Loading...
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * Simple NUMA memory policy for the Linux kernel.
4 *
5 * Copyright 2003,2004 Andi Kleen, SuSE Labs.
6 * (C) Copyright 2005 Christoph Lameter, Silicon Graphics, Inc.
7 *
8 * NUMA policy allows the user to give hints in which node(s) memory should
9 * be allocated.
10 *
11 * Support four policies per VMA and per process:
12 *
13 * The VMA policy has priority over the process policy for a page fault.
14 *
15 * interleave Allocate memory interleaved over a set of nodes,
16 * with normal fallback if it fails.
17 * For VMA based allocations this interleaves based on the
18 * offset into the backing object or offset into the mapping
19 * for anonymous memory. For process policy an process counter
20 * is used.
21 *
22 * bind Only allocate memory on a specific set of nodes,
23 * no fallback.
24 * FIXME: memory is allocated starting with the first node
25 * to the last. It would be better if bind would truly restrict
26 * the allocation to memory nodes instead
27 *
28 * preferred Try a specific node first before normal fallback.
29 * As a special case NUMA_NO_NODE here means do the allocation
30 * on the local CPU. This is normally identical to default,
31 * but useful to set in a VMA when you have a non default
32 * process policy.
33 *
34 * default Allocate on the local node first, or when on a VMA
35 * use the process policy. This is what Linux always did
36 * in a NUMA aware kernel and still does by, ahem, default.
37 *
38 * The process policy is applied for most non interrupt memory allocations
39 * in that process' context. Interrupts ignore the policies and always
40 * try to allocate on the local CPU. The VMA policy is only applied for memory
41 * allocations for a VMA in the VM.
42 *
43 * Currently there are a few corner cases in swapping where the policy
44 * is not applied, but the majority should be handled. When process policy
45 * is used it is not remembered over swap outs/swap ins.
46 *
47 * Only the highest zone in the zone hierarchy gets policied. Allocations
48 * requesting a lower zone just use default policy. This implies that
49 * on systems with highmem kernel lowmem allocation don't get policied.
50 * Same with GFP_DMA allocations.
51 *
52 * For shmfs/tmpfs/hugetlbfs shared memory the policy is shared between
53 * all users and remembered even when nobody has memory mapped.
54 */
55
56/* Notebook:
57 fix mmap readahead to honour policy and enable policy for any page cache
58 object
59 statistics for bigpages
60 global policy for page cache? currently it uses process policy. Requires
61 first item above.
62 handle mremap for shared memory (currently ignored for the policy)
63 grows down?
64 make bind policy root only? It can trigger oom much faster and the
65 kernel is not always grateful with that.
66*/
67
68#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
69
70#include <linux/mempolicy.h>
71#include <linux/pagewalk.h>
72#include <linux/highmem.h>
73#include <linux/hugetlb.h>
74#include <linux/kernel.h>
75#include <linux/sched.h>
76#include <linux/sched/mm.h>
77#include <linux/sched/numa_balancing.h>
78#include <linux/sched/task.h>
79#include <linux/nodemask.h>
80#include <linux/cpuset.h>
81#include <linux/slab.h>
82#include <linux/string.h>
83#include <linux/export.h>
84#include <linux/nsproxy.h>
85#include <linux/interrupt.h>
86#include <linux/init.h>
87#include <linux/compat.h>
88#include <linux/ptrace.h>
89#include <linux/swap.h>
90#include <linux/seq_file.h>
91#include <linux/proc_fs.h>
92#include <linux/migrate.h>
93#include <linux/ksm.h>
94#include <linux/rmap.h>
95#include <linux/security.h>
96#include <linux/syscalls.h>
97#include <linux/ctype.h>
98#include <linux/mm_inline.h>
99#include <linux/mmu_notifier.h>
100#include <linux/printk.h>
101#include <linux/swapops.h>
102
103#include <asm/tlbflush.h>
104#include <linux/uaccess.h>
105
106#include "internal.h"
107
108/* Internal flags */
109#define MPOL_MF_DISCONTIG_OK (MPOL_MF_INTERNAL << 0) /* Skip checks for continuous vmas */
110#define MPOL_MF_INVERT (MPOL_MF_INTERNAL << 1) /* Invert check for nodemask */
111
112static struct kmem_cache *policy_cache;
113static struct kmem_cache *sn_cache;
114
115/* Highest zone. An specific allocation for a zone below that is not
116 policied. */
117enum zone_type policy_zone = 0;
118
119/*
120 * run-time system-wide default policy => local allocation
121 */
122static struct mempolicy default_policy = {
123 .refcnt = ATOMIC_INIT(1), /* never free it */
124 .mode = MPOL_PREFERRED,
125 .flags = MPOL_F_LOCAL,
126};
127
128static struct mempolicy preferred_node_policy[MAX_NUMNODES];
129
130struct mempolicy *get_task_policy(struct task_struct *p)
131{
132 struct mempolicy *pol = p->mempolicy;
133 int node;
134
135 if (pol)
136 return pol;
137
138 node = numa_node_id();
139 if (node != NUMA_NO_NODE) {
140 pol = &preferred_node_policy[node];
141 /* preferred_node_policy is not initialised early in boot */
142 if (pol->mode)
143 return pol;
144 }
145
146 return &default_policy;
147}
148
149static const struct mempolicy_operations {
150 int (*create)(struct mempolicy *pol, const nodemask_t *nodes);
151 void (*rebind)(struct mempolicy *pol, const nodemask_t *nodes);
152} mpol_ops[MPOL_MAX];
153
154static inline int mpol_store_user_nodemask(const struct mempolicy *pol)
155{
156 return pol->flags & MPOL_MODE_FLAGS;
157}
158
159static void mpol_relative_nodemask(nodemask_t *ret, const nodemask_t *orig,
160 const nodemask_t *rel)
161{
162 nodemask_t tmp;
163 nodes_fold(tmp, *orig, nodes_weight(*rel));
164 nodes_onto(*ret, tmp, *rel);
165}
166
167static int mpol_new_interleave(struct mempolicy *pol, const nodemask_t *nodes)
168{
169 if (nodes_empty(*nodes))
170 return -EINVAL;
171 pol->v.nodes = *nodes;
172 return 0;
173}
174
175static int mpol_new_preferred(struct mempolicy *pol, const nodemask_t *nodes)
176{
177 if (!nodes)
178 pol->flags |= MPOL_F_LOCAL; /* local allocation */
179 else if (nodes_empty(*nodes))
180 return -EINVAL; /* no allowed nodes */
181 else
182 pol->v.preferred_node = first_node(*nodes);
183 return 0;
184}
185
186static int mpol_new_bind(struct mempolicy *pol, const nodemask_t *nodes)
187{
188 if (nodes_empty(*nodes))
189 return -EINVAL;
190 pol->v.nodes = *nodes;
191 return 0;
192}
193
194/*
195 * mpol_set_nodemask is called after mpol_new() to set up the nodemask, if
196 * any, for the new policy. mpol_new() has already validated the nodes
197 * parameter with respect to the policy mode and flags. But, we need to
198 * handle an empty nodemask with MPOL_PREFERRED here.
199 *
200 * Must be called holding task's alloc_lock to protect task's mems_allowed
201 * and mempolicy. May also be called holding the mmap_semaphore for write.
202 */
203static int mpol_set_nodemask(struct mempolicy *pol,
204 const nodemask_t *nodes, struct nodemask_scratch *nsc)
205{
206 int ret;
207
208 /* if mode is MPOL_DEFAULT, pol is NULL. This is right. */
209 if (pol == NULL)
210 return 0;
211 /* Check N_MEMORY */
212 nodes_and(nsc->mask1,
213 cpuset_current_mems_allowed, node_states[N_MEMORY]);
214
215 VM_BUG_ON(!nodes);
216 if (pol->mode == MPOL_PREFERRED && nodes_empty(*nodes))
217 nodes = NULL; /* explicit local allocation */
218 else {
219 if (pol->flags & MPOL_F_RELATIVE_NODES)
220 mpol_relative_nodemask(&nsc->mask2, nodes, &nsc->mask1);
221 else
222 nodes_and(nsc->mask2, *nodes, nsc->mask1);
223
224 if (mpol_store_user_nodemask(pol))
225 pol->w.user_nodemask = *nodes;
226 else
227 pol->w.cpuset_mems_allowed =
228 cpuset_current_mems_allowed;
229 }
230
231 if (nodes)
232 ret = mpol_ops[pol->mode].create(pol, &nsc->mask2);
233 else
234 ret = mpol_ops[pol->mode].create(pol, NULL);
235 return ret;
236}
237
238/*
239 * This function just creates a new policy, does some check and simple
240 * initialization. You must invoke mpol_set_nodemask() to set nodes.
241 */
242static struct mempolicy *mpol_new(unsigned short mode, unsigned short flags,
243 nodemask_t *nodes)
244{
245 struct mempolicy *policy;
246
247 pr_debug("setting mode %d flags %d nodes[0] %lx\n",
248 mode, flags, nodes ? nodes_addr(*nodes)[0] : NUMA_NO_NODE);
249
250 if (mode == MPOL_DEFAULT) {
251 if (nodes && !nodes_empty(*nodes))
252 return ERR_PTR(-EINVAL);
253 return NULL;
254 }
255 VM_BUG_ON(!nodes);
256
257 /*
258 * MPOL_PREFERRED cannot be used with MPOL_F_STATIC_NODES or
259 * MPOL_F_RELATIVE_NODES if the nodemask is empty (local allocation).
260 * All other modes require a valid pointer to a non-empty nodemask.
261 */
262 if (mode == MPOL_PREFERRED) {
263 if (nodes_empty(*nodes)) {
264 if (((flags & MPOL_F_STATIC_NODES) ||
265 (flags & MPOL_F_RELATIVE_NODES)))
266 return ERR_PTR(-EINVAL);
267 }
268 } else if (mode == MPOL_LOCAL) {
269 if (!nodes_empty(*nodes) ||
270 (flags & MPOL_F_STATIC_NODES) ||
271 (flags & MPOL_F_RELATIVE_NODES))
272 return ERR_PTR(-EINVAL);
273 mode = MPOL_PREFERRED;
274 } else if (nodes_empty(*nodes))
275 return ERR_PTR(-EINVAL);
276 policy = kmem_cache_alloc(policy_cache, GFP_KERNEL);
277 if (!policy)
278 return ERR_PTR(-ENOMEM);
279 atomic_set(&policy->refcnt, 1);
280 policy->mode = mode;
281 policy->flags = flags;
282
283 return policy;
284}
285
286/* Slow path of a mpol destructor. */
287void __mpol_put(struct mempolicy *p)
288{
289 if (!atomic_dec_and_test(&p->refcnt))
290 return;
291 kmem_cache_free(policy_cache, p);
292}
293
294static void mpol_rebind_default(struct mempolicy *pol, const nodemask_t *nodes)
295{
296}
297
298static void mpol_rebind_nodemask(struct mempolicy *pol, const nodemask_t *nodes)
299{
300 nodemask_t tmp;
301
302 if (pol->flags & MPOL_F_STATIC_NODES)
303 nodes_and(tmp, pol->w.user_nodemask, *nodes);
304 else if (pol->flags & MPOL_F_RELATIVE_NODES)
305 mpol_relative_nodemask(&tmp, &pol->w.user_nodemask, nodes);
306 else {
307 nodes_remap(tmp, pol->v.nodes,pol->w.cpuset_mems_allowed,
308 *nodes);
309 pol->w.cpuset_mems_allowed = *nodes;
310 }
311
312 if (nodes_empty(tmp))
313 tmp = *nodes;
314
315 pol->v.nodes = tmp;
316}
317
318static void mpol_rebind_preferred(struct mempolicy *pol,
319 const nodemask_t *nodes)
320{
321 nodemask_t tmp;
322
323 if (pol->flags & MPOL_F_STATIC_NODES) {
324 int node = first_node(pol->w.user_nodemask);
325
326 if (node_isset(node, *nodes)) {
327 pol->v.preferred_node = node;
328 pol->flags &= ~MPOL_F_LOCAL;
329 } else
330 pol->flags |= MPOL_F_LOCAL;
331 } else if (pol->flags & MPOL_F_RELATIVE_NODES) {
332 mpol_relative_nodemask(&tmp, &pol->w.user_nodemask, nodes);
333 pol->v.preferred_node = first_node(tmp);
334 } else if (!(pol->flags & MPOL_F_LOCAL)) {
335 pol->v.preferred_node = node_remap(pol->v.preferred_node,
336 pol->w.cpuset_mems_allowed,
337 *nodes);
338 pol->w.cpuset_mems_allowed = *nodes;
339 }
340}
341
342/*
343 * mpol_rebind_policy - Migrate a policy to a different set of nodes
344 *
345 * Per-vma policies are protected by mmap_sem. Allocations using per-task
346 * policies are protected by task->mems_allowed_seq to prevent a premature
347 * OOM/allocation failure due to parallel nodemask modification.
348 */
349static void mpol_rebind_policy(struct mempolicy *pol, const nodemask_t *newmask)
350{
351 if (!pol)
352 return;
353 if (!mpol_store_user_nodemask(pol) && !(pol->flags & MPOL_F_LOCAL) &&
354 nodes_equal(pol->w.cpuset_mems_allowed, *newmask))
355 return;
356
357 mpol_ops[pol->mode].rebind(pol, newmask);
358}
359
360/*
361 * Wrapper for mpol_rebind_policy() that just requires task
362 * pointer, and updates task mempolicy.
363 *
364 * Called with task's alloc_lock held.
365 */
366
367void mpol_rebind_task(struct task_struct *tsk, const nodemask_t *new)
368{
369 mpol_rebind_policy(tsk->mempolicy, new);
370}
371
372/*
373 * Rebind each vma in mm to new nodemask.
374 *
375 * Call holding a reference to mm. Takes mm->mmap_sem during call.
376 */
377
378void mpol_rebind_mm(struct mm_struct *mm, nodemask_t *new)
379{
380 struct vm_area_struct *vma;
381
382 down_write(&mm->mmap_sem);
383 for (vma = mm->mmap; vma; vma = vma->vm_next)
384 mpol_rebind_policy(vma->vm_policy, new);
385 up_write(&mm->mmap_sem);
386}
387
388static const struct mempolicy_operations mpol_ops[MPOL_MAX] = {
389 [MPOL_DEFAULT] = {
390 .rebind = mpol_rebind_default,
391 },
392 [MPOL_INTERLEAVE] = {
393 .create = mpol_new_interleave,
394 .rebind = mpol_rebind_nodemask,
395 },
396 [MPOL_PREFERRED] = {
397 .create = mpol_new_preferred,
398 .rebind = mpol_rebind_preferred,
399 },
400 [MPOL_BIND] = {
401 .create = mpol_new_bind,
402 .rebind = mpol_rebind_nodemask,
403 },
404};
405
406static int migrate_page_add(struct page *page, struct list_head *pagelist,
407 unsigned long flags);
408
409struct queue_pages {
410 struct list_head *pagelist;
411 unsigned long flags;
412 nodemask_t *nmask;
413 struct vm_area_struct *prev;
414};
415
416/*
417 * Check if the page's nid is in qp->nmask.
418 *
419 * If MPOL_MF_INVERT is set in qp->flags, check if the nid is
420 * in the invert of qp->nmask.
421 */
422static inline bool queue_pages_required(struct page *page,
423 struct queue_pages *qp)
424{
425 int nid = page_to_nid(page);
426 unsigned long flags = qp->flags;
427
428 return node_isset(nid, *qp->nmask) == !(flags & MPOL_MF_INVERT);
429}
430
431/*
432 * queue_pages_pmd() has four possible return values:
433 * 0 - pages are placed on the right node or queued successfully.
434 * 1 - there is unmovable page, and MPOL_MF_MOVE* & MPOL_MF_STRICT were
435 * specified.
436 * 2 - THP was split.
437 * -EIO - is migration entry or only MPOL_MF_STRICT was specified and an
438 * existing page was already on a node that does not follow the
439 * policy.
440 */
441static int queue_pages_pmd(pmd_t *pmd, spinlock_t *ptl, unsigned long addr,
442 unsigned long end, struct mm_walk *walk)
443{
444 int ret = 0;
445 struct page *page;
446 struct queue_pages *qp = walk->private;
447 unsigned long flags;
448
449 if (unlikely(is_pmd_migration_entry(*pmd))) {
450 ret = -EIO;
451 goto unlock;
452 }
453 page = pmd_page(*pmd);
454 if (is_huge_zero_page(page)) {
455 spin_unlock(ptl);
456 __split_huge_pmd(walk->vma, pmd, addr, false, NULL);
457 ret = 2;
458 goto out;
459 }
460 if (!queue_pages_required(page, qp))
461 goto unlock;
462
463 flags = qp->flags;
464 /* go to thp migration */
465 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) {
466 if (!vma_migratable(walk->vma) ||
467 migrate_page_add(page, qp->pagelist, flags)) {
468 ret = 1;
469 goto unlock;
470 }
471 } else
472 ret = -EIO;
473unlock:
474 spin_unlock(ptl);
475out:
476 return ret;
477}
478
479/*
480 * Scan through pages checking if pages follow certain conditions,
481 * and move them to the pagelist if they do.
482 *
483 * queue_pages_pte_range() has three possible return values:
484 * 0 - pages are placed on the right node or queued successfully.
485 * 1 - there is unmovable page, and MPOL_MF_MOVE* & MPOL_MF_STRICT were
486 * specified.
487 * -EIO - only MPOL_MF_STRICT was specified and an existing page was already
488 * on a node that does not follow the policy.
489 */
490static int queue_pages_pte_range(pmd_t *pmd, unsigned long addr,
491 unsigned long end, struct mm_walk *walk)
492{
493 struct vm_area_struct *vma = walk->vma;
494 struct page *page;
495 struct queue_pages *qp = walk->private;
496 unsigned long flags = qp->flags;
497 int ret;
498 bool has_unmovable = false;
499 pte_t *pte;
500 spinlock_t *ptl;
501
502 ptl = pmd_trans_huge_lock(pmd, vma);
503 if (ptl) {
504 ret = queue_pages_pmd(pmd, ptl, addr, end, walk);
505 if (ret != 2)
506 return ret;
507 }
508 /* THP was split, fall through to pte walk */
509
510 if (pmd_trans_unstable(pmd))
511 return 0;
512
513 pte = pte_offset_map_lock(walk->mm, pmd, addr, &ptl);
514 for (; addr != end; pte++, addr += PAGE_SIZE) {
515 if (!pte_present(*pte))
516 continue;
517 page = vm_normal_page(vma, addr, *pte);
518 if (!page)
519 continue;
520 /*
521 * vm_normal_page() filters out zero pages, but there might
522 * still be PageReserved pages to skip, perhaps in a VDSO.
523 */
524 if (PageReserved(page))
525 continue;
526 if (!queue_pages_required(page, qp))
527 continue;
528 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) {
529 /* MPOL_MF_STRICT must be specified if we get here */
530 if (!vma_migratable(vma)) {
531 has_unmovable = true;
532 break;
533 }
534
535 /*
536 * Do not abort immediately since there may be
537 * temporary off LRU pages in the range. Still
538 * need migrate other LRU pages.
539 */
540 if (migrate_page_add(page, qp->pagelist, flags))
541 has_unmovable = true;
542 } else
543 break;
544 }
545 pte_unmap_unlock(pte - 1, ptl);
546 cond_resched();
547
548 if (has_unmovable)
549 return 1;
550
551 return addr != end ? -EIO : 0;
552}
553
554static int queue_pages_hugetlb(pte_t *pte, unsigned long hmask,
555 unsigned long addr, unsigned long end,
556 struct mm_walk *walk)
557{
558#ifdef CONFIG_HUGETLB_PAGE
559 struct queue_pages *qp = walk->private;
560 unsigned long flags = qp->flags;
561 struct page *page;
562 spinlock_t *ptl;
563 pte_t entry;
564
565 ptl = huge_pte_lock(hstate_vma(walk->vma), walk->mm, pte);
566 entry = huge_ptep_get(pte);
567 if (!pte_present(entry))
568 goto unlock;
569 page = pte_page(entry);
570 if (!queue_pages_required(page, qp))
571 goto unlock;
572 /* With MPOL_MF_MOVE, we migrate only unshared hugepage. */
573 if (flags & (MPOL_MF_MOVE_ALL) ||
574 (flags & MPOL_MF_MOVE && page_mapcount(page) == 1))
575 isolate_huge_page(page, qp->pagelist);
576unlock:
577 spin_unlock(ptl);
578#else
579 BUG();
580#endif
581 return 0;
582}
583
584#ifdef CONFIG_NUMA_BALANCING
585/*
586 * This is used to mark a range of virtual addresses to be inaccessible.
587 * These are later cleared by a NUMA hinting fault. Depending on these
588 * faults, pages may be migrated for better NUMA placement.
589 *
590 * This is assuming that NUMA faults are handled using PROT_NONE. If
591 * an architecture makes a different choice, it will need further
592 * changes to the core.
593 */
594unsigned long change_prot_numa(struct vm_area_struct *vma,
595 unsigned long addr, unsigned long end)
596{
597 int nr_updated;
598
599 nr_updated = change_protection(vma, addr, end, PAGE_NONE, 0, 1);
600 if (nr_updated)
601 count_vm_numa_events(NUMA_PTE_UPDATES, nr_updated);
602
603 return nr_updated;
604}
605#else
606static unsigned long change_prot_numa(struct vm_area_struct *vma,
607 unsigned long addr, unsigned long end)
608{
609 return 0;
610}
611#endif /* CONFIG_NUMA_BALANCING */
612
613static int queue_pages_test_walk(unsigned long start, unsigned long end,
614 struct mm_walk *walk)
615{
616 struct vm_area_struct *vma = walk->vma;
617 struct queue_pages *qp = walk->private;
618 unsigned long endvma = vma->vm_end;
619 unsigned long flags = qp->flags;
620
621 /*
622 * Need check MPOL_MF_STRICT to return -EIO if possible
623 * regardless of vma_migratable
624 */
625 if (!vma_migratable(vma) &&
626 !(flags & MPOL_MF_STRICT))
627 return 1;
628
629 if (endvma > end)
630 endvma = end;
631 if (vma->vm_start > start)
632 start = vma->vm_start;
633
634 if (!(flags & MPOL_MF_DISCONTIG_OK)) {
635 if (!vma->vm_next && vma->vm_end < end)
636 return -EFAULT;
637 if (qp->prev && qp->prev->vm_end < vma->vm_start)
638 return -EFAULT;
639 }
640
641 qp->prev = vma;
642
643 if (flags & MPOL_MF_LAZY) {
644 /* Similar to task_numa_work, skip inaccessible VMAs */
645 if (!is_vm_hugetlb_page(vma) &&
646 (vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)) &&
647 !(vma->vm_flags & VM_MIXEDMAP))
648 change_prot_numa(vma, start, endvma);
649 return 1;
650 }
651
652 /* queue pages from current vma */
653 if (flags & MPOL_MF_VALID)
654 return 0;
655 return 1;
656}
657
658static const struct mm_walk_ops queue_pages_walk_ops = {
659 .hugetlb_entry = queue_pages_hugetlb,
660 .pmd_entry = queue_pages_pte_range,
661 .test_walk = queue_pages_test_walk,
662};
663
664/*
665 * Walk through page tables and collect pages to be migrated.
666 *
667 * If pages found in a given range are on a set of nodes (determined by
668 * @nodes and @flags,) it's isolated and queued to the pagelist which is
669 * passed via @private.
670 *
671 * queue_pages_range() has three possible return values:
672 * 1 - there is unmovable page, but MPOL_MF_MOVE* & MPOL_MF_STRICT were
673 * specified.
674 * 0 - queue pages successfully or no misplaced page.
675 * errno - i.e. misplaced pages with MPOL_MF_STRICT specified (-EIO) or
676 * memory range specified by nodemask and maxnode points outside
677 * your accessible address space (-EFAULT)
678 */
679static int
680queue_pages_range(struct mm_struct *mm, unsigned long start, unsigned long end,
681 nodemask_t *nodes, unsigned long flags,
682 struct list_head *pagelist)
683{
684 struct queue_pages qp = {
685 .pagelist = pagelist,
686 .flags = flags,
687 .nmask = nodes,
688 .prev = NULL,
689 };
690
691 return walk_page_range(mm, start, end, &queue_pages_walk_ops, &qp);
692}
693
694/*
695 * Apply policy to a single VMA
696 * This must be called with the mmap_sem held for writing.
