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 six 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 * weighted interleave
23 * Allocate memory interleaved over a set of nodes based on
24 * a set of weights (per-node), with normal fallback if it
25 * fails. Otherwise operates the same as interleave.
26 * Example: nodeset(0,1) & weights (2,1) - 2 pages allocated
27 * on node 0 for every 1 page allocated on node 1.
28 *
29 * bind Only allocate memory on a specific set of nodes,
30 * no fallback.
31 * FIXME: memory is allocated starting with the first node
32 * to the last. It would be better if bind would truly restrict
33 * the allocation to memory nodes instead
34 *
35 * preferred Try a specific node first before normal fallback.
36 * As a special case NUMA_NO_NODE here means do the allocation
37 * on the local CPU. This is normally identical to default,
38 * but useful to set in a VMA when you have a non default
39 * process policy.
40 *
41 * preferred many Try a set of nodes first before normal fallback. This is
42 * similar to preferred without the special case.
43 *
44 * default Allocate on the local node first, or when on a VMA
45 * use the process policy. This is what Linux always did
46 * in a NUMA aware kernel and still does by, ahem, default.
47 *
48 * The process policy is applied for most non interrupt memory allocations
49 * in that process' context. Interrupts ignore the policies and always
50 * try to allocate on the local CPU. The VMA policy is only applied for memory
51 * allocations for a VMA in the VM.
52 *
53 * Currently there are a few corner cases in swapping where the policy
54 * is not applied, but the majority should be handled. When process policy
55 * is used it is not remembered over swap outs/swap ins.
56 *
57 * Only the highest zone in the zone hierarchy gets policied. Allocations
58 * requesting a lower zone just use default policy. This implies that
59 * on systems with highmem kernel lowmem allocation don't get policied.
60 * Same with GFP_DMA allocations.
61 *
62 * For shmem/tmpfs shared memory the policy is shared between
63 * all users and remembered even when nobody has memory mapped.
64 */
65
66/* Notebook:
67 fix mmap readahead to honour policy and enable policy for any page cache
68 object
69 statistics for bigpages
70 global policy for page cache? currently it uses process policy. Requires
71 first item above.
72 handle mremap for shared memory (currently ignored for the policy)
73 grows down?
74 make bind policy root only? It can trigger oom much faster and the
75 kernel is not always grateful with that.
76*/
77
78#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
79
80#include <linux/mempolicy.h>
81#include <linux/pagewalk.h>
82#include <linux/highmem.h>
83#include <linux/hugetlb.h>
84#include <linux/kernel.h>
85#include <linux/sched.h>
86#include <linux/sched/mm.h>
87#include <linux/sched/numa_balancing.h>
88#include <linux/sched/task.h>
89#include <linux/nodemask.h>
90#include <linux/cpuset.h>
91#include <linux/slab.h>
92#include <linux/string.h>
93#include <linux/export.h>
94#include <linux/nsproxy.h>
95#include <linux/interrupt.h>
96#include <linux/init.h>
97#include <linux/compat.h>
98#include <linux/ptrace.h>
99#include <linux/swap.h>
100#include <linux/seq_file.h>
101#include <linux/proc_fs.h>
102#include <linux/migrate.h>
103#include <linux/ksm.h>
104#include <linux/rmap.h>
105#include <linux/security.h>
106#include <linux/syscalls.h>
107#include <linux/ctype.h>
108#include <linux/mm_inline.h>
109#include <linux/mmu_notifier.h>
110#include <linux/printk.h>
111#include <linux/swapops.h>
112
113#include <asm/tlbflush.h>
114#include <asm/tlb.h>
115#include <linux/uaccess.h>
116
117#include "internal.h"
118
119/* Internal flags */
120#define MPOL_MF_DISCONTIG_OK (MPOL_MF_INTERNAL << 0) /* Skip checks for continuous vmas */
121#define MPOL_MF_INVERT (MPOL_MF_INTERNAL << 1) /* Invert check for nodemask */
122#define MPOL_MF_WRLOCK (MPOL_MF_INTERNAL << 2) /* Write-lock walked vmas */
123
124static struct kmem_cache *policy_cache;
125static struct kmem_cache *sn_cache;
126
127/* Highest zone. An specific allocation for a zone below that is not
128 policied. */
129enum zone_type policy_zone = 0;
130
131/*
132 * run-time system-wide default policy => local allocation
133 */
134static struct mempolicy default_policy = {
135 .refcnt = ATOMIC_INIT(1), /* never free it */
136 .mode = MPOL_LOCAL,
137};
138
139static struct mempolicy preferred_node_policy[MAX_NUMNODES];
140
141/*
142 * iw_table is the sysfs-set interleave weight table, a value of 0 denotes
143 * system-default value should be used. A NULL iw_table also denotes that
144 * system-default values should be used. Until the system-default table
145 * is implemented, the system-default is always 1.
146 *
147 * iw_table is RCU protected
148 */
149static u8 __rcu *iw_table;
150static DEFINE_MUTEX(iw_table_lock);
151
152static u8 get_il_weight(int node)
153{
154 u8 *table;
155 u8 weight;
156
157 rcu_read_lock();
158 table = rcu_dereference(iw_table);
159 /* if no iw_table, use system default */
160 weight = table ? table[node] : 1;
161 /* if value in iw_table is 0, use system default */
162 weight = weight ? weight : 1;
163 rcu_read_unlock();
164 return weight;
165}
166
167/**
168 * numa_nearest_node - Find nearest node by state
169 * @node: Node id to start the search
170 * @state: State to filter the search
171 *
172 * Lookup the closest node by distance if @nid is not in state.
173 *
174 * Return: this @node if it is in state, otherwise the closest node by distance
175 */
176int numa_nearest_node(int node, unsigned int state)
177{
178 int min_dist = INT_MAX, dist, n, min_node;
179
180 if (state >= NR_NODE_STATES)
181 return -EINVAL;
182
183 if (node == NUMA_NO_NODE || node_state(node, state))
184 return node;
185
186 min_node = node;
187 for_each_node_state(n, state) {
188 dist = node_distance(node, n);
189 if (dist < min_dist) {
190 min_dist = dist;
191 min_node = n;
192 }
193 }
194
195 return min_node;
196}
197EXPORT_SYMBOL_GPL(numa_nearest_node);
198
199struct mempolicy *get_task_policy(struct task_struct *p)
200{
201 struct mempolicy *pol = p->mempolicy;
202 int node;
203
204 if (pol)
205 return pol;
206
207 node = numa_node_id();
208 if (node != NUMA_NO_NODE) {
209 pol = &preferred_node_policy[node];
210 /* preferred_node_policy is not initialised early in boot */
211 if (pol->mode)
212 return pol;
213 }
214
215 return &default_policy;
216}
217
218static const struct mempolicy_operations {
219 int (*create)(struct mempolicy *pol, const nodemask_t *nodes);
220 void (*rebind)(struct mempolicy *pol, const nodemask_t *nodes);
221} mpol_ops[MPOL_MAX];
222
223static inline int mpol_store_user_nodemask(const struct mempolicy *pol)
224{
225 return pol->flags & MPOL_MODE_FLAGS;
226}
227
228static void mpol_relative_nodemask(nodemask_t *ret, const nodemask_t *orig,
229 const nodemask_t *rel)
230{
231 nodemask_t tmp;
232 nodes_fold(tmp, *orig, nodes_weight(*rel));
233 nodes_onto(*ret, tmp, *rel);
234}
235
236static int mpol_new_nodemask(struct mempolicy *pol, const nodemask_t *nodes)
237{
238 if (nodes_empty(*nodes))
239 return -EINVAL;
240 pol->nodes = *nodes;
241 return 0;
242}
243
244static int mpol_new_preferred(struct mempolicy *pol, const nodemask_t *nodes)
245{
246 if (nodes_empty(*nodes))
247 return -EINVAL;
248
249 nodes_clear(pol->nodes);
250 node_set(first_node(*nodes), pol->nodes);
251 return 0;
252}
253
254/*
255 * mpol_set_nodemask is called after mpol_new() to set up the nodemask, if
256 * any, for the new policy. mpol_new() has already validated the nodes
257 * parameter with respect to the policy mode and flags.
258 *
259 * Must be called holding task's alloc_lock to protect task's mems_allowed
260 * and mempolicy. May also be called holding the mmap_lock for write.
261 */
262static int mpol_set_nodemask(struct mempolicy *pol,
263 const nodemask_t *nodes, struct nodemask_scratch *nsc)
264{
265 int ret;
266
267 /*
268 * Default (pol==NULL) resp. local memory policies are not a
269 * subject of any remapping. They also do not need any special
270 * constructor.
271 */
272 if (!pol || pol->mode == MPOL_LOCAL)
273 return 0;
274
275 /* Check N_MEMORY */
276 nodes_and(nsc->mask1,
277 cpuset_current_mems_allowed, node_states[N_MEMORY]);
278
279 VM_BUG_ON(!nodes);
280
281 if (pol->flags & MPOL_F_RELATIVE_NODES)
282 mpol_relative_nodemask(&nsc->mask2, nodes, &nsc->mask1);
283 else
284 nodes_and(nsc->mask2, *nodes, nsc->mask1);
285
286 if (mpol_store_user_nodemask(pol))
287 pol->w.user_nodemask = *nodes;
288 else
289 pol->w.cpuset_mems_allowed = cpuset_current_mems_allowed;
290
291 ret = mpol_ops[pol->mode].create(pol, &nsc->mask2);
292 return ret;
293}
294
295/*
296 * This function just creates a new policy, does some check and simple
297 * initialization. You must invoke mpol_set_nodemask() to set nodes.
298 */
299static struct mempolicy *mpol_new(unsigned short mode, unsigned short flags,
300 nodemask_t *nodes)
301{
302 struct mempolicy *policy;
303
304 if (mode == MPOL_DEFAULT) {
305 if (nodes && !nodes_empty(*nodes))
306 return ERR_PTR(-EINVAL);
307 return NULL;
308 }
309 VM_BUG_ON(!nodes);
310
311 /*
312 * MPOL_PREFERRED cannot be used with MPOL_F_STATIC_NODES or
313 * MPOL_F_RELATIVE_NODES if the nodemask is empty (local allocation).
314 * All other modes require a valid pointer to a non-empty nodemask.
315 */
316 if (mode == MPOL_PREFERRED) {
317 if (nodes_empty(*nodes)) {
318 if (((flags & MPOL_F_STATIC_NODES) ||
319 (flags & MPOL_F_RELATIVE_NODES)))
320 return ERR_PTR(-EINVAL);
321
322 mode = MPOL_LOCAL;
323 }
324 } else if (mode == MPOL_LOCAL) {
325 if (!nodes_empty(*nodes) ||
326 (flags & MPOL_F_STATIC_NODES) ||
327 (flags & MPOL_F_RELATIVE_NODES))
328 return ERR_PTR(-EINVAL);
329 } else if (nodes_empty(*nodes))
330 return ERR_PTR(-EINVAL);
331
332 policy = kmem_cache_alloc(policy_cache, GFP_KERNEL);
333 if (!policy)
334 return ERR_PTR(-ENOMEM);
335 atomic_set(&policy->refcnt, 1);
336 policy->mode = mode;
337 policy->flags = flags;
338 policy->home_node = NUMA_NO_NODE;
339
340 return policy;
341}
342
343/* Slow path of a mpol destructor. */
344void __mpol_put(struct mempolicy *pol)
345{
346 if (!atomic_dec_and_test(&pol->refcnt))
347 return;
348 kmem_cache_free(policy_cache, pol);
349}
350
351static void mpol_rebind_default(struct mempolicy *pol, const nodemask_t *nodes)
352{
353}
354
355static void mpol_rebind_nodemask(struct mempolicy *pol, const nodemask_t *nodes)
356{
357 nodemask_t tmp;
358
359 if (pol->flags & MPOL_F_STATIC_NODES)
360 nodes_and(tmp, pol->w.user_nodemask, *nodes);
361 else if (pol->flags & MPOL_F_RELATIVE_NODES)
362 mpol_relative_nodemask(&tmp, &pol->w.user_nodemask, nodes);
363 else {
364 nodes_remap(tmp, pol->nodes, pol->w.cpuset_mems_allowed,
365 *nodes);
366 pol->w.cpuset_mems_allowed = *nodes;
367 }
368
369 if (nodes_empty(tmp))
370 tmp = *nodes;
371
372 pol->nodes = tmp;
373}
374
375static void mpol_rebind_preferred(struct mempolicy *pol,
376 const nodemask_t *nodes)
377{
378 pol->w.cpuset_mems_allowed = *nodes;
379}
380
381/*
382 * mpol_rebind_policy - Migrate a policy to a different set of nodes
383 *
384 * Per-vma policies are protected by mmap_lock. Allocations using per-task
385 * policies are protected by task->mems_allowed_seq to prevent a premature
386 * OOM/allocation failure due to parallel nodemask modification.
387 */
388static void mpol_rebind_policy(struct mempolicy *pol, const nodemask_t *newmask)
389{
390 if (!pol || pol->mode == MPOL_LOCAL)
391 return;
392 if (!mpol_store_user_nodemask(pol) &&
393 nodes_equal(pol->w.cpuset_mems_allowed, *newmask))
394 return;
395
396 mpol_ops[pol->mode].rebind(pol, newmask);
397}
398
399/*
400 * Wrapper for mpol_rebind_policy() that just requires task
401 * pointer, and updates task mempolicy.
402 *
403 * Called with task's alloc_lock held.
404 */
405void mpol_rebind_task(struct task_struct *tsk, const nodemask_t *new)
406{
407 mpol_rebind_policy(tsk->mempolicy, new);
408}
409
410/*
411 * Rebind each vma in mm to new nodemask.
412 *
413 * Call holding a reference to mm. Takes mm->mmap_lock during call.
414 */
415void mpol_rebind_mm(struct mm_struct *mm, nodemask_t *new)
416{
417 struct vm_area_struct *vma;
418 VMA_ITERATOR(vmi, mm, 0);
419
420 mmap_write_lock(mm);
421 for_each_vma(vmi, vma) {
422 vma_start_write(vma);
423 mpol_rebind_policy(vma->vm_policy, new);
424 }
425 mmap_write_unlock(mm);
426}
427
428static const struct mempolicy_operations mpol_ops[MPOL_MAX] = {
429 [MPOL_DEFAULT] = {
430 .rebind = mpol_rebind_default,
431 },
432 [MPOL_INTERLEAVE] = {
433 .create = mpol_new_nodemask,
434 .rebind = mpol_rebind_nodemask,
435 },
436 [MPOL_PREFERRED] = {
437 .create = mpol_new_preferred,
438 .rebind = mpol_rebind_preferred,
439 },
440 [MPOL_BIND] = {
441 .create = mpol_new_nodemask,
442 .rebind = mpol_rebind_nodemask,
443 },
444 [MPOL_LOCAL] = {
445 .rebind = mpol_rebind_default,
446 },
447 [MPOL_PREFERRED_MANY] = {
448 .create = mpol_new_nodemask,
449 .rebind = mpol_rebind_preferred,
450 },
451 [MPOL_WEIGHTED_INTERLEAVE] = {
452 .create = mpol_new_nodemask,
453 .rebind = mpol_rebind_nodemask,
454 },
455};
456
457static bool migrate_folio_add(struct folio *folio, struct list_head *foliolist,
458 unsigned long flags);
459static nodemask_t *policy_nodemask(gfp_t gfp, struct mempolicy *pol,
460 pgoff_t ilx, int *nid);
461
462static bool strictly_unmovable(unsigned long flags)
463{
464 /*
465 * STRICT without MOVE flags lets do_mbind() fail immediately with -EIO
466 * if any misplaced page is found.
