Loading...
1/*
2 * Simple NUMA memory policy for the Linux kernel.
3 *
4 * Copyright 2003,2004 Andi Kleen, SuSE Labs.
5 * (C) Copyright 2005 Christoph Lameter, Silicon Graphics, Inc.
6 * Subject to the GNU Public License, version 2.
7 *
8 * NUMA policy allows the user to give hints in which node(s) memory should
9 * be allocated.
10 *
11 * Support four policies per VMA and per process:
12 *
13 * The VMA policy has priority over the process policy for a page fault.
14 *
15 * interleave Allocate memory interleaved over a set of nodes,
16 * with normal fallback if it fails.
17 * For VMA based allocations this interleaves based on the
18 * offset into the backing object or offset into the mapping
19 * for anonymous memory. For process policy an process counter
20 * is used.
21 *
22 * bind Only allocate memory on a specific set of nodes,
23 * no fallback.
24 * FIXME: memory is allocated starting with the first node
25 * to the last. It would be better if bind would truly restrict
26 * the allocation to memory nodes instead
27 *
28 * preferred Try a specific node first before normal fallback.
29 * As a special case NUMA_NO_NODE here means do the allocation
30 * on the local CPU. This is normally identical to default,
31 * but useful to set in a VMA when you have a non default
32 * process policy.
33 *
34 * default Allocate on the local node first, or when on a VMA
35 * use the process policy. This is what Linux always did
36 * in a NUMA aware kernel and still does by, ahem, default.
37 *
38 * The process policy is applied for most non interrupt memory allocations
39 * in that process' context. Interrupts ignore the policies and always
40 * try to allocate on the local CPU. The VMA policy is only applied for memory
41 * allocations for a VMA in the VM.
42 *
43 * Currently there are a few corner cases in swapping where the policy
44 * is not applied, but the majority should be handled. When process policy
45 * is used it is not remembered over swap outs/swap ins.
46 *
47 * Only the highest zone in the zone hierarchy gets policied. Allocations
48 * requesting a lower zone just use default policy. This implies that
49 * on systems with highmem kernel lowmem allocation don't get policied.
50 * Same with GFP_DMA allocations.
51 *
52 * For shmfs/tmpfs/hugetlbfs shared memory the policy is shared between
53 * all users and remembered even when nobody has memory mapped.
54 */
55
56/* Notebook:
57 fix mmap readahead to honour policy and enable policy for any page cache
58 object
59 statistics for bigpages
60 global policy for page cache? currently it uses process policy. Requires
61 first item above.
62 handle mremap for shared memory (currently ignored for the policy)
63 grows down?
64 make bind policy root only? It can trigger oom much faster and the
65 kernel is not always grateful with that.
66*/
67
68#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
69
70#include <linux/mempolicy.h>
71#include <linux/mm.h>
72#include <linux/highmem.h>
73#include <linux/hugetlb.h>
74#include <linux/kernel.h>
75#include <linux/sched.h>
76#include <linux/nodemask.h>
77#include <linux/cpuset.h>
78#include <linux/slab.h>
79#include <linux/string.h>
80#include <linux/export.h>
81#include <linux/nsproxy.h>
82#include <linux/interrupt.h>
83#include <linux/init.h>
84#include <linux/compat.h>
85#include <linux/swap.h>
86#include <linux/seq_file.h>
87#include <linux/proc_fs.h>
88#include <linux/migrate.h>
89#include <linux/ksm.h>
90#include <linux/rmap.h>
91#include <linux/security.h>
92#include <linux/syscalls.h>
93#include <linux/ctype.h>
94#include <linux/mm_inline.h>
95#include <linux/mmu_notifier.h>
96#include <linux/printk.h>
97
98#include <asm/tlbflush.h>
99#include <asm/uaccess.h>
100#include <linux/random.h>
101
102#include "internal.h"
103
104/* Internal flags */
105#define MPOL_MF_DISCONTIG_OK (MPOL_MF_INTERNAL << 0) /* Skip checks for continuous vmas */
106#define MPOL_MF_INVERT (MPOL_MF_INTERNAL << 1) /* Invert check for nodemask */
107
108static struct kmem_cache *policy_cache;
109static struct kmem_cache *sn_cache;
110
111/* Highest zone. An specific allocation for a zone below that is not
112 policied. */
113enum zone_type policy_zone = 0;
114
115/*
116 * run-time system-wide default policy => local allocation
117 */
118static struct mempolicy default_policy = {
119 .refcnt = ATOMIC_INIT(1), /* never free it */
120 .mode = MPOL_PREFERRED,
121 .flags = MPOL_F_LOCAL,
122};
123
124static struct mempolicy preferred_node_policy[MAX_NUMNODES];
125
126struct mempolicy *get_task_policy(struct task_struct *p)
127{
128 struct mempolicy *pol = p->mempolicy;
129 int node;
130
131 if (pol)
132 return pol;
133
134 node = numa_node_id();
135 if (node != NUMA_NO_NODE) {
136 pol = &preferred_node_policy[node];
137 /* preferred_node_policy is not initialised early in boot */
138 if (pol->mode)
139 return pol;
140 }
141
142 return &default_policy;
143}
144
145static const struct mempolicy_operations {
146 int (*create)(struct mempolicy *pol, const nodemask_t *nodes);
147 /*
148 * If read-side task has no lock to protect task->mempolicy, write-side
149 * task will rebind the task->mempolicy by two step. The first step is
150 * setting all the newly nodes, and the second step is cleaning all the
151 * disallowed nodes. In this way, we can avoid finding no node to alloc
152 * page.
153 * If we have a lock to protect task->mempolicy in read-side, we do
154 * rebind directly.
155 *
156 * step:
157 * MPOL_REBIND_ONCE - do rebind work at once
158 * MPOL_REBIND_STEP1 - set all the newly nodes
159 * MPOL_REBIND_STEP2 - clean all the disallowed nodes
160 */
161 void (*rebind)(struct mempolicy *pol, const nodemask_t *nodes,
162 enum mpol_rebind_step step);
163} mpol_ops[MPOL_MAX];
164
165static inline int mpol_store_user_nodemask(const struct mempolicy *pol)
166{
167 return pol->flags & MPOL_MODE_FLAGS;
168}
169
170static void mpol_relative_nodemask(nodemask_t *ret, const nodemask_t *orig,
171 const nodemask_t *rel)
172{
173 nodemask_t tmp;
174 nodes_fold(tmp, *orig, nodes_weight(*rel));
175 nodes_onto(*ret, tmp, *rel);
176}
177
178static int mpol_new_interleave(struct mempolicy *pol, const nodemask_t *nodes)
179{
180 if (nodes_empty(*nodes))
181 return -EINVAL;
182 pol->v.nodes = *nodes;
183 return 0;
184}
185
186static int mpol_new_preferred(struct mempolicy *pol, const nodemask_t *nodes)
187{
188 if (!nodes)
189 pol->flags |= MPOL_F_LOCAL; /* local allocation */
190 else if (nodes_empty(*nodes))
191 return -EINVAL; /* no allowed nodes */
192 else
193 pol->v.preferred_node = first_node(*nodes);
194 return 0;
195}
196
197static int mpol_new_bind(struct mempolicy *pol, const nodemask_t *nodes)
198{
199 if (nodes_empty(*nodes))
200 return -EINVAL;
201 pol->v.nodes = *nodes;
202 return 0;
203}
204
205/*
206 * mpol_set_nodemask is called after mpol_new() to set up the nodemask, if
207 * any, for the new policy. mpol_new() has already validated the nodes
208 * parameter with respect to the policy mode and flags. But, we need to
209 * handle an empty nodemask with MPOL_PREFERRED here.
210 *
211 * Must be called holding task's alloc_lock to protect task's mems_allowed
212 * and mempolicy. May also be called holding the mmap_semaphore for write.
213 */
214static int mpol_set_nodemask(struct mempolicy *pol,
215 const nodemask_t *nodes, struct nodemask_scratch *nsc)
216{
217 int ret;
218
219 /* if mode is MPOL_DEFAULT, pol is NULL. This is right. */
220 if (pol == NULL)
221 return 0;
222 /* Check N_MEMORY */
223 nodes_and(nsc->mask1,
224 cpuset_current_mems_allowed, node_states[N_MEMORY]);
225
226 VM_BUG_ON(!nodes);
227 if (pol->mode == MPOL_PREFERRED && nodes_empty(*nodes))
228 nodes = NULL; /* explicit local allocation */
229 else {
230 if (pol->flags & MPOL_F_RELATIVE_NODES)
231 mpol_relative_nodemask(&nsc->mask2, nodes, &nsc->mask1);
232 else
233 nodes_and(nsc->mask2, *nodes, nsc->mask1);
234
235 if (mpol_store_user_nodemask(pol))
236 pol->w.user_nodemask = *nodes;
237 else
238 pol->w.cpuset_mems_allowed =
239 cpuset_current_mems_allowed;
240 }
241
242 if (nodes)
243 ret = mpol_ops[pol->mode].create(pol, &nsc->mask2);
244 else
245 ret = mpol_ops[pol->mode].create(pol, NULL);
246 return ret;
247}
248
249/*
250 * This function just creates a new policy, does some check and simple
251 * initialization. You must invoke mpol_set_nodemask() to set nodes.
252 */
253static struct mempolicy *mpol_new(unsigned short mode, unsigned short flags,
254 nodemask_t *nodes)
255{
256 struct mempolicy *policy;
257
258 pr_debug("setting mode %d flags %d nodes[0] %lx\n",
259 mode, flags, nodes ? nodes_addr(*nodes)[0] : NUMA_NO_NODE);
260
261 if (mode == MPOL_DEFAULT) {
262 if (nodes && !nodes_empty(*nodes))
263 return ERR_PTR(-EINVAL);
264 return NULL;
265 }
266 VM_BUG_ON(!nodes);
267
268 /*
269 * MPOL_PREFERRED cannot be used with MPOL_F_STATIC_NODES or
270 * MPOL_F_RELATIVE_NODES if the nodemask is empty (local allocation).
271 * All other modes require a valid pointer to a non-empty nodemask.
272 */
273 if (mode == MPOL_PREFERRED) {
274 if (nodes_empty(*nodes)) {
275 if (((flags & MPOL_F_STATIC_NODES) ||
276 (flags & MPOL_F_RELATIVE_NODES)))
277 return ERR_PTR(-EINVAL);
278 }
279 } else if (mode == MPOL_LOCAL) {
280 if (!nodes_empty(*nodes))
281 return ERR_PTR(-EINVAL);
282 mode = MPOL_PREFERRED;
283 } else if (nodes_empty(*nodes))
284 return ERR_PTR(-EINVAL);
285 policy = kmem_cache_alloc(policy_cache, GFP_KERNEL);
286 if (!policy)
287 return ERR_PTR(-ENOMEM);
288 atomic_set(&policy->refcnt, 1);
289 policy->mode = mode;
290 policy->flags = flags;
291
292 return policy;
293}
294
295/* Slow path of a mpol destructor. */
296void __mpol_put(struct mempolicy *p)
297{
298 if (!atomic_dec_and_test(&p->refcnt))
299 return;
300 kmem_cache_free(policy_cache, p);
301}
302
303static void mpol_rebind_default(struct mempolicy *pol, const nodemask_t *nodes,
304 enum mpol_rebind_step step)
305{
306}
307
308/*
309 * step:
310 * MPOL_REBIND_ONCE - do rebind work at once
311 * MPOL_REBIND_STEP1 - set all the newly nodes
312 * MPOL_REBIND_STEP2 - clean all the disallowed nodes
313 */
314static void mpol_rebind_nodemask(struct mempolicy *pol, const nodemask_t *nodes,
315 enum mpol_rebind_step step)
316{
317 nodemask_t tmp;
318
319 if (pol->flags & MPOL_F_STATIC_NODES)
320 nodes_and(tmp, pol->w.user_nodemask, *nodes);
321 else if (pol->flags & MPOL_F_RELATIVE_NODES)
322 mpol_relative_nodemask(&tmp, &pol->w.user_nodemask, nodes);
323 else {
324 /*
325 * if step == 1, we use ->w.cpuset_mems_allowed to cache the
326 * result
327 */
328 if (step == MPOL_REBIND_ONCE || step == MPOL_REBIND_STEP1) {
329 nodes_remap(tmp, pol->v.nodes,
330 pol->w.cpuset_mems_allowed, *nodes);
331 pol->w.cpuset_mems_allowed = step ? tmp : *nodes;
332 } else if (step == MPOL_REBIND_STEP2) {
333 tmp = pol->w.cpuset_mems_allowed;
334 pol->w.cpuset_mems_allowed = *nodes;
335 } else
336 BUG();
337 }
338
339 if (nodes_empty(tmp))
340 tmp = *nodes;
341
342 if (step == MPOL_REBIND_STEP1)
343 nodes_or(pol->v.nodes, pol->v.nodes, tmp);
344 else if (step == MPOL_REBIND_ONCE || step == MPOL_REBIND_STEP2)
345 pol->v.nodes = tmp;
346 else
347 BUG();
348
349 if (!node_isset(current->il_next, tmp)) {
350 current->il_next = next_node(current->il_next, tmp);
351 if (current->il_next >= MAX_NUMNODES)
352 current->il_next = first_node(tmp);
353 if (current->il_next >= MAX_NUMNODES)
354 current->il_next = numa_node_id();
355 }
356}
357
358static void mpol_rebind_preferred(struct mempolicy *pol,
359 const nodemask_t *nodes,
360 enum mpol_rebind_step step)
361{
362 nodemask_t tmp;
363
364 if (pol->flags & MPOL_F_STATIC_NODES) {
365 int node = first_node(pol->w.user_nodemask);
366
367 if (node_isset(node, *nodes)) {
368 pol->v.preferred_node = node;
369 pol->flags &= ~MPOL_F_LOCAL;
370 } else
371 pol->flags |= MPOL_F_LOCAL;
372 } else if (pol->flags & MPOL_F_RELATIVE_NODES) {
373 mpol_relative_nodemask(&tmp, &pol->w.user_nodemask, nodes);
374 pol->v.preferred_node = first_node(tmp);
375 } else if (!(pol->flags & MPOL_F_LOCAL)) {
376 pol->v.preferred_node = node_remap(pol->v.preferred_node,
377 pol->w.cpuset_mems_allowed,
378 *nodes);
379 pol->w.cpuset_mems_allowed = *nodes;
380 }
381}
382
383/*
384 * mpol_rebind_policy - Migrate a policy to a different set of nodes
385 *
386 * If read-side task has no lock to protect task->mempolicy, write-side
387 * task will rebind the task->mempolicy by two step. The first step is
388 * setting all the newly nodes, and the second step is cleaning all the
389 * disallowed nodes. In this way, we can avoid finding no node to alloc
390 * page.
391 * If we have a lock to protect task->mempolicy in read-side, we do
392 * rebind directly.
393 *
394 * step:
395 * MPOL_REBIND_ONCE - do rebind work at once
396 * MPOL_REBIND_STEP1 - set all the newly nodes
397 * MPOL_REBIND_STEP2 - clean all the disallowed nodes
398 */
399static void mpol_rebind_policy(struct mempolicy *pol, const nodemask_t *newmask,
400 enum mpol_rebind_step step)
401{
402 if (!pol)
403 return;
404 if (!mpol_store_user_nodemask(pol) && step == MPOL_REBIND_ONCE &&
405 nodes_equal(pol->w.cpuset_mems_allowed, *newmask))
406 return;
407
408 if (step == MPOL_REBIND_STEP1 && (pol->flags & MPOL_F_REBINDING))
409 return;
410
411 if (step == MPOL_REBIND_STEP2 && !(pol->flags & MPOL_F_REBINDING))
412 BUG();
413
414 if (step == MPOL_REBIND_STEP1)
415 pol->flags |= MPOL_F_REBINDING;
416 else if (step == MPOL_REBIND_STEP2)
417 pol->flags &= ~MPOL_F_REBINDING;
418 else if (step >= MPOL_REBIND_NSTEP)
419 BUG();
420
421 mpol_ops[pol->mode].rebind(pol, newmask, step);
422}
423
424/*
425 * Wrapper for mpol_rebind_policy() that just requires task
426 * pointer, and updates task mempolicy.
427 *
428 * Called with task's alloc_lock held.
429 */
430
431void mpol_rebind_task(struct task_struct *tsk, const nodemask_t *new,
432 enum mpol_rebind_step step)
433{
434 mpol_rebind_policy(tsk->mempolicy, new, step);
435}
436
437/*
438 * Rebind each vma in mm to new nodemask.
439 *
440 * Call holding a reference to mm. Takes mm->mmap_sem during call.
441 */
442
443void mpol_rebind_mm(struct mm_struct *mm, nodemask_t *new)
444{
445 struct vm_area_struct *vma;
446
447 down_write(&mm->mmap_sem);
448 for (vma = mm->mmap; vma; vma = vma->vm_next)
449 mpol_rebind_policy(vma->vm_policy, new, MPOL_REBIND_ONCE);
450 up_write(&mm->mmap_sem);
451}
452
453static const struct mempolicy_operations mpol_ops[MPOL_MAX] = {
454 [MPOL_DEFAULT] = {
455 .rebind = mpol_rebind_default,
456 },
457 [MPOL_INTERLEAVE] = {
458 .create = mpol_new_interleave,
459 .rebind = mpol_rebind_nodemask,
460 },
461 [MPOL_PREFERRED] = {
462 .create = mpol_new_preferred,
463 .rebind = mpol_rebind_preferred,
464 },
465 [MPOL_BIND] = {
466 .create = mpol_new_bind,
467 .rebind = mpol_rebind_nodemask,
468 },
469};
470
471static void migrate_page_add(struct page *page, struct list_head *pagelist,
472 unsigned long flags);
473
474struct queue_pages {
475 struct list_head *pagelist;
476 unsigned long flags;
477 nodemask_t *nmask;
478 struct vm_area_struct *prev;
479};
480
481/*
482 * Scan through pages checking if pages follow certain conditions,
483 * and move them to the pagelist if they do.
484 */
485static int queue_pages_pte_range(pmd_t *pmd, unsigned long addr,
486 unsigned long end, struct mm_walk *walk)
487{
488 struct vm_area_struct *vma = walk->vma;
489 struct page *page;
490 struct queue_pages *qp = walk->private;
491 unsigned long flags = qp->flags;
492 int nid, ret;
493 pte_t *pte;
494 spinlock_t *ptl;
495
496 if (pmd_trans_huge(*pmd)) {
497 ptl = pmd_lock(walk->mm, pmd);
498 if (pmd_trans_huge(*pmd)) {
499 page = pmd_page(*pmd);
500 if (is_huge_zero_page(page)) {
501 spin_unlock(ptl);
502 split_huge_pmd(vma, pmd, addr);
503 } else {
504 get_page(page);
505 spin_unlock(ptl);
506 lock_page(page);
507 ret = split_huge_page(page);
508 unlock_page(page);
509 put_page(page);
510 if (ret)
511 return 0;
512 }
513 } else {
514 spin_unlock(ptl);
515 }
516 }
517
518retry:
519 pte = pte_offset_map_lock(walk->mm, pmd, addr, &ptl);
520 for (; addr != end; pte++, addr += PAGE_SIZE) {
521 if (!pte_present(*pte))
522 continue;
523 page = vm_normal_page(vma, addr, *pte);
524 if (!page)
525 continue;
526 /*
527 * vm_normal_page() filters out zero pages, but there might
528 * still be PageReserved pages to skip, perhaps in a VDSO.
529 */
530 if (PageReserved(page))
531 continue;
532 nid = page_to_nid(page);
533 if (node_isset(nid, *qp->nmask) == !!(flags & MPOL_MF_INVERT))
534 continue;
535 if (PageTransCompound(page) && PageAnon(page)) {
536 get_page(page);
537 pte_unmap_unlock(pte, ptl);
538 lock_page(page);
539 ret = split_huge_page(page);
540 unlock_page(page);
541 put_page(page);
542 /* Failed to split -- skip. */
543 if (ret) {
544 pte = pte_offset_map_lock(walk->mm, pmd,
545 addr, &ptl);
546 continue;
547 }
548 goto retry;
549 }
550
551 migrate_page_add(page, qp->pagelist, flags);
552 }
553 pte_unmap_unlock(pte - 1, ptl);
554 cond_resched();
555 return 0;
556}
557
558static int queue_pages_hugetlb(pte_t *pte, unsigned long hmask,
559 unsigned long addr, unsigned long end,
560 struct mm_walk *walk)
561{
562#ifdef CONFIG_HUGETLB_PAGE
563 struct queue_pages *qp = walk->private;
564 unsigned long flags = qp->flags;
565 int nid;
566 struct page *page;
567 spinlock_t *ptl;
568 pte_t entry;
569
570 ptl = huge_pte_lock(hstate_vma(walk->vma), walk->mm, pte);
571 entry = huge_ptep_get(pte);
572 if (!pte_present(entry))
573 goto unlock;
574 page = pte_page(entry);
575 nid = page_to_nid(page);
576 if (node_isset(nid, *qp->nmask) == !!(flags & MPOL_MF_INVERT))
577 goto unlock;
578 /* With MPOL_MF_MOVE, we migrate only unshared hugepage. */
579 if (flags & (MPOL_MF_MOVE_ALL) ||
580 (flags & MPOL_MF_MOVE && page_mapcount(page) == 1))
581 isolate_huge_page(page, qp->pagelist);
582unlock:
583 spin_unlock(ptl);
584#else
585 BUG();
586#endif
587 return 0;
588}
589
590#ifdef CONFIG_NUMA_BALANCING
591/*
592 * This is used to mark a range of virtual addresses to be inaccessible.
593 * These are later cleared by a NUMA hinting fault. Depending on these
594 * faults, pages may be migrated for better NUMA placement.
595 *
596 * This is assuming that NUMA faults are handled using PROT_NONE. If
597 * an architecture makes a different choice, it will need further
598 * changes to the core.
599 */
600unsigned long change_prot_numa(struct vm_area_struct *vma,
601 unsigned long addr, unsigned long end)
602{
603 int nr_updated;
604
605 nr_updated = change_protection(vma, addr, end, PAGE_NONE, 0, 1);
606 if (nr_updated)
607 count_vm_numa_events(NUMA_PTE_UPDATES, nr_updated);
608
609 return nr_updated;
610}
611#else
612static unsigned long change_prot_numa(struct vm_area_struct *vma,
613 unsigned long addr, unsigned long end)
614{
615 return 0;
616}
617#endif /* CONFIG_NUMA_BALANCING */
618
619static int queue_pages_test_walk(unsigned long start, unsigned long end,
620 struct mm_walk *walk)
621{
622 struct vm_area_struct *vma = walk->vma;
623 struct queue_pages *qp = walk->private;
624 unsigned long endvma = vma->vm_end;
625 unsigned long flags = qp->flags;
626
627 if (!vma_migratable(vma))
628 return 1;
629
630 if (endvma > end)
631 endvma = end;
632 if (vma->vm_start > start)
633 start = vma->vm_start;
634
635 if (!(flags & MPOL_MF_DISCONTIG_OK)) {
636 if (!vma->vm_next && vma->vm_end < end)
637 return -EFAULT;
638 if (qp->prev && qp->prev->vm_end < vma->vm_start)
639 return -EFAULT;
640 }
641
642 qp->prev = vma;
643
644 if (flags & MPOL_MF_LAZY) {
645 /* Similar to task_numa_work, skip inaccessible VMAs */
646 if (!is_vm_hugetlb_page(vma) &&
647 (vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)) &&
648 !(vma->vm_flags & VM_MIXEDMAP))
649 change_prot_numa(vma, start, endvma);
650 return 1;
651 }
652
653 /* queue pages from current vma */
654 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))
655 return 0;
656 return 1;
657}
658
659/*
660 * Walk through page tables and collect pages to be migrated.
661 *
662 * If pages found in a given range are on a set of nodes (determined by
663 * @nodes and @flags,) it's isolated and queued to the pagelist which is
664 * passed via @private.)
665 */
666static int
667queue_pages_range(struct mm_struct *mm, unsigned long start, unsigned long end,
668 nodemask_t *nodes, unsigned long flags,
669 struct list_head *pagelist)
670{
671 struct queue_pages qp = {
672 .pagelist = pagelist,
673 .flags = flags,
674 .nmask = nodes,
675 .prev = NULL,
676 };
677 struct mm_walk queue_pages_walk = {
678 .hugetlb_entry = queue_pages_hugetlb,
679 .pmd_entry = queue_pages_pte_range,
680 .test_walk = queue_pages_test_walk,
681 .mm = mm,
682 .private = &qp,
683 };
684
685 return walk_page_range(start, end, &queue_pages_walk);
686}
687
688/*
689 * Apply policy to a single VMA
690 * This must be called with the mmap_sem held for writing.
