mempolicy: support optional mode flags
[linux-2.6/kvm.git] / mm / mempolicy.c
blob1f6ff9c1bbc3ec99d5ad6b8db04859eb234ee96d
1 /*
2 * Simple NUMA memory policy for the Linux kernel.
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.
8 * NUMA policy allows the user to give hints in which node(s) memory should
9 * be allocated.
11 * Support four policies per VMA and per process:
13 * The VMA policy has priority over the process policy for a page fault.
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.
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
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.
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.
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.
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.
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.
52 * For shmfs/tmpfs/hugetlbfs shared memory the policy is shared between
53 * all users and remembered even when nobody has memory mapped.
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 could replace all the switch()es with a mempolicy_ops structure.
69 #include <linux/mempolicy.h>
70 #include <linux/mm.h>
71 #include <linux/highmem.h>
72 #include <linux/hugetlb.h>
73 #include <linux/kernel.h>
74 #include <linux/sched.h>
75 #include <linux/nodemask.h>
76 #include <linux/cpuset.h>
77 #include <linux/gfp.h>
78 #include <linux/slab.h>
79 #include <linux/string.h>
80 #include <linux/module.h>
81 #include <linux/nsproxy.h>
82 #include <linux/interrupt.h>
83 #include <linux/init.h>
84 #include <linux/compat.h>
85 #include <linux/swap.h>
86 #include <linux/seq_file.h>
87 #include <linux/proc_fs.h>
88 #include <linux/migrate.h>
89 #include <linux/rmap.h>
90 #include <linux/security.h>
91 #include <linux/syscalls.h>
93 #include <asm/tlbflush.h>
94 #include <asm/uaccess.h>
96 /* Internal flags */
97 #define MPOL_MF_DISCONTIG_OK (MPOL_MF_INTERNAL << 0) /* Skip checks for continuous vmas */
98 #define MPOL_MF_INVERT (MPOL_MF_INTERNAL << 1) /* Invert check for nodemask */
99 #define MPOL_MF_STATS (MPOL_MF_INTERNAL << 2) /* Gather statistics */
101 static struct kmem_cache *policy_cache;
102 static struct kmem_cache *sn_cache;
104 /* Highest zone. An specific allocation for a zone below that is not
105 policied. */
106 enum zone_type policy_zone = 0;
108 struct mempolicy default_policy = {
109 .refcnt = ATOMIC_INIT(1), /* never free it */
110 .policy = MPOL_DEFAULT,
113 static void mpol_rebind_policy(struct mempolicy *pol,
114 const nodemask_t *newmask);
116 /* Do sanity checking on a policy */
117 static int mpol_check_policy(unsigned short mode, nodemask_t *nodes)
119 int was_empty, is_empty;
121 if (!nodes)
122 return 0;
125 * "Contextualize" the in-coming nodemast for cpusets:
126 * Remember whether in-coming nodemask was empty, If not,
127 * restrict the nodes to the allowed nodes in the cpuset.
128 * This is guaranteed to be a subset of nodes with memory.
130 cpuset_update_task_memory_state();
131 is_empty = was_empty = nodes_empty(*nodes);
132 if (!was_empty) {
133 nodes_and(*nodes, *nodes, cpuset_current_mems_allowed);
134 is_empty = nodes_empty(*nodes); /* after "contextualization" */
137 switch (mode) {
138 case MPOL_DEFAULT:
140 * require caller to specify an empty nodemask
141 * before "contextualization"
143 if (!was_empty)
144 return -EINVAL;
145 break;
146 case MPOL_BIND:
147 case MPOL_INTERLEAVE:
149 * require at least 1 valid node after "contextualization"
151 if (is_empty)
152 return -EINVAL;
153 break;
154 case MPOL_PREFERRED:
156 * Did caller specify invalid nodes?
157 * Don't silently accept this as "local allocation".
159 if (!was_empty && is_empty)
160 return -EINVAL;
161 break;
162 default:
163 BUG();
165 return 0;
168 /* Check that the nodemask contains at least one populated zone */
169 static int is_valid_nodemask(nodemask_t *nodemask)
171 int nd, k;
173 /* Check that there is something useful in this mask */
174 k = policy_zone;
176 for_each_node_mask(nd, *nodemask) {
177 struct zone *z;
179 for (k = 0; k <= policy_zone; k++) {
180 z = &NODE_DATA(nd)->node_zones[k];
181 if (z->present_pages > 0)
182 return 1;
186 return 0;
189 /* Create a new policy */
190 static struct mempolicy *mpol_new(unsigned short mode, unsigned short flags,
191 nodemask_t *nodes)
193 struct mempolicy *policy;
195 pr_debug("setting mode %d flags %d nodes[0] %lx\n",
196 mode, flags, nodes ? nodes_addr(*nodes)[0] : -1);
198 if (mode == MPOL_DEFAULT)
199 return NULL;
200 policy = kmem_cache_alloc(policy_cache, GFP_KERNEL);
201 if (!policy)
202 return ERR_PTR(-ENOMEM);
203 atomic_set(&policy->refcnt, 1);
204 switch (mode) {
205 case MPOL_INTERLEAVE:
206 policy->v.nodes = *nodes;
207 if (nodes_weight(policy->v.nodes) == 0) {
208 kmem_cache_free(policy_cache, policy);
209 return ERR_PTR(-EINVAL);
211 break;
212 case MPOL_PREFERRED:
213 policy->v.preferred_node = first_node(*nodes);
214 if (policy->v.preferred_node >= MAX_NUMNODES)
215 policy->v.preferred_node = -1;
216 break;
217 case MPOL_BIND:
218 if (!is_valid_nodemask(nodes)) {
219 kmem_cache_free(policy_cache, policy);
220 return ERR_PTR(-EINVAL);
222 policy->v.nodes = *nodes;
223 break;
224 default:
225 BUG();
227 policy->policy = mode;
228 policy->flags = flags;
229 policy->cpuset_mems_allowed = cpuset_mems_allowed(current);
230 return policy;
233 static void gather_stats(struct page *, void *, int pte_dirty);
234 static void migrate_page_add(struct page *page, struct list_head *pagelist,
235 unsigned long flags);
237 /* Scan through pages checking if pages follow certain conditions. */
238 static int check_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
239 unsigned long addr, unsigned long end,
240 const nodemask_t *nodes, unsigned long flags,
241 void *private)
243 pte_t *orig_pte;
244 pte_t *pte;
245 spinlock_t *ptl;
247 orig_pte = pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
248 do {
249 struct page *page;
250 int nid;
252 if (!pte_present(*pte))
253 continue;
254 page = vm_normal_page(vma, addr, *pte);
255 if (!page)
256 continue;
258 * The check for PageReserved here is important to avoid
259 * handling zero pages and other pages that may have been
260 * marked special by the system.
262 * If the PageReserved would not be checked here then f.e.
263 * the location of the zero page could have an influence
264 * on MPOL_MF_STRICT, zero pages would be counted for
265 * the per node stats, and there would be useless attempts
266 * to put zero pages on the migration list.
