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/*
2 * Simple NUMA memory policy for the Linux kernel.
3 *
4 * Copyright 2003,2004 Andi Kleen, SuSE Labs.
5 * (C) Copyright 2005 Christoph Lameter, Silicon Graphics, Inc.
6 * Subject to the GNU Public License, version 2.
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#include <linux/mempolicy.h>
69#include <linux/mm.h>
70#include <linux/highmem.h>
71#include <linux/hugetlb.h>
72#include <linux/kernel.h>
73#include <linux/sched.h>
74#include <linux/nodemask.h>
75#include <linux/cpuset.h>
76#include <linux/slab.h>
77#include <linux/string.h>
78#include <linux/export.h>
79#include <linux/nsproxy.h>
80#include <linux/interrupt.h>
81#include <linux/init.h>
82#include <linux/compat.h>
83#include <linux/swap.h>
84#include <linux/seq_file.h>
85#include <linux/proc_fs.h>
86#include <linux/migrate.h>
87#include <linux/ksm.h>
88#include <linux/rmap.h>
89#include <linux/security.h>
90#include <linux/syscalls.h>
91#include <linux/ctype.h>
92#include <linux/mm_inline.h>
93#include <linux/mmu_notifier.h>
94
95#include <asm/tlbflush.h>
96#include <asm/uaccess.h>
97#include <linux/random.h>
98
99#include "internal.h"
100
101/* Internal flags */
102#define MPOL_MF_DISCONTIG_OK (MPOL_MF_INTERNAL << 0) /* Skip checks for continuous vmas */
103#define MPOL_MF_INVERT (MPOL_MF_INTERNAL << 1) /* Invert check for nodemask */
104
105static struct kmem_cache *policy_cache;
106static struct kmem_cache *sn_cache;
107
108/* Highest zone. An specific allocation for a zone below that is not
109 policied. */
110enum zone_type policy_zone = 0;
111
112/*
113 * run-time system-wide default policy => local allocation
114 */
115static struct mempolicy default_policy = {
116 .refcnt = ATOMIC_INIT(1), /* never free it */
117 .mode = MPOL_PREFERRED,
118 .flags = MPOL_F_LOCAL,
119};
120
121static struct mempolicy preferred_node_policy[MAX_NUMNODES];
122
123static struct mempolicy *get_task_policy(struct task_struct *p)
124{
125 struct mempolicy *pol = p->mempolicy;
126
127 if (!pol) {
128 int node = numa_node_id();
129
130 if (node != NUMA_NO_NODE) {
131 pol = &preferred_node_policy[node];
132 /*
133 * preferred_node_policy is not initialised early in
134 * boot
135 */
136 if (!pol->mode)
137 pol = NULL;
138 }
139 }
140
141 return pol;
142}
143
144static const struct mempolicy_operations {
145 int (*create)(struct mempolicy *pol, const nodemask_t *nodes);
146 /*
147 * If read-side task has no lock to protect task->mempolicy, write-side
148 * task will rebind the task->mempolicy by two step. The first step is
149 * setting all the newly nodes, and the second step is cleaning all the
150 * disallowed nodes. In this way, we can avoid finding no node to alloc
151 * page.
152 * If we have a lock to protect task->mempolicy in read-side, we do
153 * rebind directly.
154 *
155 * step:
156 * MPOL_REBIND_ONCE - do rebind work at once
157 * MPOL_REBIND_STEP1 - set all the newly nodes
158 * MPOL_REBIND_STEP2 - clean all the disallowed nodes
159 */
160 void (*rebind)(struct mempolicy *pol, const nodemask_t *nodes,
161 enum mpol_rebind_step step);
162} mpol_ops[MPOL_MAX];
163
164/* Check that the nodemask contains at least one populated zone */
165static int is_valid_nodemask(const nodemask_t *nodemask)
166{
167 return nodes_intersects(*nodemask, node_states[N_MEMORY]);
168}
169
170static inline int mpol_store_user_nodemask(const struct mempolicy *pol)
171{
172 return pol->flags & MPOL_MODE_FLAGS;
173}
174
175static void mpol_relative_nodemask(nodemask_t *ret, const nodemask_t *orig,
176 const nodemask_t *rel)
177{
178 nodemask_t tmp;
179 nodes_fold(tmp, *orig, nodes_weight(*rel));
180 nodes_onto(*ret, tmp, *rel);
181}
182
183static int mpol_new_interleave(struct mempolicy *pol, const nodemask_t *nodes)
184{
185 if (nodes_empty(*nodes))
186 return -EINVAL;
187 pol->v.nodes = *nodes;
188 return 0;
189}
190
191static int mpol_new_preferred(struct mempolicy *pol, const nodemask_t *nodes)
192{
193 if (!nodes)
194 pol->flags |= MPOL_F_LOCAL; /* local allocation */
195 else if (nodes_empty(*nodes))
196 return -EINVAL; /* no allowed nodes */
197 else
198 pol->v.preferred_node = first_node(*nodes);
199 return 0;
200}
201
202static int mpol_new_bind(struct mempolicy *pol, const nodemask_t *nodes)
203{
204 if (!is_valid_nodemask(nodes))
205 return -EINVAL;
206 pol->v.nodes = *nodes;
207 return 0;
208}
209
210/*
211 * mpol_set_nodemask is called after mpol_new() to set up the nodemask, if
212 * any, for the new policy. mpol_new() has already validated the nodes
213 * parameter with respect to the policy mode and flags. But, we need to
214 * handle an empty nodemask with MPOL_PREFERRED here.
215 *
216 * Must be called holding task's alloc_lock to protect task's mems_allowed
217 * and mempolicy. May also be called holding the mmap_semaphore for write.
218 */
219static int mpol_set_nodemask(struct mempolicy *pol,
220 const nodemask_t *nodes, struct nodemask_scratch *nsc)
221{
222 int ret;
223
224 /* if mode is MPOL_DEFAULT, pol is NULL. This is right. */
225 if (pol == NULL)
226 return 0;
227 /* Check N_MEMORY */
228 nodes_and(nsc->mask1,
229 cpuset_current_mems_allowed, node_states[N_MEMORY]);
230
231 VM_BUG_ON(!nodes);
232 if (pol->mode == MPOL_PREFERRED && nodes_empty(*nodes))
233 nodes = NULL; /* explicit local allocation */
234 else {
235 if (pol->flags & MPOL_F_RELATIVE_NODES)
236 mpol_relative_nodemask(&nsc->mask2, nodes,&nsc->mask1);
237 else
238 nodes_and(nsc->mask2, *nodes, nsc->mask1);
239
240 if (mpol_store_user_nodemask(pol))
241 pol->w.user_nodemask = *nodes;
242 else
243 pol->w.cpuset_mems_allowed =
244 cpuset_current_mems_allowed;
245 }
246
247 if (nodes)
248 ret = mpol_ops[pol->mode].create(pol, &nsc->mask2);
249 else
250 ret = mpol_ops[pol->mode].create(pol, NULL);
251 return ret;
252}
253
254/*
255 * This function just creates a new policy, does some check and simple
256 * initialization. You must invoke mpol_set_nodemask() to set nodes.
257 */
258static struct mempolicy *mpol_new(unsigned short mode, unsigned short flags,
259 nodemask_t *nodes)
260{
261 struct mempolicy *policy;
262
263 pr_debug("setting mode %d flags %d nodes[0] %lx\n",
264 mode, flags, nodes ? nodes_addr(*nodes)[0] : NUMA_NO_NODE);
265
266 if (mode == MPOL_DEFAULT) {
267 if (nodes && !nodes_empty(*nodes))
268 return ERR_PTR(-EINVAL);
269 return NULL;
270 }
271 VM_BUG_ON(!nodes);
272
273 /*
274 * MPOL_PREFERRED cannot be used with MPOL_F_STATIC_NODES or
275 * MPOL_F_RELATIVE_NODES if the nodemask is empty (local allocation).
276 * All other modes require a valid pointer to a non-empty nodemask.
277 */
278 if (mode == MPOL_PREFERRED) {
279 if (nodes_empty(*nodes)) {
280 if (((flags & MPOL_F_STATIC_NODES) ||
281 (flags & MPOL_F_RELATIVE_NODES)))
282 return ERR_PTR(-EINVAL);
283 }
284 } else if (mode == MPOL_LOCAL) {
285 if (!nodes_empty(*nodes))
286 return ERR_PTR(-EINVAL);
287 mode = MPOL_PREFERRED;
288 } else if (nodes_empty(*nodes))
289 return ERR_PTR(-EINVAL);
290 policy = kmem_cache_alloc(policy_cache, GFP_KERNEL);
291 if (!policy)
292 return ERR_PTR(-ENOMEM);
293 atomic_set(&policy->refcnt, 1);
294 policy->mode = mode;
295 policy->flags = flags;
296
297 return policy;
298}
299
300/* Slow path of a mpol destructor. */
301void __mpol_put(struct mempolicy *p)
302{
303 if (!atomic_dec_and_test(&p->refcnt))
304 return;
305 kmem_cache_free(policy_cache, p);
306}
307
308static void mpol_rebind_default(struct mempolicy *pol, const nodemask_t *nodes,
309 enum mpol_rebind_step step)
310{
311}
312
313/*
314 * step:
315 * MPOL_REBIND_ONCE - do rebind work at once
316 * MPOL_REBIND_STEP1 - set all the newly nodes
317 * MPOL_REBIND_STEP2 - clean all the disallowed nodes
318 */
319static void mpol_rebind_nodemask(struct mempolicy *pol, const nodemask_t *nodes,
320 enum mpol_rebind_step step)
321{
322 nodemask_t tmp;
323
324 if (pol->flags & MPOL_F_STATIC_NODES)
325 nodes_and(tmp, pol->w.user_nodemask, *nodes);
326 else if (pol->flags & MPOL_F_RELATIVE_NODES)
327 mpol_relative_nodemask(&tmp, &pol->w.user_nodemask, nodes);
328 else {
329 /*
330 * if step == 1, we use ->w.cpuset_mems_allowed to cache the
331 * result
332 */
333 if (step == MPOL_REBIND_ONCE || step == MPOL_REBIND_STEP1) {
334 nodes_remap(tmp, pol->v.nodes,
335 pol->w.cpuset_mems_allowed, *nodes);
336 pol->w.cpuset_mems_allowed = step ? tmp : *nodes;
337 } else if (step == MPOL_REBIND_STEP2) {
338 tmp = pol->w.cpuset_mems_allowed;
339 pol->w.cpuset_mems_allowed = *nodes;
340 } else
341 BUG();
342 }
343
344 if (nodes_empty(tmp))
345 tmp = *nodes;
346
347 if (step == MPOL_REBIND_STEP1)
348 nodes_or(pol->v.nodes, pol->v.nodes, tmp);
349 else if (step == MPOL_REBIND_ONCE || step == MPOL_REBIND_STEP2)
350 pol->v.nodes = tmp;
351 else
352 BUG();
353
354 if (!node_isset(current->il_next, tmp)) {
355 current->il_next = next_node(current->il_next, tmp);
356 if (current->il_next >= MAX_NUMNODES)
357 current->il_next = first_node(tmp);
358 if (current->il_next >= MAX_NUMNODES)
359 current->il_next = numa_node_id();
360 }
361}
362
363static void mpol_rebind_preferred(struct mempolicy *pol,
364 const nodemask_t *nodes,
365 enum mpol_rebind_step step)
366{
367 nodemask_t tmp;
368
369 if (pol->flags & MPOL_F_STATIC_NODES) {
370 int node = first_node(pol->w.user_nodemask);
371
372 if (node_isset(node, *nodes)) {
373 pol->v.preferred_node = node;
374 pol->flags &= ~MPOL_F_LOCAL;
375 } else
376 pol->flags |= MPOL_F_LOCAL;
377 } else if (pol->flags & MPOL_F_RELATIVE_NODES) {
378 mpol_relative_nodemask(&tmp, &pol->w.user_nodemask, nodes);
379 pol->v.preferred_node = first_node(tmp);
380 } else if (!(pol->flags & MPOL_F_LOCAL)) {
381 pol->v.preferred_node = node_remap(pol->v.preferred_node,
382 pol->w.cpuset_mems_allowed,
383 *nodes);
384 pol->w.cpuset_mems_allowed = *nodes;
385 }
386}
387
388/*
389 * mpol_rebind_policy - Migrate a policy to a different set of nodes
390 *
391 * If read-side task has no lock to protect task->mempolicy, write-side
392 * task will rebind the task->mempolicy by two step. The first step is
393 * setting all the newly nodes, and the second step is cleaning all the
394 * disallowed nodes. In this way, we can avoid finding no node to alloc
395 * page.