691 */
692static int vma_replace_policy(struct vm_area_struct *vma,
693 struct mempolicy *pol)
694{
695 int err;
696 struct mempolicy *old;
697 struct mempolicy *new;
698
699 pr_debug("vma %lx-%lx/%lx vm_ops %p vm_file %p set_policy %p\n",
700 vma->vm_start, vma->vm_end, vma->vm_pgoff,
701 vma->vm_ops, vma->vm_file,
702 vma->vm_ops ? vma->vm_ops->set_policy : NULL);
703
704 new = mpol_dup(pol);
705 if (IS_ERR(new))
706 return PTR_ERR(new);
707
708 if (vma->vm_ops && vma->vm_ops->set_policy) {
709 err = vma->vm_ops->set_policy(vma, new);
710 if (err)
711 goto err_out;
712 }
713
714 old = vma->vm_policy;
715 vma->vm_policy = new; /* protected by mmap_sem */
716 mpol_put(old);
717
718 return 0;
719 err_out:
720 mpol_put(new);
721 return err;
722}
723
724/* Step 2: apply policy to a range and do splits. */
725static int mbind_range(struct mm_struct *mm, unsigned long start,
726 unsigned long end, struct mempolicy *new_pol)
727{
728 struct vm_area_struct *next;
729 struct vm_area_struct *prev;
730 struct vm_area_struct *vma;
731 int err = 0;
732 pgoff_t pgoff;
733 unsigned long vmstart;
734 unsigned long vmend;
735
736 vma = find_vma(mm, start);
737 if (!vma || vma->vm_start > start)
738 return -EFAULT;
739
740 prev = vma->vm_prev;
741 if (start > vma->vm_start)
742 prev = vma;
743
744 for (; vma && vma->vm_start < end; prev = vma, vma = next) {
745 next = vma->vm_next;
746 vmstart = max(start, vma->vm_start);
747 vmend = min(end, vma->vm_end);
748
749 if (mpol_equal(vma_policy(vma), new_pol))
750 continue;
751
752 pgoff = vma->vm_pgoff +
753 ((vmstart - vma->vm_start) >> PAGE_SHIFT);
754 prev = vma_merge(mm, prev, vmstart, vmend, vma->vm_flags,
755 vma->anon_vma, vma->vm_file, pgoff,
756 new_pol, vma->vm_userfaultfd_ctx);
757 if (prev) {
758 vma = prev;
759 next = vma->vm_next;
760 if (mpol_equal(vma_policy(vma), new_pol))
761 continue;
762 /* vma_merge() joined vma && vma->next, case 8 */
763 goto replace;
764 }
765 if (vma->vm_start != vmstart) {
766 err = split_vma(vma->vm_mm, vma, vmstart, 1);
767 if (err)
768 goto out;
769 }
770 if (vma->vm_end != vmend) {
771 err = split_vma(vma->vm_mm, vma, vmend, 0);
772 if (err)
773 goto out;
774 }
775 replace:
776 err = vma_replace_policy(vma, new_pol);
777 if (err)
778 goto out;
779 }
780
781 out:
782 return err;
783}
784
785/* Set the process memory policy */
786static long do_set_mempolicy(unsigned short mode, unsigned short flags,
787 nodemask_t *nodes)
788{
789 struct mempolicy *new, *old;
790 NODEMASK_SCRATCH(scratch);
791 int ret;
792
793 if (!scratch)
794 return -ENOMEM;
795
796 new = mpol_new(mode, flags, nodes);
797 if (IS_ERR(new)) {
798 ret = PTR_ERR(new);
799 goto out;
800 }
801
802 task_lock(current);
803 ret = mpol_set_nodemask(new, nodes, scratch);
804 if (ret) {
805 task_unlock(current);
806 mpol_put(new);
807 goto out;
808 }
809 old = current->mempolicy;
810 current->mempolicy = new;
811 if (new && new->mode == MPOL_INTERLEAVE &&
812 nodes_weight(new->v.nodes))
813 current->il_next = first_node(new->v.nodes);
814 task_unlock(current);
815 mpol_put(old);
816 ret = 0;
817out:
818 NODEMASK_SCRATCH_FREE(scratch);
819 return ret;
820}
821
822/*
823 * Return nodemask for policy for get_mempolicy() query
824 *
825 * Called with task's alloc_lock held
826 */
827static void get_policy_nodemask(struct mempolicy *p, nodemask_t *nodes)
828{
829 nodes_clear(*nodes);
830 if (p == &default_policy)
831 return;
832
833 switch (p->mode) {
834 case MPOL_BIND:
835 /* Fall through */
836 case MPOL_INTERLEAVE:
837 *nodes = p->v.nodes;
838 break;
839 case MPOL_PREFERRED:
840 if (!(p->flags & MPOL_F_LOCAL))
841 node_set(p->v.preferred_node, *nodes);
842 /* else return empty node mask for local allocation */
843 break;
844 default:
845 BUG();
846 }
847}
848
849static int lookup_node(unsigned long addr)
850{
851 struct page *p;
852 int err;
853
854 err = get_user_pages(addr & PAGE_MASK, 1, 0, 0, &p, NULL);
855 if (err >= 0) {
856 err = page_to_nid(p);
857 put_page(p);
858 }
859 return err;
860}
861
862/* Retrieve NUMA policy */
863static long do_get_mempolicy(int *policy, nodemask_t *nmask,
864 unsigned long addr, unsigned long flags)
865{
866 int err;
867 struct mm_struct *mm = current->mm;
868 struct vm_area_struct *vma = NULL;
869 struct mempolicy *pol = current->mempolicy;
870
871 if (flags &
872 ~(unsigned long)(MPOL_F_NODE|MPOL_F_ADDR|MPOL_F_MEMS_ALLOWED))
873 return -EINVAL;
874
875 if (flags & MPOL_F_MEMS_ALLOWED) {
876 if (flags & (MPOL_F_NODE|MPOL_F_ADDR))
877 return -EINVAL;
878 *policy = 0; /* just so it's initialized */
879 task_lock(current);
880 *nmask = cpuset_current_mems_allowed;
881 task_unlock(current);
882 return 0;
883 }
884
885 if (flags & MPOL_F_ADDR) {
886 /*
887 * Do NOT fall back to task policy if the
888 * vma/shared policy at addr is NULL. We
889 * want to return MPOL_DEFAULT in this case.
890 */
891 down_read(&mm->mmap_sem);
892 vma = find_vma_intersection(mm, addr, addr+1);
893 if (!vma) {
894 up_read(&mm->mmap_sem);
895 return -EFAULT;
896 }
897 if (vma->vm_ops && vma->vm_ops->get_policy)
898 pol = vma->vm_ops->get_policy(vma, addr);
899 else
900 pol = vma->vm_policy;
901 } else if (addr)
902 return -EINVAL;
903
904 if (!pol)
905 pol = &default_policy; /* indicates default behavior */
906
907 if (flags & MPOL_F_NODE) {
908 if (flags & MPOL_F_ADDR) {
909 err = lookup_node(addr);
910 if (err < 0)
911 goto out;
912 *policy = err;
913 } else if (pol == current->mempolicy &&
914 pol->mode == MPOL_INTERLEAVE) {
915 *policy = current->il_next;
916 } else {
917 err = -EINVAL;
918 goto out;
919 }
920 } else {
921 *policy = pol == &default_policy ? MPOL_DEFAULT :
922 pol->mode;
923 /*
924 * Internal mempolicy flags must be masked off before exposing
925 * the policy to userspace.
926 */
927 *policy |= (pol->flags & MPOL_MODE_FLAGS);
928 }
929
930 if (vma) {
931 up_read(¤t->mm->mmap_sem);
932 vma = NULL;
933 }
934
935 err = 0;
936 if (nmask) {
937 if (mpol_store_user_nodemask(pol)) {
938 *nmask = pol->w.user_nodemask;
939 } else {
940 task_lock(current);
941 get_policy_nodemask(pol, nmask);
942 task_unlock(current);
943 }
944 }
945
946 out:
947 mpol_cond_put(pol);
948 if (vma)
949 up_read(¤t->mm->mmap_sem);
950 return err;
951}
952
953#ifdef CONFIG_MIGRATION
954/*
955 * page migration
956 */
957static void migrate_page_add(struct page *page, struct list_head *pagelist,
958 unsigned long flags)
959{
960 /*
961 * Avoid migrating a page that is shared with others.
962 */
963 if ((flags & MPOL_MF_MOVE_ALL) || page_mapcount(page) == 1) {
964 if (!isolate_lru_page(page)) {
965 list_add_tail(&page->lru, pagelist);
966 inc_zone_page_state(page, NR_ISOLATED_ANON +
967 page_is_file_cache(page));
968 }
969 }
970}
971
972static struct page *new_node_page(struct page *page, unsigned long node, int **x)
973{
974 if (PageHuge(page))
975 return alloc_huge_page_node(page_hstate(compound_head(page)),
976 node);
977 else
978 return __alloc_pages_node(node, GFP_HIGHUSER_MOVABLE |
979 __GFP_THISNODE, 0);
980}
981
982/*
983 * Migrate pages from one node to a target node.
984 * Returns error or the number of pages not migrated.
985 */
986static int migrate_to_node(struct mm_struct *mm, int source, int dest,
987 int flags)
988{
989 nodemask_t nmask;
990 LIST_HEAD(pagelist);
991 int err = 0;
992
993 nodes_clear(nmask);
994 node_set(source, nmask);
995
996 /*
997 * This does not "check" the range but isolates all pages that
998 * need migration. Between passing in the full user address
999 * space range and MPOL_MF_DISCONTIG_OK, this call can not fail.
1000 */
1001 VM_BUG_ON(!(flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)));
1002 queue_pages_range(mm, mm->mmap->vm_start, mm->task_size, &nmask,
1003 flags | MPOL_MF_DISCONTIG_OK, &pagelist);
1004
1005 if (!list_empty(&pagelist)) {
1006 err = migrate_pages(&pagelist, new_node_page, NULL, dest,
1007 MIGRATE_SYNC, MR_SYSCALL);
1008 if (err)
1009 putback_movable_pages(&pagelist);
1010 }
1011
1012 return err;
1013}
1014
1015/*
1016 * Move pages between the two nodesets so as to preserve the physical
1017 * layout as much as possible.
1018 *
1019 * Returns the number of page that could not be moved.
1020 */
1021int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
1022 const nodemask_t *to, int flags)
1023{
1024 int busy = 0;
1025 int err;
1026 nodemask_t tmp;
1027
1028 err = migrate_prep();
1029 if (err)
1030 return err;
1031
1032 down_read(&mm->mmap_sem);
1033
1034 /*
1035 * Find a 'source' bit set in 'tmp' whose corresponding 'dest'
1036 * bit in 'to' is not also set in 'tmp'. Clear the found 'source'
1037 * bit in 'tmp', and return that <source, dest> pair for migration.
1038 * The pair of nodemasks 'to' and 'from' define the map.
1039 *
1040 * If no pair of bits is found that way, fallback to picking some
1041 * pair of 'source' and 'dest' bits that are not the same. If the
1042 * 'source' and 'dest' bits are the same, this represents a node
1043 * that will be migrating to itself, so no pages need move.
1044 *
1045 * If no bits are left in 'tmp', or if all remaining bits left
1046 * in 'tmp' correspond to the same bit in 'to', return false
1047 * (nothing left to migrate).
1048 *
1049 * This lets us pick a pair of nodes to migrate between, such that
1050 * if possible the dest node is not already occupied by some other
1051 * source node, minimizing the risk of overloading the memory on a
1052 * node that would happen if we migrated incoming memory to a node
1053 * before migrating outgoing memory source that same node.
1054 *
1055 * A single scan of tmp is sufficient. As we go, we remember the
1056 * most recent <s, d> pair that moved (s != d). If we find a pair
1057 * that not only moved, but what's better, moved to an empty slot
1058 * (d is not set in tmp), then we break out then, with that pair.
1059 * Otherwise when we finish scanning from_tmp, we at least have the
1060 * most recent <s, d> pair that moved. If we get all the way through
1061 * the scan of tmp without finding any node that moved, much less
1062 * moved to an empty node, then there is nothing left worth migrating.
1063 */
1064
1065 tmp = *from;
1066 while (!nodes_empty(tmp)) {
1067 int s,d;
1068 int source = NUMA_NO_NODE;
1069 int dest = 0;
1070
1071 for_each_node_mask(s, tmp) {
1072
1073 /*
1074 * do_migrate_pages() tries to maintain the relative
1075 * node relationship of the pages established between
1076 * threads and memory areas.
1077 *
1078 * However if the number of source nodes is not equal to
1079 * the number of destination nodes we can not preserve
1080 * this node relative relationship. In that case, skip
1081 * copying memory from a node that is in the destination
1082 * mask.
1083 *
1084 * Example: [2,3,4] -> [3,4,5] moves everything.
1085 * [0-7] - > [3,4,5] moves only 0,1,2,6,7.
1086 */
1087
1088 if ((nodes_weight(*from) != nodes_weight(*to)) &&
1089 (node_isset(s, *to)))
1090 continue;
1091
1092 d = node_remap(s, *from, *to);
1093 if (s == d)
1094 continue;
1095
1096 source = s; /* Node moved. Memorize */
1097 dest = d;
1098
1099 /* dest not in remaining from nodes? */
1100 if (!node_isset(dest, tmp))
1101 break;
1102 }
1103 if (source == NUMA_NO_NODE)
1104 break;
1105
1106 node_clear(source, tmp);
1107 err = migrate_to_node(mm, source, dest, flags);
1108 if (err > 0)
1109 busy += err;
1110 if (err < 0)
1111 break;
1112 }
1113 up_read(&mm->mmap_sem);
1114 if (err < 0)
1115 return err;
1116 return busy;
1117
1118}
1119
1120/*
1121 * Allocate a new page for page migration based on vma policy.
1122 * Start by assuming the page is mapped by the same vma as contains @start.
1123 * Search forward from there, if not. N.B., this assumes that the
1124 * list of pages handed to migrate_pages()--which is how we get here--
1125 * is in virtual address order.
1126 */
1127static struct page *new_page(struct page *page, unsigned long start, int **x)
1128{
1129 struct vm_area_struct *vma;
1130 unsigned long uninitialized_var(address);
1131
1132 vma = find_vma(current->mm, start);
1133 while (vma) {
1134 address = page_address_in_vma(page, vma);
1135 if (address != -EFAULT)
1136 break;
1137 vma = vma->vm_next;
1138 }
1139
1140 if (PageHuge(page)) {
1141 BUG_ON(!vma);
1142 return alloc_huge_page_noerr(vma, address, 1);
1143 }
1144 /*
1145 * if !vma, alloc_page_vma() will use task or system default policy
1146 */
1147 return alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
1148}
1149#else
1150
1151static void migrate_page_add(struct page *page, struct list_head *pagelist,
1152 unsigned long flags)
1153{
1154}
1155
1156int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
1157 const nodemask_t *to, int flags)
1158{
1159 return -ENOSYS;
1160}
1161
1162static struct page *new_page(struct page *page, unsigned long start, int **x)
1163{
1164 return NULL;
1165}
1166#endif
1167
1168static long do_mbind(unsigned long start, unsigned long len,
1169 unsigned short mode, unsigned short mode_flags,
1170 nodemask_t *nmask, unsigned long flags)
1171{
1172 struct mm_struct *mm = current->mm;
1173 struct mempolicy *new;
1174 unsigned long end;
1175 int err;
1176 LIST_HEAD(pagelist);
1177
1178 if (flags & ~(unsigned long)MPOL_MF_VALID)
1179 return -EINVAL;
1180 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1181 return -EPERM;
1182
1183 if (start & ~PAGE_MASK)
1184 return -EINVAL;
1185
1186 if (mode == MPOL_DEFAULT)
1187 flags &= ~MPOL_MF_STRICT;
1188
1189 len = (len + PAGE_SIZE - 1) & PAGE_MASK;
1190 end = start + len;
1191
1192 if (end < start)
1193 return -EINVAL;
1194 if (end == start)
1195 return 0;
1196
1197 new = mpol_new(mode, mode_flags, nmask);
1198 if (IS_ERR(new))
1199 return PTR_ERR(new);
1200
1201 if (flags & MPOL_MF_LAZY)
1202 new->flags |= MPOL_F_MOF;
1203
1204 /*
1205 * If we are using the default policy then operation
1206 * on discontinuous address spaces is okay after all
1207 */
1208 if (!new)
1209 flags |= MPOL_MF_DISCONTIG_OK;
1210
1211 pr_debug("mbind %lx-%lx mode:%d flags:%d nodes:%lx\n",
1212 start, start + len, mode, mode_flags,
1213 nmask ? nodes_addr(*nmask)[0] : NUMA_NO_NODE);
1214
1215 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) {
1216
1217 err = migrate_prep();
1218 if (err)
1219 goto mpol_out;
1220 }
1221 {
1222 NODEMASK_SCRATCH(scratch);
1223 if (scratch) {
1224 down_write(&mm->mmap_sem);
1225 task_lock(current);
1226 err = mpol_set_nodemask(new, nmask, scratch);
1227 task_unlock(current);
1228 if (err)
1229 up_write(&mm->mmap_sem);
1230 } else
1231 err = -ENOMEM;
1232 NODEMASK_SCRATCH_FREE(scratch);
1233 }
1234 if (err)
1235 goto mpol_out;
1236
1237 err = queue_pages_range(mm, start, end, nmask,
1238 flags | MPOL_MF_INVERT, &pagelist);
1239 if (!err)
1240 err = mbind_range(mm, start, end, new);
1241
1242 if (!err) {
1243 int nr_failed = 0;
1244
1245 if (!list_empty(&pagelist)) {
1246 WARN_ON_ONCE(flags & MPOL_MF_LAZY);
1247 nr_failed = migrate_pages(&pagelist, new_page, NULL,
1248 start, MIGRATE_SYNC, MR_MEMPOLICY_MBIND);
1249 if (nr_failed)
1250 putback_movable_pages(&pagelist);
1251 }
1252
1253 if (nr_failed && (flags & MPOL_MF_STRICT))
1254 err = -EIO;
1255 } else
1256 putback_movable_pages(&pagelist);
1257
1258 up_write(&mm->mmap_sem);
1259 mpol_out:
1260 mpol_put(new);
1261 return err;
1262}
1263
1264/*
1265 * User space interface with variable sized bitmaps for nodelists.
1266 */
1267
1268/* Copy a node mask from user space. */
1269static int get_nodes(nodemask_t *nodes, const unsigned long __user *nmask,
1270 unsigned long maxnode)
1271{
1272 unsigned long k;
1273 unsigned long nlongs;
1274 unsigned long endmask;
1275
1276 --maxnode;
1277 nodes_clear(*nodes);
1278 if (maxnode == 0 || !nmask)
1279 return 0;
1280 if (maxnode > PAGE_SIZE*BITS_PER_BYTE)
1281 return -EINVAL;
1282
1283 nlongs = BITS_TO_LONGS(maxnode);
1284 if ((maxnode % BITS_PER_LONG) == 0)
1285 endmask = ~0UL;
1286 else
1287 endmask = (1UL << (maxnode % BITS_PER_LONG)) - 1;
1288
1289 /* When the user specified more nodes than supported just check
1290 if the non supported part is all zero. */
1291 if (nlongs > BITS_TO_LONGS(MAX_NUMNODES)) {
1292 if (nlongs > PAGE_SIZE/sizeof(long))
1293 return -EINVAL;
1294 for (k = BITS_TO_LONGS(MAX_NUMNODES); k < nlongs; k++) {
1295 unsigned long t;
1296 if (get_user(t, nmask + k))
1297 return -EFAULT;
1298 if (k == nlongs - 1) {
1299 if (t & endmask)
1300 return -EINVAL;
1301 } else if (t)
1302 return -EINVAL;
1303 }
1304 nlongs = BITS_TO_LONGS(MAX_NUMNODES);
1305 endmask = ~0UL;
1306 }
1307
1308 if (copy_from_user(nodes_addr(*nodes), nmask, nlongs*sizeof(unsigned long)))
1309 return -EFAULT;
1310 nodes_addr(*nodes)[nlongs-1] &= endmask;
1311 return 0;
1312}
1313
1314/* Copy a kernel node mask to user space */
1315static int copy_nodes_to_user(unsigned long __user *mask, unsigned long maxnode,
1316 nodemask_t *nodes)
1317{
1318 unsigned long copy = ALIGN(maxnode-1, 64) / 8;
1319 const int nbytes = BITS_TO_LONGS(MAX_NUMNODES) * sizeof(long);
1320
1321 if (copy > nbytes) {
1322 if (copy > PAGE_SIZE)
1323 return -EINVAL;
1324 if (clear_user((char __user *)mask + nbytes, copy - nbytes))
1325 return -EFAULT;
1326 copy = nbytes;
1327 }
1328 return copy_to_user(mask, nodes_addr(*nodes), copy) ? -EFAULT : 0;
1329}
1330
1331SYSCALL_DEFINE6(mbind, unsigned long, start, unsigned long, len,
1332 unsigned long, mode, const unsigned long __user *, nmask,
1333 unsigned long, maxnode, unsigned, flags)
1334{
1335 nodemask_t nodes;
1336 int err;
1337 unsigned short mode_flags;
1338
1339 mode_flags = mode & MPOL_MODE_FLAGS;
1340 mode &= ~MPOL_MODE_FLAGS;
1341 if (mode >= MPOL_MAX)
1342 return -EINVAL;
1343 if ((mode_flags & MPOL_F_STATIC_NODES) &&
1344 (mode_flags & MPOL_F_RELATIVE_NODES))
1345 return -EINVAL;
1346 err = get_nodes(&nodes, nmask, maxnode);
1347 if (err)
1348 return err;
1349 return do_mbind(start, len, mode, mode_flags, &nodes, flags);
1350}
1351
1352/* Set the process memory policy */
1353SYSCALL_DEFINE3(set_mempolicy, int, mode, const unsigned long __user *, nmask,
1354 unsigned long, maxnode)
1355{
1356 int err;
1357 nodemask_t nodes;
1358 unsigned short flags;
1359
1360 flags = mode & MPOL_MODE_FLAGS;
1361 mode &= ~MPOL_MODE_FLAGS;
1362 if ((unsigned int)mode >= MPOL_MAX)
1363 return -EINVAL;
1364 if ((flags & MPOL_F_STATIC_NODES) && (flags & MPOL_F_RELATIVE_NODES))
1365 return -EINVAL;
1366 err = get_nodes(&nodes, nmask, maxnode);
1367 if (err)
1368 return err;
1369 return do_set_mempolicy(mode, flags, &nodes);
1370}
1371
1372SYSCALL_DEFINE4(migrate_pages, pid_t, pid, unsigned long, maxnode,
1373 const unsigned long __user *, old_nodes,
1374 const unsigned long __user *, new_nodes)
1375{
1376 const struct cred *cred = current_cred(), *tcred;
1377 struct mm_struct *mm = NULL;
1378 struct task_struct *task;
1379 nodemask_t task_nodes;
1380 int err;
1381 nodemask_t *old;
1382 nodemask_t *new;
1383 NODEMASK_SCRATCH(scratch);
1384
1385 if (!scratch)
1386 return -ENOMEM;
1387
1388 old = &scratch->mask1;
1389 new = &scratch->mask2;
1390
1391 err = get_nodes(old, old_nodes, maxnode);
1392 if (err)
1393 goto out;
1394
1395 err = get_nodes(new, new_nodes, maxnode);
1396 if (err)
1397 goto out;
1398
1399 /* Find the mm_struct */
1400 rcu_read_lock();
1401 task = pid ? find_task_by_vpid(pid) : current;
1402 if (!task) {
1403 rcu_read_unlock();
1404 err = -ESRCH;
1405 goto out;
1406 }
1407 get_task_struct(task);
1408
1409 err = -EINVAL;
1410
1411 /*
1412 * Check if this process has the right to modify the specified
1413 * process. The right exists if the process has administrative
1414 * capabilities, superuser privileges or the same
1415 * userid as the target process.
