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 node -1 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
94#include <asm/tlbflush.h>
95#include <asm/uaccess.h>
96#include <linux/random.h>
97
98#include "internal.h"
99
100/* Internal flags */
101#define MPOL_MF_DISCONTIG_OK (MPOL_MF_INTERNAL << 0) /* Skip checks for continuous vmas */
102#define MPOL_MF_INVERT (MPOL_MF_INTERNAL << 1) /* Invert check for nodemask */
103
104static struct kmem_cache *policy_cache;
105static struct kmem_cache *sn_cache;
106
107/* Highest zone. An specific allocation for a zone below that is not
108 policied. */
109enum zone_type policy_zone = 0;
110
111/*
112 * run-time system-wide default policy => local allocation
113 */
114static struct mempolicy default_policy = {
115 .refcnt = ATOMIC_INIT(1), /* never free it */
116 .mode = MPOL_PREFERRED,
117 .flags = MPOL_F_LOCAL,
118};
119
120static const struct mempolicy_operations {
121 int (*create)(struct mempolicy *pol, const nodemask_t *nodes);
122 /*
123 * If read-side task has no lock to protect task->mempolicy, write-side
124 * task will rebind the task->mempolicy by two step. The first step is
125 * setting all the newly nodes, and the second step is cleaning all the
126 * disallowed nodes. In this way, we can avoid finding no node to alloc
127 * page.
128 * If we have a lock to protect task->mempolicy in read-side, we do
129 * rebind directly.
130 *
131 * step:
132 * MPOL_REBIND_ONCE - do rebind work at once
133 * MPOL_REBIND_STEP1 - set all the newly nodes
134 * MPOL_REBIND_STEP2 - clean all the disallowed nodes
135 */
136 void (*rebind)(struct mempolicy *pol, const nodemask_t *nodes,
137 enum mpol_rebind_step step);
138} mpol_ops[MPOL_MAX];
139
140/* Check that the nodemask contains at least one populated zone */
141static int is_valid_nodemask(const nodemask_t *nodemask)
142{
143 int nd, k;
144
145 for_each_node_mask(nd, *nodemask) {
146 struct zone *z;
147
148 for (k = 0; k <= policy_zone; k++) {
149 z = &NODE_DATA(nd)->node_zones[k];
150 if (z->present_pages > 0)
151 return 1;
152 }
153 }
154
155 return 0;
156}
157
158static inline int mpol_store_user_nodemask(const struct mempolicy *pol)
159{
160 return pol->flags & MPOL_MODE_FLAGS;
161}
162
163static void mpol_relative_nodemask(nodemask_t *ret, const nodemask_t *orig,
164 const nodemask_t *rel)
165{
166 nodemask_t tmp;
167 nodes_fold(tmp, *orig, nodes_weight(*rel));
168 nodes_onto(*ret, tmp, *rel);
169}
170
171static int mpol_new_interleave(struct mempolicy *pol, const nodemask_t *nodes)
172{
173 if (nodes_empty(*nodes))
174 return -EINVAL;
175 pol->v.nodes = *nodes;
176 return 0;
177}
178
179static int mpol_new_preferred(struct mempolicy *pol, const nodemask_t *nodes)
180{
181 if (!nodes)
182 pol->flags |= MPOL_F_LOCAL; /* local allocation */
183 else if (nodes_empty(*nodes))
184 return -EINVAL; /* no allowed nodes */
185 else
186 pol->v.preferred_node = first_node(*nodes);
187 return 0;
188}
189
190static int mpol_new_bind(struct mempolicy *pol, const nodemask_t *nodes)
191{
192 if (!is_valid_nodemask(nodes))
193 return -EINVAL;
194 pol->v.nodes = *nodes;
195 return 0;
196}
197
198/*
199 * mpol_set_nodemask is called after mpol_new() to set up the nodemask, if
200 * any, for the new policy. mpol_new() has already validated the nodes
201 * parameter with respect to the policy mode and flags. But, we need to
202 * handle an empty nodemask with MPOL_PREFERRED here.
203 *
204 * Must be called holding task's alloc_lock to protect task's mems_allowed
205 * and mempolicy. May also be called holding the mmap_semaphore for write.
206 */
207static int mpol_set_nodemask(struct mempolicy *pol,
208 const nodemask_t *nodes, struct nodemask_scratch *nsc)
209{
210 int ret;
211
212 /* if mode is MPOL_DEFAULT, pol is NULL. This is right. */
213 if (pol == NULL)
214 return 0;
215 /* Check N_HIGH_MEMORY */
216 nodes_and(nsc->mask1,
217 cpuset_current_mems_allowed, node_states[N_HIGH_MEMORY]);
218
219 VM_BUG_ON(!nodes);
220 if (pol->mode == MPOL_PREFERRED && nodes_empty(*nodes))
221 nodes = NULL; /* explicit local allocation */
222 else {
223 if (pol->flags & MPOL_F_RELATIVE_NODES)
224 mpol_relative_nodemask(&nsc->mask2, nodes,&nsc->mask1);
225 else
226 nodes_and(nsc->mask2, *nodes, nsc->mask1);
227
228 if (mpol_store_user_nodemask(pol))
229 pol->w.user_nodemask = *nodes;
230 else
231 pol->w.cpuset_mems_allowed =
232 cpuset_current_mems_allowed;
233 }
234
235 if (nodes)
236 ret = mpol_ops[pol->mode].create(pol, &nsc->mask2);
237 else
238 ret = mpol_ops[pol->mode].create(pol, NULL);
239 return ret;
240}
241
242/*
243 * This function just creates a new policy, does some check and simple
244 * initialization. You must invoke mpol_set_nodemask() to set nodes.
245 */
246static struct mempolicy *mpol_new(unsigned short mode, unsigned short flags,
247 nodemask_t *nodes)
248{
249 struct mempolicy *policy;
250
251 pr_debug("setting mode %d flags %d nodes[0] %lx\n",
252 mode, flags, nodes ? nodes_addr(*nodes)[0] : -1);
253
254 if (mode == MPOL_DEFAULT) {
255 if (nodes && !nodes_empty(*nodes))
256 return ERR_PTR(-EINVAL);
257 return NULL; /* simply delete any existing policy */
258 }
259 VM_BUG_ON(!nodes);
260
261 /*
262 * MPOL_PREFERRED cannot be used with MPOL_F_STATIC_NODES or
263 * MPOL_F_RELATIVE_NODES if the nodemask is empty (local allocation).
264 * All other modes require a valid pointer to a non-empty nodemask.
265 */
266 if (mode == MPOL_PREFERRED) {
267 if (nodes_empty(*nodes)) {
268 if (((flags & MPOL_F_STATIC_NODES) ||
269 (flags & MPOL_F_RELATIVE_NODES)))
270 return ERR_PTR(-EINVAL);
271 }
272 } else if (nodes_empty(*nodes))
273 return ERR_PTR(-EINVAL);
274 policy = kmem_cache_alloc(policy_cache, GFP_KERNEL);
275 if (!policy)
276 return ERR_PTR(-ENOMEM);
277 atomic_set(&policy->refcnt, 1);
278 policy->mode = mode;
279 policy->flags = flags;
280
281 return policy;
282}
283
284/* Slow path of a mpol destructor. */
285void __mpol_put(struct mempolicy *p)
286{
287 if (!atomic_dec_and_test(&p->refcnt))
288 return;
289 kmem_cache_free(policy_cache, p);
290}
291
292static void mpol_rebind_default(struct mempolicy *pol, const nodemask_t *nodes,
293 enum mpol_rebind_step step)
294{
295}
296
297/*
298 * step:
299 * MPOL_REBIND_ONCE - do rebind work at once
300 * MPOL_REBIND_STEP1 - set all the newly nodes
301 * MPOL_REBIND_STEP2 - clean all the disallowed nodes
302 */
303static void mpol_rebind_nodemask(struct mempolicy *pol, const nodemask_t *nodes,
304 enum mpol_rebind_step step)
305{
306 nodemask_t tmp;
307
308 if (pol->flags & MPOL_F_STATIC_NODES)
309 nodes_and(tmp, pol->w.user_nodemask, *nodes);
310 else if (pol->flags & MPOL_F_RELATIVE_NODES)
311 mpol_relative_nodemask(&tmp, &pol->w.user_nodemask, nodes);
312 else {
313 /*
314 * if step == 1, we use ->w.cpuset_mems_allowed to cache the
315 * result
316 */
317 if (step == MPOL_REBIND_ONCE || step == MPOL_REBIND_STEP1) {
318 nodes_remap(tmp, pol->v.nodes,
319 pol->w.cpuset_mems_allowed, *nodes);
320 pol->w.cpuset_mems_allowed = step ? tmp : *nodes;
321 } else if (step == MPOL_REBIND_STEP2) {
322 tmp = pol->w.cpuset_mems_allowed;
323 pol->w.cpuset_mems_allowed = *nodes;
324 } else
325 BUG();
326 }
327
328 if (nodes_empty(tmp))
329 tmp = *nodes;
330
331 if (step == MPOL_REBIND_STEP1)
332 nodes_or(pol->v.nodes, pol->v.nodes, tmp);
333 else if (step == MPOL_REBIND_ONCE || step == MPOL_REBIND_STEP2)
334 pol->v.nodes = tmp;
335 else
336 BUG();
337
338 if (!node_isset(current->il_next, tmp)) {
339 current->il_next = next_node(current->il_next, tmp);
340 if (current->il_next >= MAX_NUMNODES)
341 current->il_next = first_node(tmp);
342 if (current->il_next >= MAX_NUMNODES)
343 current->il_next = numa_node_id();
344 }
345}
346
347static void mpol_rebind_preferred(struct mempolicy *pol,
348 const nodemask_t *nodes,
349 enum mpol_rebind_step step)
350{
351 nodemask_t tmp;
352
353 if (pol->flags & MPOL_F_STATIC_NODES) {
354 int node = first_node(pol->w.user_nodemask);
355
356 if (node_isset(node, *nodes)) {
357 pol->v.preferred_node = node;
358 pol->flags &= ~MPOL_F_LOCAL;
359 } else
360 pol->flags |= MPOL_F_LOCAL;
361 } else if (pol->flags & MPOL_F_RELATIVE_NODES) {
362 mpol_relative_nodemask(&tmp, &pol->w.user_nodemask, nodes);
363 pol->v.preferred_node = first_node(tmp);
364 } else if (!(pol->flags & MPOL_F_LOCAL)) {
365 pol->v.preferred_node = node_remap(pol->v.preferred_node,
366 pol->w.cpuset_mems_allowed,
367 *nodes);
368 pol->w.cpuset_mems_allowed = *nodes;
369 }
370}
371
372/*
373 * mpol_rebind_policy - Migrate a policy to a different set of nodes
374 *
375 * If read-side task has no lock to protect task->mempolicy, write-side
376 * task will rebind the task->mempolicy by two step. The first step is
377 * setting all the newly nodes, and the second step is cleaning all the
378 * disallowed nodes. In this way, we can avoid finding no node to alloc
379 * page.
