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