1416 */
1417 tcred = __task_cred(task);
1418 if (!uid_eq(cred->euid, tcred->suid) && !uid_eq(cred->euid, tcred->uid) &&
1419 !uid_eq(cred->uid, tcred->suid) && !uid_eq(cred->uid, tcred->uid) &&
1420 !capable(CAP_SYS_NICE)) {
1421 rcu_read_unlock();
1422 err = -EPERM;
1423 goto out_put;
1424 }
1425 rcu_read_unlock();
1426
1427 task_nodes = cpuset_mems_allowed(task);
1428 /* Is the user allowed to access the target nodes? */
1429 if (!nodes_subset(*new, task_nodes) && !capable(CAP_SYS_NICE)) {
1430 err = -EPERM;
1431 goto out_put;
1432 }
1433
1434 if (!nodes_subset(*new, node_states[N_MEMORY])) {
1435 err = -EINVAL;
1436 goto out_put;
1437 }
1438
1439 err = security_task_movememory(task);
1440 if (err)
1441 goto out_put;
1442
1443 mm = get_task_mm(task);
1444 put_task_struct(task);
1445
1446 if (!mm) {
1447 err = -EINVAL;
1448 goto out;
1449 }
1450
1451 err = do_migrate_pages(mm, old, new,
1452 capable(CAP_SYS_NICE) ? MPOL_MF_MOVE_ALL : MPOL_MF_MOVE);
1453
1454 mmput(mm);
1455out:
1456 NODEMASK_SCRATCH_FREE(scratch);
1457
1458 return err;
1459
1460out_put:
1461 put_task_struct(task);
1462 goto out;
1463
1464}
1465
1466
1467/* Retrieve NUMA policy */
1468SYSCALL_DEFINE5(get_mempolicy, int __user *, policy,
1469 unsigned long __user *, nmask, unsigned long, maxnode,
1470 unsigned long, addr, unsigned long, flags)
1471{
1472 int err;
1473 int uninitialized_var(pval);
1474 nodemask_t nodes;
1475
1476 if (nmask != NULL && maxnode < MAX_NUMNODES)
1477 return -EINVAL;
1478
1479 err = do_get_mempolicy(&pval, &nodes, addr, flags);
1480
1481 if (err)
1482 return err;
1483
1484 if (policy && put_user(pval, policy))
1485 return -EFAULT;
1486
1487 if (nmask)
1488 err = copy_nodes_to_user(nmask, maxnode, &nodes);
1489
1490 return err;
1491}
1492
1493#ifdef CONFIG_COMPAT
1494
1495COMPAT_SYSCALL_DEFINE5(get_mempolicy, int __user *, policy,
1496 compat_ulong_t __user *, nmask,
1497 compat_ulong_t, maxnode,
1498 compat_ulong_t, addr, compat_ulong_t, flags)
1499{
1500 long err;
1501 unsigned long __user *nm = NULL;
1502 unsigned long nr_bits, alloc_size;
1503 DECLARE_BITMAP(bm, MAX_NUMNODES);
1504
1505 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1506 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1507
1508 if (nmask)
1509 nm = compat_alloc_user_space(alloc_size);
1510
1511 err = sys_get_mempolicy(policy, nm, nr_bits+1, addr, flags);
1512
1513 if (!err && nmask) {
1514 unsigned long copy_size;
1515 copy_size = min_t(unsigned long, sizeof(bm), alloc_size);
1516 err = copy_from_user(bm, nm, copy_size);
1517 /* ensure entire bitmap is zeroed */
1518 err |= clear_user(nmask, ALIGN(maxnode-1, 8) / 8);
1519 err |= compat_put_bitmap(nmask, bm, nr_bits);
1520 }
1521
1522 return err;
1523}
1524
1525COMPAT_SYSCALL_DEFINE3(set_mempolicy, int, mode, compat_ulong_t __user *, nmask,
1526 compat_ulong_t, maxnode)
1527{
1528 long err = 0;
1529 unsigned long __user *nm = NULL;
1530 unsigned long nr_bits, alloc_size;
1531 DECLARE_BITMAP(bm, MAX_NUMNODES);
1532
1533 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1534 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1535
1536 if (nmask) {
1537 err = compat_get_bitmap(bm, nmask, nr_bits);
1538 nm = compat_alloc_user_space(alloc_size);
1539 err |= copy_to_user(nm, bm, alloc_size);
1540 }
1541
1542 if (err)
1543 return -EFAULT;
1544
1545 return sys_set_mempolicy(mode, nm, nr_bits+1);
1546}
1547
1548COMPAT_SYSCALL_DEFINE6(mbind, compat_ulong_t, start, compat_ulong_t, len,
1549 compat_ulong_t, mode, compat_ulong_t __user *, nmask,
1550 compat_ulong_t, maxnode, compat_ulong_t, flags)
1551{
1552 long err = 0;
1553 unsigned long __user *nm = NULL;
1554 unsigned long nr_bits, alloc_size;
1555 nodemask_t bm;
1556
1557 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1558 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1559
1560 if (nmask) {
1561 err = compat_get_bitmap(nodes_addr(bm), nmask, nr_bits);
1562 nm = compat_alloc_user_space(alloc_size);
1563 err |= copy_to_user(nm, nodes_addr(bm), alloc_size);
1564 }
1565
1566 if (err)
1567 return -EFAULT;
1568
1569 return sys_mbind(start, len, mode, nm, nr_bits+1, flags);
1570}
1571
1572#endif
1573
1574struct mempolicy *__get_vma_policy(struct vm_area_struct *vma,
1575 unsigned long addr)
1576{
1577 struct mempolicy *pol = NULL;
1578
1579 if (vma) {
1580 if (vma->vm_ops && vma->vm_ops->get_policy) {
1581 pol = vma->vm_ops->get_policy(vma, addr);
1582 } else if (vma->vm_policy) {
1583 pol = vma->vm_policy;
1584
1585 /*
1586 * shmem_alloc_page() passes MPOL_F_SHARED policy with
1587 * a pseudo vma whose vma->vm_ops=NULL. Take a reference
1588 * count on these policies which will be dropped by
1589 * mpol_cond_put() later
1590 */
1591 if (mpol_needs_cond_ref(pol))
1592 mpol_get(pol);
1593 }
1594 }
1595
1596 return pol;
1597}
1598
1599/*
1600 * get_vma_policy(@vma, @addr)
1601 * @vma: virtual memory area whose policy is sought
1602 * @addr: address in @vma for shared policy lookup
1603 *
1604 * Returns effective policy for a VMA at specified address.
1605 * Falls back to current->mempolicy or system default policy, as necessary.
1606 * Shared policies [those marked as MPOL_F_SHARED] require an extra reference
1607 * count--added by the get_policy() vm_op, as appropriate--to protect against
1608 * freeing by another task. It is the caller's responsibility to free the
1609 * extra reference for shared policies.
1610 */
1611static struct mempolicy *get_vma_policy(struct vm_area_struct *vma,
1612 unsigned long addr)
1613{
1614 struct mempolicy *pol = __get_vma_policy(vma, addr);
1615
1616 if (!pol)
1617 pol = get_task_policy(current);
1618
1619 return pol;
1620}
1621
1622bool vma_policy_mof(struct vm_area_struct *vma)
1623{
1624 struct mempolicy *pol;
1625
1626 if (vma->vm_ops && vma->vm_ops->get_policy) {
1627 bool ret = false;
1628
1629 pol = vma->vm_ops->get_policy(vma, vma->vm_start);
1630 if (pol && (pol->flags & MPOL_F_MOF))
1631 ret = true;
1632 mpol_cond_put(pol);
1633
1634 return ret;
1635 }
1636
1637 pol = vma->vm_policy;
1638 if (!pol)
1639 pol = get_task_policy(current);
1640
1641 return pol->flags & MPOL_F_MOF;
1642}
1643
1644static int apply_policy_zone(struct mempolicy *policy, enum zone_type zone)
1645{
1646 enum zone_type dynamic_policy_zone = policy_zone;
1647
1648 BUG_ON(dynamic_policy_zone == ZONE_MOVABLE);
1649
1650 /*
1651 * if policy->v.nodes has movable memory only,
1652 * we apply policy when gfp_zone(gfp) = ZONE_MOVABLE only.
1653 *
1654 * policy->v.nodes is intersect with node_states[N_MEMORY].
1655 * so if the following test faile, it implies
1656 * policy->v.nodes has movable memory only.
1657 */
1658 if (!nodes_intersects(policy->v.nodes, node_states[N_HIGH_MEMORY]))
1659 dynamic_policy_zone = ZONE_MOVABLE;
1660
1661 return zone >= dynamic_policy_zone;
1662}
1663
1664/*
1665 * Return a nodemask representing a mempolicy for filtering nodes for
1666 * page allocation
1667 */
1668static nodemask_t *policy_nodemask(gfp_t gfp, struct mempolicy *policy)
1669{
1670 /* Lower zones don't get a nodemask applied for MPOL_BIND */
1671 if (unlikely(policy->mode == MPOL_BIND) &&
1672 apply_policy_zone(policy, gfp_zone(gfp)) &&
1673 cpuset_nodemask_valid_mems_allowed(&policy->v.nodes))
1674 return &policy->v.nodes;
1675
1676 return NULL;
1677}
1678
1679/* Return a zonelist indicated by gfp for node representing a mempolicy */
1680static struct zonelist *policy_zonelist(gfp_t gfp, struct mempolicy *policy,
1681 int nd)
1682{
1683 switch (policy->mode) {
1684 case MPOL_PREFERRED:
1685 if (!(policy->flags & MPOL_F_LOCAL))
1686 nd = policy->v.preferred_node;
1687 break;
1688 case MPOL_BIND:
1689 /*
1690 * Normally, MPOL_BIND allocations are node-local within the
1691 * allowed nodemask. However, if __GFP_THISNODE is set and the
1692 * current node isn't part of the mask, we use the zonelist for
1693 * the first node in the mask instead.
1694 */
1695 if (unlikely(gfp & __GFP_THISNODE) &&
1696 unlikely(!node_isset(nd, policy->v.nodes)))
1697 nd = first_node(policy->v.nodes);
1698 break;
1699 default:
1700 BUG();
1701 }
1702 return node_zonelist(nd, gfp);
1703}
1704
1705/* Do dynamic interleaving for a process */
1706static unsigned interleave_nodes(struct mempolicy *policy)
1707{
1708 unsigned nid, next;
1709 struct task_struct *me = current;
1710
1711 nid = me->il_next;
1712 next = next_node(nid, policy->v.nodes);
1713 if (next >= MAX_NUMNODES)
1714 next = first_node(policy->v.nodes);
1715 if (next < MAX_NUMNODES)
1716 me->il_next = next;
1717 return nid;
1718}
1719
1720/*
1721 * Depending on the memory policy provide a node from which to allocate the
1722 * next slab entry.
1723 */
1724unsigned int mempolicy_slab_node(void)
1725{
1726 struct mempolicy *policy;
1727 int node = numa_mem_id();
1728
1729 if (in_interrupt())
1730 return node;
1731
1732 policy = current->mempolicy;
1733 if (!policy || policy->flags & MPOL_F_LOCAL)
1734 return node;
1735
1736 switch (policy->mode) {
1737 case MPOL_PREFERRED:
1738 /*
1739 * handled MPOL_F_LOCAL above
1740 */
1741 return policy->v.preferred_node;
1742
1743 case MPOL_INTERLEAVE:
1744 return interleave_nodes(policy);
1745
1746 case MPOL_BIND: {
1747 /*
1748 * Follow bind policy behavior and start allocation at the
1749 * first node.
1750 */
1751 struct zonelist *zonelist;
1752 struct zone *zone;
1753 enum zone_type highest_zoneidx = gfp_zone(GFP_KERNEL);
1754 zonelist = &NODE_DATA(node)->node_zonelists[0];
1755 (void)first_zones_zonelist(zonelist, highest_zoneidx,
1756 &policy->v.nodes,
1757 &zone);
1758 return zone ? zone->node : node;
1759 }
1760
1761 default:
1762 BUG();
1763 }
1764}
1765
1766/* Do static interleaving for a VMA with known offset. */
1767static unsigned offset_il_node(struct mempolicy *pol,
1768 struct vm_area_struct *vma, unsigned long off)
1769{
1770 unsigned nnodes = nodes_weight(pol->v.nodes);
1771 unsigned target;
1772 int c;
1773 int nid = NUMA_NO_NODE;
1774
1775 if (!nnodes)
1776 return numa_node_id();
1777 target = (unsigned int)off % nnodes;
1778 c = 0;
1779 do {
1780 nid = next_node(nid, pol->v.nodes);
1781 c++;
1782 } while (c <= target);
1783 return nid;
1784}
1785
1786/* Determine a node number for interleave */
1787static inline unsigned interleave_nid(struct mempolicy *pol,
1788 struct vm_area_struct *vma, unsigned long addr, int shift)
1789{
1790 if (vma) {
1791 unsigned long off;
1792
1793 /*
1794 * for small pages, there is no difference between
1795 * shift and PAGE_SHIFT, so the bit-shift is safe.
1796 * for huge pages, since vm_pgoff is in units of small
1797 * pages, we need to shift off the always 0 bits to get
1798 * a useful offset.
1799 */
1800 BUG_ON(shift < PAGE_SHIFT);
1801 off = vma->vm_pgoff >> (shift - PAGE_SHIFT);
1802 off += (addr - vma->vm_start) >> shift;
1803 return offset_il_node(pol, vma, off);
1804 } else
1805 return interleave_nodes(pol);
1806}
1807
1808/*
1809 * Return the bit number of a random bit set in the nodemask.
1810 * (returns NUMA_NO_NODE if nodemask is empty)
1811 */
1812int node_random(const nodemask_t *maskp)
1813{
1814 int w, bit = NUMA_NO_NODE;
1815
1816 w = nodes_weight(*maskp);
1817 if (w)
1818 bit = bitmap_ord_to_pos(maskp->bits,
1819 get_random_int() % w, MAX_NUMNODES);
1820 return bit;
1821}
1822
1823#ifdef CONFIG_HUGETLBFS
1824/*
1825 * huge_zonelist(@vma, @addr, @gfp_flags, @mpol)
1826 * @vma: virtual memory area whose policy is sought
1827 * @addr: address in @vma for shared policy lookup and interleave policy
1828 * @gfp_flags: for requested zone
1829 * @mpol: pointer to mempolicy pointer for reference counted mempolicy
1830 * @nodemask: pointer to nodemask pointer for MPOL_BIND nodemask
1831 *
1832 * Returns a zonelist suitable for a huge page allocation and a pointer
1833 * to the struct mempolicy for conditional unref after allocation.
1834 * If the effective policy is 'BIND, returns a pointer to the mempolicy's
1835 * @nodemask for filtering the zonelist.
1836 *
1837 * Must be protected by read_mems_allowed_begin()
1838 */
1839struct zonelist *huge_zonelist(struct vm_area_struct *vma, unsigned long addr,
1840 gfp_t gfp_flags, struct mempolicy **mpol,
1841 nodemask_t **nodemask)
1842{
1843 struct zonelist *zl;
1844
1845 *mpol = get_vma_policy(vma, addr);
1846 *nodemask = NULL; /* assume !MPOL_BIND */
1847
1848 if (unlikely((*mpol)->mode == MPOL_INTERLEAVE)) {
1849 zl = node_zonelist(interleave_nid(*mpol, vma, addr,
1850 huge_page_shift(hstate_vma(vma))), gfp_flags);
1851 } else {
1852 zl = policy_zonelist(gfp_flags, *mpol, numa_node_id());
1853 if ((*mpol)->mode == MPOL_BIND)
1854 *nodemask = &(*mpol)->v.nodes;
1855 }
1856 return zl;
1857}
1858
1859/*
1860 * init_nodemask_of_mempolicy
1861 *
1862 * If the current task's mempolicy is "default" [NULL], return 'false'
1863 * to indicate default policy. Otherwise, extract the policy nodemask
1864 * for 'bind' or 'interleave' policy into the argument nodemask, or
1865 * initialize the argument nodemask to contain the single node for
1866 * 'preferred' or 'local' policy and return 'true' to indicate presence
1867 * of non-default mempolicy.
1868 *
1869 * We don't bother with reference counting the mempolicy [mpol_get/put]
1870 * because the current task is examining it's own mempolicy and a task's
1871 * mempolicy is only ever changed by the task itself.
1872 *
1873 * N.B., it is the caller's responsibility to free a returned nodemask.
1874 */
1875bool init_nodemask_of_mempolicy(nodemask_t *mask)
1876{
1877 struct mempolicy *mempolicy;
1878 int nid;
1879
1880 if (!(mask && current->mempolicy))
1881 return false;
1882
1883 task_lock(current);
1884 mempolicy = current->mempolicy;
1885 switch (mempolicy->mode) {
1886 case MPOL_PREFERRED:
1887 if (mempolicy->flags & MPOL_F_LOCAL)
1888 nid = numa_node_id();
1889 else
1890 nid = mempolicy->v.preferred_node;
1891 init_nodemask_of_node(mask, nid);
1892 break;
1893
1894 case MPOL_BIND:
1895 /* Fall through */
1896 case MPOL_INTERLEAVE:
1897 *mask = mempolicy->v.nodes;
1898 break;
1899
1900 default:
1901 BUG();
1902 }
1903 task_unlock(current);
1904
1905 return true;
1906}
1907#endif
1908
1909/*
1910 * mempolicy_nodemask_intersects
1911 *
1912 * If tsk's mempolicy is "default" [NULL], return 'true' to indicate default
1913 * policy. Otherwise, check for intersection between mask and the policy
1914 * nodemask for 'bind' or 'interleave' policy. For 'perferred' or 'local'
1915 * policy, always return true since it may allocate elsewhere on fallback.
1916 *
1917 * Takes task_lock(tsk) to prevent freeing of its mempolicy.
1918 */
1919bool mempolicy_nodemask_intersects(struct task_struct *tsk,
1920 const nodemask_t *mask)
1921{
1922 struct mempolicy *mempolicy;
1923 bool ret = true;
1924
1925 if (!mask)
1926 return ret;
1927 task_lock(tsk);
1928 mempolicy = tsk->mempolicy;
1929 if (!mempolicy)
1930 goto out;
1931
1932 switch (mempolicy->mode) {
1933 case MPOL_PREFERRED:
1934 /*
1935 * MPOL_PREFERRED and MPOL_F_LOCAL are only preferred nodes to
1936 * allocate from, they may fallback to other nodes when oom.
1937 * Thus, it's possible for tsk to have allocated memory from
1938 * nodes in mask.
1939 */
1940 break;
1941 case MPOL_BIND:
1942 case MPOL_INTERLEAVE:
1943 ret = nodes_intersects(mempolicy->v.nodes, *mask);
1944 break;
1945 default:
1946 BUG();
1947 }
1948out:
1949 task_unlock(tsk);
1950 return ret;
1951}
1952
1953/* Allocate a page in interleaved policy.
1954 Own path because it needs to do special accounting. */
1955static struct page *alloc_page_interleave(gfp_t gfp, unsigned order,
1956 unsigned nid)
1957{
1958 struct zonelist *zl;
1959 struct page *page;
1960
1961 zl = node_zonelist(nid, gfp);
1962 page = __alloc_pages(gfp, order, zl);
1963 if (page && page_zone(page) == zonelist_zone(&zl->_zonerefs[0]))
1964 inc_zone_page_state(page, NUMA_INTERLEAVE_HIT);
1965 return page;
1966}
1967
1968/**
1969 * alloc_pages_vma - Allocate a page for a VMA.
1970 *
1971 * @gfp:
1972 * %GFP_USER user allocation.
1973 * %GFP_KERNEL kernel allocations,
1974 * %GFP_HIGHMEM highmem/user allocations,
1975 * %GFP_FS allocation should not call back into a file system.
1976 * %GFP_ATOMIC don't sleep.
1977 *
1978 * @order:Order of the GFP allocation.
1979 * @vma: Pointer to VMA or NULL if not available.
1980 * @addr: Virtual Address of the allocation. Must be inside the VMA.
1981 * @node: Which node to prefer for allocation (modulo policy).
1982 * @hugepage: for hugepages try only the preferred node if possible
1983 *
1984 * This function allocates a page from the kernel page pool and applies
1985 * a NUMA policy associated with the VMA or the current process.
1986 * When VMA is not NULL caller must hold down_read on the mmap_sem of the
1987 * mm_struct of the VMA to prevent it from going away. Should be used for
1988 * all allocations for pages that will be mapped into user space. Returns
1989 * NULL when no page can be allocated.
1990 */
1991struct page *
1992alloc_pages_vma(gfp_t gfp, int order, struct vm_area_struct *vma,
1993 unsigned long addr, int node, bool hugepage)
1994{
1995 struct mempolicy *pol;
1996 struct page *page;
1997 unsigned int cpuset_mems_cookie;
1998 struct zonelist *zl;
1999 nodemask_t *nmask;
2000
2001retry_cpuset:
2002 pol = get_vma_policy(vma, addr);
2003 cpuset_mems_cookie = read_mems_allowed_begin();
2004
2005 if (pol->mode == MPOL_INTERLEAVE) {
2006 unsigned nid;
2007
2008 nid = interleave_nid(pol, vma, addr, PAGE_SHIFT + order);
2009 mpol_cond_put(pol);
2010 page = alloc_page_interleave(gfp, order, nid);
2011 goto out;
2012 }
2013
2014 if (unlikely(IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE) && hugepage)) {
2015 int hpage_node = node;
2016
2017 /*
2018 * For hugepage allocation and non-interleave policy which
2019 * allows the current node (or other explicitly preferred
2020 * node) we only try to allocate from the current/preferred
2021 * node and don't fall back to other nodes, as the cost of
2022 * remote accesses would likely offset THP benefits.
2023 *
2024 * If the policy is interleave, or does not allow the current
2025 * node in its nodemask, we allocate the standard way.
2026 */
2027 if (pol->mode == MPOL_PREFERRED &&
2028 !(pol->flags & MPOL_F_LOCAL))
2029 hpage_node = pol->v.preferred_node;
2030
2031 nmask = policy_nodemask(gfp, pol);
2032 if (!nmask || node_isset(hpage_node, *nmask)) {
2033 mpol_cond_put(pol);
2034 page = __alloc_pages_node(hpage_node,
2035 gfp | __GFP_THISNODE, order);
2036 goto out;
2037 }
2038 }
2039
2040 nmask = policy_nodemask(gfp, pol);
2041 zl = policy_zonelist(gfp, pol, node);
2042 mpol_cond_put(pol);
2043 page = __alloc_pages_nodemask(gfp, order, zl, nmask);
2044out:
2045 if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie)))
2046 goto retry_cpuset;
2047 return page;
2048}
2049
2050/**
2051 * alloc_pages_current - Allocate pages.
2052 *
2053 * @gfp:
2054 * %GFP_USER user allocation,
2055 * %GFP_KERNEL kernel allocation,
2056 * %GFP_HIGHMEM highmem allocation,
2057 * %GFP_FS don't call back into a file system.
2058 * %GFP_ATOMIC don't sleep.
2059 * @order: Power of two of allocation size in pages. 0 is a single page.
2060 *
2061 * Allocate a page from the kernel page pool. When not in
2062 * interrupt context and apply the current process NUMA policy.
2063 * Returns NULL when no page can be allocated.
2064 *
2065 * Don't call cpuset_update_task_memory_state() unless
2066 * 1) it's ok to take cpuset_sem (can WAIT), and
2067 * 2) allocating for current task (not interrupt).
2068 */
2069struct page *alloc_pages_current(gfp_t gfp, unsigned order)
2070{
2071 struct mempolicy *pol = &default_policy;
2072 struct page *page;
2073 unsigned int cpuset_mems_cookie;
2074
2075 if (!in_interrupt() && !(gfp & __GFP_THISNODE))
2076 pol = get_task_policy(current);
2077
2078retry_cpuset:
2079 cpuset_mems_cookie = read_mems_allowed_begin();
2080
2081 /*
2082 * No reference counting needed for current->mempolicy
2083 * nor system default_policy
2084 */
2085 if (pol->mode == MPOL_INTERLEAVE)
2086 page = alloc_page_interleave(gfp, order, interleave_nodes(pol));
2087 else
2088 page = __alloc_pages_nodemask(gfp, order,
2089 policy_zonelist(gfp, pol, numa_node_id()),
2090 policy_nodemask(gfp, pol));
2091
2092 if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie)))
2093 goto retry_cpuset;
2094
2095 return page;
2096}
2097EXPORT_SYMBOL(alloc_pages_current);
2098
2099int vma_dup_policy(struct vm_area_struct *src, struct vm_area_struct *dst)
2100{
2101 struct mempolicy *pol = mpol_dup(vma_policy(src));
2102
2103 if (IS_ERR(pol))
2104 return PTR_ERR(pol);
2105 dst->vm_policy = pol;
2106 return 0;
2107}
2108
2109/*
2110 * If mpol_dup() sees current->cpuset == cpuset_being_rebound, then it
2111 * rebinds the mempolicy its copying by calling mpol_rebind_policy()
2112 * with the mems_allowed returned by cpuset_mems_allowed(). This
2113 * keeps mempolicies cpuset relative after its cpuset moves. See
2114 * further kernel/cpuset.c update_nodemask().
2115 *
2116 * current's mempolicy may be rebinded by the other task(the task that changes
2117 * cpuset's mems), so we needn't do rebind work for current task.
