1364 } else if (t)
1365 return -EINVAL;
1366 }
1367 nlongs = BITS_TO_LONGS(MAX_NUMNODES);
1368 endmask = ~0UL;
1369 }
1370
1371 if (maxnode > MAX_NUMNODES && MAX_NUMNODES % BITS_PER_LONG != 0) {
1372 unsigned long valid_mask = endmask;
1373
1374 valid_mask &= ~((1UL << (MAX_NUMNODES % BITS_PER_LONG)) - 1);
1375 if (get_user(t, nmask + nlongs - 1))
1376 return -EFAULT;
1377 if (t & valid_mask)
1378 return -EINVAL;
1379 }
1380
1381 if (copy_from_user(nodes_addr(*nodes), nmask, nlongs*sizeof(unsigned long)))
1382 return -EFAULT;
1383 nodes_addr(*nodes)[nlongs-1] &= endmask;
1384 return 0;
1385}
1386
1387/* Copy a kernel node mask to user space */
1388static int copy_nodes_to_user(unsigned long __user *mask, unsigned long maxnode,
1389 nodemask_t *nodes)
1390{
1391 unsigned long copy = ALIGN(maxnode-1, 64) / 8;
1392 unsigned int nbytes = BITS_TO_LONGS(nr_node_ids) * sizeof(long);
1393
1394 if (copy > nbytes) {
1395 if (copy > PAGE_SIZE)
1396 return -EINVAL;
1397 if (clear_user((char __user *)mask + nbytes, copy - nbytes))
1398 return -EFAULT;
1399 copy = nbytes;
1400 }
1401 return copy_to_user(mask, nodes_addr(*nodes), copy) ? -EFAULT : 0;
1402}
1403
1404static long kernel_mbind(unsigned long start, unsigned long len,
1405 unsigned long mode, const unsigned long __user *nmask,
1406 unsigned long maxnode, unsigned int flags)
1407{
1408 nodemask_t nodes;
1409 int err;
1410 unsigned short mode_flags;
1411
1412 start = untagged_addr(start);
1413 mode_flags = mode & MPOL_MODE_FLAGS;
1414 mode &= ~MPOL_MODE_FLAGS;
1415 if (mode >= MPOL_MAX)
1416 return -EINVAL;
1417 if ((mode_flags & MPOL_F_STATIC_NODES) &&
1418 (mode_flags & MPOL_F_RELATIVE_NODES))
1419 return -EINVAL;
1420 err = get_nodes(&nodes, nmask, maxnode);
1421 if (err)
1422 return err;
1423 return do_mbind(start, len, mode, mode_flags, &nodes, flags);
1424}
1425
1426SYSCALL_DEFINE6(mbind, unsigned long, start, unsigned long, len,
1427 unsigned long, mode, const unsigned long __user *, nmask,
1428 unsigned long, maxnode, unsigned int, flags)
1429{
1430 return kernel_mbind(start, len, mode, nmask, maxnode, flags);
1431}
1432
1433/* Set the process memory policy */
1434static long kernel_set_mempolicy(int mode, const unsigned long __user *nmask,
1435 unsigned long maxnode)
1436{
1437 int err;
1438 nodemask_t nodes;
1439 unsigned short flags;
1440
1441 flags = mode & MPOL_MODE_FLAGS;
1442 mode &= ~MPOL_MODE_FLAGS;
1443 if ((unsigned int)mode >= MPOL_MAX)
1444 return -EINVAL;
1445 if ((flags & MPOL_F_STATIC_NODES) && (flags & MPOL_F_RELATIVE_NODES))
1446 return -EINVAL;
1447 err = get_nodes(&nodes, nmask, maxnode);
1448 if (err)
1449 return err;
1450 return do_set_mempolicy(mode, flags, &nodes);
1451}
1452
1453SYSCALL_DEFINE3(set_mempolicy, int, mode, const unsigned long __user *, nmask,
1454 unsigned long, maxnode)
1455{
1456 return kernel_set_mempolicy(mode, nmask, maxnode);
1457}
1458
1459static int kernel_migrate_pages(pid_t pid, unsigned long maxnode,
1460 const unsigned long __user *old_nodes,
1461 const unsigned long __user *new_nodes)
1462{
1463 struct mm_struct *mm = NULL;
1464 struct task_struct *task;
1465 nodemask_t task_nodes;
1466 int err;
1467 nodemask_t *old;
1468 nodemask_t *new;
1469 NODEMASK_SCRATCH(scratch);
1470
1471 if (!scratch)
1472 return -ENOMEM;
1473
1474 old = &scratch->mask1;
1475 new = &scratch->mask2;
1476
1477 err = get_nodes(old, old_nodes, maxnode);
1478 if (err)
1479 goto out;
1480
1481 err = get_nodes(new, new_nodes, maxnode);
1482 if (err)
1483 goto out;
1484
1485 /* Find the mm_struct */
1486 rcu_read_lock();
1487 task = pid ? find_task_by_vpid(pid) : current;
1488 if (!task) {
1489 rcu_read_unlock();
1490 err = -ESRCH;
1491 goto out;
1492 }
1493 get_task_struct(task);
1494
1495 err = -EINVAL;
1496
1497 /*
1498 * Check if this process has the right to modify the specified process.
1499 * Use the regular "ptrace_may_access()" checks.
1500 */
1501 if (!ptrace_may_access(task, PTRACE_MODE_READ_REALCREDS)) {
1502 rcu_read_unlock();
1503 err = -EPERM;
1504 goto out_put;
1505 }
1506 rcu_read_unlock();
1507
1508 task_nodes = cpuset_mems_allowed(task);
1509 /* Is the user allowed to access the target nodes? */
1510 if (!nodes_subset(*new, task_nodes) && !capable(CAP_SYS_NICE)) {
1511 err = -EPERM;
1512 goto out_put;
1513 }
1514
1515 task_nodes = cpuset_mems_allowed(current);
1516 nodes_and(*new, *new, task_nodes);
1517 if (nodes_empty(*new))
1518 goto out_put;
1519
1520 err = security_task_movememory(task);
1521 if (err)
1522 goto out_put;
1523
1524 mm = get_task_mm(task);
1525 put_task_struct(task);
1526
1527 if (!mm) {
1528 err = -EINVAL;
1529 goto out;
1530 }
1531
1532 err = do_migrate_pages(mm, old, new,
1533 capable(CAP_SYS_NICE) ? MPOL_MF_MOVE_ALL : MPOL_MF_MOVE);
1534
1535 mmput(mm);
1536out:
1537 NODEMASK_SCRATCH_FREE(scratch);
1538
1539 return err;
1540
1541out_put:
1542 put_task_struct(task);
1543 goto out;
1544
1545}
1546
1547SYSCALL_DEFINE4(migrate_pages, pid_t, pid, unsigned long, maxnode,
1548 const unsigned long __user *, old_nodes,
1549 const unsigned long __user *, new_nodes)
1550{
1551 return kernel_migrate_pages(pid, maxnode, old_nodes, new_nodes);
1552}
1553
1554
1555/* Retrieve NUMA policy */
1556static int kernel_get_mempolicy(int __user *policy,
1557 unsigned long __user *nmask,
1558 unsigned long maxnode,
1559 unsigned long addr,
1560 unsigned long flags)
1561{
1562 int err;
1563 int uninitialized_var(pval);
1564 nodemask_t nodes;
1565
1566 addr = untagged_addr(addr);
1567
1568 if (nmask != NULL && maxnode < nr_node_ids)
1569 return -EINVAL;
1570
1571 err = do_get_mempolicy(&pval, &nodes, addr, flags);
1572
1573 if (err)
1574 return err;
1575
1576 if (policy && put_user(pval, policy))
1577 return -EFAULT;
1578
1579 if (nmask)
1580 err = copy_nodes_to_user(nmask, maxnode, &nodes);
1581
1582 return err;
1583}
1584
1585SYSCALL_DEFINE5(get_mempolicy, int __user *, policy,
1586 unsigned long __user *, nmask, unsigned long, maxnode,
1587 unsigned long, addr, unsigned long, flags)
1588{
1589 return kernel_get_mempolicy(policy, nmask, maxnode, addr, flags);
1590}
1591
1592#ifdef CONFIG_COMPAT
1593
1594COMPAT_SYSCALL_DEFINE5(get_mempolicy, int __user *, policy,
1595 compat_ulong_t __user *, nmask,
1596 compat_ulong_t, maxnode,
1597 compat_ulong_t, addr, compat_ulong_t, flags)
1598{
1599 long err;
1600 unsigned long __user *nm = NULL;
1601 unsigned long nr_bits, alloc_size;
1602 DECLARE_BITMAP(bm, MAX_NUMNODES);
1603
1604 nr_bits = min_t(unsigned long, maxnode-1, nr_node_ids);
1605 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1606
1607 if (nmask)
1608 nm = compat_alloc_user_space(alloc_size);
1609
1610 err = kernel_get_mempolicy(policy, nm, nr_bits+1, addr, flags);
1611
1612 if (!err && nmask) {
1613 unsigned long copy_size;
1614 copy_size = min_t(unsigned long, sizeof(bm), alloc_size);
1615 err = copy_from_user(bm, nm, copy_size);
1616 /* ensure entire bitmap is zeroed */
1617 err |= clear_user(nmask, ALIGN(maxnode-1, 8) / 8);
1618 err |= compat_put_bitmap(nmask, bm, nr_bits);
1619 }
1620
1621 return err;
1622}
1623
1624COMPAT_SYSCALL_DEFINE3(set_mempolicy, int, mode, compat_ulong_t __user *, nmask,
1625 compat_ulong_t, maxnode)
1626{
1627 unsigned long __user *nm = NULL;
1628 unsigned long nr_bits, alloc_size;
1629 DECLARE_BITMAP(bm, MAX_NUMNODES);
1630
1631 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1632 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1633
1634 if (nmask) {
1635 if (compat_get_bitmap(bm, nmask, nr_bits))
1636 return -EFAULT;
1637 nm = compat_alloc_user_space(alloc_size);
1638 if (copy_to_user(nm, bm, alloc_size))
1639 return -EFAULT;
1640 }
1641
1642 return kernel_set_mempolicy(mode, nm, nr_bits+1);
1643}
1644
1645COMPAT_SYSCALL_DEFINE6(mbind, compat_ulong_t, start, compat_ulong_t, len,
1646 compat_ulong_t, mode, compat_ulong_t __user *, nmask,
1647 compat_ulong_t, maxnode, compat_ulong_t, flags)
1648{
1649 unsigned long __user *nm = NULL;
1650 unsigned long nr_bits, alloc_size;
1651 nodemask_t bm;
1652
1653 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1654 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1655
1656 if (nmask) {
1657 if (compat_get_bitmap(nodes_addr(bm), nmask, nr_bits))
1658 return -EFAULT;
1659 nm = compat_alloc_user_space(alloc_size);
1660 if (copy_to_user(nm, nodes_addr(bm), alloc_size))
1661 return -EFAULT;
1662 }
1663
1664 return kernel_mbind(start, len, mode, nm, nr_bits+1, flags);
1665}
1666
1667COMPAT_SYSCALL_DEFINE4(migrate_pages, compat_pid_t, pid,
1668 compat_ulong_t, maxnode,
1669 const compat_ulong_t __user *, old_nodes,
1670 const compat_ulong_t __user *, new_nodes)
1671{
1672 unsigned long __user *old = NULL;
1673 unsigned long __user *new = NULL;
1674 nodemask_t tmp_mask;
1675 unsigned long nr_bits;
1676 unsigned long size;
1677
1678 nr_bits = min_t(unsigned long, maxnode - 1, MAX_NUMNODES);
1679 size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1680 if (old_nodes) {
1681 if (compat_get_bitmap(nodes_addr(tmp_mask), old_nodes, nr_bits))
1682 return -EFAULT;
1683 old = compat_alloc_user_space(new_nodes ? size * 2 : size);
1684 if (new_nodes)
1685 new = old + size / sizeof(unsigned long);
1686 if (copy_to_user(old, nodes_addr(tmp_mask), size))
1687 return -EFAULT;
1688 }
1689 if (new_nodes) {
1690 if (compat_get_bitmap(nodes_addr(tmp_mask), new_nodes, nr_bits))
1691 return -EFAULT;
1692 if (new == NULL)
1693 new = compat_alloc_user_space(size);
1694 if (copy_to_user(new, nodes_addr(tmp_mask), size))
1695 return -EFAULT;
1696 }
1697 return kernel_migrate_pages(pid, nr_bits + 1, old, new);
1698}
1699
1700#endif /* CONFIG_COMPAT */
1701
1702struct mempolicy *__get_vma_policy(struct vm_area_struct *vma,
1703 unsigned long addr)
1704{
1705 struct mempolicy *pol = NULL;
1706
1707 if (vma) {
1708 if (vma->vm_ops && vma->vm_ops->get_policy) {
1709 pol = vma->vm_ops->get_policy(vma, addr);
1710 } else if (vma->vm_policy) {
1711 pol = vma->vm_policy;
1712
1713 /*
1714 * shmem_alloc_page() passes MPOL_F_SHARED policy with
1715 * a pseudo vma whose vma->vm_ops=NULL. Take a reference
1716 * count on these policies which will be dropped by
1717 * mpol_cond_put() later
1718 */
1719 if (mpol_needs_cond_ref(pol))
1720 mpol_get(pol);
1721 }
1722 }
1723
1724 return pol;
1725}
1726
1727/*
1728 * get_vma_policy(@vma, @addr)
1729 * @vma: virtual memory area whose policy is sought
1730 * @addr: address in @vma for shared policy lookup
1731 *
1732 * Returns effective policy for a VMA at specified address.
1733 * Falls back to current->mempolicy or system default policy, as necessary.
1734 * Shared policies [those marked as MPOL_F_SHARED] require an extra reference
1735 * count--added by the get_policy() vm_op, as appropriate--to protect against
1736 * freeing by another task. It is the caller's responsibility to free the
1737 * extra reference for shared policies.
1738 */
1739static struct mempolicy *get_vma_policy(struct vm_area_struct *vma,
1740 unsigned long addr)
1741{
1742 struct mempolicy *pol = __get_vma_policy(vma, addr);
1743
1744 if (!pol)
1745 pol = get_task_policy(current);
1746
1747 return pol;
1748}
1749
1750bool vma_policy_mof(struct vm_area_struct *vma)
1751{
1752 struct mempolicy *pol;
1753
1754 if (vma->vm_ops && vma->vm_ops->get_policy) {
1755 bool ret = false;
1756
1757 pol = vma->vm_ops->get_policy(vma, vma->vm_start);
1758 if (pol && (pol->flags & MPOL_F_MOF))
1759 ret = true;
1760 mpol_cond_put(pol);
1761
1762 return ret;
1763 }
1764
1765 pol = vma->vm_policy;
1766 if (!pol)
1767 pol = get_task_policy(current);
1768
1769 return pol->flags & MPOL_F_MOF;
1770}
1771
1772static int apply_policy_zone(struct mempolicy *policy, enum zone_type zone)
1773{
1774 enum zone_type dynamic_policy_zone = policy_zone;
1775
1776 BUG_ON(dynamic_policy_zone == ZONE_MOVABLE);
1777
1778 /*
1779 * if policy->v.nodes has movable memory only,
1780 * we apply policy when gfp_zone(gfp) = ZONE_MOVABLE only.
1781 *
1782 * policy->v.nodes is intersect with node_states[N_MEMORY].
1783 * so if the following test faile, it implies
1784 * policy->v.nodes has movable memory only.
1785 */
1786 if (!nodes_intersects(policy->v.nodes, node_states[N_HIGH_MEMORY]))
1787 dynamic_policy_zone = ZONE_MOVABLE;
1788
1789 return zone >= dynamic_policy_zone;
1790}
1791
1792/*
1793 * Return a nodemask representing a mempolicy for filtering nodes for
1794 * page allocation
1795 */
1796static nodemask_t *policy_nodemask(gfp_t gfp, struct mempolicy *policy)
1797{
1798 /* Lower zones don't get a nodemask applied for MPOL_BIND */
1799 if (unlikely(policy->mode == MPOL_BIND) &&
1800 apply_policy_zone(policy, gfp_zone(gfp)) &&
1801 cpuset_nodemask_valid_mems_allowed(&policy->v.nodes))
1802 return &policy->v.nodes;
1803
1804 return NULL;
1805}
1806
1807/* Return the node id preferred by the given mempolicy, or the given id */
1808static int policy_node(gfp_t gfp, struct mempolicy *policy,
1809 int nd)
1810{
1811 if (policy->mode == MPOL_PREFERRED && !(policy->flags & MPOL_F_LOCAL))
1812 nd = policy->v.preferred_node;
1813 else {
1814 /*
1815 * __GFP_THISNODE shouldn't even be used with the bind policy
1816 * because we might easily break the expectation to stay on the
1817 * requested node and not break the policy.
1818 */
1819 WARN_ON_ONCE(policy->mode == MPOL_BIND && (gfp & __GFP_THISNODE));
1820 }
1821
1822 return nd;
1823}
1824
1825/* Do dynamic interleaving for a process */
1826static unsigned interleave_nodes(struct mempolicy *policy)
1827{
1828 unsigned next;
1829 struct task_struct *me = current;
1830
1831 next = next_node_in(me->il_prev, policy->v.nodes);
1832 if (next < MAX_NUMNODES)
1833 me->il_prev = next;
1834 return next;
1835}
1836
1837/*
1838 * Depending on the memory policy provide a node from which to allocate the
1839 * next slab entry.
1840 */
1841unsigned int mempolicy_slab_node(void)
1842{
1843 struct mempolicy *policy;
1844 int node = numa_mem_id();
1845
1846 if (in_interrupt())
1847 return node;
1848
1849 policy = current->mempolicy;
1850 if (!policy || policy->flags & MPOL_F_LOCAL)
1851 return node;
1852
1853 switch (policy->mode) {
1854 case MPOL_PREFERRED:
1855 /*
1856 * handled MPOL_F_LOCAL above
1857 */
1858 return policy->v.preferred_node;
1859
1860 case MPOL_INTERLEAVE:
1861 return interleave_nodes(policy);
1862
1863 case MPOL_BIND: {
1864 struct zoneref *z;
1865
1866 /*
1867 * Follow bind policy behavior and start allocation at the
1868 * first node.
1869 */
1870 struct zonelist *zonelist;
1871 enum zone_type highest_zoneidx = gfp_zone(GFP_KERNEL);
1872 zonelist = &NODE_DATA(node)->node_zonelists[ZONELIST_FALLBACK];
1873 z = first_zones_zonelist(zonelist, highest_zoneidx,
1874 &policy->v.nodes);
1875 return z->zone ? zone_to_nid(z->zone) : node;
1876 }
1877
1878 default:
1879 BUG();
1880 }
1881}
1882
1883/*
1884 * Do static interleaving for a VMA with known offset @n. Returns the n'th
1885 * node in pol->v.nodes (starting from n=0), wrapping around if n exceeds the
1886 * number of present nodes.
1887 */
1888static unsigned offset_il_node(struct mempolicy *pol, unsigned long n)
1889{
1890 unsigned nnodes = nodes_weight(pol->v.nodes);
1891 unsigned target;
1892 int i;
1893 int nid;
1894
1895 if (!nnodes)
1896 return numa_node_id();
1897 target = (unsigned int)n % nnodes;
1898 nid = first_node(pol->v.nodes);
1899 for (i = 0; i < target; i++)
1900 nid = next_node(nid, pol->v.nodes);
1901 return nid;
1902}
1903
1904/* Determine a node number for interleave */
1905static inline unsigned interleave_nid(struct mempolicy *pol,
1906 struct vm_area_struct *vma, unsigned long addr, int shift)
1907{
1908 if (vma) {
1909 unsigned long off;
1910
1911 /*
1912 * for small pages, there is no difference between
1913 * shift and PAGE_SHIFT, so the bit-shift is safe.
1914 * for huge pages, since vm_pgoff is in units of small
1915 * pages, we need to shift off the always 0 bits to get
1916 * a useful offset.