268 if (PageReserved(page))
269 continue;
270 nid = page_to_nid(page);
271 if (node_isset(nid, *nodes) == !!(flags & MPOL_MF_INVERT))
272 continue;
274 if (flags & MPOL_MF_STATS)
275 gather_stats(page, private, pte_dirty(*pte));
276 else if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))
277 migrate_page_add(page, private, flags);
278 else
279 break;
280 } while (pte++, addr += PAGE_SIZE, addr != end);
281 pte_unmap_unlock(orig_pte, ptl);
282 return addr != end;
285 static inline int check_pmd_range(struct vm_area_struct *vma, pud_t *pud,
286 unsigned long addr, unsigned long end,
287 const nodemask_t *nodes, unsigned long flags,
288 void *private)
290 pmd_t *pmd;
291 unsigned long next;
293 pmd = pmd_offset(pud, addr);
294 do {
295 next = pmd_addr_end(addr, end);
296 if (pmd_none_or_clear_bad(pmd))
297 continue;
298 if (check_pte_range(vma, pmd, addr, next, nodes,
299 flags, private))
300 return -EIO;
301 } while (pmd++, addr = next, addr != end);
302 return 0;
305 static inline int check_pud_range(struct vm_area_struct *vma, pgd_t *pgd,
306 unsigned long addr, unsigned long end,
307 const nodemask_t *nodes, unsigned long flags,
308 void *private)
310 pud_t *pud;
311 unsigned long next;
313 pud = pud_offset(pgd, addr);
314 do {
315 next = pud_addr_end(addr, end);
316 if (pud_none_or_clear_bad(pud))
317 continue;
318 if (check_pmd_range(vma, pud, addr, next, nodes,
319 flags, private))
320 return -EIO;
321 } while (pud++, addr = next, addr != end);
322 return 0;
325 static inline int check_pgd_range(struct vm_area_struct *vma,
326 unsigned long addr, unsigned long end,
327 const nodemask_t *nodes, unsigned long flags,
328 void *private)
330 pgd_t *pgd;
331 unsigned long next;
333 pgd = pgd_offset(vma->vm_mm, addr);
334 do {
335 next = pgd_addr_end(addr, end);
336 if (pgd_none_or_clear_bad(pgd))
337 continue;
338 if (check_pud_range(vma, pgd, addr, next, nodes,
339 flags, private))
340 return -EIO;
341 } while (pgd++, addr = next, addr != end);
342 return 0;
346 * Check if all pages in a range are on a set of nodes.
347 * If pagelist != NULL then isolate pages from the LRU and
348 * put them on the pagelist.
350 static struct vm_area_struct *
351 check_range(struct mm_struct *mm, unsigned long start, unsigned long end,
352 const nodemask_t *nodes, unsigned long flags, void *private)
354 int err;
355 struct vm_area_struct *first, *vma, *prev;
357 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) {
359 err = migrate_prep();
360 if (err)
361 return ERR_PTR(err);
364 first = find_vma(mm, start);
365 if (!first)
366 return ERR_PTR(-EFAULT);
367 prev = NULL;
368 for (vma = first; vma && vma->vm_start < end; vma = vma->vm_next) {
369 if (!(flags & MPOL_MF_DISCONTIG_OK)) {
370 if (!vma->vm_next && vma->vm_end < end)
371 return ERR_PTR(-EFAULT);
372 if (prev && prev->vm_end < vma->vm_start)
373 return ERR_PTR(-EFAULT);
375 if (!is_vm_hugetlb_page(vma) &&
376 ((flags & MPOL_MF_STRICT) ||
377 ((flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) &&
378 vma_migratable(vma)))) {
379 unsigned long endvma = vma->vm_end;
381 if (endvma > end)
382 endvma = end;
383 if (vma->vm_start > start)
384 start = vma->vm_start;
385 err = check_pgd_range(vma, start, endvma, nodes,
386 flags, private);
387 if (err) {
388 first = ERR_PTR(err);
389 break;
392 prev = vma;
394 return first;
397 /* Apply policy to a single VMA */
398 static int policy_vma(struct vm_area_struct *vma, struct mempolicy *new)
400 int err = 0;
401 struct mempolicy *old = vma->vm_policy;
403 pr_debug("vma %lx-%lx/%lx vm_ops %p vm_file %p set_policy %p\n",
404 vma->vm_start, vma->vm_end, vma->vm_pgoff,
405 vma->vm_ops, vma->vm_file,
406 vma->vm_ops ? vma->vm_ops->set_policy : NULL);
408 if (vma->vm_ops && vma->vm_ops->set_policy)
409 err = vma->vm_ops->set_policy(vma, new);
410 if (!err) {
411 mpol_get(new);
412 vma->vm_policy = new;
413 mpol_free(old);
415 return err;
418 /* Step 2: apply policy to a range and do splits. */
419 static int mbind_range(struct vm_area_struct *vma, unsigned long start,
420 unsigned long end, struct mempolicy *new)
422 struct vm_area_struct *next;
423 int err;
425 err = 0;
426 for (; vma && vma->vm_start < end; vma = next) {
427 next = vma->vm_next;
428 if (vma->vm_start < start)
429 err = split_vma(vma->vm_mm, vma, start, 1);
430 if (!err && vma->vm_end > end)
431 err = split_vma(vma->vm_mm, vma, end, 0);
432 if (!err)
433 err = policy_vma(vma, new);
434 if (err)
435 break;
437 return err;
441 * Update task->flags PF_MEMPOLICY bit: set iff non-default
442 * mempolicy. Allows more rapid checking of this (combined perhaps
443 * with other PF_* flag bits) on memory allocation hot code paths.
445 * If called from outside this file, the task 'p' should -only- be
446 * a newly forked child not yet visible on the task list, because
447 * manipulating the task flags of a visible task is not safe.
449 * The above limitation is why this routine has the funny name
450 * mpol_fix_fork_child_flag().
452 * It is also safe to call this with a task pointer of current,
453 * which the static wrapper mpol_set_task_struct_flag() does,
454 * for use within this file.
457 void mpol_fix_fork_child_flag(struct task_struct *p)
459 if (p->mempolicy)
460 p->flags |= PF_MEMPOLICY;
461 else
462 p->flags &= ~PF_MEMPOLICY;
465 static void mpol_set_task_struct_flag(void)
467 mpol_fix_fork_child_flag(current);
470 /* Set the process memory policy */
471 static long do_set_mempolicy(unsigned short mode, unsigned short flags,
472 nodemask_t *nodes)
474 struct mempolicy *new;
476 if (mpol_check_policy(mode, nodes))
477 return -EINVAL;
478 new = mpol_new(mode, flags, nodes);
479 if (IS_ERR(new))
480 return PTR_ERR(new);
481 mpol_free(current->mempolicy);
482 current->mempolicy = new;
483 mpol_set_task_struct_flag();
484 if (new && new->policy == MPOL_INTERLEAVE)
485 current->il_next = first_node(new->v.nodes);
486 return 0;
489 /* Fill a zone bitmap for a policy */
490 static void get_zonemask(struct mempolicy *p, nodemask_t *nodes)
492 nodes_clear(*nodes);
493 switch (p->policy) {
494 case MPOL_DEFAULT:
495 break;
496 case MPOL_BIND:
497 /* Fall through */
498 case MPOL_INTERLEAVE:
499 *nodes = p->v.nodes;
500 break;
501 case MPOL_PREFERRED:
502 /* or use current node instead of memory_map? */
503 if (p->v.preferred_node < 0)
504 *nodes = node_states[N_HIGH_MEMORY];
505 else
506 node_set(p->v.preferred_node, *nodes);
507 break;
508 default:
509 BUG();
513 static int lookup_node(struct mm_struct *mm, unsigned long addr)
515 struct page *p;
516 int err;
518 err = get_user_pages(current, mm, addr & PAGE_MASK, 1, 0, 0, &p, NULL);
519 if (err >= 0) {
520 err = page_to_nid(p);
521 put_page(p);
523 return err;
526 /* Retrieve NUMA policy */
527 static long do_get_mempolicy(int *policy, nodemask_t *nmask,
528 unsigned long addr, unsigned long flags)
530 int err;
531 struct mm_struct *mm = current->mm;
532 struct vm_area_struct *vma = NULL;
533 struct mempolicy *pol = current->mempolicy;
535 cpuset_update_task_memory_state();
536 if (flags &
537 ~(unsigned long)(MPOL_F_NODE|MPOL_F_ADDR|MPOL_F_MEMS_ALLOWED))
538 return -EINVAL;
540 if (flags & MPOL_F_MEMS_ALLOWED) {
541 if (flags & (MPOL_F_NODE|MPOL_F_ADDR))
542 return -EINVAL;
543 *policy = 0; /* just so it's initialized */
544 *nmask = cpuset_current_mems_allowed;
545 return 0;
548 if (flags & MPOL_F_ADDR) {
549 down_read(&mm->mmap_sem);
550 vma = find_vma_intersection(mm, addr, addr+1);
551 if (!vma) {
552 up_read(&mm->mmap_sem);
553 return -EFAULT;
555 if (vma->vm_ops && vma->vm_ops->get_policy)
556 pol = vma->vm_ops->get_policy(vma, addr);
557 else
558 pol = vma->vm_policy;
559 } else if (addr)
560 return -EINVAL;
562 if (!pol)
563 pol = &default_policy;
565 if (flags & MPOL_F_NODE) {
566 if (flags & MPOL_F_ADDR) {
567 err = lookup_node(mm, addr);
568 if (err < 0)
569 goto out;
570 *policy = err;
571 } else if (pol == current->mempolicy &&
572 pol->policy == MPOL_INTERLEAVE) {
573 *policy = current->il_next;
574 } else {
575 err = -EINVAL;
576 goto out;
578 } else
579 *policy = pol->policy | pol->flags;
581 if (vma) {
582 up_read(&current->mm->mmap_sem);
583 vma = NULL;
586 err = 0;
587 if (nmask)
588 get_zonemask(pol, nmask);
590 out:
591 if (vma)
592 up_read(&current->mm->mmap_sem);
593 return err;
596 #ifdef CONFIG_MIGRATION
598 * page migration
600 static void migrate_page_add(struct page *page, struct list_head *pagelist,
601 unsigned long flags)
604 * Avoid migrating a page that is shared with others.
