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