Loading...
1/*
2 * Simple NUMA memory policy for the Linux kernel.
3 *
4 * Copyright 2003,2004 Andi Kleen, SuSE Labs.
5 * (C) Copyright 2005 Christoph Lameter, Silicon Graphics, Inc.
6 * Subject to the GNU Public License, version 2.
7 *
8 * NUMA policy allows the user to give hints in which node(s) memory should
9 * be allocated.
10 *
11 * Support four policies per VMA and per process:
12 *
13 * The VMA policy has priority over the process policy for a page fault.
14 *
15 * interleave Allocate memory interleaved over a set of nodes,
16 * with normal fallback if it fails.
17 * For VMA based allocations this interleaves based on the
18 * offset into the backing object or offset into the mapping
19 * for anonymous memory. For process policy an process counter
20 * is used.
21 *
22 * bind Only allocate memory on a specific set of nodes,
23 * no fallback.
24 * FIXME: memory is allocated starting with the first node
25 * to the last. It would be better if bind would truly restrict
26 * the allocation to memory nodes instead
27 *
28 * preferred Try a specific node first before normal fallback.
29 * As a special case NUMA_NO_NODE here means do the allocation
30 * on the local CPU. This is normally identical to default,
31 * but useful to set in a VMA when you have a non default
32 * process policy.
33 *
34 * default Allocate on the local node first, or when on a VMA
35 * use the process policy. This is what Linux always did
36 * in a NUMA aware kernel and still does by, ahem, default.
37 *
38 * The process policy is applied for most non interrupt memory allocations
39 * in that process' context. Interrupts ignore the policies and always
40 * try to allocate on the local CPU. The VMA policy is only applied for memory
41 * allocations for a VMA in the VM.
42 *
43 * Currently there are a few corner cases in swapping where the policy
44 * is not applied, but the majority should be handled. When process policy
45 * is used it is not remembered over swap outs/swap ins.
46 *
47 * Only the highest zone in the zone hierarchy gets policied. Allocations
48 * requesting a lower zone just use default policy. This implies that
49 * on systems with highmem kernel lowmem allocation don't get policied.
50 * Same with GFP_DMA allocations.
51 *
52 * For shmfs/tmpfs/hugetlbfs shared memory the policy is shared between
53 * all users and remembered even when nobody has memory mapped.
54 */
55
56/* Notebook:
57 fix mmap readahead to honour policy and enable policy for any page cache
58 object
59 statistics for bigpages
60 global policy for page cache? currently it uses process policy. Requires
61 first item above.
62 handle mremap for shared memory (currently ignored for the policy)
63 grows down?
64 make bind policy root only? It can trigger oom much faster and the
65 kernel is not always grateful with that.
66*/
67
68#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
69
70#include <linux/mempolicy.h>
71#include <linux/mm.h>
72#include <linux/highmem.h>
73#include <linux/hugetlb.h>
74#include <linux/kernel.h>
75#include <linux/sched.h>
76#include <linux/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 = tmp;
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) &&
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 void 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
431static int queue_pages_pmd(pmd_t *pmd, spinlock_t *ptl, unsigned long addr,
432 unsigned long end, struct mm_walk *walk)
433{
434 int ret = 0;
435 struct page *page;
436 struct queue_pages *qp = walk->private;
437 unsigned long flags;
438
439 if (unlikely(is_pmd_migration_entry(*pmd))) {
440 ret = 1;
441 goto unlock;
442 }
443 page = pmd_page(*pmd);
444 if (is_huge_zero_page(page)) {
445 spin_unlock(ptl);
446 __split_huge_pmd(walk->vma, pmd, addr, false, NULL);
447 goto out;
448 }
449 if (!queue_pages_required(page, qp)) {
450 ret = 1;
451 goto unlock;
452 }
453
454 ret = 1;
455 flags = qp->flags;
456 /* go to thp migration */
457 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))
458 migrate_page_add(page, qp->pagelist, flags);
459unlock:
460 spin_unlock(ptl);
461out:
462 return ret;
463}
464
465/*
466 * Scan through pages checking if pages follow certain conditions,
467 * and move them to the pagelist if they do.
468 */
469static int queue_pages_pte_range(pmd_t *pmd, unsigned long addr,
470 unsigned long end, struct mm_walk *walk)
471{
472 struct vm_area_struct *vma = walk->vma;
473 struct page *page;
474 struct queue_pages *qp = walk->private;
475 unsigned long flags = qp->flags;
476 int ret;
477 pte_t *pte;
478 spinlock_t *ptl;
479
480 ptl = pmd_trans_huge_lock(pmd, vma);
481 if (ptl) {
482 ret = queue_pages_pmd(pmd, ptl, addr, end, walk);
483 if (ret)
484 return 0;
485 }
486
487 if (pmd_trans_unstable(pmd))
488 return 0;
489
490 pte = pte_offset_map_lock(walk->mm, pmd, addr, &ptl);
491 for (; addr != end; pte++, addr += PAGE_SIZE) {
492 if (!pte_present(*pte))
493 continue;
494 page = vm_normal_page(vma, addr, *pte);
495 if (!page)
496 continue;
497 /*
498 * vm_normal_page() filters out zero pages, but there might
499 * still be PageReserved pages to skip, perhaps in a VDSO.
500 */
501 if (PageReserved(page))
502 continue;
503 if (!queue_pages_required(page, qp))
504 continue;
505 migrate_page_add(page, qp->pagelist, flags);
506 }
507 pte_unmap_unlock(pte - 1, ptl);
508 cond_resched();
509 return 0;
510}
511
512static int queue_pages_hugetlb(pte_t *pte, unsigned long hmask,
513 unsigned long addr, unsigned long end,
514 struct mm_walk *walk)
515{
516#ifdef CONFIG_HUGETLB_PAGE
517 struct queue_pages *qp = walk->private;
518 unsigned long flags = qp->flags;
519 struct page *page;
520 spinlock_t *ptl;
521 pte_t entry;
522
523 ptl = huge_pte_lock(hstate_vma(walk->vma), walk->mm, pte);
524 entry = huge_ptep_get(pte);
525 if (!pte_present(entry))
526 goto unlock;
527 page = pte_page(entry);
528 if (!queue_pages_required(page, qp))
529 goto unlock;
530 /* With MPOL_MF_MOVE, we migrate only unshared hugepage. */
531 if (flags & (MPOL_MF_MOVE_ALL) ||
532 (flags & MPOL_MF_MOVE && page_mapcount(page) == 1))
533 isolate_huge_page(page, qp->pagelist);
534unlock:
535 spin_unlock(ptl);
536#else
537 BUG();
538#endif
539 return 0;
540}
541
542#ifdef CONFIG_NUMA_BALANCING
543/*
544 * This is used to mark a range of virtual addresses to be inaccessible.
545 * These are later cleared by a NUMA hinting fault. Depending on these
546 * faults, pages may be migrated for better NUMA placement.
547 *
548 * This is assuming that NUMA faults are handled using PROT_NONE. If
549 * an architecture makes a different choice, it will need further
550 * changes to the core.
551 */
552unsigned long change_prot_numa(struct vm_area_struct *vma,
553 unsigned long addr, unsigned long end)
554{
555 int nr_updated;
556
557 nr_updated = change_protection(vma, addr, end, PAGE_NONE, 0, 1);
558 if (nr_updated)
559 count_vm_numa_events(NUMA_PTE_UPDATES, nr_updated);
560
561 return nr_updated;
562}
563#else
564static unsigned long change_prot_numa(struct vm_area_struct *vma,
565 unsigned long addr, unsigned long end)
566{
567 return 0;
568}
569#endif /* CONFIG_NUMA_BALANCING */
570
571static int queue_pages_test_walk(unsigned long start, unsigned long end,
572 struct mm_walk *walk)
573{
574 struct vm_area_struct *vma = walk->vma;
575 struct queue_pages *qp = walk->private;
576 unsigned long endvma = vma->vm_end;
577 unsigned long flags = qp->flags;
578
579 if (!vma_migratable(vma))
580 return 1;
581
582 if (endvma > end)
583 endvma = end;
584 if (vma->vm_start > start)
585 start = vma->vm_start;
586
587 if (!(flags & MPOL_MF_DISCONTIG_OK)) {
588 if (!vma->vm_next && vma->vm_end < end)
589 return -EFAULT;
590 if (qp->prev && qp->prev->vm_end < vma->vm_start)
591 return -EFAULT;
592 }
593
594 qp->prev = vma;
595
596 if (flags & MPOL_MF_LAZY) {
597 /* Similar to task_numa_work, skip inaccessible VMAs */
598 if (!is_vm_hugetlb_page(vma) &&
599 (vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)) &&
600 !(vma->vm_flags & VM_MIXEDMAP))
601 change_prot_numa(vma, start, endvma);
602 return 1;
603 }
604
605 /* queue pages from current vma */
606 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))
607 return 0;
608 return 1;
609}
610
611/*
612 * Walk through page tables and collect pages to be migrated.
613 *
614 * If pages found in a given range are on a set of nodes (determined by
615 * @nodes and @flags,) it's isolated and queued to the pagelist which is
616 * passed via @private.)
617 */
618static int
619queue_pages_range(struct mm_struct *mm, unsigned long start, unsigned long end,
620 nodemask_t *nodes, unsigned long flags,
621 struct list_head *pagelist)
622{
623 struct queue_pages qp = {
624 .pagelist = pagelist,
625 .flags = flags,
626 .nmask = nodes,
627 .prev = NULL,
628 };
629 struct mm_walk queue_pages_walk = {
630 .hugetlb_entry = queue_pages_hugetlb,
631 .pmd_entry = queue_pages_pte_range,
632 .test_walk = queue_pages_test_walk,
633 .mm = mm,
634 .private = &qp,
635 };
636
637 return walk_page_range(start, end, &queue_pages_walk);
638}
639
640/*
641 * Apply policy to a single VMA
642 * This must be called with the mmap_sem held for writing.
643 */
644static int vma_replace_policy(struct vm_area_struct *vma,
645 struct mempolicy *pol)
646{
647 int err;
648 struct mempolicy *old;
649 struct mempolicy *new;
650
651 pr_debug("vma %lx-%lx/%lx vm_ops %p vm_file %p set_policy %p\n",
652 vma->vm_start, vma->vm_end, vma->vm_pgoff,
653 vma->vm_ops, vma->vm_file,
654 vma->vm_ops ? vma->vm_ops->set_policy : NULL);
655
656 new = mpol_dup(pol);
657 if (IS_ERR(new))
658 return PTR_ERR(new);
659
660 if (vma->vm_ops && vma->vm_ops->set_policy) {
661 err = vma->vm_ops->set_policy(vma, new);
662 if (err)
663 goto err_out;
664 }
665
666 old = vma->vm_policy;
667 vma->vm_policy = new; /* protected by mmap_sem */
668 mpol_put(old);
669
670 return 0;
671 err_out:
672 mpol_put(new);
673 return err;
674}
675
676/* Step 2: apply policy to a range and do splits. */
677static int mbind_range(struct mm_struct *mm, unsigned long start,
678 unsigned long end, struct mempolicy *new_pol)
679{
680 struct vm_area_struct *next;
681 struct vm_area_struct *prev;
682 struct vm_area_struct *vma;
683 int err = 0;
684 pgoff_t pgoff;
685 unsigned long vmstart;
686 unsigned long vmend;
687
688 vma = find_vma(mm, start);
689 if (!vma || vma->vm_start > start)
690 return -EFAULT;
691
692 prev = vma->vm_prev;
693 if (start > vma->vm_start)
694 prev = vma;
695
696 for (; vma && vma->vm_start < end; prev = vma, vma = next) {
697 next = vma->vm_next;
698 vmstart = max(start, vma->vm_start);
699 vmend = min(end, vma->vm_end);
700
701 if (mpol_equal(vma_policy(vma), new_pol))
702 continue;
703
704 pgoff = vma->vm_pgoff +
705 ((vmstart - vma->vm_start) >> PAGE_SHIFT);
706 prev = vma_merge(mm, prev, vmstart, vmend, vma->vm_flags,
707 vma->anon_vma, vma->vm_file, pgoff,
708 new_pol, vma->vm_userfaultfd_ctx);
709 if (prev) {
710 vma = prev;
711 next = vma->vm_next;
712 if (mpol_equal(vma_policy(vma), new_pol))
713 continue;
714 /* vma_merge() joined vma && vma->next, case 8 */
715 goto replace;
716 }
717 if (vma->vm_start != vmstart) {
718 err = split_vma(vma->vm_mm, vma, vmstart, 1);
719 if (err)
720 goto out;
721 }
722 if (vma->vm_end != vmend) {
723 err = split_vma(vma->vm_mm, vma, vmend, 0);
724 if (err)
725 goto out;
726 }
727 replace:
728 err = vma_replace_policy(vma, new_pol);
729 if (err)
730 goto out;
731 }
732
733 out:
734 return err;
735}
736
737/* Set the process memory policy */
738static long do_set_mempolicy(unsigned short mode, unsigned short flags,
739 nodemask_t *nodes)
740{
741 struct mempolicy *new, *old;
742 NODEMASK_SCRATCH(scratch);
743 int ret;
744
745 if (!scratch)
746 return -ENOMEM;
747
748 new = mpol_new(mode, flags, nodes);
749 if (IS_ERR(new)) {
750 ret = PTR_ERR(new);
751 goto out;
752 }
753
754 task_lock(current);
755 ret = mpol_set_nodemask(new, nodes, scratch);
756 if (ret) {
757 task_unlock(current);
758 mpol_put(new);
759 goto out;
760 }
761 old = current->mempolicy;
762 current->mempolicy = new;
763 if (new && new->mode == MPOL_INTERLEAVE)
764 current->il_prev = MAX_NUMNODES-1;
765 task_unlock(current);
766 mpol_put(old);
767 ret = 0;
768out:
769 NODEMASK_SCRATCH_FREE(scratch);
770 return ret;
771}
772
773/*
774 * Return nodemask for policy for get_mempolicy() query
775 *
776 * Called with task's alloc_lock held
777 */
778static void get_policy_nodemask(struct mempolicy *p, nodemask_t *nodes)
779{
780 nodes_clear(*nodes);
781 if (p == &default_policy)
782 return;
783
784 switch (p->mode) {
785 case MPOL_BIND:
786 /* Fall through */
787 case MPOL_INTERLEAVE:
788 *nodes = p->v.nodes;
789 break;
790 case MPOL_PREFERRED:
791 if (!(p->flags & MPOL_F_LOCAL))
792 node_set(p->v.preferred_node, *nodes);
793 /* else return empty node mask for local allocation */
794 break;
795 default:
796 BUG();
797 }
798}
799
800static int lookup_node(unsigned long addr)
801{
802 struct page *p;
803 int err;
804
805 err = get_user_pages(addr & PAGE_MASK, 1, 0, &p, NULL);
806 if (err >= 0) {
807 err = page_to_nid(p);
808 put_page(p);
809 }
810 return err;
811}
812
813/* Retrieve NUMA policy */
814static long do_get_mempolicy(int *policy, nodemask_t *nmask,
815 unsigned long addr, unsigned long flags)
816{
817 int err;
818 struct mm_struct *mm = current->mm;
819 struct vm_area_struct *vma = NULL;
820 struct mempolicy *pol = current->mempolicy;
821
822 if (flags &
823 ~(unsigned long)(MPOL_F_NODE|MPOL_F_ADDR|MPOL_F_MEMS_ALLOWED))
824 return -EINVAL;
825
826 if (flags & MPOL_F_MEMS_ALLOWED) {
827 if (flags & (MPOL_F_NODE|MPOL_F_ADDR))
828 return -EINVAL;
829 *policy = 0; /* just so it's initialized */
830 task_lock(current);
831 *nmask = cpuset_current_mems_allowed;
832 task_unlock(current);
833 return 0;
834 }
835
836 if (flags & MPOL_F_ADDR) {
837 /*
838 * Do NOT fall back to task policy if the
839 * vma/shared policy at addr is NULL. We
840 * want to return MPOL_DEFAULT in this case.
841 */
842 down_read(&mm->mmap_sem);
843 vma = find_vma_intersection(mm, addr, addr+1);
844 if (!vma) {
845 up_read(&mm->mmap_sem);
846 return -EFAULT;
847 }
848 if (vma->vm_ops && vma->vm_ops->get_policy)
849 pol = vma->vm_ops->get_policy(vma, addr);
850 else
851 pol = vma->vm_policy;
852 } else if (addr)
853 return -EINVAL;
854
855 if (!pol)
856 pol = &default_policy; /* indicates default behavior */
857
858 if (flags & MPOL_F_NODE) {
859 if (flags & MPOL_F_ADDR) {
860 err = lookup_node(addr);
861 if (err < 0)
862 goto out;
863 *policy = err;
864 } else if (pol == current->mempolicy &&
865 pol->mode == MPOL_INTERLEAVE) {
866 *policy = next_node_in(current->il_prev, pol->v.nodes);
867 } else {
868 err = -EINVAL;
869 goto out;
870 }
871 } else {
872 *policy = pol == &default_policy ? MPOL_DEFAULT :
873 pol->mode;
874 /*
875 * Internal mempolicy flags must be masked off before exposing
876 * the policy to userspace.
877 */
878 *policy |= (pol->flags & MPOL_MODE_FLAGS);
879 }
880
881 err = 0;
882 if (nmask) {
883 if (mpol_store_user_nodemask(pol)) {
884 *nmask = pol->w.user_nodemask;
885 } else {
886 task_lock(current);
887 get_policy_nodemask(pol, nmask);
888 task_unlock(current);
889 }
890 }
891
892 out:
893 mpol_cond_put(pol);
894 if (vma)
895 up_read(¤t->mm->mmap_sem);
896 return err;
897}
898
899#ifdef CONFIG_MIGRATION
900/*
901 * page migration, thp tail pages can be passed.
902 */
903static void migrate_page_add(struct page *page, struct list_head *pagelist,
904 unsigned long flags)
905{
906 struct page *head = compound_head(page);
907 /*
908 * Avoid migrating a page that is shared with others.
909 */
910 if ((flags & MPOL_MF_MOVE_ALL) || page_mapcount(head) == 1) {
911 if (!isolate_lru_page(head)) {
912 list_add_tail(&head->lru, pagelist);
913 mod_node_page_state(page_pgdat(head),
914 NR_ISOLATED_ANON + page_is_file_cache(head),
915 hpage_nr_pages(head));
916 }
917 }
918}
919
920/* page allocation callback for NUMA node migration */
921struct page *alloc_new_node_page(struct page *page, unsigned long node)
922{
923 if (PageHuge(page))
924 return alloc_huge_page_node(page_hstate(compound_head(page)),
925 node);
926 else if (PageTransHuge(page)) {
927 struct page *thp;
928
929 thp = alloc_pages_node(node,
930 (GFP_TRANSHUGE | __GFP_THISNODE),
931 HPAGE_PMD_ORDER);
932 if (!thp)
933 return NULL;
934 prep_transhuge_page(thp);
935 return thp;
936 } else
937 return __alloc_pages_node(node, GFP_HIGHUSER_MOVABLE |
938 __GFP_THISNODE, 0);
939}
940
941/*
942 * Migrate pages from one node to a target node.
943 * Returns error or the number of pages not migrated.
944 */
945static int migrate_to_node(struct mm_struct *mm, int source, int dest,
946 int flags)
947{
948 nodemask_t nmask;
949 LIST_HEAD(pagelist);
950 int err = 0;
951
952 nodes_clear(nmask);
953 node_set(source, nmask);
954
955 /*
956 * This does not "check" the range but isolates all pages that
957 * need migration. Between passing in the full user address
958 * space range and MPOL_MF_DISCONTIG_OK, this call can not fail.
959 */
960 VM_BUG_ON(!(flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)));
961 queue_pages_range(mm, mm->mmap->vm_start, mm->task_size, &nmask,
962 flags | MPOL_MF_DISCONTIG_OK, &pagelist);
963
964 if (!list_empty(&pagelist)) {
965 err = migrate_pages(&pagelist, alloc_new_node_page, NULL, dest,
966 MIGRATE_SYNC, MR_SYSCALL);
967 if (err)
968 putback_movable_pages(&pagelist);
969 }
970
971 return err;
972}
973
974/*
975 * Move pages between the two nodesets so as to preserve the physical
976 * layout as much as possible.
977 *
978 * Returns the number of page that could not be moved.
979 */
980int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
981 const nodemask_t *to, int flags)
982{
983 int busy = 0;
984 int err;
985 nodemask_t tmp;
986
987 err = migrate_prep();
988 if (err)
989 return err;
990
991 down_read(&mm->mmap_sem);
992
993 /*
994 * Find a 'source' bit set in 'tmp' whose corresponding 'dest'
995 * bit in 'to' is not also set in 'tmp'. Clear the found 'source'
996 * bit in 'tmp', and return that <source, dest> pair for migration.
997 * The pair of nodemasks 'to' and 'from' define the map.
998 *
999 * If no pair of bits is found that way, fallback to picking some
1000 * pair of 'source' and 'dest' bits that are not the same. If the
1001 * 'source' and 'dest' bits are the same, this represents a node
1002 * that will be migrating to itself, so no pages need move.
1003 *
1004 * If no bits are left in 'tmp', or if all remaining bits left
1005 * in 'tmp' correspond to the same bit in 'to', return false
1006 * (nothing left to migrate).
1007 *
1008 * This lets us pick a pair of nodes to migrate between, such that
1009 * if possible the dest node is not already occupied by some other
1010 * source node, minimizing the risk of overloading the memory on a
1011 * node that would happen if we migrated incoming memory to a node
1012 * before migrating outgoing memory source that same node.
1013 *
1014 * A single scan of tmp is sufficient. As we go, we remember the
1015 * most recent <s, d> pair that moved (s != d). If we find a pair
1016 * that not only moved, but what's better, moved to an empty slot
1017 * (d is not set in tmp), then we break out then, with that pair.
1018 * Otherwise when we finish scanning from_tmp, we at least have the
1019 * most recent <s, d> pair that moved. If we get all the way through
1020 * the scan of tmp without finding any node that moved, much less
1021 * moved to an empty node, then there is nothing left worth migrating.