606 if ((flags & MPOL_MF_MOVE_ALL) || page_mapcount(page) == 1)
607 isolate_lru_page(page, pagelist);
610 static struct page *new_node_page(struct page *page, unsigned long node, int **x)
612 return alloc_pages_node(node, GFP_HIGHUSER_MOVABLE, 0);
616 * Migrate pages from one node to a target node.
617 * Returns error or the number of pages not migrated.
619 static int migrate_to_node(struct mm_struct *mm, int source, int dest,
620 int flags)
622 nodemask_t nmask;
623 LIST_HEAD(pagelist);
624 int err = 0;
626 nodes_clear(nmask);
627 node_set(source, nmask);
629 check_range(mm, mm->mmap->vm_start, TASK_SIZE, &nmask,
630 flags | MPOL_MF_DISCONTIG_OK, &pagelist);
632 if (!list_empty(&pagelist))
633 err = migrate_pages(&pagelist, new_node_page, dest);
635 return err;
639 * Move pages between the two nodesets so as to preserve the physical
640 * layout as much as possible.
642 * Returns the number of page that could not be moved.
644 int do_migrate_pages(struct mm_struct *mm,
645 const nodemask_t *from_nodes, const nodemask_t *to_nodes, int flags)
647 LIST_HEAD(pagelist);
648 int busy = 0;
649 int err = 0;
650 nodemask_t tmp;
652 down_read(&mm->mmap_sem);
654 err = migrate_vmas(mm, from_nodes, to_nodes, flags);
655 if (err)
656 goto out;
659 * Find a 'source' bit set in 'tmp' whose corresponding 'dest'
660 * bit in 'to' is not also set in 'tmp'. Clear the found 'source'
661 * bit in 'tmp', and return that <source, dest> pair for migration.
662 * The pair of nodemasks 'to' and 'from' define the map.
664 * If no pair of bits is found that way, fallback to picking some
665 * pair of 'source' and 'dest' bits that are not the same. If the
666 * 'source' and 'dest' bits are the same, this represents a node
667 * that will be migrating to itself, so no pages need move.
669 * If no bits are left in 'tmp', or if all remaining bits left
670 * in 'tmp' correspond to the same bit in 'to', return false
671 * (nothing left to migrate).
673 * This lets us pick a pair of nodes to migrate between, such that
674 * if possible the dest node is not already occupied by some other
675 * source node, minimizing the risk of overloading the memory on a
676 * node that would happen if we migrated incoming memory to a node
677 * before migrating outgoing memory source that same node.
679 * A single scan of tmp is sufficient. As we go, we remember the
680 * most recent <s, d> pair that moved (s != d). If we find a pair
681 * that not only moved, but what's better, moved to an empty slot
682 * (d is not set in tmp), then we break out then, with that pair.
683 * Otherwise when we finish scannng from_tmp, we at least have the
684 * most recent <s, d> pair that moved. If we get all the way through
685 * the scan of tmp without finding any node that moved, much less
686 * moved to an empty node, then there is nothing left worth migrating.
689 tmp = *from_nodes;
690 while (!nodes_empty(tmp)) {
691 int s,d;
692 int source = -1;
693 int dest = 0;
695 for_each_node_mask(s, tmp) {
696 d = node_remap(s, *from_nodes, *to_nodes);
697 if (s == d)
698 continue;
700 source = s; /* Node moved. Memorize */
701 dest = d;
703 /* dest not in remaining from nodes? */
704 if (!node_isset(dest, tmp))
705 break;
707 if (source == -1)
708 break;
710 node_clear(source, tmp);
711 err = migrate_to_node(mm, source, dest, flags);
712 if (err > 0)
713 busy += err;
714 if (err < 0)
715 break;
717 out:
718 up_read(&mm->mmap_sem);
719 if (err < 0)
720 return err;
721 return busy;
726 * Allocate a new page for page migration based on vma policy.
727 * Start assuming that page is mapped by vma pointed to by @private.
728 * Search forward from there, if not. N.B., this assumes that the
729 * list of pages handed to migrate_pages()--which is how we get here--
730 * is in virtual address order.
732 static struct page *new_vma_page(struct page *page, unsigned long private, int **x)
734 struct vm_area_struct *vma = (struct vm_area_struct *)private;
735 unsigned long uninitialized_var(address);
737 while (vma) {
738 address = page_address_in_vma(page, vma);
739 if (address != -EFAULT)
740 break;
741 vma = vma->vm_next;
745 * if !vma, alloc_page_vma() will use task or system default policy
747 return alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
749 #else
751 static void migrate_page_add(struct page *page, struct list_head *pagelist,
752 unsigned long flags)
756 int do_migrate_pages(struct mm_struct *mm,
757 const nodemask_t *from_nodes, const nodemask_t *to_nodes, int flags)
759 return -ENOSYS;
762 static struct page *new_vma_page(struct page *page, unsigned long private, int **x)
764 return NULL;
766 #endif
768 static long do_mbind(unsigned long start, unsigned long len,
769 unsigned short mode, unsigned short mode_flags,
770 nodemask_t *nmask, unsigned long flags)
772 struct vm_area_struct *vma;
773 struct mm_struct *mm = current->mm;
774 struct mempolicy *new;
775 unsigned long end;
776 int err;
777 LIST_HEAD(pagelist);
779 if (flags & ~(unsigned long)(MPOL_MF_STRICT |
780 MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))
781 return -EINVAL;
782 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
783 return -EPERM;
785 if (start & ~PAGE_MASK)
786 return -EINVAL;
788 if (mode == MPOL_DEFAULT)
789 flags &= ~MPOL_MF_STRICT;
791 len = (len + PAGE_SIZE - 1) & PAGE_MASK;
792 end = start + len;
794 if (end < start)
795 return -EINVAL;
796 if (end == start)
797 return 0;
799 if (mpol_check_policy(mode, nmask))
800 return -EINVAL;
802 new = mpol_new(mode, mode_flags, nmask);
803 if (IS_ERR(new))
804 return PTR_ERR(new);
807 * If we are using the default policy then operation
808 * on discontinuous address spaces is okay after all
810 if (!new)
811 flags |= MPOL_MF_DISCONTIG_OK;
813 pr_debug("mbind %lx-%lx mode:%d flags:%d nodes:%lx\n",
814 start, start + len, mode, mode_flags,
815 nmask ? nodes_addr(*nmask)[0] : -1);
817 down_write(&mm->mmap_sem);
818 vma = check_range(mm, start, end, nmask,
819 flags | MPOL_MF_INVERT, &pagelist);
821 err = PTR_ERR(vma);
822 if (!IS_ERR(vma)) {
823 int nr_failed = 0;
825 err = mbind_range(vma, start, end, new);
827 if (!list_empty(&pagelist))
828 nr_failed = migrate_pages(&pagelist, new_vma_page,
829 (unsigned long)vma);
831 if (!err && nr_failed && (flags & MPOL_MF_STRICT))
832 err = -EIO;
835 up_write(&mm->mmap_sem);
836 mpol_free(new);
837 return err;
841 * User space interface with variable sized bitmaps for nodelists.
