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// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * Simple NUMA memory policy for the Linux kernel.
4 *
5 * Copyright 2003,2004 Andi Kleen, SuSE Labs.
6 * (C) Copyright 2005 Christoph Lameter, Silicon Graphics, Inc.
7 *
8 * NUMA policy allows the user to give hints in which node(s) memory should
9 * be allocated.
10 *
11 * Support four policies per VMA and per process:
12 *
13 * The VMA policy has priority over the process policy for a page fault.
14 *
15 * interleave Allocate memory interleaved over a set of nodes,
16 * with normal fallback if it fails.
17 * For VMA based allocations this interleaves based on the
18 * offset into the backing object or offset into the mapping
19 * for anonymous memory. For process policy an process counter
20 * is used.
21 *
22 * bind Only allocate memory on a specific set of nodes,
23 * no fallback.
24 * FIXME: memory is allocated starting with the first node
25 * to the last. It would be better if bind would truly restrict
26 * the allocation to memory nodes instead
27 *
28 * preferred Try a specific node first before normal fallback.
29 * As a special case NUMA_NO_NODE here means do the allocation
30 * on the local CPU. This is normally identical to default,
31 * but useful to set in a VMA when you have a non default
32 * process policy.
33 *
34 * default Allocate on the local node first, or when on a VMA
35 * use the process policy. This is what Linux always did
36 * in a NUMA aware kernel and still does by, ahem, default.
37 *
38 * The process policy is applied for most non interrupt memory allocations
39 * in that process' context. Interrupts ignore the policies and always
40 * try to allocate on the local CPU. The VMA policy is only applied for memory
41 * allocations for a VMA in the VM.
42 *
43 * Currently there are a few corner cases in swapping where the policy
44 * is not applied, but the majority should be handled. When process policy
45 * is used it is not remembered over swap outs/swap ins.
46 *
47 * Only the highest zone in the zone hierarchy gets policied. Allocations
48 * requesting a lower zone just use default policy. This implies that
49 * on systems with highmem kernel lowmem allocation don't get policied.
50 * Same with GFP_DMA allocations.
51 *
52 * For shmfs/tmpfs/hugetlbfs shared memory the policy is shared between
53 * all users and remembered even when nobody has memory mapped.
54 */
55
56/* Notebook:
57 fix mmap readahead to honour policy and enable policy for any page cache
58 object
59 statistics for bigpages
60 global policy for page cache? currently it uses process policy. Requires
61 first item above.
62 handle mremap for shared memory (currently ignored for the policy)
63 grows down?
64 make bind policy root only? It can trigger oom much faster and the
65 kernel is not always grateful with that.
66*/
67
68#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
69
70#include <linux/mempolicy.h>
71#include <linux/pagewalk.h>
72#include <linux/highmem.h>
73#include <linux/hugetlb.h>
74#include <linux/kernel.h>
75#include <linux/sched.h>
76#include <linux/sched/mm.h>
77#include <linux/sched/numa_balancing.h>
78#include <linux/sched/task.h>
79#include <linux/nodemask.h>
80#include <linux/cpuset.h>
81#include <linux/slab.h>
82#include <linux/string.h>
83#include <linux/export.h>
84#include <linux/nsproxy.h>
85#include <linux/interrupt.h>
86#include <linux/init.h>
87#include <linux/compat.h>
88#include <linux/ptrace.h>
89#include <linux/swap.h>
90#include <linux/seq_file.h>
91#include <linux/proc_fs.h>
92#include <linux/migrate.h>
93#include <linux/ksm.h>
94#include <linux/rmap.h>
95#include <linux/security.h>
96#include <linux/syscalls.h>
97#include <linux/ctype.h>
98#include <linux/mm_inline.h>
99#include <linux/mmu_notifier.h>
100#include <linux/printk.h>
101#include <linux/swapops.h>
102
103#include <asm/tlbflush.h>
104#include <linux/uaccess.h>
105
106#include "internal.h"
107
108/* Internal flags */
109#define MPOL_MF_DISCONTIG_OK (MPOL_MF_INTERNAL << 0) /* Skip checks for continuous vmas */
110#define MPOL_MF_INVERT (MPOL_MF_INTERNAL << 1) /* Invert check for nodemask */
111
112static struct kmem_cache *policy_cache;
113static struct kmem_cache *sn_cache;
114
115/* Highest zone. An specific allocation for a zone below that is not
116 policied. */
117enum zone_type policy_zone = 0;
118
119/*
120 * run-time system-wide default policy => local allocation
121 */
122static struct mempolicy default_policy = {
123 .refcnt = ATOMIC_INIT(1), /* never free it */
124 .mode = MPOL_PREFERRED,
125 .flags = MPOL_F_LOCAL,
126};
127
128static struct mempolicy preferred_node_policy[MAX_NUMNODES];
129
130/**
131 * numa_map_to_online_node - Find closest online node
132 * @node: Node id to start the search
133 *
134 * Lookup the next closest node by distance if @nid is not online.
135 */
136int numa_map_to_online_node(int node)
137{
138 int min_dist = INT_MAX, dist, n, min_node;
139
140 if (node == NUMA_NO_NODE || node_online(node))
141 return node;
142
143 min_node = node;
144 for_each_online_node(n) {
145 dist = node_distance(node, n);
146 if (dist < min_dist) {
147 min_dist = dist;
148 min_node = n;
149 }
150 }
151
152 return min_node;
153}
154EXPORT_SYMBOL_GPL(numa_map_to_online_node);
155
156struct mempolicy *get_task_policy(struct task_struct *p)
157{
158 struct mempolicy *pol = p->mempolicy;
159 int node;
160
161 if (pol)
162 return pol;
163
164 node = numa_node_id();
165 if (node != NUMA_NO_NODE) {
166 pol = &preferred_node_policy[node];
167 /* preferred_node_policy is not initialised early in boot */
168 if (pol->mode)
169 return pol;
170 }
171
172 return &default_policy;
173}
174
175static const struct mempolicy_operations {
176 int (*create)(struct mempolicy *pol, const nodemask_t *nodes);
177 void (*rebind)(struct mempolicy *pol, const nodemask_t *nodes);
178} mpol_ops[MPOL_MAX];
179
180static inline int mpol_store_user_nodemask(const struct mempolicy *pol)
181{
182 return pol->flags & MPOL_MODE_FLAGS;
183}
184
185static void mpol_relative_nodemask(nodemask_t *ret, const nodemask_t *orig,
186 const nodemask_t *rel)
187{
188 nodemask_t tmp;
189 nodes_fold(tmp, *orig, nodes_weight(*rel));
190 nodes_onto(*ret, tmp, *rel);
191}
192
193static int mpol_new_interleave(struct mempolicy *pol, const nodemask_t *nodes)
194{
195 if (nodes_empty(*nodes))
196 return -EINVAL;
197 pol->v.nodes = *nodes;
198 return 0;
199}
200
201static int mpol_new_preferred(struct mempolicy *pol, const nodemask_t *nodes)
202{
203 if (!nodes)
204 pol->flags |= MPOL_F_LOCAL; /* local allocation */
205 else if (nodes_empty(*nodes))
206 return -EINVAL; /* no allowed nodes */
207 else
208 pol->v.preferred_node = first_node(*nodes);
209 return 0;
210}
211
212static int mpol_new_bind(struct mempolicy *pol, const nodemask_t *nodes)
213{
214 if (nodes_empty(*nodes))
215 return -EINVAL;
216 pol->v.nodes = *nodes;
217 return 0;
218}
219
220/*
221 * mpol_set_nodemask is called after mpol_new() to set up the nodemask, if
222 * any, for the new policy. mpol_new() has already validated the nodes
223 * parameter with respect to the policy mode and flags. But, we need to
224 * handle an empty nodemask with MPOL_PREFERRED here.
225 *
226 * Must be called holding task's alloc_lock to protect task's mems_allowed
227 * and mempolicy. May also be called holding the mmap_lock for write.
228 */
229static int mpol_set_nodemask(struct mempolicy *pol,
230 const nodemask_t *nodes, struct nodemask_scratch *nsc)
231{
232 int ret;
233
234 /* if mode is MPOL_DEFAULT, pol is NULL. This is right. */
235 if (pol == NULL)
236 return 0;
237 /* Check N_MEMORY */
238 nodes_and(nsc->mask1,
239 cpuset_current_mems_allowed, node_states[N_MEMORY]);
240
241 VM_BUG_ON(!nodes);
242 if (pol->mode == MPOL_PREFERRED && nodes_empty(*nodes))
243 nodes = NULL; /* explicit local allocation */
244 else {
245 if (pol->flags & MPOL_F_RELATIVE_NODES)
246 mpol_relative_nodemask(&nsc->mask2, nodes, &nsc->mask1);
247 else
248 nodes_and(nsc->mask2, *nodes, nsc->mask1);
249
250 if (mpol_store_user_nodemask(pol))
251 pol->w.user_nodemask = *nodes;
252 else
253 pol->w.cpuset_mems_allowed =
254 cpuset_current_mems_allowed;
255 }
256
257 if (nodes)
258 ret = mpol_ops[pol->mode].create(pol, &nsc->mask2);
259 else
260 ret = mpol_ops[pol->mode].create(pol, NULL);
261 return ret;
262}
263
264/*
265 * This function just creates a new policy, does some check and simple
266 * initialization. You must invoke mpol_set_nodemask() to set nodes.
267 */
268static struct mempolicy *mpol_new(unsigned short mode, unsigned short flags,
269 nodemask_t *nodes)
270{
271 struct mempolicy *policy;
272
273 pr_debug("setting mode %d flags %d nodes[0] %lx\n",
274 mode, flags, nodes ? nodes_addr(*nodes)[0] : NUMA_NO_NODE);
275
276 if (mode == MPOL_DEFAULT) {
277 if (nodes && !nodes_empty(*nodes))
278 return ERR_PTR(-EINVAL);
279 return NULL;
280 }
281 VM_BUG_ON(!nodes);
282
283 /*
284 * MPOL_PREFERRED cannot be used with MPOL_F_STATIC_NODES or
285 * MPOL_F_RELATIVE_NODES if the nodemask is empty (local allocation).
286 * All other modes require a valid pointer to a non-empty nodemask.
287 */
288 if (mode == MPOL_PREFERRED) {
289 if (nodes_empty(*nodes)) {
290 if (((flags & MPOL_F_STATIC_NODES) ||
291 (flags & MPOL_F_RELATIVE_NODES)))
292 return ERR_PTR(-EINVAL);
293 }
294 } else if (mode == MPOL_LOCAL) {
295 if (!nodes_empty(*nodes) ||
296 (flags & MPOL_F_STATIC_NODES) ||
297 (flags & MPOL_F_RELATIVE_NODES))
298 return ERR_PTR(-EINVAL);
299 mode = MPOL_PREFERRED;
300 } else if (nodes_empty(*nodes))
301 return ERR_PTR(-EINVAL);
302 policy = kmem_cache_alloc(policy_cache, GFP_KERNEL);
303 if (!policy)
304 return ERR_PTR(-ENOMEM);
305 atomic_set(&policy->refcnt, 1);
306 policy->mode = mode;
307 policy->flags = flags;
308
309 return policy;
310}
311
312/* Slow path of a mpol destructor. */
313void __mpol_put(struct mempolicy *p)
314{
315 if (!atomic_dec_and_test(&p->refcnt))
316 return;
317 kmem_cache_free(policy_cache, p);
318}
319
320static void mpol_rebind_default(struct mempolicy *pol, const nodemask_t *nodes)
321{
322}
323
324static void mpol_rebind_nodemask(struct mempolicy *pol, const nodemask_t *nodes)
325{
326 nodemask_t tmp;
327
328 if (pol->flags & MPOL_F_STATIC_NODES)
329 nodes_and(tmp, pol->w.user_nodemask, *nodes);
330 else if (pol->flags & MPOL_F_RELATIVE_NODES)
331 mpol_relative_nodemask(&tmp, &pol->w.user_nodemask, nodes);
332 else {
333 nodes_remap(tmp, pol->v.nodes,pol->w.cpuset_mems_allowed,
334 *nodes);
335 pol->w.cpuset_mems_allowed = *nodes;
336 }
337
338 if (nodes_empty(tmp))
339 tmp = *nodes;
340
341 pol->v.nodes = tmp;
342}
343
344static void mpol_rebind_preferred(struct mempolicy *pol,
345 const nodemask_t *nodes)
346{
347 nodemask_t tmp;
348
349 if (pol->flags & MPOL_F_STATIC_NODES) {
350 int node = first_node(pol->w.user_nodemask);
351
352 if (node_isset(node, *nodes)) {
353 pol->v.preferred_node = node;
354 pol->flags &= ~MPOL_F_LOCAL;
355 } else
356 pol->flags |= MPOL_F_LOCAL;
357 } else if (pol->flags & MPOL_F_RELATIVE_NODES) {
358 mpol_relative_nodemask(&tmp, &pol->w.user_nodemask, nodes);
359 pol->v.preferred_node = first_node(tmp);
360 } else if (!(pol->flags & MPOL_F_LOCAL)) {
361 pol->v.preferred_node = node_remap(pol->v.preferred_node,
362 pol->w.cpuset_mems_allowed,
363 *nodes);
364 pol->w.cpuset_mems_allowed = *nodes;
365 }
366}
367
368/*
369 * mpol_rebind_policy - Migrate a policy to a different set of nodes
370 *
371 * Per-vma policies are protected by mmap_lock. Allocations using per-task
372 * policies are protected by task->mems_allowed_seq to prevent a premature
373 * OOM/allocation failure due to parallel nodemask modification.
