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