Staging: comedi: remove some RT code that lingered
[linux-2.6/linux-acpi-2.6/ibm-acpi-2.6.git] / mm / mempolicy.c
blobe08e2c4da63a5dc45bfcfb9f56cd440d557d0772
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.
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/gfp.h>
77 #include <linux/slab.h>
78 #include <linux/string.h>
79 #include <linux/module.h>
80 #include <linux/nsproxy.h>
81 #include <linux/interrupt.h>
82 #include <linux/init.h>
83 #include <linux/compat.h>
84 #include <linux/swap.h>
85 #include <linux/seq_file.h>
86 #include <linux/proc_fs.h>
87 #include <linux/migrate.h>
88 #include <linux/rmap.h>
89 #include <linux/security.h>
90 #include <linux/syscalls.h>
91 #include <linux/ctype.h>
93 #include <asm/tlbflush.h>
94 #include <asm/uaccess.h>
96 #include "internal.h"
98 /* Internal flags */
99 #define MPOL_MF_DISCONTIG_OK (MPOL_MF_INTERNAL << 0) /* Skip checks for continuous vmas */
100 #define MPOL_MF_INVERT (MPOL_MF_INTERNAL << 1) /* Invert check for nodemask */
101 #define MPOL_MF_STATS (MPOL_MF_INTERNAL << 2) /* Gather statistics */
103 static struct kmem_cache *policy_cache;
104 static struct kmem_cache *sn_cache;
106 /* Highest zone. An specific allocation for a zone below that is not
107 policied. */
108 enum zone_type policy_zone = 0;
111 * run-time system-wide default policy => local allocation
113 struct mempolicy default_policy = {
114 .refcnt = ATOMIC_INIT(1), /* never free it */
115 .mode = MPOL_PREFERRED,
116 .flags = MPOL_F_LOCAL,
119 static const struct mempolicy_operations {
120 int (*create)(struct mempolicy *pol, const nodemask_t *nodes);
121 void (*rebind)(struct mempolicy *pol, const nodemask_t *nodes);
122 } mpol_ops[MPOL_MAX];
124 /* Check that the nodemask contains at least one populated zone */
125 static int is_valid_nodemask(const nodemask_t *nodemask)
127 int nd, k;
129 /* Check that there is something useful in this mask */
130 k = policy_zone;
132 for_each_node_mask(nd, *nodemask) {
133 struct zone *z;
135 for (k = 0; k <= policy_zone; k++) {
136 z = &NODE_DATA(nd)->node_zones[k];
137 if (z->present_pages > 0)
138 return 1;
142 return 0;
145 static inline int mpol_store_user_nodemask(const struct mempolicy *pol)
147 return pol->flags & (MPOL_F_STATIC_NODES | MPOL_F_RELATIVE_NODES);
150 static void mpol_relative_nodemask(nodemask_t *ret, const nodemask_t *orig,
151 const nodemask_t *rel)
153 nodemask_t tmp;
154 nodes_fold(tmp, *orig, nodes_weight(*rel));
155 nodes_onto(*ret, tmp, *rel);
158 static int mpol_new_interleave(struct mempolicy *pol, const nodemask_t *nodes)
160 if (nodes_empty(*nodes))
161 return -EINVAL;
162 pol->v.nodes = *nodes;
163 return 0;
166 static int mpol_new_preferred(struct mempolicy *pol, const nodemask_t *nodes)
168 if (!nodes)
169 pol->flags |= MPOL_F_LOCAL; /* local allocation */
170 else if (nodes_empty(*nodes))
171 return -EINVAL; /* no allowed nodes */
172 else
173 pol->v.preferred_node = first_node(*nodes);
174 return 0;
177 static int mpol_new_bind(struct mempolicy *pol, const nodemask_t *nodes)
179 if (!is_valid_nodemask(nodes))
180 return -EINVAL;
181 pol->v.nodes = *nodes;
182 return 0;
186 * mpol_set_nodemask is called after mpol_new() to set up the nodemask, if
187 * any, for the new policy. mpol_new() has already validated the nodes
188 * parameter with respect to the policy mode and flags. But, we need to
189 * handle an empty nodemask with MPOL_PREFERRED here.
191 * Must be called holding task's alloc_lock to protect task's mems_allowed
192 * and mempolicy. May also be called holding the mmap_semaphore for write.
194 static int mpol_set_nodemask(struct mempolicy *pol, const nodemask_t *nodes)
196 nodemask_t cpuset_context_nmask;
197 int ret;
199 /* if mode is MPOL_DEFAULT, pol is NULL. This is right. */
200 if (pol == NULL)
201 return 0;
203 VM_BUG_ON(!nodes);
204 if (pol->mode == MPOL_PREFERRED && nodes_empty(*nodes))
205 nodes = NULL; /* explicit local allocation */
206 else {
207 if (pol->flags & MPOL_F_RELATIVE_NODES)
208 mpol_relative_nodemask(&cpuset_context_nmask, nodes,
209 &cpuset_current_mems_allowed);
210 else
211 nodes_and(cpuset_context_nmask, *nodes,
212 cpuset_current_mems_allowed);
213 if (mpol_store_user_nodemask(pol))
214 pol->w.user_nodemask = *nodes;
215 else
216 pol->w.cpuset_mems_allowed =
217 cpuset_current_mems_allowed;
220 ret = mpol_ops[pol->mode].create(pol,
221 nodes ? &cpuset_context_nmask : NULL);
222 return ret;
226 * This function just creates a new policy, does some check and simple
227 * initialization. You must invoke mpol_set_nodemask() to set nodes.
229 static struct mempolicy *mpol_new(unsigned short mode, unsigned short flags,
230 nodemask_t *nodes)
232 struct mempolicy *policy;
234 pr_debug("setting mode %d flags %d nodes[0] %lx\n",
235 mode, flags, nodes ? nodes_addr(*nodes)[0] : -1);
237 if (mode == MPOL_DEFAULT) {
238 if (nodes && !nodes_empty(*nodes))
239 return ERR_PTR(-EINVAL);
240 return NULL; /* simply delete any existing policy */
242 VM_BUG_ON(!nodes);
245 * MPOL_PREFERRED cannot be used with MPOL_F_STATIC_NODES or
246 * MPOL_F_RELATIVE_NODES if the nodemask is empty (local allocation).
247 * All other modes require a valid pointer to a non-empty nodemask.
249 if (mode == MPOL_PREFERRED) {
250 if (nodes_empty(*nodes)) {
251 if (((flags & MPOL_F_STATIC_NODES) ||
252 (flags & MPOL_F_RELATIVE_NODES)))
253 return ERR_PTR(-EINVAL);
255 } else if (nodes_empty(*nodes))
256 return ERR_PTR(-EINVAL);
257 policy = kmem_cache_alloc(policy_cache, GFP_KERNEL);
258 if (!policy)
259 return ERR_PTR(-ENOMEM);
260 atomic_set(&policy->refcnt, 1);
261 policy->mode = mode;
262 policy->flags = flags;
264 return policy;
267 /* Slow path of a mpol destructor. */
268 void __mpol_put(struct mempolicy *p)
270 if (!atomic_dec_and_test(&p->refcnt))
271 return;
272 kmem_cache_free(policy_cache, p);
275 static void mpol_rebind_default(struct mempolicy *pol, const nodemask_t *nodes)
279 static void mpol_rebind_nodemask(struct mempolicy *pol,
280 const nodemask_t *nodes)
282 nodemask_t tmp;
284 if (pol->flags & MPOL_F_STATIC_NODES)
285 nodes_and(tmp, pol->w.user_nodemask, *nodes);
286 else if (pol->flags & MPOL_F_RELATIVE_NODES)
287 mpol_relative_nodemask(&tmp, &pol->w.user_nodemask, nodes);
288 else {
289 nodes_remap(tmp, pol->v.nodes, pol->w.cpuset_mems_allowed,
290 *nodes);
291 pol->w.cpuset_mems_allowed = *nodes;
294 pol->v.nodes = tmp;
295 if (!node_isset(current->il_next, tmp)) {
296 current->il_next = next_node(current->il_next, tmp);
297 if (current->il_next >= MAX_NUMNODES)
298 current->il_next = first_node(tmp);
299 if (current->il_next >= MAX_NUMNODES)
300 current->il_next = numa_node_id();
304 static void mpol_rebind_preferred(struct mempolicy *pol,
305 const nodemask_t *nodes)
307 nodemask_t tmp;
309 if (pol->flags & MPOL_F_STATIC_NODES) {
310 int node = first_node(pol->w.user_nodemask);
312 if (node_isset(node, *nodes)) {
313 pol->v.preferred_node = node;
314 pol->flags &= ~MPOL_F_LOCAL;
315 } else
316 pol->flags |= MPOL_F_LOCAL;
317 } else if (pol->flags & MPOL_F_RELATIVE_NODES) {
318 mpol_relative_nodemask(&tmp, &pol->w.user_nodemask, nodes);
319 pol->v.preferred_node = first_node(tmp);
320 } else if (!(pol->flags & MPOL_F_LOCAL)) {
321 pol->v.preferred_node = node_remap(pol->v.preferred_node,
322 pol->w.cpuset_mems_allowed,
323 *nodes);
324 pol->w.cpuset_mems_allowed = *nodes;
328 /* Migrate a policy to a different set of nodes */
329 static void mpol_rebind_policy(struct mempolicy *pol,
330 const nodemask_t *newmask)
332 if (!pol)
333 return;
334 if (!mpol_store_user_nodemask(pol) &&
335 nodes_equal(pol->w.cpuset_mems_allowed, *newmask))
336 return;
337 mpol_ops[pol->mode].rebind(pol, newmask);
341 * Wrapper for mpol_rebind_policy() that just requires task
342 * pointer, and updates task mempolicy.