1022 */
1023
1024 tmp = *from;
1025 while (!nodes_empty(tmp)) {
1026 int s,d;
1027 int source = NUMA_NO_NODE;
1028 int dest = 0;
1029
1030 for_each_node_mask(s, tmp) {
1031
1032 /*
1033 * do_migrate_pages() tries to maintain the relative
1034 * node relationship of the pages established between
1035 * threads and memory areas.
1036 *
1037 * However if the number of source nodes is not equal to
1038 * the number of destination nodes we can not preserve
1039 * this node relative relationship. In that case, skip
1040 * copying memory from a node that is in the destination
1041 * mask.
1042 *
1043 * Example: [2,3,4] -> [3,4,5] moves everything.
1044 * [0-7] - > [3,4,5] moves only 0,1,2,6,7.
1045 */
1046
1047 if ((nodes_weight(*from) != nodes_weight(*to)) &&
1048 (node_isset(s, *to)))
1049 continue;
1050
1051 d = node_remap(s, *from, *to);
1052 if (s == d)
1053 continue;
1054
1055 source = s; /* Node moved. Memorize */
1056 dest = d;
1057
1058 /* dest not in remaining from nodes? */
1059 if (!node_isset(dest, tmp))
1060 break;
1061 }
1062 if (source == NUMA_NO_NODE)
1063 break;
1064
1065 node_clear(source, tmp);
1066 err = migrate_to_node(mm, source, dest, flags);
1067 if (err > 0)
1068 busy += err;
1069 if (err < 0)
1070 break;
1071 }
1072 up_read(&mm->mmap_sem);
1073 if (err < 0)
1074 return err;
1075 return busy;
1076
1077}
1078
1079/*
1080 * Allocate a new page for page migration based on vma policy.
1081 * Start by assuming the page is mapped by the same vma as contains @start.
1082 * Search forward from there, if not. N.B., this assumes that the
1083 * list of pages handed to migrate_pages()--which is how we get here--
1084 * is in virtual address order.
1085 */
1086static struct page *new_page(struct page *page, unsigned long start)
1087{
1088 struct vm_area_struct *vma;
1089 unsigned long uninitialized_var(address);
1090
1091 vma = find_vma(current->mm, start);
1092 while (vma) {
1093 address = page_address_in_vma(page, vma);
1094 if (address != -EFAULT)
1095 break;
1096 vma = vma->vm_next;
1097 }
1098
1099 if (PageHuge(page)) {
1100 return alloc_huge_page_vma(page_hstate(compound_head(page)),
1101 vma, address);
1102 } else if (PageTransHuge(page)) {
1103 struct page *thp;
1104
1105 thp = alloc_hugepage_vma(GFP_TRANSHUGE, vma, address,
1106 HPAGE_PMD_ORDER);
1107 if (!thp)
1108 return NULL;
1109 prep_transhuge_page(thp);
1110 return thp;
1111 }
1112 /*
1113 * if !vma, alloc_page_vma() will use task or system default policy
1114 */
1115 return alloc_page_vma(GFP_HIGHUSER_MOVABLE | __GFP_RETRY_MAYFAIL,
1116 vma, address);
1117}
1118#else
1119
1120static void migrate_page_add(struct page *page, struct list_head *pagelist,
1121 unsigned long flags)
1122{
1123}
1124
1125int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
1126 const nodemask_t *to, int flags)
1127{
1128 return -ENOSYS;
1129}
1130
1131static struct page *new_page(struct page *page, unsigned long start)
1132{
1133 return NULL;
1134}
1135#endif
1136
1137static long do_mbind(unsigned long start, unsigned long len,
1138 unsigned short mode, unsigned short mode_flags,
1139 nodemask_t *nmask, unsigned long flags)
1140{
1141 struct mm_struct *mm = current->mm;
1142 struct mempolicy *new;
1143 unsigned long end;
1144 int err;
1145 LIST_HEAD(pagelist);
1146
1147 if (flags & ~(unsigned long)MPOL_MF_VALID)
1148 return -EINVAL;
1149 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1150 return -EPERM;
1151
1152 if (start & ~PAGE_MASK)
1153 return -EINVAL;
1154
1155 if (mode == MPOL_DEFAULT)
1156 flags &= ~MPOL_MF_STRICT;
1157
1158 len = (len + PAGE_SIZE - 1) & PAGE_MASK;
1159 end = start + len;
1160
1161 if (end < start)
1162 return -EINVAL;
1163 if (end == start)
1164 return 0;
1165
1166 new = mpol_new(mode, mode_flags, nmask);
1167 if (IS_ERR(new))
1168 return PTR_ERR(new);
1169
1170 if (flags & MPOL_MF_LAZY)
1171 new->flags |= MPOL_F_MOF;
1172
1173 /*
1174 * If we are using the default policy then operation
1175 * on discontinuous address spaces is okay after all
1176 */
1177 if (!new)
1178 flags |= MPOL_MF_DISCONTIG_OK;
1179
1180 pr_debug("mbind %lx-%lx mode:%d flags:%d nodes:%lx\n",
1181 start, start + len, mode, mode_flags,
1182 nmask ? nodes_addr(*nmask)[0] : NUMA_NO_NODE);
1183
1184 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) {
1185
1186 err = migrate_prep();
1187 if (err)
1188 goto mpol_out;
1189 }
1190 {
1191 NODEMASK_SCRATCH(scratch);
1192 if (scratch) {
1193 down_write(&mm->mmap_sem);
1194 task_lock(current);
1195 err = mpol_set_nodemask(new, nmask, scratch);
1196 task_unlock(current);
1197 if (err)
1198 up_write(&mm->mmap_sem);
1199 } else
1200 err = -ENOMEM;
1201 NODEMASK_SCRATCH_FREE(scratch);
1202 }
1203 if (err)
1204 goto mpol_out;
1205
1206 err = queue_pages_range(mm, start, end, nmask,
1207 flags | MPOL_MF_INVERT, &pagelist);
1208 if (!err)
1209 err = mbind_range(mm, start, end, new);
1210
1211 if (!err) {
1212 int nr_failed = 0;
1213
1214 if (!list_empty(&pagelist)) {
1215 WARN_ON_ONCE(flags & MPOL_MF_LAZY);
1216 nr_failed = migrate_pages(&pagelist, new_page, NULL,
1217 start, MIGRATE_SYNC, MR_MEMPOLICY_MBIND);
1218 if (nr_failed)
1219 putback_movable_pages(&pagelist);
1220 }
1221
1222 if (nr_failed && (flags & MPOL_MF_STRICT))
1223 err = -EIO;
1224 } else
1225 putback_movable_pages(&pagelist);
1226
1227 up_write(&mm->mmap_sem);
1228 mpol_out:
1229 mpol_put(new);
1230 return err;
1231}
1232
1233/*
1234 * User space interface with variable sized bitmaps for nodelists.
1235 */
1236
1237/* Copy a node mask from user space. */
1238static int get_nodes(nodemask_t *nodes, const unsigned long __user *nmask,
1239 unsigned long maxnode)
1240{
1241 unsigned long k;
1242 unsigned long t;
1243 unsigned long nlongs;
1244 unsigned long endmask;
1245
1246 --maxnode;
1247 nodes_clear(*nodes);
1248 if (maxnode == 0 || !nmask)
1249 return 0;
1250 if (maxnode > PAGE_SIZE*BITS_PER_BYTE)
1251 return -EINVAL;
1252
1253 nlongs = BITS_TO_LONGS(maxnode);
1254 if ((maxnode % BITS_PER_LONG) == 0)
1255 endmask = ~0UL;
1256 else
1257 endmask = (1UL << (maxnode % BITS_PER_LONG)) - 1;
1258
1259 /*
1260 * When the user specified more nodes than supported just check
1261 * if the non supported part is all zero.
1262 *
1263 * If maxnode have more longs than MAX_NUMNODES, check
1264 * the bits in that area first. And then go through to
1265 * check the rest bits which equal or bigger than MAX_NUMNODES.
1266 * Otherwise, just check bits [MAX_NUMNODES, maxnode).
1267 */
1268 if (nlongs > BITS_TO_LONGS(MAX_NUMNODES)) {
1269 for (k = BITS_TO_LONGS(MAX_NUMNODES); k < nlongs; k++) {
1270 if (get_user(t, nmask + k))
1271 return -EFAULT;
1272 if (k == nlongs - 1) {
1273 if (t & endmask)
1274 return -EINVAL;
1275 } else if (t)
1276 return -EINVAL;
1277 }
1278 nlongs = BITS_TO_LONGS(MAX_NUMNODES);
1279 endmask = ~0UL;
1280 }
1281
1282 if (maxnode > MAX_NUMNODES && MAX_NUMNODES % BITS_PER_LONG != 0) {
1283 unsigned long valid_mask = endmask;
1284
1285 valid_mask &= ~((1UL << (MAX_NUMNODES % BITS_PER_LONG)) - 1);
1286 if (get_user(t, nmask + nlongs - 1))
1287 return -EFAULT;
1288 if (t & valid_mask)
1289 return -EINVAL;
1290 }
1291
1292 if (copy_from_user(nodes_addr(*nodes), nmask, nlongs*sizeof(unsigned long)))
1293 return -EFAULT;
1294 nodes_addr(*nodes)[nlongs-1] &= endmask;
1295 return 0;
1296}
1297
1298/* Copy a kernel node mask to user space */
1299static int copy_nodes_to_user(unsigned long __user *mask, unsigned long maxnode,
1300 nodemask_t *nodes)
1301{
1302 unsigned long copy = ALIGN(maxnode-1, 64) / 8;
1303 const int nbytes = BITS_TO_LONGS(MAX_NUMNODES) * sizeof(long);
1304
1305 if (copy > nbytes) {
1306 if (copy > PAGE_SIZE)
1307 return -EINVAL;
1308 if (clear_user((char __user *)mask + nbytes, copy - nbytes))
1309 return -EFAULT;
1310 copy = nbytes;
1311 }
1312 return copy_to_user(mask, nodes_addr(*nodes), copy) ? -EFAULT : 0;
1313}
1314
1315static long kernel_mbind(unsigned long start, unsigned long len,
1316 unsigned long mode, const unsigned long __user *nmask,
1317 unsigned long maxnode, unsigned int flags)
1318{
1319 nodemask_t nodes;
1320 int err;
1321 unsigned short mode_flags;
1322
1323 mode_flags = mode & MPOL_MODE_FLAGS;
1324 mode &= ~MPOL_MODE_FLAGS;
1325 if (mode >= MPOL_MAX)
1326 return -EINVAL;
1327 if ((mode_flags & MPOL_F_STATIC_NODES) &&
1328 (mode_flags & MPOL_F_RELATIVE_NODES))
1329 return -EINVAL;
1330 err = get_nodes(&nodes, nmask, maxnode);
1331 if (err)
1332 return err;
1333 return do_mbind(start, len, mode, mode_flags, &nodes, flags);
1334}
1335
1336SYSCALL_DEFINE6(mbind, unsigned long, start, unsigned long, len,
1337 unsigned long, mode, const unsigned long __user *, nmask,
1338 unsigned long, maxnode, unsigned int, flags)
1339{
1340 return kernel_mbind(start, len, mode, nmask, maxnode, flags);
1341}
1342
1343/* Set the process memory policy */
1344static long kernel_set_mempolicy(int mode, const unsigned long __user *nmask,
1345 unsigned long maxnode)
1346{
1347 int err;
1348 nodemask_t nodes;
1349 unsigned short flags;
1350
1351 flags = mode & MPOL_MODE_FLAGS;
1352 mode &= ~MPOL_MODE_FLAGS;
1353 if ((unsigned int)mode >= MPOL_MAX)
1354 return -EINVAL;
1355 if ((flags & MPOL_F_STATIC_NODES) && (flags & MPOL_F_RELATIVE_NODES))
1356 return -EINVAL;
1357 err = get_nodes(&nodes, nmask, maxnode);
1358 if (err)
1359 return err;
1360 return do_set_mempolicy(mode, flags, &nodes);
1361}
1362
1363SYSCALL_DEFINE3(set_mempolicy, int, mode, const unsigned long __user *, nmask,
1364 unsigned long, maxnode)
1365{
1366 return kernel_set_mempolicy(mode, nmask, maxnode);
1367}
1368
1369static int kernel_migrate_pages(pid_t pid, unsigned long maxnode,
1370 const unsigned long __user *old_nodes,
1371 const unsigned long __user *new_nodes)
1372{
1373 struct mm_struct *mm = NULL;
1374 struct task_struct *task;
1375 nodemask_t task_nodes;
1376 int err;
1377 nodemask_t *old;
1378 nodemask_t *new;
1379 NODEMASK_SCRATCH(scratch);
1380
1381 if (!scratch)
1382 return -ENOMEM;
1383
1384 old = &scratch->mask1;
1385 new = &scratch->mask2;
1386
1387 err = get_nodes(old, old_nodes, maxnode);
1388 if (err)
1389 goto out;
1390
1391 err = get_nodes(new, new_nodes, maxnode);
1392 if (err)
1393 goto out;
1394
1395 /* Find the mm_struct */
1396 rcu_read_lock();
1397 task = pid ? find_task_by_vpid(pid) : current;
1398 if (!task) {
1399 rcu_read_unlock();
1400 err = -ESRCH;
1401 goto out;
1402 }
1403 get_task_struct(task);
1404
1405 err = -EINVAL;
1406
1407 /*
1408 * Check if this process has the right to modify the specified process.
1409 * Use the regular "ptrace_may_access()" checks.
1410 */
1411 if (!ptrace_may_access(task, PTRACE_MODE_READ_REALCREDS)) {
1412 rcu_read_unlock();
1413 err = -EPERM;
1414 goto out_put;
1415 }
1416 rcu_read_unlock();
1417
1418 task_nodes = cpuset_mems_allowed(task);
1419 /* Is the user allowed to access the target nodes? */
1420 if (!nodes_subset(*new, task_nodes) && !capable(CAP_SYS_NICE)) {
1421 err = -EPERM;
1422 goto out_put;
1423 }
1424
1425 task_nodes = cpuset_mems_allowed(current);
1426 nodes_and(*new, *new, task_nodes);
1427 if (nodes_empty(*new))
1428 goto out_put;
1429
1430 nodes_and(*new, *new, node_states[N_MEMORY]);
1431 if (nodes_empty(*new))
1432 goto out_put;
1433
1434 err = security_task_movememory(task);
1435 if (err)
1436 goto out_put;
1437
1438 mm = get_task_mm(task);
1439 put_task_struct(task);
1440
1441 if (!mm) {
1442 err = -EINVAL;
1443 goto out;
1444 }
1445
1446 err = do_migrate_pages(mm, old, new,
1447 capable(CAP_SYS_NICE) ? MPOL_MF_MOVE_ALL : MPOL_MF_MOVE);
1448
1449 mmput(mm);
1450out:
1451 NODEMASK_SCRATCH_FREE(scratch);
1452
1453 return err;
1454
1455out_put:
1456 put_task_struct(task);
1457 goto out;
1458
1459}
1460
1461SYSCALL_DEFINE4(migrate_pages, pid_t, pid, unsigned long, maxnode,
1462 const unsigned long __user *, old_nodes,
1463 const unsigned long __user *, new_nodes)
1464{
1465 return kernel_migrate_pages(pid, maxnode, old_nodes, new_nodes);
1466}
1467
1468
1469/* Retrieve NUMA policy */
1470static int kernel_get_mempolicy(int __user *policy,
1471 unsigned long __user *nmask,
1472 unsigned long maxnode,
1473 unsigned long addr,
1474 unsigned long flags)
1475{
1476 int err;
1477 int uninitialized_var(pval);
1478 nodemask_t nodes;
1479
1480 if (nmask != NULL && maxnode < MAX_NUMNODES)
1481 return -EINVAL;
1482
1483 err = do_get_mempolicy(&pval, &nodes, addr, flags);
1484
1485 if (err)
1486 return err;
1487
1488 if (policy && put_user(pval, policy))
1489 return -EFAULT;
1490
1491 if (nmask)
1492 err = copy_nodes_to_user(nmask, maxnode, &nodes);
1493
1494 return err;
1495}
1496
1497SYSCALL_DEFINE5(get_mempolicy, int __user *, policy,
1498 unsigned long __user *, nmask, unsigned long, maxnode,
1499 unsigned long, addr, unsigned long, flags)
1500{
1501 return kernel_get_mempolicy(policy, nmask, maxnode, addr, flags);
1502}
1503
1504#ifdef CONFIG_COMPAT
1505
1506COMPAT_SYSCALL_DEFINE5(get_mempolicy, int __user *, policy,
1507 compat_ulong_t __user *, nmask,
1508 compat_ulong_t, maxnode,
1509 compat_ulong_t, addr, compat_ulong_t, flags)
1510{
1511 long err;
1512 unsigned long __user *nm = NULL;
1513 unsigned long nr_bits, alloc_size;
1514 DECLARE_BITMAP(bm, MAX_NUMNODES);
1515
1516 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1517 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1518
1519 if (nmask)
1520 nm = compat_alloc_user_space(alloc_size);
1521
1522 err = kernel_get_mempolicy(policy, nm, nr_bits+1, addr, flags);
1523
1524 if (!err && nmask) {
1525 unsigned long copy_size;
1526 copy_size = min_t(unsigned long, sizeof(bm), alloc_size);
1527 err = copy_from_user(bm, nm, copy_size);
1528 /* ensure entire bitmap is zeroed */
1529 err |= clear_user(nmask, ALIGN(maxnode-1, 8) / 8);
1530 err |= compat_put_bitmap(nmask, bm, nr_bits);
1531 }
1532
1533 return err;
1534}
1535
1536COMPAT_SYSCALL_DEFINE3(set_mempolicy, int, mode, compat_ulong_t __user *, nmask,
1537 compat_ulong_t, maxnode)
1538{
1539 unsigned long __user *nm = NULL;
1540 unsigned long nr_bits, alloc_size;
1541 DECLARE_BITMAP(bm, MAX_NUMNODES);
1542
1543 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1544 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1545
1546 if (nmask) {
1547 if (compat_get_bitmap(bm, nmask, nr_bits))
1548 return -EFAULT;
1549 nm = compat_alloc_user_space(alloc_size);
1550 if (copy_to_user(nm, bm, alloc_size))
1551 return -EFAULT;
1552 }
1553
1554 return kernel_set_mempolicy(mode, nm, nr_bits+1);
1555}
1556
1557COMPAT_SYSCALL_DEFINE6(mbind, compat_ulong_t, start, compat_ulong_t, len,
1558 compat_ulong_t, mode, compat_ulong_t __user *, nmask,
1559 compat_ulong_t, maxnode, compat_ulong_t, flags)
1560{
1561 unsigned long __user *nm = NULL;
1562 unsigned long nr_bits, alloc_size;
1563 nodemask_t bm;
1564
1565 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1566 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1567
1568 if (nmask) {
1569 if (compat_get_bitmap(nodes_addr(bm), nmask, nr_bits))
1570 return -EFAULT;
1571 nm = compat_alloc_user_space(alloc_size);
1572 if (copy_to_user(nm, nodes_addr(bm), alloc_size))
1573 return -EFAULT;
1574 }
1575
1576 return kernel_mbind(start, len, mode, nm, nr_bits+1, flags);
1577}
1578
1579COMPAT_SYSCALL_DEFINE4(migrate_pages, compat_pid_t, pid,
1580 compat_ulong_t, maxnode,
1581 const compat_ulong_t __user *, old_nodes,
1582 const compat_ulong_t __user *, new_nodes)
1583{
1584 unsigned long __user *old = NULL;
1585 unsigned long __user *new = NULL;
1586 nodemask_t tmp_mask;
1587 unsigned long nr_bits;
1588 unsigned long size;
1589
1590 nr_bits = min_t(unsigned long, maxnode - 1, MAX_NUMNODES);
1591 size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1592 if (old_nodes) {
1593 if (compat_get_bitmap(nodes_addr(tmp_mask), old_nodes, nr_bits))
1594 return -EFAULT;
1595 old = compat_alloc_user_space(new_nodes ? size * 2 : size);
1596 if (new_nodes)
1597 new = old + size / sizeof(unsigned long);
1598 if (copy_to_user(old, nodes_addr(tmp_mask), size))
1599 return -EFAULT;
1600 }
1601 if (new_nodes) {
1602 if (compat_get_bitmap(nodes_addr(tmp_mask), new_nodes, nr_bits))
1603 return -EFAULT;
1604 if (new == NULL)
1605 new = compat_alloc_user_space(size);
1606 if (copy_to_user(new, nodes_addr(tmp_mask), size))
1607 return -EFAULT;
1608 }
1609 return kernel_migrate_pages(pid, nr_bits + 1, old, new);
1610}
1611
1612#endif /* CONFIG_COMPAT */
1613
1614struct mempolicy *__get_vma_policy(struct vm_area_struct *vma,
1615 unsigned long addr)
1616{
1617 struct mempolicy *pol = NULL;
1618
1619 if (vma) {
1620 if (vma->vm_ops && vma->vm_ops->get_policy) {
1621 pol = vma->vm_ops->get_policy(vma, addr);
1622 } else if (vma->vm_policy) {
1623 pol = vma->vm_policy;
1624
1625 /*
1626 * shmem_alloc_page() passes MPOL_F_SHARED policy with
1627 * a pseudo vma whose vma->vm_ops=NULL. Take a reference
1628 * count on these policies which will be dropped by
1629 * mpol_cond_put() later
1630 */
1631 if (mpol_needs_cond_ref(pol))
1632 mpol_get(pol);
1633 }
1634 }
1635
1636 return pol;
1637}
1638
1639/*
1640 * get_vma_policy(@vma, @addr)
1641 * @vma: virtual memory area whose policy is sought
1642 * @addr: address in @vma for shared policy lookup
1643 *
1644 * Returns effective policy for a VMA at specified address.
1645 * Falls back to current->mempolicy or system default policy, as necessary.
1646 * Shared policies [those marked as MPOL_F_SHARED] require an extra reference
1647 * count--added by the get_policy() vm_op, as appropriate--to protect against
1648 * freeing by another task. It is the caller's responsibility to free the
1649 * extra reference for shared policies.