2118 */
2119
2120/* Slow path of a mempolicy duplicate */
2121struct mempolicy *__mpol_dup(struct mempolicy *old)
2122{
2123 struct mempolicy *new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
2124
2125 if (!new)
2126 return ERR_PTR(-ENOMEM);
2127
2128 /* task's mempolicy is protected by alloc_lock */
2129 if (old == current->mempolicy) {
2130 task_lock(current);
2131 *new = *old;
2132 task_unlock(current);
2133 } else
2134 *new = *old;
2135
2136 if (current_cpuset_is_being_rebound()) {
2137 nodemask_t mems = cpuset_mems_allowed(current);
2138 if (new->flags & MPOL_F_REBINDING)
2139 mpol_rebind_policy(new, &mems, MPOL_REBIND_STEP2);
2140 else
2141 mpol_rebind_policy(new, &mems, MPOL_REBIND_ONCE);
2142 }
2143 atomic_set(&new->refcnt, 1);
2144 return new;
2145}
2146
2147/* Slow path of a mempolicy comparison */
2148bool __mpol_equal(struct mempolicy *a, struct mempolicy *b)
2149{
2150 if (!a || !b)
2151 return false;
2152 if (a->mode != b->mode)
2153 return false;
2154 if (a->flags != b->flags)
2155 return false;
2156 if (mpol_store_user_nodemask(a))
2157 if (!nodes_equal(a->w.user_nodemask, b->w.user_nodemask))
2158 return false;
2159
2160 switch (a->mode) {
2161 case MPOL_BIND:
2162 /* Fall through */
2163 case MPOL_INTERLEAVE:
2164 return !!nodes_equal(a->v.nodes, b->v.nodes);
2165 case MPOL_PREFERRED:
2166 return a->v.preferred_node == b->v.preferred_node;
2167 default:
2168 BUG();
2169 return false;
2170 }
2171}
2172
2173/*
2174 * Shared memory backing store policy support.
2175 *
2176 * Remember policies even when nobody has shared memory mapped.
2177 * The policies are kept in Red-Black tree linked from the inode.
2178 * They are protected by the sp->lock rwlock, which should be held
2179 * for any accesses to the tree.
2180 */
2181
2182/*
2183 * lookup first element intersecting start-end. Caller holds sp->lock for
2184 * reading or for writing
2185 */
2186static struct sp_node *
2187sp_lookup(struct shared_policy *sp, unsigned long start, unsigned long end)
2188{
2189 struct rb_node *n = sp->root.rb_node;
2190
2191 while (n) {
2192 struct sp_node *p = rb_entry(n, struct sp_node, nd);
2193
2194 if (start >= p->end)
2195 n = n->rb_right;
2196 else if (end <= p->start)
2197 n = n->rb_left;
2198 else
2199 break;
2200 }
2201 if (!n)
2202 return NULL;
2203 for (;;) {
2204 struct sp_node *w = NULL;
2205 struct rb_node *prev = rb_prev(n);
2206 if (!prev)
2207 break;
2208 w = rb_entry(prev, struct sp_node, nd);
2209 if (w->end <= start)
2210 break;
2211 n = prev;
2212 }
2213 return rb_entry(n, struct sp_node, nd);
2214}
2215
2216/*
2217 * Insert a new shared policy into the list. Caller holds sp->lock for
2218 * writing.
2219 */
2220static void sp_insert(struct shared_policy *sp, struct sp_node *new)
2221{
2222 struct rb_node **p = &sp->root.rb_node;
2223 struct rb_node *parent = NULL;
2224 struct sp_node *nd;
2225
2226 while (*p) {
2227 parent = *p;
2228 nd = rb_entry(parent, struct sp_node, nd);
2229 if (new->start < nd->start)
2230 p = &(*p)->rb_left;
2231 else if (new->end > nd->end)
2232 p = &(*p)->rb_right;
2233 else
2234 BUG();
2235 }
2236 rb_link_node(&new->nd, parent, p);
2237 rb_insert_color(&new->nd, &sp->root);
2238 pr_debug("inserting %lx-%lx: %d\n", new->start, new->end,
2239 new->policy ? new->policy->mode : 0);
2240}
2241
2242/* Find shared policy intersecting idx */
2243struct mempolicy *
2244mpol_shared_policy_lookup(struct shared_policy *sp, unsigned long idx)
2245{
2246 struct mempolicy *pol = NULL;
2247 struct sp_node *sn;
2248
2249 if (!sp->root.rb_node)
2250 return NULL;
2251 read_lock(&sp->lock);
2252 sn = sp_lookup(sp, idx, idx+1);
2253 if (sn) {
2254 mpol_get(sn->policy);
2255 pol = sn->policy;
2256 }
2257 read_unlock(&sp->lock);
2258 return pol;
2259}
2260
2261static void sp_free(struct sp_node *n)
2262{
2263 mpol_put(n->policy);
2264 kmem_cache_free(sn_cache, n);
2265}
2266
2267/**
2268 * mpol_misplaced - check whether current page node is valid in policy
2269 *
2270 * @page: page to be checked
2271 * @vma: vm area where page mapped
2272 * @addr: virtual address where page mapped
2273 *
2274 * Lookup current policy node id for vma,addr and "compare to" page's
2275 * node id.
2276 *
2277 * Returns:
2278 * -1 - not misplaced, page is in the right node
2279 * node - node id where the page should be
2280 *
2281 * Policy determination "mimics" alloc_page_vma().
2282 * Called from fault path where we know the vma and faulting address.
2283 */
2284int mpol_misplaced(struct page *page, struct vm_area_struct *vma, unsigned long addr)
2285{
2286 struct mempolicy *pol;
2287 struct zone *zone;
2288 int curnid = page_to_nid(page);
2289 unsigned long pgoff;
2290 int thiscpu = raw_smp_processor_id();
2291 int thisnid = cpu_to_node(thiscpu);
2292 int polnid = -1;
2293 int ret = -1;
2294
2295 BUG_ON(!vma);
2296
2297 pol = get_vma_policy(vma, addr);
2298 if (!(pol->flags & MPOL_F_MOF))
2299 goto out;
2300
2301 switch (pol->mode) {
2302 case MPOL_INTERLEAVE:
2303 BUG_ON(addr >= vma->vm_end);
2304 BUG_ON(addr < vma->vm_start);
2305
2306 pgoff = vma->vm_pgoff;
2307 pgoff += (addr - vma->vm_start) >> PAGE_SHIFT;
2308 polnid = offset_il_node(pol, vma, pgoff);
2309 break;
2310
2311 case MPOL_PREFERRED:
2312 if (pol->flags & MPOL_F_LOCAL)
2313 polnid = numa_node_id();
2314 else
2315 polnid = pol->v.preferred_node;
2316 break;
2317
2318 case MPOL_BIND:
2319 /*
2320 * allows binding to multiple nodes.
2321 * use current page if in policy nodemask,
2322 * else select nearest allowed node, if any.
2323 * If no allowed nodes, use current [!misplaced].
2324 */
2325 if (node_isset(curnid, pol->v.nodes))
2326 goto out;
2327 (void)first_zones_zonelist(
2328 node_zonelist(numa_node_id(), GFP_HIGHUSER),
2329 gfp_zone(GFP_HIGHUSER),
2330 &pol->v.nodes, &zone);
2331 polnid = zone->node;
2332 break;
2333
2334 default:
2335 BUG();
2336 }
2337
2338 /* Migrate the page towards the node whose CPU is referencing it */
2339 if (pol->flags & MPOL_F_MORON) {
2340 polnid = thisnid;
2341
2342 if (!should_numa_migrate_memory(current, page, curnid, thiscpu))
2343 goto out;
2344 }
2345
2346 if (curnid != polnid)
2347 ret = polnid;
2348out:
2349 mpol_cond_put(pol);
2350
2351 return ret;
2352}
2353
2354static void sp_delete(struct shared_policy *sp, struct sp_node *n)
2355{
2356 pr_debug("deleting %lx-l%lx\n", n->start, n->end);
2357 rb_erase(&n->nd, &sp->root);
2358 sp_free(n);
2359}
2360
2361static void sp_node_init(struct sp_node *node, unsigned long start,
2362 unsigned long end, struct mempolicy *pol)
2363{
2364 node->start = start;
2365 node->end = end;
2366 node->policy = pol;
2367}
2368
2369static struct sp_node *sp_alloc(unsigned long start, unsigned long end,
2370 struct mempolicy *pol)
2371{
2372 struct sp_node *n;
2373 struct mempolicy *newpol;
2374
2375 n = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2376 if (!n)
2377 return NULL;
2378
2379 newpol = mpol_dup(pol);
2380 if (IS_ERR(newpol)) {
2381 kmem_cache_free(sn_cache, n);
2382 return NULL;
2383 }
2384 newpol->flags |= MPOL_F_SHARED;
2385 sp_node_init(n, start, end, newpol);
2386
2387 return n;
2388}
2389
2390/* Replace a policy range. */
2391static int shared_policy_replace(struct shared_policy *sp, unsigned long start,
2392 unsigned long end, struct sp_node *new)
2393{
2394 struct sp_node *n;
2395 struct sp_node *n_new = NULL;
2396 struct mempolicy *mpol_new = NULL;
2397 int ret = 0;
2398
2399restart:
2400 write_lock(&sp->lock);
2401 n = sp_lookup(sp, start, end);
2402 /* Take care of old policies in the same range. */
2403 while (n && n->start < end) {
2404 struct rb_node *next = rb_next(&n->nd);
2405 if (n->start >= start) {
2406 if (n->end <= end)
2407 sp_delete(sp, n);
2408 else
2409 n->start = end;
2410 } else {
2411 /* Old policy spanning whole new range. */
2412 if (n->end > end) {
2413 if (!n_new)
2414 goto alloc_new;
2415
2416 *mpol_new = *n->policy;
2417 atomic_set(&mpol_new->refcnt, 1);
2418 sp_node_init(n_new, end, n->end, mpol_new);
2419 n->end = start;
2420 sp_insert(sp, n_new);
2421 n_new = NULL;
2422 mpol_new = NULL;
2423 break;
2424 } else
2425 n->end = start;
2426 }
2427 if (!next)
2428 break;
2429 n = rb_entry(next, struct sp_node, nd);
2430 }
2431 if (new)
2432 sp_insert(sp, new);
2433 write_unlock(&sp->lock);
2434 ret = 0;
2435
2436err_out:
2437 if (mpol_new)
2438 mpol_put(mpol_new);
2439 if (n_new)
2440 kmem_cache_free(sn_cache, n_new);
2441
2442 return ret;
2443
2444alloc_new:
2445 write_unlock(&sp->lock);
2446 ret = -ENOMEM;
2447 n_new = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2448 if (!n_new)
2449 goto err_out;
2450 mpol_new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
2451 if (!mpol_new)
2452 goto err_out;
2453 goto restart;
2454}
2455
2456/**
2457 * mpol_shared_policy_init - initialize shared policy for inode
2458 * @sp: pointer to inode shared policy
2459 * @mpol: struct mempolicy to install
2460 *
2461 * Install non-NULL @mpol in inode's shared policy rb-tree.
2462 * On entry, the current task has a reference on a non-NULL @mpol.
2463 * This must be released on exit.
2464 * This is called at get_inode() calls and we can use GFP_KERNEL.
2465 */
2466void mpol_shared_policy_init(struct shared_policy *sp, struct mempolicy *mpol)
2467{
2468 int ret;
2469
2470 sp->root = RB_ROOT; /* empty tree == default mempolicy */
2471 rwlock_init(&sp->lock);
2472
2473 if (mpol) {
2474 struct vm_area_struct pvma;
2475 struct mempolicy *new;
2476 NODEMASK_SCRATCH(scratch);
2477
2478 if (!scratch)
2479 goto put_mpol;
2480 /* contextualize the tmpfs mount point mempolicy */
2481 new = mpol_new(mpol->mode, mpol->flags, &mpol->w.user_nodemask);
2482 if (IS_ERR(new))
2483 goto free_scratch; /* no valid nodemask intersection */
2484
2485 task_lock(current);
2486 ret = mpol_set_nodemask(new, &mpol->w.user_nodemask, scratch);
2487 task_unlock(current);
2488 if (ret)
2489 goto put_new;
2490
2491 /* Create pseudo-vma that contains just the policy */
2492 memset(&pvma, 0, sizeof(struct vm_area_struct));
2493 pvma.vm_end = TASK_SIZE; /* policy covers entire file */
2494 mpol_set_shared_policy(sp, &pvma, new); /* adds ref */
2495
2496put_new:
2497 mpol_put(new); /* drop initial ref */
2498free_scratch:
2499 NODEMASK_SCRATCH_FREE(scratch);
2500put_mpol:
2501 mpol_put(mpol); /* drop our incoming ref on sb mpol */
2502 }
2503}
2504
2505int mpol_set_shared_policy(struct shared_policy *info,
2506 struct vm_area_struct *vma, struct mempolicy *npol)
2507{
2508 int err;
2509 struct sp_node *new = NULL;
2510 unsigned long sz = vma_pages(vma);
2511
2512 pr_debug("set_shared_policy %lx sz %lu %d %d %lx\n",
2513 vma->vm_pgoff,
2514 sz, npol ? npol->mode : -1,
2515 npol ? npol->flags : -1,
2516 npol ? nodes_addr(npol->v.nodes)[0] : NUMA_NO_NODE);
2517
2518 if (npol) {
2519 new = sp_alloc(vma->vm_pgoff, vma->vm_pgoff + sz, npol);
2520 if (!new)
2521 return -ENOMEM;
2522 }
2523 err = shared_policy_replace(info, vma->vm_pgoff, vma->vm_pgoff+sz, new);
2524 if (err && new)
2525 sp_free(new);
2526 return err;
2527}
2528
2529/* Free a backing policy store on inode delete. */
2530void mpol_free_shared_policy(struct shared_policy *p)
2531{
2532 struct sp_node *n;
2533 struct rb_node *next;
2534
2535 if (!p->root.rb_node)
2536 return;
2537 write_lock(&p->lock);
2538 next = rb_first(&p->root);
2539 while (next) {
2540 n = rb_entry(next, struct sp_node, nd);
2541 next = rb_next(&n->nd);
2542 sp_delete(p, n);
2543 }
2544 write_unlock(&p->lock);
2545}
2546
2547#ifdef CONFIG_NUMA_BALANCING
2548static int __initdata numabalancing_override;
2549
2550static void __init check_numabalancing_enable(void)
2551{
2552 bool numabalancing_default = false;
2553
2554 if (IS_ENABLED(CONFIG_NUMA_BALANCING_DEFAULT_ENABLED))
2555 numabalancing_default = true;
2556
2557 /* Parsed by setup_numabalancing. override == 1 enables, -1 disables */
2558 if (numabalancing_override)
2559 set_numabalancing_state(numabalancing_override == 1);
2560
2561 if (num_online_nodes() > 1 && !numabalancing_override) {
2562 pr_info("%s automatic NUMA balancing. Configure with numa_balancing= or the kernel.numa_balancing sysctl\n",
2563 numabalancing_default ? "Enabling" : "Disabling");
2564 set_numabalancing_state(numabalancing_default);
2565 }
2566}
2567
2568static int __init setup_numabalancing(char *str)
2569{
2570 int ret = 0;
2571 if (!str)
2572 goto out;
2573
2574 if (!strcmp(str, "enable")) {
2575 numabalancing_override = 1;
2576 ret = 1;
2577 } else if (!strcmp(str, "disable")) {
2578 numabalancing_override = -1;
2579 ret = 1;
2580 }
2581out:
2582 if (!ret)
2583 pr_warn("Unable to parse numa_balancing=\n");
2584
2585 return ret;
2586}
2587__setup("numa_balancing=", setup_numabalancing);
2588#else
2589static inline void __init check_numabalancing_enable(void)
2590{
2591}
2592#endif /* CONFIG_NUMA_BALANCING */
2593
2594/* assumes fs == KERNEL_DS */
2595void __init numa_policy_init(void)
2596{
2597 nodemask_t interleave_nodes;
2598 unsigned long largest = 0;
2599 int nid, prefer = 0;
2600
2601 policy_cache = kmem_cache_create("numa_policy",
2602 sizeof(struct mempolicy),
2603 0, SLAB_PANIC, NULL);
2604
2605 sn_cache = kmem_cache_create("shared_policy_node",
2606 sizeof(struct sp_node),
2607 0, SLAB_PANIC, NULL);
2608
2609 for_each_node(nid) {
2610 preferred_node_policy[nid] = (struct mempolicy) {
2611 .refcnt = ATOMIC_INIT(1),
2612 .mode = MPOL_PREFERRED,
2613 .flags = MPOL_F_MOF | MPOL_F_MORON,
2614 .v = { .preferred_node = nid, },
2615 };
2616 }
2617
2618 /*
2619 * Set interleaving policy for system init. Interleaving is only
2620 * enabled across suitably sized nodes (default is >= 16MB), or
2621 * fall back to the largest node if they're all smaller.
2622 */
2623 nodes_clear(interleave_nodes);
2624 for_each_node_state(nid, N_MEMORY) {
2625 unsigned long total_pages = node_present_pages(nid);
2626
2627 /* Preserve the largest node */
2628 if (largest < total_pages) {
2629 largest = total_pages;
2630 prefer = nid;
2631 }
2632
2633 /* Interleave this node? */
2634 if ((total_pages << PAGE_SHIFT) >= (16 << 20))
2635 node_set(nid, interleave_nodes);
2636 }
2637
2638 /* All too small, use the largest */
2639 if (unlikely(nodes_empty(interleave_nodes)))
2640 node_set(prefer, interleave_nodes);
2641
2642 if (do_set_mempolicy(MPOL_INTERLEAVE, 0, &interleave_nodes))
2643 pr_err("%s: interleaving failed\n", __func__);
2644
2645 check_numabalancing_enable();
2646}
2647
2648/* Reset policy of current process to default */
2649void numa_default_policy(void)
2650{
2651 do_set_mempolicy(MPOL_DEFAULT, 0, NULL);
2652}
2653
2654/*
2655 * Parse and format mempolicy from/to strings
2656 */
2657
2658/*
2659 * "local" is implemented internally by MPOL_PREFERRED with MPOL_F_LOCAL flag.
2660 */
2661static const char * const policy_modes[] =
2662{
2663 [MPOL_DEFAULT] = "default",
2664 [MPOL_PREFERRED] = "prefer",
2665 [MPOL_BIND] = "bind",
2666 [MPOL_INTERLEAVE] = "interleave",
2667 [MPOL_LOCAL] = "local",
2668};
2669
2670
2671#ifdef CONFIG_TMPFS
2672/**
2673 * mpol_parse_str - parse string to mempolicy, for tmpfs mpol mount option.
2674 * @str: string containing mempolicy to parse
2675 * @mpol: pointer to struct mempolicy pointer, returned on success.
2676 *
2677 * Format of input:
2678 * <mode>[=<flags>][:<nodelist>]
2679 *
2680 * On success, returns 0, else 1
2681 */
2682int mpol_parse_str(char *str, struct mempolicy **mpol)
2683{
2684 struct mempolicy *new = NULL;
2685 unsigned short mode;
2686 unsigned short mode_flags;
2687 nodemask_t nodes;
2688 char *nodelist = strchr(str, ':');
2689 char *flags = strchr(str, '=');
2690 int err = 1;
2691
2692 if (nodelist) {
2693 /* NUL-terminate mode or flags string */
2694 *nodelist++ = '\0';
2695 if (nodelist_parse(nodelist, nodes))
2696 goto out;
2697 if (!nodes_subset(nodes, node_states[N_MEMORY]))
2698 goto out;
2699 } else
2700 nodes_clear(nodes);
2701
2702 if (flags)
2703 *flags++ = '\0'; /* terminate mode string */
2704
2705 for (mode = 0; mode < MPOL_MAX; mode++) {
2706 if (!strcmp(str, policy_modes[mode])) {
2707 break;
2708 }
2709 }
2710 if (mode >= MPOL_MAX)
2711 goto out;
2712
2713 switch (mode) {
2714 case MPOL_PREFERRED:
2715 /*
2716 * Insist on a nodelist of one node only
2717 */
2718 if (nodelist) {
2719 char *rest = nodelist;
2720 while (isdigit(*rest))
2721 rest++;
2722 if (*rest)
2723 goto out;
2724 }
2725 break;
2726 case MPOL_INTERLEAVE:
2727 /*
2728 * Default to online nodes with memory if no nodelist
2729 */
2730 if (!nodelist)
2731 nodes = node_states[N_MEMORY];
2732 break;
2733 case MPOL_LOCAL:
2734 /*
2735 * Don't allow a nodelist; mpol_new() checks flags
2736 */
2737 if (nodelist)
2738 goto out;
2739 mode = MPOL_PREFERRED;
2740 break;
2741 case MPOL_DEFAULT:
2742 /*
2743 * Insist on a empty nodelist
2744 */
2745 if (!nodelist)
2746 err = 0;
2747 goto out;
2748 case MPOL_BIND:
2749 /*
2750 * Insist on a nodelist
2751 */
2752 if (!nodelist)
2753 goto out;
2754 }
2755
2756 mode_flags = 0;
2757 if (flags) {
2758 /*
2759 * Currently, we only support two mutually exclusive
2760 * mode flags.
2761 */
2762 if (!strcmp(flags, "static"))
2763 mode_flags |= MPOL_F_STATIC_NODES;
2764 else if (!strcmp(flags, "relative"))
2765 mode_flags |= MPOL_F_RELATIVE_NODES;
2766 else
2767 goto out;
2768 }
2769
2770 new = mpol_new(mode, mode_flags, &nodes);
2771 if (IS_ERR(new))
2772 goto out;
2773
2774 /*
2775 * Save nodes for mpol_to_str() to show the tmpfs mount options
2776 * for /proc/mounts, /proc/pid/mounts and /proc/pid/mountinfo.
2777 */
2778 if (mode != MPOL_PREFERRED)
2779 new->v.nodes = nodes;
2780 else if (nodelist)
2781 new->v.preferred_node = first_node(nodes);
2782 else
2783 new->flags |= MPOL_F_LOCAL;
2784
2785 /*
2786 * Save nodes for contextualization: this will be used to "clone"
2787 * the mempolicy in a specific context [cpuset] at a later time.
2788 */
2789 new->w.user_nodemask = nodes;
2790
2791 err = 0;
2792
2793out:
2794 /* Restore string for error message */
2795 if (nodelist)
2796 *--nodelist = ':';
2797 if (flags)
2798 *--flags = '=';
2799 if (!err)
2800 *mpol = new;
2801 return err;
2802}
2803#endif /* CONFIG_TMPFS */
2804
2805/**
2806 * mpol_to_str - format a mempolicy structure for printing
2807 * @buffer: to contain formatted mempolicy string
2808 * @maxlen: length of @buffer
2809 * @pol: pointer to mempolicy to be formatted
2810 *
2811 * Convert @pol into a string. If @buffer is too short, truncate the string.
2812 * Recommend a @maxlen of at least 32 for the longest mode, "interleave", the
2813 * longest flag, "relative", and to display at least a few node ids.
2814 */
2815void mpol_to_str(char *buffer, int maxlen, struct mempolicy *pol)
2816{
2817 char *p = buffer;
2818 nodemask_t nodes = NODE_MASK_NONE;
2819 unsigned short mode = MPOL_DEFAULT;
2820 unsigned short flags = 0;
2821
2822 if (pol && pol != &default_policy && !(pol->flags & MPOL_F_MORON)) {
2823 mode = pol->mode;
2824 flags = pol->flags;
2825 }
2826
2827 switch (mode) {
2828 case MPOL_DEFAULT:
2829 break;
2830 case MPOL_PREFERRED:
2831 if (flags & MPOL_F_LOCAL)
2832 mode = MPOL_LOCAL;
2833 else
2834 node_set(pol->v.preferred_node, nodes);
2835 break;
2836 case MPOL_BIND:
2837 case MPOL_INTERLEAVE:
2838 nodes = pol->v.nodes;
2839 break;
2840 default:
2841 WARN_ON_ONCE(1);
2842 snprintf(p, maxlen, "unknown");
2843 return;
2844 }
2845
2846 p += snprintf(p, maxlen, "%s", policy_modes[mode]);
2847
2848 if (flags & MPOL_MODE_FLAGS) {
2849 p += snprintf(p, buffer + maxlen - p, "=");
2850
2851 /*
2852 * Currently, the only defined flags are mutually exclusive
2853 */
2854 if (flags & MPOL_F_STATIC_NODES)
2855 p += snprintf(p, buffer + maxlen - p, "static");
2856 else if (flags & MPOL_F_RELATIVE_NODES)
2857 p += snprintf(p, buffer + maxlen - p, "relative");
2858 }
2859
2860 if (!nodes_empty(nodes))
2861 p += scnprintf(p, buffer + maxlen - p, ":%*pbl",
2862 nodemask_pr_args(&nodes));
2863}
1/*
2 * Simple NUMA memory policy for the Linux kernel.
3 *
4 * Copyright 2003,2004 Andi Kleen, SuSE Labs.
5 * (C) Copyright 2005 Christoph Lameter, Silicon Graphics, Inc.