344 * Called with task's alloc_lock held.
347 void mpol_rebind_task(struct task_struct *tsk, const nodemask_t *new)
349 mpol_rebind_policy(tsk->mempolicy, new);
353 * Rebind each vma in mm to new nodemask.
355 * Call holding a reference to mm. Takes mm->mmap_sem during call.
358 void mpol_rebind_mm(struct mm_struct *mm, nodemask_t *new)
360 struct vm_area_struct *vma;
362 down_write(&mm->mmap_sem);
363 for (vma = mm->mmap; vma; vma = vma->vm_next)
364 mpol_rebind_policy(vma->vm_policy, new);
365 up_write(&mm->mmap_sem);
368 static const struct mempolicy_operations mpol_ops[MPOL_MAX] = {
369 [MPOL_DEFAULT] = {
370 .rebind = mpol_rebind_default,
372 [MPOL_INTERLEAVE] = {
373 .create = mpol_new_interleave,
374 .rebind = mpol_rebind_nodemask,
376 [MPOL_PREFERRED] = {
377 .create = mpol_new_preferred,
378 .rebind = mpol_rebind_preferred,
380 [MPOL_BIND] = {
381 .create = mpol_new_bind,
382 .rebind = mpol_rebind_nodemask,
386 static void gather_stats(struct page *, void *, int pte_dirty);
387 static void migrate_page_add(struct page *page, struct list_head *pagelist,
388 unsigned long flags);
390 /* Scan through pages checking if pages follow certain conditions. */
391 static int check_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
392 unsigned long addr, unsigned long end,
393 const nodemask_t *nodes, unsigned long flags,
394 void *private)
396 pte_t *orig_pte;
397 pte_t *pte;
398 spinlock_t *ptl;
400 orig_pte = pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
401 do {
402 struct page *page;
403 int nid;
405 if (!pte_present(*pte))
406 continue;
407 page = vm_normal_page(vma, addr, *pte);
408 if (!page)
409 continue;
411 * The check for PageReserved here is important to avoid
412 * handling zero pages and other pages that may have been
413 * marked special by the system.
415 * If the PageReserved would not be checked here then f.e.
416 * the location of the zero page could have an influence
417 * on MPOL_MF_STRICT, zero pages would be counted for
418 * the per node stats, and there would be useless attempts
419 * to put zero pages on the migration list.
421 if (PageReserved(page))
422 continue;
423 nid = page_to_nid(page);
424 if (node_isset(nid, *nodes) == !!(flags & MPOL_MF_INVERT))
425 continue;
427 if (flags & MPOL_MF_STATS)
428 gather_stats(page, private, pte_dirty(*pte));
429 else if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))
430 migrate_page_add(page, private, flags);
431 else
432 break;
433 } while (pte++, addr += PAGE_SIZE, addr != end);
434 pte_unmap_unlock(orig_pte, ptl);
435 return addr != end;
438 static inline int check_pmd_range(struct vm_area_struct *vma, pud_t *pud,
439 unsigned long addr, unsigned long end,
440 const nodemask_t *nodes, unsigned long flags,
441 void *private)
443 pmd_t *pmd;
444 unsigned long next;
446 pmd = pmd_offset(pud, addr);
447 do {
448 next = pmd_addr_end(addr, end);
449 if (pmd_none_or_clear_bad(pmd))
450 continue;
451 if (check_pte_range(vma, pmd, addr, next, nodes,
452 flags, private))
453 return -EIO;
454 } while (pmd++, addr = next, addr != end);
455 return 0;
458 static inline int check_pud_range(struct vm_area_struct *vma, pgd_t *pgd,
459 unsigned long addr, unsigned long end,
460 const nodemask_t *nodes, unsigned long flags,
461 void *private)
463 pud_t *pud;
464 unsigned long next;
466 pud = pud_offset(pgd, addr);
467 do {
468 next = pud_addr_end(addr, end);
469 if (pud_none_or_clear_bad(pud))
470 continue;
471 if (check_pmd_range(vma, pud, addr, next, nodes,
472 flags, private))
473 return -EIO;
474 } while (pud++, addr = next, addr != end);
475 return 0;
478 static inline int check_pgd_range(struct vm_area_struct *vma,
479 unsigned long addr, unsigned long end,
480 const nodemask_t *nodes, unsigned long flags,
481 void *private)
483 pgd_t *pgd;
484 unsigned long next;
486 pgd = pgd_offset(vma->vm_mm, addr);
487 do {
488 next = pgd_addr_end(addr, end);
489 if (pgd_none_or_clear_bad(pgd))
490 continue;
491 if (check_pud_range(vma, pgd, addr, next, nodes,
492 flags, private))
493 return -EIO;
494 } while (pgd++, addr = next, addr != end);
495 return 0;
499 * Check if all pages in a range are on a set of nodes.
500 * If pagelist != NULL then isolate pages from the LRU and
501 * put them on the pagelist.
503 static struct vm_area_struct *
504 check_range(struct mm_struct *mm, unsigned long start, unsigned long end,
505 const nodemask_t *nodes, unsigned long flags, void *private)
507 int err;
508 struct vm_area_struct *first, *vma, *prev;
511 first = find_vma(mm, start);
512 if (!first)
513 return ERR_PTR(-EFAULT);
514 prev = NULL;
515 for (vma = first; vma && vma->vm_start < end; vma = vma->vm_next) {
516 if (!(flags & MPOL_MF_DISCONTIG_OK)) {
517 if (!vma->vm_next && vma->vm_end < end)
518 return ERR_PTR(-EFAULT);
519 if (prev && prev->vm_end < vma->vm_start)
520 return ERR_PTR(-EFAULT);
522 if (!is_vm_hugetlb_page(vma) &&
523 ((flags & MPOL_MF_STRICT) ||
524 ((flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) &&
525 vma_migratable(vma)))) {
526 unsigned long endvma = vma->vm_end;
528 if (endvma > end)
529 endvma = end;
530 if (vma->vm_start > start)
531 start = vma->vm_start;
532 err = check_pgd_range(vma, start, endvma, nodes,
533 flags, private);
534 if (err) {
535 first = ERR_PTR(err);
536 break;
539 prev = vma;
541 return first;
544 /* Apply policy to a single VMA */
545 static int policy_vma(struct vm_area_struct *vma, struct mempolicy *new)
547 int err = 0;
548 struct mempolicy *old = vma->vm_policy;
550 pr_debug("vma %lx-%lx/%lx vm_ops %p vm_file %p set_policy %p\n",
551 vma->vm_start, vma->vm_end, vma->vm_pgoff,
552 vma->vm_ops, vma->vm_file,
553 vma->vm_ops ? vma->vm_ops->set_policy : NULL);
555 if (vma->vm_ops && vma->vm_ops->set_policy)
556 err = vma->vm_ops->set_policy(vma, new);
557 if (!err) {
558 mpol_get(new);
559 vma->vm_policy = new;
560 mpol_put(old);
562 return err;
565 /* Step 2: apply policy to a range and do splits. */
566 static int mbind_range(struct vm_area_struct *vma, unsigned long start,
567 unsigned long end, struct mempolicy *new)
569 struct vm_area_struct *next;
570 int err;
572 err = 0;
573 for (; vma && vma->vm_start < end; vma = next) {
574 next = vma->vm_next;
575 if (vma->vm_start < start)
576 err = split_vma(vma->vm_mm, vma, start, 1);
577 if (!err && vma->vm_end > end)
578 err = split_vma(vma->vm_mm, vma, end, 0);
579 if (!err)
580 err = policy_vma(vma, new);
581 if (err)
582 break;
584 return err;
588 * Update task->flags PF_MEMPOLICY bit: set iff non-default
589 * mempolicy. Allows more rapid checking of this (combined perhaps
590 * with other PF_* flag bits) on memory allocation hot code paths.
592 * If called from outside this file, the task 'p' should -only- be
593 * a newly forked child not yet visible on the task list, because
594 * manipulating the task flags of a visible task is not safe.
596 * The above limitation is why this routine has the funny name
597 * mpol_fix_fork_child_flag().
599 * It is also safe to call this with a task pointer of current,
600 * which the static wrapper mpol_set_task_struct_flag() does,
601 * for use within this file.
604 void mpol_fix_fork_child_flag(struct task_struct *p)
606 if (p->mempolicy)
607 p->flags |= PF_MEMPOLICY;
608 else
609 p->flags &= ~PF_MEMPOLICY;
612 static void mpol_set_task_struct_flag(void)
614 mpol_fix_fork_child_flag(current);
617 /* Set the process memory policy */
618 static long do_set_mempolicy(unsigned short mode, unsigned short flags,
619 nodemask_t *nodes)
621 struct mempolicy *new, *old;
622 struct mm_struct *mm = current->mm;
623 int ret;
625 new = mpol_new(mode, flags, nodes);
626 if (IS_ERR(new))
627 return PTR_ERR(new);
630 * prevent changing our mempolicy while show_numa_maps()
631 * is using it.
632 * Note: do_set_mempolicy() can be called at init time
633 * with no 'mm'.