380 * If we have a lock to protect task->mempolicy in read-side, we do
381 * rebind directly.
382 *
383 * step:
384 * MPOL_REBIND_ONCE - do rebind work at once
385 * MPOL_REBIND_STEP1 - set all the newly nodes
386 * MPOL_REBIND_STEP2 - clean all the disallowed nodes
387 */
388static void mpol_rebind_policy(struct mempolicy *pol, const nodemask_t *newmask,
389 enum mpol_rebind_step step)
390{
391 if (!pol)
392 return;
393 if (!mpol_store_user_nodemask(pol) && step == MPOL_REBIND_ONCE &&
394 nodes_equal(pol->w.cpuset_mems_allowed, *newmask))
395 return;
396
397 if (step == MPOL_REBIND_STEP1 && (pol->flags & MPOL_F_REBINDING))
398 return;
399
400 if (step == MPOL_REBIND_STEP2 && !(pol->flags & MPOL_F_REBINDING))
401 BUG();
402
403 if (step == MPOL_REBIND_STEP1)
404 pol->flags |= MPOL_F_REBINDING;
405 else if (step == MPOL_REBIND_STEP2)
406 pol->flags &= ~MPOL_F_REBINDING;
407 else if (step >= MPOL_REBIND_NSTEP)
408 BUG();
409
410 mpol_ops[pol->mode].rebind(pol, newmask, step);
411}
412
413/*
414 * Wrapper for mpol_rebind_policy() that just requires task
415 * pointer, and updates task mempolicy.
416 *
417 * Called with task's alloc_lock held.
418 */
419
420void mpol_rebind_task(struct task_struct *tsk, const nodemask_t *new,
421 enum mpol_rebind_step step)
422{
423 mpol_rebind_policy(tsk->mempolicy, new, step);
424}
425
426/*
427 * Rebind each vma in mm to new nodemask.
428 *
429 * Call holding a reference to mm. Takes mm->mmap_sem during call.
430 */
431
432void mpol_rebind_mm(struct mm_struct *mm, nodemask_t *new)
433{
434 struct vm_area_struct *vma;
435
436 down_write(&mm->mmap_sem);
437 for (vma = mm->mmap; vma; vma = vma->vm_next)
438 mpol_rebind_policy(vma->vm_policy, new, MPOL_REBIND_ONCE);
439 up_write(&mm->mmap_sem);
440}
441
442static const struct mempolicy_operations mpol_ops[MPOL_MAX] = {
443 [MPOL_DEFAULT] = {
444 .rebind = mpol_rebind_default,
445 },
446 [MPOL_INTERLEAVE] = {
447 .create = mpol_new_interleave,
448 .rebind = mpol_rebind_nodemask,
449 },
450 [MPOL_PREFERRED] = {
451 .create = mpol_new_preferred,
452 .rebind = mpol_rebind_preferred,
453 },
454 [MPOL_BIND] = {
455 .create = mpol_new_bind,
456 .rebind = mpol_rebind_nodemask,
457 },
458};
459
460static void migrate_page_add(struct page *page, struct list_head *pagelist,
461 unsigned long flags);
462
463/* Scan through pages checking if pages follow certain conditions. */
464static int check_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
465 unsigned long addr, unsigned long end,
466 const nodemask_t *nodes, unsigned long flags,
467 void *private)
468{
469 pte_t *orig_pte;
470 pte_t *pte;
471 spinlock_t *ptl;
472
473 orig_pte = pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
474 do {
475 struct page *page;
476 int nid;
477
478 if (!pte_present(*pte))
479 continue;
480 page = vm_normal_page(vma, addr, *pte);
481 if (!page)
482 continue;
483 /*
484 * vm_normal_page() filters out zero pages, but there might
485 * still be PageReserved pages to skip, perhaps in a VDSO.
486 * And we cannot move PageKsm pages sensibly or safely yet.
487 */
488 if (PageReserved(page) || PageKsm(page))
489 continue;
490 nid = page_to_nid(page);
491 if (node_isset(nid, *nodes) == !!(flags & MPOL_MF_INVERT))
492 continue;
493
494 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))
495 migrate_page_add(page, private, flags);
496 else
497 break;
498 } while (pte++, addr += PAGE_SIZE, addr != end);
499 pte_unmap_unlock(orig_pte, ptl);
500 return addr != end;
501}
502
503static inline int check_pmd_range(struct vm_area_struct *vma, pud_t *pud,
504 unsigned long addr, unsigned long end,
505 const nodemask_t *nodes, unsigned long flags,
506 void *private)
507{
508 pmd_t *pmd;
509 unsigned long next;
510
511 pmd = pmd_offset(pud, addr);
512 do {
513 next = pmd_addr_end(addr, end);
514 split_huge_page_pmd(vma->vm_mm, pmd);
515 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
516 continue;
517 if (check_pte_range(vma, pmd, addr, next, nodes,
518 flags, private))
519 return -EIO;
520 } while (pmd++, addr = next, addr != end);
521 return 0;
522}
523
524static inline int check_pud_range(struct vm_area_struct *vma, pgd_t *pgd,
525 unsigned long addr, unsigned long end,
526 const nodemask_t *nodes, unsigned long flags,
527 void *private)
528{
529 pud_t *pud;
530 unsigned long next;
531
532 pud = pud_offset(pgd, addr);
533 do {
534 next = pud_addr_end(addr, end);
535 if (pud_none_or_clear_bad(pud))
536 continue;
537 if (check_pmd_range(vma, pud, addr, next, nodes,
538 flags, private))
539 return -EIO;
540 } while (pud++, addr = next, addr != end);
541 return 0;
542}
543
544static inline int check_pgd_range(struct vm_area_struct *vma,
545 unsigned long addr, unsigned long end,
546 const nodemask_t *nodes, unsigned long flags,
547 void *private)
548{
549 pgd_t *pgd;
550 unsigned long next;
551
552 pgd = pgd_offset(vma->vm_mm, addr);
553 do {
554 next = pgd_addr_end(addr, end);
555 if (pgd_none_or_clear_bad(pgd))
556 continue;
557 if (check_pud_range(vma, pgd, addr, next, nodes,
558 flags, private))
559 return -EIO;
560 } while (pgd++, addr = next, addr != end);
561 return 0;
562}
563
564/*
565 * Check if all pages in a range are on a set of nodes.
566 * If pagelist != NULL then isolate pages from the LRU and
567 * put them on the pagelist.
568 */
569static struct vm_area_struct *
570check_range(struct mm_struct *mm, unsigned long start, unsigned long end,
571 const nodemask_t *nodes, unsigned long flags, void *private)
572{
573 int err;
574 struct vm_area_struct *first, *vma, *prev;
575
576
577 first = find_vma(mm, start);
578 if (!first)
579 return ERR_PTR(-EFAULT);
580 prev = NULL;
581 for (vma = first; vma && vma->vm_start < end; vma = vma->vm_next) {
582 if (!(flags & MPOL_MF_DISCONTIG_OK)) {
583 if (!vma->vm_next && vma->vm_end < end)
584 return ERR_PTR(-EFAULT);
585 if (prev && prev->vm_end < vma->vm_start)
586 return ERR_PTR(-EFAULT);
587 }
588 if (!is_vm_hugetlb_page(vma) &&
589 ((flags & MPOL_MF_STRICT) ||
590 ((flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) &&
591 vma_migratable(vma)))) {
592 unsigned long endvma = vma->vm_end;
593
594 if (endvma > end)
595 endvma = end;
596 if (vma->vm_start > start)
597 start = vma->vm_start;
598 err = check_pgd_range(vma, start, endvma, nodes,
599 flags, private);
600 if (err) {
601 first = ERR_PTR(err);
602 break;
603 }
604 }
605 prev = vma;
606 }
607 return first;
608}
609
610/* Step 2: apply policy to a range and do splits. */
611static int mbind_range(struct mm_struct *mm, unsigned long start,
612 unsigned long end, struct mempolicy *new_pol)
613{
614 struct vm_area_struct *next;
615 struct vm_area_struct *prev;
616 struct vm_area_struct *vma;
617 int err = 0;
618 pgoff_t pgoff;
619 unsigned long vmstart;
620 unsigned long vmend;
621
622 vma = find_vma(mm, start);
623 if (!vma || vma->vm_start > start)
624 return -EFAULT;
625
626 prev = vma->vm_prev;
627 if (start > vma->vm_start)
628 prev = vma;
629
630 for (; vma && vma->vm_start < end; prev = vma, vma = next) {
631 next = vma->vm_next;
632 vmstart = max(start, vma->vm_start);
633 vmend = min(end, vma->vm_end);
634
635 if (mpol_equal(vma_policy(vma), new_pol))
636 continue;
637
638 pgoff = vma->vm_pgoff +
639 ((vmstart - vma->vm_start) >> PAGE_SHIFT);
640 prev = vma_merge(mm, prev, vmstart, vmend, vma->vm_flags,
641 vma->anon_vma, vma->vm_file, pgoff,
642 new_pol);
643 if (prev) {
644 vma = prev;
645 next = vma->vm_next;
646 continue;
647 }
648 if (vma->vm_start != vmstart) {
649 err = split_vma(vma->vm_mm, vma, vmstart, 1);
650 if (err)
651 goto out;
652 }
653 if (vma->vm_end != vmend) {
654 err = split_vma(vma->vm_mm, vma, vmend, 0);
655 if (err)
656 goto out;
657 }
658
659 /*
660 * Apply policy to a single VMA. The reference counting of
661 * policy for vma_policy linkages has already been handled by
662 * vma_merge and split_vma as necessary. If this is a shared
663 * policy then ->set_policy will increment the reference count
664 * for an sp node.
665 */
666 pr_debug("vma %lx-%lx/%lx vm_ops %p vm_file %p set_policy %p\n",
667 vma->vm_start, vma->vm_end, vma->vm_pgoff,
668 vma->vm_ops, vma->vm_file,
669 vma->vm_ops ? vma->vm_ops->set_policy : NULL);
670 if (vma->vm_ops && vma->vm_ops->set_policy) {
671 err = vma->vm_ops->set_policy(vma, new_pol);
672 if (err)
673 goto out;
674 }
675 }
676
677 out:
678 return err;
679}
680
681/*
682 * Update task->flags PF_MEMPOLICY bit: set iff non-default
683 * mempolicy. Allows more rapid checking of this (combined perhaps
684 * with other PF_* flag bits) on memory allocation hot code paths.
685 *
686 * If called from outside this file, the task 'p' should -only- be
687 * a newly forked child not yet visible on the task list, because
688 * manipulating the task flags of a visible task is not safe.
689 *
690 * The above limitation is why this routine has the funny name
691 * mpol_fix_fork_child_flag().
692 *
693 * It is also safe to call this with a task pointer of current,
694 * which the static wrapper mpol_set_task_struct_flag() does,
695 * for use within this file.