374 */
375static void mpol_rebind_policy(struct mempolicy *pol, const nodemask_t *newmask)
376{
377 if (!pol)
378 return;
379 if (!mpol_store_user_nodemask(pol) && !(pol->flags & MPOL_F_LOCAL) &&
380 nodes_equal(pol->w.cpuset_mems_allowed, *newmask))
381 return;
382
383 mpol_ops[pol->mode].rebind(pol, newmask);
384}
385
386/*
387 * Wrapper for mpol_rebind_policy() that just requires task
388 * pointer, and updates task mempolicy.
389 *
390 * Called with task's alloc_lock held.
391 */
392
393void mpol_rebind_task(struct task_struct *tsk, const nodemask_t *new)
394{
395 mpol_rebind_policy(tsk->mempolicy, new);
396}
397
398/*
399 * Rebind each vma in mm to new nodemask.
400 *
401 * Call holding a reference to mm. Takes mm->mmap_lock during call.
402 */
403
404void mpol_rebind_mm(struct mm_struct *mm, nodemask_t *new)
405{
406 struct vm_area_struct *vma;
407
408 mmap_write_lock(mm);
409 for (vma = mm->mmap; vma; vma = vma->vm_next)
410 mpol_rebind_policy(vma->vm_policy, new);
411 mmap_write_unlock(mm);
412}
413
414static const struct mempolicy_operations mpol_ops[MPOL_MAX] = {
415 [MPOL_DEFAULT] = {
416 .rebind = mpol_rebind_default,
417 },
418 [MPOL_INTERLEAVE] = {
419 .create = mpol_new_interleave,
420 .rebind = mpol_rebind_nodemask,
421 },
422 [MPOL_PREFERRED] = {
423 .create = mpol_new_preferred,
424 .rebind = mpol_rebind_preferred,
425 },
426 [MPOL_BIND] = {
427 .create = mpol_new_bind,
428 .rebind = mpol_rebind_nodemask,
429 },
430};
431
432static int migrate_page_add(struct page *page, struct list_head *pagelist,
433 unsigned long flags);
434
435struct queue_pages {
436 struct list_head *pagelist;
437 unsigned long flags;
438 nodemask_t *nmask;
439 unsigned long start;
440 unsigned long end;
441 struct vm_area_struct *first;
442};
443
444/*
445 * Check if the page's nid is in qp->nmask.
446 *
447 * If MPOL_MF_INVERT is set in qp->flags, check if the nid is
448 * in the invert of qp->nmask.
449 */
450static inline bool queue_pages_required(struct page *page,
451 struct queue_pages *qp)
452{
453 int nid = page_to_nid(page);
454 unsigned long flags = qp->flags;
455
456 return node_isset(nid, *qp->nmask) == !(flags & MPOL_MF_INVERT);
457}
458
459/*
460 * queue_pages_pmd() has four possible return values:
461 * 0 - pages are placed on the right node or queued successfully.
462 * 1 - there is unmovable page, and MPOL_MF_MOVE* & MPOL_MF_STRICT were
463 * specified.
464 * 2 - THP was split.
465 * -EIO - is migration entry or only MPOL_MF_STRICT was specified and an
466 * existing page was already on a node that does not follow the
467 * policy.
468 */
469static int queue_pages_pmd(pmd_t *pmd, spinlock_t *ptl, unsigned long addr,
470 unsigned long end, struct mm_walk *walk)
471 __releases(ptl)
472{
473 int ret = 0;
474 struct page *page;
475 struct queue_pages *qp = walk->private;
476 unsigned long flags;
477
478 if (unlikely(is_pmd_migration_entry(*pmd))) {
479 ret = -EIO;
480 goto unlock;
481 }
482 page = pmd_page(*pmd);
483 if (is_huge_zero_page(page)) {
484 spin_unlock(ptl);
485 __split_huge_pmd(walk->vma, pmd, addr, false, NULL);
486 ret = 2;
487 goto out;
488 }
489 if (!queue_pages_required(page, qp))
490 goto unlock;
491
492 flags = qp->flags;
493 /* go to thp migration */
494 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) {
495 if (!vma_migratable(walk->vma) ||
496 migrate_page_add(page, qp->pagelist, flags)) {
497 ret = 1;
498 goto unlock;
499 }
500 } else
501 ret = -EIO;
502unlock:
503 spin_unlock(ptl);
504out:
505 return ret;
506}
507
508/*
509 * Scan through pages checking if pages follow certain conditions,
510 * and move them to the pagelist if they do.
511 *
512 * queue_pages_pte_range() has three possible return values:
513 * 0 - pages are placed on the right node or queued successfully.
514 * 1 - there is unmovable page, and MPOL_MF_MOVE* & MPOL_MF_STRICT were
515 * specified.
516 * -EIO - only MPOL_MF_STRICT was specified and an existing page was already
517 * on a node that does not follow the policy.
518 */
519static int queue_pages_pte_range(pmd_t *pmd, unsigned long addr,
520 unsigned long end, struct mm_walk *walk)
521{
522 struct vm_area_struct *vma = walk->vma;
523 struct page *page;
524 struct queue_pages *qp = walk->private;
525 unsigned long flags = qp->flags;
526 int ret;
527 bool has_unmovable = false;
528 pte_t *pte;
529 spinlock_t *ptl;
530
531 ptl = pmd_trans_huge_lock(pmd, vma);
532 if (ptl) {
533 ret = queue_pages_pmd(pmd, ptl, addr, end, walk);
534 if (ret != 2)
535 return ret;
536 }
537 /* THP was split, fall through to pte walk */
538
539 if (pmd_trans_unstable(pmd))
540 return 0;
541
542 pte = pte_offset_map_lock(walk->mm, pmd, addr, &ptl);
543 for (; addr != end; pte++, addr += PAGE_SIZE) {
544 if (!pte_present(*pte))
545 continue;
546 page = vm_normal_page(vma, addr, *pte);
547 if (!page)
548 continue;
549 /*
550 * vm_normal_page() filters out zero pages, but there might
551 * still be PageReserved pages to skip, perhaps in a VDSO.
552 */
553 if (PageReserved(page))
554 continue;
555 if (!queue_pages_required(page, qp))
556 continue;
557 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) {
558 /* MPOL_MF_STRICT must be specified if we get here */
559 if (!vma_migratable(vma)) {
560 has_unmovable = true;
561 break;
562 }
563
564 /*
565 * Do not abort immediately since there may be
566 * temporary off LRU pages in the range. Still
567 * need migrate other LRU pages.
568 */
569 if (migrate_page_add(page, qp->pagelist, flags))
570 has_unmovable = true;
571 } else
572 break;
573 }
574 pte_unmap_unlock(pte - 1, ptl);
575 cond_resched();
576
577 if (has_unmovable)
578 return 1;
579
580 return addr != end ? -EIO : 0;
581}
582
583static int queue_pages_hugetlb(pte_t *pte, unsigned long hmask,
584 unsigned long addr, unsigned long end,
585 struct mm_walk *walk)
586{
587 int ret = 0;
588#ifdef CONFIG_HUGETLB_PAGE
589 struct queue_pages *qp = walk->private;
590 unsigned long flags = (qp->flags & MPOL_MF_VALID);
591 struct page *page;
592 spinlock_t *ptl;
593 pte_t entry;
594
595 ptl = huge_pte_lock(hstate_vma(walk->vma), walk->mm, pte);
596 entry = huge_ptep_get(pte);
597 if (!pte_present(entry))
598 goto unlock;
599 page = pte_page(entry);
600 if (!queue_pages_required(page, qp))
601 goto unlock;
602
603 if (flags == MPOL_MF_STRICT) {
604 /*
605 * STRICT alone means only detecting misplaced page and no
606 * need to further check other vma.
607 */
608 ret = -EIO;
609 goto unlock;
610 }
611
612 if (!vma_migratable(walk->vma)) {
613 /*
614 * Must be STRICT with MOVE*, otherwise .test_walk() have
615 * stopped walking current vma.
616 * Detecting misplaced page but allow migrating pages which
617 * have been queued.
618 */
619 ret = 1;
620 goto unlock;
621 }
622
623 /* With MPOL_MF_MOVE, we migrate only unshared hugepage. */
624 if (flags & (MPOL_MF_MOVE_ALL) ||
625 (flags & MPOL_MF_MOVE && page_mapcount(page) == 1)) {
626 if (!isolate_huge_page(page, qp->pagelist) &&
627 (flags & MPOL_MF_STRICT))
628 /*
629 * Failed to isolate page but allow migrating pages
630 * which have been queued.
631 */
632 ret = 1;
633 }
634unlock:
635 spin_unlock(ptl);
636#else
637 BUG();
638#endif
639 return ret;
640}
641
642#ifdef CONFIG_NUMA_BALANCING
643/*
644 * This is used to mark a range of virtual addresses to be inaccessible.
645 * These are later cleared by a NUMA hinting fault. Depending on these
646 * faults, pages may be migrated for better NUMA placement.
647 *
648 * This is assuming that NUMA faults are handled using PROT_NONE. If
649 * an architecture makes a different choice, it will need further
650 * changes to the core.
651 */
652unsigned long change_prot_numa(struct vm_area_struct *vma,
653 unsigned long addr, unsigned long end)
654{
655 int nr_updated;
656
657 nr_updated = change_protection(vma, addr, end, PAGE_NONE, MM_CP_PROT_NUMA);
658 if (nr_updated)
659 count_vm_numa_events(NUMA_PTE_UPDATES, nr_updated);
660
661 return nr_updated;
662}
663#else
664static unsigned long change_prot_numa(struct vm_area_struct *vma,
665 unsigned long addr, unsigned long end)
666{
667 return 0;
668}
669#endif /* CONFIG_NUMA_BALANCING */
670
671static int queue_pages_test_walk(unsigned long start, unsigned long end,
672 struct mm_walk *walk)
673{
674 struct vm_area_struct *vma = walk->vma;
675 struct queue_pages *qp = walk->private;
676 unsigned long endvma = vma->vm_end;
677 unsigned long flags = qp->flags;
678
679 /* range check first */
680 VM_BUG_ON_VMA((vma->vm_start > start) || (vma->vm_end < end), vma);
681
682 if (!qp->first) {
683 qp->first = vma;
684 if (!(flags & MPOL_MF_DISCONTIG_OK) &&
685 (qp->start < vma->vm_start))
686 /* hole at head side of range */
687 return -EFAULT;
688 }
689 if (!(flags & MPOL_MF_DISCONTIG_OK) &&
690 ((vma->vm_end < qp->end) &&
691 (!vma->vm_next || vma->vm_end < vma->vm_next->vm_start)))
692 /* hole at middle or tail of range */
693 return -EFAULT;
694
695 /*
696 * Need check MPOL_MF_STRICT to return -EIO if possible
697 * regardless of vma_migratable
698 */
699 if (!vma_migratable(vma) &&
700 !(flags & MPOL_MF_STRICT))
701 return 1;
702
703 if (endvma > end)
704 endvma = end;
705
706 if (flags & MPOL_MF_LAZY) {
707 /* Similar to task_numa_work, skip inaccessible VMAs */
708 if (!is_vm_hugetlb_page(vma) && vma_is_accessible(vma) &&
709 !(vma->vm_flags & VM_MIXEDMAP))
710 change_prot_numa(vma, start, endvma);
711 return 1;
712 }
713
714 /* queue pages from current vma */
715 if (flags & MPOL_MF_VALID)
716 return 0;
717 return 1;
718}
719
720static const struct mm_walk_ops queue_pages_walk_ops = {
721 .hugetlb_entry = queue_pages_hugetlb,
722 .pmd_entry = queue_pages_pte_range,
723 .test_walk = queue_pages_test_walk,
724};
725
726/*
727 * Walk through page tables and collect pages to be migrated.
728 *
729 * If pages found in a given range are on a set of nodes (determined by
730 * @nodes and @flags,) it's isolated and queued to the pagelist which is
731 * passed via @private.
732 *
733 * queue_pages_range() has three possible return values:
734 * 1 - there is unmovable page, but MPOL_MF_MOVE* & MPOL_MF_STRICT were
735 * specified.
736 * 0 - queue pages successfully or no misplaced page.
737 * errno - i.e. misplaced pages with MPOL_MF_STRICT specified (-EIO) or
738 * memory range specified by nodemask and maxnode points outside
739 * your accessible address space (-EFAULT)
740 */
741static int
742queue_pages_range(struct mm_struct *mm, unsigned long start, unsigned long end,
743 nodemask_t *nodes, unsigned long flags,
744 struct list_head *pagelist)
745{
746 int err;
747 struct queue_pages qp = {
748 .pagelist = pagelist,
749 .flags = flags,
750 .nmask = nodes,
751 .start = start,
752 .end = end,
753 .first = NULL,
754 };
755
756 err = walk_page_range(mm, start, end, &queue_pages_walk_ops, &qp);
757
758 if (!qp.first)
759 /* whole range in hole */
760 err = -EFAULT;
761
762 return err;
763}
764
765/*
766 * Apply policy to a single VMA
767 * This must be called with the mmap_lock held for writing.