467 */
468 return (flags & (MPOL_MF_STRICT | MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) ==
469 MPOL_MF_STRICT;
470}
471
472struct migration_mpol { /* for alloc_migration_target_by_mpol() */
473 struct mempolicy *pol;
474 pgoff_t ilx;
475};
476
477struct queue_pages {
478 struct list_head *pagelist;
479 unsigned long flags;
480 nodemask_t *nmask;
481 unsigned long start;
482 unsigned long end;
483 struct vm_area_struct *first;
484 struct folio *large; /* note last large folio encountered */
485 long nr_failed; /* could not be isolated at this time */
486};
487
488/*
489 * Check if the folio's nid is in qp->nmask.
490 *
491 * If MPOL_MF_INVERT is set in qp->flags, check if the nid is
492 * in the invert of qp->nmask.
493 */
494static inline bool queue_folio_required(struct folio *folio,
495 struct queue_pages *qp)
496{
497 int nid = folio_nid(folio);
498 unsigned long flags = qp->flags;
499
500 return node_isset(nid, *qp->nmask) == !(flags & MPOL_MF_INVERT);
501}
502
503static void queue_folios_pmd(pmd_t *pmd, struct mm_walk *walk)
504{
505 struct folio *folio;
506 struct queue_pages *qp = walk->private;
507
508 if (unlikely(is_pmd_migration_entry(*pmd))) {
509 qp->nr_failed++;
510 return;
511 }
512 folio = pmd_folio(*pmd);
513 if (is_huge_zero_folio(folio)) {
514 walk->action = ACTION_CONTINUE;
515 return;
516 }
517 if (!queue_folio_required(folio, qp))
518 return;
519 if (!(qp->flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) ||
520 !vma_migratable(walk->vma) ||
521 !migrate_folio_add(folio, qp->pagelist, qp->flags))
522 qp->nr_failed++;
523}
524
525/*
526 * Scan through folios, checking if they satisfy the required conditions,
527 * moving them from LRU to local pagelist for migration if they do (or not).
528 *
529 * queue_folios_pte_range() has two possible return values:
530 * 0 - continue walking to scan for more, even if an existing folio on the
531 * wrong node could not be isolated and queued for migration.
532 * -EIO - only MPOL_MF_STRICT was specified, without MPOL_MF_MOVE or ..._ALL,
533 * and an existing folio was on a node that does not follow the policy.
534 */
535static int queue_folios_pte_range(pmd_t *pmd, unsigned long addr,
536 unsigned long end, struct mm_walk *walk)
537{
538 struct vm_area_struct *vma = walk->vma;
539 struct folio *folio;
540 struct queue_pages *qp = walk->private;
541 unsigned long flags = qp->flags;
542 pte_t *pte, *mapped_pte;
543 pte_t ptent;
544 spinlock_t *ptl;
545
546 ptl = pmd_trans_huge_lock(pmd, vma);
547 if (ptl) {
548 queue_folios_pmd(pmd, walk);
549 spin_unlock(ptl);
550 goto out;
551 }
552
553 mapped_pte = pte = pte_offset_map_lock(walk->mm, pmd, addr, &ptl);
554 if (!pte) {
555 walk->action = ACTION_AGAIN;
556 return 0;
557 }
558 for (; addr != end; pte++, addr += PAGE_SIZE) {
559 ptent = ptep_get(pte);
560 if (pte_none(ptent))
561 continue;
562 if (!pte_present(ptent)) {
563 if (is_migration_entry(pte_to_swp_entry(ptent)))
564 qp->nr_failed++;
565 continue;
566 }
567 folio = vm_normal_folio(vma, addr, ptent);
568 if (!folio || folio_is_zone_device(folio))
569 continue;
570 /*
571 * vm_normal_folio() filters out zero pages, but there might
572 * still be reserved folios to skip, perhaps in a VDSO.
573 */
574 if (folio_test_reserved(folio))
575 continue;
576 if (!queue_folio_required(folio, qp))
577 continue;
578 if (folio_test_large(folio)) {
579 /*
580 * A large folio can only be isolated from LRU once,
581 * but may be mapped by many PTEs (and Copy-On-Write may
582 * intersperse PTEs of other, order 0, folios). This is
583 * a common case, so don't mistake it for failure (but
584 * there can be other cases of multi-mapped pages which
585 * this quick check does not help to filter out - and a
586 * search of the pagelist might grow to be prohibitive).
587 *
588 * migrate_pages(&pagelist) returns nr_failed folios, so
589 * check "large" now so that queue_pages_range() returns
590 * a comparable nr_failed folios. This does imply that
591 * if folio could not be isolated for some racy reason
592 * at its first PTE, later PTEs will not give it another
593 * chance of isolation; but keeps the accounting simple.
594 */
595 if (folio == qp->large)
596 continue;
597 qp->large = folio;
598 }
599 if (!(flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) ||
600 !vma_migratable(vma) ||
601 !migrate_folio_add(folio, qp->pagelist, flags)) {
602 qp->nr_failed++;
603 if (strictly_unmovable(flags))
604 break;
605 }
606 }
607 pte_unmap_unlock(mapped_pte, ptl);
608 cond_resched();
609out:
610 if (qp->nr_failed && strictly_unmovable(flags))
611 return -EIO;
612 return 0;
613}
614
615static int queue_folios_hugetlb(pte_t *pte, unsigned long hmask,
616 unsigned long addr, unsigned long end,
617 struct mm_walk *walk)
618{
619#ifdef CONFIG_HUGETLB_PAGE
620 struct queue_pages *qp = walk->private;
621 unsigned long flags = qp->flags;
622 struct folio *folio;
623 spinlock_t *ptl;
624 pte_t entry;
625
626 ptl = huge_pte_lock(hstate_vma(walk->vma), walk->mm, pte);
627 entry = huge_ptep_get(walk->mm, addr, pte);
628 if (!pte_present(entry)) {
629 if (unlikely(is_hugetlb_entry_migration(entry)))
630 qp->nr_failed++;
631 goto unlock;
632 }
633 folio = pfn_folio(pte_pfn(entry));
634 if (!queue_folio_required(folio, qp))
635 goto unlock;
636 if (!(flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) ||
637 !vma_migratable(walk->vma)) {
638 qp->nr_failed++;
639 goto unlock;
640 }
641 /*
642 * Unless MPOL_MF_MOVE_ALL, we try to avoid migrating a shared folio.
643 * Choosing not to migrate a shared folio is not counted as a failure.
644 *
645 * See folio_likely_mapped_shared() on possible imprecision when we
646 * cannot easily detect if a folio is shared.
647 */
648 if ((flags & MPOL_MF_MOVE_ALL) ||
649 (!folio_likely_mapped_shared(folio) && !hugetlb_pmd_shared(pte)))
650 if (!isolate_hugetlb(folio, qp->pagelist))
651 qp->nr_failed++;
652unlock:
653 spin_unlock(ptl);
654 if (qp->nr_failed && strictly_unmovable(flags))
655 return -EIO;
656#endif
657 return 0;
658}
659
660#ifdef CONFIG_NUMA_BALANCING
661/*
662 * This is used to mark a range of virtual addresses to be inaccessible.
663 * These are later cleared by a NUMA hinting fault. Depending on these
664 * faults, pages may be migrated for better NUMA placement.
665 *
666 * This is assuming that NUMA faults are handled using PROT_NONE. If
667 * an architecture makes a different choice, it will need further
668 * changes to the core.
669 */
670unsigned long change_prot_numa(struct vm_area_struct *vma,
671 unsigned long addr, unsigned long end)
672{
673 struct mmu_gather tlb;
674 long nr_updated;
675
676 tlb_gather_mmu(&tlb, vma->vm_mm);
677
678 nr_updated = change_protection(&tlb, vma, addr, end, MM_CP_PROT_NUMA);
679 if (nr_updated > 0) {
680 count_vm_numa_events(NUMA_PTE_UPDATES, nr_updated);
681 count_memcg_events_mm(vma->vm_mm, NUMA_PTE_UPDATES, nr_updated);
682 }
683
684 tlb_finish_mmu(&tlb);
685
686 return nr_updated;
687}
688#endif /* CONFIG_NUMA_BALANCING */
689
690static int queue_pages_test_walk(unsigned long start, unsigned long end,
691 struct mm_walk *walk)
692{
693 struct vm_area_struct *next, *vma = walk->vma;
694 struct queue_pages *qp = walk->private;
695 unsigned long flags = qp->flags;
696
697 /* range check first */
698 VM_BUG_ON_VMA(!range_in_vma(vma, start, end), vma);
699
700 if (!qp->first) {
701 qp->first = vma;
702 if (!(flags & MPOL_MF_DISCONTIG_OK) &&
703 (qp->start < vma->vm_start))
704 /* hole at head side of range */
705 return -EFAULT;
706 }
707 next = find_vma(vma->vm_mm, vma->vm_end);
708 if (!(flags & MPOL_MF_DISCONTIG_OK) &&
709 ((vma->vm_end < qp->end) &&
710 (!next || vma->vm_end < next->vm_start)))
711 /* hole at middle or tail of range */
712 return -EFAULT;
713
714 /*
715 * Need check MPOL_MF_STRICT to return -EIO if possible
716 * regardless of vma_migratable
717 */
718 if (!vma_migratable(vma) &&
719 !(flags & MPOL_MF_STRICT))
720 return 1;
721
722 /*
723 * Check page nodes, and queue pages to move, in the current vma.
724 * But if no moving, and no strict checking, the scan can be skipped.
725 */
726 if (flags & (MPOL_MF_STRICT | MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))
727 return 0;
728 return 1;
729}
730
731static const struct mm_walk_ops queue_pages_walk_ops = {
732 .hugetlb_entry = queue_folios_hugetlb,
733 .pmd_entry = queue_folios_pte_range,
734 .test_walk = queue_pages_test_walk,
735 .walk_lock = PGWALK_RDLOCK,
736};
737
738static const struct mm_walk_ops queue_pages_lock_vma_walk_ops = {
739 .hugetlb_entry = queue_folios_hugetlb,
740 .pmd_entry = queue_folios_pte_range,
741 .test_walk = queue_pages_test_walk,
742 .walk_lock = PGWALK_WRLOCK,
743};
744
745/*
746 * Walk through page tables and collect pages to be migrated.
747 *
748 * If pages found in a given range are not on the required set of @nodes,
749 * and migration is allowed, they are isolated and queued to @pagelist.
750 *
751 * queue_pages_range() may return:
752 * 0 - all pages already on the right node, or successfully queued for moving
753 * (or neither strict checking nor moving requested: only range checking).
754 * >0 - this number of misplaced folios could not be queued for moving
755 * (a hugetlbfs page or a transparent huge page being counted as 1).
756 * -EIO - a misplaced page found, when MPOL_MF_STRICT specified without MOVEs.
757 * -EFAULT - a hole in the memory range, when MPOL_MF_DISCONTIG_OK unspecified.
758 */
759static long
760queue_pages_range(struct mm_struct *mm, unsigned long start, unsigned long end,
761 nodemask_t *nodes, unsigned long flags,
762 struct list_head *pagelist)
763{
764 int err;
765 struct queue_pages qp = {
766 .pagelist = pagelist,
767 .flags = flags,
768 .nmask = nodes,
769 .start = start,
770 .end = end,
771 .first = NULL,
772 };
773 const struct mm_walk_ops *ops = (flags & MPOL_MF_WRLOCK) ?
774 &queue_pages_lock_vma_walk_ops : &queue_pages_walk_ops;
775
776 err = walk_page_range(mm, start, end, ops, &qp);
777
778 if (!qp.first)
779 /* whole range in hole */
780 err = -EFAULT;
781
782 return err ? : qp.nr_failed;
783}
784
785/*
786 * Apply policy to a single VMA
787 * This must be called with the mmap_lock held for writing.
788 */
789static int vma_replace_policy(struct vm_area_struct *vma,
790 struct mempolicy *pol)
791{
792 int err;
793 struct mempolicy *old;
794 struct mempolicy *new;
795
796 vma_assert_write_locked(vma);
797
798 new = mpol_dup(pol);
799 if (IS_ERR(new))
800 return PTR_ERR(new);
801
802 if (vma->vm_ops && vma->vm_ops->set_policy) {
803 err = vma->vm_ops->set_policy(vma, new);
804 if (err)
805 goto err_out;
806 }
807
808 old = vma->vm_policy;
809 vma->vm_policy = new; /* protected by mmap_lock */
810 mpol_put(old);
811
812 return 0;
813 err_out:
814 mpol_put(new);
815 return err;
816}
817
818/* Split or merge the VMA (if required) and apply the new policy */
819static int mbind_range(struct vma_iterator *vmi, struct vm_area_struct *vma,
820 struct vm_area_struct **prev, unsigned long start,
821 unsigned long end, struct mempolicy *new_pol)
822{
823 unsigned long vmstart, vmend;
824
825 vmend = min(end, vma->vm_end);
826 if (start > vma->vm_start) {
827 *prev = vma;
828 vmstart = start;
829 } else {
830 vmstart = vma->vm_start;
831 }
832
833 if (mpol_equal(vma->vm_policy, new_pol)) {
834 *prev = vma;
835 return 0;
836 }
837
838 vma = vma_modify_policy(vmi, *prev, vma, vmstart, vmend, new_pol);
839 if (IS_ERR(vma))
840 return PTR_ERR(vma);
841
842 *prev = vma;
843 return vma_replace_policy(vma, new_pol);
844}
845
846/* Set the process memory policy */
847static long do_set_mempolicy(unsigned short mode, unsigned short flags,
848 nodemask_t *nodes)
849{
850 struct mempolicy *new, *old;
851 NODEMASK_SCRATCH(scratch);
852 int ret;
853
854 if (!scratch)
855 return -ENOMEM;
856
857 new = mpol_new(mode, flags, nodes);
858 if (IS_ERR(new)) {
859 ret = PTR_ERR(new);
860 goto out;
861 }
862
863 task_lock(current);
864 ret = mpol_set_nodemask(new, nodes, scratch);
865 if (ret) {
866 task_unlock(current);
867 mpol_put(new);
868 goto out;
869 }
870
871 old = current->mempolicy;
872 current->mempolicy = new;
873 if (new && (new->mode == MPOL_INTERLEAVE ||
874 new->mode == MPOL_WEIGHTED_INTERLEAVE)) {
875 current->il_prev = MAX_NUMNODES-1;
876 current->il_weight = 0;
877 }
878 task_unlock(current);
879 mpol_put(old);
880 ret = 0;
881out:
882 NODEMASK_SCRATCH_FREE(scratch);
883 return ret;
884}
885
886/*
887 * Return nodemask for policy for get_mempolicy() query
888 *
889 * Called with task's alloc_lock held
890 */
891static void get_policy_nodemask(struct mempolicy *pol, nodemask_t *nodes)
892{
893 nodes_clear(*nodes);
894 if (pol == &default_policy)
895 return;
896
897 switch (pol->mode) {
898 case MPOL_BIND:
899 case MPOL_INTERLEAVE:
900 case MPOL_PREFERRED:
901 case MPOL_PREFERRED_MANY:
902 case MPOL_WEIGHTED_INTERLEAVE:
903 *nodes = pol->nodes;
904 break;
905 case MPOL_LOCAL:
906 /* return empty node mask for local allocation */
907 break;
908 default:
909 BUG();
910 }
911}
912
913static int lookup_node(struct mm_struct *mm, unsigned long addr)
914{
915 struct page *p = NULL;
916 int ret;
917
918 ret = get_user_pages_fast(addr & PAGE_MASK, 1, 0, &p);
919 if (ret > 0) {
920 ret = page_to_nid(p);
921 put_page(p);
922 }
923 return ret;
924}
925
926/* Retrieve NUMA policy */
927static long do_get_mempolicy(int *policy, nodemask_t *nmask,
928 unsigned long addr, unsigned long flags)
929{
930 int err;
931 struct mm_struct *mm = current->mm;
932 struct vm_area_struct *vma = NULL;
933 struct mempolicy *pol = current->mempolicy, *pol_refcount = NULL;
934
935 if (flags &
936 ~(unsigned long)(MPOL_F_NODE|MPOL_F_ADDR|MPOL_F_MEMS_ALLOWED))
937 return -EINVAL;
938
939 if (flags & MPOL_F_MEMS_ALLOWED) {
940 if (flags & (MPOL_F_NODE|MPOL_F_ADDR))
941 return -EINVAL;
942 *policy = 0; /* just so it's initialized */
943 task_lock(current);
944 *nmask = cpuset_current_mems_allowed;
945 task_unlock(current);
946 return 0;
947 }
948
949 if (flags & MPOL_F_ADDR) {
950 pgoff_t ilx; /* ignored here */
951 /*
952 * Do NOT fall back to task policy if the
953 * vma/shared policy at addr is NULL. We
954 * want to return MPOL_DEFAULT in this case.
