mmc: SDHI should depend on SUPERH || ARCH_SHMOBILE
[cris-mirror.git] / mm / mempolicy.c
blob959a8b8c7350c28ffa6cacced61b4c24977479da
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/slab.h>
77 #include <linux/string.h>
78 #include <linux/module.h>
79 #include <linux/nsproxy.h>
80 #include <linux/interrupt.h>
81 #include <linux/init.h>
82 #include <linux/compat.h>
83 #include <linux/swap.h>
84 #include <linux/seq_file.h>
85 #include <linux/proc_fs.h>
86 #include <linux/migrate.h>
87 #include <linux/ksm.h>
88 #include <linux/rmap.h>
89 #include <linux/security.h>
90 #include <linux/syscalls.h>
91 #include <linux/ctype.h>
92 #include <linux/mm_inline.h>
94 #include <asm/tlbflush.h>
95 #include <asm/uaccess.h>
97 #include "internal.h"
99 /* Internal flags */
100 #define MPOL_MF_DISCONTIG_OK (MPOL_MF_INTERNAL << 0) /* Skip checks for continuous vmas */
101 #define MPOL_MF_INVERT (MPOL_MF_INTERNAL << 1) /* Invert check for nodemask */
102 #define MPOL_MF_STATS (MPOL_MF_INTERNAL << 2) /* Gather statistics */
104 static struct kmem_cache *policy_cache;
105 static struct kmem_cache *sn_cache;
107 /* Highest zone. An specific allocation for a zone below that is not
108 policied. */
109 enum zone_type policy_zone = 0;
112 * run-time system-wide default policy => local allocation
114 struct mempolicy default_policy = {
115 .refcnt = ATOMIC_INIT(1), /* never free it */
116 .mode = MPOL_PREFERRED,
117 .flags = MPOL_F_LOCAL,
120 static const struct mempolicy_operations {
121 int (*create)(struct mempolicy *pol, const nodemask_t *nodes);
123 * If read-side task has no lock to protect task->mempolicy, write-side
124 * task will rebind the task->mempolicy by two step. The first step is
125 * setting all the newly nodes, and the second step is cleaning all the
126 * disallowed nodes. In this way, we can avoid finding no node to alloc
127 * page.
128 * If we have a lock to protect task->mempolicy in read-side, we do
129 * rebind directly.
131 * step:
132 * MPOL_REBIND_ONCE - do rebind work at once
133 * MPOL_REBIND_STEP1 - set all the newly nodes
134 * MPOL_REBIND_STEP2 - clean all the disallowed nodes
136 void (*rebind)(struct mempolicy *pol, const nodemask_t *nodes,
137 enum mpol_rebind_step step);
138 } mpol_ops[MPOL_MAX];
140 /* Check that the nodemask contains at least one populated zone */
141 static int is_valid_nodemask(const nodemask_t *nodemask)
143 int nd, k;
145 for_each_node_mask(nd, *nodemask) {
146 struct zone *z;
148 for (k = 0; k <= policy_zone; k++) {
149 z = &NODE_DATA(nd)->node_zones[k];
150 if (z->present_pages > 0)
151 return 1;
155 return 0;
158 static inline int mpol_store_user_nodemask(const struct mempolicy *pol)
160 return pol->flags & MPOL_MODE_FLAGS;
163 static void mpol_relative_nodemask(nodemask_t *ret, const nodemask_t *orig,
164 const nodemask_t *rel)
166 nodemask_t tmp;
167 nodes_fold(tmp, *orig, nodes_weight(*rel));
168 nodes_onto(*ret, tmp, *rel);
171 static int mpol_new_interleave(struct mempolicy *pol, const nodemask_t *nodes)
173 if (nodes_empty(*nodes))
174 return -EINVAL;
175 pol->v.nodes = *nodes;
176 return 0;
179 static int mpol_new_preferred(struct mempolicy *pol, const nodemask_t *nodes)
181 if (!nodes)
182 pol->flags |= MPOL_F_LOCAL; /* local allocation */
183 else if (nodes_empty(*nodes))
184 return -EINVAL; /* no allowed nodes */
185 else
186 pol->v.preferred_node = first_node(*nodes);
187 return 0;
190 static int mpol_new_bind(struct mempolicy *pol, const nodemask_t *nodes)
192 if (!is_valid_nodemask(nodes))
193 return -EINVAL;
194 pol->v.nodes = *nodes;
195 return 0;
199 * mpol_set_nodemask is called after mpol_new() to set up the nodemask, if
200 * any, for the new policy. mpol_new() has already validated the nodes
201 * parameter with respect to the policy mode and flags. But, we need to
202 * handle an empty nodemask with MPOL_PREFERRED here.
204 * Must be called holding task's alloc_lock to protect task's mems_allowed
205 * and mempolicy. May also be called holding the mmap_semaphore for write.
207 static int mpol_set_nodemask(struct mempolicy *pol,
208 const nodemask_t *nodes, struct nodemask_scratch *nsc)
210 int ret;
212 /* if mode is MPOL_DEFAULT, pol is NULL. This is right. */
213 if (pol == NULL)
214 return 0;
215 /* Check N_HIGH_MEMORY */
216 nodes_and(nsc->mask1,
217 cpuset_current_mems_allowed, node_states[N_HIGH_MEMORY]);
219 VM_BUG_ON(!nodes);
220 if (pol->mode == MPOL_PREFERRED && nodes_empty(*nodes))
221 nodes = NULL; /* explicit local allocation */
222 else {
223 if (pol->flags & MPOL_F_RELATIVE_NODES)
224 mpol_relative_nodemask(&nsc->mask2, nodes,&nsc->mask1);
225 else
226 nodes_and(nsc->mask2, *nodes, nsc->mask1);
228 if (mpol_store_user_nodemask(pol))
229 pol->w.user_nodemask = *nodes;
230 else
231 pol->w.cpuset_mems_allowed =
232 cpuset_current_mems_allowed;
235 if (nodes)
236 ret = mpol_ops[pol->mode].create(pol, &nsc->mask2);
237 else
238 ret = mpol_ops[pol->mode].create(pol, NULL);
239 return ret;
243 * This function just creates a new policy, does some check and simple
244 * initialization. You must invoke mpol_set_nodemask() to set nodes.
246 static struct mempolicy *mpol_new(unsigned short mode, unsigned short flags,
247 nodemask_t *nodes)
249 struct mempolicy *policy;
251 pr_debug("setting mode %d flags %d nodes[0] %lx\n",
252 mode, flags, nodes ? nodes_addr(*nodes)[0] : -1);
254 if (mode == MPOL_DEFAULT) {
255 if (nodes && !nodes_empty(*nodes))
256 return ERR_PTR(-EINVAL);
257 return NULL; /* simply delete any existing policy */
259 VM_BUG_ON(!nodes);
262 * MPOL_PREFERRED cannot be used with MPOL_F_STATIC_NODES or
263 * MPOL_F_RELATIVE_NODES if the nodemask is empty (local allocation).
264 * All other modes require a valid pointer to a non-empty nodemask.
266 if (mode == MPOL_PREFERRED) {
267 if (nodes_empty(*nodes)) {
268 if (((flags & MPOL_F_STATIC_NODES) ||
269 (flags & MPOL_F_RELATIVE_NODES)))
270 return ERR_PTR(-EINVAL);
272 } else if (nodes_empty(*nodes))
273 return ERR_PTR(-EINVAL);
274 policy = kmem_cache_alloc(policy_cache, GFP_KERNEL);
275 if (!policy)
276 return ERR_PTR(-ENOMEM);
277 atomic_set(&policy->refcnt, 1);
278 policy->mode = mode;
279 policy->flags = flags;
281 return policy;
284 /* Slow path of a mpol destructor. */
285 void __mpol_put(struct mempolicy *p)
287 if (!atomic_dec_and_test(&p->refcnt))
288 return;
289 kmem_cache_free(policy_cache, p);
292 static void mpol_rebind_default(struct mempolicy *pol, const nodemask_t *nodes,
293 enum mpol_rebind_step step)
298 * step:
299 * MPOL_REBIND_ONCE - do rebind work at once
300 * MPOL_REBIND_STEP1 - set all the newly nodes
301 * MPOL_REBIND_STEP2 - clean all the disallowed nodes
303 static void mpol_rebind_nodemask(struct mempolicy *pol, const nodemask_t *nodes,
304 enum mpol_rebind_step step)
306 nodemask_t tmp;
308 if (pol->flags & MPOL_F_STATIC_NODES)
309 nodes_and(tmp, pol->w.user_nodemask, *nodes);
310 else if (pol->flags & MPOL_F_RELATIVE_NODES)
311 mpol_relative_nodemask(&tmp, &pol->w.user_nodemask, nodes);
312 else {
314 * if step == 1, we use ->w.cpuset_mems_allowed to cache the
315 * result
317 if (step == MPOL_REBIND_ONCE || step == MPOL_REBIND_STEP1) {
318 nodes_remap(tmp, pol->v.nodes,
319 pol->w.cpuset_mems_allowed, *nodes);
320 pol->w.cpuset_mems_allowed = step ? tmp : *nodes;
321 } else if (step == MPOL_REBIND_STEP2) {
322 tmp = pol->w.cpuset_mems_allowed;
323 pol->w.cpuset_mems_allowed = *nodes;
324 } else
325 BUG();
328 if (nodes_empty(tmp))
329 tmp = *nodes;
331 if (step == MPOL_REBIND_STEP1)
332 nodes_or(pol->v.nodes, pol->v.nodes, tmp);
333 else if (step == MPOL_REBIND_ONCE || step == MPOL_REBIND_STEP2)
334 pol->v.nodes = tmp;
335 else
336 BUG();
338 if (!node_isset(current->il_next, tmp)) {
339 current->il_next = next_node(current->il_next, tmp);
340 if (current->il_next >= MAX_NUMNODES)
341 current->il_next = first_node(tmp);
342 if (current->il_next >= MAX_NUMNODES)
343 current->il_next = numa_node_id();
347 static void mpol_rebind_preferred(struct mempolicy *pol,
348 const nodemask_t *nodes,
349 enum mpol_rebind_step step)
351 nodemask_t tmp;
353 if (pol->flags & MPOL_F_STATIC_NODES) {
354 int node = first_node(pol->w.user_nodemask);
356 if (node_isset(node, *nodes)) {
357 pol->v.preferred_node = node;
358 pol->flags &= ~MPOL_F_LOCAL;
359 } else
360 pol->flags |= MPOL_F_LOCAL;
361 } else if (pol->flags & MPOL_F_RELATIVE_NODES) {
362 mpol_relative_nodemask(&tmp, &pol->w.user_nodemask, nodes);
363 pol->v.preferred_node = first_node(tmp);
364 } else if (!(pol->flags & MPOL_F_LOCAL)) {
365 pol->v.preferred_node = node_remap(pol->v.preferred_node,
366 pol->w.cpuset_mems_allowed,
367 *nodes);
368 pol->w.cpuset_mems_allowed = *nodes;
373 * mpol_rebind_policy - Migrate a policy to a different set of nodes
375 * If read-side task has no lock to protect task->mempolicy, write-side
376 * task will rebind the task->mempolicy by two step. The first step is
377 * setting all the newly nodes, and the second step is cleaning all the
378 * disallowed nodes. In this way, we can avoid finding no node to alloc
379 * page.