396 * If we have a lock to protect task->mempolicy in read-side, we do
397 * rebind directly.
398 *
399 * step:
400 * MPOL_REBIND_ONCE - do rebind work at once
401 * MPOL_REBIND_STEP1 - set all the newly nodes
402 * MPOL_REBIND_STEP2 - clean all the disallowed nodes
403 */
404static void mpol_rebind_policy(struct mempolicy *pol, const nodemask_t *newmask,
405 enum mpol_rebind_step step)
406{
407 if (!pol)
408 return;
409 if (!mpol_store_user_nodemask(pol) && step == MPOL_REBIND_ONCE &&
410 nodes_equal(pol->w.cpuset_mems_allowed, *newmask))
411 return;
412
413 if (step == MPOL_REBIND_STEP1 && (pol->flags & MPOL_F_REBINDING))
414 return;
415
416 if (step == MPOL_REBIND_STEP2 && !(pol->flags & MPOL_F_REBINDING))
417 BUG();
418
419 if (step == MPOL_REBIND_STEP1)
420 pol->flags |= MPOL_F_REBINDING;
421 else if (step == MPOL_REBIND_STEP2)
422 pol->flags &= ~MPOL_F_REBINDING;
423 else if (step >= MPOL_REBIND_NSTEP)
424 BUG();
425
426 mpol_ops[pol->mode].rebind(pol, newmask, step);
427}
428
429/*
430 * Wrapper for mpol_rebind_policy() that just requires task
431 * pointer, and updates task mempolicy.
432 *
433 * Called with task's alloc_lock held.
434 */
435
436void mpol_rebind_task(struct task_struct *tsk, const nodemask_t *new,
437 enum mpol_rebind_step step)
438{
439 mpol_rebind_policy(tsk->mempolicy, new, step);
440}
441
442/*
443 * Rebind each vma in mm to new nodemask.
444 *
445 * Call holding a reference to mm. Takes mm->mmap_sem during call.
446 */
447
448void mpol_rebind_mm(struct mm_struct *mm, nodemask_t *new)
449{
450 struct vm_area_struct *vma;
451
452 down_write(&mm->mmap_sem);
453 for (vma = mm->mmap; vma; vma = vma->vm_next)
454 mpol_rebind_policy(vma->vm_policy, new, MPOL_REBIND_ONCE);
455 up_write(&mm->mmap_sem);
456}
457
458static const struct mempolicy_operations mpol_ops[MPOL_MAX] = {
459 [MPOL_DEFAULT] = {
460 .rebind = mpol_rebind_default,
461 },
462 [MPOL_INTERLEAVE] = {
463 .create = mpol_new_interleave,
464 .rebind = mpol_rebind_nodemask,
465 },
466 [MPOL_PREFERRED] = {
467 .create = mpol_new_preferred,
468 .rebind = mpol_rebind_preferred,
469 },
470 [MPOL_BIND] = {
471 .create = mpol_new_bind,
472 .rebind = mpol_rebind_nodemask,
473 },
474};
475
476static void migrate_page_add(struct page *page, struct list_head *pagelist,
477 unsigned long flags);
478
479/*
480 * Scan through pages checking if pages follow certain conditions,
481 * and move them to the pagelist if they do.
482 */
483static int queue_pages_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
484 unsigned long addr, unsigned long end,
485 const nodemask_t *nodes, unsigned long flags,
486 void *private)
487{
488 pte_t *orig_pte;
489 pte_t *pte;
490 spinlock_t *ptl;
491
492 orig_pte = pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
493 do {
494 struct page *page;
495 int nid;
496
497 if (!pte_present(*pte))
498 continue;
499 page = vm_normal_page(vma, addr, *pte);
500 if (!page)
501 continue;
502 /*
503 * vm_normal_page() filters out zero pages, but there might
504 * still be PageReserved pages to skip, perhaps in a VDSO.
505 */
506 if (PageReserved(page))
507 continue;
508 nid = page_to_nid(page);
509 if (node_isset(nid, *nodes) == !!(flags & MPOL_MF_INVERT))
510 continue;
511
512 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))
513 migrate_page_add(page, private, flags);
514 else
515 break;
516 } while (pte++, addr += PAGE_SIZE, addr != end);
517 pte_unmap_unlock(orig_pte, ptl);
518 return addr != end;
519}
520
521static void queue_pages_hugetlb_pmd_range(struct vm_area_struct *vma,
522 pmd_t *pmd, const nodemask_t *nodes, unsigned long flags,
523 void *private)
524{
525#ifdef CONFIG_HUGETLB_PAGE
526 int nid;
527 struct page *page;
528 spinlock_t *ptl;
529 pte_t entry;
530
531 ptl = huge_pte_lock(hstate_vma(vma), vma->vm_mm, (pte_t *)pmd);
532 entry = huge_ptep_get((pte_t *)pmd);
533 if (!pte_present(entry))
534 goto unlock;
535 page = pte_page(entry);
536 nid = page_to_nid(page);
537 if (node_isset(nid, *nodes) == !!(flags & MPOL_MF_INVERT))
538 goto unlock;
539 /* With MPOL_MF_MOVE, we migrate only unshared hugepage. */
540 if (flags & (MPOL_MF_MOVE_ALL) ||
541 (flags & MPOL_MF_MOVE && page_mapcount(page) == 1))
542 isolate_huge_page(page, private);
543unlock:
544 spin_unlock(ptl);
545#else
546 BUG();
547#endif
548}
549
550static inline int queue_pages_pmd_range(struct vm_area_struct *vma, pud_t *pud,
551 unsigned long addr, unsigned long end,
552 const nodemask_t *nodes, unsigned long flags,
553 void *private)
554{
555 pmd_t *pmd;
556 unsigned long next;
557
558 pmd = pmd_offset(pud, addr);
559 do {
560 next = pmd_addr_end(addr, end);
561 if (!pmd_present(*pmd))
562 continue;
563 if (pmd_huge(*pmd) && is_vm_hugetlb_page(vma)) {
564 queue_pages_hugetlb_pmd_range(vma, pmd, nodes,
565 flags, private);
566 continue;
567 }
568 split_huge_page_pmd(vma, addr, pmd);
569 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
570 continue;
571 if (queue_pages_pte_range(vma, pmd, addr, next, nodes,
572 flags, private))
573 return -EIO;
574 } while (pmd++, addr = next, addr != end);
575 return 0;
576}
577
578static inline int queue_pages_pud_range(struct vm_area_struct *vma, pgd_t *pgd,
579 unsigned long addr, unsigned long end,
580 const nodemask_t *nodes, unsigned long flags,
581 void *private)
582{
583 pud_t *pud;
584 unsigned long next;
585
586 pud = pud_offset(pgd, addr);
587 do {
588 next = pud_addr_end(addr, end);
589 if (pud_huge(*pud) && is_vm_hugetlb_page(vma))
590 continue;
591 if (pud_none_or_clear_bad(pud))
592 continue;
593 if (queue_pages_pmd_range(vma, pud, addr, next, nodes,
594 flags, private))
595 return -EIO;
596 } while (pud++, addr = next, addr != end);
597 return 0;
598}
599
600static inline int queue_pages_pgd_range(struct vm_area_struct *vma,
601 unsigned long addr, unsigned long end,
602 const nodemask_t *nodes, unsigned long flags,
603 void *private)
604{
605 pgd_t *pgd;
606 unsigned long next;
607
608 pgd = pgd_offset(vma->vm_mm, addr);
609 do {
610 next = pgd_addr_end(addr, end);
611 if (pgd_none_or_clear_bad(pgd))
612 continue;
613 if (queue_pages_pud_range(vma, pgd, addr, next, nodes,
614 flags, private))
615 return -EIO;
616 } while (pgd++, addr = next, addr != end);
617 return 0;
618}
619
620#ifdef CONFIG_NUMA_BALANCING
621/*
622 * This is used to mark a range of virtual addresses to be inaccessible.
623 * These are later cleared by a NUMA hinting fault. Depending on these
624 * faults, pages may be migrated for better NUMA placement.
625 *
626 * This is assuming that NUMA faults are handled using PROT_NONE. If
627 * an architecture makes a different choice, it will need further
628 * changes to the core.
629 */
630unsigned long change_prot_numa(struct vm_area_struct *vma,
631 unsigned long addr, unsigned long end)
632{
633 int nr_updated;
634
635 nr_updated = change_protection(vma, addr, end, vma->vm_page_prot, 0, 1);
636 if (nr_updated)
637 count_vm_numa_events(NUMA_PTE_UPDATES, nr_updated);
638
639 return nr_updated;
640}
641#else
642static unsigned long change_prot_numa(struct vm_area_struct *vma,
643 unsigned long addr, unsigned long end)
644{
645 return 0;
646}
647#endif /* CONFIG_NUMA_BALANCING */
648
649/*
650 * Walk through page tables and collect pages to be migrated.
651 *
652 * If pages found in a given range are on a set of nodes (determined by
653 * @nodes and @flags,) it's isolated and queued to the pagelist which is
654 * passed via @private.)
655 */
656static struct vm_area_struct *
657queue_pages_range(struct mm_struct *mm, unsigned long start, unsigned long end,
658 const nodemask_t *nodes, unsigned long flags, void *private)
659{
660 int err;
661 struct vm_area_struct *first, *vma, *prev;
662
663
664 first = find_vma(mm, start);
665 if (!first)
666 return ERR_PTR(-EFAULT);
667 prev = NULL;
668 for (vma = first; vma && vma->vm_start < end; vma = vma->vm_next) {
669 unsigned long endvma = vma->vm_end;
670
671 if (endvma > end)
672 endvma = end;
673 if (vma->vm_start > start)
674 start = vma->vm_start;
675
676 if (!(flags & MPOL_MF_DISCONTIG_OK)) {
677 if (!vma->vm_next && vma->vm_end < end)
678 return ERR_PTR(-EFAULT);
679 if (prev && prev->vm_end < vma->vm_start)
680 return ERR_PTR(-EFAULT);
681 }
682
683 if (flags & MPOL_MF_LAZY) {
684 change_prot_numa(vma, start, endvma);
685 goto next;
686 }
687
688 if ((flags & MPOL_MF_STRICT) ||
689 ((flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) &&
690 vma_migratable(vma))) {
691
692 err = queue_pages_pgd_range(vma, start, endvma, nodes,
693 flags, private);
694 if (err) {
695 first = ERR_PTR(err);
696 break;
697 }
698 }
699next:
700 prev = vma;
701 }
702 return first;
703}
704
705/*
706 * Apply policy to a single VMA
707 * This must be called with the mmap_sem held for writing.