844 /* Copy a node mask from user space. */
845 static int get_nodes(nodemask_t *nodes, const unsigned long __user *nmask,
846 unsigned long maxnode)
848 unsigned long k;
849 unsigned long nlongs;
850 unsigned long endmask;
852 --maxnode;
853 nodes_clear(*nodes);
854 if (maxnode == 0 || !nmask)
855 return 0;
856 if (maxnode > PAGE_SIZE*BITS_PER_BYTE)
857 return -EINVAL;
859 nlongs = BITS_TO_LONGS(maxnode);
860 if ((maxnode % BITS_PER_LONG) == 0)
861 endmask = ~0UL;
862 else
863 endmask = (1UL << (maxnode % BITS_PER_LONG)) - 1;
865 /* When the user specified more nodes than supported just check
866 if the non supported part is all zero. */
867 if (nlongs > BITS_TO_LONGS(MAX_NUMNODES)) {
868 if (nlongs > PAGE_SIZE/sizeof(long))
869 return -EINVAL;
870 for (k = BITS_TO_LONGS(MAX_NUMNODES); k < nlongs; k++) {
871 unsigned long t;
872 if (get_user(t, nmask + k))
873 return -EFAULT;
874 if (k == nlongs - 1) {
875 if (t & endmask)
876 return -EINVAL;
877 } else if (t)
878 return -EINVAL;
880 nlongs = BITS_TO_LONGS(MAX_NUMNODES);
881 endmask = ~0UL;
884 if (copy_from_user(nodes_addr(*nodes), nmask, nlongs*sizeof(unsigned long)))
885 return -EFAULT;
886 nodes_addr(*nodes)[nlongs-1] &= endmask;
887 return 0;
890 /* Copy a kernel node mask to user space */
891 static int copy_nodes_to_user(unsigned long __user *mask, unsigned long maxnode,
892 nodemask_t *nodes)
894 unsigned long copy = ALIGN(maxnode-1, 64) / 8;
895 const int nbytes = BITS_TO_LONGS(MAX_NUMNODES) * sizeof(long);
897 if (copy > nbytes) {
898 if (copy > PAGE_SIZE)
899 return -EINVAL;
900 if (clear_user((char __user *)mask + nbytes, copy - nbytes))
901 return -EFAULT;
902 copy = nbytes;
904 return copy_to_user(mask, nodes_addr(*nodes), copy) ? -EFAULT : 0;
907 asmlinkage long sys_mbind(unsigned long start, unsigned long len,
908 unsigned long mode,
909 unsigned long __user *nmask, unsigned long maxnode,
910 unsigned flags)
912 nodemask_t nodes;
913 int err;
914 unsigned short mode_flags;
916 mode_flags = mode & MPOL_MODE_FLAGS;
917 mode &= ~MPOL_MODE_FLAGS;
918 if (mode >= MPOL_MAX)
919 return -EINVAL;
920 err = get_nodes(&nodes, nmask, maxnode);
921 if (err)
922 return err;
923 return do_mbind(start, len, mode, mode_flags, &nodes, flags);
926 /* Set the process memory policy */
927 asmlinkage long sys_set_mempolicy(int mode, unsigned long __user *nmask,
928 unsigned long maxnode)
930 int err;
931 nodemask_t nodes;
932 unsigned short flags;
934 flags = mode & MPOL_MODE_FLAGS;
935 mode &= ~MPOL_MODE_FLAGS;
936 if ((unsigned int)mode >= MPOL_MAX)
937 return -EINVAL;
938 err = get_nodes(&nodes, nmask, maxnode);
939 if (err)
940 return err;
941 return do_set_mempolicy(mode, flags, &nodes);
944 asmlinkage long sys_migrate_pages(pid_t pid, unsigned long maxnode,
945 const unsigned long __user *old_nodes,
946 const unsigned long __user *new_nodes)
948 struct mm_struct *mm;
949 struct task_struct *task;
950 nodemask_t old;
951 nodemask_t new;
952 nodemask_t task_nodes;
953 int err;
955 err = get_nodes(&old, old_nodes, maxnode);
956 if (err)
957 return err;
959 err = get_nodes(&new, new_nodes, maxnode);
960 if (err)
961 return err;
963 /* Find the mm_struct */
964 read_lock(&tasklist_lock);
965 task = pid ? find_task_by_vpid(pid) : current;
966 if (!task) {
967 read_unlock(&tasklist_lock);
968 return -ESRCH;
970 mm = get_task_mm(task);
971 read_unlock(&tasklist_lock);
973 if (!mm)
974 return -EINVAL;
977 * Check if this process has the right to modify the specified
978 * process. The right exists if the process has administrative
979 * capabilities, superuser privileges or the same
980 * userid as the target process.
982 if ((current->euid != task->suid) && (current->euid != task->uid) &&
983 (current->uid != task->suid) && (current->uid != task->uid) &&
984 !capable(CAP_SYS_NICE)) {
985 err = -EPERM;
986 goto out;
989 task_nodes = cpuset_mems_allowed(task);
990 /* Is the user allowed to access the target nodes? */
991 if (!nodes_subset(new, task_nodes) && !capable(CAP_SYS_NICE)) {
992 err = -EPERM;
993 goto out;
996 if (!nodes_subset(new, node_states[N_HIGH_MEMORY])) {
997 err = -EINVAL;
998 goto out;
1001 err = security_task_movememory(task);
1002 if (err)
1003 goto out;
1005 err = do_migrate_pages(mm, &old, &new,
1006 capable(CAP_SYS_NICE) ? MPOL_MF_MOVE_ALL : MPOL_MF_MOVE);
1007 out:
1008 mmput(mm);
1009 return err;
1013 /* Retrieve NUMA policy */
1014 asmlinkage long sys_get_mempolicy(int __user *policy,
1015 unsigned long __user *nmask,
1016 unsigned long maxnode,
1017 unsigned long addr, unsigned long flags)
1019 int err;
1020 int uninitialized_var(pval);
1021 nodemask_t nodes;
1023 if (nmask != NULL && maxnode < MAX_NUMNODES)
1024 return -EINVAL;
1026 err = do_get_mempolicy(&pval, &nodes, addr, flags);
1028 if (err)
1029 return err;
1031 if (policy && put_user(pval, policy))
1032 return -EFAULT;
1034 if (nmask)
1035 err = copy_nodes_to_user(nmask, maxnode, &nodes);
1037 return err;
1040 #ifdef CONFIG_COMPAT
1042 asmlinkage long compat_sys_get_mempolicy(int __user *policy,
1043 compat_ulong_t __user *nmask,
1044 compat_ulong_t maxnode,
1045 compat_ulong_t addr, compat_ulong_t flags)
1047 long err;
1048 unsigned long __user *nm = NULL;
1049 unsigned long nr_bits, alloc_size;
1050 DECLARE_BITMAP(bm, MAX_NUMNODES);
1052 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1053 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1055 if (nmask)
1056 nm = compat_alloc_user_space(alloc_size);
1058 err = sys_get_mempolicy(policy, nm, nr_bits+1, addr, flags);
1060 if (!err && nmask) {
1061 err = copy_from_user(bm, nm, alloc_size);
1062 /* ensure entire bitmap is zeroed */
1063 err |= clear_user(nmask, ALIGN(maxnode-1, 8) / 8);
1064 err |= compat_put_bitmap(nmask, bm, nr_bits);
1067 return err;
1070 asmlinkage long compat_sys_set_mempolicy(int mode, compat_ulong_t __user *nmask,
1071 compat_ulong_t maxnode)
1073 long err = 0;
1074 unsigned long __user *nm = NULL;
1075 unsigned long nr_bits, alloc_size;
1076 DECLARE_BITMAP(bm, MAX_NUMNODES);
1078 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1079 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1081 if (nmask) {
1082 err = compat_get_bitmap(bm, nmask, nr_bits);
1083 nm = compat_alloc_user_space(alloc_size);
1084 err |= copy_to_user(nm, bm, alloc_size);
1087 if (err)
1088 return -EFAULT;
1090 return sys_set_mempolicy(mode, nm, nr_bits+1);
1093 asmlinkage long compat_sys_mbind(compat_ulong_t start, compat_ulong_t len,
1094 compat_ulong_t mode, compat_ulong_t __user *nmask,
1095 compat_ulong_t maxnode, compat_ulong_t flags)
1097 long err = 0;
1098 unsigned long __user *nm = NULL;
1099 unsigned long nr_bits, alloc_size;
1100 nodemask_t bm;
1102 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1103 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1105 if (nmask) {
1106 err = compat_get_bitmap(nodes_addr(bm), nmask, nr_bits);
1107 nm = compat_alloc_user_space(alloc_size);
1108 err |= copy_to_user(nm, nodes_addr(bm), alloc_size);
1111 if (err)
1112 return -EFAULT;
1114 return sys_mbind(start, len, mode, nm, nr_bits+1, flags);
1117 #endif
1120 * get_vma_policy(@task, @vma, @addr)
1121 * @task - task for fallback if vma policy == default
1122 * @vma - virtual memory area whose policy is sought
1123 * @addr - address in @vma for shared policy lookup
1125 * Returns effective policy for a VMA at specified address.