6 * Subject to the GNU Public License, version 2.
7 *
8 * NUMA policy allows the user to give hints in which node(s) memory should
9 * be allocated.
10 *
11 * Support four policies per VMA and per process:
12 *
13 * The VMA policy has priority over the process policy for a page fault.
14 *
15 * interleave Allocate memory interleaved over a set of nodes,
16 * with normal fallback if it fails.
17 * For VMA based allocations this interleaves based on the
18 * offset into the backing object or offset into the mapping
19 * for anonymous memory. For process policy an process counter
20 * is used.
21 *
22 * bind Only allocate memory on a specific set of nodes,
23 * no fallback.
24 * FIXME: memory is allocated starting with the first node
25 * to the last. It would be better if bind would truly restrict
26 * the allocation to memory nodes instead
27 *
28 * preferred Try a specific node first before normal fallback.
29 * As a special case NUMA_NO_NODE here means do the allocation
30 * on the local CPU. This is normally identical to default,
31 * but useful to set in a VMA when you have a non default
32 * process policy.
33 *
34 * default Allocate on the local node first, or when on a VMA
35 * use the process policy. This is what Linux always did
36 * in a NUMA aware kernel and still does by, ahem, default.
37 *
38 * The process policy is applied for most non interrupt memory allocations
39 * in that process' context. Interrupts ignore the policies and always
40 * try to allocate on the local CPU. The VMA policy is only applied for memory
41 * allocations for a VMA in the VM.
42 *
43 * Currently there are a few corner cases in swapping where the policy
44 * is not applied, but the majority should be handled. When process policy
45 * is used it is not remembered over swap outs/swap ins.
46 *
47 * Only the highest zone in the zone hierarchy gets policied. Allocations
48 * requesting a lower zone just use default policy. This implies that
49 * on systems with highmem kernel lowmem allocation don't get policied.
50 * Same with GFP_DMA allocations.
51 *
52 * For shmfs/tmpfs/hugetlbfs shared memory the policy is shared between
53 * all users and remembered even when nobody has memory mapped.
54 */
55
56/* Notebook:
57 fix mmap readahead to honour policy and enable policy for any page cache
58 object
59 statistics for bigpages
60 global policy for page cache? currently it uses process policy. Requires
61 first item above.
62 handle mremap for shared memory (currently ignored for the policy)
63 grows down?
64 make bind policy root only? It can trigger oom much faster and the
65 kernel is not always grateful with that.
66*/
67
68#include <linux/mempolicy.h>
69#include <linux/mm.h>
70#include <linux/highmem.h>
71#include <linux/hugetlb.h>
72#include <linux/kernel.h>
73#include <linux/sched.h>
74#include <linux/nodemask.h>
75#include <linux/cpuset.h>
76#include <linux/slab.h>
77#include <linux/string.h>
78#include <linux/export.h>
79#include <linux/nsproxy.h>
80#include <linux/interrupt.h>
81#include <linux/init.h>
82#include <linux/compat.h>
83#include <linux/swap.h>
84#include <linux/seq_file.h>
85#include <linux/proc_fs.h>
86#include <linux/migrate.h>
87#include <linux/ksm.h>
88#include <linux/rmap.h>
89#include <linux/security.h>
90#include <linux/syscalls.h>
91#include <linux/ctype.h>
92#include <linux/mm_inline.h>
93#include <linux/mmu_notifier.h>
94
95#include <asm/tlbflush.h>
96#include <asm/uaccess.h>
97#include <linux/random.h>
98
99#include "internal.h"
100
101/* Internal flags */
102#define MPOL_MF_DISCONTIG_OK (MPOL_MF_INTERNAL << 0) /* Skip checks for continuous vmas */
103#define MPOL_MF_INVERT (MPOL_MF_INTERNAL << 1) /* Invert check for nodemask */
104
105static struct kmem_cache *policy_cache;
106static struct kmem_cache *sn_cache;
107
108/* Highest zone. An specific allocation for a zone below that is not
109 policied. */
110enum zone_type policy_zone = 0;
111
112/*
113 * run-time system-wide default policy => local allocation
114 */
115static struct mempolicy default_policy = {
116 .refcnt = ATOMIC_INIT(1), /* never free it */
117 .mode = MPOL_PREFERRED,
118 .flags = MPOL_F_LOCAL,
119};
120
121static struct mempolicy preferred_node_policy[MAX_NUMNODES];
122
123static struct mempolicy *get_task_policy(struct task_struct *p)
124{
125 struct mempolicy *pol = p->mempolicy;
126
127 if (!pol) {
128 int node = numa_node_id();
129
130 if (node != NUMA_NO_NODE) {
131 pol = &preferred_node_policy[node];
132 /*
133 * preferred_node_policy is not initialised early in
134 * boot
135 */
136 if (!pol->mode)
137 pol = NULL;
138 }
139 }
140
141 return pol;
142}
143
144static const struct mempolicy_operations {
145 int (*create)(struct mempolicy *pol, const nodemask_t *nodes);
146 /*
147 * If read-side task has no lock to protect task->mempolicy, write-side
148 * task will rebind the task->mempolicy by two step. The first step is
149 * setting all the newly nodes, and the second step is cleaning all the
150 * disallowed nodes. In this way, we can avoid finding no node to alloc
151 * page.
152 * If we have a lock to protect task->mempolicy in read-side, we do
153 * rebind directly.
154 *
155 * step:
156 * MPOL_REBIND_ONCE - do rebind work at once
157 * MPOL_REBIND_STEP1 - set all the newly nodes
158 * MPOL_REBIND_STEP2 - clean all the disallowed nodes
159 */
160 void (*rebind)(struct mempolicy *pol, const nodemask_t *nodes,
161 enum mpol_rebind_step step);
162} mpol_ops[MPOL_MAX];
163
164/* Check that the nodemask contains at least one populated zone */
165static int is_valid_nodemask(const nodemask_t *nodemask)
166{
167 return nodes_intersects(*nodemask, node_states[N_MEMORY]);
168}
169
170static inline int mpol_store_user_nodemask(const struct mempolicy *pol)
171{
172 return pol->flags & MPOL_MODE_FLAGS;
173}
174
175static void mpol_relative_nodemask(nodemask_t *ret, const nodemask_t *orig,
176 const nodemask_t *rel)
177{
178 nodemask_t tmp;
179 nodes_fold(tmp, *orig, nodes_weight(*rel));
180 nodes_onto(*ret, tmp, *rel);
181}
182
183static int mpol_new_interleave(struct mempolicy *pol, const nodemask_t *nodes)
184{
185 if (nodes_empty(*nodes))
186 return -EINVAL;
187 pol->v.nodes = *nodes;
188 return 0;
189}
190
191static int mpol_new_preferred(struct mempolicy *pol, const nodemask_t *nodes)
192{
193 if (!nodes)
194 pol->flags |= MPOL_F_LOCAL; /* local allocation */
195 else if (nodes_empty(*nodes))
196 return -EINVAL; /* no allowed nodes */
197 else
198 pol->v.preferred_node = first_node(*nodes);
199 return 0;
200}
201
202static int mpol_new_bind(struct mempolicy *pol, const nodemask_t *nodes)
203{
204 if (!is_valid_nodemask(nodes))
205 return -EINVAL;
206 pol->v.nodes = *nodes;
207 return 0;
208}
209
210/*
211 * mpol_set_nodemask is called after mpol_new() to set up the nodemask, if
212 * any, for the new policy. mpol_new() has already validated the nodes
213 * parameter with respect to the policy mode and flags. But, we need to
214 * handle an empty nodemask with MPOL_PREFERRED here.
215 *
216 * Must be called holding task's alloc_lock to protect task's mems_allowed
217 * and mempolicy. May also be called holding the mmap_semaphore for write.
218 */
219static int mpol_set_nodemask(struct mempolicy *pol,
220 const nodemask_t *nodes, struct nodemask_scratch *nsc)
221{
222 int ret;
223
224 /* if mode is MPOL_DEFAULT, pol is NULL. This is right. */
225 if (pol == NULL)
226 return 0;
227 /* Check N_MEMORY */
228 nodes_and(nsc->mask1,
229 cpuset_current_mems_allowed, node_states[N_MEMORY]);
230
231 VM_BUG_ON(!nodes);
232 if (pol->mode == MPOL_PREFERRED && nodes_empty(*nodes))
233 nodes = NULL; /* explicit local allocation */
234 else {
235 if (pol->flags & MPOL_F_RELATIVE_NODES)
236 mpol_relative_nodemask(&nsc->mask2, nodes,&nsc->mask1);
237 else
238 nodes_and(nsc->mask2, *nodes, nsc->mask1);
239
240 if (mpol_store_user_nodemask(pol))
241 pol->w.user_nodemask = *nodes;
242 else
243 pol->w.cpuset_mems_allowed =
244 cpuset_current_mems_allowed;
245 }
246
247 if (nodes)
248 ret = mpol_ops[pol->mode].create(pol, &nsc->mask2);
249 else
250 ret = mpol_ops[pol->mode].create(pol, NULL);
251 return ret;
252}
253
254/*
255 * This function just creates a new policy, does some check and simple
256 * initialization. You must invoke mpol_set_nodemask() to set nodes.
257 */
258static struct mempolicy *mpol_new(unsigned short mode, unsigned short flags,
259 nodemask_t *nodes)
260{
261 struct mempolicy *policy;
262
263 pr_debug("setting mode %d flags %d nodes[0] %lx\n",
264 mode, flags, nodes ? nodes_addr(*nodes)[0] : NUMA_NO_NODE);
265
266 if (mode == MPOL_DEFAULT) {
267 if (nodes && !nodes_empty(*nodes))
268 return ERR_PTR(-EINVAL);
269 return NULL;
270 }
271 VM_BUG_ON(!nodes);
272
273 /*
274 * MPOL_PREFERRED cannot be used with MPOL_F_STATIC_NODES or
275 * MPOL_F_RELATIVE_NODES if the nodemask is empty (local allocation).
276 * All other modes require a valid pointer to a non-empty nodemask.
277 */
278 if (mode == MPOL_PREFERRED) {
279 if (nodes_empty(*nodes)) {
280 if (((flags & MPOL_F_STATIC_NODES) ||
281 (flags & MPOL_F_RELATIVE_NODES)))
282 return ERR_PTR(-EINVAL);
283 }
284 } else if (mode == MPOL_LOCAL) {
285 if (!nodes_empty(*nodes))
286 return ERR_PTR(-EINVAL);
287 mode = MPOL_PREFERRED;
288 } else if (nodes_empty(*nodes))
289 return ERR_PTR(-EINVAL);
290 policy = kmem_cache_alloc(policy_cache, GFP_KERNEL);
291 if (!policy)
292 return ERR_PTR(-ENOMEM);
293 atomic_set(&policy->refcnt, 1);
294 policy->mode = mode;
295 policy->flags = flags;
296
297 return policy;
298}
299
300/* Slow path of a mpol destructor. */
301void __mpol_put(struct mempolicy *p)
302{
303 if (!atomic_dec_and_test(&p->refcnt))
304 return;
305 kmem_cache_free(policy_cache, p);
306}
307
308static void mpol_rebind_default(struct mempolicy *pol, const nodemask_t *nodes,
309 enum mpol_rebind_step step)
310{
311}
312
313/*
314 * step:
315 * MPOL_REBIND_ONCE - do rebind work at once
316 * MPOL_REBIND_STEP1 - set all the newly nodes
317 * MPOL_REBIND_STEP2 - clean all the disallowed nodes
318 */
319static void mpol_rebind_nodemask(struct mempolicy *pol, const nodemask_t *nodes,
320 enum mpol_rebind_step step)
321{
322 nodemask_t tmp;
323
324 if (pol->flags & MPOL_F_STATIC_NODES)
325 nodes_and(tmp, pol->w.user_nodemask, *nodes);
326 else if (pol->flags & MPOL_F_RELATIVE_NODES)
327 mpol_relative_nodemask(&tmp, &pol->w.user_nodemask, nodes);
328 else {
329 /*
330 * if step == 1, we use ->w.cpuset_mems_allowed to cache the
331 * result
332 */
333 if (step == MPOL_REBIND_ONCE || step == MPOL_REBIND_STEP1) {
334 nodes_remap(tmp, pol->v.nodes,
335 pol->w.cpuset_mems_allowed, *nodes);
336 pol->w.cpuset_mems_allowed = step ? tmp : *nodes;
337 } else if (step == MPOL_REBIND_STEP2) {
338 tmp = pol->w.cpuset_mems_allowed;
339 pol->w.cpuset_mems_allowed = *nodes;
340 } else
341 BUG();
342 }
343
344 if (nodes_empty(tmp))
345 tmp = *nodes;
346
347 if (step == MPOL_REBIND_STEP1)
348 nodes_or(pol->v.nodes, pol->v.nodes, tmp);
349 else if (step == MPOL_REBIND_ONCE || step == MPOL_REBIND_STEP2)
350 pol->v.nodes = tmp;
351 else
352 BUG();
353
354 if (!node_isset(current->il_next, tmp)) {
355 current->il_next = next_node(current->il_next, tmp);
356 if (current->il_next >= MAX_NUMNODES)
357 current->il_next = first_node(tmp);
358 if (current->il_next >= MAX_NUMNODES)
359 current->il_next = numa_node_id();
360 }
361}
362
363static void mpol_rebind_preferred(struct mempolicy *pol,
364 const nodemask_t *nodes,
365 enum mpol_rebind_step step)
366{
367 nodemask_t tmp;
368
369 if (pol->flags & MPOL_F_STATIC_NODES) {
370 int node = first_node(pol->w.user_nodemask);
371
372 if (node_isset(node, *nodes)) {
373 pol->v.preferred_node = node;
374 pol->flags &= ~MPOL_F_LOCAL;
375 } else
376 pol->flags |= MPOL_F_LOCAL;
377 } else if (pol->flags & MPOL_F_RELATIVE_NODES) {
378 mpol_relative_nodemask(&tmp, &pol->w.user_nodemask, nodes);
379 pol->v.preferred_node = first_node(tmp);
380 } else if (!(pol->flags & MPOL_F_LOCAL)) {
381 pol->v.preferred_node = node_remap(pol->v.preferred_node,
382 pol->w.cpuset_mems_allowed,
383 *nodes);
384 pol->w.cpuset_mems_allowed = *nodes;
385 }
386}
387
388/*
389 * mpol_rebind_policy - Migrate a policy to a different set of nodes
390 *
391 * If read-side task has no lock to protect task->mempolicy, write-side
392 * task will rebind the task->mempolicy by two step. The first step is
393 * setting all the newly nodes, and the second step is cleaning all the
394 * disallowed nodes. In this way, we can avoid finding no node to alloc
395 * page.
396 * If we have a lock to protect task->mempolicy in read-side, we do
397 * rebind directly.
398 *
399 * step:
400 * MPOL_REBIND_ONCE - do rebind work at once
401 * MPOL_REBIND_STEP1 - set all the newly nodes
402 * MPOL_REBIND_STEP2 - clean all the disallowed nodes
403 */
404static void mpol_rebind_policy(struct mempolicy *pol, const nodemask_t *newmask,
405 enum mpol_rebind_step step)
406{
407 if (!pol)
408 return;
409 if (!mpol_store_user_nodemask(pol) && step == MPOL_REBIND_ONCE &&
410 nodes_equal(pol->w.cpuset_mems_allowed, *newmask))
411 return;
412
413 if (step == MPOL_REBIND_STEP1 && (pol->flags & MPOL_F_REBINDING))
414 return;
415
416 if (step == MPOL_REBIND_STEP2 && !(pol->flags & MPOL_F_REBINDING))
417 BUG();
418
419 if (step == MPOL_REBIND_STEP1)
420 pol->flags |= MPOL_F_REBINDING;
421 else if (step == MPOL_REBIND_STEP2)
422 pol->flags &= ~MPOL_F_REBINDING;
423 else if (step >= MPOL_REBIND_NSTEP)
424 BUG();
425
426 mpol_ops[pol->mode].rebind(pol, newmask, step);
427}
428
429/*
430 * Wrapper for mpol_rebind_policy() that just requires task
431 * pointer, and updates task mempolicy.
432 *
433 * Called with task's alloc_lock held.
434 */
435
436void mpol_rebind_task(struct task_struct *tsk, const nodemask_t *new,
437 enum mpol_rebind_step step)
438{
439 mpol_rebind_policy(tsk->mempolicy, new, step);
440}
441
442/*
443 * Rebind each vma in mm to new nodemask.
444 *
445 * Call holding a reference to mm. Takes mm->mmap_sem during call.
446 */
447
448void mpol_rebind_mm(struct mm_struct *mm, nodemask_t *new)
449{
450 struct vm_area_struct *vma;
451
452 down_write(&mm->mmap_sem);
453 for (vma = mm->mmap; vma; vma = vma->vm_next)
454 mpol_rebind_policy(vma->vm_policy, new, MPOL_REBIND_ONCE);
455 up_write(&mm->mmap_sem);
456}
457
458static const struct mempolicy_operations mpol_ops[MPOL_MAX] = {
459 [MPOL_DEFAULT] = {
460 .rebind = mpol_rebind_default,
461 },
462 [MPOL_INTERLEAVE] = {
463 .create = mpol_new_interleave,
464 .rebind = mpol_rebind_nodemask,
465 },
466 [MPOL_PREFERRED] = {
467 .create = mpol_new_preferred,
468 .rebind = mpol_rebind_preferred,
469 },
470 [MPOL_BIND] = {
471 .create = mpol_new_bind,
472 .rebind = mpol_rebind_nodemask,
473 },
474};
475
476static void migrate_page_add(struct page *page, struct list_head *pagelist,
477 unsigned long flags);
478
479/*
480 * Scan through pages checking if pages follow certain conditions,
481 * and move them to the pagelist if they do.
482 */
483static int queue_pages_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
484 unsigned long addr, unsigned long end,
485 const nodemask_t *nodes, unsigned long flags,
486 void *private)
487{
488 pte_t *orig_pte;
489 pte_t *pte;
490 spinlock_t *ptl;
491
492 orig_pte = pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
493 do {
494 struct page *page;
495 int nid;
496
497 if (!pte_present(*pte))
498 continue;
499 page = vm_normal_page(vma, addr, *pte);
500 if (!page)
501 continue;
502 /*
503 * vm_normal_page() filters out zero pages, but there might
504 * still be PageReserved pages to skip, perhaps in a VDSO.
505 */
506 if (PageReserved(page))
507 continue;
508 nid = page_to_nid(page);
509 if (node_isset(nid, *nodes) == !!(flags & MPOL_MF_INVERT))
510 continue;
511
512 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))
513 migrate_page_add(page, private, flags);
514 else
515 break;
516 } while (pte++, addr += PAGE_SIZE, addr != end);
517 pte_unmap_unlock(orig_pte, ptl);
518 return addr != end;
519}
520
521static void queue_pages_hugetlb_pmd_range(struct vm_area_struct *vma,
522 pmd_t *pmd, const nodemask_t *nodes, unsigned long flags,
523 void *private)
524{
525#ifdef CONFIG_HUGETLB_PAGE
526 int nid;
527 struct page *page;
528 spinlock_t *ptl;
529 pte_t entry;
530
531 ptl = huge_pte_lock(hstate_vma(vma), vma->vm_mm, (pte_t *)pmd);
532 entry = huge_ptep_get((pte_t *)pmd);
533 if (!pte_present(entry))
534 goto unlock;
535 page = pte_page(entry);
536 nid = page_to_nid(page);
537 if (node_isset(nid, *nodes) == !!(flags & MPOL_MF_INVERT))
538 goto unlock;
539 /* With MPOL_MF_MOVE, we migrate only unshared hugepage. */
540 if (flags & (MPOL_MF_MOVE_ALL) ||
541 (flags & MPOL_MF_MOVE && page_mapcount(page) == 1))
542 isolate_huge_page(page, private);
543unlock:
544 spin_unlock(ptl);
545#else
546 BUG();
547#endif
548}
549
550static inline int queue_pages_pmd_range(struct vm_area_struct *vma, pud_t *pud,
551 unsigned long addr, unsigned long end,
552 const nodemask_t *nodes, unsigned long flags,
553 void *private)
554{
555 pmd_t *pmd;
556 unsigned long next;
557
558 pmd = pmd_offset(pud, addr);
559 do {
560 next = pmd_addr_end(addr, end);
561 if (!pmd_present(*pmd))
562 continue;
563 if (pmd_huge(*pmd) && is_vm_hugetlb_page(vma)) {
564 queue_pages_hugetlb_pmd_range(vma, pmd, nodes,
565 flags, private);
566 continue;
567 }
568 split_huge_page_pmd(vma, addr, pmd);
569 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
570 continue;
571 if (queue_pages_pte_range(vma, pmd, addr, next, nodes,
572 flags, private))
573 return -EIO;
574 } while (pmd++, addr = next, addr != end);
575 return 0;
576}
577
578static inline int queue_pages_pud_range(struct vm_area_struct *vma, pgd_t *pgd,
579 unsigned long addr, unsigned long end,
580 const nodemask_t *nodes, unsigned long flags,
581 void *private)
582{
583 pud_t *pud;
584 unsigned long next;
585
586 pud = pud_offset(pgd, addr);
587 do {
588 next = pud_addr_end(addr, end);
589 if (pud_huge(*pud) && is_vm_hugetlb_page(vma))
590 continue;
591 if (pud_none_or_clear_bad(pud))
592 continue;
593 if (queue_pages_pmd_range(vma, pud, addr, next, nodes,
594 flags, private))
595 return -EIO;
596 } while (pud++, addr = next, addr != end);
597 return 0;
598}
599
600static inline int queue_pages_pgd_range(struct vm_area_struct *vma,
601 unsigned long addr, unsigned long end,
602 const nodemask_t *nodes, unsigned long flags,
603 void *private)
604{
605 pgd_t *pgd;
606 unsigned long next;
607
608 pgd = pgd_offset(vma->vm_mm, addr);
609 do {
610 next = pgd_addr_end(addr, end);
611 if (pgd_none_or_clear_bad(pgd))
612 continue;
613 if (queue_pages_pud_range(vma, pgd, addr, next, nodes,
614 flags, private))
615 return -EIO;
616 } while (pgd++, addr = next, addr != end);
617 return 0;
618}
619
620#ifdef CONFIG_NUMA_BALANCING
621/*
622 * This is used to mark a range of virtual addresses to be inaccessible.
623 * These are later cleared by a NUMA hinting fault. Depending on these
624 * faults, pages may be migrated for better NUMA placement.
625 *
626 * This is assuming that NUMA faults are handled using PROT_NONE. If
627 * an architecture makes a different choice, it will need further
628 * changes to the core.
629 */
630unsigned long change_prot_numa(struct vm_area_struct *vma,
631 unsigned long addr, unsigned long end)
632{
633 int nr_updated;
634
635 nr_updated = change_protection(vma, addr, end, vma->vm_page_prot, 0, 1);
636 if (nr_updated)
637 count_vm_numa_events(NUMA_PTE_UPDATES, nr_updated);
638
639 return nr_updated;
640}
641#else
642static unsigned long change_prot_numa(struct vm_area_struct *vma,
643 unsigned long addr, unsigned long end)
644{
645 return 0;
646}
647#endif /* CONFIG_NUMA_BALANCING */
648
649/*
650 * Walk through page tables and collect pages to be migrated.
651 *
652 * If pages found in a given range are on a set of nodes (determined by
653 * @nodes and @flags,) it's isolated and queued to the pagelist which is
654 * passed via @private.)
655 */
656static struct vm_area_struct *
657queue_pages_range(struct mm_struct *mm, unsigned long start, unsigned long end,
658 const nodemask_t *nodes, unsigned long flags, void *private)
659{
660 int err;
661 struct vm_area_struct *first, *vma, *prev;
662
663
664 first = find_vma(mm, start);
665 if (!first)
666 return ERR_PTR(-EFAULT);
667 prev = NULL;
668 for (vma = first; vma && vma->vm_start < end; vma = vma->vm_next) {
669 unsigned long endvma = vma->vm_end;
670
671 if (endvma > end)
672 endvma = end;
673 if (vma->vm_start > start)
674 start = vma->vm_start;
675
676 if (!(flags & MPOL_MF_DISCONTIG_OK)) {
677 if (!vma->vm_next && vma->vm_end < end)
678 return ERR_PTR(-EFAULT);
679 if (prev && prev->vm_end < vma->vm_start)
680 return ERR_PTR(-EFAULT);
681 }
682
683 if (flags & MPOL_MF_LAZY) {
684 change_prot_numa(vma, start, endvma);
685 goto next;
686 }
687
688 if ((flags & MPOL_MF_STRICT) ||
689 ((flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) &&
690 vma_migratable(vma))) {
691
692 err = queue_pages_pgd_range(vma, start, endvma, nodes,
693 flags, private);
694 if (err) {
695 first = ERR_PTR(err);
696 break;
697 }
698 }
699next:
700 prev = vma;
701 }
702 return first;
703}
704
705/*
706 * Apply policy to a single VMA
707 * This must be called with the mmap_sem held for writing.