1650 */
1651static struct mempolicy *get_vma_policy(struct vm_area_struct *vma,
1652 unsigned long addr)
1653{
1654 struct mempolicy *pol = __get_vma_policy(vma, addr);
1655
1656 if (!pol)
1657 pol = get_task_policy(current);
1658
1659 return pol;
1660}
1661
1662bool vma_policy_mof(struct vm_area_struct *vma)
1663{
1664 struct mempolicy *pol;
1665
1666 if (vma->vm_ops && vma->vm_ops->get_policy) {
1667 bool ret = false;
1668
1669 pol = vma->vm_ops->get_policy(vma, vma->vm_start);
1670 if (pol && (pol->flags & MPOL_F_MOF))
1671 ret = true;
1672 mpol_cond_put(pol);
1673
1674 return ret;
1675 }
1676
1677 pol = vma->vm_policy;
1678 if (!pol)
1679 pol = get_task_policy(current);
1680
1681 return pol->flags & MPOL_F_MOF;
1682}
1683
1684static int apply_policy_zone(struct mempolicy *policy, enum zone_type zone)
1685{
1686 enum zone_type dynamic_policy_zone = policy_zone;
1687
1688 BUG_ON(dynamic_policy_zone == ZONE_MOVABLE);
1689
1690 /*
1691 * if policy->v.nodes has movable memory only,
1692 * we apply policy when gfp_zone(gfp) = ZONE_MOVABLE only.
1693 *
1694 * policy->v.nodes is intersect with node_states[N_MEMORY].
1695 * so if the following test faile, it implies
1696 * policy->v.nodes has movable memory only.
1697 */
1698 if (!nodes_intersects(policy->v.nodes, node_states[N_HIGH_MEMORY]))
1699 dynamic_policy_zone = ZONE_MOVABLE;
1700
1701 return zone >= dynamic_policy_zone;
1702}
1703
1704/*
1705 * Return a nodemask representing a mempolicy for filtering nodes for
1706 * page allocation
1707 */
1708static nodemask_t *policy_nodemask(gfp_t gfp, struct mempolicy *policy)
1709{
1710 /* Lower zones don't get a nodemask applied for MPOL_BIND */
1711 if (unlikely(policy->mode == MPOL_BIND) &&
1712 apply_policy_zone(policy, gfp_zone(gfp)) &&
1713 cpuset_nodemask_valid_mems_allowed(&policy->v.nodes))
1714 return &policy->v.nodes;
1715
1716 return NULL;
1717}
1718
1719/* Return the node id preferred by the given mempolicy, or the given id */
1720static int policy_node(gfp_t gfp, struct mempolicy *policy,
1721 int nd)
1722{
1723 if (policy->mode == MPOL_PREFERRED && !(policy->flags & MPOL_F_LOCAL))
1724 nd = policy->v.preferred_node;
1725 else {
1726 /*
1727 * __GFP_THISNODE shouldn't even be used with the bind policy
1728 * because we might easily break the expectation to stay on the
1729 * requested node and not break the policy.
1730 */
1731 WARN_ON_ONCE(policy->mode == MPOL_BIND && (gfp & __GFP_THISNODE));
1732 }
1733
1734 return nd;
1735}
1736
1737/* Do dynamic interleaving for a process */
1738static unsigned interleave_nodes(struct mempolicy *policy)
1739{
1740 unsigned next;
1741 struct task_struct *me = current;
1742
1743 next = next_node_in(me->il_prev, policy->v.nodes);
1744 if (next < MAX_NUMNODES)
1745 me->il_prev = next;
1746 return next;
1747}
1748
1749/*
1750 * Depending on the memory policy provide a node from which to allocate the
1751 * next slab entry.
1752 */
1753unsigned int mempolicy_slab_node(void)
1754{
1755 struct mempolicy *policy;
1756 int node = numa_mem_id();
1757
1758 if (in_interrupt())
1759 return node;
1760
1761 policy = current->mempolicy;
1762 if (!policy || policy->flags & MPOL_F_LOCAL)
1763 return node;
1764
1765 switch (policy->mode) {
1766 case MPOL_PREFERRED:
1767 /*
1768 * handled MPOL_F_LOCAL above
1769 */
1770 return policy->v.preferred_node;
1771
1772 case MPOL_INTERLEAVE:
1773 return interleave_nodes(policy);
1774
1775 case MPOL_BIND: {
1776 struct zoneref *z;
1777
1778 /*
1779 * Follow bind policy behavior and start allocation at the
1780 * first node.
1781 */
1782 struct zonelist *zonelist;
1783 enum zone_type highest_zoneidx = gfp_zone(GFP_KERNEL);
1784 zonelist = &NODE_DATA(node)->node_zonelists[ZONELIST_FALLBACK];
1785 z = first_zones_zonelist(zonelist, highest_zoneidx,
1786 &policy->v.nodes);
1787 return z->zone ? z->zone->node : node;
1788 }
1789
1790 default:
1791 BUG();
1792 }
1793}
1794
1795/*
1796 * Do static interleaving for a VMA with known offset @n. Returns the n'th
1797 * node in pol->v.nodes (starting from n=0), wrapping around if n exceeds the
1798 * number of present nodes.
1799 */
1800static unsigned offset_il_node(struct mempolicy *pol, unsigned long n)
1801{
1802 unsigned nnodes = nodes_weight(pol->v.nodes);
1803 unsigned target;
1804 int i;
1805 int nid;
1806
1807 if (!nnodes)
1808 return numa_node_id();
1809 target = (unsigned int)n % nnodes;
1810 nid = first_node(pol->v.nodes);
1811 for (i = 0; i < target; i++)
1812 nid = next_node(nid, pol->v.nodes);
1813 return nid;
1814}
1815
1816/* Determine a node number for interleave */
1817static inline unsigned interleave_nid(struct mempolicy *pol,
1818 struct vm_area_struct *vma, unsigned long addr, int shift)
1819{
1820 if (vma) {
1821 unsigned long off;
1822
1823 /*
1824 * for small pages, there is no difference between
1825 * shift and PAGE_SHIFT, so the bit-shift is safe.
1826 * for huge pages, since vm_pgoff is in units of small
1827 * pages, we need to shift off the always 0 bits to get
1828 * a useful offset.
1829 */
1830 BUG_ON(shift < PAGE_SHIFT);
1831 off = vma->vm_pgoff >> (shift - PAGE_SHIFT);
1832 off += (addr - vma->vm_start) >> shift;
1833 return offset_il_node(pol, off);
1834 } else
1835 return interleave_nodes(pol);
1836}
1837
1838#ifdef CONFIG_HUGETLBFS
1839/*
1840 * huge_node(@vma, @addr, @gfp_flags, @mpol)
1841 * @vma: virtual memory area whose policy is sought
1842 * @addr: address in @vma for shared policy lookup and interleave policy
1843 * @gfp_flags: for requested zone
1844 * @mpol: pointer to mempolicy pointer for reference counted mempolicy
1845 * @nodemask: pointer to nodemask pointer for MPOL_BIND nodemask
1846 *
1847 * Returns a nid suitable for a huge page allocation and a pointer
1848 * to the struct mempolicy for conditional unref after allocation.
1849 * If the effective policy is 'BIND, returns a pointer to the mempolicy's
1850 * @nodemask for filtering the zonelist.
1851 *
1852 * Must be protected by read_mems_allowed_begin()
1853 */
1854int huge_node(struct vm_area_struct *vma, unsigned long addr, gfp_t gfp_flags,
1855 struct mempolicy **mpol, nodemask_t **nodemask)
1856{
1857 int nid;
1858
1859 *mpol = get_vma_policy(vma, addr);
1860 *nodemask = NULL; /* assume !MPOL_BIND */
1861
1862 if (unlikely((*mpol)->mode == MPOL_INTERLEAVE)) {
1863 nid = interleave_nid(*mpol, vma, addr,
1864 huge_page_shift(hstate_vma(vma)));
1865 } else {
1866 nid = policy_node(gfp_flags, *mpol, numa_node_id());
1867 if ((*mpol)->mode == MPOL_BIND)
1868 *nodemask = &(*mpol)->v.nodes;
1869 }
1870 return nid;
1871}
1872
1873/*
1874 * init_nodemask_of_mempolicy
1875 *
1876 * If the current task's mempolicy is "default" [NULL], return 'false'
1877 * to indicate default policy. Otherwise, extract the policy nodemask
1878 * for 'bind' or 'interleave' policy into the argument nodemask, or
1879 * initialize the argument nodemask to contain the single node for
1880 * 'preferred' or 'local' policy and return 'true' to indicate presence
1881 * of non-default mempolicy.
1882 *
1883 * We don't bother with reference counting the mempolicy [mpol_get/put]
1884 * because the current task is examining it's own mempolicy and a task's
1885 * mempolicy is only ever changed by the task itself.
1886 *
1887 * N.B., it is the caller's responsibility to free a returned nodemask.
1888 */
1889bool init_nodemask_of_mempolicy(nodemask_t *mask)
1890{
1891 struct mempolicy *mempolicy;
1892 int nid;
1893
1894 if (!(mask && current->mempolicy))
1895 return false;
1896
1897 task_lock(current);
1898 mempolicy = current->mempolicy;
1899 switch (mempolicy->mode) {
1900 case MPOL_PREFERRED:
1901 if (mempolicy->flags & MPOL_F_LOCAL)
1902 nid = numa_node_id();
1903 else
1904 nid = mempolicy->v.preferred_node;
1905 init_nodemask_of_node(mask, nid);
1906 break;
1907
1908 case MPOL_BIND:
1909 /* Fall through */
1910 case MPOL_INTERLEAVE:
1911 *mask = mempolicy->v.nodes;
1912 break;
1913
1914 default:
1915 BUG();
1916 }
1917 task_unlock(current);
1918
1919 return true;
1920}
1921#endif
1922
1923/*
1924 * mempolicy_nodemask_intersects
1925 *
1926 * If tsk's mempolicy is "default" [NULL], return 'true' to indicate default
1927 * policy. Otherwise, check for intersection between mask and the policy
1928 * nodemask for 'bind' or 'interleave' policy. For 'perferred' or 'local'
1929 * policy, always return true since it may allocate elsewhere on fallback.
1930 *
1931 * Takes task_lock(tsk) to prevent freeing of its mempolicy.
1932 */
1933bool mempolicy_nodemask_intersects(struct task_struct *tsk,
1934 const nodemask_t *mask)
1935{
1936 struct mempolicy *mempolicy;
1937 bool ret = true;
1938
1939 if (!mask)
1940 return ret;
1941 task_lock(tsk);
1942 mempolicy = tsk->mempolicy;
1943 if (!mempolicy)
1944 goto out;
1945
1946 switch (mempolicy->mode) {
1947 case MPOL_PREFERRED:
1948 /*
1949 * MPOL_PREFERRED and MPOL_F_LOCAL are only preferred nodes to
1950 * allocate from, they may fallback to other nodes when oom.
1951 * Thus, it's possible for tsk to have allocated memory from
1952 * nodes in mask.
1953 */
1954 break;
1955 case MPOL_BIND:
1956 case MPOL_INTERLEAVE:
1957 ret = nodes_intersects(mempolicy->v.nodes, *mask);
1958 break;
1959 default:
1960 BUG();
1961 }
1962out:
1963 task_unlock(tsk);
1964 return ret;
1965}
1966
1967/* Allocate a page in interleaved policy.
1968 Own path because it needs to do special accounting. */
1969static struct page *alloc_page_interleave(gfp_t gfp, unsigned order,
1970 unsigned nid)
1971{
1972 struct page *page;
1973
1974 page = __alloc_pages(gfp, order, nid);
1975 /* skip NUMA_INTERLEAVE_HIT counter update if numa stats is disabled */
1976 if (!static_branch_likely(&vm_numa_stat_key))
1977 return page;
1978 if (page && page_to_nid(page) == nid) {
1979 preempt_disable();
1980 __inc_numa_state(page_zone(page), NUMA_INTERLEAVE_HIT);
1981 preempt_enable();
1982 }
1983 return page;
1984}
1985
1986/**
1987 * alloc_pages_vma - Allocate a page for a VMA.
1988 *
1989 * @gfp:
1990 * %GFP_USER user allocation.
1991 * %GFP_KERNEL kernel allocations,
1992 * %GFP_HIGHMEM highmem/user allocations,
1993 * %GFP_FS allocation should not call back into a file system.
1994 * %GFP_ATOMIC don't sleep.
1995 *
1996 * @order:Order of the GFP allocation.
1997 * @vma: Pointer to VMA or NULL if not available.
1998 * @addr: Virtual Address of the allocation. Must be inside the VMA.
1999 * @node: Which node to prefer for allocation (modulo policy).
2000 * @hugepage: for hugepages try only the preferred node if possible
2001 *
2002 * This function allocates a page from the kernel page pool and applies
2003 * a NUMA policy associated with the VMA or the current process.
2004 * When VMA is not NULL caller must hold down_read on the mmap_sem of the
2005 * mm_struct of the VMA to prevent it from going away. Should be used for
2006 * all allocations for pages that will be mapped into user space. Returns
2007 * NULL when no page can be allocated.
2008 */
2009struct page *
2010alloc_pages_vma(gfp_t gfp, int order, struct vm_area_struct *vma,
2011 unsigned long addr, int node, bool hugepage)
2012{
2013 struct mempolicy *pol;
2014 struct page *page;
2015 int preferred_nid;
2016 nodemask_t *nmask;
2017
2018 pol = get_vma_policy(vma, addr);
2019
2020 if (pol->mode == MPOL_INTERLEAVE) {
2021 unsigned nid;
2022
2023 nid = interleave_nid(pol, vma, addr, PAGE_SHIFT + order);
2024 mpol_cond_put(pol);
2025 page = alloc_page_interleave(gfp, order, nid);
2026 goto out;
2027 }
2028
2029 if (unlikely(IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE) && hugepage)) {
2030 int hpage_node = node;
2031
2032 /*
2033 * For hugepage allocation and non-interleave policy which
2034 * allows the current node (or other explicitly preferred
2035 * node) we only try to allocate from the current/preferred
2036 * node and don't fall back to other nodes, as the cost of
2037 * remote accesses would likely offset THP benefits.
2038 *
2039 * If the policy is interleave, or does not allow the current
2040 * node in its nodemask, we allocate the standard way.
2041 */
2042 if (pol->mode == MPOL_PREFERRED &&
2043 !(pol->flags & MPOL_F_LOCAL))
2044 hpage_node = pol->v.preferred_node;
2045
2046 nmask = policy_nodemask(gfp, pol);
2047 if (!nmask || node_isset(hpage_node, *nmask)) {
2048 mpol_cond_put(pol);
2049 page = __alloc_pages_node(hpage_node,
2050 gfp | __GFP_THISNODE, order);
2051 goto out;
2052 }
2053 }
2054
2055 nmask = policy_nodemask(gfp, pol);
2056 preferred_nid = policy_node(gfp, pol, node);
2057 page = __alloc_pages_nodemask(gfp, order, preferred_nid, nmask);
2058 mpol_cond_put(pol);
2059out:
2060 return page;
2061}
2062
2063/**
2064 * alloc_pages_current - Allocate pages.
2065 *
2066 * @gfp:
2067 * %GFP_USER user allocation,
2068 * %GFP_KERNEL kernel allocation,
2069 * %GFP_HIGHMEM highmem allocation,
2070 * %GFP_FS don't call back into a file system.
2071 * %GFP_ATOMIC don't sleep.
2072 * @order: Power of two of allocation size in pages. 0 is a single page.
2073 *
2074 * Allocate a page from the kernel page pool. When not in
2075 * interrupt context and apply the current process NUMA policy.
2076 * Returns NULL when no page can be allocated.
2077 */
2078struct page *alloc_pages_current(gfp_t gfp, unsigned order)
2079{
2080 struct mempolicy *pol = &default_policy;
2081 struct page *page;
2082
2083 if (!in_interrupt() && !(gfp & __GFP_THISNODE))
2084 pol = get_task_policy(current);
2085
2086 /*
2087 * No reference counting needed for current->mempolicy
2088 * nor system default_policy
2089 */
2090 if (pol->mode == MPOL_INTERLEAVE)
2091 page = alloc_page_interleave(gfp, order, interleave_nodes(pol));
2092 else
2093 page = __alloc_pages_nodemask(gfp, order,
2094 policy_node(gfp, pol, numa_node_id()),
2095 policy_nodemask(gfp, pol));
2096
2097 return page;
2098}
2099EXPORT_SYMBOL(alloc_pages_current);
2100
2101int vma_dup_policy(struct vm_area_struct *src, struct vm_area_struct *dst)
2102{
2103 struct mempolicy *pol = mpol_dup(vma_policy(src));
2104
2105 if (IS_ERR(pol))
2106 return PTR_ERR(pol);
2107 dst->vm_policy = pol;
2108 return 0;
2109}
2110
2111/*
2112 * If mpol_dup() sees current->cpuset == cpuset_being_rebound, then it
2113 * rebinds the mempolicy its copying by calling mpol_rebind_policy()
2114 * with the mems_allowed returned by cpuset_mems_allowed(). This
2115 * keeps mempolicies cpuset relative after its cpuset moves. See
2116 * further kernel/cpuset.c update_nodemask().
2117 *
2118 * current's mempolicy may be rebinded by the other task(the task that changes
2119 * cpuset's mems), so we needn't do rebind work for current task.
2120 */
2121
2122/* Slow path of a mempolicy duplicate */
2123struct mempolicy *__mpol_dup(struct mempolicy *old)
2124{
2125 struct mempolicy *new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
2126
2127 if (!new)
2128 return ERR_PTR(-ENOMEM);
2129
2130 /* task's mempolicy is protected by alloc_lock */
2131 if (old == current->mempolicy) {
2132 task_lock(current);
2133 *new = *old;
2134 task_unlock(current);
2135 } else
2136 *new = *old;
2137
2138 if (current_cpuset_is_being_rebound()) {
2139 nodemask_t mems = cpuset_mems_allowed(current);
2140 mpol_rebind_policy(new, &mems);
2141 }
2142 atomic_set(&new->refcnt, 1);
2143 return new;
2144}
2145
2146/* Slow path of a mempolicy comparison */
2147bool __mpol_equal(struct mempolicy *a, struct mempolicy *b)
2148{
2149 if (!a || !b)
2150 return false;
2151 if (a->mode != b->mode)
2152 return false;
2153 if (a->flags != b->flags)
2154 return false;
2155 if (mpol_store_user_nodemask(a))
2156 if (!nodes_equal(a->w.user_nodemask, b->w.user_nodemask))
2157 return false;
2158
2159 switch (a->mode) {
2160 case MPOL_BIND:
2161 /* Fall through */
2162 case MPOL_INTERLEAVE:
2163 return !!nodes_equal(a->v.nodes, b->v.nodes);
2164 case MPOL_PREFERRED:
2165 /* a's ->flags is the same as b's */
2166 if (a->flags & MPOL_F_LOCAL)
2167 return true;
2168 return a->v.preferred_node == b->v.preferred_node;
2169 default:
2170 BUG();
2171 return false;
2172 }
2173}
2174
2175/*
2176 * Shared memory backing store policy support.
2177 *
2178 * Remember policies even when nobody has shared memory mapped.
2179 * The policies are kept in Red-Black tree linked from the inode.
2180 * They are protected by the sp->lock rwlock, which should be held
2181 * for any accesses to the tree.
2182 */
2183
2184/*
2185 * lookup first element intersecting start-end. Caller holds sp->lock for
2186 * reading or for writing
2187 */
2188static struct sp_node *
2189sp_lookup(struct shared_policy *sp, unsigned long start, unsigned long end)
2190{
2191 struct rb_node *n = sp->root.rb_node;
2192
2193 while (n) {
2194 struct sp_node *p = rb_entry(n, struct sp_node, nd);
2195
2196 if (start >= p->end)
2197 n = n->rb_right;
2198 else if (end <= p->start)
2199 n = n->rb_left;
2200 else
2201 break;
2202 }
2203 if (!n)
2204 return NULL;
2205 for (;;) {
2206 struct sp_node *w = NULL;
2207 struct rb_node *prev = rb_prev(n);
2208 if (!prev)
2209 break;
2210 w = rb_entry(prev, struct sp_node, nd);
2211 if (w->end <= start)
2212 break;
2213 n = prev;
2214 }
2215 return rb_entry(n, struct sp_node, nd);
2216}
2217
2218/*
2219 * Insert a new shared policy into the list. Caller holds sp->lock for
2220 * writing.
2221 */
2222static void sp_insert(struct shared_policy *sp, struct sp_node *new)
2223{
2224 struct rb_node **p = &sp->root.rb_node;
2225 struct rb_node *parent = NULL;
2226 struct sp_node *nd;
2227
2228 while (*p) {
2229 parent = *p;
2230 nd = rb_entry(parent, struct sp_node, nd);
2231 if (new->start < nd->start)
2232 p = &(*p)->rb_left;
2233 else if (new->end > nd->end)
2234 p = &(*p)->rb_right;
2235 else
2236 BUG();
2237 }
2238 rb_link_node(&new->nd, parent, p);
2239 rb_insert_color(&new->nd, &sp->root);
2240 pr_debug("inserting %lx-%lx: %d\n", new->start, new->end,
2241 new->policy ? new->policy->mode : 0);
2242}
2243
2244/* Find shared policy intersecting idx */
2245struct mempolicy *
2246mpol_shared_policy_lookup(struct shared_policy *sp, unsigned long idx)
2247{
2248 struct mempolicy *pol = NULL;
2249 struct sp_node *sn;
2250
2251 if (!sp->root.rb_node)
2252 return NULL;
2253 read_lock(&sp->lock);
2254 sn = sp_lookup(sp, idx, idx+1);
2255 if (sn) {
2256 mpol_get(sn->policy);
2257 pol = sn->policy;
2258 }
2259 read_unlock(&sp->lock);
2260 return pol;
2261}
2262
2263static void sp_free(struct sp_node *n)
2264{
2265 mpol_put(n->policy);
2266 kmem_cache_free(sn_cache, n);
2267}
2268
2269/**
2270 * mpol_misplaced - check whether current page node is valid in policy
2271 *
2272 * @page: page to be checked
2273 * @vma: vm area where page mapped
2274 * @addr: virtual address where page mapped
2275 *
2276 * Lookup current policy node id for vma,addr and "compare to" page's
2277 * node id.
