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
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
22 * bind Only allocate memory on a specific set of nodes,
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
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.
57 fix mmap readahead to honour policy and enable policy for any page cache
59 statistics for bigpages
60 global policy for page cache? currently it uses process policy. Requires
62 handle mremap for shared memory (currently ignored for the policy)
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>
70 #include <linux/highmem.h>
71 #include <linux/hugetlb.h>
72 #include <linux/kernel.h>
73 #include <linux/sched.h>
74 #include <linux/nodemask.h>
75 #include <linux/cpuset.h>
76 #include <linux/slab.h>
77 #include <linux/string.h>
78 #include <linux/export.h>
79 #include <linux/nsproxy.h>
80 #include <linux/interrupt.h>
81 #include <linux/init.h>
82 #include <linux/compat.h>
83 #include <linux/swap.h>
84 #include <linux/seq_file.h>
85 #include <linux/proc_fs.h>
86 #include <linux/migrate.h>
87 #include <linux/ksm.h>
88 #include <linux/rmap.h>
89 #include <linux/security.h>
90 #include <linux/syscalls.h>
91 #include <linux/ctype.h>
92 #include <linux/mm_inline.h>
94 #include <asm/tlbflush.h>
95 #include <asm/uaccess.h>
96 #include <linux/random.h>
101 #define MPOL_MF_DISCONTIG_OK (MPOL_MF_INTERNAL << 0) /* Skip checks for continuous vmas */
102 #define MPOL_MF_INVERT (MPOL_MF_INTERNAL << 1) /* Invert check for nodemask */
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
109 enum zone_type policy_zone = 0;
112 * run-time system-wide default policy => local allocation
114 static struct mempolicy default_policy = {
115 .refcnt = ATOMIC_INIT(1), /* never free it */
116 .mode = MPOL_PREFERRED,
117 .flags = MPOL_F_LOCAL,
120 static struct mempolicy preferred_node_policy[MAX_NUMNODES];
122 static struct mempolicy *get_task_policy(struct task_struct *p)
124 struct mempolicy *pol = p->mempolicy;
128 node = tsk_home_node(p);
130 pol = &preferred_node_policy[node];
136 static const struct mempolicy_operations {
137 int (*create)(struct mempolicy *pol, const nodemask_t *nodes);
139 * If read-side task has no lock to protect task->mempolicy, write-side
140 * task will rebind the task->mempolicy by two step. The first step is
141 * setting all the newly nodes, and the second step is cleaning all the
142 * disallowed nodes. In this way, we can avoid finding no node to alloc
144 * If we have a lock to protect task->mempolicy in read-side, we do
148 * MPOL_REBIND_ONCE - do rebind work at once
149 * MPOL_REBIND_STEP1 - set all the newly nodes
150 * MPOL_REBIND_STEP2 - clean all the disallowed nodes
152 void (*rebind)(struct mempolicy *pol, const nodemask_t *nodes,
153 enum mpol_rebind_step step);
154 } mpol_ops[MPOL_MAX];
156 /* Check that the nodemask contains at least one populated zone */
157 static int is_valid_nodemask(const nodemask_t *nodemask)
161 for_each_node_mask(nd, *nodemask) {
164 for (k = 0; k <= policy_zone; k++) {
165 z = &NODE_DATA(nd)->node_zones[k];
166 if (z->present_pages > 0)
174 static inline int mpol_store_user_nodemask(const struct mempolicy *pol)
176 return pol->flags & MPOL_MODE_FLAGS;
179 static void mpol_relative_nodemask(nodemask_t *ret, const nodemask_t *orig,
180 const nodemask_t *rel)
183 nodes_fold(tmp, *orig, nodes_weight(*rel));
184 nodes_onto(*ret, tmp, *rel);
187 static int mpol_new_interleave(struct mempolicy *pol, const nodemask_t *nodes)
189 if (nodes_empty(*nodes))
191 pol->v.nodes = *nodes;
195 static int mpol_new_preferred(struct mempolicy *pol, const nodemask_t *nodes)
198 pol->flags |= MPOL_F_LOCAL; /* local allocation */
199 else if (nodes_empty(*nodes))
200 return -EINVAL; /* no allowed nodes */
202 pol->v.preferred_node = first_node(*nodes);
206 static int mpol_new_bind(struct mempolicy *pol, const nodemask_t *nodes)
208 if (!is_valid_nodemask(nodes))
210 pol->v.nodes = *nodes;
215 * mpol_set_nodemask is called after mpol_new() to set up the nodemask, if
216 * any, for the new policy. mpol_new() has already validated the nodes
217 * parameter with respect to the policy mode and flags. But, we need to
218 * handle an empty nodemask with MPOL_PREFERRED here.
220 * Must be called holding task's alloc_lock to protect task's mems_allowed
221 * and mempolicy. May also be called holding the mmap_semaphore for write.
223 static int mpol_set_nodemask(struct mempolicy *pol,
224 const nodemask_t *nodes, struct nodemask_scratch *nsc)
228 /* if mode is MPOL_DEFAULT, pol is NULL. This is right. */
231 /* Check N_HIGH_MEMORY */
232 nodes_and(nsc->mask1,
233 cpuset_current_mems_allowed, node_states[N_HIGH_MEMORY]);
236 if (pol->mode == MPOL_PREFERRED && nodes_empty(*nodes))
237 nodes = NULL; /* explicit local allocation */
239 if (pol->flags & MPOL_F_RELATIVE_NODES)
240 mpol_relative_nodemask(&nsc->mask2, nodes,&nsc->mask1);
242 nodes_and(nsc->mask2, *nodes, nsc->mask1);
244 if (mpol_store_user_nodemask(pol))
245 pol->w.user_nodemask = *nodes;
247 pol->w.cpuset_mems_allowed =
248 cpuset_current_mems_allowed;
252 ret = mpol_ops[pol->mode].create(pol, &nsc->mask2);
254 ret = mpol_ops[pol->mode].create(pol, NULL);
259 * This function just creates a new policy, does some check and simple
260 * initialization. You must invoke mpol_set_nodemask() to set nodes.
262 static struct mempolicy *mpol_new(unsigned short mode, unsigned short flags,
265 struct mempolicy *policy;
267 pr_debug("setting mode %d flags %d nodes[0] %lx\n",
268 mode, flags, nodes ? nodes_addr(*nodes)[0] : -1);
270 if (mode == MPOL_DEFAULT || mode == MPOL_NOOP) {
271 if (nodes && !nodes_empty(*nodes))
272 return ERR_PTR(-EINVAL);
278 * MPOL_PREFERRED cannot be used with MPOL_F_STATIC_NODES or
279 * MPOL_F_RELATIVE_NODES if the nodemask is empty (local allocation).
280 * All other modes require a valid pointer to a non-empty nodemask.
