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 shmfs/tmpfs/hugetlbfs 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
116static struct kmem_cache *policy_cache;
117static struct kmem_cache *sn_cache;
118
119/* Highest zone. An specific allocation for a zone below that is not
120 policied. */
121enum zone_type policy_zone = 0;
122
123/*
124 * run-time system-wide default policy => local allocation
125 */
126static struct mempolicy default_policy = {
127 .refcnt = ATOMIC_INIT(1), /* never free it */
128 .mode = MPOL_LOCAL,
129};
130
131static struct mempolicy preferred_node_policy[MAX_NUMNODES];
132
133/**
134 * numa_map_to_online_node - Find closest online node
135 * @node: Node id to start the search
136 *
137 * Lookup the next closest node by distance if @nid is not online.
138 *
139 * Return: this @node if it is online, otherwise the closest node by distance
140 */
141int numa_map_to_online_node(int node)
142{
143 int min_dist = INT_MAX, dist, n, min_node;
144
145 if (node == NUMA_NO_NODE || node_online(node))
146 return node;
147
148 min_node = node;
149 for_each_online_node(n) {
150 dist = node_distance(node, n);
151 if (dist < min_dist) {
152 min_dist = dist;
153 min_node = n;
154 }
155 }
156
157 return min_node;
158}
159EXPORT_SYMBOL_GPL(numa_map_to_online_node);
160
161struct mempolicy *get_task_policy(struct task_struct *p)
162{
163 struct mempolicy *pol = p->mempolicy;
164 int node;
165
166 if (pol)
167 return pol;
168
169 node = numa_node_id();
170 if (node != NUMA_NO_NODE) {
171 pol = &preferred_node_policy[node];
172 /* preferred_node_policy is not initialised early in boot */
173 if (pol->mode)
174 return pol;
175 }
176
177 return &default_policy;
178}
179
180static const struct mempolicy_operations {
181 int (*create)(struct mempolicy *pol, const nodemask_t *nodes);
182 void (*rebind)(struct mempolicy *pol, const nodemask_t *nodes);
183} mpol_ops[MPOL_MAX];
184
185static inline int mpol_store_user_nodemask(const struct mempolicy *pol)
186{
187 return pol->flags & MPOL_MODE_FLAGS;
188}
189
190static void mpol_relative_nodemask(nodemask_t *ret, const nodemask_t *orig,
191 const nodemask_t *rel)
192{
193 nodemask_t tmp;
194 nodes_fold(tmp, *orig, nodes_weight(*rel));
195 nodes_onto(*ret, tmp, *rel);
196}
197
198static int mpol_new_nodemask(struct mempolicy *pol, const nodemask_t *nodes)
199{
200 if (nodes_empty(*nodes))
201 return -EINVAL;
202 pol->nodes = *nodes;
203 return 0;
204}
205
206static int mpol_new_preferred(struct mempolicy *pol, const nodemask_t *nodes)
207{
208 if (nodes_empty(*nodes))
209 return -EINVAL;
210
211 nodes_clear(pol->nodes);
212 node_set(first_node(*nodes), pol->nodes);
213 return 0;
214}
215
216/*
217 * mpol_set_nodemask is called after mpol_new() to set up the nodemask, if
218 * any, for the new policy. mpol_new() has already validated the nodes
219 * parameter with respect to the policy mode and flags.
220 *
221 * Must be called holding task's alloc_lock to protect task's mems_allowed
222 * and mempolicy. May also be called holding the mmap_lock for write.
223 */
224static int mpol_set_nodemask(struct mempolicy *pol,
225 const nodemask_t *nodes, struct nodemask_scratch *nsc)
226{
227 int ret;
228
229 /*
230 * Default (pol==NULL) resp. local memory policies are not a
231 * subject of any remapping. They also do not need any special
232 * constructor.
233 */
234 if (!pol || pol->mode == MPOL_LOCAL)
235 return 0;
236
237 /* Check N_MEMORY */
238 nodes_and(nsc->mask1,
239 cpuset_current_mems_allowed, node_states[N_MEMORY]);
240
241 VM_BUG_ON(!nodes);
242
243 if (pol->flags & MPOL_F_RELATIVE_NODES)
244 mpol_relative_nodemask(&nsc->mask2, nodes, &nsc->mask1);
245 else
246 nodes_and(nsc->mask2, *nodes, nsc->mask1);
247
248 if (mpol_store_user_nodemask(pol))
249 pol->w.user_nodemask = *nodes;
250 else
251 pol->w.cpuset_mems_allowed = cpuset_current_mems_allowed;
252
253 ret = mpol_ops[pol->mode].create(pol, &nsc->mask2);
254 return ret;
255}
256
257/*
258 * This function just creates a new policy, does some check and simple
259 * initialization. You must invoke mpol_set_nodemask() to set nodes.
260 */
261static struct mempolicy *mpol_new(unsigned short mode, unsigned short flags,
262 nodemask_t *nodes)
263{
264 struct mempolicy *policy;
265
266 pr_debug("setting mode %d flags %d nodes[0] %lx\n",
267 mode, flags, nodes ? nodes_addr(*nodes)[0] : NUMA_NO_NODE);
268
269 if (mode == MPOL_DEFAULT) {
270 if (nodes && !nodes_empty(*nodes))
271 return ERR_PTR(-EINVAL);
272 return NULL;
273 }
274 VM_BUG_ON(!nodes);
275
276 /*
277 * MPOL_PREFERRED cannot be used with MPOL_F_STATIC_NODES or
278 * MPOL_F_RELATIVE_NODES if the nodemask is empty (local allocation).
279 * All other modes require a valid pointer to a non-empty nodemask.
280 */
281 if (mode == MPOL_PREFERRED) {
282 if (nodes_empty(*nodes)) {
283 if (((flags & MPOL_F_STATIC_NODES) ||
284 (flags & MPOL_F_RELATIVE_NODES)))
285 return ERR_PTR(-EINVAL);
286
287 mode = MPOL_LOCAL;
288 }
289 } else if (mode == MPOL_LOCAL) {
290 if (!nodes_empty(*nodes) ||
291 (flags & MPOL_F_STATIC_NODES) ||
292 (flags & MPOL_F_RELATIVE_NODES))
293 return ERR_PTR(-EINVAL);
294 } else if (nodes_empty(*nodes))
295 return ERR_PTR(-EINVAL);
296 policy = kmem_cache_alloc(policy_cache, GFP_KERNEL);
297 if (!policy)
298 return ERR_PTR(-ENOMEM);
299 atomic_set(&policy->refcnt, 1);
300 policy->mode = mode;
301 policy->flags = flags;
302 policy->home_node = NUMA_NO_NODE;
303
304 return policy;
305}
306
307/* Slow path of a mpol destructor. */
308void __mpol_put(struct mempolicy *p)
309{
310 if (!atomic_dec_and_test(&p->refcnt))
311 return;
312 kmem_cache_free(policy_cache, p);
313}
314
315static void mpol_rebind_default(struct mempolicy *pol, const nodemask_t *nodes)
316{
317}
318
319static void mpol_rebind_nodemask(struct mempolicy *pol, const nodemask_t *nodes)
320{
321 nodemask_t tmp;
322
323 if (pol->flags & MPOL_F_STATIC_NODES)
324 nodes_and(tmp, pol->w.user_nodemask, *nodes);
325 else if (pol->flags & MPOL_F_RELATIVE_NODES)
326 mpol_relative_nodemask(&tmp, &pol->w.user_nodemask, nodes);
327 else {
328 nodes_remap(tmp, pol->nodes, pol->w.cpuset_mems_allowed,
329 *nodes);
330 pol->w.cpuset_mems_allowed = *nodes;
331 }
332
333 if (nodes_empty(tmp))
334 tmp = *nodes;
335
336 pol->nodes = tmp;
337}
338
339static void mpol_rebind_preferred(struct mempolicy *pol,
340 const nodemask_t *nodes)
341{
342 pol->w.cpuset_mems_allowed = *nodes;
343}
344
345/*
346 * mpol_rebind_policy - Migrate a policy to a different set of nodes
347 *
348 * Per-vma policies are protected by mmap_lock. Allocations using per-task
349 * policies are protected by task->mems_allowed_seq to prevent a premature
350 * OOM/allocation failure due to parallel nodemask modification.
351 */
352static void mpol_rebind_policy(struct mempolicy *pol, const nodemask_t *newmask)
353{
354 if (!pol || pol->mode == MPOL_LOCAL)
355 return;
356 if (!mpol_store_user_nodemask(pol) &&
357 nodes_equal(pol->w.cpuset_mems_allowed, *newmask))
358 return;
359
360 mpol_ops[pol->mode].rebind(pol, newmask);
361}
362
363/*
364 * Wrapper for mpol_rebind_policy() that just requires task
365 * pointer, and updates task mempolicy.
366 *
367 * Called with task's alloc_lock held.
368 */
369
370void mpol_rebind_task(struct task_struct *tsk, const nodemask_t *new)
371{
372 mpol_rebind_policy(tsk->mempolicy, new);
373}
374
375/*
376 * Rebind each vma in mm to new nodemask.
377 *
378 * Call holding a reference to mm. Takes mm->mmap_lock during call.
379 */
380
381void mpol_rebind_mm(struct mm_struct *mm, nodemask_t *new)
382{
383 struct vm_area_struct *vma;
384 VMA_ITERATOR(vmi, mm, 0);
385
386 mmap_write_lock(mm);
387 for_each_vma(vmi, vma)
388 mpol_rebind_policy(vma->vm_policy, new);
389 mmap_write_unlock(mm);
390}
391
392static const struct mempolicy_operations mpol_ops[MPOL_MAX] = {
393 [MPOL_DEFAULT] = {
394 .rebind = mpol_rebind_default,
395 },
396 [MPOL_INTERLEAVE] = {
397 .create = mpol_new_nodemask,
398 .rebind = mpol_rebind_nodemask,
399 },
400 [MPOL_PREFERRED] = {
401 .create = mpol_new_preferred,
402 .rebind = mpol_rebind_preferred,
403 },
404 [MPOL_BIND] = {
405 .create = mpol_new_nodemask,
406 .rebind = mpol_rebind_nodemask,
407 },
408 [MPOL_LOCAL] = {
409 .rebind = mpol_rebind_default,
410 },
411 [MPOL_PREFERRED_MANY] = {
412 .create = mpol_new_nodemask,
413 .rebind = mpol_rebind_preferred,
414 },
415};
416
417static int migrate_page_add(struct page *page, struct list_head *pagelist,
418 unsigned long flags);
419
420struct queue_pages {
421 struct list_head *pagelist;
422 unsigned long flags;
423 nodemask_t *nmask;
424 unsigned long start;
425 unsigned long end;
426 struct vm_area_struct *first;
427};
428
429/*
430 * Check if the page's nid is in qp->nmask.
431 *
432 * If MPOL_MF_INVERT is set in qp->flags, check if the nid is
433 * in the invert of qp->nmask.
434 */
435static inline bool queue_pages_required(struct page *page,
436 struct queue_pages *qp)
437{
438 int nid = page_to_nid(page);
439 unsigned long flags = qp->flags;
440
441 return node_isset(nid, *qp->nmask) == !(flags & MPOL_MF_INVERT);
442}
443
444/*
445 * queue_pages_pmd() has three possible return values:
446 * 0 - pages are placed on the right node or queued successfully, or
447 * special page is met, i.e. huge zero page.
448 * 1 - there is unmovable page, and MPOL_MF_MOVE* & MPOL_MF_STRICT were
449 * specified.
450 * -EIO - is migration entry or only MPOL_MF_STRICT was specified and an
451 * existing page was already on a node that does not follow the
452 * policy.
453 */
454static int queue_pages_pmd(pmd_t *pmd, spinlock_t *ptl, unsigned long addr,
455 unsigned long end, struct mm_walk *walk)
456 __releases(ptl)
457{
458 int ret = 0;
459 struct page *page;
460 struct queue_pages *qp = walk->private;
461 unsigned long flags;
462
463 if (unlikely(is_pmd_migration_entry(*pmd))) {
464 ret = -EIO;
465 goto unlock;
466 }
467 page = pmd_page(*pmd);
468 if (is_huge_zero_page(page)) {
469 walk->action = ACTION_CONTINUE;
470 goto unlock;
471 }
472 if (!queue_pages_required(page, qp))
473 goto unlock;
474
475 flags = qp->flags;
476 /* go to thp migration */
477 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) {
478 if (!vma_migratable(walk->vma) ||
479 migrate_page_add(page, qp->pagelist, flags)) {
480 ret = 1;
481 goto unlock;
482 }
483 } else
484 ret = -EIO;
485unlock:
486 spin_unlock(ptl);
487 return ret;
488}
489
490/*
491 * Scan through pages checking if pages follow certain conditions,
492 * and move them to the pagelist if they do.
493 *
494 * queue_pages_pte_range() has three possible return values:
495 * 0 - pages are placed on the right node or queued successfully, or
496 * special page is met, i.e. zero page.
497 * 1 - there is unmovable page, and MPOL_MF_MOVE* & MPOL_MF_STRICT were
498 * specified.