1126 * Falls back to @task or system default policy, as necessary.
1127 * Returned policy has extra reference count if shared, vma,
1128 * or some other task's policy [show_numa_maps() can pass
1129 * @task != current]. It is the caller's responsibility to
1130 * free the reference in these cases.
1132 static struct mempolicy * get_vma_policy(struct task_struct *task,
1133 struct vm_area_struct *vma, unsigned long addr)
1135 struct mempolicy *pol = task->mempolicy;
1136 int shared_pol = 0;
1138 if (vma) {
1139 if (vma->vm_ops && vma->vm_ops->get_policy) {
1140 pol = vma->vm_ops->get_policy(vma, addr);
1141 shared_pol = 1; /* if pol non-NULL, add ref below */
1142 } else if (vma->vm_policy &&
1143 vma->vm_policy->policy != MPOL_DEFAULT)
1144 pol = vma->vm_policy;
1146 if (!pol)
1147 pol = &default_policy;
1148 else if (!shared_pol && pol != current->mempolicy)
1149 mpol_get(pol); /* vma or other task's policy */
1150 return pol;
1153 /* Return a nodemask representing a mempolicy */
1154 static nodemask_t *nodemask_policy(gfp_t gfp, struct mempolicy *policy)
1156 /* Lower zones don't get a nodemask applied for MPOL_BIND */
1157 if (unlikely(policy->policy == MPOL_BIND) &&
1158 gfp_zone(gfp) >= policy_zone &&
1159 cpuset_nodemask_valid_mems_allowed(&policy->v.nodes))
1160 return &policy->v.nodes;
1162 return NULL;
1165 /* Return a zonelist representing a mempolicy */
1166 static struct zonelist *zonelist_policy(gfp_t gfp, struct mempolicy *policy)
1168 int nd;
1170 switch (policy->policy) {
1171 case MPOL_PREFERRED:
1172 nd = policy->v.preferred_node;
1173 if (nd < 0)
1174 nd = numa_node_id();
1175 break;
1176 case MPOL_BIND:
1178 * Normally, MPOL_BIND allocations node-local are node-local
1179 * within the allowed nodemask. However, if __GFP_THISNODE is
1180 * set and the current node is part of the mask, we use the
1181 * the zonelist for the first node in the mask instead.
1183 nd = numa_node_id();
1184 if (unlikely(gfp & __GFP_THISNODE) &&
1185 unlikely(!node_isset(nd, policy->v.nodes)))
1186 nd = first_node(policy->v.nodes);
1187 break;
1188 case MPOL_INTERLEAVE: /* should not happen */
1189 case MPOL_DEFAULT:
1190 nd = numa_node_id();
1191 break;
1192 default:
1193 nd = 0;
1194 BUG();
1196 return node_zonelist(nd, gfp);
1199 /* Do dynamic interleaving for a process */
1200 static unsigned interleave_nodes(struct mempolicy *policy)
1202 unsigned nid, next;
1203 struct task_struct *me = current;
1205 nid = me->il_next;
1206 next = next_node(nid, policy->v.nodes);
1207 if (next >= MAX_NUMNODES)
1208 next = first_node(policy->v.nodes);
1209 me->il_next = next;
1210 return nid;
1214 * Depending on the memory policy provide a node from which to allocate the
1215 * next slab entry.
1217 unsigned slab_node(struct mempolicy *policy)
1219 unsigned short pol = policy ? policy->policy : MPOL_DEFAULT;
1221 switch (pol) {
1222 case MPOL_INTERLEAVE:
1223 return interleave_nodes(policy);
1225 case MPOL_BIND: {
1227 * Follow bind policy behavior and start allocation at the
1228 * first node.
1230 struct zonelist *zonelist;
1231 struct zone *zone;
1232 enum zone_type highest_zoneidx = gfp_zone(GFP_KERNEL);
1233 zonelist = &NODE_DATA(numa_node_id())->node_zonelists[0];
1234 (void)first_zones_zonelist(zonelist, highest_zoneidx,
1235 &policy->v.nodes,
1236 &zone);
1237 return zone->node;
1240 case MPOL_PREFERRED:
1241 if (policy->v.preferred_node >= 0)
1242 return policy->v.preferred_node;
1243 /* Fall through */
1245 default:
1246 return numa_node_id();
1250 /* Do static interleaving for a VMA with known offset. */
1251 static unsigned offset_il_node(struct mempolicy *pol,
1252 struct vm_area_struct *vma, unsigned long off)
1254 unsigned nnodes = nodes_weight(pol->v.nodes);
1255 unsigned target = (unsigned)off % nnodes;
1256 int c;
1257 int nid = -1;
1259 c = 0;
1260 do {
1261 nid = next_node(nid, pol->v.nodes);
1262 c++;
1263 } while (c <= target);
1264 return nid;
1267 /* Determine a node number for interleave */
1268 static inline unsigned interleave_nid(struct mempolicy *pol,
1269 struct vm_area_struct *vma, unsigned long addr, int shift)
1271 if (vma) {
1272 unsigned long off;
1275 * for small pages, there is no difference between
1276 * shift and PAGE_SHIFT, so the bit-shift is safe.
1277 * for huge pages, since vm_pgoff is in units of small
1278 * pages, we need to shift off the always 0 bits to get
1279 * a useful offset.
1281 BUG_ON(shift < PAGE_SHIFT);
1282 off = vma->vm_pgoff >> (shift - PAGE_SHIFT);
1283 off += (addr - vma->vm_start) >> shift;
1284 return offset_il_node(pol, vma, off);
1285 } else
1286 return interleave_nodes(pol);
1289 #ifdef CONFIG_HUGETLBFS
1291 * huge_zonelist(@vma, @addr, @gfp_flags, @mpol)
1292 * @vma = virtual memory area whose policy is sought
1293 * @addr = address in @vma for shared policy lookup and interleave policy
1294 * @gfp_flags = for requested zone
1295 * @mpol = pointer to mempolicy pointer for reference counted mempolicy
1296 * @nodemask = pointer to nodemask pointer for MPOL_BIND nodemask
1298 * Returns a zonelist suitable for a huge page allocation.
1299 * If the effective policy is 'BIND, returns pointer to local node's zonelist,
1300 * and a pointer to the mempolicy's @nodemask for filtering the zonelist.
1301 * If it is also a policy for which get_vma_policy() returns an extra
1302 * reference, we must hold that reference until after the allocation.
1303 * In that case, return policy via @mpol so hugetlb allocation can drop
1304 * the reference. For non-'BIND referenced policies, we can/do drop the
1305 * reference here, so the caller doesn't need to know about the special case
1306 * for default and current task policy.