2278 *
2279 * Returns:
2280 * -1 - not misplaced, page is in the right node
2281 * node - node id where the page should be
2282 *
2283 * Policy determination "mimics" alloc_page_vma().
2284 * Called from fault path where we know the vma and faulting address.
2285 */
2286int mpol_misplaced(struct page *page, struct vm_area_struct *vma, unsigned long addr)
2287{
2288 struct mempolicy *pol;
2289 struct zoneref *z;
2290 int curnid = page_to_nid(page);
2291 unsigned long pgoff;
2292 int thiscpu = raw_smp_processor_id();
2293 int thisnid = cpu_to_node(thiscpu);
2294 int polnid = -1;
2295 int ret = -1;
2296
2297 pol = get_vma_policy(vma, addr);
2298 if (!(pol->flags & MPOL_F_MOF))
2299 goto out;
2300
2301 switch (pol->mode) {
2302 case MPOL_INTERLEAVE:
2303 pgoff = vma->vm_pgoff;
2304 pgoff += (addr - vma->vm_start) >> PAGE_SHIFT;
2305 polnid = offset_il_node(pol, pgoff);
2306 break;
2307
2308 case MPOL_PREFERRED:
2309 if (pol->flags & MPOL_F_LOCAL)
2310 polnid = numa_node_id();
2311 else
2312 polnid = pol->v.preferred_node;
2313 break;
2314
2315 case MPOL_BIND:
2316
2317 /*
2318 * allows binding to multiple nodes.
2319 * use current page if in policy nodemask,
2320 * else select nearest allowed node, if any.
2321 * If no allowed nodes, use current [!misplaced].
2322 */
2323 if (node_isset(curnid, pol->v.nodes))
2324 goto out;
2325 z = first_zones_zonelist(
2326 node_zonelist(numa_node_id(), GFP_HIGHUSER),
2327 gfp_zone(GFP_HIGHUSER),
2328 &pol->v.nodes);
2329 polnid = z->zone->node;
2330 break;
2331
2332 default:
2333 BUG();
2334 }
2335
2336 /* Migrate the page towards the node whose CPU is referencing it */
2337 if (pol->flags & MPOL_F_MORON) {
2338 polnid = thisnid;
2339
2340 if (!should_numa_migrate_memory(current, page, curnid, thiscpu))
2341 goto out;
2342 }
2343
2344 if (curnid != polnid)
2345 ret = polnid;
2346out:
2347 mpol_cond_put(pol);
2348
2349 return ret;
2350}
2351
2352/*
2353 * Drop the (possibly final) reference to task->mempolicy. It needs to be
2354 * dropped after task->mempolicy is set to NULL so that any allocation done as
2355 * part of its kmem_cache_free(), such as by KASAN, doesn't reference a freed
2356 * policy.
2357 */
2358void mpol_put_task_policy(struct task_struct *task)
2359{
2360 struct mempolicy *pol;
2361
2362 task_lock(task);
2363 pol = task->mempolicy;
2364 task->mempolicy = NULL;
2365 task_unlock(task);
2366 mpol_put(pol);
2367}
2368
2369static void sp_delete(struct shared_policy *sp, struct sp_node *n)
2370{
2371 pr_debug("deleting %lx-l%lx\n", n->start, n->end);
2372 rb_erase(&n->nd, &sp->root);
2373 sp_free(n);
2374}
2375
2376static void sp_node_init(struct sp_node *node, unsigned long start,
2377 unsigned long end, struct mempolicy *pol)
2378{
2379 node->start = start;
2380 node->end = end;
2381 node->policy = pol;
2382}
2383
2384static struct sp_node *sp_alloc(unsigned long start, unsigned long end,
2385 struct mempolicy *pol)
2386{
2387 struct sp_node *n;
2388 struct mempolicy *newpol;
2389
2390 n = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2391 if (!n)
2392 return NULL;
2393
2394 newpol = mpol_dup(pol);
2395 if (IS_ERR(newpol)) {
2396 kmem_cache_free(sn_cache, n);
2397 return NULL;
2398 }
2399 newpol->flags |= MPOL_F_SHARED;
2400 sp_node_init(n, start, end, newpol);
2401
2402 return n;
2403}
2404
2405/* Replace a policy range. */
2406static int shared_policy_replace(struct shared_policy *sp, unsigned long start,
2407 unsigned long end, struct sp_node *new)
2408{
2409 struct sp_node *n;
2410 struct sp_node *n_new = NULL;
2411 struct mempolicy *mpol_new = NULL;
2412 int ret = 0;
2413
2414restart:
2415 write_lock(&sp->lock);
2416 n = sp_lookup(sp, start, end);
2417 /* Take care of old policies in the same range. */
2418 while (n && n->start < end) {
2419 struct rb_node *next = rb_next(&n->nd);
2420 if (n->start >= start) {
2421 if (n->end <= end)
2422 sp_delete(sp, n);
2423 else
2424 n->start = end;
2425 } else {
2426 /* Old policy spanning whole new range. */
2427 if (n->end > end) {
2428 if (!n_new)
2429 goto alloc_new;
2430
2431 *mpol_new = *n->policy;
2432 atomic_set(&mpol_new->refcnt, 1);
2433 sp_node_init(n_new, end, n->end, mpol_new);
2434 n->end = start;
2435 sp_insert(sp, n_new);
2436 n_new = NULL;
2437 mpol_new = NULL;
2438 break;
2439 } else
2440 n->end = start;
2441 }
2442 if (!next)
2443 break;
2444 n = rb_entry(next, struct sp_node, nd);
2445 }
2446 if (new)
2447 sp_insert(sp, new);
2448 write_unlock(&sp->lock);
2449 ret = 0;
2450
2451err_out:
2452 if (mpol_new)
2453 mpol_put(mpol_new);
2454 if (n_new)
2455 kmem_cache_free(sn_cache, n_new);
2456
2457 return ret;
2458
2459alloc_new:
2460 write_unlock(&sp->lock);
2461 ret = -ENOMEM;
2462 n_new = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2463 if (!n_new)
2464 goto err_out;
2465 mpol_new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
2466 if (!mpol_new)
2467 goto err_out;
2468 goto restart;
2469}
2470
2471/**
2472 * mpol_shared_policy_init - initialize shared policy for inode
2473 * @sp: pointer to inode shared policy
2474 * @mpol: struct mempolicy to install
2475 *
2476 * Install non-NULL @mpol in inode's shared policy rb-tree.
2477 * On entry, the current task has a reference on a non-NULL @mpol.
2478 * This must be released on exit.
2479 * This is called at get_inode() calls and we can use GFP_KERNEL.
2480 */
2481void mpol_shared_policy_init(struct shared_policy *sp, struct mempolicy *mpol)
2482{
2483 int ret;
2484
2485 sp->root = RB_ROOT; /* empty tree == default mempolicy */
2486 rwlock_init(&sp->lock);
2487
2488 if (mpol) {
2489 struct vm_area_struct pvma;
2490 struct mempolicy *new;
2491 NODEMASK_SCRATCH(scratch);
2492
2493 if (!scratch)
2494 goto put_mpol;
2495 /* contextualize the tmpfs mount point mempolicy */
2496 new = mpol_new(mpol->mode, mpol->flags, &mpol->w.user_nodemask);
2497 if (IS_ERR(new))
2498 goto free_scratch; /* no valid nodemask intersection */
2499
2500 task_lock(current);
2501 ret = mpol_set_nodemask(new, &mpol->w.user_nodemask, scratch);
2502 task_unlock(current);
2503 if (ret)
2504 goto put_new;
2505
2506 /* Create pseudo-vma that contains just the policy */
2507 memset(&pvma, 0, sizeof(struct vm_area_struct));
2508 pvma.vm_end = TASK_SIZE; /* policy covers entire file */
2509 mpol_set_shared_policy(sp, &pvma, new); /* adds ref */
2510
2511put_new:
2512 mpol_put(new); /* drop initial ref */
2513free_scratch:
2514 NODEMASK_SCRATCH_FREE(scratch);
2515put_mpol:
2516 mpol_put(mpol); /* drop our incoming ref on sb mpol */
2517 }
2518}
2519
2520int mpol_set_shared_policy(struct shared_policy *info,
2521 struct vm_area_struct *vma, struct mempolicy *npol)
2522{
2523 int err;
2524 struct sp_node *new = NULL;
2525 unsigned long sz = vma_pages(vma);
2526
2527 pr_debug("set_shared_policy %lx sz %lu %d %d %lx\n",
2528 vma->vm_pgoff,
2529 sz, npol ? npol->mode : -1,
2530 npol ? npol->flags : -1,
2531 npol ? nodes_addr(npol->v.nodes)[0] : NUMA_NO_NODE);
2532
2533 if (npol) {
2534 new = sp_alloc(vma->vm_pgoff, vma->vm_pgoff + sz, npol);
2535 if (!new)
2536 return -ENOMEM;
2537 }
2538 err = shared_policy_replace(info, vma->vm_pgoff, vma->vm_pgoff+sz, new);
2539 if (err && new)
2540 sp_free(new);
2541 return err;
2542}
2543
2544/* Free a backing policy store on inode delete. */
2545void mpol_free_shared_policy(struct shared_policy *p)
2546{
2547 struct sp_node *n;
2548 struct rb_node *next;
2549
2550 if (!p->root.rb_node)
2551 return;
2552 write_lock(&p->lock);
2553 next = rb_first(&p->root);
2554 while (next) {
2555 n = rb_entry(next, struct sp_node, nd);
2556 next = rb_next(&n->nd);
2557 sp_delete(p, n);
2558 }
2559 write_unlock(&p->lock);
2560}
2561
2562#ifdef CONFIG_NUMA_BALANCING
2563static int __initdata numabalancing_override;
2564
2565static void __init check_numabalancing_enable(void)
2566{
2567 bool numabalancing_default = false;
2568
2569 if (IS_ENABLED(CONFIG_NUMA_BALANCING_DEFAULT_ENABLED))
2570 numabalancing_default = true;
2571
2572 /* Parsed by setup_numabalancing. override == 1 enables, -1 disables */
2573 if (numabalancing_override)
2574 set_numabalancing_state(numabalancing_override == 1);
2575
2576 if (num_online_nodes() > 1 && !numabalancing_override) {
2577 pr_info("%s automatic NUMA balancing. Configure with numa_balancing= or the kernel.numa_balancing sysctl\n",
2578 numabalancing_default ? "Enabling" : "Disabling");
2579 set_numabalancing_state(numabalancing_default);
2580 }
2581}
2582
2583static int __init setup_numabalancing(char *str)
2584{
2585 int ret = 0;
2586 if (!str)
2587 goto out;
2588
2589 if (!strcmp(str, "enable")) {
2590 numabalancing_override = 1;
2591 ret = 1;
2592 } else if (!strcmp(str, "disable")) {
2593 numabalancing_override = -1;
2594 ret = 1;
2595 }
2596out:
2597 if (!ret)
2598 pr_warn("Unable to parse numa_balancing=\n");
2599
2600 return ret;
2601}
2602__setup("numa_balancing=", setup_numabalancing);
2603#else
2604static inline void __init check_numabalancing_enable(void)
2605{
2606}
2607#endif /* CONFIG_NUMA_BALANCING */
2608
2609/* assumes fs == KERNEL_DS */
2610void __init numa_policy_init(void)
2611{
2612 nodemask_t interleave_nodes;
2613 unsigned long largest = 0;
2614 int nid, prefer = 0;
2615
2616 policy_cache = kmem_cache_create("numa_policy",
2617 sizeof(struct mempolicy),
2618 0, SLAB_PANIC, NULL);
2619
2620 sn_cache = kmem_cache_create("shared_policy_node",
2621 sizeof(struct sp_node),
2622 0, SLAB_PANIC, NULL);
2623
2624 for_each_node(nid) {
2625 preferred_node_policy[nid] = (struct mempolicy) {
2626 .refcnt = ATOMIC_INIT(1),
2627 .mode = MPOL_PREFERRED,
2628 .flags = MPOL_F_MOF | MPOL_F_MORON,
2629 .v = { .preferred_node = nid, },
2630 };
2631 }
2632
2633 /*
2634 * Set interleaving policy for system init. Interleaving is only
2635 * enabled across suitably sized nodes (default is >= 16MB), or
2636 * fall back to the largest node if they're all smaller.
2637 */
2638 nodes_clear(interleave_nodes);
2639 for_each_node_state(nid, N_MEMORY) {
2640 unsigned long total_pages = node_present_pages(nid);
2641
2642 /* Preserve the largest node */
2643 if (largest < total_pages) {
2644 largest = total_pages;
2645 prefer = nid;
2646 }
2647
2648 /* Interleave this node? */
2649 if ((total_pages << PAGE_SHIFT) >= (16 << 20))
2650 node_set(nid, interleave_nodes);
2651 }
2652
2653 /* All too small, use the largest */
2654 if (unlikely(nodes_empty(interleave_nodes)))
2655 node_set(prefer, interleave_nodes);
2656
2657 if (do_set_mempolicy(MPOL_INTERLEAVE, 0, &interleave_nodes))
2658 pr_err("%s: interleaving failed\n", __func__);
2659
2660 check_numabalancing_enable();
2661}
2662
2663/* Reset policy of current process to default */
2664void numa_default_policy(void)
2665{
2666 do_set_mempolicy(MPOL_DEFAULT, 0, NULL);
2667}
2668
2669/*
2670 * Parse and format mempolicy from/to strings
2671 */
2672
2673/*
2674 * "local" is implemented internally by MPOL_PREFERRED with MPOL_F_LOCAL flag.
2675 */
2676static const char * const policy_modes[] =
2677{
2678 [MPOL_DEFAULT] = "default",
2679 [MPOL_PREFERRED] = "prefer",
2680 [MPOL_BIND] = "bind",
2681 [MPOL_INTERLEAVE] = "interleave",
2682 [MPOL_LOCAL] = "local",
2683};
2684
2685
2686#ifdef CONFIG_TMPFS
2687/**
2688 * mpol_parse_str - parse string to mempolicy, for tmpfs mpol mount option.
2689 * @str: string containing mempolicy to parse
2690 * @mpol: pointer to struct mempolicy pointer, returned on success.
2691 *
2692 * Format of input:
2693 * <mode>[=<flags>][:<nodelist>]
2694 *
2695 * On success, returns 0, else 1
2696 */
2697int mpol_parse_str(char *str, struct mempolicy **mpol)
2698{
2699 struct mempolicy *new = NULL;
2700 unsigned short mode;
2701 unsigned short mode_flags;
2702 nodemask_t nodes;
2703 char *nodelist = strchr(str, ':');
2704 char *flags = strchr(str, '=');
2705 int err = 1;
2706
2707 if (nodelist) {
2708 /* NUL-terminate mode or flags string */
2709 *nodelist++ = '\0';
2710 if (nodelist_parse(nodelist, nodes))
2711 goto out;
2712 if (!nodes_subset(nodes, node_states[N_MEMORY]))
2713 goto out;
2714 } else
2715 nodes_clear(nodes);
2716
2717 if (flags)
2718 *flags++ = '\0'; /* terminate mode string */
2719
2720 for (mode = 0; mode < MPOL_MAX; mode++) {
2721 if (!strcmp(str, policy_modes[mode])) {
2722 break;
2723 }
2724 }
2725 if (mode >= MPOL_MAX)
2726 goto out;
2727
2728 switch (mode) {
2729 case MPOL_PREFERRED:
2730 /*
2731 * Insist on a nodelist of one node only
2732 */
2733 if (nodelist) {
2734 char *rest = nodelist;
2735 while (isdigit(*rest))
2736 rest++;
2737 if (*rest)
2738 goto out;
2739 }
2740 break;
2741 case MPOL_INTERLEAVE:
2742 /*
2743 * Default to online nodes with memory if no nodelist
2744 */
2745 if (!nodelist)
2746 nodes = node_states[N_MEMORY];
2747 break;
2748 case MPOL_LOCAL:
2749 /*
2750 * Don't allow a nodelist; mpol_new() checks flags
2751 */
2752 if (nodelist)
2753 goto out;
2754 mode = MPOL_PREFERRED;
2755 break;
2756 case MPOL_DEFAULT:
2757 /*
2758 * Insist on a empty nodelist
2759 */
2760 if (!nodelist)
2761 err = 0;
2762 goto out;
2763 case MPOL_BIND:
2764 /*
2765 * Insist on a nodelist
2766 */
2767 if (!nodelist)
2768 goto out;
2769 }
2770
2771 mode_flags = 0;
2772 if (flags) {
2773 /*
2774 * Currently, we only support two mutually exclusive
2775 * mode flags.
2776 */
2777 if (!strcmp(flags, "static"))
2778 mode_flags |= MPOL_F_STATIC_NODES;
2779 else if (!strcmp(flags, "relative"))
2780 mode_flags |= MPOL_F_RELATIVE_NODES;
2781 else
2782 goto out;
2783 }
2784
2785 new = mpol_new(mode, mode_flags, &nodes);
2786 if (IS_ERR(new))
2787 goto out;
2788
2789 /*
2790 * Save nodes for mpol_to_str() to show the tmpfs mount options
2791 * for /proc/mounts, /proc/pid/mounts and /proc/pid/mountinfo.
2792 */
2793 if (mode != MPOL_PREFERRED)
2794 new->v.nodes = nodes;
2795 else if (nodelist)
2796 new->v.preferred_node = first_node(nodes);
2797 else
2798 new->flags |= MPOL_F_LOCAL;
2799
2800 /*
2801 * Save nodes for contextualization: this will be used to "clone"
2802 * the mempolicy in a specific context [cpuset] at a later time.
2803 */
2804 new->w.user_nodemask = nodes;
2805
2806 err = 0;
2807
2808out:
2809 /* Restore string for error message */
2810 if (nodelist)
2811 *--nodelist = ':';
2812 if (flags)
2813 *--flags = '=';
2814 if (!err)
2815 *mpol = new;
2816 return err;
2817}
2818#endif /* CONFIG_TMPFS */
2819
2820/**
2821 * mpol_to_str - format a mempolicy structure for printing
2822 * @buffer: to contain formatted mempolicy string
2823 * @maxlen: length of @buffer
2824 * @pol: pointer to mempolicy to be formatted
2825 *
2826 * Convert @pol into a string. If @buffer is too short, truncate the string.
2827 * Recommend a @maxlen of at least 32 for the longest mode, "interleave", the
2828 * longest flag, "relative", and to display at least a few node ids.
2829 */
2830void mpol_to_str(char *buffer, int maxlen, struct mempolicy *pol)
2831{
2832 char *p = buffer;
2833 nodemask_t nodes = NODE_MASK_NONE;
2834 unsigned short mode = MPOL_DEFAULT;
2835 unsigned short flags = 0;
2836
2837 if (pol && pol != &default_policy && !(pol->flags & MPOL_F_MORON)) {
2838 mode = pol->mode;
2839 flags = pol->flags;
2840 }
2841
2842 switch (mode) {
2843 case MPOL_DEFAULT:
2844 break;
2845 case MPOL_PREFERRED:
2846 if (flags & MPOL_F_LOCAL)
2847 mode = MPOL_LOCAL;
2848 else
2849 node_set(pol->v.preferred_node, nodes);
2850 break;
2851 case MPOL_BIND:
2852 case MPOL_INTERLEAVE:
2853 nodes = pol->v.nodes;
2854 break;
2855 default:
2856 WARN_ON_ONCE(1);
2857 snprintf(p, maxlen, "unknown");
2858 return;
2859 }
2860
2861 p += snprintf(p, maxlen, "%s", policy_modes[mode]);
2862
2863 if (flags & MPOL_MODE_FLAGS) {
2864 p += snprintf(p, buffer + maxlen - p, "=");
2865
2866 /*
2867 * Currently, the only defined flags are mutually exclusive
2868 */
2869 if (flags & MPOL_F_STATIC_NODES)
2870 p += snprintf(p, buffer + maxlen - p, "static");
2871 else if (flags & MPOL_F_RELATIVE_NODES)
2872 p += snprintf(p, buffer + maxlen - p, "relative");
2873 }
2874
2875 if (!nodes_empty(nodes))
2876 p += scnprintf(p, buffer + maxlen - p, ":%*pbl",
2877 nodemask_pr_args(&nodes));
2878}
1/*
2 * Simple NUMA memory policy for the Linux kernel.
3 *
4 * Copyright 2003,2004 Andi Kleen, SuSE Labs.
5 * (C) Copyright 2005 Christoph Lameter, Silicon Graphics, Inc.
6 * Subject to the GNU Public License, version 2.
7 *
8 * NUMA policy allows the user to give hints in which node(s) memory should
9 * be allocated.
10 *
11 * Support four policies per VMA and per process:
12 *
13 * The VMA policy has priority over the process policy for a page fault.
14 *
15 * interleave Allocate memory interleaved over a set of nodes,
16 * with normal fallback if it fails.
17 * For VMA based allocations this interleaves based on the
18 * offset into the backing object or offset into the mapping
19 * for anonymous memory. For process policy an process counter
20 * is used.
21 *
22 * bind Only allocate memory on a specific set of nodes,
23 * no fallback.
24 * FIXME: memory is allocated starting with the first node
25 * to the last. It would be better if bind would truly restrict
26 * the allocation to memory nodes instead
27 *
28 * preferred Try a specific node first before normal fallback.
29 * As a special case node -1 here means do the allocation
30 * on the local CPU. This is normally identical to default,
31 * but useful to set in a VMA when you have a non default
32 * process policy.
33 *
34 * default Allocate on the local node first, or when on a VMA
35 * use the process policy. This is what Linux always did
36 * in a NUMA aware kernel and still does by, ahem, default.
37 *
38 * The process policy is applied for most non interrupt memory allocations
39 * in that process' context. Interrupts ignore the policies and always
40 * try to allocate on the local CPU. The VMA policy is only applied for memory
41 * allocations for a VMA in the VM.