697 */
698static int vma_replace_policy(struct vm_area_struct *vma,
699 struct mempolicy *pol)
700{
701 int err;
702 struct mempolicy *old;
703 struct mempolicy *new;
704
705 pr_debug("vma %lx-%lx/%lx vm_ops %p vm_file %p set_policy %p\n",
706 vma->vm_start, vma->vm_end, vma->vm_pgoff,
707 vma->vm_ops, vma->vm_file,
708 vma->vm_ops ? vma->vm_ops->set_policy : NULL);
709
710 new = mpol_dup(pol);
711 if (IS_ERR(new))
712 return PTR_ERR(new);
713
714 if (vma->vm_ops && vma->vm_ops->set_policy) {
715 err = vma->vm_ops->set_policy(vma, new);
716 if (err)
717 goto err_out;
718 }
719
720 old = vma->vm_policy;
721 vma->vm_policy = new; /* protected by mmap_sem */
722 mpol_put(old);
723
724 return 0;
725 err_out:
726 mpol_put(new);
727 return err;
728}
729
730/* Step 2: apply policy to a range and do splits. */
731static int mbind_range(struct mm_struct *mm, unsigned long start,
732 unsigned long end, struct mempolicy *new_pol)
733{
734 struct vm_area_struct *next;
735 struct vm_area_struct *prev;
736 struct vm_area_struct *vma;
737 int err = 0;
738 pgoff_t pgoff;
739 unsigned long vmstart;
740 unsigned long vmend;
741
742 vma = find_vma(mm, start);
743 if (!vma || vma->vm_start > start)
744 return -EFAULT;
745
746 prev = vma->vm_prev;
747 if (start > vma->vm_start)
748 prev = vma;
749
750 for (; vma && vma->vm_start < end; prev = vma, vma = next) {
751 next = vma->vm_next;
752 vmstart = max(start, vma->vm_start);
753 vmend = min(end, vma->vm_end);
754
755 if (mpol_equal(vma_policy(vma), new_pol))
756 continue;
757
758 pgoff = vma->vm_pgoff +
759 ((vmstart - vma->vm_start) >> PAGE_SHIFT);
760 prev = vma_merge(mm, prev, vmstart, vmend, vma->vm_flags,
761 vma->anon_vma, vma->vm_file, pgoff,
762 new_pol, vma->vm_userfaultfd_ctx);
763 if (prev) {
764 vma = prev;
765 next = vma->vm_next;
766 if (mpol_equal(vma_policy(vma), new_pol))
767 continue;
768 /* vma_merge() joined vma && vma->next, case 8 */
769 goto replace;
770 }
771 if (vma->vm_start != vmstart) {
772 err = split_vma(vma->vm_mm, vma, vmstart, 1);
773 if (err)
774 goto out;
775 }
776 if (vma->vm_end != vmend) {
777 err = split_vma(vma->vm_mm, vma, vmend, 0);
778 if (err)
779 goto out;
780 }
781 replace:
782 err = vma_replace_policy(vma, new_pol);
783 if (err)
784 goto out;
785 }
786
787 out:
788 return err;
789}
790
791/* Set the process memory policy */
792static long do_set_mempolicy(unsigned short mode, unsigned short flags,
793 nodemask_t *nodes)
794{
795 struct mempolicy *new, *old;
796 NODEMASK_SCRATCH(scratch);
797 int ret;
798
799 if (!scratch)
800 return -ENOMEM;
801
802 new = mpol_new(mode, flags, nodes);
803 if (IS_ERR(new)) {
804 ret = PTR_ERR(new);
805 goto out;
806 }
807
808 task_lock(current);
809 ret = mpol_set_nodemask(new, nodes, scratch);
810 if (ret) {
811 task_unlock(current);
812 mpol_put(new);
813 goto out;
814 }
815 old = current->mempolicy;
816 current->mempolicy = new;
817 if (new && new->mode == MPOL_INTERLEAVE)
818 current->il_prev = MAX_NUMNODES-1;
819 task_unlock(current);
820 mpol_put(old);
821 ret = 0;
822out:
823 NODEMASK_SCRATCH_FREE(scratch);
824 return ret;
825}
826
827/*
828 * Return nodemask for policy for get_mempolicy() query
829 *
830 * Called with task's alloc_lock held
831 */
832static void get_policy_nodemask(struct mempolicy *p, nodemask_t *nodes)
833{
834 nodes_clear(*nodes);
835 if (p == &default_policy)
836 return;
837
838 switch (p->mode) {
839 case MPOL_BIND:
840 /* Fall through */
841 case MPOL_INTERLEAVE:
842 *nodes = p->v.nodes;
843 break;
844 case MPOL_PREFERRED:
845 if (!(p->flags & MPOL_F_LOCAL))
846 node_set(p->v.preferred_node, *nodes);
847 /* else return empty node mask for local allocation */
848 break;
849 default:
850 BUG();
851 }
852}
853
854static int lookup_node(struct mm_struct *mm, unsigned long addr)
855{
856 struct page *p;
857 int err;
858
859 int locked = 1;
860 err = get_user_pages_locked(addr & PAGE_MASK, 1, 0, &p, &locked);
861 if (err >= 0) {
862 err = page_to_nid(p);
863 put_page(p);
864 }
865 if (locked)
866 up_read(&mm->mmap_sem);
867 return err;
868}
869
870/* Retrieve NUMA policy */
871static long do_get_mempolicy(int *policy, nodemask_t *nmask,
872 unsigned long addr, unsigned long flags)
873{
874 int err;
875 struct mm_struct *mm = current->mm;
876 struct vm_area_struct *vma = NULL;
877 struct mempolicy *pol = current->mempolicy, *pol_refcount = NULL;
878
879 if (flags &
880 ~(unsigned long)(MPOL_F_NODE|MPOL_F_ADDR|MPOL_F_MEMS_ALLOWED))
881 return -EINVAL;
882
883 if (flags & MPOL_F_MEMS_ALLOWED) {
884 if (flags & (MPOL_F_NODE|MPOL_F_ADDR))
885 return -EINVAL;
886 *policy = 0; /* just so it's initialized */
887 task_lock(current);
888 *nmask = cpuset_current_mems_allowed;
889 task_unlock(current);
890 return 0;
891 }
892
893 if (flags & MPOL_F_ADDR) {
894 /*
895 * Do NOT fall back to task policy if the
896 * vma/shared policy at addr is NULL. We
897 * want to return MPOL_DEFAULT in this case.
898 */
899 down_read(&mm->mmap_sem);
900 vma = find_vma_intersection(mm, addr, addr+1);
901 if (!vma) {
902 up_read(&mm->mmap_sem);
903 return -EFAULT;
904 }
905 if (vma->vm_ops && vma->vm_ops->get_policy)
906 pol = vma->vm_ops->get_policy(vma, addr);
907 else
908 pol = vma->vm_policy;
909 } else if (addr)
910 return -EINVAL;
911
912 if (!pol)
913 pol = &default_policy; /* indicates default behavior */
914
915 if (flags & MPOL_F_NODE) {
916 if (flags & MPOL_F_ADDR) {
917 /*
918 * Take a refcount on the mpol, lookup_node()
919 * wil drop the mmap_sem, so after calling
920 * lookup_node() only "pol" remains valid, "vma"
921 * is stale.
922 */
923 pol_refcount = pol;
924 vma = NULL;
925 mpol_get(pol);
926 err = lookup_node(mm, addr);
927 if (err < 0)
928 goto out;
929 *policy = err;
930 } else if (pol == current->mempolicy &&
931 pol->mode == MPOL_INTERLEAVE) {
932 *policy = next_node_in(current->il_prev, pol->v.nodes);
933 } else {
934 err = -EINVAL;
935 goto out;
936 }
937 } else {
938 *policy = pol == &default_policy ? MPOL_DEFAULT :
939 pol->mode;
940 /*
941 * Internal mempolicy flags must be masked off before exposing
942 * the policy to userspace.
943 */
944 *policy |= (pol->flags & MPOL_MODE_FLAGS);
945 }
946
947 err = 0;
948 if (nmask) {
949 if (mpol_store_user_nodemask(pol)) {
950 *nmask = pol->w.user_nodemask;
951 } else {
952 task_lock(current);
953 get_policy_nodemask(pol, nmask);
954 task_unlock(current);
955 }
956 }
957
958 out:
959 mpol_cond_put(pol);
960 if (vma)
961 up_read(&mm->mmap_sem);
962 if (pol_refcount)
963 mpol_put(pol_refcount);
964 return err;
965}
966
967#ifdef CONFIG_MIGRATION
968/*
969 * page migration, thp tail pages can be passed.
970 */
971static int migrate_page_add(struct page *page, struct list_head *pagelist,
972 unsigned long flags)
973{
974 struct page *head = compound_head(page);
975 /*
976 * Avoid migrating a page that is shared with others.
977 */
978 if ((flags & MPOL_MF_MOVE_ALL) || page_mapcount(head) == 1) {
979 if (!isolate_lru_page(head)) {
980 list_add_tail(&head->lru, pagelist);
981 mod_node_page_state(page_pgdat(head),
982 NR_ISOLATED_ANON + page_is_file_cache(head),
983 hpage_nr_pages(head));
984 } else if (flags & MPOL_MF_STRICT) {
985 /*
986 * Non-movable page may reach here. And, there may be
987 * temporary off LRU pages or non-LRU movable pages.
988 * Treat them as unmovable pages since they can't be
989 * isolated, so they can't be moved at the moment. It
990 * should return -EIO for this case too.
991 */
992 return -EIO;
993 }
994 }
995
996 return 0;
997}
998
999/* page allocation callback for NUMA node migration */
1000struct page *alloc_new_node_page(struct page *page, unsigned long node)
1001{
1002 if (PageHuge(page))
1003 return alloc_huge_page_node(page_hstate(compound_head(page)),
1004 node);
1005 else if (PageTransHuge(page)) {
1006 struct page *thp;
1007
1008 thp = alloc_pages_node(node,
1009 (GFP_TRANSHUGE | __GFP_THISNODE),
1010 HPAGE_PMD_ORDER);
1011 if (!thp)
1012 return NULL;
1013 prep_transhuge_page(thp);
1014 return thp;
1015 } else
1016 return __alloc_pages_node(node, GFP_HIGHUSER_MOVABLE |
1017 __GFP_THISNODE, 0);
1018}
1019
1020/*
1021 * Migrate pages from one node to a target node.
1022 * Returns error or the number of pages not migrated.
1023 */
1024static int migrate_to_node(struct mm_struct *mm, int source, int dest,
1025 int flags)
1026{
1027 nodemask_t nmask;
1028 LIST_HEAD(pagelist);
1029 int err = 0;
1030
1031 nodes_clear(nmask);
1032 node_set(source, nmask);
1033
1034 /*
1035 * This does not "check" the range but isolates all pages that
1036 * need migration. Between passing in the full user address
1037 * space range and MPOL_MF_DISCONTIG_OK, this call can not fail.
1038 */
1039 VM_BUG_ON(!(flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)));
1040 queue_pages_range(mm, mm->mmap->vm_start, mm->task_size, &nmask,
1041 flags | MPOL_MF_DISCONTIG_OK, &pagelist);
1042
1043 if (!list_empty(&pagelist)) {
1044 err = migrate_pages(&pagelist, alloc_new_node_page, NULL, dest,
1045 MIGRATE_SYNC, MR_SYSCALL);
1046 if (err)
1047 putback_movable_pages(&pagelist);
1048 }
1049
1050 return err;
1051}
1052
1053/*
1054 * Move pages between the two nodesets so as to preserve the physical
1055 * layout as much as possible.
1056 *
1057 * Returns the number of page that could not be moved.
1058 */
1059int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
1060 const nodemask_t *to, int flags)
1061{
1062 int busy = 0;
1063 int err;
1064 nodemask_t tmp;
1065
1066 err = migrate_prep();
1067 if (err)
1068 return err;
1069
1070 down_read(&mm->mmap_sem);
1071
1072 /*
1073 * Find a 'source' bit set in 'tmp' whose corresponding 'dest'
1074 * bit in 'to' is not also set in 'tmp'. Clear the found 'source'
1075 * bit in 'tmp', and return that <source, dest> pair for migration.
1076 * The pair of nodemasks 'to' and 'from' define the map.
1077 *
1078 * If no pair of bits is found that way, fallback to picking some
1079 * pair of 'source' and 'dest' bits that are not the same. If the
1080 * 'source' and 'dest' bits are the same, this represents a node
1081 * that will be migrating to itself, so no pages need move.
1082 *
1083 * If no bits are left in 'tmp', or if all remaining bits left
1084 * in 'tmp' correspond to the same bit in 'to', return false
1085 * (nothing left to migrate).
1086 *
1087 * This lets us pick a pair of nodes to migrate between, such that
1088 * if possible the dest node is not already occupied by some other
1089 * source node, minimizing the risk of overloading the memory on a
1090 * node that would happen if we migrated incoming memory to a node
1091 * before migrating outgoing memory source that same node.
1092 *
1093 * A single scan of tmp is sufficient. As we go, we remember the
1094 * most recent <s, d> pair that moved (s != d). If we find a pair
1095 * that not only moved, but what's better, moved to an empty slot
1096 * (d is not set in tmp), then we break out then, with that pair.
1097 * Otherwise when we finish scanning from_tmp, we at least have the
1098 * most recent <s, d> pair that moved. If we get all the way through
1099 * the scan of tmp without finding any node that moved, much less
1100 * moved to an empty node, then there is nothing left worth migrating.
1101 */
1102
1103 tmp = *from;
1104 while (!nodes_empty(tmp)) {
1105 int s,d;
1106 int source = NUMA_NO_NODE;
1107 int dest = 0;
1108
1109 for_each_node_mask(s, tmp) {
1110
1111 /*
1112 * do_migrate_pages() tries to maintain the relative
1113 * node relationship of the pages established between
1114 * threads and memory areas.
1115 *
1116 * However if the number of source nodes is not equal to
1117 * the number of destination nodes we can not preserve
1118 * this node relative relationship. In that case, skip
1119 * copying memory from a node that is in the destination
1120 * mask.
1121 *
1122 * Example: [2,3,4] -> [3,4,5] moves everything.
1123 * [0-7] - > [3,4,5] moves only 0,1,2,6,7.
1124 */
1125
1126 if ((nodes_weight(*from) != nodes_weight(*to)) &&
1127 (node_isset(s, *to)))
1128 continue;
1129
1130 d = node_remap(s, *from, *to);
1131 if (s == d)
1132 continue;
1133
1134 source = s; /* Node moved. Memorize */
1135 dest = d;
1136
1137 /* dest not in remaining from nodes? */
1138 if (!node_isset(dest, tmp))
1139 break;
1140 }
1141 if (source == NUMA_NO_NODE)
1142 break;
1143
1144 node_clear(source, tmp);
1145 err = migrate_to_node(mm, source, dest, flags);
1146 if (err > 0)
1147 busy += err;
1148 if (err < 0)
1149 break;
1150 }
1151 up_read(&mm->mmap_sem);
1152 if (err < 0)
1153 return err;
1154 return busy;
1155
1156}
1157
1158/*
1159 * Allocate a new page for page migration based on vma policy.
1160 * Start by assuming the page is mapped by the same vma as contains @start.
1161 * Search forward from there, if not. N.B., this assumes that the
1162 * list of pages handed to migrate_pages()--which is how we get here--
1163 * is in virtual address order.
1164 */
1165static struct page *new_page(struct page *page, unsigned long start)
1166{
1167 struct vm_area_struct *vma;
1168 unsigned long uninitialized_var(address);
1169
1170 vma = find_vma(current->mm, start);
1171 while (vma) {
1172 address = page_address_in_vma(page, vma);
1173 if (address != -EFAULT)
1174 break;
1175 vma = vma->vm_next;
1176 }
1177
1178 if (PageHuge(page)) {
1179 return alloc_huge_page_vma(page_hstate(compound_head(page)),
1180 vma, address);
1181 } else if (PageTransHuge(page)) {
1182 struct page *thp;
1183
1184 thp = alloc_hugepage_vma(GFP_TRANSHUGE, vma, address,
1185 HPAGE_PMD_ORDER);
1186 if (!thp)
1187 return NULL;
1188 prep_transhuge_page(thp);
1189 return thp;
1190 }
1191 /*
1192 * if !vma, alloc_page_vma() will use task or system default policy
1193 */
1194 return alloc_page_vma(GFP_HIGHUSER_MOVABLE | __GFP_RETRY_MAYFAIL,
1195 vma, address);
1196}
1197#else
1198
1199static int migrate_page_add(struct page *page, struct list_head *pagelist,
1200 unsigned long flags)
1201{
1202 return -EIO;
1203}
1204
1205int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
1206 const nodemask_t *to, int flags)
1207{
1208 return -ENOSYS;
1209}
1210
1211static struct page *new_page(struct page *page, unsigned long start)
1212{
1213 return NULL;
1214}
1215#endif
1216
1217static long do_mbind(unsigned long start, unsigned long len,
1218 unsigned short mode, unsigned short mode_flags,
1219 nodemask_t *nmask, unsigned long flags)
1220{
1221 struct mm_struct *mm = current->mm;
1222 struct mempolicy *new;
1223 unsigned long end;
1224 int err;
1225 int ret;
1226 LIST_HEAD(pagelist);
1227
1228 if (flags & ~(unsigned long)MPOL_MF_VALID)
1229 return -EINVAL;
1230 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1231 return -EPERM;
1232
1233 if (start & ~PAGE_MASK)
1234 return -EINVAL;
1235
1236 if (mode == MPOL_DEFAULT)
1237 flags &= ~MPOL_MF_STRICT;
1238
1239 len = (len + PAGE_SIZE - 1) & PAGE_MASK;
1240 end = start + len;
1241
1242 if (end < start)
1243 return -EINVAL;
1244 if (end == start)
1245 return 0;
1246
1247 new = mpol_new(mode, mode_flags, nmask);
1248 if (IS_ERR(new))
1249 return PTR_ERR(new);
1250
1251 if (flags & MPOL_MF_LAZY)
1252 new->flags |= MPOL_F_MOF;
1253
1254 /*
1255 * If we are using the default policy then operation
1256 * on discontinuous address spaces is okay after all
1257 */
1258 if (!new)
1259 flags |= MPOL_MF_DISCONTIG_OK;
1260
1261 pr_debug("mbind %lx-%lx mode:%d flags:%d nodes:%lx\n",
1262 start, start + len, mode, mode_flags,
1263 nmask ? nodes_addr(*nmask)[0] : NUMA_NO_NODE);
1264
1265 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) {
1266
1267 err = migrate_prep();
1268 if (err)
1269 goto mpol_out;
1270 }
1271 {
1272 NODEMASK_SCRATCH(scratch);
1273 if (scratch) {
1274 down_write(&mm->mmap_sem);
1275 task_lock(current);
1276 err = mpol_set_nodemask(new, nmask, scratch);
1277 task_unlock(current);
1278 if (err)
1279 up_write(&mm->mmap_sem);
1280 } else
1281 err = -ENOMEM;
1282 NODEMASK_SCRATCH_FREE(scratch);
1283 }
1284 if (err)
1285 goto mpol_out;
1286
1287 ret = queue_pages_range(mm, start, end, nmask,
1288 flags | MPOL_MF_INVERT, &pagelist);
1289
1290 if (ret < 0) {
1291 err = ret;
1292 goto up_out;
1293 }
1294
1295 err = mbind_range(mm, start, end, new);
1296
1297 if (!err) {
1298 int nr_failed = 0;
1299
1300 if (!list_empty(&pagelist)) {
1301 WARN_ON_ONCE(flags & MPOL_MF_LAZY);
1302 nr_failed = migrate_pages(&pagelist, new_page, NULL,
1303 start, MIGRATE_SYNC, MR_MEMPOLICY_MBIND);
1304 if (nr_failed)
1305 putback_movable_pages(&pagelist);
1306 }
1307
1308 if ((ret > 0) || (nr_failed && (flags & MPOL_MF_STRICT)))
1309 err = -EIO;
1310 } else {
1311up_out:
1312 if (!list_empty(&pagelist))
1313 putback_movable_pages(&pagelist);
1314 }
1315
1316 up_write(&mm->mmap_sem);
1317mpol_out:
1318 mpol_put(new);
1319 return err;
1320}
1321
1322/*
1323 * User space interface with variable sized bitmaps for nodelists.
1324 */
1325
1326/* Copy a node mask from user space. */
1327static int get_nodes(nodemask_t *nodes, const unsigned long __user *nmask,
1328 unsigned long maxnode)
1329{
1330 unsigned long k;
1331 unsigned long t;
1332 unsigned long nlongs;
1333 unsigned long endmask;
1334
1335 --maxnode;
1336 nodes_clear(*nodes);
1337 if (maxnode == 0 || !nmask)
1338 return 0;
1339 if (maxnode > PAGE_SIZE*BITS_PER_BYTE)
1340 return -EINVAL;
1341
1342 nlongs = BITS_TO_LONGS(maxnode);
1343 if ((maxnode % BITS_PER_LONG) == 0)
1344 endmask = ~0UL;
1345 else
1346 endmask = (1UL << (maxnode % BITS_PER_LONG)) - 1;
1347
1348 /*
1349 * When the user specified more nodes than supported just check
1350 * if the non supported part is all zero.
1351 *
1352 * If maxnode have more longs than MAX_NUMNODES, check
1353 * the bits in that area first. And then go through to
1354 * check the rest bits which equal or bigger than MAX_NUMNODES.
1355 * Otherwise, just check bits [MAX_NUMNODES, maxnode).
1356 */
1357 if (nlongs > BITS_TO_LONGS(MAX_NUMNODES)) {
1358 for (k = BITS_TO_LONGS(MAX_NUMNODES); k < nlongs; k++) {
1359 if (get_user(t, nmask + k))
1360 return -EFAULT;
1361 if (k == nlongs - 1) {
1362 if (t & endmask)
1363 return -EINVAL;
1364 } else if (t)
1365 return -EINVAL;
1366 }
1367 nlongs = BITS_TO_LONGS(MAX_NUMNODES);
1368 endmask = ~0UL;
1369 }
1370
1371 if (maxnode > MAX_NUMNODES && MAX_NUMNODES % BITS_PER_LONG != 0) {
1372 unsigned long valid_mask = endmask;
1373
1374 valid_mask &= ~((1UL << (MAX_NUMNODES % BITS_PER_LONG)) - 1);
1375 if (get_user(t, nmask + nlongs - 1))
1376 return -EFAULT;
1377 if (t & valid_mask)
1378 return -EINVAL;
1379 }
1380
1381 if (copy_from_user(nodes_addr(*nodes), nmask, nlongs*sizeof(unsigned long)))
1382 return -EFAULT;
1383 nodes_addr(*nodes)[nlongs-1] &= endmask;
1384 return 0;
1385}
1386
1387/* Copy a kernel node mask to user space */
1388static int copy_nodes_to_user(unsigned long __user *mask, unsigned long maxnode,
1389 nodemask_t *nodes)
1390{
1391 unsigned long copy = ALIGN(maxnode-1, 64) / 8;
1392 unsigned int nbytes = BITS_TO_LONGS(nr_node_ids) * sizeof(long);
1393
1394 if (copy > nbytes) {
1395 if (copy > PAGE_SIZE)
1396 return -EINVAL;
1397 if (clear_user((char __user *)mask + nbytes, copy - nbytes))
1398 return -EFAULT;
1399 copy = nbytes;
1400 }
1401 return copy_to_user(mask, nodes_addr(*nodes), copy) ? -EFAULT : 0;
1402}
1403
1404static long kernel_mbind(unsigned long start, unsigned long len,
1405 unsigned long mode, const unsigned long __user *nmask,
1406 unsigned long maxnode, unsigned int flags)
1407{
1408 nodemask_t nodes;
1409 int err;
1410 unsigned short mode_flags;
1411
1412 start = untagged_addr(start);
1413 mode_flags = mode & MPOL_MODE_FLAGS;
1414 mode &= ~MPOL_MODE_FLAGS;
1415 if (mode >= MPOL_MAX)
1416 return -EINVAL;
1417 if ((mode_flags & MPOL_F_STATIC_NODES) &&
1418 (mode_flags & MPOL_F_RELATIVE_NODES))
1419 return -EINVAL;
1420 err = get_nodes(&nodes, nmask, maxnode);
1421 if (err)
1422 return err;
1423 return do_mbind(start, len, mode, mode_flags, &nodes, flags);
1424}
1425
1426SYSCALL_DEFINE6(mbind, unsigned long, start, unsigned long, len,
1427 unsigned long, mode, const unsigned long __user *, nmask,
1428 unsigned long, maxnode, unsigned int, flags)
1429{
1430 return kernel_mbind(start, len, mode, nmask, maxnode, flags);
1431}
1432
1433/* Set the process memory policy */
1434static long kernel_set_mempolicy(int mode, const unsigned long __user *nmask,
1435 unsigned long maxnode)
1436{
1437 int err;
1438 nodemask_t nodes;
1439 unsigned short flags;
1440
1441 flags = mode & MPOL_MODE_FLAGS;
1442 mode &= ~MPOL_MODE_FLAGS;
1443 if ((unsigned int)mode >= MPOL_MAX)
1444 return -EINVAL;
1445 if ((flags & MPOL_F_STATIC_NODES) && (flags & MPOL_F_RELATIVE_NODES))
1446 return -EINVAL;
1447 err = get_nodes(&nodes, nmask, maxnode);
1448 if (err)
1449 return err;
1450 return do_set_mempolicy(mode, flags, &nodes);
1451}
1452
1453SYSCALL_DEFINE3(set_mempolicy, int, mode, const unsigned long __user *, nmask,
1454 unsigned long, maxnode)
1455{
1456 return kernel_set_mempolicy(mode, nmask, maxnode);
1457}
1458
1459static int kernel_migrate_pages(pid_t pid, unsigned long maxnode,
1460 const unsigned long __user *old_nodes,
1461 const unsigned long __user *new_nodes)
1462{
1463 struct mm_struct *mm = NULL;
1464 struct task_struct *task;
1465 nodemask_t task_nodes;
1466 int err;
1467 nodemask_t *old;
1468 nodemask_t *new;
1469 NODEMASK_SCRATCH(scratch);
1470
1471 if (!scratch)
1472 return -ENOMEM;
1473
1474 old = &scratch->mask1;
1475 new = &scratch->mask2;
1476
1477 err = get_nodes(old, old_nodes, maxnode);
1478 if (err)
1479 goto out;
1480
1481 err = get_nodes(new, new_nodes, maxnode);
1482 if (err)
1483 goto out;
1484
1485 /* Find the mm_struct */
1486 rcu_read_lock();
1487 task = pid ? find_task_by_vpid(pid) : current;
1488 if (!task) {
1489 rcu_read_unlock();
1490 err = -ESRCH;
1491 goto out;
1492 }
1493 get_task_struct(task);
1494
1495 err = -EINVAL;
1496
1497 /*
1498 * Check if this process has the right to modify the specified process.