696 */
697
698void mpol_fix_fork_child_flag(struct task_struct *p)
699{
700 if (p->mempolicy)
701 p->flags |= PF_MEMPOLICY;
702 else
703 p->flags &= ~PF_MEMPOLICY;
704}
705
706static void mpol_set_task_struct_flag(void)
707{
708 mpol_fix_fork_child_flag(current);
709}
710
711/* Set the process memory policy */
712static long do_set_mempolicy(unsigned short mode, unsigned short flags,
713 nodemask_t *nodes)
714{
715 struct mempolicy *new, *old;
716 struct mm_struct *mm = current->mm;
717 NODEMASK_SCRATCH(scratch);
718 int ret;
719
720 if (!scratch)
721 return -ENOMEM;
722
723 new = mpol_new(mode, flags, nodes);
724 if (IS_ERR(new)) {
725 ret = PTR_ERR(new);
726 goto out;
727 }
728 /*
729 * prevent changing our mempolicy while show_numa_maps()
730 * is using it.
731 * Note: do_set_mempolicy() can be called at init time
732 * with no 'mm'.
733 */
734 if (mm)
735 down_write(&mm->mmap_sem);
736 task_lock(current);
737 ret = mpol_set_nodemask(new, nodes, scratch);
738 if (ret) {
739 task_unlock(current);
740 if (mm)
741 up_write(&mm->mmap_sem);
742 mpol_put(new);
743 goto out;
744 }
745 old = current->mempolicy;
746 current->mempolicy = new;
747 mpol_set_task_struct_flag();
748 if (new && new->mode == MPOL_INTERLEAVE &&
749 nodes_weight(new->v.nodes))
750 current->il_next = first_node(new->v.nodes);
751 task_unlock(current);
752 if (mm)
753 up_write(&mm->mmap_sem);
754
755 mpol_put(old);
756 ret = 0;
757out:
758 NODEMASK_SCRATCH_FREE(scratch);
759 return ret;
760}
761
762/*
763 * Return nodemask for policy for get_mempolicy() query
764 *
765 * Called with task's alloc_lock held
766 */
767static void get_policy_nodemask(struct mempolicy *p, nodemask_t *nodes)
768{
769 nodes_clear(*nodes);
770 if (p == &default_policy)
771 return;
772
773 switch (p->mode) {
774 case MPOL_BIND:
775 /* Fall through */
776 case MPOL_INTERLEAVE:
777 *nodes = p->v.nodes;
778 break;
779 case MPOL_PREFERRED:
780 if (!(p->flags & MPOL_F_LOCAL))
781 node_set(p->v.preferred_node, *nodes);
782 /* else return empty node mask for local allocation */
783 break;
784 default:
785 BUG();
786 }
787}
788
789static int lookup_node(struct mm_struct *mm, unsigned long addr)
790{
791 struct page *p;
792 int err;
793
794 err = get_user_pages(current, mm, addr & PAGE_MASK, 1, 0, 0, &p, NULL);
795 if (err >= 0) {
796 err = page_to_nid(p);
797 put_page(p);
798 }
799 return err;
800}
801
802/* Retrieve NUMA policy */
803static long do_get_mempolicy(int *policy, nodemask_t *nmask,
804 unsigned long addr, unsigned long flags)
805{
806 int err;
807 struct mm_struct *mm = current->mm;
808 struct vm_area_struct *vma = NULL;
809 struct mempolicy *pol = current->mempolicy;
810
811 if (flags &
812 ~(unsigned long)(MPOL_F_NODE|MPOL_F_ADDR|MPOL_F_MEMS_ALLOWED))
813 return -EINVAL;
814
815 if (flags & MPOL_F_MEMS_ALLOWED) {
816 if (flags & (MPOL_F_NODE|MPOL_F_ADDR))
817 return -EINVAL;
818 *policy = 0; /* just so it's initialized */
819 task_lock(current);
820 *nmask = cpuset_current_mems_allowed;
821 task_unlock(current);
822 return 0;
823 }
824
825 if (flags & MPOL_F_ADDR) {
826 /*
827 * Do NOT fall back to task policy if the
828 * vma/shared policy at addr is NULL. We
829 * want to return MPOL_DEFAULT in this case.
830 */
831 down_read(&mm->mmap_sem);
832 vma = find_vma_intersection(mm, addr, addr+1);
833 if (!vma) {
834 up_read(&mm->mmap_sem);
835 return -EFAULT;
836 }
837 if (vma->vm_ops && vma->vm_ops->get_policy)
838 pol = vma->vm_ops->get_policy(vma, addr);
839 else
840 pol = vma->vm_policy;
841 } else if (addr)
842 return -EINVAL;
843
844 if (!pol)
845 pol = &default_policy; /* indicates default behavior */
846
847 if (flags & MPOL_F_NODE) {
848 if (flags & MPOL_F_ADDR) {
849 err = lookup_node(mm, addr);
850 if (err < 0)
851 goto out;
852 *policy = err;
853 } else if (pol == current->mempolicy &&
854 pol->mode == MPOL_INTERLEAVE) {
855 *policy = current->il_next;
856 } else {
857 err = -EINVAL;
858 goto out;
859 }
860 } else {
861 *policy = pol == &default_policy ? MPOL_DEFAULT :
862 pol->mode;
863 /*
864 * Internal mempolicy flags must be masked off before exposing
865 * the policy to userspace.
866 */
867 *policy |= (pol->flags & MPOL_MODE_FLAGS);
868 }
869
870 if (vma) {
871 up_read(¤t->mm->mmap_sem);
872 vma = NULL;
873 }
874
875 err = 0;
876 if (nmask) {
877 if (mpol_store_user_nodemask(pol)) {
878 *nmask = pol->w.user_nodemask;
879 } else {
880 task_lock(current);
881 get_policy_nodemask(pol, nmask);
882 task_unlock(current);
883 }
884 }
885
886 out:
887 mpol_cond_put(pol);
888 if (vma)
889 up_read(¤t->mm->mmap_sem);
890 return err;
891}
892
893#ifdef CONFIG_MIGRATION
894/*
895 * page migration
896 */
897static void migrate_page_add(struct page *page, struct list_head *pagelist,
898 unsigned long flags)
899{
900 /*
901 * Avoid migrating a page that is shared with others.
902 */
903 if ((flags & MPOL_MF_MOVE_ALL) || page_mapcount(page) == 1) {
904 if (!isolate_lru_page(page)) {
905 list_add_tail(&page->lru, pagelist);
906 inc_zone_page_state(page, NR_ISOLATED_ANON +
907 page_is_file_cache(page));
908 }
909 }
910}
911
912static struct page *new_node_page(struct page *page, unsigned long node, int **x)
913{
914 return alloc_pages_exact_node(node, GFP_HIGHUSER_MOVABLE, 0);
915}
916
917/*
918 * Migrate pages from one node to a target node.
919 * Returns error or the number of pages not migrated.
920 */
921static int migrate_to_node(struct mm_struct *mm, int source, int dest,
922 int flags)
923{
924 nodemask_t nmask;
925 LIST_HEAD(pagelist);
926 int err = 0;
927 struct vm_area_struct *vma;
928
929 nodes_clear(nmask);
930 node_set(source, nmask);
931
932 vma = check_range(mm, mm->mmap->vm_start, mm->task_size, &nmask,
933 flags | MPOL_MF_DISCONTIG_OK, &pagelist);
934 if (IS_ERR(vma))
935 return PTR_ERR(vma);
936
937 if (!list_empty(&pagelist)) {
938 err = migrate_pages(&pagelist, new_node_page, dest,
939 false, MIGRATE_SYNC);
940 if (err)
941 putback_lru_pages(&pagelist);
942 }
943
944 return err;
945}
946
947/*
948 * Move pages between the two nodesets so as to preserve the physical
949 * layout as much as possible.
950 *
951 * Returns the number of page that could not be moved.
952 */
953int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
954 const nodemask_t *to, int flags)
955{
956 int busy = 0;
957 int err;
958 nodemask_t tmp;
959
960 err = migrate_prep();
961 if (err)
962 return err;
963
964 down_read(&mm->mmap_sem);
965
966 err = migrate_vmas(mm, from, to, flags);
967 if (err)
968 goto out;
969
970 /*
971 * Find a 'source' bit set in 'tmp' whose corresponding 'dest'
972 * bit in 'to' is not also set in 'tmp'. Clear the found 'source'
973 * bit in 'tmp', and return that <source, dest> pair for migration.
974 * The pair of nodemasks 'to' and 'from' define the map.
975 *
976 * If no pair of bits is found that way, fallback to picking some
977 * pair of 'source' and 'dest' bits that are not the same. If the
978 * 'source' and 'dest' bits are the same, this represents a node
979 * that will be migrating to itself, so no pages need move.
980 *
981 * If no bits are left in 'tmp', or if all remaining bits left
982 * in 'tmp' correspond to the same bit in 'to', return false
983 * (nothing left to migrate).
984 *
985 * This lets us pick a pair of nodes to migrate between, such that
986 * if possible the dest node is not already occupied by some other
987 * source node, minimizing the risk of overloading the memory on a
988 * node that would happen if we migrated incoming memory to a node
989 * before migrating outgoing memory source that same node.
990 *
991 * A single scan of tmp is sufficient. As we go, we remember the
992 * most recent <s, d> pair that moved (s != d). If we find a pair
993 * that not only moved, but what's better, moved to an empty slot
994 * (d is not set in tmp), then we break out then, with that pair.
995 * Otherwise when we finish scanning from_tmp, we at least have the
996 * most recent <s, d> pair that moved. If we get all the way through
997 * the scan of tmp without finding any node that moved, much less
998 * moved to an empty node, then there is nothing left worth migrating.
999 */
1000
1001 tmp = *from;
1002 while (!nodes_empty(tmp)) {
1003 int s,d;
1004 int source = -1;
1005 int dest = 0;
1006
1007 for_each_node_mask(s, tmp) {
1008
1009 /*
1010 * do_migrate_pages() tries to maintain the relative
1011 * node relationship of the pages established between
1012 * threads and memory areas.
1013 *
1014 * However if the number of source nodes is not equal to
1015 * the number of destination nodes we can not preserve
1016 * this node relative relationship. In that case, skip
1017 * copying memory from a node that is in the destination
1018 * mask.
1019 *
1020 * Example: [2,3,4] -> [3,4,5] moves everything.
1021 * [0-7] - > [3,4,5] moves only 0,1,2,6,7.
1022 */
1023
1024 if ((nodes_weight(*from) != nodes_weight(*to)) &&
1025 (node_isset(s, *to)))
1026 continue;
1027
1028 d = node_remap(s, *from, *to);
1029 if (s == d)
1030 continue;
1031
1032 source = s; /* Node moved. Memorize */
1033 dest = d;
1034
1035 /* dest not in remaining from nodes? */
1036 if (!node_isset(dest, tmp))
1037 break;
1038 }
1039 if (source == -1)
1040 break;
1041
1042 node_clear(source, tmp);
1043 err = migrate_to_node(mm, source, dest, flags);
1044 if (err > 0)
1045 busy += err;
1046 if (err < 0)
1047 break;
1048 }
1049out:
1050 up_read(&mm->mmap_sem);
1051 if (err < 0)
1052 return err;
1053 return busy;
1054
1055}
1056
1057/*
1058 * Allocate a new page for page migration based on vma policy.