42 *
43 * Currently there are a few corner cases in swapping where the policy
44 * is not applied, but the majority should be handled. When process policy
45 * is used it is not remembered over swap outs/swap ins.
46 *
47 * Only the highest zone in the zone hierarchy gets policied. Allocations
48 * requesting a lower zone just use default policy. This implies that
49 * on systems with highmem kernel lowmem allocation don't get policied.
50 * Same with GFP_DMA allocations.
51 *
52 * For shmfs/tmpfs/hugetlbfs shared memory the policy is shared between
53 * all users and remembered even when nobody has memory mapped.
54 */
55
56/* Notebook:
57 fix mmap readahead to honour policy and enable policy for any page cache
58 object
59 statistics for bigpages
60 global policy for page cache? currently it uses process policy. Requires
61 first item above.
62 handle mremap for shared memory (currently ignored for the policy)
63 grows down?
64 make bind policy root only? It can trigger oom much faster and the
65 kernel is not always grateful with that.
66*/
67
68#include <linux/mempolicy.h>
69#include <linux/mm.h>
70#include <linux/highmem.h>
71#include <linux/hugetlb.h>
72#include <linux/kernel.h>
73#include <linux/sched.h>
74#include <linux/nodemask.h>
75#include <linux/cpuset.h>
76#include <linux/slab.h>
77#include <linux/string.h>
78#include <linux/export.h>
79#include <linux/nsproxy.h>
80#include <linux/interrupt.h>
81#include <linux/init.h>
82#include <linux/compat.h>
83#include <linux/swap.h>
84#include <linux/seq_file.h>
85#include <linux/proc_fs.h>
86#include <linux/migrate.h>
87#include <linux/ksm.h>
88#include <linux/rmap.h>
89#include <linux/security.h>
90#include <linux/syscalls.h>
91#include <linux/ctype.h>
92#include <linux/mm_inline.h>
93
94#include <asm/tlbflush.h>
95#include <asm/uaccess.h>
96#include <linux/random.h>
97
98#include "internal.h"
99
100/* Internal flags */
101#define MPOL_MF_DISCONTIG_OK (MPOL_MF_INTERNAL << 0) /* Skip checks for continuous vmas */
102#define MPOL_MF_INVERT (MPOL_MF_INTERNAL << 1) /* Invert check for nodemask */
103
104static struct kmem_cache *policy_cache;
105static struct kmem_cache *sn_cache;
106
107/* Highest zone. An specific allocation for a zone below that is not
108 policied. */
109enum zone_type policy_zone = 0;
110
111/*
112 * run-time system-wide default policy => local allocation
113 */
114static struct mempolicy default_policy = {
115 .refcnt = ATOMIC_INIT(1), /* never free it */
116 .mode = MPOL_PREFERRED,
117 .flags = MPOL_F_LOCAL,
118};
119
120static const struct mempolicy_operations {
121 int (*create)(struct mempolicy *pol, const nodemask_t *nodes);
122 /*
123 * If read-side task has no lock to protect task->mempolicy, write-side
124 * task will rebind the task->mempolicy by two step. The first step is
125 * setting all the newly nodes, and the second step is cleaning all the
126 * disallowed nodes. In this way, we can avoid finding no node to alloc
127 * page.
128 * If we have a lock to protect task->mempolicy in read-side, we do
129 * rebind directly.
130 *
131 * step:
132 * MPOL_REBIND_ONCE - do rebind work at once
133 * MPOL_REBIND_STEP1 - set all the newly nodes
134 * MPOL_REBIND_STEP2 - clean all the disallowed nodes
135 */
136 void (*rebind)(struct mempolicy *pol, const nodemask_t *nodes,
137 enum mpol_rebind_step step);
138} mpol_ops[MPOL_MAX];
139
140/* Check that the nodemask contains at least one populated zone */
141static int is_valid_nodemask(const nodemask_t *nodemask)
142{
143 int nd, k;
144
145 for_each_node_mask(nd, *nodemask) {
146 struct zone *z;
147
148 for (k = 0; k <= policy_zone; k++) {
149 z = &NODE_DATA(nd)->node_zones[k];
150 if (z->present_pages > 0)
151 return 1;
152 }
153 }
154
155 return 0;
156}
157
158static inline int mpol_store_user_nodemask(const struct mempolicy *pol)
159{
160 return pol->flags & MPOL_MODE_FLAGS;
161}
162
163static void mpol_relative_nodemask(nodemask_t *ret, const nodemask_t *orig,
164 const nodemask_t *rel)
165{
166 nodemask_t tmp;
167 nodes_fold(tmp, *orig, nodes_weight(*rel));
168 nodes_onto(*ret, tmp, *rel);
169}
170
171static int mpol_new_interleave(struct mempolicy *pol, const nodemask_t *nodes)
172{
173 if (nodes_empty(*nodes))
174 return -EINVAL;
175 pol->v.nodes = *nodes;
176 return 0;
177}
178
179static int mpol_new_preferred(struct mempolicy *pol, const nodemask_t *nodes)
180{
181 if (!nodes)
182 pol->flags |= MPOL_F_LOCAL; /* local allocation */
183 else if (nodes_empty(*nodes))
184 return -EINVAL; /* no allowed nodes */
185 else
186 pol->v.preferred_node = first_node(*nodes);
187 return 0;
188}
189
190static int mpol_new_bind(struct mempolicy *pol, const nodemask_t *nodes)
191{
192 if (!is_valid_nodemask(nodes))
193 return -EINVAL;
194 pol->v.nodes = *nodes;
195 return 0;
196}
197
198/*
199 * mpol_set_nodemask is called after mpol_new() to set up the nodemask, if
200 * any, for the new policy. mpol_new() has already validated the nodes
201 * parameter with respect to the policy mode and flags. But, we need to
202 * handle an empty nodemask with MPOL_PREFERRED here.
203 *
204 * Must be called holding task's alloc_lock to protect task's mems_allowed
205 * and mempolicy. May also be called holding the mmap_semaphore for write.
206 */
207static int mpol_set_nodemask(struct mempolicy *pol,
208 const nodemask_t *nodes, struct nodemask_scratch *nsc)
209{
210 int ret;
211
212 /* if mode is MPOL_DEFAULT, pol is NULL. This is right. */
213 if (pol == NULL)
214 return 0;
215 /* Check N_HIGH_MEMORY */
216 nodes_and(nsc->mask1,
217 cpuset_current_mems_allowed, node_states[N_HIGH_MEMORY]);
218
219 VM_BUG_ON(!nodes);
220 if (pol->mode == MPOL_PREFERRED && nodes_empty(*nodes))
221 nodes = NULL; /* explicit local allocation */
222 else {
223 if (pol->flags & MPOL_F_RELATIVE_NODES)
224 mpol_relative_nodemask(&nsc->mask2, nodes,&nsc->mask1);
225 else
226 nodes_and(nsc->mask2, *nodes, nsc->mask1);
227
228 if (mpol_store_user_nodemask(pol))
229 pol->w.user_nodemask = *nodes;
230 else
231 pol->w.cpuset_mems_allowed =
232 cpuset_current_mems_allowed;
233 }
234
235 if (nodes)
236 ret = mpol_ops[pol->mode].create(pol, &nsc->mask2);
237 else
238 ret = mpol_ops[pol->mode].create(pol, NULL);
239 return ret;
240}
241
242/*
243 * This function just creates a new policy, does some check and simple
244 * initialization. You must invoke mpol_set_nodemask() to set nodes.
245 */
246static struct mempolicy *mpol_new(unsigned short mode, unsigned short flags,
247 nodemask_t *nodes)
248{
249 struct mempolicy *policy;
250
251 pr_debug("setting mode %d flags %d nodes[0] %lx\n",
252 mode, flags, nodes ? nodes_addr(*nodes)[0] : -1);
253
254 if (mode == MPOL_DEFAULT) {
255 if (nodes && !nodes_empty(*nodes))
256 return ERR_PTR(-EINVAL);
257 return NULL; /* simply delete any existing policy */
258 }
259 VM_BUG_ON(!nodes);
260
261 /*
262 * MPOL_PREFERRED cannot be used with MPOL_F_STATIC_NODES or
263 * MPOL_F_RELATIVE_NODES if the nodemask is empty (local allocation).
264 * All other modes require a valid pointer to a non-empty nodemask.
265 */
266 if (mode == MPOL_PREFERRED) {
267 if (nodes_empty(*nodes)) {
268 if (((flags & MPOL_F_STATIC_NODES) ||
269 (flags & MPOL_F_RELATIVE_NODES)))
270 return ERR_PTR(-EINVAL);
271 }
272 } else if (nodes_empty(*nodes))
273 return ERR_PTR(-EINVAL);
274 policy = kmem_cache_alloc(policy_cache, GFP_KERNEL);
275 if (!policy)
276 return ERR_PTR(-ENOMEM);
277 atomic_set(&policy->refcnt, 1);
278 policy->mode = mode;
279 policy->flags = flags;
280
281 return policy;
282}
283
284/* Slow path of a mpol destructor. */
285void __mpol_put(struct mempolicy *p)
286{
287 if (!atomic_dec_and_test(&p->refcnt))
288 return;
289 kmem_cache_free(policy_cache, p);
290}
291
292static void mpol_rebind_default(struct mempolicy *pol, const nodemask_t *nodes,
293 enum mpol_rebind_step step)
294{
295}
296
297/*
298 * step:
299 * MPOL_REBIND_ONCE - do rebind work at once
300 * MPOL_REBIND_STEP1 - set all the newly nodes
301 * MPOL_REBIND_STEP2 - clean all the disallowed nodes
302 */
303static void mpol_rebind_nodemask(struct mempolicy *pol, const nodemask_t *nodes,
304 enum mpol_rebind_step step)
305{
306 nodemask_t tmp;
307
308 if (pol->flags & MPOL_F_STATIC_NODES)
309 nodes_and(tmp, pol->w.user_nodemask, *nodes);
310 else if (pol->flags & MPOL_F_RELATIVE_NODES)
311 mpol_relative_nodemask(&tmp, &pol->w.user_nodemask, nodes);
312 else {
313 /*
314 * if step == 1, we use ->w.cpuset_mems_allowed to cache the
315 * result
316 */
317 if (step == MPOL_REBIND_ONCE || step == MPOL_REBIND_STEP1) {
318 nodes_remap(tmp, pol->v.nodes,
319 pol->w.cpuset_mems_allowed, *nodes);
320 pol->w.cpuset_mems_allowed = step ? tmp : *nodes;
321 } else if (step == MPOL_REBIND_STEP2) {
322 tmp = pol->w.cpuset_mems_allowed;
323 pol->w.cpuset_mems_allowed = *nodes;
324 } else
325 BUG();
326 }
327
328 if (nodes_empty(tmp))
329 tmp = *nodes;
330
331 if (step == MPOL_REBIND_STEP1)
332 nodes_or(pol->v.nodes, pol->v.nodes, tmp);
333 else if (step == MPOL_REBIND_ONCE || step == MPOL_REBIND_STEP2)
334 pol->v.nodes = tmp;
335 else
336 BUG();
337
338 if (!node_isset(current->il_next, tmp)) {
339 current->il_next = next_node(current->il_next, tmp);
340 if (current->il_next >= MAX_NUMNODES)
341 current->il_next = first_node(tmp);
342 if (current->il_next >= MAX_NUMNODES)
343 current->il_next = numa_node_id();
344 }
345}
346
347static void mpol_rebind_preferred(struct mempolicy *pol,
348 const nodemask_t *nodes,
349 enum mpol_rebind_step step)
350{
351 nodemask_t tmp;
352
353 if (pol->flags & MPOL_F_STATIC_NODES) {
354 int node = first_node(pol->w.user_nodemask);
355
356 if (node_isset(node, *nodes)) {
357 pol->v.preferred_node = node;
358 pol->flags &= ~MPOL_F_LOCAL;
359 } else
360 pol->flags |= MPOL_F_LOCAL;
361 } else if (pol->flags & MPOL_F_RELATIVE_NODES) {
362 mpol_relative_nodemask(&tmp, &pol->w.user_nodemask, nodes);
363 pol->v.preferred_node = first_node(tmp);
364 } else if (!(pol->flags & MPOL_F_LOCAL)) {
365 pol->v.preferred_node = node_remap(pol->v.preferred_node,
366 pol->w.cpuset_mems_allowed,
367 *nodes);
368 pol->w.cpuset_mems_allowed = *nodes;
369 }
370}
371
372/*
373 * mpol_rebind_policy - Migrate a policy to a different set of nodes
374 *
375 * If read-side task has no lock to protect task->mempolicy, write-side
376 * task will rebind the task->mempolicy by two step. The first step is
377 * setting all the newly nodes, and the second step is cleaning all the
378 * disallowed nodes. In this way, we can avoid finding no node to alloc
379 * page.
380 * If we have a lock to protect task->mempolicy in read-side, we do
381 * rebind directly.
382 *
383 * step:
384 * MPOL_REBIND_ONCE - do rebind work at once
385 * MPOL_REBIND_STEP1 - set all the newly nodes
386 * MPOL_REBIND_STEP2 - clean all the disallowed nodes
387 */
388static void mpol_rebind_policy(struct mempolicy *pol, const nodemask_t *newmask,
389 enum mpol_rebind_step step)
390{
391 if (!pol)
392 return;
393 if (!mpol_store_user_nodemask(pol) && step == MPOL_REBIND_ONCE &&
394 nodes_equal(pol->w.cpuset_mems_allowed, *newmask))
395 return;
396
397 if (step == MPOL_REBIND_STEP1 && (pol->flags & MPOL_F_REBINDING))
398 return;
399
400 if (step == MPOL_REBIND_STEP2 && !(pol->flags & MPOL_F_REBINDING))
401 BUG();
402
403 if (step == MPOL_REBIND_STEP1)
404 pol->flags |= MPOL_F_REBINDING;
405 else if (step == MPOL_REBIND_STEP2)
406 pol->flags &= ~MPOL_F_REBINDING;
407 else if (step >= MPOL_REBIND_NSTEP)
408 BUG();
409
410 mpol_ops[pol->mode].rebind(pol, newmask, step);
411}
412
413/*
414 * Wrapper for mpol_rebind_policy() that just requires task
415 * pointer, and updates task mempolicy.
416 *
417 * Called with task's alloc_lock held.
418 */
419
420void mpol_rebind_task(struct task_struct *tsk, const nodemask_t *new,
421 enum mpol_rebind_step step)
422{
423 mpol_rebind_policy(tsk->mempolicy, new, step);
424}
425
426/*
427 * Rebind each vma in mm to new nodemask.
428 *
429 * Call holding a reference to mm. Takes mm->mmap_sem during call.
430 */
431
432void mpol_rebind_mm(struct mm_struct *mm, nodemask_t *new)
433{
434 struct vm_area_struct *vma;
435
436 down_write(&mm->mmap_sem);
437 for (vma = mm->mmap; vma; vma = vma->vm_next)
438 mpol_rebind_policy(vma->vm_policy, new, MPOL_REBIND_ONCE);
439 up_write(&mm->mmap_sem);
440}
441
442static const struct mempolicy_operations mpol_ops[MPOL_MAX] = {
443 [MPOL_DEFAULT] = {
444 .rebind = mpol_rebind_default,
445 },
446 [MPOL_INTERLEAVE] = {
447 .create = mpol_new_interleave,
448 .rebind = mpol_rebind_nodemask,
449 },
450 [MPOL_PREFERRED] = {
451 .create = mpol_new_preferred,
452 .rebind = mpol_rebind_preferred,
453 },
454 [MPOL_BIND] = {
455 .create = mpol_new_bind,
456 .rebind = mpol_rebind_nodemask,
457 },
458};
459
460static void migrate_page_add(struct page *page, struct list_head *pagelist,
461 unsigned long flags);
462
463/* Scan through pages checking if pages follow certain conditions. */
464static int check_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
465 unsigned long addr, unsigned long end,
466 const nodemask_t *nodes, unsigned long flags,
467 void *private)
468{
469 pte_t *orig_pte;
470 pte_t *pte;
471 spinlock_t *ptl;
472
473 orig_pte = pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
474 do {
475 struct page *page;
476 int nid;
477
478 if (!pte_present(*pte))
479 continue;
480 page = vm_normal_page(vma, addr, *pte);
481 if (!page)
482 continue;
483 /*
484 * vm_normal_page() filters out zero pages, but there might
485 * still be PageReserved pages to skip, perhaps in a VDSO.
486 * And we cannot move PageKsm pages sensibly or safely yet.
487 */
488 if (PageReserved(page) || PageKsm(page))
489 continue;
490 nid = page_to_nid(page);
491 if (node_isset(nid, *nodes) == !!(flags & MPOL_MF_INVERT))
492 continue;
493
494 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))
495 migrate_page_add(page, private, flags);
496 else
497 break;
498 } while (pte++, addr += PAGE_SIZE, addr != end);
499 pte_unmap_unlock(orig_pte, ptl);
500 return addr != end;
501}
502
503static inline int check_pmd_range(struct vm_area_struct *vma, pud_t *pud,
504 unsigned long addr, unsigned long end,
505 const nodemask_t *nodes, unsigned long flags,
506 void *private)
507{
508 pmd_t *pmd;
509 unsigned long next;
510
511 pmd = pmd_offset(pud, addr);
512 do {
513 next = pmd_addr_end(addr, end);
514 split_huge_page_pmd(vma->vm_mm, pmd);
515 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
516 continue;
517 if (check_pte_range(vma, pmd, addr, next, nodes,
518 flags, private))
519 return -EIO;
520 } while (pmd++, addr = next, addr != end);
521 return 0;
522}
523
524static inline int check_pud_range(struct vm_area_struct *vma, pgd_t *pgd,
525 unsigned long addr, unsigned long end,
526 const nodemask_t *nodes, unsigned long flags,
527 void *private)
528{
529 pud_t *pud;
530 unsigned long next;
531
532 pud = pud_offset(pgd, addr);
533 do {
534 next = pud_addr_end(addr, end);
535 if (pud_none_or_clear_bad(pud))
536 continue;
537 if (check_pmd_range(vma, pud, addr, next, nodes,
538 flags, private))
539 return -EIO;
540 } while (pud++, addr = next, addr != end);
541 return 0;
542}
543
544static inline int check_pgd_range(struct vm_area_struct *vma,
545 unsigned long addr, unsigned long end,
546 const nodemask_t *nodes, unsigned long flags,
547 void *private)
548{
549 pgd_t *pgd;
550 unsigned long next;
551
552 pgd = pgd_offset(vma->vm_mm, addr);
553 do {
554 next = pgd_addr_end(addr, end);
555 if (pgd_none_or_clear_bad(pgd))
556 continue;
557 if (check_pud_range(vma, pgd, addr, next, nodes,
558 flags, private))
559 return -EIO;
560 } while (pgd++, addr = next, addr != end);
561 return 0;
562}
563
564/*
565 * Check if all pages in a range are on a set of nodes.
566 * If pagelist != NULL then isolate pages from the LRU and
567 * put them on the pagelist.
568 */
569static struct vm_area_struct *
570check_range(struct mm_struct *mm, unsigned long start, unsigned long end,
571 const nodemask_t *nodes, unsigned long flags, void *private)
572{
573 int err;
574 struct vm_area_struct *first, *vma, *prev;
575
576
577 first = find_vma(mm, start);
578 if (!first)
579 return ERR_PTR(-EFAULT);
580 prev = NULL;
581 for (vma = first; vma && vma->vm_start < end; vma = vma->vm_next) {
582 if (!(flags & MPOL_MF_DISCONTIG_OK)) {
583 if (!vma->vm_next && vma->vm_end < end)
584 return ERR_PTR(-EFAULT);
585 if (prev && prev->vm_end < vma->vm_start)
586 return ERR_PTR(-EFAULT);
587 }
588 if (!is_vm_hugetlb_page(vma) &&
589 ((flags & MPOL_MF_STRICT) ||
590 ((flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) &&
591 vma_migratable(vma)))) {
592 unsigned long endvma = vma->vm_end;
593
594 if (endvma > end)
595 endvma = end;
596 if (vma->vm_start > start)
597 start = vma->vm_start;
598 err = check_pgd_range(vma, start, endvma, nodes,
599 flags, private);
600 if (err) {
601 first = ERR_PTR(err);
602 break;
603 }
604 }
605 prev = vma;
606 }
607 return first;
608}
609
610/* Step 2: apply policy to a range and do splits. */
611static int mbind_range(struct mm_struct *mm, unsigned long start,
612 unsigned long end, struct mempolicy *new_pol)
613{
614 struct vm_area_struct *next;
615 struct vm_area_struct *prev;
616 struct vm_area_struct *vma;
617 int err = 0;
618 pgoff_t pgoff;
619 unsigned long vmstart;
620 unsigned long vmend;
621
622 vma = find_vma(mm, start);
623 if (!vma || vma->vm_start > start)
624 return -EFAULT;
625
626 prev = vma->vm_prev;
627 if (start > vma->vm_start)
628 prev = vma;
629
630 for (; vma && vma->vm_start < end; prev = vma, vma = next) {
631 next = vma->vm_next;
632 vmstart = max(start, vma->vm_start);
633 vmend = min(end, vma->vm_end);
634
635 if (mpol_equal(vma_policy(vma), new_pol))
636 continue;
637
638 pgoff = vma->vm_pgoff +
639 ((vmstart - vma->vm_start) >> PAGE_SHIFT);
640 prev = vma_merge(mm, prev, vmstart, vmend, vma->vm_flags,
641 vma->anon_vma, vma->vm_file, pgoff,
642 new_pol);
643 if (prev) {
644 vma = prev;
645 next = vma->vm_next;
646 continue;
647 }
648 if (vma->vm_start != vmstart) {
649 err = split_vma(vma->vm_mm, vma, vmstart, 1);
650 if (err)
651 goto out;
652 }
653 if (vma->vm_end != vmend) {
654 err = split_vma(vma->vm_mm, vma, vmend, 0);
655 if (err)
656 goto out;
657 }
658
659 /*
660 * Apply policy to a single VMA. The reference counting of
661 * policy for vma_policy linkages has already been handled by
662 * vma_merge and split_vma as necessary. If this is a shared
663 * policy then ->set_policy will increment the reference count
664 * for an sp node.