1499 * Use the regular "ptrace_may_access()" checks.
1500 */
1501 if (!ptrace_may_access(task, PTRACE_MODE_READ_REALCREDS)) {
1502 rcu_read_unlock();
1503 err = -EPERM;
1504 goto out_put;
1505 }
1506 rcu_read_unlock();
1507
1508 task_nodes = cpuset_mems_allowed(task);
1509 /* Is the user allowed to access the target nodes? */
1510 if (!nodes_subset(*new, task_nodes) && !capable(CAP_SYS_NICE)) {
1511 err = -EPERM;
1512 goto out_put;
1513 }
1514
1515 task_nodes = cpuset_mems_allowed(current);
1516 nodes_and(*new, *new, task_nodes);
1517 if (nodes_empty(*new))
1518 goto out_put;
1519
1520 err = security_task_movememory(task);
1521 if (err)
1522 goto out_put;
1523
1524 mm = get_task_mm(task);
1525 put_task_struct(task);
1526
1527 if (!mm) {
1528 err = -EINVAL;
1529 goto out;
1530 }
1531
1532 err = do_migrate_pages(mm, old, new,
1533 capable(CAP_SYS_NICE) ? MPOL_MF_MOVE_ALL : MPOL_MF_MOVE);
1534
1535 mmput(mm);
1536out:
1537 NODEMASK_SCRATCH_FREE(scratch);
1538
1539 return err;
1540
1541out_put:
1542 put_task_struct(task);
1543 goto out;
1544
1545}
1546
1547SYSCALL_DEFINE4(migrate_pages, pid_t, pid, unsigned long, maxnode,
1548 const unsigned long __user *, old_nodes,
1549 const unsigned long __user *, new_nodes)
1550{
1551 return kernel_migrate_pages(pid, maxnode, old_nodes, new_nodes);
1552}
1553
1554
1555/* Retrieve NUMA policy */
1556static int kernel_get_mempolicy(int __user *policy,
1557 unsigned long __user *nmask,
1558 unsigned long maxnode,
1559 unsigned long addr,
1560 unsigned long flags)
1561{
1562 int err;
1563 int uninitialized_var(pval);
1564 nodemask_t nodes;
1565
1566 addr = untagged_addr(addr);
1567
1568 if (nmask != NULL && maxnode < nr_node_ids)
1569 return -EINVAL;
1570
1571 err = do_get_mempolicy(&pval, &nodes, addr, flags);
1572
1573 if (err)
1574 return err;
1575
1576 if (policy && put_user(pval, policy))
1577 return -EFAULT;
1578
1579 if (nmask)
1580 err = copy_nodes_to_user(nmask, maxnode, &nodes);
1581
1582 return err;
1583}
1584
1585SYSCALL_DEFINE5(get_mempolicy, int __user *, policy,
1586 unsigned long __user *, nmask, unsigned long, maxnode,
1587 unsigned long, addr, unsigned long, flags)
1588{
1589 return kernel_get_mempolicy(policy, nmask, maxnode, addr, flags);
1590}
1591
1592#ifdef CONFIG_COMPAT
1593
1594COMPAT_SYSCALL_DEFINE5(get_mempolicy, int __user *, policy,
1595 compat_ulong_t __user *, nmask,
1596 compat_ulong_t, maxnode,
1597 compat_ulong_t, addr, compat_ulong_t, flags)
1598{
1599 long err;
1600 unsigned long __user *nm = NULL;
1601 unsigned long nr_bits, alloc_size;
1602 DECLARE_BITMAP(bm, MAX_NUMNODES);
1603
1604 nr_bits = min_t(unsigned long, maxnode-1, nr_node_ids);
1605 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1606
1607 if (nmask)
1608 nm = compat_alloc_user_space(alloc_size);
1609
1610 err = kernel_get_mempolicy(policy, nm, nr_bits+1, addr, flags);
1611
1612 if (!err && nmask) {
1613 unsigned long copy_size;
1614 copy_size = min_t(unsigned long, sizeof(bm), alloc_size);
1615 err = copy_from_user(bm, nm, copy_size);
1616 /* ensure entire bitmap is zeroed */
1617 err |= clear_user(nmask, ALIGN(maxnode-1, 8) / 8);
1618 err |= compat_put_bitmap(nmask, bm, nr_bits);
1619 }
1620
1621 return err;
1622}
1623
1624COMPAT_SYSCALL_DEFINE3(set_mempolicy, int, mode, compat_ulong_t __user *, nmask,
1625 compat_ulong_t, maxnode)
1626{
1627 unsigned long __user *nm = NULL;
1628 unsigned long nr_bits, alloc_size;
1629 DECLARE_BITMAP(bm, MAX_NUMNODES);
1630
1631 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1632 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1633
1634 if (nmask) {
1635 if (compat_get_bitmap(bm, nmask, nr_bits))
1636 return -EFAULT;
1637 nm = compat_alloc_user_space(alloc_size);
1638 if (copy_to_user(nm, bm, alloc_size))
1639 return -EFAULT;
1640 }
1641
1642 return kernel_set_mempolicy(mode, nm, nr_bits+1);
1643}
1644
1645COMPAT_SYSCALL_DEFINE6(mbind, compat_ulong_t, start, compat_ulong_t, len,
1646 compat_ulong_t, mode, compat_ulong_t __user *, nmask,
1647 compat_ulong_t, maxnode, compat_ulong_t, flags)
1648{
1649 unsigned long __user *nm = NULL;
1650 unsigned long nr_bits, alloc_size;
1651 nodemask_t bm;
1652
1653 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1654 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1655
1656 if (nmask) {
1657 if (compat_get_bitmap(nodes_addr(bm), nmask, nr_bits))
1658 return -EFAULT;
1659 nm = compat_alloc_user_space(alloc_size);
1660 if (copy_to_user(nm, nodes_addr(bm), alloc_size))
1661 return -EFAULT;
1662 }
1663
1664 return kernel_mbind(start, len, mode, nm, nr_bits+1, flags);
1665}
1666
1667COMPAT_SYSCALL_DEFINE4(migrate_pages, compat_pid_t, pid,
1668 compat_ulong_t, maxnode,
1669 const compat_ulong_t __user *, old_nodes,
1670 const compat_ulong_t __user *, new_nodes)
1671{
1672 unsigned long __user *old = NULL;
1673 unsigned long __user *new = NULL;
1674 nodemask_t tmp_mask;
1675 unsigned long nr_bits;
1676 unsigned long size;
1677
1678 nr_bits = min_t(unsigned long, maxnode - 1, MAX_NUMNODES);
1679 size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1680 if (old_nodes) {
1681 if (compat_get_bitmap(nodes_addr(tmp_mask), old_nodes, nr_bits))
1682 return -EFAULT;
1683 old = compat_alloc_user_space(new_nodes ? size * 2 : size);
1684 if (new_nodes)
1685 new = old + size / sizeof(unsigned long);
1686 if (copy_to_user(old, nodes_addr(tmp_mask), size))
1687 return -EFAULT;
1688 }
1689 if (new_nodes) {
1690 if (compat_get_bitmap(nodes_addr(tmp_mask), new_nodes, nr_bits))
1691 return -EFAULT;
1692 if (new == NULL)
1693 new = compat_alloc_user_space(size);
1694 if (copy_to_user(new, nodes_addr(tmp_mask), size))
1695 return -EFAULT;
1696 }
1697 return kernel_migrate_pages(pid, nr_bits + 1, old, new);
1698}
1699
1700#endif /* CONFIG_COMPAT */
1701
1702struct mempolicy *__get_vma_policy(struct vm_area_struct *vma,
1703 unsigned long addr)
1704{
1705 struct mempolicy *pol = NULL;
1706
1707 if (vma) {
1708 if (vma->vm_ops && vma->vm_ops->get_policy) {
1709 pol = vma->vm_ops->get_policy(vma, addr);
1710 } else if (vma->vm_policy) {
1711 pol = vma->vm_policy;
1712
1713 /*
1714 * shmem_alloc_page() passes MPOL_F_SHARED policy with
1715 * a pseudo vma whose vma->vm_ops=NULL. Take a reference
1716 * count on these policies which will be dropped by
1717 * mpol_cond_put() later
1718 */
1719 if (mpol_needs_cond_ref(pol))
1720 mpol_get(pol);
1721 }
1722 }
1723
1724 return pol;
1725}
1726
1727/*
1728 * get_vma_policy(@vma, @addr)
1729 * @vma: virtual memory area whose policy is sought
1730 * @addr: address in @vma for shared policy lookup
1731 *
1732 * Returns effective policy for a VMA at specified address.
1733 * Falls back to current->mempolicy or system default policy, as necessary.
1734 * Shared policies [those marked as MPOL_F_SHARED] require an extra reference
1735 * count--added by the get_policy() vm_op, as appropriate--to protect against
1736 * freeing by another task. It is the caller's responsibility to free the
1737 * extra reference for shared policies.
1738 */
1739static struct mempolicy *get_vma_policy(struct vm_area_struct *vma,
1740 unsigned long addr)
1741{
1742 struct mempolicy *pol = __get_vma_policy(vma, addr);
1743
1744 if (!pol)
1745 pol = get_task_policy(current);
1746
1747 return pol;
1748}
1749
1750bool vma_policy_mof(struct vm_area_struct *vma)
1751{
1752 struct mempolicy *pol;
1753
1754 if (vma->vm_ops && vma->vm_ops->get_policy) {
1755 bool ret = false;
1756
1757 pol = vma->vm_ops->get_policy(vma, vma->vm_start);
1758 if (pol && (pol->flags & MPOL_F_MOF))
1759 ret = true;
1760 mpol_cond_put(pol);
1761
1762 return ret;
1763 }
1764
1765 pol = vma->vm_policy;
1766 if (!pol)
1767 pol = get_task_policy(current);
1768
1769 return pol->flags & MPOL_F_MOF;
1770}
1771
1772static int apply_policy_zone(struct mempolicy *policy, enum zone_type zone)
1773{
1774 enum zone_type dynamic_policy_zone = policy_zone;
1775
1776 BUG_ON(dynamic_policy_zone == ZONE_MOVABLE);
1777
1778 /*
1779 * if policy->v.nodes has movable memory only,
1780 * we apply policy when gfp_zone(gfp) = ZONE_MOVABLE only.
1781 *
1782 * policy->v.nodes is intersect with node_states[N_MEMORY].
1783 * so if the following test faile, it implies
1784 * policy->v.nodes has movable memory only.
1785 */
1786 if (!nodes_intersects(policy->v.nodes, node_states[N_HIGH_MEMORY]))
1787 dynamic_policy_zone = ZONE_MOVABLE;
1788
1789 return zone >= dynamic_policy_zone;
1790}
1791
1792/*
1793 * Return a nodemask representing a mempolicy for filtering nodes for
1794 * page allocation
1795 */
1796static nodemask_t *policy_nodemask(gfp_t gfp, struct mempolicy *policy)
1797{
1798 /* Lower zones don't get a nodemask applied for MPOL_BIND */
1799 if (unlikely(policy->mode == MPOL_BIND) &&
1800 apply_policy_zone(policy, gfp_zone(gfp)) &&
1801 cpuset_nodemask_valid_mems_allowed(&policy->v.nodes))
1802 return &policy->v.nodes;
1803
1804 return NULL;
1805}
1806
1807/* Return the node id preferred by the given mempolicy, or the given id */
1808static int policy_node(gfp_t gfp, struct mempolicy *policy,
1809 int nd)
1810{
1811 if (policy->mode == MPOL_PREFERRED && !(policy->flags & MPOL_F_LOCAL))
1812 nd = policy->v.preferred_node;
1813 else {
1814 /*
1815 * __GFP_THISNODE shouldn't even be used with the bind policy
1816 * because we might easily break the expectation to stay on the
1817 * requested node and not break the policy.
1818 */
1819 WARN_ON_ONCE(policy->mode == MPOL_BIND && (gfp & __GFP_THISNODE));
1820 }
1821
1822 return nd;
1823}
1824
1825/* Do dynamic interleaving for a process */
1826static unsigned interleave_nodes(struct mempolicy *policy)
1827{
1828 unsigned next;
1829 struct task_struct *me = current;
1830
1831 next = next_node_in(me->il_prev, policy->v.nodes);
1832 if (next < MAX_NUMNODES)
1833 me->il_prev = next;
1834 return next;
1835}
1836
1837/*
1838 * Depending on the memory policy provide a node from which to allocate the
1839 * next slab entry.
1840 */
1841unsigned int mempolicy_slab_node(void)
1842{
1843 struct mempolicy *policy;
1844 int node = numa_mem_id();
1845
1846 if (in_interrupt())
1847 return node;
1848
1849 policy = current->mempolicy;
1850 if (!policy || policy->flags & MPOL_F_LOCAL)
1851 return node;
1852
1853 switch (policy->mode) {
1854 case MPOL_PREFERRED:
1855 /*
1856 * handled MPOL_F_LOCAL above
1857 */
1858 return policy->v.preferred_node;
1859
1860 case MPOL_INTERLEAVE:
1861 return interleave_nodes(policy);
1862
1863 case MPOL_BIND: {
1864 struct zoneref *z;
1865
1866 /*
1867 * Follow bind policy behavior and start allocation at the
1868 * first node.
1869 */
1870 struct zonelist *zonelist;
1871 enum zone_type highest_zoneidx = gfp_zone(GFP_KERNEL);
1872 zonelist = &NODE_DATA(node)->node_zonelists[ZONELIST_FALLBACK];
1873 z = first_zones_zonelist(zonelist, highest_zoneidx,
1874 &policy->v.nodes);
1875 return z->zone ? zone_to_nid(z->zone) : node;
1876 }
1877
1878 default:
1879 BUG();
1880 }
1881}
1882
1883/*
1884 * Do static interleaving for a VMA with known offset @n. Returns the n'th
1885 * node in pol->v.nodes (starting from n=0), wrapping around if n exceeds the
1886 * number of present nodes.
1887 */
1888static unsigned offset_il_node(struct mempolicy *pol, unsigned long n)
1889{
1890 unsigned nnodes = nodes_weight(pol->v.nodes);
1891 unsigned target;
1892 int i;
1893 int nid;
1894
1895 if (!nnodes)
1896 return numa_node_id();
1897 target = (unsigned int)n % nnodes;
1898 nid = first_node(pol->v.nodes);
1899 for (i = 0; i < target; i++)
1900 nid = next_node(nid, pol->v.nodes);
1901 return nid;
1902}
1903
1904/* Determine a node number for interleave */
1905static inline unsigned interleave_nid(struct mempolicy *pol,
1906 struct vm_area_struct *vma, unsigned long addr, int shift)
1907{
1908 if (vma) {
1909 unsigned long off;
1910
1911 /*
1912 * for small pages, there is no difference between
1913 * shift and PAGE_SHIFT, so the bit-shift is safe.
1914 * for huge pages, since vm_pgoff is in units of small
1915 * pages, we need to shift off the always 0 bits to get
1916 * a useful offset.
1917 */
1918 BUG_ON(shift < PAGE_SHIFT);
1919 off = vma->vm_pgoff >> (shift - PAGE_SHIFT);
1920 off += (addr - vma->vm_start) >> shift;
1921 return offset_il_node(pol, off);
1922 } else
1923 return interleave_nodes(pol);
1924}
1925
1926#ifdef CONFIG_HUGETLBFS
1927/*
1928 * huge_node(@vma, @addr, @gfp_flags, @mpol)
1929 * @vma: virtual memory area whose policy is sought
1930 * @addr: address in @vma for shared policy lookup and interleave policy
1931 * @gfp_flags: for requested zone
1932 * @mpol: pointer to mempolicy pointer for reference counted mempolicy
1933 * @nodemask: pointer to nodemask pointer for MPOL_BIND nodemask
1934 *
1935 * Returns a nid suitable for a huge page allocation and a pointer
1936 * to the struct mempolicy for conditional unref after allocation.
1937 * If the effective policy is 'BIND, returns a pointer to the mempolicy's
1938 * @nodemask for filtering the zonelist.
1939 *
1940 * Must be protected by read_mems_allowed_begin()
1941 */
1942int huge_node(struct vm_area_struct *vma, unsigned long addr, gfp_t gfp_flags,
1943 struct mempolicy **mpol, nodemask_t **nodemask)
1944{
1945 int nid;
1946
1947 *mpol = get_vma_policy(vma, addr);
1948 *nodemask = NULL; /* assume !MPOL_BIND */
1949
1950 if (unlikely((*mpol)->mode == MPOL_INTERLEAVE)) {
1951 nid = interleave_nid(*mpol, vma, addr,
1952 huge_page_shift(hstate_vma(vma)));
1953 } else {
1954 nid = policy_node(gfp_flags, *mpol, numa_node_id());
1955 if ((*mpol)->mode == MPOL_BIND)
1956 *nodemask = &(*mpol)->v.nodes;
1957 }
1958 return nid;
1959}
1960
1961/*
1962 * init_nodemask_of_mempolicy
1963 *
1964 * If the current task's mempolicy is "default" [NULL], return 'false'
1965 * to indicate default policy. Otherwise, extract the policy nodemask
1966 * for 'bind' or 'interleave' policy into the argument nodemask, or
1967 * initialize the argument nodemask to contain the single node for
1968 * 'preferred' or 'local' policy and return 'true' to indicate presence
1969 * of non-default mempolicy.
1970 *
1971 * We don't bother with reference counting the mempolicy [mpol_get/put]
1972 * because the current task is examining it's own mempolicy and a task's
1973 * mempolicy is only ever changed by the task itself.
1974 *
1975 * N.B., it is the caller's responsibility to free a returned nodemask.
1976 */
1977bool init_nodemask_of_mempolicy(nodemask_t *mask)
1978{
1979 struct mempolicy *mempolicy;
1980 int nid;
1981
1982 if (!(mask && current->mempolicy))
1983 return false;
1984
1985 task_lock(current);
1986 mempolicy = current->mempolicy;
1987 switch (mempolicy->mode) {
1988 case MPOL_PREFERRED:
1989 if (mempolicy->flags & MPOL_F_LOCAL)
1990 nid = numa_node_id();
1991 else
1992 nid = mempolicy->v.preferred_node;
1993 init_nodemask_of_node(mask, nid);
1994 break;
1995
1996 case MPOL_BIND:
1997 /* Fall through */
1998 case MPOL_INTERLEAVE:
1999 *mask = mempolicy->v.nodes;
2000 break;
2001
2002 default:
2003 BUG();
2004 }
2005 task_unlock(current);
2006
2007 return true;
2008}
2009#endif
2010
2011/*
2012 * mempolicy_nodemask_intersects
2013 *
2014 * If tsk's mempolicy is "default" [NULL], return 'true' to indicate default
2015 * policy. Otherwise, check for intersection between mask and the policy
2016 * nodemask for 'bind' or 'interleave' policy. For 'perferred' or 'local'
2017 * policy, always return true since it may allocate elsewhere on fallback.
2018 *
2019 * Takes task_lock(tsk) to prevent freeing of its mempolicy.
2020 */
2021bool mempolicy_nodemask_intersects(struct task_struct *tsk,
2022 const nodemask_t *mask)
2023{
2024 struct mempolicy *mempolicy;
2025 bool ret = true;
2026
2027 if (!mask)
2028 return ret;
2029 task_lock(tsk);
2030 mempolicy = tsk->mempolicy;
2031 if (!mempolicy)
2032 goto out;
2033
2034 switch (mempolicy->mode) {
2035 case MPOL_PREFERRED:
2036 /*
2037 * MPOL_PREFERRED and MPOL_F_LOCAL are only preferred nodes to
2038 * allocate from, they may fallback to other nodes when oom.
2039 * Thus, it's possible for tsk to have allocated memory from
2040 * nodes in mask.
2041 */
2042 break;
2043 case MPOL_BIND:
2044 case MPOL_INTERLEAVE:
2045 ret = nodes_intersects(mempolicy->v.nodes, *mask);
2046 break;
2047 default:
2048 BUG();
2049 }
2050out:
2051 task_unlock(tsk);
2052 return ret;
2053}
2054
2055/* Allocate a page in interleaved policy.
2056 Own path because it needs to do special accounting. */
2057static struct page *alloc_page_interleave(gfp_t gfp, unsigned order,
2058 unsigned nid)
2059{
2060 struct page *page;
2061
2062 page = __alloc_pages(gfp, order, nid);
2063 /* skip NUMA_INTERLEAVE_HIT counter update if numa stats is disabled */
2064 if (!static_branch_likely(&vm_numa_stat_key))
2065 return page;
2066 if (page && page_to_nid(page) == nid) {
2067 preempt_disable();
2068 __inc_numa_state(page_zone(page), NUMA_INTERLEAVE_HIT);
2069 preempt_enable();
2070 }
2071 return page;
2072}
2073
2074/**
2075 * alloc_pages_vma - Allocate a page for a VMA.
2076 *
2077 * @gfp:
2078 * %GFP_USER user allocation.
2079 * %GFP_KERNEL kernel allocations,
2080 * %GFP_HIGHMEM highmem/user allocations,
2081 * %GFP_FS allocation should not call back into a file system.
2082 * %GFP_ATOMIC don't sleep.
2083 *
2084 * @order:Order of the GFP allocation.
2085 * @vma: Pointer to VMA or NULL if not available.
2086 * @addr: Virtual Address of the allocation. Must be inside the VMA.
2087 * @node: Which node to prefer for allocation (modulo policy).
2088 * @hugepage: for hugepages try only the preferred node if possible
2089 *
2090 * This function allocates a page from the kernel page pool and applies
2091 * a NUMA policy associated with the VMA or the current process.
2092 * When VMA is not NULL caller must hold down_read on the mmap_sem of the
2093 * mm_struct of the VMA to prevent it from going away. Should be used for
2094 * all allocations for pages that will be mapped into user space. Returns
2095 * NULL when no page can be allocated.
2096 */
2097struct page *
2098alloc_pages_vma(gfp_t gfp, int order, struct vm_area_struct *vma,
2099 unsigned long addr, int node, bool hugepage)
2100{
2101 struct mempolicy *pol;
2102 struct page *page;
2103 int preferred_nid;
2104 nodemask_t *nmask;
2105
2106 pol = get_vma_policy(vma, addr);
2107
2108 if (pol->mode == MPOL_INTERLEAVE) {
2109 unsigned nid;
2110
2111 nid = interleave_nid(pol, vma, addr, PAGE_SHIFT + order);
2112 mpol_cond_put(pol);
2113 page = alloc_page_interleave(gfp, order, nid);
2114 goto out;
2115 }
2116
2117 if (unlikely(IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE) && hugepage)) {
2118 int hpage_node = node;
2119
2120 /*
2121 * For hugepage allocation and non-interleave policy which
2122 * allows the current node (or other explicitly preferred
2123 * node) we only try to allocate from the current/preferred
2124 * node and don't fall back to other nodes, as the cost of
2125 * remote accesses would likely offset THP benefits.
2126 *
2127 * If the policy is interleave, or does not allow the current
2128 * node in its nodemask, we allocate the standard way.
2129 */
2130 if (pol->mode == MPOL_PREFERRED && !(pol->flags & MPOL_F_LOCAL))
2131 hpage_node = pol->v.preferred_node;
2132
2133 nmask = policy_nodemask(gfp, pol);
2134 if (!nmask || node_isset(hpage_node, *nmask)) {
2135 mpol_cond_put(pol);
2136 page = __alloc_pages_node(hpage_node,
2137 gfp | __GFP_THISNODE, order);
2138
2139 /*
2140 * If hugepage allocations are configured to always
2141 * synchronous compact or the vma has been madvised
2142 * to prefer hugepage backing, retry allowing remote
2143 * memory as well.
2144 */
2145 if (!page && (gfp & __GFP_DIRECT_RECLAIM))
2146 page = __alloc_pages_node(hpage_node,
2147 gfp | __GFP_NORETRY, order);
2148
2149 goto out;
2150 }
2151 }
2152
2153 nmask = policy_nodemask(gfp, pol);
2154 preferred_nid = policy_node(gfp, pol, node);
2155 page = __alloc_pages_nodemask(gfp, order, preferred_nid, nmask);
2156 mpol_cond_put(pol);
2157out:
2158 return page;
2159}
2160EXPORT_SYMBOL(alloc_pages_vma);
2161
2162/**
2163 * alloc_pages_current - Allocate pages.
2164 *
2165 * @gfp:
2166 * %GFP_USER user allocation,
2167 * %GFP_KERNEL kernel allocation,
2168 * %GFP_HIGHMEM highmem allocation,
2169 * %GFP_FS don't call back into a file system.
2170 * %GFP_ATOMIC don't sleep.
2171 * @order: Power of two of allocation size in pages. 0 is a single page.
2172 *
2173 * Allocate a page from the kernel page pool. When not in
2174 * interrupt context and apply the current process NUMA policy.
2175 * Returns NULL when no page can be allocated.