768 */
769static int vma_replace_policy(struct vm_area_struct *vma,
770 struct mempolicy *pol)
771{
772 int err;
773 struct mempolicy *old;
774 struct mempolicy *new;
775
776 pr_debug("vma %lx-%lx/%lx vm_ops %p vm_file %p set_policy %p\n",
777 vma->vm_start, vma->vm_end, vma->vm_pgoff,
778 vma->vm_ops, vma->vm_file,
779 vma->vm_ops ? vma->vm_ops->set_policy : NULL);
780
781 new = mpol_dup(pol);
782 if (IS_ERR(new))
783 return PTR_ERR(new);
784
785 if (vma->vm_ops && vma->vm_ops->set_policy) {
786 err = vma->vm_ops->set_policy(vma, new);
787 if (err)
788 goto err_out;
789 }
790
791 old = vma->vm_policy;
792 vma->vm_policy = new; /* protected by mmap_lock */
793 mpol_put(old);
794
795 return 0;
796 err_out:
797 mpol_put(new);
798 return err;
799}
800
801/* Step 2: apply policy to a range and do splits. */
802static int mbind_range(struct mm_struct *mm, unsigned long start,
803 unsigned long end, struct mempolicy *new_pol)
804{
805 struct vm_area_struct *next;
806 struct vm_area_struct *prev;
807 struct vm_area_struct *vma;
808 int err = 0;
809 pgoff_t pgoff;
810 unsigned long vmstart;
811 unsigned long vmend;
812
813 vma = find_vma(mm, start);
814 VM_BUG_ON(!vma);
815
816 prev = vma->vm_prev;
817 if (start > vma->vm_start)
818 prev = vma;
819
820 for (; vma && vma->vm_start < end; prev = vma, vma = next) {
821 next = vma->vm_next;
822 vmstart = max(start, vma->vm_start);
823 vmend = min(end, vma->vm_end);
824
825 if (mpol_equal(vma_policy(vma), new_pol))
826 continue;
827
828 pgoff = vma->vm_pgoff +
829 ((vmstart - vma->vm_start) >> PAGE_SHIFT);
830 prev = vma_merge(mm, prev, vmstart, vmend, vma->vm_flags,
831 vma->anon_vma, vma->vm_file, pgoff,
832 new_pol, vma->vm_userfaultfd_ctx);
833 if (prev) {
834 vma = prev;
835 next = vma->vm_next;
836 if (mpol_equal(vma_policy(vma), new_pol))
837 continue;
838 /* vma_merge() joined vma && vma->next, case 8 */
839 goto replace;
840 }
841 if (vma->vm_start != vmstart) {
842 err = split_vma(vma->vm_mm, vma, vmstart, 1);
843 if (err)
844 goto out;
845 }
846 if (vma->vm_end != vmend) {
847 err = split_vma(vma->vm_mm, vma, vmend, 0);
848 if (err)
849 goto out;
850 }
851 replace:
852 err = vma_replace_policy(vma, new_pol);
853 if (err)
854 goto out;
855 }
856
857 out:
858 return err;
859}
860
861/* Set the process memory policy */
862static long do_set_mempolicy(unsigned short mode, unsigned short flags,
863 nodemask_t *nodes)
864{
865 struct mempolicy *new, *old;
866 NODEMASK_SCRATCH(scratch);
867 int ret;
868
869 if (!scratch)
870 return -ENOMEM;
871
872 new = mpol_new(mode, flags, nodes);
873 if (IS_ERR(new)) {
874 ret = PTR_ERR(new);
875 goto out;
876 }
877
878 task_lock(current);
879 ret = mpol_set_nodemask(new, nodes, scratch);
880 if (ret) {
881 task_unlock(current);
882 mpol_put(new);
883 goto out;
884 }
885 old = current->mempolicy;
886 current->mempolicy = new;
887 if (new && new->mode == MPOL_INTERLEAVE)
888 current->il_prev = MAX_NUMNODES-1;
889 task_unlock(current);
890 mpol_put(old);
891 ret = 0;
892out:
893 NODEMASK_SCRATCH_FREE(scratch);
894 return ret;
895}
896
897/*
898 * Return nodemask for policy for get_mempolicy() query
899 *
900 * Called with task's alloc_lock held
901 */
902static void get_policy_nodemask(struct mempolicy *p, nodemask_t *nodes)
903{
904 nodes_clear(*nodes);
905 if (p == &default_policy)
906 return;
907
908 switch (p->mode) {
909 case MPOL_BIND:
910 case MPOL_INTERLEAVE:
911 *nodes = p->v.nodes;
912 break;
913 case MPOL_PREFERRED:
914 if (!(p->flags & MPOL_F_LOCAL))
915 node_set(p->v.preferred_node, *nodes);
916 /* else return empty node mask for local allocation */
917 break;
918 default:
919 BUG();
920 }
921}
922
923static int lookup_node(struct mm_struct *mm, unsigned long addr)
924{
925 struct page *p = NULL;
926 int err;
927
928 int locked = 1;
929 err = get_user_pages_locked(addr & PAGE_MASK, 1, 0, &p, &locked);
930 if (err > 0) {
931 err = page_to_nid(p);
932 put_page(p);
933 }
934 if (locked)
935 mmap_read_unlock(mm);
936 return err;
937}
938
939/* Retrieve NUMA policy */
940static long do_get_mempolicy(int *policy, nodemask_t *nmask,
941 unsigned long addr, unsigned long flags)
942{
943 int err;
944 struct mm_struct *mm = current->mm;
945 struct vm_area_struct *vma = NULL;
946 struct mempolicy *pol = current->mempolicy, *pol_refcount = NULL;
947
948 if (flags &
949 ~(unsigned long)(MPOL_F_NODE|MPOL_F_ADDR|MPOL_F_MEMS_ALLOWED))
950 return -EINVAL;
951
952 if (flags & MPOL_F_MEMS_ALLOWED) {
953 if (flags & (MPOL_F_NODE|MPOL_F_ADDR))
954 return -EINVAL;
955 *policy = 0; /* just so it's initialized */
956 task_lock(current);
957 *nmask = cpuset_current_mems_allowed;
958 task_unlock(current);
959 return 0;
960 }
961
962 if (flags & MPOL_F_ADDR) {
963 /*
964 * Do NOT fall back to task policy if the
965 * vma/shared policy at addr is NULL. We
966 * want to return MPOL_DEFAULT in this case.
967 */
968 mmap_read_lock(mm);
969 vma = find_vma_intersection(mm, addr, addr+1);
970 if (!vma) {
971 mmap_read_unlock(mm);
972 return -EFAULT;
973 }
974 if (vma->vm_ops && vma->vm_ops->get_policy)
975 pol = vma->vm_ops->get_policy(vma, addr);
976 else
977 pol = vma->vm_policy;
978 } else if (addr)
979 return -EINVAL;
980
981 if (!pol)
982 pol = &default_policy; /* indicates default behavior */
983
984 if (flags & MPOL_F_NODE) {
985 if (flags & MPOL_F_ADDR) {
986 /*
987 * Take a refcount on the mpol, lookup_node()
988 * wil drop the mmap_lock, so after calling
989 * lookup_node() only "pol" remains valid, "vma"
990 * is stale.
991 */
992 pol_refcount = pol;
993 vma = NULL;
994 mpol_get(pol);
995 err = lookup_node(mm, addr);
996 if (err < 0)
997 goto out;
998 *policy = err;
999 } else if (pol == current->mempolicy &&
1000 pol->mode == MPOL_INTERLEAVE) {
1001 *policy = next_node_in(current->il_prev, pol->v.nodes);
1002 } else {
1003 err = -EINVAL;
1004 goto out;
1005 }
1006 } else {
1007 *policy = pol == &default_policy ? MPOL_DEFAULT :
1008 pol->mode;
1009 /*
1010 * Internal mempolicy flags must be masked off before exposing
1011 * the policy to userspace.
1012 */
1013 *policy |= (pol->flags & MPOL_MODE_FLAGS);
1014 }
1015
1016 err = 0;
1017 if (nmask) {
1018 if (mpol_store_user_nodemask(pol)) {
1019 *nmask = pol->w.user_nodemask;
1020 } else {
1021 task_lock(current);
1022 get_policy_nodemask(pol, nmask);
1023 task_unlock(current);
1024 }
1025 }
1026
1027 out:
1028 mpol_cond_put(pol);
1029 if (vma)
1030 mmap_read_unlock(mm);
1031 if (pol_refcount)
1032 mpol_put(pol_refcount);
1033 return err;
1034}
1035
1036#ifdef CONFIG_MIGRATION
1037/*
1038 * page migration, thp tail pages can be passed.
1039 */
1040static int migrate_page_add(struct page *page, struct list_head *pagelist,
1041 unsigned long flags)
1042{
1043 struct page *head = compound_head(page);
1044 /*
1045 * Avoid migrating a page that is shared with others.
1046 */
1047 if ((flags & MPOL_MF_MOVE_ALL) || page_mapcount(head) == 1) {
1048 if (!isolate_lru_page(head)) {
1049 list_add_tail(&head->lru, pagelist);
1050 mod_node_page_state(page_pgdat(head),
1051 NR_ISOLATED_ANON + page_is_file_lru(head),
1052 thp_nr_pages(head));
1053 } else if (flags & MPOL_MF_STRICT) {
1054 /*
1055 * Non-movable page may reach here. And, there may be
1056 * temporary off LRU pages or non-LRU movable pages.
1057 * Treat them as unmovable pages since they can't be
1058 * isolated, so they can't be moved at the moment. It
1059 * should return -EIO for this case too.
1060 */
1061 return -EIO;
1062 }
1063 }
1064
1065 return 0;
1066}
1067
1068/*
1069 * Migrate pages from one node to a target node.
1070 * Returns error or the number of pages not migrated.
1071 */
1072static int migrate_to_node(struct mm_struct *mm, int source, int dest,
1073 int flags)
1074{
1075 nodemask_t nmask;
1076 LIST_HEAD(pagelist);
1077 int err = 0;
1078 struct migration_target_control mtc = {
1079 .nid = dest,
1080 .gfp_mask = GFP_HIGHUSER_MOVABLE | __GFP_THISNODE,
1081 };
1082
1083 nodes_clear(nmask);
1084 node_set(source, nmask);
1085
1086 /*
1087 * This does not "check" the range but isolates all pages that
1088 * need migration. Between passing in the full user address
1089 * space range and MPOL_MF_DISCONTIG_OK, this call can not fail.
1090 */
1091 VM_BUG_ON(!(flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)));
1092 queue_pages_range(mm, mm->mmap->vm_start, mm->task_size, &nmask,
1093 flags | MPOL_MF_DISCONTIG_OK, &pagelist);
1094
1095 if (!list_empty(&pagelist)) {
1096 err = migrate_pages(&pagelist, alloc_migration_target, NULL,
1097 (unsigned long)&mtc, MIGRATE_SYNC, MR_SYSCALL);
1098 if (err)
1099 putback_movable_pages(&pagelist);
1100 }
1101
1102 return err;
1103}
1104
1105/*
1106 * Move pages between the two nodesets so as to preserve the physical
1107 * layout as much as possible.
1108 *
1109 * Returns the number of page that could not be moved.
1110 */
1111int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
1112 const nodemask_t *to, int flags)
1113{
1114 int busy = 0;
1115 int err;
1116 nodemask_t tmp;
1117
1118 err = migrate_prep();
1119 if (err)
1120 return err;
1121
1122 mmap_read_lock(mm);
1123
1124 /*
1125 * Find a 'source' bit set in 'tmp' whose corresponding 'dest'
1126 * bit in 'to' is not also set in 'tmp'. Clear the found 'source'
1127 * bit in 'tmp', and return that <source, dest> pair for migration.
1128 * The pair of nodemasks 'to' and 'from' define the map.
1129 *
1130 * If no pair of bits is found that way, fallback to picking some
1131 * pair of 'source' and 'dest' bits that are not the same. If the
1132 * 'source' and 'dest' bits are the same, this represents a node
1133 * that will be migrating to itself, so no pages need move.
1134 *
1135 * If no bits are left in 'tmp', or if all remaining bits left
1136 * in 'tmp' correspond to the same bit in 'to', return false
1137 * (nothing left to migrate).
1138 *
1139 * This lets us pick a pair of nodes to migrate between, such that
1140 * if possible the dest node is not already occupied by some other
1141 * source node, minimizing the risk of overloading the memory on a
1142 * node that would happen if we migrated incoming memory to a node
1143 * before migrating outgoing memory source that same node.
1144 *
1145 * A single scan of tmp is sufficient. As we go, we remember the
1146 * most recent <s, d> pair that moved (s != d). If we find a pair
1147 * that not only moved, but what's better, moved to an empty slot
1148 * (d is not set in tmp), then we break out then, with that pair.
1149 * Otherwise when we finish scanning from_tmp, we at least have the
1150 * most recent <s, d> pair that moved. If we get all the way through
1151 * the scan of tmp without finding any node that moved, much less
1152 * moved to an empty node, then there is nothing left worth migrating.
1153 */
1154
1155 tmp = *from;
1156 while (!nodes_empty(tmp)) {
1157 int s,d;
1158 int source = NUMA_NO_NODE;
1159 int dest = 0;
1160
1161 for_each_node_mask(s, tmp) {
1162
1163 /*
1164 * do_migrate_pages() tries to maintain the relative
1165 * node relationship of the pages established between
1166 * threads and memory areas.