1059 * Start assuming that page is mapped by vma pointed to by @private.
1060 * Search forward from there, if not. N.B., this assumes that the
1061 * list of pages handed to migrate_pages()--which is how we get here--
1062 * is in virtual address order.
1063 */
1064static struct page *new_vma_page(struct page *page, unsigned long private, int **x)
1065{
1066 struct vm_area_struct *vma = (struct vm_area_struct *)private;
1067 unsigned long uninitialized_var(address);
1068
1069 while (vma) {
1070 address = page_address_in_vma(page, vma);
1071 if (address != -EFAULT)
1072 break;
1073 vma = vma->vm_next;
1074 }
1075
1076 /*
1077 * if !vma, alloc_page_vma() will use task or system default policy
1078 */
1079 return alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
1080}
1081#else
1082
1083static void migrate_page_add(struct page *page, struct list_head *pagelist,
1084 unsigned long flags)
1085{
1086}
1087
1088int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
1089 const nodemask_t *to, int flags)
1090{
1091 return -ENOSYS;
1092}
1093
1094static struct page *new_vma_page(struct page *page, unsigned long private, int **x)
1095{
1096 return NULL;
1097}
1098#endif
1099
1100static long do_mbind(unsigned long start, unsigned long len,
1101 unsigned short mode, unsigned short mode_flags,
1102 nodemask_t *nmask, unsigned long flags)
1103{
1104 struct vm_area_struct *vma;
1105 struct mm_struct *mm = current->mm;
1106 struct mempolicy *new;
1107 unsigned long end;
1108 int err;
1109 LIST_HEAD(pagelist);
1110
1111 if (flags & ~(unsigned long)(MPOL_MF_STRICT |
1112 MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))
1113 return -EINVAL;
1114 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1115 return -EPERM;
1116
1117 if (start & ~PAGE_MASK)
1118 return -EINVAL;
1119
1120 if (mode == MPOL_DEFAULT)
1121 flags &= ~MPOL_MF_STRICT;
1122
1123 len = (len + PAGE_SIZE - 1) & PAGE_MASK;
1124 end = start + len;
1125
1126 if (end < start)
1127 return -EINVAL;
1128 if (end == start)
1129 return 0;
1130
1131 new = mpol_new(mode, mode_flags, nmask);
1132 if (IS_ERR(new))
1133 return PTR_ERR(new);
1134
1135 /*
1136 * If we are using the default policy then operation
1137 * on discontinuous address spaces is okay after all
1138 */
1139 if (!new)
1140 flags |= MPOL_MF_DISCONTIG_OK;
1141
1142 pr_debug("mbind %lx-%lx mode:%d flags:%d nodes:%lx\n",
1143 start, start + len, mode, mode_flags,
1144 nmask ? nodes_addr(*nmask)[0] : -1);
1145
1146 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) {
1147
1148 err = migrate_prep();
1149 if (err)
1150 goto mpol_out;
1151 }
1152 {
1153 NODEMASK_SCRATCH(scratch);
1154 if (scratch) {
1155 down_write(&mm->mmap_sem);
1156 task_lock(current);
1157 err = mpol_set_nodemask(new, nmask, scratch);
1158 task_unlock(current);
1159 if (err)
1160 up_write(&mm->mmap_sem);
1161 } else
1162 err = -ENOMEM;
1163 NODEMASK_SCRATCH_FREE(scratch);
1164 }
1165 if (err)
1166 goto mpol_out;
1167
1168 vma = check_range(mm, start, end, nmask,
1169 flags | MPOL_MF_INVERT, &pagelist);
1170
1171 err = PTR_ERR(vma);
1172 if (!IS_ERR(vma)) {
1173 int nr_failed = 0;
1174
1175 err = mbind_range(mm, start, end, new);
1176
1177 if (!list_empty(&pagelist)) {
1178 nr_failed = migrate_pages(&pagelist, new_vma_page,
1179 (unsigned long)vma,
1180 false, MIGRATE_SYNC);
1181 if (nr_failed)
1182 putback_lru_pages(&pagelist);
1183 }
1184
1185 if (!err && nr_failed && (flags & MPOL_MF_STRICT))
1186 err = -EIO;
1187 } else
1188 putback_lru_pages(&pagelist);
1189
1190 up_write(&mm->mmap_sem);
1191 mpol_out:
1192 mpol_put(new);
1193 return err;
1194}
1195
1196/*
1197 * User space interface with variable sized bitmaps for nodelists.
1198 */
1199
1200/* Copy a node mask from user space. */
1201static int get_nodes(nodemask_t *nodes, const unsigned long __user *nmask,
1202 unsigned long maxnode)
1203{
1204 unsigned long k;
1205 unsigned long nlongs;
1206 unsigned long endmask;
1207
1208 --maxnode;
1209 nodes_clear(*nodes);
1210 if (maxnode == 0 || !nmask)
1211 return 0;
1212 if (maxnode > PAGE_SIZE*BITS_PER_BYTE)
1213 return -EINVAL;
1214
1215 nlongs = BITS_TO_LONGS(maxnode);
1216 if ((maxnode % BITS_PER_LONG) == 0)
1217 endmask = ~0UL;
1218 else
1219 endmask = (1UL << (maxnode % BITS_PER_LONG)) - 1;
1220
1221 /* When the user specified more nodes than supported just check
1222 if the non supported part is all zero. */
1223 if (nlongs > BITS_TO_LONGS(MAX_NUMNODES)) {
1224 if (nlongs > PAGE_SIZE/sizeof(long))
1225 return -EINVAL;
1226 for (k = BITS_TO_LONGS(MAX_NUMNODES); k < nlongs; k++) {
1227 unsigned long t;
1228 if (get_user(t, nmask + k))
1229 return -EFAULT;
1230 if (k == nlongs - 1) {
1231 if (t & endmask)
1232 return -EINVAL;
1233 } else if (t)
1234 return -EINVAL;
1235 }
1236 nlongs = BITS_TO_LONGS(MAX_NUMNODES);
1237 endmask = ~0UL;
1238 }
1239
1240 if (copy_from_user(nodes_addr(*nodes), nmask, nlongs*sizeof(unsigned long)))
1241 return -EFAULT;
1242 nodes_addr(*nodes)[nlongs-1] &= endmask;
1243 return 0;
1244}
1245
1246/* Copy a kernel node mask to user space */
1247static int copy_nodes_to_user(unsigned long __user *mask, unsigned long maxnode,
1248 nodemask_t *nodes)
1249{
1250 unsigned long copy = ALIGN(maxnode-1, 64) / 8;
1251 const int nbytes = BITS_TO_LONGS(MAX_NUMNODES) * sizeof(long);
1252
1253 if (copy > nbytes) {
1254 if (copy > PAGE_SIZE)
1255 return -EINVAL;
1256 if (clear_user((char __user *)mask + nbytes, copy - nbytes))
1257 return -EFAULT;
1258 copy = nbytes;
1259 }
1260 return copy_to_user(mask, nodes_addr(*nodes), copy) ? -EFAULT : 0;
1261}
1262
1263SYSCALL_DEFINE6(mbind, unsigned long, start, unsigned long, len,
1264 unsigned long, mode, unsigned long __user *, nmask,
1265 unsigned long, maxnode, unsigned, flags)
1266{
1267 nodemask_t nodes;
1268 int err;
1269 unsigned short mode_flags;
1270
1271 mode_flags = mode & MPOL_MODE_FLAGS;
1272 mode &= ~MPOL_MODE_FLAGS;
1273 if (mode >= MPOL_MAX)
1274 return -EINVAL;
1275 if ((mode_flags & MPOL_F_STATIC_NODES) &&
1276 (mode_flags & MPOL_F_RELATIVE_NODES))
1277 return -EINVAL;
1278 err = get_nodes(&nodes, nmask, maxnode);
1279 if (err)
1280 return err;
1281 return do_mbind(start, len, mode, mode_flags, &nodes, flags);
1282}
1283
1284/* Set the process memory policy */
1285SYSCALL_DEFINE3(set_mempolicy, int, mode, unsigned long __user *, nmask,
1286 unsigned long, maxnode)
1287{
1288 int err;
1289 nodemask_t nodes;
1290 unsigned short flags;
1291
1292 flags = mode & MPOL_MODE_FLAGS;
1293 mode &= ~MPOL_MODE_FLAGS;
1294 if ((unsigned int)mode >= MPOL_MAX)
1295 return -EINVAL;
1296 if ((flags & MPOL_F_STATIC_NODES) && (flags & MPOL_F_RELATIVE_NODES))
1297 return -EINVAL;
1298 err = get_nodes(&nodes, nmask, maxnode);
1299 if (err)
1300 return err;
1301 return do_set_mempolicy(mode, flags, &nodes);
1302}
1303
1304SYSCALL_DEFINE4(migrate_pages, pid_t, pid, unsigned long, maxnode,
1305 const unsigned long __user *, old_nodes,
1306 const unsigned long __user *, new_nodes)
1307{
1308 const struct cred *cred = current_cred(), *tcred;
1309 struct mm_struct *mm = NULL;
1310 struct task_struct *task;
1311 nodemask_t task_nodes;
1312 int err;
1313 nodemask_t *old;
1314 nodemask_t *new;
1315 NODEMASK_SCRATCH(scratch);
1316
1317 if (!scratch)
1318 return -ENOMEM;
1319
1320 old = &scratch->mask1;
1321 new = &scratch->mask2;
1322
1323 err = get_nodes(old, old_nodes, maxnode);
1324 if (err)
1325 goto out;
1326
1327 err = get_nodes(new, new_nodes, maxnode);
1328 if (err)
1329 goto out;
1330
1331 /* Find the mm_struct */
1332 rcu_read_lock();
1333 task = pid ? find_task_by_vpid(pid) : current;
1334 if (!task) {
1335 rcu_read_unlock();
1336 err = -ESRCH;
1337 goto out;
1338 }
1339 get_task_struct(task);
1340
1341 err = -EINVAL;
1342
1343 /*
1344 * Check if this process has the right to modify the specified
1345 * process. The right exists if the process has administrative
1346 * capabilities, superuser privileges or the same
1347 * userid as the target process.