635 if (mm)
636 down_write(&mm->mmap_sem);
637 task_lock(current);
638 ret = mpol_set_nodemask(new, nodes);
639 if (ret) {
640 task_unlock(current);
641 if (mm)
642 up_write(&mm->mmap_sem);
643 mpol_put(new);
644 return ret;
646 old = current->mempolicy;
647 current->mempolicy = new;
648 mpol_set_task_struct_flag();
649 if (new && new->mode == MPOL_INTERLEAVE &&
650 nodes_weight(new->v.nodes))
651 current->il_next = first_node(new->v.nodes);
652 task_unlock(current);
653 if (mm)
654 up_write(&mm->mmap_sem);
656 mpol_put(old);
657 return 0;
661 * Return nodemask for policy for get_mempolicy() query
663 * Called with task's alloc_lock held
665 static void get_policy_nodemask(struct mempolicy *p, nodemask_t *nodes)
667 nodes_clear(*nodes);
668 if (p == &default_policy)
669 return;
671 switch (p->mode) {
672 case MPOL_BIND:
673 /* Fall through */
674 case MPOL_INTERLEAVE:
675 *nodes = p->v.nodes;
676 break;
677 case MPOL_PREFERRED:
678 if (!(p->flags & MPOL_F_LOCAL))
679 node_set(p->v.preferred_node, *nodes);
680 /* else return empty node mask for local allocation */
681 break;
682 default:
683 BUG();
687 static int lookup_node(struct mm_struct *mm, unsigned long addr)
689 struct page *p;
690 int err;
692 err = get_user_pages(current, mm, addr & PAGE_MASK, 1, 0, 0, &p, NULL);
693 if (err >= 0) {
694 err = page_to_nid(p);
695 put_page(p);
697 return err;
700 /* Retrieve NUMA policy */
701 static long do_get_mempolicy(int *policy, nodemask_t *nmask,
702 unsigned long addr, unsigned long flags)
704 int err;
705 struct mm_struct *mm = current->mm;
706 struct vm_area_struct *vma = NULL;
707 struct mempolicy *pol = current->mempolicy;
709 if (flags &
710 ~(unsigned long)(MPOL_F_NODE|MPOL_F_ADDR|MPOL_F_MEMS_ALLOWED))
711 return -EINVAL;
713 if (flags & MPOL_F_MEMS_ALLOWED) {
714 if (flags & (MPOL_F_NODE|MPOL_F_ADDR))
715 return -EINVAL;
716 *policy = 0; /* just so it's initialized */
717 task_lock(current);
718 *nmask = cpuset_current_mems_allowed;
719 task_unlock(current);
720 return 0;
723 if (flags & MPOL_F_ADDR) {
725 * Do NOT fall back to task policy if the
726 * vma/shared policy at addr is NULL. We
727 * want to return MPOL_DEFAULT in this case.
729 down_read(&mm->mmap_sem);
730 vma = find_vma_intersection(mm, addr, addr+1);
731 if (!vma) {
732 up_read(&mm->mmap_sem);
733 return -EFAULT;
735 if (vma->vm_ops && vma->vm_ops->get_policy)
736 pol = vma->vm_ops->get_policy(vma, addr);
737 else
738 pol = vma->vm_policy;
739 } else if (addr)
740 return -EINVAL;
742 if (!pol)
743 pol = &default_policy; /* indicates default behavior */
745 if (flags & MPOL_F_NODE) {
746 if (flags & MPOL_F_ADDR) {
747 err = lookup_node(mm, addr);
748 if (err < 0)
749 goto out;
750 *policy = err;
751 } else if (pol == current->mempolicy &&
752 pol->mode == MPOL_INTERLEAVE) {
753 *policy = current->il_next;
754 } else {
755 err = -EINVAL;
756 goto out;
758 } else {
759 *policy = pol == &default_policy ? MPOL_DEFAULT :
760 pol->mode;
762 * Internal mempolicy flags must be masked off before exposing
763 * the policy to userspace.
765 *policy |= (pol->flags & MPOL_MODE_FLAGS);
768 if (vma) {
769 up_read(&current->mm->mmap_sem);
770 vma = NULL;
773 err = 0;
774 if (nmask) {
775 task_lock(current);
776 get_policy_nodemask(pol, nmask);
777 task_unlock(current);
780 out:
781 mpol_cond_put(pol);
782 if (vma)
783 up_read(&current->mm->mmap_sem);
784 return err;
787 #ifdef CONFIG_MIGRATION
789 * page migration
791 static void migrate_page_add(struct page *page, struct list_head *pagelist,
792 unsigned long flags)
795 * Avoid migrating a page that is shared with others.
797 if ((flags & MPOL_MF_MOVE_ALL) || page_mapcount(page) == 1) {
798 if (!isolate_lru_page(page)) {
799 list_add_tail(&page->lru, pagelist);
804 static struct page *new_node_page(struct page *page, unsigned long node, int **x)
806 return alloc_pages_exact_node(node, GFP_HIGHUSER_MOVABLE, 0);
810 * Migrate pages from one node to a target node.
811 * Returns error or the number of pages not migrated.
813 static int migrate_to_node(struct mm_struct *mm, int source, int dest,
814 int flags)
816 nodemask_t nmask;
817 LIST_HEAD(pagelist);
818 int err = 0;
820 nodes_clear(nmask);
821 node_set(source, nmask);
823 check_range(mm, mm->mmap->vm_start, TASK_SIZE, &nmask,
824 flags | MPOL_MF_DISCONTIG_OK, &pagelist);
826 if (!list_empty(&pagelist))
827 err = migrate_pages(&pagelist, new_node_page, dest);
829 return err;
833 * Move pages between the two nodesets so as to preserve the physical
834 * layout as much as possible.
836 * Returns the number of page that could not be moved.
838 int do_migrate_pages(struct mm_struct *mm,
839 const nodemask_t *from_nodes, const nodemask_t *to_nodes, int flags)
841 int busy = 0;
842 int err;
843 nodemask_t tmp;
845 err = migrate_prep();
846 if (err)
847 return err;
849 down_read(&mm->mmap_sem);
851 err = migrate_vmas(mm, from_nodes, to_nodes, flags);
852 if (err)
853 goto out;
856 * Find a 'source' bit set in 'tmp' whose corresponding 'dest'
857 * bit in 'to' is not also set in 'tmp'. Clear the found 'source'
858 * bit in 'tmp', and return that <source, dest> pair for migration.
859 * The pair of nodemasks 'to' and 'from' define the map.
861 * If no pair of bits is found that way, fallback to picking some
862 * pair of 'source' and 'dest' bits that are not the same. If the
863 * 'source' and 'dest' bits are the same, this represents a node
864 * that will be migrating to itself, so no pages need move.
866 * If no bits are left in 'tmp', or if all remaining bits left
867 * in 'tmp' correspond to the same bit in 'to', return false
868 * (nothing left to migrate).
870 * This lets us pick a pair of nodes to migrate between, such that
871 * if possible the dest node is not already occupied by some other
872 * source node, minimizing the risk of overloading the memory on a
873 * node that would happen if we migrated incoming memory to a node
874 * before migrating outgoing memory source that same node.
876 * A single scan of tmp is sufficient. As we go, we remember the
877 * most recent <s, d> pair that moved (s != d). If we find a pair
878 * that not only moved, but what's better, moved to an empty slot
879 * (d is not set in tmp), then we break out then, with that pair.
880 * Otherwise when we finish scannng from_tmp, we at least have the
881 * most recent <s, d> pair that moved. If we get all the way through
882 * the scan of tmp without finding any node that moved, much less
883 * moved to an empty node, then there is nothing left worth migrating.
886 tmp = *from_nodes;
887 while (!nodes_empty(tmp)) {
888 int s,d;
889 int source = -1;
890 int dest = 0;
892 for_each_node_mask(s, tmp) {
893 d = node_remap(s, *from_nodes, *to_nodes);
894 if (s == d)
895 continue;
897 source = s; /* Node moved. Memorize */
898 dest = d;
900 /* dest not in remaining from nodes? */
901 if (!node_isset(dest, tmp))
902 break;
904 if (source == -1)
905 break;
907 node_clear(source, tmp);
908 err = migrate_to_node(mm, source, dest, flags);
909 if (err > 0)
910 busy += err;
911 if (err < 0)
912 break;
914 out:
915 up_read(&mm->mmap_sem);
916 if (err < 0)
917 return err;
918 return busy;
923 * Allocate a new page for page migration based on vma policy.
924 * Start assuming that page is mapped by vma pointed to by @private.
925 * Search forward from there, if not. N.B., this assumes that the
926 * list of pages handed to migrate_pages()--which is how we get here--
927 * is in virtual address order.