2176 */
2177struct page *alloc_pages_current(gfp_t gfp, unsigned order)
2178{
2179 struct mempolicy *pol = &default_policy;
2180 struct page *page;
2181
2182 if (!in_interrupt() && !(gfp & __GFP_THISNODE))
2183 pol = get_task_policy(current);
2184
2185 /*
2186 * No reference counting needed for current->mempolicy
2187 * nor system default_policy
2188 */
2189 if (pol->mode == MPOL_INTERLEAVE)
2190 page = alloc_page_interleave(gfp, order, interleave_nodes(pol));
2191 else
2192 page = __alloc_pages_nodemask(gfp, order,
2193 policy_node(gfp, pol, numa_node_id()),
2194 policy_nodemask(gfp, pol));
2195
2196 return page;
2197}
2198EXPORT_SYMBOL(alloc_pages_current);
2199
2200int vma_dup_policy(struct vm_area_struct *src, struct vm_area_struct *dst)
2201{
2202 struct mempolicy *pol = mpol_dup(vma_policy(src));
2203
2204 if (IS_ERR(pol))
2205 return PTR_ERR(pol);
2206 dst->vm_policy = pol;
2207 return 0;
2208}
2209
2210/*
2211 * If mpol_dup() sees current->cpuset == cpuset_being_rebound, then it
2212 * rebinds the mempolicy its copying by calling mpol_rebind_policy()
2213 * with the mems_allowed returned by cpuset_mems_allowed(). This
2214 * keeps mempolicies cpuset relative after its cpuset moves. See
2215 * further kernel/cpuset.c update_nodemask().
2216 *
2217 * current's mempolicy may be rebinded by the other task(the task that changes
2218 * cpuset's mems), so we needn't do rebind work for current task.
2219 */
2220
2221/* Slow path of a mempolicy duplicate */
2222struct mempolicy *__mpol_dup(struct mempolicy *old)
2223{
2224 struct mempolicy *new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
2225
2226 if (!new)
2227 return ERR_PTR(-ENOMEM);
2228
2229 /* task's mempolicy is protected by alloc_lock */
2230 if (old == current->mempolicy) {
2231 task_lock(current);
2232 *new = *old;
2233 task_unlock(current);
2234 } else
2235 *new = *old;
2236
2237 if (current_cpuset_is_being_rebound()) {
2238 nodemask_t mems = cpuset_mems_allowed(current);
2239 mpol_rebind_policy(new, &mems);
2240 }
2241 atomic_set(&new->refcnt, 1);
2242 return new;
2243}
2244
2245/* Slow path of a mempolicy comparison */
2246bool __mpol_equal(struct mempolicy *a, struct mempolicy *b)
2247{
2248 if (!a || !b)
2249 return false;
2250 if (a->mode != b->mode)
2251 return false;
2252 if (a->flags != b->flags)
2253 return false;
2254 if (mpol_store_user_nodemask(a))
2255 if (!nodes_equal(a->w.user_nodemask, b->w.user_nodemask))
2256 return false;
2257
2258 switch (a->mode) {
2259 case MPOL_BIND:
2260 /* Fall through */
2261 case MPOL_INTERLEAVE:
2262 return !!nodes_equal(a->v.nodes, b->v.nodes);
2263 case MPOL_PREFERRED:
2264 /* a's ->flags is the same as b's */
2265 if (a->flags & MPOL_F_LOCAL)
2266 return true;
2267 return a->v.preferred_node == b->v.preferred_node;
2268 default:
2269 BUG();
2270 return false;
2271 }
2272}
2273
2274/*
2275 * Shared memory backing store policy support.
2276 *
2277 * Remember policies even when nobody has shared memory mapped.
2278 * The policies are kept in Red-Black tree linked from the inode.
2279 * They are protected by the sp->lock rwlock, which should be held
2280 * for any accesses to the tree.
2281 */
2282
2283/*
2284 * lookup first element intersecting start-end. Caller holds sp->lock for
2285 * reading or for writing
2286 */
2287static struct sp_node *
2288sp_lookup(struct shared_policy *sp, unsigned long start, unsigned long end)
2289{
2290 struct rb_node *n = sp->root.rb_node;
2291
2292 while (n) {
2293 struct sp_node *p = rb_entry(n, struct sp_node, nd);
2294
2295 if (start >= p->end)
2296 n = n->rb_right;
2297 else if (end <= p->start)
2298 n = n->rb_left;
2299 else
2300 break;
2301 }
2302 if (!n)
2303 return NULL;
2304 for (;;) {
2305 struct sp_node *w = NULL;
2306 struct rb_node *prev = rb_prev(n);
2307 if (!prev)
2308 break;
2309 w = rb_entry(prev, struct sp_node, nd);
2310 if (w->end <= start)
2311 break;
2312 n = prev;
2313 }
2314 return rb_entry(n, struct sp_node, nd);
2315}
2316
2317/*
2318 * Insert a new shared policy into the list. Caller holds sp->lock for
2319 * writing.
2320 */
2321static void sp_insert(struct shared_policy *sp, struct sp_node *new)
2322{
2323 struct rb_node **p = &sp->root.rb_node;
2324 struct rb_node *parent = NULL;
2325 struct sp_node *nd;
2326
2327 while (*p) {
2328 parent = *p;
2329 nd = rb_entry(parent, struct sp_node, nd);
2330 if (new->start < nd->start)
2331 p = &(*p)->rb_left;
2332 else if (new->end > nd->end)
2333 p = &(*p)->rb_right;
2334 else
2335 BUG();
2336 }
2337 rb_link_node(&new->nd, parent, p);
2338 rb_insert_color(&new->nd, &sp->root);
2339 pr_debug("inserting %lx-%lx: %d\n", new->start, new->end,
2340 new->policy ? new->policy->mode : 0);
2341}
2342
2343/* Find shared policy intersecting idx */
2344struct mempolicy *
2345mpol_shared_policy_lookup(struct shared_policy *sp, unsigned long idx)
2346{
2347 struct mempolicy *pol = NULL;
2348 struct sp_node *sn;
2349
2350 if (!sp->root.rb_node)
2351 return NULL;
2352 read_lock(&sp->lock);
2353 sn = sp_lookup(sp, idx, idx+1);
2354 if (sn) {
2355 mpol_get(sn->policy);
2356 pol = sn->policy;
2357 }
2358 read_unlock(&sp->lock);
2359 return pol;
2360}
2361
2362static void sp_free(struct sp_node *n)
2363{
2364 mpol_put(n->policy);
2365 kmem_cache_free(sn_cache, n);
2366}
2367
2368/**
2369 * mpol_misplaced - check whether current page node is valid in policy
2370 *
2371 * @page: page to be checked
2372 * @vma: vm area where page mapped
2373 * @addr: virtual address where page mapped
2374 *
2375 * Lookup current policy node id for vma,addr and "compare to" page's
2376 * node id.
2377 *
2378 * Returns:
2379 * -1 - not misplaced, page is in the right node
2380 * node - node id where the page should be
2381 *
2382 * Policy determination "mimics" alloc_page_vma().
2383 * Called from fault path where we know the vma and faulting address.
2384 */
2385int mpol_misplaced(struct page *page, struct vm_area_struct *vma, unsigned long addr)
2386{
2387 struct mempolicy *pol;
2388 struct zoneref *z;
2389 int curnid = page_to_nid(page);
2390 unsigned long pgoff;
2391 int thiscpu = raw_smp_processor_id();
2392 int thisnid = cpu_to_node(thiscpu);
2393 int polnid = NUMA_NO_NODE;
2394 int ret = -1;
2395
2396 pol = get_vma_policy(vma, addr);
2397 if (!(pol->flags & MPOL_F_MOF))
2398 goto out;
2399
2400 switch (pol->mode) {
2401 case MPOL_INTERLEAVE:
2402 pgoff = vma->vm_pgoff;
2403 pgoff += (addr - vma->vm_start) >> PAGE_SHIFT;
2404 polnid = offset_il_node(pol, pgoff);
2405 break;
2406
2407 case MPOL_PREFERRED:
2408 if (pol->flags & MPOL_F_LOCAL)
2409 polnid = numa_node_id();
2410 else
2411 polnid = pol->v.preferred_node;
2412 break;
2413
2414 case MPOL_BIND:
2415
2416 /*
2417 * allows binding to multiple nodes.
2418 * use current page if in policy nodemask,
2419 * else select nearest allowed node, if any.
2420 * If no allowed nodes, use current [!misplaced].
2421 */
2422 if (node_isset(curnid, pol->v.nodes))
2423 goto out;
2424 z = first_zones_zonelist(
2425 node_zonelist(numa_node_id(), GFP_HIGHUSER),
2426 gfp_zone(GFP_HIGHUSER),
2427 &pol->v.nodes);
2428 polnid = zone_to_nid(z->zone);
2429 break;
2430
2431 default:
2432 BUG();
2433 }
2434
2435 /* Migrate the page towards the node whose CPU is referencing it */
2436 if (pol->flags & MPOL_F_MORON) {
2437 polnid = thisnid;
2438
2439 if (!should_numa_migrate_memory(current, page, curnid, thiscpu))
2440 goto out;
2441 }
2442
2443 if (curnid != polnid)
2444 ret = polnid;
2445out:
2446 mpol_cond_put(pol);
2447
2448 return ret;
2449}
2450
2451/*
2452 * Drop the (possibly final) reference to task->mempolicy. It needs to be
2453 * dropped after task->mempolicy is set to NULL so that any allocation done as
2454 * part of its kmem_cache_free(), such as by KASAN, doesn't reference a freed
2455 * policy.
2456 */
2457void mpol_put_task_policy(struct task_struct *task)
2458{
2459 struct mempolicy *pol;
2460
2461 task_lock(task);
2462 pol = task->mempolicy;
2463 task->mempolicy = NULL;
2464 task_unlock(task);
2465 mpol_put(pol);
2466}
2467
2468static void sp_delete(struct shared_policy *sp, struct sp_node *n)
2469{
2470 pr_debug("deleting %lx-l%lx\n", n->start, n->end);
2471 rb_erase(&n->nd, &sp->root);
2472 sp_free(n);
2473}
2474
2475static void sp_node_init(struct sp_node *node, unsigned long start,
2476 unsigned long end, struct mempolicy *pol)
2477{
2478 node->start = start;
2479 node->end = end;
2480 node->policy = pol;
2481}
2482
2483static struct sp_node *sp_alloc(unsigned long start, unsigned long end,
2484 struct mempolicy *pol)
2485{
2486 struct sp_node *n;
2487 struct mempolicy *newpol;
2488
2489 n = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2490 if (!n)
2491 return NULL;
2492
2493 newpol = mpol_dup(pol);
2494 if (IS_ERR(newpol)) {
2495 kmem_cache_free(sn_cache, n);
2496 return NULL;
2497 }
2498 newpol->flags |= MPOL_F_SHARED;
2499 sp_node_init(n, start, end, newpol);
2500
2501 return n;
2502}
2503
2504/* Replace a policy range. */
2505static int shared_policy_replace(struct shared_policy *sp, unsigned long start,
2506 unsigned long end, struct sp_node *new)
2507{
2508 struct sp_node *n;
2509 struct sp_node *n_new = NULL;
2510 struct mempolicy *mpol_new = NULL;
2511 int ret = 0;
2512
2513restart:
2514 write_lock(&sp->lock);
2515 n = sp_lookup(sp, start, end);
2516 /* Take care of old policies in the same range. */
2517 while (n && n->start < end) {
2518 struct rb_node *next = rb_next(&n->nd);
2519 if (n->start >= start) {
2520 if (n->end <= end)
2521 sp_delete(sp, n);
2522 else
2523 n->start = end;
2524 } else {
2525 /* Old policy spanning whole new range. */
2526 if (n->end > end) {
2527 if (!n_new)
2528 goto alloc_new;
2529
2530 *mpol_new = *n->policy;
2531 atomic_set(&mpol_new->refcnt, 1);
2532 sp_node_init(n_new, end, n->end, mpol_new);
2533 n->end = start;
2534 sp_insert(sp, n_new);
2535 n_new = NULL;
2536 mpol_new = NULL;
2537 break;
2538 } else
2539 n->end = start;
2540 }
2541 if (!next)
2542 break;
2543 n = rb_entry(next, struct sp_node, nd);
2544 }
2545 if (new)
2546 sp_insert(sp, new);
2547 write_unlock(&sp->lock);
2548 ret = 0;
2549
2550err_out:
2551 if (mpol_new)
2552 mpol_put(mpol_new);
2553 if (n_new)
2554 kmem_cache_free(sn_cache, n_new);
2555
2556 return ret;
2557
2558alloc_new:
2559 write_unlock(&sp->lock);
2560 ret = -ENOMEM;
2561 n_new = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2562 if (!n_new)
2563 goto err_out;
2564 mpol_new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
2565 if (!mpol_new)
2566 goto err_out;
2567 goto restart;
2568}
2569
2570/**
2571 * mpol_shared_policy_init - initialize shared policy for inode
2572 * @sp: pointer to inode shared policy
2573 * @mpol: struct mempolicy to install
2574 *
2575 * Install non-NULL @mpol in inode's shared policy rb-tree.
2576 * On entry, the current task has a reference on a non-NULL @mpol.
2577 * This must be released on exit.
2578 * This is called at get_inode() calls and we can use GFP_KERNEL.
2579 */
2580void mpol_shared_policy_init(struct shared_policy *sp, struct mempolicy *mpol)
2581{
2582 int ret;
2583
2584 sp->root = RB_ROOT; /* empty tree == default mempolicy */
2585 rwlock_init(&sp->lock);
2586
2587 if (mpol) {
2588 struct vm_area_struct pvma;
2589 struct mempolicy *new;
2590 NODEMASK_SCRATCH(scratch);
2591
2592 if (!scratch)
2593 goto put_mpol;
2594 /* contextualize the tmpfs mount point mempolicy */
2595 new = mpol_new(mpol->mode, mpol->flags, &mpol->w.user_nodemask);
2596 if (IS_ERR(new))
2597 goto free_scratch; /* no valid nodemask intersection */
2598
2599 task_lock(current);
2600 ret = mpol_set_nodemask(new, &mpol->w.user_nodemask, scratch);
2601 task_unlock(current);
2602 if (ret)
2603 goto put_new;
2604
2605 /* Create pseudo-vma that contains just the policy */
2606 vma_init(&pvma, NULL);
2607 pvma.vm_end = TASK_SIZE; /* policy covers entire file */
2608 mpol_set_shared_policy(sp, &pvma, new); /* adds ref */
2609
2610put_new:
2611 mpol_put(new); /* drop initial ref */
2612free_scratch:
2613 NODEMASK_SCRATCH_FREE(scratch);
2614put_mpol:
2615 mpol_put(mpol); /* drop our incoming ref on sb mpol */
2616 }
2617}
2618
2619int mpol_set_shared_policy(struct shared_policy *info,
2620 struct vm_area_struct *vma, struct mempolicy *npol)
2621{
2622 int err;
2623 struct sp_node *new = NULL;
2624 unsigned long sz = vma_pages(vma);
2625
2626 pr_debug("set_shared_policy %lx sz %lu %d %d %lx\n",
2627 vma->vm_pgoff,
2628 sz, npol ? npol->mode : -1,
2629 npol ? npol->flags : -1,
2630 npol ? nodes_addr(npol->v.nodes)[0] : NUMA_NO_NODE);
2631
2632 if (npol) {
2633 new = sp_alloc(vma->vm_pgoff, vma->vm_pgoff + sz, npol);
2634 if (!new)
2635 return -ENOMEM;
2636 }
2637 err = shared_policy_replace(info, vma->vm_pgoff, vma->vm_pgoff+sz, new);
2638 if (err && new)
2639 sp_free(new);
2640 return err;
2641}
2642
2643/* Free a backing policy store on inode delete. */
2644void mpol_free_shared_policy(struct shared_policy *p)
2645{
2646 struct sp_node *n;
2647 struct rb_node *next;
2648
2649 if (!p->root.rb_node)
2650 return;
2651 write_lock(&p->lock);
2652 next = rb_first(&p->root);
2653 while (next) {
2654 n = rb_entry(next, struct sp_node, nd);
2655 next = rb_next(&n->nd);
2656 sp_delete(p, n);
2657 }
2658 write_unlock(&p->lock);
2659}
2660
2661#ifdef CONFIG_NUMA_BALANCING
2662static int __initdata numabalancing_override;
2663
2664static void __init check_numabalancing_enable(void)
2665{
2666 bool numabalancing_default = false;
2667
2668 if (IS_ENABLED(CONFIG_NUMA_BALANCING_DEFAULT_ENABLED))
2669 numabalancing_default = true;
2670
2671 /* Parsed by setup_numabalancing. override == 1 enables, -1 disables */
2672 if (numabalancing_override)
2673 set_numabalancing_state(numabalancing_override == 1);
2674
2675 if (num_online_nodes() > 1 && !numabalancing_override) {
2676 pr_info("%s automatic NUMA balancing. Configure with numa_balancing= or the kernel.numa_balancing sysctl\n",
2677 numabalancing_default ? "Enabling" : "Disabling");
2678 set_numabalancing_state(numabalancing_default);
2679 }
2680}
2681
2682static int __init setup_numabalancing(char *str)
2683{
2684 int ret = 0;
2685 if (!str)
2686 goto out;
2687
2688 if (!strcmp(str, "enable")) {
2689 numabalancing_override = 1;
2690 ret = 1;
2691 } else if (!strcmp(str, "disable")) {
2692 numabalancing_override = -1;
2693 ret = 1;
2694 }
2695out:
2696 if (!ret)
2697 pr_warn("Unable to parse numa_balancing=\n");
2698
2699 return ret;
2700}
2701__setup("numa_balancing=", setup_numabalancing);
2702#else
2703static inline void __init check_numabalancing_enable(void)
2704{
2705}
2706#endif /* CONFIG_NUMA_BALANCING */
2707
2708/* assumes fs == KERNEL_DS */
2709void __init numa_policy_init(void)
2710{
2711 nodemask_t interleave_nodes;
2712 unsigned long largest = 0;
2713 int nid, prefer = 0;
2714
2715 policy_cache = kmem_cache_create("numa_policy",
2716 sizeof(struct mempolicy),
2717 0, SLAB_PANIC, NULL);
2718
2719 sn_cache = kmem_cache_create("shared_policy_node",
2720 sizeof(struct sp_node),
2721 0, SLAB_PANIC, NULL);
2722
2723 for_each_node(nid) {
2724 preferred_node_policy[nid] = (struct mempolicy) {
2725 .refcnt = ATOMIC_INIT(1),
2726 .mode = MPOL_PREFERRED,
2727 .flags = MPOL_F_MOF | MPOL_F_MORON,
2728 .v = { .preferred_node = nid, },
2729 };
2730 }
2731
2732 /*
2733 * Set interleaving policy for system init. Interleaving is only
2734 * enabled across suitably sized nodes (default is >= 16MB), or
2735 * fall back to the largest node if they're all smaller.
2736 */
2737 nodes_clear(interleave_nodes);
2738 for_each_node_state(nid, N_MEMORY) {
2739 unsigned long total_pages = node_present_pages(nid);
2740
2741 /* Preserve the largest node */
2742 if (largest < total_pages) {
2743 largest = total_pages;
2744 prefer = nid;
2745 }
2746
2747 /* Interleave this node? */
2748 if ((total_pages << PAGE_SHIFT) >= (16 << 20))
2749 node_set(nid, interleave_nodes);
2750 }
2751
2752 /* All too small, use the largest */
2753 if (unlikely(nodes_empty(interleave_nodes)))
2754 node_set(prefer, interleave_nodes);
2755
2756 if (do_set_mempolicy(MPOL_INTERLEAVE, 0, &interleave_nodes))
2757 pr_err("%s: interleaving failed\n", __func__);
2758
2759 check_numabalancing_enable();
2760}
2761
2762/* Reset policy of current process to default */
2763void numa_default_policy(void)
2764{
2765 do_set_mempolicy(MPOL_DEFAULT, 0, NULL);
2766}
2767
2768/*
2769 * Parse and format mempolicy from/to strings
2770 */
2771
2772/*
2773 * "local" is implemented internally by MPOL_PREFERRED with MPOL_F_LOCAL flag.
2774 */
2775static const char * const policy_modes[] =
2776{
2777 [MPOL_DEFAULT] = "default",
2778 [MPOL_PREFERRED] = "prefer",
2779 [MPOL_BIND] = "bind",
2780 [MPOL_INTERLEAVE] = "interleave",
2781 [MPOL_LOCAL] = "local",
2782};
2783
2784
2785#ifdef CONFIG_TMPFS
2786/**
2787 * mpol_parse_str - parse string to mempolicy, for tmpfs mpol mount option.
2788 * @str: string containing mempolicy to parse
2789 * @mpol: pointer to struct mempolicy pointer, returned on success.
2790 *
2791 * Format of input:
2792 * <mode>[=<flags>][:<nodelist>]
2793 *
2794 * On success, returns 0, else 1
2795 */
2796int mpol_parse_str(char *str, struct mempolicy **mpol)
2797{
2798 struct mempolicy *new = NULL;
2799 unsigned short mode_flags;
2800 nodemask_t nodes;
2801 char *nodelist = strchr(str, ':');
2802 char *flags = strchr(str, '=');
2803 int err = 1, mode;
2804
2805 if (nodelist) {
2806 /* NUL-terminate mode or flags string */
2807 *nodelist++ = '\0';
2808 if (nodelist_parse(nodelist, nodes))
2809 goto out;
2810 if (!nodes_subset(nodes, node_states[N_MEMORY]))
2811 goto out;
2812 } else
2813 nodes_clear(nodes);
2814
2815 if (flags)
2816 *flags++ = '\0'; /* terminate mode string */
2817
2818 mode = match_string(policy_modes, MPOL_MAX, str);
2819 if (mode < 0)
2820 goto out;
2821
2822 switch (mode) {
2823 case MPOL_PREFERRED:
2824 /*
2825 * Insist on a nodelist of one node only
2826 */
2827 if (nodelist) {
2828 char *rest = nodelist;
2829 while (isdigit(*rest))
2830 rest++;
2831 if (*rest)
2832 goto out;
2833 }
2834 break;
2835 case MPOL_INTERLEAVE:
2836 /*
2837 * Default to online nodes with memory if no nodelist
2838 */
2839 if (!nodelist)
2840 nodes = node_states[N_MEMORY];
2841 break;
2842 case MPOL_LOCAL:
2843 /*
2844 * Don't allow a nodelist; mpol_new() checks flags
2845 */
2846 if (nodelist)
2847 goto out;
2848 mode = MPOL_PREFERRED;
2849 break;
2850 case MPOL_DEFAULT:
2851 /*
2852 * Insist on a empty nodelist
2853 */
2854 if (!nodelist)
2855 err = 0;
2856 goto out;
2857 case MPOL_BIND:
2858 /*
2859 * Insist on a nodelist
2860 */
2861 if (!nodelist)
2862 goto out;
2863 }
2864
2865 mode_flags = 0;
2866 if (flags) {
2867 /*
2868 * Currently, we only support two mutually exclusive
2869 * mode flags.
2870 */
2871 if (!strcmp(flags, "static"))
2872 mode_flags |= MPOL_F_STATIC_NODES;
2873 else if (!strcmp(flags, "relative"))
2874 mode_flags |= MPOL_F_RELATIVE_NODES;
2875 else
2876 goto out;
2877 }
2878
2879 new = mpol_new(mode, mode_flags, &nodes);
2880 if (IS_ERR(new))
2881 goto out;
2882
2883 /*
2884 * Save nodes for mpol_to_str() to show the tmpfs mount options
2885 * for /proc/mounts, /proc/pid/mounts and /proc/pid/mountinfo.
2886 */
2887 if (mode != MPOL_PREFERRED)
2888 new->v.nodes = nodes;
2889 else if (nodelist)
2890 new->v.preferred_node = first_node(nodes);
2891 else
2892 new->flags |= MPOL_F_LOCAL;
2893
2894 /*
2895 * Save nodes for contextualization: this will be used to "clone"
2896 * the mempolicy in a specific context [cpuset] at a later time.
2897 */
2898 new->w.user_nodemask = nodes;
2899
2900 err = 0;
2901
2902out:
2903 /* Restore string for error message */
2904 if (nodelist)
2905 *--nodelist = ':';
2906 if (flags)
2907 *--flags = '=';
2908 if (!err)
2909 *mpol = new;
2910 return err;
2911}
2912#endif /* CONFIG_TMPFS */
2913
2914/**
2915 * mpol_to_str - format a mempolicy structure for printing
2916 * @buffer: to contain formatted mempolicy string
2917 * @maxlen: length of @buffer
2918 * @pol: pointer to mempolicy to be formatted
2919 *
2920 * Convert @pol into a string. If @buffer is too short, truncate the string.
2921 * Recommend a @maxlen of at least 32 for the longest mode, "interleave", the
2922 * longest flag, "relative", and to display at least a few node ids.