282 if (mode == MPOL_PREFERRED) {
283 if (nodes_empty(*nodes)) {
284 if (((flags & MPOL_F_STATIC_NODES) ||
285 (flags & MPOL_F_RELATIVE_NODES)))
286 return ERR_PTR(-EINVAL);
288 } else if (mode == MPOL_LOCAL) {
289 if (!nodes_empty(*nodes))
290 return ERR_PTR(-EINVAL);
291 mode = MPOL_PREFERRED;
292 } else if (nodes_empty(*nodes))
293 return ERR_PTR(-EINVAL);
294 policy = kmem_cache_alloc(policy_cache, GFP_KERNEL);
296 return ERR_PTR(-ENOMEM);
297 atomic_set(&policy->refcnt, 1);
299 policy->flags = flags;
304 /* Slow path of a mpol destructor. */
305 void __mpol_put(struct mempolicy *p)
307 if (!atomic_dec_and_test(&p->refcnt))
309 kmem_cache_free(policy_cache, p);
312 static void mpol_rebind_default(struct mempolicy *pol, const nodemask_t *nodes,
313 enum mpol_rebind_step step)
319 * MPOL_REBIND_ONCE - do rebind work at once
320 * MPOL_REBIND_STEP1 - set all the newly nodes
321 * MPOL_REBIND_STEP2 - clean all the disallowed nodes
323 static void mpol_rebind_nodemask(struct mempolicy *pol, const nodemask_t *nodes,
324 enum mpol_rebind_step step)
328 if (pol->flags & MPOL_F_STATIC_NODES)
329 nodes_and(tmp, pol->w.user_nodemask, *nodes);
330 else if (pol->flags & MPOL_F_RELATIVE_NODES)
331 mpol_relative_nodemask(&tmp, &pol->w.user_nodemask, nodes);
334 * if step == 1, we use ->w.cpuset_mems_allowed to cache the
337 if (step == MPOL_REBIND_ONCE || step == MPOL_REBIND_STEP1) {
338 nodes_remap(tmp, pol->v.nodes,
339 pol->w.cpuset_mems_allowed, *nodes);
340 pol->w.cpuset_mems_allowed = step ? tmp : *nodes;
341 } else if (step == MPOL_REBIND_STEP2) {
342 tmp = pol->w.cpuset_mems_allowed;
343 pol->w.cpuset_mems_allowed = *nodes;
348 if (nodes_empty(tmp))
351 if (step == MPOL_REBIND_STEP1)
352 nodes_or(pol->v.nodes, pol->v.nodes, tmp);
353 else if (step == MPOL_REBIND_ONCE || step == MPOL_REBIND_STEP2)
358 if (!node_isset(current->il_next, tmp)) {
359 current->il_next = next_node(current->il_next, tmp);
360 if (current->il_next >= MAX_NUMNODES)
361 current->il_next = first_node(tmp);
362 if (current->il_next >= MAX_NUMNODES)
363 current->il_next = numa_node_id();
367 static void mpol_rebind_preferred(struct mempolicy *pol,
368 const nodemask_t *nodes,
369 enum mpol_rebind_step step)
373 if (pol->flags & MPOL_F_STATIC_NODES) {
374 int node = first_node(pol->w.user_nodemask);
376 if (node_isset(node, *nodes)) {
377 pol->v.preferred_node = node;
378 pol->flags &= ~MPOL_F_LOCAL;
380 pol->flags |= MPOL_F_LOCAL;
381 } else if (pol->flags & MPOL_F_RELATIVE_NODES) {
382 mpol_relative_nodemask(&tmp, &pol->w.user_nodemask, nodes);
383 pol->v.preferred_node = first_node(tmp);
384 } else if (!(pol->flags & MPOL_F_LOCAL)) {
385 pol->v.preferred_node = node_remap(pol->v.preferred_node,
386 pol->w.cpuset_mems_allowed,
388 pol->w.cpuset_mems_allowed = *nodes;
393 * mpol_rebind_policy - Migrate a policy to a different set of nodes
395 * If read-side task has no lock to protect task->mempolicy, write-side
396 * task will rebind the task->mempolicy by two step. The first step is
397 * setting all the newly nodes, and the second step is cleaning all the
398 * disallowed nodes. In this way, we can avoid finding no node to alloc
400 * If we have a lock to protect task->mempolicy in read-side, we do
404 * MPOL_REBIND_ONCE - do rebind work at once
405 * MPOL_REBIND_STEP1 - set all the newly nodes
406 * MPOL_REBIND_STEP2 - clean all the disallowed nodes
408 static void mpol_rebind_policy(struct mempolicy *pol, const nodemask_t *newmask,
409 enum mpol_rebind_step step)
413 if (!mpol_store_user_nodemask(pol) && step == MPOL_REBIND_ONCE &&
414 nodes_equal(pol->w.cpuset_mems_allowed, *newmask))
417 if (step == MPOL_REBIND_STEP1 && (pol->flags & MPOL_F_REBINDING))
420 if (step == MPOL_REBIND_STEP2 && !(pol->flags & MPOL_F_REBINDING))
423 if (step == MPOL_REBIND_STEP1)
424 pol->flags |= MPOL_F_REBINDING;
425 else if (step == MPOL_REBIND_STEP2)
426 pol->flags &= ~MPOL_F_REBINDING;
427 else if (step >= MPOL_REBIND_NSTEP)
430 mpol_ops[pol->mode].rebind(pol, newmask, step);
434 * Wrapper for mpol_rebind_policy() that just requires task
435 * pointer, and updates task mempolicy.
437 * Called with task's alloc_lock held.
440 void mpol_rebind_task(struct task_struct *tsk, const nodemask_t *new,
441 enum mpol_rebind_step step)
443 mpol_rebind_policy(tsk->mempolicy, new, step);
447 * Rebind each vma in mm to new nodemask.
449 * Call holding a reference to mm. Takes mm->mmap_sem during call.
452 void mpol_rebind_mm(struct mm_struct *mm, nodemask_t *new)
454 struct vm_area_struct *vma;
456 down_write(&mm->mmap_sem);
457 for (vma = mm->mmap; vma; vma = vma->vm_next)
458 mpol_rebind_policy(vma->vm_policy, new, MPOL_REBIND_ONCE);
459 up_write(&mm->mmap_sem);
462 static const struct mempolicy_operations mpol_ops[MPOL_MAX] = {
464 .rebind = mpol_rebind_default,
466 [MPOL_INTERLEAVE] = {
467 .create = mpol_new_interleave,
468 .rebind = mpol_rebind_nodemask,
471 .create = mpol_new_preferred,
472 .rebind = mpol_rebind_preferred,
475 .create = mpol_new_bind,
476 .rebind = mpol_rebind_nodemask,
480 static void migrate_page_add(struct page *page, struct list_head *pagelist,
481 unsigned long flags);
483 /* Scan through pages checking if pages follow certain conditions. */
484 static int check_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
485 unsigned long addr, unsigned long end,
486 const nodemask_t *nodes, unsigned long flags,
493 orig_pte = pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
498 if (!pte_present(*pte))
500 page = vm_normal_page(vma, addr, *pte);
504 * vm_normal_page() filters out zero pages, but there might
505 * still be PageReserved pages to skip, perhaps in a VDSO.
506 * And we cannot move PageKsm pages sensibly or safely yet.
508 if (PageReserved(page) || PageKsm(page))
510 nid = page_to_nid(page);
511 if (node_isset(nid, *nodes) == !!(flags & MPOL_MF_INVERT))
514 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))
515 migrate_page_add(page, private, flags);
518 } while (pte++, addr += PAGE_SIZE, addr != end);
519 pte_unmap_unlock(orig_pte, ptl);
523 static inline int check_pmd_range(struct vm_area_struct *vma, pud_t *pud,
524 unsigned long addr, unsigned long end,
525 const nodemask_t *nodes, unsigned long flags,
531 pmd = pmd_offset(pud, addr);
533 next = pmd_addr_end(addr, end);
534 split_huge_page_pmd(vma->vm_mm, pmd);
535 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
537 if (check_pte_range(vma, pmd, addr, next, nodes,
540 } while (pmd++, addr = next, addr != end);
544 static inline int check_pud_range(struct vm_area_struct *vma, pgd_t *pgd,
545 unsigned long addr, unsigned long end,
546 const nodemask_t *nodes, unsigned long flags,
552 pud = pud_offset(pgd, addr);
554 next = pud_addr_end(addr, end);
555 if (pud_none_or_clear_bad(pud))
557 if (check_pmd_range(vma, pud, addr, next, nodes,
560 } while (pud++, addr = next, addr != end);
564 static inline int check_pgd_range(struct vm_area_struct *vma,
565 unsigned long addr, unsigned long end,
566 const nodemask_t *nodes, unsigned long flags,
572 pgd = pgd_offset(vma->vm_mm, addr);
574 next = pgd_addr_end(addr, end);
575 if (pgd_none_or_clear_bad(pgd))
577 if (check_pud_range(vma, pgd, addr, next, nodes,
580 } while (pgd++, addr = next, addr != end);
585 change_prot_none(struct vm_area_struct *vma, unsigned long start, unsigned long end)
587 change_protection(vma, start, end, vma_prot_none(vma), 0);
591 * Check if all pages in a range are on a set of nodes.
592 * If pagelist != NULL then isolate pages from the LRU and
593 * put them on the pagelist.
595 static struct vm_area_struct *
596 check_range(struct mm_struct *mm, unsigned long start, unsigned long end,
597 const nodemask_t *nodes, unsigned long flags, void *private)
600 struct vm_area_struct *first, *vma, *prev;
603 first = find_vma(mm, start);
605 return ERR_PTR(-EFAULT);
607 for (vma = first; vma && vma->vm_start < end; vma = vma->vm_next) {
608 unsigned long endvma = vma->vm_end;
612 if (vma->vm_start > start)
613 start = vma->vm_start;
615 if (!(flags & MPOL_MF_DISCONTIG_OK)) {
616 if (!vma->vm_next && vma->vm_end < end)
617 return ERR_PTR(-EFAULT);
618 if (prev && prev->vm_end < vma->vm_start)
619 return ERR_PTR(-EFAULT);
622 if (is_vm_hugetlb_page(vma))
625 if (flags & MPOL_MF_LAZY) {
626 change_prot_none(vma, start, endvma);
630 if ((flags & MPOL_MF_STRICT) ||
631 ((flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) &&
632 vma_migratable(vma))) {
634 err = check_pgd_range(vma, start, endvma, nodes,
637 first = ERR_PTR(err);
648 * Apply policy to a single VMA
649 * This must be called with the mmap_sem held for writing.