380 * If we have a lock to protect task->mempolicy in read-side, we do
381 * rebind directly.
383 * step:
384 * MPOL_REBIND_ONCE - do rebind work at once
385 * MPOL_REBIND_STEP1 - set all the newly nodes
386 * MPOL_REBIND_STEP2 - clean all the disallowed nodes
388 static void mpol_rebind_policy(struct mempolicy *pol, const nodemask_t *newmask,
389 enum mpol_rebind_step step)
391 if (!pol)
392 return;
393 if (!mpol_store_user_nodemask(pol) && step == 0 &&
394 nodes_equal(pol->w.cpuset_mems_allowed, *newmask))
395 return;
397 if (step == MPOL_REBIND_STEP1 && (pol->flags & MPOL_F_REBINDING))
398 return;
400 if (step == MPOL_REBIND_STEP2 && !(pol->flags & MPOL_F_REBINDING))
401 BUG();
403 if (step == MPOL_REBIND_STEP1)
404 pol->flags |= MPOL_F_REBINDING;
405 else if (step == MPOL_REBIND_STEP2)
406 pol->flags &= ~MPOL_F_REBINDING;
407 else if (step >= MPOL_REBIND_NSTEP)
408 BUG();
410 mpol_ops[pol->mode].rebind(pol, newmask, step);
414 * Wrapper for mpol_rebind_policy() that just requires task
415 * pointer, and updates task mempolicy.
417 * Called with task's alloc_lock held.
420 void mpol_rebind_task(struct task_struct *tsk, const nodemask_t *new,
421 enum mpol_rebind_step step)
423 mpol_rebind_policy(tsk->mempolicy, new, step);
427 * Rebind each vma in mm to new nodemask.
429 * Call holding a reference to mm. Takes mm->mmap_sem during call.
432 void mpol_rebind_mm(struct mm_struct *mm, nodemask_t *new)
434 struct vm_area_struct *vma;
436 down_write(&mm->mmap_sem);
437 for (vma = mm->mmap; vma; vma = vma->vm_next)
438 mpol_rebind_policy(vma->vm_policy, new, MPOL_REBIND_ONCE);
439 up_write(&mm->mmap_sem);
442 static const struct mempolicy_operations mpol_ops[MPOL_MAX] = {
443 [MPOL_DEFAULT] = {
444 .rebind = mpol_rebind_default,
446 [MPOL_INTERLEAVE] = {
447 .create = mpol_new_interleave,
448 .rebind = mpol_rebind_nodemask,
450 [MPOL_PREFERRED] = {
451 .create = mpol_new_preferred,
452 .rebind = mpol_rebind_preferred,
454 [MPOL_BIND] = {
455 .create = mpol_new_bind,
456 .rebind = mpol_rebind_nodemask,
460 static void gather_stats(struct page *, void *, int pte_dirty);
461 static void migrate_page_add(struct page *page, struct list_head *pagelist,
462 unsigned long flags);
464 /* Scan through pages checking if pages follow certain conditions. */
465 static int check_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
466 unsigned long addr, unsigned long end,
467 const nodemask_t *nodes, unsigned long flags,
468 void *private)
470 pte_t *orig_pte;
471 pte_t *pte;
472 spinlock_t *ptl;
474 orig_pte = pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
475 do {
476 struct page *page;
477 int nid;
479 if (!pte_present(*pte))
480 continue;
481 page = vm_normal_page(vma, addr, *pte);
482 if (!page)
483 continue;
485 * vm_normal_page() filters out zero pages, but there might
486 * still be PageReserved pages to skip, perhaps in a VDSO.
487 * And we cannot move PageKsm pages sensibly or safely yet.
489 if (PageReserved(page) || PageKsm(page))
490 continue;
491 nid = page_to_nid(page);
492 if (node_isset(nid, *nodes) == !!(flags & MPOL_MF_INVERT))
493 continue;
495 if (flags & MPOL_MF_STATS)
496 gather_stats(page, private, pte_dirty(*pte));
497 else if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))
498 migrate_page_add(page, private, flags);
499 else
500 break;
501 } while (pte++, addr += PAGE_SIZE, addr != end);
502 pte_unmap_unlock(orig_pte, ptl);
503 return addr != end;
506 static inline int check_pmd_range(struct vm_area_struct *vma, pud_t *pud,
507 unsigned long addr, unsigned long end,
508 const nodemask_t *nodes, unsigned long flags,
509 void *private)
511 pmd_t *pmd;
512 unsigned long next;
514 pmd = pmd_offset(pud, addr);
515 do {
516 next = pmd_addr_end(addr, end);
517 split_huge_page_pmd(vma->vm_mm, pmd);
518 if (pmd_none_or_clear_bad(pmd))
519 continue;
520 if (check_pte_range(vma, pmd, addr, next, nodes,
521 flags, private))
522 return -EIO;
523 } while (pmd++, addr = next, addr != end);
524 return 0;
527 static inline int check_pud_range(struct vm_area_struct *vma, pgd_t *pgd,
528 unsigned long addr, unsigned long end,
529 const nodemask_t *nodes, unsigned long flags,
530 void *private)
532 pud_t *pud;
533 unsigned long next;
535 pud = pud_offset(pgd, addr);
536 do {
537 next = pud_addr_end(addr, end);
538 if (pud_none_or_clear_bad(pud))
539 continue;
540 if (check_pmd_range(vma, pud, addr, next, nodes,
541 flags, private))
542 return -EIO;
543 } while (pud++, addr = next, addr != end);
544 return 0;
547 static inline int check_pgd_range(struct vm_area_struct *vma,
548 unsigned long addr, unsigned long end,
549 const nodemask_t *nodes, unsigned long flags,
550 void *private)
552 pgd_t *pgd;
553 unsigned long next;
555 pgd = pgd_offset(vma->vm_mm, addr);
556 do {
557 next = pgd_addr_end(addr, end);
558 if (pgd_none_or_clear_bad(pgd))
559 continue;
560 if (check_pud_range(vma, pgd, addr, next, nodes,
561 flags, private))
562 return -EIO;
563 } while (pgd++, addr = next, addr != end);
564 return 0;
568 * Check if all pages in a range are on a set of nodes.
569 * If pagelist != NULL then isolate pages from the LRU and
570 * put them on the pagelist.
572 static struct vm_area_struct *
573 check_range(struct mm_struct *mm, unsigned long start, unsigned long end,
574 const nodemask_t *nodes, unsigned long flags, void *private)
576 int err;
577 struct vm_area_struct *first, *vma, *prev;
580 first = find_vma(mm, start);
581 if (!first)
582 return ERR_PTR(-EFAULT);
583 prev = NULL;
584 for (vma = first; vma && vma->vm_start < end; vma = vma->vm_next) {
585 if (!(flags & MPOL_MF_DISCONTIG_OK)) {
586 if (!vma->vm_next && vma->vm_end < end)
587 return ERR_PTR(-EFAULT);
588 if (prev && prev->vm_end < vma->vm_start)
589 return ERR_PTR(-EFAULT);
591 if (!is_vm_hugetlb_page(vma) &&
592 ((flags & MPOL_MF_STRICT) ||
593 ((flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) &&
594 vma_migratable(vma)))) {
595 unsigned long endvma = vma->vm_end;
597 if (endvma > end)
598 endvma = end;
599 if (vma->vm_start > start)
600 start = vma->vm_start;
601 err = check_pgd_range(vma, start, endvma, nodes,
602 flags, private);
603 if (err) {
604 first = ERR_PTR(err);
605 break;
608 prev = vma;
610 return first;
613 /* Apply policy to a single VMA */
614 static int policy_vma(struct vm_area_struct *vma, struct mempolicy *new)
616 int err = 0;
617 struct mempolicy *old = vma->vm_policy;
619 pr_debug("vma %lx-%lx/%lx vm_ops %p vm_file %p set_policy %p\n",
620 vma->vm_start, vma->vm_end, vma->vm_pgoff,
621 vma->vm_ops, vma->vm_file,
622 vma->vm_ops ? vma->vm_ops->set_policy : NULL);
624 if (vma->vm_ops && vma->vm_ops->set_policy)
625 err = vma->vm_ops->set_policy(vma, new);
626 if (!err) {
627 mpol_get(new);
628 vma->vm_policy = new;
629 mpol_put(old);
631 return err;
634 /* Step 2: apply policy to a range and do splits. */
635 static int mbind_range(struct mm_struct *mm, unsigned long start,
636 unsigned long end, struct mempolicy *new_pol)
638 struct vm_area_struct *next;
639 struct vm_area_struct *prev;
640 struct vm_area_struct *vma;
641 int err = 0;
642 pgoff_t pgoff;
643 unsigned long vmstart;
644 unsigned long vmend;
646 vma = find_vma_prev(mm, start, &prev);
647 if (!vma || vma->vm_start > start)
648 return -EFAULT;
650 for (; vma && vma->vm_start < end; prev = vma, vma = next) {
651 next = vma->vm_next;
652 vmstart = max(start, vma->vm_start);
653 vmend = min(end, vma->vm_end);
655 pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT);
656 prev = vma_merge(mm, prev, vmstart, vmend, vma->vm_flags,
657 vma->anon_vma, vma->vm_file, pgoff, new_pol);
658 if (prev) {
659 vma = prev;
660 next = vma->vm_next;
661 continue;
663 if (vma->vm_start != vmstart) {
664 err = split_vma(vma->vm_mm, vma, vmstart, 1);
665 if (err)
666 goto out;
668 if (vma->vm_end != vmend) {
669 err = split_vma(vma->vm_mm, vma, vmend, 0);
670 if (err)
671 goto out;
673 err = policy_vma(vma, new_pol);
674 if (err)
675 goto out;
678 out:
679 return err;
683 * Update task->flags PF_MEMPOLICY bit: set iff non-default
684 * mempolicy. Allows more rapid checking of this (combined perhaps
685 * with other PF_* flag bits) on memory allocation hot code paths.
687 * If called from outside this file, the task 'p' should -only- be
688 * a newly forked child not yet visible on the task list, because
689 * manipulating the task flags of a visible task is not safe.
691 * The above limitation is why this routine has the funny name
692 * mpol_fix_fork_child_flag().