955 */
956 mmap_read_lock(mm);
957 vma = vma_lookup(mm, addr);
958 if (!vma) {
959 mmap_read_unlock(mm);
960 return -EFAULT;
961 }
962 pol = __get_vma_policy(vma, addr, &ilx);
963 } else if (addr)
964 return -EINVAL;
965
966 if (!pol)
967 pol = &default_policy; /* indicates default behavior */
968
969 if (flags & MPOL_F_NODE) {
970 if (flags & MPOL_F_ADDR) {
971 /*
972 * Take a refcount on the mpol, because we are about to
973 * drop the mmap_lock, after which only "pol" remains
974 * valid, "vma" is stale.
975 */
976 pol_refcount = pol;
977 vma = NULL;
978 mpol_get(pol);
979 mmap_read_unlock(mm);
980 err = lookup_node(mm, addr);
981 if (err < 0)
982 goto out;
983 *policy = err;
984 } else if (pol == current->mempolicy &&
985 pol->mode == MPOL_INTERLEAVE) {
986 *policy = next_node_in(current->il_prev, pol->nodes);
987 } else if (pol == current->mempolicy &&
988 pol->mode == MPOL_WEIGHTED_INTERLEAVE) {
989 if (current->il_weight)
990 *policy = current->il_prev;
991 else
992 *policy = next_node_in(current->il_prev,
993 pol->nodes);
994 } else {
995 err = -EINVAL;
996 goto out;
997 }
998 } else {
999 *policy = pol == &default_policy ? MPOL_DEFAULT :
1000 pol->mode;
1001 /*
1002 * Internal mempolicy flags must be masked off before exposing
1003 * the policy to userspace.
1004 */
1005 *policy |= (pol->flags & MPOL_MODE_FLAGS);
1006 }
1007
1008 err = 0;
1009 if (nmask) {
1010 if (mpol_store_user_nodemask(pol)) {
1011 *nmask = pol->w.user_nodemask;
1012 } else {
1013 task_lock(current);
1014 get_policy_nodemask(pol, nmask);
1015 task_unlock(current);
1016 }
1017 }
1018
1019 out:
1020 mpol_cond_put(pol);
1021 if (vma)
1022 mmap_read_unlock(mm);
1023 if (pol_refcount)
1024 mpol_put(pol_refcount);
1025 return err;
1026}
1027
1028#ifdef CONFIG_MIGRATION
1029static bool migrate_folio_add(struct folio *folio, struct list_head *foliolist,
1030 unsigned long flags)
1031{
1032 /*
1033 * Unless MPOL_MF_MOVE_ALL, we try to avoid migrating a shared folio.
1034 * Choosing not to migrate a shared folio is not counted as a failure.
1035 *
1036 * See folio_likely_mapped_shared() on possible imprecision when we
1037 * cannot easily detect if a folio is shared.
1038 */
1039 if ((flags & MPOL_MF_MOVE_ALL) || !folio_likely_mapped_shared(folio)) {
1040 if (folio_isolate_lru(folio)) {
1041 list_add_tail(&folio->lru, foliolist);
1042 node_stat_mod_folio(folio,
1043 NR_ISOLATED_ANON + folio_is_file_lru(folio),
1044 folio_nr_pages(folio));
1045 } else {
1046 /*
1047 * Non-movable folio may reach here. And, there may be
1048 * temporary off LRU folios or non-LRU movable folios.
1049 * Treat them as unmovable folios since they can't be
1050 * isolated, so they can't be moved at the moment.
1051 */
1052 return false;
1053 }
1054 }
1055 return true;
1056}
1057
1058/*
1059 * Migrate pages from one node to a target node.
1060 * Returns error or the number of pages not migrated.
1061 */
1062static long migrate_to_node(struct mm_struct *mm, int source, int dest,
1063 int flags)
1064{
1065 nodemask_t nmask;
1066 struct vm_area_struct *vma;
1067 LIST_HEAD(pagelist);
1068 long nr_failed;
1069 long err = 0;
1070 struct migration_target_control mtc = {
1071 .nid = dest,
1072 .gfp_mask = GFP_HIGHUSER_MOVABLE | __GFP_THISNODE,
1073 .reason = MR_SYSCALL,
1074 };
1075
1076 nodes_clear(nmask);
1077 node_set(source, nmask);
1078
1079 VM_BUG_ON(!(flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)));
1080
1081 mmap_read_lock(mm);
1082 vma = find_vma(mm, 0);
1083 if (unlikely(!vma)) {
1084 mmap_read_unlock(mm);
1085 return 0;
1086 }
1087
1088 /*
1089 * This does not migrate the range, but isolates all pages that
1090 * need migration. Between passing in the full user address
1091 * space range and MPOL_MF_DISCONTIG_OK, this call cannot fail,
1092 * but passes back the count of pages which could not be isolated.
1093 */
1094 nr_failed = queue_pages_range(mm, vma->vm_start, mm->task_size, &nmask,
1095 flags | MPOL_MF_DISCONTIG_OK, &pagelist);
1096 mmap_read_unlock(mm);
1097
1098 if (!list_empty(&pagelist)) {
1099 err = migrate_pages(&pagelist, alloc_migration_target, NULL,
1100 (unsigned long)&mtc, MIGRATE_SYNC, MR_SYSCALL, NULL);
1101 if (err)
1102 putback_movable_pages(&pagelist);
1103 }
1104
1105 if (err >= 0)
1106 err += nr_failed;
1107 return err;
1108}
1109
1110/*
1111 * Move pages between the two nodesets so as to preserve the physical
1112 * layout as much as possible.
1113 *
1114 * Returns the number of page that could not be moved.
1115 */
1116int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
1117 const nodemask_t *to, int flags)
1118{
1119 long nr_failed = 0;
1120 long err = 0;
1121 nodemask_t tmp;
1122
1123 lru_cache_disable();
1124
1125 /*
1126 * Find a 'source' bit set in 'tmp' whose corresponding 'dest'
1127 * bit in 'to' is not also set in 'tmp'. Clear the found 'source'
1128 * bit in 'tmp', and return that <source, dest> pair for migration.
1129 * The pair of nodemasks 'to' and 'from' define the map.
1130 *
1131 * If no pair of bits is found that way, fallback to picking some
1132 * pair of 'source' and 'dest' bits that are not the same. If the
1133 * 'source' and 'dest' bits are the same, this represents a node
1134 * that will be migrating to itself, so no pages need move.
1135 *
1136 * If no bits are left in 'tmp', or if all remaining bits left
1137 * in 'tmp' correspond to the same bit in 'to', return false
1138 * (nothing left to migrate).
1139 *
1140 * This lets us pick a pair of nodes to migrate between, such that
1141 * if possible the dest node is not already occupied by some other
1142 * source node, minimizing the risk of overloading the memory on a
1143 * node that would happen if we migrated incoming memory to a node
1144 * before migrating outgoing memory source that same node.
1145 *
1146 * A single scan of tmp is sufficient. As we go, we remember the
1147 * most recent <s, d> pair that moved (s != d). If we find a pair
1148 * that not only moved, but what's better, moved to an empty slot
1149 * (d is not set in tmp), then we break out then, with that pair.
1150 * Otherwise when we finish scanning from_tmp, we at least have the
1151 * most recent <s, d> pair that moved. If we get all the way through
1152 * the scan of tmp without finding any node that moved, much less
1153 * moved to an empty node, then there is nothing left worth migrating.
1154 */
1155
1156 tmp = *from;
1157 while (!nodes_empty(tmp)) {
1158 int s, d;
1159 int source = NUMA_NO_NODE;
1160 int dest = 0;
1161
1162 for_each_node_mask(s, tmp) {
1163
1164 /*
1165 * do_migrate_pages() tries to maintain the relative
1166 * node relationship of the pages established between
1167 * threads and memory areas.
1168 *
1169 * However if the number of source nodes is not equal to
1170 * the number of destination nodes we can not preserve
1171 * this node relative relationship. In that case, skip
1172 * copying memory from a node that is in the destination
1173 * mask.
1174 *
1175 * Example: [2,3,4] -> [3,4,5] moves everything.
1176 * [0-7] - > [3,4,5] moves only 0,1,2,6,7.
1177 */
1178
1179 if ((nodes_weight(*from) != nodes_weight(*to)) &&
1180 (node_isset(s, *to)))
1181 continue;
1182
1183 d = node_remap(s, *from, *to);
1184 if (s == d)
1185 continue;
1186
1187 source = s; /* Node moved. Memorize */
1188 dest = d;
1189
1190 /* dest not in remaining from nodes? */
1191 if (!node_isset(dest, tmp))
1192 break;
1193 }
1194 if (source == NUMA_NO_NODE)
1195 break;
1196
1197 node_clear(source, tmp);
1198 err = migrate_to_node(mm, source, dest, flags);
1199 if (err > 0)
1200 nr_failed += err;
1201 if (err < 0)
1202 break;
1203 }
1204
1205 lru_cache_enable();
1206 if (err < 0)
1207 return err;
1208 return (nr_failed < INT_MAX) ? nr_failed : INT_MAX;
1209}
1210
1211/*
1212 * Allocate a new folio for page migration, according to NUMA mempolicy.
1213 */
1214static struct folio *alloc_migration_target_by_mpol(struct folio *src,
1215 unsigned long private)
1216{
1217 struct migration_mpol *mmpol = (struct migration_mpol *)private;
1218 struct mempolicy *pol = mmpol->pol;
1219 pgoff_t ilx = mmpol->ilx;
1220 unsigned int order;
1221 int nid = numa_node_id();
1222 gfp_t gfp;
1223
1224 order = folio_order(src);
1225 ilx += src->index >> order;
1226
1227 if (folio_test_hugetlb(src)) {
1228 nodemask_t *nodemask;
1229 struct hstate *h;
1230
1231 h = folio_hstate(src);
1232 gfp = htlb_alloc_mask(h);
1233 nodemask = policy_nodemask(gfp, pol, ilx, &nid);
1234 return alloc_hugetlb_folio_nodemask(h, nid, nodemask, gfp,
1235 htlb_allow_alloc_fallback(MR_MEMPOLICY_MBIND));
1236 }
1237
1238 if (folio_test_large(src))
1239 gfp = GFP_TRANSHUGE;
1240 else
1241 gfp = GFP_HIGHUSER_MOVABLE | __GFP_RETRY_MAYFAIL | __GFP_COMP;
1242
1243 return folio_alloc_mpol(gfp, order, pol, ilx, nid);
1244}
1245#else
1246
1247static bool migrate_folio_add(struct folio *folio, struct list_head *foliolist,
1248 unsigned long flags)
1249{
1250 return false;
1251}
1252
1253int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
1254 const nodemask_t *to, int flags)
1255{
1256 return -ENOSYS;
1257}
1258
1259static struct folio *alloc_migration_target_by_mpol(struct folio *src,
1260 unsigned long private)
1261{
1262 return NULL;
1263}
1264#endif
1265
1266static long do_mbind(unsigned long start, unsigned long len,
1267 unsigned short mode, unsigned short mode_flags,
1268 nodemask_t *nmask, unsigned long flags)
1269{
1270 struct mm_struct *mm = current->mm;
1271 struct vm_area_struct *vma, *prev;
1272 struct vma_iterator vmi;
1273 struct migration_mpol mmpol;
1274 struct mempolicy *new;
1275 unsigned long end;
1276 long err;
1277 long nr_failed;
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 = PAGE_ALIGN(len);
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 /*
1304 * If we are using the default policy then operation
1305 * on discontinuous address spaces is okay after all
1306 */
1307 if (!new)
1308 flags |= MPOL_MF_DISCONTIG_OK;
1309
1310 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))
1311 lru_cache_disable();
1312 {
1313 NODEMASK_SCRATCH(scratch);
1314 if (scratch) {
1315 mmap_write_lock(mm);
1316 err = mpol_set_nodemask(new, nmask, scratch);
1317 if (err)
1318 mmap_write_unlock(mm);
1319 } else
1320 err = -ENOMEM;
1321 NODEMASK_SCRATCH_FREE(scratch);
1322 }
1323 if (err)
1324 goto mpol_out;
1325
1326 /*
1327 * Lock the VMAs before scanning for pages to migrate,
1328 * to ensure we don't miss a concurrently inserted page.
1329 */
1330 nr_failed = queue_pages_range(mm, start, end, nmask,
1331 flags | MPOL_MF_INVERT | MPOL_MF_WRLOCK, &pagelist);
1332
1333 if (nr_failed < 0) {
1334 err = nr_failed;
1335 nr_failed = 0;
1336 } else {
1337 vma_iter_init(&vmi, mm, start);
1338 prev = vma_prev(&vmi);
1339 for_each_vma_range(vmi, vma, end) {
1340 err = mbind_range(&vmi, vma, &prev, start, end, new);
1341 if (err)
1342 break;
1343 }
1344 }
1345
1346 if (!err && !list_empty(&pagelist)) {
1347 /* Convert MPOL_DEFAULT's NULL to task or default policy */
1348 if (!new) {
1349 new = get_task_policy(current);
1350 mpol_get(new);
1351 }
1352 mmpol.pol = new;
1353 mmpol.ilx = 0;
1354
1355 /*
1356 * In the interleaved case, attempt to allocate on exactly the
1357 * targeted nodes, for the first VMA to be migrated; for later
1358 * VMAs, the nodes will still be interleaved from the targeted
1359 * nodemask, but one by one may be selected differently.
1360 */
1361 if (new->mode == MPOL_INTERLEAVE ||
1362 new->mode == MPOL_WEIGHTED_INTERLEAVE) {
1363 struct folio *folio;
1364 unsigned int order;
1365 unsigned long addr = -EFAULT;
1366
1367 list_for_each_entry(folio, &pagelist, lru) {
1368 if (!folio_test_ksm(folio))
1369 break;
1370 }
1371 if (!list_entry_is_head(folio, &pagelist, lru)) {
1372 vma_iter_init(&vmi, mm, start);
1373 for_each_vma_range(vmi, vma, end) {
1374 addr = page_address_in_vma(folio,
1375 folio_page(folio, 0), vma);
1376 if (addr != -EFAULT)
1377 break;
1378 }
1379 }
1380 if (addr != -EFAULT) {
1381 order = folio_order(folio);
1382 /* We already know the pol, but not the ilx */
1383 mpol_cond_put(get_vma_policy(vma, addr, order,
1384 &mmpol.ilx));
1385 /* Set base from which to increment by index */
1386 mmpol.ilx -= folio->index >> order;
1387 }
1388 }
1389 }
1390
1391 mmap_write_unlock(mm);
1392
1393 if (!err && !list_empty(&pagelist)) {
1394 nr_failed |= migrate_pages(&pagelist,
1395 alloc_migration_target_by_mpol, NULL,
1396 (unsigned long)&mmpol, MIGRATE_SYNC,
1397 MR_MEMPOLICY_MBIND, NULL);
1398 }
1399
1400 if (nr_failed && (flags & MPOL_MF_STRICT))
1401 err = -EIO;
1402 if (!list_empty(&pagelist))
1403 putback_movable_pages(&pagelist);
1404mpol_out:
1405 mpol_put(new);
1406 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))
1407 lru_cache_enable();
1408 return err;
1409}
1410
1411/*
1412 * User space interface with variable sized bitmaps for nodelists.