2923 */
2924void mpol_to_str(char *buffer, int maxlen, struct mempolicy *pol)
2925{
2926 char *p = buffer;
2927 nodemask_t nodes = NODE_MASK_NONE;
2928 unsigned short mode = MPOL_DEFAULT;
2929 unsigned short flags = 0;
2930
2931 if (pol && pol != &default_policy && !(pol->flags & MPOL_F_MORON)) {
2932 mode = pol->mode;
2933 flags = pol->flags;
2934 }
2935
2936 switch (mode) {
2937 case MPOL_DEFAULT:
2938 break;
2939 case MPOL_PREFERRED:
2940 if (flags & MPOL_F_LOCAL)
2941 mode = MPOL_LOCAL;
2942 else
2943 node_set(pol->v.preferred_node, nodes);
2944 break;
2945 case MPOL_BIND:
2946 case MPOL_INTERLEAVE:
2947 nodes = pol->v.nodes;
2948 break;
2949 default:
2950 WARN_ON_ONCE(1);
2951 snprintf(p, maxlen, "unknown");
2952 return;
2953 }
2954
2955 p += snprintf(p, maxlen, "%s", policy_modes[mode]);
2956
2957 if (flags & MPOL_MODE_FLAGS) {
2958 p += snprintf(p, buffer + maxlen - p, "=");
2959
2960 /*
2961 * Currently, the only defined flags are mutually exclusive
2962 */
2963 if (flags & MPOL_F_STATIC_NODES)
2964 p += snprintf(p, buffer + maxlen - p, "static");
2965 else if (flags & MPOL_F_RELATIVE_NODES)
2966 p += snprintf(p, buffer + maxlen - p, "relative");
2967 }
2968
2969 if (!nodes_empty(nodes))
2970 p += scnprintf(p, buffer + maxlen - p, ":%*pbl",
2971 nodemask_pr_args(&nodes));
2972}
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * Simple NUMA memory policy for the Linux kernel.
4 *
5 * Copyright 2003,2004 Andi Kleen, SuSE Labs.
6 * (C) Copyright 2005 Christoph Lameter, Silicon Graphics, Inc.
7 *
8 * NUMA policy allows the user to give hints in which node(s) memory should
9 * be allocated.
10 *
11 * Support four policies per VMA and per process:
12 *
13 * The VMA policy has priority over the process policy for a page fault.
14 *
15 * interleave Allocate memory interleaved over a set of nodes,
16 * with normal fallback if it fails.
17 * For VMA based allocations this interleaves based on the
18 * offset into the backing object or offset into the mapping
19 * for anonymous memory. For process policy an process counter
20 * is used.
21 *
22 * 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 = pfn_folio(pmd_pfn(*pmd));
513 if (is_huge_zero_page(&folio->page)) {
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(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 * To check if the folio is shared, ideally we want to make sure
646 * every page is mapped to the same process. Doing that is very
647 * expensive, so check the estimated sharers of the folio instead.
648 */
649 if ((flags & MPOL_MF_MOVE_ALL) ||
650 (folio_estimated_sharers(folio) == 1 && !hugetlb_pmd_shared(pte)))
651 if (!isolate_hugetlb(folio, qp->pagelist))
652 qp->nr_failed++;
653unlock:
654 spin_unlock(ptl);
655 if (qp->nr_failed && strictly_unmovable(flags))
656 return -EIO;
657#endif
658 return 0;
659}
660
661#ifdef CONFIG_NUMA_BALANCING
662/*
663 * This is used to mark a range of virtual addresses to be inaccessible.
664 * These are later cleared by a NUMA hinting fault. Depending on these
665 * faults, pages may be migrated for better NUMA placement.
666 *
667 * This is assuming that NUMA faults are handled using PROT_NONE. If
668 * an architecture makes a different choice, it will need further
669 * changes to the core.
670 */
671unsigned long change_prot_numa(struct vm_area_struct *vma,
672 unsigned long addr, unsigned long end)
673{
674 struct mmu_gather tlb;
675 long nr_updated;
676
677 tlb_gather_mmu(&tlb, vma->vm_mm);
678
679 nr_updated = change_protection(&tlb, vma, addr, end, MM_CP_PROT_NUMA);
680 if (nr_updated > 0)
681 count_vm_numa_events(NUMA_PTE_UPDATES, nr_updated);
682
683 tlb_finish_mmu(&tlb);
684
685 return nr_updated;
686}
687#endif /* CONFIG_NUMA_BALANCING */
688
689static int queue_pages_test_walk(unsigned long start, unsigned long end,
690 struct mm_walk *walk)
691{
692 struct vm_area_struct *next, *vma = walk->vma;
693 struct queue_pages *qp = walk->private;
694 unsigned long flags = qp->flags;
695
696 /* range check first */
697 VM_BUG_ON_VMA(!range_in_vma(vma, start, end), vma);
698
699 if (!qp->first) {
700 qp->first = vma;
701 if (!(flags & MPOL_MF_DISCONTIG_OK) &&
702 (qp->start < vma->vm_start))
703 /* hole at head side of range */
704 return -EFAULT;
705 }
706 next = find_vma(vma->vm_mm, vma->vm_end);
707 if (!(flags & MPOL_MF_DISCONTIG_OK) &&
708 ((vma->vm_end < qp->end) &&
709 (!next || vma->vm_end < next->vm_start)))
710 /* hole at middle or tail of range */
711 return -EFAULT;
712
713 /*
714 * Need check MPOL_MF_STRICT to return -EIO if possible
715 * regardless of vma_migratable
716 */
717 if (!vma_migratable(vma) &&
718 !(flags & MPOL_MF_STRICT))
719 return 1;
720
721 /*
722 * Check page nodes, and queue pages to move, in the current vma.
723 * But if no moving, and no strict checking, the scan can be skipped.
724 */
725 if (flags & (MPOL_MF_STRICT | MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))
726 return 0;
727 return 1;
728}
729
730static const struct mm_walk_ops queue_pages_walk_ops = {
731 .hugetlb_entry = queue_folios_hugetlb,
732 .pmd_entry = queue_folios_pte_range,
733 .test_walk = queue_pages_test_walk,
734 .walk_lock = PGWALK_RDLOCK,
735};
736
737static const struct mm_walk_ops queue_pages_lock_vma_walk_ops = {
738 .hugetlb_entry = queue_folios_hugetlb,
739 .pmd_entry = queue_folios_pte_range,
740 .test_walk = queue_pages_test_walk,
741 .walk_lock = PGWALK_WRLOCK,
742};
743
744/*
745 * Walk through page tables and collect pages to be migrated.
746 *
747 * If pages found in a given range are not on the required set of @nodes,
748 * and migration is allowed, they are isolated and queued to @pagelist.
749 *
750 * queue_pages_range() may return:
751 * 0 - all pages already on the right node, or successfully queued for moving
752 * (or neither strict checking nor moving requested: only range checking).
753 * >0 - this number of misplaced folios could not be queued for moving
754 * (a hugetlbfs page or a transparent huge page being counted as 1).
755 * -EIO - a misplaced page found, when MPOL_MF_STRICT specified without MOVEs.
756 * -EFAULT - a hole in the memory range, when MPOL_MF_DISCONTIG_OK unspecified.
757 */
758static long
759queue_pages_range(struct mm_struct *mm, unsigned long start, unsigned long end,
760 nodemask_t *nodes, unsigned long flags,
761 struct list_head *pagelist)
762{
763 int err;
764 struct queue_pages qp = {
765 .pagelist = pagelist,
766 .flags = flags,
767 .nmask = nodes,
768 .start = start,
769 .end = end,
770 .first = NULL,
771 };
772 const struct mm_walk_ops *ops = (flags & MPOL_MF_WRLOCK) ?
773 &queue_pages_lock_vma_walk_ops : &queue_pages_walk_ops;
774
775 err = walk_page_range(mm, start, end, ops, &qp);
776
777 if (!qp.first)
778 /* whole range in hole */
779 err = -EFAULT;
780
781 return err ? : qp.nr_failed;
782}
783
784/*
785 * Apply policy to a single VMA
786 * This must be called with the mmap_lock held for writing.
787 */
788static int vma_replace_policy(struct vm_area_struct *vma,
789 struct mempolicy *pol)
790{
791 int err;
792 struct mempolicy *old;
793 struct mempolicy *new;
794
795 vma_assert_write_locked(vma);
796
797 new = mpol_dup(pol);
798 if (IS_ERR(new))
799 return PTR_ERR(new);
800
801 if (vma->vm_ops && vma->vm_ops->set_policy) {
802 err = vma->vm_ops->set_policy(vma, new);
803 if (err)
804 goto err_out;
805 }
806
807 old = vma->vm_policy;
808 vma->vm_policy = new; /* protected by mmap_lock */
809 mpol_put(old);
810
811 return 0;
812 err_out:
813 mpol_put(new);
814 return err;
815}
816
817/* Split or merge the VMA (if required) and apply the new policy */
818static int mbind_range(struct vma_iterator *vmi, struct vm_area_struct *vma,
819 struct vm_area_struct **prev, unsigned long start,
820 unsigned long end, struct mempolicy *new_pol)
821{
822 unsigned long vmstart, vmend;
823
824 vmend = min(end, vma->vm_end);
825 if (start > vma->vm_start) {
826 *prev = vma;
827 vmstart = start;
828 } else {
829 vmstart = vma->vm_start;
830 }
831
832 if (mpol_equal(vma->vm_policy, new_pol)) {
833 *prev = vma;
834 return 0;
835 }
836
837 vma = vma_modify_policy(vmi, *prev, vma, vmstart, vmend, new_pol);
838 if (IS_ERR(vma))
839 return PTR_ERR(vma);
840
841 *prev = vma;
842 return vma_replace_policy(vma, new_pol);
843}
844
845/* Set the process memory policy */
846static long do_set_mempolicy(unsigned short mode, unsigned short flags,
847 nodemask_t *nodes)
848{
849 struct mempolicy *new, *old;
850 NODEMASK_SCRATCH(scratch);
851 int ret;
852
853 if (!scratch)
854 return -ENOMEM;
855
856 new = mpol_new(mode, flags, nodes);
857 if (IS_ERR(new)) {
858 ret = PTR_ERR(new);
859 goto out;
860 }
861
862 task_lock(current);
863 ret = mpol_set_nodemask(new, nodes, scratch);
864 if (ret) {
865 task_unlock(current);
866 mpol_put(new);
867 goto out;
868 }
869
870 old = current->mempolicy;
871 current->mempolicy = new;
872 if (new && (new->mode == MPOL_INTERLEAVE ||
873 new->mode == MPOL_WEIGHTED_INTERLEAVE)) {
874 current->il_prev = MAX_NUMNODES-1;
875 current->il_weight = 0;
876 }
877 task_unlock(current);
878 mpol_put(old);
879 ret = 0;
880out:
881 NODEMASK_SCRATCH_FREE(scratch);
882 return ret;
883}
884
885/*
886 * Return nodemask for policy for get_mempolicy() query
887 *
888 * Called with task's alloc_lock held
889 */
890static void get_policy_nodemask(struct mempolicy *pol, nodemask_t *nodes)
891{
892 nodes_clear(*nodes);
893 if (pol == &default_policy)
894 return;
895
896 switch (pol->mode) {
897 case MPOL_BIND:
898 case MPOL_INTERLEAVE:
899 case MPOL_PREFERRED:
900 case MPOL_PREFERRED_MANY:
901 case MPOL_WEIGHTED_INTERLEAVE:
902 *nodes = pol->nodes;
903 break;
904 case MPOL_LOCAL:
905 /* return empty node mask for local allocation */
906 break;
907 default:
908 BUG();
909 }
910}
911
912static int lookup_node(struct mm_struct *mm, unsigned long addr)
913{
914 struct page *p = NULL;
915 int ret;
916
917 ret = get_user_pages_fast(addr & PAGE_MASK, 1, 0, &p);
918 if (ret > 0) {
919 ret = page_to_nid(p);
920 put_page(p);
921 }
922 return ret;
923}
924
925/* Retrieve NUMA policy */
926static long do_get_mempolicy(int *policy, nodemask_t *nmask,
927 unsigned long addr, unsigned long flags)
928{
929 int err;
930 struct mm_struct *mm = current->mm;
931 struct vm_area_struct *vma = NULL;
932 struct mempolicy *pol = current->mempolicy, *pol_refcount = NULL;
933
934 if (flags &
935 ~(unsigned long)(MPOL_F_NODE|MPOL_F_ADDR|MPOL_F_MEMS_ALLOWED))
936 return -EINVAL;
937
938 if (flags & MPOL_F_MEMS_ALLOWED) {
939 if (flags & (MPOL_F_NODE|MPOL_F_ADDR))
940 return -EINVAL;
941 *policy = 0; /* just so it's initialized */
942 task_lock(current);
943 *nmask = cpuset_current_mems_allowed;
944 task_unlock(current);
945 return 0;
946 }
947
948 if (flags & MPOL_F_ADDR) {
949 pgoff_t ilx; /* ignored here */
950 /*
951 * Do NOT fall back to task policy if the
952 * vma/shared policy at addr is NULL. We
953 * want to return MPOL_DEFAULT in this case.
954 */
955 mmap_read_lock(mm);
956 vma = vma_lookup(mm, addr);
957 if (!vma) {
958 mmap_read_unlock(mm);
959 return -EFAULT;
960 }
961 pol = __get_vma_policy(vma, addr, &ilx);
962 } else if (addr)
963 return -EINVAL;
964
965 if (!pol)
966 pol = &default_policy; /* indicates default behavior */
967
968 if (flags & MPOL_F_NODE) {
969 if (flags & MPOL_F_ADDR) {
970 /*
971 * Take a refcount on the mpol, because we are about to
972 * drop the mmap_lock, after which only "pol" remains
973 * valid, "vma" is stale.
974 */
975 pol_refcount = pol;
976 vma = NULL;
977 mpol_get(pol);
978 mmap_read_unlock(mm);
979 err = lookup_node(mm, addr);
980 if (err < 0)
981 goto out;
982 *policy = err;
983 } else if (pol == current->mempolicy &&
984 pol->mode == MPOL_INTERLEAVE) {
985 *policy = next_node_in(current->il_prev, pol->nodes);
986 } else if (pol == current->mempolicy &&
987 pol->mode == MPOL_WEIGHTED_INTERLEAVE) {
988 if (current->il_weight)
989 *policy = current->il_prev;
990 else
991 *policy = next_node_in(current->il_prev,
992 pol->nodes);
993 } else {
994 err = -EINVAL;
995 goto out;
996 }
997 } else {
998 *policy = pol == &default_policy ? MPOL_DEFAULT :
999 pol->mode;
1000 /*
1001 * Internal mempolicy flags must be masked off before exposing
1002 * the policy to userspace.
1003 */
1004 *policy |= (pol->flags & MPOL_MODE_FLAGS);
1005 }
1006
1007 err = 0;
1008 if (nmask) {
1009 if (mpol_store_user_nodemask(pol)) {
1010 *nmask = pol->w.user_nodemask;
1011 } else {
1012 task_lock(current);
1013 get_policy_nodemask(pol, nmask);
1014 task_unlock(current);
1015 }
1016 }
1017
1018 out:
1019 mpol_cond_put(pol);
1020 if (vma)
1021 mmap_read_unlock(mm);
1022 if (pol_refcount)
1023 mpol_put(pol_refcount);
1024 return err;
1025}
1026
1027#ifdef CONFIG_MIGRATION
1028static bool migrate_folio_add(struct folio *folio, struct list_head *foliolist,
1029 unsigned long flags)
1030{
1031 /*
1032 * Unless MPOL_MF_MOVE_ALL, we try to avoid migrating a shared folio.
1033 * Choosing not to migrate a shared folio is not counted as a failure.
1034 *
1035 * To check if the folio is shared, ideally we want to make sure
1036 * every page is mapped to the same process. Doing that is very
1037 * expensive, so check the estimated sharers of the folio instead.
1038 */
1039 if ((flags & MPOL_MF_MOVE_ALL) || folio_estimated_sharers(folio) == 1) {
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 };
1074
1075 nodes_clear(nmask);
1076 node_set(source, nmask);
1077
1078 VM_BUG_ON(!(flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)));
1079
1080 mmap_read_lock(mm);
1081 vma = find_vma(mm, 0);
1082
1083 /*
1084 * This does not migrate the range, but isolates all pages that
1085 * need migration. Between passing in the full user address
1086 * space range and MPOL_MF_DISCONTIG_OK, this call cannot fail,
1087 * but passes back the count of pages which could not be isolated.
1088 */
1089 nr_failed = queue_pages_range(mm, vma->vm_start, mm->task_size, &nmask,
1090 flags | MPOL_MF_DISCONTIG_OK, &pagelist);
1091 mmap_read_unlock(mm);
1092
1093 if (!list_empty(&pagelist)) {
1094 err = migrate_pages(&pagelist, alloc_migration_target, NULL,
1095 (unsigned long)&mtc, MIGRATE_SYNC, MR_SYSCALL, NULL);
1096 if (err)
1097 putback_movable_pages(&pagelist);
1098 }
1099
1100 if (err >= 0)
1101 err += nr_failed;
1102 return err;
1103}
1104
1105/*
1106 * Move pages between the two nodesets so as to preserve the physical
1107 * layout as much as possible.
1108 *
1109 * Returns the number of page that could not be moved.
1110 */
1111int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
1112 const nodemask_t *to, int flags)
1113{
1114 long nr_failed = 0;
1115 long err = 0;
1116 nodemask_t tmp;
1117
1118 lru_cache_disable();
1119
1120 /*
1121 * Find a 'source' bit set in 'tmp' whose corresponding 'dest'
1122 * bit in 'to' is not also set in 'tmp'. Clear the found 'source'
1123 * bit in 'tmp', and return that <source, dest> pair for migration.
1124 * The pair of nodemasks 'to' and 'from' define the map.
1125 *
1126 * If no pair of bits is found that way, fallback to picking some
1127 * pair of 'source' and 'dest' bits that are not the same. If the
1128 * 'source' and 'dest' bits are the same, this represents a node
1129 * that will be migrating to itself, so no pages need move.
1130 *
1131 * If no bits are left in 'tmp', or if all remaining bits left
1132 * in 'tmp' correspond to the same bit in 'to', return false
1133 * (nothing left to migrate).
1134 *
1135 * This lets us pick a pair of nodes to migrate between, such that
1136 * if possible the dest node is not already occupied by some other
1137 * source node, minimizing the risk of overloading the memory on a
1138 * node that would happen if we migrated incoming memory to a node
1139 * before migrating outgoing memory source that same node.
1140 *
1141 * A single scan of tmp is sufficient. As we go, we remember the
1142 * most recent <s, d> pair that moved (s != d). If we find a pair
1143 * that not only moved, but what's better, moved to an empty slot
1144 * (d is not set in tmp), then we break out then, with that pair.
1145 * Otherwise when we finish scanning from_tmp, we at least have the
1146 * most recent <s, d> pair that moved. If we get all the way through
1147 * the scan of tmp without finding any node that moved, much less
1148 * moved to an empty node, then there is nothing left worth migrating.
1149 */
1150
1151 tmp = *from;
1152 while (!nodes_empty(tmp)) {
1153 int s, d;
1154 int source = NUMA_NO_NODE;
1155 int dest = 0;
1156
1157 for_each_node_mask(s, tmp) {
1158
1159 /*
1160 * do_migrate_pages() tries to maintain the relative
1161 * node relationship of the pages established between
1162 * threads and memory areas.
1163 *
1164 * However if the number of source nodes is not equal to
1165 * the number of destination nodes we can not preserve
1166 * this node relative relationship. In that case, skip
1167 * copying memory from a node that is in the destination
1168 * mask.
1169 *
1170 * Example: [2,3,4] -> [3,4,5] moves everything.
1171 * [0-7] - > [3,4,5] moves only 0,1,2,6,7.
1172 */
1173
1174 if ((nodes_weight(*from) != nodes_weight(*to)) &&
1175 (node_isset(s, *to)))
1176 continue;
1177
1178 d = node_remap(s, *from, *to);
1179 if (s == d)
1180 continue;
1181
1182 source = s; /* Node moved. Memorize */
1183 dest = d;
1184
1185 /* dest not in remaining from nodes? */
1186 if (!node_isset(dest, tmp))
1187 break;
1188 }
1189 if (source == NUMA_NO_NODE)
1190 break;
1191
1192 node_clear(source, tmp);
1193 err = migrate_to_node(mm, source, dest, flags);
1194 if (err > 0)
1195 nr_failed += err;
1196 if (err < 0)
1197 break;
1198 }
1199
1200 lru_cache_enable();
1201 if (err < 0)
1202 return err;
1203 return (nr_failed < INT_MAX) ? nr_failed : INT_MAX;
1204}
1205
1206/*
1207 * Allocate a new folio for page migration, according to NUMA mempolicy.
1208 */
1209static struct folio *alloc_migration_target_by_mpol(struct folio *src,
1210 unsigned long private)
1211{
1212 struct migration_mpol *mmpol = (struct migration_mpol *)private;
1213 struct mempolicy *pol = mmpol->pol;
1214 pgoff_t ilx = mmpol->ilx;
1215 struct page *page;
1216 unsigned int order;
1217 int nid = numa_node_id();
1218 gfp_t gfp;
1219
1220 order = folio_order(src);
1221 ilx += src->index >> order;
1222
1223 if (folio_test_hugetlb(src)) {
1224 nodemask_t *nodemask;
1225 struct hstate *h;
1226
1227 h = folio_hstate(src);
1228 gfp = htlb_alloc_mask(h);
1229 nodemask = policy_nodemask(gfp, pol, ilx, &nid);
1230 return alloc_hugetlb_folio_nodemask(h, nid, nodemask, gfp);
1231 }
1232
1233 if (folio_test_large(src))
1234 gfp = GFP_TRANSHUGE;
1235 else
1236 gfp = GFP_HIGHUSER_MOVABLE | __GFP_RETRY_MAYFAIL | __GFP_COMP;
1237
1238 page = alloc_pages_mpol(gfp, order, pol, ilx, nid);
1239 return page_rmappable_folio(page);
1240}
1241#else
1242
1243static bool migrate_folio_add(struct folio *folio, struct list_head *foliolist,
1244 unsigned long flags)
1245{
1246 return false;
1247}
1248
1249int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
1250 const nodemask_t *to, int flags)
1251{
1252 return -ENOSYS;
1253}
1254
1255static struct folio *alloc_migration_target_by_mpol(struct folio *src,
1256 unsigned long private)
1257{
1258 return NULL;
1259}
1260#endif
1261
1262static long do_mbind(unsigned long start, unsigned long len,
1263 unsigned short mode, unsigned short mode_flags,
1264 nodemask_t *nmask, unsigned long flags)
1265{
1266 struct mm_struct *mm = current->mm;
1267 struct vm_area_struct *vma, *prev;
1268 struct vma_iterator vmi;
1269 struct migration_mpol mmpol;
1270 struct mempolicy *new;
1271 unsigned long end;
1272 long err;
1273 long nr_failed;
1274 LIST_HEAD(pagelist);
1275
1276 if (flags & ~(unsigned long)MPOL_MF_VALID)
1277 return -EINVAL;
1278 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1279 return -EPERM;
1280
1281 if (start & ~PAGE_MASK)
1282 return -EINVAL;
1283
1284 if (mode == MPOL_DEFAULT)
1285 flags &= ~MPOL_MF_STRICT;
1286
1287 len = PAGE_ALIGN(len);
1288 end = start + len;
1289
1290 if (end < start)
1291 return -EINVAL;
1292 if (end == start)
1293 return 0;
1294
1295 new = mpol_new(mode, mode_flags, nmask);
1296 if (IS_ERR(new))
1297 return PTR_ERR(new);
1298
1299 /*
1300 * If we are using the default policy then operation
1301 * on discontinuous address spaces is okay after all
1302 */
1303 if (!new)
1304 flags |= MPOL_MF_DISCONTIG_OK;
1305
1306 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))
1307 lru_cache_disable();
1308 {
1309 NODEMASK_SCRATCH(scratch);
1310 if (scratch) {
1311 mmap_write_lock(mm);
1312 err = mpol_set_nodemask(new, nmask, scratch);
1313 if (err)
1314 mmap_write_unlock(mm);
1315 } else
1316 err = -ENOMEM;
1317 NODEMASK_SCRATCH_FREE(scratch);
1318 }
1319 if (err)
1320 goto mpol_out;
1321
1322 /*
1323 * Lock the VMAs before scanning for pages to migrate,
1324 * to ensure we don't miss a concurrently inserted page.
1325 */
1326 nr_failed = queue_pages_range(mm, start, end, nmask,
1327 flags | MPOL_MF_INVERT | MPOL_MF_WRLOCK, &pagelist);
1328
1329 if (nr_failed < 0) {
1330 err = nr_failed;
1331 nr_failed = 0;
1332 } else {
1333 vma_iter_init(&vmi, mm, start);
1334 prev = vma_prev(&vmi);
1335 for_each_vma_range(vmi, vma, end) {
1336 err = mbind_range(&vmi, vma, &prev, start, end, new);
1337 if (err)
1338 break;
1339 }
1340 }
1341
1342 if (!err && !list_empty(&pagelist)) {
1343 /* Convert MPOL_DEFAULT's NULL to task or default policy */
1344 if (!new) {
1345 new = get_task_policy(current);
1346 mpol_get(new);
1347 }
1348 mmpol.pol = new;
1349 mmpol.ilx = 0;
1350
1351 /*
1352 * In the interleaved case, attempt to allocate on exactly the
1353 * targeted nodes, for the first VMA to be migrated; for later
1354 * VMAs, the nodes will still be interleaved from the targeted
1355 * nodemask, but one by one may be selected differently.