1308 struct zonelist *huge_zonelist(struct vm_area_struct *vma, unsigned long addr,
1309 gfp_t gfp_flags, struct mempolicy **mpol,
1310 nodemask_t **nodemask)
1312 struct mempolicy *pol = get_vma_policy(current, vma, addr);
1313 struct zonelist *zl;
1315 *mpol = NULL; /* probably no unref needed */
1316 *nodemask = NULL; /* assume !MPOL_BIND */
1317 if (pol->policy == MPOL_BIND) {
1318 *nodemask = &pol->v.nodes;
1319 } else if (pol->policy == MPOL_INTERLEAVE) {
1320 unsigned nid;
1322 nid = interleave_nid(pol, vma, addr, HPAGE_SHIFT);
1323 if (unlikely(pol != &default_policy &&
1324 pol != current->mempolicy))
1325 __mpol_free(pol); /* finished with pol */
1326 return node_zonelist(nid, gfp_flags);
1329 zl = zonelist_policy(GFP_HIGHUSER, pol);
1330 if (unlikely(pol != &default_policy && pol != current->mempolicy)) {
1331 if (pol->policy != MPOL_BIND)
1332 __mpol_free(pol); /* finished with pol */
1333 else
1334 *mpol = pol; /* unref needed after allocation */
1336 return zl;
1338 #endif
1340 /* Allocate a page in interleaved policy.
1341 Own path because it needs to do special accounting. */
1342 static struct page *alloc_page_interleave(gfp_t gfp, unsigned order,
1343 unsigned nid)
1345 struct zonelist *zl;
1346 struct page *page;
1348 zl = node_zonelist(nid, gfp);
1349 page = __alloc_pages(gfp, order, zl);
1350 if (page && page_zone(page) == zonelist_zone(&zl->_zonerefs[0]))
1351 inc_zone_page_state(page, NUMA_INTERLEAVE_HIT);
1352 return page;
1356 * alloc_page_vma - Allocate a page for a VMA.
1358 * @gfp:
1359 * %GFP_USER user allocation.
1360 * %GFP_KERNEL kernel allocations,
1361 * %GFP_HIGHMEM highmem/user allocations,
1362 * %GFP_FS allocation should not call back into a file system.
1363 * %GFP_ATOMIC don't sleep.
1365 * @vma: Pointer to VMA or NULL if not available.
1366 * @addr: Virtual Address of the allocation. Must be inside the VMA.
1368 * This function allocates a page from the kernel page pool and applies
1369 * a NUMA policy associated with the VMA or the current process.
1370 * When VMA is not NULL caller must hold down_read on the mmap_sem of the
1371 * mm_struct of the VMA to prevent it from going away. Should be used for
1372 * all allocations for pages that will be mapped into
1373 * user space. Returns NULL when no page can be allocated.
1375 * Should be called with the mm_sem of the vma hold.
1377 struct page *
1378 alloc_page_vma(gfp_t gfp, struct vm_area_struct *vma, unsigned long addr)
1380 struct mempolicy *pol = get_vma_policy(current, vma, addr);
1381 struct zonelist *zl;
1383 cpuset_update_task_memory_state();
1385 if (unlikely(pol->policy == MPOL_INTERLEAVE)) {
1386 unsigned nid;
1388 nid = interleave_nid(pol, vma, addr, PAGE_SHIFT);
1389 if (unlikely(pol != &default_policy &&
1390 pol != current->mempolicy))
1391 __mpol_free(pol); /* finished with pol */
1392 return alloc_page_interleave(gfp, 0, nid);
1394 zl = zonelist_policy(gfp, pol);
1395 if (pol != &default_policy && pol != current->mempolicy) {
1397 * slow path: ref counted policy -- shared or vma
1399 struct page *page = __alloc_pages_nodemask(gfp, 0,
1400 zl, nodemask_policy(gfp, pol));
1401 __mpol_free(pol);
1402 return page;
1405 * fast path: default or task policy
1407 return __alloc_pages_nodemask(gfp, 0, zl, nodemask_policy(gfp, pol));
1411 * alloc_pages_current - Allocate pages.
1413 * @gfp:
1414 * %GFP_USER user allocation,
1415 * %GFP_KERNEL kernel allocation,
1416 * %GFP_HIGHMEM highmem allocation,
1417 * %GFP_FS don't call back into a file system.
1418 * %GFP_ATOMIC don't sleep.
1419 * @order: Power of two of allocation size in pages. 0 is a single page.
1421 * Allocate a page from the kernel page pool. When not in
1422 * interrupt context and apply the current process NUMA policy.
1423 * Returns NULL when no page can be allocated.
1425 * Don't call cpuset_update_task_memory_state() unless
1426 * 1) it's ok to take cpuset_sem (can WAIT), and
1427 * 2) allocating for current task (not interrupt).
1429 struct page *alloc_pages_current(gfp_t gfp, unsigned order)
1431 struct mempolicy *pol = current->mempolicy;
1433 if ((gfp & __GFP_WAIT) && !in_interrupt())
1434 cpuset_update_task_memory_state();
1435 if (!pol || in_interrupt() || (gfp & __GFP_THISNODE))
1436 pol = &default_policy;
1437 if (pol->policy == MPOL_INTERLEAVE)
1438 return alloc_page_interleave(gfp, order, interleave_nodes(pol));
1439 return __alloc_pages_nodemask(gfp, order,
1440 zonelist_policy(gfp, pol), nodemask_policy(gfp, pol));
1442 EXPORT_SYMBOL(alloc_pages_current);
1445 * If mpol_copy() sees current->cpuset == cpuset_being_rebound, then it
1446 * rebinds the mempolicy its copying by calling mpol_rebind_policy()
1447 * with the mems_allowed returned by cpuset_mems_allowed(). This
1448 * keeps mempolicies cpuset relative after its cpuset moves. See
1449 * further kernel/cpuset.c update_nodemask().
1452 /* Slow path of a mempolicy copy */
1453 struct mempolicy *__mpol_copy(struct mempolicy *old)
1455 struct mempolicy *new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
1457 if (!new)
1458 return ERR_PTR(-ENOMEM);
1459 if (current_cpuset_is_being_rebound()) {
1460 nodemask_t mems = cpuset_mems_allowed(current);
1461 mpol_rebind_policy(old, &mems);
1463 *new = *old;
1464 atomic_set(&new->refcnt, 1);
1465 return new;
1468 /* Slow path of a mempolicy comparison */
1469 int __mpol_equal(struct mempolicy *a, struct mempolicy *b)
1471 if (!a || !b)
1472 return 0;
1473 if (a->policy != b->policy)
1474 return 0;
1475 switch (a->policy) {
1476 case MPOL_DEFAULT:
1477 return 1;
1478 case MPOL_BIND:
1479 /* Fall through */
1480 case MPOL_INTERLEAVE:
1481 return nodes_equal(a->v.nodes, b->v.nodes);
1482 case MPOL_PREFERRED:
1483 return a->v.preferred_node == b->v.preferred_node;
1484 default:
1485 BUG();
1486 return 0;
1490 /* Slow path of a mpol destructor. */
1491 void __mpol_free(struct mempolicy *p)
1493 if (!atomic_dec_and_test(&p->refcnt))
1494 return;
1495 p->policy = MPOL_DEFAULT;
1496 kmem_cache_free(policy_cache, p);
1500 * Shared memory backing store policy support.
1502 * Remember policies even when nobody has shared memory mapped.
1503 * The policies are kept in Red-Black tree linked from the inode.
1504 * They are protected by the sp->lock spinlock, which should be held
1505 * for any accesses to the tree.