1413 */
1414static int get_bitmap(unsigned long *mask, const unsigned long __user *nmask,
1415 unsigned long maxnode)
1416{
1417 unsigned long nlongs = BITS_TO_LONGS(maxnode);
1418 int ret;
1419
1420 if (in_compat_syscall())
1421 ret = compat_get_bitmap(mask,
1422 (const compat_ulong_t __user *)nmask,
1423 maxnode);
1424 else
1425 ret = copy_from_user(mask, nmask,
1426 nlongs * sizeof(unsigned long));
1427
1428 if (ret)
1429 return -EFAULT;
1430
1431 if (maxnode % BITS_PER_LONG)
1432 mask[nlongs - 1] &= (1UL << (maxnode % BITS_PER_LONG)) - 1;
1433
1434 return 0;
1435}
1436
1437/* Copy a node mask from user space. */
1438static int get_nodes(nodemask_t *nodes, const unsigned long __user *nmask,
1439 unsigned long maxnode)
1440{
1441 --maxnode;
1442 nodes_clear(*nodes);
1443 if (maxnode == 0 || !nmask)
1444 return 0;
1445 if (maxnode > PAGE_SIZE*BITS_PER_BYTE)
1446 return -EINVAL;
1447
1448 /*
1449 * When the user specified more nodes than supported just check
1450 * if the non supported part is all zero, one word at a time,
1451 * starting at the end.
1452 */
1453 while (maxnode > MAX_NUMNODES) {
1454 unsigned long bits = min_t(unsigned long, maxnode, BITS_PER_LONG);
1455 unsigned long t;
1456
1457 if (get_bitmap(&t, &nmask[(maxnode - 1) / BITS_PER_LONG], bits))
1458 return -EFAULT;
1459
1460 if (maxnode - bits >= MAX_NUMNODES) {
1461 maxnode -= bits;
1462 } else {
1463 maxnode = MAX_NUMNODES;
1464 t &= ~((1UL << (MAX_NUMNODES % BITS_PER_LONG)) - 1);
1465 }
1466 if (t)
1467 return -EINVAL;
1468 }
1469
1470 return get_bitmap(nodes_addr(*nodes), nmask, maxnode);
1471}
1472
1473/* Copy a kernel node mask to user space */
1474static int copy_nodes_to_user(unsigned long __user *mask, unsigned long maxnode,
1475 nodemask_t *nodes)
1476{
1477 unsigned long copy = ALIGN(maxnode-1, 64) / 8;
1478 unsigned int nbytes = BITS_TO_LONGS(nr_node_ids) * sizeof(long);
1479 bool compat = in_compat_syscall();
1480
1481 if (compat)
1482 nbytes = BITS_TO_COMPAT_LONGS(nr_node_ids) * sizeof(compat_long_t);
1483
1484 if (copy > nbytes) {
1485 if (copy > PAGE_SIZE)
1486 return -EINVAL;
1487 if (clear_user((char __user *)mask + nbytes, copy - nbytes))
1488 return -EFAULT;
1489 copy = nbytes;
1490 maxnode = nr_node_ids;
1491 }
1492
1493 if (compat)
1494 return compat_put_bitmap((compat_ulong_t __user *)mask,
1495 nodes_addr(*nodes), maxnode);
1496
1497 return copy_to_user(mask, nodes_addr(*nodes), copy) ? -EFAULT : 0;
1498}
1499
1500/* Basic parameter sanity check used by both mbind() and set_mempolicy() */
1501static inline int sanitize_mpol_flags(int *mode, unsigned short *flags)
1502{
1503 *flags = *mode & MPOL_MODE_FLAGS;
1504 *mode &= ~MPOL_MODE_FLAGS;
1505
1506 if ((unsigned int)(*mode) >= MPOL_MAX)
1507 return -EINVAL;
1508 if ((*flags & MPOL_F_STATIC_NODES) && (*flags & MPOL_F_RELATIVE_NODES))
1509 return -EINVAL;
1510 if (*flags & MPOL_F_NUMA_BALANCING) {
1511 if (*mode == MPOL_BIND || *mode == MPOL_PREFERRED_MANY)
1512 *flags |= (MPOL_F_MOF | MPOL_F_MORON);
1513 else
1514 return -EINVAL;
1515 }
1516 return 0;
1517}
1518
1519static long kernel_mbind(unsigned long start, unsigned long len,
1520 unsigned long mode, const unsigned long __user *nmask,
1521 unsigned long maxnode, unsigned int flags)
1522{
1523 unsigned short mode_flags;
1524 nodemask_t nodes;
1525 int lmode = mode;
1526 int err;
1527
1528 start = untagged_addr(start);
1529 err = sanitize_mpol_flags(&lmode, &mode_flags);
1530 if (err)
1531 return err;
1532
1533 err = get_nodes(&nodes, nmask, maxnode);
1534 if (err)
1535 return err;
1536
1537 return do_mbind(start, len, lmode, mode_flags, &nodes, flags);
1538}
1539
1540SYSCALL_DEFINE4(set_mempolicy_home_node, unsigned long, start, unsigned long, len,
1541 unsigned long, home_node, unsigned long, flags)
1542{
1543 struct mm_struct *mm = current->mm;
1544 struct vm_area_struct *vma, *prev;
1545 struct mempolicy *new, *old;
1546 unsigned long end;
1547 int err = -ENOENT;
1548 VMA_ITERATOR(vmi, mm, start);
1549
1550 start = untagged_addr(start);
1551 if (start & ~PAGE_MASK)
1552 return -EINVAL;
1553 /*
1554 * flags is used for future extension if any.
1555 */
1556 if (flags != 0)
1557 return -EINVAL;
1558
1559 /*
1560 * Check home_node is online to avoid accessing uninitialized
1561 * NODE_DATA.
1562 */
1563 if (home_node >= MAX_NUMNODES || !node_online(home_node))
1564 return -EINVAL;
1565
1566 len = PAGE_ALIGN(len);
1567 end = start + len;
1568
1569 if (end < start)
1570 return -EINVAL;
1571 if (end == start)
1572 return 0;
1573 mmap_write_lock(mm);
1574 prev = vma_prev(&vmi);
1575 for_each_vma_range(vmi, vma, end) {
1576 /*
1577 * If any vma in the range got policy other than MPOL_BIND
1578 * or MPOL_PREFERRED_MANY we return error. We don't reset
1579 * the home node for vmas we already updated before.
1580 */
1581 old = vma_policy(vma);
1582 if (!old) {
1583 prev = vma;
1584 continue;
1585 }
1586 if (old->mode != MPOL_BIND && old->mode != MPOL_PREFERRED_MANY) {
1587 err = -EOPNOTSUPP;
1588 break;
1589 }
1590 new = mpol_dup(old);
1591 if (IS_ERR(new)) {
1592 err = PTR_ERR(new);
1593 break;
1594 }
1595
1596 vma_start_write(vma);
1597 new->home_node = home_node;
1598 err = mbind_range(&vmi, vma, &prev, start, end, new);
1599 mpol_put(new);
1600 if (err)
1601 break;
1602 }
1603 mmap_write_unlock(mm);
1604 return err;
1605}
1606
1607SYSCALL_DEFINE6(mbind, unsigned long, start, unsigned long, len,
1608 unsigned long, mode, const unsigned long __user *, nmask,
1609 unsigned long, maxnode, unsigned int, flags)
1610{
1611 return kernel_mbind(start, len, mode, nmask, maxnode, flags);
1612}
1613
1614/* Set the process memory policy */
1615static long kernel_set_mempolicy(int mode, const unsigned long __user *nmask,
1616 unsigned long maxnode)
1617{
1618 unsigned short mode_flags;
1619 nodemask_t nodes;
1620 int lmode = mode;
1621 int err;
1622
1623 err = sanitize_mpol_flags(&lmode, &mode_flags);
1624 if (err)
1625 return err;
1626
1627 err = get_nodes(&nodes, nmask, maxnode);
1628 if (err)
1629 return err;
1630
1631 return do_set_mempolicy(lmode, mode_flags, &nodes);
1632}
1633
1634SYSCALL_DEFINE3(set_mempolicy, int, mode, const unsigned long __user *, nmask,
1635 unsigned long, maxnode)
1636{
1637 return kernel_set_mempolicy(mode, nmask, maxnode);
1638}
1639
1640static int kernel_migrate_pages(pid_t pid, unsigned long maxnode,
1641 const unsigned long __user *old_nodes,
1642 const unsigned long __user *new_nodes)
1643{
1644 struct mm_struct *mm = NULL;
1645 struct task_struct *task;
1646 nodemask_t task_nodes;
1647 int err;
1648 nodemask_t *old;
1649 nodemask_t *new;
1650 NODEMASK_SCRATCH(scratch);
1651
1652 if (!scratch)
1653 return -ENOMEM;
1654
1655 old = &scratch->mask1;
1656 new = &scratch->mask2;
1657
1658 err = get_nodes(old, old_nodes, maxnode);
1659 if (err)
1660 goto out;
1661
1662 err = get_nodes(new, new_nodes, maxnode);
1663 if (err)
1664 goto out;
1665
1666 /* Find the mm_struct */
1667 rcu_read_lock();
1668 task = pid ? find_task_by_vpid(pid) : current;
1669 if (!task) {
1670 rcu_read_unlock();
1671 err = -ESRCH;
1672 goto out;
1673 }
1674 get_task_struct(task);
1675
1676 err = -EINVAL;
1677
1678 /*
1679 * Check if this process has the right to modify the specified process.
1680 * Use the regular "ptrace_may_access()" checks.
1681 */
1682 if (!ptrace_may_access(task, PTRACE_MODE_READ_REALCREDS)) {
1683 rcu_read_unlock();
1684 err = -EPERM;
1685 goto out_put;
1686 }
1687 rcu_read_unlock();
1688
1689 task_nodes = cpuset_mems_allowed(task);
1690 /* Is the user allowed to access the target nodes? */
1691 if (!nodes_subset(*new, task_nodes) && !capable(CAP_SYS_NICE)) {
1692 err = -EPERM;
1693 goto out_put;
1694 }
1695
1696 task_nodes = cpuset_mems_allowed(current);
1697 nodes_and(*new, *new, task_nodes);
1698 if (nodes_empty(*new))
1699 goto out_put;
1700
1701 err = security_task_movememory(task);
1702 if (err)
1703 goto out_put;
1704
1705 mm = get_task_mm(task);
1706 put_task_struct(task);
1707
1708 if (!mm) {
1709 err = -EINVAL;
1710 goto out;
1711 }
1712
1713 err = do_migrate_pages(mm, old, new,
1714 capable(CAP_SYS_NICE) ? MPOL_MF_MOVE_ALL : MPOL_MF_MOVE);
1715
1716 mmput(mm);
1717out:
1718 NODEMASK_SCRATCH_FREE(scratch);
1719
1720 return err;
1721
1722out_put:
1723 put_task_struct(task);
1724 goto out;
1725}
1726
1727SYSCALL_DEFINE4(migrate_pages, pid_t, pid, unsigned long, maxnode,
1728 const unsigned long __user *, old_nodes,
1729 const unsigned long __user *, new_nodes)
1730{
1731 return kernel_migrate_pages(pid, maxnode, old_nodes, new_nodes);
1732}
1733
1734/* Retrieve NUMA policy */
1735static int kernel_get_mempolicy(int __user *policy,
1736 unsigned long __user *nmask,
1737 unsigned long maxnode,
1738 unsigned long addr,
1739 unsigned long flags)
1740{
1741 int err;
1742 int pval;
1743 nodemask_t nodes;
1744
1745 if (nmask != NULL && maxnode < nr_node_ids)
1746 return -EINVAL;
1747
1748 addr = untagged_addr(addr);
1749
1750 err = do_get_mempolicy(&pval, &nodes, addr, flags);
1751
1752 if (err)
1753 return err;
1754
1755 if (policy && put_user(pval, policy))
1756 return -EFAULT;
1757
1758 if (nmask)
1759 err = copy_nodes_to_user(nmask, maxnode, &nodes);
1760
1761 return err;
1762}
1763
1764SYSCALL_DEFINE5(get_mempolicy, int __user *, policy,
1765 unsigned long __user *, nmask, unsigned long, maxnode,
1766 unsigned long, addr, unsigned long, flags)
1767{
1768 return kernel_get_mempolicy(policy, nmask, maxnode, addr, flags);
1769}
1770
1771bool vma_migratable(struct vm_area_struct *vma)
1772{
1773 if (vma->vm_flags & (VM_IO | VM_PFNMAP))
1774 return false;
1775
1776 /*
1777 * DAX device mappings require predictable access latency, so avoid
1778 * incurring periodic faults.
1779 */
1780 if (vma_is_dax(vma))
1781 return false;
1782
1783 if (is_vm_hugetlb_page(vma) &&
1784 !hugepage_migration_supported(hstate_vma(vma)))
1785 return false;
1786
1787 /*
1788 * Migration allocates pages in the highest zone. If we cannot
1789 * do so then migration (at least from node to node) is not
1790 * possible.
1791 */
1792 if (vma->vm_file &&
1793 gfp_zone(mapping_gfp_mask(vma->vm_file->f_mapping))
1794 < policy_zone)
1795 return false;
1796 return true;
1797}
1798
1799struct mempolicy *__get_vma_policy(struct vm_area_struct *vma,
1800 unsigned long addr, pgoff_t *ilx)
1801{
1802 *ilx = 0;
1803 return (vma->vm_ops && vma->vm_ops->get_policy) ?
1804 vma->vm_ops->get_policy(vma, addr, ilx) : vma->vm_policy;
1805}
1806
1807/*
1808 * get_vma_policy(@vma, @addr, @order, @ilx)
1809 * @vma: virtual memory area whose policy is sought
1810 * @addr: address in @vma for shared policy lookup
1811 * @order: 0, or appropriate huge_page_order for interleaving
1812 * @ilx: interleave index (output), for use only when MPOL_INTERLEAVE or
1813 * MPOL_WEIGHTED_INTERLEAVE
1814 *
1815 * Returns effective policy for a VMA at specified address.
1816 * Falls back to current->mempolicy or system default policy, as necessary.
1817 * Shared policies [those marked as MPOL_F_SHARED] require an extra reference
1818 * count--added by the get_policy() vm_op, as appropriate--to protect against
1819 * freeing by another task. It is the caller's responsibility to free the
1820 * extra reference for shared policies.