1917 */
1918 BUG_ON(shift < PAGE_SHIFT);
1919 off = vma->vm_pgoff >> (shift - PAGE_SHIFT);
1920 off += (addr - vma->vm_start) >> shift;
1921 return offset_il_node(pol, off);
1922 } else
1923 return interleave_nodes(pol);
1924}
1925
1926#ifdef CONFIG_HUGETLBFS
1927/*
1928 * huge_node(@vma, @addr, @gfp_flags, @mpol)
1929 * @vma: virtual memory area whose policy is sought
1930 * @addr: address in @vma for shared policy lookup and interleave policy
1931 * @gfp_flags: for requested zone
1932 * @mpol: pointer to mempolicy pointer for reference counted mempolicy
1933 * @nodemask: pointer to nodemask pointer for MPOL_BIND nodemask
1934 *
1935 * Returns a nid suitable for a huge page allocation and a pointer
1936 * to the struct mempolicy for conditional unref after allocation.
1937 * If the effective policy is 'BIND, returns a pointer to the mempolicy's
1938 * @nodemask for filtering the zonelist.
1939 *
1940 * Must be protected by read_mems_allowed_begin()
1941 */
1942int huge_node(struct vm_area_struct *vma, unsigned long addr, gfp_t gfp_flags,
1943 struct mempolicy **mpol, nodemask_t **nodemask)
1944{
1945 int nid;
1946
1947 *mpol = get_vma_policy(vma, addr);
1948 *nodemask = NULL; /* assume !MPOL_BIND */
1949
1950 if (unlikely((*mpol)->mode == MPOL_INTERLEAVE)) {
1951 nid = interleave_nid(*mpol, vma, addr,
1952 huge_page_shift(hstate_vma(vma)));
1953 } else {
1954 nid = policy_node(gfp_flags, *mpol, numa_node_id());
1955 if ((*mpol)->mode == MPOL_BIND)
1956 *nodemask = &(*mpol)->v.nodes;
1957 }
1958 return nid;
1959}
1960
1961/*
1962 * init_nodemask_of_mempolicy
1963 *
1964 * If the current task's mempolicy is "default" [NULL], return 'false'
1965 * to indicate default policy. Otherwise, extract the policy nodemask
1966 * for 'bind' or 'interleave' policy into the argument nodemask, or
1967 * initialize the argument nodemask to contain the single node for
1968 * 'preferred' or 'local' policy and return 'true' to indicate presence
1969 * of non-default mempolicy.
1970 *
1971 * We don't bother with reference counting the mempolicy [mpol_get/put]
1972 * because the current task is examining it's own mempolicy and a task's
1973 * mempolicy is only ever changed by the task itself.
1974 *
1975 * N.B., it is the caller's responsibility to free a returned nodemask.
1976 */
1977bool init_nodemask_of_mempolicy(nodemask_t *mask)
1978{
1979 struct mempolicy *mempolicy;
1980 int nid;
1981
1982 if (!(mask && current->mempolicy))
1983 return false;
1984
1985 task_lock(current);
1986 mempolicy = current->mempolicy;
1987 switch (mempolicy->mode) {
1988 case MPOL_PREFERRED:
1989 if (mempolicy->flags & MPOL_F_LOCAL)
1990 nid = numa_node_id();
1991 else
1992 nid = mempolicy->v.preferred_node;
1993 init_nodemask_of_node(mask, nid);
1994 break;
1995
1996 case MPOL_BIND:
1997 /* Fall through */
1998 case MPOL_INTERLEAVE:
1999 *mask = mempolicy->v.nodes;
2000 break;
2001
2002 default:
2003 BUG();
2004 }
2005 task_unlock(current);
2006
2007 return true;
2008}
2009#endif
2010
2011/*
2012 * mempolicy_nodemask_intersects
2013 *
2014 * If tsk's mempolicy is "default" [NULL], return 'true' to indicate default
2015 * policy. Otherwise, check for intersection between mask and the policy
2016 * nodemask for 'bind' or 'interleave' policy. For 'perferred' or 'local'
2017 * policy, always return true since it may allocate elsewhere on fallback.
2018 *
2019 * Takes task_lock(tsk) to prevent freeing of its mempolicy.
2020 */
2021bool mempolicy_nodemask_intersects(struct task_struct *tsk,
2022 const nodemask_t *mask)
2023{
2024 struct mempolicy *mempolicy;
2025 bool ret = true;
2026
2027 if (!mask)
2028 return ret;
2029 task_lock(tsk);
2030 mempolicy = tsk->mempolicy;
2031 if (!mempolicy)
2032 goto out;
2033
2034 switch (mempolicy->mode) {
2035 case MPOL_PREFERRED:
2036 /*
2037 * MPOL_PREFERRED and MPOL_F_LOCAL are only preferred nodes to
2038 * allocate from, they may fallback to other nodes when oom.
2039 * Thus, it's possible for tsk to have allocated memory from
2040 * nodes in mask.
2041 */
2042 break;
2043 case MPOL_BIND:
2044 case MPOL_INTERLEAVE:
2045 ret = nodes_intersects(mempolicy->v.nodes, *mask);
2046 break;
2047 default:
2048 BUG();
2049 }
2050out:
2051 task_unlock(tsk);
2052 return ret;
2053}
2054
2055/* Allocate a page in interleaved policy.
2056 Own path because it needs to do special accounting. */
2057static struct page *alloc_page_interleave(gfp_t gfp, unsigned order,
2058 unsigned nid)
2059{
2060 struct page *page;
2061
2062 page = __alloc_pages(gfp, order, nid);
2063 /* skip NUMA_INTERLEAVE_HIT counter update if numa stats is disabled */
2064 if (!static_branch_likely(&vm_numa_stat_key))
2065 return page;
2066 if (page && page_to_nid(page) == nid) {
2067 preempt_disable();
2068 __inc_numa_state(page_zone(page), NUMA_INTERLEAVE_HIT);
2069 preempt_enable();
2070 }
2071 return page;
2072}
2073
2074/**
2075 * alloc_pages_vma - Allocate a page for a VMA.
2076 *
2077 * @gfp:
2078 * %GFP_USER user allocation.
2079 * %GFP_KERNEL kernel allocations,
2080 * %GFP_HIGHMEM highmem/user allocations,
2081 * %GFP_FS allocation should not call back into a file system.
2082 * %GFP_ATOMIC don't sleep.
2083 *
2084 * @order:Order of the GFP allocation.
2085 * @vma: Pointer to VMA or NULL if not available.
2086 * @addr: Virtual Address of the allocation. Must be inside the VMA.
2087 * @node: Which node to prefer for allocation (modulo policy).
2088 * @hugepage: for hugepages try only the preferred node if possible
2089 *
2090 * This function allocates a page from the kernel page pool and applies
2091 * a NUMA policy associated with the VMA or the current process.
2092 * When VMA is not NULL caller must hold down_read on the mmap_sem of the
2093 * mm_struct of the VMA to prevent it from going away. Should be used for
2094 * all allocations for pages that will be mapped into user space. Returns
2095 * NULL when no page can be allocated.
2096 */
2097struct page *
2098alloc_pages_vma(gfp_t gfp, int order, struct vm_area_struct *vma,
2099 unsigned long addr, int node, bool hugepage)
2100{
2101 struct mempolicy *pol;
2102 struct page *page;
2103 int preferred_nid;
2104 nodemask_t *nmask;
2105
2106 pol = get_vma_policy(vma, addr);
2107
2108 if (pol->mode == MPOL_INTERLEAVE) {
2109 unsigned nid;
2110
2111 nid = interleave_nid(pol, vma, addr, PAGE_SHIFT + order);
2112 mpol_cond_put(pol);
2113 page = alloc_page_interleave(gfp, order, nid);
2114 goto out;
2115 }
2116
2117 if (unlikely(IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE) && hugepage)) {
2118 int hpage_node = node;
2119
2120 /*
2121 * For hugepage allocation and non-interleave policy which
2122 * allows the current node (or other explicitly preferred
2123 * node) we only try to allocate from the current/preferred
2124 * node and don't fall back to other nodes, as the cost of
2125 * remote accesses would likely offset THP benefits.
2126 *
2127 * If the policy is interleave, or does not allow the current
2128 * node in its nodemask, we allocate the standard way.
2129 */
2130 if (pol->mode == MPOL_PREFERRED && !(pol->flags & MPOL_F_LOCAL))
2131 hpage_node = pol->v.preferred_node;
2132
2133 nmask = policy_nodemask(gfp, pol);
2134 if (!nmask || node_isset(hpage_node, *nmask)) {
2135 mpol_cond_put(pol);
2136 page = __alloc_pages_node(hpage_node,
2137 gfp | __GFP_THISNODE, order);
2138
2139 /*
2140 * If hugepage allocations are configured to always
2141 * synchronous compact or the vma has been madvised
2142 * to prefer hugepage backing, retry allowing remote
2143 * memory as well.
2144 */
2145 if (!page && (gfp & __GFP_DIRECT_RECLAIM))
2146 page = __alloc_pages_node(hpage_node,
2147 gfp | __GFP_NORETRY, order);
2148
2149 goto out;
2150 }
2151 }
2152
2153 nmask = policy_nodemask(gfp, pol);
2154 preferred_nid = policy_node(gfp, pol, node);
2155 page = __alloc_pages_nodemask(gfp, order, preferred_nid, nmask);
2156 mpol_cond_put(pol);
2157out:
2158 return page;
2159}
2160EXPORT_SYMBOL(alloc_pages_vma);
2161
2162/**
2163 * alloc_pages_current - Allocate pages.
2164 *
2165 * @gfp:
2166 * %GFP_USER user allocation,
2167 * %GFP_KERNEL kernel allocation,
2168 * %GFP_HIGHMEM highmem allocation,
2169 * %GFP_FS don't call back into a file system.
2170 * %GFP_ATOMIC don't sleep.
2171 * @order: Power of two of allocation size in pages. 0 is a single page.
2172 *
2173 * Allocate a page from the kernel page pool. When not in
2174 * interrupt context and apply the current process NUMA policy.
2175 * Returns NULL when no page can be allocated.
2176 */
2177struct page *alloc_pages_current(gfp_t gfp, unsigned order)
2178{
2179 struct mempolicy *pol = &default_policy;
2180 struct page *page;
2181
2182 if (!in_interrupt() && !(gfp & __GFP_THISNODE))
2183 pol = get_task_policy(current);
2184
2185 /*
2186 * No reference counting needed for current->mempolicy
2187 * nor system default_policy
2188 */
2189 if (pol->mode == MPOL_INTERLEAVE)
2190 page = alloc_page_interleave(gfp, order, interleave_nodes(pol));
2191 else
2192 page = __alloc_pages_nodemask(gfp, order,
2193 policy_node(gfp, pol, numa_node_id()),
2194 policy_nodemask(gfp, pol));
2195
2196 return page;
2197}
2198EXPORT_SYMBOL(alloc_pages_current);
2199
2200int vma_dup_policy(struct vm_area_struct *src, struct vm_area_struct *dst)
2201{
2202 struct mempolicy *pol = mpol_dup(vma_policy(src));
2203
2204 if (IS_ERR(pol))
2205 return PTR_ERR(pol);
2206 dst->vm_policy = pol;
2207 return 0;
2208}
2209
2210/*
2211 * If mpol_dup() sees current->cpuset == cpuset_being_rebound, then it
2212 * rebinds the mempolicy its copying by calling mpol_rebind_policy()
2213 * with the mems_allowed returned by cpuset_mems_allowed(). This
2214 * keeps mempolicies cpuset relative after its cpuset moves. See
2215 * further kernel/cpuset.c update_nodemask().
2216 *
2217 * current's mempolicy may be rebinded by the other task(the task that changes
2218 * cpuset's mems), so we needn't do rebind work for current task.
2219 */
2220
2221/* Slow path of a mempolicy duplicate */
2222struct mempolicy *__mpol_dup(struct mempolicy *old)
2223{
2224 struct mempolicy *new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
2225
2226 if (!new)
2227 return ERR_PTR(-ENOMEM);
2228
2229 /* task's mempolicy is protected by alloc_lock */
2230 if (old == current->mempolicy) {
2231 task_lock(current);
2232 *new = *old;
2233 task_unlock(current);
2234 } else
2235 *new = *old;
2236
2237 if (current_cpuset_is_being_rebound()) {
2238 nodemask_t mems = cpuset_mems_allowed(current);
2239 mpol_rebind_policy(new, &mems);
2240 }
2241 atomic_set(&new->refcnt, 1);
2242 return new;
2243}
2244
2245/* Slow path of a mempolicy comparison */
2246bool __mpol_equal(struct mempolicy *a, struct mempolicy *b)
2247{
2248 if (!a || !b)
2249 return false;
2250 if (a->mode != b->mode)
2251 return false;
2252 if (a->flags != b->flags)
2253 return false;
2254 if (mpol_store_user_nodemask(a))
2255 if (!nodes_equal(a->w.user_nodemask, b->w.user_nodemask))
2256 return false;
2257
2258 switch (a->mode) {
2259 case MPOL_BIND:
2260 /* Fall through */
2261 case MPOL_INTERLEAVE:
2262 return !!nodes_equal(a->v.nodes, b->v.nodes);
2263 case MPOL_PREFERRED:
2264 /* a's ->flags is the same as b's */
2265 if (a->flags & MPOL_F_LOCAL)
2266 return true;
2267 return a->v.preferred_node == b->v.preferred_node;
2268 default:
2269 BUG();
2270 return false;
2271 }
2272}
2273
2274/*
2275 * Shared memory backing store policy support.
2276 *
2277 * Remember policies even when nobody has shared memory mapped.
2278 * The policies are kept in Red-Black tree linked from the inode.
2279 * They are protected by the sp->lock rwlock, which should be held
2280 * for any accesses to the tree.
2281 */
2282
2283/*
2284 * lookup first element intersecting start-end. Caller holds sp->lock for
2285 * reading or for writing
2286 */
2287static struct sp_node *
2288sp_lookup(struct shared_policy *sp, unsigned long start, unsigned long end)
2289{
2290 struct rb_node *n = sp->root.rb_node;
2291
2292 while (n) {
2293 struct sp_node *p = rb_entry(n, struct sp_node, nd);
2294
2295 if (start >= p->end)
2296 n = n->rb_right;
2297 else if (end <= p->start)
2298 n = n->rb_left;
2299 else
2300 break;
2301 }
2302 if (!n)
2303 return NULL;
2304 for (;;) {
2305 struct sp_node *w = NULL;
2306 struct rb_node *prev = rb_prev(n);
2307 if (!prev)
2308 break;
2309 w = rb_entry(prev, struct sp_node, nd);
2310 if (w->end <= start)
2311 break;
2312 n = prev;
2313 }
2314 return rb_entry(n, struct sp_node, nd);
2315}
2316
2317/*
2318 * Insert a new shared policy into the list. Caller holds sp->lock for
2319 * writing.
2320 */
2321static void sp_insert(struct shared_policy *sp, struct sp_node *new)
2322{
2323 struct rb_node **p = &sp->root.rb_node;
2324 struct rb_node *parent = NULL;
2325 struct sp_node *nd;
2326
2327 while (*p) {
2328 parent = *p;
2329 nd = rb_entry(parent, struct sp_node, nd);
2330 if (new->start < nd->start)
2331 p = &(*p)->rb_left;
2332 else if (new->end > nd->end)
2333 p = &(*p)->rb_right;
2334 else
2335 BUG();
2336 }
2337 rb_link_node(&new->nd, parent, p);
2338 rb_insert_color(&new->nd, &sp->root);
2339 pr_debug("inserting %lx-%lx: %d\n", new->start, new->end,
2340 new->policy ? new->policy->mode : 0);
2341}
2342
2343/* Find shared policy intersecting idx */
2344struct mempolicy *
2345mpol_shared_policy_lookup(struct shared_policy *sp, unsigned long idx)
2346{
2347 struct mempolicy *pol = NULL;
2348 struct sp_node *sn;
2349
2350 if (!sp->root.rb_node)
2351 return NULL;
2352 read_lock(&sp->lock);
2353 sn = sp_lookup(sp, idx, idx+1);
2354 if (sn) {
2355 mpol_get(sn->policy);
2356 pol = sn->policy;
2357 }
2358 read_unlock(&sp->lock);
2359 return pol;
2360}
2361
2362static void sp_free(struct sp_node *n)
2363{
2364 mpol_put(n->policy);
2365 kmem_cache_free(sn_cache, n);
2366}
2367
2368/**
2369 * mpol_misplaced - check whether current page node is valid in policy
2370 *
2371 * @page: page to be checked
2372 * @vma: vm area where page mapped
2373 * @addr: virtual address where page mapped
2374 *
2375 * Lookup current policy node id for vma,addr and "compare to" page's
2376 * node id.
2377 *
2378 * Returns:
2379 * -1 - not misplaced, page is in the right node
2380 * node - node id where the page should be
2381 *
2382 * Policy determination "mimics" alloc_page_vma().
2383 * Called from fault path where we know the vma and faulting address.
2384 */
2385int mpol_misplaced(struct page *page, struct vm_area_struct *vma, unsigned long addr)
2386{
2387 struct mempolicy *pol;
2388 struct zoneref *z;
2389 int curnid = page_to_nid(page);
2390 unsigned long pgoff;
2391 int thiscpu = raw_smp_processor_id();
2392 int thisnid = cpu_to_node(thiscpu);
2393 int polnid = NUMA_NO_NODE;
2394 int ret = -1;
2395
2396 pol = get_vma_policy(vma, addr);
2397 if (!(pol->flags & MPOL_F_MOF))
2398 goto out;
2399
2400 switch (pol->mode) {
2401 case MPOL_INTERLEAVE:
2402 pgoff = vma->vm_pgoff;
2403 pgoff += (addr - vma->vm_start) >> PAGE_SHIFT;
2404 polnid = offset_il_node(pol, pgoff);
2405 break;
2406
2407 case MPOL_PREFERRED:
2408 if (pol->flags & MPOL_F_LOCAL)
2409 polnid = numa_node_id();
2410 else
2411 polnid = pol->v.preferred_node;
2412 break;
2413
2414 case MPOL_BIND:
2415
2416 /*
2417 * allows binding to multiple nodes.
2418 * use current page if in policy nodemask,
2419 * else select nearest allowed node, if any.
2420 * If no allowed nodes, use current [!misplaced].
2421 */
2422 if (node_isset(curnid, pol->v.nodes))
2423 goto out;
2424 z = first_zones_zonelist(
2425 node_zonelist(numa_node_id(), GFP_HIGHUSER),
2426 gfp_zone(GFP_HIGHUSER),
2427 &pol->v.nodes);
2428 polnid = zone_to_nid(z->zone);
2429 break;
2430
2431 default:
2432 BUG();
2433 }
2434
2435 /* Migrate the page towards the node whose CPU is referencing it */
2436 if (pol->flags & MPOL_F_MORON) {
2437 polnid = thisnid;
2438
2439 if (!should_numa_migrate_memory(current, page, curnid, thiscpu))
2440 goto out;
2441 }
2442
2443 if (curnid != polnid)
2444 ret = polnid;
2445out:
2446 mpol_cond_put(pol);
2447
2448 return ret;
2449}
2450
2451/*
2452 * Drop the (possibly final) reference to task->mempolicy. It needs to be
2453 * dropped after task->mempolicy is set to NULL so that any allocation done as
2454 * part of its kmem_cache_free(), such as by KASAN, doesn't reference a freed
2455 * policy.