929 static struct page *new_vma_page(struct page *page, unsigned long private, int **x)
931 struct vm_area_struct *vma = (struct vm_area_struct *)private;
932 unsigned long uninitialized_var(address);
934 while (vma) {
935 address = page_address_in_vma(page, vma);
936 if (address != -EFAULT)
937 break;
938 vma = vma->vm_next;
942 * if !vma, alloc_page_vma() will use task or system default policy
944 return alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
946 #else
948 static void migrate_page_add(struct page *page, struct list_head *pagelist,
949 unsigned long flags)
953 int do_migrate_pages(struct mm_struct *mm,
954 const nodemask_t *from_nodes, const nodemask_t *to_nodes, int flags)
956 return -ENOSYS;
959 static struct page *new_vma_page(struct page *page, unsigned long private, int **x)
961 return NULL;
963 #endif
965 static long do_mbind(unsigned long start, unsigned long len,
966 unsigned short mode, unsigned short mode_flags,
967 nodemask_t *nmask, unsigned long flags)
969 struct vm_area_struct *vma;
970 struct mm_struct *mm = current->mm;
971 struct mempolicy *new;
972 unsigned long end;
973 int err;
974 LIST_HEAD(pagelist);
976 if (flags & ~(unsigned long)(MPOL_MF_STRICT |
977 MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))
978 return -EINVAL;
979 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
980 return -EPERM;
982 if (start & ~PAGE_MASK)
983 return -EINVAL;
985 if (mode == MPOL_DEFAULT)
986 flags &= ~MPOL_MF_STRICT;
988 len = (len + PAGE_SIZE - 1) & PAGE_MASK;
989 end = start + len;
991 if (end < start)
992 return -EINVAL;
993 if (end == start)
994 return 0;
996 new = mpol_new(mode, mode_flags, nmask);
997 if (IS_ERR(new))
998 return PTR_ERR(new);
1001 * If we are using the default policy then operation
1002 * on discontinuous address spaces is okay after all
1004 if (!new)
1005 flags |= MPOL_MF_DISCONTIG_OK;
1007 pr_debug("mbind %lx-%lx mode:%d flags:%d nodes:%lx\n",
1008 start, start + len, mode, mode_flags,
1009 nmask ? nodes_addr(*nmask)[0] : -1);
1011 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) {
1013 err = migrate_prep();
1014 if (err)
1015 return err;
1017 down_write(&mm->mmap_sem);
1018 task_lock(current);
1019 err = mpol_set_nodemask(new, nmask);
1020 task_unlock(current);
1021 if (err) {
1022 up_write(&mm->mmap_sem);
1023 mpol_put(new);
1024 return err;
1026 vma = check_range(mm, start, end, nmask,
1027 flags | MPOL_MF_INVERT, &pagelist);
1029 err = PTR_ERR(vma);
1030 if (!IS_ERR(vma)) {
1031 int nr_failed = 0;
1033 err = mbind_range(vma, start, end, new);
1035 if (!list_empty(&pagelist))
1036 nr_failed = migrate_pages(&pagelist, new_vma_page,
1037 (unsigned long)vma);
1039 if (!err && nr_failed && (flags & MPOL_MF_STRICT))
1040 err = -EIO;
1043 up_write(&mm->mmap_sem);
1044 mpol_put(new);
1045 return err;
1049 * User space interface with variable sized bitmaps for nodelists.
1052 /* Copy a node mask from user space. */
1053 static int get_nodes(nodemask_t *nodes, const unsigned long __user *nmask,
1054 unsigned long maxnode)
1056 unsigned long k;
1057 unsigned long nlongs;
1058 unsigned long endmask;
1060 --maxnode;
1061 nodes_clear(*nodes);
1062 if (maxnode == 0 || !nmask)
1063 return 0;
1064 if (maxnode > PAGE_SIZE*BITS_PER_BYTE)
1065 return -EINVAL;
1067 nlongs = BITS_TO_LONGS(maxnode);
1068 if ((maxnode % BITS_PER_LONG) == 0)
1069 endmask = ~0UL;
1070 else
1071 endmask = (1UL << (maxnode % BITS_PER_LONG)) - 1;
1073 /* When the user specified more nodes than supported just check
1074 if the non supported part is all zero. */
1075 if (nlongs > BITS_TO_LONGS(MAX_NUMNODES)) {
1076 if (nlongs > PAGE_SIZE/sizeof(long))
1077 return -EINVAL;
1078 for (k = BITS_TO_LONGS(MAX_NUMNODES); k < nlongs; k++) {
1079 unsigned long t;
1080 if (get_user(t, nmask + k))
1081 return -EFAULT;
1082 if (k == nlongs - 1) {
1083 if (t & endmask)
1084 return -EINVAL;
1085 } else if (t)
1086 return -EINVAL;
1088 nlongs = BITS_TO_LONGS(MAX_NUMNODES);
1089 endmask = ~0UL;
1092 if (copy_from_user(nodes_addr(*nodes), nmask, nlongs*sizeof(unsigned long)))
1093 return -EFAULT;
1094 nodes_addr(*nodes)[nlongs-1] &= endmask;
1095 return 0;
1098 /* Copy a kernel node mask to user space */
1099 static int copy_nodes_to_user(unsigned long __user *mask, unsigned long maxnode,
1100 nodemask_t *nodes)
1102 unsigned long copy = ALIGN(maxnode-1, 64) / 8;
1103 const int nbytes = BITS_TO_LONGS(MAX_NUMNODES) * sizeof(long);
1105 if (copy > nbytes) {
1106 if (copy > PAGE_SIZE)
1107 return -EINVAL;
1108 if (clear_user((char __user *)mask + nbytes, copy - nbytes))
1109 return -EFAULT;
1110 copy = nbytes;
1112 return copy_to_user(mask, nodes_addr(*nodes), copy) ? -EFAULT : 0;
1115 SYSCALL_DEFINE6(mbind, unsigned long, start, unsigned long, len,
1116 unsigned long, mode, unsigned long __user *, nmask,
1117 unsigned long, maxnode, unsigned, flags)
1119 nodemask_t nodes;
1120 int err;
1121 unsigned short mode_flags;
1123 mode_flags = mode & MPOL_MODE_FLAGS;
1124 mode &= ~MPOL_MODE_FLAGS;
1125 if (mode >= MPOL_MAX)
1126 return -EINVAL;
1127 if ((mode_flags & MPOL_F_STATIC_NODES) &&
1128 (mode_flags & MPOL_F_RELATIVE_NODES))
1129 return -EINVAL;
1130 err = get_nodes(&nodes, nmask, maxnode);
1131 if (err)
1132 return err;
1133 return do_mbind(start, len, mode, mode_flags, &nodes, flags);
1136 /* Set the process memory policy */
1137 SYSCALL_DEFINE3(set_mempolicy, int, mode, unsigned long __user *, nmask,
1138 unsigned long, maxnode)
1140 int err;
1141 nodemask_t nodes;
1142 unsigned short flags;
1144 flags = mode & MPOL_MODE_FLAGS;
1145 mode &= ~MPOL_MODE_FLAGS;
1146 if ((unsigned int)mode >= MPOL_MAX)
1147 return -EINVAL;
1148 if ((flags & MPOL_F_STATIC_NODES) && (flags & MPOL_F_RELATIVE_NODES))
1149 return -EINVAL;
1150 err = get_nodes(&nodes, nmask, maxnode);
1151 if (err)
1152 return err;
1153 return do_set_mempolicy(mode, flags, &nodes);
1156 SYSCALL_DEFINE4(migrate_pages, pid_t, pid, unsigned long, maxnode,
1157 const unsigned long __user *, old_nodes,
1158 const unsigned long __user *, new_nodes)
1160 const struct cred *cred = current_cred(), *tcred;
1161 struct mm_struct *mm;
1162 struct task_struct *task;
1163 nodemask_t old;
1164 nodemask_t new;
1165 nodemask_t task_nodes;
1166 int err;
1168 err = get_nodes(&old, old_nodes, maxnode);
1169 if (err)
1170 return err;
1172 err = get_nodes(&new, new_nodes, maxnode);
1173 if (err)
1174 return err;
1176 /* Find the mm_struct */
1177 read_lock(&tasklist_lock);
1178 task = pid ? find_task_by_vpid(pid) : current;
1179 if (!task) {
1180 read_unlock(&tasklist_lock);
1181 return -ESRCH;
1183 mm = get_task_mm(task);
1184 read_unlock(&tasklist_lock);
1186 if (!mm)
1187 return -EINVAL;
1190 * Check if this process has the right to modify the specified
1191 * process. The right exists if the process has administrative
1192 * capabilities, superuser privileges or the same
1193 * userid as the target process.