651 static int vma_replace_policy(struct vm_area_struct *vma,
652 struct mempolicy *pol)
655 struct mempolicy *old;
656 struct mempolicy *new;
658 pr_debug("vma %lx-%lx/%lx vm_ops %p vm_file %p set_policy %p\n",
659 vma->vm_start, vma->vm_end, vma->vm_pgoff,
660 vma->vm_ops, vma->vm_file,
661 vma->vm_ops ? vma->vm_ops->set_policy : NULL);
667 if (vma->vm_ops && vma->vm_ops->set_policy) {
668 err = vma->vm_ops->set_policy(vma, new);
673 old = vma->vm_policy;
674 vma->vm_policy = new; /* protected by mmap_sem */
683 /* Step 2: apply policy to a range and do splits. */
684 static int mbind_range(struct mm_struct *mm, unsigned long start,
685 unsigned long end, struct mempolicy *new_pol)
687 struct vm_area_struct *next;
688 struct vm_area_struct *prev;
689 struct vm_area_struct *vma;
692 unsigned long vmstart;
695 vma = find_vma(mm, start);
696 if (!vma || vma->vm_start > start)
700 if (start > vma->vm_start)
703 for (; vma && vma->vm_start < end; prev = vma, vma = next) {
705 vmstart = max(start, vma->vm_start);
706 vmend = min(end, vma->vm_end);
708 if (mpol_equal(vma_policy(vma), new_pol))
711 pgoff = vma->vm_pgoff +
712 ((vmstart - vma->vm_start) >> PAGE_SHIFT);
713 prev = vma_merge(mm, prev, vmstart, vmend, vma->vm_flags,
714 vma->anon_vma, vma->vm_file, pgoff,
721 if (vma->vm_start != vmstart) {
722 err = split_vma(vma->vm_mm, vma, vmstart, 1);
726 if (vma->vm_end != vmend) {
727 err = split_vma(vma->vm_mm, vma, vmend, 0);
731 err = vma_replace_policy(vma, new_pol);
741 * Update task->flags PF_MEMPOLICY bit: set iff non-default
742 * mempolicy. Allows more rapid checking of this (combined perhaps
743 * with other PF_* flag bits) on memory allocation hot code paths.
745 * If called from outside this file, the task 'p' should -only- be
746 * a newly forked child not yet visible on the task list, because
747 * manipulating the task flags of a visible task is not safe.
749 * The above limitation is why this routine has the funny name
750 * mpol_fix_fork_child_flag().
752 * It is also safe to call this with a task pointer of current,
753 * which the static wrapper mpol_set_task_struct_flag() does,
754 * for use within this file.
757 void mpol_fix_fork_child_flag(struct task_struct *p)
760 p->flags |= PF_MEMPOLICY;
762 p->flags &= ~PF_MEMPOLICY;
765 static void mpol_set_task_struct_flag(void)
767 mpol_fix_fork_child_flag(current);
770 /* Set the process memory policy */
771 static long do_set_mempolicy(unsigned short mode, unsigned short flags,
774 struct mempolicy *new, *old;
775 struct mm_struct *mm = current->mm;
776 NODEMASK_SCRATCH(scratch);
782 new = mpol_new(mode, flags, nodes);
788 * prevent changing our mempolicy while show_numa_maps()
790 * Note: do_set_mempolicy() can be called at init time
794 down_write(&mm->mmap_sem);
796 ret = mpol_set_nodemask(new, nodes, scratch);
798 task_unlock(current);
800 up_write(&mm->mmap_sem);
804 old = current->mempolicy;
805 current->mempolicy = new;
806 mpol_set_task_struct_flag();
807 if (new && new->mode == MPOL_INTERLEAVE &&
808 nodes_weight(new->v.nodes))
809 current->il_next = first_node(new->v.nodes);
810 task_unlock(current);
812 up_write(&mm->mmap_sem);
817 NODEMASK_SCRATCH_FREE(scratch);
822 * Return nodemask for policy for get_mempolicy() query
824 * Called with task's alloc_lock held
826 static void get_policy_nodemask(struct mempolicy *p, nodemask_t *nodes)
829 if (p == &default_policy)
835 case MPOL_INTERLEAVE:
839 if (!(p->flags & MPOL_F_LOCAL))
840 node_set(p->v.preferred_node, *nodes);
841 /* else return empty node mask for local allocation */
848 static int lookup_node(struct mm_struct *mm, unsigned long addr)
853 err = get_user_pages(current, mm, addr & PAGE_MASK, 1, 0, 0, &p, NULL);
855 err = page_to_nid(p);
861 /* Retrieve NUMA policy */
862 static long do_get_mempolicy(int *policy, nodemask_t *nmask,
863 unsigned long addr, unsigned long flags)
866 struct mm_struct *mm = current->mm;
867 struct vm_area_struct *vma = NULL;
868 struct mempolicy *pol = current->mempolicy;
871 ~(unsigned long)(MPOL_F_NODE|MPOL_F_ADDR|MPOL_F_MEMS_ALLOWED))
874 if (flags & MPOL_F_MEMS_ALLOWED) {
875 if (flags & (MPOL_F_NODE|MPOL_F_ADDR))
877 *policy = 0; /* just so it's initialized */
879 *nmask = cpuset_current_mems_allowed;
880 task_unlock(current);
884 if (flags & MPOL_F_ADDR) {
886 * Do NOT fall back to task policy if the
887 * vma/shared policy at addr is NULL. We
888 * want to return MPOL_DEFAULT in this case.
890 down_read(&mm->mmap_sem);
891 vma = find_vma_intersection(mm, addr, addr+1);
893 up_read(&mm->mmap_sem);
896 if (vma->vm_ops && vma->vm_ops->get_policy)
897 pol = vma->vm_ops->get_policy(vma, addr);
899 pol = vma->vm_policy;
904 pol = &default_policy; /* indicates default behavior */
906 if (flags & MPOL_F_NODE) {
907 if (flags & MPOL_F_ADDR) {
908 err = lookup_node(mm, addr);
912 } else if (pol == current->mempolicy &&
913 pol->mode == MPOL_INTERLEAVE) {
914 *policy = current->il_next;
920 *policy = pol == &default_policy ? MPOL_DEFAULT :
923 * Internal mempolicy flags must be masked off before exposing
924 * the policy to userspace.
926 *policy |= (pol->flags & MPOL_MODE_FLAGS);
930 up_read(¤t->mm->mmap_sem);
936 if (mpol_store_user_nodemask(pol)) {
937 *nmask = pol->w.user_nodemask;
940 get_policy_nodemask(pol, nmask);
941 task_unlock(current);
948 up_read(¤t->mm->mmap_sem);
952 #ifdef CONFIG_MIGRATION
956 static void migrate_page_add(struct page *page, struct list_head *pagelist,
960 * Avoid migrating a page that is shared with others.
962 if ((flags & MPOL_MF_MOVE_ALL) || page_mapcount(page) == 1) {
963 if (!isolate_lru_page(page)) {
964 list_add_tail(&page->lru, pagelist);
965 inc_zone_page_state(page, NR_ISOLATED_ANON +
966 page_is_file_cache(page));
971 static struct page *new_node_page(struct page *page, unsigned long node, int **x)
973 return alloc_pages_exact_node(node, GFP_HIGHUSER_MOVABLE, 0);
977 * Migrate pages from one node to a target node.
978 * Returns error or the number of pages not migrated.
980 static int migrate_to_node(struct mm_struct *mm, int source, int dest,
988 node_set(source, nmask);
991 * This does not "check" the range but isolates all pages that
992 * need migration. Between passing in the full user address
993 * space range and MPOL_MF_DISCONTIG_OK, this call can not fail.
995 VM_BUG_ON(!(flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)));
996 check_range(mm, mm->mmap->vm_start, mm->task_size, &nmask,
997 flags | MPOL_MF_DISCONTIG_OK, &pagelist);
999 if (!list_empty(&pagelist)) {
1000 err = migrate_pages(&pagelist, new_node_page, dest,
1001 false, MIGRATE_SYNC);
1003 putback_lru_pages(&pagelist);
1010 * Move pages between the two nodesets so as to preserve the physical
1011 * layout as much as possible.
1013 * Returns the number of page that could not be moved.
1015 int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
1016 const nodemask_t *to, int flags)
1022 err = migrate_prep();
1026 down_read(&mm->mmap_sem);
1028 err = migrate_vmas(mm, from, to, flags);
1033 * Find a 'source' bit set in 'tmp' whose corresponding 'dest'
1034 * bit in 'to' is not also set in 'tmp'. Clear the found 'source'
1035 * bit in 'tmp', and return that <source, dest> pair for migration.
1036 * The pair of nodemasks 'to' and 'from' define the map.
1038 * If no pair of bits is found that way, fallback to picking some
1039 * pair of 'source' and 'dest' bits that are not the same. If the
1040 * 'source' and 'dest' bits are the same, this represents a node
1041 * that will be migrating to itself, so no pages need move.
1043 * If no bits are left in 'tmp', or if all remaining bits left
1044 * in 'tmp' correspond to the same bit in 'to', return false
1045 * (nothing left to migrate).
1047 * This lets us pick a pair of nodes to migrate between, such that
1048 * if possible the dest node is not already occupied by some other
1049 * source node, minimizing the risk of overloading the memory on a
1050 * node that would happen if we migrated incoming memory to a node
1051 * before migrating outgoing memory source that same node.
1053 * A single scan of tmp is sufficient. As we go, we remember the
1054 * most recent <s, d> pair that moved (s != d). If we find a pair
1055 * that not only moved, but what's better, moved to an empty slot
1056 * (d is not set in tmp), then we break out then, with that pair.
1057 * Otherwise when we finish scanning from_tmp, we at least have the
1058 * most recent <s, d> pair that moved. If we get all the way through
1059 * the scan of tmp without finding any node that moved, much less
1060 * moved to an empty node, then there is nothing left worth migrating.