694 * It is also safe to call this with a task pointer of current,
695 * which the static wrapper mpol_set_task_struct_flag() does,
696 * for use within this file.
699 void mpol_fix_fork_child_flag(struct task_struct *p)
701 if (p->mempolicy)
702 p->flags |= PF_MEMPOLICY;
703 else
704 p->flags &= ~PF_MEMPOLICY;
707 static void mpol_set_task_struct_flag(void)
709 mpol_fix_fork_child_flag(current);
712 /* Set the process memory policy */
713 static long do_set_mempolicy(unsigned short mode, unsigned short flags,
714 nodemask_t *nodes)
716 struct mempolicy *new, *old;
717 struct mm_struct *mm = current->mm;
718 NODEMASK_SCRATCH(scratch);
719 int ret;
721 if (!scratch)
722 return -ENOMEM;
724 new = mpol_new(mode, flags, nodes);
725 if (IS_ERR(new)) {
726 ret = PTR_ERR(new);
727 goto out;
730 * prevent changing our mempolicy while show_numa_maps()
731 * is using it.
732 * Note: do_set_mempolicy() can be called at init time
733 * with no 'mm'.
735 if (mm)
736 down_write(&mm->mmap_sem);
737 task_lock(current);
738 ret = mpol_set_nodemask(new, nodes, scratch);
739 if (ret) {
740 task_unlock(current);
741 if (mm)
742 up_write(&mm->mmap_sem);
743 mpol_put(new);
744 goto out;
746 old = current->mempolicy;
747 current->mempolicy = new;
748 mpol_set_task_struct_flag();
749 if (new && new->mode == MPOL_INTERLEAVE &&
750 nodes_weight(new->v.nodes))
751 current->il_next = first_node(new->v.nodes);
752 task_unlock(current);
753 if (mm)
754 up_write(&mm->mmap_sem);
756 mpol_put(old);
757 ret = 0;
758 out:
759 NODEMASK_SCRATCH_FREE(scratch);
760 return ret;
764 * Return nodemask for policy for get_mempolicy() query
766 * Called with task's alloc_lock held
768 static void get_policy_nodemask(struct mempolicy *p, nodemask_t *nodes)
770 nodes_clear(*nodes);
771 if (p == &default_policy)
772 return;
774 switch (p->mode) {
775 case MPOL_BIND:
776 /* Fall through */
777 case MPOL_INTERLEAVE:
778 *nodes = p->v.nodes;
779 break;
780 case MPOL_PREFERRED:
781 if (!(p->flags & MPOL_F_LOCAL))
782 node_set(p->v.preferred_node, *nodes);
783 /* else return empty node mask for local allocation */
784 break;
785 default:
786 BUG();
790 static int lookup_node(struct mm_struct *mm, unsigned long addr)
792 struct page *p;
793 int err;
795 err = get_user_pages(current, mm, addr & PAGE_MASK, 1, 0, 0, &p, NULL);
796 if (err >= 0) {
797 err = page_to_nid(p);
798 put_page(p);
800 return err;
803 /* Retrieve NUMA policy */
804 static long do_get_mempolicy(int *policy, nodemask_t *nmask,
805 unsigned long addr, unsigned long flags)
807 int err;
808 struct mm_struct *mm = current->mm;
809 struct vm_area_struct *vma = NULL;
810 struct mempolicy *pol = current->mempolicy;
812 if (flags &
813 ~(unsigned long)(MPOL_F_NODE|MPOL_F_ADDR|MPOL_F_MEMS_ALLOWED))
814 return -EINVAL;
816 if (flags & MPOL_F_MEMS_ALLOWED) {
817 if (flags & (MPOL_F_NODE|MPOL_F_ADDR))
818 return -EINVAL;
819 *policy = 0; /* just so it's initialized */
820 task_lock(current);
821 *nmask = cpuset_current_mems_allowed;
822 task_unlock(current);
823 return 0;
826 if (flags & MPOL_F_ADDR) {
828 * Do NOT fall back to task policy if the
829 * vma/shared policy at addr is NULL. We
830 * want to return MPOL_DEFAULT in this case.
832 down_read(&mm->mmap_sem);
833 vma = find_vma_intersection(mm, addr, addr+1);
834 if (!vma) {
835 up_read(&mm->mmap_sem);
836 return -EFAULT;
838 if (vma->vm_ops && vma->vm_ops->get_policy)
839 pol = vma->vm_ops->get_policy(vma, addr);
840 else
841 pol = vma->vm_policy;
842 } else if (addr)
843 return -EINVAL;
845 if (!pol)
846 pol = &default_policy; /* indicates default behavior */
848 if (flags & MPOL_F_NODE) {
849 if (flags & MPOL_F_ADDR) {
850 err = lookup_node(mm, addr);
851 if (err < 0)
852 goto out;
853 *policy = err;
854 } else if (pol == current->mempolicy &&
855 pol->mode == MPOL_INTERLEAVE) {
856 *policy = current->il_next;
857 } else {
858 err = -EINVAL;
859 goto out;
861 } else {
862 *policy = pol == &default_policy ? MPOL_DEFAULT :
863 pol->mode;
865 * Internal mempolicy flags must be masked off before exposing
866 * the policy to userspace.
868 *policy |= (pol->flags & MPOL_MODE_FLAGS);
871 if (vma) {
872 up_read(&current->mm->mmap_sem);
873 vma = NULL;
876 err = 0;
877 if (nmask) {
878 if (mpol_store_user_nodemask(pol)) {
879 *nmask = pol->w.user_nodemask;
880 } else {
881 task_lock(current);
882 get_policy_nodemask(pol, nmask);
883 task_unlock(current);
887 out:
888 mpol_cond_put(pol);
889 if (vma)
890 up_read(&current->mm->mmap_sem);
891 return err;
894 #ifdef CONFIG_MIGRATION
896 * page migration
898 static void migrate_page_add(struct page *page, struct list_head *pagelist,
899 unsigned long flags)
902 * Avoid migrating a page that is shared with others.
904 if ((flags & MPOL_MF_MOVE_ALL) || page_mapcount(page) == 1) {
905 if (!isolate_lru_page(page)) {
906 list_add_tail(&page->lru, pagelist);
907 inc_zone_page_state(page, NR_ISOLATED_ANON +
908 page_is_file_cache(page));
913 static struct page *new_node_page(struct page *page, unsigned long node, int **x)
915 return alloc_pages_exact_node(node, GFP_HIGHUSER_MOVABLE, 0);
919 * Migrate pages from one node to a target node.
920 * Returns error or the number of pages not migrated.
922 static int migrate_to_node(struct mm_struct *mm, int source, int dest,
923 int flags)
925 nodemask_t nmask;
926 LIST_HEAD(pagelist);
927 int err = 0;
928 struct vm_area_struct *vma;
930 nodes_clear(nmask);
931 node_set(source, nmask);
933 vma = check_range(mm, mm->mmap->vm_start, mm->task_size, &nmask,
934 flags | MPOL_MF_DISCONTIG_OK, &pagelist);
935 if (IS_ERR(vma))
936 return PTR_ERR(vma);
938 if (!list_empty(&pagelist)) {
939 err = migrate_pages(&pagelist, new_node_page, dest,
940 false, true);
941 if (err)
942 putback_lru_pages(&pagelist);
945 return err;
949 * Move pages between the two nodesets so as to preserve the physical
950 * layout as much as possible.
952 * Returns the number of page that could not be moved.
954 int do_migrate_pages(struct mm_struct *mm,
955 const nodemask_t *from_nodes, const nodemask_t *to_nodes, int flags)
957 int busy = 0;
958 int err;
959 nodemask_t tmp;
961 err = migrate_prep();
962 if (err)
963 return err;
965 down_read(&mm->mmap_sem);
967 err = migrate_vmas(mm, from_nodes, to_nodes, flags);
968 if (err)
969 goto out;
972 * Find a 'source' bit set in 'tmp' whose corresponding 'dest'
973 * bit in 'to' is not also set in 'tmp'. Clear the found 'source'
974 * bit in 'tmp', and return that <source, dest> pair for migration.
975 * The pair of nodemasks 'to' and 'from' define the map.
977 * If no pair of bits is found that way, fallback to picking some
978 * pair of 'source' and 'dest' bits that are not the same. If the
979 * 'source' and 'dest' bits are the same, this represents a node
980 * that will be migrating to itself, so no pages need move.
982 * If no bits are left in 'tmp', or if all remaining bits left
983 * in 'tmp' correspond to the same bit in 'to', return false
984 * (nothing left to migrate).
986 * This lets us pick a pair of nodes to migrate between, such that
987 * if possible the dest node is not already occupied by some other
988 * source node, minimizing the risk of overloading the memory on a
989 * node that would happen if we migrated incoming memory to a node
990 * before migrating outgoing memory source that same node.
992 * A single scan of tmp is sufficient. As we go, we remember the
993 * most recent <s, d> pair that moved (s != d). If we find a pair
994 * that not only moved, but what's better, moved to an empty slot
995 * (d is not set in tmp), then we break out then, with that pair.
996 * Otherwise when we finish scanning from_tmp, we at least have the
997 * most recent <s, d> pair that moved. If we get all the way through
998 * the scan of tmp without finding any node that moved, much less
999 * moved to an empty node, then there is nothing left worth migrating.
1002 tmp = *from_nodes;
1003 while (!nodes_empty(tmp)) {
1004 int s,d;
1005 int source = -1;
1006 int dest = 0;
1008 for_each_node_mask(s, tmp) {
1009 d = node_remap(s, *from_nodes, *to_nodes);
1010 if (s == d)
1011 continue;
1013 source = s; /* Node moved. Memorize */
1014 dest = d;
1016 /* dest not in remaining from nodes? */
1017 if (!node_isset(dest, tmp))
1018 break;
1020 if (source == -1)
1021 break;
1023 node_clear(source, tmp);
1024 err = migrate_to_node(mm, source, dest, flags);
1025 if (err > 0)
1026 busy += err;
1027 if (err < 0)
1028 break;
1030 out:
1031 up_read(&mm->mmap_sem);
1032 if (err < 0)
1033 return err;
1034 return busy;
1039 * Allocate a new page for page migration based on vma policy.
1040 * Start assuming that page is mapped by vma pointed to by @private.
1041 * Search forward from there, if not. N.B., this assumes that the
1042 * list of pages handed to migrate_pages()--which is how we get here--
1043 * is in virtual address order.