1508 /* lookup first element intersecting start-end */
1509 /* Caller holds sp->lock */
1510 static struct sp_node *
1511 sp_lookup(struct shared_policy *sp, unsigned long start, unsigned long end)
1513 struct rb_node *n = sp->root.rb_node;
1515 while (n) {
1516 struct sp_node *p = rb_entry(n, struct sp_node, nd);
1518 if (start >= p->end)
1519 n = n->rb_right;
1520 else if (end <= p->start)
1521 n = n->rb_left;
1522 else
1523 break;
1525 if (!n)
1526 return NULL;
1527 for (;;) {
1528 struct sp_node *w = NULL;
1529 struct rb_node *prev = rb_prev(n);
1530 if (!prev)
1531 break;
1532 w = rb_entry(prev, struct sp_node, nd);
1533 if (w->end <= start)
1534 break;
1535 n = prev;
1537 return rb_entry(n, struct sp_node, nd);
1540 /* Insert a new shared policy into the list. */
1541 /* Caller holds sp->lock */
1542 static void sp_insert(struct shared_policy *sp, struct sp_node *new)
1544 struct rb_node **p = &sp->root.rb_node;
1545 struct rb_node *parent = NULL;
1546 struct sp_node *nd;
1548 while (*p) {
1549 parent = *p;
1550 nd = rb_entry(parent, struct sp_node, nd);
1551 if (new->start < nd->start)
1552 p = &(*p)->rb_left;
1553 else if (new->end > nd->end)
1554 p = &(*p)->rb_right;
1555 else
1556 BUG();
1558 rb_link_node(&new->nd, parent, p);
1559 rb_insert_color(&new->nd, &sp->root);
1560 pr_debug("inserting %lx-%lx: %d\n", new->start, new->end,
1561 new->policy ? new->policy->policy : 0);
1564 /* Find shared policy intersecting idx */
1565 struct mempolicy *
1566 mpol_shared_policy_lookup(struct shared_policy *sp, unsigned long idx)
1568 struct mempolicy *pol = NULL;
1569 struct sp_node *sn;
1571 if (!sp->root.rb_node)
1572 return NULL;
1573 spin_lock(&sp->lock);
1574 sn = sp_lookup(sp, idx, idx+1);
1575 if (sn) {
1576 mpol_get(sn->policy);
1577 pol = sn->policy;
1579 spin_unlock(&sp->lock);
1580 return pol;
1583 static void sp_delete(struct shared_policy *sp, struct sp_node *n)
1585 pr_debug("deleting %lx-l%lx\n", n->start, n->end);
1586 rb_erase(&n->nd, &sp->root);
1587 mpol_free(n->policy);
1588 kmem_cache_free(sn_cache, n);
1591 static struct sp_node *sp_alloc(unsigned long start, unsigned long end,
1592 struct mempolicy *pol)
1594 struct sp_node *n = kmem_cache_alloc(sn_cache, GFP_KERNEL);
1596 if (!n)
1597 return NULL;
1598 n->start = start;
1599 n->end = end;
1600 mpol_get(pol);
1601 n->policy = pol;
1602 return n;
1605 /* Replace a policy range. */
1606 static int shared_policy_replace(struct shared_policy *sp, unsigned long start,
1607 unsigned long end, struct sp_node *new)
1609 struct sp_node *n, *new2 = NULL;
1611 restart:
1612 spin_lock(&sp->lock);
1613 n = sp_lookup(sp, start, end);
1614 /* Take care of old policies in the same range. */
1615 while (n && n->start < end) {
1616 struct rb_node *next = rb_next(&n->nd);
1617 if (n->start >= start) {
1618 if (n->end <= end)
1619 sp_delete(sp, n);
1620 else
1621 n->start = end;
1622 } else {
1623 /* Old policy spanning whole new range. */
1624 if (n->end > end) {
1625 if (!new2) {
1626 spin_unlock(&sp->lock);
1627 new2 = sp_alloc(end, n->end, n->policy);
1628 if (!new2)
1629 return -ENOMEM;
1630 goto restart;
1632 n->end = start;
1633 sp_insert(sp, new2);
1634 new2 = NULL;
1635 break;
1636 } else
1637 n->end = start;
1639 if (!next)
1640 break;
1641 n = rb_entry(next, struct sp_node, nd);
1643 if (new)
1644 sp_insert(sp, new);
1645 spin_unlock(&sp->lock);
1646 if (new2) {
1647 mpol_free(new2->policy);
1648 kmem_cache_free(sn_cache, new2);
1650 return 0;
1653 void mpol_shared_policy_init(struct shared_policy *info, unsigned short policy,
1654 unsigned short flags, nodemask_t *policy_nodes)
1656 info->root = RB_ROOT;
1657 spin_lock_init(&info->lock);
1659 if (policy != MPOL_DEFAULT) {
1660 struct mempolicy *newpol;
1662 /* Falls back to MPOL_DEFAULT on any error */
1663 newpol = mpol_new(policy, flags, policy_nodes);
1664 if (!IS_ERR(newpol)) {
1665 /* Create pseudo-vma that contains just the policy */
1666 struct vm_area_struct pvma;
1668 memset(&pvma, 0, sizeof(struct vm_area_struct));
1669 /* Policy covers entire file */
1670 pvma.vm_end = TASK_SIZE;
1671 mpol_set_shared_policy(info, &pvma, newpol);
1672 mpol_free(newpol);
1677 int mpol_set_shared_policy(struct shared_policy *info,
1678 struct vm_area_struct *vma, struct mempolicy *npol)
1680 int err;
1681 struct sp_node *new = NULL;
1682 unsigned long sz = vma_pages(vma);
1684 pr_debug("set_shared_policy %lx sz %lu %d %d %lx\n",
1685 vma->vm_pgoff,
1686 sz, npol ? npol->policy : -1,
1687 npol ? npol->flags : -1,
1688 npol ? nodes_addr(npol->v.nodes)[0] : -1);
1690 if (npol) {
1691 new = sp_alloc(vma->vm_pgoff, vma->vm_pgoff + sz, npol);
1692 if (!new)
1693 return -ENOMEM;
1695 err = shared_policy_replace(info, vma->vm_pgoff, vma->vm_pgoff+sz, new);
1696 if (err && new)
1697 kmem_cache_free(sn_cache, new);
1698 return err;
1701 /* Free a backing policy store on inode delete. */
1702 void mpol_free_shared_policy(struct shared_policy *p)
1704 struct sp_node *n;
1705 struct rb_node *next;
1707 if (!p->root.rb_node)
1708 return;
1709 spin_lock(&p->lock);
1710 next = rb_first(&p->root);
1711 while (next) {
1712 n = rb_entry(next, struct sp_node, nd);
1713 next = rb_next(&n->nd);
1714 rb_erase(&n->nd, &p->root);
1715 mpol_free(n->policy);
1716 kmem_cache_free(sn_cache, n);
1718 spin_unlock(&p->lock);
1721 /* assumes fs == KERNEL_DS */
1722 void __init numa_policy_init(void)
1724 nodemask_t interleave_nodes;
1725 unsigned long largest = 0;
1726 int nid, prefer = 0;
1728 policy_cache = kmem_cache_create("numa_policy",
1729 sizeof(struct mempolicy),
1730 0, SLAB_PANIC, NULL);
1732 sn_cache = kmem_cache_create("shared_policy_node",
1733 sizeof(struct sp_node),
1734 0, SLAB_PANIC, NULL);
1737 * Set interleaving policy for system init. Interleaving is only
1738 * enabled across suitably sized nodes (default is >= 16MB), or
1739 * fall back to the largest node if they're all smaller.
1741 nodes_clear(interleave_nodes);
1742 for_each_node_state(nid, N_HIGH_MEMORY) {
1743 unsigned long total_pages = node_present_pages(nid);
1745 /* Preserve the largest node */
1746 if (largest < total_pages) {
1747 largest = total_pages;
1748 prefer = nid;
1751 /* Interleave this node? */
1752 if ((total_pages << PAGE_SHIFT) >= (16 << 20))
1753 node_set(nid, interleave_nodes);
1756 /* All too small, use the largest */
1757 if (unlikely(nodes_empty(interleave_nodes)))
1758 node_set(prefer, interleave_nodes);
1760 if (do_set_mempolicy(MPOL_INTERLEAVE, 0, &interleave_nodes))
1761 printk("numa_policy_init: interleaving failed\n");
1764 /* Reset policy of current process to default */
1765 void numa_default_policy(void)
1767 do_set_mempolicy(MPOL_DEFAULT, 0, NULL);
1770 /* Migrate a policy to a different set of nodes */
1771 static void mpol_rebind_policy(struct mempolicy *pol,
1772 const nodemask_t *newmask)
1774 nodemask_t *mpolmask;
1775 nodemask_t tmp;
1777 if (!pol)
1778 return;
1779 mpolmask = &pol->cpuset_mems_allowed;
1780 if (nodes_equal(*mpolmask, *newmask))
1781 return;
1783 switch (pol->policy) {
1784 case MPOL_DEFAULT:
1785 break;
1786 case MPOL_BIND:
1787 /* Fall through */
1788 case MPOL_INTERLEAVE:
1789 nodes_remap(tmp, pol->v.nodes, *mpolmask, *newmask);
1790 pol->v.nodes = tmp;
1791 *mpolmask = *newmask;
1792 current->il_next = node_remap(current->il_next,
1793 *mpolmask, *newmask);
1794 break;
1795 case MPOL_PREFERRED:
1796 pol->v.preferred_node = node_remap(pol->v.preferred_node,
1797 *mpolmask, *newmask);
1798 *mpolmask = *newmask;
1799 break;
1800 default:
1801 BUG();
1802 break;
1807 * Wrapper for mpol_rebind_policy() that just requires task
1808 * pointer, and updates task mempolicy.