499 * -EIO - only MPOL_MF_STRICT was specified and an existing page was already
500 * on a node that does not follow the policy.
501 */
502static int queue_pages_pte_range(pmd_t *pmd, unsigned long addr,
503 unsigned long end, struct mm_walk *walk)
504{
505 struct vm_area_struct *vma = walk->vma;
506 struct page *page;
507 struct queue_pages *qp = walk->private;
508 unsigned long flags = qp->flags;
509 bool has_unmovable = false;
510 pte_t *pte, *mapped_pte;
511 spinlock_t *ptl;
512
513 ptl = pmd_trans_huge_lock(pmd, vma);
514 if (ptl)
515 return queue_pages_pmd(pmd, ptl, addr, end, walk);
516
517 if (pmd_trans_unstable(pmd))
518 return 0;
519
520 mapped_pte = pte = pte_offset_map_lock(walk->mm, pmd, addr, &ptl);
521 for (; addr != end; pte++, addr += PAGE_SIZE) {
522 if (!pte_present(*pte))
523 continue;
524 page = vm_normal_page(vma, addr, *pte);
525 if (!page || is_zone_device_page(page))
526 continue;
527 /*
528 * vm_normal_page() filters out zero pages, but there might
529 * still be PageReserved pages to skip, perhaps in a VDSO.
530 */
531 if (PageReserved(page))
532 continue;
533 if (!queue_pages_required(page, qp))
534 continue;
535 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) {
536 /* MPOL_MF_STRICT must be specified if we get here */
537 if (!vma_migratable(vma)) {
538 has_unmovable = true;
539 break;
540 }
541
542 /*
543 * Do not abort immediately since there may be
544 * temporary off LRU pages in the range. Still
545 * need migrate other LRU pages.
546 */
547 if (migrate_page_add(page, qp->pagelist, flags))
548 has_unmovable = true;
549 } else
550 break;
551 }
552 pte_unmap_unlock(mapped_pte, ptl);
553 cond_resched();
554
555 if (has_unmovable)
556 return 1;
557
558 return addr != end ? -EIO : 0;
559}
560
561static int queue_pages_hugetlb(pte_t *pte, unsigned long hmask,
562 unsigned long addr, unsigned long end,
563 struct mm_walk *walk)
564{
565 int ret = 0;
566#ifdef CONFIG_HUGETLB_PAGE
567 struct queue_pages *qp = walk->private;
568 unsigned long flags = (qp->flags & MPOL_MF_VALID);
569 struct page *page;
570 spinlock_t *ptl;
571 pte_t entry;
572
573 ptl = huge_pte_lock(hstate_vma(walk->vma), walk->mm, pte);
574 entry = huge_ptep_get(pte);
575 if (!pte_present(entry))
576 goto unlock;
577 page = pte_page(entry);
578 if (!queue_pages_required(page, qp))
579 goto unlock;
580
581 if (flags == MPOL_MF_STRICT) {
582 /*
583 * STRICT alone means only detecting misplaced page and no
584 * need to further check other vma.
585 */
586 ret = -EIO;
587 goto unlock;
588 }
589
590 if (!vma_migratable(walk->vma)) {
591 /*
592 * Must be STRICT with MOVE*, otherwise .test_walk() have
593 * stopped walking current vma.
594 * Detecting misplaced page but allow migrating pages which
595 * have been queued.
596 */
597 ret = 1;
598 goto unlock;
599 }
600
601 /* With MPOL_MF_MOVE, we migrate only unshared hugepage. */
602 if (flags & (MPOL_MF_MOVE_ALL) ||
603 (flags & MPOL_MF_MOVE && page_mapcount(page) == 1 &&
604 !hugetlb_pmd_shared(pte))) {
605 if (isolate_hugetlb(page, qp->pagelist) &&
606 (flags & MPOL_MF_STRICT))
607 /*
608 * Failed to isolate page but allow migrating pages
609 * which have been queued.
610 */
611 ret = 1;
612 }
613unlock:
614 spin_unlock(ptl);
615#else
616 BUG();
617#endif
618 return ret;
619}
620
621#ifdef CONFIG_NUMA_BALANCING
622/*
623 * This is used to mark a range of virtual addresses to be inaccessible.
624 * These are later cleared by a NUMA hinting fault. Depending on these
625 * faults, pages may be migrated for better NUMA placement.
626 *
627 * This is assuming that NUMA faults are handled using PROT_NONE. If
628 * an architecture makes a different choice, it will need further
629 * changes to the core.
630 */
631unsigned long change_prot_numa(struct vm_area_struct *vma,
632 unsigned long addr, unsigned long end)
633{
634 struct mmu_gather tlb;
635 int nr_updated;
636
637 tlb_gather_mmu(&tlb, vma->vm_mm);
638
639 nr_updated = change_protection(&tlb, vma, addr, end, PAGE_NONE,
640 MM_CP_PROT_NUMA);
641 if (nr_updated)
642 count_vm_numa_events(NUMA_PTE_UPDATES, nr_updated);
643
644 tlb_finish_mmu(&tlb);
645
646 return nr_updated;
647}
648#else
649static unsigned long change_prot_numa(struct vm_area_struct *vma,
650 unsigned long addr, unsigned long end)
651{
652 return 0;
653}
654#endif /* CONFIG_NUMA_BALANCING */
655
656static int queue_pages_test_walk(unsigned long start, unsigned long end,
657 struct mm_walk *walk)
658{
659 struct vm_area_struct *next, *vma = walk->vma;
660 struct queue_pages *qp = walk->private;
661 unsigned long endvma = vma->vm_end;
662 unsigned long flags = qp->flags;
663
664 /* range check first */
665 VM_BUG_ON_VMA(!range_in_vma(vma, start, end), vma);
666
667 if (!qp->first) {
668 qp->first = vma;
669 if (!(flags & MPOL_MF_DISCONTIG_OK) &&
670 (qp->start < vma->vm_start))
671 /* hole at head side of range */
672 return -EFAULT;
673 }
674 next = find_vma(vma->vm_mm, vma->vm_end);
675 if (!(flags & MPOL_MF_DISCONTIG_OK) &&
676 ((vma->vm_end < qp->end) &&
677 (!next || vma->vm_end < next->vm_start)))
678 /* hole at middle or tail of range */
679 return -EFAULT;
680
681 /*
682 * Need check MPOL_MF_STRICT to return -EIO if possible
683 * regardless of vma_migratable
684 */
685 if (!vma_migratable(vma) &&
686 !(flags & MPOL_MF_STRICT))
687 return 1;
688
689 if (endvma > end)
690 endvma = end;
691
692 if (flags & MPOL_MF_LAZY) {
693 /* Similar to task_numa_work, skip inaccessible VMAs */
694 if (!is_vm_hugetlb_page(vma) && vma_is_accessible(vma) &&
695 !(vma->vm_flags & VM_MIXEDMAP))
696 change_prot_numa(vma, start, endvma);
697 return 1;
698 }
699
700 /* queue pages from current vma */
701 if (flags & MPOL_MF_VALID)
702 return 0;
703 return 1;
704}
705
706static const struct mm_walk_ops queue_pages_walk_ops = {
707 .hugetlb_entry = queue_pages_hugetlb,
708 .pmd_entry = queue_pages_pte_range,
709 .test_walk = queue_pages_test_walk,
710};
711
712/*
713 * Walk through page tables and collect pages to be migrated.
714 *
715 * If pages found in a given range are on a set of nodes (determined by
716 * @nodes and @flags,) it's isolated and queued to the pagelist which is
717 * passed via @private.
718 *
719 * queue_pages_range() has three possible return values:
720 * 1 - there is unmovable page, but MPOL_MF_MOVE* & MPOL_MF_STRICT were
721 * specified.
722 * 0 - queue pages successfully or no misplaced page.
723 * errno - i.e. misplaced pages with MPOL_MF_STRICT specified (-EIO) or
724 * memory range specified by nodemask and maxnode points outside
725 * your accessible address space (-EFAULT)
726 */
727static int
728queue_pages_range(struct mm_struct *mm, unsigned long start, unsigned long end,
729 nodemask_t *nodes, unsigned long flags,
730 struct list_head *pagelist)
731{
732 int err;
733 struct queue_pages qp = {
734 .pagelist = pagelist,
735 .flags = flags,
736 .nmask = nodes,
737 .start = start,
738 .end = end,
739 .first = NULL,
740 };
741
742 err = walk_page_range(mm, start, end, &queue_pages_walk_ops, &qp);
743
744 if (!qp.first)
745 /* whole range in hole */
746 err = -EFAULT;
747
748 return err;
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 pr_debug("vma %lx-%lx/%lx vm_ops %p vm_file %p set_policy %p\n",
763 vma->vm_start, vma->vm_end, vma->vm_pgoff,
764 vma->vm_ops, vma->vm_file,
765 vma->vm_ops ? vma->vm_ops->set_policy : NULL);
766
767 new = mpol_dup(pol);
768 if (IS_ERR(new))
769 return PTR_ERR(new);
770
771 if (vma->vm_ops && vma->vm_ops->set_policy) {
772 err = vma->vm_ops->set_policy(vma, new);
773 if (err)
774 goto err_out;
775 }
776
777 old = vma->vm_policy;
778 vma->vm_policy = new; /* protected by mmap_lock */
779 mpol_put(old);
780
781 return 0;
782 err_out:
783 mpol_put(new);
784 return err;
785}
786
787/* Step 2: apply policy to a range and do splits. */
788static int mbind_range(struct mm_struct *mm, unsigned long start,
789 unsigned long end, struct mempolicy *new_pol)
790{
791 MA_STATE(mas, &mm->mm_mt, start, start);
792 struct vm_area_struct *prev;
793 struct vm_area_struct *vma;
794 int err = 0;
795 pgoff_t pgoff;
796
797 prev = mas_prev(&mas, 0);
798 if (unlikely(!prev))
799 mas_set(&mas, start);
800
801 vma = mas_find(&mas, end - 1);
802 if (WARN_ON(!vma))
803 return 0;
804
805 if (start > vma->vm_start)
806 prev = vma;
807
808 for (; vma; vma = mas_next(&mas, end - 1)) {
809 unsigned long vmstart = max(start, vma->vm_start);
810 unsigned long vmend = min(end, vma->vm_end);
811
812 if (mpol_equal(vma_policy(vma), new_pol))
813 goto next;
814
815 pgoff = vma->vm_pgoff +
816 ((vmstart - vma->vm_start) >> PAGE_SHIFT);
817 prev = vma_merge(mm, prev, vmstart, vmend, vma->vm_flags,
818 vma->anon_vma, vma->vm_file, pgoff,
819 new_pol, vma->vm_userfaultfd_ctx,
820 anon_vma_name(vma));
821 if (prev) {
822 /* vma_merge() invalidated the mas */
823 mas_pause(&mas);
824 vma = prev;
825 goto replace;
826 }
827 if (vma->vm_start != vmstart) {
828 err = split_vma(vma->vm_mm, vma, vmstart, 1);
829 if (err)
830 goto out;
831 /* split_vma() invalidated the mas */
832 mas_pause(&mas);
833 }
834 if (vma->vm_end != vmend) {
835 err = split_vma(vma->vm_mm, vma, vmend, 0);
836 if (err)
837 goto out;
838 /* split_vma() invalidated the mas */
839 mas_pause(&mas);
840 }
841replace:
842 err = vma_replace_policy(vma, new_pol);
843 if (err)
844 goto out;
845next:
846 prev = vma;
847 }
848
849out:
850 return err;
851}
852
853/* Set the process memory policy */
854static long do_set_mempolicy(unsigned short mode, unsigned short flags,
855 nodemask_t *nodes)
856{
857 struct mempolicy *new, *old;
858 NODEMASK_SCRATCH(scratch);
859 int ret;
860
861 if (!scratch)
862 return -ENOMEM;
863
864 new = mpol_new(mode, flags, nodes);
865 if (IS_ERR(new)) {
866 ret = PTR_ERR(new);
867 goto out;
868 }
869
870 task_lock(current);
871 ret = mpol_set_nodemask(new, nodes, scratch);
872 if (ret) {
873 task_unlock(current);
874 mpol_put(new);
875 goto out;
876 }
877
878 old = current->mempolicy;
879 current->mempolicy = new;
880 if (new && new->mode == MPOL_INTERLEAVE)
881 current->il_prev = MAX_NUMNODES-1;
882 task_unlock(current);
883 mpol_put(old);
884 ret = 0;
885out:
886 NODEMASK_SCRATCH_FREE(scratch);
887 return ret;
888}
889
890/*
891 * Return nodemask for policy for get_mempolicy() query
892 *
893 * Called with task's alloc_lock held
894 */
895static void get_policy_nodemask(struct mempolicy *p, nodemask_t *nodes)
896{
897 nodes_clear(*nodes);
898 if (p == &default_policy)
899 return;
900
901 switch (p->mode) {
902 case MPOL_BIND:
903 case MPOL_INTERLEAVE:
904 case MPOL_PREFERRED:
905 case MPOL_PREFERRED_MANY:
906 *nodes = p->nodes;
907 break;
908 case MPOL_LOCAL:
909 /* return empty node mask for local allocation */
910 break;
911 default:
912 BUG();
913 }
914}
915
916static int lookup_node(struct mm_struct *mm, unsigned long addr)
917{
918 struct page *p = NULL;
919 int ret;
920
921 ret = get_user_pages_fast(addr & PAGE_MASK, 1, 0, &p);
922 if (ret > 0) {
923 ret = page_to_nid(p);
924 put_page(p);
925 }
926 return ret;
927}
928
929/* Retrieve NUMA policy */
930static long do_get_mempolicy(int *policy, nodemask_t *nmask,
931 unsigned long addr, unsigned long flags)
932{
933 int err;
934 struct mm_struct *mm = current->mm;
935 struct vm_area_struct *vma = NULL;
936 struct mempolicy *pol = current->mempolicy, *pol_refcount = NULL;
937
938 if (flags &
939 ~(unsigned long)(MPOL_F_NODE|MPOL_F_ADDR|MPOL_F_MEMS_ALLOWED))
940 return -EINVAL;
941
942 if (flags & MPOL_F_MEMS_ALLOWED) {
943 if (flags & (MPOL_F_NODE|MPOL_F_ADDR))
944 return -EINVAL;
945 *policy = 0; /* just so it's initialized */
946 task_lock(current);
947 *nmask = cpuset_current_mems_allowed;
948 task_unlock(current);
949 return 0;
950 }
951
952 if (flags & MPOL_F_ADDR) {
953 /*
954 * Do NOT fall back to task policy if the
955 * vma/shared policy at addr is NULL. We
956 * want to return MPOL_DEFAULT in this case.