1821 */
1822struct mempolicy *get_vma_policy(struct vm_area_struct *vma,
1823 unsigned long addr, int order, pgoff_t *ilx)
1824{
1825 struct mempolicy *pol;
1826
1827 pol = __get_vma_policy(vma, addr, ilx);
1828 if (!pol)
1829 pol = get_task_policy(current);
1830 if (pol->mode == MPOL_INTERLEAVE ||
1831 pol->mode == MPOL_WEIGHTED_INTERLEAVE) {
1832 *ilx += vma->vm_pgoff >> order;
1833 *ilx += (addr - vma->vm_start) >> (PAGE_SHIFT + order);
1834 }
1835 return pol;
1836}
1837
1838bool vma_policy_mof(struct vm_area_struct *vma)
1839{
1840 struct mempolicy *pol;
1841
1842 if (vma->vm_ops && vma->vm_ops->get_policy) {
1843 bool ret = false;
1844 pgoff_t ilx; /* ignored here */
1845
1846 pol = vma->vm_ops->get_policy(vma, vma->vm_start, &ilx);
1847 if (pol && (pol->flags & MPOL_F_MOF))
1848 ret = true;
1849 mpol_cond_put(pol);
1850
1851 return ret;
1852 }
1853
1854 pol = vma->vm_policy;
1855 if (!pol)
1856 pol = get_task_policy(current);
1857
1858 return pol->flags & MPOL_F_MOF;
1859}
1860
1861bool apply_policy_zone(struct mempolicy *policy, enum zone_type zone)
1862{
1863 enum zone_type dynamic_policy_zone = policy_zone;
1864
1865 BUG_ON(dynamic_policy_zone == ZONE_MOVABLE);
1866
1867 /*
1868 * if policy->nodes has movable memory only,
1869 * we apply policy when gfp_zone(gfp) = ZONE_MOVABLE only.
1870 *
1871 * policy->nodes is intersect with node_states[N_MEMORY].
1872 * so if the following test fails, it implies
1873 * policy->nodes has movable memory only.
1874 */
1875 if (!nodes_intersects(policy->nodes, node_states[N_HIGH_MEMORY]))
1876 dynamic_policy_zone = ZONE_MOVABLE;
1877
1878 return zone >= dynamic_policy_zone;
1879}
1880
1881static unsigned int weighted_interleave_nodes(struct mempolicy *policy)
1882{
1883 unsigned int node;
1884 unsigned int cpuset_mems_cookie;
1885
1886retry:
1887 /* to prevent miscount use tsk->mems_allowed_seq to detect rebind */
1888 cpuset_mems_cookie = read_mems_allowed_begin();
1889 node = current->il_prev;
1890 if (!current->il_weight || !node_isset(node, policy->nodes)) {
1891 node = next_node_in(node, policy->nodes);
1892 if (read_mems_allowed_retry(cpuset_mems_cookie))
1893 goto retry;
1894 if (node == MAX_NUMNODES)
1895 return node;
1896 current->il_prev = node;
1897 current->il_weight = get_il_weight(node);
1898 }
1899 current->il_weight--;
1900 return node;
1901}
1902
1903/* Do dynamic interleaving for a process */
1904static unsigned int interleave_nodes(struct mempolicy *policy)
1905{
1906 unsigned int nid;
1907 unsigned int cpuset_mems_cookie;
1908
1909 /* to prevent miscount, use tsk->mems_allowed_seq to detect rebind */
1910 do {
1911 cpuset_mems_cookie = read_mems_allowed_begin();
1912 nid = next_node_in(current->il_prev, policy->nodes);
1913 } while (read_mems_allowed_retry(cpuset_mems_cookie));
1914
1915 if (nid < MAX_NUMNODES)
1916 current->il_prev = nid;
1917 return nid;
1918}
1919
1920/*
1921 * Depending on the memory policy provide a node from which to allocate the
1922 * next slab entry.
1923 */
1924unsigned int mempolicy_slab_node(void)
1925{
1926 struct mempolicy *policy;
1927 int node = numa_mem_id();
1928
1929 if (!in_task())
1930 return node;
1931
1932 policy = current->mempolicy;
1933 if (!policy)
1934 return node;
1935
1936 switch (policy->mode) {
1937 case MPOL_PREFERRED:
1938 return first_node(policy->nodes);
1939
1940 case MPOL_INTERLEAVE:
1941 return interleave_nodes(policy);
1942
1943 case MPOL_WEIGHTED_INTERLEAVE:
1944 return weighted_interleave_nodes(policy);
1945
1946 case MPOL_BIND:
1947 case MPOL_PREFERRED_MANY:
1948 {
1949 struct zoneref *z;
1950
1951 /*
1952 * Follow bind policy behavior and start allocation at the
1953 * first node.
1954 */
1955 struct zonelist *zonelist;
1956 enum zone_type highest_zoneidx = gfp_zone(GFP_KERNEL);
1957 zonelist = &NODE_DATA(node)->node_zonelists[ZONELIST_FALLBACK];
1958 z = first_zones_zonelist(zonelist, highest_zoneidx,
1959 &policy->nodes);
1960 return zonelist_zone(z) ? zonelist_node_idx(z) : node;
1961 }
1962 case MPOL_LOCAL:
1963 return node;
1964
1965 default:
1966 BUG();
1967 }
1968}
1969
1970static unsigned int read_once_policy_nodemask(struct mempolicy *pol,
1971 nodemask_t *mask)
1972{
1973 /*
1974 * barrier stabilizes the nodemask locally so that it can be iterated
1975 * over safely without concern for changes. Allocators validate node
1976 * selection does not violate mems_allowed, so this is safe.
1977 */
1978 barrier();
1979 memcpy(mask, &pol->nodes, sizeof(nodemask_t));
1980 barrier();
1981 return nodes_weight(*mask);
1982}
1983
1984static unsigned int weighted_interleave_nid(struct mempolicy *pol, pgoff_t ilx)
1985{
1986 nodemask_t nodemask;
1987 unsigned int target, nr_nodes;
1988 u8 *table;
1989 unsigned int weight_total = 0;
1990 u8 weight;
1991 int nid;
1992
1993 nr_nodes = read_once_policy_nodemask(pol, &nodemask);
1994 if (!nr_nodes)
1995 return numa_node_id();
1996
1997 rcu_read_lock();
1998 table = rcu_dereference(iw_table);
1999 /* calculate the total weight */
2000 for_each_node_mask(nid, nodemask) {
2001 /* detect system default usage */
2002 weight = table ? table[nid] : 1;
2003 weight = weight ? weight : 1;
2004 weight_total += weight;
2005 }
2006
2007 /* Calculate the node offset based on totals */
2008 target = ilx % weight_total;
2009 nid = first_node(nodemask);
2010 while (target) {
2011 /* detect system default usage */
2012 weight = table ? table[nid] : 1;
2013 weight = weight ? weight : 1;
2014 if (target < weight)
2015 break;
2016 target -= weight;
2017 nid = next_node_in(nid, nodemask);
2018 }
2019 rcu_read_unlock();
2020 return nid;
2021}
2022
2023/*
2024 * Do static interleaving for interleave index @ilx. Returns the ilx'th
2025 * node in pol->nodes (starting from ilx=0), wrapping around if ilx
2026 * exceeds the number of present nodes.
2027 */
2028static unsigned int interleave_nid(struct mempolicy *pol, pgoff_t ilx)
2029{
2030 nodemask_t nodemask;
2031 unsigned int target, nnodes;
2032 int i;
2033 int nid;
2034
2035 nnodes = read_once_policy_nodemask(pol, &nodemask);
2036 if (!nnodes)
2037 return numa_node_id();
2038 target = ilx % nnodes;
2039 nid = first_node(nodemask);
2040 for (i = 0; i < target; i++)
2041 nid = next_node(nid, nodemask);
2042 return nid;
2043}
2044
2045/*
2046 * Return a nodemask representing a mempolicy for filtering nodes for
2047 * page allocation, together with preferred node id (or the input node id).
2048 */
2049static nodemask_t *policy_nodemask(gfp_t gfp, struct mempolicy *pol,
2050 pgoff_t ilx, int *nid)
2051{
2052 nodemask_t *nodemask = NULL;
2053
2054 switch (pol->mode) {
2055 case MPOL_PREFERRED:
2056 /* Override input node id */
2057 *nid = first_node(pol->nodes);
2058 break;
2059 case MPOL_PREFERRED_MANY:
2060 nodemask = &pol->nodes;
2061 if (pol->home_node != NUMA_NO_NODE)
2062 *nid = pol->home_node;
2063 break;
2064 case MPOL_BIND:
2065 /* Restrict to nodemask (but not on lower zones) */
2066 if (apply_policy_zone(pol, gfp_zone(gfp)) &&
2067 cpuset_nodemask_valid_mems_allowed(&pol->nodes))
2068 nodemask = &pol->nodes;
2069 if (pol->home_node != NUMA_NO_NODE)
2070 *nid = pol->home_node;
2071 /*
2072 * __GFP_THISNODE shouldn't even be used with the bind policy
2073 * because we might easily break the expectation to stay on the
2074 * requested node and not break the policy.
2075 */
2076 WARN_ON_ONCE(gfp & __GFP_THISNODE);
2077 break;
2078 case MPOL_INTERLEAVE:
2079 /* Override input node id */
2080 *nid = (ilx == NO_INTERLEAVE_INDEX) ?
2081 interleave_nodes(pol) : interleave_nid(pol, ilx);
2082 break;
2083 case MPOL_WEIGHTED_INTERLEAVE:
2084 *nid = (ilx == NO_INTERLEAVE_INDEX) ?
2085 weighted_interleave_nodes(pol) :
2086 weighted_interleave_nid(pol, ilx);
2087 break;
2088 }
2089
2090 return nodemask;
2091}
2092
2093#ifdef CONFIG_HUGETLBFS
2094/*
2095 * huge_node(@vma, @addr, @gfp_flags, @mpol)
2096 * @vma: virtual memory area whose policy is sought
2097 * @addr: address in @vma for shared policy lookup and interleave policy
2098 * @gfp_flags: for requested zone
2099 * @mpol: pointer to mempolicy pointer for reference counted mempolicy
2100 * @nodemask: pointer to nodemask pointer for 'bind' and 'prefer-many' policy
2101 *
2102 * Returns a nid suitable for a huge page allocation and a pointer
2103 * to the struct mempolicy for conditional unref after allocation.
2104 * If the effective policy is 'bind' or 'prefer-many', returns a pointer
2105 * to the mempolicy's @nodemask for filtering the zonelist.
2106 */
2107int huge_node(struct vm_area_struct *vma, unsigned long addr, gfp_t gfp_flags,
2108 struct mempolicy **mpol, nodemask_t **nodemask)
2109{
2110 pgoff_t ilx;
2111 int nid;
2112
2113 nid = numa_node_id();
2114 *mpol = get_vma_policy(vma, addr, hstate_vma(vma)->order, &ilx);
2115 *nodemask = policy_nodemask(gfp_flags, *mpol, ilx, &nid);
2116 return nid;
2117}
2118
2119/*
2120 * init_nodemask_of_mempolicy
2121 *
2122 * If the current task's mempolicy is "default" [NULL], return 'false'
2123 * to indicate default policy. Otherwise, extract the policy nodemask
2124 * for 'bind' or 'interleave' policy into the argument nodemask, or
2125 * initialize the argument nodemask to contain the single node for
2126 * 'preferred' or 'local' policy and return 'true' to indicate presence
2127 * of non-default mempolicy.
2128 *
2129 * We don't bother with reference counting the mempolicy [mpol_get/put]
2130 * because the current task is examining it's own mempolicy and a task's
2131 * mempolicy is only ever changed by the task itself.
2132 *
2133 * N.B., it is the caller's responsibility to free a returned nodemask.
2134 */
2135bool init_nodemask_of_mempolicy(nodemask_t *mask)
2136{
2137 struct mempolicy *mempolicy;
2138
2139 if (!(mask && current->mempolicy))
2140 return false;
2141
2142 task_lock(current);
2143 mempolicy = current->mempolicy;
2144 switch (mempolicy->mode) {
2145 case MPOL_PREFERRED:
2146 case MPOL_PREFERRED_MANY:
2147 case MPOL_BIND:
2148 case MPOL_INTERLEAVE:
2149 case MPOL_WEIGHTED_INTERLEAVE:
2150 *mask = mempolicy->nodes;
2151 break;
2152
2153 case MPOL_LOCAL:
2154 init_nodemask_of_node(mask, numa_node_id());
2155 break;
2156
2157 default:
2158 BUG();
2159 }
2160 task_unlock(current);
2161
2162 return true;
2163}
2164#endif
2165
2166/*
2167 * mempolicy_in_oom_domain
2168 *
2169 * If tsk's mempolicy is "bind", check for intersection between mask and
2170 * the policy nodemask. Otherwise, return true for all other policies
2171 * including "interleave", as a tsk with "interleave" policy may have
2172 * memory allocated from all nodes in system.
2173 *
2174 * Takes task_lock(tsk) to prevent freeing of its mempolicy.
2175 */
2176bool mempolicy_in_oom_domain(struct task_struct *tsk,
2177 const nodemask_t *mask)
2178{
2179 struct mempolicy *mempolicy;
2180 bool ret = true;
2181
2182 if (!mask)
2183 return ret;
2184
2185 task_lock(tsk);
2186 mempolicy = tsk->mempolicy;
2187 if (mempolicy && mempolicy->mode == MPOL_BIND)
2188 ret = nodes_intersects(mempolicy->nodes, *mask);
2189 task_unlock(tsk);
2190
2191 return ret;
2192}
2193
2194static struct page *alloc_pages_preferred_many(gfp_t gfp, unsigned int order,
2195 int nid, nodemask_t *nodemask)
2196{
2197 struct page *page;
2198 gfp_t preferred_gfp;
2199
2200 /*
2201 * This is a two pass approach. The first pass will only try the
2202 * preferred nodes but skip the direct reclaim and allow the
2203 * allocation to fail, while the second pass will try all the
2204 * nodes in system.
2205 */
2206 preferred_gfp = gfp | __GFP_NOWARN;
2207 preferred_gfp &= ~(__GFP_DIRECT_RECLAIM | __GFP_NOFAIL);
2208 page = __alloc_pages_noprof(preferred_gfp, order, nid, nodemask);
2209 if (!page)
2210 page = __alloc_pages_noprof(gfp, order, nid, NULL);
2211
2212 return page;
2213}
2214
2215/**
2216 * alloc_pages_mpol - Allocate pages according to NUMA mempolicy.
2217 * @gfp: GFP flags.
2218 * @order: Order of the page allocation.
2219 * @pol: Pointer to the NUMA mempolicy.
2220 * @ilx: Index for interleave mempolicy (also distinguishes alloc_pages()).
2221 * @nid: Preferred node (usually numa_node_id() but @mpol may override it).
2222 *
2223 * Return: The page on success or NULL if allocation fails.
2224 */
2225struct page *alloc_pages_mpol_noprof(gfp_t gfp, unsigned int order,
2226 struct mempolicy *pol, pgoff_t ilx, int nid)
2227{
2228 nodemask_t *nodemask;
2229 struct page *page;
2230
2231 nodemask = policy_nodemask(gfp, pol, ilx, &nid);
2232
2233 if (pol->mode == MPOL_PREFERRED_MANY)
2234 return alloc_pages_preferred_many(gfp, order, nid, nodemask);
2235
2236 if (IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE) &&
2237 /* filter "hugepage" allocation, unless from alloc_pages() */
2238 order == HPAGE_PMD_ORDER && ilx != NO_INTERLEAVE_INDEX) {
2239 /*
2240 * For hugepage allocation and non-interleave policy which
2241 * allows the current node (or other explicitly preferred
2242 * node) we only try to allocate from the current/preferred
2243 * node and don't fall back to other nodes, as the cost of
2244 * remote accesses would likely offset THP benefits.
2245 *
2246 * If the policy is interleave or does not allow the current
2247 * node in its nodemask, we allocate the standard way.