1195 rcu_read_lock();
1196 tcred = __task_cred(task);
1197 if (cred->euid != tcred->suid && cred->euid != tcred->uid &&
1198 cred->uid != tcred->suid && cred->uid != tcred->uid &&
1199 !capable(CAP_SYS_NICE)) {
1200 rcu_read_unlock();
1201 err = -EPERM;
1202 goto out;
1204 rcu_read_unlock();
1206 task_nodes = cpuset_mems_allowed(task);
1207 /* Is the user allowed to access the target nodes? */
1208 if (!nodes_subset(new, task_nodes) && !capable(CAP_SYS_NICE)) {
1209 err = -EPERM;
1210 goto out;
1213 if (!nodes_subset(new, node_states[N_HIGH_MEMORY])) {
1214 err = -EINVAL;
1215 goto out;
1218 err = security_task_movememory(task);
1219 if (err)
1220 goto out;
1222 err = do_migrate_pages(mm, &old, &new,
1223 capable(CAP_SYS_NICE) ? MPOL_MF_MOVE_ALL : MPOL_MF_MOVE);
1224 out:
1225 mmput(mm);
1226 return err;
1230 /* Retrieve NUMA policy */
1231 SYSCALL_DEFINE5(get_mempolicy, int __user *, policy,
1232 unsigned long __user *, nmask, unsigned long, maxnode,
1233 unsigned long, addr, unsigned long, flags)
1235 int err;
1236 int uninitialized_var(pval);
1237 nodemask_t nodes;
1239 if (nmask != NULL && maxnode < MAX_NUMNODES)
1240 return -EINVAL;
1242 err = do_get_mempolicy(&pval, &nodes, addr, flags);
1244 if (err)
1245 return err;
1247 if (policy && put_user(pval, policy))
1248 return -EFAULT;
1250 if (nmask)
1251 err = copy_nodes_to_user(nmask, maxnode, &nodes);
1253 return err;
1256 #ifdef CONFIG_COMPAT
1258 asmlinkage long compat_sys_get_mempolicy(int __user *policy,
1259 compat_ulong_t __user *nmask,
1260 compat_ulong_t maxnode,
1261 compat_ulong_t addr, compat_ulong_t flags)
1263 long err;
1264 unsigned long __user *nm = NULL;
1265 unsigned long nr_bits, alloc_size;
1266 DECLARE_BITMAP(bm, MAX_NUMNODES);
1268 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1269 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1271 if (nmask)
1272 nm = compat_alloc_user_space(alloc_size);
1274 err = sys_get_mempolicy(policy, nm, nr_bits+1, addr, flags);
1276 if (!err && nmask) {
1277 err = copy_from_user(bm, nm, alloc_size);
1278 /* ensure entire bitmap is zeroed */
1279 err |= clear_user(nmask, ALIGN(maxnode-1, 8) / 8);
1280 err |= compat_put_bitmap(nmask, bm, nr_bits);
1283 return err;
1286 asmlinkage long compat_sys_set_mempolicy(int mode, compat_ulong_t __user *nmask,
1287 compat_ulong_t maxnode)
1289 long err = 0;
1290 unsigned long __user *nm = NULL;
1291 unsigned long nr_bits, alloc_size;
1292 DECLARE_BITMAP(bm, MAX_NUMNODES);
1294 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1295 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1297 if (nmask) {
1298 err = compat_get_bitmap(bm, nmask, nr_bits);
1299 nm = compat_alloc_user_space(alloc_size);
1300 err |= copy_to_user(nm, bm, alloc_size);
1303 if (err)
1304 return -EFAULT;
1306 return sys_set_mempolicy(mode, nm, nr_bits+1);
1309 asmlinkage long compat_sys_mbind(compat_ulong_t start, compat_ulong_t len,
1310 compat_ulong_t mode, compat_ulong_t __user *nmask,
1311 compat_ulong_t maxnode, compat_ulong_t flags)
1313 long err = 0;
1314 unsigned long __user *nm = NULL;
1315 unsigned long nr_bits, alloc_size;
1316 nodemask_t bm;
1318 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1319 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1321 if (nmask) {
1322 err = compat_get_bitmap(nodes_addr(bm), nmask, nr_bits);
1323 nm = compat_alloc_user_space(alloc_size);
1324 err |= copy_to_user(nm, nodes_addr(bm), alloc_size);
1327 if (err)
1328 return -EFAULT;
1330 return sys_mbind(start, len, mode, nm, nr_bits+1, flags);
1333 #endif
1336 * get_vma_policy(@task, @vma, @addr)
1337 * @task - task for fallback if vma policy == default
1338 * @vma - virtual memory area whose policy is sought
1339 * @addr - address in @vma for shared policy lookup
1341 * Returns effective policy for a VMA at specified address.
1342 * Falls back to @task or system default policy, as necessary.
1343 * Current or other task's task mempolicy and non-shared vma policies
1344 * are protected by the task's mmap_sem, which must be held for read by
1345 * the caller.
1346 * Shared policies [those marked as MPOL_F_SHARED] require an extra reference
1347 * count--added by the get_policy() vm_op, as appropriate--to protect against
1348 * freeing by another task. It is the caller's responsibility to free the
1349 * extra reference for shared policies.
1351 static struct mempolicy *get_vma_policy(struct task_struct *task,
1352 struct vm_area_struct *vma, unsigned long addr)
1354 struct mempolicy *pol = task->mempolicy;
1356 if (vma) {
1357 if (vma->vm_ops && vma->vm_ops->get_policy) {
1358 struct mempolicy *vpol = vma->vm_ops->get_policy(vma,
1359 addr);
1360 if (vpol)
1361 pol = vpol;
1362 } else if (vma->vm_policy)
1363 pol = vma->vm_policy;
1365 if (!pol)
1366 pol = &default_policy;
1367 return pol;
1371 * Return a nodemask representing a mempolicy for filtering nodes for
1372 * page allocation
1374 static nodemask_t *policy_nodemask(gfp_t gfp, struct mempolicy *policy)
1376 /* Lower zones don't get a nodemask applied for MPOL_BIND */
1377 if (unlikely(policy->mode == MPOL_BIND) &&
1378 gfp_zone(gfp) >= policy_zone &&
1379 cpuset_nodemask_valid_mems_allowed(&policy->v.nodes))
1380 return &policy->v.nodes;
1382 return NULL;
1385 /* Return a zonelist indicated by gfp for node representing a mempolicy */
1386 static struct zonelist *policy_zonelist(gfp_t gfp, struct mempolicy *policy)
1388 int nd = numa_node_id();
1390 switch (policy->mode) {
1391 case MPOL_PREFERRED:
1392 if (!(policy->flags & MPOL_F_LOCAL))
1393 nd = policy->v.preferred_node;
1394 break;
1395 case MPOL_BIND:
1397 * Normally, MPOL_BIND allocations are node-local within the
1398 * allowed nodemask. However, if __GFP_THISNODE is set and the
1399 * current node is part of the mask, we use the zonelist for
1400 * the first node in the mask instead.
1402 if (unlikely(gfp & __GFP_THISNODE) &&
1403 unlikely(!node_isset(nd, policy->v.nodes)))
1404 nd = first_node(policy->v.nodes);
1405 break;
1406 case MPOL_INTERLEAVE: /* should not happen */
1407 break;
1408 default:
1409 BUG();
1411 return node_zonelist(nd, gfp);
1414 /* Do dynamic interleaving for a process */
1415 static unsigned interleave_nodes(struct mempolicy *policy)
1417 unsigned nid, next;
1418 struct task_struct *me = current;
1420 nid = me->il_next;
1421 next = next_node(nid, policy->v.nodes);
1422 if (next >= MAX_NUMNODES)
1423 next = first_node(policy->v.nodes);
1424 if (next < MAX_NUMNODES)
1425 me->il_next = next;
1426 return nid;
1430 * Depending on the memory policy provide a node from which to allocate the
1431 * next slab entry.
1432 * @policy must be protected by freeing by the caller. If @policy is
1433 * the current task's mempolicy, this protection is implicit, as only the
1434 * task can change it's policy. The system default policy requires no
1435 * such protection.
1437 unsigned slab_node(struct mempolicy *policy)
1439 if (!policy || policy->flags & MPOL_F_LOCAL)
1440 return numa_node_id();
1442 switch (policy->mode) {
1443 case MPOL_PREFERRED:
1445 * handled MPOL_F_LOCAL above
1447 return policy->v.preferred_node;
1449 case MPOL_INTERLEAVE:
1450 return interleave_nodes(policy);
1452 case MPOL_BIND: {
1454 * Follow bind policy behavior and start allocation at the
1455 * first node.
1457 struct zonelist *zonelist;
1458 struct zone *zone;
1459 enum zone_type highest_zoneidx = gfp_zone(GFP_KERNEL);
1460 zonelist = &NODE_DATA(numa_node_id())->node_zonelists[0];
1461 (void)first_zones_zonelist(zonelist, highest_zoneidx,
1462 &policy->v.nodes,
1463 &zone);
1464 return zone->node;
1467 default:
1468 BUG();
1472 /* Do static interleaving for a VMA with known offset. */
1473 static unsigned offset_il_node(struct mempolicy *pol,
1474 struct vm_area_struct *vma, unsigned long off)
1476 unsigned nnodes = nodes_weight(pol->v.nodes);
1477 unsigned target;
1478 int c;
1479 int nid = -1;
1481 if (!nnodes)
1482 return numa_node_id();
1483 target = (unsigned int)off % nnodes;
1484 c = 0;
1485 do {
1486 nid = next_node(nid, pol->v.nodes);
1487 c++;
1488 } while (c <= target);
1489 return nid;
1492 /* Determine a node number for interleave */
1493 static inline unsigned interleave_nid(struct mempolicy *pol,
1494 struct vm_area_struct *vma, unsigned long addr, int shift)
1496 if (vma) {
1497 unsigned long off;
1500 * for small pages, there is no difference between
1501 * shift and PAGE_SHIFT, so the bit-shift is safe.
1502 * for huge pages, since vm_pgoff is in units of small
1503 * pages, we need to shift off the always 0 bits to get
1504 * a useful offset.
1506 BUG_ON(shift < PAGE_SHIFT);
1507 off = vma->vm_pgoff >> (shift - PAGE_SHIFT);
1508 off += (addr - vma->vm_start) >> shift;
1509 return offset_il_node(pol, vma, off);
1510 } else
1511 return interleave_nodes(pol);
1514 #ifdef CONFIG_HUGETLBFS
1516 * huge_zonelist(@vma, @addr, @gfp_flags, @mpol)
1517 * @vma = virtual memory area whose policy is sought
1518 * @addr = address in @vma for shared policy lookup and interleave policy
1519 * @gfp_flags = for requested zone
1520 * @mpol = pointer to mempolicy pointer for reference counted mempolicy
1521 * @nodemask = pointer to nodemask pointer for MPOL_BIND nodemask
1523 * Returns a zonelist suitable for a huge page allocation and a pointer
1524 * to the struct mempolicy for conditional unref after allocation.
1525 * If the effective policy is 'BIND, returns a pointer to the mempolicy's
1526 * @nodemask for filtering the zonelist.