708 */
709static int vma_replace_policy(struct vm_area_struct *vma,
710 struct mempolicy *pol)
711{
712 int err;
713 struct mempolicy *old;
714 struct mempolicy *new;
715
716 pr_debug("vma %lx-%lx/%lx vm_ops %p vm_file %p set_policy %p\n",
717 vma->vm_start, vma->vm_end, vma->vm_pgoff,
718 vma->vm_ops, vma->vm_file,
719 vma->vm_ops ? vma->vm_ops->set_policy : NULL);
720
721 new = mpol_dup(pol);
722 if (IS_ERR(new))
723 return PTR_ERR(new);
724
725 if (vma->vm_ops && vma->vm_ops->set_policy) {
726 err = vma->vm_ops->set_policy(vma, new);
727 if (err)
728 goto err_out;
729 }
730
731 old = vma->vm_policy;
732 vma->vm_policy = new; /* protected by mmap_sem */
733 mpol_put(old);
734
735 return 0;
736 err_out:
737 mpol_put(new);
738 return err;
739}
740
741/* Step 2: apply policy to a range and do splits. */
742static int mbind_range(struct mm_struct *mm, unsigned long start,
743 unsigned long end, struct mempolicy *new_pol)
744{
745 struct vm_area_struct *next;
746 struct vm_area_struct *prev;
747 struct vm_area_struct *vma;
748 int err = 0;
749 pgoff_t pgoff;
750 unsigned long vmstart;
751 unsigned long vmend;
752
753 vma = find_vma(mm, start);
754 if (!vma || vma->vm_start > start)
755 return -EFAULT;
756
757 prev = vma->vm_prev;
758 if (start > vma->vm_start)
759 prev = vma;
760
761 for (; vma && vma->vm_start < end; prev = vma, vma = next) {
762 next = vma->vm_next;
763 vmstart = max(start, vma->vm_start);
764 vmend = min(end, vma->vm_end);
765
766 if (mpol_equal(vma_policy(vma), new_pol))
767 continue;
768
769 pgoff = vma->vm_pgoff +
770 ((vmstart - vma->vm_start) >> PAGE_SHIFT);
771 prev = vma_merge(mm, prev, vmstart, vmend, vma->vm_flags,
772 vma->anon_vma, vma->vm_file, pgoff,
773 new_pol);
774 if (prev) {
775 vma = prev;
776 next = vma->vm_next;
777 if (mpol_equal(vma_policy(vma), new_pol))
778 continue;
779 /* vma_merge() joined vma && vma->next, case 8 */
780 goto replace;
781 }
782 if (vma->vm_start != vmstart) {
783 err = split_vma(vma->vm_mm, vma, vmstart, 1);
784 if (err)
785 goto out;
786 }
787 if (vma->vm_end != vmend) {
788 err = split_vma(vma->vm_mm, vma, vmend, 0);
789 if (err)
790 goto out;
791 }
792 replace:
793 err = vma_replace_policy(vma, new_pol);
794 if (err)
795 goto out;
796 }
797
798 out:
799 return err;
800}
801
802/* Set the process memory policy */
803static long do_set_mempolicy(unsigned short mode, unsigned short flags,
804 nodemask_t *nodes)
805{
806 struct mempolicy *new, *old;
807 struct mm_struct *mm = current->mm;
808 NODEMASK_SCRATCH(scratch);
809 int ret;
810
811 if (!scratch)
812 return -ENOMEM;
813
814 new = mpol_new(mode, flags, nodes);
815 if (IS_ERR(new)) {
816 ret = PTR_ERR(new);
817 goto out;
818 }
819 /*
820 * prevent changing our mempolicy while show_numa_maps()
821 * is using it.
822 * Note: do_set_mempolicy() can be called at init time
823 * with no 'mm'.
824 */
825 if (mm)
826 down_write(&mm->mmap_sem);
827 task_lock(current);
828 ret = mpol_set_nodemask(new, nodes, scratch);
829 if (ret) {
830 task_unlock(current);
831 if (mm)
832 up_write(&mm->mmap_sem);
833 mpol_put(new);
834 goto out;
835 }
836 old = current->mempolicy;
837 current->mempolicy = new;
838 if (new && new->mode == MPOL_INTERLEAVE &&
839 nodes_weight(new->v.nodes))
840 current->il_next = first_node(new->v.nodes);
841 task_unlock(current);
842 if (mm)
843 up_write(&mm->mmap_sem);
844
845 mpol_put(old);
846 ret = 0;
847out:
848 NODEMASK_SCRATCH_FREE(scratch);
849 return ret;
850}
851
852/*
853 * Return nodemask for policy for get_mempolicy() query
854 *
855 * Called with task's alloc_lock held
856 */
857static void get_policy_nodemask(struct mempolicy *p, nodemask_t *nodes)
858{
859 nodes_clear(*nodes);
860 if (p == &default_policy)
861 return;
862
863 switch (p->mode) {
864 case MPOL_BIND:
865 /* Fall through */
866 case MPOL_INTERLEAVE:
867 *nodes = p->v.nodes;
868 break;
869 case MPOL_PREFERRED:
870 if (!(p->flags & MPOL_F_LOCAL))
871 node_set(p->v.preferred_node, *nodes);
872 /* else return empty node mask for local allocation */
873 break;
874 default:
875 BUG();
876 }
877}
878
879static int lookup_node(struct mm_struct *mm, unsigned long addr)
880{
881 struct page *p;
882 int err;
883
884 err = get_user_pages(current, mm, addr & PAGE_MASK, 1, 0, 0, &p, NULL);
885 if (err >= 0) {
886 err = page_to_nid(p);
887 put_page(p);
888 }
889 return err;
890}
891
892/* Retrieve NUMA policy */
893static long do_get_mempolicy(int *policy, nodemask_t *nmask,
894 unsigned long addr, unsigned long flags)
895{
896 int err;
897 struct mm_struct *mm = current->mm;
898 struct vm_area_struct *vma = NULL;
899 struct mempolicy *pol = current->mempolicy;
900
901 if (flags &
902 ~(unsigned long)(MPOL_F_NODE|MPOL_F_ADDR|MPOL_F_MEMS_ALLOWED))
903 return -EINVAL;
904
905 if (flags & MPOL_F_MEMS_ALLOWED) {
906 if (flags & (MPOL_F_NODE|MPOL_F_ADDR))
907 return -EINVAL;
908 *policy = 0; /* just so it's initialized */
909 task_lock(current);
910 *nmask = cpuset_current_mems_allowed;
911 task_unlock(current);
912 return 0;
913 }
914
915 if (flags & MPOL_F_ADDR) {
916 /*
917 * Do NOT fall back to task policy if the
918 * vma/shared policy at addr is NULL. We
919 * want to return MPOL_DEFAULT in this case.
920 */
921 down_read(&mm->mmap_sem);
922 vma = find_vma_intersection(mm, addr, addr+1);
923 if (!vma) {
924 up_read(&mm->mmap_sem);
925 return -EFAULT;
926 }
927 if (vma->vm_ops && vma->vm_ops->get_policy)
928 pol = vma->vm_ops->get_policy(vma, addr);
929 else
930 pol = vma->vm_policy;
931 } else if (addr)
932 return -EINVAL;
933
934 if (!pol)
935 pol = &default_policy; /* indicates default behavior */
936
937 if (flags & MPOL_F_NODE) {
938 if (flags & MPOL_F_ADDR) {
939 err = lookup_node(mm, addr);
940 if (err < 0)
941 goto out;
942 *policy = err;
943 } else if (pol == current->mempolicy &&
944 pol->mode == MPOL_INTERLEAVE) {
945 *policy = current->il_next;
946 } else {
947 err = -EINVAL;
948 goto out;
949 }
950 } else {
951 *policy = pol == &default_policy ? MPOL_DEFAULT :
952 pol->mode;
953 /*
954 * Internal mempolicy flags must be masked off before exposing
955 * the policy to userspace.
956 */
957 *policy |= (pol->flags & MPOL_MODE_FLAGS);
958 }
959
960 if (vma) {
961 up_read(¤t->mm->mmap_sem);
962 vma = NULL;
963 }
964
965 err = 0;
966 if (nmask) {
967 if (mpol_store_user_nodemask(pol)) {
968 *nmask = pol->w.user_nodemask;
969 } else {
970 task_lock(current);
971 get_policy_nodemask(pol, nmask);
972 task_unlock(current);
973 }
974 }
975
976 out:
977 mpol_cond_put(pol);
978 if (vma)
979 up_read(¤t->mm->mmap_sem);
980 return err;
981}
982
983#ifdef CONFIG_MIGRATION
984/*
985 * page migration
986 */
987static void migrate_page_add(struct page *page, struct list_head *pagelist,
988 unsigned long flags)
989{
990 /*
991 * Avoid migrating a page that is shared with others.
992 */
993 if ((flags & MPOL_MF_MOVE_ALL) || page_mapcount(page) == 1) {
994 if (!isolate_lru_page(page)) {
995 list_add_tail(&page->lru, pagelist);
996 inc_zone_page_state(page, NR_ISOLATED_ANON +
997 page_is_file_cache(page));
998 }
999 }
1000}
1001
1002static struct page *new_node_page(struct page *page, unsigned long node, int **x)
1003{
1004 if (PageHuge(page))
1005 return alloc_huge_page_node(page_hstate(compound_head(page)),
1006 node);
1007 else
1008 return alloc_pages_exact_node(node, GFP_HIGHUSER_MOVABLE, 0);
1009}
1010
1011/*
1012 * Migrate pages from one node to a target node.
1013 * Returns error or the number of pages not migrated.
1014 */
1015static int migrate_to_node(struct mm_struct *mm, int source, int dest,
1016 int flags)
1017{
1018 nodemask_t nmask;
1019 LIST_HEAD(pagelist);
1020 int err = 0;
1021
1022 nodes_clear(nmask);
1023 node_set(source, nmask);
1024
1025 /*
1026 * This does not "check" the range but isolates all pages that
1027 * need migration. Between passing in the full user address
1028 * space range and MPOL_MF_DISCONTIG_OK, this call can not fail.
1029 */
1030 VM_BUG_ON(!(flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)));
1031 queue_pages_range(mm, mm->mmap->vm_start, mm->task_size, &nmask,
1032 flags | MPOL_MF_DISCONTIG_OK, &pagelist);
1033
1034 if (!list_empty(&pagelist)) {
1035 err = migrate_pages(&pagelist, new_node_page, dest,
1036 MIGRATE_SYNC, MR_SYSCALL);
1037 if (err)
1038 putback_movable_pages(&pagelist);
1039 }
1040
1041 return err;
1042}
1043
1044/*
1045 * Move pages between the two nodesets so as to preserve the physical
1046 * layout as much as possible.
1047 *
1048 * Returns the number of page that could not be moved.
1049 */
1050int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
1051 const nodemask_t *to, int flags)
1052{
1053 int busy = 0;
1054 int err;
1055 nodemask_t tmp;
1056
1057 err = migrate_prep();
1058 if (err)
1059 return err;
1060
1061 down_read(&mm->mmap_sem);
1062
1063 err = migrate_vmas(mm, from, to, flags);
1064 if (err)
1065 goto out;
1066
1067 /*
1068 * Find a 'source' bit set in 'tmp' whose corresponding 'dest'
1069 * bit in 'to' is not also set in 'tmp'. Clear the found 'source'
1070 * bit in 'tmp', and return that <source, dest> pair for migration.
1071 * The pair of nodemasks 'to' and 'from' define the map.
1072 *
1073 * If no pair of bits is found that way, fallback to picking some
1074 * pair of 'source' and 'dest' bits that are not the same. If the
1075 * 'source' and 'dest' bits are the same, this represents a node
1076 * that will be migrating to itself, so no pages need move.
1077 *
1078 * If no bits are left in 'tmp', or if all remaining bits left
1079 * in 'tmp' correspond to the same bit in 'to', return false
1080 * (nothing left to migrate).
1081 *
1082 * This lets us pick a pair of nodes to migrate between, such that
1083 * if possible the dest node is not already occupied by some other
1084 * source node, minimizing the risk of overloading the memory on a
1085 * node that would happen if we migrated incoming memory to a node
1086 * before migrating outgoing memory source that same node.
1087 *
1088 * A single scan of tmp is sufficient. As we go, we remember the
1089 * most recent <s, d> pair that moved (s != d). If we find a pair
1090 * that not only moved, but what's better, moved to an empty slot
1091 * (d is not set in tmp), then we break out then, with that pair.
1092 * Otherwise when we finish scanning from_tmp, we at least have the
1093 * most recent <s, d> pair that moved. If we get all the way through
1094 * the scan of tmp without finding any node that moved, much less
1095 * moved to an empty node, then there is nothing left worth migrating.
1096 */
1097
1098 tmp = *from;
1099 while (!nodes_empty(tmp)) {
1100 int s,d;
1101 int source = NUMA_NO_NODE;
1102 int dest = 0;
1103
1104 for_each_node_mask(s, tmp) {
1105
1106 /*
1107 * do_migrate_pages() tries to maintain the relative
1108 * node relationship of the pages established between
1109 * threads and memory areas.
1110 *
1111 * However if the number of source nodes is not equal to
1112 * the number of destination nodes we can not preserve
1113 * this node relative relationship. In that case, skip
1114 * copying memory from a node that is in the destination
1115 * mask.
1116 *
1117 * Example: [2,3,4] -> [3,4,5] moves everything.
1118 * [0-7] - > [3,4,5] moves only 0,1,2,6,7.