1064 while (!nodes_empty(tmp)) {
1069 for_each_node_mask(s, tmp) {
1072 * do_migrate_pages() tries to maintain the relative
1073 * node relationship of the pages established between
1074 * threads and memory areas.
1076 * However if the number of source nodes is not equal to
1077 * the number of destination nodes we can not preserve
1078 * this node relative relationship. In that case, skip
1079 * copying memory from a node that is in the destination
1082 * Example: [2,3,4] -> [3,4,5] moves everything.
1083 * [0-7] - > [3,4,5] moves only 0,1,2,6,7.
1086 if ((nodes_weight(*from) != nodes_weight(*to)) &&
1087 (node_isset(s, *to)))
1090 d = node_remap(s, *from, *to);
1094 source = s; /* Node moved. Memorize */
1097 /* dest not in remaining from nodes? */
1098 if (!node_isset(dest, tmp))
1104 node_clear(source, tmp);
1105 err = migrate_to_node(mm, source, dest, flags);
1112 up_read(&mm->mmap_sem);
1120 * Allocate a new page for page migration based on vma policy.
1121 * Start assuming that page is mapped by vma pointed to by @private.
1122 * Search forward from there, if not. N.B., this assumes that the
1123 * list of pages handed to migrate_pages()--which is how we get here--
1124 * is in virtual address order.
1126 static struct page *new_vma_page(struct page *page, unsigned long private, int **x)
1128 struct vm_area_struct *vma = (struct vm_area_struct *)private;
1129 unsigned long uninitialized_var(address);
1132 address = page_address_in_vma(page, vma);
1133 if (address != -EFAULT)
1139 * if !vma, alloc_page_vma() will use task or system default policy
1141 return alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
1145 static void migrate_page_add(struct page *page, struct list_head *pagelist,
1146 unsigned long flags)
1150 int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
1151 const nodemask_t *to, int flags)
1156 static struct page *new_vma_page(struct page *page, unsigned long private, int **x)
1162 static long do_mbind(unsigned long start, unsigned long len,
1163 unsigned short mode, unsigned short mode_flags,
1164 nodemask_t *nmask, unsigned long flags)
1166 struct vm_area_struct *vma;
1167 struct mm_struct *mm = current->mm;
1168 struct mempolicy *new;
1171 LIST_HEAD(pagelist);
1173 if (flags & ~(unsigned long)MPOL_MF_VALID)
1175 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1178 if (start & ~PAGE_MASK)
1181 if (mode == MPOL_DEFAULT || mode == MPOL_NOOP)
1182 flags &= ~MPOL_MF_STRICT;
1184 len = (len + PAGE_SIZE - 1) & PAGE_MASK;
1192 new = mpol_new(mode, mode_flags, nmask);
1194 return PTR_ERR(new);
1196 if (flags & MPOL_MF_LAZY)
1197 new->flags |= MPOL_F_MOF;
1200 * If we are using the default policy then operation
1201 * on discontinuous address spaces is okay after all
1204 flags |= MPOL_MF_DISCONTIG_OK;
1206 pr_debug("mbind %lx-%lx mode:%d flags:%d nodes:%lx\n",
1207 start, start + len, mode, mode_flags,
1208 nmask ? nodes_addr(*nmask)[0] : -1);
1210 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) {
1212 err = migrate_prep();
1217 NODEMASK_SCRATCH(scratch);
1219 down_write(&mm->mmap_sem);
1221 err = mpol_set_nodemask(new, nmask, scratch);
1222 task_unlock(current);
1224 up_write(&mm->mmap_sem);
1227 NODEMASK_SCRATCH_FREE(scratch);
1232 vma = check_range(mm, start, end, nmask,
1233 flags | MPOL_MF_INVERT, &pagelist);
1235 err = PTR_ERR(vma); /* maybe ... */
1236 if (!IS_ERR(vma) && mode != MPOL_NOOP)
1237 err = mbind_range(mm, start, end, new);
1242 if (!list_empty(&pagelist)) {
1243 WARN_ON_ONCE(flags & MPOL_MF_LAZY);
1244 nr_failed = migrate_pages(&pagelist, new_vma_page,
1246 false, MIGRATE_SYNC);
1248 putback_lru_pages(&pagelist);
1251 if (nr_failed && (flags & MPOL_MF_STRICT))
1254 putback_lru_pages(&pagelist);
1256 up_write(&mm->mmap_sem);
1262 static void lazy_migrate_vma(struct vm_area_struct *vma)
1264 if (!vma_migratable(vma))
1267 change_prot_none(vma, vma->vm_start, vma->vm_end);
1270 void lazy_migrate_process(struct mm_struct *mm)
1272 struct vm_area_struct *vma;
1274 down_read(&mm->mmap_sem);
1275 for (vma = mm->mmap; vma; vma = vma->vm_next)
1276 lazy_migrate_vma(vma);
1277 up_read(&mm->mmap_sem);
1281 * User space interface with variable sized bitmaps for nodelists.
1284 /* Copy a node mask from user space. */
1285 static int get_nodes(nodemask_t *nodes, const unsigned long __user *nmask,
1286 unsigned long maxnode)
1289 unsigned long nlongs;
1290 unsigned long endmask;
1293 nodes_clear(*nodes);
1294 if (maxnode == 0 || !nmask)
1296 if (maxnode > PAGE_SIZE*BITS_PER_BYTE)
1299 nlongs = BITS_TO_LONGS(maxnode);
1300 if ((maxnode % BITS_PER_LONG) == 0)
1303 endmask = (1UL << (maxnode % BITS_PER_LONG)) - 1;
1305 /* When the user specified more nodes than supported just check
1306 if the non supported part is all zero. */
1307 if (nlongs > BITS_TO_LONGS(MAX_NUMNODES)) {
1308 if (nlongs > PAGE_SIZE/sizeof(long))
1310 for (k = BITS_TO_LONGS(MAX_NUMNODES); k < nlongs; k++) {
1312 if (get_user(t, nmask + k))
1314 if (k == nlongs - 1) {
1320 nlongs = BITS_TO_LONGS(MAX_NUMNODES);
1324 if (copy_from_user(nodes_addr(*nodes), nmask, nlongs*sizeof(unsigned long)))
1326 nodes_addr(*nodes)[nlongs-1] &= endmask;
1330 /* Copy a kernel node mask to user space */
1331 static int copy_nodes_to_user(unsigned long __user *mask, unsigned long maxnode,
1334 unsigned long copy = ALIGN(maxnode-1, 64) / 8;
1335 const int nbytes = BITS_TO_LONGS(MAX_NUMNODES) * sizeof(long);
1337 if (copy > nbytes) {
1338 if (copy > PAGE_SIZE)
1340 if (clear_user((char __user *)mask + nbytes, copy - nbytes))
1344 return copy_to_user(mask, nodes_addr(*nodes), copy) ? -EFAULT : 0;
1347 SYSCALL_DEFINE6(mbind, unsigned long, start, unsigned long, len,
1348 unsigned long, mode, unsigned long __user *, nmask,
1349 unsigned long, maxnode, unsigned, flags)
1353 unsigned short mode_flags;
1355 mode_flags = mode & MPOL_MODE_FLAGS;
1356 mode &= ~MPOL_MODE_FLAGS;
1357 if (mode >= MPOL_MAX)
1359 if ((mode_flags & MPOL_F_STATIC_NODES) &&
1360 (mode_flags & MPOL_F_RELATIVE_NODES))
1362 err = get_nodes(&nodes, nmask, maxnode);
1365 return do_mbind(start, len, mode, mode_flags, &nodes, flags);
1368 /* Set the process memory policy */
1369 SYSCALL_DEFINE3(set_mempolicy, int, mode, unsigned long __user *, nmask,
1370 unsigned long, maxnode)
1374 unsigned short flags;
1376 flags = mode & MPOL_MODE_FLAGS;
1377 mode &= ~MPOL_MODE_FLAGS;
1378 if ((unsigned int)mode >= MPOL_MAX)
1380 if ((flags & MPOL_F_STATIC_NODES) && (flags & MPOL_F_RELATIVE_NODES))
1382 err = get_nodes(&nodes, nmask, maxnode);
1385 return do_set_mempolicy(mode, flags, &nodes);
1388 SYSCALL_DEFINE4(migrate_pages, pid_t, pid, unsigned long, maxnode,
1389 const unsigned long __user *, old_nodes,
1390 const unsigned long __user *, new_nodes)
1392 const struct cred *cred = current_cred(), *tcred;
1393 struct mm_struct *mm = NULL;
1394 struct task_struct *task;
1395 nodemask_t task_nodes;
1399 NODEMASK_SCRATCH(scratch);
1404 old = &scratch->mask1;
1405 new = &scratch->mask2;
1407 err = get_nodes(old, old_nodes, maxnode);
1411 err = get_nodes(new, new_nodes, maxnode);
1415 /* Find the mm_struct */
1417 task = pid ? find_task_by_vpid(pid) : current;
1423 get_task_struct(task);
1428 * Check if this process has the right to modify the specified
1429 * process. The right exists if the process has administrative
1430 * capabilities, superuser privileges or the same
1431 * userid as the target process.