1528 struct zonelist *huge_zonelist(struct vm_area_struct *vma, unsigned long addr,
1529 gfp_t gfp_flags, struct mempolicy **mpol,
1530 nodemask_t **nodemask)
1532 struct zonelist *zl;
1534 *mpol = get_vma_policy(current, vma, addr);
1535 *nodemask = NULL; /* assume !MPOL_BIND */
1537 if (unlikely((*mpol)->mode == MPOL_INTERLEAVE)) {
1538 zl = node_zonelist(interleave_nid(*mpol, vma, addr,
1539 huge_page_shift(hstate_vma(vma))), gfp_flags);
1540 } else {
1541 zl = policy_zonelist(gfp_flags, *mpol);
1542 if ((*mpol)->mode == MPOL_BIND)
1543 *nodemask = &(*mpol)->v.nodes;
1545 return zl;
1547 #endif
1549 /* Allocate a page in interleaved policy.
1550 Own path because it needs to do special accounting. */
1551 static struct page *alloc_page_interleave(gfp_t gfp, unsigned order,
1552 unsigned nid)
1554 struct zonelist *zl;
1555 struct page *page;
1557 zl = node_zonelist(nid, gfp);
1558 page = __alloc_pages(gfp, order, zl);
1559 if (page && page_zone(page) == zonelist_zone(&zl->_zonerefs[0]))
1560 inc_zone_page_state(page, NUMA_INTERLEAVE_HIT);
1561 return page;
1565 * alloc_page_vma - Allocate a page for a VMA.
1567 * @gfp:
1568 * %GFP_USER user allocation.
1569 * %GFP_KERNEL kernel allocations,
1570 * %GFP_HIGHMEM highmem/user allocations,
1571 * %GFP_FS allocation should not call back into a file system.
1572 * %GFP_ATOMIC don't sleep.
1574 * @vma: Pointer to VMA or NULL if not available.
1575 * @addr: Virtual Address of the allocation. Must be inside the VMA.
1577 * This function allocates a page from the kernel page pool and applies
1578 * a NUMA policy associated with the VMA or the current process.
1579 * When VMA is not NULL caller must hold down_read on the mmap_sem of the
1580 * mm_struct of the VMA to prevent it from going away. Should be used for
1581 * all allocations for pages that will be mapped into
1582 * user space. Returns NULL when no page can be allocated.
1584 * Should be called with the mm_sem of the vma hold.
1586 struct page *
1587 alloc_page_vma(gfp_t gfp, struct vm_area_struct *vma, unsigned long addr)
1589 struct mempolicy *pol = get_vma_policy(current, vma, addr);
1590 struct zonelist *zl;
1592 if (unlikely(pol->mode == MPOL_INTERLEAVE)) {
1593 unsigned nid;
1595 nid = interleave_nid(pol, vma, addr, PAGE_SHIFT);
1596 mpol_cond_put(pol);
1597 return alloc_page_interleave(gfp, 0, nid);
1599 zl = policy_zonelist(gfp, pol);
1600 if (unlikely(mpol_needs_cond_ref(pol))) {
1602 * slow path: ref counted shared policy
1604 struct page *page = __alloc_pages_nodemask(gfp, 0,
1605 zl, policy_nodemask(gfp, pol));
1606 __mpol_put(pol);
1607 return page;
1610 * fast path: default or task policy
1612 return __alloc_pages_nodemask(gfp, 0, zl, policy_nodemask(gfp, pol));
1616 * alloc_pages_current - Allocate pages.
1618 * @gfp:
1619 * %GFP_USER user allocation,
1620 * %GFP_KERNEL kernel allocation,
1621 * %GFP_HIGHMEM highmem allocation,
1622 * %GFP_FS don't call back into a file system.
1623 * %GFP_ATOMIC don't sleep.
1624 * @order: Power of two of allocation size in pages. 0 is a single page.
1626 * Allocate a page from the kernel page pool. When not in
1627 * interrupt context and apply the current process NUMA policy.
1628 * Returns NULL when no page can be allocated.
1630 * Don't call cpuset_update_task_memory_state() unless
1631 * 1) it's ok to take cpuset_sem (can WAIT), and
1632 * 2) allocating for current task (not interrupt).
1634 struct page *alloc_pages_current(gfp_t gfp, unsigned order)
1636 struct mempolicy *pol = current->mempolicy;
1638 if (!pol || in_interrupt() || (gfp & __GFP_THISNODE))
1639 pol = &default_policy;
1642 * No reference counting needed for current->mempolicy
1643 * nor system default_policy
1645 if (pol->mode == MPOL_INTERLEAVE)
1646 return alloc_page_interleave(gfp, order, interleave_nodes(pol));
1647 return __alloc_pages_nodemask(gfp, order,
1648 policy_zonelist(gfp, pol), policy_nodemask(gfp, pol));
1650 EXPORT_SYMBOL(alloc_pages_current);
1653 * If mpol_dup() sees current->cpuset == cpuset_being_rebound, then it
1654 * rebinds the mempolicy its copying by calling mpol_rebind_policy()
1655 * with the mems_allowed returned by cpuset_mems_allowed(). This
1656 * keeps mempolicies cpuset relative after its cpuset moves. See
1657 * further kernel/cpuset.c update_nodemask().
1660 /* Slow path of a mempolicy duplicate */
1661 struct mempolicy *__mpol_dup(struct mempolicy *old)
1663 struct mempolicy *new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
1665 if (!new)
1666 return ERR_PTR(-ENOMEM);
1667 if (current_cpuset_is_being_rebound()) {
1668 nodemask_t mems = cpuset_mems_allowed(current);
1669 mpol_rebind_policy(old, &mems);
1671 *new = *old;
1672 atomic_set(&new->refcnt, 1);
1673 return new;
1677 * If *frompol needs [has] an extra ref, copy *frompol to *tompol ,
1678 * eliminate the * MPOL_F_* flags that require conditional ref and
1679 * [NOTE!!!] drop the extra ref. Not safe to reference *frompol directly
1680 * after return. Use the returned value.
1682 * Allows use of a mempolicy for, e.g., multiple allocations with a single
1683 * policy lookup, even if the policy needs/has extra ref on lookup.
1684 * shmem_readahead needs this.
1686 struct mempolicy *__mpol_cond_copy(struct mempolicy *tompol,
1687 struct mempolicy *frompol)
1689 if (!mpol_needs_cond_ref(frompol))
1690 return frompol;
1692 *tompol = *frompol;
1693 tompol->flags &= ~MPOL_F_SHARED; /* copy doesn't need unref */
1694 __mpol_put(frompol);
1695 return tompol;
1698 static int mpol_match_intent(const struct mempolicy *a,
1699 const struct mempolicy *b)
1701 if (a->flags != b->flags)
1702 return 0;
1703 if (!mpol_store_user_nodemask(a))
1704 return 1;
1705 return nodes_equal(a->w.user_nodemask, b->w.user_nodemask);
1708 /* Slow path of a mempolicy comparison */
1709 int __mpol_equal(struct mempolicy *a, struct mempolicy *b)
1711 if (!a || !b)
1712 return 0;
1713 if (a->mode != b->mode)
1714 return 0;
1715 if (a->mode != MPOL_DEFAULT && !mpol_match_intent(a, b))
1716 return 0;
1717 switch (a->mode) {
1718 case MPOL_BIND:
1719 /* Fall through */
1720 case MPOL_INTERLEAVE:
1721 return nodes_equal(a->v.nodes, b->v.nodes);
1722 case MPOL_PREFERRED:
1723 return a->v.preferred_node == b->v.preferred_node &&
1724 a->flags == b->flags;
1725 default:
1726 BUG();
1727 return 0;
1732 * Shared memory backing store policy support.
1734 * Remember policies even when nobody has shared memory mapped.
1735 * The policies are kept in Red-Black tree linked from the inode.
1736 * They are protected by the sp->lock spinlock, which should be held
1737 * for any accesses to the tree.