1356 */
1357 if (new->mode == MPOL_INTERLEAVE ||
1358 new->mode == MPOL_WEIGHTED_INTERLEAVE) {
1359 struct folio *folio;
1360 unsigned int order;
1361 unsigned long addr = -EFAULT;
1362
1363 list_for_each_entry(folio, &pagelist, lru) {
1364 if (!folio_test_ksm(folio))
1365 break;
1366 }
1367 if (!list_entry_is_head(folio, &pagelist, lru)) {
1368 vma_iter_init(&vmi, mm, start);
1369 for_each_vma_range(vmi, vma, end) {
1370 addr = page_address_in_vma(
1371 folio_page(folio, 0), vma);
1372 if (addr != -EFAULT)
1373 break;
1374 }
1375 }
1376 if (addr != -EFAULT) {
1377 order = folio_order(folio);
1378 /* We already know the pol, but not the ilx */
1379 mpol_cond_put(get_vma_policy(vma, addr, order,
1380 &mmpol.ilx));
1381 /* Set base from which to increment by index */
1382 mmpol.ilx -= folio->index >> order;
1383 }
1384 }
1385 }
1386
1387 mmap_write_unlock(mm);
1388
1389 if (!err && !list_empty(&pagelist)) {
1390 nr_failed |= migrate_pages(&pagelist,
1391 alloc_migration_target_by_mpol, NULL,
1392 (unsigned long)&mmpol, MIGRATE_SYNC,
1393 MR_MEMPOLICY_MBIND, NULL);
1394 }
1395
1396 if (nr_failed && (flags & MPOL_MF_STRICT))
1397 err = -EIO;
1398 if (!list_empty(&pagelist))
1399 putback_movable_pages(&pagelist);
1400mpol_out:
1401 mpol_put(new);
1402 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))
1403 lru_cache_enable();
1404 return err;
1405}
1406
1407/*
1408 * User space interface with variable sized bitmaps for nodelists.
1409 */
1410static int get_bitmap(unsigned long *mask, const unsigned long __user *nmask,
1411 unsigned long maxnode)
1412{
1413 unsigned long nlongs = BITS_TO_LONGS(maxnode);
1414 int ret;
1415
1416 if (in_compat_syscall())
1417 ret = compat_get_bitmap(mask,
1418 (const compat_ulong_t __user *)nmask,
1419 maxnode);
1420 else
1421 ret = copy_from_user(mask, nmask,
1422 nlongs * sizeof(unsigned long));
1423
1424 if (ret)
1425 return -EFAULT;
1426
1427 if (maxnode % BITS_PER_LONG)
1428 mask[nlongs - 1] &= (1UL << (maxnode % BITS_PER_LONG)) - 1;
1429
1430 return 0;
1431}
1432
1433/* Copy a node mask from user space. */
1434static int get_nodes(nodemask_t *nodes, const unsigned long __user *nmask,
1435 unsigned long maxnode)
1436{
1437 --maxnode;
1438 nodes_clear(*nodes);
1439 if (maxnode == 0 || !nmask)
1440 return 0;
1441 if (maxnode > PAGE_SIZE*BITS_PER_BYTE)
1442 return -EINVAL;
1443
1444 /*
1445 * When the user specified more nodes than supported just check
1446 * if the non supported part is all zero, one word at a time,
1447 * starting at the end.
1448 */
1449 while (maxnode > MAX_NUMNODES) {
1450 unsigned long bits = min_t(unsigned long, maxnode, BITS_PER_LONG);
1451 unsigned long t;
1452
1453 if (get_bitmap(&t, &nmask[(maxnode - 1) / BITS_PER_LONG], bits))
1454 return -EFAULT;
1455
1456 if (maxnode - bits >= MAX_NUMNODES) {
1457 maxnode -= bits;
1458 } else {
1459 maxnode = MAX_NUMNODES;
1460 t &= ~((1UL << (MAX_NUMNODES % BITS_PER_LONG)) - 1);
1461 }
1462 if (t)
1463 return -EINVAL;
1464 }
1465
1466 return get_bitmap(nodes_addr(*nodes), nmask, maxnode);
1467}
1468
1469/* Copy a kernel node mask to user space */
1470static int copy_nodes_to_user(unsigned long __user *mask, unsigned long maxnode,
1471 nodemask_t *nodes)
1472{
1473 unsigned long copy = ALIGN(maxnode-1, 64) / 8;
1474 unsigned int nbytes = BITS_TO_LONGS(nr_node_ids) * sizeof(long);
1475 bool compat = in_compat_syscall();
1476
1477 if (compat)
1478 nbytes = BITS_TO_COMPAT_LONGS(nr_node_ids) * sizeof(compat_long_t);
1479
1480 if (copy > nbytes) {
1481 if (copy > PAGE_SIZE)
1482 return -EINVAL;
1483 if (clear_user((char __user *)mask + nbytes, copy - nbytes))
1484 return -EFAULT;
1485 copy = nbytes;
1486 maxnode = nr_node_ids;
1487 }
1488
1489 if (compat)
1490 return compat_put_bitmap((compat_ulong_t __user *)mask,
1491 nodes_addr(*nodes), maxnode);
1492
1493 return copy_to_user(mask, nodes_addr(*nodes), copy) ? -EFAULT : 0;
1494}
1495
1496/* Basic parameter sanity check used by both mbind() and set_mempolicy() */
1497static inline int sanitize_mpol_flags(int *mode, unsigned short *flags)
1498{
1499 *flags = *mode & MPOL_MODE_FLAGS;
1500 *mode &= ~MPOL_MODE_FLAGS;
1501
1502 if ((unsigned int)(*mode) >= MPOL_MAX)
1503 return -EINVAL;
1504 if ((*flags & MPOL_F_STATIC_NODES) && (*flags & MPOL_F_RELATIVE_NODES))
1505 return -EINVAL;
1506 if (*flags & MPOL_F_NUMA_BALANCING) {
1507 if (*mode != MPOL_BIND)
1508 return -EINVAL;
1509 *flags |= (MPOL_F_MOF | MPOL_F_MORON);
1510 }
1511 return 0;
1512}
1513
1514static long kernel_mbind(unsigned long start, unsigned long len,
1515 unsigned long mode, const unsigned long __user *nmask,
1516 unsigned long maxnode, unsigned int flags)
1517{
1518 unsigned short mode_flags;
1519 nodemask_t nodes;
1520 int lmode = mode;
1521 int err;
1522
1523 start = untagged_addr(start);
1524 err = sanitize_mpol_flags(&lmode, &mode_flags);
1525 if (err)
1526 return err;
1527
1528 err = get_nodes(&nodes, nmask, maxnode);
1529 if (err)
1530 return err;
1531
1532 return do_mbind(start, len, lmode, mode_flags, &nodes, flags);
1533}
1534
1535SYSCALL_DEFINE4(set_mempolicy_home_node, unsigned long, start, unsigned long, len,
1536 unsigned long, home_node, unsigned long, flags)
1537{
1538 struct mm_struct *mm = current->mm;
1539 struct vm_area_struct *vma, *prev;
1540 struct mempolicy *new, *old;
1541 unsigned long end;
1542 int err = -ENOENT;
1543 VMA_ITERATOR(vmi, mm, start);
1544
1545 start = untagged_addr(start);
1546 if (start & ~PAGE_MASK)
1547 return -EINVAL;
1548 /*
1549 * flags is used for future extension if any.
1550 */
1551 if (flags != 0)
1552 return -EINVAL;
1553
1554 /*
1555 * Check home_node is online to avoid accessing uninitialized
1556 * NODE_DATA.
1557 */
1558 if (home_node >= MAX_NUMNODES || !node_online(home_node))
1559 return -EINVAL;
1560
1561 len = PAGE_ALIGN(len);
1562 end = start + len;
1563
1564 if (end < start)
1565 return -EINVAL;
1566 if (end == start)
1567 return 0;
1568 mmap_write_lock(mm);
1569 prev = vma_prev(&vmi);
1570 for_each_vma_range(vmi, vma, end) {
1571 /*
1572 * If any vma in the range got policy other than MPOL_BIND
1573 * or MPOL_PREFERRED_MANY we return error. We don't reset
1574 * the home node for vmas we already updated before.
1575 */
1576 old = vma_policy(vma);
1577 if (!old) {
1578 prev = vma;
1579 continue;
1580 }
1581 if (old->mode != MPOL_BIND && old->mode != MPOL_PREFERRED_MANY) {
1582 err = -EOPNOTSUPP;
1583 break;
1584 }
1585 new = mpol_dup(old);
1586 if (IS_ERR(new)) {
1587 err = PTR_ERR(new);
1588 break;
1589 }
1590
1591 vma_start_write(vma);
1592 new->home_node = home_node;
1593 err = mbind_range(&vmi, vma, &prev, start, end, new);
1594 mpol_put(new);
1595 if (err)
1596 break;
1597 }
1598 mmap_write_unlock(mm);
1599 return err;
1600}
1601
1602SYSCALL_DEFINE6(mbind, unsigned long, start, unsigned long, len,
1603 unsigned long, mode, const unsigned long __user *, nmask,
1604 unsigned long, maxnode, unsigned int, flags)
1605{
1606 return kernel_mbind(start, len, mode, nmask, maxnode, flags);
1607}
1608
1609/* Set the process memory policy */
1610static long kernel_set_mempolicy(int mode, const unsigned long __user *nmask,
1611 unsigned long maxnode)
1612{
1613 unsigned short mode_flags;
1614 nodemask_t nodes;
1615 int lmode = mode;
1616 int err;
1617
1618 err = sanitize_mpol_flags(&lmode, &mode_flags);
1619 if (err)
1620 return err;
1621
1622 err = get_nodes(&nodes, nmask, maxnode);
1623 if (err)
1624 return err;
1625
1626 return do_set_mempolicy(lmode, mode_flags, &nodes);
1627}
1628
1629SYSCALL_DEFINE3(set_mempolicy, int, mode, const unsigned long __user *, nmask,
1630 unsigned long, maxnode)
1631{
1632 return kernel_set_mempolicy(mode, nmask, maxnode);
1633}
1634
1635static int kernel_migrate_pages(pid_t pid, unsigned long maxnode,
1636 const unsigned long __user *old_nodes,
1637 const unsigned long __user *new_nodes)
1638{
1639 struct mm_struct *mm = NULL;
1640 struct task_struct *task;
1641 nodemask_t task_nodes;
1642 int err;
1643 nodemask_t *old;
1644 nodemask_t *new;
1645 NODEMASK_SCRATCH(scratch);
1646
1647 if (!scratch)
1648 return -ENOMEM;
1649
1650 old = &scratch->mask1;
1651 new = &scratch->mask2;
1652
1653 err = get_nodes(old, old_nodes, maxnode);
1654 if (err)
1655 goto out;
1656
1657 err = get_nodes(new, new_nodes, maxnode);
1658 if (err)
1659 goto out;
1660
1661 /* Find the mm_struct */
1662 rcu_read_lock();
1663 task = pid ? find_task_by_vpid(pid) : current;
1664 if (!task) {
1665 rcu_read_unlock();
1666 err = -ESRCH;
1667 goto out;
1668 }
1669 get_task_struct(task);
1670
1671 err = -EINVAL;
1672
1673 /*
1674 * Check if this process has the right to modify the specified process.
1675 * Use the regular "ptrace_may_access()" checks.
1676 */
1677 if (!ptrace_may_access(task, PTRACE_MODE_READ_REALCREDS)) {
1678 rcu_read_unlock();
1679 err = -EPERM;
1680 goto out_put;
1681 }
1682 rcu_read_unlock();
1683
1684 task_nodes = cpuset_mems_allowed(task);
1685 /* Is the user allowed to access the target nodes? */
1686 if (!nodes_subset(*new, task_nodes) && !capable(CAP_SYS_NICE)) {
1687 err = -EPERM;
1688 goto out_put;
1689 }
1690
1691 task_nodes = cpuset_mems_allowed(current);
1692 nodes_and(*new, *new, task_nodes);
1693 if (nodes_empty(*new))
1694 goto out_put;
1695
1696 err = security_task_movememory(task);
1697 if (err)
1698 goto out_put;
1699
1700 mm = get_task_mm(task);
1701 put_task_struct(task);
1702
1703 if (!mm) {
1704 err = -EINVAL;
1705 goto out;
1706 }
1707
1708 err = do_migrate_pages(mm, old, new,
1709 capable(CAP_SYS_NICE) ? MPOL_MF_MOVE_ALL : MPOL_MF_MOVE);
1710
1711 mmput(mm);
1712out:
1713 NODEMASK_SCRATCH_FREE(scratch);
1714
1715 return err;
1716
1717out_put:
1718 put_task_struct(task);
1719 goto out;
1720}
1721
1722SYSCALL_DEFINE4(migrate_pages, pid_t, pid, unsigned long, maxnode,
1723 const unsigned long __user *, old_nodes,
1724 const unsigned long __user *, new_nodes)
1725{
1726 return kernel_migrate_pages(pid, maxnode, old_nodes, new_nodes);
1727}
1728
1729/* Retrieve NUMA policy */
1730static int kernel_get_mempolicy(int __user *policy,
1731 unsigned long __user *nmask,
1732 unsigned long maxnode,
1733 unsigned long addr,
1734 unsigned long flags)
1735{
1736 int err;
1737 int pval;
1738 nodemask_t nodes;
1739
1740 if (nmask != NULL && maxnode < nr_node_ids)
1741 return -EINVAL;
1742
1743 addr = untagged_addr(addr);
1744
1745 err = do_get_mempolicy(&pval, &nodes, addr, flags);
1746
1747 if (err)
1748 return err;
1749
1750 if (policy && put_user(pval, policy))
1751 return -EFAULT;
1752
1753 if (nmask)
1754 err = copy_nodes_to_user(nmask, maxnode, &nodes);
1755
1756 return err;
1757}
1758
1759SYSCALL_DEFINE5(get_mempolicy, int __user *, policy,
1760 unsigned long __user *, nmask, unsigned long, maxnode,
1761 unsigned long, addr, unsigned long, flags)
1762{
1763 return kernel_get_mempolicy(policy, nmask, maxnode, addr, flags);
1764}
1765
1766bool vma_migratable(struct vm_area_struct *vma)
1767{
1768 if (vma->vm_flags & (VM_IO | VM_PFNMAP))
1769 return false;
1770
1771 /*
1772 * DAX device mappings require predictable access latency, so avoid
1773 * incurring periodic faults.
1774 */
1775 if (vma_is_dax(vma))
1776 return false;
1777
1778 if (is_vm_hugetlb_page(vma) &&
1779 !hugepage_migration_supported(hstate_vma(vma)))
1780 return false;
1781
1782 /*
1783 * Migration allocates pages in the highest zone. If we cannot
1784 * do so then migration (at least from node to node) is not
1785 * possible.
1786 */
1787 if (vma->vm_file &&
1788 gfp_zone(mapping_gfp_mask(vma->vm_file->f_mapping))
1789 < policy_zone)
1790 return false;
1791 return true;
1792}
1793
1794struct mempolicy *__get_vma_policy(struct vm_area_struct *vma,
1795 unsigned long addr, pgoff_t *ilx)
1796{
1797 *ilx = 0;
1798 return (vma->vm_ops && vma->vm_ops->get_policy) ?
1799 vma->vm_ops->get_policy(vma, addr, ilx) : vma->vm_policy;
1800}
1801
1802/*
1803 * get_vma_policy(@vma, @addr, @order, @ilx)
1804 * @vma: virtual memory area whose policy is sought
1805 * @addr: address in @vma for shared policy lookup
1806 * @order: 0, or appropriate huge_page_order for interleaving
1807 * @ilx: interleave index (output), for use only when MPOL_INTERLEAVE or
1808 * MPOL_WEIGHTED_INTERLEAVE
1809 *
1810 * Returns effective policy for a VMA at specified address.
1811 * Falls back to current->mempolicy or system default policy, as necessary.
1812 * Shared policies [those marked as MPOL_F_SHARED] require an extra reference
1813 * count--added by the get_policy() vm_op, as appropriate--to protect against
1814 * freeing by another task. It is the caller's responsibility to free the
1815 * extra reference for shared policies.
1816 */
1817struct mempolicy *get_vma_policy(struct vm_area_struct *vma,
1818 unsigned long addr, int order, pgoff_t *ilx)
1819{
1820 struct mempolicy *pol;
1821
1822 pol = __get_vma_policy(vma, addr, ilx);
1823 if (!pol)
1824 pol = get_task_policy(current);
1825 if (pol->mode == MPOL_INTERLEAVE ||
1826 pol->mode == MPOL_WEIGHTED_INTERLEAVE) {
1827 *ilx += vma->vm_pgoff >> order;
1828 *ilx += (addr - vma->vm_start) >> (PAGE_SHIFT + order);
1829 }
1830 return pol;
1831}
1832
1833bool vma_policy_mof(struct vm_area_struct *vma)
1834{
1835 struct mempolicy *pol;
1836
1837 if (vma->vm_ops && vma->vm_ops->get_policy) {
1838 bool ret = false;
1839 pgoff_t ilx; /* ignored here */
1840
1841 pol = vma->vm_ops->get_policy(vma, vma->vm_start, &ilx);
1842 if (pol && (pol->flags & MPOL_F_MOF))
1843 ret = true;
1844 mpol_cond_put(pol);
1845
1846 return ret;
1847 }
1848
1849 pol = vma->vm_policy;
1850 if (!pol)
1851 pol = get_task_policy(current);
1852
1853 return pol->flags & MPOL_F_MOF;
1854}
1855
1856bool apply_policy_zone(struct mempolicy *policy, enum zone_type zone)
1857{
1858 enum zone_type dynamic_policy_zone = policy_zone;
1859
1860 BUG_ON(dynamic_policy_zone == ZONE_MOVABLE);
1861
1862 /*
1863 * if policy->nodes has movable memory only,
1864 * we apply policy when gfp_zone(gfp) = ZONE_MOVABLE only.
1865 *
1866 * policy->nodes is intersect with node_states[N_MEMORY].
1867 * so if the following test fails, it implies
1868 * policy->nodes has movable memory only.
1869 */
1870 if (!nodes_intersects(policy->nodes, node_states[N_HIGH_MEMORY]))
1871 dynamic_policy_zone = ZONE_MOVABLE;
1872
1873 return zone >= dynamic_policy_zone;
1874}
1875
1876static unsigned int weighted_interleave_nodes(struct mempolicy *policy)
1877{
1878 unsigned int node;
1879 unsigned int cpuset_mems_cookie;
1880
1881retry:
1882 /* to prevent miscount use tsk->mems_allowed_seq to detect rebind */
1883 cpuset_mems_cookie = read_mems_allowed_begin();
1884 node = current->il_prev;
1885 if (!current->il_weight || !node_isset(node, policy->nodes)) {
1886 node = next_node_in(node, policy->nodes);
1887 if (read_mems_allowed_retry(cpuset_mems_cookie))
1888 goto retry;
1889 if (node == MAX_NUMNODES)
1890 return node;
1891 current->il_prev = node;
1892 current->il_weight = get_il_weight(node);
1893 }
1894 current->il_weight--;
1895 return node;
1896}
1897
1898/* Do dynamic interleaving for a process */
1899static unsigned int interleave_nodes(struct mempolicy *policy)
1900{
1901 unsigned int nid;
1902 unsigned int cpuset_mems_cookie;
1903
1904 /* to prevent miscount, use tsk->mems_allowed_seq to detect rebind */
1905 do {
1906 cpuset_mems_cookie = read_mems_allowed_begin();
1907 nid = next_node_in(current->il_prev, policy->nodes);
1908 } while (read_mems_allowed_retry(cpuset_mems_cookie));
1909
1910 if (nid < MAX_NUMNODES)
1911 current->il_prev = nid;
1912 return nid;
1913}
1914
1915/*
1916 * Depending on the memory policy provide a node from which to allocate the
1917 * next slab entry.
1918 */
1919unsigned int mempolicy_slab_node(void)
1920{
1921 struct mempolicy *policy;
1922 int node = numa_mem_id();
1923
1924 if (!in_task())
1925 return node;
1926
1927 policy = current->mempolicy;
1928 if (!policy)
1929 return node;
1930
1931 switch (policy->mode) {
1932 case MPOL_PREFERRED:
1933 return first_node(policy->nodes);
1934
1935 case MPOL_INTERLEAVE:
1936 return interleave_nodes(policy);
1937
1938 case MPOL_WEIGHTED_INTERLEAVE:
1939 return weighted_interleave_nodes(policy);
1940
1941 case MPOL_BIND:
1942 case MPOL_PREFERRED_MANY:
1943 {
1944 struct zoneref *z;
1945
1946 /*
1947 * Follow bind policy behavior and start allocation at the
1948 * first node.
1949 */
1950 struct zonelist *zonelist;
1951 enum zone_type highest_zoneidx = gfp_zone(GFP_KERNEL);
1952 zonelist = &NODE_DATA(node)->node_zonelists[ZONELIST_FALLBACK];
1953 z = first_zones_zonelist(zonelist, highest_zoneidx,
1954 &policy->nodes);
1955 return z->zone ? zone_to_nid(z->zone) : node;
1956 }
1957 case MPOL_LOCAL:
1958 return node;
1959
1960 default:
1961 BUG();
1962 }
1963}
1964
1965static unsigned int read_once_policy_nodemask(struct mempolicy *pol,
1966 nodemask_t *mask)
1967{
1968 /*
1969 * barrier stabilizes the nodemask locally so that it can be iterated
1970 * over safely without concern for changes. Allocators validate node
1971 * selection does not violate mems_allowed, so this is safe.
1972 */
1973 barrier();
1974 memcpy(mask, &pol->nodes, sizeof(nodemask_t));
1975 barrier();
1976 return nodes_weight(*mask);
1977}
1978
1979static unsigned int weighted_interleave_nid(struct mempolicy *pol, pgoff_t ilx)
1980{
1981 nodemask_t nodemask;
1982 unsigned int target, nr_nodes;
1983 u8 *table;
1984 unsigned int weight_total = 0;
1985 u8 weight;
1986 int nid;
1987
1988 nr_nodes = read_once_policy_nodemask(pol, &nodemask);
1989 if (!nr_nodes)
1990 return numa_node_id();
1991
1992 rcu_read_lock();
1993 table = rcu_dereference(iw_table);
1994 /* calculate the total weight */
1995 for_each_node_mask(nid, nodemask) {
1996 /* detect system default usage */
1997 weight = table ? table[nid] : 1;
1998 weight = weight ? weight : 1;
1999 weight_total += weight;
2000 }
2001
2002 /* Calculate the node offset based on totals */
2003 target = ilx % weight_total;
2004 nid = first_node(nodemask);
2005 while (target) {
2006 /* detect system default usage */
2007 weight = table ? table[nid] : 1;
2008 weight = weight ? weight : 1;
2009 if (target < weight)
2010 break;
2011 target -= weight;
2012 nid = next_node_in(nid, nodemask);
2013 }
2014 rcu_read_unlock();
2015 return nid;
2016}
2017
2018/*
2019 * Do static interleaving for interleave index @ilx. Returns the ilx'th
2020 * node in pol->nodes (starting from ilx=0), wrapping around if ilx
2021 * exceeds the number of present nodes.
2022 */
2023static unsigned int interleave_nid(struct mempolicy *pol, pgoff_t ilx)
2024{
2025 nodemask_t nodemask;
2026 unsigned int target, nnodes;
2027 int i;
2028 int nid;
2029
2030 nnodes = read_once_policy_nodemask(pol, &nodemask);
2031 if (!nnodes)
2032 return numa_node_id();
2033 target = ilx % nnodes;
2034 nid = first_node(nodemask);
2035 for (i = 0; i < target; i++)
2036 nid = next_node(nid, nodemask);
2037 return nid;
2038}
2039
2040/*
2041 * Return a nodemask representing a mempolicy for filtering nodes for
2042 * page allocation, together with preferred node id (or the input node id).
2043 */
2044static nodemask_t *policy_nodemask(gfp_t gfp, struct mempolicy *pol,
2045 pgoff_t ilx, int *nid)
2046{
2047 nodemask_t *nodemask = NULL;
2048
2049 switch (pol->mode) {
2050 case MPOL_PREFERRED:
2051 /* Override input node id */
2052 *nid = first_node(pol->nodes);
2053 break;
2054 case MPOL_PREFERRED_MANY:
2055 nodemask = &pol->nodes;
2056 if (pol->home_node != NUMA_NO_NODE)
2057 *nid = pol->home_node;
2058 break;
2059 case MPOL_BIND:
2060 /* Restrict to nodemask (but not on lower zones) */
2061 if (apply_policy_zone(pol, gfp_zone(gfp)) &&
2062 cpuset_nodemask_valid_mems_allowed(&pol->nodes))
2063 nodemask = &pol->nodes;
2064 if (pol->home_node != NUMA_NO_NODE)
2065 *nid = pol->home_node;
2066 /*
2067 * __GFP_THISNODE shouldn't even be used with the bind policy
2068 * because we might easily break the expectation to stay on the
2069 * requested node and not break the policy.
2070 */
2071 WARN_ON_ONCE(gfp & __GFP_THISNODE);
2072 break;
2073 case MPOL_INTERLEAVE:
2074 /* Override input node id */
2075 *nid = (ilx == NO_INTERLEAVE_INDEX) ?
2076 interleave_nodes(pol) : interleave_nid(pol, ilx);
2077 break;
2078 case MPOL_WEIGHTED_INTERLEAVE:
2079 *nid = (ilx == NO_INTERLEAVE_INDEX) ?
2080 weighted_interleave_nodes(pol) :
2081 weighted_interleave_nid(pol, ilx);
2082 break;
2083 }
2084
2085 return nodemask;
2086}
2087
2088#ifdef CONFIG_HUGETLBFS
2089/*
2090 * huge_node(@vma, @addr, @gfp_flags, @mpol)
2091 * @vma: virtual memory area whose policy is sought
2092 * @addr: address in @vma for shared policy lookup and interleave policy
2093 * @gfp_flags: for requested zone
2094 * @mpol: pointer to mempolicy pointer for reference counted mempolicy
2095 * @nodemask: pointer to nodemask pointer for 'bind' and 'prefer-many' policy
2096 *
2097 * Returns a nid suitable for a huge page allocation and a pointer
2098 * to the struct mempolicy for conditional unref after allocation.