665 */
666 pr_debug("vma %lx-%lx/%lx vm_ops %p vm_file %p set_policy %p\n",
667 vma->vm_start, vma->vm_end, vma->vm_pgoff,
668 vma->vm_ops, vma->vm_file,
669 vma->vm_ops ? vma->vm_ops->set_policy : NULL);
670 if (vma->vm_ops && vma->vm_ops->set_policy) {
671 err = vma->vm_ops->set_policy(vma, new_pol);
672 if (err)
673 goto out;
674 }
675 }
676
677 out:
678 return err;
679}
680
681/*
682 * Update task->flags PF_MEMPOLICY bit: set iff non-default
683 * mempolicy. Allows more rapid checking of this (combined perhaps
684 * with other PF_* flag bits) on memory allocation hot code paths.
685 *
686 * If called from outside this file, the task 'p' should -only- be
687 * a newly forked child not yet visible on the task list, because
688 * manipulating the task flags of a visible task is not safe.
689 *
690 * The above limitation is why this routine has the funny name
691 * mpol_fix_fork_child_flag().
692 *
693 * It is also safe to call this with a task pointer of current,
694 * which the static wrapper mpol_set_task_struct_flag() does,
695 * for use within this file.
696 */
697
698void mpol_fix_fork_child_flag(struct task_struct *p)
699{
700 if (p->mempolicy)
701 p->flags |= PF_MEMPOLICY;
702 else
703 p->flags &= ~PF_MEMPOLICY;
704}
705
706static void mpol_set_task_struct_flag(void)
707{
708 mpol_fix_fork_child_flag(current);
709}
710
711/* Set the process memory policy */
712static long do_set_mempolicy(unsigned short mode, unsigned short flags,
713 nodemask_t *nodes)
714{
715 struct mempolicy *new, *old;
716 struct mm_struct *mm = current->mm;
717 NODEMASK_SCRATCH(scratch);
718 int ret;
719
720 if (!scratch)
721 return -ENOMEM;
722
723 new = mpol_new(mode, flags, nodes);
724 if (IS_ERR(new)) {
725 ret = PTR_ERR(new);
726 goto out;
727 }
728 /*
729 * prevent changing our mempolicy while show_numa_maps()
730 * is using it.
731 * Note: do_set_mempolicy() can be called at init time
732 * with no 'mm'.
733 */
734 if (mm)
735 down_write(&mm->mmap_sem);
736 task_lock(current);
737 ret = mpol_set_nodemask(new, nodes, scratch);
738 if (ret) {
739 task_unlock(current);
740 if (mm)
741 up_write(&mm->mmap_sem);
742 mpol_put(new);
743 goto out;
744 }
745 old = current->mempolicy;
746 current->mempolicy = new;
747 mpol_set_task_struct_flag();
748 if (new && new->mode == MPOL_INTERLEAVE &&
749 nodes_weight(new->v.nodes))
750 current->il_next = first_node(new->v.nodes);
751 task_unlock(current);
752 if (mm)
753 up_write(&mm->mmap_sem);
754
755 mpol_put(old);
756 ret = 0;
757out:
758 NODEMASK_SCRATCH_FREE(scratch);
759 return ret;
760}
761
762/*
763 * Return nodemask for policy for get_mempolicy() query
764 *
765 * Called with task's alloc_lock held
766 */
767static void get_policy_nodemask(struct mempolicy *p, nodemask_t *nodes)
768{
769 nodes_clear(*nodes);
770 if (p == &default_policy)
771 return;
772
773 switch (p->mode) {
774 case MPOL_BIND:
775 /* Fall through */
776 case MPOL_INTERLEAVE:
777 *nodes = p->v.nodes;
778 break;
779 case MPOL_PREFERRED:
780 if (!(p->flags & MPOL_F_LOCAL))
781 node_set(p->v.preferred_node, *nodes);
782 /* else return empty node mask for local allocation */
783 break;
784 default:
785 BUG();
786 }
787}
788
789static int lookup_node(struct mm_struct *mm, unsigned long addr)
790{
791 struct page *p;
792 int err;
793
794 err = get_user_pages(current, mm, addr & PAGE_MASK, 1, 0, 0, &p, NULL);
795 if (err >= 0) {
796 err = page_to_nid(p);
797 put_page(p);
798 }
799 return err;
800}
801
802/* Retrieve NUMA policy */
803static long do_get_mempolicy(int *policy, nodemask_t *nmask,
804 unsigned long addr, unsigned long flags)
805{
806 int err;
807 struct mm_struct *mm = current->mm;
808 struct vm_area_struct *vma = NULL;
809 struct mempolicy *pol = current->mempolicy;
810
811 if (flags &
812 ~(unsigned long)(MPOL_F_NODE|MPOL_F_ADDR|MPOL_F_MEMS_ALLOWED))
813 return -EINVAL;
814
815 if (flags & MPOL_F_MEMS_ALLOWED) {
816 if (flags & (MPOL_F_NODE|MPOL_F_ADDR))
817 return -EINVAL;
818 *policy = 0; /* just so it's initialized */
819 task_lock(current);
820 *nmask = cpuset_current_mems_allowed;
821 task_unlock(current);
822 return 0;
823 }
824
825 if (flags & MPOL_F_ADDR) {
826 /*
827 * Do NOT fall back to task policy if the
828 * vma/shared policy at addr is NULL. We
829 * want to return MPOL_DEFAULT in this case.
830 */
831 down_read(&mm->mmap_sem);
832 vma = find_vma_intersection(mm, addr, addr+1);
833 if (!vma) {
834 up_read(&mm->mmap_sem);
835 return -EFAULT;
836 }
837 if (vma->vm_ops && vma->vm_ops->get_policy)
838 pol = vma->vm_ops->get_policy(vma, addr);
839 else
840 pol = vma->vm_policy;
841 } else if (addr)
842 return -EINVAL;
843
844 if (!pol)
845 pol = &default_policy; /* indicates default behavior */
846
847 if (flags & MPOL_F_NODE) {
848 if (flags & MPOL_F_ADDR) {
849 err = lookup_node(mm, addr);
850 if (err < 0)
851 goto out;
852 *policy = err;
853 } else if (pol == current->mempolicy &&
854 pol->mode == MPOL_INTERLEAVE) {
855 *policy = current->il_next;
856 } else {
857 err = -EINVAL;
858 goto out;
859 }
860 } else {
861 *policy = pol == &default_policy ? MPOL_DEFAULT :
862 pol->mode;
863 /*
864 * Internal mempolicy flags must be masked off before exposing
865 * the policy to userspace.
866 */
867 *policy |= (pol->flags & MPOL_MODE_FLAGS);
868 }
869
870 if (vma) {
871 up_read(¤t->mm->mmap_sem);
872 vma = NULL;
873 }
874
875 err = 0;
876 if (nmask) {
877 if (mpol_store_user_nodemask(pol)) {
878 *nmask = pol->w.user_nodemask;
879 } else {
880 task_lock(current);
881 get_policy_nodemask(pol, nmask);
882 task_unlock(current);
883 }
884 }
885
886 out:
887 mpol_cond_put(pol);
888 if (vma)
889 up_read(¤t->mm->mmap_sem);
890 return err;
891}
892
893#ifdef CONFIG_MIGRATION
894/*
895 * page migration
896 */
897static void migrate_page_add(struct page *page, struct list_head *pagelist,
898 unsigned long flags)
899{
900 /*
901 * Avoid migrating a page that is shared with others.
902 */
903 if ((flags & MPOL_MF_MOVE_ALL) || page_mapcount(page) == 1) {
904 if (!isolate_lru_page(page)) {
905 list_add_tail(&page->lru, pagelist);
906 inc_zone_page_state(page, NR_ISOLATED_ANON +
907 page_is_file_cache(page));
908 }
909 }
910}
911
912static struct page *new_node_page(struct page *page, unsigned long node, int **x)
913{
914 return alloc_pages_exact_node(node, GFP_HIGHUSER_MOVABLE, 0);
915}
916
917/*
918 * Migrate pages from one node to a target node.
919 * Returns error or the number of pages not migrated.
920 */
921static int migrate_to_node(struct mm_struct *mm, int source, int dest,
922 int flags)
923{
924 nodemask_t nmask;
925 LIST_HEAD(pagelist);
926 int err = 0;
927 struct vm_area_struct *vma;
928
929 nodes_clear(nmask);
930 node_set(source, nmask);
931
932 vma = check_range(mm, mm->mmap->vm_start, mm->task_size, &nmask,
933 flags | MPOL_MF_DISCONTIG_OK, &pagelist);
934 if (IS_ERR(vma))
935 return PTR_ERR(vma);
936
937 if (!list_empty(&pagelist)) {
938 err = migrate_pages(&pagelist, new_node_page, dest,
939 false, MIGRATE_SYNC);
940 if (err)
941 putback_lru_pages(&pagelist);
942 }
943
944 return err;
945}
946
947/*
948 * Move pages between the two nodesets so as to preserve the physical
949 * layout as much as possible.
950 *
951 * Returns the number of page that could not be moved.
952 */
953int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
954 const nodemask_t *to, int flags)
955{
956 int busy = 0;
957 int err;
958 nodemask_t tmp;
959
960 err = migrate_prep();
961 if (err)
962 return err;
963
964 down_read(&mm->mmap_sem);
965
966 err = migrate_vmas(mm, from, to, flags);
967 if (err)
968 goto out;
969
970 /*
971 * Find a 'source' bit set in 'tmp' whose corresponding 'dest'
972 * bit in 'to' is not also set in 'tmp'. Clear the found 'source'
973 * bit in 'tmp', and return that <source, dest> pair for migration.
974 * The pair of nodemasks 'to' and 'from' define the map.
975 *
976 * If no pair of bits is found that way, fallback to picking some
977 * pair of 'source' and 'dest' bits that are not the same. If the
978 * 'source' and 'dest' bits are the same, this represents a node
979 * that will be migrating to itself, so no pages need move.
980 *
981 * If no bits are left in 'tmp', or if all remaining bits left
982 * in 'tmp' correspond to the same bit in 'to', return false
983 * (nothing left to migrate).
984 *
985 * This lets us pick a pair of nodes to migrate between, such that
986 * if possible the dest node is not already occupied by some other
987 * source node, minimizing the risk of overloading the memory on a
988 * node that would happen if we migrated incoming memory to a node
989 * before migrating outgoing memory source that same node.
990 *
991 * A single scan of tmp is sufficient. As we go, we remember the
992 * most recent <s, d> pair that moved (s != d). If we find a pair
993 * that not only moved, but what's better, moved to an empty slot
994 * (d is not set in tmp), then we break out then, with that pair.
995 * Otherwise when we finish scanning from_tmp, we at least have the
996 * most recent <s, d> pair that moved. If we get all the way through
997 * the scan of tmp without finding any node that moved, much less
998 * moved to an empty node, then there is nothing left worth migrating.
999 */
1000
1001 tmp = *from;
1002 while (!nodes_empty(tmp)) {
1003 int s,d;
1004 int source = -1;
1005 int dest = 0;
1006
1007 for_each_node_mask(s, tmp) {
1008
1009 /*
1010 * do_migrate_pages() tries to maintain the relative
1011 * node relationship of the pages established between
1012 * threads and memory areas.
1013 *
1014 * However if the number of source nodes is not equal to
1015 * the number of destination nodes we can not preserve
1016 * this node relative relationship. In that case, skip
1017 * copying memory from a node that is in the destination
1018 * mask.
1019 *
1020 * Example: [2,3,4] -> [3,4,5] moves everything.
1021 * [0-7] - > [3,4,5] moves only 0,1,2,6,7.
1022 */
1023
1024 if ((nodes_weight(*from) != nodes_weight(*to)) &&
1025 (node_isset(s, *to)))
1026 continue;
1027
1028 d = node_remap(s, *from, *to);
1029 if (s == d)
1030 continue;
1031
1032 source = s; /* Node moved. Memorize */
1033 dest = d;
1034
1035 /* dest not in remaining from nodes? */
1036 if (!node_isset(dest, tmp))
1037 break;
1038 }
1039 if (source == -1)
1040 break;
1041
1042 node_clear(source, tmp);
1043 err = migrate_to_node(mm, source, dest, flags);
1044 if (err > 0)
1045 busy += err;
1046 if (err < 0)
1047 break;
1048 }
1049out:
1050 up_read(&mm->mmap_sem);
1051 if (err < 0)
1052 return err;
1053 return busy;
1054
1055}
1056
1057/*
1058 * Allocate a new page for page migration based on vma policy.
1059 * Start assuming that page is mapped by vma pointed to by @private.
1060 * Search forward from there, if not. N.B., this assumes that the
1061 * list of pages handed to migrate_pages()--which is how we get here--
1062 * is in virtual address order.
1063 */
1064static struct page *new_vma_page(struct page *page, unsigned long private, int **x)
1065{
1066 struct vm_area_struct *vma = (struct vm_area_struct *)private;
1067 unsigned long uninitialized_var(address);
1068
1069 while (vma) {
1070 address = page_address_in_vma(page, vma);
1071 if (address != -EFAULT)
1072 break;
1073 vma = vma->vm_next;
1074 }
1075
1076 /*
1077 * if !vma, alloc_page_vma() will use task or system default policy
1078 */
1079 return alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
1080}
1081#else
1082
1083static void migrate_page_add(struct page *page, struct list_head *pagelist,
1084 unsigned long flags)
1085{
1086}
1087
1088int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
1089 const nodemask_t *to, int flags)
1090{
1091 return -ENOSYS;
1092}
1093
1094static struct page *new_vma_page(struct page *page, unsigned long private, int **x)
1095{
1096 return NULL;
1097}
1098#endif
1099
1100static long do_mbind(unsigned long start, unsigned long len,
1101 unsigned short mode, unsigned short mode_flags,
1102 nodemask_t *nmask, unsigned long flags)
1103{
1104 struct vm_area_struct *vma;
1105 struct mm_struct *mm = current->mm;
1106 struct mempolicy *new;
1107 unsigned long end;
1108 int err;
1109 LIST_HEAD(pagelist);
1110
1111 if (flags & ~(unsigned long)(MPOL_MF_STRICT |
1112 MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))
1113 return -EINVAL;
1114 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1115 return -EPERM;
1116
1117 if (start & ~PAGE_MASK)
1118 return -EINVAL;
1119
1120 if (mode == MPOL_DEFAULT)
1121 flags &= ~MPOL_MF_STRICT;
1122
1123 len = (len + PAGE_SIZE - 1) & PAGE_MASK;
1124 end = start + len;
1125
1126 if (end < start)
1127 return -EINVAL;
1128 if (end == start)
1129 return 0;
1130
1131 new = mpol_new(mode, mode_flags, nmask);
1132 if (IS_ERR(new))
1133 return PTR_ERR(new);
1134
1135 /*
1136 * If we are using the default policy then operation
1137 * on discontinuous address spaces is okay after all
1138 */
1139 if (!new)
1140 flags |= MPOL_MF_DISCONTIG_OK;
1141
1142 pr_debug("mbind %lx-%lx mode:%d flags:%d nodes:%lx\n",
1143 start, start + len, mode, mode_flags,
1144 nmask ? nodes_addr(*nmask)[0] : -1);
1145
1146 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) {
1147
1148 err = migrate_prep();
1149 if (err)
1150 goto mpol_out;
1151 }
1152 {
1153 NODEMASK_SCRATCH(scratch);
1154 if (scratch) {
1155 down_write(&mm->mmap_sem);
1156 task_lock(current);
1157 err = mpol_set_nodemask(new, nmask, scratch);
1158 task_unlock(current);
1159 if (err)
1160 up_write(&mm->mmap_sem);
1161 } else
1162 err = -ENOMEM;
1163 NODEMASK_SCRATCH_FREE(scratch);
1164 }
1165 if (err)
1166 goto mpol_out;
1167
1168 vma = check_range(mm, start, end, nmask,
1169 flags | MPOL_MF_INVERT, &pagelist);
1170
1171 err = PTR_ERR(vma);
1172 if (!IS_ERR(vma)) {
1173 int nr_failed = 0;
1174
1175 err = mbind_range(mm, start, end, new);
1176
1177 if (!list_empty(&pagelist)) {
1178 nr_failed = migrate_pages(&pagelist, new_vma_page,
1179 (unsigned long)vma,
1180 false, MIGRATE_SYNC);
1181 if (nr_failed)
1182 putback_lru_pages(&pagelist);
1183 }
1184
1185 if (!err && nr_failed && (flags & MPOL_MF_STRICT))
1186 err = -EIO;
1187 } else
1188 putback_lru_pages(&pagelist);
1189
1190 up_write(&mm->mmap_sem);
1191 mpol_out:
1192 mpol_put(new);
1193 return err;
1194}
1195
1196/*
1197 * User space interface with variable sized bitmaps for nodelists.
1198 */
1199
1200/* Copy a node mask from user space. */
1201static int get_nodes(nodemask_t *nodes, const unsigned long __user *nmask,
1202 unsigned long maxnode)
1203{
1204 unsigned long k;
1205 unsigned long nlongs;
1206 unsigned long endmask;
1207
1208 --maxnode;
1209 nodes_clear(*nodes);
1210 if (maxnode == 0 || !nmask)
1211 return 0;
1212 if (maxnode > PAGE_SIZE*BITS_PER_BYTE)
1213 return -EINVAL;
1214
1215 nlongs = BITS_TO_LONGS(maxnode);
1216 if ((maxnode % BITS_PER_LONG) == 0)
1217 endmask = ~0UL;
1218 else
1219 endmask = (1UL << (maxnode % BITS_PER_LONG)) - 1;
1220
1221 /* When the user specified more nodes than supported just check
1222 if the non supported part is all zero. */
1223 if (nlongs > BITS_TO_LONGS(MAX_NUMNODES)) {
1224 if (nlongs > PAGE_SIZE/sizeof(long))
1225 return -EINVAL;
1226 for (k = BITS_TO_LONGS(MAX_NUMNODES); k < nlongs; k++) {
1227 unsigned long t;
1228 if (get_user(t, nmask + k))
1229 return -EFAULT;
1230 if (k == nlongs - 1) {
1231 if (t & endmask)
1232 return -EINVAL;
1233 } else if (t)
1234 return -EINVAL;
1235 }
1236 nlongs = BITS_TO_LONGS(MAX_NUMNODES);
1237 endmask = ~0UL;
1238 }
1239
1240 if (copy_from_user(nodes_addr(*nodes), nmask, nlongs*sizeof(unsigned long)))
1241 return -EFAULT;
1242 nodes_addr(*nodes)[nlongs-1] &= endmask;
1243 return 0;
1244}
1245
1246/* Copy a kernel node mask to user space */
1247static int copy_nodes_to_user(unsigned long __user *mask, unsigned long maxnode,
1248 nodemask_t *nodes)
1249{
1250 unsigned long copy = ALIGN(maxnode-1, 64) / 8;
1251 const int nbytes = BITS_TO_LONGS(MAX_NUMNODES) * sizeof(long);
1252
1253 if (copy > nbytes) {
1254 if (copy > PAGE_SIZE)
1255 return -EINVAL;
1256 if (clear_user((char __user *)mask + nbytes, copy - nbytes))
1257 return -EFAULT;
1258 copy = nbytes;
1259 }
1260 return copy_to_user(mask, nodes_addr(*nodes), copy) ? -EFAULT : 0;
1261}
1262
1263SYSCALL_DEFINE6(mbind, unsigned long, start, unsigned long, len,
1264 unsigned long, mode, unsigned long __user *, nmask,
1265 unsigned long, maxnode, unsigned, flags)
1266{
1267 nodemask_t nodes;
1268 int err;
1269 unsigned short mode_flags;
1270
1271 mode_flags = mode & MPOL_MODE_FLAGS;
1272 mode &= ~MPOL_MODE_FLAGS;
1273 if (mode >= MPOL_MAX)
1274 return -EINVAL;
1275 if ((mode_flags & MPOL_F_STATIC_NODES) &&
1276 (mode_flags & MPOL_F_RELATIVE_NODES))
1277 return -EINVAL;
1278 err = get_nodes(&nodes, nmask, maxnode);
1279 if (err)
1280 return err;
1281 return do_mbind(start, len, mode, mode_flags, &nodes, flags);
1282}
1283
1284/* Set the process memory policy */
1285SYSCALL_DEFINE3(set_mempolicy, int, mode, unsigned long __user *, nmask,
1286 unsigned long, maxnode)
1287{
1288 int err;
1289 nodemask_t nodes;
1290 unsigned short flags;
1291
1292 flags = mode & MPOL_MODE_FLAGS;
1293 mode &= ~MPOL_MODE_FLAGS;
1294 if ((unsigned int)mode >= MPOL_MAX)
1295 return -EINVAL;
1296 if ((flags & MPOL_F_STATIC_NODES) && (flags & MPOL_F_RELATIVE_NODES))
1297 return -EINVAL;
1298 err = get_nodes(&nodes, nmask, maxnode);
1299 if (err)
1300 return err;
1301 return do_set_mempolicy(mode, flags, &nodes);
1302}
1303
1304SYSCALL_DEFINE4(migrate_pages, pid_t, pid, unsigned long, maxnode,
1305 const unsigned long __user *, old_nodes,
1306 const unsigned long __user *, new_nodes)
1307{
1308 const struct cred *cred = current_cred(), *tcred;
1309 struct mm_struct *mm = NULL;
1310 struct task_struct *task;
1311 nodemask_t task_nodes;
1312 int err;
1313 nodemask_t *old;
1314 nodemask_t *new;
1315 NODEMASK_SCRATCH(scratch);
1316
1317 if (!scratch)
1318 return -ENOMEM;
1319
1320 old = &scratch->mask1;
1321 new = &scratch->mask2;
1322
1323 err = get_nodes(old, old_nodes, maxnode);
1324 if (err)
1325 goto out;
1326
1327 err = get_nodes(new, new_nodes, maxnode);
1328 if (err)
1329 goto out;
1330
1331 /* Find the mm_struct */
1332 rcu_read_lock();
1333 task = pid ? find_task_by_vpid(pid) : current;
1334 if (!task) {
1335 rcu_read_unlock();
1336 err = -ESRCH;
1337 goto out;
1338 }
1339 get_task_struct(task);
1340
1341 err = -EINVAL;
1342
1343 /*
1344 * Check if this process has the right to modify the specified
1345 * process. The right exists if the process has administrative
1346 * capabilities, superuser privileges or the same
1347 * userid as the target process.