708 */
709static int vma_replace_policy(struct vm_area_struct *vma,
710 struct mempolicy *pol)
711{
712 int err;
713 struct mempolicy *old;
714 struct mempolicy *new;
715
716 pr_debug("vma %lx-%lx/%lx vm_ops %p vm_file %p set_policy %p\n",
717 vma->vm_start, vma->vm_end, vma->vm_pgoff,
718 vma->vm_ops, vma->vm_file,
719 vma->vm_ops ? vma->vm_ops->set_policy : NULL);
720
721 new = mpol_dup(pol);
722 if (IS_ERR(new))
723 return PTR_ERR(new);
724
725 if (vma->vm_ops && vma->vm_ops->set_policy) {
726 err = vma->vm_ops->set_policy(vma, new);
727 if (err)
728 goto err_out;
729 }
730
731 old = vma->vm_policy;
732 vma->vm_policy = new; /* protected by mmap_sem */
733 mpol_put(old);
734
735 return 0;
736 err_out:
737 mpol_put(new);
738 return err;
739}
740
741/* Step 2: apply policy to a range and do splits. */
742static int mbind_range(struct mm_struct *mm, unsigned long start,
743 unsigned long end, struct mempolicy *new_pol)
744{
745 struct vm_area_struct *next;
746 struct vm_area_struct *prev;
747 struct vm_area_struct *vma;
748 int err = 0;
749 pgoff_t pgoff;
750 unsigned long vmstart;
751 unsigned long vmend;
752
753 vma = find_vma(mm, start);
754 if (!vma || vma->vm_start > start)
755 return -EFAULT;
756
757 prev = vma->vm_prev;
758 if (start > vma->vm_start)
759 prev = vma;
760
761 for (; vma && vma->vm_start < end; prev = vma, vma = next) {
762 next = vma->vm_next;
763 vmstart = max(start, vma->vm_start);
764 vmend = min(end, vma->vm_end);
765
766 if (mpol_equal(vma_policy(vma), new_pol))
767 continue;
768
769 pgoff = vma->vm_pgoff +
770 ((vmstart - vma->vm_start) >> PAGE_SHIFT);
771 prev = vma_merge(mm, prev, vmstart, vmend, vma->vm_flags,
772 vma->anon_vma, vma->vm_file, pgoff,
773 new_pol);
774 if (prev) {
775 vma = prev;
776 next = vma->vm_next;
777 if (mpol_equal(vma_policy(vma), new_pol))
778 continue;
779 /* vma_merge() joined vma && vma->next, case 8 */
780 goto replace;
781 }
782 if (vma->vm_start != vmstart) {
783 err = split_vma(vma->vm_mm, vma, vmstart, 1);
784 if (err)
785 goto out;
786 }
787 if (vma->vm_end != vmend) {
788 err = split_vma(vma->vm_mm, vma, vmend, 0);
789 if (err)
790 goto out;
791 }
792 replace:
793 err = vma_replace_policy(vma, new_pol);
794 if (err)
795 goto out;
796 }
797
798 out:
799 return err;
800}
801
802/* Set the process memory policy */
803static long do_set_mempolicy(unsigned short mode, unsigned short flags,
804 nodemask_t *nodes)
805{
806 struct mempolicy *new, *old;
807 struct mm_struct *mm = current->mm;
808 NODEMASK_SCRATCH(scratch);
809 int ret;
810
811 if (!scratch)
812 return -ENOMEM;
813
814 new = mpol_new(mode, flags, nodes);
815 if (IS_ERR(new)) {
816 ret = PTR_ERR(new);
817 goto out;
818 }
819 /*
820 * prevent changing our mempolicy while show_numa_maps()
821 * is using it.
822 * Note: do_set_mempolicy() can be called at init time
823 * with no 'mm'.
824 */
825 if (mm)
826 down_write(&mm->mmap_sem);
827 task_lock(current);
828 ret = mpol_set_nodemask(new, nodes, scratch);
829 if (ret) {
830 task_unlock(current);
831 if (mm)
832 up_write(&mm->mmap_sem);
833 mpol_put(new);
834 goto out;
835 }
836 old = current->mempolicy;
837 current->mempolicy = new;
838 if (new && new->mode == MPOL_INTERLEAVE &&
839 nodes_weight(new->v.nodes))
840 current->il_next = first_node(new->v.nodes);
841 task_unlock(current);
842 if (mm)
843 up_write(&mm->mmap_sem);
844
845 mpol_put(old);
846 ret = 0;
847out:
848 NODEMASK_SCRATCH_FREE(scratch);
849 return ret;
850}
851
852/*
853 * Return nodemask for policy for get_mempolicy() query
854 *
855 * Called with task's alloc_lock held
856 */
857static void get_policy_nodemask(struct mempolicy *p, nodemask_t *nodes)
858{
859 nodes_clear(*nodes);
860 if (p == &default_policy)
861 return;
862
863 switch (p->mode) {
864 case MPOL_BIND:
865 /* Fall through */
866 case MPOL_INTERLEAVE:
867 *nodes = p->v.nodes;
868 break;
869 case MPOL_PREFERRED:
870 if (!(p->flags & MPOL_F_LOCAL))
871 node_set(p->v.preferred_node, *nodes);
872 /* else return empty node mask for local allocation */
873 break;
874 default:
875 BUG();
876 }
877}
878
879static int lookup_node(struct mm_struct *mm, unsigned long addr)
880{
881 struct page *p;
882 int err;
883
884 err = get_user_pages(current, mm, addr & PAGE_MASK, 1, 0, 0, &p, NULL);
885 if (err >= 0) {
886 err = page_to_nid(p);
887 put_page(p);
888 }
889 return err;
890}
891
892/* Retrieve NUMA policy */
893static long do_get_mempolicy(int *policy, nodemask_t *nmask,
894 unsigned long addr, unsigned long flags)
895{
896 int err;
897 struct mm_struct *mm = current->mm;
898 struct vm_area_struct *vma = NULL;
899 struct mempolicy *pol = current->mempolicy;
900
901 if (flags &
902 ~(unsigned long)(MPOL_F_NODE|MPOL_F_ADDR|MPOL_F_MEMS_ALLOWED))
903 return -EINVAL;
904
905 if (flags & MPOL_F_MEMS_ALLOWED) {
906 if (flags & (MPOL_F_NODE|MPOL_F_ADDR))
907 return -EINVAL;
908 *policy = 0; /* just so it's initialized */
909 task_lock(current);
910 *nmask = cpuset_current_mems_allowed;
911 task_unlock(current);
912 return 0;
913 }
914
915 if (flags & MPOL_F_ADDR) {
916 /*
917 * Do NOT fall back to task policy if the
918 * vma/shared policy at addr is NULL. We
919 * want to return MPOL_DEFAULT in this case.
920 */
921 down_read(&mm->mmap_sem);
922 vma = find_vma_intersection(mm, addr, addr+1);
923 if (!vma) {
924 up_read(&mm->mmap_sem);
925 return -EFAULT;
926 }
927 if (vma->vm_ops && vma->vm_ops->get_policy)
928 pol = vma->vm_ops->get_policy(vma, addr);
929 else
930 pol = vma->vm_policy;
931 } else if (addr)
932 return -EINVAL;
933
934 if (!pol)
935 pol = &default_policy; /* indicates default behavior */
936
937 if (flags & MPOL_F_NODE) {
938 if (flags & MPOL_F_ADDR) {
939 err = lookup_node(mm, addr);
940 if (err < 0)
941 goto out;
942 *policy = err;
943 } else if (pol == current->mempolicy &&
944 pol->mode == MPOL_INTERLEAVE) {
945 *policy = current->il_next;
946 } else {
947 err = -EINVAL;
948 goto out;
949 }
950 } else {
951 *policy = pol == &default_policy ? MPOL_DEFAULT :
952 pol->mode;
953 /*
954 * Internal mempolicy flags must be masked off before exposing
955 * the policy to userspace.
956 */
957 *policy |= (pol->flags & MPOL_MODE_FLAGS);
958 }
959
960 if (vma) {
961 up_read(¤t->mm->mmap_sem);
962 vma = NULL;
963 }
964
965 err = 0;
966 if (nmask) {
967 if (mpol_store_user_nodemask(pol)) {
968 *nmask = pol->w.user_nodemask;
969 } else {
970 task_lock(current);
971 get_policy_nodemask(pol, nmask);
972 task_unlock(current);
973 }
974 }
975
976 out:
977 mpol_cond_put(pol);
978 if (vma)
979 up_read(¤t->mm->mmap_sem);
980 return err;
981}
982
983#ifdef CONFIG_MIGRATION
984/*
985 * page migration
986 */
987static void migrate_page_add(struct page *page, struct list_head *pagelist,
988 unsigned long flags)
989{
990 /*
991 * Avoid migrating a page that is shared with others.
992 */
993 if ((flags & MPOL_MF_MOVE_ALL) || page_mapcount(page) == 1) {
994 if (!isolate_lru_page(page)) {
995 list_add_tail(&page->lru, pagelist);
996 inc_zone_page_state(page, NR_ISOLATED_ANON +
997 page_is_file_cache(page));
998 }
999 }
1000}
1001
1002static struct page *new_node_page(struct page *page, unsigned long node, int **x)
1003{
1004 if (PageHuge(page))
1005 return alloc_huge_page_node(page_hstate(compound_head(page)),
1006 node);
1007 else
1008 return alloc_pages_exact_node(node, GFP_HIGHUSER_MOVABLE, 0);
1009}
1010
1011/*
1012 * Migrate pages from one node to a target node.
1013 * Returns error or the number of pages not migrated.
1014 */
1015static int migrate_to_node(struct mm_struct *mm, int source, int dest,
1016 int flags)
1017{
1018 nodemask_t nmask;
1019 LIST_HEAD(pagelist);
1020 int err = 0;
1021
1022 nodes_clear(nmask);
1023 node_set(source, nmask);
1024
1025 /*
1026 * This does not "check" the range but isolates all pages that
1027 * need migration. Between passing in the full user address
1028 * space range and MPOL_MF_DISCONTIG_OK, this call can not fail.
1029 */
1030 VM_BUG_ON(!(flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)));
1031 queue_pages_range(mm, mm->mmap->vm_start, mm->task_size, &nmask,
1032 flags | MPOL_MF_DISCONTIG_OK, &pagelist);
1033
1034 if (!list_empty(&pagelist)) {
1035 err = migrate_pages(&pagelist, new_node_page, dest,
1036 MIGRATE_SYNC, MR_SYSCALL);
1037 if (err)
1038 putback_movable_pages(&pagelist);
1039 }
1040
1041 return err;
1042}
1043
1044/*
1045 * Move pages between the two nodesets so as to preserve the physical
1046 * layout as much as possible.
1047 *
1048 * Returns the number of page that could not be moved.
1049 */
1050int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
1051 const nodemask_t *to, int flags)
1052{
1053 int busy = 0;
1054 int err;
1055 nodemask_t tmp;
1056
1057 err = migrate_prep();
1058 if (err)
1059 return err;
1060
1061 down_read(&mm->mmap_sem);
1062
1063 err = migrate_vmas(mm, from, to, flags);
1064 if (err)
1065 goto out;
1066
1067 /*
1068 * Find a 'source' bit set in 'tmp' whose corresponding 'dest'
1069 * bit in 'to' is not also set in 'tmp'. Clear the found 'source'
1070 * bit in 'tmp', and return that <source, dest> pair for migration.
1071 * The pair of nodemasks 'to' and 'from' define the map.
1072 *
1073 * If no pair of bits is found that way, fallback to picking some
1074 * pair of 'source' and 'dest' bits that are not the same. If the
1075 * 'source' and 'dest' bits are the same, this represents a node
1076 * that will be migrating to itself, so no pages need move.
1077 *
1078 * If no bits are left in 'tmp', or if all remaining bits left
1079 * in 'tmp' correspond to the same bit in 'to', return false
1080 * (nothing left to migrate).
1081 *
1082 * This lets us pick a pair of nodes to migrate between, such that
1083 * if possible the dest node is not already occupied by some other
1084 * source node, minimizing the risk of overloading the memory on a
1085 * node that would happen if we migrated incoming memory to a node
1086 * before migrating outgoing memory source that same node.
1087 *
1088 * A single scan of tmp is sufficient. As we go, we remember the
1089 * most recent <s, d> pair that moved (s != d). If we find a pair
1090 * that not only moved, but what's better, moved to an empty slot
1091 * (d is not set in tmp), then we break out then, with that pair.
1092 * Otherwise when we finish scanning from_tmp, we at least have the
1093 * most recent <s, d> pair that moved. If we get all the way through
1094 * the scan of tmp without finding any node that moved, much less
1095 * moved to an empty node, then there is nothing left worth migrating.
1096 */
1097
1098 tmp = *from;
1099 while (!nodes_empty(tmp)) {
1100 int s,d;
1101 int source = NUMA_NO_NODE;
1102 int dest = 0;
1103
1104 for_each_node_mask(s, tmp) {
1105
1106 /*
1107 * do_migrate_pages() tries to maintain the relative
1108 * node relationship of the pages established between
1109 * threads and memory areas.
1110 *
1111 * However if the number of source nodes is not equal to
1112 * the number of destination nodes we can not preserve
1113 * this node relative relationship. In that case, skip
1114 * copying memory from a node that is in the destination
1115 * mask.
1116 *
1117 * Example: [2,3,4] -> [3,4,5] moves everything.
1118 * [0-7] - > [3,4,5] moves only 0,1,2,6,7.
1119 */
1120
1121 if ((nodes_weight(*from) != nodes_weight(*to)) &&
1122 (node_isset(s, *to)))
1123 continue;
1124
1125 d = node_remap(s, *from, *to);
1126 if (s == d)
1127 continue;
1128
1129 source = s; /* Node moved. Memorize */
1130 dest = d;
1131
1132 /* dest not in remaining from nodes? */
1133 if (!node_isset(dest, tmp))
1134 break;
1135 }
1136 if (source == NUMA_NO_NODE)
1137 break;
1138
1139 node_clear(source, tmp);
1140 err = migrate_to_node(mm, source, dest, flags);
1141 if (err > 0)
1142 busy += err;
1143 if (err < 0)
1144 break;
1145 }
1146out:
1147 up_read(&mm->mmap_sem);
1148 if (err < 0)
1149 return err;
1150 return busy;
1151
1152}
1153
1154/*
1155 * Allocate a new page for page migration based on vma policy.
1156 * Start assuming that page is mapped by vma pointed to by @private.
1157 * Search forward from there, if not. N.B., this assumes that the
1158 * list of pages handed to migrate_pages()--which is how we get here--
1159 * is in virtual address order.
1160 */
1161static struct page *new_vma_page(struct page *page, unsigned long private, int **x)
1162{
1163 struct vm_area_struct *vma = (struct vm_area_struct *)private;
1164 unsigned long uninitialized_var(address);
1165
1166 while (vma) {
1167 address = page_address_in_vma(page, vma);
1168 if (address != -EFAULT)
1169 break;
1170 vma = vma->vm_next;
1171 }
1172
1173 if (PageHuge(page)) {
1174 BUG_ON(!vma);
1175 return alloc_huge_page_noerr(vma, address, 1);
1176 }
1177 /*
1178 * if !vma, alloc_page_vma() will use task or system default policy
1179 */
1180 return alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
1181}
1182#else
1183
1184static void migrate_page_add(struct page *page, struct list_head *pagelist,
1185 unsigned long flags)
1186{
1187}
1188
1189int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
1190 const nodemask_t *to, int flags)
1191{
1192 return -ENOSYS;
1193}
1194
1195static struct page *new_vma_page(struct page *page, unsigned long private, int **x)
1196{
1197 return NULL;
1198}
1199#endif
1200
1201static long do_mbind(unsigned long start, unsigned long len,
1202 unsigned short mode, unsigned short mode_flags,
1203 nodemask_t *nmask, unsigned long flags)
1204{
1205 struct vm_area_struct *vma;
1206 struct mm_struct *mm = current->mm;
1207 struct mempolicy *new;
1208 unsigned long end;
1209 int err;
1210 LIST_HEAD(pagelist);
1211
1212 if (flags & ~(unsigned long)MPOL_MF_VALID)
1213 return -EINVAL;
1214 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1215 return -EPERM;
1216
1217 if (start & ~PAGE_MASK)
1218 return -EINVAL;
1219
1220 if (mode == MPOL_DEFAULT)
1221 flags &= ~MPOL_MF_STRICT;
1222
1223 len = (len + PAGE_SIZE - 1) & PAGE_MASK;
1224 end = start + len;
1225
1226 if (end < start)
1227 return -EINVAL;
1228 if (end == start)
1229 return 0;
1230
1231 new = mpol_new(mode, mode_flags, nmask);
1232 if (IS_ERR(new))
1233 return PTR_ERR(new);
1234
1235 if (flags & MPOL_MF_LAZY)
1236 new->flags |= MPOL_F_MOF;
1237
1238 /*
1239 * If we are using the default policy then operation
1240 * on discontinuous address spaces is okay after all
1241 */
1242 if (!new)
1243 flags |= MPOL_MF_DISCONTIG_OK;
1244
1245 pr_debug("mbind %lx-%lx mode:%d flags:%d nodes:%lx\n",
1246 start, start + len, mode, mode_flags,
1247 nmask ? nodes_addr(*nmask)[0] : NUMA_NO_NODE);
1248
1249 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) {
1250
1251 err = migrate_prep();
1252 if (err)
1253 goto mpol_out;
1254 }
1255 {
1256 NODEMASK_SCRATCH(scratch);
1257 if (scratch) {
1258 down_write(&mm->mmap_sem);
1259 task_lock(current);
1260 err = mpol_set_nodemask(new, nmask, scratch);
1261 task_unlock(current);
1262 if (err)
1263 up_write(&mm->mmap_sem);
1264 } else
1265 err = -ENOMEM;
1266 NODEMASK_SCRATCH_FREE(scratch);
1267 }
1268 if (err)
1269 goto mpol_out;
1270
1271 vma = queue_pages_range(mm, start, end, nmask,
1272 flags | MPOL_MF_INVERT, &pagelist);
1273
1274 err = PTR_ERR(vma); /* maybe ... */
1275 if (!IS_ERR(vma))
1276 err = mbind_range(mm, start, end, new);
1277
1278 if (!err) {
1279 int nr_failed = 0;
1280
1281 if (!list_empty(&pagelist)) {
1282 WARN_ON_ONCE(flags & MPOL_MF_LAZY);
1283 nr_failed = migrate_pages(&pagelist, new_vma_page,
1284 (unsigned long)vma,
1285 MIGRATE_SYNC, MR_MEMPOLICY_MBIND);
1286 if (nr_failed)
1287 putback_movable_pages(&pagelist);
1288 }
1289
1290 if (nr_failed && (flags & MPOL_MF_STRICT))
1291 err = -EIO;
1292 } else
1293 putback_movable_pages(&pagelist);
1294
1295 up_write(&mm->mmap_sem);
1296 mpol_out:
1297 mpol_put(new);
1298 return err;
1299}
1300
1301/*
1302 * User space interface with variable sized bitmaps for nodelists.
1303 */
1304
1305/* Copy a node mask from user space. */
1306static int get_nodes(nodemask_t *nodes, const unsigned long __user *nmask,
1307 unsigned long maxnode)
1308{
1309 unsigned long k;
1310 unsigned long nlongs;
1311 unsigned long endmask;
1312
1313 --maxnode;
1314 nodes_clear(*nodes);
1315 if (maxnode == 0 || !nmask)
1316 return 0;
1317 if (maxnode > PAGE_SIZE*BITS_PER_BYTE)
1318 return -EINVAL;
1319
1320 nlongs = BITS_TO_LONGS(maxnode);
1321 if ((maxnode % BITS_PER_LONG) == 0)
1322 endmask = ~0UL;
1323 else
1324 endmask = (1UL << (maxnode % BITS_PER_LONG)) - 1;
1325
1326 /* When the user specified more nodes than supported just check
1327 if the non supported part is all zero. */
1328 if (nlongs > BITS_TO_LONGS(MAX_NUMNODES)) {
1329 if (nlongs > PAGE_SIZE/sizeof(long))
1330 return -EINVAL;
1331 for (k = BITS_TO_LONGS(MAX_NUMNODES); k < nlongs; k++) {
1332 unsigned long t;
1333 if (get_user(t, nmask + k))
1334 return -EFAULT;
1335 if (k == nlongs - 1) {
1336 if (t & endmask)
1337 return -EINVAL;
1338 } else if (t)
1339 return -EINVAL;
1340 }
1341 nlongs = BITS_TO_LONGS(MAX_NUMNODES);
1342 endmask = ~0UL;
1343 }
1344
1345 if (copy_from_user(nodes_addr(*nodes), nmask, nlongs*sizeof(unsigned long)))
1346 return -EFAULT;
1347 nodes_addr(*nodes)[nlongs-1] &= endmask;
1348 return 0;
1349}
1350
1351/* Copy a kernel node mask to user space */
1352static int copy_nodes_to_user(unsigned long __user *mask, unsigned long maxnode,
1353 nodemask_t *nodes)
1354{
1355 unsigned long copy = ALIGN(maxnode-1, 64) / 8;
1356 const int nbytes = BITS_TO_LONGS(MAX_NUMNODES) * sizeof(long);
1357
1358 if (copy > nbytes) {
1359 if (copy > PAGE_SIZE)
1360 return -EINVAL;
1361 if (clear_user((char __user *)mask + nbytes, copy - nbytes))
1362 return -EFAULT;
1363 copy = nbytes;
1364 }
1365 return copy_to_user(mask, nodes_addr(*nodes), copy) ? -EFAULT : 0;
1366}
1367
1368SYSCALL_DEFINE6(mbind, unsigned long, start, unsigned long, len,
1369 unsigned long, mode, unsigned long __user *, nmask,
1370 unsigned long, maxnode, unsigned, flags)
1371{
1372 nodemask_t nodes;
1373 int err;
1374 unsigned short mode_flags;
1375
1376 mode_flags = mode & MPOL_MODE_FLAGS;
1377 mode &= ~MPOL_MODE_FLAGS;
1378 if (mode >= MPOL_MAX)
1379 return -EINVAL;
1380 if ((mode_flags & MPOL_F_STATIC_NODES) &&
1381 (mode_flags & MPOL_F_RELATIVE_NODES))
1382 return -EINVAL;
1383 err = get_nodes(&nodes, nmask, maxnode);
1384 if (err)
1385 return err;
1386 return do_mbind(start, len, mode, mode_flags, &nodes, flags);
1387}
1388
1389/* Set the process memory policy */
1390SYSCALL_DEFINE3(set_mempolicy, int, mode, unsigned long __user *, nmask,
1391 unsigned long, maxnode)
1392{
1393 int err;
1394 nodemask_t nodes;
1395 unsigned short flags;
1396
1397 flags = mode & MPOL_MODE_FLAGS;
1398 mode &= ~MPOL_MODE_FLAGS;
1399 if ((unsigned int)mode >= MPOL_MAX)
1400 return -EINVAL;
1401 if ((flags & MPOL_F_STATIC_NODES) && (flags & MPOL_F_RELATIVE_NODES))
1402 return -EINVAL;
1403 err = get_nodes(&nodes, nmask, maxnode);
1404 if (err)
1405 return err;
1406 return do_set_mempolicy(mode, flags, &nodes);
1407}
1408
1409SYSCALL_DEFINE4(migrate_pages, pid_t, pid, unsigned long, maxnode,
1410 const unsigned long __user *, old_nodes,
1411 const unsigned long __user *, new_nodes)
1412{
1413 const struct cred *cred = current_cred(), *tcred;
1414 struct mm_struct *mm = NULL;
1415 struct task_struct *task;
1416 nodemask_t task_nodes;
1417 int err;
1418 nodemask_t *old;
1419 nodemask_t *new;
1420 NODEMASK_SCRATCH(scratch);
1421
1422 if (!scratch)
1423 return -ENOMEM;
1424
1425 old = &scratch->mask1;
1426 new = &scratch->mask2;
1427
1428 err = get_nodes(old, old_nodes, maxnode);
1429 if (err)
1430 goto out;
1431
1432 err = get_nodes(new, new_nodes, maxnode);
1433 if (err)
1434 goto out;
1435
1436 /* Find the mm_struct */
1437 rcu_read_lock();
1438 task = pid ? find_task_by_vpid(pid) : current;
1439 if (!task) {
1440 rcu_read_unlock();
1441 err = -ESRCH;
1442 goto out;
1443 }
1444 get_task_struct(task);
1445
1446 err = -EINVAL;
1447
1448 /*
1449 * Check if this process has the right to modify the specified
1450 * process. The right exists if the process has administrative
1451 * capabilities, superuser privileges or the same
1452 * userid as the target process.