1045 static struct page *new_vma_page(struct page *page, unsigned long private, int **x)
1047 struct vm_area_struct *vma = (struct vm_area_struct *)private;
1048 unsigned long uninitialized_var(address);
1050 while (vma) {
1051 address = page_address_in_vma(page, vma);
1052 if (address != -EFAULT)
1053 break;
1054 vma = vma->vm_next;
1058 * if !vma, alloc_page_vma() will use task or system default policy
1060 return alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
1062 #else
1064 static void migrate_page_add(struct page *page, struct list_head *pagelist,
1065 unsigned long flags)
1069 int do_migrate_pages(struct mm_struct *mm,
1070 const nodemask_t *from_nodes, const nodemask_t *to_nodes, int flags)
1072 return -ENOSYS;
1075 static struct page *new_vma_page(struct page *page, unsigned long private, int **x)
1077 return NULL;
1079 #endif
1081 static long do_mbind(unsigned long start, unsigned long len,
1082 unsigned short mode, unsigned short mode_flags,
1083 nodemask_t *nmask, unsigned long flags)
1085 struct vm_area_struct *vma;
1086 struct mm_struct *mm = current->mm;
1087 struct mempolicy *new;
1088 unsigned long end;
1089 int err;
1090 LIST_HEAD(pagelist);
1092 if (flags & ~(unsigned long)(MPOL_MF_STRICT |
1093 MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))
1094 return -EINVAL;
1095 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1096 return -EPERM;
1098 if (start & ~PAGE_MASK)
1099 return -EINVAL;
1101 if (mode == MPOL_DEFAULT)
1102 flags &= ~MPOL_MF_STRICT;
1104 len = (len + PAGE_SIZE - 1) & PAGE_MASK;
1105 end = start + len;
1107 if (end < start)
1108 return -EINVAL;
1109 if (end == start)
1110 return 0;
1112 new = mpol_new(mode, mode_flags, nmask);
1113 if (IS_ERR(new))
1114 return PTR_ERR(new);
1117 * If we are using the default policy then operation
1118 * on discontinuous address spaces is okay after all
1120 if (!new)
1121 flags |= MPOL_MF_DISCONTIG_OK;
1123 pr_debug("mbind %lx-%lx mode:%d flags:%d nodes:%lx\n",
1124 start, start + len, mode, mode_flags,
1125 nmask ? nodes_addr(*nmask)[0] : -1);
1127 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) {
1129 err = migrate_prep();
1130 if (err)
1131 goto mpol_out;
1134 NODEMASK_SCRATCH(scratch);
1135 if (scratch) {
1136 down_write(&mm->mmap_sem);
1137 task_lock(current);
1138 err = mpol_set_nodemask(new, nmask, scratch);
1139 task_unlock(current);
1140 if (err)
1141 up_write(&mm->mmap_sem);
1142 } else
1143 err = -ENOMEM;
1144 NODEMASK_SCRATCH_FREE(scratch);
1146 if (err)
1147 goto mpol_out;
1149 vma = check_range(mm, start, end, nmask,
1150 flags | MPOL_MF_INVERT, &pagelist);
1152 err = PTR_ERR(vma);
1153 if (!IS_ERR(vma)) {
1154 int nr_failed = 0;
1156 err = mbind_range(mm, start, end, new);
1158 if (!list_empty(&pagelist)) {
1159 nr_failed = migrate_pages(&pagelist, new_vma_page,
1160 (unsigned long)vma,
1161 false, true);
1162 if (nr_failed)
1163 putback_lru_pages(&pagelist);
1166 if (!err && nr_failed && (flags & MPOL_MF_STRICT))
1167 err = -EIO;
1168 } else
1169 putback_lru_pages(&pagelist);
1171 up_write(&mm->mmap_sem);
1172 mpol_out:
1173 mpol_put(new);
1174 return err;
1178 * User space interface with variable sized bitmaps for nodelists.
1181 /* Copy a node mask from user space. */
1182 static int get_nodes(nodemask_t *nodes, const unsigned long __user *nmask,
1183 unsigned long maxnode)
1185 unsigned long k;
1186 unsigned long nlongs;
1187 unsigned long endmask;
1189 --maxnode;
1190 nodes_clear(*nodes);
1191 if (maxnode == 0 || !nmask)
1192 return 0;
1193 if (maxnode > PAGE_SIZE*BITS_PER_BYTE)
1194 return -EINVAL;
1196 nlongs = BITS_TO_LONGS(maxnode);
1197 if ((maxnode % BITS_PER_LONG) == 0)
1198 endmask = ~0UL;
1199 else
1200 endmask = (1UL << (maxnode % BITS_PER_LONG)) - 1;
1202 /* When the user specified more nodes than supported just check
1203 if the non supported part is all zero. */
1204 if (nlongs > BITS_TO_LONGS(MAX_NUMNODES)) {
1205 if (nlongs > PAGE_SIZE/sizeof(long))
1206 return -EINVAL;
1207 for (k = BITS_TO_LONGS(MAX_NUMNODES); k < nlongs; k++) {
1208 unsigned long t;
1209 if (get_user(t, nmask + k))
1210 return -EFAULT;
1211 if (k == nlongs - 1) {
1212 if (t & endmask)
1213 return -EINVAL;
1214 } else if (t)
1215 return -EINVAL;
1217 nlongs = BITS_TO_LONGS(MAX_NUMNODES);
1218 endmask = ~0UL;
1221 if (copy_from_user(nodes_addr(*nodes), nmask, nlongs*sizeof(unsigned long)))
1222 return -EFAULT;
1223 nodes_addr(*nodes)[nlongs-1] &= endmask;
1224 return 0;
1227 /* Copy a kernel node mask to user space */
1228 static int copy_nodes_to_user(unsigned long __user *mask, unsigned long maxnode,
1229 nodemask_t *nodes)
1231 unsigned long copy = ALIGN(maxnode-1, 64) / 8;
1232 const int nbytes = BITS_TO_LONGS(MAX_NUMNODES) * sizeof(long);
1234 if (copy > nbytes) {
1235 if (copy > PAGE_SIZE)
1236 return -EINVAL;
1237 if (clear_user((char __user *)mask + nbytes, copy - nbytes))
1238 return -EFAULT;
1239 copy = nbytes;
1241 return copy_to_user(mask, nodes_addr(*nodes), copy) ? -EFAULT : 0;
1244 SYSCALL_DEFINE6(mbind, unsigned long, start, unsigned long, len,
1245 unsigned long, mode, unsigned long __user *, nmask,
1246 unsigned long, maxnode, unsigned, flags)
1248 nodemask_t nodes;
1249 int err;
1250 unsigned short mode_flags;
1252 mode_flags = mode & MPOL_MODE_FLAGS;
1253 mode &= ~MPOL_MODE_FLAGS;
1254 if (mode >= MPOL_MAX)
1255 return -EINVAL;
1256 if ((mode_flags & MPOL_F_STATIC_NODES) &&
1257 (mode_flags & MPOL_F_RELATIVE_NODES))
1258 return -EINVAL;
1259 err = get_nodes(&nodes, nmask, maxnode);
1260 if (err)
1261 return err;
1262 return do_mbind(start, len, mode, mode_flags, &nodes, flags);
1265 /* Set the process memory policy */
1266 SYSCALL_DEFINE3(set_mempolicy, int, mode, unsigned long __user *, nmask,
1267 unsigned long, maxnode)
1269 int err;
1270 nodemask_t nodes;
1271 unsigned short flags;
1273 flags = mode & MPOL_MODE_FLAGS;
1274 mode &= ~MPOL_MODE_FLAGS;
1275 if ((unsigned int)mode >= MPOL_MAX)
1276 return -EINVAL;
1277 if ((flags & MPOL_F_STATIC_NODES) && (flags & MPOL_F_RELATIVE_NODES))
1278 return -EINVAL;
1279 err = get_nodes(&nodes, nmask, maxnode);
1280 if (err)
1281 return err;
1282 return do_set_mempolicy(mode, flags, &nodes);
1285 SYSCALL_DEFINE4(migrate_pages, pid_t, pid, unsigned long, maxnode,
1286 const unsigned long __user *, old_nodes,
1287 const unsigned long __user *, new_nodes)
1289 const struct cred *cred = current_cred(), *tcred;
1290 struct mm_struct *mm = NULL;
1291 struct task_struct *task;
1292 nodemask_t task_nodes;
1293 int err;
1294 nodemask_t *old;
1295 nodemask_t *new;
1296 NODEMASK_SCRATCH(scratch);
1298 if (!scratch)
1299 return -ENOMEM;
1301 old = &scratch->mask1;
1302 new = &scratch->mask2;
1304 err = get_nodes(old, old_nodes, maxnode);
1305 if (err)
1306 goto out;
1308 err = get_nodes(new, new_nodes, maxnode);
1309 if (err)
1310 goto out;
1312 /* Find the mm_struct */
1313 rcu_read_lock();
1314 task = pid ? find_task_by_vpid(pid) : current;
1315 if (!task) {
1316 rcu_read_unlock();
1317 err = -ESRCH;
1318 goto out;
1320 mm = get_task_mm(task);
1321 rcu_read_unlock();
1323 err = -EINVAL;
1324 if (!mm)
1325 goto out;
1328 * Check if this process has the right to modify the specified
1329 * process. The right exists if the process has administrative
1330 * capabilities, superuser privileges or the same
1331 * userid as the target process.