1119 */
1120
1121 if ((nodes_weight(*from) != nodes_weight(*to)) &&
1122 (node_isset(s, *to)))
1123 continue;
1124
1125 d = node_remap(s, *from, *to);
1126 if (s == d)
1127 continue;
1128
1129 source = s; /* Node moved. Memorize */
1130 dest = d;
1131
1132 /* dest not in remaining from nodes? */
1133 if (!node_isset(dest, tmp))
1134 break;
1135 }
1136 if (source == NUMA_NO_NODE)
1137 break;
1138
1139 node_clear(source, tmp);
1140 err = migrate_to_node(mm, source, dest, flags);
1141 if (err > 0)
1142 busy += err;
1143 if (err < 0)
1144 break;
1145 }
1146out:
1147 up_read(&mm->mmap_sem);
1148 if (err < 0)
1149 return err;
1150 return busy;
1151
1152}
1153
1154/*
1155 * Allocate a new page for page migration based on vma policy.
1156 * Start assuming that page is mapped by vma pointed to by @private.
1157 * Search forward from there, if not. N.B., this assumes that the
1158 * list of pages handed to migrate_pages()--which is how we get here--
1159 * is in virtual address order.
1160 */
1161static struct page *new_vma_page(struct page *page, unsigned long private, int **x)
1162{
1163 struct vm_area_struct *vma = (struct vm_area_struct *)private;
1164 unsigned long uninitialized_var(address);
1165
1166 while (vma) {
1167 address = page_address_in_vma(page, vma);
1168 if (address != -EFAULT)
1169 break;
1170 vma = vma->vm_next;
1171 }
1172
1173 if (PageHuge(page)) {
1174 BUG_ON(!vma);
1175 return alloc_huge_page_noerr(vma, address, 1);
1176 }
1177 /*
1178 * if !vma, alloc_page_vma() will use task or system default policy
1179 */
1180 return alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
1181}
1182#else
1183
1184static void migrate_page_add(struct page *page, struct list_head *pagelist,
1185 unsigned long flags)
1186{
1187}
1188
1189int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
1190 const nodemask_t *to, int flags)
1191{
1192 return -ENOSYS;
1193}
1194
1195static struct page *new_vma_page(struct page *page, unsigned long private, int **x)
1196{
1197 return NULL;
1198}
1199#endif
1200
1201static long do_mbind(unsigned long start, unsigned long len,
1202 unsigned short mode, unsigned short mode_flags,
1203 nodemask_t *nmask, unsigned long flags)
1204{
1205 struct vm_area_struct *vma;
1206 struct mm_struct *mm = current->mm;
1207 struct mempolicy *new;
1208 unsigned long end;
1209 int err;
1210 LIST_HEAD(pagelist);
1211
1212 if (flags & ~(unsigned long)MPOL_MF_VALID)
1213 return -EINVAL;
1214 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1215 return -EPERM;
1216
1217 if (start & ~PAGE_MASK)
1218 return -EINVAL;
1219
1220 if (mode == MPOL_DEFAULT)
1221 flags &= ~MPOL_MF_STRICT;
1222
1223 len = (len + PAGE_SIZE - 1) & PAGE_MASK;
1224 end = start + len;
1225
1226 if (end < start)
1227 return -EINVAL;
1228 if (end == start)
1229 return 0;
1230
1231 new = mpol_new(mode, mode_flags, nmask);
1232 if (IS_ERR(new))
1233 return PTR_ERR(new);
1234
1235 if (flags & MPOL_MF_LAZY)
1236 new->flags |= MPOL_F_MOF;
1237
1238 /*
1239 * If we are using the default policy then operation
1240 * on discontinuous address spaces is okay after all
1241 */
1242 if (!new)
1243 flags |= MPOL_MF_DISCONTIG_OK;
1244
1245 pr_debug("mbind %lx-%lx mode:%d flags:%d nodes:%lx\n",
1246 start, start + len, mode, mode_flags,
1247 nmask ? nodes_addr(*nmask)[0] : NUMA_NO_NODE);
1248
1249 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) {
1250
1251 err = migrate_prep();
1252 if (err)
1253 goto mpol_out;
1254 }
1255 {
1256 NODEMASK_SCRATCH(scratch);
1257 if (scratch) {
1258 down_write(&mm->mmap_sem);
1259 task_lock(current);
1260 err = mpol_set_nodemask(new, nmask, scratch);
1261 task_unlock(current);
1262 if (err)
1263 up_write(&mm->mmap_sem);
1264 } else
1265 err = -ENOMEM;
1266 NODEMASK_SCRATCH_FREE(scratch);
1267 }
1268 if (err)
1269 goto mpol_out;
1270
1271 vma = queue_pages_range(mm, start, end, nmask,
1272 flags | MPOL_MF_INVERT, &pagelist);
1273
1274 err = PTR_ERR(vma); /* maybe ... */
1275 if (!IS_ERR(vma))
1276 err = mbind_range(mm, start, end, new);
1277
1278 if (!err) {
1279 int nr_failed = 0;
1280
1281 if (!list_empty(&pagelist)) {
1282 WARN_ON_ONCE(flags & MPOL_MF_LAZY);
1283 nr_failed = migrate_pages(&pagelist, new_vma_page,
1284 (unsigned long)vma,
1285 MIGRATE_SYNC, MR_MEMPOLICY_MBIND);
1286 if (nr_failed)
1287 putback_movable_pages(&pagelist);
1288 }
1289
1290 if (nr_failed && (flags & MPOL_MF_STRICT))
1291 err = -EIO;
1292 } else
1293 putback_movable_pages(&pagelist);
1294
1295 up_write(&mm->mmap_sem);
1296 mpol_out:
1297 mpol_put(new);
1298 return err;
1299}
1300
1301/*
1302 * User space interface with variable sized bitmaps for nodelists.
1303 */
1304
1305/* Copy a node mask from user space. */
1306static int get_nodes(nodemask_t *nodes, const unsigned long __user *nmask,
1307 unsigned long maxnode)
1308{
1309 unsigned long k;
1310 unsigned long nlongs;
1311 unsigned long endmask;
1312
1313 --maxnode;
1314 nodes_clear(*nodes);
1315 if (maxnode == 0 || !nmask)
1316 return 0;
1317 if (maxnode > PAGE_SIZE*BITS_PER_BYTE)
1318 return -EINVAL;
1319
1320 nlongs = BITS_TO_LONGS(maxnode);
1321 if ((maxnode % BITS_PER_LONG) == 0)
1322 endmask = ~0UL;
1323 else
1324 endmask = (1UL << (maxnode % BITS_PER_LONG)) - 1;
1325
1326 /* When the user specified more nodes than supported just check
1327 if the non supported part is all zero. */
1328 if (nlongs > BITS_TO_LONGS(MAX_NUMNODES)) {
1329 if (nlongs > PAGE_SIZE/sizeof(long))
1330 return -EINVAL;
1331 for (k = BITS_TO_LONGS(MAX_NUMNODES); k < nlongs; k++) {
1332 unsigned long t;
1333 if (get_user(t, nmask + k))
1334 return -EFAULT;
1335 if (k == nlongs - 1) {
1336 if (t & endmask)
1337 return -EINVAL;
1338 } else if (t)
1339 return -EINVAL;
1340 }
1341 nlongs = BITS_TO_LONGS(MAX_NUMNODES);
1342 endmask = ~0UL;
1343 }
1344
1345 if (copy_from_user(nodes_addr(*nodes), nmask, nlongs*sizeof(unsigned long)))
1346 return -EFAULT;
1347 nodes_addr(*nodes)[nlongs-1] &= endmask;
1348 return 0;
1349}
1350
1351/* Copy a kernel node mask to user space */
1352static int copy_nodes_to_user(unsigned long __user *mask, unsigned long maxnode,
1353 nodemask_t *nodes)
1354{
1355 unsigned long copy = ALIGN(maxnode-1, 64) / 8;
1356 const int nbytes = BITS_TO_LONGS(MAX_NUMNODES) * sizeof(long);
1357
1358 if (copy > nbytes) {
1359 if (copy > PAGE_SIZE)
1360 return -EINVAL;
1361 if (clear_user((char __user *)mask + nbytes, copy - nbytes))
1362 return -EFAULT;
1363 copy = nbytes;
1364 }
1365 return copy_to_user(mask, nodes_addr(*nodes), copy) ? -EFAULT : 0;
1366}
1367
1368SYSCALL_DEFINE6(mbind, unsigned long, start, unsigned long, len,
1369 unsigned long, mode, unsigned long __user *, nmask,
1370 unsigned long, maxnode, unsigned, flags)
1371{
1372 nodemask_t nodes;
1373 int err;
1374 unsigned short mode_flags;
1375
1376 mode_flags = mode & MPOL_MODE_FLAGS;
1377 mode &= ~MPOL_MODE_FLAGS;
1378 if (mode >= MPOL_MAX)
1379 return -EINVAL;
1380 if ((mode_flags & MPOL_F_STATIC_NODES) &&
1381 (mode_flags & MPOL_F_RELATIVE_NODES))
1382 return -EINVAL;
1383 err = get_nodes(&nodes, nmask, maxnode);
1384 if (err)
1385 return err;
1386 return do_mbind(start, len, mode, mode_flags, &nodes, flags);
1387}
1388
1389/* Set the process memory policy */
1390SYSCALL_DEFINE3(set_mempolicy, int, mode, unsigned long __user *, nmask,
1391 unsigned long, maxnode)
1392{
1393 int err;
1394 nodemask_t nodes;
1395 unsigned short flags;
1396
1397 flags = mode & MPOL_MODE_FLAGS;
1398 mode &= ~MPOL_MODE_FLAGS;
1399 if ((unsigned int)mode >= MPOL_MAX)
1400 return -EINVAL;
1401 if ((flags & MPOL_F_STATIC_NODES) && (flags & MPOL_F_RELATIVE_NODES))
1402 return -EINVAL;
1403 err = get_nodes(&nodes, nmask, maxnode);
1404 if (err)
1405 return err;
1406 return do_set_mempolicy(mode, flags, &nodes);
1407}
1408
1409SYSCALL_DEFINE4(migrate_pages, pid_t, pid, unsigned long, maxnode,
1410 const unsigned long __user *, old_nodes,
1411 const unsigned long __user *, new_nodes)
1412{
1413 const struct cred *cred = current_cred(), *tcred;
1414 struct mm_struct *mm = NULL;
1415 struct task_struct *task;
1416 nodemask_t task_nodes;
1417 int err;
1418 nodemask_t *old;
1419 nodemask_t *new;
1420 NODEMASK_SCRATCH(scratch);
1421
1422 if (!scratch)
1423 return -ENOMEM;
1424
1425 old = &scratch->mask1;
1426 new = &scratch->mask2;
1427
1428 err = get_nodes(old, old_nodes, maxnode);
1429 if (err)
1430 goto out;
1431
1432 err = get_nodes(new, new_nodes, maxnode);
1433 if (err)
1434 goto out;
1435
1436 /* Find the mm_struct */
1437 rcu_read_lock();
1438 task = pid ? find_task_by_vpid(pid) : current;
1439 if (!task) {
1440 rcu_read_unlock();
1441 err = -ESRCH;
1442 goto out;
1443 }
1444 get_task_struct(task);
1445
1446 err = -EINVAL;
1447
1448 /*
1449 * Check if this process has the right to modify the specified
1450 * process. The right exists if the process has administrative
1451 * capabilities, superuser privileges or the same
1452 * userid as the target process.