957 */
958 mmap_read_lock(mm);
959 vma = vma_lookup(mm, addr);
960 if (!vma) {
961 mmap_read_unlock(mm);
962 return -EFAULT;
963 }
964 if (vma->vm_ops && vma->vm_ops->get_policy)
965 pol = vma->vm_ops->get_policy(vma, addr);
966 else
967 pol = vma->vm_policy;
968 } else if (addr)
969 return -EINVAL;
970
971 if (!pol)
972 pol = &default_policy; /* indicates default behavior */
973
974 if (flags & MPOL_F_NODE) {
975 if (flags & MPOL_F_ADDR) {
976 /*
977 * Take a refcount on the mpol, because we are about to
978 * drop the mmap_lock, after which only "pol" remains
979 * valid, "vma" is stale.
980 */
981 pol_refcount = pol;
982 vma = NULL;
983 mpol_get(pol);
984 mmap_read_unlock(mm);
985 err = lookup_node(mm, addr);
986 if (err < 0)
987 goto out;
988 *policy = err;
989 } else if (pol == current->mempolicy &&
990 pol->mode == MPOL_INTERLEAVE) {
991 *policy = next_node_in(current->il_prev, pol->nodes);
992 } else {
993 err = -EINVAL;
994 goto out;
995 }
996 } else {
997 *policy = pol == &default_policy ? MPOL_DEFAULT :
998 pol->mode;
999 /*
1000 * Internal mempolicy flags must be masked off before exposing
1001 * the policy to userspace.
1002 */
1003 *policy |= (pol->flags & MPOL_MODE_FLAGS);
1004 }
1005
1006 err = 0;
1007 if (nmask) {
1008 if (mpol_store_user_nodemask(pol)) {
1009 *nmask = pol->w.user_nodemask;
1010 } else {
1011 task_lock(current);
1012 get_policy_nodemask(pol, nmask);
1013 task_unlock(current);
1014 }
1015 }
1016
1017 out:
1018 mpol_cond_put(pol);
1019 if (vma)
1020 mmap_read_unlock(mm);
1021 if (pol_refcount)
1022 mpol_put(pol_refcount);
1023 return err;
1024}
1025
1026#ifdef CONFIG_MIGRATION
1027/*
1028 * page migration, thp tail pages can be passed.
1029 */
1030static int migrate_page_add(struct page *page, struct list_head *pagelist,
1031 unsigned long flags)
1032{
1033 struct page *head = compound_head(page);
1034 /*
1035 * Avoid migrating a page that is shared with others.
1036 */
1037 if ((flags & MPOL_MF_MOVE_ALL) || page_mapcount(head) == 1) {
1038 if (!isolate_lru_page(head)) {
1039 list_add_tail(&head->lru, pagelist);
1040 mod_node_page_state(page_pgdat(head),
1041 NR_ISOLATED_ANON + page_is_file_lru(head),
1042 thp_nr_pages(head));
1043 } else if (flags & MPOL_MF_STRICT) {
1044 /*
1045 * Non-movable page may reach here. And, there may be
1046 * temporary off LRU pages or non-LRU movable pages.
1047 * Treat them as unmovable pages since they can't be
1048 * isolated, so they can't be moved at the moment. It
1049 * should return -EIO for this case too.
1050 */
1051 return -EIO;
1052 }
1053 }
1054
1055 return 0;
1056}
1057
1058/*
1059 * Migrate pages from one node to a target node.
1060 * Returns error or the number of pages not migrated.
1061 */
1062static int migrate_to_node(struct mm_struct *mm, int source, int dest,
1063 int flags)
1064{
1065 nodemask_t nmask;
1066 struct vm_area_struct *vma;
1067 LIST_HEAD(pagelist);
1068 int err = 0;
1069 struct migration_target_control mtc = {
1070 .nid = dest,
1071 .gfp_mask = GFP_HIGHUSER_MOVABLE | __GFP_THISNODE,
1072 };
1073
1074 nodes_clear(nmask);
1075 node_set(source, nmask);
1076
1077 /*
1078 * This does not "check" the range but isolates all pages that
1079 * need migration. Between passing in the full user address
1080 * space range and MPOL_MF_DISCONTIG_OK, this call can not fail.
1081 */
1082 vma = find_vma(mm, 0);
1083 VM_BUG_ON(!(flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)));
1084 queue_pages_range(mm, vma->vm_start, mm->task_size, &nmask,
1085 flags | MPOL_MF_DISCONTIG_OK, &pagelist);
1086
1087 if (!list_empty(&pagelist)) {
1088 err = migrate_pages(&pagelist, alloc_migration_target, NULL,
1089 (unsigned long)&mtc, MIGRATE_SYNC, MR_SYSCALL, NULL);
1090 if (err)
1091 putback_movable_pages(&pagelist);
1092 }
1093
1094 return err;
1095}
1096
1097/*
1098 * Move pages between the two nodesets so as to preserve the physical
1099 * layout as much as possible.
1100 *
1101 * Returns the number of page that could not be moved.
1102 */
1103int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
1104 const nodemask_t *to, int flags)
1105{
1106 int busy = 0;
1107 int err = 0;
1108 nodemask_t tmp;
1109
1110 lru_cache_disable();
1111
1112 mmap_read_lock(mm);
1113
1114 /*
1115 * Find a 'source' bit set in 'tmp' whose corresponding 'dest'
1116 * bit in 'to' is not also set in 'tmp'. Clear the found 'source'
1117 * bit in 'tmp', and return that <source, dest> pair for migration.
1118 * The pair of nodemasks 'to' and 'from' define the map.
1119 *
1120 * If no pair of bits is found that way, fallback to picking some
1121 * pair of 'source' and 'dest' bits that are not the same. If the
1122 * 'source' and 'dest' bits are the same, this represents a node
1123 * that will be migrating to itself, so no pages need move.
1124 *
1125 * If no bits are left in 'tmp', or if all remaining bits left
1126 * in 'tmp' correspond to the same bit in 'to', return false
1127 * (nothing left to migrate).
1128 *
1129 * This lets us pick a pair of nodes to migrate between, such that
1130 * if possible the dest node is not already occupied by some other
1131 * source node, minimizing the risk of overloading the memory on a
1132 * node that would happen if we migrated incoming memory to a node
1133 * before migrating outgoing memory source that same node.
1134 *
1135 * A single scan of tmp is sufficient. As we go, we remember the
1136 * most recent <s, d> pair that moved (s != d). If we find a pair
1137 * that not only moved, but what's better, moved to an empty slot
1138 * (d is not set in tmp), then we break out then, with that pair.
1139 * Otherwise when we finish scanning from_tmp, we at least have the
1140 * most recent <s, d> pair that moved. If we get all the way through
1141 * the scan of tmp without finding any node that moved, much less
1142 * moved to an empty node, then there is nothing left worth migrating.
1143 */
1144
1145 tmp = *from;
1146 while (!nodes_empty(tmp)) {
1147 int s, d;
1148 int source = NUMA_NO_NODE;
1149 int dest = 0;
1150
1151 for_each_node_mask(s, tmp) {
1152
1153 /*
1154 * do_migrate_pages() tries to maintain the relative
1155 * node relationship of the pages established between
1156 * threads and memory areas.
1157 *
1158 * However if the number of source nodes is not equal to
1159 * the number of destination nodes we can not preserve
1160 * this node relative relationship. In that case, skip
1161 * copying memory from a node that is in the destination
1162 * mask.
1163 *
1164 * Example: [2,3,4] -> [3,4,5] moves everything.
1165 * [0-7] - > [3,4,5] moves only 0,1,2,6,7.
1166 */
1167
1168 if ((nodes_weight(*from) != nodes_weight(*to)) &&
1169 (node_isset(s, *to)))
1170 continue;
1171
1172 d = node_remap(s, *from, *to);
1173 if (s == d)
1174 continue;
1175
1176 source = s; /* Node moved. Memorize */
1177 dest = d;
1178
1179 /* dest not in remaining from nodes? */
1180 if (!node_isset(dest, tmp))
1181 break;
1182 }
1183 if (source == NUMA_NO_NODE)
1184 break;
1185
1186 node_clear(source, tmp);
1187 err = migrate_to_node(mm, source, dest, flags);
1188 if (err > 0)
1189 busy += err;
1190 if (err < 0)
1191 break;
1192 }
1193 mmap_read_unlock(mm);
1194
1195 lru_cache_enable();
1196 if (err < 0)
1197 return err;
1198 return busy;
1199
1200}
1201
1202/*
1203 * Allocate a new page for page migration based on vma policy.
1204 * Start by assuming the page is mapped by the same vma as contains @start.
1205 * Search forward from there, if not. N.B., this assumes that the
1206 * list of pages handed to migrate_pages()--which is how we get here--
1207 * is in virtual address order.
1208 */
1209static struct page *new_page(struct page *page, unsigned long start)
1210{
1211 struct folio *dst, *src = page_folio(page);
1212 struct vm_area_struct *vma;
1213 unsigned long address;
1214 VMA_ITERATOR(vmi, current->mm, start);
1215 gfp_t gfp = GFP_HIGHUSER_MOVABLE | __GFP_RETRY_MAYFAIL;
1216
1217 for_each_vma(vmi, vma) {
1218 address = page_address_in_vma(page, vma);
1219 if (address != -EFAULT)
1220 break;
1221 }
1222
1223 if (folio_test_hugetlb(src))
1224 return alloc_huge_page_vma(page_hstate(&src->page),
1225 vma, address);
1226
1227 if (folio_test_large(src))
1228 gfp = GFP_TRANSHUGE;
1229
1230 /*
1231 * if !vma, vma_alloc_folio() will use task or system default policy
1232 */
1233 dst = vma_alloc_folio(gfp, folio_order(src), vma, address,
1234 folio_test_large(src));
1235 return &dst->page;
1236}
1237#else
1238
1239static int migrate_page_add(struct page *page, struct list_head *pagelist,
1240 unsigned long flags)
1241{
1242 return -EIO;
1243}
1244
1245int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
1246 const nodemask_t *to, int flags)
1247{
1248 return -ENOSYS;
1249}
1250
1251static struct page *new_page(struct page *page, unsigned long start)
1252{
1253 return NULL;
1254}
1255#endif
1256
1257static long do_mbind(unsigned long start, unsigned long len,
1258 unsigned short mode, unsigned short mode_flags,
1259 nodemask_t *nmask, unsigned long flags)
1260{
1261 struct mm_struct *mm = current->mm;
1262 struct mempolicy *new;
1263 unsigned long end;
1264 int err;
1265 int ret;
1266 LIST_HEAD(pagelist);
1267
1268 if (flags & ~(unsigned long)MPOL_MF_VALID)
1269 return -EINVAL;
1270 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1271 return -EPERM;
1272
1273 if (start & ~PAGE_MASK)
1274 return -EINVAL;
1275
1276 if (mode == MPOL_DEFAULT)
1277 flags &= ~MPOL_MF_STRICT;
1278
1279 len = PAGE_ALIGN(len);
1280 end = start + len;
1281
1282 if (end < start)
1283 return -EINVAL;
1284 if (end == start)
1285 return 0;
1286
1287 new = mpol_new(mode, mode_flags, nmask);
1288 if (IS_ERR(new))
1289 return PTR_ERR(new);
1290
1291 if (flags & MPOL_MF_LAZY)
1292 new->flags |= MPOL_F_MOF;
1293
1294 /*
1295 * If we are using the default policy then operation
1296 * on discontinuous address spaces is okay after all
1297 */
1298 if (!new)
1299 flags |= MPOL_MF_DISCONTIG_OK;
1300
1301 pr_debug("mbind %lx-%lx mode:%d flags:%d nodes:%lx\n",
1302 start, start + len, mode, mode_flags,
1303 nmask ? nodes_addr(*nmask)[0] : NUMA_NO_NODE);
1304
1305 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) {
1306
1307 lru_cache_disable();
1308 }
1309 {
1310 NODEMASK_SCRATCH(scratch);
1311 if (scratch) {
1312 mmap_write_lock(mm);
1313 err = mpol_set_nodemask(new, nmask, scratch);
1314 if (err)
1315 mmap_write_unlock(mm);
1316 } else
1317 err = -ENOMEM;
1318 NODEMASK_SCRATCH_FREE(scratch);
1319 }
1320 if (err)
1321 goto mpol_out;
1322
1323 ret = queue_pages_range(mm, start, end, nmask,
1324 flags | MPOL_MF_INVERT, &pagelist);
1325
1326 if (ret < 0) {
1327 err = ret;
1328 goto up_out;
1329 }
1330
1331 err = mbind_range(mm, start, end, new);
1332
1333 if (!err) {
1334 int nr_failed = 0;
1335
1336 if (!list_empty(&pagelist)) {
1337 WARN_ON_ONCE(flags & MPOL_MF_LAZY);
1338 nr_failed = migrate_pages(&pagelist, new_page, NULL,
1339 start, MIGRATE_SYNC, MR_MEMPOLICY_MBIND, NULL);
1340 if (nr_failed)
1341 putback_movable_pages(&pagelist);
1342 }
1343
1344 if ((ret > 0) || (nr_failed && (flags & MPOL_MF_STRICT)))
1345 err = -EIO;
1346 } else {
1347up_out:
1348 if (!list_empty(&pagelist))
1349 putback_movable_pages(&pagelist);
1350 }
1351
1352 mmap_write_unlock(mm);
1353mpol_out:
1354 mpol_put(new);
1355 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))
1356 lru_cache_enable();
1357 return err;
1358}
1359
1360/*
1361 * User space interface with variable sized bitmaps for nodelists.