1348 */
1349 tcred = __task_cred(task);
1350 if (!uid_eq(cred->euid, tcred->suid) && !uid_eq(cred->euid, tcred->uid) &&
1351 !uid_eq(cred->uid, tcred->suid) && !uid_eq(cred->uid, tcred->uid) &&
1352 !capable(CAP_SYS_NICE)) {
1353 rcu_read_unlock();
1354 err = -EPERM;
1355 goto out_put;
1356 }
1357 rcu_read_unlock();
1358
1359 task_nodes = cpuset_mems_allowed(task);
1360 /* Is the user allowed to access the target nodes? */
1361 if (!nodes_subset(*new, task_nodes) && !capable(CAP_SYS_NICE)) {
1362 err = -EPERM;
1363 goto out_put;
1364 }
1365
1366 if (!nodes_subset(*new, node_states[N_HIGH_MEMORY])) {
1367 err = -EINVAL;
1368 goto out_put;
1369 }
1370
1371 err = security_task_movememory(task);
1372 if (err)
1373 goto out_put;
1374
1375 mm = get_task_mm(task);
1376 put_task_struct(task);
1377
1378 if (!mm) {
1379 err = -EINVAL;
1380 goto out;
1381 }
1382
1383 err = do_migrate_pages(mm, old, new,
1384 capable(CAP_SYS_NICE) ? MPOL_MF_MOVE_ALL : MPOL_MF_MOVE);
1385
1386 mmput(mm);
1387out:
1388 NODEMASK_SCRATCH_FREE(scratch);
1389
1390 return err;
1391
1392out_put:
1393 put_task_struct(task);
1394 goto out;
1395
1396}
1397
1398
1399/* Retrieve NUMA policy */
1400SYSCALL_DEFINE5(get_mempolicy, int __user *, policy,
1401 unsigned long __user *, nmask, unsigned long, maxnode,
1402 unsigned long, addr, unsigned long, flags)
1403{
1404 int err;
1405 int uninitialized_var(pval);
1406 nodemask_t nodes;
1407
1408 if (nmask != NULL && maxnode < MAX_NUMNODES)
1409 return -EINVAL;
1410
1411 err = do_get_mempolicy(&pval, &nodes, addr, flags);
1412
1413 if (err)
1414 return err;
1415
1416 if (policy && put_user(pval, policy))
1417 return -EFAULT;
1418
1419 if (nmask)
1420 err = copy_nodes_to_user(nmask, maxnode, &nodes);
1421
1422 return err;
1423}
1424
1425#ifdef CONFIG_COMPAT
1426
1427asmlinkage long compat_sys_get_mempolicy(int __user *policy,
1428 compat_ulong_t __user *nmask,
1429 compat_ulong_t maxnode,
1430 compat_ulong_t addr, compat_ulong_t flags)
1431{
1432 long err;
1433 unsigned long __user *nm = NULL;
1434 unsigned long nr_bits, alloc_size;
1435 DECLARE_BITMAP(bm, MAX_NUMNODES);
1436
1437 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1438 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1439
1440 if (nmask)
1441 nm = compat_alloc_user_space(alloc_size);
1442
1443 err = sys_get_mempolicy(policy, nm, nr_bits+1, addr, flags);
1444
1445 if (!err && nmask) {
1446 unsigned long copy_size;
1447 copy_size = min_t(unsigned long, sizeof(bm), alloc_size);
1448 err = copy_from_user(bm, nm, copy_size);
1449 /* ensure entire bitmap is zeroed */
1450 err |= clear_user(nmask, ALIGN(maxnode-1, 8) / 8);
1451 err |= compat_put_bitmap(nmask, bm, nr_bits);
1452 }
1453
1454 return err;
1455}
1456
1457asmlinkage long compat_sys_set_mempolicy(int mode, compat_ulong_t __user *nmask,
1458 compat_ulong_t maxnode)
1459{
1460 long err = 0;
1461 unsigned long __user *nm = NULL;
1462 unsigned long nr_bits, alloc_size;
1463 DECLARE_BITMAP(bm, MAX_NUMNODES);
1464
1465 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1466 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1467
1468 if (nmask) {
1469 err = compat_get_bitmap(bm, nmask, nr_bits);
1470 nm = compat_alloc_user_space(alloc_size);
1471 err |= copy_to_user(nm, bm, alloc_size);
1472 }
1473
1474 if (err)
1475 return -EFAULT;
1476
1477 return sys_set_mempolicy(mode, nm, nr_bits+1);
1478}
1479
1480asmlinkage long compat_sys_mbind(compat_ulong_t start, compat_ulong_t len,
1481 compat_ulong_t mode, compat_ulong_t __user *nmask,
1482 compat_ulong_t maxnode, compat_ulong_t flags)
1483{
1484 long err = 0;
1485 unsigned long __user *nm = NULL;
1486 unsigned long nr_bits, alloc_size;
1487 nodemask_t bm;
1488
1489 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1490 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1491
1492 if (nmask) {
1493 err = compat_get_bitmap(nodes_addr(bm), nmask, nr_bits);
1494 nm = compat_alloc_user_space(alloc_size);
1495 err |= copy_to_user(nm, nodes_addr(bm), alloc_size);
1496 }
1497
1498 if (err)
1499 return -EFAULT;
1500
1501 return sys_mbind(start, len, mode, nm, nr_bits+1, flags);
1502}
1503
1504#endif
1505
1506/*
1507 * get_vma_policy(@task, @vma, @addr)
1508 * @task - task for fallback if vma policy == default
1509 * @vma - virtual memory area whose policy is sought
1510 * @addr - address in @vma for shared policy lookup
1511 *
1512 * Returns effective policy for a VMA at specified address.
1513 * Falls back to @task or system default policy, as necessary.
1514 * Current or other task's task mempolicy and non-shared vma policies
1515 * are protected by the task's mmap_sem, which must be held for read by
1516 * the caller.
1517 * Shared policies [those marked as MPOL_F_SHARED] require an extra reference
1518 * count--added by the get_policy() vm_op, as appropriate--to protect against
1519 * freeing by another task. It is the caller's responsibility to free the
1520 * extra reference for shared policies.
1521 */
1522struct mempolicy *get_vma_policy(struct task_struct *task,
1523 struct vm_area_struct *vma, unsigned long addr)
1524{
1525 struct mempolicy *pol = task->mempolicy;
1526
1527 if (vma) {
1528 if (vma->vm_ops && vma->vm_ops->get_policy) {
1529 struct mempolicy *vpol = vma->vm_ops->get_policy(vma,
1530 addr);
1531 if (vpol)
1532 pol = vpol;
1533 } else if (vma->vm_policy)
1534 pol = vma->vm_policy;
1535 }
1536 if (!pol)
1537 pol = &default_policy;
1538 return pol;
1539}
1540
1541/*
1542 * Return a nodemask representing a mempolicy for filtering nodes for
1543 * page allocation
1544 */
1545static nodemask_t *policy_nodemask(gfp_t gfp, struct mempolicy *policy)
1546{
1547 /* Lower zones don't get a nodemask applied for MPOL_BIND */
1548 if (unlikely(policy->mode == MPOL_BIND) &&
1549 gfp_zone(gfp) >= policy_zone &&
1550 cpuset_nodemask_valid_mems_allowed(&policy->v.nodes))
1551 return &policy->v.nodes;
1552
1553 return NULL;
1554}
1555
1556/* Return a zonelist indicated by gfp for node representing a mempolicy */
1557static struct zonelist *policy_zonelist(gfp_t gfp, struct mempolicy *policy,
1558 int nd)
1559{
1560 switch (policy->mode) {
1561 case MPOL_PREFERRED:
1562 if (!(policy->flags & MPOL_F_LOCAL))
1563 nd = policy->v.preferred_node;
1564 break;
1565 case MPOL_BIND:
1566 /*
1567 * Normally, MPOL_BIND allocations are node-local within the
1568 * allowed nodemask. However, if __GFP_THISNODE is set and the
1569 * current node isn't part of the mask, we use the zonelist for
1570 * the first node in the mask instead.
1571 */
1572 if (unlikely(gfp & __GFP_THISNODE) &&
1573 unlikely(!node_isset(nd, policy->v.nodes)))
1574 nd = first_node(policy->v.nodes);
1575 break;
1576 default:
1577 BUG();
1578 }
1579 return node_zonelist(nd, gfp);
1580}
1581
1582/* Do dynamic interleaving for a process */
1583static unsigned interleave_nodes(struct mempolicy *policy)
1584{
1585 unsigned nid, next;
1586 struct task_struct *me = current;
1587
1588 nid = me->il_next;
1589 next = next_node(nid, policy->v.nodes);
1590 if (next >= MAX_NUMNODES)
1591 next = first_node(policy->v.nodes);
1592 if (next < MAX_NUMNODES)
1593 me->il_next = next;
1594 return nid;
1595}
1596
1597/*
1598 * Depending on the memory policy provide a node from which to allocate the
1599 * next slab entry.
1600 * @policy must be protected by freeing by the caller. If @policy is
1601 * the current task's mempolicy, this protection is implicit, as only the
1602 * task can change it's policy. The system default policy requires no
1603 * such protection.
1604 */
1605unsigned slab_node(struct mempolicy *policy)
1606{
1607 if (!policy || policy->flags & MPOL_F_LOCAL)
1608 return numa_node_id();
1609
1610 switch (policy->mode) {
1611 case MPOL_PREFERRED:
1612 /*
1613 * handled MPOL_F_LOCAL above
1614 */
1615 return policy->v.preferred_node;
1616
1617 case MPOL_INTERLEAVE:
1618 return interleave_nodes(policy);
1619
1620 case MPOL_BIND: {
1621 /*
1622 * Follow bind policy behavior and start allocation at the
1623 * first node.
1624 */
1625 struct zonelist *zonelist;
1626 struct zone *zone;
1627 enum zone_type highest_zoneidx = gfp_zone(GFP_KERNEL);
1628 zonelist = &NODE_DATA(numa_node_id())->node_zonelists[0];
1629 (void)first_zones_zonelist(zonelist, highest_zoneidx,
1630 &policy->v.nodes,
1631 &zone);
1632 return zone ? zone->node : numa_node_id();
1633 }
1634
1635 default:
1636 BUG();
1637 }
1638}
1639
1640/* Do static interleaving for a VMA with known offset. */
1641static unsigned offset_il_node(struct mempolicy *pol,
1642 struct vm_area_struct *vma, unsigned long off)
1643{
1644 unsigned nnodes = nodes_weight(pol->v.nodes);
1645 unsigned target;
1646 int c;
1647 int nid = -1;
1648
1649 if (!nnodes)
1650 return numa_node_id();
1651 target = (unsigned int)off % nnodes;
1652 c = 0;
1653 do {
1654 nid = next_node(nid, pol->v.nodes);
1655 c++;
1656 } while (c <= target);
1657 return nid;
1658}
1659
1660/* Determine a node number for interleave */
1661static inline unsigned interleave_nid(struct mempolicy *pol,
1662 struct vm_area_struct *vma, unsigned long addr, int shift)
1663{
1664 if (vma) {
1665 unsigned long off;
1666
1667 /*
1668 * for small pages, there is no difference between
1669 * shift and PAGE_SHIFT, so the bit-shift is safe.
1670 * for huge pages, since vm_pgoff is in units of small
1671 * pages, we need to shift off the always 0 bits to get
1672 * a useful offset.
1673 */
1674 BUG_ON(shift < PAGE_SHIFT);
1675 off = vma->vm_pgoff >> (shift - PAGE_SHIFT);
1676 off += (addr - vma->vm_start) >> shift;
1677 return offset_il_node(pol, vma, off);
1678 } else
1679 return interleave_nodes(pol);
1680}
1681
1682/*
1683 * Return the bit number of a random bit set in the nodemask.
1684 * (returns -1 if nodemask is empty)
1685 */
1686int node_random(const nodemask_t *maskp)
1687{
1688 int w, bit = -1;
1689
1690 w = nodes_weight(*maskp);
1691 if (w)
1692 bit = bitmap_ord_to_pos(maskp->bits,
1693 get_random_int() % w, MAX_NUMNODES);
1694 return bit;
1695}
1696
1697#ifdef CONFIG_HUGETLBFS
1698/*
1699 * huge_zonelist(@vma, @addr, @gfp_flags, @mpol)
1700 * @vma = virtual memory area whose policy is sought
1701 * @addr = address in @vma for shared policy lookup and interleave policy
1702 * @gfp_flags = for requested zone
1703 * @mpol = pointer to mempolicy pointer for reference counted mempolicy
1704 * @nodemask = pointer to nodemask pointer for MPOL_BIND nodemask
1705 *
1706 * Returns a zonelist suitable for a huge page allocation and a pointer
1707 * to the struct mempolicy for conditional unref after allocation.
1708 * If the effective policy is 'BIND, returns a pointer to the mempolicy's
1709 * @nodemask for filtering the zonelist.
1710 *
1711 * Must be protected by get_mems_allowed()
1712 */
1713struct zonelist *huge_zonelist(struct vm_area_struct *vma, unsigned long addr,
1714 gfp_t gfp_flags, struct mempolicy **mpol,
1715 nodemask_t **nodemask)
1716{
1717 struct zonelist *zl;
1718
1719 *mpol = get_vma_policy(current, vma, addr);
1720 *nodemask = NULL; /* assume !MPOL_BIND */
1721
1722 if (unlikely((*mpol)->mode == MPOL_INTERLEAVE)) {
1723 zl = node_zonelist(interleave_nid(*mpol, vma, addr,
1724 huge_page_shift(hstate_vma(vma))), gfp_flags);
1725 } else {
1726 zl = policy_zonelist(gfp_flags, *mpol, numa_node_id());
1727 if ((*mpol)->mode == MPOL_BIND)
1728 *nodemask = &(*mpol)->v.nodes;
1729 }
1730 return zl;
1731}
1732
1733/*
1734 * init_nodemask_of_mempolicy
1735 *
1736 * If the current task's mempolicy is "default" [NULL], return 'false'
1737 * to indicate default policy. Otherwise, extract the policy nodemask
1738 * for 'bind' or 'interleave' policy into the argument nodemask, or
1739 * initialize the argument nodemask to contain the single node for
1740 * 'preferred' or 'local' policy and return 'true' to indicate presence
1741 * of non-default mempolicy.
1742 *
1743 * We don't bother with reference counting the mempolicy [mpol_get/put]
1744 * because the current task is examining it's own mempolicy and a task's
1745 * mempolicy is only ever changed by the task itself.
1746 *
1747 * N.B., it is the caller's responsibility to free a returned nodemask.
1748 */
1749bool init_nodemask_of_mempolicy(nodemask_t *mask)
1750{
1751 struct mempolicy *mempolicy;
1752 int nid;
1753
1754 if (!(mask && current->mempolicy))
1755 return false;
1756
1757 task_lock(current);
1758 mempolicy = current->mempolicy;
1759 switch (mempolicy->mode) {
1760 case MPOL_PREFERRED:
1761 if (mempolicy->flags & MPOL_F_LOCAL)
1762 nid = numa_node_id();
1763 else
1764 nid = mempolicy->v.preferred_node;
1765 init_nodemask_of_node(mask, nid);
1766 break;
1767
1768 case MPOL_BIND:
1769 /* Fall through */
1770 case MPOL_INTERLEAVE:
1771 *mask = mempolicy->v.nodes;
1772 break;
1773
1774 default:
1775 BUG();
1776 }
1777 task_unlock(current);
1778
1779 return true;
1780}
1781#endif
1782
1783/*
1784 * mempolicy_nodemask_intersects
1785 *
1786 * If tsk's mempolicy is "default" [NULL], return 'true' to indicate default
1787 * policy. Otherwise, check for intersection between mask and the policy
1788 * nodemask for 'bind' or 'interleave' policy. For 'perferred' or 'local'
1789 * policy, always return true since it may allocate elsewhere on fallback.
1790 *
1791 * Takes task_lock(tsk) to prevent freeing of its mempolicy.
1792 */
1793bool mempolicy_nodemask_intersects(struct task_struct *tsk,
1794 const nodemask_t *mask)
1795{
1796 struct mempolicy *mempolicy;
1797 bool ret = true;
1798
1799 if (!mask)
1800 return ret;
1801 task_lock(tsk);
1802 mempolicy = tsk->mempolicy;
1803 if (!mempolicy)
1804 goto out;
1805
1806 switch (mempolicy->mode) {
1807 case MPOL_PREFERRED:
1808 /*
1809 * MPOL_PREFERRED and MPOL_F_LOCAL are only preferred nodes to
1810 * allocate from, they may fallback to other nodes when oom.
1811 * Thus, it's possible for tsk to have allocated memory from
1812 * nodes in mask.
1813 */
1814 break;
1815 case MPOL_BIND:
1816 case MPOL_INTERLEAVE:
1817 ret = nodes_intersects(mempolicy->v.nodes, *mask);
1818 break;
1819 default:
1820 BUG();
1821 }
1822out:
1823 task_unlock(tsk);
1824 return ret;
1825}
1826
1827/* Allocate a page in interleaved policy.
1828 Own path because it needs to do special accounting. */
1829static struct page *alloc_page_interleave(gfp_t gfp, unsigned order,
1830 unsigned nid)
1831{
1832 struct zonelist *zl;
1833 struct page *page;
1834
1835 zl = node_zonelist(nid, gfp);
1836 page = __alloc_pages(gfp, order, zl);
1837 if (page && page_zone(page) == zonelist_zone(&zl->_zonerefs[0]))
1838 inc_zone_page_state(page, NUMA_INTERLEAVE_HIT);
1839 return page;
1840}
1841
1842/**
1843 * alloc_pages_vma - Allocate a page for a VMA.
1844 *
1845 * @gfp:
1846 * %GFP_USER user allocation.
1847 * %GFP_KERNEL kernel allocations,
1848 * %GFP_HIGHMEM highmem/user allocations,
1849 * %GFP_FS allocation should not call back into a file system.
1850 * %GFP_ATOMIC don't sleep.
1851 *
1852 * @order:Order of the GFP allocation.
1853 * @vma: Pointer to VMA or NULL if not available.
1854 * @addr: Virtual Address of the allocation. Must be inside the VMA.
1855 *
1856 * This function allocates a page from the kernel page pool and applies
1857 * a NUMA policy associated with the VMA or the current process.
1858 * When VMA is not NULL caller must hold down_read on the mmap_sem of the
1859 * mm_struct of the VMA to prevent it from going away. Should be used for
1860 * all allocations for pages that will be mapped into
1861 * user space. Returns NULL when no page can be allocated.
1862 *
1863 * Should be called with the mm_sem of the vma hold.
1864 */
1865struct page *
1866alloc_pages_vma(gfp_t gfp, int order, struct vm_area_struct *vma,
1867 unsigned long addr, int node)
1868{
1869 struct mempolicy *pol;
1870 struct zonelist *zl;
1871 struct page *page;
1872 unsigned int cpuset_mems_cookie;
1873
1874retry_cpuset:
1875 pol = get_vma_policy(current, vma, addr);
1876 cpuset_mems_cookie = get_mems_allowed();
1877
1878 if (unlikely(pol->mode == MPOL_INTERLEAVE)) {
1879 unsigned nid;
1880
1881 nid = interleave_nid(pol, vma, addr, PAGE_SHIFT + order);
1882 mpol_cond_put(pol);
1883 page = alloc_page_interleave(gfp, order, nid);
1884 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page))
1885 goto retry_cpuset;
1886
1887 return page;
1888 }
1889 zl = policy_zonelist(gfp, pol, node);
1890 if (unlikely(mpol_needs_cond_ref(pol))) {
1891 /*
1892 * slow path: ref counted shared policy
1893 */
1894 struct page *page = __alloc_pages_nodemask(gfp, order,
1895 zl, policy_nodemask(gfp, pol));
1896 __mpol_put(pol);
1897 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page))
1898 goto retry_cpuset;
1899 return page;
1900 }
1901 /*
1902 * fast path: default or task policy
1903 */
1904 page = __alloc_pages_nodemask(gfp, order, zl,
1905 policy_nodemask(gfp, pol));
1906 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page))
1907 goto retry_cpuset;
1908 return page;
1909}
1910
1911/**
1912 * alloc_pages_current - Allocate pages.
1913 *
1914 * @gfp:
1915 * %GFP_USER user allocation,
1916 * %GFP_KERNEL kernel allocation,
1917 * %GFP_HIGHMEM highmem allocation,
1918 * %GFP_FS don't call back into a file system.
1919 * %GFP_ATOMIC don't sleep.
1920 * @order: Power of two of allocation size in pages. 0 is a single page.
1921 *
1922 * Allocate a page from the kernel page pool. When not in
1923 * interrupt context and apply the current process NUMA policy.
1924 * Returns NULL when no page can be allocated.
1925 *
1926 * Don't call cpuset_update_task_memory_state() unless
1927 * 1) it's ok to take cpuset_sem (can WAIT), and
1928 * 2) allocating for current task (not interrupt).