1167 *
1168 * However if the number of source nodes is not equal to
1169 * the number of destination nodes we can not preserve
1170 * this node relative relationship. In that case, skip
1171 * copying memory from a node that is in the destination
1172 * mask.
1173 *
1174 * Example: [2,3,4] -> [3,4,5] moves everything.
1175 * [0-7] - > [3,4,5] moves only 0,1,2,6,7.
1176 */
1177
1178 if ((nodes_weight(*from) != nodes_weight(*to)) &&
1179 (node_isset(s, *to)))
1180 continue;
1181
1182 d = node_remap(s, *from, *to);
1183 if (s == d)
1184 continue;
1185
1186 source = s; /* Node moved. Memorize */
1187 dest = d;
1188
1189 /* dest not in remaining from nodes? */
1190 if (!node_isset(dest, tmp))
1191 break;
1192 }
1193 if (source == NUMA_NO_NODE)
1194 break;
1195
1196 node_clear(source, tmp);
1197 err = migrate_to_node(mm, source, dest, flags);
1198 if (err > 0)
1199 busy += err;
1200 if (err < 0)
1201 break;
1202 }
1203 mmap_read_unlock(mm);
1204 if (err < 0)
1205 return err;
1206 return busy;
1207
1208}
1209
1210/*
1211 * Allocate a new page for page migration based on vma policy.
1212 * Start by assuming the page is mapped by the same vma as contains @start.
1213 * Search forward from there, if not. N.B., this assumes that the
1214 * list of pages handed to migrate_pages()--which is how we get here--
1215 * is in virtual address order.
1216 */
1217static struct page *new_page(struct page *page, unsigned long start)
1218{
1219 struct vm_area_struct *vma;
1220 unsigned long address;
1221
1222 vma = find_vma(current->mm, start);
1223 while (vma) {
1224 address = page_address_in_vma(page, vma);
1225 if (address != -EFAULT)
1226 break;
1227 vma = vma->vm_next;
1228 }
1229
1230 if (PageHuge(page)) {
1231 return alloc_huge_page_vma(page_hstate(compound_head(page)),
1232 vma, address);
1233 } else if (PageTransHuge(page)) {
1234 struct page *thp;
1235
1236 thp = alloc_hugepage_vma(GFP_TRANSHUGE, vma, address,
1237 HPAGE_PMD_ORDER);
1238 if (!thp)
1239 return NULL;
1240 prep_transhuge_page(thp);
1241 return thp;
1242 }
1243 /*
1244 * if !vma, alloc_page_vma() will use task or system default policy
1245 */
1246 return alloc_page_vma(GFP_HIGHUSER_MOVABLE | __GFP_RETRY_MAYFAIL,
1247 vma, address);
1248}
1249#else
1250
1251static int migrate_page_add(struct page *page, struct list_head *pagelist,
1252 unsigned long flags)
1253{
1254 return -EIO;
1255}
1256
1257int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
1258 const nodemask_t *to, int flags)
1259{
1260 return -ENOSYS;
1261}
1262
1263static struct page *new_page(struct page *page, unsigned long start)
1264{
1265 return NULL;
1266}
1267#endif
1268
1269static long do_mbind(unsigned long start, unsigned long len,
1270 unsigned short mode, unsigned short mode_flags,
1271 nodemask_t *nmask, unsigned long flags)
1272{
1273 struct mm_struct *mm = current->mm;
1274 struct mempolicy *new;
1275 unsigned long end;
1276 int err;
1277 int ret;
1278 LIST_HEAD(pagelist);
1279
1280 if (flags & ~(unsigned long)MPOL_MF_VALID)
1281 return -EINVAL;
1282 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1283 return -EPERM;
1284
1285 if (start & ~PAGE_MASK)
1286 return -EINVAL;
1287
1288 if (mode == MPOL_DEFAULT)
1289 flags &= ~MPOL_MF_STRICT;
1290
1291 len = (len + PAGE_SIZE - 1) & PAGE_MASK;
1292 end = start + len;
1293
1294 if (end < start)
1295 return -EINVAL;
1296 if (end == start)
1297 return 0;
1298
1299 new = mpol_new(mode, mode_flags, nmask);
1300 if (IS_ERR(new))
1301 return PTR_ERR(new);
1302
1303 if (flags & MPOL_MF_LAZY)
1304 new->flags |= MPOL_F_MOF;
1305
1306 /*
1307 * If we are using the default policy then operation
1308 * on discontinuous address spaces is okay after all
1309 */
1310 if (!new)
1311 flags |= MPOL_MF_DISCONTIG_OK;
1312
1313 pr_debug("mbind %lx-%lx mode:%d flags:%d nodes:%lx\n",
1314 start, start + len, mode, mode_flags,
1315 nmask ? nodes_addr(*nmask)[0] : NUMA_NO_NODE);
1316
1317 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) {
1318
1319 err = migrate_prep();
1320 if (err)
1321 goto mpol_out;
1322 }
1323 {
1324 NODEMASK_SCRATCH(scratch);
1325 if (scratch) {
1326 mmap_write_lock(mm);
1327 task_lock(current);
1328 err = mpol_set_nodemask(new, nmask, scratch);
1329 task_unlock(current);
1330 if (err)
1331 mmap_write_unlock(mm);
1332 } else
1333 err = -ENOMEM;
1334 NODEMASK_SCRATCH_FREE(scratch);
1335 }
1336 if (err)
1337 goto mpol_out;
1338
1339 ret = queue_pages_range(mm, start, end, nmask,
1340 flags | MPOL_MF_INVERT, &pagelist);
1341
1342 if (ret < 0) {
1343 err = ret;
1344 goto up_out;
1345 }
1346
1347 err = mbind_range(mm, start, end, new);
1348
1349 if (!err) {
1350 int nr_failed = 0;
1351
1352 if (!list_empty(&pagelist)) {
1353 WARN_ON_ONCE(flags & MPOL_MF_LAZY);
1354 nr_failed = migrate_pages(&pagelist, new_page, NULL,
1355 start, MIGRATE_SYNC, MR_MEMPOLICY_MBIND);
1356 if (nr_failed)
1357 putback_movable_pages(&pagelist);
1358 }
1359
1360 if ((ret > 0) || (nr_failed && (flags & MPOL_MF_STRICT)))
1361 err = -EIO;
1362 } else {
1363up_out:
1364 if (!list_empty(&pagelist))
1365 putback_movable_pages(&pagelist);
1366 }
1367
1368 mmap_write_unlock(mm);
1369mpol_out:
1370 mpol_put(new);
1371 return err;
1372}
1373
1374/*
1375 * User space interface with variable sized bitmaps for nodelists.
1376 */
1377
1378/* Copy a node mask from user space. */
1379static int get_nodes(nodemask_t *nodes, const unsigned long __user *nmask,
1380 unsigned long maxnode)
1381{
1382 unsigned long k;
1383 unsigned long t;
1384 unsigned long nlongs;
1385 unsigned long endmask;
1386
1387 --maxnode;
1388 nodes_clear(*nodes);
1389 if (maxnode == 0 || !nmask)
1390 return 0;
1391 if (maxnode > PAGE_SIZE*BITS_PER_BYTE)
1392 return -EINVAL;
1393
1394 nlongs = BITS_TO_LONGS(maxnode);
1395 if ((maxnode % BITS_PER_LONG) == 0)
1396 endmask = ~0UL;
1397 else
1398 endmask = (1UL << (maxnode % BITS_PER_LONG)) - 1;
1399
1400 /*
1401 * When the user specified more nodes than supported just check
1402 * if the non supported part is all zero.
1403 *
1404 * If maxnode have more longs than MAX_NUMNODES, check
1405 * the bits in that area first. And then go through to
1406 * check the rest bits which equal or bigger than MAX_NUMNODES.
1407 * Otherwise, just check bits [MAX_NUMNODES, maxnode).
1408 */
1409 if (nlongs > BITS_TO_LONGS(MAX_NUMNODES)) {
1410 for (k = BITS_TO_LONGS(MAX_NUMNODES); k < nlongs; k++) {
1411 if (get_user(t, nmask + k))
1412 return -EFAULT;
1413 if (k == nlongs - 1) {
1414 if (t & endmask)
1415 return -EINVAL;
1416 } else if (t)
1417 return -EINVAL;
1418 }
1419 nlongs = BITS_TO_LONGS(MAX_NUMNODES);
1420 endmask = ~0UL;
1421 }
1422
1423 if (maxnode > MAX_NUMNODES && MAX_NUMNODES % BITS_PER_LONG != 0) {
1424 unsigned long valid_mask = endmask;
1425
1426 valid_mask &= ~((1UL << (MAX_NUMNODES % BITS_PER_LONG)) - 1);
1427 if (get_user(t, nmask + nlongs - 1))
1428 return -EFAULT;
1429 if (t & valid_mask)
1430 return -EINVAL;
1431 }
1432
1433 if (copy_from_user(nodes_addr(*nodes), nmask, nlongs*sizeof(unsigned long)))
1434 return -EFAULT;
1435 nodes_addr(*nodes)[nlongs-1] &= endmask;
1436 return 0;
1437}
1438
1439/* Copy a kernel node mask to user space */
1440static int copy_nodes_to_user(unsigned long __user *mask, unsigned long maxnode,
1441 nodemask_t *nodes)
1442{
1443 unsigned long copy = ALIGN(maxnode-1, 64) / 8;
1444 unsigned int nbytes = BITS_TO_LONGS(nr_node_ids) * sizeof(long);
1445
1446 if (copy > nbytes) {
1447 if (copy > PAGE_SIZE)
1448 return -EINVAL;
1449 if (clear_user((char __user *)mask + nbytes, copy - nbytes))
1450 return -EFAULT;
1451 copy = nbytes;
1452 }
1453 return copy_to_user(mask, nodes_addr(*nodes), copy) ? -EFAULT : 0;
1454}
1455
1456static long kernel_mbind(unsigned long start, unsigned long len,
1457 unsigned long mode, const unsigned long __user *nmask,
1458 unsigned long maxnode, unsigned int flags)
1459{
1460 nodemask_t nodes;
1461 int err;
1462 unsigned short mode_flags;
1463
1464 start = untagged_addr(start);
1465 mode_flags = mode & MPOL_MODE_FLAGS;
1466 mode &= ~MPOL_MODE_FLAGS;
1467 if (mode >= MPOL_MAX)
1468 return -EINVAL;
1469 if ((mode_flags & MPOL_F_STATIC_NODES) &&
1470 (mode_flags & MPOL_F_RELATIVE_NODES))
1471 return -EINVAL;
1472 err = get_nodes(&nodes, nmask, maxnode);
1473 if (err)
1474 return err;
1475 return do_mbind(start, len, mode, mode_flags, &nodes, flags);
1476}
1477
1478SYSCALL_DEFINE6(mbind, unsigned long, start, unsigned long, len,
1479 unsigned long, mode, const unsigned long __user *, nmask,
1480 unsigned long, maxnode, unsigned int, flags)
1481{
1482 return kernel_mbind(start, len, mode, nmask, maxnode, flags);
1483}
1484
1485/* Set the process memory policy */
1486static long kernel_set_mempolicy(int mode, const unsigned long __user *nmask,
1487 unsigned long maxnode)
1488{
1489 int err;
1490 nodemask_t nodes;
1491 unsigned short flags;
1492
1493 flags = mode & MPOL_MODE_FLAGS;
1494 mode &= ~MPOL_MODE_FLAGS;
1495 if ((unsigned int)mode >= MPOL_MAX)
1496 return -EINVAL;
1497 if ((flags & MPOL_F_STATIC_NODES) && (flags & MPOL_F_RELATIVE_NODES))
1498 return -EINVAL;
1499 err = get_nodes(&nodes, nmask, maxnode);
1500 if (err)
1501 return err;
1502 return do_set_mempolicy(mode, flags, &nodes);
1503}
1504
1505SYSCALL_DEFINE3(set_mempolicy, int, mode, const unsigned long __user *, nmask,
1506 unsigned long, maxnode)
1507{
1508 return kernel_set_mempolicy(mode, nmask, maxnode);
1509}
1510
1511static int kernel_migrate_pages(pid_t pid, unsigned long maxnode,
1512 const unsigned long __user *old_nodes,
1513 const unsigned long __user *new_nodes)
1514{
1515 struct mm_struct *mm = NULL;
1516 struct task_struct *task;
1517 nodemask_t task_nodes;
1518 int err;
1519 nodemask_t *old;
1520 nodemask_t *new;
1521 NODEMASK_SCRATCH(scratch);
1522
1523 if (!scratch)
1524 return -ENOMEM;
1525
1526 old = &scratch->mask1;
1527 new = &scratch->mask2;
1528
1529 err = get_nodes(old, old_nodes, maxnode);
1530 if (err)
1531 goto out;
1532
1533 err = get_nodes(new, new_nodes, maxnode);
1534 if (err)
1535 goto out;
1536
1537 /* Find the mm_struct */
1538 rcu_read_lock();
1539 task = pid ? find_task_by_vpid(pid) : current;
1540 if (!task) {
1541 rcu_read_unlock();
1542 err = -ESRCH;
1543 goto out;
1544 }
1545 get_task_struct(task);
1546
1547 err = -EINVAL;
1548
1549 /*
1550 * Check if this process has the right to modify the specified process.