2248 */
2249 if (pol->mode != MPOL_INTERLEAVE &&
2250 pol->mode != MPOL_WEIGHTED_INTERLEAVE &&
2251 (!nodemask || node_isset(nid, *nodemask))) {
2252 /*
2253 * First, try to allocate THP only on local node, but
2254 * don't reclaim unnecessarily, just compact.
2255 */
2256 page = __alloc_pages_node_noprof(nid,
2257 gfp | __GFP_THISNODE | __GFP_NORETRY, order);
2258 if (page || !(gfp & __GFP_DIRECT_RECLAIM))
2259 return page;
2260 /*
2261 * If hugepage allocations are configured to always
2262 * synchronous compact or the vma has been madvised
2263 * to prefer hugepage backing, retry allowing remote
2264 * memory with both reclaim and compact as well.
2265 */
2266 }
2267 }
2268
2269 page = __alloc_pages_noprof(gfp, order, nid, nodemask);
2270
2271 if (unlikely(pol->mode == MPOL_INTERLEAVE ||
2272 pol->mode == MPOL_WEIGHTED_INTERLEAVE) && page) {
2273 /* skip NUMA_INTERLEAVE_HIT update if numa stats is disabled */
2274 if (static_branch_likely(&vm_numa_stat_key) &&
2275 page_to_nid(page) == nid) {
2276 preempt_disable();
2277 __count_numa_event(page_zone(page), NUMA_INTERLEAVE_HIT);
2278 preempt_enable();
2279 }
2280 }
2281
2282 return page;
2283}
2284
2285struct folio *folio_alloc_mpol_noprof(gfp_t gfp, unsigned int order,
2286 struct mempolicy *pol, pgoff_t ilx, int nid)
2287{
2288 return page_rmappable_folio(alloc_pages_mpol_noprof(gfp | __GFP_COMP,
2289 order, pol, ilx, nid));
2290}
2291
2292/**
2293 * vma_alloc_folio - Allocate a folio for a VMA.
2294 * @gfp: GFP flags.
2295 * @order: Order of the folio.
2296 * @vma: Pointer to VMA.
2297 * @addr: Virtual address of the allocation. Must be inside @vma.
2298 *
2299 * Allocate a folio for a specific address in @vma, using the appropriate
2300 * NUMA policy. The caller must hold the mmap_lock of the mm_struct of the
2301 * VMA to prevent it from going away. Should be used for all allocations
2302 * for folios that will be mapped into user space, excepting hugetlbfs, and
2303 * excepting where direct use of alloc_pages_mpol() is more appropriate.
2304 *
2305 * Return: The folio on success or NULL if allocation fails.
2306 */
2307struct folio *vma_alloc_folio_noprof(gfp_t gfp, int order, struct vm_area_struct *vma,
2308 unsigned long addr)
2309{
2310 struct mempolicy *pol;
2311 pgoff_t ilx;
2312 struct folio *folio;
2313
2314 if (vma->vm_flags & VM_DROPPABLE)
2315 gfp |= __GFP_NOWARN;
2316
2317 pol = get_vma_policy(vma, addr, order, &ilx);
2318 folio = folio_alloc_mpol_noprof(gfp, order, pol, ilx, numa_node_id());
2319 mpol_cond_put(pol);
2320 return folio;
2321}
2322EXPORT_SYMBOL(vma_alloc_folio_noprof);
2323
2324/**
2325 * alloc_pages - Allocate pages.
2326 * @gfp: GFP flags.
2327 * @order: Power of two of number of pages to allocate.
2328 *
2329 * Allocate 1 << @order contiguous pages. The physical address of the
2330 * first page is naturally aligned (eg an order-3 allocation will be aligned
2331 * to a multiple of 8 * PAGE_SIZE bytes). The NUMA policy of the current
2332 * process is honoured when in process context.
2333 *
2334 * Context: Can be called from any context, providing the appropriate GFP
2335 * flags are used.
2336 * Return: The page on success or NULL if allocation fails.
2337 */
2338struct page *alloc_pages_noprof(gfp_t gfp, unsigned int order)
2339{
2340 struct mempolicy *pol = &default_policy;
2341
2342 /*
2343 * No reference counting needed for current->mempolicy
2344 * nor system default_policy
2345 */
2346 if (!in_interrupt() && !(gfp & __GFP_THISNODE))
2347 pol = get_task_policy(current);
2348
2349 return alloc_pages_mpol_noprof(gfp, order, pol, NO_INTERLEAVE_INDEX,
2350 numa_node_id());
2351}
2352EXPORT_SYMBOL(alloc_pages_noprof);
2353
2354struct folio *folio_alloc_noprof(gfp_t gfp, unsigned int order)
2355{
2356 return page_rmappable_folio(alloc_pages_noprof(gfp | __GFP_COMP, order));
2357}
2358EXPORT_SYMBOL(folio_alloc_noprof);
2359
2360static unsigned long alloc_pages_bulk_array_interleave(gfp_t gfp,
2361 struct mempolicy *pol, unsigned long nr_pages,
2362 struct page **page_array)
2363{
2364 int nodes;
2365 unsigned long nr_pages_per_node;
2366 int delta;
2367 int i;
2368 unsigned long nr_allocated;
2369 unsigned long total_allocated = 0;
2370
2371 nodes = nodes_weight(pol->nodes);
2372 nr_pages_per_node = nr_pages / nodes;
2373 delta = nr_pages - nodes * nr_pages_per_node;
2374
2375 for (i = 0; i < nodes; i++) {
2376 if (delta) {
2377 nr_allocated = alloc_pages_bulk_noprof(gfp,
2378 interleave_nodes(pol), NULL,
2379 nr_pages_per_node + 1, NULL,
2380 page_array);
2381 delta--;
2382 } else {
2383 nr_allocated = alloc_pages_bulk_noprof(gfp,
2384 interleave_nodes(pol), NULL,
2385 nr_pages_per_node, NULL, page_array);
2386 }
2387
2388 page_array += nr_allocated;
2389 total_allocated += nr_allocated;
2390 }
2391
2392 return total_allocated;
2393}
2394
2395static unsigned long alloc_pages_bulk_array_weighted_interleave(gfp_t gfp,
2396 struct mempolicy *pol, unsigned long nr_pages,
2397 struct page **page_array)
2398{
2399 struct task_struct *me = current;
2400 unsigned int cpuset_mems_cookie;
2401 unsigned long total_allocated = 0;
2402 unsigned long nr_allocated = 0;
2403 unsigned long rounds;
2404 unsigned long node_pages, delta;
2405 u8 *table, *weights, weight;
2406 unsigned int weight_total = 0;
2407 unsigned long rem_pages = nr_pages;
2408 nodemask_t nodes;
2409 int nnodes, node;
2410 int resume_node = MAX_NUMNODES - 1;
2411 u8 resume_weight = 0;
2412 int prev_node;
2413 int i;
2414
2415 if (!nr_pages)
2416 return 0;
2417
2418 /* read the nodes onto the stack, retry if done during rebind */
2419 do {
2420 cpuset_mems_cookie = read_mems_allowed_begin();
2421 nnodes = read_once_policy_nodemask(pol, &nodes);
2422 } while (read_mems_allowed_retry(cpuset_mems_cookie));
2423
2424 /* if the nodemask has become invalid, we cannot do anything */
2425 if (!nnodes)
2426 return 0;
2427
2428 /* Continue allocating from most recent node and adjust the nr_pages */
2429 node = me->il_prev;
2430 weight = me->il_weight;
2431 if (weight && node_isset(node, nodes)) {
2432 node_pages = min(rem_pages, weight);
2433 nr_allocated = __alloc_pages_bulk(gfp, node, NULL, node_pages,
2434 NULL, page_array);
2435 page_array += nr_allocated;
2436 total_allocated += nr_allocated;
2437 /* if that's all the pages, no need to interleave */
2438 if (rem_pages <= weight) {
2439 me->il_weight -= rem_pages;
2440 return total_allocated;
2441 }
2442 /* Otherwise we adjust remaining pages, continue from there */
2443 rem_pages -= weight;
2444 }
2445 /* clear active weight in case of an allocation failure */
2446 me->il_weight = 0;
2447 prev_node = node;
2448
2449 /* create a local copy of node weights to operate on outside rcu */
2450 weights = kzalloc(nr_node_ids, GFP_KERNEL);
2451 if (!weights)
2452 return total_allocated;
2453
2454 rcu_read_lock();
2455 table = rcu_dereference(iw_table);
2456 if (table)
2457 memcpy(weights, table, nr_node_ids);
2458 rcu_read_unlock();
2459
2460 /* calculate total, detect system default usage */
2461 for_each_node_mask(node, nodes) {
2462 if (!weights[node])
2463 weights[node] = 1;
2464 weight_total += weights[node];
2465 }
2466
2467 /*
2468 * Calculate rounds/partial rounds to minimize __alloc_pages_bulk calls.
2469 * Track which node weighted interleave should resume from.
2470 *
2471 * if (rounds > 0) and (delta == 0), resume_node will always be
2472 * the node following prev_node and its weight.
2473 */
2474 rounds = rem_pages / weight_total;
2475 delta = rem_pages % weight_total;
2476 resume_node = next_node_in(prev_node, nodes);
2477 resume_weight = weights[resume_node];
2478 for (i = 0; i < nnodes; i++) {
2479 node = next_node_in(prev_node, nodes);
2480 weight = weights[node];
2481 node_pages = weight * rounds;
2482 /* If a delta exists, add this node's portion of the delta */
2483 if (delta > weight) {
2484 node_pages += weight;
2485 delta -= weight;
2486 } else if (delta) {
2487 /* when delta is depleted, resume from that node */
2488 node_pages += delta;
2489 resume_node = node;
2490 resume_weight = weight - delta;
2491 delta = 0;
2492 }
2493 /* node_pages can be 0 if an allocation fails and rounds == 0 */
2494 if (!node_pages)
2495 break;
2496 nr_allocated = __alloc_pages_bulk(gfp, node, NULL, node_pages,
2497 NULL, page_array);
2498 page_array += nr_allocated;
2499 total_allocated += nr_allocated;
2500 if (total_allocated == nr_pages)
2501 break;
2502 prev_node = node;
2503 }
2504 me->il_prev = resume_node;
2505 me->il_weight = resume_weight;
2506 kfree(weights);
2507 return total_allocated;
2508}
2509
2510static unsigned long alloc_pages_bulk_array_preferred_many(gfp_t gfp, int nid,
2511 struct mempolicy *pol, unsigned long nr_pages,
2512 struct page **page_array)
2513{
2514 gfp_t preferred_gfp;
2515 unsigned long nr_allocated = 0;
2516
2517 preferred_gfp = gfp | __GFP_NOWARN;
2518 preferred_gfp &= ~(__GFP_DIRECT_RECLAIM | __GFP_NOFAIL);
2519
2520 nr_allocated = alloc_pages_bulk_noprof(preferred_gfp, nid, &pol->nodes,
2521 nr_pages, NULL, page_array);
2522
2523 if (nr_allocated < nr_pages)
2524 nr_allocated += alloc_pages_bulk_noprof(gfp, numa_node_id(), NULL,
2525 nr_pages - nr_allocated, NULL,
2526 page_array + nr_allocated);
2527 return nr_allocated;
2528}
2529
2530/* alloc pages bulk and mempolicy should be considered at the
2531 * same time in some situation such as vmalloc.
2532 *
2533 * It can accelerate memory allocation especially interleaving
2534 * allocate memory.
2535 */
2536unsigned long alloc_pages_bulk_array_mempolicy_noprof(gfp_t gfp,
2537 unsigned long nr_pages, struct page **page_array)
2538{
2539 struct mempolicy *pol = &default_policy;
2540 nodemask_t *nodemask;
2541 int nid;
2542
2543 if (!in_interrupt() && !(gfp & __GFP_THISNODE))
2544 pol = get_task_policy(current);
2545
2546 if (pol->mode == MPOL_INTERLEAVE)
2547 return alloc_pages_bulk_array_interleave(gfp, pol,
2548 nr_pages, page_array);
2549
2550 if (pol->mode == MPOL_WEIGHTED_INTERLEAVE)
2551 return alloc_pages_bulk_array_weighted_interleave(
2552 gfp, pol, nr_pages, page_array);
2553
2554 if (pol->mode == MPOL_PREFERRED_MANY)
2555 return alloc_pages_bulk_array_preferred_many(gfp,
2556 numa_node_id(), pol, nr_pages, page_array);
2557
2558 nid = numa_node_id();
2559 nodemask = policy_nodemask(gfp, pol, NO_INTERLEAVE_INDEX, &nid);
2560 return alloc_pages_bulk_noprof(gfp, nid, nodemask,
2561 nr_pages, NULL, page_array);
2562}
2563
2564int vma_dup_policy(struct vm_area_struct *src, struct vm_area_struct *dst)
2565{
2566 struct mempolicy *pol = mpol_dup(src->vm_policy);
2567
2568 if (IS_ERR(pol))
2569 return PTR_ERR(pol);
2570 dst->vm_policy = pol;
2571 return 0;
2572}
2573
2574/*
2575 * If mpol_dup() sees current->cpuset == cpuset_being_rebound, then it
2576 * rebinds the mempolicy its copying by calling mpol_rebind_policy()
2577 * with the mems_allowed returned by cpuset_mems_allowed(). This
2578 * keeps mempolicies cpuset relative after its cpuset moves. See
2579 * further kernel/cpuset.c update_nodemask().
2580 *
2581 * current's mempolicy may be rebinded by the other task(the task that changes
2582 * cpuset's mems), so we needn't do rebind work for current task.
2583 */
2584
2585/* Slow path of a mempolicy duplicate */
2586struct mempolicy *__mpol_dup(struct mempolicy *old)
2587{
2588 struct mempolicy *new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
2589
2590 if (!new)
2591 return ERR_PTR(-ENOMEM);
2592
2593 /* task's mempolicy is protected by alloc_lock */
2594 if (old == current->mempolicy) {
2595 task_lock(current);
2596 *new = *old;
2597 task_unlock(current);
2598 } else
2599 *new = *old;
2600
2601 if (current_cpuset_is_being_rebound()) {
2602 nodemask_t mems = cpuset_mems_allowed(current);
2603 mpol_rebind_policy(new, &mems);
2604 }
2605 atomic_set(&new->refcnt, 1);
2606 return new;
2607}
2608
2609/* Slow path of a mempolicy comparison */
2610bool __mpol_equal(struct mempolicy *a, struct mempolicy *b)
2611{
2612 if (!a || !b)
2613 return false;
2614 if (a->mode != b->mode)
2615 return false;
2616 if (a->flags != b->flags)
2617 return false;
2618 if (a->home_node != b->home_node)
2619 return false;
2620 if (mpol_store_user_nodemask(a))
2621 if (!nodes_equal(a->w.user_nodemask, b->w.user_nodemask))
2622 return false;
2623
2624 switch (a->mode) {
2625 case MPOL_BIND:
2626 case MPOL_INTERLEAVE:
2627 case MPOL_PREFERRED:
2628 case MPOL_PREFERRED_MANY:
2629 case MPOL_WEIGHTED_INTERLEAVE:
2630 return !!nodes_equal(a->nodes, b->nodes);
2631 case MPOL_LOCAL:
2632 return true;
2633 default:
2634 BUG();
2635 return false;
2636 }
2637}
2638
2639/*
2640 * Shared memory backing store policy support.
2641 *
2642 * Remember policies even when nobody has shared memory mapped.
2643 * The policies are kept in Red-Black tree linked from the inode.
2644 * They are protected by the sp->lock rwlock, which should be held
2645 * for any accesses to the tree.