1453 */
1454 tcred = __task_cred(task);
1455 if (!uid_eq(cred->euid, tcred->suid) && !uid_eq(cred->euid, tcred->uid) &&
1456 !uid_eq(cred->uid, tcred->suid) && !uid_eq(cred->uid, tcred->uid) &&
1457 !capable(CAP_SYS_NICE)) {
1458 rcu_read_unlock();
1459 err = -EPERM;
1460 goto out_put;
1461 }
1462 rcu_read_unlock();
1463
1464 task_nodes = cpuset_mems_allowed(task);
1465 /* Is the user allowed to access the target nodes? */
1466 if (!nodes_subset(*new, task_nodes) && !capable(CAP_SYS_NICE)) {
1467 err = -EPERM;
1468 goto out_put;
1469 }
1470
1471 if (!nodes_subset(*new, node_states[N_MEMORY])) {
1472 err = -EINVAL;
1473 goto out_put;
1474 }
1475
1476 err = security_task_movememory(task);
1477 if (err)
1478 goto out_put;
1479
1480 mm = get_task_mm(task);
1481 put_task_struct(task);
1482
1483 if (!mm) {
1484 err = -EINVAL;
1485 goto out;
1486 }
1487
1488 err = do_migrate_pages(mm, old, new,
1489 capable(CAP_SYS_NICE) ? MPOL_MF_MOVE_ALL : MPOL_MF_MOVE);
1490
1491 mmput(mm);
1492out:
1493 NODEMASK_SCRATCH_FREE(scratch);
1494
1495 return err;
1496
1497out_put:
1498 put_task_struct(task);
1499 goto out;
1500
1501}
1502
1503
1504/* Retrieve NUMA policy */
1505SYSCALL_DEFINE5(get_mempolicy, int __user *, policy,
1506 unsigned long __user *, nmask, unsigned long, maxnode,
1507 unsigned long, addr, unsigned long, flags)
1508{
1509 int err;
1510 int uninitialized_var(pval);
1511 nodemask_t nodes;
1512
1513 if (nmask != NULL && maxnode < MAX_NUMNODES)
1514 return -EINVAL;
1515
1516 err = do_get_mempolicy(&pval, &nodes, addr, flags);
1517
1518 if (err)
1519 return err;
1520
1521 if (policy && put_user(pval, policy))
1522 return -EFAULT;
1523
1524 if (nmask)
1525 err = copy_nodes_to_user(nmask, maxnode, &nodes);
1526
1527 return err;
1528}
1529
1530#ifdef CONFIG_COMPAT
1531
1532COMPAT_SYSCALL_DEFINE5(get_mempolicy, int __user *, policy,
1533 compat_ulong_t __user *, nmask,
1534 compat_ulong_t, maxnode,
1535 compat_ulong_t, addr, compat_ulong_t, flags)
1536{
1537 long err;
1538 unsigned long __user *nm = NULL;
1539 unsigned long nr_bits, alloc_size;
1540 DECLARE_BITMAP(bm, MAX_NUMNODES);
1541
1542 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1543 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1544
1545 if (nmask)
1546 nm = compat_alloc_user_space(alloc_size);
1547
1548 err = sys_get_mempolicy(policy, nm, nr_bits+1, addr, flags);
1549
1550 if (!err && nmask) {
1551 unsigned long copy_size;
1552 copy_size = min_t(unsigned long, sizeof(bm), alloc_size);
1553 err = copy_from_user(bm, nm, copy_size);
1554 /* ensure entire bitmap is zeroed */
1555 err |= clear_user(nmask, ALIGN(maxnode-1, 8) / 8);
1556 err |= compat_put_bitmap(nmask, bm, nr_bits);
1557 }
1558
1559 return err;
1560}
1561
1562COMPAT_SYSCALL_DEFINE3(set_mempolicy, int, mode, compat_ulong_t __user *, nmask,
1563 compat_ulong_t, maxnode)
1564{
1565 long err = 0;
1566 unsigned long __user *nm = NULL;
1567 unsigned long nr_bits, alloc_size;
1568 DECLARE_BITMAP(bm, MAX_NUMNODES);
1569
1570 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1571 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1572
1573 if (nmask) {
1574 err = compat_get_bitmap(bm, nmask, nr_bits);
1575 nm = compat_alloc_user_space(alloc_size);
1576 err |= copy_to_user(nm, bm, alloc_size);
1577 }
1578
1579 if (err)
1580 return -EFAULT;
1581
1582 return sys_set_mempolicy(mode, nm, nr_bits+1);
1583}
1584
1585COMPAT_SYSCALL_DEFINE6(mbind, compat_ulong_t, start, compat_ulong_t, len,
1586 compat_ulong_t, mode, compat_ulong_t __user *, nmask,
1587 compat_ulong_t, maxnode, compat_ulong_t, flags)
1588{
1589 long err = 0;
1590 unsigned long __user *nm = NULL;
1591 unsigned long nr_bits, alloc_size;
1592 nodemask_t bm;
1593
1594 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1595 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1596
1597 if (nmask) {
1598 err = compat_get_bitmap(nodes_addr(bm), nmask, nr_bits);
1599 nm = compat_alloc_user_space(alloc_size);
1600 err |= copy_to_user(nm, nodes_addr(bm), alloc_size);
1601 }
1602
1603 if (err)
1604 return -EFAULT;
1605
1606 return sys_mbind(start, len, mode, nm, nr_bits+1, flags);
1607}
1608
1609#endif
1610
1611/*
1612 * get_vma_policy(@task, @vma, @addr)
1613 * @task - task for fallback if vma policy == default
1614 * @vma - virtual memory area whose policy is sought
1615 * @addr - address in @vma for shared policy lookup
1616 *
1617 * Returns effective policy for a VMA at specified address.
1618 * Falls back to @task or system default policy, as necessary.
1619 * Current or other task's task mempolicy and non-shared vma policies must be
1620 * protected by task_lock(task) by the caller.
1621 * Shared policies [those marked as MPOL_F_SHARED] require an extra reference
1622 * count--added by the get_policy() vm_op, as appropriate--to protect against
1623 * freeing by another task. It is the caller's responsibility to free the
1624 * extra reference for shared policies.
1625 */
1626struct mempolicy *get_vma_policy(struct task_struct *task,
1627 struct vm_area_struct *vma, unsigned long addr)
1628{
1629 struct mempolicy *pol = get_task_policy(task);
1630
1631 if (vma) {
1632 if (vma->vm_ops && vma->vm_ops->get_policy) {
1633 struct mempolicy *vpol = vma->vm_ops->get_policy(vma,
1634 addr);
1635 if (vpol)
1636 pol = vpol;
1637 } else if (vma->vm_policy) {
1638 pol = vma->vm_policy;
1639
1640 /*
1641 * shmem_alloc_page() passes MPOL_F_SHARED policy with
1642 * a pseudo vma whose vma->vm_ops=NULL. Take a reference
1643 * count on these policies which will be dropped by
1644 * mpol_cond_put() later
1645 */
1646 if (mpol_needs_cond_ref(pol))
1647 mpol_get(pol);
1648 }
1649 }
1650 if (!pol)
1651 pol = &default_policy;
1652 return pol;
1653}
1654
1655bool vma_policy_mof(struct task_struct *task, struct vm_area_struct *vma)
1656{
1657 struct mempolicy *pol = get_task_policy(task);
1658 if (vma) {
1659 if (vma->vm_ops && vma->vm_ops->get_policy) {
1660 bool ret = false;
1661
1662 pol = vma->vm_ops->get_policy(vma, vma->vm_start);
1663 if (pol && (pol->flags & MPOL_F_MOF))
1664 ret = true;
1665 mpol_cond_put(pol);
1666
1667 return ret;
1668 } else if (vma->vm_policy) {
1669 pol = vma->vm_policy;
1670 }
1671 }
1672
1673 if (!pol)
1674 return default_policy.flags & MPOL_F_MOF;
1675
1676 return pol->flags & MPOL_F_MOF;
1677}
1678
1679static int apply_policy_zone(struct mempolicy *policy, enum zone_type zone)
1680{
1681 enum zone_type dynamic_policy_zone = policy_zone;
1682
1683 BUG_ON(dynamic_policy_zone == ZONE_MOVABLE);
1684
1685 /*
1686 * if policy->v.nodes has movable memory only,
1687 * we apply policy when gfp_zone(gfp) = ZONE_MOVABLE only.
1688 *
1689 * policy->v.nodes is intersect with node_states[N_MEMORY].
1690 * so if the following test faile, it implies
1691 * policy->v.nodes has movable memory only.
1692 */
1693 if (!nodes_intersects(policy->v.nodes, node_states[N_HIGH_MEMORY]))
1694 dynamic_policy_zone = ZONE_MOVABLE;
1695
1696 return zone >= dynamic_policy_zone;
1697}
1698
1699/*
1700 * Return a nodemask representing a mempolicy for filtering nodes for
1701 * page allocation
1702 */
1703static nodemask_t *policy_nodemask(gfp_t gfp, struct mempolicy *policy)
1704{
1705 /* Lower zones don't get a nodemask applied for MPOL_BIND */
1706 if (unlikely(policy->mode == MPOL_BIND) &&
1707 apply_policy_zone(policy, gfp_zone(gfp)) &&
1708 cpuset_nodemask_valid_mems_allowed(&policy->v.nodes))
1709 return &policy->v.nodes;
1710
1711 return NULL;
1712}
1713
1714/* Return a zonelist indicated by gfp for node representing a mempolicy */
1715static struct zonelist *policy_zonelist(gfp_t gfp, struct mempolicy *policy,
1716 int nd)
1717{
1718 switch (policy->mode) {
1719 case MPOL_PREFERRED:
1720 if (!(policy->flags & MPOL_F_LOCAL))
1721 nd = policy->v.preferred_node;
1722 break;
1723 case MPOL_BIND:
1724 /*
1725 * Normally, MPOL_BIND allocations are node-local within the
1726 * allowed nodemask. However, if __GFP_THISNODE is set and the
1727 * current node isn't part of the mask, we use the zonelist for
1728 * the first node in the mask instead.
1729 */
1730 if (unlikely(gfp & __GFP_THISNODE) &&
1731 unlikely(!node_isset(nd, policy->v.nodes)))
1732 nd = first_node(policy->v.nodes);
1733 break;
1734 default:
1735 BUG();
1736 }
1737 return node_zonelist(nd, gfp);
1738}
1739
1740/* Do dynamic interleaving for a process */
1741static unsigned interleave_nodes(struct mempolicy *policy)
1742{
1743 unsigned nid, next;
1744 struct task_struct *me = current;
1745
1746 nid = me->il_next;
1747 next = next_node(nid, policy->v.nodes);
1748 if (next >= MAX_NUMNODES)
1749 next = first_node(policy->v.nodes);
1750 if (next < MAX_NUMNODES)
1751 me->il_next = next;
1752 return nid;
1753}
1754
1755/*
1756 * Depending on the memory policy provide a node from which to allocate the
1757 * next slab entry.
1758 */
1759unsigned int mempolicy_slab_node(void)
1760{
1761 struct mempolicy *policy;
1762 int node = numa_mem_id();
1763
1764 if (in_interrupt())
1765 return node;
1766
1767 policy = current->mempolicy;
1768 if (!policy || policy->flags & MPOL_F_LOCAL)
1769 return node;
1770
1771 switch (policy->mode) {
1772 case MPOL_PREFERRED:
1773 /*
1774 * handled MPOL_F_LOCAL above
1775 */
1776 return policy->v.preferred_node;
1777
1778 case MPOL_INTERLEAVE:
1779 return interleave_nodes(policy);
1780
1781 case MPOL_BIND: {
1782 /*
1783 * Follow bind policy behavior and start allocation at the
1784 * first node.
1785 */
1786 struct zonelist *zonelist;
1787 struct zone *zone;
1788 enum zone_type highest_zoneidx = gfp_zone(GFP_KERNEL);
1789 zonelist = &NODE_DATA(node)->node_zonelists[0];
1790 (void)first_zones_zonelist(zonelist, highest_zoneidx,
1791 &policy->v.nodes,
1792 &zone);
1793 return zone ? zone->node : node;
1794 }
1795
1796 default:
1797 BUG();
1798 }
1799}
1800
1801/* Do static interleaving for a VMA with known offset. */
1802static unsigned offset_il_node(struct mempolicy *pol,
1803 struct vm_area_struct *vma, unsigned long off)
1804{
1805 unsigned nnodes = nodes_weight(pol->v.nodes);
1806 unsigned target;
1807 int c;
1808 int nid = NUMA_NO_NODE;
1809
1810 if (!nnodes)
1811 return numa_node_id();
1812 target = (unsigned int)off % nnodes;
1813 c = 0;
1814 do {
1815 nid = next_node(nid, pol->v.nodes);
1816 c++;
1817 } while (c <= target);
1818 return nid;
1819}
1820
1821/* Determine a node number for interleave */
1822static inline unsigned interleave_nid(struct mempolicy *pol,
1823 struct vm_area_struct *vma, unsigned long addr, int shift)
1824{
1825 if (vma) {
1826 unsigned long off;
1827
1828 /*
1829 * for small pages, there is no difference between
1830 * shift and PAGE_SHIFT, so the bit-shift is safe.