2099 * If the effective policy is 'bind' or 'prefer-many', returns a pointer
2100 * to the mempolicy's @nodemask for filtering the zonelist.
2101 */
2102int huge_node(struct vm_area_struct *vma, unsigned long addr, gfp_t gfp_flags,
2103 struct mempolicy **mpol, nodemask_t **nodemask)
2104{
2105 pgoff_t ilx;
2106 int nid;
2107
2108 nid = numa_node_id();
2109 *mpol = get_vma_policy(vma, addr, hstate_vma(vma)->order, &ilx);
2110 *nodemask = policy_nodemask(gfp_flags, *mpol, ilx, &nid);
2111 return nid;
2112}
2113
2114/*
2115 * init_nodemask_of_mempolicy
2116 *
2117 * If the current task's mempolicy is "default" [NULL], return 'false'
2118 * to indicate default policy. Otherwise, extract the policy nodemask
2119 * for 'bind' or 'interleave' policy into the argument nodemask, or
2120 * initialize the argument nodemask to contain the single node for
2121 * 'preferred' or 'local' policy and return 'true' to indicate presence
2122 * of non-default mempolicy.
2123 *
2124 * We don't bother with reference counting the mempolicy [mpol_get/put]
2125 * because the current task is examining it's own mempolicy and a task's
2126 * mempolicy is only ever changed by the task itself.
2127 *
2128 * N.B., it is the caller's responsibility to free a returned nodemask.
2129 */
2130bool init_nodemask_of_mempolicy(nodemask_t *mask)
2131{
2132 struct mempolicy *mempolicy;
2133
2134 if (!(mask && current->mempolicy))
2135 return false;
2136
2137 task_lock(current);
2138 mempolicy = current->mempolicy;
2139 switch (mempolicy->mode) {
2140 case MPOL_PREFERRED:
2141 case MPOL_PREFERRED_MANY:
2142 case MPOL_BIND:
2143 case MPOL_INTERLEAVE:
2144 case MPOL_WEIGHTED_INTERLEAVE:
2145 *mask = mempolicy->nodes;
2146 break;
2147
2148 case MPOL_LOCAL:
2149 init_nodemask_of_node(mask, numa_node_id());
2150 break;
2151
2152 default:
2153 BUG();
2154 }
2155 task_unlock(current);
2156
2157 return true;
2158}
2159#endif
2160
2161/*
2162 * mempolicy_in_oom_domain
2163 *
2164 * If tsk's mempolicy is "bind", check for intersection between mask and
2165 * the policy nodemask. Otherwise, return true for all other policies
2166 * including "interleave", as a tsk with "interleave" policy may have
2167 * memory allocated from all nodes in system.
2168 *
2169 * Takes task_lock(tsk) to prevent freeing of its mempolicy.
2170 */
2171bool mempolicy_in_oom_domain(struct task_struct *tsk,
2172 const nodemask_t *mask)
2173{
2174 struct mempolicy *mempolicy;
2175 bool ret = true;
2176
2177 if (!mask)
2178 return ret;
2179
2180 task_lock(tsk);
2181 mempolicy = tsk->mempolicy;
2182 if (mempolicy && mempolicy->mode == MPOL_BIND)
2183 ret = nodes_intersects(mempolicy->nodes, *mask);
2184 task_unlock(tsk);
2185
2186 return ret;
2187}
2188
2189static struct page *alloc_pages_preferred_many(gfp_t gfp, unsigned int order,
2190 int nid, nodemask_t *nodemask)
2191{
2192 struct page *page;
2193 gfp_t preferred_gfp;
2194
2195 /*
2196 * This is a two pass approach. The first pass will only try the
2197 * preferred nodes but skip the direct reclaim and allow the
2198 * allocation to fail, while the second pass will try all the
2199 * nodes in system.
2200 */
2201 preferred_gfp = gfp | __GFP_NOWARN;
2202 preferred_gfp &= ~(__GFP_DIRECT_RECLAIM | __GFP_NOFAIL);
2203 page = __alloc_pages(preferred_gfp, order, nid, nodemask);
2204 if (!page)
2205 page = __alloc_pages(gfp, order, nid, NULL);
2206
2207 return page;
2208}
2209
2210/**
2211 * alloc_pages_mpol - Allocate pages according to NUMA mempolicy.
2212 * @gfp: GFP flags.
2213 * @order: Order of the page allocation.
2214 * @pol: Pointer to the NUMA mempolicy.
2215 * @ilx: Index for interleave mempolicy (also distinguishes alloc_pages()).
2216 * @nid: Preferred node (usually numa_node_id() but @mpol may override it).
2217 *
2218 * Return: The page on success or NULL if allocation fails.
2219 */
2220struct page *alloc_pages_mpol(gfp_t gfp, unsigned int order,
2221 struct mempolicy *pol, pgoff_t ilx, int nid)
2222{
2223 nodemask_t *nodemask;
2224 struct page *page;
2225
2226 nodemask = policy_nodemask(gfp, pol, ilx, &nid);
2227
2228 if (pol->mode == MPOL_PREFERRED_MANY)
2229 return alloc_pages_preferred_many(gfp, order, nid, nodemask);
2230
2231 if (IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE) &&
2232 /* filter "hugepage" allocation, unless from alloc_pages() */
2233 order == HPAGE_PMD_ORDER && ilx != NO_INTERLEAVE_INDEX) {
2234 /*
2235 * For hugepage allocation and non-interleave policy which
2236 * allows the current node (or other explicitly preferred
2237 * node) we only try to allocate from the current/preferred
2238 * node and don't fall back to other nodes, as the cost of
2239 * remote accesses would likely offset THP benefits.
2240 *
2241 * If the policy is interleave or does not allow the current
2242 * node in its nodemask, we allocate the standard way.
2243 */
2244 if (pol->mode != MPOL_INTERLEAVE &&
2245 pol->mode != MPOL_WEIGHTED_INTERLEAVE &&
2246 (!nodemask || node_isset(nid, *nodemask))) {
2247 /*
2248 * First, try to allocate THP only on local node, but
2249 * don't reclaim unnecessarily, just compact.
2250 */
2251 page = __alloc_pages_node(nid,
2252 gfp | __GFP_THISNODE | __GFP_NORETRY, order);
2253 if (page || !(gfp & __GFP_DIRECT_RECLAIM))
2254 return page;
2255 /*
2256 * If hugepage allocations are configured to always
2257 * synchronous compact or the vma has been madvised
2258 * to prefer hugepage backing, retry allowing remote
2259 * memory with both reclaim and compact as well.
2260 */
2261 }
2262 }
2263
2264 page = __alloc_pages(gfp, order, nid, nodemask);
2265
2266 if (unlikely(pol->mode == MPOL_INTERLEAVE) && page) {
2267 /* skip NUMA_INTERLEAVE_HIT update if numa stats is disabled */
2268 if (static_branch_likely(&vm_numa_stat_key) &&
2269 page_to_nid(page) == nid) {
2270 preempt_disable();
2271 __count_numa_event(page_zone(page), NUMA_INTERLEAVE_HIT);
2272 preempt_enable();
2273 }
2274 }
2275
2276 return page;
2277}
2278
2279/**
2280 * vma_alloc_folio - Allocate a folio for a VMA.
2281 * @gfp: GFP flags.
2282 * @order: Order of the folio.
2283 * @vma: Pointer to VMA.
2284 * @addr: Virtual address of the allocation. Must be inside @vma.
2285 * @hugepage: Unused (was: For hugepages try only preferred node if possible).
2286 *
2287 * Allocate a folio for a specific address in @vma, using the appropriate
2288 * NUMA policy. The caller must hold the mmap_lock of the mm_struct of the
2289 * VMA to prevent it from going away. Should be used for all allocations
2290 * for folios that will be mapped into user space, excepting hugetlbfs, and
2291 * excepting where direct use of alloc_pages_mpol() is more appropriate.
2292 *
2293 * Return: The folio on success or NULL if allocation fails.
2294 */
2295struct folio *vma_alloc_folio(gfp_t gfp, int order, struct vm_area_struct *vma,
2296 unsigned long addr, bool hugepage)
2297{
2298 struct mempolicy *pol;
2299 pgoff_t ilx;
2300 struct page *page;
2301
2302 pol = get_vma_policy(vma, addr, order, &ilx);
2303 page = alloc_pages_mpol(gfp | __GFP_COMP, order,
2304 pol, ilx, numa_node_id());
2305 mpol_cond_put(pol);
2306 return page_rmappable_folio(page);
2307}
2308EXPORT_SYMBOL(vma_alloc_folio);
2309
2310/**
2311 * alloc_pages - Allocate pages.
2312 * @gfp: GFP flags.
2313 * @order: Power of two of number of pages to allocate.
2314 *
2315 * Allocate 1 << @order contiguous pages. The physical address of the
2316 * first page is naturally aligned (eg an order-3 allocation will be aligned
2317 * to a multiple of 8 * PAGE_SIZE bytes). The NUMA policy of the current
2318 * process is honoured when in process context.
2319 *
2320 * Context: Can be called from any context, providing the appropriate GFP
2321 * flags are used.
2322 * Return: The page on success or NULL if allocation fails.
2323 */
2324struct page *alloc_pages(gfp_t gfp, unsigned int order)
2325{
2326 struct mempolicy *pol = &default_policy;
2327
2328 /*
2329 * No reference counting needed for current->mempolicy
2330 * nor system default_policy
2331 */
2332 if (!in_interrupt() && !(gfp & __GFP_THISNODE))
2333 pol = get_task_policy(current);
2334
2335 return alloc_pages_mpol(gfp, order,
2336 pol, NO_INTERLEAVE_INDEX, numa_node_id());
2337}
2338EXPORT_SYMBOL(alloc_pages);
2339
2340struct folio *folio_alloc(gfp_t gfp, unsigned int order)
2341{
2342 return page_rmappable_folio(alloc_pages(gfp | __GFP_COMP, order));
2343}
2344EXPORT_SYMBOL(folio_alloc);
2345
2346static unsigned long alloc_pages_bulk_array_interleave(gfp_t gfp,
2347 struct mempolicy *pol, unsigned long nr_pages,
2348 struct page **page_array)
2349{
2350 int nodes;
2351 unsigned long nr_pages_per_node;
2352 int delta;
2353 int i;
2354 unsigned long nr_allocated;
2355 unsigned long total_allocated = 0;
2356
2357 nodes = nodes_weight(pol->nodes);
2358 nr_pages_per_node = nr_pages / nodes;
2359 delta = nr_pages - nodes * nr_pages_per_node;
2360
2361 for (i = 0; i < nodes; i++) {
2362 if (delta) {
2363 nr_allocated = __alloc_pages_bulk(gfp,
2364 interleave_nodes(pol), NULL,
2365 nr_pages_per_node + 1, NULL,
2366 page_array);
2367 delta--;
2368 } else {
2369 nr_allocated = __alloc_pages_bulk(gfp,
2370 interleave_nodes(pol), NULL,
2371 nr_pages_per_node, NULL, page_array);
2372 }
2373
2374 page_array += nr_allocated;
2375 total_allocated += nr_allocated;
2376 }
2377
2378 return total_allocated;
2379}
2380
2381static unsigned long alloc_pages_bulk_array_weighted_interleave(gfp_t gfp,
2382 struct mempolicy *pol, unsigned long nr_pages,
2383 struct page **page_array)
2384{
2385 struct task_struct *me = current;
2386 unsigned int cpuset_mems_cookie;
2387 unsigned long total_allocated = 0;
2388 unsigned long nr_allocated = 0;
2389 unsigned long rounds;
2390 unsigned long node_pages, delta;
2391 u8 *table, *weights, weight;
2392 unsigned int weight_total = 0;
2393 unsigned long rem_pages = nr_pages;
2394 nodemask_t nodes;
2395 int nnodes, node;
2396 int resume_node = MAX_NUMNODES - 1;
2397 u8 resume_weight = 0;
2398 int prev_node;
2399 int i;
2400
2401 if (!nr_pages)
2402 return 0;
2403
2404 /* read the nodes onto the stack, retry if done during rebind */
2405 do {
2406 cpuset_mems_cookie = read_mems_allowed_begin();
2407 nnodes = read_once_policy_nodemask(pol, &nodes);
2408 } while (read_mems_allowed_retry(cpuset_mems_cookie));
2409
2410 /* if the nodemask has become invalid, we cannot do anything */
2411 if (!nnodes)
2412 return 0;
2413
2414 /* Continue allocating from most recent node and adjust the nr_pages */
2415 node = me->il_prev;
2416 weight = me->il_weight;
2417 if (weight && node_isset(node, nodes)) {
2418 node_pages = min(rem_pages, weight);
2419 nr_allocated = __alloc_pages_bulk(gfp, node, NULL, node_pages,
2420 NULL, page_array);
2421 page_array += nr_allocated;
2422 total_allocated += nr_allocated;
2423 /* if that's all the pages, no need to interleave */
2424 if (rem_pages <= weight) {
2425 me->il_weight -= rem_pages;
2426 return total_allocated;
2427 }
2428 /* Otherwise we adjust remaining pages, continue from there */
2429 rem_pages -= weight;
2430 }
2431 /* clear active weight in case of an allocation failure */
2432 me->il_weight = 0;
2433 prev_node = node;
2434
2435 /* create a local copy of node weights to operate on outside rcu */
2436 weights = kzalloc(nr_node_ids, GFP_KERNEL);
2437 if (!weights)
2438 return total_allocated;
2439
2440 rcu_read_lock();
2441 table = rcu_dereference(iw_table);
2442 if (table)
2443 memcpy(weights, table, nr_node_ids);
2444 rcu_read_unlock();
2445
2446 /* calculate total, detect system default usage */
2447 for_each_node_mask(node, nodes) {
2448 if (!weights[node])
2449 weights[node] = 1;
2450 weight_total += weights[node];
2451 }
2452
2453 /*
2454 * Calculate rounds/partial rounds to minimize __alloc_pages_bulk calls.
2455 * Track which node weighted interleave should resume from.
2456 *
2457 * if (rounds > 0) and (delta == 0), resume_node will always be
2458 * the node following prev_node and its weight.
2459 */
2460 rounds = rem_pages / weight_total;
2461 delta = rem_pages % weight_total;
2462 resume_node = next_node_in(prev_node, nodes);
2463 resume_weight = weights[resume_node];
2464 for (i = 0; i < nnodes; i++) {
2465 node = next_node_in(prev_node, nodes);
2466 weight = weights[node];
2467 node_pages = weight * rounds;
2468 /* If a delta exists, add this node's portion of the delta */
2469 if (delta > weight) {
2470 node_pages += weight;
2471 delta -= weight;
2472 } else if (delta) {
2473 /* when delta is depleted, resume from that node */
2474 node_pages += delta;
2475 resume_node = node;
2476 resume_weight = weight - delta;
2477 delta = 0;
2478 }
2479 /* node_pages can be 0 if an allocation fails and rounds == 0 */
2480 if (!node_pages)
2481 break;
2482 nr_allocated = __alloc_pages_bulk(gfp, node, NULL, node_pages,
2483 NULL, page_array);
2484 page_array += nr_allocated;
2485 total_allocated += nr_allocated;
2486 if (total_allocated == nr_pages)
2487 break;
2488 prev_node = node;
2489 }
2490 me->il_prev = resume_node;
2491 me->il_weight = resume_weight;
2492 kfree(weights);
2493 return total_allocated;
2494}
2495
2496static unsigned long alloc_pages_bulk_array_preferred_many(gfp_t gfp, int nid,
2497 struct mempolicy *pol, unsigned long nr_pages,
2498 struct page **page_array)
2499{
2500 gfp_t preferred_gfp;
2501 unsigned long nr_allocated = 0;
2502
2503 preferred_gfp = gfp | __GFP_NOWARN;
2504 preferred_gfp &= ~(__GFP_DIRECT_RECLAIM | __GFP_NOFAIL);
2505
2506 nr_allocated = __alloc_pages_bulk(preferred_gfp, nid, &pol->nodes,
2507 nr_pages, NULL, page_array);
2508
2509 if (nr_allocated < nr_pages)
2510 nr_allocated += __alloc_pages_bulk(gfp, numa_node_id(), NULL,
2511 nr_pages - nr_allocated, NULL,
2512 page_array + nr_allocated);
2513 return nr_allocated;
2514}
2515
2516/* alloc pages bulk and mempolicy should be considered at the
2517 * same time in some situation such as vmalloc.
2518 *
2519 * It can accelerate memory allocation especially interleaving
2520 * allocate memory.
2521 */
2522unsigned long alloc_pages_bulk_array_mempolicy(gfp_t gfp,
2523 unsigned long nr_pages, struct page **page_array)
2524{
2525 struct mempolicy *pol = &default_policy;
2526 nodemask_t *nodemask;
2527 int nid;
2528
2529 if (!in_interrupt() && !(gfp & __GFP_THISNODE))
2530 pol = get_task_policy(current);
2531
2532 if (pol->mode == MPOL_INTERLEAVE)
2533 return alloc_pages_bulk_array_interleave(gfp, pol,
2534 nr_pages, page_array);
2535
2536 if (pol->mode == MPOL_WEIGHTED_INTERLEAVE)
2537 return alloc_pages_bulk_array_weighted_interleave(
2538 gfp, pol, nr_pages, page_array);
2539
2540 if (pol->mode == MPOL_PREFERRED_MANY)
2541 return alloc_pages_bulk_array_preferred_many(gfp,
2542 numa_node_id(), pol, nr_pages, page_array);
2543
2544 nid = numa_node_id();
2545 nodemask = policy_nodemask(gfp, pol, NO_INTERLEAVE_INDEX, &nid);
2546 return __alloc_pages_bulk(gfp, nid, nodemask,
2547 nr_pages, NULL, page_array);
2548}
2549
2550int vma_dup_policy(struct vm_area_struct *src, struct vm_area_struct *dst)
2551{
2552 struct mempolicy *pol = mpol_dup(src->vm_policy);
2553
2554 if (IS_ERR(pol))
2555 return PTR_ERR(pol);
2556 dst->vm_policy = pol;
2557 return 0;
2558}
2559
2560/*
2561 * If mpol_dup() sees current->cpuset == cpuset_being_rebound, then it
2562 * rebinds the mempolicy its copying by calling mpol_rebind_policy()
2563 * with the mems_allowed returned by cpuset_mems_allowed(). This
2564 * keeps mempolicies cpuset relative after its cpuset moves. See
2565 * further kernel/cpuset.c update_nodemask().
2566 *
2567 * current's mempolicy may be rebinded by the other task(the task that changes
2568 * cpuset's mems), so we needn't do rebind work for current task.
2569 */
2570
2571/* Slow path of a mempolicy duplicate */
2572struct mempolicy *__mpol_dup(struct mempolicy *old)
2573{
2574 struct mempolicy *new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
2575
2576 if (!new)
2577 return ERR_PTR(-ENOMEM);
2578
2579 /* task's mempolicy is protected by alloc_lock */
2580 if (old == current->mempolicy) {
2581 task_lock(current);
2582 *new = *old;
2583 task_unlock(current);
2584 } else
2585 *new = *old;
2586
2587 if (current_cpuset_is_being_rebound()) {
2588 nodemask_t mems = cpuset_mems_allowed(current);
2589 mpol_rebind_policy(new, &mems);
2590 }
2591 atomic_set(&new->refcnt, 1);
2592 return new;
2593}
2594
2595/* Slow path of a mempolicy comparison */
2596bool __mpol_equal(struct mempolicy *a, struct mempolicy *b)
2597{
2598 if (!a || !b)
2599 return false;
2600 if (a->mode != b->mode)
2601 return false;
2602 if (a->flags != b->flags)
2603 return false;
2604 if (a->home_node != b->home_node)
2605 return false;
2606 if (mpol_store_user_nodemask(a))
2607 if (!nodes_equal(a->w.user_nodemask, b->w.user_nodemask))
2608 return false;
2609
2610 switch (a->mode) {
2611 case MPOL_BIND:
2612 case MPOL_INTERLEAVE:
2613 case MPOL_PREFERRED:
2614 case MPOL_PREFERRED_MANY:
2615 case MPOL_WEIGHTED_INTERLEAVE:
2616 return !!nodes_equal(a->nodes, b->nodes);
2617 case MPOL_LOCAL:
2618 return true;
2619 default:
2620 BUG();
2621 return false;
2622 }
2623}
2624
2625/*
2626 * Shared memory backing store policy support.
2627 *
2628 * Remember policies even when nobody has shared memory mapped.
2629 * The policies are kept in Red-Black tree linked from the inode.
2630 * They are protected by the sp->lock rwlock, which should be held
2631 * for any accesses to the tree.
2632 */
2633
2634/*
2635 * lookup first element intersecting start-end. Caller holds sp->lock for
2636 * reading or for writing
2637 */
2638static struct sp_node *sp_lookup(struct shared_policy *sp,
2639 pgoff_t start, pgoff_t end)
2640{
2641 struct rb_node *n = sp->root.rb_node;
2642
2643 while (n) {
2644 struct sp_node *p = rb_entry(n, struct sp_node, nd);
2645
2646 if (start >= p->end)
2647 n = n->rb_right;
2648 else if (end <= p->start)
2649 n = n->rb_left;
2650 else
2651 break;
2652 }
2653 if (!n)
2654 return NULL;
2655 for (;;) {
2656 struct sp_node *w = NULL;
2657 struct rb_node *prev = rb_prev(n);
2658 if (!prev)
2659 break;
2660 w = rb_entry(prev, struct sp_node, nd);
2661 if (w->end <= start)
2662 break;
2663 n = prev;
2664 }
2665 return rb_entry(n, struct sp_node, nd);
2666}
2667
2668/*
2669 * Insert a new shared policy into the list. Caller holds sp->lock for
2670 * writing.
2671 */
2672static void sp_insert(struct shared_policy *sp, struct sp_node *new)
2673{
2674 struct rb_node **p = &sp->root.rb_node;
2675 struct rb_node *parent = NULL;
2676 struct sp_node *nd;
2677
2678 while (*p) {
2679 parent = *p;
2680 nd = rb_entry(parent, struct sp_node, nd);
2681 if (new->start < nd->start)
2682 p = &(*p)->rb_left;
2683 else if (new->end > nd->end)
2684 p = &(*p)->rb_right;
2685 else
2686 BUG();
2687 }
2688 rb_link_node(&new->nd, parent, p);
2689 rb_insert_color(&new->nd, &sp->root);
2690}
2691
2692/* Find shared policy intersecting idx */
2693struct mempolicy *mpol_shared_policy_lookup(struct shared_policy *sp,
2694 pgoff_t idx)
2695{
2696 struct mempolicy *pol = NULL;
2697 struct sp_node *sn;
2698
2699 if (!sp->root.rb_node)
2700 return NULL;
2701 read_lock(&sp->lock);
2702 sn = sp_lookup(sp, idx, idx+1);
2703 if (sn) {
2704 mpol_get(sn->policy);
2705 pol = sn->policy;
2706 }
2707 read_unlock(&sp->lock);
2708 return pol;
2709}
2710
2711static void sp_free(struct sp_node *n)
2712{
2713 mpol_put(n->policy);
2714 kmem_cache_free(sn_cache, n);
2715}
2716
2717/**
2718 * mpol_misplaced - check whether current folio node is valid in policy
2719 *
2720 * @folio: folio to be checked
2721 * @vma: vm area where folio mapped
2722 * @addr: virtual address in @vma for shared policy lookup and interleave policy
2723 *
2724 * Lookup current policy node id for vma,addr and "compare to" folio's
2725 * node id. Policy determination "mimics" alloc_page_vma().
2726 * Called from fault path where we know the vma and faulting address.
2727 *
2728 * Return: NUMA_NO_NODE if the page is in a node that is valid for this
2729 * policy, or a suitable node ID to allocate a replacement folio from.
2730 */
2731int mpol_misplaced(struct folio *folio, struct vm_area_struct *vma,
2732 unsigned long addr)
2733{
2734 struct mempolicy *pol;
2735 pgoff_t ilx;
2736 struct zoneref *z;
2737 int curnid = folio_nid(folio);
2738 int thiscpu = raw_smp_processor_id();
2739 int thisnid = cpu_to_node(thiscpu);
2740 int polnid = NUMA_NO_NODE;
2741 int ret = NUMA_NO_NODE;
2742
2743 pol = get_vma_policy(vma, addr, folio_order(folio), &ilx);
2744 if (!(pol->flags & MPOL_F_MOF))
2745 goto out;
2746
2747 switch (pol->mode) {
2748 case MPOL_INTERLEAVE:
2749 polnid = interleave_nid(pol, ilx);
2750 break;
2751
2752 case MPOL_WEIGHTED_INTERLEAVE:
2753 polnid = weighted_interleave_nid(pol, ilx);
2754 break;
2755
2756 case MPOL_PREFERRED:
2757 if (node_isset(curnid, pol->nodes))
2758 goto out;
2759 polnid = first_node(pol->nodes);
2760 break;
2761
2762 case MPOL_LOCAL:
2763 polnid = numa_node_id();
2764 break;
2765
2766 case MPOL_BIND:
2767 /* Optimize placement among multiple nodes via NUMA balancing */
2768 if (pol->flags & MPOL_F_MORON) {
2769 if (node_isset(thisnid, pol->nodes))
2770 break;
2771 goto out;
2772 }
2773 fallthrough;
2774
2775 case MPOL_PREFERRED_MANY:
2776 /*
2777 * use current page if in policy nodemask,
2778 * else select nearest allowed node, if any.