1740 /* lookup first element intersecting start-end */
1741 /* Caller holds sp->lock */
1742 static struct sp_node *
1743 sp_lookup(struct shared_policy *sp, unsigned long start, unsigned long end)
1745 struct rb_node *n = sp->root.rb_node;
1747 while (n) {
1748 struct sp_node *p = rb_entry(n, struct sp_node, nd);
1750 if (start >= p->end)
1751 n = n->rb_right;
1752 else if (end <= p->start)
1753 n = n->rb_left;
1754 else
1755 break;
1757 if (!n)
1758 return NULL;
1759 for (;;) {
1760 struct sp_node *w = NULL;
1761 struct rb_node *prev = rb_prev(n);
1762 if (!prev)
1763 break;
1764 w = rb_entry(prev, struct sp_node, nd);
1765 if (w->end <= start)
1766 break;
1767 n = prev;
1769 return rb_entry(n, struct sp_node, nd);
1772 /* Insert a new shared policy into the list. */
1773 /* Caller holds sp->lock */
1774 static void sp_insert(struct shared_policy *sp, struct sp_node *new)
1776 struct rb_node **p = &sp->root.rb_node;
1777 struct rb_node *parent = NULL;
1778 struct sp_node *nd;
1780 while (*p) {
1781 parent = *p;
1782 nd = rb_entry(parent, struct sp_node, nd);
1783 if (new->start < nd->start)
1784 p = &(*p)->rb_left;
1785 else if (new->end > nd->end)
1786 p = &(*p)->rb_right;
1787 else
1788 BUG();
1790 rb_link_node(&new->nd, parent, p);
1791 rb_insert_color(&new->nd, &sp->root);
1792 pr_debug("inserting %lx-%lx: %d\n", new->start, new->end,
1793 new->policy ? new->policy->mode : 0);
1796 /* Find shared policy intersecting idx */
1797 struct mempolicy *
1798 mpol_shared_policy_lookup(struct shared_policy *sp, unsigned long idx)
1800 struct mempolicy *pol = NULL;
1801 struct sp_node *sn;
1803 if (!sp->root.rb_node)
1804 return NULL;
1805 spin_lock(&sp->lock);
1806 sn = sp_lookup(sp, idx, idx+1);
1807 if (sn) {
1808 mpol_get(sn->policy);
1809 pol = sn->policy;
1811 spin_unlock(&sp->lock);
1812 return pol;
1815 static void sp_delete(struct shared_policy *sp, struct sp_node *n)
1817 pr_debug("deleting %lx-l%lx\n", n->start, n->end);
1818 rb_erase(&n->nd, &sp->root);
1819 mpol_put(n->policy);
1820 kmem_cache_free(sn_cache, n);
1823 static struct sp_node *sp_alloc(unsigned long start, unsigned long end,
1824 struct mempolicy *pol)
1826 struct sp_node *n = kmem_cache_alloc(sn_cache, GFP_KERNEL);
1828 if (!n)
1829 return NULL;
1830 n->start = start;
1831 n->end = end;
1832 mpol_get(pol);
1833 pol->flags |= MPOL_F_SHARED; /* for unref */
1834 n->policy = pol;
1835 return n;
1838 /* Replace a policy range. */
1839 static int shared_policy_replace(struct shared_policy *sp, unsigned long start,
1840 unsigned long end, struct sp_node *new)
1842 struct sp_node *n, *new2 = NULL;
1844 restart:
1845 spin_lock(&sp->lock);
1846 n = sp_lookup(sp, start, end);
1847 /* Take care of old policies in the same range. */
1848 while (n && n->start < end) {
1849 struct rb_node *next = rb_next(&n->nd);
1850 if (n->start >= start) {
1851 if (n->end <= end)
1852 sp_delete(sp, n);
1853 else
1854 n->start = end;
1855 } else {
1856 /* Old policy spanning whole new range. */
1857 if (n->end > end) {
1858 if (!new2) {
1859 spin_unlock(&sp->lock);
1860 new2 = sp_alloc(end, n->end, n->policy);
1861 if (!new2)
1862 return -ENOMEM;
1863 goto restart;
1865 n->end = start;
1866 sp_insert(sp, new2);
1867 new2 = NULL;
1868 break;
1869 } else
1870 n->end = start;
1872 if (!next)
1873 break;
1874 n = rb_entry(next, struct sp_node, nd);
1876 if (new)
1877 sp_insert(sp, new);
1878 spin_unlock(&sp->lock);
1879 if (new2) {
1880 mpol_put(new2->policy);
1881 kmem_cache_free(sn_cache, new2);
1883 return 0;
1887 * mpol_shared_policy_init - initialize shared policy for inode
1888 * @sp: pointer to inode shared policy
1889 * @mpol: struct mempolicy to install
1891 * Install non-NULL @mpol in inode's shared policy rb-tree.
1892 * On entry, the current task has a reference on a non-NULL @mpol.
1893 * This must be released on exit.
1895 void mpol_shared_policy_init(struct shared_policy *sp, struct mempolicy *mpol)
1897 int ret;
1899 sp->root = RB_ROOT; /* empty tree == default mempolicy */
1900 spin_lock_init(&sp->lock);
1902 if (mpol) {
1903 struct vm_area_struct pvma;
1904 struct mempolicy *new;
1906 /* contextualize the tmpfs mount point mempolicy */
1907 new = mpol_new(mpol->mode, mpol->flags, &mpol->w.user_nodemask);
1908 if (IS_ERR(new)) {
1909 mpol_put(mpol); /* drop our ref on sb mpol */
1910 return; /* no valid nodemask intersection */
1913 task_lock(current);
1914 ret = mpol_set_nodemask(new, &mpol->w.user_nodemask);
1915 task_unlock(current);
1916 mpol_put(mpol); /* drop our ref on sb mpol */
1917 if (ret) {
1918 mpol_put(new);
1919 return;
1922 /* Create pseudo-vma that contains just the policy */
1923 memset(&pvma, 0, sizeof(struct vm_area_struct));
1924 pvma.vm_end = TASK_SIZE; /* policy covers entire file */
1925 mpol_set_shared_policy(sp, &pvma, new); /* adds ref */
1926 mpol_put(new); /* drop initial ref */
1930 int mpol_set_shared_policy(struct shared_policy *info,
1931 struct vm_area_struct *vma, struct mempolicy *npol)
1933 int err;
1934 struct sp_node *new = NULL;
1935 unsigned long sz = vma_pages(vma);
1937 pr_debug("set_shared_policy %lx sz %lu %d %d %lx\n",
1938 vma->vm_pgoff,
1939 sz, npol ? npol->mode : -1,
1940 npol ? npol->flags : -1,
1941 npol ? nodes_addr(npol->v.nodes)[0] : -1);
1943 if (npol) {
1944 new = sp_alloc(vma->vm_pgoff, vma->vm_pgoff + sz, npol);
1945 if (!new)
1946 return -ENOMEM;
1948 err = shared_policy_replace(info, vma->vm_pgoff, vma->vm_pgoff+sz, new);
1949 if (err && new)
1950 kmem_cache_free(sn_cache, new);
1951 return err;
1954 /* Free a backing policy store on inode delete. */
1955 void mpol_free_shared_policy(struct shared_policy *p)
1957 struct sp_node *n;
1958 struct rb_node *next;
1960 if (!p->root.rb_node)
1961 return;
1962 spin_lock(&p->lock);
1963 next = rb_first(&p->root);
1964 while (next) {
1965 n = rb_entry(next, struct sp_node, nd);
1966 next = rb_next(&n->nd);
1967 rb_erase(&n->nd, &p->root);
1968 mpol_put(n->policy);
1969 kmem_cache_free(sn_cache, n);
1971 spin_unlock(&p->lock);
1974 /* assumes fs == KERNEL_DS */
1975 void __init numa_policy_init(void)
1977 nodemask_t interleave_nodes;
1978 unsigned long largest = 0;
1979 int nid, prefer = 0;
1981 policy_cache = kmem_cache_create("numa_policy",
1982 sizeof(struct mempolicy),
1983 0, SLAB_PANIC, NULL);
1985 sn_cache = kmem_cache_create("shared_policy_node",
1986 sizeof(struct sp_node),
1987 0, SLAB_PANIC, NULL);
1990 * Set interleaving policy for system init. Interleaving is only
1991 * enabled across suitably sized nodes (default is >= 16MB), or
1992 * fall back to the largest node if they're all smaller.
1994 nodes_clear(interleave_nodes);
1995 for_each_node_state(nid, N_HIGH_MEMORY) {
1996 unsigned long total_pages = node_present_pages(nid);
1998 /* Preserve the largest node */
1999 if (largest < total_pages) {
2000 largest = total_pages;
2001 prefer = nid;
2004 /* Interleave this node? */
2005 if ((total_pages << PAGE_SHIFT) >= (16 << 20))
2006 node_set(nid, interleave_nodes);
2009 /* All too small, use the largest */
2010 if (unlikely(nodes_empty(interleave_nodes)))
2011 node_set(prefer, interleave_nodes);
2013 if (do_set_mempolicy(MPOL_INTERLEAVE, 0, &interleave_nodes))
2014 printk("numa_policy_init: interleaving failed\n");
2017 /* Reset policy of current process to default */
2018 void numa_default_policy(void)
2020 do_set_mempolicy(MPOL_DEFAULT, 0, NULL);
2024 * Parse and format mempolicy from/to strings
2028 * "local" is pseudo-policy: MPOL_PREFERRED with MPOL_F_LOCAL flag
2029 * Used only for mpol_parse_str() and mpol_to_str()
2031 #define MPOL_LOCAL (MPOL_INTERLEAVE + 1)
2032 static const char * const policy_types[] =
2033 { "default", "prefer", "bind", "interleave", "local" };
2036 #ifdef CONFIG_TMPFS
2038 * mpol_parse_str - parse string to mempolicy
2039 * @str: string containing mempolicy to parse
2040 * @mpol: pointer to struct mempolicy pointer, returned on success.
2041 * @no_context: flag whether to "contextualize" the mempolicy
2043 * Format of input:
2044 * <mode>[=<flags>][:<nodelist>]
2046 * if @no_context is true, save the input nodemask in w.user_nodemask in
2047 * the returned mempolicy. This will be used to "clone" the mempolicy in
2048 * a specific context [cpuset] at a later time. Used to parse tmpfs mpol
2049 * mount option. Note that if 'static' or 'relative' mode flags were
2050 * specified, the input nodemask will already have been saved. Saving
2051 * it again is redundant, but safe.
2053 * On success, returns 0, else 1
2055 int mpol_parse_str(char *str, struct mempolicy **mpol, int no_context)
2057 struct mempolicy *new = NULL;
2058 unsigned short uninitialized_var(mode);
2059 unsigned short uninitialized_var(mode_flags);
2060 nodemask_t nodes;
2061 char *nodelist = strchr(str, ':');
2062 char *flags = strchr(str, '=');
2063 int i;
2064 int err = 1;
2066 if (nodelist) {
2067 /* NUL-terminate mode or flags string */
2068 *nodelist++ = '\0';
2069 if (nodelist_parse(nodelist, nodes))
2070 goto out;
2071 if (!nodes_subset(nodes, node_states[N_HIGH_MEMORY]))
2072 goto out;
2073 } else
2074 nodes_clear(nodes);
2076 if (flags)
2077 *flags++ = '\0'; /* terminate mode string */
2079 for (i = 0; i <= MPOL_LOCAL; i++) {
2080 if (!strcmp(str, policy_types[i])) {
2081 mode = i;
2082 break;
2085 if (i > MPOL_LOCAL)
2086 goto out;
2088 switch (mode) {
2089 case MPOL_PREFERRED:
2091 * Insist on a nodelist of one node only
2093 if (nodelist) {
2094 char *rest = nodelist;
2095 while (isdigit(*rest))
2096 rest++;
2097 if (!*rest)
2098 err = 0;
2100 break;
2101 case MPOL_INTERLEAVE:
2103 * Default to online nodes with memory if no nodelist
2105 if (!nodelist)
2106 nodes = node_states[N_HIGH_MEMORY];
2107 err = 0;
2108 break;
2109 case MPOL_LOCAL:
2111 * Don't allow a nodelist; mpol_new() checks flags
2113 if (nodelist)
2114 goto out;
2115 mode = MPOL_PREFERRED;
2116 break;
2119 * case MPOL_BIND: mpol_new() enforces non-empty nodemask.