1831 * for huge pages, since vm_pgoff is in units of small
1832 * pages, we need to shift off the always 0 bits to get
1833 * a useful offset.
1834 */
1835 BUG_ON(shift < PAGE_SHIFT);
1836 off = vma->vm_pgoff >> (shift - PAGE_SHIFT);
1837 off += (addr - vma->vm_start) >> shift;
1838 return offset_il_node(pol, vma, off);
1839 } else
1840 return interleave_nodes(pol);
1841}
1842
1843/*
1844 * Return the bit number of a random bit set in the nodemask.
1845 * (returns NUMA_NO_NODE if nodemask is empty)
1846 */
1847int node_random(const nodemask_t *maskp)
1848{
1849 int w, bit = NUMA_NO_NODE;
1850
1851 w = nodes_weight(*maskp);
1852 if (w)
1853 bit = bitmap_ord_to_pos(maskp->bits,
1854 get_random_int() % w, MAX_NUMNODES);
1855 return bit;
1856}
1857
1858#ifdef CONFIG_HUGETLBFS
1859/*
1860 * huge_zonelist(@vma, @addr, @gfp_flags, @mpol)
1861 * @vma = virtual memory area whose policy is sought
1862 * @addr = address in @vma for shared policy lookup and interleave policy
1863 * @gfp_flags = for requested zone
1864 * @mpol = pointer to mempolicy pointer for reference counted mempolicy
1865 * @nodemask = pointer to nodemask pointer for MPOL_BIND nodemask
1866 *
1867 * Returns a zonelist suitable for a huge page allocation and a pointer
1868 * to the struct mempolicy for conditional unref after allocation.
1869 * If the effective policy is 'BIND, returns a pointer to the mempolicy's
1870 * @nodemask for filtering the zonelist.
1871 *
1872 * Must be protected by read_mems_allowed_begin()
1873 */
1874struct zonelist *huge_zonelist(struct vm_area_struct *vma, unsigned long addr,
1875 gfp_t gfp_flags, struct mempolicy **mpol,
1876 nodemask_t **nodemask)
1877{
1878 struct zonelist *zl;
1879
1880 *mpol = get_vma_policy(current, vma, addr);
1881 *nodemask = NULL; /* assume !MPOL_BIND */
1882
1883 if (unlikely((*mpol)->mode == MPOL_INTERLEAVE)) {
1884 zl = node_zonelist(interleave_nid(*mpol, vma, addr,
1885 huge_page_shift(hstate_vma(vma))), gfp_flags);
1886 } else {
1887 zl = policy_zonelist(gfp_flags, *mpol, numa_node_id());
1888 if ((*mpol)->mode == MPOL_BIND)
1889 *nodemask = &(*mpol)->v.nodes;
1890 }
1891 return zl;
1892}
1893
1894/*
1895 * init_nodemask_of_mempolicy
1896 *
1897 * If the current task's mempolicy is "default" [NULL], return 'false'
1898 * to indicate default policy. Otherwise, extract the policy nodemask
1899 * for 'bind' or 'interleave' policy into the argument nodemask, or
1900 * initialize the argument nodemask to contain the single node for
1901 * 'preferred' or 'local' policy and return 'true' to indicate presence
1902 * of non-default mempolicy.
1903 *
1904 * We don't bother with reference counting the mempolicy [mpol_get/put]
1905 * because the current task is examining it's own mempolicy and a task's
1906 * mempolicy is only ever changed by the task itself.
1907 *
1908 * N.B., it is the caller's responsibility to free a returned nodemask.
1909 */
1910bool init_nodemask_of_mempolicy(nodemask_t *mask)
1911{
1912 struct mempolicy *mempolicy;
1913 int nid;
1914
1915 if (!(mask && current->mempolicy))
1916 return false;
1917
1918 task_lock(current);
1919 mempolicy = current->mempolicy;
1920 switch (mempolicy->mode) {
1921 case MPOL_PREFERRED:
1922 if (mempolicy->flags & MPOL_F_LOCAL)
1923 nid = numa_node_id();
1924 else
1925 nid = mempolicy->v.preferred_node;
1926 init_nodemask_of_node(mask, nid);
1927 break;
1928
1929 case MPOL_BIND:
1930 /* Fall through */
1931 case MPOL_INTERLEAVE:
1932 *mask = mempolicy->v.nodes;
1933 break;
1934
1935 default:
1936 BUG();
1937 }
1938 task_unlock(current);
1939
1940 return true;
1941}
1942#endif
1943
1944/*
1945 * mempolicy_nodemask_intersects
1946 *
1947 * If tsk's mempolicy is "default" [NULL], return 'true' to indicate default
1948 * policy. Otherwise, check for intersection between mask and the policy
1949 * nodemask for 'bind' or 'interleave' policy. For 'perferred' or 'local'
1950 * policy, always return true since it may allocate elsewhere on fallback.
1951 *
1952 * Takes task_lock(tsk) to prevent freeing of its mempolicy.
1953 */
1954bool mempolicy_nodemask_intersects(struct task_struct *tsk,
1955 const nodemask_t *mask)
1956{
1957 struct mempolicy *mempolicy;
1958 bool ret = true;
1959
1960 if (!mask)
1961 return ret;
1962 task_lock(tsk);
1963 mempolicy = tsk->mempolicy;
1964 if (!mempolicy)
1965 goto out;
1966
1967 switch (mempolicy->mode) {
1968 case MPOL_PREFERRED:
1969 /*
1970 * MPOL_PREFERRED and MPOL_F_LOCAL are only preferred nodes to
1971 * allocate from, they may fallback to other nodes when oom.
1972 * Thus, it's possible for tsk to have allocated memory from
1973 * nodes in mask.
1974 */
1975 break;
1976 case MPOL_BIND:
1977 case MPOL_INTERLEAVE:
1978 ret = nodes_intersects(mempolicy->v.nodes, *mask);
1979 break;
1980 default:
1981 BUG();
1982 }
1983out:
1984 task_unlock(tsk);
1985 return ret;
1986}
1987
1988/* Allocate a page in interleaved policy.
1989 Own path because it needs to do special accounting. */
1990static struct page *alloc_page_interleave(gfp_t gfp, unsigned order,
1991 unsigned nid)
1992{
1993 struct zonelist *zl;
1994 struct page *page;
1995
1996 zl = node_zonelist(nid, gfp);
1997 page = __alloc_pages(gfp, order, zl);
1998 if (page && page_zone(page) == zonelist_zone(&zl->_zonerefs[0]))
1999 inc_zone_page_state(page, NUMA_INTERLEAVE_HIT);
2000 return page;
2001}
2002
2003/**
2004 * alloc_pages_vma - Allocate a page for a VMA.
2005 *
2006 * @gfp:
2007 * %GFP_USER user allocation.
2008 * %GFP_KERNEL kernel allocations,
2009 * %GFP_HIGHMEM highmem/user allocations,
2010 * %GFP_FS allocation should not call back into a file system.
2011 * %GFP_ATOMIC don't sleep.
2012 *
2013 * @order:Order of the GFP allocation.
2014 * @vma: Pointer to VMA or NULL if not available.
2015 * @addr: Virtual Address of the allocation. Must be inside the VMA.
2016 *
2017 * This function allocates a page from the kernel page pool and applies
2018 * a NUMA policy associated with the VMA or the current process.
2019 * When VMA is not NULL caller must hold down_read on the mmap_sem of the
2020 * mm_struct of the VMA to prevent it from going away. Should be used for
2021 * all allocations for pages that will be mapped into
2022 * user space. Returns NULL when no page can be allocated.
2023 *
2024 * Should be called with the mm_sem of the vma hold.
2025 */
2026struct page *
2027alloc_pages_vma(gfp_t gfp, int order, struct vm_area_struct *vma,
2028 unsigned long addr, int node)
2029{
2030 struct mempolicy *pol;
2031 struct page *page;
2032 unsigned int cpuset_mems_cookie;
2033
2034retry_cpuset:
2035 pol = get_vma_policy(current, vma, addr);
2036 cpuset_mems_cookie = read_mems_allowed_begin();
2037
2038 if (unlikely(pol->mode == MPOL_INTERLEAVE)) {
2039 unsigned nid;
2040
2041 nid = interleave_nid(pol, vma, addr, PAGE_SHIFT + order);
2042 mpol_cond_put(pol);
2043 page = alloc_page_interleave(gfp, order, nid);
2044 if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie)))
2045 goto retry_cpuset;
2046
2047 return page;
2048 }
2049 page = __alloc_pages_nodemask(gfp, order,
2050 policy_zonelist(gfp, pol, node),
2051 policy_nodemask(gfp, pol));
2052 if (unlikely(mpol_needs_cond_ref(pol)))
2053 __mpol_put(pol);
2054 if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie)))
2055 goto retry_cpuset;
2056 return page;
2057}
2058
2059/**
2060 * alloc_pages_current - Allocate pages.
2061 *
2062 * @gfp:
2063 * %GFP_USER user allocation,
2064 * %GFP_KERNEL kernel allocation,
2065 * %GFP_HIGHMEM highmem allocation,
2066 * %GFP_FS don't call back into a file system.
2067 * %GFP_ATOMIC don't sleep.
2068 * @order: Power of two of allocation size in pages. 0 is a single page.
2069 *
2070 * Allocate a page from the kernel page pool. When not in
2071 * interrupt context and apply the current process NUMA policy.
2072 * Returns NULL when no page can be allocated.
2073 *
2074 * Don't call cpuset_update_task_memory_state() unless
2075 * 1) it's ok to take cpuset_sem (can WAIT), and
2076 * 2) allocating for current task (not interrupt).
2077 */
2078struct page *alloc_pages_current(gfp_t gfp, unsigned order)
2079{
2080 struct mempolicy *pol = get_task_policy(current);
2081 struct page *page;
2082 unsigned int cpuset_mems_cookie;
2083
2084 if (!pol || in_interrupt() || (gfp & __GFP_THISNODE))
2085 pol = &default_policy;
2086
2087retry_cpuset:
2088 cpuset_mems_cookie = read_mems_allowed_begin();
2089
2090 /*
2091 * No reference counting needed for current->mempolicy
2092 * nor system default_policy
2093 */
2094 if (pol->mode == MPOL_INTERLEAVE)
2095 page = alloc_page_interleave(gfp, order, interleave_nodes(pol));
2096 else
2097 page = __alloc_pages_nodemask(gfp, order,
2098 policy_zonelist(gfp, pol, numa_node_id()),
2099 policy_nodemask(gfp, pol));
2100
2101 if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie)))
2102 goto retry_cpuset;
2103
2104 return page;
2105}
2106EXPORT_SYMBOL(alloc_pages_current);
2107
2108int vma_dup_policy(struct vm_area_struct *src, struct vm_area_struct *dst)
2109{
2110 struct mempolicy *pol = mpol_dup(vma_policy(src));
2111
2112 if (IS_ERR(pol))
2113 return PTR_ERR(pol);
2114 dst->vm_policy = pol;
2115 return 0;
2116}
2117
2118/*
2119 * If mpol_dup() sees current->cpuset == cpuset_being_rebound, then it
2120 * rebinds the mempolicy its copying by calling mpol_rebind_policy()
2121 * with the mems_allowed returned by cpuset_mems_allowed(). This
2122 * keeps mempolicies cpuset relative after its cpuset moves. See
2123 * further kernel/cpuset.c update_nodemask().
2124 *
2125 * current's mempolicy may be rebinded by the other task(the task that changes
2126 * cpuset's mems), so we needn't do rebind work for current task.