1433 tcred = __task_cred(task);
1434 if (!uid_eq(cred->euid, tcred->suid) && !uid_eq(cred->euid, tcred->uid) &&
1435 !uid_eq(cred->uid, tcred->suid) && !uid_eq(cred->uid, tcred->uid) &&
1436 !capable(CAP_SYS_NICE)) {
1443 task_nodes = cpuset_mems_allowed(task);
1444 /* Is the user allowed to access the target nodes? */
1445 if (!nodes_subset(*new, task_nodes) && !capable(CAP_SYS_NICE)) {
1450 if (!nodes_subset(*new, node_states[N_HIGH_MEMORY])) {
1455 err = security_task_movememory(task);
1459 mm = get_task_mm(task);
1460 put_task_struct(task);
1467 err = do_migrate_pages(mm, old, new,
1468 capable(CAP_SYS_NICE) ? MPOL_MF_MOVE_ALL : MPOL_MF_MOVE);
1472 NODEMASK_SCRATCH_FREE(scratch);
1477 put_task_struct(task);
1483 /* Retrieve NUMA policy */
1484 SYSCALL_DEFINE5(get_mempolicy, int __user *, policy,
1485 unsigned long __user *, nmask, unsigned long, maxnode,
1486 unsigned long, addr, unsigned long, flags)
1489 int uninitialized_var(pval);
1492 if (nmask != NULL && maxnode < MAX_NUMNODES)
1495 err = do_get_mempolicy(&pval, &nodes, addr, flags);
1500 if (policy && put_user(pval, policy))
1504 err = copy_nodes_to_user(nmask, maxnode, &nodes);
1509 #ifdef CONFIG_COMPAT
1511 asmlinkage long compat_sys_get_mempolicy(int __user *policy,
1512 compat_ulong_t __user *nmask,
1513 compat_ulong_t maxnode,
1514 compat_ulong_t addr, compat_ulong_t flags)
1517 unsigned long __user *nm = NULL;
1518 unsigned long nr_bits, alloc_size;
1519 DECLARE_BITMAP(bm, MAX_NUMNODES);
1521 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1522 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1525 nm = compat_alloc_user_space(alloc_size);
1527 err = sys_get_mempolicy(policy, nm, nr_bits+1, addr, flags);
1529 if (!err && nmask) {
1530 unsigned long copy_size;
1531 copy_size = min_t(unsigned long, sizeof(bm), alloc_size);
1532 err = copy_from_user(bm, nm, copy_size);
1533 /* ensure entire bitmap is zeroed */
1534 err |= clear_user(nmask, ALIGN(maxnode-1, 8) / 8);
1535 err |= compat_put_bitmap(nmask, bm, nr_bits);
1541 asmlinkage long compat_sys_set_mempolicy(int mode, compat_ulong_t __user *nmask,
1542 compat_ulong_t maxnode)
1545 unsigned long __user *nm = NULL;
1546 unsigned long nr_bits, alloc_size;
1547 DECLARE_BITMAP(bm, MAX_NUMNODES);
1549 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1550 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1553 err = compat_get_bitmap(bm, nmask, nr_bits);
1554 nm = compat_alloc_user_space(alloc_size);
1555 err |= copy_to_user(nm, bm, alloc_size);
1561 return sys_set_mempolicy(mode, nm, nr_bits+1);
1564 asmlinkage long compat_sys_mbind(compat_ulong_t start, compat_ulong_t len,
1565 compat_ulong_t mode, compat_ulong_t __user *nmask,
1566 compat_ulong_t maxnode, compat_ulong_t flags)
1569 unsigned long __user *nm = NULL;
1570 unsigned long nr_bits, alloc_size;
1573 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1574 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1577 err = compat_get_bitmap(nodes_addr(bm), nmask, nr_bits);
1578 nm = compat_alloc_user_space(alloc_size);
1579 err |= copy_to_user(nm, nodes_addr(bm), alloc_size);
1585 return sys_mbind(start, len, mode, nm, nr_bits+1, flags);
1591 * get_vma_policy(@task, @vma, @addr)
1592 * @task - task for fallback if vma policy == default
1593 * @vma - virtual memory area whose policy is sought
1594 * @addr - address in @vma for shared policy lookup
1596 * Returns effective policy for a VMA at specified address.
1597 * Falls back to @task or system default policy, as necessary.
1598 * Current or other task's task mempolicy and non-shared vma policies must be
1599 * protected by task_lock(task) by the caller.
1600 * Shared policies [those marked as MPOL_F_SHARED] require an extra reference
1601 * count--added by the get_policy() vm_op, as appropriate--to protect against
1602 * freeing by another task. It is the caller's responsibility to free the
1603 * extra reference for shared policies.
1605 struct mempolicy *get_vma_policy(struct task_struct *task,
1606 struct vm_area_struct *vma, unsigned long addr)
1608 struct mempolicy *pol = get_task_policy(task);
1611 if (vma->vm_ops && vma->vm_ops->get_policy) {
1612 struct mempolicy *vpol = vma->vm_ops->get_policy(vma,
1616 } else if (vma->vm_policy) {
1617 pol = vma->vm_policy;
1620 * shmem_alloc_page() passes MPOL_F_SHARED policy with
1621 * a pseudo vma whose vma->vm_ops=NULL. Take a reference
1622 * count on these policies which will be dropped by
1623 * mpol_cond_put() later
1625 if (mpol_needs_cond_ref(pol))
1630 pol = &default_policy;
1635 * Return a nodemask representing a mempolicy for filtering nodes for
1638 static nodemask_t *policy_nodemask(gfp_t gfp, struct mempolicy *policy)
1640 /* Lower zones don't get a nodemask applied for MPOL_BIND */
1641 if (unlikely(policy->mode == MPOL_BIND) &&
1642 gfp_zone(gfp) >= policy_zone &&
1643 cpuset_nodemask_valid_mems_allowed(&policy->v.nodes))
1644 return &policy->v.nodes;
1649 /* Do dynamic interleaving for a process */
1650 static unsigned interleave_nodes(struct mempolicy *policy)
1653 struct task_struct *me = current;
1656 next = next_node(nid, policy->v.nodes);
1657 if (next >= MAX_NUMNODES)
1658 next = first_node(policy->v.nodes);
1659 if (next < MAX_NUMNODES)
1664 /* Return a zonelist indicated by gfp for node representing a mempolicy */
1665 static struct zonelist *policy_zonelist(gfp_t gfp, struct mempolicy *policy,
1668 switch (policy->mode) {
1669 case MPOL_INTERLEAVE:
1670 nd = interleave_nodes(policy);
1672 case MPOL_PREFERRED:
1673 if (!(policy->flags & MPOL_F_LOCAL))
1674 nd = policy->v.preferred_node;
1678 * Normally, MPOL_BIND allocations are node-local within the
1679 * allowed nodemask. However, if __GFP_THISNODE is set and the
1680 * current node isn't part of the mask, we use the zonelist for
1681 * the first node in the mask instead.
1683 if (unlikely(gfp & __GFP_THISNODE) &&
1684 unlikely(!node_isset(nd, policy->v.nodes)))
1685 nd = first_node(policy->v.nodes);
1690 return node_zonelist(nd, gfp);
1694 * Depending on the memory policy provide a node from which to allocate the
1696 * @policy must be protected by freeing by the caller. If @policy is
1697 * the current task's mempolicy, this protection is implicit, as only the
1698 * task can change it's policy. The system default policy requires no
1701 unsigned slab_node(void)
1703 struct mempolicy *policy;
1706 return numa_node_id();
1708 policy = current->mempolicy;
1709 if (!policy || policy->flags & MPOL_F_LOCAL)
1710 return numa_node_id();
1712 switch (policy->mode) {
1713 case MPOL_PREFERRED:
1715 * handled MPOL_F_LOCAL above
1717 return policy->v.preferred_node;
1719 case MPOL_INTERLEAVE:
1720 return interleave_nodes(policy);
1724 * Follow bind policy behavior and start allocation at the
1727 struct zonelist *zonelist;
1729 enum zone_type highest_zoneidx = gfp_zone(GFP_KERNEL);
1730 zonelist = &NODE_DATA(numa_node_id())->node_zonelists[0];
1731 (void)first_zones_zonelist(zonelist, highest_zoneidx,
1734 return zone ? zone->node : numa_node_id();
1742 /* Do static interleaving for a VMA with known offset. */
1743 static unsigned offset_il_node(struct mempolicy *pol,
1744 struct vm_area_struct *vma, unsigned long off)
1746 unsigned nnodes = nodes_weight(pol->v.nodes);
1752 return numa_node_id();
1753 target = (unsigned int)off % nnodes;
1756 nid = next_node(nid, pol->v.nodes);
1758 } while (c <= target);
1762 /* Determine a node number for interleave */
1763 static inline unsigned interleave_nid(struct mempolicy *pol,
1764 struct vm_area_struct *vma, unsigned long addr, int shift)
1770 * for small pages, there is no difference between
1771 * shift and PAGE_SHIFT, so the bit-shift is safe.