2456 */
2457void mpol_put_task_policy(struct task_struct *task)
2458{
2459 struct mempolicy *pol;
2460
2461 task_lock(task);
2462 pol = task->mempolicy;
2463 task->mempolicy = NULL;
2464 task_unlock(task);
2465 mpol_put(pol);
2466}
2467
2468static void sp_delete(struct shared_policy *sp, struct sp_node *n)
2469{
2470 pr_debug("deleting %lx-l%lx\n", n->start, n->end);
2471 rb_erase(&n->nd, &sp->root);
2472 sp_free(n);
2473}
2474
2475static void sp_node_init(struct sp_node *node, unsigned long start,
2476 unsigned long end, struct mempolicy *pol)
2477{
2478 node->start = start;
2479 node->end = end;
2480 node->policy = pol;
2481}
2482
2483static struct sp_node *sp_alloc(unsigned long start, unsigned long end,
2484 struct mempolicy *pol)
2485{
2486 struct sp_node *n;
2487 struct mempolicy *newpol;
2488
2489 n = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2490 if (!n)
2491 return NULL;
2492
2493 newpol = mpol_dup(pol);
2494 if (IS_ERR(newpol)) {
2495 kmem_cache_free(sn_cache, n);
2496 return NULL;
2497 }
2498 newpol->flags |= MPOL_F_SHARED;
2499 sp_node_init(n, start, end, newpol);
2500
2501 return n;
2502}
2503
2504/* Replace a policy range. */
2505static int shared_policy_replace(struct shared_policy *sp, unsigned long start,
2506 unsigned long end, struct sp_node *new)
2507{
2508 struct sp_node *n;
2509 struct sp_node *n_new = NULL;
2510 struct mempolicy *mpol_new = NULL;
2511 int ret = 0;
2512
2513restart:
2514 write_lock(&sp->lock);
2515 n = sp_lookup(sp, start, end);
2516 /* Take care of old policies in the same range. */
2517 while (n && n->start < end) {
2518 struct rb_node *next = rb_next(&n->nd);
2519 if (n->start >= start) {
2520 if (n->end <= end)
2521 sp_delete(sp, n);
2522 else
2523 n->start = end;
2524 } else {
2525 /* Old policy spanning whole new range. */
2526 if (n->end > end) {
2527 if (!n_new)
2528 goto alloc_new;
2529
2530 *mpol_new = *n->policy;
2531 atomic_set(&mpol_new->refcnt, 1);
2532 sp_node_init(n_new, end, n->end, mpol_new);
2533 n->end = start;
2534 sp_insert(sp, n_new);
2535 n_new = NULL;
2536 mpol_new = NULL;
2537 break;
2538 } else
2539 n->end = start;
2540 }
2541 if (!next)
2542 break;
2543 n = rb_entry(next, struct sp_node, nd);
2544 }
2545 if (new)
2546 sp_insert(sp, new);
2547 write_unlock(&sp->lock);
2548 ret = 0;
2549
2550err_out:
2551 if (mpol_new)
2552 mpol_put(mpol_new);
2553 if (n_new)
2554 kmem_cache_free(sn_cache, n_new);
2555
2556 return ret;
2557
2558alloc_new:
2559 write_unlock(&sp->lock);
2560 ret = -ENOMEM;
2561 n_new = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2562 if (!n_new)
2563 goto err_out;
2564 mpol_new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
2565 if (!mpol_new)
2566 goto err_out;
2567 goto restart;
2568}
2569
2570/**
2571 * mpol_shared_policy_init - initialize shared policy for inode
2572 * @sp: pointer to inode shared policy
2573 * @mpol: struct mempolicy to install
2574 *
2575 * Install non-NULL @mpol in inode's shared policy rb-tree.
2576 * On entry, the current task has a reference on a non-NULL @mpol.
2577 * This must be released on exit.
2578 * This is called at get_inode() calls and we can use GFP_KERNEL.
2579 */
2580void mpol_shared_policy_init(struct shared_policy *sp, struct mempolicy *mpol)
2581{
2582 int ret;
2583
2584 sp->root = RB_ROOT; /* empty tree == default mempolicy */
2585 rwlock_init(&sp->lock);
2586
2587 if (mpol) {
2588 struct vm_area_struct pvma;
2589 struct mempolicy *new;
2590 NODEMASK_SCRATCH(scratch);
2591
2592 if (!scratch)
2593 goto put_mpol;
2594 /* contextualize the tmpfs mount point mempolicy */
2595 new = mpol_new(mpol->mode, mpol->flags, &mpol->w.user_nodemask);
2596 if (IS_ERR(new))
2597 goto free_scratch; /* no valid nodemask intersection */
2598
2599 task_lock(current);
2600 ret = mpol_set_nodemask(new, &mpol->w.user_nodemask, scratch);
2601 task_unlock(current);
2602 if (ret)
2603 goto put_new;
2604
2605 /* Create pseudo-vma that contains just the policy */
2606 vma_init(&pvma, NULL);
2607 pvma.vm_end = TASK_SIZE; /* policy covers entire file */
2608 mpol_set_shared_policy(sp, &pvma, new); /* adds ref */
2609
2610put_new:
2611 mpol_put(new); /* drop initial ref */
2612free_scratch:
2613 NODEMASK_SCRATCH_FREE(scratch);
2614put_mpol:
2615 mpol_put(mpol); /* drop our incoming ref on sb mpol */
2616 }
2617}
2618
2619int mpol_set_shared_policy(struct shared_policy *info,
2620 struct vm_area_struct *vma, struct mempolicy *npol)
2621{
2622 int err;
2623 struct sp_node *new = NULL;
2624 unsigned long sz = vma_pages(vma);
2625
2626 pr_debug("set_shared_policy %lx sz %lu %d %d %lx\n",
2627 vma->vm_pgoff,
2628 sz, npol ? npol->mode : -1,
2629 npol ? npol->flags : -1,
2630 npol ? nodes_addr(npol->v.nodes)[0] : NUMA_NO_NODE);
2631
2632 if (npol) {
2633 new = sp_alloc(vma->vm_pgoff, vma->vm_pgoff + sz, npol);
2634 if (!new)
2635 return -ENOMEM;
2636 }
2637 err = shared_policy_replace(info, vma->vm_pgoff, vma->vm_pgoff+sz, new);
2638 if (err && new)
2639 sp_free(new);
2640 return err;
2641}
2642
2643/* Free a backing policy store on inode delete. */
2644void mpol_free_shared_policy(struct shared_policy *p)
2645{
2646 struct sp_node *n;
2647 struct rb_node *next;
2648
2649 if (!p->root.rb_node)
2650 return;
2651 write_lock(&p->lock);
2652 next = rb_first(&p->root);
2653 while (next) {
2654 n = rb_entry(next, struct sp_node, nd);
2655 next = rb_next(&n->nd);
2656 sp_delete(p, n);
2657 }
2658 write_unlock(&p->lock);
2659}
2660
2661#ifdef CONFIG_NUMA_BALANCING
2662static int __initdata numabalancing_override;
2663
2664static void __init check_numabalancing_enable(void)
2665{
2666 bool numabalancing_default = false;
2667
2668 if (IS_ENABLED(CONFIG_NUMA_BALANCING_DEFAULT_ENABLED))
2669 numabalancing_default = true;
2670
2671 /* Parsed by setup_numabalancing. override == 1 enables, -1 disables */
2672 if (numabalancing_override)
2673 set_numabalancing_state(numabalancing_override == 1);
2674
2675 if (num_online_nodes() > 1 && !numabalancing_override) {
2676 pr_info("%s automatic NUMA balancing. Configure with numa_balancing= or the kernel.numa_balancing sysctl\n",
2677 numabalancing_default ? "Enabling" : "Disabling");
2678 set_numabalancing_state(numabalancing_default);
2679 }
2680}
2681
2682static int __init setup_numabalancing(char *str)
2683{
2684 int ret = 0;
2685 if (!str)
2686 goto out;
2687
2688 if (!strcmp(str, "enable")) {
2689 numabalancing_override = 1;
2690 ret = 1;
2691 } else if (!strcmp(str, "disable")) {
2692 numabalancing_override = -1;
2693 ret = 1;
2694 }
2695out:
2696 if (!ret)
2697 pr_warn("Unable to parse numa_balancing=\n");
2698
2699 return ret;
2700}
2701__setup("numa_balancing=", setup_numabalancing);
2702#else
2703static inline void __init check_numabalancing_enable(void)
2704{
2705}
2706#endif /* CONFIG_NUMA_BALANCING */
2707
2708/* assumes fs == KERNEL_DS */
2709void __init numa_policy_init(void)
2710{
2711 nodemask_t interleave_nodes;
2712 unsigned long largest = 0;
2713 int nid, prefer = 0;
2714
2715 policy_cache = kmem_cache_create("numa_policy",
2716 sizeof(struct mempolicy),
2717 0, SLAB_PANIC, NULL);
2718
2719 sn_cache = kmem_cache_create("shared_policy_node",
2720 sizeof(struct sp_node),
2721 0, SLAB_PANIC, NULL);
2722
2723 for_each_node(nid) {
2724 preferred_node_policy[nid] = (struct mempolicy) {
2725 .refcnt = ATOMIC_INIT(1),
2726 .mode = MPOL_PREFERRED,
2727 .flags = MPOL_F_MOF | MPOL_F_MORON,
2728 .v = { .preferred_node = nid, },
2729 };
2730 }
2731
2732 /*
2733 * Set interleaving policy for system init. Interleaving is only
2734 * enabled across suitably sized nodes (default is >= 16MB), or
2735 * fall back to the largest node if they're all smaller.
2736 */
2737 nodes_clear(interleave_nodes);
2738 for_each_node_state(nid, N_MEMORY) {
2739 unsigned long total_pages = node_present_pages(nid);
2740
2741 /* Preserve the largest node */
2742 if (largest < total_pages) {
2743 largest = total_pages;
2744 prefer = nid;
2745 }
2746
2747 /* Interleave this node? */
2748 if ((total_pages << PAGE_SHIFT) >= (16 << 20))
2749 node_set(nid, interleave_nodes);
2750 }
2751
2752 /* All too small, use the largest */
2753 if (unlikely(nodes_empty(interleave_nodes)))
2754 node_set(prefer, interleave_nodes);
2755
2756 if (do_set_mempolicy(MPOL_INTERLEAVE, 0, &interleave_nodes))
2757 pr_err("%s: interleaving failed\n", __func__);
2758
2759 check_numabalancing_enable();
2760}
2761
2762/* Reset policy of current process to default */
2763void numa_default_policy(void)
2764{
2765 do_set_mempolicy(MPOL_DEFAULT, 0, NULL);
2766}
2767
2768/*
2769 * Parse and format mempolicy from/to strings
2770 */
2771
2772/*
2773 * "local" is implemented internally by MPOL_PREFERRED with MPOL_F_LOCAL flag.
2774 */
2775static const char * const policy_modes[] =
2776{
2777 [MPOL_DEFAULT] = "default",
2778 [MPOL_PREFERRED] = "prefer",
2779 [MPOL_BIND] = "bind",
2780 [MPOL_INTERLEAVE] = "interleave",
2781 [MPOL_LOCAL] = "local",
2782};
2783
2784
2785#ifdef CONFIG_TMPFS
2786/**
2787 * mpol_parse_str - parse string to mempolicy, for tmpfs mpol mount option.
2788 * @str: string containing mempolicy to parse
2789 * @mpol: pointer to struct mempolicy pointer, returned on success.
2790 *
2791 * Format of input:
2792 * <mode>[=<flags>][:<nodelist>]
2793 *
2794 * On success, returns 0, else 1
2795 */
2796int mpol_parse_str(char *str, struct mempolicy **mpol)
2797{
2798 struct mempolicy *new = NULL;
2799 unsigned short mode_flags;
2800 nodemask_t nodes;
2801 char *nodelist = strchr(str, ':');
2802 char *flags = strchr(str, '=');
2803 int err = 1, mode;
2804
2805 if (nodelist) {
2806 /* NUL-terminate mode or flags string */
2807 *nodelist++ = '\0';
2808 if (nodelist_parse(nodelist, nodes))
2809 goto out;
2810 if (!nodes_subset(nodes, node_states[N_MEMORY]))
2811 goto out;
2812 } else
2813 nodes_clear(nodes);
2814
2815 if (flags)
2816 *flags++ = '\0'; /* terminate mode string */
2817
2818 mode = match_string(policy_modes, MPOL_MAX, str);
2819 if (mode < 0)
2820 goto out;
2821
2822 switch (mode) {
2823 case MPOL_PREFERRED:
2824 /*
2825 * Insist on a nodelist of one node only
2826 */
2827 if (nodelist) {
2828 char *rest = nodelist;
2829 while (isdigit(*rest))
2830 rest++;
2831 if (*rest)
2832 goto out;
2833 }
2834 break;
2835 case MPOL_INTERLEAVE:
2836 /*
2837 * Default to online nodes with memory if no nodelist
2838 */
2839 if (!nodelist)
2840 nodes = node_states[N_MEMORY];
2841 break;
2842 case MPOL_LOCAL:
2843 /*
2844 * Don't allow a nodelist; mpol_new() checks flags
2845 */
2846 if (nodelist)
2847 goto out;
2848 mode = MPOL_PREFERRED;
2849 break;
2850 case MPOL_DEFAULT:
2851 /*
2852 * Insist on a empty nodelist
2853 */
2854 if (!nodelist)
2855 err = 0;
2856 goto out;
2857 case MPOL_BIND:
2858 /*
2859 * Insist on a nodelist
2860 */
2861 if (!nodelist)
2862 goto out;
2863 }
2864
2865 mode_flags = 0;
2866 if (flags) {
2867 /*
2868 * Currently, we only support two mutually exclusive
2869 * mode flags.
2870 */
2871 if (!strcmp(flags, "static"))
2872 mode_flags |= MPOL_F_STATIC_NODES;
2873 else if (!strcmp(flags, "relative"))
2874 mode_flags |= MPOL_F_RELATIVE_NODES;
2875 else
2876 goto out;
2877 }
2878
2879 new = mpol_new(mode, mode_flags, &nodes);
2880 if (IS_ERR(new))
2881 goto out;
2882
2883 /*
2884 * Save nodes for mpol_to_str() to show the tmpfs mount options
2885 * for /proc/mounts, /proc/pid/mounts and /proc/pid/mountinfo.
2886 */
2887 if (mode != MPOL_PREFERRED)
2888 new->v.nodes = nodes;
2889 else if (nodelist)
2890 new->v.preferred_node = first_node(nodes);
2891 else
2892 new->flags |= MPOL_F_LOCAL;
2893
2894 /*
2895 * Save nodes for contextualization: this will be used to "clone"
2896 * the mempolicy in a specific context [cpuset] at a later time.
2897 */
2898 new->w.user_nodemask = nodes;
2899
2900 err = 0;
2901
2902out:
2903 /* Restore string for error message */
2904 if (nodelist)
2905 *--nodelist = ':';
2906 if (flags)
2907 *--flags = '=';
2908 if (!err)
2909 *mpol = new;
2910 return err;
2911}
2912#endif /* CONFIG_TMPFS */
2913
2914/**
2915 * mpol_to_str - format a mempolicy structure for printing
2916 * @buffer: to contain formatted mempolicy string
2917 * @maxlen: length of @buffer
2918 * @pol: pointer to mempolicy to be formatted
2919 *
2920 * Convert @pol into a string. If @buffer is too short, truncate the string.
2921 * Recommend a @maxlen of at least 32 for the longest mode, "interleave", the
2922 * longest flag, "relative", and to display at least a few node ids.
2923 */
2924void mpol_to_str(char *buffer, int maxlen, struct mempolicy *pol)
2925{
2926 char *p = buffer;
2927 nodemask_t nodes = NODE_MASK_NONE;
2928 unsigned short mode = MPOL_DEFAULT;
2929 unsigned short flags = 0;
2930
2931 if (pol && pol != &default_policy && !(pol->flags & MPOL_F_MORON)) {
2932 mode = pol->mode;
2933 flags = pol->flags;
2934 }
2935
2936 switch (mode) {
2937 case MPOL_DEFAULT:
2938 break;
2939 case MPOL_PREFERRED:
2940 if (flags & MPOL_F_LOCAL)
2941 mode = MPOL_LOCAL;
2942 else
2943 node_set(pol->v.preferred_node, nodes);
2944 break;
2945 case MPOL_BIND:
2946 case MPOL_INTERLEAVE:
2947 nodes = pol->v.nodes;
2948 break;
2949 default:
2950 WARN_ON_ONCE(1);
2951 snprintf(p, maxlen, "unknown");
2952 return;
2953 }
2954
2955 p += snprintf(p, maxlen, "%s", policy_modes[mode]);
2956
2957 if (flags & MPOL_MODE_FLAGS) {
2958 p += snprintf(p, buffer + maxlen - p, "=");
2959
2960 /*
2961 * Currently, the only defined flags are mutually exclusive
2962 */
2963 if (flags & MPOL_F_STATIC_NODES)
2964 p += snprintf(p, buffer + maxlen - p, "static");
2965 else if (flags & MPOL_F_RELATIVE_NODES)
2966 p += snprintf(p, buffer + maxlen - p, "relative");
2967 }
2968
2969 if (!nodes_empty(nodes))
2970 p += scnprintf(p, buffer + maxlen - p, ":%*pbl",
2971 nodemask_pr_args(&nodes));
2972}
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 * preferred many Try a set of nodes first before normal fallback. This is
35 * similar to preferred without the special case.
36 *
37 * default Allocate on the local node first, or when on a VMA
38 * use the process policy. This is what Linux always did
39 * in a NUMA aware kernel and still does by, ahem, default.
40 *
41 * The process policy is applied for most non interrupt memory allocations
42 * in that process' context. Interrupts ignore the policies and always
43 * try to allocate on the local CPU. The VMA policy is only applied for memory
44 * allocations for a VMA in the VM.
45 *
46 * Currently there are a few corner cases in swapping where the policy
47 * is not applied, but the majority should be handled. When process policy
48 * is used it is not remembered over swap outs/swap ins.
49 *
50 * Only the highest zone in the zone hierarchy gets policied. Allocations
51 * requesting a lower zone just use default policy. This implies that
52 * on systems with highmem kernel lowmem allocation don't get policied.
53 * Same with GFP_DMA allocations.
54 *
55 * For shmem/tmpfs shared memory the policy is shared between
56 * all users and remembered even when nobody has memory mapped.
57 */
58
59/* Notebook:
60 fix mmap readahead to honour policy and enable policy for any page cache
61 object
62 statistics for bigpages
63 global policy for page cache? currently it uses process policy. Requires
64 first item above.
65 handle mremap for shared memory (currently ignored for the policy)
66 grows down?
67 make bind policy root only? It can trigger oom much faster and the
68 kernel is not always grateful with that.
69*/
70
71#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
72
73#include <linux/mempolicy.h>
74#include <linux/pagewalk.h>
75#include <linux/highmem.h>
76#include <linux/hugetlb.h>
77#include <linux/kernel.h>
78#include <linux/sched.h>
79#include <linux/sched/mm.h>
80#include <linux/sched/numa_balancing.h>
81#include <linux/sched/task.h>
82#include <linux/nodemask.h>
83#include <linux/cpuset.h>
84#include <linux/slab.h>
85#include <linux/string.h>
86#include <linux/export.h>
87#include <linux/nsproxy.h>
88#include <linux/interrupt.h>
89#include <linux/init.h>
90#include <linux/compat.h>
91#include <linux/ptrace.h>
92#include <linux/swap.h>
93#include <linux/seq_file.h>
94#include <linux/proc_fs.h>
95#include <linux/migrate.h>
96#include <linux/ksm.h>
97#include <linux/rmap.h>
98#include <linux/security.h>
99#include <linux/syscalls.h>
100#include <linux/ctype.h>
101#include <linux/mm_inline.h>
102#include <linux/mmu_notifier.h>
103#include <linux/printk.h>
104#include <linux/swapops.h>
105
106#include <asm/tlbflush.h>
107#include <asm/tlb.h>
108#include <linux/uaccess.h>
109
110#include "internal.h"
111
112/* Internal flags */
113#define MPOL_MF_DISCONTIG_OK (MPOL_MF_INTERNAL << 0) /* Skip checks for continuous vmas */
114#define MPOL_MF_INVERT (MPOL_MF_INTERNAL << 1) /* Invert check for nodemask */
115#define MPOL_MF_WRLOCK (MPOL_MF_INTERNAL << 2) /* Write-lock walked vmas */
116
117static struct kmem_cache *policy_cache;
118static struct kmem_cache *sn_cache;
119
120/* Highest zone. An specific allocation for a zone below that is not
121 policied. */
122enum zone_type policy_zone = 0;
123
124/*
125 * run-time system-wide default policy => local allocation
126 */
127static struct mempolicy default_policy = {
128 .refcnt = ATOMIC_INIT(1), /* never free it */
129 .mode = MPOL_LOCAL,
130};
131
132static struct mempolicy preferred_node_policy[MAX_NUMNODES];
133
134/**
135 * numa_nearest_node - Find nearest node by state
136 * @node: Node id to start the search
137 * @state: State to filter the search
138 *
139 * Lookup the closest node by distance if @nid is not in state.