1929 */
1930struct page *alloc_pages_current(gfp_t gfp, unsigned order)
1931{
1932 struct mempolicy *pol = current->mempolicy;
1933 struct page *page;
1934 unsigned int cpuset_mems_cookie;
1935
1936 if (!pol || in_interrupt() || (gfp & __GFP_THISNODE))
1937 pol = &default_policy;
1938
1939retry_cpuset:
1940 cpuset_mems_cookie = get_mems_allowed();
1941
1942 /*
1943 * No reference counting needed for current->mempolicy
1944 * nor system default_policy
1945 */
1946 if (pol->mode == MPOL_INTERLEAVE)
1947 page = alloc_page_interleave(gfp, order, interleave_nodes(pol));
1948 else
1949 page = __alloc_pages_nodemask(gfp, order,
1950 policy_zonelist(gfp, pol, numa_node_id()),
1951 policy_nodemask(gfp, pol));
1952
1953 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page))
1954 goto retry_cpuset;
1955
1956 return page;
1957}
1958EXPORT_SYMBOL(alloc_pages_current);
1959
1960/*
1961 * If mpol_dup() sees current->cpuset == cpuset_being_rebound, then it
1962 * rebinds the mempolicy its copying by calling mpol_rebind_policy()
1963 * with the mems_allowed returned by cpuset_mems_allowed(). This
1964 * keeps mempolicies cpuset relative after its cpuset moves. See
1965 * further kernel/cpuset.c update_nodemask().
1966 *
1967 * current's mempolicy may be rebinded by the other task(the task that changes
1968 * cpuset's mems), so we needn't do rebind work for current task.
1969 */
1970
1971/* Slow path of a mempolicy duplicate */
1972struct mempolicy *__mpol_dup(struct mempolicy *old)
1973{
1974 struct mempolicy *new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
1975
1976 if (!new)
1977 return ERR_PTR(-ENOMEM);
1978
1979 /* task's mempolicy is protected by alloc_lock */
1980 if (old == current->mempolicy) {
1981 task_lock(current);
1982 *new = *old;
1983 task_unlock(current);
1984 } else
1985 *new = *old;
1986
1987 rcu_read_lock();
1988 if (current_cpuset_is_being_rebound()) {
1989 nodemask_t mems = cpuset_mems_allowed(current);
1990 if (new->flags & MPOL_F_REBINDING)
1991 mpol_rebind_policy(new, &mems, MPOL_REBIND_STEP2);
1992 else
1993 mpol_rebind_policy(new, &mems, MPOL_REBIND_ONCE);
1994 }
1995 rcu_read_unlock();
1996 atomic_set(&new->refcnt, 1);
1997 return new;
1998}
1999
2000/*
2001 * If *frompol needs [has] an extra ref, copy *frompol to *tompol ,
2002 * eliminate the * MPOL_F_* flags that require conditional ref and
2003 * [NOTE!!!] drop the extra ref. Not safe to reference *frompol directly
2004 * after return. Use the returned value.
2005 *
2006 * Allows use of a mempolicy for, e.g., multiple allocations with a single
2007 * policy lookup, even if the policy needs/has extra ref on lookup.
2008 * shmem_readahead needs this.
2009 */
2010struct mempolicy *__mpol_cond_copy(struct mempolicy *tompol,
2011 struct mempolicy *frompol)
2012{
2013 if (!mpol_needs_cond_ref(frompol))
2014 return frompol;
2015
2016 *tompol = *frompol;
2017 tompol->flags &= ~MPOL_F_SHARED; /* copy doesn't need unref */
2018 __mpol_put(frompol);
2019 return tompol;
2020}
2021
2022/* Slow path of a mempolicy comparison */
2023bool __mpol_equal(struct mempolicy *a, struct mempolicy *b)
2024{
2025 if (!a || !b)
2026 return false;
2027 if (a->mode != b->mode)
2028 return false;
2029 if (a->flags != b->flags)
2030 return false;
2031 if (mpol_store_user_nodemask(a))
2032 if (!nodes_equal(a->w.user_nodemask, b->w.user_nodemask))
2033 return false;
2034
2035 switch (a->mode) {
2036 case MPOL_BIND:
2037 /* Fall through */
2038 case MPOL_INTERLEAVE:
2039 return !!nodes_equal(a->v.nodes, b->v.nodes);
2040 case MPOL_PREFERRED:
2041 return a->v.preferred_node == b->v.preferred_node;
2042 default:
2043 BUG();
2044 return false;
2045 }
2046}
2047
2048/*
2049 * Shared memory backing store policy support.
2050 *
2051 * Remember policies even when nobody has shared memory mapped.
2052 * The policies are kept in Red-Black tree linked from the inode.
2053 * They are protected by the sp->lock spinlock, which should be held
2054 * for any accesses to the tree.
2055 */
2056
2057/* lookup first element intersecting start-end */
2058/* Caller holds sp->lock */
2059static struct sp_node *
2060sp_lookup(struct shared_policy *sp, unsigned long start, unsigned long end)
2061{
2062 struct rb_node *n = sp->root.rb_node;
2063
2064 while (n) {
2065 struct sp_node *p = rb_entry(n, struct sp_node, nd);
2066
2067 if (start >= p->end)
2068 n = n->rb_right;
2069 else if (end <= p->start)
2070 n = n->rb_left;
2071 else
2072 break;
2073 }
2074 if (!n)
2075 return NULL;
2076 for (;;) {
2077 struct sp_node *w = NULL;
2078 struct rb_node *prev = rb_prev(n);
2079 if (!prev)
2080 break;
2081 w = rb_entry(prev, struct sp_node, nd);
2082 if (w->end <= start)
2083 break;
2084 n = prev;
2085 }
2086 return rb_entry(n, struct sp_node, nd);
2087}
2088
2089/* Insert a new shared policy into the list. */
2090/* Caller holds sp->lock */
2091static void sp_insert(struct shared_policy *sp, struct sp_node *new)
2092{
2093 struct rb_node **p = &sp->root.rb_node;
2094 struct rb_node *parent = NULL;
2095 struct sp_node *nd;
2096
2097 while (*p) {
2098 parent = *p;
2099 nd = rb_entry(parent, struct sp_node, nd);
2100 if (new->start < nd->start)
2101 p = &(*p)->rb_left;
2102 else if (new->end > nd->end)
2103 p = &(*p)->rb_right;
2104 else
2105 BUG();
2106 }
2107 rb_link_node(&new->nd, parent, p);
2108 rb_insert_color(&new->nd, &sp->root);
2109 pr_debug("inserting %lx-%lx: %d\n", new->start, new->end,
2110 new->policy ? new->policy->mode : 0);
2111}
2112
2113/* Find shared policy intersecting idx */
2114struct mempolicy *
2115mpol_shared_policy_lookup(struct shared_policy *sp, unsigned long idx)
2116{
2117 struct mempolicy *pol = NULL;
2118 struct sp_node *sn;
2119
2120 if (!sp->root.rb_node)
2121 return NULL;
2122 spin_lock(&sp->lock);
2123 sn = sp_lookup(sp, idx, idx+1);
2124 if (sn) {
2125 mpol_get(sn->policy);
2126 pol = sn->policy;
2127 }
2128 spin_unlock(&sp->lock);
2129 return pol;
2130}
2131
2132static void sp_delete(struct shared_policy *sp, struct sp_node *n)
2133{
2134 pr_debug("deleting %lx-l%lx\n", n->start, n->end);
2135 rb_erase(&n->nd, &sp->root);
2136 mpol_put(n->policy);
2137 kmem_cache_free(sn_cache, n);
2138}
2139
2140static struct sp_node *sp_alloc(unsigned long start, unsigned long end,
2141 struct mempolicy *pol)
2142{
2143 struct sp_node *n = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2144
2145 if (!n)
2146 return NULL;
2147 n->start = start;
2148 n->end = end;
2149 mpol_get(pol);
2150 pol->flags |= MPOL_F_SHARED; /* for unref */
2151 n->policy = pol;
2152 return n;
2153}
2154
2155/* Replace a policy range. */
2156static int shared_policy_replace(struct shared_policy *sp, unsigned long start,
2157 unsigned long end, struct sp_node *new)
2158{
2159 struct sp_node *n, *new2 = NULL;
2160
2161restart:
2162 spin_lock(&sp->lock);
2163 n = sp_lookup(sp, start, end);
2164 /* Take care of old policies in the same range. */
2165 while (n && n->start < end) {
2166 struct rb_node *next = rb_next(&n->nd);
2167 if (n->start >= start) {
2168 if (n->end <= end)
2169 sp_delete(sp, n);
2170 else
2171 n->start = end;
2172 } else {
2173 /* Old policy spanning whole new range. */
2174 if (n->end > end) {
2175 if (!new2) {
2176 spin_unlock(&sp->lock);
2177 new2 = sp_alloc(end, n->end, n->policy);
2178 if (!new2)
2179 return -ENOMEM;
2180 goto restart;
2181 }
2182 n->end = start;
2183 sp_insert(sp, new2);
2184 new2 = NULL;
2185 break;
2186 } else
2187 n->end = start;
2188 }
2189 if (!next)
2190 break;
2191 n = rb_entry(next, struct sp_node, nd);
2192 }
2193 if (new)
2194 sp_insert(sp, new);
2195 spin_unlock(&sp->lock);
2196 if (new2) {
2197 mpol_put(new2->policy);
2198 kmem_cache_free(sn_cache, new2);
2199 }
2200 return 0;
2201}
2202
2203/**
2204 * mpol_shared_policy_init - initialize shared policy for inode
2205 * @sp: pointer to inode shared policy
2206 * @mpol: struct mempolicy to install
2207 *
2208 * Install non-NULL @mpol in inode's shared policy rb-tree.
2209 * On entry, the current task has a reference on a non-NULL @mpol.
2210 * This must be released on exit.
2211 * This is called at get_inode() calls and we can use GFP_KERNEL.
2212 */
2213void mpol_shared_policy_init(struct shared_policy *sp, struct mempolicy *mpol)
2214{
2215 int ret;
2216
2217 sp->root = RB_ROOT; /* empty tree == default mempolicy */
2218 spin_lock_init(&sp->lock);
2219
2220 if (mpol) {
2221 struct vm_area_struct pvma;
2222 struct mempolicy *new;
2223 NODEMASK_SCRATCH(scratch);
2224
2225 if (!scratch)
2226 goto put_mpol;
2227 /* contextualize the tmpfs mount point mempolicy */
2228 new = mpol_new(mpol->mode, mpol->flags, &mpol->w.user_nodemask);
2229 if (IS_ERR(new))
2230 goto free_scratch; /* no valid nodemask intersection */
2231
2232 task_lock(current);
2233 ret = mpol_set_nodemask(new, &mpol->w.user_nodemask, scratch);
2234 task_unlock(current);
2235 if (ret)
2236 goto put_new;
2237
2238 /* Create pseudo-vma that contains just the policy */
2239 memset(&pvma, 0, sizeof(struct vm_area_struct));
2240 pvma.vm_end = TASK_SIZE; /* policy covers entire file */
2241 mpol_set_shared_policy(sp, &pvma, new); /* adds ref */
2242
2243put_new:
2244 mpol_put(new); /* drop initial ref */
2245free_scratch:
2246 NODEMASK_SCRATCH_FREE(scratch);
2247put_mpol:
2248 mpol_put(mpol); /* drop our incoming ref on sb mpol */
2249 }
2250}
2251
2252int mpol_set_shared_policy(struct shared_policy *info,
2253 struct vm_area_struct *vma, struct mempolicy *npol)
2254{
2255 int err;
2256 struct sp_node *new = NULL;
2257 unsigned long sz = vma_pages(vma);
2258
2259 pr_debug("set_shared_policy %lx sz %lu %d %d %lx\n",
2260 vma->vm_pgoff,
2261 sz, npol ? npol->mode : -1,
2262 npol ? npol->flags : -1,
2263 npol ? nodes_addr(npol->v.nodes)[0] : -1);
2264
2265 if (npol) {
2266 new = sp_alloc(vma->vm_pgoff, vma->vm_pgoff + sz, npol);
2267 if (!new)
2268 return -ENOMEM;
2269 }
2270 err = shared_policy_replace(info, vma->vm_pgoff, vma->vm_pgoff+sz, new);
2271 if (err && new)
2272 kmem_cache_free(sn_cache, new);
2273 return err;
2274}
2275
2276/* Free a backing policy store on inode delete. */
2277void mpol_free_shared_policy(struct shared_policy *p)
2278{
2279 struct sp_node *n;
2280 struct rb_node *next;
2281
2282 if (!p->root.rb_node)
2283 return;
2284 spin_lock(&p->lock);
2285 next = rb_first(&p->root);
2286 while (next) {
2287 n = rb_entry(next, struct sp_node, nd);
2288 next = rb_next(&n->nd);
2289 rb_erase(&n->nd, &p->root);
2290 mpol_put(n->policy);
2291 kmem_cache_free(sn_cache, n);
2292 }
2293 spin_unlock(&p->lock);
2294}
2295
2296/* assumes fs == KERNEL_DS */
2297void __init numa_policy_init(void)
2298{
2299 nodemask_t interleave_nodes;
2300 unsigned long largest = 0;
2301 int nid, prefer = 0;
2302
2303 policy_cache = kmem_cache_create("numa_policy",
2304 sizeof(struct mempolicy),
2305 0, SLAB_PANIC, NULL);
2306
2307 sn_cache = kmem_cache_create("shared_policy_node",
2308 sizeof(struct sp_node),
2309 0, SLAB_PANIC, NULL);
2310
2311 /*
2312 * Set interleaving policy for system init. Interleaving is only
2313 * enabled across suitably sized nodes (default is >= 16MB), or
2314 * fall back to the largest node if they're all smaller.
2315 */
2316 nodes_clear(interleave_nodes);
2317 for_each_node_state(nid, N_HIGH_MEMORY) {
2318 unsigned long total_pages = node_present_pages(nid);
2319
2320 /* Preserve the largest node */
2321 if (largest < total_pages) {
2322 largest = total_pages;
2323 prefer = nid;
2324 }
2325
2326 /* Interleave this node? */
2327 if ((total_pages << PAGE_SHIFT) >= (16 << 20))
2328 node_set(nid, interleave_nodes);
2329 }
2330
2331 /* All too small, use the largest */
2332 if (unlikely(nodes_empty(interleave_nodes)))
2333 node_set(prefer, interleave_nodes);
2334
2335 if (do_set_mempolicy(MPOL_INTERLEAVE, 0, &interleave_nodes))
2336 printk("numa_policy_init: interleaving failed\n");
2337}
2338
2339/* Reset policy of current process to default */
2340void numa_default_policy(void)
2341{
2342 do_set_mempolicy(MPOL_DEFAULT, 0, NULL);
2343}
2344
2345/*
2346 * Parse and format mempolicy from/to strings
2347 */
2348
2349/*
2350 * "local" is pseudo-policy: MPOL_PREFERRED with MPOL_F_LOCAL flag
2351 * Used only for mpol_parse_str() and mpol_to_str()
2352 */
2353#define MPOL_LOCAL MPOL_MAX
2354static const char * const policy_modes[] =
2355{
2356 [MPOL_DEFAULT] = "default",
2357 [MPOL_PREFERRED] = "prefer",
2358 [MPOL_BIND] = "bind",
2359 [MPOL_INTERLEAVE] = "interleave",
2360 [MPOL_LOCAL] = "local"
2361};
2362
2363
2364#ifdef CONFIG_TMPFS
2365/**
2366 * mpol_parse_str - parse string to mempolicy
2367 * @str: string containing mempolicy to parse
2368 * @mpol: pointer to struct mempolicy pointer, returned on success.
2369 * @no_context: flag whether to "contextualize" the mempolicy
2370 *
2371 * Format of input:
2372 * <mode>[=<flags>][:<nodelist>]
2373 *
2374 * if @no_context is true, save the input nodemask in w.user_nodemask in
2375 * the returned mempolicy. This will be used to "clone" the mempolicy in
2376 * a specific context [cpuset] at a later time. Used to parse tmpfs mpol
2377 * mount option. Note that if 'static' or 'relative' mode flags were
2378 * specified, the input nodemask will already have been saved. Saving
2379 * it again is redundant, but safe.
2380 *
2381 * On success, returns 0, else 1
2382 */
2383int mpol_parse_str(char *str, struct mempolicy **mpol, int no_context)
2384{
2385 struct mempolicy *new = NULL;
2386 unsigned short mode;
2387 unsigned short uninitialized_var(mode_flags);
2388 nodemask_t nodes;
2389 char *nodelist = strchr(str, ':');
2390 char *flags = strchr(str, '=');
2391 int err = 1;
2392
2393 if (nodelist) {
2394 /* NUL-terminate mode or flags string */
2395 *nodelist++ = '\0';
2396 if (nodelist_parse(nodelist, nodes))
2397 goto out;
2398 if (!nodes_subset(nodes, node_states[N_HIGH_MEMORY]))
2399 goto out;
2400 } else
2401 nodes_clear(nodes);
2402
2403 if (flags)
2404 *flags++ = '\0'; /* terminate mode string */
2405
2406 for (mode = 0; mode <= MPOL_LOCAL; mode++) {
2407 if (!strcmp(str, policy_modes[mode])) {
2408 break;
2409 }
2410 }
2411 if (mode > MPOL_LOCAL)
2412 goto out;
2413
2414 switch (mode) {
2415 case MPOL_PREFERRED:
2416 /*
2417 * Insist on a nodelist of one node only
2418 */
2419 if (nodelist) {
2420 char *rest = nodelist;
2421 while (isdigit(*rest))
2422 rest++;
2423 if (*rest)
2424 goto out;
2425 }
2426 break;
2427 case MPOL_INTERLEAVE:
2428 /*
2429 * Default to online nodes with memory if no nodelist
2430 */
2431 if (!nodelist)
2432 nodes = node_states[N_HIGH_MEMORY];
2433 break;
2434 case MPOL_LOCAL:
2435 /*
2436 * Don't allow a nodelist; mpol_new() checks flags
2437 */
2438 if (nodelist)
2439 goto out;
2440 mode = MPOL_PREFERRED;
2441 break;
2442 case MPOL_DEFAULT:
2443 /*
2444 * Insist on a empty nodelist
2445 */
2446 if (!nodelist)
2447 err = 0;
2448 goto out;
2449 case MPOL_BIND:
2450 /*
2451 * Insist on a nodelist
2452 */
2453 if (!nodelist)
2454 goto out;
2455 }
2456
2457 mode_flags = 0;
2458 if (flags) {
2459 /*
2460 * Currently, we only support two mutually exclusive
2461 * mode flags.
2462 */
2463 if (!strcmp(flags, "static"))
2464 mode_flags |= MPOL_F_STATIC_NODES;
2465 else if (!strcmp(flags, "relative"))
2466 mode_flags |= MPOL_F_RELATIVE_NODES;
2467 else
2468 goto out;
2469 }
2470
2471 new = mpol_new(mode, mode_flags, &nodes);
2472 if (IS_ERR(new))
2473 goto out;
2474
2475 if (no_context) {
2476 /* save for contextualization */
2477 new->w.user_nodemask = nodes;
2478 } else {
2479 int ret;
2480 NODEMASK_SCRATCH(scratch);
2481 if (scratch) {
2482 task_lock(current);
2483 ret = mpol_set_nodemask(new, &nodes, scratch);
2484 task_unlock(current);
2485 } else
2486 ret = -ENOMEM;
2487 NODEMASK_SCRATCH_FREE(scratch);
2488 if (ret) {
2489 mpol_put(new);
2490 goto out;
2491 }
2492 }
2493 err = 0;
2494
2495out:
2496 /* Restore string for error message */
2497 if (nodelist)
2498 *--nodelist = ':';
2499 if (flags)
2500 *--flags = '=';
2501 if (!err)
2502 *mpol = new;
2503 return err;
2504}
2505#endif /* CONFIG_TMPFS */
2506
2507/**
2508 * mpol_to_str - format a mempolicy structure for printing
2509 * @buffer: to contain formatted mempolicy string
2510 * @maxlen: length of @buffer
2511 * @pol: pointer to mempolicy to be formatted
2512 * @no_context: "context free" mempolicy - use nodemask in w.user_nodemask
2513 *
2514 * Convert a mempolicy into a string.
2515 * Returns the number of characters in buffer (if positive)
2516 * or an error (negative)
2517 */
2518int mpol_to_str(char *buffer, int maxlen, struct mempolicy *pol, int no_context)
2519{
2520 char *p = buffer;
2521 int l;
2522 nodemask_t nodes;
2523 unsigned short mode;
2524 unsigned short flags = pol ? pol->flags : 0;
2525
2526 /*
2527 * Sanity check: room for longest mode, flag and some nodes
2528 */
2529 VM_BUG_ON(maxlen < strlen("interleave") + strlen("relative") + 16);
2530
2531 if (!pol || pol == &default_policy)
2532 mode = MPOL_DEFAULT;
2533 else
2534 mode = pol->mode;
2535
2536 switch (mode) {
2537 case MPOL_DEFAULT:
2538 nodes_clear(nodes);
2539 break;
2540
2541 case MPOL_PREFERRED:
2542 nodes_clear(nodes);
2543 if (flags & MPOL_F_LOCAL)
2544 mode = MPOL_LOCAL; /* pseudo-policy */
2545 else
2546 node_set(pol->v.preferred_node, nodes);
2547 break;
2548
2549 case MPOL_BIND:
2550 /* Fall through */
2551 case MPOL_INTERLEAVE:
2552 if (no_context)
2553 nodes = pol->w.user_nodemask;
2554 else
2555 nodes = pol->v.nodes;
2556 break;
2557
2558 default:
2559 return -EINVAL;
2560 }
2561
2562 l = strlen(policy_modes[mode]);
2563 if (buffer + maxlen < p + l + 1)
2564 return -ENOSPC;
2565
2566 strcpy(p, policy_modes[mode]);
2567 p += l;
2568
2569 if (flags & MPOL_MODE_FLAGS) {
2570 if (buffer + maxlen < p + 2)
2571 return -ENOSPC;
2572 *p++ = '=';
2573
2574 /*
2575 * Currently, the only defined flags are mutually exclusive
2576 */
2577 if (flags & MPOL_F_STATIC_NODES)
2578 p += snprintf(p, buffer + maxlen - p, "static");
2579 else if (flags & MPOL_F_RELATIVE_NODES)
2580 p += snprintf(p, buffer + maxlen - p, "relative");
2581 }
2582
2583 if (!nodes_empty(nodes)) {
2584 if (buffer + maxlen < p + 2)
2585 return -ENOSPC;
2586 *p++ = ':';
2587 p += nodelist_scnprintf(p, buffer + maxlen - p, nodes);
2588 }
2589 return p - buffer;
2590}