1362 */
1363static int get_bitmap(unsigned long *mask, const unsigned long __user *nmask,
1364 unsigned long maxnode)
1365{
1366 unsigned long nlongs = BITS_TO_LONGS(maxnode);
1367 int ret;
1368
1369 if (in_compat_syscall())
1370 ret = compat_get_bitmap(mask,
1371 (const compat_ulong_t __user *)nmask,
1372 maxnode);
1373 else
1374 ret = copy_from_user(mask, nmask,
1375 nlongs * sizeof(unsigned long));
1376
1377 if (ret)
1378 return -EFAULT;
1379
1380 if (maxnode % BITS_PER_LONG)
1381 mask[nlongs - 1] &= (1UL << (maxnode % BITS_PER_LONG)) - 1;
1382
1383 return 0;
1384}
1385
1386/* Copy a node mask from user space. */
1387static int get_nodes(nodemask_t *nodes, const unsigned long __user *nmask,
1388 unsigned long maxnode)
1389{
1390 --maxnode;
1391 nodes_clear(*nodes);
1392 if (maxnode == 0 || !nmask)
1393 return 0;
1394 if (maxnode > PAGE_SIZE*BITS_PER_BYTE)
1395 return -EINVAL;
1396
1397 /*
1398 * When the user specified more nodes than supported just check
1399 * if the non supported part is all zero, one word at a time,
1400 * starting at the end.
1401 */
1402 while (maxnode > MAX_NUMNODES) {
1403 unsigned long bits = min_t(unsigned long, maxnode, BITS_PER_LONG);
1404 unsigned long t;
1405
1406 if (get_bitmap(&t, &nmask[(maxnode - 1) / BITS_PER_LONG], bits))
1407 return -EFAULT;
1408
1409 if (maxnode - bits >= MAX_NUMNODES) {
1410 maxnode -= bits;
1411 } else {
1412 maxnode = MAX_NUMNODES;
1413 t &= ~((1UL << (MAX_NUMNODES % BITS_PER_LONG)) - 1);
1414 }
1415 if (t)
1416 return -EINVAL;
1417 }
1418
1419 return get_bitmap(nodes_addr(*nodes), nmask, maxnode);
1420}
1421
1422/* Copy a kernel node mask to user space */
1423static int copy_nodes_to_user(unsigned long __user *mask, unsigned long maxnode,
1424 nodemask_t *nodes)
1425{
1426 unsigned long copy = ALIGN(maxnode-1, 64) / 8;
1427 unsigned int nbytes = BITS_TO_LONGS(nr_node_ids) * sizeof(long);
1428 bool compat = in_compat_syscall();
1429
1430 if (compat)
1431 nbytes = BITS_TO_COMPAT_LONGS(nr_node_ids) * sizeof(compat_long_t);
1432
1433 if (copy > nbytes) {
1434 if (copy > PAGE_SIZE)
1435 return -EINVAL;
1436 if (clear_user((char __user *)mask + nbytes, copy - nbytes))
1437 return -EFAULT;
1438 copy = nbytes;
1439 maxnode = nr_node_ids;
1440 }
1441
1442 if (compat)
1443 return compat_put_bitmap((compat_ulong_t __user *)mask,
1444 nodes_addr(*nodes), maxnode);
1445
1446 return copy_to_user(mask, nodes_addr(*nodes), copy) ? -EFAULT : 0;
1447}
1448
1449/* Basic parameter sanity check used by both mbind() and set_mempolicy() */
1450static inline int sanitize_mpol_flags(int *mode, unsigned short *flags)
1451{
1452 *flags = *mode & MPOL_MODE_FLAGS;
1453 *mode &= ~MPOL_MODE_FLAGS;
1454
1455 if ((unsigned int)(*mode) >= MPOL_MAX)
1456 return -EINVAL;
1457 if ((*flags & MPOL_F_STATIC_NODES) && (*flags & MPOL_F_RELATIVE_NODES))
1458 return -EINVAL;
1459 if (*flags & MPOL_F_NUMA_BALANCING) {
1460 if (*mode != MPOL_BIND)
1461 return -EINVAL;
1462 *flags |= (MPOL_F_MOF | MPOL_F_MORON);
1463 }
1464 return 0;
1465}
1466
1467static long kernel_mbind(unsigned long start, unsigned long len,
1468 unsigned long mode, const unsigned long __user *nmask,
1469 unsigned long maxnode, unsigned int flags)
1470{
1471 unsigned short mode_flags;
1472 nodemask_t nodes;
1473 int lmode = mode;
1474 int err;
1475
1476 start = untagged_addr(start);
1477 err = sanitize_mpol_flags(&lmode, &mode_flags);
1478 if (err)
1479 return err;
1480
1481 err = get_nodes(&nodes, nmask, maxnode);
1482 if (err)
1483 return err;
1484
1485 return do_mbind(start, len, lmode, mode_flags, &nodes, flags);
1486}
1487
1488SYSCALL_DEFINE4(set_mempolicy_home_node, unsigned long, start, unsigned long, len,
1489 unsigned long, home_node, unsigned long, flags)
1490{
1491 struct mm_struct *mm = current->mm;
1492 struct vm_area_struct *vma;
1493 struct mempolicy *new;
1494 unsigned long vmstart;
1495 unsigned long vmend;
1496 unsigned long end;
1497 int err = -ENOENT;
1498 VMA_ITERATOR(vmi, mm, start);
1499
1500 start = untagged_addr(start);
1501 if (start & ~PAGE_MASK)
1502 return -EINVAL;
1503 /*
1504 * flags is used for future extension if any.
1505 */
1506 if (flags != 0)
1507 return -EINVAL;
1508
1509 /*
1510 * Check home_node is online to avoid accessing uninitialized
1511 * NODE_DATA.
1512 */
1513 if (home_node >= MAX_NUMNODES || !node_online(home_node))
1514 return -EINVAL;
1515
1516 len = PAGE_ALIGN(len);
1517 end = start + len;
1518
1519 if (end < start)
1520 return -EINVAL;
1521 if (end == start)
1522 return 0;
1523 mmap_write_lock(mm);
1524 for_each_vma_range(vmi, vma, end) {
1525 vmstart = max(start, vma->vm_start);
1526 vmend = min(end, vma->vm_end);
1527 new = mpol_dup(vma_policy(vma));
1528 if (IS_ERR(new)) {
1529 err = PTR_ERR(new);
1530 break;
1531 }
1532 /*
1533 * Only update home node if there is an existing vma policy
1534 */
1535 if (!new)
1536 continue;
1537
1538 /*
1539 * If any vma in the range got policy other than MPOL_BIND
1540 * or MPOL_PREFERRED_MANY we return error. We don't reset
1541 * the home node for vmas we already updated before.
1542 */
1543 if (new->mode != MPOL_BIND && new->mode != MPOL_PREFERRED_MANY) {
1544 mpol_put(new);
1545 err = -EOPNOTSUPP;
1546 break;
1547 }
1548
1549 new->home_node = home_node;
1550 err = mbind_range(mm, vmstart, vmend, new);
1551 mpol_put(new);
1552 if (err)
1553 break;
1554 }
1555 mmap_write_unlock(mm);
1556 return err;
1557}
1558
1559SYSCALL_DEFINE6(mbind, unsigned long, start, unsigned long, len,
1560 unsigned long, mode, const unsigned long __user *, nmask,
1561 unsigned long, maxnode, unsigned int, flags)
1562{
1563 return kernel_mbind(start, len, mode, nmask, maxnode, flags);
1564}
1565
1566/* Set the process memory policy */
1567static long kernel_set_mempolicy(int mode, const unsigned long __user *nmask,
1568 unsigned long maxnode)
1569{
1570 unsigned short mode_flags;
1571 nodemask_t nodes;
1572 int lmode = mode;
1573 int err;
1574
1575 err = sanitize_mpol_flags(&lmode, &mode_flags);
1576 if (err)
1577 return err;
1578
1579 err = get_nodes(&nodes, nmask, maxnode);
1580 if (err)
1581 return err;
1582
1583 return do_set_mempolicy(lmode, mode_flags, &nodes);
1584}
1585
1586SYSCALL_DEFINE3(set_mempolicy, int, mode, const unsigned long __user *, nmask,
1587 unsigned long, maxnode)
1588{
1589 return kernel_set_mempolicy(mode, nmask, maxnode);
1590}
1591
1592static int kernel_migrate_pages(pid_t pid, unsigned long maxnode,
1593 const unsigned long __user *old_nodes,
1594 const unsigned long __user *new_nodes)
1595{
1596 struct mm_struct *mm = NULL;
1597 struct task_struct *task;
1598 nodemask_t task_nodes;
1599 int err;
1600 nodemask_t *old;
1601 nodemask_t *new;
1602 NODEMASK_SCRATCH(scratch);
1603
1604 if (!scratch)
1605 return -ENOMEM;
1606
1607 old = &scratch->mask1;
1608 new = &scratch->mask2;
1609
1610 err = get_nodes(old, old_nodes, maxnode);
1611 if (err)
1612 goto out;
1613
1614 err = get_nodes(new, new_nodes, maxnode);
1615 if (err)
1616 goto out;
1617
1618 /* Find the mm_struct */
1619 rcu_read_lock();
1620 task = pid ? find_task_by_vpid(pid) : current;
1621 if (!task) {
1622 rcu_read_unlock();
1623 err = -ESRCH;
1624 goto out;
1625 }
1626 get_task_struct(task);
1627
1628 err = -EINVAL;
1629
1630 /*
1631 * Check if this process has the right to modify the specified process.
1632 * Use the regular "ptrace_may_access()" checks.