140 *
141 * Return: this @node if it is in state, otherwise the closest node by distance
142 */
143int numa_nearest_node(int node, unsigned int state)
144{
145 int min_dist = INT_MAX, dist, n, min_node;
146
147 if (state >= NR_NODE_STATES)
148 return -EINVAL;
149
150 if (node == NUMA_NO_NODE || node_state(node, state))
151 return node;
152
153 min_node = node;
154 for_each_node_state(n, state) {
155 dist = node_distance(node, n);
156 if (dist < min_dist) {
157 min_dist = dist;
158 min_node = n;
159 }
160 }
161
162 return min_node;
163}
164EXPORT_SYMBOL_GPL(numa_nearest_node);
165
166struct mempolicy *get_task_policy(struct task_struct *p)
167{
168 struct mempolicy *pol = p->mempolicy;
169 int node;
170
171 if (pol)
172 return pol;
173
174 node = numa_node_id();
175 if (node != NUMA_NO_NODE) {
176 pol = &preferred_node_policy[node];
177 /* preferred_node_policy is not initialised early in boot */
178 if (pol->mode)
179 return pol;
180 }
181
182 return &default_policy;
183}
184
185static const struct mempolicy_operations {
186 int (*create)(struct mempolicy *pol, const nodemask_t *nodes);
187 void (*rebind)(struct mempolicy *pol, const nodemask_t *nodes);
188} mpol_ops[MPOL_MAX];
189
190static inline int mpol_store_user_nodemask(const struct mempolicy *pol)
191{
192 return pol->flags & MPOL_MODE_FLAGS;
193}
194
195static void mpol_relative_nodemask(nodemask_t *ret, const nodemask_t *orig,
196 const nodemask_t *rel)
197{
198 nodemask_t tmp;
199 nodes_fold(tmp, *orig, nodes_weight(*rel));
200 nodes_onto(*ret, tmp, *rel);
201}
202
203static int mpol_new_nodemask(struct mempolicy *pol, const nodemask_t *nodes)
204{
205 if (nodes_empty(*nodes))
206 return -EINVAL;
207 pol->nodes = *nodes;
208 return 0;
209}
210
211static int mpol_new_preferred(struct mempolicy *pol, const nodemask_t *nodes)
212{
213 if (nodes_empty(*nodes))
214 return -EINVAL;
215
216 nodes_clear(pol->nodes);
217 node_set(first_node(*nodes), pol->nodes);
218 return 0;
219}
220
221/*
222 * mpol_set_nodemask is called after mpol_new() to set up the nodemask, if
223 * any, for the new policy. mpol_new() has already validated the nodes
224 * parameter with respect to the policy mode and flags.
225 *
226 * Must be called holding task's alloc_lock to protect task's mems_allowed
227 * and mempolicy. May also be called holding the mmap_lock for write.
228 */
229static int mpol_set_nodemask(struct mempolicy *pol,
230 const nodemask_t *nodes, struct nodemask_scratch *nsc)
231{
232 int ret;
233
234 /*
235 * Default (pol==NULL) resp. local memory policies are not a
236 * subject of any remapping. They also do not need any special
237 * constructor.
238 */
239 if (!pol || pol->mode == MPOL_LOCAL)
240 return 0;
241
242 /* Check N_MEMORY */
243 nodes_and(nsc->mask1,
244 cpuset_current_mems_allowed, node_states[N_MEMORY]);
245
246 VM_BUG_ON(!nodes);
247
248 if (pol->flags & MPOL_F_RELATIVE_NODES)
249 mpol_relative_nodemask(&nsc->mask2, nodes, &nsc->mask1);
250 else
251 nodes_and(nsc->mask2, *nodes, nsc->mask1);
252
253 if (mpol_store_user_nodemask(pol))
254 pol->w.user_nodemask = *nodes;
255 else
256 pol->w.cpuset_mems_allowed = cpuset_current_mems_allowed;
257
258 ret = mpol_ops[pol->mode].create(pol, &nsc->mask2);
259 return ret;
260}
261
262/*
263 * This function just creates a new policy, does some check and simple
264 * initialization. You must invoke mpol_set_nodemask() to set nodes.
265 */
266static struct mempolicy *mpol_new(unsigned short mode, unsigned short flags,
267 nodemask_t *nodes)
268{
269 struct mempolicy *policy;
270
271 if (mode == MPOL_DEFAULT) {
272 if (nodes && !nodes_empty(*nodes))
273 return ERR_PTR(-EINVAL);
274 return NULL;
275 }
276 VM_BUG_ON(!nodes);
277
278 /*
279 * MPOL_PREFERRED cannot be used with MPOL_F_STATIC_NODES or
280 * MPOL_F_RELATIVE_NODES if the nodemask is empty (local allocation).
281 * All other modes require a valid pointer to a non-empty nodemask.
282 */
283 if (mode == MPOL_PREFERRED) {
284 if (nodes_empty(*nodes)) {
285 if (((flags & MPOL_F_STATIC_NODES) ||
286 (flags & MPOL_F_RELATIVE_NODES)))
287 return ERR_PTR(-EINVAL);
288
289 mode = MPOL_LOCAL;
290 }
291 } else if (mode == MPOL_LOCAL) {
292 if (!nodes_empty(*nodes) ||
293 (flags & MPOL_F_STATIC_NODES) ||
294 (flags & MPOL_F_RELATIVE_NODES))
295 return ERR_PTR(-EINVAL);
296 } else if (nodes_empty(*nodes))
297 return ERR_PTR(-EINVAL);
298
299 policy = kmem_cache_alloc(policy_cache, GFP_KERNEL);
300 if (!policy)
301 return ERR_PTR(-ENOMEM);
302 atomic_set(&policy->refcnt, 1);
303 policy->mode = mode;
304 policy->flags = flags;
305 policy->home_node = NUMA_NO_NODE;
306
307 return policy;
308}
309
310/* Slow path of a mpol destructor. */
311void __mpol_put(struct mempolicy *pol)
312{
313 if (!atomic_dec_and_test(&pol->refcnt))
314 return;
315 kmem_cache_free(policy_cache, pol);
316}
317
318static void mpol_rebind_default(struct mempolicy *pol, const nodemask_t *nodes)
319{
320}
321
322static void mpol_rebind_nodemask(struct mempolicy *pol, const nodemask_t *nodes)
323{
324 nodemask_t tmp;
325
326 if (pol->flags & MPOL_F_STATIC_NODES)
327 nodes_and(tmp, pol->w.user_nodemask, *nodes);
328 else if (pol->flags & MPOL_F_RELATIVE_NODES)
329 mpol_relative_nodemask(&tmp, &pol->w.user_nodemask, nodes);
330 else {
331 nodes_remap(tmp, pol->nodes, pol->w.cpuset_mems_allowed,
332 *nodes);
333 pol->w.cpuset_mems_allowed = *nodes;
334 }
335
336 if (nodes_empty(tmp))
337 tmp = *nodes;
338
339 pol->nodes = tmp;
340}
341
342static void mpol_rebind_preferred(struct mempolicy *pol,
343 const nodemask_t *nodes)
344{
345 pol->w.cpuset_mems_allowed = *nodes;
346}
347
348/*
349 * mpol_rebind_policy - Migrate a policy to a different set of nodes
350 *
351 * Per-vma policies are protected by mmap_lock. Allocations using per-task
352 * policies are protected by task->mems_allowed_seq to prevent a premature
353 * OOM/allocation failure due to parallel nodemask modification.
354 */
355static void mpol_rebind_policy(struct mempolicy *pol, const nodemask_t *newmask)
356{
357 if (!pol || pol->mode == MPOL_LOCAL)
358 return;
359 if (!mpol_store_user_nodemask(pol) &&
360 nodes_equal(pol->w.cpuset_mems_allowed, *newmask))
361 return;
362
363 mpol_ops[pol->mode].rebind(pol, newmask);
364}
365
366/*
367 * Wrapper for mpol_rebind_policy() that just requires task
368 * pointer, and updates task mempolicy.
369 *
370 * Called with task's alloc_lock held.
371 */
372void mpol_rebind_task(struct task_struct *tsk, const nodemask_t *new)
373{
374 mpol_rebind_policy(tsk->mempolicy, new);
375}
376
377/*
378 * Rebind each vma in mm to new nodemask.
379 *
380 * Call holding a reference to mm. Takes mm->mmap_lock during call.
381 */
382void mpol_rebind_mm(struct mm_struct *mm, nodemask_t *new)
383{
384 struct vm_area_struct *vma;
385 VMA_ITERATOR(vmi, mm, 0);
386
387 mmap_write_lock(mm);
388 for_each_vma(vmi, vma) {
389 vma_start_write(vma);
390 mpol_rebind_policy(vma->vm_policy, new);
391 }
392 mmap_write_unlock(mm);
393}
394
395static const struct mempolicy_operations mpol_ops[MPOL_MAX] = {
396 [MPOL_DEFAULT] = {
397 .rebind = mpol_rebind_default,
398 },
399 [MPOL_INTERLEAVE] = {
400 .create = mpol_new_nodemask,
401 .rebind = mpol_rebind_nodemask,
402 },
403 [MPOL_PREFERRED] = {
404 .create = mpol_new_preferred,
405 .rebind = mpol_rebind_preferred,
406 },
407 [MPOL_BIND] = {
408 .create = mpol_new_nodemask,
409 .rebind = mpol_rebind_nodemask,
410 },
411 [MPOL_LOCAL] = {
412 .rebind = mpol_rebind_default,
413 },
414 [MPOL_PREFERRED_MANY] = {
415 .create = mpol_new_nodemask,
416 .rebind = mpol_rebind_preferred,
417 },
418};
419
420static bool migrate_folio_add(struct folio *folio, struct list_head *foliolist,
421 unsigned long flags);
422static nodemask_t *policy_nodemask(gfp_t gfp, struct mempolicy *pol,
423 pgoff_t ilx, int *nid);
424
425static bool strictly_unmovable(unsigned long flags)
426{
427 /*
428 * STRICT without MOVE flags lets do_mbind() fail immediately with -EIO
429 * if any misplaced page is found.
430 */
431 return (flags & (MPOL_MF_STRICT | MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) ==
432 MPOL_MF_STRICT;
433}
434
435struct migration_mpol { /* for alloc_migration_target_by_mpol() */
436 struct mempolicy *pol;
437 pgoff_t ilx;
438};
439
440struct queue_pages {
441 struct list_head *pagelist;
442 unsigned long flags;
443 nodemask_t *nmask;
444 unsigned long start;
445 unsigned long end;
446 struct vm_area_struct *first;
447 struct folio *large; /* note last large folio encountered */
448 long nr_failed; /* could not be isolated at this time */
449};
450
451/*
452 * Check if the folio's nid is in qp->nmask.
453 *
454 * If MPOL_MF_INVERT is set in qp->flags, check if the nid is
455 * in the invert of qp->nmask.
456 */
457static inline bool queue_folio_required(struct folio *folio,
458 struct queue_pages *qp)
459{
460 int nid = folio_nid(folio);
461 unsigned long flags = qp->flags;
462
463 return node_isset(nid, *qp->nmask) == !(flags & MPOL_MF_INVERT);
464}
465
466static void queue_folios_pmd(pmd_t *pmd, struct mm_walk *walk)
467{
468 struct folio *folio;
469 struct queue_pages *qp = walk->private;
470
471 if (unlikely(is_pmd_migration_entry(*pmd))) {
472 qp->nr_failed++;
473 return;
474 }
475 folio = pfn_folio(pmd_pfn(*pmd));
476 if (is_huge_zero_page(&folio->page)) {
477 walk->action = ACTION_CONTINUE;
478 return;
479 }
480 if (!queue_folio_required(folio, qp))
481 return;
482 if (!(qp->flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) ||
483 !vma_migratable(walk->vma) ||
484 !migrate_folio_add(folio, qp->pagelist, qp->flags))
485 qp->nr_failed++;
486}
487
488/*
489 * Scan through folios, checking if they satisfy the required conditions,
490 * moving them from LRU to local pagelist for migration if they do (or not).
491 *
492 * queue_folios_pte_range() has two possible return values:
493 * 0 - continue walking to scan for more, even if an existing folio on the
494 * wrong node could not be isolated and queued for migration.
495 * -EIO - only MPOL_MF_STRICT was specified, without MPOL_MF_MOVE or ..._ALL,
496 * and an existing folio was on a node that does not follow the policy.
497 */
498static int queue_folios_pte_range(pmd_t *pmd, unsigned long addr,
499 unsigned long end, struct mm_walk *walk)
500{
501 struct vm_area_struct *vma = walk->vma;
502 struct folio *folio;
503 struct queue_pages *qp = walk->private;
504 unsigned long flags = qp->flags;
505 pte_t *pte, *mapped_pte;
506 pte_t ptent;
507 spinlock_t *ptl;
508
509 ptl = pmd_trans_huge_lock(pmd, vma);
510 if (ptl) {
511 queue_folios_pmd(pmd, walk);
512 spin_unlock(ptl);
513 goto out;
514 }
515
516 mapped_pte = pte = pte_offset_map_lock(walk->mm, pmd, addr, &ptl);
517 if (!pte) {
518 walk->action = ACTION_AGAIN;
519 return 0;
520 }
521 for (; addr != end; pte++, addr += PAGE_SIZE) {
522 ptent = ptep_get(pte);
523 if (pte_none(ptent))
524 continue;
525 if (!pte_present(ptent)) {
526 if (is_migration_entry(pte_to_swp_entry(ptent)))
527 qp->nr_failed++;
528 continue;
529 }
530 folio = vm_normal_folio(vma, addr, ptent);
531 if (!folio || folio_is_zone_device(folio))
532 continue;
533 /*
534 * vm_normal_folio() filters out zero pages, but there might
535 * still be reserved folios to skip, perhaps in a VDSO.
536 */
537 if (folio_test_reserved(folio))
538 continue;
539 if (!queue_folio_required(folio, qp))
540 continue;
541 if (folio_test_large(folio)) {
542 /*
543 * A large folio can only be isolated from LRU once,
544 * but may be mapped by many PTEs (and Copy-On-Write may
545 * intersperse PTEs of other, order 0, folios). This is
546 * a common case, so don't mistake it for failure (but
547 * there can be other cases of multi-mapped pages which
548 * this quick check does not help to filter out - and a
549 * search of the pagelist might grow to be prohibitive).
550 *
551 * migrate_pages(&pagelist) returns nr_failed folios, so
552 * check "large" now so that queue_pages_range() returns
553 * a comparable nr_failed folios. This does imply that
554 * if folio could not be isolated for some racy reason
555 * at its first PTE, later PTEs will not give it another
556 * chance of isolation; but keeps the accounting simple.
557 */
558 if (folio == qp->large)
559 continue;
560 qp->large = folio;
561 }
562 if (!(flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) ||
563 !vma_migratable(vma) ||
564 !migrate_folio_add(folio, qp->pagelist, flags)) {
565 qp->nr_failed++;
566 if (strictly_unmovable(flags))
567 break;
568 }
569 }
570 pte_unmap_unlock(mapped_pte, ptl);
571 cond_resched();
572out:
573 if (qp->nr_failed && strictly_unmovable(flags))
574 return -EIO;
575 return 0;
576}
577
578static int queue_folios_hugetlb(pte_t *pte, unsigned long hmask,
579 unsigned long addr, unsigned long end,
580 struct mm_walk *walk)
581{
582#ifdef CONFIG_HUGETLB_PAGE
583 struct queue_pages *qp = walk->private;
584 unsigned long flags = qp->flags;
585 struct folio *folio;
586 spinlock_t *ptl;
587 pte_t entry;
588
589 ptl = huge_pte_lock(hstate_vma(walk->vma), walk->mm, pte);
590 entry = huge_ptep_get(pte);
591 if (!pte_present(entry)) {
592 if (unlikely(is_hugetlb_entry_migration(entry)))
593 qp->nr_failed++;
594 goto unlock;
595 }
596 folio = pfn_folio(pte_pfn(entry));
597 if (!queue_folio_required(folio, qp))
598 goto unlock;
599 if (!(flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) ||
600 !vma_migratable(walk->vma)) {
601 qp->nr_failed++;
602 goto unlock;
603 }
604 /*
605 * Unless MPOL_MF_MOVE_ALL, we try to avoid migrating a shared folio.
606 * Choosing not to migrate a shared folio is not counted as a failure.
607 *
608 * To check if the folio is shared, ideally we want to make sure
609 * every page is mapped to the same process. Doing that is very
610 * expensive, so check the estimated sharers of the folio instead.
611 */
612 if ((flags & MPOL_MF_MOVE_ALL) ||
613 (folio_estimated_sharers(folio) == 1 && !hugetlb_pmd_shared(pte)))
614 if (!isolate_hugetlb(folio, qp->pagelist))
615 qp->nr_failed++;
616unlock:
617 spin_unlock(ptl);
618 if (qp->nr_failed && strictly_unmovable(flags))
619 return -EIO;
620#endif
621 return 0;
622}
623
624#ifdef CONFIG_NUMA_BALANCING
625/*
626 * This is used to mark a range of virtual addresses to be inaccessible.
627 * These are later cleared by a NUMA hinting fault. Depending on these
628 * faults, pages may be migrated for better NUMA placement.
629 *
630 * This is assuming that NUMA faults are handled using PROT_NONE. If
631 * an architecture makes a different choice, it will need further
632 * changes to the core.
633 */
634unsigned long change_prot_numa(struct vm_area_struct *vma,
635 unsigned long addr, unsigned long end)
636{
637 struct mmu_gather tlb;
638 long nr_updated;
639
640 tlb_gather_mmu(&tlb, vma->vm_mm);
641
642 nr_updated = change_protection(&tlb, vma, addr, end, MM_CP_PROT_NUMA);
643 if (nr_updated > 0)
644 count_vm_numa_events(NUMA_PTE_UPDATES, nr_updated);
645
646 tlb_finish_mmu(&tlb);
647
648 return nr_updated;
649}
650#endif /* CONFIG_NUMA_BALANCING */
651
652static int queue_pages_test_walk(unsigned long start, unsigned long end,
653 struct mm_walk *walk)
654{
655 struct vm_area_struct *next, *vma = walk->vma;
656 struct queue_pages *qp = walk->private;
657 unsigned long endvma = vma->vm_end;
658 unsigned long flags = qp->flags;
659
660 /* range check first */
661 VM_BUG_ON_VMA(!range_in_vma(vma, start, end), vma);
662
663 if (!qp->first) {
664 qp->first = vma;
665 if (!(flags & MPOL_MF_DISCONTIG_OK) &&
666 (qp->start < vma->vm_start))
667 /* hole at head side of range */
668 return -EFAULT;
669 }
670 next = find_vma(vma->vm_mm, vma->vm_end);
671 if (!(flags & MPOL_MF_DISCONTIG_OK) &&
672 ((vma->vm_end < qp->end) &&
673 (!next || vma->vm_end < next->vm_start)))
674 /* hole at middle or tail of range */
675 return -EFAULT;
676
677 /*
678 * Need check MPOL_MF_STRICT to return -EIO if possible
679 * regardless of vma_migratable
680 */
681 if (!vma_migratable(vma) &&
682 !(flags & MPOL_MF_STRICT))
683 return 1;
684
685 if (endvma > end)
686 endvma = end;
687
688 /*
689 * Check page nodes, and queue pages to move, in the current vma.
690 * But if no moving, and no strict checking, the scan can be skipped.
691 */
692 if (flags & (MPOL_MF_STRICT | MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))
693 return 0;
694 return 1;
695}
696
697static const struct mm_walk_ops queue_pages_walk_ops = {
698 .hugetlb_entry = queue_folios_hugetlb,
699 .pmd_entry = queue_folios_pte_range,
700 .test_walk = queue_pages_test_walk,
701 .walk_lock = PGWALK_RDLOCK,
702};
703
704static const struct mm_walk_ops queue_pages_lock_vma_walk_ops = {
705 .hugetlb_entry = queue_folios_hugetlb,
706 .pmd_entry = queue_folios_pte_range,
707 .test_walk = queue_pages_test_walk,
708 .walk_lock = PGWALK_WRLOCK,
709};
710
711/*
712 * Walk through page tables and collect pages to be migrated.
713 *
714 * If pages found in a given range are not on the required set of @nodes,
715 * and migration is allowed, they are isolated and queued to @pagelist.
716 *
717 * queue_pages_range() may return:
718 * 0 - all pages already on the right node, or successfully queued for moving
719 * (or neither strict checking nor moving requested: only range checking).
720 * >0 - this number of misplaced folios could not be queued for moving
721 * (a hugetlbfs page or a transparent huge page being counted as 1).
722 * -EIO - a misplaced page found, when MPOL_MF_STRICT specified without MOVEs.