1811 void mpol_rebind_task(struct task_struct *tsk, const nodemask_t *new)
1813 mpol_rebind_policy(tsk->mempolicy, new);
1817 * Rebind each vma in mm to new nodemask.
1819 * Call holding a reference to mm. Takes mm->mmap_sem during call.
1822 void mpol_rebind_mm(struct mm_struct *mm, nodemask_t *new)
1824 struct vm_area_struct *vma;
1826 down_write(&mm->mmap_sem);
1827 for (vma = mm->mmap; vma; vma = vma->vm_next)
1828 mpol_rebind_policy(vma->vm_policy, new);
1829 up_write(&mm->mmap_sem);
1833 * Display pages allocated per node and memory policy via /proc.
1836 static const char * const policy_types[] =
1837 { "default", "prefer", "bind", "interleave" };
1840 * Convert a mempolicy into a string.
1841 * Returns the number of characters in buffer (if positive)
1842 * or an error (negative)
1844 static inline int mpol_to_str(char *buffer, int maxlen, struct mempolicy *pol)
1846 char *p = buffer;
1847 int l;
1848 nodemask_t nodes;
1849 unsigned short mode = pol ? pol->policy : MPOL_DEFAULT;
1851 switch (mode) {
1852 case MPOL_DEFAULT:
1853 nodes_clear(nodes);
1854 break;
1856 case MPOL_PREFERRED:
1857 nodes_clear(nodes);
1858 node_set(pol->v.preferred_node, nodes);
1859 break;
1861 case MPOL_BIND:
1862 /* Fall through */
1863 case MPOL_INTERLEAVE:
1864 nodes = pol->v.nodes;
1865 break;
1867 default:
1868 BUG();
1869 return -EFAULT;
1872 l = strlen(policy_types[mode]);
1873 if (buffer + maxlen < p + l + 1)
1874 return -ENOSPC;
1876 strcpy(p, policy_types[mode]);
1877 p += l;
1879 if (!nodes_empty(nodes)) {
1880 if (buffer + maxlen < p + 2)
1881 return -ENOSPC;
1882 *p++ = '=';
1883 p += nodelist_scnprintf(p, buffer + maxlen - p, nodes);
1885 return p - buffer;
1888 struct numa_maps {
1889 unsigned long pages;
1890 unsigned long anon;
1891 unsigned long active;
1892 unsigned long writeback;
1893 unsigned long mapcount_max;
1894 unsigned long dirty;
1895 unsigned long swapcache;
1896 unsigned long node[MAX_NUMNODES];
1899 static void gather_stats(struct page *page, void *private, int pte_dirty)
1901 struct numa_maps *md = private;
1902 int count = page_mapcount(page);
1904 md->pages++;
1905 if (pte_dirty || PageDirty(page))
1906 md->dirty++;
1908 if (PageSwapCache(page))
1909 md->swapcache++;
1911 if (PageActive(page))
1912 md->active++;
1914 if (PageWriteback(page))
1915 md->writeback++;
1917 if (PageAnon(page))
1918 md->anon++;
1920 if (count > md->mapcount_max)
1921 md->mapcount_max = count;
1923 md->node[page_to_nid(page)]++;
1926 #ifdef CONFIG_HUGETLB_PAGE
1927 static void check_huge_range(struct vm_area_struct *vma,
1928 unsigned long start, unsigned long end,
1929 struct numa_maps *md)
1931 unsigned long addr;
1932 struct page *page;
1934 for (addr = start; addr < end; addr += HPAGE_SIZE) {
1935 pte_t *ptep = huge_pte_offset(vma->vm_mm, addr & HPAGE_MASK);
1936 pte_t pte;
1938 if (!ptep)
1939 continue;
1941 pte = *ptep;
1942 if (pte_none(pte))
1943 continue;
1945 page = pte_page(pte);
1946 if (!page)
1947 continue;
1949 gather_stats(page, md, pte_dirty(*ptep));
1952 #else
1953 static inline void check_huge_range(struct vm_area_struct *vma,
1954 unsigned long start, unsigned long end,
1955 struct numa_maps *md)
1958 #endif
1960 int show_numa_map(struct seq_file *m, void *v)
1962 struct proc_maps_private *priv = m->private;
1963 struct vm_area_struct *vma = v;
1964 struct numa_maps *md;
1965 struct file *file = vma->vm_file;
1966 struct mm_struct *mm = vma->vm_mm;
1967 struct mempolicy *pol;
1968 int n;
1969 char buffer[50];
1971 if (!mm)
1972 return 0;
1974 md = kzalloc(sizeof(struct numa_maps), GFP_KERNEL);
1975 if (!md)
1976 return 0;
1978 pol = get_vma_policy(priv->task, vma, vma->vm_start);
1979 mpol_to_str(buffer, sizeof(buffer), pol);
1981 * unref shared or other task's mempolicy
1983 if (pol != &default_policy && pol != current->mempolicy)
1984 __mpol_free(pol);
1986 seq_printf(m, "%08lx %s", vma->vm_start, buffer);
1988 if (file) {
1989 seq_printf(m, " file=");
1990 seq_path(m, &file->f_path, "\n\t= ");
1991 } else if (vma->vm_start <= mm->brk && vma->vm_end >= mm->start_brk) {
1992 seq_printf(m, " heap");
1993 } else if (vma->vm_start <= mm->start_stack &&
1994 vma->vm_end >= mm->start_stack) {
1995 seq_printf(m, " stack");
1998 if (is_vm_hugetlb_page(vma)) {
1999 check_huge_range(vma, vma->vm_start, vma->vm_end, md);
2000 seq_printf(m, " huge");
2001 } else {
2002 check_pgd_range(vma, vma->vm_start, vma->vm_end,
2003 &node_states[N_HIGH_MEMORY], MPOL_MF_STATS, md);
2006 if (!md->pages)
2007 goto out;
2009 if (md->anon)
2010 seq_printf(m," anon=%lu",md->anon);
2012 if (md->dirty)
2013 seq_printf(m," dirty=%lu",md->dirty);
2015 if (md->pages != md->anon && md->pages != md->dirty)
2016 seq_printf(m, " mapped=%lu", md->pages);
2018 if (md->mapcount_max > 1)
2019 seq_printf(m, " mapmax=%lu", md->mapcount_max);
2021 if (md->swapcache)
2022 seq_printf(m," swapcache=%lu", md->swapcache);
2024 if (md->active < md->pages && !is_vm_hugetlb_page(vma))
2025 seq_printf(m," active=%lu", md->active);
2027 if (md->writeback)
2028 seq_printf(m," writeback=%lu", md->writeback);
2030 for_each_node_state(n, N_HIGH_MEMORY)
2031 if (md->node[n])
2032 seq_printf(m, " N%d=%lu", n, md->node[n]);
2033 out:
2034 seq_putc(m, '\n');
2035 kfree(md);
2037 if (m->count < m->size)
2038 m->version = (vma != priv->tail_vma) ? vma->vm_start : 0;
2039 return 0;