2646 */
2647
2648/*
2649 * lookup first element intersecting start-end. Caller holds sp->lock for
2650 * reading or for writing
2651 */
2652static struct sp_node *sp_lookup(struct shared_policy *sp,
2653 pgoff_t start, pgoff_t end)
2654{
2655 struct rb_node *n = sp->root.rb_node;
2656
2657 while (n) {
2658 struct sp_node *p = rb_entry(n, struct sp_node, nd);
2659
2660 if (start >= p->end)
2661 n = n->rb_right;
2662 else if (end <= p->start)
2663 n = n->rb_left;
2664 else
2665 break;
2666 }
2667 if (!n)
2668 return NULL;
2669 for (;;) {
2670 struct sp_node *w = NULL;
2671 struct rb_node *prev = rb_prev(n);
2672 if (!prev)
2673 break;
2674 w = rb_entry(prev, struct sp_node, nd);
2675 if (w->end <= start)
2676 break;
2677 n = prev;
2678 }
2679 return rb_entry(n, struct sp_node, nd);
2680}
2681
2682/*
2683 * Insert a new shared policy into the list. Caller holds sp->lock for
2684 * writing.
2685 */
2686static void sp_insert(struct shared_policy *sp, struct sp_node *new)
2687{
2688 struct rb_node **p = &sp->root.rb_node;
2689 struct rb_node *parent = NULL;
2690 struct sp_node *nd;
2691
2692 while (*p) {
2693 parent = *p;
2694 nd = rb_entry(parent, struct sp_node, nd);
2695 if (new->start < nd->start)
2696 p = &(*p)->rb_left;
2697 else if (new->end > nd->end)
2698 p = &(*p)->rb_right;
2699 else
2700 BUG();
2701 }
2702 rb_link_node(&new->nd, parent, p);
2703 rb_insert_color(&new->nd, &sp->root);
2704}
2705
2706/* Find shared policy intersecting idx */
2707struct mempolicy *mpol_shared_policy_lookup(struct shared_policy *sp,
2708 pgoff_t idx)
2709{
2710 struct mempolicy *pol = NULL;
2711 struct sp_node *sn;
2712
2713 if (!sp->root.rb_node)
2714 return NULL;
2715 read_lock(&sp->lock);
2716 sn = sp_lookup(sp, idx, idx+1);
2717 if (sn) {
2718 mpol_get(sn->policy);
2719 pol = sn->policy;
2720 }
2721 read_unlock(&sp->lock);
2722 return pol;
2723}
2724
2725static void sp_free(struct sp_node *n)
2726{
2727 mpol_put(n->policy);
2728 kmem_cache_free(sn_cache, n);
2729}
2730
2731/**
2732 * mpol_misplaced - check whether current folio node is valid in policy
2733 *
2734 * @folio: folio to be checked
2735 * @vmf: structure describing the fault
2736 * @addr: virtual address in @vma for shared policy lookup and interleave policy
2737 *
2738 * Lookup current policy node id for vma,addr and "compare to" folio's
2739 * node id. Policy determination "mimics" alloc_page_vma().
2740 * Called from fault path where we know the vma and faulting address.
2741 *
2742 * Return: NUMA_NO_NODE if the page is in a node that is valid for this
2743 * policy, or a suitable node ID to allocate a replacement folio from.
2744 */
2745int mpol_misplaced(struct folio *folio, struct vm_fault *vmf,
2746 unsigned long addr)
2747{
2748 struct mempolicy *pol;
2749 pgoff_t ilx;
2750 struct zoneref *z;
2751 int curnid = folio_nid(folio);
2752 struct vm_area_struct *vma = vmf->vma;
2753 int thiscpu = raw_smp_processor_id();
2754 int thisnid = numa_node_id();
2755 int polnid = NUMA_NO_NODE;
2756 int ret = NUMA_NO_NODE;
2757
2758 /*
2759 * Make sure ptl is held so that we don't preempt and we
2760 * have a stable smp processor id
2761 */
2762 lockdep_assert_held(vmf->ptl);
2763 pol = get_vma_policy(vma, addr, folio_order(folio), &ilx);
2764 if (!(pol->flags & MPOL_F_MOF))
2765 goto out;
2766
2767 switch (pol->mode) {
2768 case MPOL_INTERLEAVE:
2769 polnid = interleave_nid(pol, ilx);
2770 break;
2771
2772 case MPOL_WEIGHTED_INTERLEAVE:
2773 polnid = weighted_interleave_nid(pol, ilx);
2774 break;
2775
2776 case MPOL_PREFERRED:
2777 if (node_isset(curnid, pol->nodes))
2778 goto out;
2779 polnid = first_node(pol->nodes);
2780 break;
2781
2782 case MPOL_LOCAL:
2783 polnid = numa_node_id();
2784 break;
2785
2786 case MPOL_BIND:
2787 case MPOL_PREFERRED_MANY:
2788 /*
2789 * Even though MPOL_PREFERRED_MANY can allocate pages outside
2790 * policy nodemask we don't allow numa migration to nodes
2791 * outside policy nodemask for now. This is done so that if we
2792 * want demotion to slow memory to happen, before allocating
2793 * from some DRAM node say 'x', we will end up using a
2794 * MPOL_PREFERRED_MANY mask excluding node 'x'. In such scenario
2795 * we should not promote to node 'x' from slow memory node.
2796 */
2797 if (pol->flags & MPOL_F_MORON) {
2798 /*
2799 * Optimize placement among multiple nodes
2800 * via NUMA balancing
2801 */
2802 if (node_isset(thisnid, pol->nodes))
2803 break;
2804 goto out;
2805 }
2806
2807 /*
2808 * use current page if in policy nodemask,
2809 * else select nearest allowed node, if any.
2810 * If no allowed nodes, use current [!misplaced].
2811 */
2812 if (node_isset(curnid, pol->nodes))
2813 goto out;
2814 z = first_zones_zonelist(
2815 node_zonelist(thisnid, GFP_HIGHUSER),
2816 gfp_zone(GFP_HIGHUSER),
2817 &pol->nodes);
2818 polnid = zonelist_node_idx(z);
2819 break;
2820
2821 default:
2822 BUG();
2823 }
2824
2825 /* Migrate the folio towards the node whose CPU is referencing it */
2826 if (pol->flags & MPOL_F_MORON) {
2827 polnid = thisnid;
2828
2829 if (!should_numa_migrate_memory(current, folio, curnid,
2830 thiscpu))
2831 goto out;
2832 }
2833
2834 if (curnid != polnid)
2835 ret = polnid;
2836out:
2837 mpol_cond_put(pol);
2838
2839 return ret;
2840}
2841
2842/*
2843 * Drop the (possibly final) reference to task->mempolicy. It needs to be
2844 * dropped after task->mempolicy is set to NULL so that any allocation done as
2845 * part of its kmem_cache_free(), such as by KASAN, doesn't reference a freed
2846 * policy.
2847 */
2848void mpol_put_task_policy(struct task_struct *task)
2849{
2850 struct mempolicy *pol;
2851
2852 task_lock(task);
2853 pol = task->mempolicy;
2854 task->mempolicy = NULL;
2855 task_unlock(task);
2856 mpol_put(pol);
2857}
2858
2859static void sp_delete(struct shared_policy *sp, struct sp_node *n)
2860{
2861 rb_erase(&n->nd, &sp->root);
2862 sp_free(n);
2863}
2864
2865static void sp_node_init(struct sp_node *node, unsigned long start,
2866 unsigned long end, struct mempolicy *pol)
2867{
2868 node->start = start;
2869 node->end = end;
2870 node->policy = pol;
2871}
2872
2873static struct sp_node *sp_alloc(unsigned long start, unsigned long end,
2874 struct mempolicy *pol)
2875{
2876 struct sp_node *n;
2877 struct mempolicy *newpol;
2878
2879 n = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2880 if (!n)
2881 return NULL;
2882
2883 newpol = mpol_dup(pol);
2884 if (IS_ERR(newpol)) {
2885 kmem_cache_free(sn_cache, n);
2886 return NULL;
2887 }
2888 newpol->flags |= MPOL_F_SHARED;
2889 sp_node_init(n, start, end, newpol);
2890
2891 return n;
2892}
2893
2894/* Replace a policy range. */
2895static int shared_policy_replace(struct shared_policy *sp, pgoff_t start,
2896 pgoff_t end, struct sp_node *new)
2897{
2898 struct sp_node *n;
2899 struct sp_node *n_new = NULL;
2900 struct mempolicy *mpol_new = NULL;
2901 int ret = 0;
2902
2903restart:
2904 write_lock(&sp->lock);
2905 n = sp_lookup(sp, start, end);
2906 /* Take care of old policies in the same range. */
2907 while (n && n->start < end) {
2908 struct rb_node *next = rb_next(&n->nd);
2909 if (n->start >= start) {
2910 if (n->end <= end)
2911 sp_delete(sp, n);
2912 else
2913 n->start = end;
2914 } else {
2915 /* Old policy spanning whole new range. */
2916 if (n->end > end) {
2917 if (!n_new)
2918 goto alloc_new;
2919
2920 *mpol_new = *n->policy;
2921 atomic_set(&mpol_new->refcnt, 1);
2922 sp_node_init(n_new, end, n->end, mpol_new);
2923 n->end = start;
2924 sp_insert(sp, n_new);
2925 n_new = NULL;
2926 mpol_new = NULL;
2927 break;
2928 } else
2929 n->end = start;
2930 }
2931 if (!next)
2932 break;
2933 n = rb_entry(next, struct sp_node, nd);
2934 }
2935 if (new)
2936 sp_insert(sp, new);
2937 write_unlock(&sp->lock);
2938 ret = 0;
2939
2940err_out:
2941 if (mpol_new)
2942 mpol_put(mpol_new);
2943 if (n_new)
2944 kmem_cache_free(sn_cache, n_new);
2945
2946 return ret;
2947
2948alloc_new:
2949 write_unlock(&sp->lock);
2950 ret = -ENOMEM;
2951 n_new = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2952 if (!n_new)
2953 goto err_out;
2954 mpol_new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
2955 if (!mpol_new)
2956 goto err_out;
2957 atomic_set(&mpol_new->refcnt, 1);
2958 goto restart;
2959}
2960
2961/**
2962 * mpol_shared_policy_init - initialize shared policy for inode
2963 * @sp: pointer to inode shared policy
2964 * @mpol: struct mempolicy to install
2965 *
2966 * Install non-NULL @mpol in inode's shared policy rb-tree.
2967 * On entry, the current task has a reference on a non-NULL @mpol.
2968 * This must be released on exit.
2969 * This is called at get_inode() calls and we can use GFP_KERNEL.
2970 */
2971void mpol_shared_policy_init(struct shared_policy *sp, struct mempolicy *mpol)
2972{
2973 int ret;
2974
2975 sp->root = RB_ROOT; /* empty tree == default mempolicy */
2976 rwlock_init(&sp->lock);
2977
2978 if (mpol) {
2979 struct sp_node *sn;
2980 struct mempolicy *npol;
2981 NODEMASK_SCRATCH(scratch);
2982
2983 if (!scratch)
2984 goto put_mpol;
2985
2986 /* contextualize the tmpfs mount point mempolicy to this file */
2987 npol = mpol_new(mpol->mode, mpol->flags, &mpol->w.user_nodemask);
2988 if (IS_ERR(npol))
2989 goto free_scratch; /* no valid nodemask intersection */
2990
2991 task_lock(current);
2992 ret = mpol_set_nodemask(npol, &mpol->w.user_nodemask, scratch);
2993 task_unlock(current);
2994 if (ret)
2995 goto put_npol;
2996
2997 /* alloc node covering entire file; adds ref to file's npol */
2998 sn = sp_alloc(0, MAX_LFS_FILESIZE >> PAGE_SHIFT, npol);
2999 if (sn)
3000 sp_insert(sp, sn);
3001put_npol:
3002 mpol_put(npol); /* drop initial ref on file's npol */
3003free_scratch:
3004 NODEMASK_SCRATCH_FREE(scratch);
3005put_mpol:
3006 mpol_put(mpol); /* drop our incoming ref on sb mpol */
3007 }
3008}
3009
3010int mpol_set_shared_policy(struct shared_policy *sp,
3011 struct vm_area_struct *vma, struct mempolicy *pol)
3012{
3013 int err;
3014 struct sp_node *new = NULL;
3015 unsigned long sz = vma_pages(vma);
3016
3017 if (pol) {
3018 new = sp_alloc(vma->vm_pgoff, vma->vm_pgoff + sz, pol);
3019 if (!new)
3020 return -ENOMEM;
3021 }
3022 err = shared_policy_replace(sp, vma->vm_pgoff, vma->vm_pgoff + sz, new);
3023 if (err && new)
3024 sp_free(new);
3025 return err;
3026}
3027
3028/* Free a backing policy store on inode delete. */
3029void mpol_free_shared_policy(struct shared_policy *sp)
3030{
3031 struct sp_node *n;
3032 struct rb_node *next;
3033
3034 if (!sp->root.rb_node)
3035 return;
3036 write_lock(&sp->lock);
3037 next = rb_first(&sp->root);
3038 while (next) {
3039 n = rb_entry(next, struct sp_node, nd);
3040 next = rb_next(&n->nd);
3041 sp_delete(sp, n);
3042 }
3043 write_unlock(&sp->lock);
3044}
3045
3046#ifdef CONFIG_NUMA_BALANCING
3047static int __initdata numabalancing_override;
3048
3049static void __init check_numabalancing_enable(void)
3050{
3051 bool numabalancing_default = false;
3052
3053 if (IS_ENABLED(CONFIG_NUMA_BALANCING_DEFAULT_ENABLED))
3054 numabalancing_default = true;
3055
3056 /* Parsed by setup_numabalancing. override == 1 enables, -1 disables */
3057 if (numabalancing_override)
3058 set_numabalancing_state(numabalancing_override == 1);
3059
3060 if (num_online_nodes() > 1 && !numabalancing_override) {
3061 pr_info("%s automatic NUMA balancing. Configure with numa_balancing= or the kernel.numa_balancing sysctl\n",
3062 numabalancing_default ? "Enabling" : "Disabling");
3063 set_numabalancing_state(numabalancing_default);
3064 }
3065}
3066
3067static int __init setup_numabalancing(char *str)
3068{
3069 int ret = 0;
3070 if (!str)
3071 goto out;
3072
3073 if (!strcmp(str, "enable")) {
3074 numabalancing_override = 1;
3075 ret = 1;
3076 } else if (!strcmp(str, "disable")) {
3077 numabalancing_override = -1;
3078 ret = 1;
3079 }
3080out:
3081 if (!ret)
3082 pr_warn("Unable to parse numa_balancing=\n");
3083
3084 return ret;
3085}
3086__setup("numa_balancing=", setup_numabalancing);
3087#else
3088static inline void __init check_numabalancing_enable(void)
3089{
3090}
3091#endif /* CONFIG_NUMA_BALANCING */
3092
3093void __init numa_policy_init(void)
3094{
3095 nodemask_t interleave_nodes;
3096 unsigned long largest = 0;
3097 int nid, prefer = 0;
3098
3099 policy_cache = kmem_cache_create("numa_policy",
3100 sizeof(struct mempolicy),
3101 0, SLAB_PANIC, NULL);
3102
3103 sn_cache = kmem_cache_create("shared_policy_node",
3104 sizeof(struct sp_node),
3105 0, SLAB_PANIC, NULL);
3106
3107 for_each_node(nid) {
3108 preferred_node_policy[nid] = (struct mempolicy) {
3109 .refcnt = ATOMIC_INIT(1),
3110 .mode = MPOL_PREFERRED,
3111 .flags = MPOL_F_MOF | MPOL_F_MORON,
3112 .nodes = nodemask_of_node(nid),
3113 };
3114 }
3115
3116 /*
3117 * Set interleaving policy for system init. Interleaving is only
3118 * enabled across suitably sized nodes (default is >= 16MB), or
3119 * fall back to the largest node if they're all smaller.