1333 rcu_read_lock();
1334 tcred = __task_cred(task);
1335 if (cred->euid != tcred->suid && cred->euid != tcred->uid &&
1336 cred->uid != tcred->suid && cred->uid != tcred->uid &&
1337 !capable(CAP_SYS_NICE)) {
1338 rcu_read_unlock();
1339 err = -EPERM;
1340 goto out;
1342 rcu_read_unlock();
1344 task_nodes = cpuset_mems_allowed(task);
1345 /* Is the user allowed to access the target nodes? */
1346 if (!nodes_subset(*new, task_nodes) && !capable(CAP_SYS_NICE)) {
1347 err = -EPERM;
1348 goto out;
1351 if (!nodes_subset(*new, node_states[N_HIGH_MEMORY])) {
1352 err = -EINVAL;
1353 goto out;
1356 err = security_task_movememory(task);
1357 if (err)
1358 goto out;
1360 err = do_migrate_pages(mm, old, new,
1361 capable(CAP_SYS_NICE) ? MPOL_MF_MOVE_ALL : MPOL_MF_MOVE);
1362 out:
1363 if (mm)
1364 mmput(mm);
1365 NODEMASK_SCRATCH_FREE(scratch);
1367 return err;
1371 /* Retrieve NUMA policy */
1372 SYSCALL_DEFINE5(get_mempolicy, int __user *, policy,
1373 unsigned long __user *, nmask, unsigned long, maxnode,
1374 unsigned long, addr, unsigned long, flags)
1376 int err;
1377 int uninitialized_var(pval);
1378 nodemask_t nodes;
1380 if (nmask != NULL && maxnode < MAX_NUMNODES)
1381 return -EINVAL;
1383 err = do_get_mempolicy(&pval, &nodes, addr, flags);
1385 if (err)
1386 return err;
1388 if (policy && put_user(pval, policy))
1389 return -EFAULT;
1391 if (nmask)
1392 err = copy_nodes_to_user(nmask, maxnode, &nodes);
1394 return err;
1397 #ifdef CONFIG_COMPAT
1399 asmlinkage long compat_sys_get_mempolicy(int __user *policy,
1400 compat_ulong_t __user *nmask,
1401 compat_ulong_t maxnode,
1402 compat_ulong_t addr, compat_ulong_t flags)
1404 long err;
1405 unsigned long __user *nm = NULL;
1406 unsigned long nr_bits, alloc_size;
1407 DECLARE_BITMAP(bm, MAX_NUMNODES);
1409 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1410 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1412 if (nmask)
1413 nm = compat_alloc_user_space(alloc_size);
1415 err = sys_get_mempolicy(policy, nm, nr_bits+1, addr, flags);
1417 if (!err && nmask) {
1418 err = copy_from_user(bm, nm, alloc_size);
1419 /* ensure entire bitmap is zeroed */
1420 err |= clear_user(nmask, ALIGN(maxnode-1, 8) / 8);
1421 err |= compat_put_bitmap(nmask, bm, nr_bits);
1424 return err;
1427 asmlinkage long compat_sys_set_mempolicy(int mode, compat_ulong_t __user *nmask,
1428 compat_ulong_t maxnode)
1430 long err = 0;
1431 unsigned long __user *nm = NULL;
1432 unsigned long nr_bits, alloc_size;
1433 DECLARE_BITMAP(bm, MAX_NUMNODES);
1435 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1436 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1438 if (nmask) {
1439 err = compat_get_bitmap(bm, nmask, nr_bits);
1440 nm = compat_alloc_user_space(alloc_size);
1441 err |= copy_to_user(nm, bm, alloc_size);
1444 if (err)
1445 return -EFAULT;
1447 return sys_set_mempolicy(mode, nm, nr_bits+1);
1450 asmlinkage long compat_sys_mbind(compat_ulong_t start, compat_ulong_t len,
1451 compat_ulong_t mode, compat_ulong_t __user *nmask,
1452 compat_ulong_t maxnode, compat_ulong_t flags)
1454 long err = 0;
1455 unsigned long __user *nm = NULL;
1456 unsigned long nr_bits, alloc_size;
1457 nodemask_t bm;
1459 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1460 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1462 if (nmask) {
1463 err = compat_get_bitmap(nodes_addr(bm), nmask, nr_bits);
1464 nm = compat_alloc_user_space(alloc_size);
1465 err |= copy_to_user(nm, nodes_addr(bm), alloc_size);
1468 if (err)
1469 return -EFAULT;
1471 return sys_mbind(start, len, mode, nm, nr_bits+1, flags);
1474 #endif
1477 * get_vma_policy(@task, @vma, @addr)
1478 * @task - task for fallback if vma policy == default
1479 * @vma - virtual memory area whose policy is sought
1480 * @addr - address in @vma for shared policy lookup
1482 * Returns effective policy for a VMA at specified address.
1483 * Falls back to @task or system default policy, as necessary.
1484 * Current or other task's task mempolicy and non-shared vma policies
1485 * are protected by the task's mmap_sem, which must be held for read by
1486 * the caller.
1487 * Shared policies [those marked as MPOL_F_SHARED] require an extra reference
1488 * count--added by the get_policy() vm_op, as appropriate--to protect against
1489 * freeing by another task. It is the caller's responsibility to free the
1490 * extra reference for shared policies.
1492 static struct mempolicy *get_vma_policy(struct task_struct *task,
1493 struct vm_area_struct *vma, unsigned long addr)
1495 struct mempolicy *pol = task->mempolicy;
1497 if (vma) {
1498 if (vma->vm_ops && vma->vm_ops->get_policy) {
1499 struct mempolicy *vpol = vma->vm_ops->get_policy(vma,
1500 addr);
1501 if (vpol)
1502 pol = vpol;
1503 } else if (vma->vm_policy)
1504 pol = vma->vm_policy;
1506 if (!pol)
1507 pol = &default_policy;
1508 return pol;
1512 * Return a nodemask representing a mempolicy for filtering nodes for
1513 * page allocation
1515 static nodemask_t *policy_nodemask(gfp_t gfp, struct mempolicy *policy)
1517 /* Lower zones don't get a nodemask applied for MPOL_BIND */
1518 if (unlikely(policy->mode == MPOL_BIND) &&
1519 gfp_zone(gfp) >= policy_zone &&
1520 cpuset_nodemask_valid_mems_allowed(&policy->v.nodes))
1521 return &policy->v.nodes;
1523 return NULL;
1526 /* Return a zonelist indicated by gfp for node representing a mempolicy */
1527 static struct zonelist *policy_zonelist(gfp_t gfp, struct mempolicy *policy,
1528 int nd)
1530 switch (policy->mode) {
1531 case MPOL_PREFERRED:
1532 if (!(policy->flags & MPOL_F_LOCAL))
1533 nd = policy->v.preferred_node;
1534 break;
1535 case MPOL_BIND:
1537 * Normally, MPOL_BIND allocations are node-local within the
1538 * allowed nodemask. However, if __GFP_THISNODE is set and the
1539 * current node isn't part of the mask, we use the zonelist for
1540 * the first node in the mask instead.
1542 if (unlikely(gfp & __GFP_THISNODE) &&
1543 unlikely(!node_isset(nd, policy->v.nodes)))
1544 nd = first_node(policy->v.nodes);
1545 break;
1546 default:
1547 BUG();
1549 return node_zonelist(nd, gfp);
1552 /* Do dynamic interleaving for a process */
1553 static unsigned interleave_nodes(struct mempolicy *policy)
1555 unsigned nid, next;
1556 struct task_struct *me = current;
1558 nid = me->il_next;
1559 next = next_node(nid, policy->v.nodes);
1560 if (next >= MAX_NUMNODES)
1561 next = first_node(policy->v.nodes);
1562 if (next < MAX_NUMNODES)
1563 me->il_next = next;
1564 return nid;
1568 * Depending on the memory policy provide a node from which to allocate the
1569 * next slab entry.
1570 * @policy must be protected by freeing by the caller. If @policy is
1571 * the current task's mempolicy, this protection is implicit, as only the
1572 * task can change it's policy. The system default policy requires no
1573 * such protection.
1575 unsigned slab_node(struct mempolicy *policy)
1577 if (!policy || policy->flags & MPOL_F_LOCAL)
1578 return numa_node_id();
1580 switch (policy->mode) {
1581 case MPOL_PREFERRED:
1583 * handled MPOL_F_LOCAL above
1585 return policy->v.preferred_node;
1587 case MPOL_INTERLEAVE:
1588 return interleave_nodes(policy);
1590 case MPOL_BIND: {
1592 * Follow bind policy behavior and start allocation at the
1593 * first node.
1595 struct zonelist *zonelist;
1596 struct zone *zone;
1597 enum zone_type highest_zoneidx = gfp_zone(GFP_KERNEL);
1598 zonelist = &NODE_DATA(numa_node_id())->node_zonelists[0];
1599 (void)first_zones_zonelist(zonelist, highest_zoneidx,
1600 &policy->v.nodes,
1601 &zone);
1602 return zone ? zone->node : numa_node_id();
1605 default:
1606 BUG();
1610 /* Do static interleaving for a VMA with known offset. */
1611 static unsigned offset_il_node(struct mempolicy *pol,
1612 struct vm_area_struct *vma, unsigned long off)
1614 unsigned nnodes = nodes_weight(pol->v.nodes);
1615 unsigned target;
1616 int c;
1617 int nid = -1;
1619 if (!nnodes)
1620 return numa_node_id();
1621 target = (unsigned int)off % nnodes;
1622 c = 0;
1623 do {
1624 nid = next_node(nid, pol->v.nodes);
1625 c++;
1626 } while (c <= target);
1627 return nid;
1630 /* Determine a node number for interleave */
1631 static inline unsigned interleave_nid(struct mempolicy *pol,
1632 struct vm_area_struct *vma, unsigned long addr, int shift)
1634 if (vma) {
1635 unsigned long off;
1638 * for small pages, there is no difference between
1639 * shift and PAGE_SHIFT, so the bit-shift is safe.
1640 * for huge pages, since vm_pgoff is in units of small
1641 * pages, we need to shift off the always 0 bits to get
1642 * a useful offset.
1644 BUG_ON(shift < PAGE_SHIFT);
1645 off = vma->vm_pgoff >> (shift - PAGE_SHIFT);
1646 off += (addr - vma->vm_start) >> shift;
1647 return offset_il_node(pol, vma, off);
1648 } else
1649 return interleave_nodes(pol);
1652 #ifdef CONFIG_HUGETLBFS
1654 * huge_zonelist(@vma, @addr, @gfp_flags, @mpol)
1655 * @vma = virtual memory area whose policy is sought
1656 * @addr = address in @vma for shared policy lookup and interleave policy
1657 * @gfp_flags = for requested zone
1658 * @mpol = pointer to mempolicy pointer for reference counted mempolicy
1659 * @nodemask = pointer to nodemask pointer for MPOL_BIND nodemask
1661 * Returns a zonelist suitable for a huge page allocation and a pointer
1662 * to the struct mempolicy for conditional unref after allocation.
1663 * If the effective policy is 'BIND, returns a pointer to the mempolicy's
1664 * @nodemask for filtering the zonelist.
1666 * Must be protected by get_mems_allowed()
1668 struct zonelist *huge_zonelist(struct vm_area_struct *vma, unsigned long addr,
1669 gfp_t gfp_flags, struct mempolicy **mpol,
1670 nodemask_t **nodemask)
1672 struct zonelist *zl;
1674 *mpol = get_vma_policy(current, vma, addr);
1675 *nodemask = NULL; /* assume !MPOL_BIND */
1677 if (unlikely((*mpol)->mode == MPOL_INTERLEAVE)) {
1678 zl = node_zonelist(interleave_nid(*mpol, vma, addr,
1679 huge_page_shift(hstate_vma(vma))), gfp_flags);
1680 } else {
1681 zl = policy_zonelist(gfp_flags, *mpol, numa_node_id());
1682 if ((*mpol)->mode == MPOL_BIND)
1683 *nodemask = &(*mpol)->v.nodes;
1685 return zl;
1689 * init_nodemask_of_mempolicy
1691 * If the current task's mempolicy is "default" [NULL], return 'false'
1692 * to indicate default policy. Otherwise, extract the policy nodemask
1693 * for 'bind' or 'interleave' policy into the argument nodemask, or
1694 * initialize the argument nodemask to contain the single node for
1695 * 'preferred' or 'local' policy and return 'true' to indicate presence
1696 * of non-default mempolicy.