1453 */
1454 tcred = __task_cred(task);
1455 if (!uid_eq(cred->euid, tcred->suid) && !uid_eq(cred->euid, tcred->uid) &&
1456 !uid_eq(cred->uid, tcred->suid) && !uid_eq(cred->uid, tcred->uid) &&
1457 !capable(CAP_SYS_NICE)) {
1458 rcu_read_unlock();
1459 err = -EPERM;
1460 goto out_put;
1461 }
1462 rcu_read_unlock();
1463
1464 task_nodes = cpuset_mems_allowed(task);
1465 /* Is the user allowed to access the target nodes? */
1466 if (!nodes_subset(*new, task_nodes) && !capable(CAP_SYS_NICE)) {
1467 err = -EPERM;
1468 goto out_put;
1469 }
1470
1471 if (!nodes_subset(*new, node_states[N_MEMORY])) {
1472 err = -EINVAL;
1473 goto out_put;
1474 }
1475
1476 err = security_task_movememory(task);
1477 if (err)
1478 goto out_put;
1479
1480 mm = get_task_mm(task);
1481 put_task_struct(task);
1482
1483 if (!mm) {
1484 err = -EINVAL;
1485 goto out;
1486 }
1487
1488 err = do_migrate_pages(mm, old, new,
1489 capable(CAP_SYS_NICE) ? MPOL_MF_MOVE_ALL : MPOL_MF_MOVE);
1490
1491 mmput(mm);
1492out:
1493 NODEMASK_SCRATCH_FREE(scratch);
1494
1495 return err;
1496
1497out_put:
1498 put_task_struct(task);
1499 goto out;
1500
1501}
1502
1503
1504/* Retrieve NUMA policy */
1505SYSCALL_DEFINE5(get_mempolicy, int __user *, policy,
1506 unsigned long __user *, nmask, unsigned long, maxnode,
1507 unsigned long, addr, unsigned long, flags)
1508{
1509 int err;
1510 int uninitialized_var(pval);
1511 nodemask_t nodes;
1512
1513 if (nmask != NULL && maxnode < MAX_NUMNODES)
1514 return -EINVAL;
1515
1516 err = do_get_mempolicy(&pval, &nodes, addr, flags);
1517
1518 if (err)
1519 return err;
1520
1521 if (policy && put_user(pval, policy))
1522 return -EFAULT;
1523
1524 if (nmask)
1525 err = copy_nodes_to_user(nmask, maxnode, &nodes);
1526
1527 return err;
1528}
1529
1530#ifdef CONFIG_COMPAT
1531
1532COMPAT_SYSCALL_DEFINE5(get_mempolicy, int __user *, policy,
1533 compat_ulong_t __user *, nmask,
1534 compat_ulong_t, maxnode,
1535 compat_ulong_t, addr, compat_ulong_t, flags)
1536{
1537 long err;
1538 unsigned long __user *nm = NULL;
1539 unsigned long nr_bits, alloc_size;
1540 DECLARE_BITMAP(bm, MAX_NUMNODES);
1541
1542 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1543 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1544
1545 if (nmask)
1546 nm = compat_alloc_user_space(alloc_size);
1547
1548 err = sys_get_mempolicy(policy, nm, nr_bits+1, addr, flags);
1549
1550 if (!err && nmask) {
1551 unsigned long copy_size;
1552 copy_size = min_t(unsigned long, sizeof(bm), alloc_size);
1553 err = copy_from_user(bm, nm, copy_size);
1554 /* ensure entire bitmap is zeroed */
1555 err |= clear_user(nmask, ALIGN(maxnode-1, 8) / 8);
1556 err |= compat_put_bitmap(nmask, bm, nr_bits);
1557 }
1558
1559 return err;
1560}
1561
1562COMPAT_SYSCALL_DEFINE3(set_mempolicy, int, mode, compat_ulong_t __user *, nmask,
1563 compat_ulong_t, maxnode)
1564{
1565 long err = 0;
1566 unsigned long __user *nm = NULL;
1567 unsigned long nr_bits, alloc_size;
1568 DECLARE_BITMAP(bm, MAX_NUMNODES);
1569
1570 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1571 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1572
1573 if (nmask) {
1574 err = compat_get_bitmap(bm, nmask, nr_bits);
1575 nm = compat_alloc_user_space(alloc_size);
1576 err |= copy_to_user(nm, bm, alloc_size);
1577 }
1578
1579 if (err)
1580 return -EFAULT;
1581
1582 return sys_set_mempolicy(mode, nm, nr_bits+1);
1583}
1584
1585COMPAT_SYSCALL_DEFINE6(mbind, compat_ulong_t, start, compat_ulong_t, len,
1586 compat_ulong_t, mode, compat_ulong_t __user *, nmask,
1587 compat_ulong_t, maxnode, compat_ulong_t, flags)
1588{
1589 long err = 0;
1590 unsigned long __user *nm = NULL;
1591 unsigned long nr_bits, alloc_size;
1592 nodemask_t bm;
1593
1594 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1595 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1596
1597 if (nmask) {
1598 err = compat_get_bitmap(nodes_addr(bm), nmask, nr_bits);
1599 nm = compat_alloc_user_space(alloc_size);
1600 err |= copy_to_user(nm, nodes_addr(bm), alloc_size);
1601 }
1602
1603 if (err)
1604 return -EFAULT;
1605
1606 return sys_mbind(start, len, mode, nm, nr_bits+1, flags);
1607}
1608
1609#endif
1610
1611/*
1612 * get_vma_policy(@task, @vma, @addr)
1613 * @task - task for fallback if vma policy == default
1614 * @vma - virtual memory area whose policy is sought
1615 * @addr - address in @vma for shared policy lookup
1616 *
1617 * Returns effective policy for a VMA at specified address.
1618 * Falls back to @task or system default policy, as necessary.
1619 * Current or other task's task mempolicy and non-shared vma policies must be
1620 * protected by task_lock(task) by the caller.
1621 * Shared policies [those marked as MPOL_F_SHARED] require an extra reference
1622 * count--added by the get_policy() vm_op, as appropriate--to protect against
1623 * freeing by another task. It is the caller's responsibility to free the
1624 * extra reference for shared policies.
1625 */
1626struct mempolicy *get_vma_policy(struct task_struct *task,
1627 struct vm_area_struct *vma, unsigned long addr)
1628{
1629 struct mempolicy *pol = get_task_policy(task);
1630
1631 if (vma) {
1632 if (vma->vm_ops && vma->vm_ops->get_policy) {
1633 struct mempolicy *vpol = vma->vm_ops->get_policy(vma,
1634 addr);
1635 if (vpol)
1636 pol = vpol;
1637 } else if (vma->vm_policy) {
1638 pol = vma->vm_policy;
1639
1640 /*
1641 * shmem_alloc_page() passes MPOL_F_SHARED policy with
1642 * a pseudo vma whose vma->vm_ops=NULL. Take a reference
1643 * count on these policies which will be dropped by
1644 * mpol_cond_put() later
1645 */
1646 if (mpol_needs_cond_ref(pol))
1647 mpol_get(pol);
1648 }
1649 }
1650 if (!pol)
1651 pol = &default_policy;
1652 return pol;
1653}
1654
1655bool vma_policy_mof(struct task_struct *task, struct vm_area_struct *vma)
1656{
1657 struct mempolicy *pol = get_task_policy(task);
1658 if (vma) {
1659 if (vma->vm_ops && vma->vm_ops->get_policy) {
1660 bool ret = false;
1661
1662 pol = vma->vm_ops->get_policy(vma, vma->vm_start);
1663 if (pol && (pol->flags & MPOL_F_MOF))
1664 ret = true;
1665 mpol_cond_put(pol);
1666
1667 return ret;
1668 } else if (vma->vm_policy) {
1669 pol = vma->vm_policy;
1670 }
1671 }
1672
1673 if (!pol)
1674 return default_policy.flags & MPOL_F_MOF;
1675
1676 return pol->flags & MPOL_F_MOF;
1677}
1678
1679static int apply_policy_zone(struct mempolicy *policy, enum zone_type zone)
1680{
1681 enum zone_type dynamic_policy_zone = policy_zone;
1682
1683 BUG_ON(dynamic_policy_zone == ZONE_MOVABLE);
1684
1685 /*
1686 * if policy->v.nodes has movable memory only,
1687 * we apply policy when gfp_zone(gfp) = ZONE_MOVABLE only.
1688 *
1689 * policy->v.nodes is intersect with node_states[N_MEMORY].
1690 * so if the following test faile, it implies
1691 * policy->v.nodes has movable memory only.
1692 */
1693 if (!nodes_intersects(policy->v.nodes, node_states[N_HIGH_MEMORY]))
1694 dynamic_policy_zone = ZONE_MOVABLE;
1695
1696 return zone >= dynamic_policy_zone;
1697}
1698
1699/*
1700 * Return a nodemask representing a mempolicy for filtering nodes for
1701 * page allocation
1702 */
1703static nodemask_t *policy_nodemask(gfp_t gfp, struct mempolicy *policy)
1704{
1705 /* Lower zones don't get a nodemask applied for MPOL_BIND */
1706 if (unlikely(policy->mode == MPOL_BIND) &&
1707 apply_policy_zone(policy, gfp_zone(gfp)) &&
1708 cpuset_nodemask_valid_mems_allowed(&policy->v.nodes))
1709 return &policy->v.nodes;
1710
1711 return NULL;
1712}
1713
1714/* Return a zonelist indicated by gfp for node representing a mempolicy */
1715static struct zonelist *policy_zonelist(gfp_t gfp, struct mempolicy *policy,
1716 int nd)
1717{
1718 switch (policy->mode) {
1719 case MPOL_PREFERRED:
1720 if (!(policy->flags & MPOL_F_LOCAL))
1721 nd = policy->v.preferred_node;
1722 break;
1723 case MPOL_BIND:
1724 /*
1725 * Normally, MPOL_BIND allocations are node-local within the
1726 * allowed nodemask. However, if __GFP_THISNODE is set and the
1727 * current node isn't part of the mask, we use the zonelist for
1728 * the first node in the mask instead.
1729 */
1730 if (unlikely(gfp & __GFP_THISNODE) &&
1731 unlikely(!node_isset(nd, policy->v.nodes)))
1732 nd = first_node(policy->v.nodes);
1733 break;
1734 default:
1735 BUG();
1736 }
1737 return node_zonelist(nd, gfp);
1738}
1739
1740/* Do dynamic interleaving for a process */
1741static unsigned interleave_nodes(struct mempolicy *policy)
1742{
1743 unsigned nid, next;
1744 struct task_struct *me = current;
1745
1746 nid = me->il_next;
1747 next = next_node(nid, policy->v.nodes);
1748 if (next >= MAX_NUMNODES)
1749 next = first_node(policy->v.nodes);
1750 if (next < MAX_NUMNODES)
1751 me->il_next = next;
1752 return nid;
1753}
1754
1755/*
1756 * Depending on the memory policy provide a node from which to allocate the
1757 * next slab entry.
1758 */
1759unsigned int mempolicy_slab_node(void)
1760{
1761 struct mempolicy *policy;
1762 int node = numa_mem_id();
1763
1764 if (in_interrupt())
1765 return node;
1766
1767 policy = current->mempolicy;
1768 if (!policy || policy->flags & MPOL_F_LOCAL)
1769 return node;
1770
1771 switch (policy->mode) {
1772 case MPOL_PREFERRED:
1773 /*
1774 * handled MPOL_F_LOCAL above
1775 */
1776 return policy->v.preferred_node;
1777
1778 case MPOL_INTERLEAVE:
1779 return interleave_nodes(policy);
1780
1781 case MPOL_BIND: {
1782 /*
1783 * Follow bind policy behavior and start allocation at the
1784 * first node.
1785 */
1786 struct zonelist *zonelist;
1787 struct zone *zone;
1788 enum zone_type highest_zoneidx = gfp_zone(GFP_KERNEL);
1789 zonelist = &NODE_DATA(node)->node_zonelists[0];
1790 (void)first_zones_zonelist(zonelist, highest_zoneidx,
1791 &policy->v.nodes,
1792 &zone);
1793 return zone ? zone->node : node;
1794 }
1795
1796 default:
1797 BUG();
1798 }
1799}
1800
1801/* Do static interleaving for a VMA with known offset. */
1802static unsigned offset_il_node(struct mempolicy *pol,
1803 struct vm_area_struct *vma, unsigned long off)
1804{
1805 unsigned nnodes = nodes_weight(pol->v.nodes);
1806 unsigned target;
1807 int c;
1808 int nid = NUMA_NO_NODE;
1809
1810 if (!nnodes)
1811 return numa_node_id();
1812 target = (unsigned int)off % nnodes;
1813 c = 0;
1814 do {
1815 nid = next_node(nid, pol->v.nodes);
1816 c++;
1817 } while (c <= target);
1818 return nid;
1819}
1820
1821/* Determine a node number for interleave */
1822static inline unsigned interleave_nid(struct mempolicy *pol,
1823 struct vm_area_struct *vma, unsigned long addr, int shift)
1824{
1825 if (vma) {
1826 unsigned long off;
1827
1828 /*
1829 * for small pages, there is no difference between
1830 * shift and PAGE_SHIFT, so the bit-shift is safe.
1831 * for huge pages, since vm_pgoff is in units of small
1832 * pages, we need to shift off the always 0 bits to get
1833 * a useful offset.
1834 */
1835 BUG_ON(shift < PAGE_SHIFT);
1836 off = vma->vm_pgoff >> (shift - PAGE_SHIFT);
1837 off += (addr - vma->vm_start) >> shift;
1838 return offset_il_node(pol, vma, off);
1839 } else
1840 return interleave_nodes(pol);
1841}
1842
1843/*
1844 * Return the bit number of a random bit set in the nodemask.
1845 * (returns NUMA_NO_NODE if nodemask is empty)
1846 */
1847int node_random(const nodemask_t *maskp)
1848{
1849 int w, bit = NUMA_NO_NODE;
1850
1851 w = nodes_weight(*maskp);
1852 if (w)
1853 bit = bitmap_ord_to_pos(maskp->bits,
1854 get_random_int() % w, MAX_NUMNODES);
1855 return bit;
1856}
1857
1858#ifdef CONFIG_HUGETLBFS
1859/*
1860 * huge_zonelist(@vma, @addr, @gfp_flags, @mpol)
1861 * @vma = virtual memory area whose policy is sought
1862 * @addr = address in @vma for shared policy lookup and interleave policy
1863 * @gfp_flags = for requested zone
1864 * @mpol = pointer to mempolicy pointer for reference counted mempolicy
1865 * @nodemask = pointer to nodemask pointer for MPOL_BIND nodemask
1866 *
1867 * Returns a zonelist suitable for a huge page allocation and a pointer
1868 * to the struct mempolicy for conditional unref after allocation.
1869 * If the effective policy is 'BIND, returns a pointer to the mempolicy's
1870 * @nodemask for filtering the zonelist.
1871 *
1872 * Must be protected by read_mems_allowed_begin()
1873 */
1874struct zonelist *huge_zonelist(struct vm_area_struct *vma, unsigned long addr,
1875 gfp_t gfp_flags, struct mempolicy **mpol,
1876 nodemask_t **nodemask)
1877{
1878 struct zonelist *zl;
1879
1880 *mpol = get_vma_policy(current, vma, addr);
1881 *nodemask = NULL; /* assume !MPOL_BIND */
1882
1883 if (unlikely((*mpol)->mode == MPOL_INTERLEAVE)) {
1884 zl = node_zonelist(interleave_nid(*mpol, vma, addr,
1885 huge_page_shift(hstate_vma(vma))), gfp_flags);
1886 } else {
1887 zl = policy_zonelist(gfp_flags, *mpol, numa_node_id());
1888 if ((*mpol)->mode == MPOL_BIND)
1889 *nodemask = &(*mpol)->v.nodes;
1890 }
1891 return zl;
1892}
1893
1894/*
1895 * init_nodemask_of_mempolicy
1896 *
1897 * If the current task's mempolicy is "default" [NULL], return 'false'
1898 * to indicate default policy. Otherwise, extract the policy nodemask
1899 * for 'bind' or 'interleave' policy into the argument nodemask, or
1900 * initialize the argument nodemask to contain the single node for
1901 * 'preferred' or 'local' policy and return 'true' to indicate presence
1902 * of non-default mempolicy.
1903 *
1904 * We don't bother with reference counting the mempolicy [mpol_get/put]
1905 * because the current task is examining it's own mempolicy and a task's
1906 * mempolicy is only ever changed by the task itself.
1907 *
1908 * N.B., it is the caller's responsibility to free a returned nodemask.
1909 */
1910bool init_nodemask_of_mempolicy(nodemask_t *mask)
1911{
1912 struct mempolicy *mempolicy;
1913 int nid;
1914
1915 if (!(mask && current->mempolicy))
1916 return false;
1917
1918 task_lock(current);
1919 mempolicy = current->mempolicy;
1920 switch (mempolicy->mode) {
1921 case MPOL_PREFERRED:
1922 if (mempolicy->flags & MPOL_F_LOCAL)
1923 nid = numa_node_id();
1924 else
1925 nid = mempolicy->v.preferred_node;
1926 init_nodemask_of_node(mask, nid);
1927 break;
1928
1929 case MPOL_BIND:
1930 /* Fall through */
1931 case MPOL_INTERLEAVE:
1932 *mask = mempolicy->v.nodes;
1933 break;
1934
1935 default:
1936 BUG();
1937 }
1938 task_unlock(current);
1939
1940 return true;
1941}
1942#endif
1943
1944/*
1945 * mempolicy_nodemask_intersects
1946 *
1947 * If tsk's mempolicy is "default" [NULL], return 'true' to indicate default
1948 * policy. Otherwise, check for intersection between mask and the policy
1949 * nodemask for 'bind' or 'interleave' policy. For 'perferred' or 'local'
1950 * policy, always return true since it may allocate elsewhere on fallback.
1951 *
1952 * Takes task_lock(tsk) to prevent freeing of its mempolicy.
1953 */
1954bool mempolicy_nodemask_intersects(struct task_struct *tsk,
1955 const nodemask_t *mask)
1956{
1957 struct mempolicy *mempolicy;
1958 bool ret = true;
1959
1960 if (!mask)
1961 return ret;
1962 task_lock(tsk);
1963 mempolicy = tsk->mempolicy;
1964 if (!mempolicy)
1965 goto out;
1966
1967 switch (mempolicy->mode) {
1968 case MPOL_PREFERRED:
1969 /*
1970 * MPOL_PREFERRED and MPOL_F_LOCAL are only preferred nodes to
1971 * allocate from, they may fallback to other nodes when oom.
1972 * Thus, it's possible for tsk to have allocated memory from
1973 * nodes in mask.
1974 */
1975 break;
1976 case MPOL_BIND:
1977 case MPOL_INTERLEAVE:
1978 ret = nodes_intersects(mempolicy->v.nodes, *mask);
1979 break;
1980 default:
1981 BUG();
1982 }
1983out:
1984 task_unlock(tsk);
1985 return ret;
1986}
1987
1988/* Allocate a page in interleaved policy.
1989 Own path because it needs to do special accounting. */
1990static struct page *alloc_page_interleave(gfp_t gfp, unsigned order,
1991 unsigned nid)
1992{
1993 struct zonelist *zl;
1994 struct page *page;
1995
1996 zl = node_zonelist(nid, gfp);
1997 page = __alloc_pages(gfp, order, zl);
1998 if (page && page_zone(page) == zonelist_zone(&zl->_zonerefs[0]))
1999 inc_zone_page_state(page, NUMA_INTERLEAVE_HIT);
2000 return page;
2001}
2002
2003/**
2004 * alloc_pages_vma - Allocate a page for a VMA.
2005 *
2006 * @gfp:
2007 * %GFP_USER user allocation.
2008 * %GFP_KERNEL kernel allocations,
2009 * %GFP_HIGHMEM highmem/user allocations,
2010 * %GFP_FS allocation should not call back into a file system.
2011 * %GFP_ATOMIC don't sleep.
2012 *
2013 * @order:Order of the GFP allocation.
2014 * @vma: Pointer to VMA or NULL if not available.
2015 * @addr: Virtual Address of the allocation. Must be inside the VMA.
2016 *
2017 * This function allocates a page from the kernel page pool and applies
2018 * a NUMA policy associated with the VMA or the current process.
2019 * When VMA is not NULL caller must hold down_read on the mmap_sem of the
2020 * mm_struct of the VMA to prevent it from going away. Should be used for
2021 * all allocations for pages that will be mapped into
2022 * user space. Returns NULL when no page can be allocated.
2023 *
2024 * Should be called with the mm_sem of the vma hold.
2025 */
2026struct page *
2027alloc_pages_vma(gfp_t gfp, int order, struct vm_area_struct *vma,
2028 unsigned long addr, int node)
2029{
2030 struct mempolicy *pol;
2031 struct page *page;
2032 unsigned int cpuset_mems_cookie;
2033
2034retry_cpuset:
2035 pol = get_vma_policy(current, vma, addr);
2036 cpuset_mems_cookie = read_mems_allowed_begin();
2037
2038 if (unlikely(pol->mode == MPOL_INTERLEAVE)) {
2039 unsigned nid;
2040
2041 nid = interleave_nid(pol, vma, addr, PAGE_SHIFT + order);
2042 mpol_cond_put(pol);
2043 page = alloc_page_interleave(gfp, order, nid);
2044 if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie)))
2045 goto retry_cpuset;
2046
2047 return page;
2048 }
2049 page = __alloc_pages_nodemask(gfp, order,
2050 policy_zonelist(gfp, pol, node),
2051 policy_nodemask(gfp, pol));
2052 if (unlikely(mpol_needs_cond_ref(pol)))
2053 __mpol_put(pol);
2054 if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie)))
2055 goto retry_cpuset;
2056 return page;
2057}
2058
2059/**
2060 * alloc_pages_current - Allocate pages.
2061 *
2062 * @gfp:
2063 * %GFP_USER user allocation,
2064 * %GFP_KERNEL kernel allocation,
2065 * %GFP_HIGHMEM highmem allocation,
2066 * %GFP_FS don't call back into a file system.
2067 * %GFP_ATOMIC don't sleep.
2068 * @order: Power of two of allocation size in pages. 0 is a single page.
2069 *
2070 * Allocate a page from the kernel page pool. When not in
2071 * interrupt context and apply the current process NUMA policy.
2072 * Returns NULL when no page can be allocated.
2073 *
2074 * Don't call cpuset_update_task_memory_state() unless
2075 * 1) it's ok to take cpuset_sem (can WAIT), and
2076 * 2) allocating for current task (not interrupt).
2077 */
2078struct page *alloc_pages_current(gfp_t gfp, unsigned order)
2079{
2080 struct mempolicy *pol = get_task_policy(current);
2081 struct page *page;
2082 unsigned int cpuset_mems_cookie;
2083
2084 if (!pol || in_interrupt() || (gfp & __GFP_THISNODE))
2085 pol = &default_policy;
2086
2087retry_cpuset:
2088 cpuset_mems_cookie = read_mems_allowed_begin();
2089
2090 /*
2091 * No reference counting needed for current->mempolicy
2092 * nor system default_policy
2093 */
2094 if (pol->mode == MPOL_INTERLEAVE)
2095 page = alloc_page_interleave(gfp, order, interleave_nodes(pol));
2096 else
2097 page = __alloc_pages_nodemask(gfp, order,
2098 policy_zonelist(gfp, pol, numa_node_id()),
2099 policy_nodemask(gfp, pol));
2100
2101 if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie)))
2102 goto retry_cpuset;
2103
2104 return page;
2105}
2106EXPORT_SYMBOL(alloc_pages_current);
2107
2108int vma_dup_policy(struct vm_area_struct *src, struct vm_area_struct *dst)
2109{
2110 struct mempolicy *pol = mpol_dup(vma_policy(src));
2111
2112 if (IS_ERR(pol))
2113 return PTR_ERR(pol);
2114 dst->vm_policy = pol;
2115 return 0;
2116}
2117
2118/*
2119 * If mpol_dup() sees current->cpuset == cpuset_being_rebound, then it
2120 * rebinds the mempolicy its copying by calling mpol_rebind_policy()
2121 * with the mems_allowed returned by cpuset_mems_allowed(). This
2122 * keeps mempolicies cpuset relative after its cpuset moves. See
2123 * further kernel/cpuset.c update_nodemask().
2124 *
2125 * current's mempolicy may be rebinded by the other task(the task that changes
2126 * cpuset's mems), so we needn't do rebind work for current task.