2779 * If no allowed nodes, use current [!misplaced].
2780 */
2781 if (node_isset(curnid, pol->nodes))
2782 goto out;
2783 z = first_zones_zonelist(
2784 node_zonelist(numa_node_id(), GFP_HIGHUSER),
2785 gfp_zone(GFP_HIGHUSER),
2786 &pol->nodes);
2787 polnid = zone_to_nid(z->zone);
2788 break;
2789
2790 default:
2791 BUG();
2792 }
2793
2794 /* Migrate the folio towards the node whose CPU is referencing it */
2795 if (pol->flags & MPOL_F_MORON) {
2796 polnid = thisnid;
2797
2798 if (!should_numa_migrate_memory(current, folio, curnid,
2799 thiscpu))
2800 goto out;
2801 }
2802
2803 if (curnid != polnid)
2804 ret = polnid;
2805out:
2806 mpol_cond_put(pol);
2807
2808 return ret;
2809}
2810
2811/*
2812 * Drop the (possibly final) reference to task->mempolicy. It needs to be
2813 * dropped after task->mempolicy is set to NULL so that any allocation done as
2814 * part of its kmem_cache_free(), such as by KASAN, doesn't reference a freed
2815 * policy.
2816 */
2817void mpol_put_task_policy(struct task_struct *task)
2818{
2819 struct mempolicy *pol;
2820
2821 task_lock(task);
2822 pol = task->mempolicy;
2823 task->mempolicy = NULL;
2824 task_unlock(task);
2825 mpol_put(pol);
2826}
2827
2828static void sp_delete(struct shared_policy *sp, struct sp_node *n)
2829{
2830 rb_erase(&n->nd, &sp->root);
2831 sp_free(n);
2832}
2833
2834static void sp_node_init(struct sp_node *node, unsigned long start,
2835 unsigned long end, struct mempolicy *pol)
2836{
2837 node->start = start;
2838 node->end = end;
2839 node->policy = pol;
2840}
2841
2842static struct sp_node *sp_alloc(unsigned long start, unsigned long end,
2843 struct mempolicy *pol)
2844{
2845 struct sp_node *n;
2846 struct mempolicy *newpol;
2847
2848 n = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2849 if (!n)
2850 return NULL;
2851
2852 newpol = mpol_dup(pol);
2853 if (IS_ERR(newpol)) {
2854 kmem_cache_free(sn_cache, n);
2855 return NULL;
2856 }
2857 newpol->flags |= MPOL_F_SHARED;
2858 sp_node_init(n, start, end, newpol);
2859
2860 return n;
2861}
2862
2863/* Replace a policy range. */
2864static int shared_policy_replace(struct shared_policy *sp, pgoff_t start,
2865 pgoff_t end, struct sp_node *new)
2866{
2867 struct sp_node *n;
2868 struct sp_node *n_new = NULL;
2869 struct mempolicy *mpol_new = NULL;
2870 int ret = 0;
2871
2872restart:
2873 write_lock(&sp->lock);
2874 n = sp_lookup(sp, start, end);
2875 /* Take care of old policies in the same range. */
2876 while (n && n->start < end) {
2877 struct rb_node *next = rb_next(&n->nd);
2878 if (n->start >= start) {
2879 if (n->end <= end)
2880 sp_delete(sp, n);
2881 else
2882 n->start = end;
2883 } else {
2884 /* Old policy spanning whole new range. */
2885 if (n->end > end) {
2886 if (!n_new)
2887 goto alloc_new;
2888
2889 *mpol_new = *n->policy;
2890 atomic_set(&mpol_new->refcnt, 1);
2891 sp_node_init(n_new, end, n->end, mpol_new);
2892 n->end = start;
2893 sp_insert(sp, n_new);
2894 n_new = NULL;
2895 mpol_new = NULL;
2896 break;
2897 } else
2898 n->end = start;
2899 }
2900 if (!next)
2901 break;
2902 n = rb_entry(next, struct sp_node, nd);
2903 }
2904 if (new)
2905 sp_insert(sp, new);
2906 write_unlock(&sp->lock);
2907 ret = 0;
2908
2909err_out:
2910 if (mpol_new)
2911 mpol_put(mpol_new);
2912 if (n_new)
2913 kmem_cache_free(sn_cache, n_new);
2914
2915 return ret;
2916
2917alloc_new:
2918 write_unlock(&sp->lock);
2919 ret = -ENOMEM;
2920 n_new = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2921 if (!n_new)
2922 goto err_out;
2923 mpol_new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
2924 if (!mpol_new)
2925 goto err_out;
2926 atomic_set(&mpol_new->refcnt, 1);
2927 goto restart;
2928}
2929
2930/**
2931 * mpol_shared_policy_init - initialize shared policy for inode
2932 * @sp: pointer to inode shared policy
2933 * @mpol: struct mempolicy to install
2934 *
2935 * Install non-NULL @mpol in inode's shared policy rb-tree.
2936 * On entry, the current task has a reference on a non-NULL @mpol.
2937 * This must be released on exit.
2938 * This is called at get_inode() calls and we can use GFP_KERNEL.
2939 */
2940void mpol_shared_policy_init(struct shared_policy *sp, struct mempolicy *mpol)
2941{
2942 int ret;
2943
2944 sp->root = RB_ROOT; /* empty tree == default mempolicy */
2945 rwlock_init(&sp->lock);
2946
2947 if (mpol) {
2948 struct sp_node *sn;
2949 struct mempolicy *npol;
2950 NODEMASK_SCRATCH(scratch);
2951
2952 if (!scratch)
2953 goto put_mpol;
2954
2955 /* contextualize the tmpfs mount point mempolicy to this file */
2956 npol = mpol_new(mpol->mode, mpol->flags, &mpol->w.user_nodemask);
2957 if (IS_ERR(npol))
2958 goto free_scratch; /* no valid nodemask intersection */
2959
2960 task_lock(current);
2961 ret = mpol_set_nodemask(npol, &mpol->w.user_nodemask, scratch);
2962 task_unlock(current);
2963 if (ret)
2964 goto put_npol;
2965
2966 /* alloc node covering entire file; adds ref to file's npol */
2967 sn = sp_alloc(0, MAX_LFS_FILESIZE >> PAGE_SHIFT, npol);
2968 if (sn)
2969 sp_insert(sp, sn);
2970put_npol:
2971 mpol_put(npol); /* drop initial ref on file's npol */
2972free_scratch:
2973 NODEMASK_SCRATCH_FREE(scratch);
2974put_mpol:
2975 mpol_put(mpol); /* drop our incoming ref on sb mpol */
2976 }
2977}
2978
2979int mpol_set_shared_policy(struct shared_policy *sp,
2980 struct vm_area_struct *vma, struct mempolicy *pol)
2981{
2982 int err;
2983 struct sp_node *new = NULL;
2984 unsigned long sz = vma_pages(vma);
2985
2986 if (pol) {
2987 new = sp_alloc(vma->vm_pgoff, vma->vm_pgoff + sz, pol);
2988 if (!new)
2989 return -ENOMEM;
2990 }
2991 err = shared_policy_replace(sp, vma->vm_pgoff, vma->vm_pgoff + sz, new);
2992 if (err && new)
2993 sp_free(new);
2994 return err;
2995}
2996
2997/* Free a backing policy store on inode delete. */
2998void mpol_free_shared_policy(struct shared_policy *sp)
2999{
3000 struct sp_node *n;
3001 struct rb_node *next;
3002
3003 if (!sp->root.rb_node)
3004 return;
3005 write_lock(&sp->lock);
3006 next = rb_first(&sp->root);
3007 while (next) {
3008 n = rb_entry(next, struct sp_node, nd);
3009 next = rb_next(&n->nd);
3010 sp_delete(sp, n);
3011 }
3012 write_unlock(&sp->lock);
3013}
3014
3015#ifdef CONFIG_NUMA_BALANCING
3016static int __initdata numabalancing_override;
3017
3018static void __init check_numabalancing_enable(void)
3019{
3020 bool numabalancing_default = false;
3021
3022 if (IS_ENABLED(CONFIG_NUMA_BALANCING_DEFAULT_ENABLED))
3023 numabalancing_default = true;
3024
3025 /* Parsed by setup_numabalancing. override == 1 enables, -1 disables */
3026 if (numabalancing_override)
3027 set_numabalancing_state(numabalancing_override == 1);
3028
3029 if (num_online_nodes() > 1 && !numabalancing_override) {
3030 pr_info("%s automatic NUMA balancing. Configure with numa_balancing= or the kernel.numa_balancing sysctl\n",
3031 numabalancing_default ? "Enabling" : "Disabling");
3032 set_numabalancing_state(numabalancing_default);
3033 }
3034}
3035
3036static int __init setup_numabalancing(char *str)
3037{
3038 int ret = 0;
3039 if (!str)
3040 goto out;
3041
3042 if (!strcmp(str, "enable")) {
3043 numabalancing_override = 1;
3044 ret = 1;
3045 } else if (!strcmp(str, "disable")) {
3046 numabalancing_override = -1;
3047 ret = 1;
3048 }
3049out:
3050 if (!ret)
3051 pr_warn("Unable to parse numa_balancing=\n");
3052
3053 return ret;
3054}
3055__setup("numa_balancing=", setup_numabalancing);
3056#else
3057static inline void __init check_numabalancing_enable(void)
3058{
3059}
3060#endif /* CONFIG_NUMA_BALANCING */
3061
3062void __init numa_policy_init(void)
3063{
3064 nodemask_t interleave_nodes;
3065 unsigned long largest = 0;
3066 int nid, prefer = 0;
3067
3068 policy_cache = kmem_cache_create("numa_policy",
3069 sizeof(struct mempolicy),
3070 0, SLAB_PANIC, NULL);
3071
3072 sn_cache = kmem_cache_create("shared_policy_node",
3073 sizeof(struct sp_node),
3074 0, SLAB_PANIC, NULL);
3075
3076 for_each_node(nid) {
3077 preferred_node_policy[nid] = (struct mempolicy) {
3078 .refcnt = ATOMIC_INIT(1),
3079 .mode = MPOL_PREFERRED,
3080 .flags = MPOL_F_MOF | MPOL_F_MORON,
3081 .nodes = nodemask_of_node(nid),
3082 };
3083 }
3084
3085 /*
3086 * Set interleaving policy for system init. Interleaving is only
3087 * enabled across suitably sized nodes (default is >= 16MB), or
3088 * fall back to the largest node if they're all smaller.
3089 */
3090 nodes_clear(interleave_nodes);
3091 for_each_node_state(nid, N_MEMORY) {
3092 unsigned long total_pages = node_present_pages(nid);
3093
3094 /* Preserve the largest node */
3095 if (largest < total_pages) {
3096 largest = total_pages;
3097 prefer = nid;
3098 }
3099
3100 /* Interleave this node? */
3101 if ((total_pages << PAGE_SHIFT) >= (16 << 20))
3102 node_set(nid, interleave_nodes);
3103 }
3104
3105 /* All too small, use the largest */
3106 if (unlikely(nodes_empty(interleave_nodes)))
3107 node_set(prefer, interleave_nodes);
3108
3109 if (do_set_mempolicy(MPOL_INTERLEAVE, 0, &interleave_nodes))
3110 pr_err("%s: interleaving failed\n", __func__);
3111
3112 check_numabalancing_enable();
3113}
3114
3115/* Reset policy of current process to default */
3116void numa_default_policy(void)
3117{
3118 do_set_mempolicy(MPOL_DEFAULT, 0, NULL);
3119}
3120
3121/*
3122 * Parse and format mempolicy from/to strings
3123 */
3124static const char * const policy_modes[] =
3125{
3126 [MPOL_DEFAULT] = "default",
3127 [MPOL_PREFERRED] = "prefer",
3128 [MPOL_BIND] = "bind",
3129 [MPOL_INTERLEAVE] = "interleave",
3130 [MPOL_WEIGHTED_INTERLEAVE] = "weighted interleave",
3131 [MPOL_LOCAL] = "local",
3132 [MPOL_PREFERRED_MANY] = "prefer (many)",
3133};
3134
3135#ifdef CONFIG_TMPFS
3136/**
3137 * mpol_parse_str - parse string to mempolicy, for tmpfs mpol mount option.
3138 * @str: string containing mempolicy to parse
3139 * @mpol: pointer to struct mempolicy pointer, returned on success.
3140 *
3141 * Format of input:
3142 * <mode>[=<flags>][:<nodelist>]
3143 *
3144 * Return: %0 on success, else %1
3145 */
3146int mpol_parse_str(char *str, struct mempolicy **mpol)
3147{
3148 struct mempolicy *new = NULL;
3149 unsigned short mode_flags;
3150 nodemask_t nodes;
3151 char *nodelist = strchr(str, ':');
3152 char *flags = strchr(str, '=');
3153 int err = 1, mode;
3154
3155 if (flags)
3156 *flags++ = '\0'; /* terminate mode string */
3157
3158 if (nodelist) {
3159 /* NUL-terminate mode or flags string */
3160 *nodelist++ = '\0';
3161 if (nodelist_parse(nodelist, nodes))
3162 goto out;
3163 if (!nodes_subset(nodes, node_states[N_MEMORY]))
3164 goto out;
3165 } else
3166 nodes_clear(nodes);
3167
3168 mode = match_string(policy_modes, MPOL_MAX, str);
3169 if (mode < 0)
3170 goto out;
3171
3172 switch (mode) {
3173 case MPOL_PREFERRED:
3174 /*
3175 * Insist on a nodelist of one node only, although later
3176 * we use first_node(nodes) to grab a single node, so here
3177 * nodelist (or nodes) cannot be empty.
3178 */
3179 if (nodelist) {
3180 char *rest = nodelist;
3181 while (isdigit(*rest))
3182 rest++;
3183 if (*rest)
3184 goto out;
3185 if (nodes_empty(nodes))
3186 goto out;
3187 }
3188 break;
3189 case MPOL_INTERLEAVE:
3190 case MPOL_WEIGHTED_INTERLEAVE:
3191 /*
3192 * Default to online nodes with memory if no nodelist
3193 */
3194 if (!nodelist)
3195 nodes = node_states[N_MEMORY];
3196 break;
3197 case MPOL_LOCAL:
3198 /*
3199 * Don't allow a nodelist; mpol_new() checks flags
3200 */
3201 if (nodelist)
3202 goto out;
3203 break;
3204 case MPOL_DEFAULT:
3205 /*
3206 * Insist on a empty nodelist
3207 */
3208 if (!nodelist)
3209 err = 0;
3210 goto out;
3211 case MPOL_PREFERRED_MANY:
3212 case MPOL_BIND:
3213 /*
3214 * Insist on a nodelist
3215 */
3216 if (!nodelist)
3217 goto out;
3218 }
3219
3220 mode_flags = 0;
3221 if (flags) {
3222 /*
3223 * Currently, we only support two mutually exclusive
3224 * mode flags.
3225 */
3226 if (!strcmp(flags, "static"))
3227 mode_flags |= MPOL_F_STATIC_NODES;
3228 else if (!strcmp(flags, "relative"))
3229 mode_flags |= MPOL_F_RELATIVE_NODES;
3230 else
3231 goto out;
3232 }
3233
3234 new = mpol_new(mode, mode_flags, &nodes);
3235 if (IS_ERR(new))
3236 goto out;
3237
3238 /*
3239 * Save nodes for mpol_to_str() to show the tmpfs mount options
3240 * for /proc/mounts, /proc/pid/mounts and /proc/pid/mountinfo.
3241 */
3242 if (mode != MPOL_PREFERRED) {
3243 new->nodes = nodes;
3244 } else if (nodelist) {
3245 nodes_clear(new->nodes);
3246 node_set(first_node(nodes), new->nodes);
3247 } else {
3248 new->mode = MPOL_LOCAL;
3249 }
3250
3251 /*
3252 * Save nodes for contextualization: this will be used to "clone"
3253 * the mempolicy in a specific context [cpuset] at a later time.
3254 */
3255 new->w.user_nodemask = nodes;
3256
3257 err = 0;
3258
3259out:
3260 /* Restore string for error message */
3261 if (nodelist)
3262 *--nodelist = ':';
3263 if (flags)
3264 *--flags = '=';
3265 if (!err)
3266 *mpol = new;
3267 return err;
3268}
3269#endif /* CONFIG_TMPFS */
3270
3271/**
3272 * mpol_to_str - format a mempolicy structure for printing
3273 * @buffer: to contain formatted mempolicy string
3274 * @maxlen: length of @buffer
3275 * @pol: pointer to mempolicy to be formatted
3276 *
3277 * Convert @pol into a string. If @buffer is too short, truncate the string.
3278 * Recommend a @maxlen of at least 32 for the longest mode, "interleave", the
3279 * longest flag, "relative", and to display at least a few node ids.
3280 */
3281void mpol_to_str(char *buffer, int maxlen, struct mempolicy *pol)
3282{
3283 char *p = buffer;
3284 nodemask_t nodes = NODE_MASK_NONE;
3285 unsigned short mode = MPOL_DEFAULT;
3286 unsigned short flags = 0;
3287
3288 if (pol && pol != &default_policy && !(pol->flags & MPOL_F_MORON)) {
3289 mode = pol->mode;
3290 flags = pol->flags;
3291 }
3292
3293 switch (mode) {
3294 case MPOL_DEFAULT:
3295 case MPOL_LOCAL:
3296 break;
3297 case MPOL_PREFERRED:
3298 case MPOL_PREFERRED_MANY:
3299 case MPOL_BIND:
3300 case MPOL_INTERLEAVE:
3301 case MPOL_WEIGHTED_INTERLEAVE:
3302 nodes = pol->nodes;
3303 break;
3304 default:
3305 WARN_ON_ONCE(1);
3306 snprintf(p, maxlen, "unknown");
3307 return;
3308 }
3309
3310 p += snprintf(p, maxlen, "%s", policy_modes[mode]);
3311
3312 if (flags & MPOL_MODE_FLAGS) {
3313 p += snprintf(p, buffer + maxlen - p, "=");
3314
3315 /*
3316 * Currently, the only defined flags are mutually exclusive
3317 */
3318 if (flags & MPOL_F_STATIC_NODES)
3319 p += snprintf(p, buffer + maxlen - p, "static");
3320 else if (flags & MPOL_F_RELATIVE_NODES)
3321 p += snprintf(p, buffer + maxlen - p, "relative");
3322 }
3323
3324 if (!nodes_empty(nodes))
3325 p += scnprintf(p, buffer + maxlen - p, ":%*pbl",
3326 nodemask_pr_args(&nodes));
3327}
3328
3329#ifdef CONFIG_SYSFS
3330struct iw_node_attr {
3331 struct kobj_attribute kobj_attr;
3332 int nid;
3333};
3334
3335static ssize_t node_show(struct kobject *kobj, struct kobj_attribute *attr,
3336 char *buf)
3337{
3338 struct iw_node_attr *node_attr;
3339 u8 weight;
3340
3341 node_attr = container_of(attr, struct iw_node_attr, kobj_attr);
3342 weight = get_il_weight(node_attr->nid);
3343 return sysfs_emit(buf, "%d\n", weight);
3344}
3345
3346static ssize_t node_store(struct kobject *kobj, struct kobj_attribute *attr,
3347 const char *buf, size_t count)
3348{
3349 struct iw_node_attr *node_attr;
3350 u8 *new;
3351 u8 *old;
3352 u8 weight = 0;
3353
3354 node_attr = container_of(attr, struct iw_node_attr, kobj_attr);
3355 if (count == 0 || sysfs_streq(buf, ""))
3356 weight = 0;
3357 else if (kstrtou8(buf, 0, &weight))
3358 return -EINVAL;
3359
3360 new = kzalloc(nr_node_ids, GFP_KERNEL);
3361 if (!new)
3362 return -ENOMEM;
3363
3364 mutex_lock(&iw_table_lock);
3365 old = rcu_dereference_protected(iw_table,
3366 lockdep_is_held(&iw_table_lock));
3367 if (old)
3368 memcpy(new, old, nr_node_ids);
3369 new[node_attr->nid] = weight;
3370 rcu_assign_pointer(iw_table, new);
3371 mutex_unlock(&iw_table_lock);
3372 synchronize_rcu();
3373 kfree(old);
3374 return count;
3375}
3376
3377static struct iw_node_attr **node_attrs;
3378
3379static void sysfs_wi_node_release(struct iw_node_attr *node_attr,
3380 struct kobject *parent)
3381{
3382 if (!node_attr)
3383 return;
3384 sysfs_remove_file(parent, &node_attr->kobj_attr.attr);
3385 kfree(node_attr->kobj_attr.attr.name);
3386 kfree(node_attr);
3387}
3388
3389static void sysfs_wi_release(struct kobject *wi_kobj)
3390{
3391 int i;
3392
3393 for (i = 0; i < nr_node_ids; i++)
3394 sysfs_wi_node_release(node_attrs[i], wi_kobj);
3395 kobject_put(wi_kobj);
3396}
3397
3398static const struct kobj_type wi_ktype = {
3399 .sysfs_ops = &kobj_sysfs_ops,
3400 .release = sysfs_wi_release,
3401};
3402
3403static int add_weight_node(int nid, struct kobject *wi_kobj)
3404{
3405 struct iw_node_attr *node_attr;
3406 char *name;
3407
3408 node_attr = kzalloc(sizeof(*node_attr), GFP_KERNEL);
3409 if (!node_attr)
3410 return -ENOMEM;
3411
3412 name = kasprintf(GFP_KERNEL, "node%d", nid);
3413 if (!name) {
3414 kfree(node_attr);
3415 return -ENOMEM;
3416 }
3417
3418 sysfs_attr_init(&node_attr->kobj_attr.attr);
3419 node_attr->kobj_attr.attr.name = name;
3420 node_attr->kobj_attr.attr.mode = 0644;
3421 node_attr->kobj_attr.show = node_show;
3422 node_attr->kobj_attr.store = node_store;
3423 node_attr->nid = nid;
3424
3425 if (sysfs_create_file(wi_kobj, &node_attr->kobj_attr.attr)) {
3426 kfree(node_attr->kobj_attr.attr.name);
3427 kfree(node_attr);
3428 pr_err("failed to add attribute to weighted_interleave\n");
3429 return -ENOMEM;
3430 }
3431
3432 node_attrs[nid] = node_attr;
3433 return 0;
3434}
3435
3436static int add_weighted_interleave_group(struct kobject *root_kobj)
3437{
3438 struct kobject *wi_kobj;
3439 int nid, err;
3440
3441 wi_kobj = kzalloc(sizeof(struct kobject), GFP_KERNEL);
3442 if (!wi_kobj)
3443 return -ENOMEM;
3444
3445 err = kobject_init_and_add(wi_kobj, &wi_ktype, root_kobj,
3446 "weighted_interleave");
3447 if (err) {
3448 kfree(wi_kobj);
3449 return err;
3450 }
3451
3452 for_each_node_state(nid, N_POSSIBLE) {
3453 err = add_weight_node(nid, wi_kobj);
3454 if (err) {
3455 pr_err("failed to add sysfs [node%d]\n", nid);
3456 break;
3457 }
3458 }
3459 if (err)
3460 kobject_put(wi_kobj);
3461 return 0;
3462}
3463
3464static void mempolicy_kobj_release(struct kobject *kobj)
3465{
3466 u8 *old;
3467
3468 mutex_lock(&iw_table_lock);
3469 old = rcu_dereference_protected(iw_table,
3470 lockdep_is_held(&iw_table_lock));
3471 rcu_assign_pointer(iw_table, NULL);
3472 mutex_unlock(&iw_table_lock);
3473 synchronize_rcu();
3474 kfree(old);
3475 kfree(node_attrs);
3476 kfree(kobj);
3477}
3478
3479static const struct kobj_type mempolicy_ktype = {
3480 .release = mempolicy_kobj_release
3481};
3482
3483static int __init mempolicy_sysfs_init(void)
3484{
3485 int err;
3486 static struct kobject *mempolicy_kobj;
3487
3488 mempolicy_kobj = kzalloc(sizeof(*mempolicy_kobj), GFP_KERNEL);
3489 if (!mempolicy_kobj) {
3490 err = -ENOMEM;
3491 goto err_out;
3492 }
3493
3494 node_attrs = kcalloc(nr_node_ids, sizeof(struct iw_node_attr *),
3495 GFP_KERNEL);
3496 if (!node_attrs) {
3497 err = -ENOMEM;
3498 goto mempol_out;
3499 }
3500
3501 err = kobject_init_and_add(mempolicy_kobj, &mempolicy_ktype, mm_kobj,
3502 "mempolicy");
3503 if (err)
3504 goto node_out;
3505
3506 err = add_weighted_interleave_group(mempolicy_kobj);
3507 if (err) {
3508 pr_err("mempolicy sysfs structure failed to initialize\n");
3509 kobject_put(mempolicy_kobj);
3510 return err;
3511 }
3512
3513 return err;
3514node_out:
3515 kfree(node_attrs);
3516mempol_out:
3517 kfree(mempolicy_kobj);
3518err_out:
3519 pr_err("failed to add mempolicy kobject to the system\n");
3520 return err;
3521}
3522
3523late_initcall(mempolicy_sysfs_init);
3524#endif /* CONFIG_SYSFS */