1551 * Use the regular "ptrace_may_access()" checks.
1552 */
1553 if (!ptrace_may_access(task, PTRACE_MODE_READ_REALCREDS)) {
1554 rcu_read_unlock();
1555 err = -EPERM;
1556 goto out_put;
1557 }
1558 rcu_read_unlock();
1559
1560 task_nodes = cpuset_mems_allowed(task);
1561 /* Is the user allowed to access the target nodes? */
1562 if (!nodes_subset(*new, task_nodes) && !capable(CAP_SYS_NICE)) {
1563 err = -EPERM;
1564 goto out_put;
1565 }
1566
1567 task_nodes = cpuset_mems_allowed(current);
1568 nodes_and(*new, *new, task_nodes);
1569 if (nodes_empty(*new))
1570 goto out_put;
1571
1572 err = security_task_movememory(task);
1573 if (err)
1574 goto out_put;
1575
1576 mm = get_task_mm(task);
1577 put_task_struct(task);
1578
1579 if (!mm) {
1580 err = -EINVAL;
1581 goto out;
1582 }
1583
1584 err = do_migrate_pages(mm, old, new,
1585 capable(CAP_SYS_NICE) ? MPOL_MF_MOVE_ALL : MPOL_MF_MOVE);
1586
1587 mmput(mm);
1588out:
1589 NODEMASK_SCRATCH_FREE(scratch);
1590
1591 return err;
1592
1593out_put:
1594 put_task_struct(task);
1595 goto out;
1596
1597}
1598
1599SYSCALL_DEFINE4(migrate_pages, pid_t, pid, unsigned long, maxnode,
1600 const unsigned long __user *, old_nodes,
1601 const unsigned long __user *, new_nodes)
1602{
1603 return kernel_migrate_pages(pid, maxnode, old_nodes, new_nodes);
1604}
1605
1606
1607/* Retrieve NUMA policy */
1608static int kernel_get_mempolicy(int __user *policy,
1609 unsigned long __user *nmask,
1610 unsigned long maxnode,
1611 unsigned long addr,
1612 unsigned long flags)
1613{
1614 int err;
1615 int pval;
1616 nodemask_t nodes;
1617
1618 if (nmask != NULL && maxnode < nr_node_ids)
1619 return -EINVAL;
1620
1621 addr = untagged_addr(addr);
1622
1623 err = do_get_mempolicy(&pval, &nodes, addr, flags);
1624
1625 if (err)
1626 return err;
1627
1628 if (policy && put_user(pval, policy))
1629 return -EFAULT;
1630
1631 if (nmask)
1632 err = copy_nodes_to_user(nmask, maxnode, &nodes);
1633
1634 return err;
1635}
1636
1637SYSCALL_DEFINE5(get_mempolicy, int __user *, policy,
1638 unsigned long __user *, nmask, unsigned long, maxnode,
1639 unsigned long, addr, unsigned long, flags)
1640{
1641 return kernel_get_mempolicy(policy, nmask, maxnode, addr, flags);
1642}
1643
1644#ifdef CONFIG_COMPAT
1645
1646COMPAT_SYSCALL_DEFINE5(get_mempolicy, int __user *, policy,
1647 compat_ulong_t __user *, nmask,
1648 compat_ulong_t, maxnode,
1649 compat_ulong_t, addr, compat_ulong_t, flags)
1650{
1651 long err;
1652 unsigned long __user *nm = NULL;
1653 unsigned long nr_bits, alloc_size;
1654 DECLARE_BITMAP(bm, MAX_NUMNODES);
1655
1656 nr_bits = min_t(unsigned long, maxnode-1, nr_node_ids);
1657 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1658
1659 if (nmask)
1660 nm = compat_alloc_user_space(alloc_size);
1661
1662 err = kernel_get_mempolicy(policy, nm, nr_bits+1, addr, flags);
1663
1664 if (!err && nmask) {
1665 unsigned long copy_size;
1666 copy_size = min_t(unsigned long, sizeof(bm), alloc_size);
1667 err = copy_from_user(bm, nm, copy_size);
1668 /* ensure entire bitmap is zeroed */
1669 err |= clear_user(nmask, ALIGN(maxnode-1, 8) / 8);
1670 err |= compat_put_bitmap(nmask, bm, nr_bits);
1671 }
1672
1673 return err;
1674}
1675
1676COMPAT_SYSCALL_DEFINE3(set_mempolicy, int, mode, compat_ulong_t __user *, nmask,
1677 compat_ulong_t, maxnode)
1678{
1679 unsigned long __user *nm = NULL;
1680 unsigned long nr_bits, alloc_size;
1681 DECLARE_BITMAP(bm, MAX_NUMNODES);
1682
1683 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1684 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1685
1686 if (nmask) {
1687 if (compat_get_bitmap(bm, nmask, nr_bits))
1688 return -EFAULT;
1689 nm = compat_alloc_user_space(alloc_size);
1690 if (copy_to_user(nm, bm, alloc_size))
1691 return -EFAULT;
1692 }
1693
1694 return kernel_set_mempolicy(mode, nm, nr_bits+1);
1695}
1696
1697COMPAT_SYSCALL_DEFINE6(mbind, compat_ulong_t, start, compat_ulong_t, len,
1698 compat_ulong_t, mode, compat_ulong_t __user *, nmask,
1699 compat_ulong_t, maxnode, compat_ulong_t, flags)
1700{
1701 unsigned long __user *nm = NULL;
1702 unsigned long nr_bits, alloc_size;
1703 nodemask_t bm;
1704
1705 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1706 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1707
1708 if (nmask) {
1709 if (compat_get_bitmap(nodes_addr(bm), nmask, nr_bits))
1710 return -EFAULT;
1711 nm = compat_alloc_user_space(alloc_size);
1712 if (copy_to_user(nm, nodes_addr(bm), alloc_size))
1713 return -EFAULT;
1714 }
1715
1716 return kernel_mbind(start, len, mode, nm, nr_bits+1, flags);
1717}
1718
1719COMPAT_SYSCALL_DEFINE4(migrate_pages, compat_pid_t, pid,
1720 compat_ulong_t, maxnode,
1721 const compat_ulong_t __user *, old_nodes,
1722 const compat_ulong_t __user *, new_nodes)
1723{
1724 unsigned long __user *old = NULL;
1725 unsigned long __user *new = NULL;
1726 nodemask_t tmp_mask;
1727 unsigned long nr_bits;
1728 unsigned long size;
1729
1730 nr_bits = min_t(unsigned long, maxnode - 1, MAX_NUMNODES);
1731 size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1732 if (old_nodes) {
1733 if (compat_get_bitmap(nodes_addr(tmp_mask), old_nodes, nr_bits))
1734 return -EFAULT;
1735 old = compat_alloc_user_space(new_nodes ? size * 2 : size);
1736 if (new_nodes)
1737 new = old + size / sizeof(unsigned long);
1738 if (copy_to_user(old, nodes_addr(tmp_mask), size))
1739 return -EFAULT;
1740 }
1741 if (new_nodes) {
1742 if (compat_get_bitmap(nodes_addr(tmp_mask), new_nodes, nr_bits))
1743 return -EFAULT;
1744 if (new == NULL)
1745 new = compat_alloc_user_space(size);
1746 if (copy_to_user(new, nodes_addr(tmp_mask), size))
1747 return -EFAULT;
1748 }
1749 return kernel_migrate_pages(pid, nr_bits + 1, old, new);
1750}
1751
1752#endif /* CONFIG_COMPAT */
1753
1754bool vma_migratable(struct vm_area_struct *vma)
1755{
1756 if (vma->vm_flags & (VM_IO | VM_PFNMAP))
1757 return false;
1758
1759 /*
1760 * DAX device mappings require predictable access latency, so avoid
1761 * incurring periodic faults.
1762 */
1763 if (vma_is_dax(vma))
1764 return false;
1765
1766 if (is_vm_hugetlb_page(vma) &&
1767 !hugepage_migration_supported(hstate_vma(vma)))
1768 return false;
1769
1770 /*
1771 * Migration allocates pages in the highest zone. If we cannot
1772 * do so then migration (at least from node to node) is not
1773 * possible.
1774 */
1775 if (vma->vm_file &&
1776 gfp_zone(mapping_gfp_mask(vma->vm_file->f_mapping))
1777 < policy_zone)
1778 return false;
1779 return true;
1780}
1781
1782struct mempolicy *__get_vma_policy(struct vm_area_struct *vma,
1783 unsigned long addr)
1784{
1785 struct mempolicy *pol = NULL;
1786
1787 if (vma) {
1788 if (vma->vm_ops && vma->vm_ops->get_policy) {
1789 pol = vma->vm_ops->get_policy(vma, addr);
1790 } else if (vma->vm_policy) {
1791 pol = vma->vm_policy;
1792
1793 /*
1794 * shmem_alloc_page() passes MPOL_F_SHARED policy with
1795 * a pseudo vma whose vma->vm_ops=NULL. Take a reference
1796 * count on these policies which will be dropped by
1797 * mpol_cond_put() later
1798 */
1799 if (mpol_needs_cond_ref(pol))
1800 mpol_get(pol);
1801 }
1802 }
1803
1804 return pol;
1805}
1806
1807/*
1808 * get_vma_policy(@vma, @addr)
1809 * @vma: virtual memory area whose policy is sought
1810 * @addr: address in @vma for shared policy lookup
1811 *
1812 * Returns effective policy for a VMA at specified address.
1813 * Falls back to current->mempolicy or system default policy, as necessary.
1814 * Shared policies [those marked as MPOL_F_SHARED] require an extra reference
1815 * count--added by the get_policy() vm_op, as appropriate--to protect against
1816 * freeing by another task. It is the caller's responsibility to free the
1817 * extra reference for shared policies.
1818 */
1819static struct mempolicy *get_vma_policy(struct vm_area_struct *vma,
1820 unsigned long addr)
1821{
1822 struct mempolicy *pol = __get_vma_policy(vma, addr);
1823
1824 if (!pol)
1825 pol = get_task_policy(current);
1826
1827 return pol;
1828}
1829
1830bool vma_policy_mof(struct vm_area_struct *vma)
1831{
1832 struct mempolicy *pol;
1833
1834 if (vma->vm_ops && vma->vm_ops->get_policy) {
1835 bool ret = false;
1836
1837 pol = vma->vm_ops->get_policy(vma, vma->vm_start);
1838 if (pol && (pol->flags & MPOL_F_MOF))
1839 ret = true;
1840 mpol_cond_put(pol);
1841
1842 return ret;
1843 }
1844
1845 pol = vma->vm_policy;
1846 if (!pol)
1847 pol = get_task_policy(current);
1848
1849 return pol->flags & MPOL_F_MOF;
1850}
1851
1852static int apply_policy_zone(struct mempolicy *policy, enum zone_type zone)
1853{
1854 enum zone_type dynamic_policy_zone = policy_zone;
1855
1856 BUG_ON(dynamic_policy_zone == ZONE_MOVABLE);
1857
1858 /*
1859 * if policy->v.nodes has movable memory only,
1860 * we apply policy when gfp_zone(gfp) = ZONE_MOVABLE only.
1861 *
1862 * policy->v.nodes is intersect with node_states[N_MEMORY].
1863 * so if the following test faile, it implies
1864 * policy->v.nodes has movable memory only.
1865 */
1866 if (!nodes_intersects(policy->v.nodes, node_states[N_HIGH_MEMORY]))
1867 dynamic_policy_zone = ZONE_MOVABLE;
1868
1869 return zone >= dynamic_policy_zone;
1870}
1871
1872/*
1873 * Return a nodemask representing a mempolicy for filtering nodes for
1874 * page allocation
1875 */
1876nodemask_t *policy_nodemask(gfp_t gfp, struct mempolicy *policy)
1877{
1878 /* Lower zones don't get a nodemask applied for MPOL_BIND */
1879 if (unlikely(policy->mode == MPOL_BIND) &&
1880 apply_policy_zone(policy, gfp_zone(gfp)) &&
1881 cpuset_nodemask_valid_mems_allowed(&policy->v.nodes))
1882 return &policy->v.nodes;
1883
1884 return NULL;
1885}
1886
1887/* Return the node id preferred by the given mempolicy, or the given id */
1888static int policy_node(gfp_t gfp, struct mempolicy *policy,
1889 int nd)
1890{
1891 if (policy->mode == MPOL_PREFERRED && !(policy->flags & MPOL_F_LOCAL))
1892 nd = policy->v.preferred_node;
1893 else {
1894 /*
1895 * __GFP_THISNODE shouldn't even be used with the bind policy
1896 * because we might easily break the expectation to stay on the
1897 * requested node and not break the policy.
1898 */
1899 WARN_ON_ONCE(policy->mode == MPOL_BIND && (gfp & __GFP_THISNODE));
1900 }
1901
1902 return nd;
1903}
1904
1905/* Do dynamic interleaving for a process */
1906static unsigned interleave_nodes(struct mempolicy *policy)
1907{
1908 unsigned next;
1909 struct task_struct *me = current;
1910
1911 next = next_node_in(me->il_prev, policy->v.nodes);
1912 if (next < MAX_NUMNODES)
1913 me->il_prev = next;
1914 return next;
1915}
1916
1917/*
1918 * Depending on the memory policy provide a node from which to allocate the
1919 * next slab entry.