1698 * We don't bother with reference counting the mempolicy [mpol_get/put]
1699 * because the current task is examining it's own mempolicy and a task's
1700 * mempolicy is only ever changed by the task itself.
1702 * N.B., it is the caller's responsibility to free a returned nodemask.
1704 bool init_nodemask_of_mempolicy(nodemask_t *mask)
1706 struct mempolicy *mempolicy;
1707 int nid;
1709 if (!(mask && current->mempolicy))
1710 return false;
1712 task_lock(current);
1713 mempolicy = current->mempolicy;
1714 switch (mempolicy->mode) {
1715 case MPOL_PREFERRED:
1716 if (mempolicy->flags & MPOL_F_LOCAL)
1717 nid = numa_node_id();
1718 else
1719 nid = mempolicy->v.preferred_node;
1720 init_nodemask_of_node(mask, nid);
1721 break;
1723 case MPOL_BIND:
1724 /* Fall through */
1725 case MPOL_INTERLEAVE:
1726 *mask = mempolicy->v.nodes;
1727 break;
1729 default:
1730 BUG();
1732 task_unlock(current);
1734 return true;
1736 #endif
1739 * mempolicy_nodemask_intersects
1741 * If tsk's mempolicy is "default" [NULL], return 'true' to indicate default
1742 * policy. Otherwise, check for intersection between mask and the policy
1743 * nodemask for 'bind' or 'interleave' policy. For 'perferred' or 'local'
1744 * policy, always return true since it may allocate elsewhere on fallback.
1746 * Takes task_lock(tsk) to prevent freeing of its mempolicy.
1748 bool mempolicy_nodemask_intersects(struct task_struct *tsk,
1749 const nodemask_t *mask)
1751 struct mempolicy *mempolicy;
1752 bool ret = true;
1754 if (!mask)
1755 return ret;
1756 task_lock(tsk);
1757 mempolicy = tsk->mempolicy;
1758 if (!mempolicy)
1759 goto out;
1761 switch (mempolicy->mode) {
1762 case MPOL_PREFERRED:
1764 * MPOL_PREFERRED and MPOL_F_LOCAL are only preferred nodes to
1765 * allocate from, they may fallback to other nodes when oom.
1766 * Thus, it's possible for tsk to have allocated memory from
1767 * nodes in mask.
1769 break;
1770 case MPOL_BIND:
1771 case MPOL_INTERLEAVE:
1772 ret = nodes_intersects(mempolicy->v.nodes, *mask);
1773 break;
1774 default:
1775 BUG();
1777 out:
1778 task_unlock(tsk);
1779 return ret;
1782 /* Allocate a page in interleaved policy.
1783 Own path because it needs to do special accounting. */
1784 static struct page *alloc_page_interleave(gfp_t gfp, unsigned order,
1785 unsigned nid)
1787 struct zonelist *zl;
1788 struct page *page;
1790 zl = node_zonelist(nid, gfp);
1791 page = __alloc_pages(gfp, order, zl);
1792 if (page && page_zone(page) == zonelist_zone(&zl->_zonerefs[0]))
1793 inc_zone_page_state(page, NUMA_INTERLEAVE_HIT);
1794 return page;
1798 * alloc_pages_vma - Allocate a page for a VMA.
1800 * @gfp:
1801 * %GFP_USER user allocation.
1802 * %GFP_KERNEL kernel allocations,
1803 * %GFP_HIGHMEM highmem/user allocations,
1804 * %GFP_FS allocation should not call back into a file system.
1805 * %GFP_ATOMIC don't sleep.
1807 * @order:Order of the GFP allocation.
1808 * @vma: Pointer to VMA or NULL if not available.
1809 * @addr: Virtual Address of the allocation. Must be inside the VMA.
1811 * This function allocates a page from the kernel page pool and applies
1812 * a NUMA policy associated with the VMA or the current process.
1813 * When VMA is not NULL caller must hold down_read on the mmap_sem of the
1814 * mm_struct of the VMA to prevent it from going away. Should be used for
1815 * all allocations for pages that will be mapped into
1816 * user space. Returns NULL when no page can be allocated.
1818 * Should be called with the mm_sem of the vma hold.
1820 struct page *
1821 alloc_pages_vma(gfp_t gfp, int order, struct vm_area_struct *vma,
1822 unsigned long addr, int node)
1824 struct mempolicy *pol = get_vma_policy(current, vma, addr);
1825 struct zonelist *zl;
1826 struct page *page;
1828 get_mems_allowed();
1829 if (unlikely(pol->mode == MPOL_INTERLEAVE)) {
1830 unsigned nid;
1832 nid = interleave_nid(pol, vma, addr, PAGE_SHIFT + order);
1833 mpol_cond_put(pol);
1834 page = alloc_page_interleave(gfp, order, nid);
1835 put_mems_allowed();
1836 return page;
1838 zl = policy_zonelist(gfp, pol, node);
1839 if (unlikely(mpol_needs_cond_ref(pol))) {
1841 * slow path: ref counted shared policy
1843 struct page *page = __alloc_pages_nodemask(gfp, order,
1844 zl, policy_nodemask(gfp, pol));
1845 __mpol_put(pol);
1846 put_mems_allowed();
1847 return page;
1850 * fast path: default or task policy
1852 page = __alloc_pages_nodemask(gfp, order, zl,
1853 policy_nodemask(gfp, pol));
1854 put_mems_allowed();
1855 return page;
1859 * alloc_pages_current - Allocate pages.
1861 * @gfp:
1862 * %GFP_USER user allocation,
1863 * %GFP_KERNEL kernel allocation,
1864 * %GFP_HIGHMEM highmem allocation,
1865 * %GFP_FS don't call back into a file system.
1866 * %GFP_ATOMIC don't sleep.
1867 * @order: Power of two of allocation size in pages. 0 is a single page.
1869 * Allocate a page from the kernel page pool. When not in
1870 * interrupt context and apply the current process NUMA policy.
1871 * Returns NULL when no page can be allocated.
1873 * Don't call cpuset_update_task_memory_state() unless
1874 * 1) it's ok to take cpuset_sem (can WAIT), and
1875 * 2) allocating for current task (not interrupt).
1877 struct page *alloc_pages_current(gfp_t gfp, unsigned order)
1879 struct mempolicy *pol = current->mempolicy;
1880 struct page *page;
1882 if (!pol || in_interrupt() || (gfp & __GFP_THISNODE))
1883 pol = &default_policy;
1885 get_mems_allowed();
1887 * No reference counting needed for current->mempolicy
1888 * nor system default_policy
1890 if (pol->mode == MPOL_INTERLEAVE)
1891 page = alloc_page_interleave(gfp, order, interleave_nodes(pol));
1892 else
1893 page = __alloc_pages_nodemask(gfp, order,
1894 policy_zonelist(gfp, pol, numa_node_id()),
1895 policy_nodemask(gfp, pol));
1896 put_mems_allowed();
1897 return page;
1899 EXPORT_SYMBOL(alloc_pages_current);
1902 * If mpol_dup() sees current->cpuset == cpuset_being_rebound, then it
1903 * rebinds the mempolicy its copying by calling mpol_rebind_policy()
1904 * with the mems_allowed returned by cpuset_mems_allowed(). This
1905 * keeps mempolicies cpuset relative after its cpuset moves. See
1906 * further kernel/cpuset.c update_nodemask().
1908 * current's mempolicy may be rebinded by the other task(the task that changes
1909 * cpuset's mems), so we needn't do rebind work for current task.
1912 /* Slow path of a mempolicy duplicate */
1913 struct mempolicy *__mpol_dup(struct mempolicy *old)
1915 struct mempolicy *new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
1917 if (!new)
1918 return ERR_PTR(-ENOMEM);
1920 /* task's mempolicy is protected by alloc_lock */
1921 if (old == current->mempolicy) {
1922 task_lock(current);
1923 *new = *old;
1924 task_unlock(current);
1925 } else
1926 *new = *old;
1928 rcu_read_lock();
1929 if (current_cpuset_is_being_rebound()) {
1930 nodemask_t mems = cpuset_mems_allowed(current);
1931 if (new->flags & MPOL_F_REBINDING)
1932 mpol_rebind_policy(new, &mems, MPOL_REBIND_STEP2);
1933 else
1934 mpol_rebind_policy(new, &mems, MPOL_REBIND_ONCE);
1936 rcu_read_unlock();
1937 atomic_set(&new->refcnt, 1);
1938 return new;
1942 * If *frompol needs [has] an extra ref, copy *frompol to *tompol ,
1943 * eliminate the * MPOL_F_* flags that require conditional ref and
1944 * [NOTE!!!] drop the extra ref. Not safe to reference *frompol directly
1945 * after return. Use the returned value.
1947 * Allows use of a mempolicy for, e.g., multiple allocations with a single
1948 * policy lookup, even if the policy needs/has extra ref on lookup.
1949 * shmem_readahead needs this.
1951 struct mempolicy *__mpol_cond_copy(struct mempolicy *tompol,
1952 struct mempolicy *frompol)
1954 if (!mpol_needs_cond_ref(frompol))
1955 return frompol;
1957 *tompol = *frompol;
1958 tompol->flags &= ~MPOL_F_SHARED; /* copy doesn't need unref */
1959 __mpol_put(frompol);
1960 return tompol;
1963 /* Slow path of a mempolicy comparison */
1964 int __mpol_equal(struct mempolicy *a, struct mempolicy *b)
1966 if (!a || !b)
1967 return 0;
1968 if (a->mode != b->mode)
1969 return 0;
1970 if (a->flags != b->flags)
1971 return 0;
1972 if (mpol_store_user_nodemask(a))
1973 if (!nodes_equal(a->w.user_nodemask, b->w.user_nodemask))
1974 return 0;
1976 switch (a->mode) {
1977 case MPOL_BIND:
1978 /* Fall through */
1979 case MPOL_INTERLEAVE:
1980 return nodes_equal(a->v.nodes, b->v.nodes);
1981 case MPOL_PREFERRED:
1982 return a->v.preferred_node == b->v.preferred_node;
1983 default:
1984 BUG();
1985 return 0;
1990 * Shared memory backing store policy support.
1992 * Remember policies even when nobody has shared memory mapped.
1993 * The policies are kept in Red-Black tree linked from the inode.
1994 * They are protected by the sp->lock spinlock, which should be held
1995 * for any accesses to the tree.