3120 */
3121 nodes_clear(interleave_nodes);
3122 for_each_node_state(nid, N_MEMORY) {
3123 unsigned long total_pages = node_present_pages(nid);
3124
3125 /* Preserve the largest node */
3126 if (largest < total_pages) {
3127 largest = total_pages;
3128 prefer = nid;
3129 }
3130
3131 /* Interleave this node? */
3132 if ((total_pages << PAGE_SHIFT) >= (16 << 20))
3133 node_set(nid, interleave_nodes);
3134 }
3135
3136 /* All too small, use the largest */
3137 if (unlikely(nodes_empty(interleave_nodes)))
3138 node_set(prefer, interleave_nodes);
3139
3140 if (do_set_mempolicy(MPOL_INTERLEAVE, 0, &interleave_nodes))
3141 pr_err("%s: interleaving failed\n", __func__);
3142
3143 check_numabalancing_enable();
3144}
3145
3146/* Reset policy of current process to default */
3147void numa_default_policy(void)
3148{
3149 do_set_mempolicy(MPOL_DEFAULT, 0, NULL);
3150}
3151
3152/*
3153 * Parse and format mempolicy from/to strings
3154 */
3155static const char * const policy_modes[] =
3156{
3157 [MPOL_DEFAULT] = "default",
3158 [MPOL_PREFERRED] = "prefer",
3159 [MPOL_BIND] = "bind",
3160 [MPOL_INTERLEAVE] = "interleave",
3161 [MPOL_WEIGHTED_INTERLEAVE] = "weighted interleave",
3162 [MPOL_LOCAL] = "local",
3163 [MPOL_PREFERRED_MANY] = "prefer (many)",
3164};
3165
3166#ifdef CONFIG_TMPFS
3167/**
3168 * mpol_parse_str - parse string to mempolicy, for tmpfs mpol mount option.
3169 * @str: string containing mempolicy to parse
3170 * @mpol: pointer to struct mempolicy pointer, returned on success.
3171 *
3172 * Format of input:
3173 * <mode>[=<flags>][:<nodelist>]
3174 *
3175 * Return: %0 on success, else %1
3176 */
3177int mpol_parse_str(char *str, struct mempolicy **mpol)
3178{
3179 struct mempolicy *new = NULL;
3180 unsigned short mode_flags;
3181 nodemask_t nodes;
3182 char *nodelist = strchr(str, ':');
3183 char *flags = strchr(str, '=');
3184 int err = 1, mode;
3185
3186 if (flags)
3187 *flags++ = '\0'; /* terminate mode string */
3188
3189 if (nodelist) {
3190 /* NUL-terminate mode or flags string */
3191 *nodelist++ = '\0';
3192 if (nodelist_parse(nodelist, nodes))
3193 goto out;
3194 if (!nodes_subset(nodes, node_states[N_MEMORY]))
3195 goto out;
3196 } else
3197 nodes_clear(nodes);
3198
3199 mode = match_string(policy_modes, MPOL_MAX, str);
3200 if (mode < 0)
3201 goto out;
3202
3203 switch (mode) {
3204 case MPOL_PREFERRED:
3205 /*
3206 * Insist on a nodelist of one node only, although later
3207 * we use first_node(nodes) to grab a single node, so here
3208 * nodelist (or nodes) cannot be empty.
3209 */
3210 if (nodelist) {
3211 char *rest = nodelist;
3212 while (isdigit(*rest))
3213 rest++;
3214 if (*rest)
3215 goto out;
3216 if (nodes_empty(nodes))
3217 goto out;
3218 }
3219 break;
3220 case MPOL_INTERLEAVE:
3221 case MPOL_WEIGHTED_INTERLEAVE:
3222 /*
3223 * Default to online nodes with memory if no nodelist
3224 */
3225 if (!nodelist)
3226 nodes = node_states[N_MEMORY];
3227 break;
3228 case MPOL_LOCAL:
3229 /*
3230 * Don't allow a nodelist; mpol_new() checks flags
3231 */
3232 if (nodelist)
3233 goto out;
3234 break;
3235 case MPOL_DEFAULT:
3236 /*
3237 * Insist on a empty nodelist
3238 */
3239 if (!nodelist)
3240 err = 0;
3241 goto out;
3242 case MPOL_PREFERRED_MANY:
3243 case MPOL_BIND:
3244 /*
3245 * Insist on a nodelist
3246 */
3247 if (!nodelist)
3248 goto out;
3249 }
3250
3251 mode_flags = 0;
3252 if (flags) {
3253 /*
3254 * Currently, we only support two mutually exclusive
3255 * mode flags.
3256 */
3257 if (!strcmp(flags, "static"))
3258 mode_flags |= MPOL_F_STATIC_NODES;
3259 else if (!strcmp(flags, "relative"))
3260 mode_flags |= MPOL_F_RELATIVE_NODES;
3261 else
3262 goto out;
3263 }
3264
3265 new = mpol_new(mode, mode_flags, &nodes);
3266 if (IS_ERR(new))
3267 goto out;
3268
3269 /*
3270 * Save nodes for mpol_to_str() to show the tmpfs mount options
3271 * for /proc/mounts, /proc/pid/mounts and /proc/pid/mountinfo.
3272 */
3273 if (mode != MPOL_PREFERRED) {
3274 new->nodes = nodes;
3275 } else if (nodelist) {
3276 nodes_clear(new->nodes);
3277 node_set(first_node(nodes), new->nodes);
3278 } else {
3279 new->mode = MPOL_LOCAL;
3280 }
3281
3282 /*
3283 * Save nodes for contextualization: this will be used to "clone"
3284 * the mempolicy in a specific context [cpuset] at a later time.
3285 */
3286 new->w.user_nodemask = nodes;
3287
3288 err = 0;
3289
3290out:
3291 /* Restore string for error message */
3292 if (nodelist)
3293 *--nodelist = ':';
3294 if (flags)
3295 *--flags = '=';
3296 if (!err)
3297 *mpol = new;
3298 return err;
3299}
3300#endif /* CONFIG_TMPFS */
3301
3302/**
3303 * mpol_to_str - format a mempolicy structure for printing
3304 * @buffer: to contain formatted mempolicy string
3305 * @maxlen: length of @buffer
3306 * @pol: pointer to mempolicy to be formatted
3307 *
3308 * Convert @pol into a string. If @buffer is too short, truncate the string.
3309 * Recommend a @maxlen of at least 51 for the longest mode, "weighted
3310 * interleave", plus the longest flag flags, "relative|balancing", and to
3311 * display at least a few node ids.
3312 */
3313void mpol_to_str(char *buffer, int maxlen, struct mempolicy *pol)
3314{
3315 char *p = buffer;
3316 nodemask_t nodes = NODE_MASK_NONE;
3317 unsigned short mode = MPOL_DEFAULT;
3318 unsigned short flags = 0;
3319
3320 if (pol &&
3321 pol != &default_policy &&
3322 !(pol >= &preferred_node_policy[0] &&
3323 pol <= &preferred_node_policy[ARRAY_SIZE(preferred_node_policy) - 1])) {
3324 mode = pol->mode;
3325 flags = pol->flags;
3326 }
3327
3328 switch (mode) {
3329 case MPOL_DEFAULT:
3330 case MPOL_LOCAL:
3331 break;
3332 case MPOL_PREFERRED:
3333 case MPOL_PREFERRED_MANY:
3334 case MPOL_BIND:
3335 case MPOL_INTERLEAVE:
3336 case MPOL_WEIGHTED_INTERLEAVE:
3337 nodes = pol->nodes;
3338 break;
3339 default:
3340 WARN_ON_ONCE(1);
3341 snprintf(p, maxlen, "unknown");
3342 return;
3343 }
3344
3345 p += snprintf(p, maxlen, "%s", policy_modes[mode]);
3346
3347 if (flags & MPOL_MODE_FLAGS) {
3348 p += snprintf(p, buffer + maxlen - p, "=");
3349
3350 /*
3351 * Static and relative are mutually exclusive.
3352 */
3353 if (flags & MPOL_F_STATIC_NODES)
3354 p += snprintf(p, buffer + maxlen - p, "static");
3355 else if (flags & MPOL_F_RELATIVE_NODES)
3356 p += snprintf(p, buffer + maxlen - p, "relative");
3357
3358 if (flags & MPOL_F_NUMA_BALANCING) {
3359 if (!is_power_of_2(flags & MPOL_MODE_FLAGS))
3360 p += snprintf(p, buffer + maxlen - p, "|");
3361 p += snprintf(p, buffer + maxlen - p, "balancing");
3362 }
3363 }
3364
3365 if (!nodes_empty(nodes))
3366 p += scnprintf(p, buffer + maxlen - p, ":%*pbl",
3367 nodemask_pr_args(&nodes));
3368}
3369
3370#ifdef CONFIG_SYSFS
3371struct iw_node_attr {
3372 struct kobj_attribute kobj_attr;
3373 int nid;
3374};
3375
3376static ssize_t node_show(struct kobject *kobj, struct kobj_attribute *attr,
3377 char *buf)
3378{
3379 struct iw_node_attr *node_attr;
3380 u8 weight;
3381
3382 node_attr = container_of(attr, struct iw_node_attr, kobj_attr);
3383 weight = get_il_weight(node_attr->nid);
3384 return sysfs_emit(buf, "%d\n", weight);
3385}
3386
3387static ssize_t node_store(struct kobject *kobj, struct kobj_attribute *attr,
3388 const char *buf, size_t count)
3389{
3390 struct iw_node_attr *node_attr;
3391 u8 *new;
3392 u8 *old;
3393 u8 weight = 0;
3394
3395 node_attr = container_of(attr, struct iw_node_attr, kobj_attr);
3396 if (count == 0 || sysfs_streq(buf, ""))
3397 weight = 0;
3398 else if (kstrtou8(buf, 0, &weight))
3399 return -EINVAL;
3400
3401 new = kzalloc(nr_node_ids, GFP_KERNEL);
3402 if (!new)
3403 return -ENOMEM;
3404
3405 mutex_lock(&iw_table_lock);
3406 old = rcu_dereference_protected(iw_table,
3407 lockdep_is_held(&iw_table_lock));
3408 if (old)
3409 memcpy(new, old, nr_node_ids);
3410 new[node_attr->nid] = weight;
3411 rcu_assign_pointer(iw_table, new);
3412 mutex_unlock(&iw_table_lock);
3413 synchronize_rcu();
3414 kfree(old);
3415 return count;
3416}
3417
3418static struct iw_node_attr **node_attrs;
3419
3420static void sysfs_wi_node_release(struct iw_node_attr *node_attr,
3421 struct kobject *parent)
3422{
3423 if (!node_attr)
3424 return;
3425 sysfs_remove_file(parent, &node_attr->kobj_attr.attr);
3426 kfree(node_attr->kobj_attr.attr.name);
3427 kfree(node_attr);
3428}
3429
3430static void sysfs_wi_release(struct kobject *wi_kobj)
3431{
3432 int i;
3433
3434 for (i = 0; i < nr_node_ids; i++)
3435 sysfs_wi_node_release(node_attrs[i], wi_kobj);
3436 kobject_put(wi_kobj);
3437}
3438
3439static const struct kobj_type wi_ktype = {
3440 .sysfs_ops = &kobj_sysfs_ops,
3441 .release = sysfs_wi_release,
3442};
3443
3444static int add_weight_node(int nid, struct kobject *wi_kobj)
3445{
3446 struct iw_node_attr *node_attr;
3447 char *name;
3448
3449 node_attr = kzalloc(sizeof(*node_attr), GFP_KERNEL);
3450 if (!node_attr)
3451 return -ENOMEM;
3452
3453 name = kasprintf(GFP_KERNEL, "node%d", nid);
3454 if (!name) {
3455 kfree(node_attr);
3456 return -ENOMEM;
3457 }
3458
3459 sysfs_attr_init(&node_attr->kobj_attr.attr);
3460 node_attr->kobj_attr.attr.name = name;
3461 node_attr->kobj_attr.attr.mode = 0644;
3462 node_attr->kobj_attr.show = node_show;
3463 node_attr->kobj_attr.store = node_store;
3464 node_attr->nid = nid;
3465
3466 if (sysfs_create_file(wi_kobj, &node_attr->kobj_attr.attr)) {
3467 kfree(node_attr->kobj_attr.attr.name);
3468 kfree(node_attr);
3469 pr_err("failed to add attribute to weighted_interleave\n");
3470 return -ENOMEM;
3471 }
3472
3473 node_attrs[nid] = node_attr;
3474 return 0;
3475}
3476
3477static int add_weighted_interleave_group(struct kobject *root_kobj)
3478{
3479 struct kobject *wi_kobj;
3480 int nid, err;
3481
3482 wi_kobj = kzalloc(sizeof(struct kobject), GFP_KERNEL);
3483 if (!wi_kobj)
3484 return -ENOMEM;
3485
3486 err = kobject_init_and_add(wi_kobj, &wi_ktype, root_kobj,
3487 "weighted_interleave");
3488 if (err) {
3489 kfree(wi_kobj);
3490 return err;
3491 }
3492
3493 for_each_node_state(nid, N_POSSIBLE) {
3494 err = add_weight_node(nid, wi_kobj);
3495 if (err) {
3496 pr_err("failed to add sysfs [node%d]\n", nid);
3497 break;
3498 }
3499 }
3500 if (err)
3501 kobject_put(wi_kobj);
3502 return 0;
3503}
3504
3505static void mempolicy_kobj_release(struct kobject *kobj)
3506{
3507 u8 *old;
3508
3509 mutex_lock(&iw_table_lock);
3510 old = rcu_dereference_protected(iw_table,
3511 lockdep_is_held(&iw_table_lock));
3512 rcu_assign_pointer(iw_table, NULL);
3513 mutex_unlock(&iw_table_lock);
3514 synchronize_rcu();
3515 kfree(old);
3516 kfree(node_attrs);
3517 kfree(kobj);
3518}
3519
3520static const struct kobj_type mempolicy_ktype = {
3521 .release = mempolicy_kobj_release
3522};
3523
3524static int __init mempolicy_sysfs_init(void)
3525{
3526 int err;
3527 static struct kobject *mempolicy_kobj;
3528
3529 mempolicy_kobj = kzalloc(sizeof(*mempolicy_kobj), GFP_KERNEL);
3530 if (!mempolicy_kobj) {
3531 err = -ENOMEM;
3532 goto err_out;
3533 }
3534
3535 node_attrs = kcalloc(nr_node_ids, sizeof(struct iw_node_attr *),
3536 GFP_KERNEL);
3537 if (!node_attrs) {
3538 err = -ENOMEM;
3539 goto mempol_out;
3540 }
3541
3542 err = kobject_init_and_add(mempolicy_kobj, &mempolicy_ktype, mm_kobj,
3543 "mempolicy");
3544 if (err)
3545 goto node_out;
3546
3547 err = add_weighted_interleave_group(mempolicy_kobj);
3548 if (err) {
3549 pr_err("mempolicy sysfs structure failed to initialize\n");
3550 kobject_put(mempolicy_kobj);
3551 return err;
3552 }
3553
3554 return err;
3555node_out:
3556 kfree(node_attrs);
3557mempol_out:
3558 kfree(mempolicy_kobj);
3559err_out:
3560 pr_err("failed to add mempolicy kobject to the system\n");
3561 return err;
3562}
3563
3564late_initcall(mempolicy_sysfs_init);
3565#endif /* CONFIG_SYSFS */