1920 */
1921unsigned int mempolicy_slab_node(void)
1922{
1923 struct mempolicy *policy;
1924 int node = numa_mem_id();
1925
1926 if (in_interrupt())
1927 return node;
1928
1929 policy = current->mempolicy;
1930 if (!policy || policy->flags & MPOL_F_LOCAL)
1931 return node;
1932
1933 switch (policy->mode) {
1934 case MPOL_PREFERRED:
1935 /*
1936 * handled MPOL_F_LOCAL above
1937 */
1938 return policy->v.preferred_node;
1939
1940 case MPOL_INTERLEAVE:
1941 return interleave_nodes(policy);
1942
1943 case MPOL_BIND: {
1944 struct zoneref *z;
1945
1946 /*
1947 * Follow bind policy behavior and start allocation at the
1948 * first node.
1949 */
1950 struct zonelist *zonelist;
1951 enum zone_type highest_zoneidx = gfp_zone(GFP_KERNEL);
1952 zonelist = &NODE_DATA(node)->node_zonelists[ZONELIST_FALLBACK];
1953 z = first_zones_zonelist(zonelist, highest_zoneidx,
1954 &policy->v.nodes);
1955 return z->zone ? zone_to_nid(z->zone) : node;
1956 }
1957
1958 default:
1959 BUG();
1960 }
1961}
1962
1963/*
1964 * Do static interleaving for a VMA with known offset @n. Returns the n'th
1965 * node in pol->v.nodes (starting from n=0), wrapping around if n exceeds the
1966 * number of present nodes.
1967 */
1968static unsigned offset_il_node(struct mempolicy *pol, unsigned long n)
1969{
1970 unsigned nnodes = nodes_weight(pol->v.nodes);
1971 unsigned target;
1972 int i;
1973 int nid;
1974
1975 if (!nnodes)
1976 return numa_node_id();
1977 target = (unsigned int)n % nnodes;
1978 nid = first_node(pol->v.nodes);
1979 for (i = 0; i < target; i++)
1980 nid = next_node(nid, pol->v.nodes);
1981 return nid;
1982}
1983
1984/* Determine a node number for interleave */
1985static inline unsigned interleave_nid(struct mempolicy *pol,
1986 struct vm_area_struct *vma, unsigned long addr, int shift)
1987{
1988 if (vma) {
1989 unsigned long off;
1990
1991 /*
1992 * for small pages, there is no difference between
1993 * shift and PAGE_SHIFT, so the bit-shift is safe.
1994 * for huge pages, since vm_pgoff is in units of small
1995 * pages, we need to shift off the always 0 bits to get
1996 * a useful offset.
1997 */
1998 BUG_ON(shift < PAGE_SHIFT);
1999 off = vma->vm_pgoff >> (shift - PAGE_SHIFT);
2000 off += (addr - vma->vm_start) >> shift;
2001 return offset_il_node(pol, off);
2002 } else
2003 return interleave_nodes(pol);
2004}
2005
2006#ifdef CONFIG_HUGETLBFS
2007/*
2008 * huge_node(@vma, @addr, @gfp_flags, @mpol)
2009 * @vma: virtual memory area whose policy is sought
2010 * @addr: address in @vma for shared policy lookup and interleave policy
2011 * @gfp_flags: for requested zone
2012 * @mpol: pointer to mempolicy pointer for reference counted mempolicy
2013 * @nodemask: pointer to nodemask pointer for MPOL_BIND nodemask
2014 *
2015 * Returns a nid suitable for a huge page allocation and a pointer
2016 * to the struct mempolicy for conditional unref after allocation.
2017 * If the effective policy is 'BIND, returns a pointer to the mempolicy's
2018 * @nodemask for filtering the zonelist.
2019 *
2020 * Must be protected by read_mems_allowed_begin()
2021 */
2022int huge_node(struct vm_area_struct *vma, unsigned long addr, gfp_t gfp_flags,
2023 struct mempolicy **mpol, nodemask_t **nodemask)
2024{
2025 int nid;
2026
2027 *mpol = get_vma_policy(vma, addr);
2028 *nodemask = NULL; /* assume !MPOL_BIND */
2029
2030 if (unlikely((*mpol)->mode == MPOL_INTERLEAVE)) {
2031 nid = interleave_nid(*mpol, vma, addr,
2032 huge_page_shift(hstate_vma(vma)));
2033 } else {
2034 nid = policy_node(gfp_flags, *mpol, numa_node_id());
2035 if ((*mpol)->mode == MPOL_BIND)
2036 *nodemask = &(*mpol)->v.nodes;
2037 }
2038 return nid;
2039}
2040
2041/*
2042 * init_nodemask_of_mempolicy
2043 *
2044 * If the current task's mempolicy is "default" [NULL], return 'false'
2045 * to indicate default policy. Otherwise, extract the policy nodemask
2046 * for 'bind' or 'interleave' policy into the argument nodemask, or
2047 * initialize the argument nodemask to contain the single node for
2048 * 'preferred' or 'local' policy and return 'true' to indicate presence
2049 * of non-default mempolicy.
2050 *
2051 * We don't bother with reference counting the mempolicy [mpol_get/put]
2052 * because the current task is examining it's own mempolicy and a task's
2053 * mempolicy is only ever changed by the task itself.
2054 *
2055 * N.B., it is the caller's responsibility to free a returned nodemask.
2056 */
2057bool init_nodemask_of_mempolicy(nodemask_t *mask)
2058{
2059 struct mempolicy *mempolicy;
2060 int nid;
2061
2062 if (!(mask && current->mempolicy))
2063 return false;
2064
2065 task_lock(current);
2066 mempolicy = current->mempolicy;
2067 switch (mempolicy->mode) {
2068 case MPOL_PREFERRED:
2069 if (mempolicy->flags & MPOL_F_LOCAL)
2070 nid = numa_node_id();
2071 else
2072 nid = mempolicy->v.preferred_node;
2073 init_nodemask_of_node(mask, nid);
2074 break;
2075
2076 case MPOL_BIND:
2077 case MPOL_INTERLEAVE:
2078 *mask = mempolicy->v.nodes;
2079 break;
2080
2081 default:
2082 BUG();
2083 }
2084 task_unlock(current);
2085
2086 return true;
2087}
2088#endif
2089
2090/*
2091 * mempolicy_nodemask_intersects
2092 *
2093 * If tsk's mempolicy is "default" [NULL], return 'true' to indicate default
2094 * policy. Otherwise, check for intersection between mask and the policy
2095 * nodemask for 'bind' or 'interleave' policy. For 'perferred' or 'local'
2096 * policy, always return true since it may allocate elsewhere on fallback.
2097 *
2098 * Takes task_lock(tsk) to prevent freeing of its mempolicy.
2099 */
2100bool mempolicy_nodemask_intersects(struct task_struct *tsk,
2101 const nodemask_t *mask)
2102{
2103 struct mempolicy *mempolicy;
2104 bool ret = true;
2105
2106 if (!mask)
2107 return ret;
2108 task_lock(tsk);
2109 mempolicy = tsk->mempolicy;
2110 if (!mempolicy)
2111 goto out;
2112
2113 switch (mempolicy->mode) {
2114 case MPOL_PREFERRED:
2115 /*
2116 * MPOL_PREFERRED and MPOL_F_LOCAL are only preferred nodes to
2117 * allocate from, they may fallback to other nodes when oom.
2118 * Thus, it's possible for tsk to have allocated memory from
2119 * nodes in mask.
2120 */
2121 break;
2122 case MPOL_BIND:
2123 case MPOL_INTERLEAVE:
2124 ret = nodes_intersects(mempolicy->v.nodes, *mask);
2125 break;
2126 default:
2127 BUG();
2128 }
2129out:
2130 task_unlock(tsk);
2131 return ret;
2132}
2133
2134/* Allocate a page in interleaved policy.
2135 Own path because it needs to do special accounting. */
2136static struct page *alloc_page_interleave(gfp_t gfp, unsigned order,
2137 unsigned nid)
2138{
2139 struct page *page;
2140
2141 page = __alloc_pages(gfp, order, nid);
2142 /* skip NUMA_INTERLEAVE_HIT counter update if numa stats is disabled */
2143 if (!static_branch_likely(&vm_numa_stat_key))
2144 return page;
2145 if (page && page_to_nid(page) == nid) {
2146 preempt_disable();
2147 __inc_numa_state(page_zone(page), NUMA_INTERLEAVE_HIT);
2148 preempt_enable();
2149 }
2150 return page;
2151}
2152
2153/**
2154 * alloc_pages_vma - Allocate a page for a VMA.
2155 *
2156 * @gfp:
2157 * %GFP_USER user allocation.
2158 * %GFP_KERNEL kernel allocations,
2159 * %GFP_HIGHMEM highmem/user allocations,
2160 * %GFP_FS allocation should not call back into a file system.
2161 * %GFP_ATOMIC don't sleep.
2162 *
2163 * @order:Order of the GFP allocation.
2164 * @vma: Pointer to VMA or NULL if not available.
2165 * @addr: Virtual Address of the allocation. Must be inside the VMA.
2166 * @node: Which node to prefer for allocation (modulo policy).
2167 * @hugepage: for hugepages try only the preferred node if possible
2168 *
2169 * This function allocates a page from the kernel page pool and applies
2170 * a NUMA policy associated with the VMA or the current process.
2171 * When VMA is not NULL caller must read-lock the mmap_lock of the
2172 * mm_struct of the VMA to prevent it from going away. Should be used for
2173 * all allocations for pages that will be mapped into user space. Returns
2174 * NULL when no page can be allocated.
2175 */
2176struct page *
2177alloc_pages_vma(gfp_t gfp, int order, struct vm_area_struct *vma,
2178 unsigned long addr, int node, bool hugepage)
2179{
2180 struct mempolicy *pol;
2181 struct page *page;
2182 int preferred_nid;
2183 nodemask_t *nmask;
2184
2185 pol = get_vma_policy(vma, addr);
2186
2187 if (pol->mode == MPOL_INTERLEAVE) {
2188 unsigned nid;
2189
2190 nid = interleave_nid(pol, vma, addr, PAGE_SHIFT + order);
2191 mpol_cond_put(pol);
2192 page = alloc_page_interleave(gfp, order, nid);
2193 goto out;
2194 }
2195
2196 if (unlikely(IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE) && hugepage)) {
2197 int hpage_node = node;
2198
2199 /*
2200 * For hugepage allocation and non-interleave policy which
2201 * allows the current node (or other explicitly preferred
2202 * node) we only try to allocate from the current/preferred
2203 * node and don't fall back to other nodes, as the cost of
2204 * remote accesses would likely offset THP benefits.
2205 *
2206 * If the policy is interleave, or does not allow the current
2207 * node in its nodemask, we allocate the standard way.
2208 */
2209 if (pol->mode == MPOL_PREFERRED && !(pol->flags & MPOL_F_LOCAL))
2210 hpage_node = pol->v.preferred_node;
2211
2212 nmask = policy_nodemask(gfp, pol);
2213 if (!nmask || node_isset(hpage_node, *nmask)) {
2214 mpol_cond_put(pol);
2215 /*
2216 * First, try to allocate THP only on local node, but
2217 * don't reclaim unnecessarily, just compact.
2218 */
2219 page = __alloc_pages_node(hpage_node,
2220 gfp | __GFP_THISNODE | __GFP_NORETRY, order);
2221
2222 /*
2223 * If hugepage allocations are configured to always
2224 * synchronous compact or the vma has been madvised
2225 * to prefer hugepage backing, retry allowing remote
2226 * memory with both reclaim and compact as well.
2227 */
2228 if (!page && (gfp & __GFP_DIRECT_RECLAIM))
2229 page = __alloc_pages_node(hpage_node,
2230 gfp, order);
2231
2232 goto out;
2233 }
2234 }
2235
2236 nmask = policy_nodemask(gfp, pol);
2237 preferred_nid = policy_node(gfp, pol, node);
2238 page = __alloc_pages_nodemask(gfp, order, preferred_nid, nmask);
2239 mpol_cond_put(pol);
2240out:
2241 return page;
2242}
2243EXPORT_SYMBOL(alloc_pages_vma);
2244
2245/**
2246 * alloc_pages_current - Allocate pages.
2247 *
2248 * @gfp:
2249 * %GFP_USER user allocation,
2250 * %GFP_KERNEL kernel allocation,
2251 * %GFP_HIGHMEM highmem allocation,
2252 * %GFP_FS don't call back into a file system.
2253 * %GFP_ATOMIC don't sleep.
2254 * @order: Power of two of allocation size in pages. 0 is a single page.
2255 *
2256 * Allocate a page from the kernel page pool. When not in
2257 * interrupt context and apply the current process NUMA policy.
2258 * Returns NULL when no page can be allocated.