1348 */
1349 tcred = __task_cred(task);
1350 if (!uid_eq(cred->euid, tcred->suid) && !uid_eq(cred->euid, tcred->uid) &&
1351 !uid_eq(cred->uid, tcred->suid) && !uid_eq(cred->uid, tcred->uid) &&
1352 !capable(CAP_SYS_NICE)) {
1353 rcu_read_unlock();
1354 err = -EPERM;
1355 goto out_put;
1356 }
1357 rcu_read_unlock();
1358
1359 task_nodes = cpuset_mems_allowed(task);
1360 /* Is the user allowed to access the target nodes? */
1361 if (!nodes_subset(*new, task_nodes) && !capable(CAP_SYS_NICE)) {
1362 err = -EPERM;
1363 goto out_put;
1364 }
1365
1366 if (!nodes_subset(*new, node_states[N_HIGH_MEMORY])) {
1367 err = -EINVAL;
1368 goto out_put;
1369 }
1370
1371 err = security_task_movememory(task);
1372 if (err)
1373 goto out_put;
1374
1375 mm = get_task_mm(task);
1376 put_task_struct(task);
1377
1378 if (!mm) {
1379 err = -EINVAL;
1380 goto out;
1381 }
1382
1383 err = do_migrate_pages(mm, old, new,
1384 capable(CAP_SYS_NICE) ? MPOL_MF_MOVE_ALL : MPOL_MF_MOVE);
1385
1386 mmput(mm);
1387out:
1388 NODEMASK_SCRATCH_FREE(scratch);
1389
1390 return err;
1391
1392out_put:
1393 put_task_struct(task);
1394 goto out;
1395
1396}
1397
1398
1399/* Retrieve NUMA policy */
1400SYSCALL_DEFINE5(get_mempolicy, int __user *, policy,
1401 unsigned long __user *, nmask, unsigned long, maxnode,
1402 unsigned long, addr, unsigned long, flags)
1403{
1404 int err;
1405 int uninitialized_var(pval);
1406 nodemask_t nodes;
1407
1408 if (nmask != NULL && maxnode < MAX_NUMNODES)
1409 return -EINVAL;
1410
1411 err = do_get_mempolicy(&pval, &nodes, addr, flags);
1412
1413 if (err)
1414 return err;
1415
1416 if (policy && put_user(pval, policy))
1417 return -EFAULT;
1418
1419 if (nmask)
1420 err = copy_nodes_to_user(nmask, maxnode, &nodes);
1421
1422 return err;
1423}
1424
1425#ifdef CONFIG_COMPAT
1426
1427asmlinkage long compat_sys_get_mempolicy(int __user *policy,
1428 compat_ulong_t __user *nmask,
1429 compat_ulong_t maxnode,
1430 compat_ulong_t addr, compat_ulong_t flags)
1431{
1432 long err;
1433 unsigned long __user *nm = NULL;
1434 unsigned long nr_bits, alloc_size;
1435 DECLARE_BITMAP(bm, MAX_NUMNODES);
1436
1437 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1438 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1439
1440 if (nmask)
1441 nm = compat_alloc_user_space(alloc_size);
1442
1443 err = sys_get_mempolicy(policy, nm, nr_bits+1, addr, flags);
1444
1445 if (!err && nmask) {
1446 unsigned long copy_size;
1447 copy_size = min_t(unsigned long, sizeof(bm), alloc_size);
1448 err = copy_from_user(bm, nm, copy_size);
1449 /* ensure entire bitmap is zeroed */
1450 err |= clear_user(nmask, ALIGN(maxnode-1, 8) / 8);
1451 err |= compat_put_bitmap(nmask, bm, nr_bits);
1452 }
1453
1454 return err;
1455}
1456
1457asmlinkage long compat_sys_set_mempolicy(int mode, compat_ulong_t __user *nmask,
1458 compat_ulong_t maxnode)
1459{
1460 long err = 0;
1461 unsigned long __user *nm = NULL;
1462 unsigned long nr_bits, alloc_size;
1463 DECLARE_BITMAP(bm, MAX_NUMNODES);
1464
1465 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1466 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1467
1468 if (nmask) {
1469 err = compat_get_bitmap(bm, nmask, nr_bits);
1470 nm = compat_alloc_user_space(alloc_size);
1471 err |= copy_to_user(nm, bm, alloc_size);
1472 }
1473
1474 if (err)
1475 return -EFAULT;
1476
1477 return sys_set_mempolicy(mode, nm, nr_bits+1);
1478}
1479
1480asmlinkage long compat_sys_mbind(compat_ulong_t start, compat_ulong_t len,
1481 compat_ulong_t mode, compat_ulong_t __user *nmask,
1482 compat_ulong_t maxnode, compat_ulong_t flags)
1483{
1484 long err = 0;
1485 unsigned long __user *nm = NULL;
1486 unsigned long nr_bits, alloc_size;
1487 nodemask_t bm;
1488
1489 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1490 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1491
1492 if (nmask) {
1493 err = compat_get_bitmap(nodes_addr(bm), nmask, nr_bits);
1494 nm = compat_alloc_user_space(alloc_size);
1495 err |= copy_to_user(nm, nodes_addr(bm), alloc_size);
1496 }
1497
1498 if (err)
1499 return -EFAULT;
1500
1501 return sys_mbind(start, len, mode, nm, nr_bits+1, flags);
1502}
1503
1504#endif
1505
1506/*
1507 * get_vma_policy(@task, @vma, @addr)
1508 * @task - task for fallback if vma policy == default
1509 * @vma - virtual memory area whose policy is sought
1510 * @addr - address in @vma for shared policy lookup
1511 *
1512 * Returns effective policy for a VMA at specified address.
1513 * Falls back to @task or system default policy, as necessary.
1514 * Current or other task's task mempolicy and non-shared vma policies
1515 * are protected by the task's mmap_sem, which must be held for read by
1516 * the caller.
1517 * Shared policies [those marked as MPOL_F_SHARED] require an extra reference
1518 * count--added by the get_policy() vm_op, as appropriate--to protect against
1519 * freeing by another task. It is the caller's responsibility to free the
1520 * extra reference for shared policies.
1521 */
1522struct mempolicy *get_vma_policy(struct task_struct *task,
1523 struct vm_area_struct *vma, unsigned long addr)
1524{
1525 struct mempolicy *pol = task->mempolicy;
1526
1527 if (vma) {
1528 if (vma->vm_ops && vma->vm_ops->get_policy) {
1529 struct mempolicy *vpol = vma->vm_ops->get_policy(vma,
1530 addr);
1531 if (vpol)
1532 pol = vpol;
1533 } else if (vma->vm_policy)
1534 pol = vma->vm_policy;
1535 }
1536 if (!pol)
1537 pol = &default_policy;
1538 return pol;
1539}
1540
1541/*
1542 * Return a nodemask representing a mempolicy for filtering nodes for
1543 * page allocation
1544 */
1545static nodemask_t *policy_nodemask(gfp_t gfp, struct mempolicy *policy)
1546{
1547 /* Lower zones don't get a nodemask applied for MPOL_BIND */
1548 if (unlikely(policy->mode == MPOL_BIND) &&
1549 gfp_zone(gfp) >= policy_zone &&
1550 cpuset_nodemask_valid_mems_allowed(&policy->v.nodes))
1551 return &policy->v.nodes;
1552
1553 return NULL;
1554}
1555
1556/* Return a zonelist indicated by gfp for node representing a mempolicy */
1557static struct zonelist *policy_zonelist(gfp_t gfp, struct mempolicy *policy,
1558 int nd)
1559{
1560 switch (policy->mode) {
1561 case MPOL_PREFERRED:
1562 if (!(policy->flags & MPOL_F_LOCAL))
1563 nd = policy->v.preferred_node;
1564 break;
1565 case MPOL_BIND:
1566 /*
1567 * Normally, MPOL_BIND allocations are node-local within the
1568 * allowed nodemask. However, if __GFP_THISNODE is set and the
1569 * current node isn't part of the mask, we use the zonelist for
1570 * the first node in the mask instead.
1571 */
1572 if (unlikely(gfp & __GFP_THISNODE) &&
1573 unlikely(!node_isset(nd, policy->v.nodes)))
1574 nd = first_node(policy->v.nodes);
1575 break;
1576 default:
1577 BUG();
1578 }
1579 return node_zonelist(nd, gfp);
1580}
1581
1582/* Do dynamic interleaving for a process */
1583static unsigned interleave_nodes(struct mempolicy *policy)
1584{
1585 unsigned nid, next;
1586 struct task_struct *me = current;
1587
1588 nid = me->il_next;
1589 next = next_node(nid, policy->v.nodes);
1590 if (next >= MAX_NUMNODES)
1591 next = first_node(policy->v.nodes);
1592 if (next < MAX_NUMNODES)
1593 me->il_next = next;
1594 return nid;
1595}
1596
1597/*
1598 * Depending on the memory policy provide a node from which to allocate the
1599 * next slab entry.
1600 * @policy must be protected by freeing by the caller. If @policy is
1601 * the current task's mempolicy, this protection is implicit, as only the
1602 * task can change it's policy. The system default policy requires no
1603 * such protection.
1604 */
1605unsigned slab_node(struct mempolicy *policy)
1606{
1607 if (!policy || policy->flags & MPOL_F_LOCAL)
1608 return numa_node_id();
1609
1610 switch (policy->mode) {
1611 case MPOL_PREFERRED:
1612 /*
1613 * handled MPOL_F_LOCAL above
1614 */
1615 return policy->v.preferred_node;
1616
1617 case MPOL_INTERLEAVE:
1618 return interleave_nodes(policy);
1619
1620 case MPOL_BIND: {
1621 /*
1622 * Follow bind policy behavior and start allocation at the
1623 * first node.
1624 */
1625 struct zonelist *zonelist;
1626 struct zone *zone;
1627 enum zone_type highest_zoneidx = gfp_zone(GFP_KERNEL);
1628 zonelist = &NODE_DATA(numa_node_id())->node_zonelists[0];
1629 (void)first_zones_zonelist(zonelist, highest_zoneidx,
1630 &policy->v.nodes,
1631 &zone);
1632 return zone ? zone->node : numa_node_id();
1633 }
1634
1635 default:
1636 BUG();
1637 }
1638}
1639
1640/* Do static interleaving for a VMA with known offset. */
1641static unsigned offset_il_node(struct mempolicy *pol,
1642 struct vm_area_struct *vma, unsigned long off)
1643{
1644 unsigned nnodes = nodes_weight(pol->v.nodes);
1645 unsigned target;
1646 int c;
1647 int nid = -1;
1648
1649 if (!nnodes)
1650 return numa_node_id();
1651 target = (unsigned int)off % nnodes;
1652 c = 0;
1653 do {
1654 nid = next_node(nid, pol->v.nodes);
1655 c++;
1656 } while (c <= target);
1657 return nid;
1658}
1659
1660/* Determine a node number for interleave */
1661static inline unsigned interleave_nid(struct mempolicy *pol,
1662 struct vm_area_struct *vma, unsigned long addr, int shift)
1663{
1664 if (vma) {
1665 unsigned long off;
1666
1667 /*
1668 * for small pages, there is no difference between
1669 * shift and PAGE_SHIFT, so the bit-shift is safe.
1670 * for huge pages, since vm_pgoff is in units of small
1671 * pages, we need to shift off the always 0 bits to get
1672 * a useful offset.