723 * -EFAULT - a hole in the memory range, when MPOL_MF_DISCONTIG_OK unspecified.
724 */
725static long
726queue_pages_range(struct mm_struct *mm, unsigned long start, unsigned long end,
727 nodemask_t *nodes, unsigned long flags,
728 struct list_head *pagelist)
729{
730 int err;
731 struct queue_pages qp = {
732 .pagelist = pagelist,
733 .flags = flags,
734 .nmask = nodes,
735 .start = start,
736 .end = end,
737 .first = NULL,
738 };
739 const struct mm_walk_ops *ops = (flags & MPOL_MF_WRLOCK) ?
740 &queue_pages_lock_vma_walk_ops : &queue_pages_walk_ops;
741
742 err = walk_page_range(mm, start, end, ops, &qp);
743
744 if (!qp.first)
745 /* whole range in hole */
746 err = -EFAULT;
747
748 return err ? : qp.nr_failed;
749}
750
751/*
752 * Apply policy to a single VMA
753 * This must be called with the mmap_lock held for writing.
754 */
755static int vma_replace_policy(struct vm_area_struct *vma,
756 struct mempolicy *pol)
757{
758 int err;
759 struct mempolicy *old;
760 struct mempolicy *new;
761
762 vma_assert_write_locked(vma);
763
764 new = mpol_dup(pol);
765 if (IS_ERR(new))
766 return PTR_ERR(new);
767
768 if (vma->vm_ops && vma->vm_ops->set_policy) {
769 err = vma->vm_ops->set_policy(vma, new);
770 if (err)
771 goto err_out;
772 }
773
774 old = vma->vm_policy;
775 vma->vm_policy = new; /* protected by mmap_lock */
776 mpol_put(old);
777
778 return 0;
779 err_out:
780 mpol_put(new);
781 return err;
782}
783
784/* Split or merge the VMA (if required) and apply the new policy */
785static int mbind_range(struct vma_iterator *vmi, struct vm_area_struct *vma,
786 struct vm_area_struct **prev, unsigned long start,
787 unsigned long end, struct mempolicy *new_pol)
788{
789 unsigned long vmstart, vmend;
790
791 vmend = min(end, vma->vm_end);
792 if (start > vma->vm_start) {
793 *prev = vma;
794 vmstart = start;
795 } else {
796 vmstart = vma->vm_start;
797 }
798
799 if (mpol_equal(vma->vm_policy, new_pol)) {
800 *prev = vma;
801 return 0;
802 }
803
804 vma = vma_modify_policy(vmi, *prev, vma, vmstart, vmend, new_pol);
805 if (IS_ERR(vma))
806 return PTR_ERR(vma);
807
808 *prev = vma;
809 return vma_replace_policy(vma, new_pol);
810}
811
812/* Set the process memory policy */
813static long do_set_mempolicy(unsigned short mode, unsigned short flags,
814 nodemask_t *nodes)
815{
816 struct mempolicy *new, *old;
817 NODEMASK_SCRATCH(scratch);
818 int ret;
819
820 if (!scratch)
821 return -ENOMEM;
822
823 new = mpol_new(mode, flags, nodes);
824 if (IS_ERR(new)) {
825 ret = PTR_ERR(new);
826 goto out;
827 }
828
829 task_lock(current);
830 ret = mpol_set_nodemask(new, nodes, scratch);
831 if (ret) {
832 task_unlock(current);
833 mpol_put(new);
834 goto out;
835 }
836
837 old = current->mempolicy;
838 current->mempolicy = new;
839 if (new && new->mode == MPOL_INTERLEAVE)
840 current->il_prev = MAX_NUMNODES-1;
841 task_unlock(current);
842 mpol_put(old);
843 ret = 0;
844out:
845 NODEMASK_SCRATCH_FREE(scratch);
846 return ret;
847}
848
849/*
850 * Return nodemask for policy for get_mempolicy() query
851 *
852 * Called with task's alloc_lock held
853 */
854static void get_policy_nodemask(struct mempolicy *pol, nodemask_t *nodes)
855{
856 nodes_clear(*nodes);
857 if (pol == &default_policy)
858 return;
859
860 switch (pol->mode) {
861 case MPOL_BIND:
862 case MPOL_INTERLEAVE:
863 case MPOL_PREFERRED:
864 case MPOL_PREFERRED_MANY:
865 *nodes = pol->nodes;
866 break;
867 case MPOL_LOCAL:
868 /* return empty node mask for local allocation */
869 break;
870 default:
871 BUG();
872 }
873}
874
875static int lookup_node(struct mm_struct *mm, unsigned long addr)
876{
877 struct page *p = NULL;
878 int ret;
879
880 ret = get_user_pages_fast(addr & PAGE_MASK, 1, 0, &p);
881 if (ret > 0) {
882 ret = page_to_nid(p);
883 put_page(p);
884 }
885 return ret;
886}
887
888/* Retrieve NUMA policy */
889static long do_get_mempolicy(int *policy, nodemask_t *nmask,
890 unsigned long addr, unsigned long flags)
891{
892 int err;
893 struct mm_struct *mm = current->mm;
894 struct vm_area_struct *vma = NULL;
895 struct mempolicy *pol = current->mempolicy, *pol_refcount = NULL;
896
897 if (flags &
898 ~(unsigned long)(MPOL_F_NODE|MPOL_F_ADDR|MPOL_F_MEMS_ALLOWED))
899 return -EINVAL;
900
901 if (flags & MPOL_F_MEMS_ALLOWED) {
902 if (flags & (MPOL_F_NODE|MPOL_F_ADDR))
903 return -EINVAL;
904 *policy = 0; /* just so it's initialized */
905 task_lock(current);
906 *nmask = cpuset_current_mems_allowed;
907 task_unlock(current);
908 return 0;
909 }
910
911 if (flags & MPOL_F_ADDR) {
912 pgoff_t ilx; /* ignored here */
913 /*
914 * Do NOT fall back to task policy if the
915 * vma/shared policy at addr is NULL. We
916 * want to return MPOL_DEFAULT in this case.
917 */
918 mmap_read_lock(mm);
919 vma = vma_lookup(mm, addr);
920 if (!vma) {
921 mmap_read_unlock(mm);
922 return -EFAULT;
923 }
924 pol = __get_vma_policy(vma, addr, &ilx);
925 } else if (addr)
926 return -EINVAL;
927
928 if (!pol)
929 pol = &default_policy; /* indicates default behavior */
930
931 if (flags & MPOL_F_NODE) {
932 if (flags & MPOL_F_ADDR) {
933 /*
934 * Take a refcount on the mpol, because we are about to
935 * drop the mmap_lock, after which only "pol" remains
936 * valid, "vma" is stale.
937 */
938 pol_refcount = pol;
939 vma = NULL;
940 mpol_get(pol);
941 mmap_read_unlock(mm);
942 err = lookup_node(mm, addr);
943 if (err < 0)
944 goto out;
945 *policy = err;
946 } else if (pol == current->mempolicy &&
947 pol->mode == MPOL_INTERLEAVE) {
948 *policy = next_node_in(current->il_prev, pol->nodes);
949 } else {
950 err = -EINVAL;
951 goto out;
952 }
953 } else {
954 *policy = pol == &default_policy ? MPOL_DEFAULT :
955 pol->mode;
956 /*
957 * Internal mempolicy flags must be masked off before exposing
958 * the policy to userspace.
959 */
960 *policy |= (pol->flags & MPOL_MODE_FLAGS);
961 }
962
963 err = 0;
964 if (nmask) {
965 if (mpol_store_user_nodemask(pol)) {
966 *nmask = pol->w.user_nodemask;
967 } else {
968 task_lock(current);
969 get_policy_nodemask(pol, nmask);
970 task_unlock(current);
971 }
972 }
973
974 out:
975 mpol_cond_put(pol);
976 if (vma)
977 mmap_read_unlock(mm);
978 if (pol_refcount)
979 mpol_put(pol_refcount);
980 return err;
981}
982
983#ifdef CONFIG_MIGRATION
984static bool migrate_folio_add(struct folio *folio, struct list_head *foliolist,
985 unsigned long flags)
986{
987 /*
988 * Unless MPOL_MF_MOVE_ALL, we try to avoid migrating a shared folio.
989 * Choosing not to migrate a shared folio is not counted as a failure.
990 *
991 * To check if the folio is shared, ideally we want to make sure
992 * every page is mapped to the same process. Doing that is very
993 * expensive, so check the estimated sharers of the folio instead.
994 */
995 if ((flags & MPOL_MF_MOVE_ALL) || folio_estimated_sharers(folio) == 1) {
996 if (folio_isolate_lru(folio)) {
997 list_add_tail(&folio->lru, foliolist);
998 node_stat_mod_folio(folio,
999 NR_ISOLATED_ANON + folio_is_file_lru(folio),
1000 folio_nr_pages(folio));
1001 } else {
1002 /*
1003 * Non-movable folio may reach here. And, there may be
1004 * temporary off LRU folios or non-LRU movable folios.
1005 * Treat them as unmovable folios since they can't be
1006 * isolated, so they can't be moved at the moment.
1007 */
1008 return false;
1009 }
1010 }
1011 return true;
1012}
1013
1014/*
1015 * Migrate pages from one node to a target node.
1016 * Returns error or the number of pages not migrated.
1017 */
1018static long migrate_to_node(struct mm_struct *mm, int source, int dest,
1019 int flags)
1020{
1021 nodemask_t nmask;
1022 struct vm_area_struct *vma;
1023 LIST_HEAD(pagelist);
1024 long nr_failed;
1025 long err = 0;
1026 struct migration_target_control mtc = {
1027 .nid = dest,
1028 .gfp_mask = GFP_HIGHUSER_MOVABLE | __GFP_THISNODE,
1029 };
1030
1031 nodes_clear(nmask);
1032 node_set(source, nmask);
1033
1034 VM_BUG_ON(!(flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)));
1035
1036 mmap_read_lock(mm);
1037 vma = find_vma(mm, 0);
1038
1039 /*
1040 * This does not migrate the range, but isolates all pages that
1041 * need migration. Between passing in the full user address
1042 * space range and MPOL_MF_DISCONTIG_OK, this call cannot fail,
1043 * but passes back the count of pages which could not be isolated.
1044 */
1045 nr_failed = queue_pages_range(mm, vma->vm_start, mm->task_size, &nmask,
1046 flags | MPOL_MF_DISCONTIG_OK, &pagelist);
1047 mmap_read_unlock(mm);
1048
1049 if (!list_empty(&pagelist)) {
1050 err = migrate_pages(&pagelist, alloc_migration_target, NULL,
1051 (unsigned long)&mtc, MIGRATE_SYNC, MR_SYSCALL, NULL);
1052 if (err)
1053 putback_movable_pages(&pagelist);
1054 }
1055
1056 if (err >= 0)
1057 err += nr_failed;
1058 return err;
1059}
1060
1061/*
1062 * Move pages between the two nodesets so as to preserve the physical
1063 * layout as much as possible.
1064 *
1065 * Returns the number of page that could not be moved.
1066 */
1067int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
1068 const nodemask_t *to, int flags)
1069{
1070 long nr_failed = 0;
1071 long err = 0;
1072 nodemask_t tmp;
1073
1074 lru_cache_disable();
1075
1076 /*
1077 * Find a 'source' bit set in 'tmp' whose corresponding 'dest'
1078 * bit in 'to' is not also set in 'tmp'. Clear the found 'source'
1079 * bit in 'tmp', and return that <source, dest> pair for migration.
1080 * The pair of nodemasks 'to' and 'from' define the map.
1081 *
1082 * If no pair of bits is found that way, fallback to picking some
1083 * pair of 'source' and 'dest' bits that are not the same. If the
1084 * 'source' and 'dest' bits are the same, this represents a node
1085 * that will be migrating to itself, so no pages need move.
1086 *
1087 * If no bits are left in 'tmp', or if all remaining bits left
1088 * in 'tmp' correspond to the same bit in 'to', return false
1089 * (nothing left to migrate).
1090 *
1091 * This lets us pick a pair of nodes to migrate between, such that
1092 * if possible the dest node is not already occupied by some other
1093 * source node, minimizing the risk of overloading the memory on a
1094 * node that would happen if we migrated incoming memory to a node
1095 * before migrating outgoing memory source that same node.
1096 *
1097 * A single scan of tmp is sufficient. As we go, we remember the
1098 * most recent <s, d> pair that moved (s != d). If we find a pair
1099 * that not only moved, but what's better, moved to an empty slot
1100 * (d is not set in tmp), then we break out then, with that pair.
1101 * Otherwise when we finish scanning from_tmp, we at least have the
1102 * most recent <s, d> pair that moved. If we get all the way through
1103 * the scan of tmp without finding any node that moved, much less
1104 * moved to an empty node, then there is nothing left worth migrating.
1105 */
1106
1107 tmp = *from;
1108 while (!nodes_empty(tmp)) {
1109 int s, d;
1110 int source = NUMA_NO_NODE;
1111 int dest = 0;
1112
1113 for_each_node_mask(s, tmp) {
1114
1115 /*
1116 * do_migrate_pages() tries to maintain the relative
1117 * node relationship of the pages established between
1118 * threads and memory areas.
1119 *
1120 * However if the number of source nodes is not equal to
1121 * the number of destination nodes we can not preserve
1122 * this node relative relationship. In that case, skip
1123 * copying memory from a node that is in the destination
1124 * mask.
1125 *
1126 * Example: [2,3,4] -> [3,4,5] moves everything.
1127 * [0-7] - > [3,4,5] moves only 0,1,2,6,7.
1128 */
1129
1130 if ((nodes_weight(*from) != nodes_weight(*to)) &&
1131 (node_isset(s, *to)))
1132 continue;
1133
1134 d = node_remap(s, *from, *to);
1135 if (s == d)
1136 continue;
1137
1138 source = s; /* Node moved. Memorize */
1139 dest = d;
1140
1141 /* dest not in remaining from nodes? */
1142 if (!node_isset(dest, tmp))
1143 break;
1144 }
1145 if (source == NUMA_NO_NODE)
1146 break;
1147
1148 node_clear(source, tmp);
1149 err = migrate_to_node(mm, source, dest, flags);
1150 if (err > 0)
1151 nr_failed += err;
1152 if (err < 0)
1153 break;
1154 }
1155
1156 lru_cache_enable();
1157 if (err < 0)
1158 return err;
1159 return (nr_failed < INT_MAX) ? nr_failed : INT_MAX;
1160}
1161
1162/*
1163 * Allocate a new folio for page migration, according to NUMA mempolicy.
1164 */
1165static struct folio *alloc_migration_target_by_mpol(struct folio *src,
1166 unsigned long private)
1167{
1168 struct migration_mpol *mmpol = (struct migration_mpol *)private;
1169 struct mempolicy *pol = mmpol->pol;
1170 pgoff_t ilx = mmpol->ilx;
1171 struct page *page;
1172 unsigned int order;
1173 int nid = numa_node_id();
1174 gfp_t gfp;
1175
1176 order = folio_order(src);
1177 ilx += src->index >> order;
1178
1179 if (folio_test_hugetlb(src)) {
1180 nodemask_t *nodemask;
1181 struct hstate *h;
1182
1183 h = folio_hstate(src);
1184 gfp = htlb_alloc_mask(h);
1185 nodemask = policy_nodemask(gfp, pol, ilx, &nid);
1186 return alloc_hugetlb_folio_nodemask(h, nid, nodemask, gfp);
1187 }
1188
1189 if (folio_test_large(src))
1190 gfp = GFP_TRANSHUGE;
1191 else
1192 gfp = GFP_HIGHUSER_MOVABLE | __GFP_RETRY_MAYFAIL | __GFP_COMP;
1193
1194 page = alloc_pages_mpol(gfp, order, pol, ilx, nid);
1195 return page_rmappable_folio(page);
1196}
1197#else
1198
1199static bool migrate_folio_add(struct folio *folio, struct list_head *foliolist,
1200 unsigned long flags)
1201{
1202 return false;
1203}
1204
1205int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
1206 const nodemask_t *to, int flags)
1207{
1208 return -ENOSYS;
1209}
1210
1211static struct folio *alloc_migration_target_by_mpol(struct folio *src,
1212 unsigned long private)
1213{
1214 return NULL;
1215}
1216#endif
1217
1218static long do_mbind(unsigned long start, unsigned long len,
1219 unsigned short mode, unsigned short mode_flags,
1220 nodemask_t *nmask, unsigned long flags)
1221{
1222 struct mm_struct *mm = current->mm;
1223 struct vm_area_struct *vma, *prev;
1224 struct vma_iterator vmi;
1225 struct migration_mpol mmpol;
1226 struct mempolicy *new;
1227 unsigned long end;
1228 long err;
1229 long nr_failed;
1230 LIST_HEAD(pagelist);
1231
1232 if (flags & ~(unsigned long)MPOL_MF_VALID)
1233 return -EINVAL;
1234 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1235 return -EPERM;
1236
1237 if (start & ~PAGE_MASK)
1238 return -EINVAL;
1239
1240 if (mode == MPOL_DEFAULT)
1241 flags &= ~MPOL_MF_STRICT;
1242
1243 len = PAGE_ALIGN(len);
1244 end = start + len;
1245
1246 if (end < start)
1247 return -EINVAL;
1248 if (end == start)
1249 return 0;
1250
1251 new = mpol_new(mode, mode_flags, nmask);
1252 if (IS_ERR(new))
1253 return PTR_ERR(new);
1254
1255 /*
1256 * If we are using the default policy then operation
1257 * on discontinuous address spaces is okay after all
1258 */
1259 if (!new)
1260 flags |= MPOL_MF_DISCONTIG_OK;
1261
1262 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))
1263 lru_cache_disable();
1264 {
1265 NODEMASK_SCRATCH(scratch);
1266 if (scratch) {
1267 mmap_write_lock(mm);
1268 err = mpol_set_nodemask(new, nmask, scratch);
1269 if (err)
1270 mmap_write_unlock(mm);
1271 } else
1272 err = -ENOMEM;
1273 NODEMASK_SCRATCH_FREE(scratch);
1274 }
1275 if (err)
1276 goto mpol_out;
1277
1278 /*
1279 * Lock the VMAs before scanning for pages to migrate,
1280 * to ensure we don't miss a concurrently inserted page.
1281 */
1282 nr_failed = queue_pages_range(mm, start, end, nmask,
1283 flags | MPOL_MF_INVERT | MPOL_MF_WRLOCK, &pagelist);
1284
1285 if (nr_failed < 0) {
1286 err = nr_failed;
1287 nr_failed = 0;
1288 } else {
1289 vma_iter_init(&vmi, mm, start);
1290 prev = vma_prev(&vmi);
1291 for_each_vma_range(vmi, vma, end) {
1292 err = mbind_range(&vmi, vma, &prev, start, end, new);
1293 if (err)
1294 break;
1295 }
1296 }
1297
1298 if (!err && !list_empty(&pagelist)) {
1299 /* Convert MPOL_DEFAULT's NULL to task or default policy */
1300 if (!new) {
1301 new = get_task_policy(current);
1302 mpol_get(new);
1303 }
1304 mmpol.pol = new;
1305 mmpol.ilx = 0;
1306
1307 /*
1308 * In the interleaved case, attempt to allocate on exactly the
1309 * targeted nodes, for the first VMA to be migrated; for later
1310 * VMAs, the nodes will still be interleaved from the targeted
1311 * nodemask, but one by one may be selected differently.
1312 */
1313 if (new->mode == MPOL_INTERLEAVE) {
1314 struct page *page;
1315 unsigned int order;
1316 unsigned long addr = -EFAULT;
1317
1318 list_for_each_entry(page, &pagelist, lru) {
1319 if (!PageKsm(page))
1320 break;
1321 }
1322 if (!list_entry_is_head(page, &pagelist, lru)) {
1323 vma_iter_init(&vmi, mm, start);
1324 for_each_vma_range(vmi, vma, end) {
1325 addr = page_address_in_vma(page, vma);
1326 if (addr != -EFAULT)
1327 break;
1328 }
1329 }
1330 if (addr != -EFAULT) {
1331 order = compound_order(page);
1332 /* We already know the pol, but not the ilx */
1333 mpol_cond_put(get_vma_policy(vma, addr, order,
1334 &mmpol.ilx));
1335 /* Set base from which to increment by index */
1336 mmpol.ilx -= page->index >> order;
1337 }
1338 }
1339 }
1340
1341 mmap_write_unlock(mm);
1342
1343 if (!err && !list_empty(&pagelist)) {
1344 nr_failed |= migrate_pages(&pagelist,
1345 alloc_migration_target_by_mpol, NULL,
1346 (unsigned long)&mmpol, MIGRATE_SYNC,
1347 MR_MEMPOLICY_MBIND, NULL);
1348 }
1349
1350 if (nr_failed && (flags & MPOL_MF_STRICT))
1351 err = -EIO;
1352 if (!list_empty(&pagelist))
1353 putback_movable_pages(&pagelist);
1354mpol_out:
1355 mpol_put(new);
1356 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))
1357 lru_cache_enable();
1358 return err;
1359}
1360
1361/*
1362 * User space interface with variable sized bitmaps for nodelists.