1772 * for huge pages, since vm_pgoff is in units of small
1773 * pages, we need to shift off the always 0 bits to get
1776 BUG_ON(shift < PAGE_SHIFT);
1777 off = vma->vm_pgoff >> (shift - PAGE_SHIFT);
1778 off += (addr - vma->vm_start) >> shift;
1779 return offset_il_node(pol, vma, off);
1781 return interleave_nodes(pol);
1785 * Return the bit number of a random bit set in the nodemask.
1786 * (returns -1 if nodemask is empty)
1788 int node_random(const nodemask_t *maskp)
1792 w = nodes_weight(*maskp);
1794 bit = bitmap_ord_to_pos(maskp->bits,
1795 get_random_int() % w, MAX_NUMNODES);
1799 #ifdef CONFIG_HUGETLBFS
1801 * huge_zonelist(@vma, @addr, @gfp_flags, @mpol)
1802 * @vma = virtual memory area whose policy is sought
1803 * @addr = address in @vma for shared policy lookup and interleave policy
1804 * @gfp_flags = for requested zone
1805 * @mpol = pointer to mempolicy pointer for reference counted mempolicy
1806 * @nodemask = pointer to nodemask pointer for MPOL_BIND nodemask
1808 * Returns a zonelist suitable for a huge page allocation and a pointer
1809 * to the struct mempolicy for conditional unref after allocation.
1810 * If the effective policy is 'BIND, returns a pointer to the mempolicy's
1811 * @nodemask for filtering the zonelist.
1813 * Must be protected by get_mems_allowed()
1815 struct zonelist *huge_zonelist(struct vm_area_struct *vma, unsigned long addr,
1816 gfp_t gfp_flags, struct mempolicy **mpol,
1817 nodemask_t **nodemask)
1819 struct zonelist *zl;
1821 *mpol = get_vma_policy(current, vma, addr);
1822 *nodemask = NULL; /* assume !MPOL_BIND */
1824 if (unlikely((*mpol)->mode == MPOL_INTERLEAVE)) {
1825 zl = node_zonelist(interleave_nid(*mpol, vma, addr,
1826 huge_page_shift(hstate_vma(vma))), gfp_flags);
1828 zl = policy_zonelist(gfp_flags, *mpol, numa_node_id());
1829 if ((*mpol)->mode == MPOL_BIND)
1830 *nodemask = &(*mpol)->v.nodes;
1836 * init_nodemask_of_mempolicy
1838 * If the current task's mempolicy is "default" [NULL], return 'false'
1839 * to indicate default policy. Otherwise, extract the policy nodemask
1840 * for 'bind' or 'interleave' policy into the argument nodemask, or
1841 * initialize the argument nodemask to contain the single node for
1842 * 'preferred' or 'local' policy and return 'true' to indicate presence
1843 * of non-default mempolicy.
1845 * We don't bother with reference counting the mempolicy [mpol_get/put]
1846 * because the current task is examining it's own mempolicy and a task's
1847 * mempolicy is only ever changed by the task itself.
1849 * N.B., it is the caller's responsibility to free a returned nodemask.
1851 bool init_nodemask_of_mempolicy(nodemask_t *mask)
1853 struct mempolicy *mempolicy;
1856 if (!(mask && current->mempolicy))
1860 mempolicy = current->mempolicy;
1861 switch (mempolicy->mode) {
1862 case MPOL_PREFERRED:
1863 if (mempolicy->flags & MPOL_F_LOCAL)
1864 nid = numa_node_id();
1866 nid = mempolicy->v.preferred_node;
1867 init_nodemask_of_node(mask, nid);
1872 case MPOL_INTERLEAVE:
1873 *mask = mempolicy->v.nodes;
1879 task_unlock(current);
1886 * mempolicy_nodemask_intersects
1888 * If tsk's mempolicy is "default" [NULL], return 'true' to indicate default
1889 * policy. Otherwise, check for intersection between mask and the policy
1890 * nodemask for 'bind' or 'interleave' policy. For 'perferred' or 'local'
1891 * policy, always return true since it may allocate elsewhere on fallback.
1893 * Takes task_lock(tsk) to prevent freeing of its mempolicy.
1895 bool mempolicy_nodemask_intersects(struct task_struct *tsk,
1896 const nodemask_t *mask)
1898 struct mempolicy *mempolicy;
1904 mempolicy = tsk->mempolicy;
1908 switch (mempolicy->mode) {
1909 case MPOL_PREFERRED:
1911 * MPOL_PREFERRED and MPOL_F_LOCAL are only preferred nodes to
1912 * allocate from, they may fallback to other nodes when oom.
1913 * Thus, it's possible for tsk to have allocated memory from
1918 case MPOL_INTERLEAVE:
1919 ret = nodes_intersects(mempolicy->v.nodes, *mask);
1930 * alloc_pages_vma - Allocate a page for a VMA.
1933 * %GFP_USER user allocation.
1934 * %GFP_KERNEL kernel allocations,
1935 * %GFP_HIGHMEM highmem/user allocations,
1936 * %GFP_FS allocation should not call back into a file system.
1937 * %GFP_ATOMIC don't sleep.
1939 * @order:Order of the GFP allocation.
1940 * @vma: Pointer to VMA or NULL if not available.
1941 * @addr: Virtual Address of the allocation. Must be inside the VMA.
1943 * This function allocates a page from the kernel page pool and applies
1944 * a NUMA policy associated with the VMA or the current process.
1945 * When VMA is not NULL caller must hold down_read on the mmap_sem of the
1946 * mm_struct of the VMA to prevent it from going away. Should be used for
1947 * all allocations for pages that will be mapped into
1948 * user space. Returns NULL when no page can be allocated.
1950 * Should be called with the mm_sem of the vma hold.
1953 alloc_pages_vma(gfp_t gfp, int order, struct vm_area_struct *vma,
1954 unsigned long addr, int node)
1956 struct mempolicy *pol;
1957 struct zonelist *zl;
1959 unsigned int cpuset_mems_cookie;
1962 pol = get_vma_policy(current, vma, addr);
1963 cpuset_mems_cookie = get_mems_allowed();
1965 zl = policy_zonelist(gfp, pol, node);
1966 if (unlikely(mpol_needs_cond_ref(pol))) {
1968 * slow path: ref counted shared policy
1970 struct page *page = __alloc_pages_nodemask(gfp, order,
1971 zl, policy_nodemask(gfp, pol));
1973 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page))
1978 * fast path: default or task policy
1980 page = __alloc_pages_nodemask(gfp, order, zl,
1981 policy_nodemask(gfp, pol));
1982 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page))
1988 * alloc_pages_current - Allocate pages.
1991 * %GFP_USER user allocation,
1992 * %GFP_KERNEL kernel allocation,
1993 * %GFP_HIGHMEM highmem allocation,
1994 * %GFP_FS don't call back into a file system.
1995 * %GFP_ATOMIC don't sleep.
1996 * @order: Power of two of allocation size in pages. 0 is a single page.
1998 * Allocate a page from the kernel page pool. When not in
1999 * interrupt context and apply the current process NUMA policy.
2000 * Returns NULL when no page can be allocated.
2002 * Don't call cpuset_update_task_memory_state() unless
2003 * 1) it's ok to take cpuset_sem (can WAIT), and
2004 * 2) allocating for current task (not interrupt).
2006 struct page *alloc_pages_current(gfp_t gfp, unsigned order)
2008 struct mempolicy *pol = get_task_policy(current);
2010 unsigned int cpuset_mems_cookie;
2012 if (!pol || in_interrupt() || (gfp & __GFP_THISNODE))
2013 pol = &default_policy;
2016 cpuset_mems_cookie = get_mems_allowed();
2019 * No reference counting needed for current->mempolicy
2020 * nor system default_policy
2022 page = __alloc_pages_nodemask(gfp, order,
2023 policy_zonelist(gfp, pol, numa_node_id()),
2024 policy_nodemask(gfp, pol));
2026 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page))
2031 EXPORT_SYMBOL(alloc_pages_current);
2034 * If mpol_dup() sees current->cpuset == cpuset_being_rebound, then it
2035 * rebinds the mempolicy its copying by calling mpol_rebind_policy()
2036 * with the mems_allowed returned by cpuset_mems_allowed(). This
2037 * keeps mempolicies cpuset relative after its cpuset moves. See
2038 * further kernel/cpuset.c update_nodemask().
2040 * current's mempolicy may be rebinded by the other task(the task that changes
2041 * cpuset's mems), so we needn't do rebind work for current task.