1673 */
1674 BUG_ON(shift < PAGE_SHIFT);
1675 off = vma->vm_pgoff >> (shift - PAGE_SHIFT);
1676 off += (addr - vma->vm_start) >> shift;
1677 return offset_il_node(pol, vma, off);
1678 } else
1679 return interleave_nodes(pol);
1680}
1681
1682/*
1683 * Return the bit number of a random bit set in the nodemask.
1684 * (returns -1 if nodemask is empty)
1685 */
1686int node_random(const nodemask_t *maskp)
1687{
1688 int w, bit = -1;
1689
1690 w = nodes_weight(*maskp);
1691 if (w)
1692 bit = bitmap_ord_to_pos(maskp->bits,
1693 get_random_int() % w, MAX_NUMNODES);
1694 return bit;
1695}
1696
1697#ifdef CONFIG_HUGETLBFS
1698/*
1699 * huge_zonelist(@vma, @addr, @gfp_flags, @mpol)
1700 * @vma = virtual memory area whose policy is sought
1701 * @addr = address in @vma for shared policy lookup and interleave policy
1702 * @gfp_flags = for requested zone
1703 * @mpol = pointer to mempolicy pointer for reference counted mempolicy
1704 * @nodemask = pointer to nodemask pointer for MPOL_BIND nodemask
1705 *
1706 * Returns a zonelist suitable for a huge page allocation and a pointer
1707 * to the struct mempolicy for conditional unref after allocation.
1708 * If the effective policy is 'BIND, returns a pointer to the mempolicy's
1709 * @nodemask for filtering the zonelist.
1710 *
1711 * Must be protected by get_mems_allowed()
1712 */
1713struct zonelist *huge_zonelist(struct vm_area_struct *vma, unsigned long addr,
1714 gfp_t gfp_flags, struct mempolicy **mpol,
1715 nodemask_t **nodemask)
1716{
1717 struct zonelist *zl;
1718
1719 *mpol = get_vma_policy(current, vma, addr);
1720 *nodemask = NULL; /* assume !MPOL_BIND */
1721
1722 if (unlikely((*mpol)->mode == MPOL_INTERLEAVE)) {
1723 zl = node_zonelist(interleave_nid(*mpol, vma, addr,
1724 huge_page_shift(hstate_vma(vma))), gfp_flags);
1725 } else {
1726 zl = policy_zonelist(gfp_flags, *mpol, numa_node_id());
1727 if ((*mpol)->mode == MPOL_BIND)
1728 *nodemask = &(*mpol)->v.nodes;
1729 }
1730 return zl;
1731}
1732
1733/*
1734 * init_nodemask_of_mempolicy
1735 *
1736 * If the current task's mempolicy is "default" [NULL], return 'false'
1737 * to indicate default policy. Otherwise, extract the policy nodemask
1738 * for 'bind' or 'interleave' policy into the argument nodemask, or
1739 * initialize the argument nodemask to contain the single node for
1740 * 'preferred' or 'local' policy and return 'true' to indicate presence
1741 * of non-default mempolicy.
1742 *
1743 * We don't bother with reference counting the mempolicy [mpol_get/put]
1744 * because the current task is examining it's own mempolicy and a task's
1745 * mempolicy is only ever changed by the task itself.
1746 *
1747 * N.B., it is the caller's responsibility to free a returned nodemask.
1748 */
1749bool init_nodemask_of_mempolicy(nodemask_t *mask)
1750{
1751 struct mempolicy *mempolicy;
1752 int nid;
1753
1754 if (!(mask && current->mempolicy))
1755 return false;
1756
1757 task_lock(current);
1758 mempolicy = current->mempolicy;
1759 switch (mempolicy->mode) {
1760 case MPOL_PREFERRED:
1761 if (mempolicy->flags & MPOL_F_LOCAL)
1762 nid = numa_node_id();
1763 else
1764 nid = mempolicy->v.preferred_node;
1765 init_nodemask_of_node(mask, nid);
1766 break;
1767
1768 case MPOL_BIND:
1769 /* Fall through */
1770 case MPOL_INTERLEAVE:
1771 *mask = mempolicy->v.nodes;
1772 break;
1773
1774 default:
1775 BUG();
1776 }
1777 task_unlock(current);
1778
1779 return true;
1780}
1781#endif
1782
1783/*
1784 * mempolicy_nodemask_intersects
1785 *
1786 * If tsk's mempolicy is "default" [NULL], return 'true' to indicate default
1787 * policy. Otherwise, check for intersection between mask and the policy
1788 * nodemask for 'bind' or 'interleave' policy. For 'perferred' or 'local'
1789 * policy, always return true since it may allocate elsewhere on fallback.
1790 *
1791 * Takes task_lock(tsk) to prevent freeing of its mempolicy.
1792 */
1793bool mempolicy_nodemask_intersects(struct task_struct *tsk,
1794 const nodemask_t *mask)
1795{
1796 struct mempolicy *mempolicy;
1797 bool ret = true;
1798
1799 if (!mask)
1800 return ret;
1801 task_lock(tsk);
1802 mempolicy = tsk->mempolicy;
1803 if (!mempolicy)
1804 goto out;
1805
1806 switch (mempolicy->mode) {
1807 case MPOL_PREFERRED:
1808 /*
1809 * MPOL_PREFERRED and MPOL_F_LOCAL are only preferred nodes to
1810 * allocate from, they may fallback to other nodes when oom.
1811 * Thus, it's possible for tsk to have allocated memory from
1812 * nodes in mask.
1813 */
1814 break;
1815 case MPOL_BIND:
1816 case MPOL_INTERLEAVE:
1817 ret = nodes_intersects(mempolicy->v.nodes, *mask);
1818 break;
1819 default:
1820 BUG();
1821 }
1822out:
1823 task_unlock(tsk);
1824 return ret;
1825}
1826
1827/* Allocate a page in interleaved policy.
1828 Own path because it needs to do special accounting. */
1829static struct page *alloc_page_interleave(gfp_t gfp, unsigned order,
1830 unsigned nid)
1831{
1832 struct zonelist *zl;
1833 struct page *page;
1834
1835 zl = node_zonelist(nid, gfp);
1836 page = __alloc_pages(gfp, order, zl);
1837 if (page && page_zone(page) == zonelist_zone(&zl->_zonerefs[0]))
1838 inc_zone_page_state(page, NUMA_INTERLEAVE_HIT);
1839 return page;
1840}
1841
1842/**
1843 * alloc_pages_vma - Allocate a page for a VMA.
1844 *
1845 * @gfp:
1846 * %GFP_USER user allocation.
1847 * %GFP_KERNEL kernel allocations,
1848 * %GFP_HIGHMEM highmem/user allocations,
1849 * %GFP_FS allocation should not call back into a file system.
1850 * %GFP_ATOMIC don't sleep.
1851 *
1852 * @order:Order of the GFP allocation.
1853 * @vma: Pointer to VMA or NULL if not available.
1854 * @addr: Virtual Address of the allocation. Must be inside the VMA.
1855 *
1856 * This function allocates a page from the kernel page pool and applies
1857 * a NUMA policy associated with the VMA or the current process.
1858 * When VMA is not NULL caller must hold down_read on the mmap_sem of the
1859 * mm_struct of the VMA to prevent it from going away. Should be used for
1860 * all allocations for pages that will be mapped into
1861 * user space. Returns NULL when no page can be allocated.
1862 *
1863 * Should be called with the mm_sem of the vma hold.
1864 */
1865struct page *
1866alloc_pages_vma(gfp_t gfp, int order, struct vm_area_struct *vma,
1867 unsigned long addr, int node)
1868{
1869 struct mempolicy *pol;
1870 struct zonelist *zl;
1871 struct page *page;
1872 unsigned int cpuset_mems_cookie;
1873
1874retry_cpuset:
1875 pol = get_vma_policy(current, vma, addr);
1876 cpuset_mems_cookie = get_mems_allowed();
1877
1878 if (unlikely(pol->mode == MPOL_INTERLEAVE)) {
1879 unsigned nid;
1880
1881 nid = interleave_nid(pol, vma, addr, PAGE_SHIFT + order);
1882 mpol_cond_put(pol);
1883 page = alloc_page_interleave(gfp, order, nid);
1884 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page))
1885 goto retry_cpuset;
1886
1887 return page;
1888 }
1889 zl = policy_zonelist(gfp, pol, node);
1890 if (unlikely(mpol_needs_cond_ref(pol))) {
1891 /*
1892 * slow path: ref counted shared policy
1893 */
1894 struct page *page = __alloc_pages_nodemask(gfp, order,
1895 zl, policy_nodemask(gfp, pol));
1896 __mpol_put(pol);
1897 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page))
1898 goto retry_cpuset;
1899 return page;
1900 }
1901 /*
1902 * fast path: default or task policy
1903 */
1904 page = __alloc_pages_nodemask(gfp, order, zl,
1905 policy_nodemask(gfp, pol));
1906 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page))
1907 goto retry_cpuset;
1908 return page;
1909}
1910
1911/**
1912 * alloc_pages_current - Allocate pages.
1913 *
1914 * @gfp:
1915 * %GFP_USER user allocation,
1916 * %GFP_KERNEL kernel allocation,
1917 * %GFP_HIGHMEM highmem allocation,
1918 * %GFP_FS don't call back into a file system.
1919 * %GFP_ATOMIC don't sleep.
1920 * @order: Power of two of allocation size in pages. 0 is a single page.
1921 *
1922 * Allocate a page from the kernel page pool. When not in
1923 * interrupt context and apply the current process NUMA policy.
1924 * Returns NULL when no page can be allocated.
1925 *
1926 * Don't call cpuset_update_task_memory_state() unless
1927 * 1) it's ok to take cpuset_sem (can WAIT), and
1928 * 2) allocating for current task (not interrupt).
1929 */
1930struct page *alloc_pages_current(gfp_t gfp, unsigned order)
1931{
1932 struct mempolicy *pol = current->mempolicy;
1933 struct page *page;
1934 unsigned int cpuset_mems_cookie;
1935
1936 if (!pol || in_interrupt() || (gfp & __GFP_THISNODE))
1937 pol = &default_policy;
1938
1939retry_cpuset:
1940 cpuset_mems_cookie = get_mems_allowed();
1941
1942 /*
1943 * No reference counting needed for current->mempolicy
1944 * nor system default_policy
1945 */
1946 if (pol->mode == MPOL_INTERLEAVE)
1947 page = alloc_page_interleave(gfp, order, interleave_nodes(pol));
1948 else
1949 page = __alloc_pages_nodemask(gfp, order,
1950 policy_zonelist(gfp, pol, numa_node_id()),
1951 policy_nodemask(gfp, pol));
1952
1953 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page))
1954 goto retry_cpuset;
1955
1956 return page;
1957}
1958EXPORT_SYMBOL(alloc_pages_current);
1959
1960/*
1961 * If mpol_dup() sees current->cpuset == cpuset_being_rebound, then it
1962 * rebinds the mempolicy its copying by calling mpol_rebind_policy()
1963 * with the mems_allowed returned by cpuset_mems_allowed(). This
1964 * keeps mempolicies cpuset relative after its cpuset moves. See
1965 * further kernel/cpuset.c update_nodemask().