1633 */
1634 if (!ptrace_may_access(task, PTRACE_MODE_READ_REALCREDS)) {
1635 rcu_read_unlock();
1636 err = -EPERM;
1637 goto out_put;
1638 }
1639 rcu_read_unlock();
1640
1641 task_nodes = cpuset_mems_allowed(task);
1642 /* Is the user allowed to access the target nodes? */
1643 if (!nodes_subset(*new, task_nodes) && !capable(CAP_SYS_NICE)) {
1644 err = -EPERM;
1645 goto out_put;
1646 }
1647
1648 task_nodes = cpuset_mems_allowed(current);
1649 nodes_and(*new, *new, task_nodes);
1650 if (nodes_empty(*new))
1651 goto out_put;
1652
1653 err = security_task_movememory(task);
1654 if (err)
1655 goto out_put;
1656
1657 mm = get_task_mm(task);
1658 put_task_struct(task);
1659
1660 if (!mm) {
1661 err = -EINVAL;
1662 goto out;
1663 }
1664
1665 err = do_migrate_pages(mm, old, new,
1666 capable(CAP_SYS_NICE) ? MPOL_MF_MOVE_ALL : MPOL_MF_MOVE);
1667
1668 mmput(mm);
1669out:
1670 NODEMASK_SCRATCH_FREE(scratch);
1671
1672 return err;
1673
1674out_put:
1675 put_task_struct(task);
1676 goto out;
1677
1678}
1679
1680SYSCALL_DEFINE4(migrate_pages, pid_t, pid, unsigned long, maxnode,
1681 const unsigned long __user *, old_nodes,
1682 const unsigned long __user *, new_nodes)
1683{
1684 return kernel_migrate_pages(pid, maxnode, old_nodes, new_nodes);
1685}
1686
1687
1688/* Retrieve NUMA policy */
1689static int kernel_get_mempolicy(int __user *policy,
1690 unsigned long __user *nmask,
1691 unsigned long maxnode,
1692 unsigned long addr,
1693 unsigned long flags)
1694{
1695 int err;
1696 int pval;
1697 nodemask_t nodes;
1698
1699 if (nmask != NULL && maxnode < nr_node_ids)
1700 return -EINVAL;
1701
1702 addr = untagged_addr(addr);
1703
1704 err = do_get_mempolicy(&pval, &nodes, addr, flags);
1705
1706 if (err)
1707 return err;
1708
1709 if (policy && put_user(pval, policy))
1710 return -EFAULT;
1711
1712 if (nmask)
1713 err = copy_nodes_to_user(nmask, maxnode, &nodes);
1714
1715 return err;
1716}
1717
1718SYSCALL_DEFINE5(get_mempolicy, int __user *, policy,
1719 unsigned long __user *, nmask, unsigned long, maxnode,
1720 unsigned long, addr, unsigned long, flags)
1721{
1722 return kernel_get_mempolicy(policy, nmask, maxnode, addr, flags);
1723}
1724
1725bool vma_migratable(struct vm_area_struct *vma)
1726{
1727 if (vma->vm_flags & (VM_IO | VM_PFNMAP))
1728 return false;
1729
1730 /*
1731 * DAX device mappings require predictable access latency, so avoid
1732 * incurring periodic faults.
1733 */
1734 if (vma_is_dax(vma))
1735 return false;
1736
1737 if (is_vm_hugetlb_page(vma) &&
1738 !hugepage_migration_supported(hstate_vma(vma)))
1739 return false;
1740
1741 /*
1742 * Migration allocates pages in the highest zone. If we cannot
1743 * do so then migration (at least from node to node) is not
1744 * possible.
1745 */
1746 if (vma->vm_file &&
1747 gfp_zone(mapping_gfp_mask(vma->vm_file->f_mapping))
1748 < policy_zone)
1749 return false;
1750 return true;
1751}
1752
1753struct mempolicy *__get_vma_policy(struct vm_area_struct *vma,
1754 unsigned long addr)
1755{
1756 struct mempolicy *pol = NULL;
1757
1758 if (vma) {
1759 if (vma->vm_ops && vma->vm_ops->get_policy) {
1760 pol = vma->vm_ops->get_policy(vma, addr);
1761 } else if (vma->vm_policy) {
1762 pol = vma->vm_policy;
1763
1764 /*
1765 * shmem_alloc_page() passes MPOL_F_SHARED policy with
1766 * a pseudo vma whose vma->vm_ops=NULL. Take a reference
1767 * count on these policies which will be dropped by
1768 * mpol_cond_put() later
1769 */
1770 if (mpol_needs_cond_ref(pol))
1771 mpol_get(pol);
1772 }
1773 }
1774
1775 return pol;
1776}
1777
1778/*
1779 * get_vma_policy(@vma, @addr)
1780 * @vma: virtual memory area whose policy is sought
1781 * @addr: address in @vma for shared policy lookup
1782 *
1783 * Returns effective policy for a VMA at specified address.
1784 * Falls back to current->mempolicy or system default policy, as necessary.
1785 * Shared policies [those marked as MPOL_F_SHARED] require an extra reference
1786 * count--added by the get_policy() vm_op, as appropriate--to protect against
1787 * freeing by another task. It is the caller's responsibility to free the
1788 * extra reference for shared policies.
1789 */
1790static struct mempolicy *get_vma_policy(struct vm_area_struct *vma,
1791 unsigned long addr)
1792{
1793 struct mempolicy *pol = __get_vma_policy(vma, addr);
1794
1795 if (!pol)
1796 pol = get_task_policy(current);
1797
1798 return pol;
1799}
1800
1801bool vma_policy_mof(struct vm_area_struct *vma)
1802{
1803 struct mempolicy *pol;
1804
1805 if (vma->vm_ops && vma->vm_ops->get_policy) {
1806 bool ret = false;
1807
1808 pol = vma->vm_ops->get_policy(vma, vma->vm_start);
1809 if (pol && (pol->flags & MPOL_F_MOF))
1810 ret = true;
1811 mpol_cond_put(pol);
1812
1813 return ret;
1814 }
1815
1816 pol = vma->vm_policy;
1817 if (!pol)
1818 pol = get_task_policy(current);
1819
1820 return pol->flags & MPOL_F_MOF;
1821}
1822
1823bool apply_policy_zone(struct mempolicy *policy, enum zone_type zone)
1824{
1825 enum zone_type dynamic_policy_zone = policy_zone;
1826
1827 BUG_ON(dynamic_policy_zone == ZONE_MOVABLE);
1828
1829 /*
1830 * if policy->nodes has movable memory only,
1831 * we apply policy when gfp_zone(gfp) = ZONE_MOVABLE only.
1832 *
1833 * policy->nodes is intersect with node_states[N_MEMORY].
1834 * so if the following test fails, it implies
1835 * policy->nodes has movable memory only.
1836 */
1837 if (!nodes_intersects(policy->nodes, node_states[N_HIGH_MEMORY]))
1838 dynamic_policy_zone = ZONE_MOVABLE;
1839
1840 return zone >= dynamic_policy_zone;
1841}
1842
1843/*
1844 * Return a nodemask representing a mempolicy for filtering nodes for
1845 * page allocation
1846 */
1847nodemask_t *policy_nodemask(gfp_t gfp, struct mempolicy *policy)
1848{
1849 int mode = policy->mode;
1850
1851 /* Lower zones don't get a nodemask applied for MPOL_BIND */
1852 if (unlikely(mode == MPOL_BIND) &&
1853 apply_policy_zone(policy, gfp_zone(gfp)) &&
1854 cpuset_nodemask_valid_mems_allowed(&policy->nodes))
1855 return &policy->nodes;
1856
1857 if (mode == MPOL_PREFERRED_MANY)
1858 return &policy->nodes;
1859
1860 return NULL;
1861}
1862
1863/*
1864 * Return the preferred node id for 'prefer' mempolicy, and return
1865 * the given id for all other policies.
1866 *
1867 * policy_node() is always coupled with policy_nodemask(), which
1868 * secures the nodemask limit for 'bind' and 'prefer-many' policy.
1869 */
1870static int policy_node(gfp_t gfp, struct mempolicy *policy, int nd)
1871{
1872 if (policy->mode == MPOL_PREFERRED) {
1873 nd = first_node(policy->nodes);
1874 } else {
1875 /*
1876 * __GFP_THISNODE shouldn't even be used with the bind policy
1877 * because we might easily break the expectation to stay on the
1878 * requested node and not break the policy.
1879 */
1880 WARN_ON_ONCE(policy->mode == MPOL_BIND && (gfp & __GFP_THISNODE));
1881 }
1882
1883 if ((policy->mode == MPOL_BIND ||
1884 policy->mode == MPOL_PREFERRED_MANY) &&
1885 policy->home_node != NUMA_NO_NODE)
1886 return policy->home_node;
1887
1888 return nd;
1889}
1890
1891/* Do dynamic interleaving for a process */
1892static unsigned interleave_nodes(struct mempolicy *policy)
1893{
1894 unsigned next;
1895 struct task_struct *me = current;
1896
1897 next = next_node_in(me->il_prev, policy->nodes);
1898 if (next < MAX_NUMNODES)
1899 me->il_prev = next;
1900 return next;
1901}
1902
1903/*
1904 * Depending on the memory policy provide a node from which to allocate the
1905 * next slab entry.
1906 */
1907unsigned int mempolicy_slab_node(void)
1908{
1909 struct mempolicy *policy;
1910 int node = numa_mem_id();
1911
1912 if (!in_task())
1913 return node;
1914
1915 policy = current->mempolicy;
1916 if (!policy)
1917 return node;
1918
1919 switch (policy->mode) {
1920 case MPOL_PREFERRED:
1921 return first_node(policy->nodes);
1922
1923 case MPOL_INTERLEAVE:
1924 return interleave_nodes(policy);
1925
1926 case MPOL_BIND:
1927 case MPOL_PREFERRED_MANY:
1928 {
1929 struct zoneref *z;
1930
1931 /*
1932 * Follow bind policy behavior and start allocation at the
1933 * first node.
1934 */
1935 struct zonelist *zonelist;
1936 enum zone_type highest_zoneidx = gfp_zone(GFP_KERNEL);
1937 zonelist = &NODE_DATA(node)->node_zonelists[ZONELIST_FALLBACK];
1938 z = first_zones_zonelist(zonelist, highest_zoneidx,
1939 &policy->nodes);
1940 return z->zone ? zone_to_nid(z->zone) : node;
1941 }
1942 case MPOL_LOCAL:
1943 return node;
1944
1945 default:
1946 BUG();
1947 }
1948}
1949
1950/*
1951 * Do static interleaving for a VMA with known offset @n. Returns the n'th
1952 * node in pol->nodes (starting from n=0), wrapping around if n exceeds the
1953 * number of present nodes.
1954 */
1955static unsigned offset_il_node(struct mempolicy *pol, unsigned long n)
1956{
1957 nodemask_t nodemask = pol->nodes;
1958 unsigned int target, nnodes;
1959 int i;
1960 int nid;
1961 /*
1962 * The barrier will stabilize the nodemask in a register or on
1963 * the stack so that it will stop changing under the code.
1964 *
1965 * Between first_node() and next_node(), pol->nodes could be changed
1966 * by other threads. So we put pol->nodes in a local stack.
1967 */
1968 barrier();
1969
1970 nnodes = nodes_weight(nodemask);
1971 if (!nnodes)
1972 return numa_node_id();
1973 target = (unsigned int)n % nnodes;
1974 nid = first_node(nodemask);
1975 for (i = 0; i < target; i++)
1976 nid = next_node(nid, nodemask);
1977 return nid;
1978}
1979
1980/* Determine a node number for interleave */
1981static inline unsigned interleave_nid(struct mempolicy *pol,
1982 struct vm_area_struct *vma, unsigned long addr, int shift)
1983{
1984 if (vma) {
1985 unsigned long off;
1986
1987 /*
1988 * for small pages, there is no difference between
1989 * shift and PAGE_SHIFT, so the bit-shift is safe.
1990 * for huge pages, since vm_pgoff is in units of small
1991 * pages, we need to shift off the always 0 bits to get
1992 * a useful offset.
1993 */
1994 BUG_ON(shift < PAGE_SHIFT);
1995 off = vma->vm_pgoff >> (shift - PAGE_SHIFT);
1996 off += (addr - vma->vm_start) >> shift;
1997 return offset_il_node(pol, off);
1998 } else
1999 return interleave_nodes(pol);
2000}
2001
2002#ifdef CONFIG_HUGETLBFS
2003/*
2004 * huge_node(@vma, @addr, @gfp_flags, @mpol)
2005 * @vma: virtual memory area whose policy is sought
2006 * @addr: address in @vma for shared policy lookup and interleave policy
2007 * @gfp_flags: for requested zone
2008 * @mpol: pointer to mempolicy pointer for reference counted mempolicy
2009 * @nodemask: pointer to nodemask pointer for 'bind' and 'prefer-many' policy
2010 *
2011 * Returns a nid suitable for a huge page allocation and a pointer
2012 * to the struct mempolicy for conditional unref after allocation.
2013 * If the effective policy is 'bind' or 'prefer-many', returns a pointer
2014 * to the mempolicy's @nodemask for filtering the zonelist.
2015 *
2016 * Must be protected by read_mems_allowed_begin()
2017 */
2018int huge_node(struct vm_area_struct *vma, unsigned long addr, gfp_t gfp_flags,
2019 struct mempolicy **mpol, nodemask_t **nodemask)
2020{
2021 int nid;
2022 int mode;
2023
2024 *mpol = get_vma_policy(vma, addr);
2025 *nodemask = NULL;
2026 mode = (*mpol)->mode;
2027
2028 if (unlikely(mode == MPOL_INTERLEAVE)) {
2029 nid = interleave_nid(*mpol, vma, addr,
2030 huge_page_shift(hstate_vma(vma)));
2031 } else {
2032 nid = policy_node(gfp_flags, *mpol, numa_node_id());
2033 if (mode == MPOL_BIND || mode == MPOL_PREFERRED_MANY)
2034 *nodemask = &(*mpol)->nodes;
2035 }
2036 return nid;
2037}
2038
2039/*
2040 * init_nodemask_of_mempolicy
2041 *
2042 * If the current task's mempolicy is "default" [NULL], return 'false'
2043 * to indicate default policy. Otherwise, extract the policy nodemask
2044 * for 'bind' or 'interleave' policy into the argument nodemask, or
2045 * initialize the argument nodemask to contain the single node for
2046 * 'preferred' or 'local' policy and return 'true' to indicate presence
2047 * of non-default mempolicy.