2044 /* Slow path of a mempolicy duplicate */
2045 struct mempolicy *__mpol_dup(struct mempolicy *old)
2047 struct mempolicy *new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
2050 return ERR_PTR(-ENOMEM);
2052 /* task's mempolicy is protected by alloc_lock */
2053 if (old == current->mempolicy) {
2056 task_unlock(current);
2061 if (current_cpuset_is_being_rebound()) {
2062 nodemask_t mems = cpuset_mems_allowed(current);
2063 if (new->flags & MPOL_F_REBINDING)
2064 mpol_rebind_policy(new, &mems, MPOL_REBIND_STEP2);
2066 mpol_rebind_policy(new, &mems, MPOL_REBIND_ONCE);
2069 atomic_set(&new->refcnt, 1);
2074 * If *frompol needs [has] an extra ref, copy *frompol to *tompol ,
2075 * eliminate the * MPOL_F_* flags that require conditional ref and
2076 * [NOTE!!!] drop the extra ref. Not safe to reference *frompol directly
2077 * after return. Use the returned value.
2079 * Allows use of a mempolicy for, e.g., multiple allocations with a single
2080 * policy lookup, even if the policy needs/has extra ref on lookup.
2081 * shmem_readahead needs this.
2083 struct mempolicy *__mpol_cond_copy(struct mempolicy *tompol,
2084 struct mempolicy *frompol)
2086 if (!mpol_needs_cond_ref(frompol))
2090 tompol->flags &= ~MPOL_F_SHARED; /* copy doesn't need unref */
2091 __mpol_put(frompol);
2095 /* Slow path of a mempolicy comparison */
2096 bool __mpol_equal(struct mempolicy *a, struct mempolicy *b)
2100 if (a->mode != b->mode)
2102 if (a->flags != b->flags)
2104 if (mpol_store_user_nodemask(a))
2105 if (!nodes_equal(a->w.user_nodemask, b->w.user_nodemask))
2111 case MPOL_INTERLEAVE:
2112 return !!nodes_equal(a->v.nodes, b->v.nodes);
2113 case MPOL_PREFERRED:
2114 return a->v.preferred_node == b->v.preferred_node;
2122 * Shared memory backing store policy support.
2124 * Remember policies even when nobody has shared memory mapped.
2125 * The policies are kept in Red-Black tree linked from the inode.
2126 * They are protected by the sp->lock spinlock, which should be held
2127 * for any accesses to the tree.
2130 /* lookup first element intersecting start-end */
2131 /* Caller holds sp->mutex */
2132 static struct sp_node *
2133 sp_lookup(struct shared_policy *sp, unsigned long start, unsigned long end)
2135 struct rb_node *n = sp->root.rb_node;
2138 struct sp_node *p = rb_entry(n, struct sp_node, nd);
2140 if (start >= p->end)
2142 else if (end <= p->start)
2150 struct sp_node *w = NULL;
2151 struct rb_node *prev = rb_prev(n);
2154 w = rb_entry(prev, struct sp_node, nd);
2155 if (w->end <= start)
2159 return rb_entry(n, struct sp_node, nd);
2162 /* Insert a new shared policy into the list. */
2163 /* Caller holds sp->lock */
2164 static void sp_insert(struct shared_policy *sp, struct sp_node *new)
2166 struct rb_node **p = &sp->root.rb_node;
2167 struct rb_node *parent = NULL;
2172 nd = rb_entry(parent, struct sp_node, nd);
2173 if (new->start < nd->start)
2175 else if (new->end > nd->end)
2176 p = &(*p)->rb_right;
2180 rb_link_node(&new->nd, parent, p);
2181 rb_insert_color(&new->nd, &sp->root);
2182 pr_debug("inserting %lx-%lx: %d\n", new->start, new->end,
2183 new->policy ? new->policy->mode : 0);
2186 /* Find shared policy intersecting idx */
2188 mpol_shared_policy_lookup(struct shared_policy *sp, unsigned long idx)
2190 struct mempolicy *pol = NULL;
2193 if (!sp->root.rb_node)
2195 mutex_lock(&sp->mutex);
2196 sn = sp_lookup(sp, idx, idx+1);
2198 mpol_get(sn->policy);
2201 mutex_unlock(&sp->mutex);
2205 static void sp_free(struct sp_node *n)
2207 mpol_put(n->policy);
2208 kmem_cache_free(sn_cache, n);
2212 * mpol_misplaced - check whether current page node is valid in policy
2214 * @page - page to be checked
2215 * @vma - vm area where page mapped
2216 * @addr - virtual address where page mapped
2217 * @multi - use multi-stage node binding
2219 * Lookup current policy node id for vma,addr and "compare to" page's
2223 * -1 - not misplaced, page is in the right node
2224 * node - node id where the page should be
2226 * Policy determination "mimics" alloc_page_vma().
2227 * Called from fault path where we know the vma and faulting address.
2229 int mpol_misplaced(struct page *page, struct vm_area_struct *vma,
2232 struct mempolicy *pol;
2234 int curnid = page_to_nid(page);
2235 unsigned long pgoff;
2241 pol = get_vma_policy(current, vma, addr);
2242 if (!(pol->flags & MPOL_F_MOF))
2245 switch (pol->mode) {
2246 case MPOL_INTERLEAVE:
2247 BUG_ON(addr >= vma->vm_end);
2248 BUG_ON(addr < vma->vm_start);
2250 pgoff = vma->vm_pgoff;
2251 pgoff += (addr - vma->vm_start) >> PAGE_SHIFT;
2252 polnid = offset_il_node(pol, vma, pgoff);
2255 case MPOL_PREFERRED:
2256 if (pol->flags & MPOL_F_LOCAL)
2257 polnid = numa_node_id();
2259 polnid = pol->v.preferred_node;
2264 * allows binding to multiple nodes.
2265 * use current page if in policy nodemask,
2266 * else select nearest allowed node, if any.
2267 * If no allowed nodes, use current [!misplaced].
2269 if (node_isset(curnid, pol->v.nodes))
2271 (void)first_zones_zonelist(
2272 node_zonelist(numa_node_id(), GFP_HIGHUSER),
2273 gfp_zone(GFP_HIGHUSER),
2274 &pol->v.nodes, &zone);
2275 polnid = zone->node;
2283 * Multi-stage node selection is used in conjunction with a periodic
2284 * migration fault to build a temporal task<->page relation. By
2285 * using a two-stage filter we remove short/unlikely relations.
2287 * Using P(p) ~ n_p / n_t as per frequentist probability, we can
2288 * equate a task's usage of a particular page (n_p) per total usage
2289 * of this page (n_t) (in a given time-span) to a probability.
2291 * Our periodic faults will then sample this probability and getting
2292 * the same result twice in a row, given these samples are fully
2293 * independent, is then given by P(n)^2, provided our sample period
2294 * is sufficiently short compared to the usage pattern.
2296 * This quadric squishes small probabilities, making it less likely
2297 * we act on an unlikely task<->page relation.
2299 if (pol->flags & MPOL_F_HOME) {
2303 * Migrate towards the current node, depends on
2304 * task_numa_placement() details.
2306 polnid = numa_node_id();
2307 last_nid = page_xchg_last_nid(page, polnid);
2308 if (last_nid != polnid)
2312 if (curnid != polnid)
2320 static void sp_delete(struct shared_policy *sp, struct sp_node *n)
2322 pr_debug("deleting %lx-l%lx\n", n->start, n->end);
2323 rb_erase(&n->nd, &sp->root);
2327 static struct sp_node *sp_alloc(unsigned long start, unsigned long end,
2328 struct mempolicy *pol)
2331 struct mempolicy *newpol;
2333 n = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2337 newpol = mpol_dup(pol);
2338 if (IS_ERR(newpol)) {
2339 kmem_cache_free(sn_cache, n);
2342 newpol->flags |= MPOL_F_SHARED;
2351 /* Replace a policy range. */
2352 static int shared_policy_replace(struct shared_policy *sp, unsigned long start,
2353 unsigned long end, struct sp_node *new)
2358 mutex_lock(&sp->mutex);
2359 n = sp_lookup(sp, start, end);
2360 /* Take care of old policies in the same range. */
2361 while (n && n->start < end) {
2362 struct rb_node *next = rb_next(&n->nd);
2363 if (n->start >= start) {
2369 /* Old policy spanning whole new range. */
2371 struct sp_node *new2;
2372 new2 = sp_alloc(end, n->end, n->policy);
2378 sp_insert(sp, new2);
2385 n = rb_entry(next, struct sp_node, nd);
2390 mutex_unlock(&sp->mutex);
2395 * mpol_shared_policy_init - initialize shared policy for inode
2396 * @sp: pointer to inode shared policy
2397 * @mpol: struct mempolicy to install
2399 * Install non-NULL @mpol in inode's shared policy rb-tree.
2400 * On entry, the current task has a reference on a non-NULL @mpol.
2401 * This must be released on exit.
2402 * This is called at get_inode() calls and we can use GFP_KERNEL.