2120 * case MPOL_DEFAULT: mpol_new() enforces empty nodemask, ignores flags.
2124 mode_flags = 0;
2125 if (flags) {
2127 * Currently, we only support two mutually exclusive
2128 * mode flags.
2130 if (!strcmp(flags, "static"))
2131 mode_flags |= MPOL_F_STATIC_NODES;
2132 else if (!strcmp(flags, "relative"))
2133 mode_flags |= MPOL_F_RELATIVE_NODES;
2134 else
2135 err = 1;
2138 new = mpol_new(mode, mode_flags, &nodes);
2139 if (IS_ERR(new))
2140 err = 1;
2141 else {
2142 int ret;
2144 task_lock(current);
2145 ret = mpol_set_nodemask(new, &nodes);
2146 task_unlock(current);
2147 if (ret)
2148 err = 1;
2149 else if (no_context) {
2150 /* save for contextualization */
2151 new->w.user_nodemask = nodes;
2155 out:
2156 /* Restore string for error message */
2157 if (nodelist)
2158 *--nodelist = ':';
2159 if (flags)
2160 *--flags = '=';
2161 if (!err)
2162 *mpol = new;
2163 return err;
2165 #endif /* CONFIG_TMPFS */
2168 * mpol_to_str - format a mempolicy structure for printing
2169 * @buffer: to contain formatted mempolicy string
2170 * @maxlen: length of @buffer
2171 * @pol: pointer to mempolicy to be formatted
2172 * @no_context: "context free" mempolicy - use nodemask in w.user_nodemask
2174 * Convert a mempolicy into a string.
2175 * Returns the number of characters in buffer (if positive)
2176 * or an error (negative)
2178 int mpol_to_str(char *buffer, int maxlen, struct mempolicy *pol, int no_context)
2180 char *p = buffer;
2181 int l;
2182 nodemask_t nodes;
2183 unsigned short mode;
2184 unsigned short flags = pol ? pol->flags : 0;
2187 * Sanity check: room for longest mode, flag and some nodes
2189 VM_BUG_ON(maxlen < strlen("interleave") + strlen("relative") + 16);
2191 if (!pol || pol == &default_policy)
2192 mode = MPOL_DEFAULT;
2193 else
2194 mode = pol->mode;
2196 switch (mode) {
2197 case MPOL_DEFAULT:
2198 nodes_clear(nodes);
2199 break;
2201 case MPOL_PREFERRED:
2202 nodes_clear(nodes);
2203 if (flags & MPOL_F_LOCAL)
2204 mode = MPOL_LOCAL; /* pseudo-policy */
2205 else
2206 node_set(pol->v.preferred_node, nodes);
2207 break;
2209 case MPOL_BIND:
2210 /* Fall through */
2211 case MPOL_INTERLEAVE:
2212 if (no_context)
2213 nodes = pol->w.user_nodemask;
2214 else
2215 nodes = pol->v.nodes;
2216 break;
2218 default:
2219 BUG();
2222 l = strlen(policy_types[mode]);
2223 if (buffer + maxlen < p + l + 1)
2224 return -ENOSPC;
2226 strcpy(p, policy_types[mode]);
2227 p += l;
2229 if (flags & MPOL_MODE_FLAGS) {
2230 if (buffer + maxlen < p + 2)
2231 return -ENOSPC;
2232 *p++ = '=';
2235 * Currently, the only defined flags are mutually exclusive
2237 if (flags & MPOL_F_STATIC_NODES)
2238 p += snprintf(p, buffer + maxlen - p, "static");
2239 else if (flags & MPOL_F_RELATIVE_NODES)
2240 p += snprintf(p, buffer + maxlen - p, "relative");
2243 if (!nodes_empty(nodes)) {
2244 if (buffer + maxlen < p + 2)
2245 return -ENOSPC;
2246 *p++ = ':';
2247 p += nodelist_scnprintf(p, buffer + maxlen - p, nodes);
2249 return p - buffer;
2252 struct numa_maps {
2253 unsigned long pages;
2254 unsigned long anon;
2255 unsigned long active;
2256 unsigned long writeback;
2257 unsigned long mapcount_max;
2258 unsigned long dirty;
2259 unsigned long swapcache;
2260 unsigned long node[MAX_NUMNODES];
2263 static void gather_stats(struct page *page, void *private, int pte_dirty)
2265 struct numa_maps *md = private;
2266 int count = page_mapcount(page);
2268 md->pages++;
2269 if (pte_dirty || PageDirty(page))
2270 md->dirty++;
2272 if (PageSwapCache(page))
2273 md->swapcache++;
2275 if (PageActive(page) || PageUnevictable(page))
2276 md->active++;
2278 if (PageWriteback(page))
2279 md->writeback++;
2281 if (PageAnon(page))
2282 md->anon++;
2284 if (count > md->mapcount_max)
2285 md->mapcount_max = count;
2287 md->node[page_to_nid(page)]++;
2290 #ifdef CONFIG_HUGETLB_PAGE
2291 static void check_huge_range(struct vm_area_struct *vma,
2292 unsigned long start, unsigned long end,
2293 struct numa_maps *md)
2295 unsigned long addr;
2296 struct page *page;
2297 struct hstate *h = hstate_vma(vma);
2298 unsigned long sz = huge_page_size(h);
2300 for (addr = start; addr < end; addr += sz) {
2301 pte_t *ptep = huge_pte_offset(vma->vm_mm,
2302 addr & huge_page_mask(h));
2303 pte_t pte;
2305 if (!ptep)
2306 continue;
2308 pte = *ptep;
2309 if (pte_none(pte))
2310 continue;
2312 page = pte_page(pte);
2313 if (!page)
2314 continue;
2316 gather_stats(page, md, pte_dirty(*ptep));
2319 #else
2320 static inline void check_huge_range(struct vm_area_struct *vma,
2321 unsigned long start, unsigned long end,
2322 struct numa_maps *md)
2325 #endif
2328 * Display pages allocated per node and memory policy via /proc.
2330 int show_numa_map(struct seq_file *m, void *v)
2332 struct proc_maps_private *priv = m->private;
2333 struct vm_area_struct *vma = v;
2334 struct numa_maps *md;
2335 struct file *file = vma->vm_file;
2336 struct mm_struct *mm = vma->vm_mm;
2337 struct mempolicy *pol;
2338 int n;
2339 char buffer[50];
2341 if (!mm)
2342 return 0;
2344 md = kzalloc(sizeof(struct numa_maps), GFP_KERNEL);
2345 if (!md)
2346 return 0;
2348 pol = get_vma_policy(priv->task, vma, vma->vm_start);
2349 mpol_to_str(buffer, sizeof(buffer), pol, 0);
2350 mpol_cond_put(pol);
2352 seq_printf(m, "%08lx %s", vma->vm_start, buffer);
2354 if (file) {
2355 seq_printf(m, " file=");
2356 seq_path(m, &file->f_path, "\n\t= ");
2357 } else if (vma->vm_start <= mm->brk && vma->vm_end >= mm->start_brk) {
2358 seq_printf(m, " heap");
2359 } else if (vma->vm_start <= mm->start_stack &&
2360 vma->vm_end >= mm->start_stack) {
2361 seq_printf(m, " stack");
2364 if (is_vm_hugetlb_page(vma)) {
2365 check_huge_range(vma, vma->vm_start, vma->vm_end, md);
2366 seq_printf(m, " huge");
2367 } else {
2368 check_pgd_range(vma, vma->vm_start, vma->vm_end,
2369 &node_states[N_HIGH_MEMORY], MPOL_MF_STATS, md);
2372 if (!md->pages)
2373 goto out;
2375 if (md->anon)
2376 seq_printf(m," anon=%lu",md->anon);
2378 if (md->dirty)
2379 seq_printf(m," dirty=%lu",md->dirty);
2381 if (md->pages != md->anon && md->pages != md->dirty)
2382 seq_printf(m, " mapped=%lu", md->pages);
2384 if (md->mapcount_max > 1)
2385 seq_printf(m, " mapmax=%lu", md->mapcount_max);
2387 if (md->swapcache)
2388 seq_printf(m," swapcache=%lu", md->swapcache);
2390 if (md->active < md->pages && !is_vm_hugetlb_page(vma))
2391 seq_printf(m," active=%lu", md->active);
2393 if (md->writeback)
2394 seq_printf(m," writeback=%lu", md->writeback);
2396 for_each_node_state(n, N_HIGH_MEMORY)
2397 if (md->node[n])
2398 seq_printf(m, " N%d=%lu", n, md->node[n]);
2399 out:
2400 seq_putc(m, '\n');
2401 kfree(md);
2403 if (m->count < m->size)
2404 m->version = (vma != priv->tail_vma) ? vma->vm_start : 0;
2405 return 0;