2048 *
2049 * We don't bother with reference counting the mempolicy [mpol_get/put]
2050 * because the current task is examining it's own mempolicy and a task's
2051 * mempolicy is only ever changed by the task itself.
2052 *
2053 * N.B., it is the caller's responsibility to free a returned nodemask.
2054 */
2055bool init_nodemask_of_mempolicy(nodemask_t *mask)
2056{
2057 struct mempolicy *mempolicy;
2058
2059 if (!(mask && current->mempolicy))
2060 return false;
2061
2062 task_lock(current);
2063 mempolicy = current->mempolicy;
2064 switch (mempolicy->mode) {
2065 case MPOL_PREFERRED:
2066 case MPOL_PREFERRED_MANY:
2067 case MPOL_BIND:
2068 case MPOL_INTERLEAVE:
2069 *mask = mempolicy->nodes;
2070 break;
2071
2072 case MPOL_LOCAL:
2073 init_nodemask_of_node(mask, numa_node_id());
2074 break;
2075
2076 default:
2077 BUG();
2078 }
2079 task_unlock(current);
2080
2081 return true;
2082}
2083#endif
2084
2085/*
2086 * mempolicy_in_oom_domain
2087 *
2088 * If tsk's mempolicy is "bind", check for intersection between mask and
2089 * the policy nodemask. Otherwise, return true for all other policies
2090 * including "interleave", as a tsk with "interleave" policy may have
2091 * memory allocated from all nodes in system.
2092 *
2093 * Takes task_lock(tsk) to prevent freeing of its mempolicy.
2094 */
2095bool mempolicy_in_oom_domain(struct task_struct *tsk,
2096 const nodemask_t *mask)
2097{
2098 struct mempolicy *mempolicy;
2099 bool ret = true;
2100
2101 if (!mask)
2102 return ret;
2103
2104 task_lock(tsk);
2105 mempolicy = tsk->mempolicy;
2106 if (mempolicy && mempolicy->mode == MPOL_BIND)
2107 ret = nodes_intersects(mempolicy->nodes, *mask);
2108 task_unlock(tsk);
2109
2110 return ret;
2111}
2112
2113/* Allocate a page in interleaved policy.
2114 Own path because it needs to do special accounting. */
2115static struct page *alloc_page_interleave(gfp_t gfp, unsigned order,
2116 unsigned nid)
2117{
2118 struct page *page;
2119
2120 page = __alloc_pages(gfp, order, nid, NULL);
2121 /* skip NUMA_INTERLEAVE_HIT counter update if numa stats is disabled */
2122 if (!static_branch_likely(&vm_numa_stat_key))
2123 return page;
2124 if (page && page_to_nid(page) == nid) {
2125 preempt_disable();
2126 __count_numa_event(page_zone(page), NUMA_INTERLEAVE_HIT);
2127 preempt_enable();
2128 }
2129 return page;
2130}
2131
2132static struct page *alloc_pages_preferred_many(gfp_t gfp, unsigned int order,
2133 int nid, struct mempolicy *pol)
2134{
2135 struct page *page;
2136 gfp_t preferred_gfp;
2137
2138 /*
2139 * This is a two pass approach. The first pass will only try the
2140 * preferred nodes but skip the direct reclaim and allow the
2141 * allocation to fail, while the second pass will try all the
2142 * nodes in system.
2143 */
2144 preferred_gfp = gfp | __GFP_NOWARN;
2145 preferred_gfp &= ~(__GFP_DIRECT_RECLAIM | __GFP_NOFAIL);
2146 page = __alloc_pages(preferred_gfp, order, nid, &pol->nodes);
2147 if (!page)
2148 page = __alloc_pages(gfp, order, nid, NULL);
2149
2150 return page;
2151}
2152
2153/**
2154 * vma_alloc_folio - Allocate a folio for a VMA.
2155 * @gfp: GFP flags.
2156 * @order: Order of the folio.
2157 * @vma: Pointer to VMA or NULL if not available.
2158 * @addr: Virtual address of the allocation. Must be inside @vma.
2159 * @hugepage: For hugepages try only the preferred node if possible.
2160 *
2161 * Allocate a folio for a specific address in @vma, using the appropriate
2162 * NUMA policy. When @vma is not NULL the caller must hold the mmap_lock
2163 * of the mm_struct of the VMA to prevent it from going away. Should be
2164 * used for all allocations for folios that will be mapped into user space.
2165 *
2166 * Return: The folio on success or NULL if allocation fails.
2167 */
2168struct folio *vma_alloc_folio(gfp_t gfp, int order, struct vm_area_struct *vma,
2169 unsigned long addr, bool hugepage)
2170{
2171 struct mempolicy *pol;
2172 int node = numa_node_id();
2173 struct folio *folio;
2174 int preferred_nid;
2175 nodemask_t *nmask;
2176
2177 pol = get_vma_policy(vma, addr);
2178
2179 if (pol->mode == MPOL_INTERLEAVE) {
2180 struct page *page;
2181 unsigned nid;
2182
2183 nid = interleave_nid(pol, vma, addr, PAGE_SHIFT + order);
2184 mpol_cond_put(pol);
2185 gfp |= __GFP_COMP;
2186 page = alloc_page_interleave(gfp, order, nid);
2187 if (page && order > 1)
2188 prep_transhuge_page(page);
2189 folio = (struct folio *)page;
2190 goto out;
2191 }
2192
2193 if (pol->mode == MPOL_PREFERRED_MANY) {
2194 struct page *page;
2195
2196 node = policy_node(gfp, pol, node);
2197 gfp |= __GFP_COMP;
2198 page = alloc_pages_preferred_many(gfp, order, node, pol);
2199 mpol_cond_put(pol);
2200 if (page && order > 1)
2201 prep_transhuge_page(page);
2202 folio = (struct folio *)page;
2203 goto out;
2204 }
2205
2206 if (unlikely(IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE) && hugepage)) {
2207 int hpage_node = node;
2208
2209 /*
2210 * For hugepage allocation and non-interleave policy which
2211 * allows the current node (or other explicitly preferred
2212 * node) we only try to allocate from the current/preferred
2213 * node and don't fall back to other nodes, as the cost of
2214 * remote accesses would likely offset THP benefits.
2215 *
2216 * If the policy is interleave or does not allow the current
2217 * node in its nodemask, we allocate the standard way.
2218 */
2219 if (pol->mode == MPOL_PREFERRED)
2220 hpage_node = first_node(pol->nodes);
2221
2222 nmask = policy_nodemask(gfp, pol);
2223 if (!nmask || node_isset(hpage_node, *nmask)) {
2224 mpol_cond_put(pol);
2225 /*
2226 * First, try to allocate THP only on local node, but
2227 * don't reclaim unnecessarily, just compact.
2228 */
2229 folio = __folio_alloc_node(gfp | __GFP_THISNODE |
2230 __GFP_NORETRY, order, hpage_node);
2231
2232 /*
2233 * If hugepage allocations are configured to always
2234 * synchronous compact or the vma has been madvised
2235 * to prefer hugepage backing, retry allowing remote
2236 * memory with both reclaim and compact as well.
2237 */
2238 if (!folio && (gfp & __GFP_DIRECT_RECLAIM))
2239 folio = __folio_alloc(gfp, order, hpage_node,
2240 nmask);
2241
2242 goto out;
2243 }
2244 }
2245
2246 nmask = policy_nodemask(gfp, pol);
2247 preferred_nid = policy_node(gfp, pol, node);
2248 folio = __folio_alloc(gfp, order, preferred_nid, nmask);
2249 mpol_cond_put(pol);
2250out:
2251 return folio;
2252}
2253EXPORT_SYMBOL(vma_alloc_folio);
2254
2255/**
2256 * alloc_pages - Allocate pages.
2257 * @gfp: GFP flags.
2258 * @order: Power of two of number of pages to allocate.
2259 *
2260 * Allocate 1 << @order contiguous pages. The physical address of the
2261 * first page is naturally aligned (eg an order-3 allocation will be aligned
2262 * to a multiple of 8 * PAGE_SIZE bytes). The NUMA policy of the current
2263 * process is honoured when in process context.
2264 *
2265 * Context: Can be called from any context, providing the appropriate GFP
2266 * flags are used.
2267 * Return: The page on success or NULL if allocation fails.
2268 */
2269struct page *alloc_pages(gfp_t gfp, unsigned order)
2270{
2271 struct mempolicy *pol = &default_policy;
2272 struct page *page;
2273
2274 if (!in_interrupt() && !(gfp & __GFP_THISNODE))
2275 pol = get_task_policy(current);
2276
2277 /*
2278 * No reference counting needed for current->mempolicy
2279 * nor system default_policy
2280 */
2281 if (pol->mode == MPOL_INTERLEAVE)
2282 page = alloc_page_interleave(gfp, order, interleave_nodes(pol));
2283 else if (pol->mode == MPOL_PREFERRED_MANY)
2284 page = alloc_pages_preferred_many(gfp, order,
2285 policy_node(gfp, pol, numa_node_id()), pol);
2286 else
2287 page = __alloc_pages(gfp, order,
2288 policy_node(gfp, pol, numa_node_id()),
2289 policy_nodemask(gfp, pol));
2290
2291 return page;
2292}
2293EXPORT_SYMBOL(alloc_pages);
2294
2295struct folio *folio_alloc(gfp_t gfp, unsigned order)
2296{
2297 struct page *page = alloc_pages(gfp | __GFP_COMP, order);
2298
2299 if (page && order > 1)
2300 prep_transhuge_page(page);
2301 return (struct folio *)page;
2302}
2303EXPORT_SYMBOL(folio_alloc);
2304
2305static unsigned long alloc_pages_bulk_array_interleave(gfp_t gfp,
2306 struct mempolicy *pol, unsigned long nr_pages,
2307 struct page **page_array)
2308{
2309 int nodes;
2310 unsigned long nr_pages_per_node;
2311 int delta;
2312 int i;
2313 unsigned long nr_allocated;
2314 unsigned long total_allocated = 0;
2315
2316 nodes = nodes_weight(pol->nodes);
2317 nr_pages_per_node = nr_pages / nodes;
2318 delta = nr_pages - nodes * nr_pages_per_node;
2319
2320 for (i = 0; i < nodes; i++) {
2321 if (delta) {
2322 nr_allocated = __alloc_pages_bulk(gfp,
2323 interleave_nodes(pol), NULL,
2324 nr_pages_per_node + 1, NULL,
2325 page_array);
2326 delta--;
2327 } else {
2328 nr_allocated = __alloc_pages_bulk(gfp,
2329 interleave_nodes(pol), NULL,
2330 nr_pages_per_node, NULL, page_array);
2331 }
2332
2333 page_array += nr_allocated;
2334 total_allocated += nr_allocated;
2335 }
2336
2337 return total_allocated;
2338}
2339
2340static unsigned long alloc_pages_bulk_array_preferred_many(gfp_t gfp, int nid,
2341 struct mempolicy *pol, unsigned long nr_pages,
2342 struct page **page_array)
2343{
2344 gfp_t preferred_gfp;
2345 unsigned long nr_allocated = 0;
2346
2347 preferred_gfp = gfp | __GFP_NOWARN;
2348 preferred_gfp &= ~(__GFP_DIRECT_RECLAIM | __GFP_NOFAIL);
2349
2350 nr_allocated = __alloc_pages_bulk(preferred_gfp, nid, &pol->nodes,
2351 nr_pages, NULL, page_array);
2352
2353 if (nr_allocated < nr_pages)
2354 nr_allocated += __alloc_pages_bulk(gfp, numa_node_id(), NULL,
2355 nr_pages - nr_allocated, NULL,
2356 page_array + nr_allocated);
2357 return nr_allocated;
2358}
2359
2360/* alloc pages bulk and mempolicy should be considered at the
2361 * same time in some situation such as vmalloc.
2362 *
2363 * It can accelerate memory allocation especially interleaving
2364 * allocate memory.
2365 */
2366unsigned long alloc_pages_bulk_array_mempolicy(gfp_t gfp,
2367 unsigned long nr_pages, struct page **page_array)
2368{
2369 struct mempolicy *pol = &default_policy;
2370
2371 if (!in_interrupt() && !(gfp & __GFP_THISNODE))
2372 pol = get_task_policy(current);
2373
2374 if (pol->mode == MPOL_INTERLEAVE)
2375 return alloc_pages_bulk_array_interleave(gfp, pol,
2376 nr_pages, page_array);
2377
2378 if (pol->mode == MPOL_PREFERRED_MANY)
2379 return alloc_pages_bulk_array_preferred_many(gfp,
2380 numa_node_id(), pol, nr_pages, page_array);
2381
2382 return __alloc_pages_bulk(gfp, policy_node(gfp, pol, numa_node_id()),
2383 policy_nodemask(gfp, pol), nr_pages, NULL,
2384 page_array);
2385}
2386
2387int vma_dup_policy(struct vm_area_struct *src, struct vm_area_struct *dst)
2388{
2389 struct mempolicy *pol = mpol_dup(vma_policy(src));
2390
2391 if (IS_ERR(pol))
2392 return PTR_ERR(pol);
2393 dst->vm_policy = pol;
2394 return 0;
2395}
2396
2397/*
2398 * If mpol_dup() sees current->cpuset == cpuset_being_rebound, then it
2399 * rebinds the mempolicy its copying by calling mpol_rebind_policy()
2400 * with the mems_allowed returned by cpuset_mems_allowed(). This
2401 * keeps mempolicies cpuset relative after its cpuset moves. See
2402 * further kernel/cpuset.c update_nodemask().