2404 void mpol_shared_policy_init(struct shared_policy *sp, struct mempolicy *mpol)
2408 sp->root = RB_ROOT; /* empty tree == default mempolicy */
2409 mutex_init(&sp->mutex);
2412 struct vm_area_struct pvma;
2413 struct mempolicy *new;
2414 NODEMASK_SCRATCH(scratch);
2418 /* contextualize the tmpfs mount point mempolicy */
2419 new = mpol_new(mpol->mode, mpol->flags, &mpol->w.user_nodemask);
2421 goto free_scratch; /* no valid nodemask intersection */
2424 ret = mpol_set_nodemask(new, &mpol->w.user_nodemask, scratch);
2425 task_unlock(current);
2429 /* Create pseudo-vma that contains just the policy */
2430 memset(&pvma, 0, sizeof(struct vm_area_struct));
2431 pvma.vm_end = TASK_SIZE; /* policy covers entire file */
2432 mpol_set_shared_policy(sp, &pvma, new); /* adds ref */
2435 mpol_put(new); /* drop initial ref */
2437 NODEMASK_SCRATCH_FREE(scratch);
2439 mpol_put(mpol); /* drop our incoming ref on sb mpol */
2443 int mpol_set_shared_policy(struct shared_policy *info,
2444 struct vm_area_struct *vma, struct mempolicy *npol)
2447 struct sp_node *new = NULL;
2448 unsigned long sz = vma_pages(vma);
2450 pr_debug("set_shared_policy %lx sz %lu %d %d %lx\n",
2452 sz, npol ? npol->mode : -1,
2453 npol ? npol->flags : -1,
2454 npol ? nodes_addr(npol->v.nodes)[0] : -1);
2457 new = sp_alloc(vma->vm_pgoff, vma->vm_pgoff + sz, npol);
2461 err = shared_policy_replace(info, vma->vm_pgoff, vma->vm_pgoff+sz, new);
2467 /* Free a backing policy store on inode delete. */
2468 void mpol_free_shared_policy(struct shared_policy *p)
2471 struct rb_node *next;
2473 if (!p->root.rb_node)
2475 mutex_lock(&p->mutex);
2476 next = rb_first(&p->root);
2478 n = rb_entry(next, struct sp_node, nd);
2479 next = rb_next(&n->nd);
2482 mutex_unlock(&p->mutex);
2485 /* assumes fs == KERNEL_DS */
2486 void __init numa_policy_init(void)
2488 nodemask_t interleave_nodes;
2489 unsigned long largest = 0;
2490 int nid, prefer = 0;
2492 policy_cache = kmem_cache_create("numa_policy",
2493 sizeof(struct mempolicy),
2494 0, SLAB_PANIC, NULL);
2496 sn_cache = kmem_cache_create("shared_policy_node",
2497 sizeof(struct sp_node),
2498 0, SLAB_PANIC, NULL);
2500 for_each_node(nid) {
2501 preferred_node_policy[nid] = (struct mempolicy) {
2502 .refcnt = ATOMIC_INIT(1),
2503 .mode = MPOL_PREFERRED,
2504 .flags = MPOL_F_MOF | MPOL_F_HOME,
2505 .v = { .preferred_node = nid, },
2510 * Set interleaving policy for system init. Interleaving is only
2511 * enabled across suitably sized nodes (default is >= 16MB), or
2512 * fall back to the largest node if they're all smaller.
2514 nodes_clear(interleave_nodes);
2515 for_each_node_state(nid, N_HIGH_MEMORY) {
2516 unsigned long total_pages = node_present_pages(nid);
2518 /* Preserve the largest node */
2519 if (largest < total_pages) {
2520 largest = total_pages;
2524 /* Interleave this node? */
2525 if ((total_pages << PAGE_SHIFT) >= (16 << 20))
2526 node_set(nid, interleave_nodes);
2529 /* All too small, use the largest */
2530 if (unlikely(nodes_empty(interleave_nodes)))
2531 node_set(prefer, interleave_nodes);
2533 if (do_set_mempolicy(MPOL_INTERLEAVE, 0, &interleave_nodes))
2534 printk("numa_policy_init: interleaving failed\n");
2537 /* Reset policy of current process to default */
2538 void numa_default_policy(void)
2540 do_set_mempolicy(MPOL_DEFAULT, 0, NULL);
2544 * Parse and format mempolicy from/to strings
2548 * "local" is pseudo-policy: MPOL_PREFERRED with MPOL_F_LOCAL flag
2549 * Used only for mpol_parse_str() and mpol_to_str()
2551 static const char * const policy_modes[] =
2553 [MPOL_DEFAULT] = "default",
2554 [MPOL_PREFERRED] = "prefer",
2555 [MPOL_BIND] = "bind",
2556 [MPOL_INTERLEAVE] = "interleave",
2557 [MPOL_LOCAL] = "local",
2558 [MPOL_NOOP] = "noop", /* should not actually be used */
2564 * mpol_parse_str - parse string to mempolicy
2565 * @str: string containing mempolicy to parse
2566 * @mpol: pointer to struct mempolicy pointer, returned on success.
2567 * @no_context: flag whether to "contextualize" the mempolicy
2570 * <mode>[=<flags>][:<nodelist>]
2572 * if @no_context is true, save the input nodemask in w.user_nodemask in
2573 * the returned mempolicy. This will be used to "clone" the mempolicy in
2574 * a specific context [cpuset] at a later time. Used to parse tmpfs mpol
2575 * mount option. Note that if 'static' or 'relative' mode flags were
2576 * specified, the input nodemask will already have been saved. Saving
2577 * it again is redundant, but safe.
2579 * On success, returns 0, else 1
2581 int mpol_parse_str(char *str, struct mempolicy **mpol, int no_context)
2583 struct mempolicy *new = NULL;
2584 unsigned short mode;
2585 unsigned short uninitialized_var(mode_flags);
2587 char *nodelist = strchr(str, ':');
2588 char *flags = strchr(str, '=');
2592 /* NUL-terminate mode or flags string */
2594 if (nodelist_parse(nodelist, nodes))
2596 if (!nodes_subset(nodes, node_states[N_HIGH_MEMORY]))
2602 *flags++ = '\0'; /* terminate mode string */
2604 for (mode = 0; mode < MPOL_MAX; mode++) {
2605 if (!strcmp(str, policy_modes[mode])) {
2609 if (mode >= MPOL_MAX || mode == MPOL_NOOP)
2613 case MPOL_PREFERRED:
2615 * Insist on a nodelist of one node only
2618 char *rest = nodelist;
2619 while (isdigit(*rest))
2625 case MPOL_INTERLEAVE:
2627 * Default to online nodes with memory if no nodelist
2630 nodes = node_states[N_HIGH_MEMORY];
2634 * Don't allow a nodelist; mpol_new() checks flags
2638 mode = MPOL_PREFERRED;
2642 * Insist on a empty nodelist
2649 * Insist on a nodelist
2658 * Currently, we only support two mutually exclusive
2661 if (!strcmp(flags, "static"))
2662 mode_flags |= MPOL_F_STATIC_NODES;
2663 else if (!strcmp(flags, "relative"))
2664 mode_flags |= MPOL_F_RELATIVE_NODES;
2669 new = mpol_new(mode, mode_flags, &nodes);
2674 /* save for contextualization */
2675 new->w.user_nodemask = nodes;
2678 NODEMASK_SCRATCH(scratch);
2681 ret = mpol_set_nodemask(new, &nodes, scratch);
2682 task_unlock(current);
2685 NODEMASK_SCRATCH_FREE(scratch);
2694 /* Restore string for error message */
2703 #endif /* CONFIG_TMPFS */
2706 * mpol_to_str - format a mempolicy structure for printing
2707 * @buffer: to contain formatted mempolicy string
2708 * @maxlen: length of @buffer
2709 * @pol: pointer to mempolicy to be formatted
2710 * @no_context: "context free" mempolicy - use nodemask in w.user_nodemask
2712 * Convert a mempolicy into a string.
2713 * Returns the number of characters in buffer (if positive)
2714 * or an error (negative)
2716 int mpol_to_str(char *buffer, int maxlen, struct mempolicy *pol, int no_context)
2721 unsigned short mode;
2722 unsigned short flags = pol ? pol->flags : 0;
2725 * Sanity check: room for longest mode, flag and some nodes
2727 VM_BUG_ON(maxlen < strlen("interleave") + strlen("relative") + 16);
2729 if (!pol || pol == &default_policy)
2730 mode = MPOL_DEFAULT;
2739 case MPOL_PREFERRED:
2741 if (flags & MPOL_F_LOCAL)
2742 mode = MPOL_LOCAL; /* pseudo-policy */
2744 node_set(pol->v.preferred_node, nodes);
2749 case MPOL_INTERLEAVE:
2751 nodes = pol->w.user_nodemask;
2753 nodes = pol->v.nodes;
2760 l = strlen(policy_modes[mode]);
2761 if (buffer + maxlen < p + l + 1)
2764 strcpy(p, policy_modes[mode]);
2767 if (flags & MPOL_MODE_FLAGS) {
2768 if (buffer + maxlen < p + 2)
2773 * Currently, the only defined flags are mutually exclusive
2775 if (flags & MPOL_F_STATIC_NODES)
2776 p += snprintf(p, buffer + maxlen - p, "static");
2777 else if (flags & MPOL_F_RELATIVE_NODES)
2778 p += snprintf(p, buffer + maxlen - p, "relative");
2781 if (!nodes_empty(nodes)) {
2782 if (buffer + maxlen < p + 2)
2785 p += nodelist_scnprintf(p, buffer + maxlen - p, nodes);