2259 */
2260struct page *alloc_pages_current(gfp_t gfp, unsigned order)
2261{
2262 struct mempolicy *pol = &default_policy;
2263 struct page *page;
2264
2265 if (!in_interrupt() && !(gfp & __GFP_THISNODE))
2266 pol = get_task_policy(current);
2267
2268 /*
2269 * No reference counting needed for current->mempolicy
2270 * nor system default_policy
2271 */
2272 if (pol->mode == MPOL_INTERLEAVE)
2273 page = alloc_page_interleave(gfp, order, interleave_nodes(pol));
2274 else
2275 page = __alloc_pages_nodemask(gfp, order,
2276 policy_node(gfp, pol, numa_node_id()),
2277 policy_nodemask(gfp, pol));
2278
2279 return page;
2280}
2281EXPORT_SYMBOL(alloc_pages_current);
2282
2283int vma_dup_policy(struct vm_area_struct *src, struct vm_area_struct *dst)
2284{
2285 struct mempolicy *pol = mpol_dup(vma_policy(src));
2286
2287 if (IS_ERR(pol))
2288 return PTR_ERR(pol);
2289 dst->vm_policy = pol;
2290 return 0;
2291}
2292
2293/*
2294 * If mpol_dup() sees current->cpuset == cpuset_being_rebound, then it
2295 * rebinds the mempolicy its copying by calling mpol_rebind_policy()
2296 * with the mems_allowed returned by cpuset_mems_allowed(). This
2297 * keeps mempolicies cpuset relative after its cpuset moves. See
2298 * further kernel/cpuset.c update_nodemask().
2299 *
2300 * current's mempolicy may be rebinded by the other task(the task that changes
2301 * cpuset's mems), so we needn't do rebind work for current task.
2302 */
2303
2304/* Slow path of a mempolicy duplicate */
2305struct mempolicy *__mpol_dup(struct mempolicy *old)
2306{
2307 struct mempolicy *new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
2308
2309 if (!new)
2310 return ERR_PTR(-ENOMEM);
2311
2312 /* task's mempolicy is protected by alloc_lock */
2313 if (old == current->mempolicy) {
2314 task_lock(current);
2315 *new = *old;
2316 task_unlock(current);
2317 } else
2318 *new = *old;
2319
2320 if (current_cpuset_is_being_rebound()) {
2321 nodemask_t mems = cpuset_mems_allowed(current);
2322 mpol_rebind_policy(new, &mems);
2323 }
2324 atomic_set(&new->refcnt, 1);
2325 return new;
2326}
2327
2328/* Slow path of a mempolicy comparison */
2329bool __mpol_equal(struct mempolicy *a, struct mempolicy *b)
2330{
2331 if (!a || !b)
2332 return false;
2333 if (a->mode != b->mode)
2334 return false;
2335 if (a->flags != b->flags)
2336 return false;
2337 if (mpol_store_user_nodemask(a))
2338 if (!nodes_equal(a->w.user_nodemask, b->w.user_nodemask))
2339 return false;
2340
2341 switch (a->mode) {
2342 case MPOL_BIND:
2343 case MPOL_INTERLEAVE:
2344 return !!nodes_equal(a->v.nodes, b->v.nodes);
2345 case MPOL_PREFERRED:
2346 /* a's ->flags is the same as b's */
2347 if (a->flags & MPOL_F_LOCAL)
2348 return true;
2349 return a->v.preferred_node == b->v.preferred_node;
2350 default:
2351 BUG();
2352 return false;
2353 }
2354}
2355
2356/*
2357 * Shared memory backing store policy support.
2358 *
2359 * Remember policies even when nobody has shared memory mapped.
2360 * The policies are kept in Red-Black tree linked from the inode.
2361 * They are protected by the sp->lock rwlock, which should be held
2362 * for any accesses to the tree.
2363 */
2364
2365/*
2366 * lookup first element intersecting start-end. Caller holds sp->lock for
2367 * reading or for writing
2368 */
2369static struct sp_node *
2370sp_lookup(struct shared_policy *sp, unsigned long start, unsigned long end)
2371{
2372 struct rb_node *n = sp->root.rb_node;
2373
2374 while (n) {
2375 struct sp_node *p = rb_entry(n, struct sp_node, nd);
2376
2377 if (start >= p->end)
2378 n = n->rb_right;
2379 else if (end <= p->start)
2380 n = n->rb_left;
2381 else
2382 break;
2383 }
2384 if (!n)
2385 return NULL;
2386 for (;;) {
2387 struct sp_node *w = NULL;
2388 struct rb_node *prev = rb_prev(n);
2389 if (!prev)
2390 break;
2391 w = rb_entry(prev, struct sp_node, nd);
2392 if (w->end <= start)
2393 break;
2394 n = prev;
2395 }
2396 return rb_entry(n, struct sp_node, nd);
2397}
2398
2399/*
2400 * Insert a new shared policy into the list. Caller holds sp->lock for
2401 * writing.
2402 */
2403static void sp_insert(struct shared_policy *sp, struct sp_node *new)
2404{
2405 struct rb_node **p = &sp->root.rb_node;
2406 struct rb_node *parent = NULL;
2407 struct sp_node *nd;
2408
2409 while (*p) {
2410 parent = *p;
2411 nd = rb_entry(parent, struct sp_node, nd);
2412 if (new->start < nd->start)
2413 p = &(*p)->rb_left;
2414 else if (new->end > nd->end)
2415 p = &(*p)->rb_right;
2416 else
2417 BUG();
2418 }
2419 rb_link_node(&new->nd, parent, p);
2420 rb_insert_color(&new->nd, &sp->root);
2421 pr_debug("inserting %lx-%lx: %d\n", new->start, new->end,
2422 new->policy ? new->policy->mode : 0);
2423}
2424
2425/* Find shared policy intersecting idx */
2426struct mempolicy *
2427mpol_shared_policy_lookup(struct shared_policy *sp, unsigned long idx)
2428{
2429 struct mempolicy *pol = NULL;
2430 struct sp_node *sn;
2431
2432 if (!sp->root.rb_node)
2433 return NULL;
2434 read_lock(&sp->lock);
2435 sn = sp_lookup(sp, idx, idx+1);
2436 if (sn) {
2437 mpol_get(sn->policy);
2438 pol = sn->policy;
2439 }
2440 read_unlock(&sp->lock);
2441 return pol;
2442}
2443
2444static void sp_free(struct sp_node *n)
2445{
2446 mpol_put(n->policy);
2447 kmem_cache_free(sn_cache, n);
2448}
2449
2450/**
2451 * mpol_misplaced - check whether current page node is valid in policy
2452 *
2453 * @page: page to be checked
2454 * @vma: vm area where page mapped
2455 * @addr: virtual address where page mapped
2456 *
2457 * Lookup current policy node id for vma,addr and "compare to" page's
2458 * node id.
2459 *
2460 * Returns:
2461 * -1 - not misplaced, page is in the right node
2462 * node - node id where the page should be
2463 *
2464 * Policy determination "mimics" alloc_page_vma().
2465 * Called from fault path where we know the vma and faulting address.
2466 */
2467int mpol_misplaced(struct page *page, struct vm_area_struct *vma, unsigned long addr)
2468{
2469 struct mempolicy *pol;
2470 struct zoneref *z;
2471 int curnid = page_to_nid(page);
2472 unsigned long pgoff;
2473 int thiscpu = raw_smp_processor_id();
2474 int thisnid = cpu_to_node(thiscpu);
2475 int polnid = NUMA_NO_NODE;
2476 int ret = -1;
2477
2478 pol = get_vma_policy(vma, addr);
2479 if (!(pol->flags & MPOL_F_MOF))
2480 goto out;
2481
2482 switch (pol->mode) {
2483 case MPOL_INTERLEAVE:
2484 pgoff = vma->vm_pgoff;
2485 pgoff += (addr - vma->vm_start) >> PAGE_SHIFT;
2486 polnid = offset_il_node(pol, pgoff);
2487 break;
2488
2489 case MPOL_PREFERRED:
2490 if (pol->flags & MPOL_F_LOCAL)
2491 polnid = numa_node_id();
2492 else
2493 polnid = pol->v.preferred_node;
2494 break;
2495
2496 case MPOL_BIND:
2497
2498 /*
2499 * allows binding to multiple nodes.
2500 * use current page if in policy nodemask,
2501 * else select nearest allowed node, if any.
2502 * If no allowed nodes, use current [!misplaced].
2503 */
2504 if (node_isset(curnid, pol->v.nodes))
2505 goto out;
2506 z = first_zones_zonelist(
2507 node_zonelist(numa_node_id(), GFP_HIGHUSER),
2508 gfp_zone(GFP_HIGHUSER),
2509 &pol->v.nodes);
2510 polnid = zone_to_nid(z->zone);
2511 break;
2512
2513 default:
2514 BUG();
2515 }
2516
2517 /* Migrate the page towards the node whose CPU is referencing it */
2518 if (pol->flags & MPOL_F_MORON) {
2519 polnid = thisnid;
2520
2521 if (!should_numa_migrate_memory(current, page, curnid, thiscpu))
2522 goto out;
2523 }
2524
2525 if (curnid != polnid)
2526 ret = polnid;
2527out:
2528 mpol_cond_put(pol);
2529
2530 return ret;
2531}
2532
2533/*
2534 * Drop the (possibly final) reference to task->mempolicy. It needs to be
2535 * dropped after task->mempolicy is set to NULL so that any allocation done as
2536 * part of its kmem_cache_free(), such as by KASAN, doesn't reference a freed
2537 * policy.
2538 */
2539void mpol_put_task_policy(struct task_struct *task)
2540{
2541 struct mempolicy *pol;
2542
2543 task_lock(task);
2544 pol = task->mempolicy;
2545 task->mempolicy = NULL;
2546 task_unlock(task);
2547 mpol_put(pol);
2548}
2549
2550static void sp_delete(struct shared_policy *sp, struct sp_node *n)
2551{
2552 pr_debug("deleting %lx-l%lx\n", n->start, n->end);
2553 rb_erase(&n->nd, &sp->root);
2554 sp_free(n);
2555}
2556
2557static void sp_node_init(struct sp_node *node, unsigned long start,
2558 unsigned long end, struct mempolicy *pol)
2559{
2560 node->start = start;
2561 node->end = end;
2562 node->policy = pol;
2563}
2564
2565static struct sp_node *sp_alloc(unsigned long start, unsigned long end,
2566 struct mempolicy *pol)
2567{
2568 struct sp_node *n;
2569 struct mempolicy *newpol;
2570
2571 n = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2572 if (!n)
2573 return NULL;
2574
2575 newpol = mpol_dup(pol);
2576 if (IS_ERR(newpol)) {
2577 kmem_cache_free(sn_cache, n);
2578 return NULL;
2579 }
2580 newpol->flags |= MPOL_F_SHARED;
2581 sp_node_init(n, start, end, newpol);
2582
2583 return n;
2584}
2585
2586/* Replace a policy range. */
2587static int shared_policy_replace(struct shared_policy *sp, unsigned long start,
2588 unsigned long end, struct sp_node *new)
2589{
2590 struct sp_node *n;
2591 struct sp_node *n_new = NULL;
2592 struct mempolicy *mpol_new = NULL;
2593 int ret = 0;
2594
2595restart:
2596 write_lock(&sp->lock);
2597 n = sp_lookup(sp, start, end);
2598 /* Take care of old policies in the same range. */
2599 while (n && n->start < end) {
2600 struct rb_node *next = rb_next(&n->nd);
2601 if (n->start >= start) {
2602 if (n->end <= end)
2603 sp_delete(sp, n);
2604 else
2605 n->start = end;
2606 } else {
2607 /* Old policy spanning whole new range. */
2608 if (n->end > end) {
2609 if (!n_new)
2610 goto alloc_new;
2611
2612 *mpol_new = *n->policy;
2613 atomic_set(&mpol_new->refcnt, 1);
2614 sp_node_init(n_new, end, n->end, mpol_new);
2615 n->end = start;
2616 sp_insert(sp, n_new);
2617 n_new = NULL;
2618 mpol_new = NULL;
2619 break;
2620 } else
2621 n->end = start;
2622 }
2623 if (!next)
2624 break;
2625 n = rb_entry(next, struct sp_node, nd);
2626 }
2627 if (new)
2628 sp_insert(sp, new);
2629 write_unlock(&sp->lock);
2630 ret = 0;
2631
2632err_out:
2633 if (mpol_new)
2634 mpol_put(mpol_new);
2635 if (n_new)
2636 kmem_cache_free(sn_cache, n_new);
2637
2638 return ret;
2639
2640alloc_new:
2641 write_unlock(&sp->lock);
2642 ret = -ENOMEM;
2643 n_new = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2644 if (!n_new)
2645 goto err_out;
2646 mpol_new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
2647 if (!mpol_new)
2648 goto err_out;
2649 goto restart;
2650}
2651
2652/**
2653 * mpol_shared_policy_init - initialize shared policy for inode
2654 * @sp: pointer to inode shared policy
2655 * @mpol: struct mempolicy to install
2656 *
2657 * Install non-NULL @mpol in inode's shared policy rb-tree.
2658 * On entry, the current task has a reference on a non-NULL @mpol.
2659 * This must be released on exit.
2660 * This is called at get_inode() calls and we can use GFP_KERNEL.