2403 *
2404 * current's mempolicy may be rebinded by the other task(the task that changes
2405 * cpuset's mems), so we needn't do rebind work for current task.
2406 */
2407
2408/* Slow path of a mempolicy duplicate */
2409struct mempolicy *__mpol_dup(struct mempolicy *old)
2410{
2411 struct mempolicy *new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
2412
2413 if (!new)
2414 return ERR_PTR(-ENOMEM);
2415
2416 /* task's mempolicy is protected by alloc_lock */
2417 if (old == current->mempolicy) {
2418 task_lock(current);
2419 *new = *old;
2420 task_unlock(current);
2421 } else
2422 *new = *old;
2423
2424 if (current_cpuset_is_being_rebound()) {
2425 nodemask_t mems = cpuset_mems_allowed(current);
2426 mpol_rebind_policy(new, &mems);
2427 }
2428 atomic_set(&new->refcnt, 1);
2429 return new;
2430}
2431
2432/* Slow path of a mempolicy comparison */
2433bool __mpol_equal(struct mempolicy *a, struct mempolicy *b)
2434{
2435 if (!a || !b)
2436 return false;
2437 if (a->mode != b->mode)
2438 return false;
2439 if (a->flags != b->flags)
2440 return false;
2441 if (a->home_node != b->home_node)
2442 return false;
2443 if (mpol_store_user_nodemask(a))
2444 if (!nodes_equal(a->w.user_nodemask, b->w.user_nodemask))
2445 return false;
2446
2447 switch (a->mode) {
2448 case MPOL_BIND:
2449 case MPOL_INTERLEAVE:
2450 case MPOL_PREFERRED:
2451 case MPOL_PREFERRED_MANY:
2452 return !!nodes_equal(a->nodes, b->nodes);
2453 case MPOL_LOCAL:
2454 return true;
2455 default:
2456 BUG();
2457 return false;
2458 }
2459}
2460
2461/*
2462 * Shared memory backing store policy support.
2463 *
2464 * Remember policies even when nobody has shared memory mapped.
2465 * The policies are kept in Red-Black tree linked from the inode.
2466 * They are protected by the sp->lock rwlock, which should be held
2467 * for any accesses to the tree.
2468 */
2469
2470/*
2471 * lookup first element intersecting start-end. Caller holds sp->lock for
2472 * reading or for writing
2473 */
2474static struct sp_node *
2475sp_lookup(struct shared_policy *sp, unsigned long start, unsigned long end)
2476{
2477 struct rb_node *n = sp->root.rb_node;
2478
2479 while (n) {
2480 struct sp_node *p = rb_entry(n, struct sp_node, nd);
2481
2482 if (start >= p->end)
2483 n = n->rb_right;
2484 else if (end <= p->start)
2485 n = n->rb_left;
2486 else
2487 break;
2488 }
2489 if (!n)
2490 return NULL;
2491 for (;;) {
2492 struct sp_node *w = NULL;
2493 struct rb_node *prev = rb_prev(n);
2494 if (!prev)
2495 break;
2496 w = rb_entry(prev, struct sp_node, nd);
2497 if (w->end <= start)
2498 break;
2499 n = prev;
2500 }
2501 return rb_entry(n, struct sp_node, nd);
2502}
2503
2504/*
2505 * Insert a new shared policy into the list. Caller holds sp->lock for
2506 * writing.
2507 */
2508static void sp_insert(struct shared_policy *sp, struct sp_node *new)
2509{
2510 struct rb_node **p = &sp->root.rb_node;
2511 struct rb_node *parent = NULL;
2512 struct sp_node *nd;
2513
2514 while (*p) {
2515 parent = *p;
2516 nd = rb_entry(parent, struct sp_node, nd);
2517 if (new->start < nd->start)
2518 p = &(*p)->rb_left;
2519 else if (new->end > nd->end)
2520 p = &(*p)->rb_right;
2521 else
2522 BUG();
2523 }
2524 rb_link_node(&new->nd, parent, p);
2525 rb_insert_color(&new->nd, &sp->root);
2526 pr_debug("inserting %lx-%lx: %d\n", new->start, new->end,
2527 new->policy ? new->policy->mode : 0);
2528}
2529
2530/* Find shared policy intersecting idx */
2531struct mempolicy *
2532mpol_shared_policy_lookup(struct shared_policy *sp, unsigned long idx)
2533{
2534 struct mempolicy *pol = NULL;
2535 struct sp_node *sn;
2536
2537 if (!sp->root.rb_node)
2538 return NULL;
2539 read_lock(&sp->lock);
2540 sn = sp_lookup(sp, idx, idx+1);
2541 if (sn) {
2542 mpol_get(sn->policy);
2543 pol = sn->policy;
2544 }
2545 read_unlock(&sp->lock);
2546 return pol;
2547}
2548
2549static void sp_free(struct sp_node *n)
2550{
2551 mpol_put(n->policy);
2552 kmem_cache_free(sn_cache, n);
2553}
2554
2555/**
2556 * mpol_misplaced - check whether current page node is valid in policy
2557 *
2558 * @page: page to be checked
2559 * @vma: vm area where page mapped
2560 * @addr: virtual address where page mapped
2561 *
2562 * Lookup current policy node id for vma,addr and "compare to" page's
2563 * node id. Policy determination "mimics" alloc_page_vma().
2564 * Called from fault path where we know the vma and faulting address.
2565 *
2566 * Return: NUMA_NO_NODE if the page is in a node that is valid for this
2567 * policy, or a suitable node ID to allocate a replacement page from.
2568 */
2569int mpol_misplaced(struct page *page, struct vm_area_struct *vma, unsigned long addr)
2570{
2571 struct mempolicy *pol;
2572 struct zoneref *z;
2573 int curnid = page_to_nid(page);
2574 unsigned long pgoff;
2575 int thiscpu = raw_smp_processor_id();
2576 int thisnid = cpu_to_node(thiscpu);
2577 int polnid = NUMA_NO_NODE;
2578 int ret = NUMA_NO_NODE;
2579
2580 pol = get_vma_policy(vma, addr);
2581 if (!(pol->flags & MPOL_F_MOF))
2582 goto out;
2583
2584 switch (pol->mode) {
2585 case MPOL_INTERLEAVE:
2586 pgoff = vma->vm_pgoff;
2587 pgoff += (addr - vma->vm_start) >> PAGE_SHIFT;
2588 polnid = offset_il_node(pol, pgoff);
2589 break;
2590
2591 case MPOL_PREFERRED:
2592 if (node_isset(curnid, pol->nodes))
2593 goto out;
2594 polnid = first_node(pol->nodes);
2595 break;
2596
2597 case MPOL_LOCAL:
2598 polnid = numa_node_id();
2599 break;
2600
2601 case MPOL_BIND:
2602 /* Optimize placement among multiple nodes via NUMA balancing */
2603 if (pol->flags & MPOL_F_MORON) {
2604 if (node_isset(thisnid, pol->nodes))
2605 break;
2606 goto out;
2607 }
2608 fallthrough;
2609
2610 case MPOL_PREFERRED_MANY:
2611 /*
2612 * use current page if in policy nodemask,
2613 * else select nearest allowed node, if any.
2614 * If no allowed nodes, use current [!misplaced].
2615 */
2616 if (node_isset(curnid, pol->nodes))
2617 goto out;
2618 z = first_zones_zonelist(
2619 node_zonelist(numa_node_id(), GFP_HIGHUSER),
2620 gfp_zone(GFP_HIGHUSER),
2621 &pol->nodes);
2622 polnid = zone_to_nid(z->zone);
2623 break;
2624
2625 default:
2626 BUG();
2627 }
2628
2629 /* Migrate the page towards the node whose CPU is referencing it */
2630 if (pol->flags & MPOL_F_MORON) {
2631 polnid = thisnid;
2632
2633 if (!should_numa_migrate_memory(current, page, curnid, thiscpu))
2634 goto out;
2635 }
2636
2637 if (curnid != polnid)
2638 ret = polnid;
2639out:
2640 mpol_cond_put(pol);
2641
2642 return ret;
2643}
2644
2645/*
2646 * Drop the (possibly final) reference to task->mempolicy. It needs to be
2647 * dropped after task->mempolicy is set to NULL so that any allocation done as
2648 * part of its kmem_cache_free(), such as by KASAN, doesn't reference a freed
2649 * policy.
2650 */
2651void mpol_put_task_policy(struct task_struct *task)
2652{
2653 struct mempolicy *pol;
2654
2655 task_lock(task);
2656 pol = task->mempolicy;
2657 task->mempolicy = NULL;
2658 task_unlock(task);
2659 mpol_put(pol);
2660}
2661
2662static void sp_delete(struct shared_policy *sp, struct sp_node *n)
2663{
2664 pr_debug("deleting %lx-l%lx\n", n->start, n->end);
2665 rb_erase(&n->nd, &sp->root);
2666 sp_free(n);
2667}
2668
2669static void sp_node_init(struct sp_node *node, unsigned long start,
2670 unsigned long end, struct mempolicy *pol)
2671{
2672 node->start = start;
2673 node->end = end;
2674 node->policy = pol;
2675}
2676
2677static struct sp_node *sp_alloc(unsigned long start, unsigned long end,
2678 struct mempolicy *pol)
2679{
2680 struct sp_node *n;
2681 struct mempolicy *newpol;
2682
2683 n = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2684 if (!n)
2685 return NULL;
2686
2687 newpol = mpol_dup(pol);
2688 if (IS_ERR(newpol)) {
2689 kmem_cache_free(sn_cache, n);
2690 return NULL;
2691 }
2692 newpol->flags |= MPOL_F_SHARED;
2693 sp_node_init(n, start, end, newpol);
2694
2695 return n;
2696}
2697
2698/* Replace a policy range. */
2699static int shared_policy_replace(struct shared_policy *sp, unsigned long start,
2700 unsigned long end, struct sp_node *new)
2701{
2702 struct sp_node *n;
2703 struct sp_node *n_new = NULL;
2704 struct mempolicy *mpol_new = NULL;
2705 int ret = 0;
2706
2707restart:
2708 write_lock(&sp->lock);
2709 n = sp_lookup(sp, start, end);
2710 /* Take care of old policies in the same range. */
2711 while (n && n->start < end) {
2712 struct rb_node *next = rb_next(&n->nd);
2713 if (n->start >= start) {
2714 if (n->end <= end)
2715 sp_delete(sp, n);
2716 else
2717 n->start = end;
2718 } else {
2719 /* Old policy spanning whole new range. */
2720 if (n->end > end) {
2721 if (!n_new)
2722 goto alloc_new;
2723
2724 *mpol_new = *n->policy;
2725 atomic_set(&mpol_new->refcnt, 1);
2726 sp_node_init(n_new, end, n->end, mpol_new);
2727 n->end = start;
2728 sp_insert(sp, n_new);
2729 n_new = NULL;
2730 mpol_new = NULL;
2731 break;
2732 } else
2733 n->end = start;
2734 }
2735 if (!next)
2736 break;
2737 n = rb_entry(next, struct sp_node, nd);
2738 }
2739 if (new)
2740 sp_insert(sp, new);
2741 write_unlock(&sp->lock);
2742 ret = 0;
2743
2744err_out:
2745 if (mpol_new)
2746 mpol_put(mpol_new);
2747 if (n_new)
2748 kmem_cache_free(sn_cache, n_new);
2749
2750 return ret;
2751
2752alloc_new:
2753 write_unlock(&sp->lock);
2754 ret = -ENOMEM;
2755 n_new = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2756 if (!n_new)
2757 goto err_out;
2758 mpol_new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
2759 if (!mpol_new)
2760 goto err_out;
2761 atomic_set(&mpol_new->refcnt, 1);
2762 goto restart;
2763}
2764
2765/**
2766 * mpol_shared_policy_init - initialize shared policy for inode
2767 * @sp: pointer to inode shared policy
2768 * @mpol: struct mempolicy to install
2769 *
2770 * Install non-NULL @mpol in inode's shared policy rb-tree.
2771 * On entry, the current task has a reference on a non-NULL @mpol.
2772 * This must be released on exit.
2773 * This is called at get_inode() calls and we can use GFP_KERNEL.
