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