1966 *
1967 * current's mempolicy may be rebinded by the other task(the task that changes
1968 * cpuset's mems), so we needn't do rebind work for current task.
1969 */
1970
1971/* Slow path of a mempolicy duplicate */
1972struct mempolicy *__mpol_dup(struct mempolicy *old)
1973{
1974 struct mempolicy *new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
1975
1976 if (!new)
1977 return ERR_PTR(-ENOMEM);
1978
1979 /* task's mempolicy is protected by alloc_lock */
1980 if (old == current->mempolicy) {
1981 task_lock(current);
1982 *new = *old;
1983 task_unlock(current);
1984 } else
1985 *new = *old;
1986
1987 rcu_read_lock();
1988 if (current_cpuset_is_being_rebound()) {
1989 nodemask_t mems = cpuset_mems_allowed(current);
1990 if (new->flags & MPOL_F_REBINDING)
1991 mpol_rebind_policy(new, &mems, MPOL_REBIND_STEP2);
1992 else
1993 mpol_rebind_policy(new, &mems, MPOL_REBIND_ONCE);
1994 }
1995 rcu_read_unlock();
1996 atomic_set(&new->refcnt, 1);
1997 return new;
1998}
1999
2000/*
2001 * If *frompol needs [has] an extra ref, copy *frompol to *tompol ,
2002 * eliminate the * MPOL_F_* flags that require conditional ref and
2003 * [NOTE!!!] drop the extra ref. Not safe to reference *frompol directly
2004 * after return. Use the returned value.
2005 *
2006 * Allows use of a mempolicy for, e.g., multiple allocations with a single
2007 * policy lookup, even if the policy needs/has extra ref on lookup.
2008 * shmem_readahead needs this.
2009 */
2010struct mempolicy *__mpol_cond_copy(struct mempolicy *tompol,
2011 struct mempolicy *frompol)
2012{
2013 if (!mpol_needs_cond_ref(frompol))
2014 return frompol;
2015
2016 *tompol = *frompol;
2017 tompol->flags &= ~MPOL_F_SHARED; /* copy doesn't need unref */
2018 __mpol_put(frompol);
2019 return tompol;
2020}
2021
2022/* Slow path of a mempolicy comparison */
2023bool __mpol_equal(struct mempolicy *a, struct mempolicy *b)
2024{
2025 if (!a || !b)
2026 return false;
2027 if (a->mode != b->mode)
2028 return false;
2029 if (a->flags != b->flags)
2030 return false;
2031 if (mpol_store_user_nodemask(a))
2032 if (!nodes_equal(a->w.user_nodemask, b->w.user_nodemask))
2033 return false;
2034
2035 switch (a->mode) {
2036 case MPOL_BIND:
2037 /* Fall through */
2038 case MPOL_INTERLEAVE:
2039 return !!nodes_equal(a->v.nodes, b->v.nodes);
2040 case MPOL_PREFERRED:
2041 return a->v.preferred_node == b->v.preferred_node;
2042 default:
2043 BUG();
2044 return false;
2045 }
2046}
2047
2048/*
2049 * Shared memory backing store policy support.
2050 *
2051 * Remember policies even when nobody has shared memory mapped.
2052 * The policies are kept in Red-Black tree linked from the inode.
2053 * They are protected by the sp->lock spinlock, which should be held
2054 * for any accesses to the tree.
2055 */
2056
2057/* lookup first element intersecting start-end */
2058/* Caller holds sp->lock */
2059static struct sp_node *
2060sp_lookup(struct shared_policy *sp, unsigned long start, unsigned long end)
2061{
2062 struct rb_node *n = sp->root.rb_node;
2063
2064 while (n) {
2065 struct sp_node *p = rb_entry(n, struct sp_node, nd);
2066
2067 if (start >= p->end)
2068 n = n->rb_right;
2069 else if (end <= p->start)
2070 n = n->rb_left;
2071 else
2072 break;
2073 }
2074 if (!n)
2075 return NULL;
2076 for (;;) {
2077 struct sp_node *w = NULL;
2078 struct rb_node *prev = rb_prev(n);
2079 if (!prev)
2080 break;
2081 w = rb_entry(prev, struct sp_node, nd);
2082 if (w->end <= start)
2083 break;
2084 n = prev;
2085 }
2086 return rb_entry(n, struct sp_node, nd);
2087}
2088
2089/* Insert a new shared policy into the list. */
2090/* Caller holds sp->lock */
2091static void sp_insert(struct shared_policy *sp, struct sp_node *new)
2092{
2093 struct rb_node **p = &sp->root.rb_node;
2094 struct rb_node *parent = NULL;
2095 struct sp_node *nd;
2096
2097 while (*p) {
2098 parent = *p;
2099 nd = rb_entry(parent, struct sp_node, nd);
2100 if (new->start < nd->start)
2101 p = &(*p)->rb_left;
2102 else if (new->end > nd->end)
2103 p = &(*p)->rb_right;
2104 else
2105 BUG();
2106 }
2107 rb_link_node(&new->nd, parent, p);
2108 rb_insert_color(&new->nd, &sp->root);
2109 pr_debug("inserting %lx-%lx: %d\n", new->start, new->end,
2110 new->policy ? new->policy->mode : 0);
2111}
2112
2113/* Find shared policy intersecting idx */
2114struct mempolicy *
2115mpol_shared_policy_lookup(struct shared_policy *sp, unsigned long idx)
2116{
2117 struct mempolicy *pol = NULL;
2118 struct sp_node *sn;
2119
2120 if (!sp->root.rb_node)
2121 return NULL;
2122 spin_lock(&sp->lock);
2123 sn = sp_lookup(sp, idx, idx+1);
2124 if (sn) {
2125 mpol_get(sn->policy);
2126 pol = sn->policy;
2127 }
2128 spin_unlock(&sp->lock);
2129 return pol;
2130}
2131
2132static void sp_delete(struct shared_policy *sp, struct sp_node *n)
2133{
2134 pr_debug("deleting %lx-l%lx\n", n->start, n->end);
2135 rb_erase(&n->nd, &sp->root);
2136 mpol_put(n->policy);
2137 kmem_cache_free(sn_cache, n);
2138}
2139
2140static struct sp_node *sp_alloc(unsigned long start, unsigned long end,
2141 struct mempolicy *pol)
2142{
2143 struct sp_node *n = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2144
2145 if (!n)
2146 return NULL;
2147 n->start = start;
2148 n->end = end;
2149 mpol_get(pol);
2150 pol->flags |= MPOL_F_SHARED; /* for unref */
2151 n->policy = pol;
2152 return n;
2153}
2154
2155/* Replace a policy range. */
2156static int shared_policy_replace(struct shared_policy *sp, unsigned long start,
2157 unsigned long end, struct sp_node *new)
2158{
2159 struct sp_node *n, *new2 = NULL;
2160
2161restart:
2162 spin_lock(&sp->lock);
2163 n = sp_lookup(sp, start, end);
2164 /* Take care of old policies in the same range. */
2165 while (n && n->start < end) {
2166 struct rb_node *next = rb_next(&n->nd);
2167 if (n->start >= start) {
2168 if (n->end <= end)
2169 sp_delete(sp, n);
2170 else
2171 n->start = end;
2172 } else {
2173 /* Old policy spanning whole new range. */
2174 if (n->end > end) {
2175 if (!new2) {
2176 spin_unlock(&sp->lock);
2177 new2 = sp_alloc(end, n->end, n->policy);
2178 if (!new2)
2179 return -ENOMEM;
2180 goto restart;
2181 }
2182 n->end = start;
2183 sp_insert(sp, new2);
2184 new2 = NULL;
2185 break;
2186 } else
2187 n->end = start;
2188 }
2189 if (!next)
2190 break;
2191 n = rb_entry(next, struct sp_node, nd);
2192 }
2193 if (new)
2194 sp_insert(sp, new);
2195 spin_unlock(&sp->lock);
2196 if (new2) {
2197 mpol_put(new2->policy);
2198 kmem_cache_free(sn_cache, new2);
2199 }
2200 return 0;
2201}
2202
2203/**
2204 * mpol_shared_policy_init - initialize shared policy for inode
2205 * @sp: pointer to inode shared policy
2206 * @mpol: struct mempolicy to install
2207 *
2208 * Install non-NULL @mpol in inode's shared policy rb-tree.
2209 * On entry, the current task has a reference on a non-NULL @mpol.
2210 * This must be released on exit.
2211 * This is called at get_inode() calls and we can use GFP_KERNEL.
2212 */
2213void mpol_shared_policy_init(struct shared_policy *sp, struct mempolicy *mpol)
2214{
2215 int ret;
2216
2217 sp->root = RB_ROOT; /* empty tree == default mempolicy */
2218 spin_lock_init(&sp->lock);
2219
2220 if (mpol) {
2221 struct vm_area_struct pvma;
2222 struct mempolicy *new;
2223 NODEMASK_SCRATCH(scratch);
2224
2225 if (!scratch)
2226 goto put_mpol;
2227 /* contextualize the tmpfs mount point mempolicy */
2228 new = mpol_new(mpol->mode, mpol->flags, &mpol->w.user_nodemask);
2229 if (IS_ERR(new))
2230 goto free_scratch; /* no valid nodemask intersection */
2231
2232 task_lock(current);
2233 ret = mpol_set_nodemask(new, &mpol->w.user_nodemask, scratch);
2234 task_unlock(current);
2235 if (ret)
2236 goto put_new;
2237
2238 /* Create pseudo-vma that contains just the policy */
2239 memset(&pvma, 0, sizeof(struct vm_area_struct));
2240 pvma.vm_end = TASK_SIZE; /* policy covers entire file */
2241 mpol_set_shared_policy(sp, &pvma, new); /* adds ref */
2242
2243put_new:
2244 mpol_put(new); /* drop initial ref */
2245free_scratch:
2246 NODEMASK_SCRATCH_FREE(scratch);
2247put_mpol:
2248 mpol_put(mpol); /* drop our incoming ref on sb mpol */
2249 }
2250}
2251
2252int mpol_set_shared_policy(struct shared_policy *info,
2253 struct vm_area_struct *vma, struct mempolicy *npol)
2254{
2255 int err;
2256 struct sp_node *new = NULL;
2257 unsigned long sz = vma_pages(vma);
2258
2259 pr_debug("set_shared_policy %lx sz %lu %d %d %lx\n",
2260 vma->vm_pgoff,
2261 sz, npol ? npol->mode : -1,
2262 npol ? npol->flags : -1,
2263 npol ? nodes_addr(npol->v.nodes)[0] : -1);
2264
2265 if (npol) {
2266 new = sp_alloc(vma->vm_pgoff, vma->vm_pgoff + sz, npol);
2267 if (!new)
2268 return -ENOMEM;
2269 }
2270 err = shared_policy_replace(info, vma->vm_pgoff, vma->vm_pgoff+sz, new);
2271 if (err && new)
2272 kmem_cache_free(sn_cache, new);
2273 return err;
2274}
2275
2276/* Free a backing policy store on inode delete. */
2277void mpol_free_shared_policy(struct shared_policy *p)
2278{
2279 struct sp_node *n;
2280 struct rb_node *next;
2281
2282 if (!p->root.rb_node)
2283 return;
2284 spin_lock(&p->lock);
2285 next = rb_first(&p->root);
2286 while (next) {
2287 n = rb_entry(next, struct sp_node, nd);
2288 next = rb_next(&n->nd);
2289 rb_erase(&n->nd, &p->root);
2290 mpol_put(n->policy);
2291 kmem_cache_free(sn_cache, n);
2292 }
2293 spin_unlock(&p->lock);
2294}
2295
2296/* assumes fs == KERNEL_DS */
2297void __init numa_policy_init(void)
2298{
2299 nodemask_t interleave_nodes;
2300 unsigned long largest = 0;
2301 int nid, prefer = 0;
2302
2303 policy_cache = kmem_cache_create("numa_policy",
2304 sizeof(struct mempolicy),
2305 0, SLAB_PANIC, NULL);
2306
2307 sn_cache = kmem_cache_create("shared_policy_node",
2308 sizeof(struct sp_node),
2309 0, SLAB_PANIC, NULL);
2310
2311 /*
2312 * Set interleaving policy for system init. Interleaving is only
2313 * enabled across suitably sized nodes (default is >= 16MB), or
2314 * fall back to the largest node if they're all smaller.