1363 */
1364static int get_bitmap(unsigned long *mask, const unsigned long __user *nmask,
1365 unsigned long maxnode)
1366{
1367 unsigned long nlongs = BITS_TO_LONGS(maxnode);
1368 int ret;
1369
1370 if (in_compat_syscall())
1371 ret = compat_get_bitmap(mask,
1372 (const compat_ulong_t __user *)nmask,
1373 maxnode);
1374 else
1375 ret = copy_from_user(mask, nmask,
1376 nlongs * sizeof(unsigned long));
1377
1378 if (ret)
1379 return -EFAULT;
1380
1381 if (maxnode % BITS_PER_LONG)
1382 mask[nlongs - 1] &= (1UL << (maxnode % BITS_PER_LONG)) - 1;
1383
1384 return 0;
1385}
1386
1387/* Copy a node mask from user space. */
1388static int get_nodes(nodemask_t *nodes, const unsigned long __user *nmask,
1389 unsigned long maxnode)
1390{
1391 --maxnode;
1392 nodes_clear(*nodes);
1393 if (maxnode == 0 || !nmask)
1394 return 0;
1395 if (maxnode > PAGE_SIZE*BITS_PER_BYTE)
1396 return -EINVAL;
1397
1398 /*
1399 * When the user specified more nodes than supported just check
1400 * if the non supported part is all zero, one word at a time,
1401 * starting at the end.
1402 */
1403 while (maxnode > MAX_NUMNODES) {
1404 unsigned long bits = min_t(unsigned long, maxnode, BITS_PER_LONG);
1405 unsigned long t;
1406
1407 if (get_bitmap(&t, &nmask[(maxnode - 1) / BITS_PER_LONG], bits))
1408 return -EFAULT;
1409
1410 if (maxnode - bits >= MAX_NUMNODES) {
1411 maxnode -= bits;
1412 } else {
1413 maxnode = MAX_NUMNODES;
1414 t &= ~((1UL << (MAX_NUMNODES % BITS_PER_LONG)) - 1);
1415 }
1416 if (t)
1417 return -EINVAL;
1418 }
1419
1420 return get_bitmap(nodes_addr(*nodes), nmask, maxnode);
1421}
1422
1423/* Copy a kernel node mask to user space */
1424static int copy_nodes_to_user(unsigned long __user *mask, unsigned long maxnode,
1425 nodemask_t *nodes)
1426{
1427 unsigned long copy = ALIGN(maxnode-1, 64) / 8;
1428 unsigned int nbytes = BITS_TO_LONGS(nr_node_ids) * sizeof(long);
1429 bool compat = in_compat_syscall();
1430
1431 if (compat)
1432 nbytes = BITS_TO_COMPAT_LONGS(nr_node_ids) * sizeof(compat_long_t);
1433
1434 if (copy > nbytes) {
1435 if (copy > PAGE_SIZE)
1436 return -EINVAL;
1437 if (clear_user((char __user *)mask + nbytes, copy - nbytes))
1438 return -EFAULT;
1439 copy = nbytes;
1440 maxnode = nr_node_ids;
1441 }
1442
1443 if (compat)
1444 return compat_put_bitmap((compat_ulong_t __user *)mask,
1445 nodes_addr(*nodes), maxnode);
1446
1447 return copy_to_user(mask, nodes_addr(*nodes), copy) ? -EFAULT : 0;
1448}
1449
1450/* Basic parameter sanity check used by both mbind() and set_mempolicy() */
1451static inline int sanitize_mpol_flags(int *mode, unsigned short *flags)
1452{
1453 *flags = *mode & MPOL_MODE_FLAGS;
1454 *mode &= ~MPOL_MODE_FLAGS;
1455
1456 if ((unsigned int)(*mode) >= MPOL_MAX)
1457 return -EINVAL;
1458 if ((*flags & MPOL_F_STATIC_NODES) && (*flags & MPOL_F_RELATIVE_NODES))
1459 return -EINVAL;
1460 if (*flags & MPOL_F_NUMA_BALANCING) {
1461 if (*mode != MPOL_BIND)
1462 return -EINVAL;
1463 *flags |= (MPOL_F_MOF | MPOL_F_MORON);
1464 }
1465 return 0;
1466}
1467
1468static long kernel_mbind(unsigned long start, unsigned long len,
1469 unsigned long mode, const unsigned long __user *nmask,
1470 unsigned long maxnode, unsigned int flags)
1471{
1472 unsigned short mode_flags;
1473 nodemask_t nodes;
1474 int lmode = mode;
1475 int err;
1476
1477 start = untagged_addr(start);
1478 err = sanitize_mpol_flags(&lmode, &mode_flags);
1479 if (err)
1480 return err;
1481
1482 err = get_nodes(&nodes, nmask, maxnode);
1483 if (err)
1484 return err;
1485
1486 return do_mbind(start, len, lmode, mode_flags, &nodes, flags);
1487}
1488
1489SYSCALL_DEFINE4(set_mempolicy_home_node, unsigned long, start, unsigned long, len,
1490 unsigned long, home_node, unsigned long, flags)
1491{
1492 struct mm_struct *mm = current->mm;
1493 struct vm_area_struct *vma, *prev;
1494 struct mempolicy *new, *old;
1495 unsigned long end;
1496 int err = -ENOENT;
1497 VMA_ITERATOR(vmi, mm, start);
1498
1499 start = untagged_addr(start);
1500 if (start & ~PAGE_MASK)
1501 return -EINVAL;
1502 /*
1503 * flags is used for future extension if any.
1504 */
1505 if (flags != 0)
1506 return -EINVAL;
1507
1508 /*
1509 * Check home_node is online to avoid accessing uninitialized
1510 * NODE_DATA.
1511 */
1512 if (home_node >= MAX_NUMNODES || !node_online(home_node))
1513 return -EINVAL;
1514
1515 len = PAGE_ALIGN(len);
1516 end = start + len;
1517
1518 if (end < start)
1519 return -EINVAL;
1520 if (end == start)
1521 return 0;
1522 mmap_write_lock(mm);
1523 prev = vma_prev(&vmi);
1524 for_each_vma_range(vmi, vma, end) {
1525 /*
1526 * If any vma in the range got policy other than MPOL_BIND
1527 * or MPOL_PREFERRED_MANY we return error. We don't reset
1528 * the home node for vmas we already updated before.
1529 */
1530 old = vma_policy(vma);
1531 if (!old) {
1532 prev = vma;
1533 continue;
1534 }
1535 if (old->mode != MPOL_BIND && old->mode != MPOL_PREFERRED_MANY) {
1536 err = -EOPNOTSUPP;
1537 break;
1538 }
1539 new = mpol_dup(old);
1540 if (IS_ERR(new)) {
1541 err = PTR_ERR(new);
1542 break;
1543 }
1544
1545 vma_start_write(vma);
1546 new->home_node = home_node;
1547 err = mbind_range(&vmi, vma, &prev, start, end, new);
1548 mpol_put(new);
1549 if (err)
1550 break;
1551 }
1552 mmap_write_unlock(mm);
1553 return err;
1554}
1555
1556SYSCALL_DEFINE6(mbind, unsigned long, start, unsigned long, len,
1557 unsigned long, mode, const unsigned long __user *, nmask,
1558 unsigned long, maxnode, unsigned int, flags)
1559{
1560 return kernel_mbind(start, len, mode, nmask, maxnode, flags);
1561}
1562
1563/* Set the process memory policy */
1564static long kernel_set_mempolicy(int mode, const unsigned long __user *nmask,
1565 unsigned long maxnode)
1566{
1567 unsigned short mode_flags;
1568 nodemask_t nodes;
1569 int lmode = mode;
1570 int err;
1571
1572 err = sanitize_mpol_flags(&lmode, &mode_flags);
1573 if (err)
1574 return err;
1575
1576 err = get_nodes(&nodes, nmask, maxnode);
1577 if (err)
1578 return err;
1579
1580 return do_set_mempolicy(lmode, mode_flags, &nodes);
1581}
1582
1583SYSCALL_DEFINE3(set_mempolicy, int, mode, const unsigned long __user *, nmask,
1584 unsigned long, maxnode)
1585{
1586 return kernel_set_mempolicy(mode, nmask, maxnode);
1587}
1588
1589static int kernel_migrate_pages(pid_t pid, unsigned long maxnode,
1590 const unsigned long __user *old_nodes,
1591 const unsigned long __user *new_nodes)
1592{
1593 struct mm_struct *mm = NULL;
1594 struct task_struct *task;
1595 nodemask_t task_nodes;
1596 int err;
1597 nodemask_t *old;
1598 nodemask_t *new;
1599 NODEMASK_SCRATCH(scratch);
1600
1601 if (!scratch)
1602 return -ENOMEM;
1603
1604 old = &scratch->mask1;
1605 new = &scratch->mask2;
1606
1607 err = get_nodes(old, old_nodes, maxnode);
1608 if (err)
1609 goto out;
1610
1611 err = get_nodes(new, new_nodes, maxnode);
1612 if (err)
1613 goto out;
1614
1615 /* Find the mm_struct */
1616 rcu_read_lock();
1617 task = pid ? find_task_by_vpid(pid) : current;
1618 if (!task) {
1619 rcu_read_unlock();
1620 err = -ESRCH;
1621 goto out;
1622 }
1623 get_task_struct(task);
1624
1625 err = -EINVAL;
1626
1627 /*
1628 * Check if this process has the right to modify the specified process.
1629 * Use the regular "ptrace_may_access()" checks.
1630 */
1631 if (!ptrace_may_access(task, PTRACE_MODE_READ_REALCREDS)) {
1632 rcu_read_unlock();
1633 err = -EPERM;
1634 goto out_put;
1635 }
1636 rcu_read_unlock();
1637
1638 task_nodes = cpuset_mems_allowed(task);
1639 /* Is the user allowed to access the target nodes? */
1640 if (!nodes_subset(*new, task_nodes) && !capable(CAP_SYS_NICE)) {
1641 err = -EPERM;
1642 goto out_put;
1643 }
1644
1645 task_nodes = cpuset_mems_allowed(current);
1646 nodes_and(*new, *new, task_nodes);
1647 if (nodes_empty(*new))
1648 goto out_put;
1649
1650 err = security_task_movememory(task);
1651 if (err)
1652 goto out_put;
1653
1654 mm = get_task_mm(task);
1655 put_task_struct(task);
1656
1657 if (!mm) {
1658 err = -EINVAL;
1659 goto out;
1660 }
1661
1662 err = do_migrate_pages(mm, old, new,
1663 capable(CAP_SYS_NICE) ? MPOL_MF_MOVE_ALL : MPOL_MF_MOVE);
1664
1665 mmput(mm);
1666out:
1667 NODEMASK_SCRATCH_FREE(scratch);
1668
1669 return err;
1670
1671out_put:
1672 put_task_struct(task);
1673 goto out;
1674}
1675
1676SYSCALL_DEFINE4(migrate_pages, pid_t, pid, unsigned long, maxnode,
1677 const unsigned long __user *, old_nodes,
1678 const unsigned long __user *, new_nodes)
1679{
1680 return kernel_migrate_pages(pid, maxnode, old_nodes, new_nodes);
1681}
1682
1683/* Retrieve NUMA policy */
1684static int kernel_get_mempolicy(int __user *policy,
1685 unsigned long __user *nmask,
1686 unsigned long maxnode,
1687 unsigned long addr,
1688 unsigned long flags)
1689{
1690 int err;
1691 int pval;
1692 nodemask_t nodes;
1693
1694 if (nmask != NULL && maxnode < nr_node_ids)
1695 return -EINVAL;
1696
1697 addr = untagged_addr(addr);
1698
1699 err = do_get_mempolicy(&pval, &nodes, addr, flags);
1700
1701 if (err)
1702 return err;
1703
1704 if (policy && put_user(pval, policy))
1705 return -EFAULT;
1706
1707 if (nmask)
1708 err = copy_nodes_to_user(nmask, maxnode, &nodes);
1709
1710 return err;
1711}
1712
1713SYSCALL_DEFINE5(get_mempolicy, int __user *, policy,
1714 unsigned long __user *, nmask, unsigned long, maxnode,
1715 unsigned long, addr, unsigned long, flags)
1716{
1717 return kernel_get_mempolicy(policy, nmask, maxnode, addr, flags);
1718}
1719
1720bool vma_migratable(struct vm_area_struct *vma)
1721{
1722 if (vma->vm_flags & (VM_IO | VM_PFNMAP))
1723 return false;
1724
1725 /*
1726 * DAX device mappings require predictable access latency, so avoid
1727 * incurring periodic faults.
1728 */
1729 if (vma_is_dax(vma))
1730 return false;
1731
1732 if (is_vm_hugetlb_page(vma) &&
1733 !hugepage_migration_supported(hstate_vma(vma)))
1734 return false;
1735
1736 /*
1737 * Migration allocates pages in the highest zone. If we cannot
1738 * do so then migration (at least from node to node) is not
1739 * possible.
1740 */
1741 if (vma->vm_file &&
1742 gfp_zone(mapping_gfp_mask(vma->vm_file->f_mapping))
1743 < policy_zone)
1744 return false;
1745 return true;
1746}
1747
1748struct mempolicy *__get_vma_policy(struct vm_area_struct *vma,
1749 unsigned long addr, pgoff_t *ilx)
1750{
1751 *ilx = 0;
1752 return (vma->vm_ops && vma->vm_ops->get_policy) ?
1753 vma->vm_ops->get_policy(vma, addr, ilx) : vma->vm_policy;
1754}
1755
1756/*
1757 * get_vma_policy(@vma, @addr, @order, @ilx)
1758 * @vma: virtual memory area whose policy is sought
1759 * @addr: address in @vma for shared policy lookup
1760 * @order: 0, or appropriate huge_page_order for interleaving
1761 * @ilx: interleave index (output), for use only when MPOL_INTERLEAVE
1762 *
1763 * Returns effective policy for a VMA at specified address.
1764 * Falls back to current->mempolicy or system default policy, as necessary.
1765 * Shared policies [those marked as MPOL_F_SHARED] require an extra reference
1766 * count--added by the get_policy() vm_op, as appropriate--to protect against
1767 * freeing by another task. It is the caller's responsibility to free the
1768 * extra reference for shared policies.
1769 */
1770struct mempolicy *get_vma_policy(struct vm_area_struct *vma,
1771 unsigned long addr, int order, pgoff_t *ilx)
1772{
1773 struct mempolicy *pol;
1774
1775 pol = __get_vma_policy(vma, addr, ilx);
1776 if (!pol)
1777 pol = get_task_policy(current);
1778 if (pol->mode == MPOL_INTERLEAVE) {
1779 *ilx += vma->vm_pgoff >> order;
1780 *ilx += (addr - vma->vm_start) >> (PAGE_SHIFT + order);
1781 }
1782 return pol;
1783}
1784
1785bool vma_policy_mof(struct vm_area_struct *vma)
1786{
1787 struct mempolicy *pol;
1788
1789 if (vma->vm_ops && vma->vm_ops->get_policy) {
1790 bool ret = false;
1791 pgoff_t ilx; /* ignored here */
1792
1793 pol = vma->vm_ops->get_policy(vma, vma->vm_start, &ilx);
1794 if (pol && (pol->flags & MPOL_F_MOF))
1795 ret = true;
1796 mpol_cond_put(pol);
1797
1798 return ret;
1799 }
1800
1801 pol = vma->vm_policy;
1802 if (!pol)
1803 pol = get_task_policy(current);
1804
1805 return pol->flags & MPOL_F_MOF;
1806}
1807
1808bool apply_policy_zone(struct mempolicy *policy, enum zone_type zone)
1809{
1810 enum zone_type dynamic_policy_zone = policy_zone;
1811
1812 BUG_ON(dynamic_policy_zone == ZONE_MOVABLE);
1813
1814 /*
1815 * if policy->nodes has movable memory only,
1816 * we apply policy when gfp_zone(gfp) = ZONE_MOVABLE only.
1817 *
1818 * policy->nodes is intersect with node_states[N_MEMORY].
1819 * so if the following test fails, it implies
1820 * policy->nodes has movable memory only.
1821 */
1822 if (!nodes_intersects(policy->nodes, node_states[N_HIGH_MEMORY]))
1823 dynamic_policy_zone = ZONE_MOVABLE;
1824
1825 return zone >= dynamic_policy_zone;
1826}
1827
1828/* Do dynamic interleaving for a process */
1829static unsigned int interleave_nodes(struct mempolicy *policy)
1830{
1831 unsigned int nid;
1832
1833 nid = next_node_in(current->il_prev, policy->nodes);
1834 if (nid < MAX_NUMNODES)
1835 current->il_prev = nid;
1836 return nid;
1837}
1838
1839/*
1840 * Depending on the memory policy provide a node from which to allocate the
1841 * next slab entry.
1842 */
1843unsigned int mempolicy_slab_node(void)
1844{
1845 struct mempolicy *policy;
1846 int node = numa_mem_id();
1847
1848 if (!in_task())
1849 return node;
1850
1851 policy = current->mempolicy;
1852 if (!policy)
1853 return node;
1854
1855 switch (policy->mode) {
1856 case MPOL_PREFERRED:
1857 return first_node(policy->nodes);
1858
1859 case MPOL_INTERLEAVE:
1860 return interleave_nodes(policy);
1861
1862 case MPOL_BIND:
1863 case MPOL_PREFERRED_MANY:
1864 {
1865 struct zoneref *z;
1866
1867 /*
1868 * Follow bind policy behavior and start allocation at the
1869 * first node.
1870 */
1871 struct zonelist *zonelist;
1872 enum zone_type highest_zoneidx = gfp_zone(GFP_KERNEL);
1873 zonelist = &NODE_DATA(node)->node_zonelists[ZONELIST_FALLBACK];
1874 z = first_zones_zonelist(zonelist, highest_zoneidx,
1875 &policy->nodes);
1876 return z->zone ? zone_to_nid(z->zone) : node;
1877 }
1878 case MPOL_LOCAL:
1879 return node;
1880
1881 default:
1882 BUG();
1883 }
1884}
1885
1886/*
1887 * Do static interleaving for interleave index @ilx. Returns the ilx'th
1888 * node in pol->nodes (starting from ilx=0), wrapping around if ilx
1889 * exceeds the number of present nodes.
1890 */
1891static unsigned int interleave_nid(struct mempolicy *pol, pgoff_t ilx)
1892{
1893 nodemask_t nodemask = pol->nodes;
1894 unsigned int target, nnodes;
1895 int i;
1896 int nid;
1897 /*
1898 * The barrier will stabilize the nodemask in a register or on
1899 * the stack so that it will stop changing under the code.
1900 *
1901 * Between first_node() and next_node(), pol->nodes could be changed
1902 * by other threads. So we put pol->nodes in a local stack.
1903 */
1904 barrier();
1905
1906 nnodes = nodes_weight(nodemask);
1907 if (!nnodes)
1908 return numa_node_id();
1909 target = ilx % nnodes;
1910 nid = first_node(nodemask);
1911 for (i = 0; i < target; i++)
1912 nid = next_node(nid, nodemask);
1913 return nid;
1914}
1915
1916/*
1917 * Return a nodemask representing a mempolicy for filtering nodes for
1918 * page allocation, together with preferred node id (or the input node id).