2127 */
2128
2129/* Slow path of a mempolicy duplicate */
2130struct mempolicy *__mpol_dup(struct mempolicy *old)
2131{
2132 struct mempolicy *new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
2133
2134 if (!new)
2135 return ERR_PTR(-ENOMEM);
2136
2137 /* task's mempolicy is protected by alloc_lock */
2138 if (old == current->mempolicy) {
2139 task_lock(current);
2140 *new = *old;
2141 task_unlock(current);
2142 } else
2143 *new = *old;
2144
2145 rcu_read_lock();
2146 if (current_cpuset_is_being_rebound()) {
2147 nodemask_t mems = cpuset_mems_allowed(current);
2148 if (new->flags & MPOL_F_REBINDING)
2149 mpol_rebind_policy(new, &mems, MPOL_REBIND_STEP2);
2150 else
2151 mpol_rebind_policy(new, &mems, MPOL_REBIND_ONCE);
2152 }
2153 rcu_read_unlock();
2154 atomic_set(&new->refcnt, 1);
2155 return new;
2156}
2157
2158/* Slow path of a mempolicy comparison */
2159bool __mpol_equal(struct mempolicy *a, struct mempolicy *b)
2160{
2161 if (!a || !b)
2162 return false;
2163 if (a->mode != b->mode)
2164 return false;
2165 if (a->flags != b->flags)
2166 return false;
2167 if (mpol_store_user_nodemask(a))
2168 if (!nodes_equal(a->w.user_nodemask, b->w.user_nodemask))
2169 return false;
2170
2171 switch (a->mode) {
2172 case MPOL_BIND:
2173 /* Fall through */
2174 case MPOL_INTERLEAVE:
2175 return !!nodes_equal(a->v.nodes, b->v.nodes);
2176 case MPOL_PREFERRED:
2177 return a->v.preferred_node == b->v.preferred_node;
2178 default:
2179 BUG();
2180 return false;
2181 }
2182}
2183
2184/*
2185 * Shared memory backing store policy support.
2186 *
2187 * Remember policies even when nobody has shared memory mapped.
2188 * The policies are kept in Red-Black tree linked from the inode.
2189 * They are protected by the sp->lock spinlock, which should be held
2190 * for any accesses to the tree.
2191 */
2192
2193/* lookup first element intersecting start-end */
2194/* Caller holds sp->lock */
2195static struct sp_node *
2196sp_lookup(struct shared_policy *sp, unsigned long start, unsigned long end)
2197{
2198 struct rb_node *n = sp->root.rb_node;
2199
2200 while (n) {
2201 struct sp_node *p = rb_entry(n, struct sp_node, nd);
2202
2203 if (start >= p->end)
2204 n = n->rb_right;
2205 else if (end <= p->start)
2206 n = n->rb_left;
2207 else
2208 break;
2209 }
2210 if (!n)
2211 return NULL;
2212 for (;;) {
2213 struct sp_node *w = NULL;
2214 struct rb_node *prev = rb_prev(n);
2215 if (!prev)
2216 break;
2217 w = rb_entry(prev, struct sp_node, nd);
2218 if (w->end <= start)
2219 break;
2220 n = prev;
2221 }
2222 return rb_entry(n, struct sp_node, nd);
2223}
2224
2225/* Insert a new shared policy into the list. */
2226/* Caller holds sp->lock */
2227static void sp_insert(struct shared_policy *sp, struct sp_node *new)
2228{
2229 struct rb_node **p = &sp->root.rb_node;
2230 struct rb_node *parent = NULL;
2231 struct sp_node *nd;
2232
2233 while (*p) {
2234 parent = *p;
2235 nd = rb_entry(parent, struct sp_node, nd);
2236 if (new->start < nd->start)
2237 p = &(*p)->rb_left;
2238 else if (new->end > nd->end)
2239 p = &(*p)->rb_right;
2240 else
2241 BUG();
2242 }
2243 rb_link_node(&new->nd, parent, p);
2244 rb_insert_color(&new->nd, &sp->root);
2245 pr_debug("inserting %lx-%lx: %d\n", new->start, new->end,
2246 new->policy ? new->policy->mode : 0);
2247}
2248
2249/* Find shared policy intersecting idx */
2250struct mempolicy *
2251mpol_shared_policy_lookup(struct shared_policy *sp, unsigned long idx)
2252{
2253 struct mempolicy *pol = NULL;
2254 struct sp_node *sn;
2255
2256 if (!sp->root.rb_node)
2257 return NULL;
2258 spin_lock(&sp->lock);
2259 sn = sp_lookup(sp, idx, idx+1);
2260 if (sn) {
2261 mpol_get(sn->policy);
2262 pol = sn->policy;
2263 }
2264 spin_unlock(&sp->lock);
2265 return pol;
2266}
2267
2268static void sp_free(struct sp_node *n)
2269{
2270 mpol_put(n->policy);
2271 kmem_cache_free(sn_cache, n);
2272}
2273
2274/**
2275 * mpol_misplaced - check whether current page node is valid in policy
2276 *
2277 * @page - page to be checked
2278 * @vma - vm area where page mapped
2279 * @addr - virtual address where page mapped
2280 *
2281 * Lookup current policy node id for vma,addr and "compare to" page's
2282 * node id.
2283 *
2284 * Returns:
2285 * -1 - not misplaced, page is in the right node
2286 * node - node id where the page should be
2287 *
2288 * Policy determination "mimics" alloc_page_vma().
2289 * Called from fault path where we know the vma and faulting address.
2290 */
2291int mpol_misplaced(struct page *page, struct vm_area_struct *vma, unsigned long addr)
2292{
2293 struct mempolicy *pol;
2294 struct zone *zone;
2295 int curnid = page_to_nid(page);
2296 unsigned long pgoff;
2297 int thiscpu = raw_smp_processor_id();
2298 int thisnid = cpu_to_node(thiscpu);
2299 int polnid = -1;
2300 int ret = -1;
2301
2302 BUG_ON(!vma);
2303
2304 pol = get_vma_policy(current, vma, addr);
2305 if (!(pol->flags & MPOL_F_MOF))
2306 goto out;
2307
2308 switch (pol->mode) {
2309 case MPOL_INTERLEAVE:
2310 BUG_ON(addr >= vma->vm_end);
2311 BUG_ON(addr < vma->vm_start);
2312
2313 pgoff = vma->vm_pgoff;
2314 pgoff += (addr - vma->vm_start) >> PAGE_SHIFT;
2315 polnid = offset_il_node(pol, vma, pgoff);
2316 break;
2317
2318 case MPOL_PREFERRED:
2319 if (pol->flags & MPOL_F_LOCAL)
2320 polnid = numa_node_id();
2321 else
2322 polnid = pol->v.preferred_node;
2323 break;
2324
2325 case MPOL_BIND:
2326 /*
2327 * allows binding to multiple nodes.
2328 * use current page if in policy nodemask,
2329 * else select nearest allowed node, if any.
2330 * If no allowed nodes, use current [!misplaced].
2331 */
2332 if (node_isset(curnid, pol->v.nodes))
2333 goto out;
2334 (void)first_zones_zonelist(
2335 node_zonelist(numa_node_id(), GFP_HIGHUSER),
2336 gfp_zone(GFP_HIGHUSER),
2337 &pol->v.nodes, &zone);
2338 polnid = zone->node;
2339 break;
2340
2341 default:
2342 BUG();
2343 }
2344
2345 /* Migrate the page towards the node whose CPU is referencing it */
2346 if (pol->flags & MPOL_F_MORON) {
2347 polnid = thisnid;
2348
2349 if (!should_numa_migrate_memory(current, page, curnid, thiscpu))
2350 goto out;
2351 }
2352
2353 if (curnid != polnid)
2354 ret = polnid;
2355out:
2356 mpol_cond_put(pol);
2357
2358 return ret;
2359}
2360
2361static void sp_delete(struct shared_policy *sp, struct sp_node *n)
2362{
2363 pr_debug("deleting %lx-l%lx\n", n->start, n->end);
2364 rb_erase(&n->nd, &sp->root);
2365 sp_free(n);
2366}
2367
2368static void sp_node_init(struct sp_node *node, unsigned long start,
2369 unsigned long end, struct mempolicy *pol)
2370{
2371 node->start = start;
2372 node->end = end;
2373 node->policy = pol;
2374}
2375
2376static struct sp_node *sp_alloc(unsigned long start, unsigned long end,
2377 struct mempolicy *pol)
2378{
2379 struct sp_node *n;
2380 struct mempolicy *newpol;
2381
2382 n = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2383 if (!n)
2384 return NULL;
2385
2386 newpol = mpol_dup(pol);
2387 if (IS_ERR(newpol)) {
2388 kmem_cache_free(sn_cache, n);
2389 return NULL;
2390 }
2391 newpol->flags |= MPOL_F_SHARED;
2392 sp_node_init(n, start, end, newpol);
2393
2394 return n;
2395}
2396
2397/* Replace a policy range. */
2398static int shared_policy_replace(struct shared_policy *sp, unsigned long start,
2399 unsigned long end, struct sp_node *new)
2400{
2401 struct sp_node *n;
2402 struct sp_node *n_new = NULL;
2403 struct mempolicy *mpol_new = NULL;
2404 int ret = 0;
2405
2406restart:
2407 spin_lock(&sp->lock);
2408 n = sp_lookup(sp, start, end);
2409 /* Take care of old policies in the same range. */
2410 while (n && n->start < end) {
2411 struct rb_node *next = rb_next(&n->nd);
2412 if (n->start >= start) {
2413 if (n->end <= end)
2414 sp_delete(sp, n);
2415 else
2416 n->start = end;
2417 } else {
2418 /* Old policy spanning whole new range. */
2419 if (n->end > end) {
2420 if (!n_new)
2421 goto alloc_new;
2422
2423 *mpol_new = *n->policy;
2424 atomic_set(&mpol_new->refcnt, 1);
2425 sp_node_init(n_new, end, n->end, mpol_new);
2426 n->end = start;
2427 sp_insert(sp, n_new);
2428 n_new = NULL;
2429 mpol_new = NULL;
2430 break;
2431 } else
2432 n->end = start;
2433 }
2434 if (!next)
2435 break;
2436 n = rb_entry(next, struct sp_node, nd);
2437 }
2438 if (new)
2439 sp_insert(sp, new);
2440 spin_unlock(&sp->lock);
2441 ret = 0;
2442
2443err_out:
2444 if (mpol_new)
2445 mpol_put(mpol_new);
2446 if (n_new)
2447 kmem_cache_free(sn_cache, n_new);
2448
2449 return ret;
2450
2451alloc_new:
2452 spin_unlock(&sp->lock);
2453 ret = -ENOMEM;
2454 n_new = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2455 if (!n_new)
2456 goto err_out;
2457 mpol_new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
2458 if (!mpol_new)
2459 goto err_out;
2460 goto restart;
2461}
2462
2463/**
2464 * mpol_shared_policy_init - initialize shared policy for inode
2465 * @sp: pointer to inode shared policy
2466 * @mpol: struct mempolicy to install
2467 *
2468 * Install non-NULL @mpol in inode's shared policy rb-tree.
2469 * On entry, the current task has a reference on a non-NULL @mpol.
2470 * This must be released on exit.
2471 * This is called at get_inode() calls and we can use GFP_KERNEL.