1998 /* lookup first element intersecting start-end */
1999 /* Caller holds sp->lock */
2000 static struct sp_node *
2001 sp_lookup(struct shared_policy *sp, unsigned long start, unsigned long end)
2003 struct rb_node *n = sp->root.rb_node;
2005 while (n) {
2006 struct sp_node *p = rb_entry(n, struct sp_node, nd);
2008 if (start >= p->end)
2009 n = n->rb_right;
2010 else if (end <= p->start)
2011 n = n->rb_left;
2012 else
2013 break;
2015 if (!n)
2016 return NULL;
2017 for (;;) {
2018 struct sp_node *w = NULL;
2019 struct rb_node *prev = rb_prev(n);
2020 if (!prev)
2021 break;
2022 w = rb_entry(prev, struct sp_node, nd);
2023 if (w->end <= start)
2024 break;
2025 n = prev;
2027 return rb_entry(n, struct sp_node, nd);
2030 /* Insert a new shared policy into the list. */
2031 /* Caller holds sp->lock */
2032 static void sp_insert(struct shared_policy *sp, struct sp_node *new)
2034 struct rb_node **p = &sp->root.rb_node;
2035 struct rb_node *parent = NULL;
2036 struct sp_node *nd;
2038 while (*p) {
2039 parent = *p;
2040 nd = rb_entry(parent, struct sp_node, nd);
2041 if (new->start < nd->start)
2042 p = &(*p)->rb_left;
2043 else if (new->end > nd->end)
2044 p = &(*p)->rb_right;
2045 else
2046 BUG();
2048 rb_link_node(&new->nd, parent, p);
2049 rb_insert_color(&new->nd, &sp->root);
2050 pr_debug("inserting %lx-%lx: %d\n", new->start, new->end,
2051 new->policy ? new->policy->mode : 0);
2054 /* Find shared policy intersecting idx */
2055 struct mempolicy *
2056 mpol_shared_policy_lookup(struct shared_policy *sp, unsigned long idx)
2058 struct mempolicy *pol = NULL;
2059 struct sp_node *sn;
2061 if (!sp->root.rb_node)
2062 return NULL;
2063 spin_lock(&sp->lock);
2064 sn = sp_lookup(sp, idx, idx+1);
2065 if (sn) {
2066 mpol_get(sn->policy);
2067 pol = sn->policy;
2069 spin_unlock(&sp->lock);
2070 return pol;
2073 static void sp_delete(struct shared_policy *sp, struct sp_node *n)
2075 pr_debug("deleting %lx-l%lx\n", n->start, n->end);
2076 rb_erase(&n->nd, &sp->root);
2077 mpol_put(n->policy);
2078 kmem_cache_free(sn_cache, n);
2081 static struct sp_node *sp_alloc(unsigned long start, unsigned long end,
2082 struct mempolicy *pol)
2084 struct sp_node *n = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2086 if (!n)
2087 return NULL;
2088 n->start = start;
2089 n->end = end;
2090 mpol_get(pol);
2091 pol->flags |= MPOL_F_SHARED; /* for unref */
2092 n->policy = pol;
2093 return n;
2096 /* Replace a policy range. */
2097 static int shared_policy_replace(struct shared_policy *sp, unsigned long start,
2098 unsigned long end, struct sp_node *new)
2100 struct sp_node *n, *new2 = NULL;
2102 restart:
2103 spin_lock(&sp->lock);
2104 n = sp_lookup(sp, start, end);
2105 /* Take care of old policies in the same range. */
2106 while (n && n->start < end) {
2107 struct rb_node *next = rb_next(&n->nd);
2108 if (n->start >= start) {
2109 if (n->end <= end)
2110 sp_delete(sp, n);
2111 else
2112 n->start = end;
2113 } else {
2114 /* Old policy spanning whole new range. */
2115 if (n->end > end) {
2116 if (!new2) {
2117 spin_unlock(&sp->lock);
2118 new2 = sp_alloc(end, n->end, n->policy);
2119 if (!new2)
2120 return -ENOMEM;
2121 goto restart;
2123 n->end = start;
2124 sp_insert(sp, new2);
2125 new2 = NULL;
2126 break;
2127 } else
2128 n->end = start;
2130 if (!next)
2131 break;
2132 n = rb_entry(next, struct sp_node, nd);
2134 if (new)
2135 sp_insert(sp, new);
2136 spin_unlock(&sp->lock);
2137 if (new2) {
2138 mpol_put(new2->policy);
2139 kmem_cache_free(sn_cache, new2);
2141 return 0;
2145 * mpol_shared_policy_init - initialize shared policy for inode
2146 * @sp: pointer to inode shared policy
2147 * @mpol: struct mempolicy to install
2149 * Install non-NULL @mpol in inode's shared policy rb-tree.
2150 * On entry, the current task has a reference on a non-NULL @mpol.
2151 * This must be released on exit.
2152 * This is called at get_inode() calls and we can use GFP_KERNEL.
2154 void mpol_shared_policy_init(struct shared_policy *sp, struct mempolicy *mpol)
2156 int ret;
2158 sp->root = RB_ROOT; /* empty tree == default mempolicy */
2159 spin_lock_init(&sp->lock);
2161 if (mpol) {
2162 struct vm_area_struct pvma;
2163 struct mempolicy *new;
2164 NODEMASK_SCRATCH(scratch);
2166 if (!scratch)
2167 goto put_mpol;
2168 /* contextualize the tmpfs mount point mempolicy */
2169 new = mpol_new(mpol->mode, mpol->flags, &mpol->w.user_nodemask);
2170 if (IS_ERR(new))
2171 goto free_scratch; /* no valid nodemask intersection */
2173 task_lock(current);
2174 ret = mpol_set_nodemask(new, &mpol->w.user_nodemask, scratch);
2175 task_unlock(current);
2176 if (ret)
2177 goto put_new;
2179 /* Create pseudo-vma that contains just the policy */
2180 memset(&pvma, 0, sizeof(struct vm_area_struct));
2181 pvma.vm_end = TASK_SIZE; /* policy covers entire file */
2182 mpol_set_shared_policy(sp, &pvma, new); /* adds ref */
2184 put_new:
2185 mpol_put(new); /* drop initial ref */
2186 free_scratch:
2187 NODEMASK_SCRATCH_FREE(scratch);
2188 put_mpol:
2189 mpol_put(mpol); /* drop our incoming ref on sb mpol */
2193 int mpol_set_shared_policy(struct shared_policy *info,
2194 struct vm_area_struct *vma, struct mempolicy *npol)
2196 int err;
2197 struct sp_node *new = NULL;
2198 unsigned long sz = vma_pages(vma);
2200 pr_debug("set_shared_policy %lx sz %lu %d %d %lx\n",
2201 vma->vm_pgoff,
2202 sz, npol ? npol->mode : -1,
2203 npol ? npol->flags : -1,
2204 npol ? nodes_addr(npol->v.nodes)[0] : -1);
2206 if (npol) {
2207 new = sp_alloc(vma->vm_pgoff, vma->vm_pgoff + sz, npol);
2208 if (!new)
2209 return -ENOMEM;
2211 err = shared_policy_replace(info, vma->vm_pgoff, vma->vm_pgoff+sz, new);
2212 if (err && new)
2213 kmem_cache_free(sn_cache, new);
2214 return err;
2217 /* Free a backing policy store on inode delete. */
2218 void mpol_free_shared_policy(struct shared_policy *p)
2220 struct sp_node *n;
2221 struct rb_node *next;
2223 if (!p->root.rb_node)
2224 return;
2225 spin_lock(&p->lock);
2226 next = rb_first(&p->root);
2227 while (next) {
2228 n = rb_entry(next, struct sp_node, nd);
2229 next = rb_next(&n->nd);
2230 rb_erase(&n->nd, &p->root);
2231 mpol_put(n->policy);
2232 kmem_cache_free(sn_cache, n);
2234 spin_unlock(&p->lock);
2237 /* assumes fs == KERNEL_DS */
2238 void __init numa_policy_init(void)
2240 nodemask_t interleave_nodes;
2241 unsigned long largest = 0;
2242 int nid, prefer = 0;
2244 policy_cache = kmem_cache_create("numa_policy",
2245 sizeof(struct mempolicy),
2246 0, SLAB_PANIC, NULL);
2248 sn_cache = kmem_cache_create("shared_policy_node",
2249 sizeof(struct sp_node),
2250 0, SLAB_PANIC, NULL);
2253 * Set interleaving policy for system init. Interleaving is only
2254 * enabled across suitably sized nodes (default is >= 16MB), or
2255 * fall back to the largest node if they're all smaller.
2257 nodes_clear(interleave_nodes);
2258 for_each_node_state(nid, N_HIGH_MEMORY) {
2259 unsigned long total_pages = node_present_pages(nid);
2261 /* Preserve the largest node */
2262 if (largest < total_pages) {
2263 largest = total_pages;
2264 prefer = nid;
2267 /* Interleave this node? */
2268 if ((total_pages << PAGE_SHIFT) >= (16 << 20))
2269 node_set(nid, interleave_nodes);
2272 /* All too small, use the largest */
2273 if (unlikely(nodes_empty(interleave_nodes)))
2274 node_set(prefer, interleave_nodes);
2276 if (do_set_mempolicy(MPOL_INTERLEAVE, 0, &interleave_nodes))
2277 printk("numa_policy_init: interleaving failed\n");
2280 /* Reset policy of current process to default */
2281 void numa_default_policy(void)
2283 do_set_mempolicy(MPOL_DEFAULT, 0, NULL);
2287 * Parse and format mempolicy from/to strings
2291 * "local" is pseudo-policy: MPOL_PREFERRED with MPOL_F_LOCAL flag
2292 * Used only for mpol_parse_str() and mpol_to_str()
2294 #define MPOL_LOCAL MPOL_MAX
2295 static const char * const policy_modes[] =
2297 [MPOL_DEFAULT] = "default",
2298 [MPOL_PREFERRED] = "prefer",
2299 [MPOL_BIND] = "bind",
2300 [MPOL_INTERLEAVE] = "interleave",
2301 [MPOL_LOCAL] = "local"
2305 #ifdef CONFIG_TMPFS
2307 * mpol_parse_str - parse string to mempolicy
2308 * @str: string containing mempolicy to parse
2309 * @mpol: pointer to struct mempolicy pointer, returned on success.
2310 * @no_context: flag whether to "contextualize" the mempolicy
2312 * Format of input:
2313 * <mode>[=<flags>][:<nodelist>]
2315 * if @no_context is true, save the input nodemask in w.user_nodemask in
2316 * the returned mempolicy. This will be used to "clone" the mempolicy in
2317 * a specific context [cpuset] at a later time. Used to parse tmpfs mpol
2318 * mount option. Note that if 'static' or 'relative' mode flags were
2319 * specified, the input nodemask will already have been saved. Saving
2320 * it again is redundant, but safe.