2774 */
2775void mpol_shared_policy_init(struct shared_policy *sp, struct mempolicy *mpol)
2776{
2777 int ret;
2778
2779 sp->root = RB_ROOT; /* empty tree == default mempolicy */
2780 rwlock_init(&sp->lock);
2781
2782 if (mpol) {
2783 struct vm_area_struct pvma;
2784 struct mempolicy *new;
2785 NODEMASK_SCRATCH(scratch);
2786
2787 if (!scratch)
2788 goto put_mpol;
2789 /* contextualize the tmpfs mount point mempolicy */
2790 new = mpol_new(mpol->mode, mpol->flags, &mpol->w.user_nodemask);
2791 if (IS_ERR(new))
2792 goto free_scratch; /* no valid nodemask intersection */
2793
2794 task_lock(current);
2795 ret = mpol_set_nodemask(new, &mpol->w.user_nodemask, scratch);
2796 task_unlock(current);
2797 if (ret)
2798 goto put_new;
2799
2800 /* Create pseudo-vma that contains just the policy */
2801 vma_init(&pvma, NULL);
2802 pvma.vm_end = TASK_SIZE; /* policy covers entire file */
2803 mpol_set_shared_policy(sp, &pvma, new); /* adds ref */
2804
2805put_new:
2806 mpol_put(new); /* drop initial ref */
2807free_scratch:
2808 NODEMASK_SCRATCH_FREE(scratch);
2809put_mpol:
2810 mpol_put(mpol); /* drop our incoming ref on sb mpol */
2811 }
2812}
2813
2814int mpol_set_shared_policy(struct shared_policy *info,
2815 struct vm_area_struct *vma, struct mempolicy *npol)
2816{
2817 int err;
2818 struct sp_node *new = NULL;
2819 unsigned long sz = vma_pages(vma);
2820
2821 pr_debug("set_shared_policy %lx sz %lu %d %d %lx\n",
2822 vma->vm_pgoff,
2823 sz, npol ? npol->mode : -1,
2824 npol ? npol->flags : -1,
2825 npol ? nodes_addr(npol->nodes)[0] : NUMA_NO_NODE);
2826
2827 if (npol) {
2828 new = sp_alloc(vma->vm_pgoff, vma->vm_pgoff + sz, npol);
2829 if (!new)
2830 return -ENOMEM;
2831 }
2832 err = shared_policy_replace(info, vma->vm_pgoff, vma->vm_pgoff+sz, new);
2833 if (err && new)
2834 sp_free(new);
2835 return err;
2836}
2837
2838/* Free a backing policy store on inode delete. */
2839void mpol_free_shared_policy(struct shared_policy *p)
2840{
2841 struct sp_node *n;
2842 struct rb_node *next;
2843
2844 if (!p->root.rb_node)
2845 return;
2846 write_lock(&p->lock);
2847 next = rb_first(&p->root);
2848 while (next) {
2849 n = rb_entry(next, struct sp_node, nd);
2850 next = rb_next(&n->nd);
2851 sp_delete(p, n);
2852 }
2853 write_unlock(&p->lock);
2854}
2855
2856#ifdef CONFIG_NUMA_BALANCING
2857static int __initdata numabalancing_override;
2858
2859static void __init check_numabalancing_enable(void)
2860{
2861 bool numabalancing_default = false;
2862
2863 if (IS_ENABLED(CONFIG_NUMA_BALANCING_DEFAULT_ENABLED))
2864 numabalancing_default = true;
2865
2866 /* Parsed by setup_numabalancing. override == 1 enables, -1 disables */
2867 if (numabalancing_override)
2868 set_numabalancing_state(numabalancing_override == 1);
2869
2870 if (num_online_nodes() > 1 && !numabalancing_override) {
2871 pr_info("%s automatic NUMA balancing. Configure with numa_balancing= or the kernel.numa_balancing sysctl\n",
2872 numabalancing_default ? "Enabling" : "Disabling");
2873 set_numabalancing_state(numabalancing_default);
2874 }
2875}
2876
2877static int __init setup_numabalancing(char *str)
2878{
2879 int ret = 0;
2880 if (!str)
2881 goto out;
2882
2883 if (!strcmp(str, "enable")) {
2884 numabalancing_override = 1;
2885 ret = 1;
2886 } else if (!strcmp(str, "disable")) {
2887 numabalancing_override = -1;
2888 ret = 1;
2889 }
2890out:
2891 if (!ret)
2892 pr_warn("Unable to parse numa_balancing=\n");
2893
2894 return ret;
2895}
2896__setup("numa_balancing=", setup_numabalancing);
2897#else
2898static inline void __init check_numabalancing_enable(void)
2899{
2900}
2901#endif /* CONFIG_NUMA_BALANCING */
2902
2903/* assumes fs == KERNEL_DS */
2904void __init numa_policy_init(void)
2905{
2906 nodemask_t interleave_nodes;
2907 unsigned long largest = 0;
2908 int nid, prefer = 0;
2909
2910 policy_cache = kmem_cache_create("numa_policy",
2911 sizeof(struct mempolicy),
2912 0, SLAB_PANIC, NULL);
2913
2914 sn_cache = kmem_cache_create("shared_policy_node",
2915 sizeof(struct sp_node),
2916 0, SLAB_PANIC, NULL);
2917
2918 for_each_node(nid) {
2919 preferred_node_policy[nid] = (struct mempolicy) {
2920 .refcnt = ATOMIC_INIT(1),
2921 .mode = MPOL_PREFERRED,
2922 .flags = MPOL_F_MOF | MPOL_F_MORON,
2923 .nodes = nodemask_of_node(nid),
2924 };
2925 }
2926
2927 /*
2928 * Set interleaving policy for system init. Interleaving is only
2929 * enabled across suitably sized nodes (default is >= 16MB), or
2930 * fall back to the largest node if they're all smaller.
2931 */
2932 nodes_clear(interleave_nodes);
2933 for_each_node_state(nid, N_MEMORY) {
2934 unsigned long total_pages = node_present_pages(nid);
2935
2936 /* Preserve the largest node */
2937 if (largest < total_pages) {
2938 largest = total_pages;
2939 prefer = nid;
2940 }
2941
2942 /* Interleave this node? */
2943 if ((total_pages << PAGE_SHIFT) >= (16 << 20))
2944 node_set(nid, interleave_nodes);
2945 }
2946
2947 /* All too small, use the largest */
2948 if (unlikely(nodes_empty(interleave_nodes)))
2949 node_set(prefer, interleave_nodes);
2950
2951 if (do_set_mempolicy(MPOL_INTERLEAVE, 0, &interleave_nodes))
2952 pr_err("%s: interleaving failed\n", __func__);
2953
2954 check_numabalancing_enable();
2955}
2956
2957/* Reset policy of current process to default */
2958void numa_default_policy(void)
2959{
2960 do_set_mempolicy(MPOL_DEFAULT, 0, NULL);
2961}
2962
2963/*
2964 * Parse and format mempolicy from/to strings
2965 */
2966
2967static const char * const policy_modes[] =
2968{
2969 [MPOL_DEFAULT] = "default",
2970 [MPOL_PREFERRED] = "prefer",
2971 [MPOL_BIND] = "bind",
2972 [MPOL_INTERLEAVE] = "interleave",
2973 [MPOL_LOCAL] = "local",
2974 [MPOL_PREFERRED_MANY] = "prefer (many)",
2975};
2976
2977
2978#ifdef CONFIG_TMPFS
2979/**
2980 * mpol_parse_str - parse string to mempolicy, for tmpfs mpol mount option.
2981 * @str: string containing mempolicy to parse
2982 * @mpol: pointer to struct mempolicy pointer, returned on success.
2983 *
2984 * Format of input:
2985 * <mode>[=<flags>][:<nodelist>]
2986 *
2987 * Return: %0 on success, else %1
2988 */
2989int mpol_parse_str(char *str, struct mempolicy **mpol)
2990{
2991 struct mempolicy *new = NULL;
2992 unsigned short mode_flags;
2993 nodemask_t nodes;
2994 char *nodelist = strchr(str, ':');
2995 char *flags = strchr(str, '=');
2996 int err = 1, mode;
2997
2998 if (flags)
2999 *flags++ = '\0'; /* terminate mode string */
3000
3001 if (nodelist) {
3002 /* NUL-terminate mode or flags string */
3003 *nodelist++ = '\0';
3004 if (nodelist_parse(nodelist, nodes))
3005 goto out;
3006 if (!nodes_subset(nodes, node_states[N_MEMORY]))
3007 goto out;
3008 } else
3009 nodes_clear(nodes);
3010
3011 mode = match_string(policy_modes, MPOL_MAX, str);
3012 if (mode < 0)
3013 goto out;
3014
3015 switch (mode) {
3016 case MPOL_PREFERRED:
3017 /*
3018 * Insist on a nodelist of one node only, although later
3019 * we use first_node(nodes) to grab a single node, so here
3020 * nodelist (or nodes) cannot be empty.
3021 */
3022 if (nodelist) {
3023 char *rest = nodelist;
3024 while (isdigit(*rest))
3025 rest++;
3026 if (*rest)
3027 goto out;
3028 if (nodes_empty(nodes))
3029 goto out;
3030 }
3031 break;
3032 case MPOL_INTERLEAVE:
3033 /*
3034 * Default to online nodes with memory if no nodelist
3035 */
3036 if (!nodelist)
3037 nodes = node_states[N_MEMORY];
3038 break;
3039 case MPOL_LOCAL:
3040 /*
3041 * Don't allow a nodelist; mpol_new() checks flags
3042 */
3043 if (nodelist)
3044 goto out;
3045 break;
3046 case MPOL_DEFAULT:
3047 /*
3048 * Insist on a empty nodelist
3049 */
3050 if (!nodelist)
3051 err = 0;
3052 goto out;
3053 case MPOL_PREFERRED_MANY:
3054 case MPOL_BIND:
3055 /*
3056 * Insist on a nodelist
3057 */
3058 if (!nodelist)
3059 goto out;
3060 }
3061
3062 mode_flags = 0;
3063 if (flags) {
3064 /*
3065 * Currently, we only support two mutually exclusive
3066 * mode flags.
3067 */
3068 if (!strcmp(flags, "static"))
3069 mode_flags |= MPOL_F_STATIC_NODES;
3070 else if (!strcmp(flags, "relative"))
3071 mode_flags |= MPOL_F_RELATIVE_NODES;
3072 else
3073 goto out;
3074 }
3075
3076 new = mpol_new(mode, mode_flags, &nodes);
3077 if (IS_ERR(new))
3078 goto out;
3079
3080 /*
3081 * Save nodes for mpol_to_str() to show the tmpfs mount options
3082 * for /proc/mounts, /proc/pid/mounts and /proc/pid/mountinfo.
3083 */
3084 if (mode != MPOL_PREFERRED) {
3085 new->nodes = nodes;
3086 } else if (nodelist) {
3087 nodes_clear(new->nodes);
3088 node_set(first_node(nodes), new->nodes);
3089 } else {
3090 new->mode = MPOL_LOCAL;
3091 }
3092
3093 /*
3094 * Save nodes for contextualization: this will be used to "clone"
3095 * the mempolicy in a specific context [cpuset] at a later time.
3096 */
3097 new->w.user_nodemask = nodes;
3098
3099 err = 0;
3100
3101out:
3102 /* Restore string for error message */
3103 if (nodelist)
3104 *--nodelist = ':';
3105 if (flags)
3106 *--flags = '=';
3107 if (!err)
3108 *mpol = new;
3109 return err;
3110}
3111#endif /* CONFIG_TMPFS */
3112
3113/**
3114 * mpol_to_str - format a mempolicy structure for printing
3115 * @buffer: to contain formatted mempolicy string
3116 * @maxlen: length of @buffer
3117 * @pol: pointer to mempolicy to be formatted
3118 *
3119 * Convert @pol into a string. If @buffer is too short, truncate the string.
3120 * Recommend a @maxlen of at least 32 for the longest mode, "interleave", the
3121 * longest flag, "relative", and to display at least a few node ids.
3122 */
3123void mpol_to_str(char *buffer, int maxlen, struct mempolicy *pol)
3124{
3125 char *p = buffer;
3126 nodemask_t nodes = NODE_MASK_NONE;
3127 unsigned short mode = MPOL_DEFAULT;
3128 unsigned short flags = 0;
3129
3130 if (pol && pol != &default_policy && !(pol->flags & MPOL_F_MORON)) {
3131 mode = pol->mode;
3132 flags = pol->flags;
3133 }
3134
3135 switch (mode) {
3136 case MPOL_DEFAULT:
3137 case MPOL_LOCAL:
3138 break;
3139 case MPOL_PREFERRED:
3140 case MPOL_PREFERRED_MANY:
3141 case MPOL_BIND:
3142 case MPOL_INTERLEAVE:
3143 nodes = pol->nodes;
3144 break;
3145 default:
3146 WARN_ON_ONCE(1);
3147 snprintf(p, maxlen, "unknown");
3148 return;
3149 }
3150
3151 p += snprintf(p, maxlen, "%s", policy_modes[mode]);
3152
3153 if (flags & MPOL_MODE_FLAGS) {
3154 p += snprintf(p, buffer + maxlen - p, "=");
3155
3156 /*
3157 * Currently, the only defined flags are mutually exclusive
3158 */
3159 if (flags & MPOL_F_STATIC_NODES)
3160 p += snprintf(p, buffer + maxlen - p, "static");
3161 else if (flags & MPOL_F_RELATIVE_NODES)
3162 p += snprintf(p, buffer + maxlen - p, "relative");
3163 }
3164
3165 if (!nodes_empty(nodes))
3166 p += scnprintf(p, buffer + maxlen - p, ":%*pbl",
3167 nodemask_pr_args(&nodes));
3168}