2127 */
2128
2129/* Slow path of a mempolicy duplicate */
2130struct mempolicy *__mpol_dup(struct mempolicy *old)
2131{
2132 struct mempolicy *new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
2133
2134 if (!new)
2135 return ERR_PTR(-ENOMEM);
2136
2137 /* task's mempolicy is protected by alloc_lock */
2138 if (old == current->mempolicy) {
2139 task_lock(current);
2140 *new = *old;
2141 task_unlock(current);
2142 } else
2143 *new = *old;
2144
2145 rcu_read_lock();
2146 if (current_cpuset_is_being_rebound()) {
2147 nodemask_t mems = cpuset_mems_allowed(current);
2148 if (new->flags & MPOL_F_REBINDING)
2149 mpol_rebind_policy(new, &mems, MPOL_REBIND_STEP2);
2150 else
2151 mpol_rebind_policy(new, &mems, MPOL_REBIND_ONCE);
2152 }
2153 rcu_read_unlock();
2154 atomic_set(&new->refcnt, 1);
2155 return new;
2156}
2157
2158/* Slow path of a mempolicy comparison */
2159bool __mpol_equal(struct mempolicy *a, struct mempolicy *b)
2160{
2161 if (!a || !b)
2162 return false;
2163 if (a->mode != b->mode)
2164 return false;
2165 if (a->flags != b->flags)
2166 return false;
2167 if (mpol_store_user_nodemask(a))
2168 if (!nodes_equal(a->w.user_nodemask, b->w.user_nodemask))
2169 return false;
2170
2171 switch (a->mode) {
2172 case MPOL_BIND:
2173 /* Fall through */
2174 case MPOL_INTERLEAVE:
2175 return !!nodes_equal(a->v.nodes, b->v.nodes);
2176 case MPOL_PREFERRED:
2177 return a->v.preferred_node == b->v.preferred_node;
2178 default:
2179 BUG();
2180 return false;
2181 }
2182}
2183
2184/*
2185 * Shared memory backing store policy support.
2186 *
2187 * Remember policies even when nobody has shared memory mapped.
2188 * The policies are kept in Red-Black tree linked from the inode.
2189 * They are protected by the sp->lock spinlock, which should be held
2190 * for any accesses to the tree.
2191 */
2192
2193/* lookup first element intersecting start-end */
2194/* Caller holds sp->lock */
2195static struct sp_node *
2196sp_lookup(struct shared_policy *sp, unsigned long start, unsigned long end)
2197{
2198 struct rb_node *n = sp->root.rb_node;
2199
2200 while (n) {
2201 struct sp_node *p = rb_entry(n, struct sp_node, nd);
2202
2203 if (start >= p->end)
2204 n = n->rb_right;
2205 else if (end <= p->start)
2206 n = n->rb_left;
2207 else
2208 break;
2209 }
2210 if (!n)
2211 return NULL;
2212 for (;;) {
2213 struct sp_node *w = NULL;
2214 struct rb_node *prev = rb_prev(n);
2215 if (!prev)
2216 break;
2217 w = rb_entry(prev, struct sp_node, nd);
2218 if (w->end <= start)
2219 break;
2220 n = prev;
2221 }
2222 return rb_entry(n, struct sp_node, nd);
2223}
2224
2225/* Insert a new shared policy into the list. */
2226/* Caller holds sp->lock */
2227static void sp_insert(struct shared_policy *sp, struct sp_node *new)
2228{
2229 struct rb_node **p = &sp->root.rb_node;
2230 struct rb_node *parent = NULL;
2231 struct sp_node *nd;
2232
2233 while (*p) {
2234 parent = *p;
2235 nd = rb_entry(parent, struct sp_node, nd);
2236 if (new->start < nd->start)
2237 p = &(*p)->rb_left;
2238 else if (new->end > nd->end)
2239 p = &(*p)->rb_right;
2240 else
2241 BUG();
2242 }
2243 rb_link_node(&new->nd, parent, p);
2244 rb_insert_color(&new->nd, &sp->root);
2245 pr_debug("inserting %lx-%lx: %d\n", new->start, new->end,
2246 new->policy ? new->policy->mode : 0);
2247}
2248
2249/* Find shared policy intersecting idx */
2250struct mempolicy *
2251mpol_shared_policy_lookup(struct shared_policy *sp, unsigned long idx)
2252{
2253 struct mempolicy *pol = NULL;
2254 struct sp_node *sn;
2255
2256 if (!sp->root.rb_node)
2257 return NULL;
2258 spin_lock(&sp->lock);
2259 sn = sp_lookup(sp, idx, idx+1);
2260 if (sn) {
2261 mpol_get(sn->policy);
2262 pol = sn->policy;
2263 }
2264 spin_unlock(&sp->lock);
2265 return pol;
2266}
2267
2268static void sp_free(struct sp_node *n)
2269{
2270 mpol_put(n->policy);
2271 kmem_cache_free(sn_cache, n);
2272}
2273
2274/**
2275 * mpol_misplaced - check whether current page node is valid in policy
2276 *
2277 * @page - page to be checked
2278 * @vma - vm area where page mapped
2279 * @addr - virtual address where page mapped
2280 *
2281 * Lookup current policy node id for vma,addr and "compare to" page's
2282 * node id.
2283 *
2284 * Returns:
2285 * -1 - not misplaced, page is in the right node
2286 * node - node id where the page should be
2287 *
2288 * Policy determination "mimics" alloc_page_vma().
2289 * Called from fault path where we know the vma and faulting address.
2290 */
2291int mpol_misplaced(struct page *page, struct vm_area_struct *vma, unsigned long addr)
2292{
2293 struct mempolicy *pol;
2294 struct zone *zone;
2295 int curnid = page_to_nid(page);
2296 unsigned long pgoff;
2297 int thiscpu = raw_smp_processor_id();
2298 int thisnid = cpu_to_node(thiscpu);
2299 int polnid = -1;
2300 int ret = -1;
2301
2302 BUG_ON(!vma);
2303
2304 pol = get_vma_policy(current, vma, addr);
2305 if (!(pol->flags & MPOL_F_MOF))
2306 goto out;
2307
2308 switch (pol->mode) {
2309 case MPOL_INTERLEAVE:
2310 BUG_ON(addr >= vma->vm_end);
2311 BUG_ON(addr < vma->vm_start);
2312
2313 pgoff = vma->vm_pgoff;
2314 pgoff += (addr - vma->vm_start) >> PAGE_SHIFT;
2315 polnid = offset_il_node(pol, vma, pgoff);
2316 break;
2317
2318 case MPOL_PREFERRED:
2319 if (pol->flags & MPOL_F_LOCAL)
2320 polnid = numa_node_id();
2321 else
2322 polnid = pol->v.preferred_node;
2323 break;
2324
2325 case MPOL_BIND:
2326 /*
2327 * allows binding to multiple nodes.
2328 * use current page if in policy nodemask,
2329 * else select nearest allowed node, if any.
2330 * If no allowed nodes, use current [!misplaced].
2331 */
2332 if (node_isset(curnid, pol->v.nodes))
2333 goto out;
2334 (void)first_zones_zonelist(
2335 node_zonelist(numa_node_id(), GFP_HIGHUSER),
2336 gfp_zone(GFP_HIGHUSER),
2337 &pol->v.nodes, &zone);
2338 polnid = zone->node;
2339 break;
2340
2341 default:
2342 BUG();
2343 }
2344
2345 /* Migrate the page towards the node whose CPU is referencing it */
2346 if (pol->flags & MPOL_F_MORON) {
2347 polnid = thisnid;
2348
2349 if (!should_numa_migrate_memory(current, page, curnid, thiscpu))
2350 goto out;
2351 }
2352
2353 if (curnid != polnid)
2354 ret = polnid;
2355out:
2356 mpol_cond_put(pol);
2357
2358 return ret;
2359}
2360
2361static void sp_delete(struct shared_policy *sp, struct sp_node *n)
2362{
2363 pr_debug("deleting %lx-l%lx\n", n->start, n->end);
2364 rb_erase(&n->nd, &sp->root);
2365 sp_free(n);
2366}
2367
2368static void sp_node_init(struct sp_node *node, unsigned long start,
2369 unsigned long end, struct mempolicy *pol)
2370{
2371 node->start = start;
2372 node->end = end;
2373 node->policy = pol;
2374}
2375
2376static struct sp_node *sp_alloc(unsigned long start, unsigned long end,
2377 struct mempolicy *pol)
2378{
2379 struct sp_node *n;
2380 struct mempolicy *newpol;
2381
2382 n = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2383 if (!n)
2384 return NULL;
2385
2386 newpol = mpol_dup(pol);
2387 if (IS_ERR(newpol)) {
2388 kmem_cache_free(sn_cache, n);
2389 return NULL;
2390 }
2391 newpol->flags |= MPOL_F_SHARED;
2392 sp_node_init(n, start, end, newpol);
2393
2394 return n;
2395}
2396
2397/* Replace a policy range. */
2398static int shared_policy_replace(struct shared_policy *sp, unsigned long start,
2399 unsigned long end, struct sp_node *new)
2400{
2401 struct sp_node *n;
2402 struct sp_node *n_new = NULL;
2403 struct mempolicy *mpol_new = NULL;
2404 int ret = 0;
2405
2406restart:
2407 spin_lock(&sp->lock);
2408 n = sp_lookup(sp, start, end);
2409 /* Take care of old policies in the same range. */
2410 while (n && n->start < end) {
2411 struct rb_node *next = rb_next(&n->nd);
2412 if (n->start >= start) {
2413 if (n->end <= end)
2414 sp_delete(sp, n);
2415 else
2416 n->start = end;
2417 } else {
2418 /* Old policy spanning whole new range. */
2419 if (n->end > end) {
2420 if (!n_new)
2421 goto alloc_new;
2422
2423 *mpol_new = *n->policy;
2424 atomic_set(&mpol_new->refcnt, 1);
2425 sp_node_init(n_new, end, n->end, mpol_new);
2426 n->end = start;
2427 sp_insert(sp, n_new);
2428 n_new = NULL;
2429 mpol_new = NULL;
2430 break;
2431 } else
2432 n->end = start;
2433 }
2434 if (!next)
2435 break;
2436 n = rb_entry(next, struct sp_node, nd);
2437 }
2438 if (new)
2439 sp_insert(sp, new);
2440 spin_unlock(&sp->lock);
2441 ret = 0;
2442
2443err_out:
2444 if (mpol_new)
2445 mpol_put(mpol_new);
2446 if (n_new)
2447 kmem_cache_free(sn_cache, n_new);
2448
2449 return ret;
2450
2451alloc_new:
2452 spin_unlock(&sp->lock);
2453 ret = -ENOMEM;
2454 n_new = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2455 if (!n_new)
2456 goto err_out;
2457 mpol_new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
2458 if (!mpol_new)
2459 goto err_out;
2460 goto restart;
2461}
2462
2463/**
2464 * mpol_shared_policy_init - initialize shared policy for inode
2465 * @sp: pointer to inode shared policy
2466 * @mpol: struct mempolicy to install
2467 *
2468 * Install non-NULL @mpol in inode's shared policy rb-tree.
2469 * On entry, the current task has a reference on a non-NULL @mpol.
2470 * This must be released on exit.
2471 * This is called at get_inode() calls and we can use GFP_KERNEL.
2472 */
2473void mpol_shared_policy_init(struct shared_policy *sp, struct mempolicy *mpol)
2474{
2475 int ret;
2476
2477 sp->root = RB_ROOT; /* empty tree == default mempolicy */
2478 spin_lock_init(&sp->lock);
2479
2480 if (mpol) {
2481 struct vm_area_struct pvma;
2482 struct mempolicy *new;
2483 NODEMASK_SCRATCH(scratch);
2484
2485 if (!scratch)
2486 goto put_mpol;
2487 /* contextualize the tmpfs mount point mempolicy */
2488 new = mpol_new(mpol->mode, mpol->flags, &mpol->w.user_nodemask);
2489 if (IS_ERR(new))
2490 goto free_scratch; /* no valid nodemask intersection */
2491
2492 task_lock(current);
2493 ret = mpol_set_nodemask(new, &mpol->w.user_nodemask, scratch);
2494 task_unlock(current);
2495 if (ret)
2496 goto put_new;
2497
2498 /* Create pseudo-vma that contains just the policy */
2499 memset(&pvma, 0, sizeof(struct vm_area_struct));
2500 pvma.vm_end = TASK_SIZE; /* policy covers entire file */
2501 mpol_set_shared_policy(sp, &pvma, new); /* adds ref */
2502
2503put_new:
2504 mpol_put(new); /* drop initial ref */
2505free_scratch:
2506 NODEMASK_SCRATCH_FREE(scratch);
2507put_mpol:
2508 mpol_put(mpol); /* drop our incoming ref on sb mpol */
2509 }
2510}
2511
2512int mpol_set_shared_policy(struct shared_policy *info,
2513 struct vm_area_struct *vma, struct mempolicy *npol)
2514{
2515 int err;
2516 struct sp_node *new = NULL;
2517 unsigned long sz = vma_pages(vma);
2518
2519 pr_debug("set_shared_policy %lx sz %lu %d %d %lx\n",
2520 vma->vm_pgoff,
2521 sz, npol ? npol->mode : -1,
2522 npol ? npol->flags : -1,
2523 npol ? nodes_addr(npol->v.nodes)[0] : NUMA_NO_NODE);
2524
2525 if (npol) {
2526 new = sp_alloc(vma->vm_pgoff, vma->vm_pgoff + sz, npol);
2527 if (!new)
2528 return -ENOMEM;
2529 }
2530 err = shared_policy_replace(info, vma->vm_pgoff, vma->vm_pgoff+sz, new);
2531 if (err && new)
2532 sp_free(new);
2533 return err;
2534}
2535
2536/* Free a backing policy store on inode delete. */
2537void mpol_free_shared_policy(struct shared_policy *p)
2538{
2539 struct sp_node *n;
2540 struct rb_node *next;
2541
2542 if (!p->root.rb_node)
2543 return;
2544 spin_lock(&p->lock);
2545 next = rb_first(&p->root);
2546 while (next) {
2547 n = rb_entry(next, struct sp_node, nd);
2548 next = rb_next(&n->nd);
2549 sp_delete(p, n);
2550 }
2551 spin_unlock(&p->lock);
2552}
2553
2554#ifdef CONFIG_NUMA_BALANCING
2555static int __initdata numabalancing_override;
2556
2557static void __init check_numabalancing_enable(void)
2558{
2559 bool numabalancing_default = false;
2560
2561 if (IS_ENABLED(CONFIG_NUMA_BALANCING_DEFAULT_ENABLED))
2562 numabalancing_default = true;
2563
2564 /* Parsed by setup_numabalancing. override == 1 enables, -1 disables */
2565 if (numabalancing_override)
2566 set_numabalancing_state(numabalancing_override == 1);
2567
2568 if (nr_node_ids > 1 && !numabalancing_override) {
2569 pr_info("%s automatic NUMA balancing. "
2570 "Configure with numa_balancing= or the "
2571 "kernel.numa_balancing sysctl",
2572 numabalancing_default ? "Enabling" : "Disabling");
2573 set_numabalancing_state(numabalancing_default);
2574 }
2575}
2576
2577static int __init setup_numabalancing(char *str)
2578{
2579 int ret = 0;
2580 if (!str)
2581 goto out;
2582
2583 if (!strcmp(str, "enable")) {
2584 numabalancing_override = 1;
2585 ret = 1;
2586 } else if (!strcmp(str, "disable")) {
2587 numabalancing_override = -1;
2588 ret = 1;
2589 }
2590out:
2591 if (!ret)
2592 pr_warn("Unable to parse numa_balancing=\n");
2593
2594 return ret;
2595}
2596__setup("numa_balancing=", setup_numabalancing);
2597#else
2598static inline void __init check_numabalancing_enable(void)
2599{
2600}
2601#endif /* CONFIG_NUMA_BALANCING */
2602
2603/* assumes fs == KERNEL_DS */
2604void __init numa_policy_init(void)
2605{
2606 nodemask_t interleave_nodes;
2607 unsigned long largest = 0;
2608 int nid, prefer = 0;
2609
2610 policy_cache = kmem_cache_create("numa_policy",
2611 sizeof(struct mempolicy),
2612 0, SLAB_PANIC, NULL);
2613
2614 sn_cache = kmem_cache_create("shared_policy_node",
2615 sizeof(struct sp_node),
2616 0, SLAB_PANIC, NULL);
2617
2618 for_each_node(nid) {
2619 preferred_node_policy[nid] = (struct mempolicy) {
2620 .refcnt = ATOMIC_INIT(1),
2621 .mode = MPOL_PREFERRED,
2622 .flags = MPOL_F_MOF | MPOL_F_MORON,
2623 .v = { .preferred_node = nid, },
2624 };
2625 }
2626
2627 /*
2628 * Set interleaving policy for system init. Interleaving is only
2629 * enabled across suitably sized nodes (default is >= 16MB), or
2630 * fall back to the largest node if they're all smaller.
2631 */
2632 nodes_clear(interleave_nodes);
2633 for_each_node_state(nid, N_MEMORY) {
2634 unsigned long total_pages = node_present_pages(nid);
2635
2636 /* Preserve the largest node */
2637 if (largest < total_pages) {
2638 largest = total_pages;
2639 prefer = nid;
2640 }
2641
2642 /* Interleave this node? */
2643 if ((total_pages << PAGE_SHIFT) >= (16 << 20))
2644 node_set(nid, interleave_nodes);
2645 }
2646
2647 /* All too small, use the largest */
2648 if (unlikely(nodes_empty(interleave_nodes)))
2649 node_set(prefer, interleave_nodes);
2650
2651 if (do_set_mempolicy(MPOL_INTERLEAVE, 0, &interleave_nodes))
2652 printk("numa_policy_init: interleaving failed\n");
2653
2654 check_numabalancing_enable();
2655}
2656
2657/* Reset policy of current process to default */
2658void numa_default_policy(void)
2659{
2660 do_set_mempolicy(MPOL_DEFAULT, 0, NULL);
2661}
2662
2663/*
2664 * Parse and format mempolicy from/to strings
2665 */
2666
2667/*
2668 * "local" is implemented internally by MPOL_PREFERRED with MPOL_F_LOCAL flag.
2669 */
2670static const char * const policy_modes[] =
2671{
2672 [MPOL_DEFAULT] = "default",
2673 [MPOL_PREFERRED] = "prefer",
2674 [MPOL_BIND] = "bind",
2675 [MPOL_INTERLEAVE] = "interleave",
2676 [MPOL_LOCAL] = "local",
2677};
2678
2679
2680#ifdef CONFIG_TMPFS
2681/**
2682 * mpol_parse_str - parse string to mempolicy, for tmpfs mpol mount option.
2683 * @str: string containing mempolicy to parse
2684 * @mpol: pointer to struct mempolicy pointer, returned on success.
2685 *
2686 * Format of input:
2687 * <mode>[=<flags>][:<nodelist>]
2688 *
2689 * On success, returns 0, else 1
2690 */
2691int mpol_parse_str(char *str, struct mempolicy **mpol)
2692{
2693 struct mempolicy *new = NULL;
2694 unsigned short mode;
2695 unsigned short mode_flags;
2696 nodemask_t nodes;
2697 char *nodelist = strchr(str, ':');
2698 char *flags = strchr(str, '=');
2699 int err = 1;
2700
2701 if (nodelist) {
2702 /* NUL-terminate mode or flags string */
2703 *nodelist++ = '\0';
2704 if (nodelist_parse(nodelist, nodes))
2705 goto out;
2706 if (!nodes_subset(nodes, node_states[N_MEMORY]))
2707 goto out;
2708 } else
2709 nodes_clear(nodes);
2710
2711 if (flags)
2712 *flags++ = '\0'; /* terminate mode string */
2713
2714 for (mode = 0; mode < MPOL_MAX; mode++) {
2715 if (!strcmp(str, policy_modes[mode])) {
2716 break;
2717 }
2718 }
2719 if (mode >= MPOL_MAX)
2720 goto out;
2721
2722 switch (mode) {
2723 case MPOL_PREFERRED:
2724 /*
2725 * Insist on a nodelist of one node only
2726 */
2727 if (nodelist) {
2728 char *rest = nodelist;
2729 while (isdigit(*rest))
2730 rest++;
2731 if (*rest)
2732 goto out;
2733 }
2734 break;
2735 case MPOL_INTERLEAVE:
2736 /*
2737 * Default to online nodes with memory if no nodelist
2738 */
2739 if (!nodelist)
2740 nodes = node_states[N_MEMORY];
2741 break;
2742 case MPOL_LOCAL:
2743 /*
2744 * Don't allow a nodelist; mpol_new() checks flags
2745 */
2746 if (nodelist)
2747 goto out;
2748 mode = MPOL_PREFERRED;
2749 break;
2750 case MPOL_DEFAULT:
2751 /*
2752 * Insist on a empty nodelist
2753 */
2754 if (!nodelist)
2755 err = 0;
2756 goto out;
2757 case MPOL_BIND:
2758 /*
2759 * Insist on a nodelist
2760 */
2761 if (!nodelist)
2762 goto out;
2763 }
2764
2765 mode_flags = 0;
2766 if (flags) {
2767 /*
2768 * Currently, we only support two mutually exclusive
2769 * mode flags.
2770 */
2771 if (!strcmp(flags, "static"))
2772 mode_flags |= MPOL_F_STATIC_NODES;
2773 else if (!strcmp(flags, "relative"))
2774 mode_flags |= MPOL_F_RELATIVE_NODES;
2775 else
2776 goto out;
2777 }
2778
2779 new = mpol_new(mode, mode_flags, &nodes);
2780 if (IS_ERR(new))
2781 goto out;
2782
2783 /*
2784 * Save nodes for mpol_to_str() to show the tmpfs mount options
2785 * for /proc/mounts, /proc/pid/mounts and /proc/pid/mountinfo.
2786 */
2787 if (mode != MPOL_PREFERRED)
2788 new->v.nodes = nodes;
2789 else if (nodelist)
2790 new->v.preferred_node = first_node(nodes);
2791 else
2792 new->flags |= MPOL_F_LOCAL;
2793
2794 /*
2795 * Save nodes for contextualization: this will be used to "clone"
2796 * the mempolicy in a specific context [cpuset] at a later time.
2797 */
2798 new->w.user_nodemask = nodes;
2799
2800 err = 0;
2801
2802out:
2803 /* Restore string for error message */
2804 if (nodelist)
2805 *--nodelist = ':';
2806 if (flags)
2807 *--flags = '=';
2808 if (!err)
2809 *mpol = new;
2810 return err;
2811}
2812#endif /* CONFIG_TMPFS */
2813
2814/**
2815 * mpol_to_str - format a mempolicy structure for printing
2816 * @buffer: to contain formatted mempolicy string
2817 * @maxlen: length of @buffer
2818 * @pol: pointer to mempolicy to be formatted
2819 *
2820 * Convert @pol into a string. If @buffer is too short, truncate the string.
2821 * Recommend a @maxlen of at least 32 for the longest mode, "interleave", the
2822 * longest flag, "relative", and to display at least a few node ids.
2823 */
2824void mpol_to_str(char *buffer, int maxlen, struct mempolicy *pol)
2825{
2826 char *p = buffer;
2827 nodemask_t nodes = NODE_MASK_NONE;
2828 unsigned short mode = MPOL_DEFAULT;
2829 unsigned short flags = 0;
2830
2831 if (pol && pol != &default_policy && !(pol->flags & MPOL_F_MORON)) {
2832 mode = pol->mode;
2833 flags = pol->flags;
2834 }
2835
2836 switch (mode) {
2837 case MPOL_DEFAULT:
2838 break;
2839 case MPOL_PREFERRED:
2840 if (flags & MPOL_F_LOCAL)
2841 mode = MPOL_LOCAL;
2842 else
2843 node_set(pol->v.preferred_node, nodes);
2844 break;
2845 case MPOL_BIND:
2846 case MPOL_INTERLEAVE:
2847 nodes = pol->v.nodes;
2848 break;
2849 default:
2850 WARN_ON_ONCE(1);
2851 snprintf(p, maxlen, "unknown");
2852 return;
2853 }
2854
2855 p += snprintf(p, maxlen, "%s", policy_modes[mode]);
2856
2857 if (flags & MPOL_MODE_FLAGS) {
2858 p += snprintf(p, buffer + maxlen - p, "=");
2859
2860 /*
2861 * Currently, the only defined flags are mutually exclusive
2862 */
2863 if (flags & MPOL_F_STATIC_NODES)
2864 p += snprintf(p, buffer + maxlen - p, "static");
2865 else if (flags & MPOL_F_RELATIVE_NODES)
2866 p += snprintf(p, buffer + maxlen - p, "relative");
2867 }
2868
2869 if (!nodes_empty(nodes)) {
2870 p += snprintf(p, buffer + maxlen - p, ":");
2871 p += nodelist_scnprintf(p, buffer + maxlen - p, nodes);
2872 }
2873}