1919 */
1920static nodemask_t *policy_nodemask(gfp_t gfp, struct mempolicy *pol,
1921 pgoff_t ilx, int *nid)
1922{
1923 nodemask_t *nodemask = NULL;
1924
1925 switch (pol->mode) {
1926 case MPOL_PREFERRED:
1927 /* Override input node id */
1928 *nid = first_node(pol->nodes);
1929 break;
1930 case MPOL_PREFERRED_MANY:
1931 nodemask = &pol->nodes;
1932 if (pol->home_node != NUMA_NO_NODE)
1933 *nid = pol->home_node;
1934 break;
1935 case MPOL_BIND:
1936 /* Restrict to nodemask (but not on lower zones) */
1937 if (apply_policy_zone(pol, gfp_zone(gfp)) &&
1938 cpuset_nodemask_valid_mems_allowed(&pol->nodes))
1939 nodemask = &pol->nodes;
1940 if (pol->home_node != NUMA_NO_NODE)
1941 *nid = pol->home_node;
1942 /*
1943 * __GFP_THISNODE shouldn't even be used with the bind policy
1944 * because we might easily break the expectation to stay on the
1945 * requested node and not break the policy.
1946 */
1947 WARN_ON_ONCE(gfp & __GFP_THISNODE);
1948 break;
1949 case MPOL_INTERLEAVE:
1950 /* Override input node id */
1951 *nid = (ilx == NO_INTERLEAVE_INDEX) ?
1952 interleave_nodes(pol) : interleave_nid(pol, ilx);
1953 break;
1954 }
1955
1956 return nodemask;
1957}
1958
1959#ifdef CONFIG_HUGETLBFS
1960/*
1961 * huge_node(@vma, @addr, @gfp_flags, @mpol)
1962 * @vma: virtual memory area whose policy is sought
1963 * @addr: address in @vma for shared policy lookup and interleave policy
1964 * @gfp_flags: for requested zone
1965 * @mpol: pointer to mempolicy pointer for reference counted mempolicy
1966 * @nodemask: pointer to nodemask pointer for 'bind' and 'prefer-many' policy
1967 *
1968 * Returns a nid suitable for a huge page allocation and a pointer
1969 * to the struct mempolicy for conditional unref after allocation.
1970 * If the effective policy is 'bind' or 'prefer-many', returns a pointer
1971 * to the mempolicy's @nodemask for filtering the zonelist.
1972 */
1973int huge_node(struct vm_area_struct *vma, unsigned long addr, gfp_t gfp_flags,
1974 struct mempolicy **mpol, nodemask_t **nodemask)
1975{
1976 pgoff_t ilx;
1977 int nid;
1978
1979 nid = numa_node_id();
1980 *mpol = get_vma_policy(vma, addr, hstate_vma(vma)->order, &ilx);
1981 *nodemask = policy_nodemask(gfp_flags, *mpol, ilx, &nid);
1982 return nid;
1983}
1984
1985/*
1986 * init_nodemask_of_mempolicy
1987 *
1988 * If the current task's mempolicy is "default" [NULL], return 'false'
1989 * to indicate default policy. Otherwise, extract the policy nodemask
1990 * for 'bind' or 'interleave' policy into the argument nodemask, or
1991 * initialize the argument nodemask to contain the single node for
1992 * 'preferred' or 'local' policy and return 'true' to indicate presence
1993 * of non-default mempolicy.
1994 *
1995 * We don't bother with reference counting the mempolicy [mpol_get/put]
1996 * because the current task is examining it's own mempolicy and a task's
1997 * mempolicy is only ever changed by the task itself.
1998 *
1999 * N.B., it is the caller's responsibility to free a returned nodemask.
2000 */
2001bool init_nodemask_of_mempolicy(nodemask_t *mask)
2002{
2003 struct mempolicy *mempolicy;
2004
2005 if (!(mask && current->mempolicy))
2006 return false;
2007
2008 task_lock(current);
2009 mempolicy = current->mempolicy;
2010 switch (mempolicy->mode) {
2011 case MPOL_PREFERRED:
2012 case MPOL_PREFERRED_MANY:
2013 case MPOL_BIND:
2014 case MPOL_INTERLEAVE:
2015 *mask = mempolicy->nodes;
2016 break;
2017
2018 case MPOL_LOCAL:
2019 init_nodemask_of_node(mask, numa_node_id());
2020 break;
2021
2022 default:
2023 BUG();
2024 }
2025 task_unlock(current);
2026
2027 return true;
2028}
2029#endif
2030
2031/*
2032 * mempolicy_in_oom_domain
2033 *
2034 * If tsk's mempolicy is "bind", check for intersection between mask and
2035 * the policy nodemask. Otherwise, return true for all other policies
2036 * including "interleave", as a tsk with "interleave" policy may have
2037 * memory allocated from all nodes in system.
2038 *
2039 * Takes task_lock(tsk) to prevent freeing of its mempolicy.
2040 */
2041bool mempolicy_in_oom_domain(struct task_struct *tsk,
2042 const nodemask_t *mask)
2043{
2044 struct mempolicy *mempolicy;
2045 bool ret = true;
2046
2047 if (!mask)
2048 return ret;
2049
2050 task_lock(tsk);
2051 mempolicy = tsk->mempolicy;
2052 if (mempolicy && mempolicy->mode == MPOL_BIND)
2053 ret = nodes_intersects(mempolicy->nodes, *mask);
2054 task_unlock(tsk);
2055
2056 return ret;
2057}
2058
2059static struct page *alloc_pages_preferred_many(gfp_t gfp, unsigned int order,
2060 int nid, nodemask_t *nodemask)
2061{
2062 struct page *page;
2063 gfp_t preferred_gfp;
2064
2065 /*
2066 * This is a two pass approach. The first pass will only try the
2067 * preferred nodes but skip the direct reclaim and allow the
2068 * allocation to fail, while the second pass will try all the
2069 * nodes in system.
2070 */
2071 preferred_gfp = gfp | __GFP_NOWARN;
2072 preferred_gfp &= ~(__GFP_DIRECT_RECLAIM | __GFP_NOFAIL);
2073 page = __alloc_pages(preferred_gfp, order, nid, nodemask);
2074 if (!page)
2075 page = __alloc_pages(gfp, order, nid, NULL);
2076
2077 return page;
2078}
2079
2080/**
2081 * alloc_pages_mpol - Allocate pages according to NUMA mempolicy.
2082 * @gfp: GFP flags.
2083 * @order: Order of the page allocation.
2084 * @pol: Pointer to the NUMA mempolicy.
2085 * @ilx: Index for interleave mempolicy (also distinguishes alloc_pages()).
2086 * @nid: Preferred node (usually numa_node_id() but @mpol may override it).
2087 *
2088 * Return: The page on success or NULL if allocation fails.
2089 */
2090struct page *alloc_pages_mpol(gfp_t gfp, unsigned int order,
2091 struct mempolicy *pol, pgoff_t ilx, int nid)
2092{
2093 nodemask_t *nodemask;
2094 struct page *page;
2095
2096 nodemask = policy_nodemask(gfp, pol, ilx, &nid);
2097
2098 if (pol->mode == MPOL_PREFERRED_MANY)
2099 return alloc_pages_preferred_many(gfp, order, nid, nodemask);
2100
2101 if (IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE) &&
2102 /* filter "hugepage" allocation, unless from alloc_pages() */
2103 order == HPAGE_PMD_ORDER && ilx != NO_INTERLEAVE_INDEX) {
2104 /*
2105 * For hugepage allocation and non-interleave policy which
2106 * allows the current node (or other explicitly preferred
2107 * node) we only try to allocate from the current/preferred
2108 * node and don't fall back to other nodes, as the cost of
2109 * remote accesses would likely offset THP benefits.
2110 *
2111 * If the policy is interleave or does not allow the current
2112 * node in its nodemask, we allocate the standard way.
2113 */
2114 if (pol->mode != MPOL_INTERLEAVE &&
2115 (!nodemask || node_isset(nid, *nodemask))) {
2116 /*
2117 * First, try to allocate THP only on local node, but
2118 * don't reclaim unnecessarily, just compact.
2119 */
2120 page = __alloc_pages_node(nid,
2121 gfp | __GFP_THISNODE | __GFP_NORETRY, order);
2122 if (page || !(gfp & __GFP_DIRECT_RECLAIM))
2123 return page;
2124 /*
2125 * If hugepage allocations are configured to always
2126 * synchronous compact or the vma has been madvised
2127 * to prefer hugepage backing, retry allowing remote
2128 * memory with both reclaim and compact as well.
2129 */
2130 }
2131 }
2132
2133 page = __alloc_pages(gfp, order, nid, nodemask);
2134
2135 if (unlikely(pol->mode == MPOL_INTERLEAVE) && page) {
2136 /* skip NUMA_INTERLEAVE_HIT update if numa stats is disabled */
2137 if (static_branch_likely(&vm_numa_stat_key) &&
2138 page_to_nid(page) == nid) {
2139 preempt_disable();
2140 __count_numa_event(page_zone(page), NUMA_INTERLEAVE_HIT);
2141 preempt_enable();
2142 }
2143 }
2144
2145 return page;
2146}
2147
2148/**
2149 * vma_alloc_folio - Allocate a folio for a VMA.
2150 * @gfp: GFP flags.
2151 * @order: Order of the folio.
2152 * @vma: Pointer to VMA.
2153 * @addr: Virtual address of the allocation. Must be inside @vma.
2154 * @hugepage: Unused (was: For hugepages try only preferred node if possible).
2155 *
2156 * Allocate a folio for a specific address in @vma, using the appropriate
2157 * NUMA policy. The caller must hold the mmap_lock of the mm_struct of the
2158 * VMA to prevent it from going away. Should be used for all allocations
2159 * for folios that will be mapped into user space, excepting hugetlbfs, and
2160 * excepting where direct use of alloc_pages_mpol() is more appropriate.
2161 *
2162 * Return: The folio on success or NULL if allocation fails.
2163 */
2164struct folio *vma_alloc_folio(gfp_t gfp, int order, struct vm_area_struct *vma,
2165 unsigned long addr, bool hugepage)
2166{
2167 struct mempolicy *pol;
2168 pgoff_t ilx;
2169 struct page *page;
2170
2171 pol = get_vma_policy(vma, addr, order, &ilx);
2172 page = alloc_pages_mpol(gfp | __GFP_COMP, order,
2173 pol, ilx, numa_node_id());
2174 mpol_cond_put(pol);
2175 return page_rmappable_folio(page);
2176}
2177EXPORT_SYMBOL(vma_alloc_folio);
2178
2179/**
2180 * alloc_pages - Allocate pages.
2181 * @gfp: GFP flags.
2182 * @order: Power of two of number of pages to allocate.
2183 *
2184 * Allocate 1 << @order contiguous pages. The physical address of the
2185 * first page is naturally aligned (eg an order-3 allocation will be aligned
2186 * to a multiple of 8 * PAGE_SIZE bytes). The NUMA policy of the current
2187 * process is honoured when in process context.
2188 *
2189 * Context: Can be called from any context, providing the appropriate GFP
2190 * flags are used.
2191 * Return: The page on success or NULL if allocation fails.
2192 */
2193struct page *alloc_pages(gfp_t gfp, unsigned int order)
2194{
2195 struct mempolicy *pol = &default_policy;
2196
2197 /*
2198 * No reference counting needed for current->mempolicy
2199 * nor system default_policy
2200 */
2201 if (!in_interrupt() && !(gfp & __GFP_THISNODE))
2202 pol = get_task_policy(current);
2203
2204 return alloc_pages_mpol(gfp, order,
2205 pol, NO_INTERLEAVE_INDEX, numa_node_id());
2206}
2207EXPORT_SYMBOL(alloc_pages);
2208
2209struct folio *folio_alloc(gfp_t gfp, unsigned int order)
2210{
2211 return page_rmappable_folio(alloc_pages(gfp | __GFP_COMP, order));
2212}
2213EXPORT_SYMBOL(folio_alloc);
2214
2215static unsigned long alloc_pages_bulk_array_interleave(gfp_t gfp,
2216 struct mempolicy *pol, unsigned long nr_pages,
2217 struct page **page_array)
2218{
2219 int nodes;
2220 unsigned long nr_pages_per_node;
2221 int delta;
2222 int i;
2223 unsigned long nr_allocated;
2224 unsigned long total_allocated = 0;
2225
2226 nodes = nodes_weight(pol->nodes);
2227 nr_pages_per_node = nr_pages / nodes;
2228 delta = nr_pages - nodes * nr_pages_per_node;
2229
2230 for (i = 0; i < nodes; i++) {
2231 if (delta) {
2232 nr_allocated = __alloc_pages_bulk(gfp,
2233 interleave_nodes(pol), NULL,
2234 nr_pages_per_node + 1, NULL,
2235 page_array);
2236 delta--;
2237 } else {
2238 nr_allocated = __alloc_pages_bulk(gfp,
2239 interleave_nodes(pol), NULL,
2240 nr_pages_per_node, NULL, page_array);
2241 }
2242
2243 page_array += nr_allocated;
2244 total_allocated += nr_allocated;
2245 }
2246
2247 return total_allocated;
2248}
2249
2250static unsigned long alloc_pages_bulk_array_preferred_many(gfp_t gfp, int nid,
2251 struct mempolicy *pol, unsigned long nr_pages,
2252 struct page **page_array)
2253{
2254 gfp_t preferred_gfp;
2255 unsigned long nr_allocated = 0;
2256
2257 preferred_gfp = gfp | __GFP_NOWARN;
2258 preferred_gfp &= ~(__GFP_DIRECT_RECLAIM | __GFP_NOFAIL);
2259
2260 nr_allocated = __alloc_pages_bulk(preferred_gfp, nid, &pol->nodes,
2261 nr_pages, NULL, page_array);
2262
2263 if (nr_allocated < nr_pages)
2264 nr_allocated += __alloc_pages_bulk(gfp, numa_node_id(), NULL,
2265 nr_pages - nr_allocated, NULL,
2266 page_array + nr_allocated);
2267 return nr_allocated;
2268}
2269
2270/* alloc pages bulk and mempolicy should be considered at the
2271 * same time in some situation such as vmalloc.
2272 *
2273 * It can accelerate memory allocation especially interleaving
2274 * allocate memory.
2275 */
2276unsigned long alloc_pages_bulk_array_mempolicy(gfp_t gfp,
2277 unsigned long nr_pages, struct page **page_array)
2278{
2279 struct mempolicy *pol = &default_policy;
2280 nodemask_t *nodemask;
2281 int nid;
2282
2283 if (!in_interrupt() && !(gfp & __GFP_THISNODE))
2284 pol = get_task_policy(current);
2285
2286 if (pol->mode == MPOL_INTERLEAVE)
2287 return alloc_pages_bulk_array_interleave(gfp, pol,
2288 nr_pages, page_array);
2289
2290 if (pol->mode == MPOL_PREFERRED_MANY)
2291 return alloc_pages_bulk_array_preferred_many(gfp,
2292 numa_node_id(), pol, nr_pages, page_array);
2293
2294 nid = numa_node_id();
2295 nodemask = policy_nodemask(gfp, pol, NO_INTERLEAVE_INDEX, &nid);
2296 return __alloc_pages_bulk(gfp, nid, nodemask,
2297 nr_pages, NULL, page_array);
2298}
2299
2300int vma_dup_policy(struct vm_area_struct *src, struct vm_area_struct *dst)
2301{
2302 struct mempolicy *pol = mpol_dup(src->vm_policy);
2303
2304 if (IS_ERR(pol))
2305 return PTR_ERR(pol);
2306 dst->vm_policy = pol;
2307 return 0;
2308}
2309
2310/*
2311 * If mpol_dup() sees current->cpuset == cpuset_being_rebound, then it
2312 * rebinds the mempolicy its copying by calling mpol_rebind_policy()
2313 * with the mems_allowed returned by cpuset_mems_allowed(). This
2314 * keeps mempolicies cpuset relative after its cpuset moves. See
2315 * further kernel/cpuset.c update_nodemask().
2316 *
2317 * current's mempolicy may be rebinded by the other task(the task that changes
2318 * cpuset's mems), so we needn't do rebind work for current task.
2319 */
2320
2321/* Slow path of a mempolicy duplicate */
2322struct mempolicy *__mpol_dup(struct mempolicy *old)
2323{
2324 struct mempolicy *new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
2325
2326 if (!new)
2327 return ERR_PTR(-ENOMEM);
2328
2329 /* task's mempolicy is protected by alloc_lock */
2330 if (old == current->mempolicy) {
2331 task_lock(current);
2332 *new = *old;
2333 task_unlock(current);
2334 } else
2335 *new = *old;
2336
2337 if (current_cpuset_is_being_rebound()) {
2338 nodemask_t mems = cpuset_mems_allowed(current);
2339 mpol_rebind_policy(new, &mems);
2340 }
2341 atomic_set(&new->refcnt, 1);
2342 return new;
2343}
2344
2345/* Slow path of a mempolicy comparison */
2346bool __mpol_equal(struct mempolicy *a, struct mempolicy *b)
2347{
2348 if (!a || !b)
2349 return false;
2350 if (a->mode != b->mode)
2351 return false;
2352 if (a->flags != b->flags)
2353 return false;
2354 if (a->home_node != b->home_node)
2355 return false;
2356 if (mpol_store_user_nodemask(a))
2357 if (!nodes_equal(a->w.user_nodemask, b->w.user_nodemask))
2358 return false;
2359
2360 switch (a->mode) {
2361 case MPOL_BIND:
2362 case MPOL_INTERLEAVE:
2363 case MPOL_PREFERRED:
2364 case MPOL_PREFERRED_MANY:
2365 return !!nodes_equal(a->nodes, b->nodes);
2366 case MPOL_LOCAL:
2367 return true;
2368 default:
2369 BUG();
2370 return false;
2371 }
2372}
2373
2374/*
2375 * Shared memory backing store policy support.
2376 *
2377 * Remember policies even when nobody has shared memory mapped.
2378 * The policies are kept in Red-Black tree linked from the inode.
2379 * They are protected by the sp->lock rwlock, which should be held
2380 * for any accesses to the tree.
2381 */
2382
2383/*
2384 * lookup first element intersecting start-end. Caller holds sp->lock for
2385 * reading or for writing
2386 */
2387static struct sp_node *sp_lookup(struct shared_policy *sp,
2388 pgoff_t start, pgoff_t end)
2389{
2390 struct rb_node *n = sp->root.rb_node;
2391
2392 while (n) {
2393 struct sp_node *p = rb_entry(n, struct sp_node, nd);
2394
2395 if (start >= p->end)
2396 n = n->rb_right;
2397 else if (end <= p->start)
2398 n = n->rb_left;
2399 else
2400 break;
2401 }
2402 if (!n)
2403 return NULL;
2404 for (;;) {
2405 struct sp_node *w = NULL;
2406 struct rb_node *prev = rb_prev(n);
2407 if (!prev)
2408 break;
2409 w = rb_entry(prev, struct sp_node, nd);
2410 if (w->end <= start)
2411 break;
2412 n = prev;
2413 }
2414 return rb_entry(n, struct sp_node, nd);
2415}
2416
2417/*
2418 * Insert a new shared policy into the list. Caller holds sp->lock for
2419 * writing.
2420 */
2421static void sp_insert(struct shared_policy *sp, struct sp_node *new)
2422{
2423 struct rb_node **p = &sp->root.rb_node;
2424 struct rb_node *parent = NULL;
2425 struct sp_node *nd;
2426
2427 while (*p) {
2428 parent = *p;
2429 nd = rb_entry(parent, struct sp_node, nd);
2430 if (new->start < nd->start)
2431 p = &(*p)->rb_left;
2432 else if (new->end > nd->end)
2433 p = &(*p)->rb_right;
2434 else
2435 BUG();
2436 }
2437 rb_link_node(&new->nd, parent, p);
2438 rb_insert_color(&new->nd, &sp->root);
2439}
2440
2441/* Find shared policy intersecting idx */
2442struct mempolicy *mpol_shared_policy_lookup(struct shared_policy *sp,
2443 pgoff_t idx)
2444{
2445 struct mempolicy *pol = NULL;
2446 struct sp_node *sn;
2447
2448 if (!sp->root.rb_node)
2449 return NULL;
2450 read_lock(&sp->lock);
2451 sn = sp_lookup(sp, idx, idx+1);
2452 if (sn) {
2453 mpol_get(sn->policy);
2454 pol = sn->policy;
2455 }
2456 read_unlock(&sp->lock);
2457 return pol;
2458}
2459
2460static void sp_free(struct sp_node *n)
2461{
2462 mpol_put(n->policy);
2463 kmem_cache_free(sn_cache, n);
2464}
2465
2466/**
2467 * mpol_misplaced - check whether current folio node is valid in policy
2468 *
2469 * @folio: folio to be checked
2470 * @vma: vm area where folio mapped
2471 * @addr: virtual address in @vma for shared policy lookup and interleave policy
2472 *
2473 * Lookup current policy node id for vma,addr and "compare to" folio's
2474 * node id. Policy determination "mimics" alloc_page_vma().