2661 */
2662void mpol_shared_policy_init(struct shared_policy *sp, struct mempolicy *mpol)
2663{
2664 int ret;
2665
2666 sp->root = RB_ROOT; /* empty tree == default mempolicy */
2667 rwlock_init(&sp->lock);
2668
2669 if (mpol) {
2670 struct vm_area_struct pvma;
2671 struct mempolicy *new;
2672 NODEMASK_SCRATCH(scratch);
2673
2674 if (!scratch)
2675 goto put_mpol;
2676 /* contextualize the tmpfs mount point mempolicy */
2677 new = mpol_new(mpol->mode, mpol->flags, &mpol->w.user_nodemask);
2678 if (IS_ERR(new))
2679 goto free_scratch; /* no valid nodemask intersection */
2680
2681 task_lock(current);
2682 ret = mpol_set_nodemask(new, &mpol->w.user_nodemask, scratch);
2683 task_unlock(current);
2684 if (ret)
2685 goto put_new;
2686
2687 /* Create pseudo-vma that contains just the policy */
2688 vma_init(&pvma, NULL);
2689 pvma.vm_end = TASK_SIZE; /* policy covers entire file */
2690 mpol_set_shared_policy(sp, &pvma, new); /* adds ref */
2691
2692put_new:
2693 mpol_put(new); /* drop initial ref */
2694free_scratch:
2695 NODEMASK_SCRATCH_FREE(scratch);
2696put_mpol:
2697 mpol_put(mpol); /* drop our incoming ref on sb mpol */
2698 }
2699}
2700
2701int mpol_set_shared_policy(struct shared_policy *info,
2702 struct vm_area_struct *vma, struct mempolicy *npol)
2703{
2704 int err;
2705 struct sp_node *new = NULL;
2706 unsigned long sz = vma_pages(vma);
2707
2708 pr_debug("set_shared_policy %lx sz %lu %d %d %lx\n",
2709 vma->vm_pgoff,
2710 sz, npol ? npol->mode : -1,
2711 npol ? npol->flags : -1,
2712 npol ? nodes_addr(npol->v.nodes)[0] : NUMA_NO_NODE);
2713
2714 if (npol) {
2715 new = sp_alloc(vma->vm_pgoff, vma->vm_pgoff + sz, npol);
2716 if (!new)
2717 return -ENOMEM;
2718 }
2719 err = shared_policy_replace(info, vma->vm_pgoff, vma->vm_pgoff+sz, new);
2720 if (err && new)
2721 sp_free(new);
2722 return err;
2723}
2724
2725/* Free a backing policy store on inode delete. */
2726void mpol_free_shared_policy(struct shared_policy *p)
2727{
2728 struct sp_node *n;
2729 struct rb_node *next;
2730
2731 if (!p->root.rb_node)
2732 return;
2733 write_lock(&p->lock);
2734 next = rb_first(&p->root);
2735 while (next) {
2736 n = rb_entry(next, struct sp_node, nd);
2737 next = rb_next(&n->nd);
2738 sp_delete(p, n);
2739 }
2740 write_unlock(&p->lock);
2741}
2742
2743#ifdef CONFIG_NUMA_BALANCING
2744static int __initdata numabalancing_override;
2745
2746static void __init check_numabalancing_enable(void)
2747{
2748 bool numabalancing_default = false;
2749
2750 if (IS_ENABLED(CONFIG_NUMA_BALANCING_DEFAULT_ENABLED))
2751 numabalancing_default = true;
2752
2753 /* Parsed by setup_numabalancing. override == 1 enables, -1 disables */
2754 if (numabalancing_override)
2755 set_numabalancing_state(numabalancing_override == 1);
2756
2757 if (num_online_nodes() > 1 && !numabalancing_override) {
2758 pr_info("%s automatic NUMA balancing. Configure with numa_balancing= or the kernel.numa_balancing sysctl\n",
2759 numabalancing_default ? "Enabling" : "Disabling");
2760 set_numabalancing_state(numabalancing_default);
2761 }
2762}
2763
2764static int __init setup_numabalancing(char *str)
2765{
2766 int ret = 0;
2767 if (!str)
2768 goto out;
2769
2770 if (!strcmp(str, "enable")) {
2771 numabalancing_override = 1;
2772 ret = 1;
2773 } else if (!strcmp(str, "disable")) {
2774 numabalancing_override = -1;
2775 ret = 1;
2776 }
2777out:
2778 if (!ret)
2779 pr_warn("Unable to parse numa_balancing=\n");
2780
2781 return ret;
2782}
2783__setup("numa_balancing=", setup_numabalancing);
2784#else
2785static inline void __init check_numabalancing_enable(void)
2786{
2787}
2788#endif /* CONFIG_NUMA_BALANCING */
2789
2790/* assumes fs == KERNEL_DS */
2791void __init numa_policy_init(void)
2792{
2793 nodemask_t interleave_nodes;
2794 unsigned long largest = 0;
2795 int nid, prefer = 0;
2796
2797 policy_cache = kmem_cache_create("numa_policy",
2798 sizeof(struct mempolicy),
2799 0, SLAB_PANIC, NULL);
2800
2801 sn_cache = kmem_cache_create("shared_policy_node",
2802 sizeof(struct sp_node),
2803 0, SLAB_PANIC, NULL);
2804
2805 for_each_node(nid) {
2806 preferred_node_policy[nid] = (struct mempolicy) {
2807 .refcnt = ATOMIC_INIT(1),
2808 .mode = MPOL_PREFERRED,
2809 .flags = MPOL_F_MOF | MPOL_F_MORON,
2810 .v = { .preferred_node = nid, },
2811 };
2812 }
2813
2814 /*
2815 * Set interleaving policy for system init. Interleaving is only
2816 * enabled across suitably sized nodes (default is >= 16MB), or
2817 * fall back to the largest node if they're all smaller.
2818 */
2819 nodes_clear(interleave_nodes);
2820 for_each_node_state(nid, N_MEMORY) {
2821 unsigned long total_pages = node_present_pages(nid);
2822
2823 /* Preserve the largest node */
2824 if (largest < total_pages) {
2825 largest = total_pages;
2826 prefer = nid;
2827 }
2828
2829 /* Interleave this node? */
2830 if ((total_pages << PAGE_SHIFT) >= (16 << 20))
2831 node_set(nid, interleave_nodes);
2832 }
2833
2834 /* All too small, use the largest */
2835 if (unlikely(nodes_empty(interleave_nodes)))
2836 node_set(prefer, interleave_nodes);
2837
2838 if (do_set_mempolicy(MPOL_INTERLEAVE, 0, &interleave_nodes))
2839 pr_err("%s: interleaving failed\n", __func__);
2840
2841 check_numabalancing_enable();
2842}
2843
2844/* Reset policy of current process to default */
2845void numa_default_policy(void)
2846{
2847 do_set_mempolicy(MPOL_DEFAULT, 0, NULL);
2848}
2849
2850/*
2851 * Parse and format mempolicy from/to strings
2852 */
2853
2854/*
2855 * "local" is implemented internally by MPOL_PREFERRED with MPOL_F_LOCAL flag.
2856 */
2857static const char * const policy_modes[] =
2858{
2859 [MPOL_DEFAULT] = "default",
2860 [MPOL_PREFERRED] = "prefer",
2861 [MPOL_BIND] = "bind",
2862 [MPOL_INTERLEAVE] = "interleave",
2863 [MPOL_LOCAL] = "local",
2864};
2865
2866
2867#ifdef CONFIG_TMPFS
2868/**
2869 * mpol_parse_str - parse string to mempolicy, for tmpfs mpol mount option.
2870 * @str: string containing mempolicy to parse
2871 * @mpol: pointer to struct mempolicy pointer, returned on success.
2872 *
2873 * Format of input:
2874 * <mode>[=<flags>][:<nodelist>]
2875 *
2876 * On success, returns 0, else 1
2877 */
2878int mpol_parse_str(char *str, struct mempolicy **mpol)
2879{
2880 struct mempolicy *new = NULL;
2881 unsigned short mode_flags;
2882 nodemask_t nodes;
2883 char *nodelist = strchr(str, ':');
2884 char *flags = strchr(str, '=');
2885 int err = 1, mode;
2886
2887 if (flags)
2888 *flags++ = '\0'; /* terminate mode string */
2889
2890 if (nodelist) {
2891 /* NUL-terminate mode or flags string */
2892 *nodelist++ = '\0';
2893 if (nodelist_parse(nodelist, nodes))
2894 goto out;
2895 if (!nodes_subset(nodes, node_states[N_MEMORY]))
2896 goto out;
2897 } else
2898 nodes_clear(nodes);
2899
2900 mode = match_string(policy_modes, MPOL_MAX, str);
2901 if (mode < 0)
2902 goto out;
2903
2904 switch (mode) {
2905 case MPOL_PREFERRED:
2906 /*
2907 * Insist on a nodelist of one node only, although later
2908 * we use first_node(nodes) to grab a single node, so here
2909 * nodelist (or nodes) cannot be empty.
2910 */
2911 if (nodelist) {
2912 char *rest = nodelist;
2913 while (isdigit(*rest))
2914 rest++;
2915 if (*rest)
2916 goto out;
2917 if (nodes_empty(nodes))
2918 goto out;
2919 }
2920 break;
2921 case MPOL_INTERLEAVE:
2922 /*
2923 * Default to online nodes with memory if no nodelist
2924 */
2925 if (!nodelist)
2926 nodes = node_states[N_MEMORY];
2927 break;
2928 case MPOL_LOCAL:
2929 /*
2930 * Don't allow a nodelist; mpol_new() checks flags
2931 */
2932 if (nodelist)
2933 goto out;
2934 mode = MPOL_PREFERRED;
2935 break;
2936 case MPOL_DEFAULT:
2937 /*
2938 * Insist on a empty nodelist
2939 */
2940 if (!nodelist)
2941 err = 0;
2942 goto out;
2943 case MPOL_BIND:
2944 /*
2945 * Insist on a nodelist
2946 */
2947 if (!nodelist)
2948 goto out;
2949 }
2950
2951 mode_flags = 0;
2952 if (flags) {
2953 /*
2954 * Currently, we only support two mutually exclusive
2955 * mode flags.
2956 */
2957 if (!strcmp(flags, "static"))
2958 mode_flags |= MPOL_F_STATIC_NODES;
2959 else if (!strcmp(flags, "relative"))
2960 mode_flags |= MPOL_F_RELATIVE_NODES;
2961 else
2962 goto out;
2963 }
2964
2965 new = mpol_new(mode, mode_flags, &nodes);
2966 if (IS_ERR(new))
2967 goto out;
2968
2969 /*
2970 * Save nodes for mpol_to_str() to show the tmpfs mount options
2971 * for /proc/mounts, /proc/pid/mounts and /proc/pid/mountinfo.
2972 */
2973 if (mode != MPOL_PREFERRED)
2974 new->v.nodes = nodes;
2975 else if (nodelist)
2976 new->v.preferred_node = first_node(nodes);
2977 else
2978 new->flags |= MPOL_F_LOCAL;
2979
2980 /*
2981 * Save nodes for contextualization: this will be used to "clone"
2982 * the mempolicy in a specific context [cpuset] at a later time.
2983 */
2984 new->w.user_nodemask = nodes;
2985
2986 err = 0;
2987
2988out:
2989 /* Restore string for error message */
2990 if (nodelist)
2991 *--nodelist = ':';
2992 if (flags)
2993 *--flags = '=';
2994 if (!err)
2995 *mpol = new;
2996 return err;
2997}
2998#endif /* CONFIG_TMPFS */
2999
3000/**
3001 * mpol_to_str - format a mempolicy structure for printing
3002 * @buffer: to contain formatted mempolicy string
3003 * @maxlen: length of @buffer
3004 * @pol: pointer to mempolicy to be formatted
3005 *
3006 * Convert @pol into a string. If @buffer is too short, truncate the string.
3007 * Recommend a @maxlen of at least 32 for the longest mode, "interleave", the
3008 * longest flag, "relative", and to display at least a few node ids.
3009 */
3010void mpol_to_str(char *buffer, int maxlen, struct mempolicy *pol)
3011{
3012 char *p = buffer;
3013 nodemask_t nodes = NODE_MASK_NONE;
3014 unsigned short mode = MPOL_DEFAULT;
3015 unsigned short flags = 0;
3016
3017 if (pol && pol != &default_policy && !(pol->flags & MPOL_F_MORON)) {
3018 mode = pol->mode;
3019 flags = pol->flags;
3020 }
3021
3022 switch (mode) {
3023 case MPOL_DEFAULT:
3024 break;
3025 case MPOL_PREFERRED:
3026 if (flags & MPOL_F_LOCAL)
3027 mode = MPOL_LOCAL;
3028 else
3029 node_set(pol->v.preferred_node, nodes);
3030 break;
3031 case MPOL_BIND:
3032 case MPOL_INTERLEAVE:
3033 nodes = pol->v.nodes;
3034 break;
3035 default:
3036 WARN_ON_ONCE(1);
3037 snprintf(p, maxlen, "unknown");
3038 return;
3039 }
3040
3041 p += snprintf(p, maxlen, "%s", policy_modes[mode]);
3042
3043 if (flags & MPOL_MODE_FLAGS) {
3044 p += snprintf(p, buffer + maxlen - p, "=");
3045
3046 /*
3047 * Currently, the only defined flags are mutually exclusive
3048 */
3049 if (flags & MPOL_F_STATIC_NODES)
3050 p += snprintf(p, buffer + maxlen - p, "static");
3051 else if (flags & MPOL_F_RELATIVE_NODES)
3052 p += snprintf(p, buffer + maxlen - p, "relative");
3053 }
3054
3055 if (!nodes_empty(nodes))
3056 p += scnprintf(p, buffer + maxlen - p, ":%*pbl",
3057 nodemask_pr_args(&nodes));
3058}