2315 */
2316 nodes_clear(interleave_nodes);
2317 for_each_node_state(nid, N_HIGH_MEMORY) {
2318 unsigned long total_pages = node_present_pages(nid);
2319
2320 /* Preserve the largest node */
2321 if (largest < total_pages) {
2322 largest = total_pages;
2323 prefer = nid;
2324 }
2325
2326 /* Interleave this node? */
2327 if ((total_pages << PAGE_SHIFT) >= (16 << 20))
2328 node_set(nid, interleave_nodes);
2329 }
2330
2331 /* All too small, use the largest */
2332 if (unlikely(nodes_empty(interleave_nodes)))
2333 node_set(prefer, interleave_nodes);
2334
2335 if (do_set_mempolicy(MPOL_INTERLEAVE, 0, &interleave_nodes))
2336 printk("numa_policy_init: interleaving failed\n");
2337}
2338
2339/* Reset policy of current process to default */
2340void numa_default_policy(void)
2341{
2342 do_set_mempolicy(MPOL_DEFAULT, 0, NULL);
2343}
2344
2345/*
2346 * Parse and format mempolicy from/to strings
2347 */
2348
2349/*
2350 * "local" is pseudo-policy: MPOL_PREFERRED with MPOL_F_LOCAL flag
2351 * Used only for mpol_parse_str() and mpol_to_str()
2352 */
2353#define MPOL_LOCAL MPOL_MAX
2354static const char * const policy_modes[] =
2355{
2356 [MPOL_DEFAULT] = "default",
2357 [MPOL_PREFERRED] = "prefer",
2358 [MPOL_BIND] = "bind",
2359 [MPOL_INTERLEAVE] = "interleave",
2360 [MPOL_LOCAL] = "local"
2361};
2362
2363
2364#ifdef CONFIG_TMPFS
2365/**
2366 * mpol_parse_str - parse string to mempolicy
2367 * @str: string containing mempolicy to parse
2368 * @mpol: pointer to struct mempolicy pointer, returned on success.
2369 * @no_context: flag whether to "contextualize" the mempolicy
2370 *
2371 * Format of input:
2372 * <mode>[=<flags>][:<nodelist>]
2373 *
2374 * if @no_context is true, save the input nodemask in w.user_nodemask in
2375 * the returned mempolicy. This will be used to "clone" the mempolicy in
2376 * a specific context [cpuset] at a later time. Used to parse tmpfs mpol
2377 * mount option. Note that if 'static' or 'relative' mode flags were
2378 * specified, the input nodemask will already have been saved. Saving
2379 * it again is redundant, but safe.
2380 *
2381 * On success, returns 0, else 1
2382 */
2383int mpol_parse_str(char *str, struct mempolicy **mpol, int no_context)
2384{
2385 struct mempolicy *new = NULL;
2386 unsigned short mode;
2387 unsigned short uninitialized_var(mode_flags);
2388 nodemask_t nodes;
2389 char *nodelist = strchr(str, ':');
2390 char *flags = strchr(str, '=');
2391 int err = 1;
2392
2393 if (nodelist) {
2394 /* NUL-terminate mode or flags string */
2395 *nodelist++ = '\0';
2396 if (nodelist_parse(nodelist, nodes))
2397 goto out;
2398 if (!nodes_subset(nodes, node_states[N_HIGH_MEMORY]))
2399 goto out;
2400 } else
2401 nodes_clear(nodes);
2402
2403 if (flags)
2404 *flags++ = '\0'; /* terminate mode string */
2405
2406 for (mode = 0; mode <= MPOL_LOCAL; mode++) {
2407 if (!strcmp(str, policy_modes[mode])) {
2408 break;
2409 }
2410 }
2411 if (mode > MPOL_LOCAL)
2412 goto out;
2413
2414 switch (mode) {
2415 case MPOL_PREFERRED:
2416 /*
2417 * Insist on a nodelist of one node only
2418 */
2419 if (nodelist) {
2420 char *rest = nodelist;
2421 while (isdigit(*rest))
2422 rest++;
2423 if (*rest)
2424 goto out;
2425 }
2426 break;
2427 case MPOL_INTERLEAVE:
2428 /*
2429 * Default to online nodes with memory if no nodelist
2430 */
2431 if (!nodelist)
2432 nodes = node_states[N_HIGH_MEMORY];
2433 break;
2434 case MPOL_LOCAL:
2435 /*
2436 * Don't allow a nodelist; mpol_new() checks flags
2437 */
2438 if (nodelist)
2439 goto out;
2440 mode = MPOL_PREFERRED;
2441 break;
2442 case MPOL_DEFAULT:
2443 /*
2444 * Insist on a empty nodelist
2445 */
2446 if (!nodelist)
2447 err = 0;
2448 goto out;
2449 case MPOL_BIND:
2450 /*
2451 * Insist on a nodelist
2452 */
2453 if (!nodelist)
2454 goto out;
2455 }
2456
2457 mode_flags = 0;
2458 if (flags) {
2459 /*
2460 * Currently, we only support two mutually exclusive
2461 * mode flags.
2462 */
2463 if (!strcmp(flags, "static"))
2464 mode_flags |= MPOL_F_STATIC_NODES;
2465 else if (!strcmp(flags, "relative"))
2466 mode_flags |= MPOL_F_RELATIVE_NODES;
2467 else
2468 goto out;
2469 }
2470
2471 new = mpol_new(mode, mode_flags, &nodes);
2472 if (IS_ERR(new))
2473 goto out;
2474
2475 if (no_context) {
2476 /* save for contextualization */
2477 new->w.user_nodemask = nodes;
2478 } else {
2479 int ret;
2480 NODEMASK_SCRATCH(scratch);
2481 if (scratch) {
2482 task_lock(current);
2483 ret = mpol_set_nodemask(new, &nodes, scratch);
2484 task_unlock(current);
2485 } else
2486 ret = -ENOMEM;
2487 NODEMASK_SCRATCH_FREE(scratch);
2488 if (ret) {
2489 mpol_put(new);
2490 goto out;
2491 }
2492 }
2493 err = 0;
2494
2495out:
2496 /* Restore string for error message */
2497 if (nodelist)
2498 *--nodelist = ':';
2499 if (flags)
2500 *--flags = '=';
2501 if (!err)
2502 *mpol = new;
2503 return err;
2504}
2505#endif /* CONFIG_TMPFS */
2506
2507/**
2508 * mpol_to_str - format a mempolicy structure for printing
2509 * @buffer: to contain formatted mempolicy string
2510 * @maxlen: length of @buffer
2511 * @pol: pointer to mempolicy to be formatted
2512 * @no_context: "context free" mempolicy - use nodemask in w.user_nodemask
2513 *
2514 * Convert a mempolicy into a string.
2515 * Returns the number of characters in buffer (if positive)
2516 * or an error (negative)
2517 */
2518int mpol_to_str(char *buffer, int maxlen, struct mempolicy *pol, int no_context)
2519{
2520 char *p = buffer;
2521 int l;
2522 nodemask_t nodes;
2523 unsigned short mode;
2524 unsigned short flags = pol ? pol->flags : 0;
2525
2526 /*
2527 * Sanity check: room for longest mode, flag and some nodes
2528 */
2529 VM_BUG_ON(maxlen < strlen("interleave") + strlen("relative") + 16);
2530
2531 if (!pol || pol == &default_policy)
2532 mode = MPOL_DEFAULT;
2533 else
2534 mode = pol->mode;
2535
2536 switch (mode) {
2537 case MPOL_DEFAULT:
2538 nodes_clear(nodes);
2539 break;
2540
2541 case MPOL_PREFERRED:
2542 nodes_clear(nodes);
2543 if (flags & MPOL_F_LOCAL)
2544 mode = MPOL_LOCAL; /* pseudo-policy */
2545 else
2546 node_set(pol->v.preferred_node, nodes);
2547 break;
2548
2549 case MPOL_BIND:
2550 /* Fall through */
2551 case MPOL_INTERLEAVE:
2552 if (no_context)
2553 nodes = pol->w.user_nodemask;
2554 else
2555 nodes = pol->v.nodes;
2556 break;
2557
2558 default:
2559 return -EINVAL;
2560 }
2561
2562 l = strlen(policy_modes[mode]);
2563 if (buffer + maxlen < p + l + 1)
2564 return -ENOSPC;
2565
2566 strcpy(p, policy_modes[mode]);
2567 p += l;
2568
2569 if (flags & MPOL_MODE_FLAGS) {
2570 if (buffer + maxlen < p + 2)
2571 return -ENOSPC;
2572 *p++ = '=';
2573
2574 /*
2575 * Currently, the only defined flags are mutually exclusive
2576 */
2577 if (flags & MPOL_F_STATIC_NODES)
2578 p += snprintf(p, buffer + maxlen - p, "static");
2579 else if (flags & MPOL_F_RELATIVE_NODES)
2580 p += snprintf(p, buffer + maxlen - p, "relative");
2581 }
2582
2583 if (!nodes_empty(nodes)) {
2584 if (buffer + maxlen < p + 2)
2585 return -ENOSPC;
2586 *p++ = ':';
2587 p += nodelist_scnprintf(p, buffer + maxlen - p, nodes);
2588 }
2589 return p - buffer;
2590}