2322 * On success, returns 0, else 1
2324 int mpol_parse_str(char *str, struct mempolicy **mpol, int no_context)
2326 struct mempolicy *new = NULL;
2327 unsigned short mode;
2328 unsigned short uninitialized_var(mode_flags);
2329 nodemask_t nodes;
2330 char *nodelist = strchr(str, ':');
2331 char *flags = strchr(str, '=');
2332 int err = 1;
2334 if (nodelist) {
2335 /* NUL-terminate mode or flags string */
2336 *nodelist++ = '\0';
2337 if (nodelist_parse(nodelist, nodes))
2338 goto out;
2339 if (!nodes_subset(nodes, node_states[N_HIGH_MEMORY]))
2340 goto out;
2341 } else
2342 nodes_clear(nodes);
2344 if (flags)
2345 *flags++ = '\0'; /* terminate mode string */
2347 for (mode = 0; mode <= MPOL_LOCAL; mode++) {
2348 if (!strcmp(str, policy_modes[mode])) {
2349 break;
2352 if (mode > MPOL_LOCAL)
2353 goto out;
2355 switch (mode) {
2356 case MPOL_PREFERRED:
2358 * Insist on a nodelist of one node only
2360 if (nodelist) {
2361 char *rest = nodelist;
2362 while (isdigit(*rest))
2363 rest++;
2364 if (*rest)
2365 goto out;
2367 break;
2368 case MPOL_INTERLEAVE:
2370 * Default to online nodes with memory if no nodelist
2372 if (!nodelist)
2373 nodes = node_states[N_HIGH_MEMORY];
2374 break;
2375 case MPOL_LOCAL:
2377 * Don't allow a nodelist; mpol_new() checks flags
2379 if (nodelist)
2380 goto out;
2381 mode = MPOL_PREFERRED;
2382 break;
2383 case MPOL_DEFAULT:
2385 * Insist on a empty nodelist
2387 if (!nodelist)
2388 err = 0;
2389 goto out;
2390 case MPOL_BIND:
2392 * Insist on a nodelist
2394 if (!nodelist)
2395 goto out;
2398 mode_flags = 0;
2399 if (flags) {
2401 * Currently, we only support two mutually exclusive
2402 * mode flags.
2404 if (!strcmp(flags, "static"))
2405 mode_flags |= MPOL_F_STATIC_NODES;
2406 else if (!strcmp(flags, "relative"))
2407 mode_flags |= MPOL_F_RELATIVE_NODES;
2408 else
2409 goto out;
2412 new = mpol_new(mode, mode_flags, &nodes);
2413 if (IS_ERR(new))
2414 goto out;
2416 if (no_context) {
2417 /* save for contextualization */
2418 new->w.user_nodemask = nodes;
2419 } else {
2420 int ret;
2421 NODEMASK_SCRATCH(scratch);
2422 if (scratch) {
2423 task_lock(current);
2424 ret = mpol_set_nodemask(new, &nodes, scratch);
2425 task_unlock(current);
2426 } else
2427 ret = -ENOMEM;
2428 NODEMASK_SCRATCH_FREE(scratch);
2429 if (ret) {
2430 mpol_put(new);
2431 goto out;
2434 err = 0;
2436 out:
2437 /* Restore string for error message */
2438 if (nodelist)
2439 *--nodelist = ':';
2440 if (flags)
2441 *--flags = '=';
2442 if (!err)
2443 *mpol = new;
2444 return err;
2446 #endif /* CONFIG_TMPFS */
2449 * mpol_to_str - format a mempolicy structure for printing
2450 * @buffer: to contain formatted mempolicy string
2451 * @maxlen: length of @buffer
2452 * @pol: pointer to mempolicy to be formatted
2453 * @no_context: "context free" mempolicy - use nodemask in w.user_nodemask
2455 * Convert a mempolicy into a string.
2456 * Returns the number of characters in buffer (if positive)
2457 * or an error (negative)
2459 int mpol_to_str(char *buffer, int maxlen, struct mempolicy *pol, int no_context)
2461 char *p = buffer;
2462 int l;
2463 nodemask_t nodes;
2464 unsigned short mode;
2465 unsigned short flags = pol ? pol->flags : 0;
2468 * Sanity check: room for longest mode, flag and some nodes
2470 VM_BUG_ON(maxlen < strlen("interleave") + strlen("relative") + 16);
2472 if (!pol || pol == &default_policy)
2473 mode = MPOL_DEFAULT;
2474 else
2475 mode = pol->mode;
2477 switch (mode) {
2478 case MPOL_DEFAULT:
2479 nodes_clear(nodes);
2480 break;
2482 case MPOL_PREFERRED:
2483 nodes_clear(nodes);
2484 if (flags & MPOL_F_LOCAL)
2485 mode = MPOL_LOCAL; /* pseudo-policy */
2486 else
2487 node_set(pol->v.preferred_node, nodes);
2488 break;
2490 case MPOL_BIND:
2491 /* Fall through */
2492 case MPOL_INTERLEAVE:
2493 if (no_context)
2494 nodes = pol->w.user_nodemask;
2495 else
2496 nodes = pol->v.nodes;
2497 break;
2499 default:
2500 BUG();
2503 l = strlen(policy_modes[mode]);
2504 if (buffer + maxlen < p + l + 1)
2505 return -ENOSPC;
2507 strcpy(p, policy_modes[mode]);
2508 p += l;
2510 if (flags & MPOL_MODE_FLAGS) {
2511 if (buffer + maxlen < p + 2)
2512 return -ENOSPC;
2513 *p++ = '=';
2516 * Currently, the only defined flags are mutually exclusive
2518 if (flags & MPOL_F_STATIC_NODES)
2519 p += snprintf(p, buffer + maxlen - p, "static");
2520 else if (flags & MPOL_F_RELATIVE_NODES)
2521 p += snprintf(p, buffer + maxlen - p, "relative");
2524 if (!nodes_empty(nodes)) {
2525 if (buffer + maxlen < p + 2)
2526 return -ENOSPC;
2527 *p++ = ':';
2528 p += nodelist_scnprintf(p, buffer + maxlen - p, nodes);
2530 return p - buffer;
2533 struct numa_maps {
2534 unsigned long pages;
2535 unsigned long anon;
2536 unsigned long active;
2537 unsigned long writeback;
2538 unsigned long mapcount_max;
2539 unsigned long dirty;
2540 unsigned long swapcache;
2541 unsigned long node[MAX_NUMNODES];
2544 static void gather_stats(struct page *page, void *private, int pte_dirty)
2546 struct numa_maps *md = private;
2547 int count = page_mapcount(page);
2549 md->pages++;
2550 if (pte_dirty || PageDirty(page))
2551 md->dirty++;
2553 if (PageSwapCache(page))
2554 md->swapcache++;
2556 if (PageActive(page) || PageUnevictable(page))
2557 md->active++;
2559 if (PageWriteback(page))
2560 md->writeback++;
2562 if (PageAnon(page))
2563 md->anon++;
2565 if (count > md->mapcount_max)
2566 md->mapcount_max = count;
2568 md->node[page_to_nid(page)]++;
2571 #ifdef CONFIG_HUGETLB_PAGE
2572 static void check_huge_range(struct vm_area_struct *vma,
2573 unsigned long start, unsigned long end,
2574 struct numa_maps *md)
2576 unsigned long addr;
2577 struct page *page;
2578 struct hstate *h = hstate_vma(vma);
2579 unsigned long sz = huge_page_size(h);
2581 for (addr = start; addr < end; addr += sz) {
2582 pte_t *ptep = huge_pte_offset(vma->vm_mm,
2583 addr & huge_page_mask(h));
2584 pte_t pte;
2586 if (!ptep)
2587 continue;
2589 pte = *ptep;
2590 if (pte_none(pte))
2591 continue;
2593 page = pte_page(pte);
2594 if (!page)
2595 continue;
2597 gather_stats(page, md, pte_dirty(*ptep));
2600 #else
2601 static inline void check_huge_range(struct vm_area_struct *vma,
2602 unsigned long start, unsigned long end,
2603 struct numa_maps *md)
2606 #endif
2609 * Display pages allocated per node and memory policy via /proc.
2611 int show_numa_map(struct seq_file *m, void *v)
2613 struct proc_maps_private *priv = m->private;
2614 struct vm_area_struct *vma = v;
2615 struct numa_maps *md;
2616 struct file *file = vma->vm_file;
2617 struct mm_struct *mm = vma->vm_mm;
2618 struct mempolicy *pol;
2619 int n;
2620 char buffer[50];
2622 if (!mm)
2623 return 0;
2625 md = kzalloc(sizeof(struct numa_maps), GFP_KERNEL);
2626 if (!md)
2627 return 0;
2629 pol = get_vma_policy(priv->task, vma, vma->vm_start);
2630 mpol_to_str(buffer, sizeof(buffer), pol, 0);
2631 mpol_cond_put(pol);
2633 seq_printf(m, "%08lx %s", vma->vm_start, buffer);
2635 if (file) {
2636 seq_printf(m, " file=");
2637 seq_path(m, &file->f_path, "\n\t= ");
2638 } else if (vma->vm_start <= mm->brk && vma->vm_end >= mm->start_brk) {
2639 seq_printf(m, " heap");
2640 } else if (vma->vm_start <= mm->start_stack &&
2641 vma->vm_end >= mm->start_stack) {
2642 seq_printf(m, " stack");
2645 if (is_vm_hugetlb_page(vma)) {
2646 check_huge_range(vma, vma->vm_start, vma->vm_end, md);
2647 seq_printf(m, " huge");
2648 } else {
2649 check_pgd_range(vma, vma->vm_start, vma->vm_end,
2650 &node_states[N_HIGH_MEMORY], MPOL_MF_STATS, md);
2653 if (!md->pages)
2654 goto out;
2656 if (md->anon)
2657 seq_printf(m," anon=%lu",md->anon);
2659 if (md->dirty)
2660 seq_printf(m," dirty=%lu",md->dirty);
2662 if (md->pages != md->anon && md->pages != md->dirty)
2663 seq_printf(m, " mapped=%lu", md->pages);
2665 if (md->mapcount_max > 1)
2666 seq_printf(m, " mapmax=%lu", md->mapcount_max);
2668 if (md->swapcache)
2669 seq_printf(m," swapcache=%lu", md->swapcache);
2671 if (md->active < md->pages && !is_vm_hugetlb_page(vma))
2672 seq_printf(m," active=%lu", md->active);
2674 if (md->writeback)
2675 seq_printf(m," writeback=%lu", md->writeback);
2677 for_each_node_state(n, N_HIGH_MEMORY)
2678 if (md->node[n])
2679 seq_printf(m, " N%d=%lu", n, md->node[n]);
2680 out:
2681 seq_putc(m, '\n');
2682 kfree(md);
2684 if (m->count < m->size)
2685 m->version = (vma != priv->tail_vma) ? vma->vm_start : 0;
2686 return 0;