2472 */
2473void mpol_shared_policy_init(struct shared_policy *sp, struct mempolicy *mpol)
2474{
2475 int ret;
2476
2477 sp->root = RB_ROOT; /* empty tree == default mempolicy */
2478 spin_lock_init(&sp->lock);
2479
2480 if (mpol) {
2481 struct vm_area_struct pvma;
2482 struct mempolicy *new;
2483 NODEMASK_SCRATCH(scratch);
2484
2485 if (!scratch)
2486 goto put_mpol;
2487 /* contextualize the tmpfs mount point mempolicy */
2488 new = mpol_new(mpol->mode, mpol->flags, &mpol->w.user_nodemask);
2489 if (IS_ERR(new))
2490 goto free_scratch; /* no valid nodemask intersection */
2491
2492 task_lock(current);
2493 ret = mpol_set_nodemask(new, &mpol->w.user_nodemask, scratch);
2494 task_unlock(current);
2495 if (ret)
2496 goto put_new;
2497
2498 /* Create pseudo-vma that contains just the policy */
2499 memset(&pvma, 0, sizeof(struct vm_area_struct));
2500 pvma.vm_end = TASK_SIZE; /* policy covers entire file */
2501 mpol_set_shared_policy(sp, &pvma, new); /* adds ref */
2502
2503put_new:
2504 mpol_put(new); /* drop initial ref */
2505free_scratch:
2506 NODEMASK_SCRATCH_FREE(scratch);
2507put_mpol:
2508 mpol_put(mpol); /* drop our incoming ref on sb mpol */
2509 }
2510}
2511
2512int mpol_set_shared_policy(struct shared_policy *info,
2513 struct vm_area_struct *vma, struct mempolicy *npol)
2514{
2515 int err;
2516 struct sp_node *new = NULL;
2517 unsigned long sz = vma_pages(vma);
2518
2519 pr_debug("set_shared_policy %lx sz %lu %d %d %lx\n",
2520 vma->vm_pgoff,
2521 sz, npol ? npol->mode : -1,
2522 npol ? npol->flags : -1,
2523 npol ? nodes_addr(npol->v.nodes)[0] : NUMA_NO_NODE);
2524
2525 if (npol) {
2526 new = sp_alloc(vma->vm_pgoff, vma->vm_pgoff + sz, npol);
2527 if (!new)
2528 return -ENOMEM;
2529 }
2530 err = shared_policy_replace(info, vma->vm_pgoff, vma->vm_pgoff+sz, new);
2531 if (err && new)
2532 sp_free(new);
2533 return err;
2534}
2535
2536/* Free a backing policy store on inode delete. */
2537void mpol_free_shared_policy(struct shared_policy *p)
2538{
2539 struct sp_node *n;
2540 struct rb_node *next;
2541
2542 if (!p->root.rb_node)
2543 return;
2544 spin_lock(&p->lock);
2545 next = rb_first(&p->root);
2546 while (next) {
2547 n = rb_entry(next, struct sp_node, nd);
2548 next = rb_next(&n->nd);
2549 sp_delete(p, n);
2550 }
2551 spin_unlock(&p->lock);
2552}
2553
2554#ifdef CONFIG_NUMA_BALANCING
2555static int __initdata numabalancing_override;
2556
2557static void __init check_numabalancing_enable(void)
2558{
2559 bool numabalancing_default = false;
2560
2561 if (IS_ENABLED(CONFIG_NUMA_BALANCING_DEFAULT_ENABLED))
2562 numabalancing_default = true;
2563
2564 /* Parsed by setup_numabalancing. override == 1 enables, -1 disables */
2565 if (numabalancing_override)
2566 set_numabalancing_state(numabalancing_override == 1);
2567
2568 if (nr_node_ids > 1 && !numabalancing_override) {
2569 pr_info("%s automatic NUMA balancing. "
2570 "Configure with numa_balancing= or the "
2571 "kernel.numa_balancing sysctl",
2572 numabalancing_default ? "Enabling" : "Disabling");
2573 set_numabalancing_state(numabalancing_default);
2574 }
2575}
2576
2577static int __init setup_numabalancing(char *str)
2578{
2579 int ret = 0;
2580 if (!str)
2581 goto out;
2582
2583 if (!strcmp(str, "enable")) {
2584 numabalancing_override = 1;
2585 ret = 1;
2586 } else if (!strcmp(str, "disable")) {
2587 numabalancing_override = -1;
2588 ret = 1;
2589 }
2590out:
2591 if (!ret)
2592 pr_warn("Unable to parse numa_balancing=\n");
2593
2594 return ret;
2595}
2596__setup("numa_balancing=", setup_numabalancing);
2597#else
2598static inline void __init check_numabalancing_enable(void)
2599{
2600}
2601#endif /* CONFIG_NUMA_BALANCING */
2602
2603/* assumes fs == KERNEL_DS */
2604void __init numa_policy_init(void)
2605{
2606 nodemask_t interleave_nodes;
2607 unsigned long largest = 0;
2608 int nid, prefer = 0;
2609
2610 policy_cache = kmem_cache_create("numa_policy",
2611 sizeof(struct mempolicy),
2612 0, SLAB_PANIC, NULL);
2613
2614 sn_cache = kmem_cache_create("shared_policy_node",
2615 sizeof(struct sp_node),
2616 0, SLAB_PANIC, NULL);
2617
2618 for_each_node(nid) {
2619 preferred_node_policy[nid] = (struct mempolicy) {
2620 .refcnt = ATOMIC_INIT(1),
2621 .mode = MPOL_PREFERRED,
2622 .flags = MPOL_F_MOF | MPOL_F_MORON,
2623 .v = { .preferred_node = nid, },
2624 };
2625 }
2626
2627 /*
2628 * Set interleaving policy for system init. Interleaving is only
2629 * enabled across suitably sized nodes (default is >= 16MB), or
2630 * fall back to the largest node if they're all smaller.
2631 */
2632 nodes_clear(interleave_nodes);
2633 for_each_node_state(nid, N_MEMORY) {
2634 unsigned long total_pages = node_present_pages(nid);
2635
2636 /* Preserve the largest node */
2637 if (largest < total_pages) {
2638 largest = total_pages;
2639 prefer = nid;
2640 }
2641
2642 /* Interleave this node? */
2643 if ((total_pages << PAGE_SHIFT) >= (16 << 20))
2644 node_set(nid, interleave_nodes);
2645 }
2646
2647 /* All too small, use the largest */
2648 if (unlikely(nodes_empty(interleave_nodes)))
2649 node_set(prefer, interleave_nodes);
2650
2651 if (do_set_mempolicy(MPOL_INTERLEAVE, 0, &interleave_nodes))
2652 printk("numa_policy_init: interleaving failed\n");
2653
2654 check_numabalancing_enable();
2655}
2656
2657/* Reset policy of current process to default */
2658void numa_default_policy(void)
2659{
2660 do_set_mempolicy(MPOL_DEFAULT, 0, NULL);
2661}
2662
2663/*
2664 * Parse and format mempolicy from/to strings
2665 */
2666
2667/*
2668 * "local" is implemented internally by MPOL_PREFERRED with MPOL_F_LOCAL flag.
2669 */
2670static const char * const policy_modes[] =
2671{
2672 [MPOL_DEFAULT] = "default",
2673 [MPOL_PREFERRED] = "prefer",
2674 [MPOL_BIND] = "bind",
2675 [MPOL_INTERLEAVE] = "interleave",
2676 [MPOL_LOCAL] = "local",
2677};
2678
2679
2680#ifdef CONFIG_TMPFS
2681/**
2682 * mpol_parse_str - parse string to mempolicy, for tmpfs mpol mount option.
2683 * @str: string containing mempolicy to parse
2684 * @mpol: pointer to struct mempolicy pointer, returned on success.
2685 *
2686 * Format of input:
2687 * <mode>[=<flags>][:<nodelist>]
2688 *
2689 * On success, returns 0, else 1
2690 */
2691int mpol_parse_str(char *str, struct mempolicy **mpol)
2692{
2693 struct mempolicy *new = NULL;
2694 unsigned short mode;
2695 unsigned short mode_flags;
2696 nodemask_t nodes;
2697 char *nodelist = strchr(str, ':');
2698 char *flags = strchr(str, '=');
2699 int err = 1;
2700
2701 if (nodelist) {
2702 /* NUL-terminate mode or flags string */
2703 *nodelist++ = '\0';
2704 if (nodelist_parse(nodelist, nodes))
2705 goto out;
2706 if (!nodes_subset(nodes, node_states[N_MEMORY]))
2707 goto out;
2708 } else
2709 nodes_clear(nodes);
2710
2711 if (flags)
2712 *flags++ = '\0'; /* terminate mode string */
2713
2714 for (mode = 0; mode < MPOL_MAX; mode++) {
2715 if (!strcmp(str, policy_modes[mode])) {
2716 break;
2717 }
2718 }
2719 if (mode >= MPOL_MAX)
2720 goto out;
2721
2722 switch (mode) {
2723 case MPOL_PREFERRED:
2724 /*
2725 * Insist on a nodelist of one node only
2726 */
2727 if (nodelist) {
2728 char *rest = nodelist;
2729 while (isdigit(*rest))
2730 rest++;
2731 if (*rest)
2732 goto out;
2733 }
2734 break;
2735 case MPOL_INTERLEAVE:
2736 /*
2737 * Default to online nodes with memory if no nodelist
2738 */
2739 if (!nodelist)
2740 nodes = node_states[N_MEMORY];
2741 break;
2742 case MPOL_LOCAL:
2743 /*
2744 * Don't allow a nodelist; mpol_new() checks flags
2745 */
2746 if (nodelist)
2747 goto out;
2748 mode = MPOL_PREFERRED;
2749 break;
2750 case MPOL_DEFAULT:
2751 /*
2752 * Insist on a empty nodelist
2753 */
2754 if (!nodelist)
2755 err = 0;
2756 goto out;
2757 case MPOL_BIND:
2758 /*
2759 * Insist on a nodelist
2760 */
2761 if (!nodelist)
2762 goto out;
2763 }
2764
2765 mode_flags = 0;
2766 if (flags) {
2767 /*
2768 * Currently, we only support two mutually exclusive
2769 * mode flags.
2770 */
2771 if (!strcmp(flags, "static"))
2772 mode_flags |= MPOL_F_STATIC_NODES;
2773 else if (!strcmp(flags, "relative"))
2774 mode_flags |= MPOL_F_RELATIVE_NODES;
2775 else
2776 goto out;
2777 }
2778
2779 new = mpol_new(mode, mode_flags, &nodes);
2780 if (IS_ERR(new))
2781 goto out;
2782
2783 /*
2784 * Save nodes for mpol_to_str() to show the tmpfs mount options
2785 * for /proc/mounts, /proc/pid/mounts and /proc/pid/mountinfo.
2786 */
2787 if (mode != MPOL_PREFERRED)
2788 new->v.nodes = nodes;
2789 else if (nodelist)
2790 new->v.preferred_node = first_node(nodes);
2791 else
2792 new->flags |= MPOL_F_LOCAL;
2793
2794 /*
2795 * Save nodes for contextualization: this will be used to "clone"
2796 * the mempolicy in a specific context [cpuset] at a later time.
2797 */
2798 new->w.user_nodemask = nodes;
2799
2800 err = 0;
2801
2802out:
2803 /* Restore string for error message */
2804 if (nodelist)
2805 *--nodelist = ':';
2806 if (flags)
2807 *--flags = '=';
2808 if (!err)
2809 *mpol = new;
2810 return err;
2811}
2812#endif /* CONFIG_TMPFS */
2813
2814/**
2815 * mpol_to_str - format a mempolicy structure for printing
2816 * @buffer: to contain formatted mempolicy string
2817 * @maxlen: length of @buffer
2818 * @pol: pointer to mempolicy to be formatted
2819 *
2820 * Convert @pol into a string. If @buffer is too short, truncate the string.
2821 * Recommend a @maxlen of at least 32 for the longest mode, "interleave", the
2822 * longest flag, "relative", and to display at least a few node ids.
2823 */
2824void mpol_to_str(char *buffer, int maxlen, struct mempolicy *pol)
2825{
2826 char *p = buffer;
2827 nodemask_t nodes = NODE_MASK_NONE;
2828 unsigned short mode = MPOL_DEFAULT;
2829 unsigned short flags = 0;
2830
2831 if (pol && pol != &default_policy && !(pol->flags & MPOL_F_MORON)) {
2832 mode = pol->mode;
2833 flags = pol->flags;
2834 }
2835
2836 switch (mode) {
2837 case MPOL_DEFAULT:
2838 break;
2839 case MPOL_PREFERRED:
2840 if (flags & MPOL_F_LOCAL)
2841 mode = MPOL_LOCAL;
2842 else
2843 node_set(pol->v.preferred_node, nodes);
2844 break;
2845 case MPOL_BIND:
2846 case MPOL_INTERLEAVE:
2847 nodes = pol->v.nodes;
2848 break;
2849 default:
2850 WARN_ON_ONCE(1);
2851 snprintf(p, maxlen, "unknown");
2852 return;
2853 }
2854
2855 p += snprintf(p, maxlen, "%s", policy_modes[mode]);
2856
2857 if (flags & MPOL_MODE_FLAGS) {
2858 p += snprintf(p, buffer + maxlen - p, "=");
2859
2860 /*
2861 * Currently, the only defined flags are mutually exclusive
2862 */
2863 if (flags & MPOL_F_STATIC_NODES)
2864 p += snprintf(p, buffer + maxlen - p, "static");
2865 else if (flags & MPOL_F_RELATIVE_NODES)
2866 p += snprintf(p, buffer + maxlen - p, "relative");
2867 }
2868
2869 if (!nodes_empty(nodes)) {
2870 p += snprintf(p, buffer + maxlen - p, ":");
2871 p += nodelist_scnprintf(p, buffer + maxlen - p, nodes);
2872 }
2873}