2475 * Called from fault path where we know the vma and faulting address.
2476 *
2477 * Return: NUMA_NO_NODE if the page is in a node that is valid for this
2478 * policy, or a suitable node ID to allocate a replacement folio from.
2479 */
2480int mpol_misplaced(struct folio *folio, struct vm_area_struct *vma,
2481 unsigned long addr)
2482{
2483 struct mempolicy *pol;
2484 pgoff_t ilx;
2485 struct zoneref *z;
2486 int curnid = folio_nid(folio);
2487 int thiscpu = raw_smp_processor_id();
2488 int thisnid = cpu_to_node(thiscpu);
2489 int polnid = NUMA_NO_NODE;
2490 int ret = NUMA_NO_NODE;
2491
2492 pol = get_vma_policy(vma, addr, folio_order(folio), &ilx);
2493 if (!(pol->flags & MPOL_F_MOF))
2494 goto out;
2495
2496 switch (pol->mode) {
2497 case MPOL_INTERLEAVE:
2498 polnid = interleave_nid(pol, ilx);
2499 break;
2500
2501 case MPOL_PREFERRED:
2502 if (node_isset(curnid, pol->nodes))
2503 goto out;
2504 polnid = first_node(pol->nodes);
2505 break;
2506
2507 case MPOL_LOCAL:
2508 polnid = numa_node_id();
2509 break;
2510
2511 case MPOL_BIND:
2512 /* Optimize placement among multiple nodes via NUMA balancing */
2513 if (pol->flags & MPOL_F_MORON) {
2514 if (node_isset(thisnid, pol->nodes))
2515 break;
2516 goto out;
2517 }
2518 fallthrough;
2519
2520 case MPOL_PREFERRED_MANY:
2521 /*
2522 * use current page if in policy nodemask,
2523 * else select nearest allowed node, if any.
2524 * If no allowed nodes, use current [!misplaced].
2525 */
2526 if (node_isset(curnid, pol->nodes))
2527 goto out;
2528 z = first_zones_zonelist(
2529 node_zonelist(numa_node_id(), GFP_HIGHUSER),
2530 gfp_zone(GFP_HIGHUSER),
2531 &pol->nodes);
2532 polnid = zone_to_nid(z->zone);
2533 break;
2534
2535 default:
2536 BUG();
2537 }
2538
2539 /* Migrate the folio towards the node whose CPU is referencing it */
2540 if (pol->flags & MPOL_F_MORON) {
2541 polnid = thisnid;
2542
2543 if (!should_numa_migrate_memory(current, folio, curnid,
2544 thiscpu))
2545 goto out;
2546 }
2547
2548 if (curnid != polnid)
2549 ret = polnid;
2550out:
2551 mpol_cond_put(pol);
2552
2553 return ret;
2554}
2555
2556/*
2557 * Drop the (possibly final) reference to task->mempolicy. It needs to be
2558 * dropped after task->mempolicy is set to NULL so that any allocation done as
2559 * part of its kmem_cache_free(), such as by KASAN, doesn't reference a freed
2560 * policy.
2561 */
2562void mpol_put_task_policy(struct task_struct *task)
2563{
2564 struct mempolicy *pol;
2565
2566 task_lock(task);
2567 pol = task->mempolicy;
2568 task->mempolicy = NULL;
2569 task_unlock(task);
2570 mpol_put(pol);
2571}
2572
2573static void sp_delete(struct shared_policy *sp, struct sp_node *n)
2574{
2575 rb_erase(&n->nd, &sp->root);
2576 sp_free(n);
2577}
2578
2579static void sp_node_init(struct sp_node *node, unsigned long start,
2580 unsigned long end, struct mempolicy *pol)
2581{
2582 node->start = start;
2583 node->end = end;
2584 node->policy = pol;
2585}
2586
2587static struct sp_node *sp_alloc(unsigned long start, unsigned long end,
2588 struct mempolicy *pol)
2589{
2590 struct sp_node *n;
2591 struct mempolicy *newpol;
2592
2593 n = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2594 if (!n)
2595 return NULL;
2596
2597 newpol = mpol_dup(pol);
2598 if (IS_ERR(newpol)) {
2599 kmem_cache_free(sn_cache, n);
2600 return NULL;
2601 }
2602 newpol->flags |= MPOL_F_SHARED;
2603 sp_node_init(n, start, end, newpol);
2604
2605 return n;
2606}
2607
2608/* Replace a policy range. */
2609static int shared_policy_replace(struct shared_policy *sp, pgoff_t start,
2610 pgoff_t end, struct sp_node *new)
2611{
2612 struct sp_node *n;
2613 struct sp_node *n_new = NULL;
2614 struct mempolicy *mpol_new = NULL;
2615 int ret = 0;
2616
2617restart:
2618 write_lock(&sp->lock);
2619 n = sp_lookup(sp, start, end);
2620 /* Take care of old policies in the same range. */
2621 while (n && n->start < end) {
2622 struct rb_node *next = rb_next(&n->nd);
2623 if (n->start >= start) {
2624 if (n->end <= end)
2625 sp_delete(sp, n);
2626 else
2627 n->start = end;
2628 } else {
2629 /* Old policy spanning whole new range. */
2630 if (n->end > end) {
2631 if (!n_new)
2632 goto alloc_new;
2633
2634 *mpol_new = *n->policy;
2635 atomic_set(&mpol_new->refcnt, 1);
2636 sp_node_init(n_new, end, n->end, mpol_new);
2637 n->end = start;
2638 sp_insert(sp, n_new);
2639 n_new = NULL;
2640 mpol_new = NULL;
2641 break;
2642 } else
2643 n->end = start;
2644 }
2645 if (!next)
2646 break;
2647 n = rb_entry(next, struct sp_node, nd);
2648 }
2649 if (new)
2650 sp_insert(sp, new);
2651 write_unlock(&sp->lock);
2652 ret = 0;
2653
2654err_out:
2655 if (mpol_new)
2656 mpol_put(mpol_new);
2657 if (n_new)
2658 kmem_cache_free(sn_cache, n_new);
2659
2660 return ret;
2661
2662alloc_new:
2663 write_unlock(&sp->lock);
2664 ret = -ENOMEM;
2665 n_new = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2666 if (!n_new)
2667 goto err_out;
2668 mpol_new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
2669 if (!mpol_new)
2670 goto err_out;
2671 atomic_set(&mpol_new->refcnt, 1);
2672 goto restart;
2673}
2674
2675/**
2676 * mpol_shared_policy_init - initialize shared policy for inode
2677 * @sp: pointer to inode shared policy
2678 * @mpol: struct mempolicy to install
2679 *
2680 * Install non-NULL @mpol in inode's shared policy rb-tree.
2681 * On entry, the current task has a reference on a non-NULL @mpol.
2682 * This must be released on exit.
2683 * This is called at get_inode() calls and we can use GFP_KERNEL.
2684 */
2685void mpol_shared_policy_init(struct shared_policy *sp, struct mempolicy *mpol)
2686{
2687 int ret;
2688
2689 sp->root = RB_ROOT; /* empty tree == default mempolicy */
2690 rwlock_init(&sp->lock);
2691
2692 if (mpol) {
2693 struct sp_node *sn;
2694 struct mempolicy *npol;
2695 NODEMASK_SCRATCH(scratch);
2696
2697 if (!scratch)
2698 goto put_mpol;
2699
2700 /* contextualize the tmpfs mount point mempolicy to this file */
2701 npol = mpol_new(mpol->mode, mpol->flags, &mpol->w.user_nodemask);
2702 if (IS_ERR(npol))
2703 goto free_scratch; /* no valid nodemask intersection */
2704
2705 task_lock(current);
2706 ret = mpol_set_nodemask(npol, &mpol->w.user_nodemask, scratch);
2707 task_unlock(current);
2708 if (ret)
2709 goto put_npol;
2710
2711 /* alloc node covering entire file; adds ref to file's npol */
2712 sn = sp_alloc(0, MAX_LFS_FILESIZE >> PAGE_SHIFT, npol);
2713 if (sn)
2714 sp_insert(sp, sn);
2715put_npol:
2716 mpol_put(npol); /* drop initial ref on file's npol */
2717free_scratch:
2718 NODEMASK_SCRATCH_FREE(scratch);
2719put_mpol:
2720 mpol_put(mpol); /* drop our incoming ref on sb mpol */
2721 }
2722}
2723
2724int mpol_set_shared_policy(struct shared_policy *sp,
2725 struct vm_area_struct *vma, struct mempolicy *pol)
2726{
2727 int err;
2728 struct sp_node *new = NULL;
2729 unsigned long sz = vma_pages(vma);
2730
2731 if (pol) {
2732 new = sp_alloc(vma->vm_pgoff, vma->vm_pgoff + sz, pol);
2733 if (!new)
2734 return -ENOMEM;
2735 }
2736 err = shared_policy_replace(sp, vma->vm_pgoff, vma->vm_pgoff + sz, new);
2737 if (err && new)
2738 sp_free(new);
2739 return err;
2740}
2741
2742/* Free a backing policy store on inode delete. */
2743void mpol_free_shared_policy(struct shared_policy *sp)
2744{
2745 struct sp_node *n;
2746 struct rb_node *next;
2747
2748 if (!sp->root.rb_node)
2749 return;
2750 write_lock(&sp->lock);
2751 next = rb_first(&sp->root);
2752 while (next) {
2753 n = rb_entry(next, struct sp_node, nd);
2754 next = rb_next(&n->nd);
2755 sp_delete(sp, n);
2756 }
2757 write_unlock(&sp->lock);
2758}
2759
2760#ifdef CONFIG_NUMA_BALANCING
2761static int __initdata numabalancing_override;
2762
2763static void __init check_numabalancing_enable(void)
2764{
2765 bool numabalancing_default = false;
2766
2767 if (IS_ENABLED(CONFIG_NUMA_BALANCING_DEFAULT_ENABLED))
2768 numabalancing_default = true;
2769
2770 /* Parsed by setup_numabalancing. override == 1 enables, -1 disables */
2771 if (numabalancing_override)
2772 set_numabalancing_state(numabalancing_override == 1);
2773
2774 if (num_online_nodes() > 1 && !numabalancing_override) {
2775 pr_info("%s automatic NUMA balancing. Configure with numa_balancing= or the kernel.numa_balancing sysctl\n",
2776 numabalancing_default ? "Enabling" : "Disabling");
2777 set_numabalancing_state(numabalancing_default);
2778 }
2779}
2780
2781static int __init setup_numabalancing(char *str)
2782{
2783 int ret = 0;
2784 if (!str)
2785 goto out;
2786
2787 if (!strcmp(str, "enable")) {
2788 numabalancing_override = 1;
2789 ret = 1;
2790 } else if (!strcmp(str, "disable")) {
2791 numabalancing_override = -1;
2792 ret = 1;
2793 }
2794out:
2795 if (!ret)
2796 pr_warn("Unable to parse numa_balancing=\n");
2797
2798 return ret;
2799}
2800__setup("numa_balancing=", setup_numabalancing);
2801#else
2802static inline void __init check_numabalancing_enable(void)
2803{
2804}
2805#endif /* CONFIG_NUMA_BALANCING */
2806
2807void __init numa_policy_init(void)
2808{
2809 nodemask_t interleave_nodes;
2810 unsigned long largest = 0;
2811 int nid, prefer = 0;
2812
2813 policy_cache = kmem_cache_create("numa_policy",
2814 sizeof(struct mempolicy),
2815 0, SLAB_PANIC, NULL);
2816
2817 sn_cache = kmem_cache_create("shared_policy_node",
2818 sizeof(struct sp_node),
2819 0, SLAB_PANIC, NULL);
2820
2821 for_each_node(nid) {
2822 preferred_node_policy[nid] = (struct mempolicy) {
2823 .refcnt = ATOMIC_INIT(1),
2824 .mode = MPOL_PREFERRED,
2825 .flags = MPOL_F_MOF | MPOL_F_MORON,
2826 .nodes = nodemask_of_node(nid),
2827 };
2828 }
2829
2830 /*
2831 * Set interleaving policy for system init. Interleaving is only
2832 * enabled across suitably sized nodes (default is >= 16MB), or
2833 * fall back to the largest node if they're all smaller.
2834 */
2835 nodes_clear(interleave_nodes);
2836 for_each_node_state(nid, N_MEMORY) {
2837 unsigned long total_pages = node_present_pages(nid);
2838
2839 /* Preserve the largest node */
2840 if (largest < total_pages) {
2841 largest = total_pages;
2842 prefer = nid;
2843 }
2844
2845 /* Interleave this node? */
2846 if ((total_pages << PAGE_SHIFT) >= (16 << 20))
2847 node_set(nid, interleave_nodes);
2848 }
2849
2850 /* All too small, use the largest */
2851 if (unlikely(nodes_empty(interleave_nodes)))
2852 node_set(prefer, interleave_nodes);
2853
2854 if (do_set_mempolicy(MPOL_INTERLEAVE, 0, &interleave_nodes))
2855 pr_err("%s: interleaving failed\n", __func__);
2856
2857 check_numabalancing_enable();
2858}
2859
2860/* Reset policy of current process to default */
2861void numa_default_policy(void)
2862{
2863 do_set_mempolicy(MPOL_DEFAULT, 0, NULL);
2864}
2865
2866/*
2867 * Parse and format mempolicy from/to strings
2868 */
2869static const char * const policy_modes[] =
2870{
2871 [MPOL_DEFAULT] = "default",
2872 [MPOL_PREFERRED] = "prefer",
2873 [MPOL_BIND] = "bind",
2874 [MPOL_INTERLEAVE] = "interleave",
2875 [MPOL_LOCAL] = "local",
2876 [MPOL_PREFERRED_MANY] = "prefer (many)",
2877};
2878
2879#ifdef CONFIG_TMPFS
2880/**
2881 * mpol_parse_str - parse string to mempolicy, for tmpfs mpol mount option.
2882 * @str: string containing mempolicy to parse
2883 * @mpol: pointer to struct mempolicy pointer, returned on success.
2884 *
2885 * Format of input:
2886 * <mode>[=<flags>][:<nodelist>]
2887 *
2888 * Return: %0 on success, else %1
2889 */
2890int mpol_parse_str(char *str, struct mempolicy **mpol)
2891{
2892 struct mempolicy *new = NULL;
2893 unsigned short mode_flags;
2894 nodemask_t nodes;
2895 char *nodelist = strchr(str, ':');
2896 char *flags = strchr(str, '=');
2897 int err = 1, mode;
2898
2899 if (flags)
2900 *flags++ = '\0'; /* terminate mode string */
2901
2902 if (nodelist) {
2903 /* NUL-terminate mode or flags string */
2904 *nodelist++ = '\0';
2905 if (nodelist_parse(nodelist, nodes))
2906 goto out;
2907 if (!nodes_subset(nodes, node_states[N_MEMORY]))
2908 goto out;
2909 } else
2910 nodes_clear(nodes);
2911
2912 mode = match_string(policy_modes, MPOL_MAX, str);
2913 if (mode < 0)
2914 goto out;
2915
2916 switch (mode) {
2917 case MPOL_PREFERRED:
2918 /*
2919 * Insist on a nodelist of one node only, although later
2920 * we use first_node(nodes) to grab a single node, so here
2921 * nodelist (or nodes) cannot be empty.
2922 */
2923 if (nodelist) {
2924 char *rest = nodelist;
2925 while (isdigit(*rest))
2926 rest++;
2927 if (*rest)
2928 goto out;
2929 if (nodes_empty(nodes))
2930 goto out;
2931 }
2932 break;
2933 case MPOL_INTERLEAVE:
2934 /*
2935 * Default to online nodes with memory if no nodelist
2936 */
2937 if (!nodelist)
2938 nodes = node_states[N_MEMORY];
2939 break;
2940 case MPOL_LOCAL:
2941 /*
2942 * Don't allow a nodelist; mpol_new() checks flags
2943 */
2944 if (nodelist)
2945 goto out;
2946 break;
2947 case MPOL_DEFAULT:
2948 /*
2949 * Insist on a empty nodelist
2950 */
2951 if (!nodelist)
2952 err = 0;
2953 goto out;
2954 case MPOL_PREFERRED_MANY:
2955 case MPOL_BIND:
2956 /*
2957 * Insist on a nodelist
2958 */
2959 if (!nodelist)
2960 goto out;
2961 }
2962
2963 mode_flags = 0;
2964 if (flags) {
2965 /*
2966 * Currently, we only support two mutually exclusive
2967 * mode flags.
2968 */
2969 if (!strcmp(flags, "static"))
2970 mode_flags |= MPOL_F_STATIC_NODES;
2971 else if (!strcmp(flags, "relative"))
2972 mode_flags |= MPOL_F_RELATIVE_NODES;
2973 else
2974 goto out;
2975 }
2976
2977 new = mpol_new(mode, mode_flags, &nodes);
2978 if (IS_ERR(new))
2979 goto out;
2980
2981 /*
2982 * Save nodes for mpol_to_str() to show the tmpfs mount options
2983 * for /proc/mounts, /proc/pid/mounts and /proc/pid/mountinfo.
2984 */
2985 if (mode != MPOL_PREFERRED) {
2986 new->nodes = nodes;
2987 } else if (nodelist) {
2988 nodes_clear(new->nodes);
2989 node_set(first_node(nodes), new->nodes);
2990 } else {
2991 new->mode = MPOL_LOCAL;
2992 }
2993
2994 /*
2995 * Save nodes for contextualization: this will be used to "clone"
2996 * the mempolicy in a specific context [cpuset] at a later time.
2997 */
2998 new->w.user_nodemask = nodes;
2999
3000 err = 0;
3001
3002out:
3003 /* Restore string for error message */
3004 if (nodelist)
3005 *--nodelist = ':';
3006 if (flags)
3007 *--flags = '=';
3008 if (!err)
3009 *mpol = new;
3010 return err;
3011}
3012#endif /* CONFIG_TMPFS */
3013
3014/**
3015 * mpol_to_str - format a mempolicy structure for printing
3016 * @buffer: to contain formatted mempolicy string
3017 * @maxlen: length of @buffer
3018 * @pol: pointer to mempolicy to be formatted
3019 *
3020 * Convert @pol into a string. If @buffer is too short, truncate the string.
3021 * Recommend a @maxlen of at least 32 for the longest mode, "interleave", the
3022 * longest flag, "relative", and to display at least a few node ids.
3023 */
3024void mpol_to_str(char *buffer, int maxlen, struct mempolicy *pol)
3025{
3026 char *p = buffer;
3027 nodemask_t nodes = NODE_MASK_NONE;
3028 unsigned short mode = MPOL_DEFAULT;
3029 unsigned short flags = 0;
3030
3031 if (pol && pol != &default_policy && !(pol->flags & MPOL_F_MORON)) {
3032 mode = pol->mode;
3033 flags = pol->flags;
3034 }
3035
3036 switch (mode) {
3037 case MPOL_DEFAULT:
3038 case MPOL_LOCAL:
3039 break;
3040 case MPOL_PREFERRED:
3041 case MPOL_PREFERRED_MANY:
3042 case MPOL_BIND:
3043 case MPOL_INTERLEAVE:
3044 nodes = pol->nodes;
3045 break;
3046 default:
3047 WARN_ON_ONCE(1);
3048 snprintf(p, maxlen, "unknown");
3049 return;
3050 }
3051
3052 p += snprintf(p, maxlen, "%s", policy_modes[mode]);
3053
3054 if (flags & MPOL_MODE_FLAGS) {
3055 p += snprintf(p, buffer + maxlen - p, "=");
3056
3057 /*
3058 * Currently, the only defined flags are mutually exclusive
3059 */
3060 if (flags & MPOL_F_STATIC_NODES)
3061 p += snprintf(p, buffer + maxlen - p, "static");
3062 else if (flags & MPOL_F_RELATIVE_NODES)
3063 p += snprintf(p, buffer + maxlen - p, "relative");
3064 }
3065
3066 if (!nodes_empty(nodes))
3067 p += scnprintf(p, buffer + maxlen - p, ":%*pbl",
3068 nodemask_pr_args(&nodes));
3069}